[Minor] Update zstd to 1.5.4

59 files changed, 15433 insertions, 5643 deletions

diff --git a/contrib/DEPENDENCY_INFO.md b/contrib/DEPENDENCY_INFO.md
index 8220b7444..49eca0992 100644
--- a/contrib/DEPENDENCY_INFO.md
+++ b/contrib/DEPENDENCY_INFO.md
@@ -25,7 +25,7 @@
 | t1ha                   | ?       | Zlib                | NO      |                    |
 | uthash                 | 1.9.8   | BSD                 | YES     |                    |
 | xxhash                 | 0.8.1   | BSD                 | NO      |                    |
-| zstd                   | 1.4.5   | BSD                 | NO      |                    |
+| zstd                   | 1.5.4   | BSD                 | YES     | build fixes only   |
 | google-ced             | 37529e6 | Apache 2            | YES     | build fixes        |
 | kann                   | ?       | MIT                 | YES     | blas/lapack changes|
 | fpconv                 | ?       | Boost               | YES     | many changes       |
diff --git a/contrib/zstd/CMakeLists.txt b/contrib/zstd/CMakeLists.txt
index 4f3f0e516..4601e53ce 100644
--- a/contrib/zstd/CMakeLists.txt
+++ b/contrib/zstd/CMakeLists.txt
@@ -9,7 +9,6 @@ SET(ZSTDSRC
 		huf_compress.c
 		huf_decompress.c
 		pool.c
-		threading.c
 		zstd_common.c
 		zstd_compress.c
 		zstd_compress_literals.c
@@ -25,4 +24,4 @@ SET(ZSTDSRC
 		zstd_opt.c)
 
 ADD_LIBRARY(rspamd-zstd STATIC ${ZSTDSRC})
-ADD_DEFINITIONS(-DZSTD_STATIC_LINKING_ONLY)
+ADD_DEFINITIONS(-DZSTD_DISABLE_ASM)
\ No newline at end of file
diff --git a/contrib/zstd/bits.h b/contrib/zstd/bits.h
new file mode 100644
index 000000000..7939f3d0f
--- /dev/null
+++ b/contrib/zstd/bits.h
@@ -0,0 +1,175 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_BITS_H
+#define ZSTD_BITS_H
+
+#include "mem.h"
+
+MEM_STATIC unsigned ZSTD_countTrailingZeros32_fallback(U32 val)
+{
+    assert(val != 0);
+    {
+        static const int DeBruijnBytePos[32] = {0, 1, 28, 2, 29, 14, 24, 3,
+                                                30, 22, 20, 15, 25, 17, 4, 8,
+                                                31, 27, 13, 23, 21, 19, 16, 7,
+                                                26, 12, 18, 6, 11, 5, 10, 9};
+        return DeBruijnBytePos[((U32) ((val & -(S32) val) * 0x077CB531U)) >> 27];
+    }
+}
+
+MEM_STATIC unsigned ZSTD_countTrailingZeros32(U32 val)
+{
+    assert(val != 0);
+#   if defined(_MSC_VER)
+#       if STATIC_BMI2 == 1
+            return _tzcnt_u32(val);
+#       else
+            if (val != 0) {
+                unsigned long r;
+                _BitScanForward(&r, val);
+                return (unsigned)r;
+            } else {
+                /* Should not reach this code path */
+                __assume(0);
+            }
+#       endif
+#   elif defined(__GNUC__) && (__GNUC__ >= 4)
+        return (unsigned)__builtin_ctz(val);
+#   else
+        return ZSTD_countTrailingZeros32_fallback(val);
+#   endif
+}
+
+MEM_STATIC unsigned ZSTD_countLeadingZeros32_fallback(U32 val) {
+    assert(val != 0);
+    {
+        static const U32 DeBruijnClz[32] = {0, 9, 1, 10, 13, 21, 2, 29,
+                                            11, 14, 16, 18, 22, 25, 3, 30,
+                                            8, 12, 20, 28, 15, 17, 24, 7,
+                                            19, 27, 23, 6, 26, 5, 4, 31};
+        val |= val >> 1;
+        val |= val >> 2;
+        val |= val >> 4;
+        val |= val >> 8;
+        val |= val >> 16;
+        return 31 - DeBruijnClz[(val * 0x07C4ACDDU) >> 27];
+    }
+}
+
+MEM_STATIC unsigned ZSTD_countLeadingZeros32(U32 val)
+{
+    assert(val != 0);
+#   if defined(_MSC_VER)
+#       if STATIC_BMI2 == 1
+            return _lzcnt_u32(val);
+#       else
+            if (val != 0) {
+                unsigned long r;
+                _BitScanReverse(&r, val);
+                return (unsigned)(31 - r);
+            } else {
+                /* Should not reach this code path */
+                __assume(0);
+            }
+#       endif
+#   elif defined(__GNUC__) && (__GNUC__ >= 4)
+        return (unsigned)__builtin_clz(val);
+#   else
+        return ZSTD_countLeadingZeros32_fallback(val);
+#   endif
+}
+
+MEM_STATIC unsigned ZSTD_countTrailingZeros64(U64 val)
+{
+    assert(val != 0);
+#   if defined(_MSC_VER) && defined(_WIN64)
+#       if STATIC_BMI2 == 1
+            return _tzcnt_u64(val);
+#       else
+            if (val != 0) {
+                unsigned long r;
+                _BitScanForward64(&r, val);
+                return (unsigned)r;
+            } else {
+                /* Should not reach this code path */
+                __assume(0);
+            }
+#       endif
+#   elif defined(__GNUC__) && (__GNUC__ >= 4) && defined(__LP64__)
+        return (unsigned)__builtin_ctzll(val);
+#   else
+        {
+            U32 mostSignificantWord = (U32)(val >> 32);
+            U32 leastSignificantWord = (U32)val;
+            if (leastSignificantWord == 0) {
+                return 32 + ZSTD_countTrailingZeros32(mostSignificantWord);
+            } else {
+                return ZSTD_countTrailingZeros32(leastSignificantWord);
+            }
+        }
+#   endif
+}
+
+MEM_STATIC unsigned ZSTD_countLeadingZeros64(U64 val)
+{
+    assert(val != 0);
+#   if defined(_MSC_VER) && defined(_WIN64)
+#       if STATIC_BMI2 == 1
+            return _lzcnt_u64(val);
+#       else
+            if (val != 0) {
+                unsigned long r;
+                _BitScanReverse64(&r, val);
+                return (unsigned)(63 - r);
+            } else {
+                /* Should not reach this code path */
+                __assume(0);
+            }
+#       endif
+#   elif defined(__GNUC__) && (__GNUC__ >= 4)
+        return (unsigned)(__builtin_clzll(val));
+#   else
+        {
+            U32 mostSignificantWord = (U32)(val >> 32);
+            U32 leastSignificantWord = (U32)val;
+            if (mostSignificantWord == 0) {
+                return 32 + ZSTD_countLeadingZeros32(leastSignificantWord);
+            } else {
+                return ZSTD_countLeadingZeros32(mostSignificantWord);
+            }
+        }
+#   endif
+}
+
+MEM_STATIC unsigned ZSTD_NbCommonBytes(size_t val)
+{
+    if (MEM_isLittleEndian()) {
+        if (MEM_64bits()) {
+            return ZSTD_countTrailingZeros64((U64)val) >> 3;
+        } else {
+            return ZSTD_countTrailingZeros32((U32)val) >> 3;
+        }
+    } else {  /* Big Endian CPU */
+        if (MEM_64bits()) {
+            return ZSTD_countLeadingZeros64((U64)val) >> 3;
+        } else {
+            return ZSTD_countLeadingZeros32((U32)val) >> 3;
+        }
+    }
+}
+
+MEM_STATIC unsigned ZSTD_highbit32(U32 val)   /* compress, dictBuilder, decodeCorpus */
+{
+    assert(val != 0);
+    return 31 - ZSTD_countLeadingZeros32(val);
+}
+
+#endif /* ZSTD_BITS_H */
diff --git a/contrib/zstd/bitstream.h b/contrib/zstd/bitstream.h
index 37b99c01e..db1b4cf13 100644
--- a/contrib/zstd/bitstream.h
+++ b/contrib/zstd/bitstream.h
@@ -1,7 +1,7 @@
 /* ******************************************************************
  * bitstream
  * Part of FSE library
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -17,7 +17,6 @@
 #if defined (__cplusplus)
 extern "C" {
 #endif
-
 /*
 *  This API consists of small unitary functions, which must be inlined for best performance.
 *  Since link-time-optimization is not available for all compilers,
@@ -31,15 +30,18 @@ extern "C" {
 #include "compiler.h"       /* UNLIKELY() */
 #include "debug.h"          /* assert(), DEBUGLOG(), RAWLOG() */
 #include "error_private.h"  /* error codes and messages */
+#include "bits.h"           /* ZSTD_highbit32 */
 
 
 /*=========================================
 *  Target specific
 =========================================*/
-#if defined(__BMI__) && defined(__GNUC__)
-#  include <immintrin.h>   /* support for bextr (experimental) */
-#elif defined(__ICCARM__)
-#  include <intrinsics.h>
+#ifndef ZSTD_NO_INTRINSICS
+#  if (defined(__BMI__) || defined(__BMI2__)) && defined(__GNUC__)
+#    include <immintrin.h>   /* support for bextr (experimental)/bzhi */
+#  elif defined(__ICCARM__)
+#    include <intrinsics.h>
+#  endif
 #endif
 
 #define STREAM_ACCUMULATOR_MIN_32  25
@@ -131,38 +133,6 @@ MEM_STATIC void BIT_flushBitsFast(BIT_CStream_t* bitC);
 MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, unsigned nbBits);
 /* faster, but works only if nbBits >= 1 */
 
-
-
-/*-**************************************************************
-*  Internal functions
-****************************************************************/
-MEM_STATIC unsigned BIT_highbit32 (U32 val)
-{
-    assert(val != 0);
-    {
-#   if defined(_MSC_VER)   /* Visual */
-        unsigned long r=0;
-        return _BitScanReverse ( &r, val ) ? (unsigned)r : 0;
-#   elif defined(__GNUC__) && (__GNUC__ >= 3)   /* Use GCC Intrinsic */
-        return __builtin_clz (val) ^ 31;
-#   elif defined(__ICCARM__)    /* IAR Intrinsic */
-        return 31 - __CLZ(val);
-#   else   /* Software version */
-        static const unsigned DeBruijnClz[32] = { 0,  9,  1, 10, 13, 21,  2, 29,
-                                                 11, 14, 16, 18, 22, 25,  3, 30,
-                                                  8, 12, 20, 28, 15, 17, 24,  7,
-                                                 19, 27, 23,  6, 26,  5,  4, 31 };
-        U32 v = val;
-        v |= v >> 1;
-        v |= v >> 2;
-        v |= v >> 4;
-        v |= v >> 8;
-        v |= v >> 16;
-        return DeBruijnClz[ (U32) (v * 0x07C4ACDDU) >> 27];
-#   endif
-    }
-}
-
 /*=====    Local Constants   =====*/
 static const unsigned BIT_mask[] = {
     0,          1,         3,         7,         0xF,       0x1F,
@@ -192,16 +162,26 @@ MEM_STATIC size_t BIT_initCStream(BIT_CStream_t* bitC,
     return 0;
 }
 
+MEM_STATIC FORCE_INLINE_ATTR size_t BIT_getLowerBits(size_t bitContainer, U32 const nbBits)
+{
+#if defined(STATIC_BMI2) && STATIC_BMI2 == 1 && !defined(ZSTD_NO_INTRINSICS)
+    return  _bzhi_u64(bitContainer, nbBits);
+#else
+    assert(nbBits < BIT_MASK_SIZE);
+    return bitContainer & BIT_mask[nbBits];
+#endif
+}
+
 /*! BIT_addBits() :
  *  can add up to 31 bits into `bitC`.
  *  Note : does not check for register overflow ! */
 MEM_STATIC void BIT_addBits(BIT_CStream_t* bitC,
                             size_t value, unsigned nbBits)
 {
-    MEM_STATIC_ASSERT(BIT_MASK_SIZE == 32);
+    DEBUG_STATIC_ASSERT(BIT_MASK_SIZE == 32);
     assert(nbBits < BIT_MASK_SIZE);
     assert(nbBits + bitC->bitPos < sizeof(bitC->bitContainer) * 8);
-    bitC->bitContainer |= (value & BIT_mask[nbBits]) << bitC->bitPos;
+    bitC->bitContainer |= BIT_getLowerBits(value, nbBits) << bitC->bitPos;
     bitC->bitPos += nbBits;
 }
 
@@ -271,7 +251,7 @@ MEM_STATIC size_t BIT_closeCStream(BIT_CStream_t* bitC)
  */
 MEM_STATIC size_t BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, size_t srcSize)
 {
-    if (srcSize < 1) { memset(bitD, 0, sizeof(*bitD)); return ERROR(srcSize_wrong); }
+    if (srcSize < 1) { ZSTD_memset(bitD, 0, sizeof(*bitD)); return ERROR(srcSize_wrong); }
 
     bitD->start = (const char*)srcBuffer;
     bitD->limitPtr = bitD->start + sizeof(bitD->bitContainer);
@@ -280,7 +260,7 @@ MEM_STATIC size_t BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, si
         bitD->ptr   = (const char*)srcBuffer + srcSize - sizeof(bitD->bitContainer);
         bitD->bitContainer = MEM_readLEST(bitD->ptr);
         { BYTE const lastByte = ((const BYTE*)srcBuffer)[srcSize-1];
-          bitD->bitsConsumed = lastByte ? 8 - BIT_highbit32(lastByte) : 0;  /* ensures bitsConsumed is always set */
+          bitD->bitsConsumed = lastByte ? 8 - ZSTD_highbit32(lastByte) : 0;  /* ensures bitsConsumed is always set */
           if (lastByte == 0) return ERROR(GENERIC); /* endMark not present */ }
     } else {
         bitD->ptr   = bitD->start;
@@ -288,27 +268,27 @@ MEM_STATIC size_t BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, si
         switch(srcSize)
         {
         case 7: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[6]) << (sizeof(bitD->bitContainer)*8 - 16);
-                /* fall-through */
+                ZSTD_FALLTHROUGH;
 
         case 6: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[5]) << (sizeof(bitD->bitContainer)*8 - 24);
-                /* fall-through */
+                ZSTD_FALLTHROUGH;
 
         case 5: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[4]) << (sizeof(bitD->bitContainer)*8 - 32);
-                /* fall-through */
+                ZSTD_FALLTHROUGH;
 
         case 4: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[3]) << 24;
-                /* fall-through */
+                ZSTD_FALLTHROUGH;
 
         case 3: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[2]) << 16;
-                /* fall-through */
+                ZSTD_FALLTHROUGH;
 
         case 2: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[1]) <<  8;
-                /* fall-through */
+                ZSTD_FALLTHROUGH;
 
         default: break;
         }
         {   BYTE const lastByte = ((const BYTE*)srcBuffer)[srcSize-1];
-            bitD->bitsConsumed = lastByte ? 8 - BIT_highbit32(lastByte) : 0;
+            bitD->bitsConsumed = lastByte ? 8 - ZSTD_highbit32(lastByte) : 0;
             if (lastByte == 0) return ERROR(corruption_detected);  /* endMark not present */
         }
         bitD->bitsConsumed += (U32)(sizeof(bitD->bitContainer) - srcSize)*8;
@@ -317,23 +297,26 @@ MEM_STATIC size_t BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, si
     return srcSize;
 }
 
-MEM_STATIC size_t BIT_getUpperBits(size_t bitContainer, U32 const start)
+MEM_STATIC FORCE_INLINE_ATTR size_t BIT_getUpperBits(size_t bitContainer, U32 const start)
 {
     return bitContainer >> start;
 }
 
-MEM_STATIC size_t BIT_getMiddleBits(size_t bitContainer, U32 const start, U32 const nbBits)
+MEM_STATIC FORCE_INLINE_ATTR size_t BIT_getMiddleBits(size_t bitContainer, U32 const start, U32 const nbBits)
 {
     U32 const regMask = sizeof(bitContainer)*8 - 1;
     /* if start > regMask, bitstream is corrupted, and result is undefined */
     assert(nbBits < BIT_MASK_SIZE);
+    /* x86 transform & ((1 << nbBits) - 1) to bzhi instruction, it is better
+     * than accessing memory. When bmi2 instruction is not present, we consider
+     * such cpus old (pre-Haswell, 2013) and their performance is not of that
+     * importance.
+     */
+#if defined(__x86_64__) || defined(_M_X86)
+    return (bitContainer >> (start & regMask)) & ((((U64)1) << nbBits) - 1);
+#else
     return (bitContainer >> (start & regMask)) & BIT_mask[nbBits];
-}
-
-MEM_STATIC size_t BIT_getLowerBits(size_t bitContainer, U32 const nbBits)
-{
-    assert(nbBits < BIT_MASK_SIZE);
-    return bitContainer & BIT_mask[nbBits];
+#endif
 }
 
 /*! BIT_lookBits() :
@@ -342,7 +325,7 @@ MEM_STATIC size_t BIT_getLowerBits(size_t bitContainer, U32 const nbBits)
  *  On 32-bits, maxNbBits==24.
  *  On 64-bits, maxNbBits==56.
  * @return : value extracted */
-MEM_STATIC size_t BIT_lookBits(const BIT_DStream_t* bitD, U32 nbBits)
+MEM_STATIC  FORCE_INLINE_ATTR size_t BIT_lookBits(const BIT_DStream_t*  bitD, U32 nbBits)
 {
     /* arbitrate between double-shift and shift+mask */
 #if 1
@@ -365,7 +348,7 @@ MEM_STATIC size_t BIT_lookBitsFast(const BIT_DStream_t* bitD, U32 nbBits)
     return (bitD->bitContainer << (bitD->bitsConsumed & regMask)) >> (((regMask+1)-nbBits) & regMask);
 }
 
-MEM_STATIC void BIT_skipBits(BIT_DStream_t* bitD, U32 nbBits)
+MEM_STATIC FORCE_INLINE_ATTR void BIT_skipBits(BIT_DStream_t* bitD, U32 nbBits)
 {
     bitD->bitsConsumed += nbBits;
 }
@@ -374,7 +357,7 @@ MEM_STATIC void BIT_skipBits(BIT_DStream_t* bitD, U32 nbBits)
  *  Read (consume) next n bits from local register and update.
  *  Pay attention to not read more than nbBits contained into local register.
  * @return : extracted value. */
-MEM_STATIC size_t BIT_readBits(BIT_DStream_t* bitD, unsigned nbBits)
+MEM_STATIC FORCE_INLINE_ATTR size_t BIT_readBits(BIT_DStream_t* bitD, unsigned nbBits)
 {
     size_t const value = BIT_lookBits(bitD, nbBits);
     BIT_skipBits(bitD, nbBits);
@@ -382,7 +365,7 @@ MEM_STATIC size_t BIT_readBits(BIT_DStream_t* bitD, unsigned nbBits)
 }
 
 /*! BIT_readBitsFast() :
- *  unsafe version; only works only if nbBits >= 1 */
+ *  unsafe version; only works if nbBits >= 1 */
 MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, unsigned nbBits)
 {
     size_t const value = BIT_lookBitsFast(bitD, nbBits);
diff --git a/contrib/zstd/clevels.h b/contrib/zstd/clevels.h
new file mode 100644
index 000000000..650a34f16
--- /dev/null
+++ b/contrib/zstd/clevels.h
@@ -0,0 +1,134 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_CLEVELS_H
+#define ZSTD_CLEVELS_H
+
+#define ZSTD_STATIC_LINKING_ONLY  /* ZSTD_compressionParameters  */
+#include "zstd.h"
+
+/*-=====  Pre-defined compression levels  =====-*/
+
+#define ZSTD_MAX_CLEVEL     22
+
+#ifdef __GNUC__
+__attribute__((__unused__))
+#endif
+
+static const ZSTD_compressionParameters ZSTD_defaultCParameters[4][ZSTD_MAX_CLEVEL+1] = {
+{   /* "default" - for any srcSize > 256 KB */
+    /* W,  C,  H,  S,  L, TL, strat */
+    { 19, 12, 13,  1,  6,  1, ZSTD_fast    },  /* base for negative levels */
+    { 19, 13, 14,  1,  7,  0, ZSTD_fast    },  /* level  1 */
+    { 20, 15, 16,  1,  6,  0, ZSTD_fast    },  /* level  2 */
+    { 21, 16, 17,  1,  5,  0, ZSTD_dfast   },  /* level  3 */
+    { 21, 18, 18,  1,  5,  0, ZSTD_dfast   },  /* level  4 */
+    { 21, 18, 19,  3,  5,  2, ZSTD_greedy  },  /* level  5 */
+    { 21, 18, 19,  3,  5,  4, ZSTD_lazy    },  /* level  6 */
+    { 21, 19, 20,  4,  5,  8, ZSTD_lazy    },  /* level  7 */
+    { 21, 19, 20,  4,  5, 16, ZSTD_lazy2   },  /* level  8 */
+    { 22, 20, 21,  4,  5, 16, ZSTD_lazy2   },  /* level  9 */
+    { 22, 21, 22,  5,  5, 16, ZSTD_lazy2   },  /* level 10 */
+    { 22, 21, 22,  6,  5, 16, ZSTD_lazy2   },  /* level 11 */
+    { 22, 22, 23,  6,  5, 32, ZSTD_lazy2   },  /* level 12 */
+    { 22, 22, 22,  4,  5, 32, ZSTD_btlazy2 },  /* level 13 */
+    { 22, 22, 23,  5,  5, 32, ZSTD_btlazy2 },  /* level 14 */
+    { 22, 23, 23,  6,  5, 32, ZSTD_btlazy2 },  /* level 15 */
+    { 22, 22, 22,  5,  5, 48, ZSTD_btopt   },  /* level 16 */
+    { 23, 23, 22,  5,  4, 64, ZSTD_btopt   },  /* level 17 */
+    { 23, 23, 22,  6,  3, 64, ZSTD_btultra },  /* level 18 */
+    { 23, 24, 22,  7,  3,256, ZSTD_btultra2},  /* level 19 */
+    { 25, 25, 23,  7,  3,256, ZSTD_btultra2},  /* level 20 */
+    { 26, 26, 24,  7,  3,512, ZSTD_btultra2},  /* level 21 */
+    { 27, 27, 25,  9,  3,999, ZSTD_btultra2},  /* level 22 */
+},
+{   /* for srcSize <= 256 KB */
+    /* W,  C,  H,  S,  L,  T, strat */
+    { 18, 12, 13,  1,  5,  1, ZSTD_fast    },  /* base for negative levels */
+    { 18, 13, 14,  1,  6,  0, ZSTD_fast    },  /* level  1 */
+    { 18, 14, 14,  1,  5,  0, ZSTD_dfast   },  /* level  2 */
+    { 18, 16, 16,  1,  4,  0, ZSTD_dfast   },  /* level  3 */
+    { 18, 16, 17,  3,  5,  2, ZSTD_greedy  },  /* level  4.*/
+    { 18, 17, 18,  5,  5,  2, ZSTD_greedy  },  /* level  5.*/
+    { 18, 18, 19,  3,  5,  4, ZSTD_lazy    },  /* level  6.*/
+    { 18, 18, 19,  4,  4,  4, ZSTD_lazy    },  /* level  7 */
+    { 18, 18, 19,  4,  4,  8, ZSTD_lazy2   },  /* level  8 */
+    { 18, 18, 19,  5,  4,  8, ZSTD_lazy2   },  /* level  9 */
+    { 18, 18, 19,  6,  4,  8, ZSTD_lazy2   },  /* level 10 */
+    { 18, 18, 19,  5,  4, 12, ZSTD_btlazy2 },  /* level 11.*/
+    { 18, 19, 19,  7,  4, 12, ZSTD_btlazy2 },  /* level 12.*/
+    { 18, 18, 19,  4,  4, 16, ZSTD_btopt   },  /* level 13 */
+    { 18, 18, 19,  4,  3, 32, ZSTD_btopt   },  /* level 14.*/
+    { 18, 18, 19,  6,  3,128, ZSTD_btopt   },  /* level 15.*/
+    { 18, 19, 19,  6,  3,128, ZSTD_btultra },  /* level 16.*/
+    { 18, 19, 19,  8,  3,256, ZSTD_btultra },  /* level 17.*/
+    { 18, 19, 19,  6,  3,128, ZSTD_btultra2},  /* level 18.*/
+    { 18, 19, 19,  8,  3,256, ZSTD_btultra2},  /* level 19.*/
+    { 18, 19, 19, 10,  3,512, ZSTD_btultra2},  /* level 20.*/
+    { 18, 19, 19, 12,  3,512, ZSTD_btultra2},  /* level 21.*/
+    { 18, 19, 19, 13,  3,999, ZSTD_btultra2},  /* level 22.*/
+},
+{   /* for srcSize <= 128 KB */
+    /* W,  C,  H,  S,  L,  T, strat */
+    { 17, 12, 12,  1,  5,  1, ZSTD_fast    },  /* base for negative levels */
+    { 17, 12, 13,  1,  6,  0, ZSTD_fast    },  /* level  1 */
+    { 17, 13, 15,  1,  5,  0, ZSTD_fast    },  /* level  2 */
+    { 17, 15, 16,  2,  5,  0, ZSTD_dfast   },  /* level  3 */
+    { 17, 17, 17,  2,  4,  0, ZSTD_dfast   },  /* level  4 */
+    { 17, 16, 17,  3,  4,  2, ZSTD_greedy  },  /* level  5 */
+    { 17, 16, 17,  3,  4,  4, ZSTD_lazy    },  /* level  6 */
+    { 17, 16, 17,  3,  4,  8, ZSTD_lazy2   },  /* level  7 */
+    { 17, 16, 17,  4,  4,  8, ZSTD_lazy2   },  /* level  8 */
+    { 17, 16, 17,  5,  4,  8, ZSTD_lazy2   },  /* level  9 */
+    { 17, 16, 17,  6,  4,  8, ZSTD_lazy2   },  /* level 10 */
+    { 17, 17, 17,  5,  4,  8, ZSTD_btlazy2 },  /* level 11 */
+    { 17, 18, 17,  7,  4, 12, ZSTD_btlazy2 },  /* level 12 */
+    { 17, 18, 17,  3,  4, 12, ZSTD_btopt   },  /* level 13.*/
+    { 17, 18, 17,  4,  3, 32, ZSTD_btopt   },  /* level 14.*/
+    { 17, 18, 17,  6,  3,256, ZSTD_btopt   },  /* level 15.*/
+    { 17, 18, 17,  6,  3,128, ZSTD_btultra },  /* level 16.*/
+    { 17, 18, 17,  8,  3,256, ZSTD_btultra },  /* level 17.*/
+    { 17, 18, 17, 10,  3,512, ZSTD_btultra },  /* level 18.*/
+    { 17, 18, 17,  5,  3,256, ZSTD_btultra2},  /* level 19.*/
+    { 17, 18, 17,  7,  3,512, ZSTD_btultra2},  /* level 20.*/
+    { 17, 18, 17,  9,  3,512, ZSTD_btultra2},  /* level 21.*/
+    { 17, 18, 17, 11,  3,999, ZSTD_btultra2},  /* level 22.*/
+},
+{   /* for srcSize <= 16 KB */
+    /* W,  C,  H,  S,  L,  T, strat */
+    { 14, 12, 13,  1,  5,  1, ZSTD_fast    },  /* base for negative levels */
+    { 14, 14, 15,  1,  5,  0, ZSTD_fast    },  /* level  1 */
+    { 14, 14, 15,  1,  4,  0, ZSTD_fast    },  /* level  2 */
+    { 14, 14, 15,  2,  4,  0, ZSTD_dfast   },  /* level  3 */
+    { 14, 14, 14,  4,  4,  2, ZSTD_greedy  },  /* level  4 */
+    { 14, 14, 14,  3,  4,  4, ZSTD_lazy    },  /* level  5.*/
+    { 14, 14, 14,  4,  4,  8, ZSTD_lazy2   },  /* level  6 */
+    { 14, 14, 14,  6,  4,  8, ZSTD_lazy2   },  /* level  7 */
+    { 14, 14, 14,  8,  4,  8, ZSTD_lazy2   },  /* level  8.*/
+    { 14, 15, 14,  5,  4,  8, ZSTD_btlazy2 },  /* level  9.*/
+    { 14, 15, 14,  9,  4,  8, ZSTD_btlazy2 },  /* level 10.*/
+    { 14, 15, 14,  3,  4, 12, ZSTD_btopt   },  /* level 11.*/
+    { 14, 15, 14,  4,  3, 24, ZSTD_btopt   },  /* level 12.*/
+    { 14, 15, 14,  5,  3, 32, ZSTD_btultra },  /* level 13.*/
+    { 14, 15, 15,  6,  3, 64, ZSTD_btultra },  /* level 14.*/
+    { 14, 15, 15,  7,  3,256, ZSTD_btultra },  /* level 15.*/
+    { 14, 15, 15,  5,  3, 48, ZSTD_btultra2},  /* level 16.*/
+    { 14, 15, 15,  6,  3,128, ZSTD_btultra2},  /* level 17.*/
+    { 14, 15, 15,  7,  3,256, ZSTD_btultra2},  /* level 18.*/
+    { 14, 15, 15,  8,  3,256, ZSTD_btultra2},  /* level 19.*/
+    { 14, 15, 15,  8,  3,512, ZSTD_btultra2},  /* level 20.*/
+    { 14, 15, 15,  9,  3,512, ZSTD_btultra2},  /* level 21.*/
+    { 14, 15, 15, 10,  3,999, ZSTD_btultra2},  /* level 22.*/
+},
+};
+
+
+
+#endif  /* ZSTD_CLEVELS_H */
diff --git a/contrib/zstd/compiler.h b/contrib/zstd/compiler.h
index 95e948352..d4f2f285d 100644
--- a/contrib/zstd/compiler.h
+++ b/contrib/zstd/compiler.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -11,6 +11,8 @@
 #ifndef ZSTD_COMPILER_H
 #define ZSTD_COMPILER_H
 
+#include "portability_macros.h"
+
 /*-*******************************************************
 *  Compiler specifics
 *********************************************************/
@@ -39,6 +41,17 @@
 #endif
 
 /**
+  On MSVC qsort requires that functions passed into it use the __cdecl calling conversion(CC).
+  This explicitly marks such functions as __cdecl so that the code will still compile
+  if a CC other than __cdecl has been made the default.
+*/
+#if  defined(_MSC_VER)
+#  define WIN_CDECL __cdecl
+#else
+#  define WIN_CDECL
+#endif
+
+/**
  * FORCE_INLINE_TEMPLATE is used to define C "templates", which take constant
  * parameters. They must be inlined for the compiler to eliminate the constant
  * branches.
@@ -79,30 +92,19 @@
 #  endif
 #endif
 
+
 /* target attribute */
-#ifndef __has_attribute
-  #define __has_attribute(x) 0  /* Compatibility with non-clang compilers. */
-#endif
 #if defined(__GNUC__) || defined(__ICCARM__)
 #  define TARGET_ATTRIBUTE(target) __attribute__((__target__(target)))
 #else
 #  define TARGET_ATTRIBUTE(target)
 #endif
 
-/* Enable runtime BMI2 dispatch based on the CPU.
- * Enabled for clang & gcc >=4.8 on x86 when BMI2 isn't enabled by default.
+/* Target attribute for BMI2 dynamic dispatch.
+ * Enable lzcnt, bmi, and bmi2.
+ * We test for bmi1 & bmi2. lzcnt is included in bmi1.
  */
-#ifndef DYNAMIC_BMI2
-  #if ((defined(__clang__) && __has_attribute(__target__)) \
-      || (defined(__GNUC__) \
-          && (__GNUC__ >= 5 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8)))) \
-      && (defined(__x86_64__) || defined(_M_X86)) \
-      && !defined(__BMI2__)
-  #  define DYNAMIC_BMI2 1
-  #else
-  #  define DYNAMIC_BMI2 0
-  #endif
-#endif
+#define BMI2_TARGET_ATTRIBUTE TARGET_ATTRIBUTE("lzcnt,bmi,bmi2")
 
 /* prefetch
  * can be disabled, by declaring NO_PREFETCH build macro */
@@ -114,12 +116,12 @@
 #    include <mmintrin.h>   /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
 #    define PREFETCH_L1(ptr)  _mm_prefetch((const char*)(ptr), _MM_HINT_T0)
 #    define PREFETCH_L2(ptr)  _mm_prefetch((const char*)(ptr), _MM_HINT_T1)
-#    elif defined(__aarch64__)
-#     define PREFETCH_L1(ptr)  __asm__ __volatile__("prfm pldl1keep, %0" ::"Q"(*(ptr)))
-#     define PREFETCH_L2(ptr)  __asm__ __volatile__("prfm pldl2keep, %0" ::"Q"(*(ptr)))
 #  elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) )
 #    define PREFETCH_L1(ptr)  __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */)
 #    define PREFETCH_L2(ptr)  __builtin_prefetch((ptr), 0 /* rw==read */, 2 /* locality */)
+#  elif defined(__aarch64__)
+#    define PREFETCH_L1(ptr)  __asm__ __volatile__("prfm pldl1keep, %0" ::"Q"(*(ptr)))
+#    define PREFETCH_L2(ptr)  __asm__ __volatile__("prfm pldl2keep, %0" ::"Q"(*(ptr)))
 #  else
 #    define PREFETCH_L1(ptr) (void)(ptr)  /* disabled */
 #    define PREFETCH_L2(ptr) (void)(ptr)  /* disabled */
@@ -138,8 +140,9 @@
 }
 
 /* vectorization
- * older GCC (pre gcc-4.3 picked as the cutoff) uses a different syntax */
-#if !defined(__INTEL_COMPILER) && !defined(__clang__) && defined(__GNUC__)
+ * older GCC (pre gcc-4.3 picked as the cutoff) uses a different syntax,
+ * and some compilers, like Intel ICC and MCST LCC, do not support it at all. */
+#if !defined(__INTEL_COMPILER) && !defined(__clang__) && defined(__GNUC__) && !defined(__LCC__)
 #  if (__GNUC__ == 4 && __GNUC_MINOR__ > 3) || (__GNUC__ >= 5)
 #    define DONT_VECTORIZE __attribute__((optimize("no-tree-vectorize")))
 #  else
@@ -162,6 +165,12 @@
 #define UNLIKELY(x) (x)
 #endif
 
+#if __has_builtin(__builtin_unreachable) || (defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 5)))
+#  define ZSTD_UNREACHABLE { assert(0), __builtin_unreachable(); }
+#else
+#  define ZSTD_UNREACHABLE { assert(0); }
+#endif
+
 /* disable warnings */
 #ifdef _MSC_VER    /* Visual Studio */
 #  include <intrin.h>                    /* For Visual 2005 */
@@ -172,4 +181,174 @@
 #  pragma warning(disable : 4324)        /* disable: C4324: padded structure */
 #endif
 
+/*Like DYNAMIC_BMI2 but for compile time determination of BMI2 support*/
+#ifndef STATIC_BMI2
+#  if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_I86))
+#    ifdef __AVX2__  //MSVC does not have a BMI2 specific flag, but every CPU that supports AVX2 also supports BMI2
+#       define STATIC_BMI2 1
+#    endif
+#  elif defined(__BMI2__) && defined(__x86_64__) && defined(__GNUC__)
+#    define STATIC_BMI2 1
+#  endif
+#endif
+
+#ifndef STATIC_BMI2
+    #define STATIC_BMI2 0
+#endif
+
+/* compile time determination of SIMD support */
+#if !defined(ZSTD_NO_INTRINSICS)
+#  if defined(__SSE2__) || defined(_M_AMD64) || (defined (_M_IX86) && defined(_M_IX86_FP) && (_M_IX86_FP >= 2))
+#    define ZSTD_ARCH_X86_SSE2
+#  endif
+#  if defined(__ARM_NEON) || defined(_M_ARM64)
+#    define ZSTD_ARCH_ARM_NEON
+#  endif
+#
+#  if defined(ZSTD_ARCH_X86_SSE2)
+#    include <emmintrin.h>
+#  elif defined(ZSTD_ARCH_ARM_NEON)
+#    include <arm_neon.h>
+#  endif
+#endif
+
+/* C-language Attributes are added in C23. */
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ > 201710L) && defined(__has_c_attribute)
+# define ZSTD_HAS_C_ATTRIBUTE(x) __has_c_attribute(x)
+#else
+# define ZSTD_HAS_C_ATTRIBUTE(x) 0
+#endif
+
+/* Only use C++ attributes in C++. Some compilers report support for C++
+ * attributes when compiling with C.
+ */
+#if defined(__cplusplus) && defined(__has_cpp_attribute)
+# define ZSTD_HAS_CPP_ATTRIBUTE(x) __has_cpp_attribute(x)
+#else
+# define ZSTD_HAS_CPP_ATTRIBUTE(x) 0
+#endif
+
+/* Define ZSTD_FALLTHROUGH macro for annotating switch case with the 'fallthrough' attribute.
+ * - C23: https://en.cppreference.com/w/c/language/attributes/fallthrough
+ * - CPP17: https://en.cppreference.com/w/cpp/language/attributes/fallthrough
+ * - Else: __attribute__((__fallthrough__))
+ */
+#ifndef ZSTD_FALLTHROUGH
+# if ZSTD_HAS_C_ATTRIBUTE(fallthrough)
+#  define ZSTD_FALLTHROUGH [[fallthrough]]
+# elif ZSTD_HAS_CPP_ATTRIBUTE(fallthrough)
+#  define ZSTD_FALLTHROUGH [[fallthrough]]
+# elif __has_attribute(__fallthrough__)
+/* Leading semicolon is to satisfy gcc-11 with -pedantic. Without the semicolon
+ * gcc complains about: a label can only be part of a statement and a declaration is not a statement.
+ */
+#  define ZSTD_FALLTHROUGH ; __attribute__((__fallthrough__))
+# else
+#  define ZSTD_FALLTHROUGH
+# endif
+#endif
+
+/*-**************************************************************
+*  Alignment check
+*****************************************************************/
+
+/* this test was initially positioned in mem.h,
+ * but this file is removed (or replaced) for linux kernel
+ * so it's now hosted in compiler.h,
+ * which remains valid for both user & kernel spaces.
+ */
+
+#ifndef ZSTD_ALIGNOF
+# if defined(__GNUC__) || defined(_MSC_VER)
+/* covers gcc, clang & MSVC */
+/* note : this section must come first, before C11,
+ * due to a limitation in the kernel source generator */
+#  define ZSTD_ALIGNOF(T) __alignof(T)
+
+# elif defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)
+/* C11 support */
+#  include <stdalign.h>
+#  define ZSTD_ALIGNOF(T) alignof(T)
+
+# else
+/* No known support for alignof() - imperfect backup */
+#  define ZSTD_ALIGNOF(T) (sizeof(void*) < sizeof(T) ? sizeof(void*) : sizeof(T))
+
+# endif
+#endif /* ZSTD_ALIGNOF */
+
+/*-**************************************************************
+*  Sanitizer
+*****************************************************************/
+
+/* Issue #3240 reports an ASAN failure on an llvm-mingw build. Out of an
+ * abundance of caution, disable our custom poisoning on mingw. */
+#ifdef __MINGW32__
+#ifndef ZSTD_ASAN_DONT_POISON_WORKSPACE
+#define ZSTD_ASAN_DONT_POISON_WORKSPACE 1
+#endif
+#ifndef ZSTD_MSAN_DONT_POISON_WORKSPACE
+#define ZSTD_MSAN_DONT_POISON_WORKSPACE 1
+#endif
+#endif
+
+#if ZSTD_MEMORY_SANITIZER && !defined(ZSTD_MSAN_DONT_POISON_WORKSPACE)
+/* Not all platforms that support msan provide sanitizers/msan_interface.h.
+ * We therefore declare the functions we need ourselves, rather than trying to
+ * include the header file... */
+#include <stddef.h>  /* size_t */
+#define ZSTD_DEPS_NEED_STDINT
+#include "zstd_deps.h"  /* intptr_t */
+
+/* Make memory region fully initialized (without changing its contents). */
+void __msan_unpoison(const volatile void *a, size_t size);
+
+/* Make memory region fully uninitialized (without changing its contents).
+   This is a legacy interface that does not update origin information. Use
+   __msan_allocated_memory() instead. */
+void __msan_poison(const volatile void *a, size_t size);
+
+/* Returns the offset of the first (at least partially) poisoned byte in the
+   memory range, or -1 if the whole range is good. */
+intptr_t __msan_test_shadow(const volatile void *x, size_t size);
+#endif
+
+#if ZSTD_ADDRESS_SANITIZER && !defined(ZSTD_ASAN_DONT_POISON_WORKSPACE)
+/* Not all platforms that support asan provide sanitizers/asan_interface.h.
+ * We therefore declare the functions we need ourselves, rather than trying to
+ * include the header file... */
+#include <stddef.h>  /* size_t */
+
+/**
+ * Marks a memory region (<c>[addr, addr+size)</c>) as unaddressable.
+ *
+ * This memory must be previously allocated by your program. Instrumented
+ * code is forbidden from accessing addresses in this region until it is
+ * unpoisoned. This function is not guaranteed to poison the entire region -
+ * it could poison only a subregion of <c>[addr, addr+size)</c> due to ASan
+ * alignment restrictions.
+ *
+ * \note This function is not thread-safe because no two threads can poison or
+ * unpoison memory in the same memory region simultaneously.
+ *
+ * \param addr Start of memory region.
+ * \param size Size of memory region. */
+void __asan_poison_memory_region(void const volatile *addr, size_t size);
+
+/**
+ * Marks a memory region (<c>[addr, addr+size)</c>) as addressable.
+ *
+ * This memory must be previously allocated by your program. Accessing
+ * addresses in this region is allowed until this region is poisoned again.
+ * This function could unpoison a super-region of <c>[addr, addr+size)</c> due
+ * to ASan alignment restrictions.
+ *
+ * \note This function is not thread-safe because no two threads can
+ * poison or unpoison memory in the same memory region simultaneously.
+ *
+ * \param addr Start of memory region.
+ * \param size Size of memory region. */
+void __asan_unpoison_memory_region(void const volatile *addr, size_t size);
+#endif
+
 #endif /* ZSTD_COMPILER_H */
diff --git a/contrib/zstd/cpu.h b/contrib/zstd/cpu.h
index 6e8a974f6..8bc34a36d 100644
--- a/contrib/zstd/cpu.h
+++ b/contrib/zstd/cpu.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2020, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -16,8 +16,6 @@
  * https://github.com/facebook/folly/blob/master/folly/CpuId.h
  */
 
-#include <string.h>
-
 #include "mem.h"
 
 #ifdef _MSC_VER
diff --git a/contrib/zstd/debug.c b/contrib/zstd/debug.c
index f303f4a2e..ebf7bfccf 100644
--- a/contrib/zstd/debug.c
+++ b/contrib/zstd/debug.c
@@ -1,7 +1,7 @@
 /* ******************************************************************
  * debug
  * Part of FSE library
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
diff --git a/contrib/zstd/debug.h b/contrib/zstd/debug.h
index ac6224888..0e9817ea6 100644
--- a/contrib/zstd/debug.h
+++ b/contrib/zstd/debug.h
@@ -1,7 +1,7 @@
 /* ******************************************************************
  * debug
  * Part of FSE library
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -51,15 +51,6 @@ extern "C" {
 #endif
 
 
-/* DEBUGFILE can be defined externally,
- * typically through compiler command line.
- * note : currently useless.
- * Value must be stderr or stdout */
-#ifndef DEBUGFILE
-#  define DEBUGFILE stderr
-#endif
-
-
 /* recommended values for DEBUGLEVEL :
  * 0 : release mode, no debug, all run-time checks disabled
  * 1 : enables assert() only, no display
@@ -76,7 +67,8 @@ extern "C" {
  */
 
 #if (DEBUGLEVEL>=1)
-#  include <assert.h>
+#  define ZSTD_DEPS_NEED_ASSERT
+#  include "zstd_deps.h"
 #else
 #  ifndef assert   /* assert may be already defined, due to prior #include <assert.h> */
 #    define assert(condition) ((void)0)   /* disable assert (default) */
@@ -84,7 +76,8 @@ extern "C" {
 #endif
 
 #if (DEBUGLEVEL>=2)
-#  include <stdio.h>
+#  define ZSTD_DEPS_NEED_IO
+#  include "zstd_deps.h"
 extern int g_debuglevel; /* the variable is only declared,
                             it actually lives in debug.c,
                             and is shared by the whole process.
@@ -92,14 +85,14 @@ extern int g_debuglevel; /* the variable is only declared,
                             It's useful when enabling very verbose levels
                             on selective conditions (such as position in src) */
 
-#  define RAWLOG(l, ...) {                                      \
-                if (l<=g_debuglevel) {                          \
-                    fprintf(stderr, __VA_ARGS__);               \
+#  define RAWLOG(l, ...) {                                       \
+                if (l<=g_debuglevel) {                           \
+                    ZSTD_DEBUG_PRINT(__VA_ARGS__);               \
             }   }
-#  define DEBUGLOG(l, ...) {                                    \
-                if (l<=g_debuglevel) {                          \
-                    fprintf(stderr, __FILE__ ": " __VA_ARGS__); \
-                    fprintf(stderr, " \n");                     \
+#  define DEBUGLOG(l, ...) {                                     \
+                if (l<=g_debuglevel) {                           \
+                    ZSTD_DEBUG_PRINT(__FILE__ ": " __VA_ARGS__); \
+                    ZSTD_DEBUG_PRINT(" \n");                     \
             }   }
 #else
 #  define RAWLOG(l, ...)      {}    /* disabled */
diff --git a/contrib/zstd/entropy_common.c b/contrib/zstd/entropy_common.c
index 9d3e4e8e3..e2173afb0 100644
--- a/contrib/zstd/entropy_common.c
+++ b/contrib/zstd/entropy_common.c
@@ -1,6 +1,6 @@
 /* ******************************************************************
  * Common functions of New Generation Entropy library
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  *  You can contact the author at :
  *  - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -19,8 +19,8 @@
 #include "error_private.h"       /* ERR_*, ERROR */
 #define FSE_STATIC_LINKING_ONLY  /* FSE_MIN_TABLELOG */
 #include "fse.h"
-#define HUF_STATIC_LINKING_ONLY  /* HUF_TABLELOG_ABSOLUTEMAX */
 #include "huf.h"
+#include "bits.h"                /* ZSDT_highbit32, ZSTD_countTrailingZeros32 */
 
 
 /*===   Version   ===*/
@@ -38,8 +38,9 @@ const char* HUF_getErrorName(size_t code) { return ERR_getErrorName(code); }
 /*-**************************************************************
 *  FSE NCount encoding-decoding
 ****************************************************************/
-size_t FSE_readNCount (short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
-                 const void* headerBuffer, size_t hbSize)
+FORCE_INLINE_TEMPLATE
+size_t FSE_readNCount_body(short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+                           const void* headerBuffer, size_t hbSize)
 {
     const BYTE* const istart = (const BYTE*) headerBuffer;
     const BYTE* const iend = istart + hbSize;
@@ -50,23 +51,23 @@ size_t FSE_readNCount (short* normalizedCounter, unsigned* maxSVPtr, unsigned* t
     U32 bitStream;
     int bitCount;
     unsigned charnum = 0;
+    unsigned const maxSV1 = *maxSVPtr + 1;
     int previous0 = 0;
 
-    if (hbSize < 4) {
-        /* This function only works when hbSize >= 4 */
-        char buffer[4];
-        memset(buffer, 0, sizeof(buffer));
-        memcpy(buffer, headerBuffer, hbSize);
+    if (hbSize < 8) {
+        /* This function only works when hbSize >= 8 */
+        char buffer[8] = {0};
+        ZSTD_memcpy(buffer, headerBuffer, hbSize);
         {   size_t const countSize = FSE_readNCount(normalizedCounter, maxSVPtr, tableLogPtr,
                                                     buffer, sizeof(buffer));
             if (FSE_isError(countSize)) return countSize;
             if (countSize > hbSize) return ERROR(corruption_detected);
             return countSize;
     }   }
-    assert(hbSize >= 4);
+    assert(hbSize >= 8);
 
     /* init */
-    memset(normalizedCounter, 0, (*maxSVPtr+1) * sizeof(normalizedCounter[0]));   /* all symbols not present in NCount have a frequency of 0 */
+    ZSTD_memset(normalizedCounter, 0, (*maxSVPtr+1) * sizeof(normalizedCounter[0]));   /* all symbols not present in NCount have a frequency of 0 */
     bitStream = MEM_readLE32(ip);
     nbBits = (bitStream & 0xF) + FSE_MIN_TABLELOG;   /* extract tableLog */
     if (nbBits > FSE_TABLELOG_ABSOLUTE_MAX) return ERROR(tableLog_tooLarge);
@@ -77,36 +78,58 @@ size_t FSE_readNCount (short* normalizedCounter, unsigned* maxSVPtr, unsigned* t
     threshold = 1<<nbBits;
     nbBits++;
 
-    while ((remaining>1) & (charnum<=*maxSVPtr)) {
+    for (;;) {
         if (previous0) {
-            unsigned n0 = charnum;
-            while ((bitStream & 0xFFFF) == 0xFFFF) {
-                n0 += 24;
-                if (ip < iend-5) {
-                    ip += 2;
-                    bitStream = MEM_readLE32(ip) >> bitCount;
+            /* Count the number of repeats. Each time the
+             * 2-bit repeat code is 0b11 there is another
+             * repeat.
+             * Avoid UB by setting the high bit to 1.
+             */
+            int repeats = ZSTD_countTrailingZeros32(~bitStream | 0x80000000) >> 1;
+            while (repeats >= 12) {
+                charnum += 3 * 12;
+                if (LIKELY(ip <= iend-7)) {
+                    ip += 3;
                 } else {
-                    bitStream >>= 16;
-                    bitCount   += 16;
-            }   }
-            while ((bitStream & 3) == 3) {
-                n0 += 3;
-                bitStream >>= 2;
-                bitCount += 2;
+                    bitCount -= (int)(8 * (iend - 7 - ip));
+                    bitCount &= 31;
+                    ip = iend - 4;
+                }
+                bitStream = MEM_readLE32(ip) >> bitCount;
+                repeats = ZSTD_countTrailingZeros32(~bitStream | 0x80000000) >> 1;
             }
-            n0 += bitStream & 3;
+            charnum += 3 * repeats;
+            bitStream >>= 2 * repeats;
+            bitCount += 2 * repeats;
+
+            /* Add the final repeat which isn't 0b11. */
+            assert((bitStream & 3) < 3);
+            charnum += bitStream & 3;
             bitCount += 2;
-            if (n0 > *maxSVPtr) return ERROR(maxSymbolValue_tooSmall);
-            while (charnum < n0) normalizedCounter[charnum++] = 0;
-            if ((ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
+
+            /* This is an error, but break and return an error
+             * at the end, because returning out of a loop makes
+             * it harder for the compiler to optimize.
+             */
+            if (charnum >= maxSV1) break;
+
+            /* We don't need to set the normalized count to 0
+             * because we already memset the whole buffer to 0.
+             */
+
+            if (LIKELY(ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
                 assert((bitCount >> 3) <= 3); /* For first condition to work */
                 ip += bitCount>>3;
                 bitCount &= 7;
-                bitStream = MEM_readLE32(ip) >> bitCount;
             } else {
-                bitStream >>= 2;
-        }   }
-        {   int const max = (2*threshold-1) - remaining;
+                bitCount -= (int)(8 * (iend - 4 - ip));
+                bitCount &= 31;
+                ip = iend - 4;
+            }
+            bitStream = MEM_readLE32(ip) >> bitCount;
+        }
+        {
+            int const max = (2*threshold-1) - remaining;
             int count;
 
             if ((bitStream & (threshold-1)) < (U32)max) {
@@ -119,24 +142,43 @@ size_t FSE_readNCount (short* normalizedCounter, unsigned* maxSVPtr, unsigned* t
             }
 
             count--;   /* extra accuracy */
-            remaining -= count < 0 ? -count : count;   /* -1 means +1 */
+            /* When it matters (small blocks), this is a
+             * predictable branch, because we don't use -1.
+             */
+            if (count >= 0) {
+                remaining -= count;
+            } else {
+                assert(count == -1);
+                remaining += count;
+            }
             normalizedCounter[charnum++] = (short)count;
             previous0 = !count;
-            while (remaining < threshold) {
-                nbBits--;
-                threshold >>= 1;
+
+            assert(threshold > 1);
+            if (remaining < threshold) {
+                /* This branch can be folded into the
+                 * threshold update condition because we
+                 * know that threshold > 1.
+                 */
+                if (remaining <= 1) break;
+                nbBits = ZSTD_highbit32(remaining) + 1;
+                threshold = 1 << (nbBits - 1);
             }
+            if (charnum >= maxSV1) break;
 
-            if ((ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
+            if (LIKELY(ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
                 ip += bitCount>>3;
                 bitCount &= 7;
             } else {
                 bitCount -= (int)(8 * (iend - 4 - ip));
+                bitCount &= 31;
                 ip = iend - 4;
             }
-            bitStream = MEM_readLE32(ip) >> (bitCount & 31);
-    }   }   /* while ((remaining>1) & (charnum<=*maxSVPtr)) */
+            bitStream = MEM_readLE32(ip) >> bitCount;
+    }   }
     if (remaining != 1) return ERROR(corruption_detected);
+    /* Only possible when there are too many zeros. */
+    if (charnum > maxSV1) return ERROR(maxSymbolValue_tooSmall);
     if (bitCount > 32) return ERROR(corruption_detected);
     *maxSVPtr = charnum-1;
 
@@ -144,6 +186,43 @@ size_t FSE_readNCount (short* normalizedCounter, unsigned* maxSVPtr, unsigned* t
     return ip-istart;
 }
 
+/* Avoids the FORCE_INLINE of the _body() function. */
+static size_t FSE_readNCount_body_default(
+        short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+        const void* headerBuffer, size_t hbSize)
+{
+    return FSE_readNCount_body(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
+}
+
+#if DYNAMIC_BMI2
+BMI2_TARGET_ATTRIBUTE static size_t FSE_readNCount_body_bmi2(
+        short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+        const void* headerBuffer, size_t hbSize)
+{
+    return FSE_readNCount_body(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
+}
+#endif
+
+size_t FSE_readNCount_bmi2(
+        short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+        const void* headerBuffer, size_t hbSize, int bmi2)
+{
+#if DYNAMIC_BMI2
+    if (bmi2) {
+        return FSE_readNCount_body_bmi2(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
+    }
+#endif
+    (void)bmi2;
+    return FSE_readNCount_body_default(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
+}
+
+size_t FSE_readNCount(
+        short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+        const void* headerBuffer, size_t hbSize)
+{
+    return FSE_readNCount_bmi2(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize, /* bmi2 */ 0);
+}
+
 
 /*! HUF_readStats() :
     Read compact Huffman tree, saved by HUF_writeCTable().
@@ -156,6 +235,17 @@ size_t HUF_readStats(BYTE* huffWeight, size_t hwSize, U32* rankStats,
                      U32* nbSymbolsPtr, U32* tableLogPtr,
                      const void* src, size_t srcSize)
 {
+    U32 wksp[HUF_READ_STATS_WORKSPACE_SIZE_U32];
+    return HUF_readStats_wksp(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, wksp, sizeof(wksp), /* flags */ 0);
+}
+
+FORCE_INLINE_TEMPLATE size_t
+HUF_readStats_body(BYTE* huffWeight, size_t hwSize, U32* rankStats,
+                   U32* nbSymbolsPtr, U32* tableLogPtr,
+                   const void* src, size_t srcSize,
+                   void* workSpace, size_t wkspSize,
+                   int bmi2)
+{
     U32 weightTotal;
     const BYTE* ip = (const BYTE*) src;
     size_t iSize;
@@ -163,7 +253,7 @@ size_t HUF_readStats(BYTE* huffWeight, size_t hwSize, U32* rankStats,
 
     if (!srcSize) return ERROR(srcSize_wrong);
     iSize = ip[0];
-    /* memset(huffWeight, 0, hwSize);   *//* is not necessary, even though some analyzer complain ... */
+    /* ZSTD_memset(huffWeight, 0, hwSize);   *//* is not necessary, even though some analyzer complain ... */
 
     if (iSize >= 128) {  /* special header */
         oSize = iSize - 127;
@@ -177,31 +267,31 @@ size_t HUF_readStats(BYTE* huffWeight, size_t hwSize, U32* rankStats,
                 huffWeight[n+1] = ip[n/2] & 15;
     }   }   }
     else  {   /* header compressed with FSE (normal case) */
-        FSE_DTable fseWorkspace[FSE_DTABLE_SIZE_U32(6)];  /* 6 is max possible tableLog for HUF header (maybe even 5, to be tested) */
         if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
-        oSize = FSE_decompress_wksp(huffWeight, hwSize-1, ip+1, iSize, fseWorkspace, 6);   /* max (hwSize-1) values decoded, as last one is implied */
+        /* max (hwSize-1) values decoded, as last one is implied */
+        oSize = FSE_decompress_wksp_bmi2(huffWeight, hwSize-1, ip+1, iSize, 6, workSpace, wkspSize, bmi2);
         if (FSE_isError(oSize)) return oSize;
     }
 
     /* collect weight stats */
-    memset(rankStats, 0, (HUF_TABLELOG_MAX + 1) * sizeof(U32));
+    ZSTD_memset(rankStats, 0, (HUF_TABLELOG_MAX + 1) * sizeof(U32));
     weightTotal = 0;
     {   U32 n; for (n=0; n<oSize; n++) {
-            if (huffWeight[n] >= HUF_TABLELOG_MAX) return ERROR(corruption_detected);
+            if (huffWeight[n] > HUF_TABLELOG_MAX) return ERROR(corruption_detected);
             rankStats[huffWeight[n]]++;
             weightTotal += (1 << huffWeight[n]) >> 1;
     }   }
     if (weightTotal == 0) return ERROR(corruption_detected);
 
     /* get last non-null symbol weight (implied, total must be 2^n) */
-    {   U32 const tableLog = BIT_highbit32(weightTotal) + 1;
+    {   U32 const tableLog = ZSTD_highbit32(weightTotal) + 1;
         if (tableLog > HUF_TABLELOG_MAX) return ERROR(corruption_detected);
         *tableLogPtr = tableLog;
         /* determine last weight */
         {   U32 const total = 1 << tableLog;
             U32 const rest = total - weightTotal;
-            U32 const verif = 1 << BIT_highbit32(rest);
-            U32 const lastWeight = BIT_highbit32(rest) + 1;
+            U32 const verif = 1 << ZSTD_highbit32(rest);
+            U32 const lastWeight = ZSTD_highbit32(rest) + 1;
             if (verif != rest) return ERROR(corruption_detected);    /* last value must be a clean power of 2 */
             huffWeight[oSize] = (BYTE)lastWeight;
             rankStats[lastWeight]++;
@@ -214,3 +304,37 @@ size_t HUF_readStats(BYTE* huffWeight, size_t hwSize, U32* rankStats,
     *nbSymbolsPtr = (U32)(oSize+1);
     return iSize+1;
 }
+
+/* Avoids the FORCE_INLINE of the _body() function. */
+static size_t HUF_readStats_body_default(BYTE* huffWeight, size_t hwSize, U32* rankStats,
+                     U32* nbSymbolsPtr, U32* tableLogPtr,
+                     const void* src, size_t srcSize,
+                     void* workSpace, size_t wkspSize)
+{
+    return HUF_readStats_body(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize, 0);
+}
+
+#if DYNAMIC_BMI2
+static BMI2_TARGET_ATTRIBUTE size_t HUF_readStats_body_bmi2(BYTE* huffWeight, size_t hwSize, U32* rankStats,
+                     U32* nbSymbolsPtr, U32* tableLogPtr,
+                     const void* src, size_t srcSize,
+                     void* workSpace, size_t wkspSize)
+{
+    return HUF_readStats_body(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize, 1);
+}
+#endif
+
+size_t HUF_readStats_wksp(BYTE* huffWeight, size_t hwSize, U32* rankStats,
+                     U32* nbSymbolsPtr, U32* tableLogPtr,
+                     const void* src, size_t srcSize,
+                     void* workSpace, size_t wkspSize,
+                     int flags)
+{
+#if DYNAMIC_BMI2
+    if (flags & HUF_flags_bmi2) {
+        return HUF_readStats_body_bmi2(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize);
+    }
+#endif
+    (void)flags;
+    return HUF_readStats_body_default(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize);
+}
diff --git a/contrib/zstd/error_private.c b/contrib/zstd/error_private.c
index cd437529c..075fc5ef4 100644
--- a/contrib/zstd/error_private.c
+++ b/contrib/zstd/error_private.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -27,9 +27,11 @@ const char* ERR_getErrorString(ERR_enum code)
     case PREFIX(version_unsupported): return "Version not supported";
     case PREFIX(frameParameter_unsupported): return "Unsupported frame parameter";
     case PREFIX(frameParameter_windowTooLarge): return "Frame requires too much memory for decoding";
-    case PREFIX(corruption_detected): return "Corrupted block detected";
+    case PREFIX(corruption_detected): return "Data corruption detected";
     case PREFIX(checksum_wrong): return "Restored data doesn't match checksum";
+    case PREFIX(literals_headerWrong): return "Header of Literals' block doesn't respect format specification";
     case PREFIX(parameter_unsupported): return "Unsupported parameter";
+    case PREFIX(parameter_combination_unsupported): return "Unsupported combination of parameters";
     case PREFIX(parameter_outOfBound): return "Parameter is out of bound";
     case PREFIX(init_missing): return "Context should be init first";
     case PREFIX(memory_allocation): return "Allocation error : not enough memory";
@@ -38,16 +40,22 @@ const char* ERR_getErrorString(ERR_enum code)
     case PREFIX(tableLog_tooLarge): return "tableLog requires too much memory : unsupported";
     case PREFIX(maxSymbolValue_tooLarge): return "Unsupported max Symbol Value : too large";
     case PREFIX(maxSymbolValue_tooSmall): return "Specified maxSymbolValue is too small";
+    case PREFIX(stabilityCondition_notRespected): return "pledged buffer stability condition is not respected";
     case PREFIX(dictionary_corrupted): return "Dictionary is corrupted";
     case PREFIX(dictionary_wrong): return "Dictionary mismatch";
     case PREFIX(dictionaryCreation_failed): return "Cannot create Dictionary from provided samples";
     case PREFIX(dstSize_tooSmall): return "Destination buffer is too small";
     case PREFIX(srcSize_wrong): return "Src size is incorrect";
     case PREFIX(dstBuffer_null): return "Operation on NULL destination buffer";
+    case PREFIX(noForwardProgress_destFull): return "Operation made no progress over multiple calls, due to output buffer being full";
+    case PREFIX(noForwardProgress_inputEmpty): return "Operation made no progress over multiple calls, due to input being empty";
         /* following error codes are not stable and may be removed or changed in a future version */
     case PREFIX(frameIndex_tooLarge): return "Frame index is too large";
     case PREFIX(seekableIO): return "An I/O error occurred when reading/seeking";
     case PREFIX(dstBuffer_wrong): return "Destination buffer is wrong";
+    case PREFIX(srcBuffer_wrong): return "Source buffer is wrong";
+    case PREFIX(sequenceProducer_failed): return "Block-level external sequence producer returned an error code";
+    case PREFIX(externalSequences_invalid): return "External sequences are not valid";
     case PREFIX(maxCode):
     default: return notErrorCode;
     }
diff --git a/contrib/zstd/error_private.h b/contrib/zstd/error_private.h
index 982cf8e9f..d99c9619e 100644
--- a/contrib/zstd/error_private.h
+++ b/contrib/zstd/error_private.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -21,8 +21,10 @@ extern "C" {
 /* ****************************************
 *  Dependencies
 ******************************************/
-#include <stddef.h>        /* size_t */
 #include "zstd_errors.h"  /* enum list */
+#include "compiler.h"
+#include "debug.h"
+#include "zstd_deps.h"       /* size_t */
 
 
 /* ****************************************
@@ -73,6 +75,83 @@ ERR_STATIC const char* ERR_getErrorName(size_t code)
     return ERR_getErrorString(ERR_getErrorCode(code));
 }
 
+/**
+ * Ignore: this is an internal helper.
+ *
+ * This is a helper function to help force C99-correctness during compilation.
+ * Under strict compilation modes, variadic macro arguments can't be empty.
+ * However, variadic function arguments can be. Using a function therefore lets
+ * us statically check that at least one (string) argument was passed,
+ * independent of the compilation flags.
+ */
+static INLINE_KEYWORD UNUSED_ATTR
+void _force_has_format_string(const char *format, ...) {
+  (void)format;
+}
+
+/**
+ * Ignore: this is an internal helper.
+ *
+ * We want to force this function invocation to be syntactically correct, but
+ * we don't want to force runtime evaluation of its arguments.
+ */
+#define _FORCE_HAS_FORMAT_STRING(...) \
+  if (0) { \
+    _force_has_format_string(__VA_ARGS__); \
+  }
+
+#define ERR_QUOTE(str) #str
+
+/**
+ * Return the specified error if the condition evaluates to true.
+ *
+ * In debug modes, prints additional information.
+ * In order to do that (particularly, printing the conditional that failed),
+ * this can't just wrap RETURN_ERROR().
+ */
+#define RETURN_ERROR_IF(cond, err, ...) \
+  if (cond) { \
+    RAWLOG(3, "%s:%d: ERROR!: check %s failed, returning %s", \
+           __FILE__, __LINE__, ERR_QUOTE(cond), ERR_QUOTE(ERROR(err))); \
+    _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
+    RAWLOG(3, ": " __VA_ARGS__); \
+    RAWLOG(3, "\n"); \
+    return ERROR(err); \
+  }
+
+/**
+ * Unconditionally return the specified error.
+ *
+ * In debug modes, prints additional information.
+ */
+#define RETURN_ERROR(err, ...) \
+  do { \
+    RAWLOG(3, "%s:%d: ERROR!: unconditional check failed, returning %s", \
+           __FILE__, __LINE__, ERR_QUOTE(ERROR(err))); \
+    _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
+    RAWLOG(3, ": " __VA_ARGS__); \
+    RAWLOG(3, "\n"); \
+    return ERROR(err); \
+  } while(0);
+
+/**
+ * If the provided expression evaluates to an error code, returns that error code.
+ *
+ * In debug modes, prints additional information.
+ */
+#define FORWARD_IF_ERROR(err, ...) \
+  do { \
+    size_t const err_code = (err); \
+    if (ERR_isError(err_code)) { \
+      RAWLOG(3, "%s:%d: ERROR!: forwarding error in %s: %s", \
+             __FILE__, __LINE__, ERR_QUOTE(err), ERR_getErrorName(err_code)); \
+      _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
+      RAWLOG(3, ": " __VA_ARGS__); \
+      RAWLOG(3, "\n"); \
+      return err_code; \
+    } \
+  } while(0);
+
 #if defined (__cplusplus)
 }
 #endif
diff --git a/contrib/zstd/fse.h b/contrib/zstd/fse.h
index ff54e70ea..02a1f0bc5 100644
--- a/contrib/zstd/fse.h
+++ b/contrib/zstd/fse.h
@@ -1,7 +1,7 @@
 /* ******************************************************************
  * FSE : Finite State Entropy codec
  * Public Prototypes declaration
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -23,7 +23,7 @@ extern "C" {
 /*-*****************************************
 *  Dependencies
 ******************************************/
-#include <stddef.h>    /* size_t, ptrdiff_t */
+#include "zstd_deps.h"    /* size_t, ptrdiff_t */
 
 
 /*-*****************************************
@@ -53,34 +53,6 @@ extern "C" {
 FSE_PUBLIC_API unsigned FSE_versionNumber(void);   /**< library version number; to be used when checking dll version */
 
 
-/*-****************************************
-*  FSE simple functions
-******************************************/
-/*! FSE_compress() :
-    Compress content of buffer 'src', of size 'srcSize', into destination buffer 'dst'.
-    'dst' buffer must be already allocated. Compression runs faster is dstCapacity >= FSE_compressBound(srcSize).
-    @return : size of compressed data (<= dstCapacity).
-    Special values : if return == 0, srcData is not compressible => Nothing is stored within dst !!!
-                     if return == 1, srcData is a single byte symbol * srcSize times. Use RLE compression instead.
-                     if FSE_isError(return), compression failed (more details using FSE_getErrorName())
-*/
-FSE_PUBLIC_API size_t FSE_compress(void* dst, size_t dstCapacity,
-                             const void* src, size_t srcSize);
-
-/*! FSE_decompress():
-    Decompress FSE data from buffer 'cSrc', of size 'cSrcSize',
-    into already allocated destination buffer 'dst', of size 'dstCapacity'.
-    @return : size of regenerated data (<= maxDstSize),
-              or an error code, which can be tested using FSE_isError() .
-
-    ** Important ** : FSE_decompress() does not decompress non-compressible nor RLE data !!!
-    Why ? : making this distinction requires a header.
-    Header management is intentionally delegated to the user layer, which can better manage special cases.
-*/
-FSE_PUBLIC_API size_t FSE_decompress(void* dst,  size_t dstCapacity,
-                               const void* cSrc, size_t cSrcSize);
-
-
 /*-*****************************************
 *  Tool functions
 ******************************************/
@@ -92,20 +64,6 @@ FSE_PUBLIC_API const char* FSE_getErrorName(size_t code);   /* provides error co
 
 
 /*-*****************************************
-*  FSE advanced functions
-******************************************/
-/*! FSE_compress2() :
-    Same as FSE_compress(), but allows the selection of 'maxSymbolValue' and 'tableLog'
-    Both parameters can be defined as '0' to mean : use default value
-    @return : size of compressed data
-    Special values : if return == 0, srcData is not compressible => Nothing is stored within cSrc !!!
-                     if return == 1, srcData is a single byte symbol * srcSize times. Use RLE compression.
-                     if FSE_isError(return), it's an error code.
-*/
-FSE_PUBLIC_API size_t FSE_compress2 (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog);
-
-
-/*-*****************************************
 *  FSE detailed API
 ******************************************/
 /*!
@@ -137,10 +95,16 @@ FSE_PUBLIC_API unsigned FSE_optimalTableLog(unsigned maxTableLog, size_t srcSize
 /*! FSE_normalizeCount():
     normalize counts so that sum(count[]) == Power_of_2 (2^tableLog)
     'normalizedCounter' is a table of short, of minimum size (maxSymbolValue+1).
+    useLowProbCount is a boolean parameter which trades off compressed size for
+    faster header decoding. When it is set to 1, the compressed data will be slightly
+    smaller. And when it is set to 0, FSE_readNCount() and FSE_buildDTable() will be
+    faster. If you are compressing a small amount of data (< 2 KB) then useLowProbCount=0
+    is a good default, since header deserialization makes a big speed difference.
+    Otherwise, useLowProbCount=1 is a good default, since the speed difference is small.
     @return : tableLog,
               or an errorCode, which can be tested using FSE_isError() */
 FSE_PUBLIC_API size_t FSE_normalizeCount(short* normalizedCounter, unsigned tableLog,
-                    const unsigned* count, size_t srcSize, unsigned maxSymbolValue);
+                    const unsigned* count, size_t srcSize, unsigned maxSymbolValue, unsigned useLowProbCount);
 
 /*! FSE_NCountWriteBound():
     Provides the maximum possible size of an FSE normalized table, given 'maxSymbolValue' and 'tableLog'.
@@ -158,8 +122,6 @@ FSE_PUBLIC_API size_t FSE_writeNCount (void* buffer, size_t bufferSize,
 /*! Constructor and Destructor of FSE_CTable.
     Note that FSE_CTable size depends on 'tableLog' and 'maxSymbolValue' */
 typedef unsigned FSE_CTable;   /* don't allocate that. It's only meant to be more restrictive than void* */
-FSE_PUBLIC_API FSE_CTable* FSE_createCTable (unsigned maxSymbolValue, unsigned tableLog);
-FSE_PUBLIC_API void        FSE_freeCTable (FSE_CTable* ct);
 
 /*! FSE_buildCTable():
     Builds `ct`, which must be already allocated, using FSE_createCTable().
@@ -228,23 +190,14 @@ FSE_PUBLIC_API size_t FSE_readNCount (short* normalizedCounter,
                            unsigned* maxSymbolValuePtr, unsigned* tableLogPtr,
                            const void* rBuffer, size_t rBuffSize);
 
-/*! Constructor and Destructor of FSE_DTable.
-    Note that its size depends on 'tableLog' */
+/*! FSE_readNCount_bmi2():
+ * Same as FSE_readNCount() but pass bmi2=1 when your CPU supports BMI2 and 0 otherwise.
+ */
+FSE_PUBLIC_API size_t FSE_readNCount_bmi2(short* normalizedCounter,
+                           unsigned* maxSymbolValuePtr, unsigned* tableLogPtr,
+                           const void* rBuffer, size_t rBuffSize, int bmi2);
+
 typedef unsigned FSE_DTable;   /* don't allocate that. It's just a way to be more restrictive than void* */
-FSE_PUBLIC_API FSE_DTable* FSE_createDTable(unsigned tableLog);
-FSE_PUBLIC_API void        FSE_freeDTable(FSE_DTable* dt);
-
-/*! FSE_buildDTable():
-    Builds 'dt', which must be already allocated, using FSE_createDTable().
-    return : 0, or an errorCode, which can be tested using FSE_isError() */
-FSE_PUBLIC_API size_t FSE_buildDTable (FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog);
-
-/*! FSE_decompress_usingDTable():
-    Decompress compressed source `cSrc` of size `cSrcSize` using `dt`
-    into `dst` which must be already allocated.
-    @return : size of regenerated data (necessarily <= `dstCapacity`),
-              or an errorCode, which can be tested using FSE_isError() */
-FSE_PUBLIC_API size_t FSE_decompress_usingDTable(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, const FSE_DTable* dt);
 
 /*!
 Tutorial :
@@ -288,12 +241,12 @@ If there is an error, the function will return an error code, which can be teste
 *******************************************/
 /* FSE buffer bounds */
 #define FSE_NCOUNTBOUND 512
-#define FSE_BLOCKBOUND(size) (size + (size>>7) + 4 /* fse states */ + sizeof(size_t) /* bitContainer */)
+#define FSE_BLOCKBOUND(size) ((size) + ((size)>>7) + 4 /* fse states */ + sizeof(size_t) /* bitContainer */)
 #define FSE_COMPRESSBOUND(size) (FSE_NCOUNTBOUND + FSE_BLOCKBOUND(size))   /* Macro version, useful for static allocation */
 
 /* It is possible to statically allocate FSE CTable/DTable as a table of FSE_CTable/FSE_DTable using below macros */
-#define FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue)   (1 + (1<<(maxTableLog-1)) + ((maxSymbolValue+1)*2))
-#define FSE_DTABLE_SIZE_U32(maxTableLog)                   (1 + (1<<maxTableLog))
+#define FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue)   (1 + (1<<((maxTableLog)-1)) + (((maxSymbolValue)+1)*2))
+#define FSE_DTABLE_SIZE_U32(maxTableLog)                   (1 + (1<<(maxTableLog)))
 
 /* or use the size to malloc() space directly. Pay attention to alignment restrictions though */
 #define FSE_CTABLE_SIZE(maxTableLog, maxSymbolValue)   (FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue) * sizeof(FSE_CTable))
@@ -307,33 +260,28 @@ If there is an error, the function will return an error code, which can be teste
 unsigned FSE_optimalTableLog_internal(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, unsigned minus);
 /**< same as FSE_optimalTableLog(), which used `minus==2` */
 
-/* FSE_compress_wksp() :
- * Same as FSE_compress2(), but using an externally allocated scratch buffer (`workSpace`).
- * FSE_WKSP_SIZE_U32() provides the minimum size required for `workSpace` as a table of FSE_CTable.
- */
-#define FSE_WKSP_SIZE_U32(maxTableLog, maxSymbolValue)   ( FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue) + ((maxTableLog > 12) ? (1 << (maxTableLog - 2)) : 1024) )
-size_t FSE_compress_wksp (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);
-
-size_t FSE_buildCTable_raw (FSE_CTable* ct, unsigned nbBits);
-/**< build a fake FSE_CTable, designed for a flat distribution, where each symbol uses nbBits */
-
 size_t FSE_buildCTable_rle (FSE_CTable* ct, unsigned char symbolValue);
 /**< build a fake FSE_CTable, designed to compress always the same symbolValue */
 
 /* FSE_buildCTable_wksp() :
  * Same as FSE_buildCTable(), but using an externally allocated scratch buffer (`workSpace`).
- * `wkspSize` must be >= `(1<<tableLog)`.
+ * `wkspSize` must be >= `FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(maxSymbolValue, tableLog)` of `unsigned`.
+ * See FSE_buildCTable_wksp() for breakdown of workspace usage.
  */
+#define FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(maxSymbolValue, tableLog) (((maxSymbolValue + 2) + (1ull << (tableLog)))/2 + sizeof(U64)/sizeof(U32) /* additional 8 bytes for potential table overwrite */)
+#define FSE_BUILD_CTABLE_WORKSPACE_SIZE(maxSymbolValue, tableLog) (sizeof(unsigned) * FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(maxSymbolValue, tableLog))
 size_t FSE_buildCTable_wksp(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);
 
-size_t FSE_buildDTable_raw (FSE_DTable* dt, unsigned nbBits);
-/**< build a fake FSE_DTable, designed to read a flat distribution where each symbol uses nbBits */
+#define FSE_BUILD_DTABLE_WKSP_SIZE(maxTableLog, maxSymbolValue) (sizeof(short) * (maxSymbolValue + 1) + (1ULL << maxTableLog) + 8)
+#define FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) ((FSE_BUILD_DTABLE_WKSP_SIZE(maxTableLog, maxSymbolValue) + sizeof(unsigned) - 1) / sizeof(unsigned))
+FSE_PUBLIC_API size_t FSE_buildDTable_wksp(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);
+/**< Same as FSE_buildDTable(), using an externally allocated `workspace` produced with `FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxSymbolValue)` */
 
-size_t FSE_buildDTable_rle (FSE_DTable* dt, unsigned char symbolValue);
-/**< build a fake FSE_DTable, designed to always generate the same symbolValue */
-
-size_t FSE_decompress_wksp(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, FSE_DTable* workSpace, unsigned maxLog);
-/**< same as FSE_decompress(), using an externally allocated `workSpace` produced with `FSE_DTABLE_SIZE_U32(maxLog)` */
+#define FSE_DECOMPRESS_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) (FSE_DTABLE_SIZE_U32(maxTableLog) + 1 + FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) + (FSE_MAX_SYMBOL_VALUE + 1) / 2 + 1)
+#define FSE_DECOMPRESS_WKSP_SIZE(maxTableLog, maxSymbolValue) (FSE_DECOMPRESS_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) * sizeof(unsigned))
+size_t FSE_decompress_wksp_bmi2(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize, int bmi2);
+/**< same as FSE_decompress(), using an externally allocated `workSpace` produced with `FSE_DECOMPRESS_WKSP_SIZE_U32(maxLog, maxSymbolValue)`.
+ * Set bmi2 to 1 if your CPU supports BMI2 or 0 if it doesn't */
 
 typedef enum {
    FSE_repeat_none,  /**< Cannot use the previous table */
@@ -529,7 +477,7 @@ MEM_STATIC void FSE_flushCState(BIT_CStream_t* bitC, const FSE_CState_t* statePt
 
 /* FSE_getMaxNbBits() :
  * Approximate maximum cost of a symbol, in bits.
- * Fractional get rounded up (i.e : a symbol with a normalized frequency of 3 gives the same result as a frequency of 2)
+ * Fractional get rounded up (i.e. a symbol with a normalized frequency of 3 gives the same result as a frequency of 2)
  * note 1 : assume symbolValue is valid (<= maxSymbolValue)
  * note 2 : if freq[symbolValue]==0, @return a fake cost of tableLog+1 bits */
 MEM_STATIC U32 FSE_getMaxNbBits(const void* symbolTTPtr, U32 symbolValue)
@@ -644,6 +592,9 @@ MEM_STATIC unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr)
 #ifndef FSE_DEFAULT_MEMORY_USAGE
 #  define FSE_DEFAULT_MEMORY_USAGE 13
 #endif
+#if (FSE_DEFAULT_MEMORY_USAGE > FSE_MAX_MEMORY_USAGE)
+#  error "FSE_DEFAULT_MEMORY_USAGE must be <= FSE_MAX_MEMORY_USAGE"
+#endif
 
 /*!FSE_MAX_SYMBOL_VALUE :
 *  Maximum symbol value authorized.
@@ -677,7 +628,7 @@ MEM_STATIC unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr)
 #  error "FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX is not supported"
 #endif
 
-#define FSE_TABLESTEP(tableSize) ((tableSize>>1) + (tableSize>>3) + 3)
+#define FSE_TABLESTEP(tableSize) (((tableSize)>>1) + ((tableSize)>>3) + 3)
 
 
 #endif /* FSE_STATIC_LINKING_ONLY */
diff --git a/contrib/zstd/fse_compress.c b/contrib/zstd/fse_compress.c
index e09b646df..3fec4a3ba 100644
--- a/contrib/zstd/fse_compress.c
+++ b/contrib/zstd/fse_compress.c
@@ -1,6 +1,6 @@
 /* ******************************************************************
  * FSE : Finite State Entropy encoder
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  *  You can contact the author at :
  *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -15,8 +15,6 @@
 /* **************************************************************
 *  Includes
 ****************************************************************/
-#include <stdlib.h>     /* malloc, free, qsort */
-#include <string.h>     /* memcpy, memset */
 #include "compiler.h"
 #include "mem.h"        /* U32, U16, etc. */
 #include "debug.h"      /* assert, DEBUGLOG */
@@ -25,6 +23,10 @@
 #define FSE_STATIC_LINKING_ONLY
 #include "fse.h"
 #include "error_private.h"
+#define ZSTD_DEPS_NEED_MALLOC
+#define ZSTD_DEPS_NEED_MATH64
+#include "zstd_deps.h"  /* ZSTD_malloc, ZSTD_free, ZSTD_memcpy, ZSTD_memset */
+#include "bits.h" /* ZSTD_highbit32 */
 
 
 /* **************************************************************
@@ -74,41 +76,85 @@ size_t FSE_buildCTable_wksp(FSE_CTable* ct,
     void* const FSCT = ((U32*)ptr) + 1 /* header */ + (tableLog ? tableSize>>1 : 1) ;
     FSE_symbolCompressionTransform* const symbolTT = (FSE_symbolCompressionTransform*) (FSCT);
     U32 const step = FSE_TABLESTEP(tableSize);
-    U32 cumul[FSE_MAX_SYMBOL_VALUE+2];
+    U32 const maxSV1 = maxSymbolValue+1;
+
+    U16* cumul = (U16*)workSpace;   /* size = maxSV1 */
+    FSE_FUNCTION_TYPE* const tableSymbol = (FSE_FUNCTION_TYPE*)(cumul + (maxSV1+1));  /* size = tableSize */
 
-    FSE_FUNCTION_TYPE* const tableSymbol = (FSE_FUNCTION_TYPE*)workSpace;
     U32 highThreshold = tableSize-1;
 
+    assert(((size_t)workSpace & 1) == 0);  /* Must be 2 bytes-aligned */
+    if (FSE_BUILD_CTABLE_WORKSPACE_SIZE(maxSymbolValue, tableLog) > wkspSize) return ERROR(tableLog_tooLarge);
     /* CTable header */
-    if (((size_t)1 << tableLog) * sizeof(FSE_FUNCTION_TYPE) > wkspSize) return ERROR(tableLog_tooLarge);
     tableU16[-2] = (U16) tableLog;
     tableU16[-1] = (U16) maxSymbolValue;
     assert(tableLog < 16);   /* required for threshold strategy to work */
 
     /* For explanations on how to distribute symbol values over the table :
-     * http://fastcompression.blogspot.fr/2014/02/fse-distributing-symbol-values.html */
+     * https://fastcompression.blogspot.fr/2014/02/fse-distributing-symbol-values.html */
 
      #ifdef __clang_analyzer__
-     memset(tableSymbol, 0, sizeof(*tableSymbol) * tableSize);   /* useless initialization, just to keep scan-build happy */
+     ZSTD_memset(tableSymbol, 0, sizeof(*tableSymbol) * tableSize);   /* useless initialization, just to keep scan-build happy */
      #endif
 
     /* symbol start positions */
     {   U32 u;
         cumul[0] = 0;
-        for (u=1; u <= maxSymbolValue+1; u++) {
+        for (u=1; u <= maxSV1; u++) {
             if (normalizedCounter[u-1]==-1) {  /* Low proba symbol */
                 cumul[u] = cumul[u-1] + 1;
                 tableSymbol[highThreshold--] = (FSE_FUNCTION_TYPE)(u-1);
             } else {
-                cumul[u] = cumul[u-1] + normalizedCounter[u-1];
+                assert(normalizedCounter[u-1] >= 0);
+                cumul[u] = cumul[u-1] + (U16)normalizedCounter[u-1];
+                assert(cumul[u] >= cumul[u-1]);  /* no overflow */
         }   }
-        cumul[maxSymbolValue+1] = tableSize+1;
+        cumul[maxSV1] = (U16)(tableSize+1);
     }
 
     /* Spread symbols */
-    {   U32 position = 0;
+    if (highThreshold == tableSize - 1) {
+        /* Case for no low prob count symbols. Lay down 8 bytes at a time
+         * to reduce branch misses since we are operating on a small block
+         */
+        BYTE* const spread = tableSymbol + tableSize; /* size = tableSize + 8 (may write beyond tableSize) */
+        {   U64 const add = 0x0101010101010101ull;
+            size_t pos = 0;
+            U64 sv = 0;
+            U32 s;
+            for (s=0; s<maxSV1; ++s, sv += add) {
+                int i;
+                int const n = normalizedCounter[s];
+                MEM_write64(spread + pos, sv);
+                for (i = 8; i < n; i += 8) {
+                    MEM_write64(spread + pos + i, sv);
+                }
+                assert(n>=0);
+                pos += (size_t)n;
+            }
+        }
+        /* Spread symbols across the table. Lack of lowprob symbols means that
+         * we don't need variable sized inner loop, so we can unroll the loop and
+         * reduce branch misses.
+         */
+        {   size_t position = 0;
+            size_t s;
+            size_t const unroll = 2; /* Experimentally determined optimal unroll */
+            assert(tableSize % unroll == 0); /* FSE_MIN_TABLELOG is 5 */
+            for (s = 0; s < (size_t)tableSize; s += unroll) {
+                size_t u;
+                for (u = 0; u < unroll; ++u) {
+                    size_t const uPosition = (position + (u * step)) & tableMask;
+                    tableSymbol[uPosition] = spread[s + u];
+                }
+                position = (position + (unroll * step)) & tableMask;
+            }
+            assert(position == 0);   /* Must have initialized all positions */
+        }
+    } else {
+        U32 position = 0;
         U32 symbol;
-        for (symbol=0; symbol<=maxSymbolValue; symbol++) {
+        for (symbol=0; symbol<maxSV1; symbol++) {
             int nbOccurrences;
             int const freq = normalizedCounter[symbol];
             for (nbOccurrences=0; nbOccurrences<freq; nbOccurrences++) {
@@ -117,7 +163,6 @@ size_t FSE_buildCTable_wksp(FSE_CTable* ct,
                 while (position > highThreshold)
                     position = (position + step) & tableMask;   /* Low proba area */
         }   }
-
         assert(position==0);  /* Must have initialized all positions */
     }
 
@@ -141,16 +186,17 @@ size_t FSE_buildCTable_wksp(FSE_CTable* ct,
             case -1:
             case  1:
                 symbolTT[s].deltaNbBits = (tableLog << 16) - (1<<tableLog);
-                symbolTT[s].deltaFindState = total - 1;
+                assert(total <= INT_MAX);
+                symbolTT[s].deltaFindState = (int)(total - 1);
                 total ++;
                 break;
             default :
-                {
-                    U32 const maxBitsOut = tableLog - BIT_highbit32 (normalizedCounter[s]-1);
-                    U32 const minStatePlus = normalizedCounter[s] << maxBitsOut;
+                assert(normalizedCounter[s] > 1);
+                {   U32 const maxBitsOut = tableLog - ZSTD_highbit32 ((U32)normalizedCounter[s]-1);
+                    U32 const minStatePlus = (U32)normalizedCounter[s] << maxBitsOut;
                     symbolTT[s].deltaNbBits = (maxBitsOut << 16) - minStatePlus;
-                    symbolTT[s].deltaFindState = total - normalizedCounter[s];
-                    total +=  normalizedCounter[s];
+                    symbolTT[s].deltaFindState = (int)(total - (unsigned)normalizedCounter[s]);
+                    total +=  (unsigned)normalizedCounter[s];
     }   }   }   }
 
 #if 0  /* debug : symbol costs */
@@ -161,31 +207,26 @@ size_t FSE_buildCTable_wksp(FSE_CTable* ct,
                 symbol, normalizedCounter[symbol],
                 FSE_getMaxNbBits(symbolTT, symbol),
                 (double)FSE_bitCost(symbolTT, tableLog, symbol, 8) / 256);
-        }
-    }
+    }   }
 #endif
 
     return 0;
 }
 
 
-size_t FSE_buildCTable(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog)
-{
-    FSE_FUNCTION_TYPE tableSymbol[FSE_MAX_TABLESIZE];   /* memset() is not necessary, even if static analyzer complain about it */
-    return FSE_buildCTable_wksp(ct, normalizedCounter, maxSymbolValue, tableLog, tableSymbol, sizeof(tableSymbol));
-}
-
-
 
 #ifndef FSE_COMMONDEFS_ONLY
 
-
 /*-**************************************************************
 *  FSE NCount encoding
 ****************************************************************/
 size_t FSE_NCountWriteBound(unsigned maxSymbolValue, unsigned tableLog)
 {
-    size_t const maxHeaderSize = (((maxSymbolValue+1) * tableLog) >> 3) + 3;
+    size_t const maxHeaderSize = (((maxSymbolValue+1) * tableLog
+                                   + 4 /* bitCount initialized at 4 */
+                                   + 2 /* first two symbols may use one additional bit each */) / 8)
+                                    + 1 /* round up to whole nb bytes */
+                                    + 2 /* additional two bytes for bitstream flush */;
     return maxSymbolValue ? maxHeaderSize : FSE_NCOUNTBOUND;  /* maxSymbolValue==0 ? use default */
 }
 
@@ -302,21 +343,11 @@ size_t FSE_writeNCount (void* buffer, size_t bufferSize,
 *  FSE Compression Code
 ****************************************************************/
 
-FSE_CTable* FSE_createCTable (unsigned maxSymbolValue, unsigned tableLog)
-{
-    size_t size;
-    if (tableLog > FSE_TABLELOG_ABSOLUTE_MAX) tableLog = FSE_TABLELOG_ABSOLUTE_MAX;
-    size = FSE_CTABLE_SIZE_U32 (tableLog, maxSymbolValue) * sizeof(U32);
-    return (FSE_CTable*)malloc(size);
-}
-
-void FSE_freeCTable (FSE_CTable* ct) { free(ct); }
-
 /* provides the minimum logSize to safely represent a distribution */
 static unsigned FSE_minTableLog(size_t srcSize, unsigned maxSymbolValue)
 {
-    U32 minBitsSrc = BIT_highbit32((U32)(srcSize)) + 1;
-    U32 minBitsSymbols = BIT_highbit32(maxSymbolValue) + 2;
+    U32 minBitsSrc = ZSTD_highbit32((U32)(srcSize)) + 1;
+    U32 minBitsSymbols = ZSTD_highbit32(maxSymbolValue) + 2;
     U32 minBits = minBitsSrc < minBitsSymbols ? minBitsSrc : minBitsSymbols;
     assert(srcSize > 1); /* Not supported, RLE should be used instead */
     return minBits;
@@ -324,7 +355,7 @@ static unsigned FSE_minTableLog(size_t srcSize, unsigned maxSymbolValue)
 
 unsigned FSE_optimalTableLog_internal(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, unsigned minus)
 {
-    U32 maxBitsSrc = BIT_highbit32((U32)(srcSize - 1)) - minus;
+    U32 maxBitsSrc = ZSTD_highbit32((U32)(srcSize - 1)) - minus;
     U32 tableLog = maxTableLog;
     U32 minBits = FSE_minTableLog(srcSize, maxSymbolValue);
     assert(srcSize > 1); /* Not supported, RLE should be used instead */
@@ -341,11 +372,10 @@ unsigned FSE_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxS
     return FSE_optimalTableLog_internal(maxTableLog, srcSize, maxSymbolValue, 2);
 }
 
-
 /* Secondary normalization method.
    To be used when primary method fails. */
 
-static size_t FSE_normalizeM2(short* norm, U32 tableLog, const unsigned* count, size_t total, U32 maxSymbolValue)
+static size_t FSE_normalizeM2(short* norm, U32 tableLog, const unsigned* count, size_t total, U32 maxSymbolValue, short lowProbCount)
 {
     short const NOT_YET_ASSIGNED = -2;
     U32 s;
@@ -362,7 +392,7 @@ static size_t FSE_normalizeM2(short* norm, U32 tableLog, const unsigned* count,
             continue;
         }
         if (count[s] <= lowThreshold) {
-            norm[s] = -1;
+            norm[s] = lowProbCount;
             distributed++;
             total -= count[s];
             continue;
@@ -414,7 +444,7 @@ static size_t FSE_normalizeM2(short* norm, U32 tableLog, const unsigned* count,
 
     {   U64 const vStepLog = 62 - tableLog;
         U64 const mid = (1ULL << (vStepLog-1)) - 1;
-        U64 const rStep = ((((U64)1<<vStepLog) * ToDistribute) + mid) / total;   /* scale on remaining */
+        U64 const rStep = ZSTD_div64((((U64)1<<vStepLog) * ToDistribute) + mid, (U32)total);   /* scale on remaining */
         U64 tmpTotal = mid;
         for (s=0; s<=maxSymbolValue; s++) {
             if (norm[s]==NOT_YET_ASSIGNED) {
@@ -431,10 +461,9 @@ static size_t FSE_normalizeM2(short* norm, U32 tableLog, const unsigned* count,
     return 0;
 }
 
-
 size_t FSE_normalizeCount (short* normalizedCounter, unsigned tableLog,
                            const unsigned* count, size_t total,
-                           unsigned maxSymbolValue)
+                           unsigned maxSymbolValue, unsigned useLowProbCount)
 {
     /* Sanity checks */
     if (tableLog==0) tableLog = FSE_DEFAULT_TABLELOG;
@@ -443,8 +472,9 @@ size_t FSE_normalizeCount (short* normalizedCounter, unsigned tableLog,
     if (tableLog < FSE_minTableLog(total, maxSymbolValue)) return ERROR(GENERIC);   /* Too small tableLog, compression potentially impossible */
 
     {   static U32 const rtbTable[] = {     0, 473195, 504333, 520860, 550000, 700000, 750000, 830000 };
+        short const lowProbCount = useLowProbCount ? -1 : 1;
         U64 const scale = 62 - tableLog;
-        U64 const step = ((U64)1<<62) / total;   /* <== here, one division ! */
+        U64 const step = ZSTD_div64((U64)1<<62, (U32)total);   /* <== here, one division ! */
         U64 const vStep = 1ULL<<(scale-20);
         int stillToDistribute = 1<<tableLog;
         unsigned s;
@@ -456,7 +486,7 @@ size_t FSE_normalizeCount (short* normalizedCounter, unsigned tableLog,
             if (count[s] == total) return 0;   /* rle special case */
             if (count[s] == 0) { normalizedCounter[s]=0; continue; }
             if (count[s] <= lowThreshold) {
-                normalizedCounter[s] = -1;
+                normalizedCounter[s] = lowProbCount;
                 stillToDistribute--;
             } else {
                 short proba = (short)((count[s]*step) >> scale);
@@ -470,7 +500,7 @@ size_t FSE_normalizeCount (short* normalizedCounter, unsigned tableLog,
         }   }
         if (-stillToDistribute >= (normalizedCounter[largest] >> 1)) {
             /* corner case, need another normalization method */
-            size_t const errorCode = FSE_normalizeM2(normalizedCounter, tableLog, count, total, maxSymbolValue);
+            size_t const errorCode = FSE_normalizeM2(normalizedCounter, tableLog, count, total, maxSymbolValue, lowProbCount);
             if (FSE_isError(errorCode)) return errorCode;
         }
         else normalizedCounter[largest] += (short)stillToDistribute;
@@ -493,40 +523,6 @@ size_t FSE_normalizeCount (short* normalizedCounter, unsigned tableLog,
     return tableLog;
 }
 
-
-/* fake FSE_CTable, for raw (uncompressed) input */
-size_t FSE_buildCTable_raw (FSE_CTable* ct, unsigned nbBits)
-{
-    const unsigned tableSize = 1 << nbBits;
-    const unsigned tableMask = tableSize - 1;
-    const unsigned maxSymbolValue = tableMask;
-    void* const ptr = ct;
-    U16* const tableU16 = ( (U16*) ptr) + 2;
-    void* const FSCT = ((U32*)ptr) + 1 /* header */ + (tableSize>>1);   /* assumption : tableLog >= 1 */
-    FSE_symbolCompressionTransform* const symbolTT = (FSE_symbolCompressionTransform*) (FSCT);
-    unsigned s;
-
-    /* Sanity checks */
-    if (nbBits < 1) return ERROR(GENERIC);             /* min size */
-
-    /* header */
-    tableU16[-2] = (U16) nbBits;
-    tableU16[-1] = (U16) maxSymbolValue;
-
-    /* Build table */
-    for (s=0; s<tableSize; s++)
-        tableU16[s] = (U16)(tableSize + s);
-
-    /* Build Symbol Transformation Table */
-    {   const U32 deltaNbBits = (nbBits << 16) - (1 << nbBits);
-        for (s=0; s<=maxSymbolValue; s++) {
-            symbolTT[s].deltaNbBits = deltaNbBits;
-            symbolTT[s].deltaFindState = s-1;
-    }   }
-
-    return 0;
-}
-
 /* fake FSE_CTable, for rle input (always same symbol) */
 size_t FSE_buildCTable_rle (FSE_CTable* ct, BYTE symbolValue)
 {
@@ -625,74 +621,4 @@ size_t FSE_compress_usingCTable (void* dst, size_t dstSize,
 
 size_t FSE_compressBound(size_t size) { return FSE_COMPRESSBOUND(size); }
 
-/* FSE_compress_wksp() :
- * Same as FSE_compress2(), but using an externally allocated scratch buffer (`workSpace`).
- * `wkspSize` size must be `(1<<tableLog)`.
- */
-size_t FSE_compress_wksp (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize)
-{
-    BYTE* const ostart = (BYTE*) dst;
-    BYTE* op = ostart;
-    BYTE* const oend = ostart + dstSize;
-
-    unsigned count[FSE_MAX_SYMBOL_VALUE+1];
-    S16   norm[FSE_MAX_SYMBOL_VALUE+1];
-    FSE_CTable* CTable = (FSE_CTable*)workSpace;
-    size_t const CTableSize = FSE_CTABLE_SIZE_U32(tableLog, maxSymbolValue);
-    void* scratchBuffer = (void*)(CTable + CTableSize);
-    size_t const scratchBufferSize = wkspSize - (CTableSize * sizeof(FSE_CTable));
-
-    /* init conditions */
-    if (wkspSize < FSE_WKSP_SIZE_U32(tableLog, maxSymbolValue)) return ERROR(tableLog_tooLarge);
-    if (srcSize <= 1) return 0;  /* Not compressible */
-    if (!maxSymbolValue) maxSymbolValue = FSE_MAX_SYMBOL_VALUE;
-    if (!tableLog) tableLog = FSE_DEFAULT_TABLELOG;
-
-    /* Scan input and build symbol stats */
-    {   CHECK_V_F(maxCount, HIST_count_wksp(count, &maxSymbolValue, src, srcSize, scratchBuffer, scratchBufferSize) );
-        if (maxCount == srcSize) return 1;   /* only a single symbol in src : rle */
-        if (maxCount == 1) return 0;         /* each symbol present maximum once => not compressible */
-        if (maxCount < (srcSize >> 7)) return 0;   /* Heuristic : not compressible enough */
-    }
-
-    tableLog = FSE_optimalTableLog(tableLog, srcSize, maxSymbolValue);
-    CHECK_F( FSE_normalizeCount(norm, tableLog, count, srcSize, maxSymbolValue) );
-
-    /* Write table description header */
-    {   CHECK_V_F(nc_err, FSE_writeNCount(op, oend-op, norm, maxSymbolValue, tableLog) );
-        op += nc_err;
-    }
-
-    /* Compress */
-    CHECK_F( FSE_buildCTable_wksp(CTable, norm, maxSymbolValue, tableLog, scratchBuffer, scratchBufferSize) );
-    {   CHECK_V_F(cSize, FSE_compress_usingCTable(op, oend - op, src, srcSize, CTable) );
-        if (cSize == 0) return 0;   /* not enough space for compressed data */
-        op += cSize;
-    }
-
-    /* check compressibility */
-    if ( (size_t)(op-ostart) >= srcSize-1 ) return 0;
-
-    return op-ostart;
-}
-
-typedef struct {
-    FSE_CTable CTable_max[FSE_CTABLE_SIZE_U32(FSE_MAX_TABLELOG, FSE_MAX_SYMBOL_VALUE)];
-    BYTE scratchBuffer[1 << FSE_MAX_TABLELOG];
-} fseWkspMax_t;
-
-size_t FSE_compress2 (void* dst, size_t dstCapacity, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog)
-{
-    fseWkspMax_t scratchBuffer;
-    DEBUG_STATIC_ASSERT(sizeof(scratchBuffer) >= FSE_WKSP_SIZE_U32(FSE_MAX_TABLELOG, FSE_MAX_SYMBOL_VALUE));   /* compilation failures here means scratchBuffer is not large enough */
-    if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);
-    return FSE_compress_wksp(dst, dstCapacity, src, srcSize, maxSymbolValue, tableLog, &scratchBuffer, sizeof(scratchBuffer));
-}
-
-size_t FSE_compress (void* dst, size_t dstCapacity, const void* src, size_t srcSize)
-{
-    return FSE_compress2(dst, dstCapacity, src, srcSize, FSE_MAX_SYMBOL_VALUE, FSE_DEFAULT_TABLELOG);
-}
-
-
 #endif   /* FSE_COMMONDEFS_ONLY */
diff --git a/contrib/zstd/fse_decompress.c b/contrib/zstd/fse_decompress.c
index bcc2223cc..1e1c9f92d 100644
--- a/contrib/zstd/fse_decompress.c
+++ b/contrib/zstd/fse_decompress.c
@@ -1,6 +1,6 @@
 /* ******************************************************************
  * FSE : Finite State Entropy decoder
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  *  You can contact the author at :
  *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -16,13 +16,15 @@
 /* **************************************************************
 *  Includes
 ****************************************************************/
-#include <stdlib.h>     /* malloc, free, qsort */
-#include <string.h>     /* memcpy, memset */
+#include "debug.h"      /* assert */
 #include "bitstream.h"
 #include "compiler.h"
 #define FSE_STATIC_LINKING_ONLY
 #include "fse.h"
 #include "error_private.h"
+#define ZSTD_DEPS_NEED_MALLOC
+#include "zstd_deps.h"
+#include "bits.h"       /* ZSTD_highbit32 */
 
 
 /* **************************************************************
@@ -54,30 +56,19 @@
 #define FSE_FUNCTION_NAME(X,Y) FSE_CAT(X,Y)
 #define FSE_TYPE_NAME(X,Y) FSE_CAT(X,Y)
 
-
-/* Function templates */
-FSE_DTable* FSE_createDTable (unsigned tableLog)
-{
-    if (tableLog > FSE_TABLELOG_ABSOLUTE_MAX) tableLog = FSE_TABLELOG_ABSOLUTE_MAX;
-    return (FSE_DTable*)malloc( FSE_DTABLE_SIZE_U32(tableLog) * sizeof (U32) );
-}
-
-void FSE_freeDTable (FSE_DTable* dt)
-{
-    free(dt);
-}
-
-size_t FSE_buildDTable(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog)
+static size_t FSE_buildDTable_internal(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize)
 {
     void* const tdPtr = dt+1;   /* because *dt is unsigned, 32-bits aligned on 32-bits */
     FSE_DECODE_TYPE* const tableDecode = (FSE_DECODE_TYPE*) (tdPtr);
-    U16 symbolNext[FSE_MAX_SYMBOL_VALUE+1];
+    U16* symbolNext = (U16*)workSpace;
+    BYTE* spread = (BYTE*)(symbolNext + maxSymbolValue + 1);
 
     U32 const maxSV1 = maxSymbolValue + 1;
     U32 const tableSize = 1 << tableLog;
     U32 highThreshold = tableSize-1;
 
     /* Sanity Checks */
+    if (FSE_BUILD_DTABLE_WKSP_SIZE(tableLog, maxSymbolValue) > wkspSize) return ERROR(maxSymbolValue_tooLarge);
     if (maxSymbolValue > FSE_MAX_SYMBOL_VALUE) return ERROR(maxSymbolValue_tooLarge);
     if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);
 
@@ -95,11 +86,57 @@ size_t FSE_buildDTable(FSE_DTable* dt, const short* normalizedCounter, unsigned
                     if (normalizedCounter[s] >= largeLimit) DTableH.fastMode=0;
                     symbolNext[s] = normalizedCounter[s];
         }   }   }
-        memcpy(dt, &DTableH, sizeof(DTableH));
+        ZSTD_memcpy(dt, &DTableH, sizeof(DTableH));
     }
 
     /* Spread symbols */
-    {   U32 const tableMask = tableSize-1;
+    if (highThreshold == tableSize - 1) {
+        size_t const tableMask = tableSize-1;
+        size_t const step = FSE_TABLESTEP(tableSize);
+        /* First lay down the symbols in order.
+         * We use a uint64_t to lay down 8 bytes at a time. This reduces branch
+         * misses since small blocks generally have small table logs, so nearly
+         * all symbols have counts <= 8. We ensure we have 8 bytes at the end of
+         * our buffer to handle the over-write.
+         */
+        {
+            U64 const add = 0x0101010101010101ull;
+            size_t pos = 0;
+            U64 sv = 0;
+            U32 s;
+            for (s=0; s<maxSV1; ++s, sv += add) {
+                int i;
+                int const n = normalizedCounter[s];
+                MEM_write64(spread + pos, sv);
+                for (i = 8; i < n; i += 8) {
+                    MEM_write64(spread + pos + i, sv);
+                }
+                pos += n;
+            }
+        }
+        /* Now we spread those positions across the table.
+         * The benefit of doing it in two stages is that we avoid the
+         * variable size inner loop, which caused lots of branch misses.
+         * Now we can run through all the positions without any branch misses.
+         * We unroll the loop twice, since that is what empirically worked best.
+         */
+        {
+            size_t position = 0;
+            size_t s;
+            size_t const unroll = 2;
+            assert(tableSize % unroll == 0); /* FSE_MIN_TABLELOG is 5 */
+            for (s = 0; s < (size_t)tableSize; s += unroll) {
+                size_t u;
+                for (u = 0; u < unroll; ++u) {
+                    size_t const uPosition = (position + (u * step)) & tableMask;
+                    tableDecode[uPosition].symbol = spread[s + u];
+                }
+                position = (position + (unroll * step)) & tableMask;
+            }
+            assert(position == 0);
+        }
+    } else {
+        U32 const tableMask = tableSize-1;
         U32 const step = FSE_TABLESTEP(tableSize);
         U32 s, position = 0;
         for (s=0; s<maxSV1; s++) {
@@ -117,62 +154,24 @@ size_t FSE_buildDTable(FSE_DTable* dt, const short* normalizedCounter, unsigned
         for (u=0; u<tableSize; u++) {
             FSE_FUNCTION_TYPE const symbol = (FSE_FUNCTION_TYPE)(tableDecode[u].symbol);
             U32 const nextState = symbolNext[symbol]++;
-            tableDecode[u].nbBits = (BYTE) (tableLog - BIT_highbit32(nextState) );
+            tableDecode[u].nbBits = (BYTE) (tableLog - ZSTD_highbit32(nextState) );
             tableDecode[u].newState = (U16) ( (nextState << tableDecode[u].nbBits) - tableSize);
     }   }
 
     return 0;
 }
 
+size_t FSE_buildDTable_wksp(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize)
+{
+    return FSE_buildDTable_internal(dt, normalizedCounter, maxSymbolValue, tableLog, workSpace, wkspSize);
+}
+
 
 #ifndef FSE_COMMONDEFS_ONLY
 
 /*-*******************************************************
 *  Decompression (Byte symbols)
 *********************************************************/
-size_t FSE_buildDTable_rle (FSE_DTable* dt, BYTE symbolValue)
-{
-    void* ptr = dt;
-    FSE_DTableHeader* const DTableH = (FSE_DTableHeader*)ptr;
-    void* dPtr = dt + 1;
-    FSE_decode_t* const cell = (FSE_decode_t*)dPtr;
-
-    DTableH->tableLog = 0;
-    DTableH->fastMode = 0;
-
-    cell->newState = 0;
-    cell->symbol = symbolValue;
-    cell->nbBits = 0;
-
-    return 0;
-}
-
-
-size_t FSE_buildDTable_raw (FSE_DTable* dt, unsigned nbBits)
-{
-    void* ptr = dt;
-    FSE_DTableHeader* const DTableH = (FSE_DTableHeader*)ptr;
-    void* dPtr = dt + 1;
-    FSE_decode_t* const dinfo = (FSE_decode_t*)dPtr;
-    const unsigned tableSize = 1 << nbBits;
-    const unsigned tableMask = tableSize - 1;
-    const unsigned maxSV1 = tableMask+1;
-    unsigned s;
-
-    /* Sanity checks */
-    if (nbBits < 1) return ERROR(GENERIC);         /* min size */
-
-    /* Build Decoding Table */
-    DTableH->tableLog = (U16)nbBits;
-    DTableH->fastMode = 1;
-    for (s=0; s<maxSV1; s++) {
-        dinfo[s].newState = 0;
-        dinfo[s].symbol = (BYTE)s;
-        dinfo[s].nbBits = (BYTE)nbBits;
-    }
-
-    return 0;
-}
 
 FORCE_INLINE_TEMPLATE size_t FSE_decompress_usingDTable_generic(
           void* dst, size_t maxDstSize,
@@ -236,51 +235,77 @@ FORCE_INLINE_TEMPLATE size_t FSE_decompress_usingDTable_generic(
     return op-ostart;
 }
 
+typedef struct {
+    short ncount[FSE_MAX_SYMBOL_VALUE + 1];
+    FSE_DTable dtable[1]; /* Dynamically sized */
+} FSE_DecompressWksp;
 
-size_t FSE_decompress_usingDTable(void* dst, size_t originalSize,
-                            const void* cSrc, size_t cSrcSize,
-                            const FSE_DTable* dt)
-{
-    const void* ptr = dt;
-    const FSE_DTableHeader* DTableH = (const FSE_DTableHeader*)ptr;
-    const U32 fastMode = DTableH->fastMode;
 
-    /* select fast mode (static) */
-    if (fastMode) return FSE_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 1);
-    return FSE_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 0);
-}
-
-
-size_t FSE_decompress_wksp(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, FSE_DTable* workSpace, unsigned maxLog)
+FORCE_INLINE_TEMPLATE size_t FSE_decompress_wksp_body(
+        void* dst, size_t dstCapacity,
+        const void* cSrc, size_t cSrcSize,
+        unsigned maxLog, void* workSpace, size_t wkspSize,
+        int bmi2)
 {
     const BYTE* const istart = (const BYTE*)cSrc;
     const BYTE* ip = istart;
-    short counting[FSE_MAX_SYMBOL_VALUE+1];
     unsigned tableLog;
     unsigned maxSymbolValue = FSE_MAX_SYMBOL_VALUE;
+    FSE_DecompressWksp* const wksp = (FSE_DecompressWksp*)workSpace;
+
+    DEBUG_STATIC_ASSERT((FSE_MAX_SYMBOL_VALUE + 1) % 2 == 0);
+    if (wkspSize < sizeof(*wksp)) return ERROR(GENERIC);
 
     /* normal FSE decoding mode */
-    size_t const NCountLength = FSE_readNCount (counting, &maxSymbolValue, &tableLog, istart, cSrcSize);
-    if (FSE_isError(NCountLength)) return NCountLength;
-    /* if (NCountLength >= cSrcSize) return ERROR(srcSize_wrong); */  /* too small input size; supposed to be already checked in NCountLength, only remaining case : NCountLength==cSrcSize */
-    if (tableLog > maxLog) return ERROR(tableLog_tooLarge);
-    ip += NCountLength;
-    cSrcSize -= NCountLength;
+    {
+        size_t const NCountLength = FSE_readNCount_bmi2(wksp->ncount, &maxSymbolValue, &tableLog, istart, cSrcSize, bmi2);
+        if (FSE_isError(NCountLength)) return NCountLength;
+        if (tableLog > maxLog) return ERROR(tableLog_tooLarge);
+        assert(NCountLength <= cSrcSize);
+        ip += NCountLength;
+        cSrcSize -= NCountLength;
+    }
 
-    CHECK_F( FSE_buildDTable (workSpace, counting, maxSymbolValue, tableLog) );
+    if (FSE_DECOMPRESS_WKSP_SIZE(tableLog, maxSymbolValue) > wkspSize) return ERROR(tableLog_tooLarge);
+    assert(sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog) <= wkspSize);
+    workSpace = (BYTE*)workSpace + sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog);
+    wkspSize -= sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog);
 
-    return FSE_decompress_usingDTable (dst, dstCapacity, ip, cSrcSize, workSpace);   /* always return, even if it is an error code */
-}
+    CHECK_F( FSE_buildDTable_internal(wksp->dtable, wksp->ncount, maxSymbolValue, tableLog, workSpace, wkspSize) );
 
+    {
+        const void* ptr = wksp->dtable;
+        const FSE_DTableHeader* DTableH = (const FSE_DTableHeader*)ptr;
+        const U32 fastMode = DTableH->fastMode;
 
-typedef FSE_DTable DTable_max_t[FSE_DTABLE_SIZE_U32(FSE_MAX_TABLELOG)];
+        /* select fast mode (static) */
+        if (fastMode) return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, wksp->dtable, 1);
+        return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, wksp->dtable, 0);
+    }
+}
 
-size_t FSE_decompress(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize)
+/* Avoids the FORCE_INLINE of the _body() function. */
+static size_t FSE_decompress_wksp_body_default(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize)
 {
-    DTable_max_t dt;   /* Static analyzer seems unable to understand this table will be properly initialized later */
-    return FSE_decompress_wksp(dst, dstCapacity, cSrc, cSrcSize, dt, FSE_MAX_TABLELOG);
+    return FSE_decompress_wksp_body(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize, 0);
 }
 
+#if DYNAMIC_BMI2
+BMI2_TARGET_ATTRIBUTE static size_t FSE_decompress_wksp_body_bmi2(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize)
+{
+    return FSE_decompress_wksp_body(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize, 1);
+}
+#endif
 
+size_t FSE_decompress_wksp_bmi2(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize, int bmi2)
+{
+#if DYNAMIC_BMI2
+    if (bmi2) {
+        return FSE_decompress_wksp_body_bmi2(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize);
+    }
+#endif
+    (void)bmi2;
+    return FSE_decompress_wksp_body_default(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize);
+}
 
 #endif   /* FSE_COMMONDEFS_ONLY */
diff --git a/contrib/zstd/hist.c b/contrib/zstd/hist.c
index c17b9725f..ebb1a1acf 100644
--- a/contrib/zstd/hist.c
+++ b/contrib/zstd/hist.c
@@ -1,7 +1,7 @@
 /* ******************************************************************
  * hist : Histogram functions
  * part of Finite State Entropy project
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  *  You can contact the author at :
  *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -34,7 +34,7 @@ unsigned HIST_count_simple(unsigned* count, unsigned* maxSymbolValuePtr,
     unsigned maxSymbolValue = *maxSymbolValuePtr;
     unsigned largestCount=0;
 
-    memset(count, 0, (maxSymbolValue+1) * sizeof(*count));
+    ZSTD_memset(count, 0, (maxSymbolValue+1) * sizeof(*count));
     if (srcSize==0) { *maxSymbolValuePtr = 0; return 0; }
 
     while (ip<end) {
@@ -60,9 +60,9 @@ typedef enum { trustInput, checkMaxSymbolValue } HIST_checkInput_e;
  * this design makes better use of OoO cpus,
  * and is noticeably faster when some values are heavily repeated.
  * But it needs some additional workspace for intermediate tables.
- * `workSpace` size must be a table of size >= HIST_WKSP_SIZE_U32.
+ * `workSpace` must be a U32 table of size >= HIST_WKSP_SIZE_U32.
  * @return : largest histogram frequency,
- *           or an error code (notably when histogram would be larger than *maxSymbolValuePtr). */
+ *           or an error code (notably when histogram's alphabet is larger than *maxSymbolValuePtr) */
 static size_t HIST_count_parallel_wksp(
                                 unsigned* count, unsigned* maxSymbolValuePtr,
                                 const void* source, size_t sourceSize,
@@ -71,22 +71,21 @@ static size_t HIST_count_parallel_wksp(
 {
     const BYTE* ip = (const BYTE*)source;
     const BYTE* const iend = ip+sourceSize;
-    unsigned maxSymbolValue = *maxSymbolValuePtr;
+    size_t const countSize = (*maxSymbolValuePtr + 1) * sizeof(*count);
     unsigned max=0;
     U32* const Counting1 = workSpace;
     U32* const Counting2 = Counting1 + 256;
     U32* const Counting3 = Counting2 + 256;
     U32* const Counting4 = Counting3 + 256;
 
-    memset(workSpace, 0, 4*256*sizeof(unsigned));
-
     /* safety checks */
+    assert(*maxSymbolValuePtr <= 255);
     if (!sourceSize) {
-        memset(count, 0, maxSymbolValue + 1);
+        ZSTD_memset(count, 0, countSize);
         *maxSymbolValuePtr = 0;
         return 0;
     }
-    if (!maxSymbolValue) maxSymbolValue = 255;            /* 0 == default */
+    ZSTD_memset(workSpace, 0, 4*256*sizeof(unsigned));
 
     /* by stripes of 16 bytes */
     {   U32 cached = MEM_read32(ip); ip += 4;
@@ -118,21 +117,18 @@ static size_t HIST_count_parallel_wksp(
     /* finish last symbols */
     while (ip<iend) Counting1[*ip++]++;
 
-    if (check) {   /* verify stats will fit into destination table */
-        U32 s; for (s=255; s>maxSymbolValue; s--) {
-            Counting1[s] += Counting2[s] + Counting3[s] + Counting4[s];
-            if (Counting1[s]) return ERROR(maxSymbolValue_tooSmall);
-    }   }
-
     {   U32 s;
-        if (maxSymbolValue > 255) maxSymbolValue = 255;
-        for (s=0; s<=maxSymbolValue; s++) {
-            count[s] = Counting1[s] + Counting2[s] + Counting3[s] + Counting4[s];
-            if (count[s] > max) max = count[s];
+        for (s=0; s<256; s++) {
+            Counting1[s] += Counting2[s] + Counting3[s] + Counting4[s];
+            if (Counting1[s] > max) max = Counting1[s];
     }   }
 
-    while (!count[maxSymbolValue]) maxSymbolValue--;
-    *maxSymbolValuePtr = maxSymbolValue;
+    {   unsigned maxSymbolValue = 255;
+        while (!Counting1[maxSymbolValue]) maxSymbolValue--;
+        if (check && maxSymbolValue > *maxSymbolValuePtr) return ERROR(maxSymbolValue_tooSmall);
+        *maxSymbolValuePtr = maxSymbolValue;
+        ZSTD_memmove(count, Counting1, countSize);   /* in case count & Counting1 are overlapping */
+    }
     return (size_t)max;
 }
 
@@ -152,14 +148,6 @@ size_t HIST_countFast_wksp(unsigned* count, unsigned* maxSymbolValuePtr,
     return HIST_count_parallel_wksp(count, maxSymbolValuePtr, source, sourceSize, trustInput, (U32*)workSpace);
 }
 
-/* fast variant (unsafe : won't check if src contains values beyond count[] limit) */
-size_t HIST_countFast(unsigned* count, unsigned* maxSymbolValuePtr,
-                     const void* source, size_t sourceSize)
-{
-    unsigned tmpCounters[HIST_WKSP_SIZE_U32];
-    return HIST_countFast_wksp(count, maxSymbolValuePtr, source, sourceSize, tmpCounters, sizeof(tmpCounters));
-}
-
 /* HIST_count_wksp() :
  * Same as HIST_count(), but using an externally provided scratch buffer.
  * `workSpace` size must be table of >= HIST_WKSP_SIZE_U32 unsigned */
@@ -175,9 +163,19 @@ size_t HIST_count_wksp(unsigned* count, unsigned* maxSymbolValuePtr,
     return HIST_countFast_wksp(count, maxSymbolValuePtr, source, sourceSize, workSpace, workSpaceSize);
 }
 
+#ifndef ZSTD_NO_UNUSED_FUNCTIONS
+/* fast variant (unsafe : won't check if src contains values beyond count[] limit) */
+size_t HIST_countFast(unsigned* count, unsigned* maxSymbolValuePtr,
+                     const void* source, size_t sourceSize)
+{
+    unsigned tmpCounters[HIST_WKSP_SIZE_U32];
+    return HIST_countFast_wksp(count, maxSymbolValuePtr, source, sourceSize, tmpCounters, sizeof(tmpCounters));
+}
+
 size_t HIST_count(unsigned* count, unsigned* maxSymbolValuePtr,
                  const void* src, size_t srcSize)
 {
     unsigned tmpCounters[HIST_WKSP_SIZE_U32];
     return HIST_count_wksp(count, maxSymbolValuePtr, src, srcSize, tmpCounters, sizeof(tmpCounters));
 }
+#endif
diff --git a/contrib/zstd/hist.h b/contrib/zstd/hist.h
index 77e3ec4fb..b8d5ba7c2 100644
--- a/contrib/zstd/hist.h
+++ b/contrib/zstd/hist.h
@@ -1,7 +1,7 @@
 /* ******************************************************************
  * hist : Histogram functions
  * part of Finite State Entropy project
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  *  You can contact the author at :
  *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -14,7 +14,7 @@
 ****************************************************************** */
 
 /* --- dependencies --- */
-#include <stddef.h>   /* size_t */
+#include "zstd_deps.h"   /* size_t */
 
 
 /* --- simple histogram functions --- */
diff --git a/contrib/zstd/huf.h b/contrib/zstd/huf.h
index ef432685d..73d1ee565 100644
--- a/contrib/zstd/huf.h
+++ b/contrib/zstd/huf.h
@@ -1,7 +1,7 @@
 /* ******************************************************************
  * huff0 huffman codec,
  * part of Finite State Entropy library
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -20,105 +20,30 @@ extern "C" {
 #define HUF_H_298734234
 
 /* *** Dependencies *** */
-#include <stddef.h>    /* size_t */
-
-
-/* *** library symbols visibility *** */
-/* Note : when linking with -fvisibility=hidden on gcc, or by default on Visual,
- *        HUF symbols remain "private" (internal symbols for library only).
- *        Set macro FSE_DLL_EXPORT to 1 if you want HUF symbols visible on DLL interface */
-#if defined(FSE_DLL_EXPORT) && (FSE_DLL_EXPORT==1) && defined(__GNUC__) && (__GNUC__ >= 4)
-#  define HUF_PUBLIC_API __attribute__ ((visibility ("default")))
-#elif defined(FSE_DLL_EXPORT) && (FSE_DLL_EXPORT==1)   /* Visual expected */
-#  define HUF_PUBLIC_API __declspec(dllexport)
-#elif defined(FSE_DLL_IMPORT) && (FSE_DLL_IMPORT==1)
-#  define HUF_PUBLIC_API __declspec(dllimport)  /* not required, just to generate faster code (saves a function pointer load from IAT and an indirect jump) */
-#else
-#  define HUF_PUBLIC_API
-#endif
-
-
-/* ========================== */
-/* ***  simple functions  *** */
-/* ========================== */
-
-/** HUF_compress() :
- *  Compress content from buffer 'src', of size 'srcSize', into buffer 'dst'.
- * 'dst' buffer must be already allocated.
- *  Compression runs faster if `dstCapacity` >= HUF_compressBound(srcSize).
- * `srcSize` must be <= `HUF_BLOCKSIZE_MAX` == 128 KB.
- * @return : size of compressed data (<= `dstCapacity`).
- *  Special values : if return == 0, srcData is not compressible => Nothing is stored within dst !!!
- *                   if HUF_isError(return), compression failed (more details using HUF_getErrorName())
- */
-HUF_PUBLIC_API size_t HUF_compress(void* dst, size_t dstCapacity,
-                             const void* src, size_t srcSize);
-
-/** HUF_decompress() :
- *  Decompress HUF data from buffer 'cSrc', of size 'cSrcSize',
- *  into already allocated buffer 'dst', of minimum size 'dstSize'.
- * `originalSize` : **must** be the ***exact*** size of original (uncompressed) data.
- *  Note : in contrast with FSE, HUF_decompress can regenerate
- *         RLE (cSrcSize==1) and uncompressed (cSrcSize==dstSize) data,
- *         because it knows size to regenerate (originalSize).
- * @return : size of regenerated data (== originalSize),
- *           or an error code, which can be tested using HUF_isError()
- */
-HUF_PUBLIC_API size_t HUF_decompress(void* dst,  size_t originalSize,
-                               const void* cSrc, size_t cSrcSize);
+#include "zstd_deps.h"    /* size_t */
+#include "mem.h"          /* U32 */
+#define FSE_STATIC_LINKING_ONLY
+#include "fse.h"
 
 
 /* ***   Tool functions *** */
-#define HUF_BLOCKSIZE_MAX (128 * 1024)                  /**< maximum input size for a single block compressed with HUF_compress */
-HUF_PUBLIC_API size_t HUF_compressBound(size_t size);   /**< maximum compressed size (worst case) */
+#define HUF_BLOCKSIZE_MAX (128 * 1024)   /**< maximum input size for a single block compressed with HUF_compress */
+size_t HUF_compressBound(size_t size);   /**< maximum compressed size (worst case) */
 
 /* Error Management */
-HUF_PUBLIC_API unsigned    HUF_isError(size_t code);       /**< tells if a return value is an error code */
-HUF_PUBLIC_API const char* HUF_getErrorName(size_t code);  /**< provides error code string (useful for debugging) */
-
-
-/* ***   Advanced function   *** */
-
-/** HUF_compress2() :
- *  Same as HUF_compress(), but offers control over `maxSymbolValue` and `tableLog`.
- * `maxSymbolValue` must be <= HUF_SYMBOLVALUE_MAX .
- * `tableLog` must be `<= HUF_TABLELOG_MAX` . */
-HUF_PUBLIC_API size_t HUF_compress2 (void* dst, size_t dstCapacity,
-                               const void* src, size_t srcSize,
-                               unsigned maxSymbolValue, unsigned tableLog);
-
-/** HUF_compress4X_wksp() :
- *  Same as HUF_compress2(), but uses externally allocated `workSpace`.
- * `workspace` must have minimum alignment of 4, and be at least as large as HUF_WORKSPACE_SIZE */
-#define HUF_WORKSPACE_SIZE ((6 << 10) + 256)
-#define HUF_WORKSPACE_SIZE_U32 (HUF_WORKSPACE_SIZE / sizeof(U32))
-HUF_PUBLIC_API size_t HUF_compress4X_wksp (void* dst, size_t dstCapacity,
-                                     const void* src, size_t srcSize,
-                                     unsigned maxSymbolValue, unsigned tableLog,
-                                     void* workSpace, size_t wkspSize);
+unsigned    HUF_isError(size_t code);       /**< tells if a return value is an error code */
+const char* HUF_getErrorName(size_t code);  /**< provides error code string (useful for debugging) */
 
-#endif   /* HUF_H_298734234 */
-
-/* ******************************************************************
- *  WARNING !!
- *  The following section contains advanced and experimental definitions
- *  which shall never be used in the context of a dynamic library,
- *  because they are not guaranteed to remain stable in the future.
- *  Only consider them in association with static linking.
- * *****************************************************************/
-#if defined(HUF_STATIC_LINKING_ONLY) && !defined(HUF_H_HUF_STATIC_LINKING_ONLY)
-#define HUF_H_HUF_STATIC_LINKING_ONLY
-
-/* *** Dependencies *** */
-#include "mem.h"   /* U32 */
 
+#define HUF_WORKSPACE_SIZE ((8 << 10) + 512 /* sorting scratch space */)
+#define HUF_WORKSPACE_SIZE_U64 (HUF_WORKSPACE_SIZE / sizeof(U64))
 
 /* *** Constants *** */
-#define HUF_TABLELOG_MAX      12      /* max runtime value of tableLog (due to static allocation); can be modified up to HUF_ABSOLUTEMAX_TABLELOG */
+#define HUF_TABLELOG_MAX      12      /* max runtime value of tableLog (due to static allocation); can be modified up to HUF_TABLELOG_ABSOLUTEMAX */
 #define HUF_TABLELOG_DEFAULT  11      /* default tableLog value when none specified */
 #define HUF_SYMBOLVALUE_MAX  255
 
-#define HUF_TABLELOG_ABSOLUTEMAX  15  /* absolute limit of HUF_MAX_TABLELOG. Beyond that value, code does not work */
+#define HUF_TABLELOG_ABSOLUTEMAX  12  /* absolute limit of HUF_MAX_TABLELOG. Beyond that value, code does not work */
 #if (HUF_TABLELOG_MAX > HUF_TABLELOG_ABSOLUTEMAX)
 #  error "HUF_TABLELOG_MAX is too large !"
 #endif
@@ -133,12 +58,12 @@ HUF_PUBLIC_API size_t HUF_compress4X_wksp (void* dst, size_t dstCapacity,
 #define HUF_COMPRESSBOUND(size) (HUF_CTABLEBOUND + HUF_BLOCKBOUND(size))   /* Macro version, useful for static allocation */
 
 /* static allocation of HUF's Compression Table */
-#define HUF_CTABLE_SIZE_U32(maxSymbolValue)   ((maxSymbolValue)+1)   /* Use tables of U32, for proper alignment */
-#define HUF_CTABLE_SIZE(maxSymbolValue)       (HUF_CTABLE_SIZE_U32(maxSymbolValue) * sizeof(U32))
+/* this is a private definition, just exposed for allocation and strict aliasing purpose. never EVER access its members directly */
+typedef size_t HUF_CElt;   /* consider it an incomplete type */
+#define HUF_CTABLE_SIZE_ST(maxSymbolValue)   ((maxSymbolValue)+2)   /* Use tables of size_t, for proper alignment */
+#define HUF_CTABLE_SIZE(maxSymbolValue)       (HUF_CTABLE_SIZE_ST(maxSymbolValue) * sizeof(size_t))
 #define HUF_CREATE_STATIC_CTABLE(name, maxSymbolValue) \
-    U32 name##hb[HUF_CTABLE_SIZE_U32(maxSymbolValue)]; \
-    void* name##hv = &(name##hb); \
-    HUF_CElt* name = (HUF_CElt*)(name##hv)   /* no final ; */
+    HUF_CElt name[HUF_CTABLE_SIZE_ST(maxSymbolValue)] /* no final ; */
 
 /* static allocation of HUF's DTable */
 typedef U32 HUF_DTable;
@@ -152,25 +77,49 @@ typedef U32 HUF_DTable;
 /* ****************************************
 *  Advanced decompression functions
 ******************************************/
-size_t HUF_decompress4X1 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< single-symbol decoder */
-#ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_decompress4X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< double-symbols decoder */
-#endif
 
-size_t HUF_decompress4X_DCtx (HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< decodes RLE and uncompressed */
-size_t HUF_decompress4X_hufOnly(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize); /**< considers RLE and uncompressed as errors */
-size_t HUF_decompress4X_hufOnly_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize); /**< considers RLE and uncompressed as errors */
-size_t HUF_decompress4X1_DCtx(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< single-symbol decoder */
-size_t HUF_decompress4X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize);   /**< single-symbol decoder */
-#ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_decompress4X2_DCtx(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< double-symbols decoder */
-size_t HUF_decompress4X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize);   /**< double-symbols decoder */
-#endif
+/**
+ * Huffman flags bitset.
+ * For all flags, 0 is the default value.
+ */
+typedef enum {
+    /**
+     * If compiled with DYNAMIC_BMI2: Set flag only if the CPU supports BMI2 at runtime.
+     * Otherwise: Ignored.
+     */
+    HUF_flags_bmi2 = (1 << 0),
+    /**
+     * If set: Test possible table depths to find the one that produces the smallest header + encoded size.
+     * If unset: Use heuristic to find the table depth.
+     */
+    HUF_flags_optimalDepth = (1 << 1),
+    /**
+     * If set: If the previous table can encode the input, always reuse the previous table.
+     * If unset: If the previous table can encode the input, reuse the previous table if it results in a smaller output.
+     */
+    HUF_flags_preferRepeat = (1 << 2),
+    /**
+     * If set: Sample the input and check if the sample is uncompressible, if it is then don't attempt to compress.
+     * If unset: Always histogram the entire input.
+     */
+    HUF_flags_suspectUncompressible = (1 << 3),
+    /**
+     * If set: Don't use assembly implementations
+     * If unset: Allow using assembly implementations
+     */
+    HUF_flags_disableAsm = (1 << 4),
+    /**
+     * If set: Don't use the fast decoding loop, always use the fallback decoding loop.
+     * If unset: Use the fast decoding loop when possible.
+     */
+    HUF_flags_disableFast = (1 << 5)
+} HUF_flags_e;
 
 
 /* ****************************************
  *  HUF detailed API
  * ****************************************/
+#define HUF_OPTIMAL_DEPTH_THRESHOLD ZSTD_btultra
 
 /*! HUF_compress() does the following:
  *  1. count symbol occurrence from source[] into table count[] using FSE_count() (exposed within "fse.h")
@@ -183,11 +132,12 @@ size_t HUF_decompress4X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
  *  For example, it's possible to compress several blocks using the same 'CTable',
  *  or to save and regenerate 'CTable' using external methods.
  */
-unsigned HUF_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue);
-typedef struct HUF_CElt_s HUF_CElt;   /* incomplete type */
-size_t HUF_buildCTable (HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue, unsigned maxNbBits);   /* @return : maxNbBits; CTable and count can overlap. In which case, CTable will overwrite count content */
-size_t HUF_writeCTable (void* dst, size_t maxDstSize, const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog);
-size_t HUF_compress4X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable);
+unsigned HUF_minTableLog(unsigned symbolCardinality);
+unsigned HUF_cardinality(const unsigned* count, unsigned maxSymbolValue);
+unsigned HUF_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, void* workSpace,
+ size_t wkspSize, HUF_CElt* table, const unsigned* count, int flags); /* table is used as scratch space for building and testing tables, not a return value */
+size_t HUF_writeCTable_wksp(void* dst, size_t maxDstSize, const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog, void* workspace, size_t workspaceSize);
+size_t HUF_compress4X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int flags);
 size_t HUF_estimateCompressedSize(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue);
 int HUF_validateCTable(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue);
 
@@ -196,22 +146,24 @@ typedef enum {
    HUF_repeat_check, /**< Can use the previous table but it must be checked. Note : The previous table must have been constructed by HUF_compress{1, 4}X_repeat */
    HUF_repeat_valid  /**< Can use the previous table and it is assumed to be valid */
  } HUF_repeat;
+
 /** HUF_compress4X_repeat() :
  *  Same as HUF_compress4X_wksp(), but considers using hufTable if *repeat != HUF_repeat_none.
  *  If it uses hufTable it does not modify hufTable or repeat.
  *  If it doesn't, it sets *repeat = HUF_repeat_none, and it sets hufTable to the table used.
- *  If preferRepeat then the old table will always be used if valid. */
+ *  If preferRepeat then the old table will always be used if valid.
+ *  If suspectUncompressible then some sampling checks will be run to potentially skip huffman coding */
 size_t HUF_compress4X_repeat(void* dst, size_t dstSize,
                        const void* src, size_t srcSize,
                        unsigned maxSymbolValue, unsigned tableLog,
                        void* workSpace, size_t wkspSize,    /**< `workSpace` must be aligned on 4-bytes boundaries, `wkspSize` must be >= HUF_WORKSPACE_SIZE */
-                       HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2);
+                       HUF_CElt* hufTable, HUF_repeat* repeat, int flags);
 
 /** HUF_buildCTable_wksp() :
  *  Same as HUF_buildCTable(), but using externally allocated scratch buffer.
  * `workSpace` must be aligned on 4-bytes boundaries, and its size must be >= HUF_CTABLE_WORKSPACE_SIZE.
  */
-#define HUF_CTABLE_WORKSPACE_SIZE_U32 (2*HUF_SYMBOLVALUE_MAX +1 +1)
+#define HUF_CTABLE_WORKSPACE_SIZE_U32 ((4 * (HUF_SYMBOLVALUE_MAX + 1)) + 192)
 #define HUF_CTABLE_WORKSPACE_SIZE (HUF_CTABLE_WORKSPACE_SIZE_U32 * sizeof(unsigned))
 size_t HUF_buildCTable_wksp (HUF_CElt* tree,
                        const unsigned* count, U32 maxSymbolValue, U32 maxNbBits,
@@ -226,15 +178,27 @@ size_t HUF_readStats(BYTE* huffWeight, size_t hwSize,
                      U32* rankStats, U32* nbSymbolsPtr, U32* tableLogPtr,
                      const void* src, size_t srcSize);
 
+/*! HUF_readStats_wksp() :
+ * Same as HUF_readStats() but takes an external workspace which must be
+ * 4-byte aligned and its size must be >= HUF_READ_STATS_WORKSPACE_SIZE.
+ * If the CPU has BMI2 support, pass bmi2=1, otherwise pass bmi2=0.
+ */
+#define HUF_READ_STATS_WORKSPACE_SIZE_U32 FSE_DECOMPRESS_WKSP_SIZE_U32(6, HUF_TABLELOG_MAX-1)
+#define HUF_READ_STATS_WORKSPACE_SIZE (HUF_READ_STATS_WORKSPACE_SIZE_U32 * sizeof(unsigned))
+size_t HUF_readStats_wksp(BYTE* huffWeight, size_t hwSize,
+                          U32* rankStats, U32* nbSymbolsPtr, U32* tableLogPtr,
+                          const void* src, size_t srcSize,
+                          void* workspace, size_t wkspSize,
+                          int flags);
+
 /** HUF_readCTable() :
  *  Loading a CTable saved with HUF_writeCTable() */
 size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void* src, size_t srcSize, unsigned *hasZeroWeights);
 
-/** HUF_getNbBits() :
+/** HUF_getNbBitsFromCTable() :
  *  Read nbBits from CTable symbolTable, for symbol `symbolValue` presumed <= HUF_SYMBOLVALUE_MAX
- *  Note 1 : is not inlined, as HUF_CElt definition is private
- *  Note 2 : const void* used, so that it can provide a statically allocated table as argument (which uses type U32) */
-U32 HUF_getNbBits(const void* symbolTable, U32 symbolValue);
+ *  Note 1 : is not inlined, as HUF_CElt definition is private */
+U32 HUF_getNbBitsFromCTable(const HUF_CElt* symbolTable, U32 symbolValue);
 
 /*
  * HUF_decompress() does the following:
@@ -260,80 +224,49 @@ U32 HUF_selectDecoder (size_t dstSize, size_t cSrcSize);
  *  a required workspace size greater than that specified in the following
  *  macro.
  */
-#define HUF_DECOMPRESS_WORKSPACE_SIZE (2 << 10)
+#define HUF_DECOMPRESS_WORKSPACE_SIZE ((2 << 10) + (1 << 9))
 #define HUF_DECOMPRESS_WORKSPACE_SIZE_U32 (HUF_DECOMPRESS_WORKSPACE_SIZE / sizeof(U32))
 
-#ifndef HUF_FORCE_DECOMPRESS_X2
-size_t HUF_readDTableX1 (HUF_DTable* DTable, const void* src, size_t srcSize);
-size_t HUF_readDTableX1_wksp (HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize);
-#endif
-#ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_readDTableX2 (HUF_DTable* DTable, const void* src, size_t srcSize);
-size_t HUF_readDTableX2_wksp (HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize);
-#endif
-
-size_t HUF_decompress4X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);
-#ifndef HUF_FORCE_DECOMPRESS_X2
-size_t HUF_decompress4X1_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);
-#endif
-#ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_decompress4X2_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);
-#endif
-
 
 /* ====================== */
 /* single stream variants */
 /* ====================== */
 
-size_t HUF_compress1X (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog);
-size_t HUF_compress1X_wksp (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);  /**< `workSpace` must be a table of at least HUF_WORKSPACE_SIZE_U32 unsigned */
-size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable);
+size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int flags);
 /** HUF_compress1X_repeat() :
  *  Same as HUF_compress1X_wksp(), but considers using hufTable if *repeat != HUF_repeat_none.
  *  If it uses hufTable it does not modify hufTable or repeat.
  *  If it doesn't, it sets *repeat = HUF_repeat_none, and it sets hufTable to the table used.
- *  If preferRepeat then the old table will always be used if valid. */
+ *  If preferRepeat then the old table will always be used if valid.
+ *  If suspectUncompressible then some sampling checks will be run to potentially skip huffman coding */
 size_t HUF_compress1X_repeat(void* dst, size_t dstSize,
                        const void* src, size_t srcSize,
                        unsigned maxSymbolValue, unsigned tableLog,
                        void* workSpace, size_t wkspSize,   /**< `workSpace` must be aligned on 4-bytes boundaries, `wkspSize` must be >= HUF_WORKSPACE_SIZE */
-                       HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2);
+                       HUF_CElt* hufTable, HUF_repeat* repeat, int flags);
 
-size_t HUF_decompress1X1 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* single-symbol decoder */
+size_t HUF_decompress1X_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags);
 #ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_decompress1X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* double-symbol decoder */
+size_t HUF_decompress1X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags);   /**< double-symbols decoder */
 #endif
 
-size_t HUF_decompress1X_DCtx (HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);
-size_t HUF_decompress1X_DCtx_wksp (HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize);
+/* BMI2 variants.
+ * If the CPU has BMI2 support, pass bmi2=1, otherwise pass bmi2=0.
+ */
+size_t HUF_decompress1X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int flags);
 #ifndef HUF_FORCE_DECOMPRESS_X2
-size_t HUF_decompress1X1_DCtx(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< single-symbol decoder */
-size_t HUF_decompress1X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize);   /**< single-symbol decoder */
-#endif
-#ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_decompress1X2_DCtx(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< double-symbols decoder */
-size_t HUF_decompress1X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize);   /**< double-symbols decoder */
+size_t HUF_decompress1X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags);
 #endif
-
-size_t HUF_decompress1X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);   /**< automatic selection of sing or double symbol decoder, based on DTable */
+size_t HUF_decompress4X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int flags);
+size_t HUF_decompress4X_hufOnly_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags);
 #ifndef HUF_FORCE_DECOMPRESS_X2
-size_t HUF_decompress1X1_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);
+size_t HUF_readDTableX1_wksp(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize, int flags);
 #endif
 #ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_decompress1X2_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);
+size_t HUF_readDTableX2_wksp(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize, int flags);
 #endif
 
-/* BMI2 variants.
- * If the CPU has BMI2 support, pass bmi2=1, otherwise pass bmi2=0.
- */
-size_t HUF_decompress1X_usingDTable_bmi2(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int bmi2);
-#ifndef HUF_FORCE_DECOMPRESS_X2
-size_t HUF_decompress1X1_DCtx_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int bmi2);
-#endif
-size_t HUF_decompress4X_usingDTable_bmi2(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int bmi2);
-size_t HUF_decompress4X_hufOnly_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int bmi2);
-
-#endif /* HUF_STATIC_LINKING_ONLY */
+#endif   /* HUF_H_298734234 */
 
 #if defined (__cplusplus)
 }
diff --git a/contrib/zstd/huf_compress.c b/contrib/zstd/huf_compress.c
index 4efffbb54..b5f09179a 100644
--- a/contrib/zstd/huf_compress.c
+++ b/contrib/zstd/huf_compress.c
@@ -1,6 +1,6 @@
 /* ******************************************************************
  * Huffman encoder, part of New Generation Entropy library
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  *  You can contact the author at :
  *  - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -23,16 +23,15 @@
 /* **************************************************************
 *  Includes
 ****************************************************************/
-#include <string.h>     /* memcpy, memset */
-#include <stdio.h>      /* printf (debug) */
+#include "zstd_deps.h"     /* ZSTD_memcpy, ZSTD_memset */
 #include "compiler.h"
 #include "bitstream.h"
 #include "hist.h"
 #define FSE_STATIC_LINKING_ONLY   /* FSE_optimalTableLog_internal */
 #include "fse.h"        /* header compression */
-#define HUF_STATIC_LINKING_ONLY
 #include "huf.h"
 #include "error_private.h"
+#include "bits.h"       /* ZSTD_highbit32 */
 
 
 /* **************************************************************
@@ -43,24 +42,111 @@
 
 
 /* **************************************************************
-*  Utils
+*  Required declarations
 ****************************************************************/
-unsigned HUF_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue)
+typedef struct nodeElt_s {
+    U32 count;
+    U16 parent;
+    BYTE byte;
+    BYTE nbBits;
+} nodeElt;
+
+
+/* **************************************************************
+*  Debug Traces
+****************************************************************/
+
+#if DEBUGLEVEL >= 2
+
+static size_t showU32(const U32* arr, size_t size)
+{
+    size_t u;
+    for (u=0; u<size; u++) {
+        RAWLOG(6, " %u", arr[u]); (void)arr;
+    }
+    RAWLOG(6, " \n");
+    return size;
+}
+
+static size_t HUF_getNbBits(HUF_CElt elt);
+
+static size_t showCTableBits(const HUF_CElt* ctable, size_t size)
+{
+    size_t u;
+    for (u=0; u<size; u++) {
+        RAWLOG(6, " %zu", HUF_getNbBits(ctable[u])); (void)ctable;
+    }
+    RAWLOG(6, " \n");
+    return size;
+
+}
+
+static size_t showHNodeSymbols(const nodeElt* hnode, size_t size)
 {
-    return FSE_optimalTableLog_internal(maxTableLog, srcSize, maxSymbolValue, 1);
+    size_t u;
+    for (u=0; u<size; u++) {
+        RAWLOG(6, " %u", hnode[u].byte); (void)hnode;
+    }
+    RAWLOG(6, " \n");
+    return size;
+}
+
+static size_t showHNodeBits(const nodeElt* hnode, size_t size)
+{
+    size_t u;
+    for (u=0; u<size; u++) {
+        RAWLOG(6, " %u", hnode[u].nbBits); (void)hnode;
+    }
+    RAWLOG(6, " \n");
+    return size;
 }
 
+#endif
+
 
 /* *******************************************************
 *  HUF : Huffman block compression
 *********************************************************/
+#define HUF_WORKSPACE_MAX_ALIGNMENT 8
+
+static void* HUF_alignUpWorkspace(void* workspace, size_t* workspaceSizePtr, size_t align)
+{
+    size_t const mask = align - 1;
+    size_t const rem = (size_t)workspace & mask;
+    size_t const add = (align - rem) & mask;
+    BYTE* const aligned = (BYTE*)workspace + add;
+    assert((align & (align - 1)) == 0); /* pow 2 */
+    assert(align <= HUF_WORKSPACE_MAX_ALIGNMENT);
+    if (*workspaceSizePtr >= add) {
+        assert(add < align);
+        assert(((size_t)aligned & mask) == 0);
+        *workspaceSizePtr -= add;
+        return aligned;
+    } else {
+        *workspaceSizePtr = 0;
+        return NULL;
+    }
+}
+
+
 /* HUF_compressWeights() :
  * Same as FSE_compress(), but dedicated to huff0's weights compression.
  * The use case needs much less stack memory.
  * Note : all elements within weightTable are supposed to be <= HUF_TABLELOG_MAX.
  */
 #define MAX_FSE_TABLELOG_FOR_HUFF_HEADER 6
-static size_t HUF_compressWeights (void* dst, size_t dstSize, const void* weightTable, size_t wtSize)
+
+typedef struct {
+    FSE_CTable CTable[FSE_CTABLE_SIZE_U32(MAX_FSE_TABLELOG_FOR_HUFF_HEADER, HUF_TABLELOG_MAX)];
+    U32 scratchBuffer[FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(HUF_TABLELOG_MAX, MAX_FSE_TABLELOG_FOR_HUFF_HEADER)];
+    unsigned count[HUF_TABLELOG_MAX+1];
+    S16 norm[HUF_TABLELOG_MAX+1];
+} HUF_CompressWeightsWksp;
+
+static size_t
+HUF_compressWeights(void* dst, size_t dstSize,
+              const void* weightTable, size_t wtSize,
+                    void* workspace, size_t workspaceSize)
 {
     BYTE* const ostart = (BYTE*) dst;
     BYTE* op = ostart;
@@ -68,33 +154,30 @@ static size_t HUF_compressWeights (void* dst, size_t dstSize, const void* weight
 
     unsigned maxSymbolValue = HUF_TABLELOG_MAX;
     U32 tableLog = MAX_FSE_TABLELOG_FOR_HUFF_HEADER;
+    HUF_CompressWeightsWksp* wksp = (HUF_CompressWeightsWksp*)HUF_alignUpWorkspace(workspace, &workspaceSize, ZSTD_ALIGNOF(U32));
 
-    FSE_CTable CTable[FSE_CTABLE_SIZE_U32(MAX_FSE_TABLELOG_FOR_HUFF_HEADER, HUF_TABLELOG_MAX)];
-    BYTE scratchBuffer[1<<MAX_FSE_TABLELOG_FOR_HUFF_HEADER];
-
-    unsigned count[HUF_TABLELOG_MAX+1];
-    S16 norm[HUF_TABLELOG_MAX+1];
+    if (workspaceSize < sizeof(HUF_CompressWeightsWksp)) return ERROR(GENERIC);
 
     /* init conditions */
     if (wtSize <= 1) return 0;  /* Not compressible */
 
     /* Scan input and build symbol stats */
-    {   unsigned const maxCount = HIST_count_simple(count, &maxSymbolValue, weightTable, wtSize);   /* never fails */
+    {   unsigned const maxCount = HIST_count_simple(wksp->count, &maxSymbolValue, weightTable, wtSize);   /* never fails */
         if (maxCount == wtSize) return 1;   /* only a single symbol in src : rle */
         if (maxCount == 1) return 0;        /* each symbol present maximum once => not compressible */
     }
 
     tableLog = FSE_optimalTableLog(tableLog, wtSize, maxSymbolValue);
-    CHECK_F( FSE_normalizeCount(norm, tableLog, count, wtSize, maxSymbolValue) );
+    CHECK_F( FSE_normalizeCount(wksp->norm, tableLog, wksp->count, wtSize, maxSymbolValue, /* useLowProbCount */ 0) );
 
     /* Write table description header */
-    {   CHECK_V_F(hSize, FSE_writeNCount(op, (size_t)(oend-op), norm, maxSymbolValue, tableLog) );
+    {   CHECK_V_F(hSize, FSE_writeNCount(op, (size_t)(oend-op), wksp->norm, maxSymbolValue, tableLog) );
         op += hSize;
     }
 
     /* Compress */
-    CHECK_F( FSE_buildCTable_wksp(CTable, norm, maxSymbolValue, tableLog, scratchBuffer, sizeof(scratchBuffer)) );
-    {   CHECK_V_F(cSize, FSE_compress_usingCTable(op, (size_t)(oend - op), weightTable, wtSize, CTable) );
+    CHECK_F( FSE_buildCTable_wksp(wksp->CTable, wksp->norm, maxSymbolValue, tableLog, wksp->scratchBuffer, sizeof(wksp->scratchBuffer)) );
+    {   CHECK_V_F(cSize, FSE_compress_usingCTable(op, (size_t)(oend - op), weightTable, wtSize, wksp->CTable) );
         if (cSize == 0) return 0;   /* not enough space for compressed data */
         op += cSize;
     }
@@ -102,35 +185,72 @@ static size_t HUF_compressWeights (void* dst, size_t dstSize, const void* weight
     return (size_t)(op-ostart);
 }
 
+static size_t HUF_getNbBits(HUF_CElt elt)
+{
+    return elt & 0xFF;
+}
+
+static size_t HUF_getNbBitsFast(HUF_CElt elt)
+{
+    return elt;
+}
+
+static size_t HUF_getValue(HUF_CElt elt)
+{
+    return elt & ~(size_t)0xFF;
+}
+
+static size_t HUF_getValueFast(HUF_CElt elt)
+{
+    return elt;
+}
 
-struct HUF_CElt_s {
-  U16  val;
-  BYTE nbBits;
-};   /* typedef'd to HUF_CElt within "huf.h" */
+static void HUF_setNbBits(HUF_CElt* elt, size_t nbBits)
+{
+    assert(nbBits <= HUF_TABLELOG_ABSOLUTEMAX);
+    *elt = nbBits;
+}
 
-/*! HUF_writeCTable() :
-    `CTable` : Huffman tree to save, using huf representation.
-    @return : size of saved CTable */
-size_t HUF_writeCTable (void* dst, size_t maxDstSize,
-                        const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog)
+static void HUF_setValue(HUF_CElt* elt, size_t value)
 {
+    size_t const nbBits = HUF_getNbBits(*elt);
+    if (nbBits > 0) {
+        assert((value >> nbBits) == 0);
+        *elt |= value << (sizeof(HUF_CElt) * 8 - nbBits);
+    }
+}
+
+typedef struct {
+    HUF_CompressWeightsWksp wksp;
     BYTE bitsToWeight[HUF_TABLELOG_MAX + 1];   /* precomputed conversion table */
     BYTE huffWeight[HUF_SYMBOLVALUE_MAX];
+} HUF_WriteCTableWksp;
+
+size_t HUF_writeCTable_wksp(void* dst, size_t maxDstSize,
+                            const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog,
+                            void* workspace, size_t workspaceSize)
+{
+    HUF_CElt const* const ct = CTable + 1;
     BYTE* op = (BYTE*)dst;
     U32 n;
+    HUF_WriteCTableWksp* wksp = (HUF_WriteCTableWksp*)HUF_alignUpWorkspace(workspace, &workspaceSize, ZSTD_ALIGNOF(U32));
+
+    HUF_STATIC_ASSERT(HUF_CTABLE_WORKSPACE_SIZE >= sizeof(HUF_WriteCTableWksp));
 
-     /* check conditions */
+    /* check conditions */
+    if (workspaceSize < sizeof(HUF_WriteCTableWksp)) return ERROR(GENERIC);
     if (maxSymbolValue > HUF_SYMBOLVALUE_MAX) return ERROR(maxSymbolValue_tooLarge);
 
     /* convert to weight */
-    bitsToWeight[0] = 0;
+    wksp->bitsToWeight[0] = 0;
     for (n=1; n<huffLog+1; n++)
-        bitsToWeight[n] = (BYTE)(huffLog + 1 - n);
+        wksp->bitsToWeight[n] = (BYTE)(huffLog + 1 - n);
     for (n=0; n<maxSymbolValue; n++)
-        huffWeight[n] = bitsToWeight[CTable[n].nbBits];
+        wksp->huffWeight[n] = wksp->bitsToWeight[HUF_getNbBits(ct[n])];
 
     /* attempt weights compression by FSE */
-    {   CHECK_V_F(hSize, HUF_compressWeights(op+1, maxDstSize-1, huffWeight, maxSymbolValue) );
+    if (maxDstSize < 1) return ERROR(dstSize_tooSmall);
+    {   CHECK_V_F(hSize, HUF_compressWeights(op+1, maxDstSize-1, wksp->huffWeight, maxSymbolValue, &wksp->wksp, sizeof(wksp->wksp)) );
         if ((hSize>1) & (hSize < maxSymbolValue/2)) {   /* FSE compressed */
             op[0] = (BYTE)hSize;
             return hSize+1;
@@ -140,9 +260,9 @@ size_t HUF_writeCTable (void* dst, size_t maxDstSize,
     if (maxSymbolValue > (256-128)) return ERROR(GENERIC);   /* should not happen : likely means source cannot be compressed */
     if (((maxSymbolValue+1)/2) + 1 > maxDstSize) return ERROR(dstSize_tooSmall);   /* not enough space within dst buffer */
     op[0] = (BYTE)(128 /*special case*/ + (maxSymbolValue-1));
-    huffWeight[maxSymbolValue] = 0;   /* to be sure it doesn't cause msan issue in final combination */
+    wksp->huffWeight[maxSymbolValue] = 0;   /* to be sure it doesn't cause msan issue in final combination */
     for (n=0; n<maxSymbolValue; n+=2)
-        op[(n/2)+1] = (BYTE)((huffWeight[n] << 4) + huffWeight[n+1]);
+        op[(n/2)+1] = (BYTE)((wksp->huffWeight[n] << 4) + wksp->huffWeight[n+1]);
     return ((maxSymbolValue+1)/2) + 1;
 }
 
@@ -153,34 +273,36 @@ size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void
     U32 rankVal[HUF_TABLELOG_ABSOLUTEMAX + 1];   /* large enough for values from 0 to 16 */
     U32 tableLog = 0;
     U32 nbSymbols = 0;
+    HUF_CElt* const ct = CTable + 1;
 
     /* get symbol weights */
     CHECK_V_F(readSize, HUF_readStats(huffWeight, HUF_SYMBOLVALUE_MAX+1, rankVal, &nbSymbols, &tableLog, src, srcSize));
+    *hasZeroWeights = (rankVal[0] > 0);
 
     /* check result */
     if (tableLog > HUF_TABLELOG_MAX) return ERROR(tableLog_tooLarge);
     if (nbSymbols > *maxSymbolValuePtr+1) return ERROR(maxSymbolValue_tooSmall);
 
+    CTable[0] = tableLog;
+
     /* Prepare base value per rank */
     {   U32 n, nextRankStart = 0;
         for (n=1; n<=tableLog; n++) {
-            U32 current = nextRankStart;
+            U32 curr = nextRankStart;
             nextRankStart += (rankVal[n] << (n-1));
-            rankVal[n] = current;
+            rankVal[n] = curr;
     }   }
 
     /* fill nbBits */
-    *hasZeroWeights = 0;
     {   U32 n; for (n=0; n<nbSymbols; n++) {
             const U32 w = huffWeight[n];
-            *hasZeroWeights |= (w == 0);
-            CTable[n].nbBits = (BYTE)(tableLog + 1 - w) & -(w != 0);
+            HUF_setNbBits(ct + n, (BYTE)(tableLog + 1 - w) & -(w != 0));
     }   }
 
     /* fill val */
     {   U16 nbPerRank[HUF_TABLELOG_MAX+2]  = {0};  /* support w=0=>n=tableLog+1 */
         U16 valPerRank[HUF_TABLELOG_MAX+2] = {0};
-        { U32 n; for (n=0; n<nbSymbols; n++) nbPerRank[CTable[n].nbBits]++; }
+        { U32 n; for (n=0; n<nbSymbols; n++) nbPerRank[HUF_getNbBits(ct[n])]++; }
         /* determine stating value per rank */
         valPerRank[tableLog+1] = 0;   /* for w==0 */
         {   U16 min = 0;
@@ -190,92 +312,150 @@ size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void
                 min >>= 1;
         }   }
         /* assign value within rank, symbol order */
-        { U32 n; for (n=0; n<nbSymbols; n++) CTable[n].val = valPerRank[CTable[n].nbBits]++; }
+        { U32 n; for (n=0; n<nbSymbols; n++) HUF_setValue(ct + n, valPerRank[HUF_getNbBits(ct[n])]++); }
     }
 
     *maxSymbolValuePtr = nbSymbols - 1;
     return readSize;
 }
 
-U32 HUF_getNbBits(const void* symbolTable, U32 symbolValue)
+U32 HUF_getNbBitsFromCTable(HUF_CElt const* CTable, U32 symbolValue)
 {
-    const HUF_CElt* table = (const HUF_CElt*)symbolTable;
+    const HUF_CElt* const ct = CTable + 1;
     assert(symbolValue <= HUF_SYMBOLVALUE_MAX);
-    return table[symbolValue].nbBits;
+    return (U32)HUF_getNbBits(ct[symbolValue]);
 }
 
 
-typedef struct nodeElt_s {
-    U32 count;
-    U16 parent;
-    BYTE byte;
-    BYTE nbBits;
-} nodeElt;
-
-static U32 HUF_setMaxHeight(nodeElt* huffNode, U32 lastNonNull, U32 maxNbBits)
+/**
+ * HUF_setMaxHeight():
+ * Try to enforce @targetNbBits on the Huffman tree described in @huffNode.
+ *
+ * It attempts to convert all nodes with nbBits > @targetNbBits
+ * to employ @targetNbBits instead. Then it adjusts the tree
+ * so that it remains a valid canonical Huffman tree.
+ *
+ * @pre               The sum of the ranks of each symbol == 2^largestBits,
+ *                    where largestBits == huffNode[lastNonNull].nbBits.
+ * @post              The sum of the ranks of each symbol == 2^largestBits,
+ *                    where largestBits is the return value (expected <= targetNbBits).
+ *
+ * @param huffNode    The Huffman tree modified in place to enforce targetNbBits.
+ *                    It's presumed sorted, from most frequent to rarest symbol.
+ * @param lastNonNull The symbol with the lowest count in the Huffman tree.
+ * @param targetNbBits  The allowed number of bits, which the Huffman tree
+ *                    may not respect. After this function the Huffman tree will
+ *                    respect targetNbBits.
+ * @return            The maximum number of bits of the Huffman tree after adjustment.
+ */
+static U32 HUF_setMaxHeight(nodeElt* huffNode, U32 lastNonNull, U32 targetNbBits)
 {
     const U32 largestBits = huffNode[lastNonNull].nbBits;
-    if (largestBits <= maxNbBits) return largestBits;   /* early exit : no elt > maxNbBits */
+    /* early exit : no elt > targetNbBits, so the tree is already valid. */
+    if (largestBits <= targetNbBits) return largestBits;
+
+    DEBUGLOG(5, "HUF_setMaxHeight (targetNbBits = %u)", targetNbBits);
 
     /* there are several too large elements (at least >= 2) */
     {   int totalCost = 0;
-        const U32 baseCost = 1 << (largestBits - maxNbBits);
+        const U32 baseCost = 1 << (largestBits - targetNbBits);
         int n = (int)lastNonNull;
 
-        while (huffNode[n].nbBits > maxNbBits) {
+        /* Adjust any ranks > targetNbBits to targetNbBits.
+         * Compute totalCost, which is how far the sum of the ranks is
+         * we are over 2^largestBits after adjust the offending ranks.
+         */
+        while (huffNode[n].nbBits > targetNbBits) {
             totalCost += baseCost - (1 << (largestBits - huffNode[n].nbBits));
-            huffNode[n].nbBits = (BYTE)maxNbBits;
-            n --;
-        }  /* n stops at huffNode[n].nbBits <= maxNbBits */
-        while (huffNode[n].nbBits == maxNbBits) n--;   /* n end at index of smallest symbol using < maxNbBits */
+            huffNode[n].nbBits = (BYTE)targetNbBits;
+            n--;
+        }
+        /* n stops at huffNode[n].nbBits <= targetNbBits */
+        assert(huffNode[n].nbBits <= targetNbBits);
+        /* n end at index of smallest symbol using < targetNbBits */
+        while (huffNode[n].nbBits == targetNbBits) --n;
 
-        /* renorm totalCost */
-        totalCost >>= (largestBits - maxNbBits);  /* note : totalCost is necessarily a multiple of baseCost */
+        /* renorm totalCost from 2^largestBits to 2^targetNbBits
+         * note : totalCost is necessarily a multiple of baseCost */
+        assert(((U32)totalCost & (baseCost - 1)) == 0);
+        totalCost >>= (largestBits - targetNbBits);
+        assert(totalCost > 0);
 
         /* repay normalized cost */
         {   U32 const noSymbol = 0xF0F0F0F0;
             U32 rankLast[HUF_TABLELOG_MAX+2];
 
-            /* Get pos of last (smallest) symbol per rank */
-            memset(rankLast, 0xF0, sizeof(rankLast));
-            {   U32 currentNbBits = maxNbBits;
+            /* Get pos of last (smallest = lowest cum. count) symbol per rank */
+            ZSTD_memset(rankLast, 0xF0, sizeof(rankLast));
+            {   U32 currentNbBits = targetNbBits;
                 int pos;
                 for (pos=n ; pos >= 0; pos--) {
                     if (huffNode[pos].nbBits >= currentNbBits) continue;
-                    currentNbBits = huffNode[pos].nbBits;   /* < maxNbBits */
-                    rankLast[maxNbBits-currentNbBits] = (U32)pos;
+                    currentNbBits = huffNode[pos].nbBits;   /* < targetNbBits */
+                    rankLast[targetNbBits-currentNbBits] = (U32)pos;
             }   }
 
             while (totalCost > 0) {
-                U32 nBitsToDecrease = BIT_highbit32((U32)totalCost) + 1;
+                /* Try to reduce the next power of 2 above totalCost because we
+                 * gain back half the rank.
+                 */
+                U32 nBitsToDecrease = ZSTD_highbit32((U32)totalCost) + 1;
                 for ( ; nBitsToDecrease > 1; nBitsToDecrease--) {
                     U32 const highPos = rankLast[nBitsToDecrease];
                     U32 const lowPos = rankLast[nBitsToDecrease-1];
                     if (highPos == noSymbol) continue;
+                    /* Decrease highPos if no symbols of lowPos or if it is
+                     * not cheaper to remove 2 lowPos than highPos.
+                     */
                     if (lowPos == noSymbol) break;
                     {   U32 const highTotal = huffNode[highPos].count;
                         U32 const lowTotal = 2 * huffNode[lowPos].count;
                         if (highTotal <= lowTotal) break;
                 }   }
                 /* only triggered when no more rank 1 symbol left => find closest one (note : there is necessarily at least one !) */
+                assert(rankLast[nBitsToDecrease] != noSymbol || nBitsToDecrease == 1);
                 /* HUF_MAX_TABLELOG test just to please gcc 5+; but it should not be necessary */
                 while ((nBitsToDecrease<=HUF_TABLELOG_MAX) && (rankLast[nBitsToDecrease] == noSymbol))
-                    nBitsToDecrease ++;
+                    nBitsToDecrease++;
+                assert(rankLast[nBitsToDecrease] != noSymbol);
+                /* Increase the number of bits to gain back half the rank cost. */
                 totalCost -= 1 << (nBitsToDecrease-1);
+                huffNode[rankLast[nBitsToDecrease]].nbBits++;
+
+                /* Fix up the new rank.
+                 * If the new rank was empty, this symbol is now its smallest.
+                 * Otherwise, this symbol will be the largest in the new rank so no adjustment.
+                 */
                 if (rankLast[nBitsToDecrease-1] == noSymbol)
-                    rankLast[nBitsToDecrease-1] = rankLast[nBitsToDecrease];   /* this rank is no longer empty */
-                huffNode[rankLast[nBitsToDecrease]].nbBits ++;
+                    rankLast[nBitsToDecrease-1] = rankLast[nBitsToDecrease];
+                /* Fix up the old rank.
+                 * If the symbol was at position 0, meaning it was the highest weight symbol in the tree,
+                 * it must be the only symbol in its rank, so the old rank now has no symbols.
+                 * Otherwise, since the Huffman nodes are sorted by count, the previous position is now
+                 * the smallest node in the rank. If the previous position belongs to a different rank,
+                 * then the rank is now empty.
+                 */
                 if (rankLast[nBitsToDecrease] == 0)    /* special case, reached largest symbol */
                     rankLast[nBitsToDecrease] = noSymbol;
                 else {
                     rankLast[nBitsToDecrease]--;
-                    if (huffNode[rankLast[nBitsToDecrease]].nbBits != maxNbBits-nBitsToDecrease)
+                    if (huffNode[rankLast[nBitsToDecrease]].nbBits != targetNbBits-nBitsToDecrease)
                         rankLast[nBitsToDecrease] = noSymbol;   /* this rank is now empty */
-            }   }   /* while (totalCost > 0) */
-
+                }
+            }   /* while (totalCost > 0) */
+
+            /* If we've removed too much weight, then we have to add it back.
+             * To avoid overshooting again, we only adjust the smallest rank.
+             * We take the largest nodes from the lowest rank 0 and move them
+             * to rank 1. There's guaranteed to be enough rank 0 symbols because
+             * TODO.
+             */
             while (totalCost < 0) {  /* Sometimes, cost correction overshoot */
-                if (rankLast[1] == noSymbol) {  /* special case : no rank 1 symbol (using maxNbBits-1); let's create one from largest rank 0 (using maxNbBits) */
-                    while (huffNode[n].nbBits == maxNbBits) n--;
+                /* special case : no rank 1 symbol (using targetNbBits-1);
+                 * let's create one from largest rank 0 (using targetNbBits).
+                 */
+                if (rankLast[1] == noSymbol) {
+                    while (huffNode[n].nbBits == targetNbBits) n--;
                     huffNode[n+1].nbBits--;
                     assert(n >= 0);
                     rankLast[1] = (U32)(n+1);
@@ -285,47 +465,178 @@ static U32 HUF_setMaxHeight(nodeElt* huffNode, U32 lastNonNull, U32 maxNbBits)
                 huffNode[ rankLast[1] + 1 ].nbBits--;
                 rankLast[1]++;
                 totalCost ++;
-    }   }   }   /* there are several too large elements (at least >= 2) */
+            }
+        }   /* repay normalized cost */
+    }   /* there are several too large elements (at least >= 2) */
 
-    return maxNbBits;
+    return targetNbBits;
 }
 
 typedef struct {
-    U32 base;
-    U32 current;
+    U16 base;
+    U16 curr;
 } rankPos;
 
-typedef nodeElt huffNodeTable[HUF_CTABLE_WORKSPACE_SIZE_U32];
+typedef nodeElt huffNodeTable[2 * (HUF_SYMBOLVALUE_MAX + 1)];
 
-#define RANK_POSITION_TABLE_SIZE 32
+/* Number of buckets available for HUF_sort() */
+#define RANK_POSITION_TABLE_SIZE 192
 
 typedef struct {
   huffNodeTable huffNodeTbl;
   rankPos rankPosition[RANK_POSITION_TABLE_SIZE];
 } HUF_buildCTable_wksp_tables;
 
-static void HUF_sort(nodeElt* huffNode, const unsigned* count, U32 maxSymbolValue, rankPos* rankPosition)
-{
+/* RANK_POSITION_DISTINCT_COUNT_CUTOFF == Cutoff point in HUF_sort() buckets for which we use log2 bucketing.
+ * Strategy is to use as many buckets as possible for representing distinct
+ * counts while using the remainder to represent all "large" counts.
+ *
+ * To satisfy this requirement for 192 buckets, we can do the following:
+ * Let buckets 0-166 represent distinct counts of [0, 166]
+ * Let buckets 166 to 192 represent all remaining counts up to RANK_POSITION_MAX_COUNT_LOG using log2 bucketing.
+ */
+#define RANK_POSITION_MAX_COUNT_LOG 32
+#define RANK_POSITION_LOG_BUCKETS_BEGIN ((RANK_POSITION_TABLE_SIZE - 1) - RANK_POSITION_MAX_COUNT_LOG - 1 /* == 158 */)
+#define RANK_POSITION_DISTINCT_COUNT_CUTOFF (RANK_POSITION_LOG_BUCKETS_BEGIN + ZSTD_highbit32(RANK_POSITION_LOG_BUCKETS_BEGIN) /* == 166 */)
+
+/* Return the appropriate bucket index for a given count. See definition of
+ * RANK_POSITION_DISTINCT_COUNT_CUTOFF for explanation of bucketing strategy.
+ */
+static U32 HUF_getIndex(U32 const count) {
+    return (count < RANK_POSITION_DISTINCT_COUNT_CUTOFF)
+        ? count
+        : ZSTD_highbit32(count) + RANK_POSITION_LOG_BUCKETS_BEGIN;
+}
+
+/* Helper swap function for HUF_quickSortPartition() */
+static void HUF_swapNodes(nodeElt* a, nodeElt* b) {
+	nodeElt tmp = *a;
+	*a = *b;
+	*b = tmp;
+}
+
+/* Returns 0 if the huffNode array is not sorted by descending count */
+MEM_STATIC int HUF_isSorted(nodeElt huffNode[], U32 const maxSymbolValue1) {
+    U32 i;
+    for (i = 1; i < maxSymbolValue1; ++i) {
+        if (huffNode[i].count > huffNode[i-1].count) {
+            return 0;
+        }
+    }
+    return 1;
+}
+
+/* Insertion sort by descending order */
+HINT_INLINE void HUF_insertionSort(nodeElt huffNode[], int const low, int const high) {
+    int i;
+    int const size = high-low+1;
+    huffNode += low;
+    for (i = 1; i < size; ++i) {
+        nodeElt const key = huffNode[i];
+        int j = i - 1;
+        while (j >= 0 && huffNode[j].count < key.count) {
+            huffNode[j + 1] = huffNode[j];
+            j--;
+        }
+        huffNode[j + 1] = key;
+    }
+}
+
+/* Pivot helper function for quicksort. */
+static int HUF_quickSortPartition(nodeElt arr[], int const low, int const high) {
+    /* Simply select rightmost element as pivot. "Better" selectors like
+     * median-of-three don't experimentally appear to have any benefit.
+     */
+    U32 const pivot = arr[high].count;
+    int i = low - 1;
+    int j = low;
+    for ( ; j < high; j++) {
+        if (arr[j].count > pivot) {
+            i++;
+            HUF_swapNodes(&arr[i], &arr[j]);
+        }
+    }
+    HUF_swapNodes(&arr[i + 1], &arr[high]);
+    return i + 1;
+}
+
+/* Classic quicksort by descending with partially iterative calls
+ * to reduce worst case callstack size.
+ */
+static void HUF_simpleQuickSort(nodeElt arr[], int low, int high) {
+    int const kInsertionSortThreshold = 8;
+    if (high - low < kInsertionSortThreshold) {
+        HUF_insertionSort(arr, low, high);
+        return;
+    }
+    while (low < high) {
+        int const idx = HUF_quickSortPartition(arr, low, high);
+        if (idx - low < high - idx) {
+            HUF_simpleQuickSort(arr, low, idx - 1);
+            low = idx + 1;
+        } else {
+            HUF_simpleQuickSort(arr, idx + 1, high);
+            high = idx - 1;
+        }
+    }
+}
+
+/**
+ * HUF_sort():
+ * Sorts the symbols [0, maxSymbolValue] by count[symbol] in decreasing order.
+ * This is a typical bucket sorting strategy that uses either quicksort or insertion sort to sort each bucket.
+ *
+ * @param[out] huffNode       Sorted symbols by decreasing count. Only members `.count` and `.byte` are filled.
+ *                            Must have (maxSymbolValue + 1) entries.
+ * @param[in]  count          Histogram of the symbols.
+ * @param[in]  maxSymbolValue Maximum symbol value.
+ * @param      rankPosition   This is a scratch workspace. Must have RANK_POSITION_TABLE_SIZE entries.
+ */
+static void HUF_sort(nodeElt huffNode[], const unsigned count[], U32 const maxSymbolValue, rankPos rankPosition[]) {
     U32 n;
+    U32 const maxSymbolValue1 = maxSymbolValue+1;
+
+    /* Compute base and set curr to base.
+     * For symbol s let lowerRank = HUF_getIndex(count[n]) and rank = lowerRank + 1.
+     * See HUF_getIndex to see bucketing strategy.
+     * We attribute each symbol to lowerRank's base value, because we want to know where
+     * each rank begins in the output, so for rank R we want to count ranks R+1 and above.
+     */
+    ZSTD_memset(rankPosition, 0, sizeof(*rankPosition) * RANK_POSITION_TABLE_SIZE);
+    for (n = 0; n < maxSymbolValue1; ++n) {
+        U32 lowerRank = HUF_getIndex(count[n]);
+        assert(lowerRank < RANK_POSITION_TABLE_SIZE - 1);
+        rankPosition[lowerRank].base++;
+    }
 
-    memset(rankPosition, 0, sizeof(*rankPosition) * RANK_POSITION_TABLE_SIZE);
-    for (n=0; n<=maxSymbolValue; n++) {
-        U32 r = BIT_highbit32(count[n] + 1);
-        rankPosition[r].base ++;
+    assert(rankPosition[RANK_POSITION_TABLE_SIZE - 1].base == 0);
+    /* Set up the rankPosition table */
+    for (n = RANK_POSITION_TABLE_SIZE - 1; n > 0; --n) {
+        rankPosition[n-1].base += rankPosition[n].base;
+        rankPosition[n-1].curr = rankPosition[n-1].base;
     }
-    for (n=30; n>0; n--) rankPosition[n-1].base += rankPosition[n].base;
-    for (n=0; n<32; n++) rankPosition[n].current = rankPosition[n].base;
-    for (n=0; n<=maxSymbolValue; n++) {
+
+    /* Insert each symbol into their appropriate bucket, setting up rankPosition table. */
+    for (n = 0; n < maxSymbolValue1; ++n) {
         U32 const c = count[n];
-        U32 const r = BIT_highbit32(c+1) + 1;
-        U32 pos = rankPosition[r].current++;
-        while ((pos > rankPosition[r].base) && (c > huffNode[pos-1].count)) {
-            huffNode[pos] = huffNode[pos-1];
-            pos--;
-        }
+        U32 const r = HUF_getIndex(c) + 1;
+        U32 const pos = rankPosition[r].curr++;
+        assert(pos < maxSymbolValue1);
         huffNode[pos].count = c;
         huffNode[pos].byte  = (BYTE)n;
     }
+
+    /* Sort each bucket. */
+    for (n = RANK_POSITION_DISTINCT_COUNT_CUTOFF; n < RANK_POSITION_TABLE_SIZE - 1; ++n) {
+        int const bucketSize = rankPosition[n].curr - rankPosition[n].base;
+        U32 const bucketStartIdx = rankPosition[n].base;
+        if (bucketSize > 1) {
+            assert(bucketStartIdx < maxSymbolValue1);
+            HUF_simpleQuickSort(huffNode + bucketStartIdx, 0, bucketSize-1);
+        }
+    }
+
+    assert(HUF_isSorted(huffNode, maxSymbolValue1));
 }
 
 
@@ -335,28 +646,21 @@ static void HUF_sort(nodeElt* huffNode, const unsigned* count, U32 maxSymbolValu
  */
 #define STARTNODE (HUF_SYMBOLVALUE_MAX+1)
 
-size_t HUF_buildCTable_wksp (HUF_CElt* tree, const unsigned* count, U32 maxSymbolValue, U32 maxNbBits, void* workSpace, size_t wkspSize)
+/* HUF_buildTree():
+ * Takes the huffNode array sorted by HUF_sort() and builds an unlimited-depth Huffman tree.
+ *
+ * @param huffNode        The array sorted by HUF_sort(). Builds the Huffman tree in this array.
+ * @param maxSymbolValue  The maximum symbol value.
+ * @return                The smallest node in the Huffman tree (by count).
+ */
+static int HUF_buildTree(nodeElt* huffNode, U32 maxSymbolValue)
 {
-    HUF_buildCTable_wksp_tables* const wksp_tables = (HUF_buildCTable_wksp_tables*)workSpace;
-    nodeElt* const huffNode0 = wksp_tables->huffNodeTbl;
-    nodeElt* const huffNode = huffNode0+1;
+    nodeElt* const huffNode0 = huffNode - 1;
     int nonNullRank;
     int lowS, lowN;
     int nodeNb = STARTNODE;
     int n, nodeRoot;
-
-    /* safety checks */
-    if (((size_t)workSpace & 3) != 0) return ERROR(GENERIC);  /* must be aligned on 4-bytes boundaries */
-    if (wkspSize < sizeof(HUF_buildCTable_wksp_tables))
-      return ERROR(workSpace_tooSmall);
-    if (maxNbBits == 0) maxNbBits = HUF_TABLELOG_DEFAULT;
-    if (maxSymbolValue > HUF_SYMBOLVALUE_MAX)
-      return ERROR(maxSymbolValue_tooLarge);
-    memset(huffNode0, 0, sizeof(huffNodeTable));
-
-    /* sort, decreasing order */
-    HUF_sort(huffNode, count, maxSymbolValue, wksp_tables->rankPosition);
-
+    DEBUGLOG(5, "HUF_buildTree (alphabet size = %u)", maxSymbolValue + 1);
     /* init for parents */
     nonNullRank = (int)maxSymbolValue;
     while(huffNode[nonNullRank].count == 0) nonNullRank--;
@@ -383,127 +687,406 @@ size_t HUF_buildCTable_wksp (HUF_CElt* tree, const unsigned* count, U32 maxSymbo
     for (n=0; n<=nonNullRank; n++)
         huffNode[n].nbBits = huffNode[ huffNode[n].parent ].nbBits + 1;
 
-    /* enforce maxTableLog */
-    maxNbBits = HUF_setMaxHeight(huffNode, (U32)nonNullRank, maxNbBits);
-
-    /* fill result into tree (val, nbBits) */
-    {   U16 nbPerRank[HUF_TABLELOG_MAX+1] = {0};
-        U16 valPerRank[HUF_TABLELOG_MAX+1] = {0};
-        int const alphabetSize = (int)(maxSymbolValue + 1);
-        if (maxNbBits > HUF_TABLELOG_MAX) return ERROR(GENERIC);   /* check fit into table */
-        for (n=0; n<=nonNullRank; n++)
-            nbPerRank[huffNode[n].nbBits]++;
-        /* determine stating value per rank */
-        {   U16 min = 0;
-            for (n=(int)maxNbBits; n>0; n--) {
-                valPerRank[n] = min;      /* get starting value within each rank */
-                min += nbPerRank[n];
-                min >>= 1;
-        }   }
-        for (n=0; n<alphabetSize; n++)
-            tree[huffNode[n].byte].nbBits = huffNode[n].nbBits;   /* push nbBits per symbol, symbol order */
-        for (n=0; n<alphabetSize; n++)
-            tree[n].val = valPerRank[tree[n].nbBits]++;   /* assign value within rank, symbol order */
-    }
+    DEBUGLOG(6, "Initial distribution of bits completed (%zu sorted symbols)", showHNodeBits(huffNode, maxSymbolValue+1));
 
-    return maxNbBits;
+    return nonNullRank;
 }
 
-/** HUF_buildCTable() :
- * @return : maxNbBits
- *  Note : count is used before tree is written, so they can safely overlap
+/**
+ * HUF_buildCTableFromTree():
+ * Build the CTable given the Huffman tree in huffNode.
+ *
+ * @param[out] CTable         The output Huffman CTable.
+ * @param      huffNode       The Huffman tree.
+ * @param      nonNullRank    The last and smallest node in the Huffman tree.
+ * @param      maxSymbolValue The maximum symbol value.
+ * @param      maxNbBits      The exact maximum number of bits used in the Huffman tree.
  */
-size_t HUF_buildCTable (HUF_CElt* tree, const unsigned* count, unsigned maxSymbolValue, unsigned maxNbBits)
+static void HUF_buildCTableFromTree(HUF_CElt* CTable, nodeElt const* huffNode, int nonNullRank, U32 maxSymbolValue, U32 maxNbBits)
 {
-    HUF_buildCTable_wksp_tables workspace;
-    return HUF_buildCTable_wksp(tree, count, maxSymbolValue, maxNbBits, &workspace, sizeof(workspace));
+    HUF_CElt* const ct = CTable + 1;
+    /* fill result into ctable (val, nbBits) */
+    int n;
+    U16 nbPerRank[HUF_TABLELOG_MAX+1] = {0};
+    U16 valPerRank[HUF_TABLELOG_MAX+1] = {0};
+    int const alphabetSize = (int)(maxSymbolValue + 1);
+    for (n=0; n<=nonNullRank; n++)
+        nbPerRank[huffNode[n].nbBits]++;
+    /* determine starting value per rank */
+    {   U16 min = 0;
+        for (n=(int)maxNbBits; n>0; n--) {
+            valPerRank[n] = min;      /* get starting value within each rank */
+            min += nbPerRank[n];
+            min >>= 1;
+    }   }
+    for (n=0; n<alphabetSize; n++)
+        HUF_setNbBits(ct + huffNode[n].byte, huffNode[n].nbBits);   /* push nbBits per symbol, symbol order */
+    for (n=0; n<alphabetSize; n++)
+        HUF_setValue(ct + n, valPerRank[HUF_getNbBits(ct[n])]++);   /* assign value within rank, symbol order */
+    CTable[0] = maxNbBits;
+}
+
+size_t
+HUF_buildCTable_wksp(HUF_CElt* CTable, const unsigned* count, U32 maxSymbolValue, U32 maxNbBits,
+                     void* workSpace, size_t wkspSize)
+{
+    HUF_buildCTable_wksp_tables* const wksp_tables =
+        (HUF_buildCTable_wksp_tables*)HUF_alignUpWorkspace(workSpace, &wkspSize, ZSTD_ALIGNOF(U32));
+    nodeElt* const huffNode0 = wksp_tables->huffNodeTbl;
+    nodeElt* const huffNode = huffNode0+1;
+    int nonNullRank;
+
+    HUF_STATIC_ASSERT(HUF_CTABLE_WORKSPACE_SIZE == sizeof(HUF_buildCTable_wksp_tables));
+
+    DEBUGLOG(5, "HUF_buildCTable_wksp (alphabet size = %u)", maxSymbolValue+1);
+
+    /* safety checks */
+    if (wkspSize < sizeof(HUF_buildCTable_wksp_tables))
+        return ERROR(workSpace_tooSmall);
+    if (maxNbBits == 0) maxNbBits = HUF_TABLELOG_DEFAULT;
+    if (maxSymbolValue > HUF_SYMBOLVALUE_MAX)
+        return ERROR(maxSymbolValue_tooLarge);
+    ZSTD_memset(huffNode0, 0, sizeof(huffNodeTable));
+
+    /* sort, decreasing order */
+    HUF_sort(huffNode, count, maxSymbolValue, wksp_tables->rankPosition);
+    DEBUGLOG(6, "sorted symbols completed (%zu symbols)", showHNodeSymbols(huffNode, maxSymbolValue+1));
+
+    /* build tree */
+    nonNullRank = HUF_buildTree(huffNode, maxSymbolValue);
+
+    /* determine and enforce maxTableLog */
+    maxNbBits = HUF_setMaxHeight(huffNode, (U32)nonNullRank, maxNbBits);
+    if (maxNbBits > HUF_TABLELOG_MAX) return ERROR(GENERIC);   /* check fit into table */
+
+    HUF_buildCTableFromTree(CTable, huffNode, nonNullRank, maxSymbolValue, maxNbBits);
+
+    return maxNbBits;
 }
 
 size_t HUF_estimateCompressedSize(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue)
 {
+    HUF_CElt const* ct = CTable + 1;
     size_t nbBits = 0;
     int s;
     for (s = 0; s <= (int)maxSymbolValue; ++s) {
-        nbBits += CTable[s].nbBits * count[s];
+        nbBits += HUF_getNbBits(ct[s]) * count[s];
     }
     return nbBits >> 3;
 }
 
 int HUF_validateCTable(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue) {
+  HUF_CElt const* ct = CTable + 1;
   int bad = 0;
   int s;
   for (s = 0; s <= (int)maxSymbolValue; ++s) {
-    bad |= (count[s] != 0) & (CTable[s].nbBits == 0);
+    bad |= (count[s] != 0) & (HUF_getNbBits(ct[s]) == 0);
   }
   return !bad;
 }
 
 size_t HUF_compressBound(size_t size) { return HUF_COMPRESSBOUND(size); }
 
+/** HUF_CStream_t:
+ * Huffman uses its own BIT_CStream_t implementation.
+ * There are three major differences from BIT_CStream_t:
+ *   1. HUF_addBits() takes a HUF_CElt (size_t) which is
+ *      the pair (nbBits, value) in the format:
+ *      format:
+ *        - Bits [0, 4)            = nbBits
+ *        - Bits [4, 64 - nbBits)  = 0
+ *        - Bits [64 - nbBits, 64) = value
+ *   2. The bitContainer is built from the upper bits and
+ *      right shifted. E.g. to add a new value of N bits
+ *      you right shift the bitContainer by N, then or in
+ *      the new value into the N upper bits.
+ *   3. The bitstream has two bit containers. You can add
+ *      bits to the second container and merge them into
+ *      the first container.
+ */
+
+#define HUF_BITS_IN_CONTAINER (sizeof(size_t) * 8)
+
+typedef struct {
+    size_t bitContainer[2];
+    size_t bitPos[2];
+
+    BYTE* startPtr;
+    BYTE* ptr;
+    BYTE* endPtr;
+} HUF_CStream_t;
+
+/**! HUF_initCStream():
+ * Initializes the bitstream.
+ * @returns 0 or an error code.
+ */
+static size_t HUF_initCStream(HUF_CStream_t* bitC,
+                                  void* startPtr, size_t dstCapacity)
+{
+    ZSTD_memset(bitC, 0, sizeof(*bitC));
+    bitC->startPtr = (BYTE*)startPtr;
+    bitC->ptr = bitC->startPtr;
+    bitC->endPtr = bitC->startPtr + dstCapacity - sizeof(bitC->bitContainer[0]);
+    if (dstCapacity <= sizeof(bitC->bitContainer[0])) return ERROR(dstSize_tooSmall);
+    return 0;
+}
+
+/*! HUF_addBits():
+ * Adds the symbol stored in HUF_CElt elt to the bitstream.
+ *
+ * @param elt   The element we're adding. This is a (nbBits, value) pair.
+ *              See the HUF_CStream_t docs for the format.
+ * @param idx   Insert into the bitstream at this idx.
+ * @param kFast This is a template parameter. If the bitstream is guaranteed
+ *              to have at least 4 unused bits after this call it may be 1,
+ *              otherwise it must be 0. HUF_addBits() is faster when fast is set.
+ */
+FORCE_INLINE_TEMPLATE void HUF_addBits(HUF_CStream_t* bitC, HUF_CElt elt, int idx, int kFast)
+{
+    assert(idx <= 1);
+    assert(HUF_getNbBits(elt) <= HUF_TABLELOG_ABSOLUTEMAX);
+    /* This is efficient on x86-64 with BMI2 because shrx
+     * only reads the low 6 bits of the register. The compiler
+     * knows this and elides the mask. When fast is set,
+     * every operation can use the same value loaded from elt.
+     */
+    bitC->bitContainer[idx] >>= HUF_getNbBits(elt);
+    bitC->bitContainer[idx] |= kFast ? HUF_getValueFast(elt) : HUF_getValue(elt);
+    /* We only read the low 8 bits of bitC->bitPos[idx] so it
+     * doesn't matter that the high bits have noise from the value.
+     */
+    bitC->bitPos[idx] += HUF_getNbBitsFast(elt);
+    assert((bitC->bitPos[idx] & 0xFF) <= HUF_BITS_IN_CONTAINER);
+    /* The last 4-bits of elt are dirty if fast is set,
+     * so we must not be overwriting bits that have already been
+     * inserted into the bit container.
+     */
+#if DEBUGLEVEL >= 1
+    {
+        size_t const nbBits = HUF_getNbBits(elt);
+        size_t const dirtyBits = nbBits == 0 ? 0 : ZSTD_highbit32((U32)nbBits) + 1;
+        (void)dirtyBits;
+        /* Middle bits are 0. */
+        assert(((elt >> dirtyBits) << (dirtyBits + nbBits)) == 0);
+        /* We didn't overwrite any bits in the bit container. */
+        assert(!kFast || (bitC->bitPos[idx] & 0xFF) <= HUF_BITS_IN_CONTAINER);
+        (void)dirtyBits;
+    }
+#endif
+}
+
+FORCE_INLINE_TEMPLATE void HUF_zeroIndex1(HUF_CStream_t* bitC)
+{
+    bitC->bitContainer[1] = 0;
+    bitC->bitPos[1] = 0;
+}
+
+/*! HUF_mergeIndex1() :
+ * Merges the bit container @ index 1 into the bit container @ index 0
+ * and zeros the bit container @ index 1.
+ */
+FORCE_INLINE_TEMPLATE void HUF_mergeIndex1(HUF_CStream_t* bitC)
+{
+    assert((bitC->bitPos[1] & 0xFF) < HUF_BITS_IN_CONTAINER);
+    bitC->bitContainer[0] >>= (bitC->bitPos[1] & 0xFF);
+    bitC->bitContainer[0] |= bitC->bitContainer[1];
+    bitC->bitPos[0] += bitC->bitPos[1];
+    assert((bitC->bitPos[0] & 0xFF) <= HUF_BITS_IN_CONTAINER);
+}
+
+/*! HUF_flushBits() :
+* Flushes the bits in the bit container @ index 0.
+*
+* @post bitPos will be < 8.
+* @param kFast If kFast is set then we must know a-priori that
+*              the bit container will not overflow.
+*/
+FORCE_INLINE_TEMPLATE void HUF_flushBits(HUF_CStream_t* bitC, int kFast)
+{
+    /* The upper bits of bitPos are noisy, so we must mask by 0xFF. */
+    size_t const nbBits = bitC->bitPos[0] & 0xFF;
+    size_t const nbBytes = nbBits >> 3;
+    /* The top nbBits bits of bitContainer are the ones we need. */
+    size_t const bitContainer = bitC->bitContainer[0] >> (HUF_BITS_IN_CONTAINER - nbBits);
+    /* Mask bitPos to account for the bytes we consumed. */
+    bitC->bitPos[0] &= 7;
+    assert(nbBits > 0);
+    assert(nbBits <= sizeof(bitC->bitContainer[0]) * 8);
+    assert(bitC->ptr <= bitC->endPtr);
+    MEM_writeLEST(bitC->ptr, bitContainer);
+    bitC->ptr += nbBytes;
+    assert(!kFast || bitC->ptr <= bitC->endPtr);
+    if (!kFast && bitC->ptr > bitC->endPtr) bitC->ptr = bitC->endPtr;
+    /* bitContainer doesn't need to be modified because the leftover
+     * bits are already the top bitPos bits. And we don't care about
+     * noise in the lower values.
+     */
+}
+
+/*! HUF_endMark()
+ * @returns The Huffman stream end mark: A 1-bit value = 1.
+ */
+static HUF_CElt HUF_endMark(void)
+{
+    HUF_CElt endMark;
+    HUF_setNbBits(&endMark, 1);
+    HUF_setValue(&endMark, 1);
+    return endMark;
+}
+
+/*! HUF_closeCStream() :
+ *  @return Size of CStream, in bytes,
+ *          or 0 if it could not fit into dstBuffer */
+static size_t HUF_closeCStream(HUF_CStream_t* bitC)
+{
+    HUF_addBits(bitC, HUF_endMark(), /* idx */ 0, /* kFast */ 0);
+    HUF_flushBits(bitC, /* kFast */ 0);
+    {
+        size_t const nbBits = bitC->bitPos[0] & 0xFF;
+        if (bitC->ptr >= bitC->endPtr) return 0; /* overflow detected */
+        return (size_t)(bitC->ptr - bitC->startPtr) + (nbBits > 0);
+    }
+}
+
 FORCE_INLINE_TEMPLATE void
-HUF_encodeSymbol(BIT_CStream_t* bitCPtr, U32 symbol, const HUF_CElt* CTable)
+HUF_encodeSymbol(HUF_CStream_t* bitCPtr, U32 symbol, const HUF_CElt* CTable, int idx, int fast)
 {
-    BIT_addBitsFast(bitCPtr, CTable[symbol].val, CTable[symbol].nbBits);
+    HUF_addBits(bitCPtr, CTable[symbol], idx, fast);
 }
 
-#define HUF_FLUSHBITS(s)  BIT_flushBits(s)
+FORCE_INLINE_TEMPLATE void
+HUF_compress1X_usingCTable_internal_body_loop(HUF_CStream_t* bitC,
+                                   const BYTE* ip, size_t srcSize,
+                                   const HUF_CElt* ct,
+                                   int kUnroll, int kFastFlush, int kLastFast)
+{
+    /* Join to kUnroll */
+    int n = (int)srcSize;
+    int rem = n % kUnroll;
+    if (rem > 0) {
+        for (; rem > 0; --rem) {
+            HUF_encodeSymbol(bitC, ip[--n], ct, 0, /* fast */ 0);
+        }
+        HUF_flushBits(bitC, kFastFlush);
+    }
+    assert(n % kUnroll == 0);
+
+    /* Join to 2 * kUnroll */
+    if (n % (2 * kUnroll)) {
+        int u;
+        for (u = 1; u < kUnroll; ++u) {
+            HUF_encodeSymbol(bitC, ip[n - u], ct, 0, 1);
+        }
+        HUF_encodeSymbol(bitC, ip[n - kUnroll], ct, 0, kLastFast);
+        HUF_flushBits(bitC, kFastFlush);
+        n -= kUnroll;
+    }
+    assert(n % (2 * kUnroll) == 0);
 
-#define HUF_FLUSHBITS_1(stream) \
-    if (sizeof((stream)->bitContainer)*8 < HUF_TABLELOG_MAX*2+7) HUF_FLUSHBITS(stream)
+    for (; n>0; n-= 2 * kUnroll) {
+        /* Encode kUnroll symbols into the bitstream @ index 0. */
+        int u;
+        for (u = 1; u < kUnroll; ++u) {
+            HUF_encodeSymbol(bitC, ip[n - u], ct, /* idx */ 0, /* fast */ 1);
+        }
+        HUF_encodeSymbol(bitC, ip[n - kUnroll], ct, /* idx */ 0, /* fast */ kLastFast);
+        HUF_flushBits(bitC, kFastFlush);
+        /* Encode kUnroll symbols into the bitstream @ index 1.
+         * This allows us to start filling the bit container
+         * without any data dependencies.
+         */
+        HUF_zeroIndex1(bitC);
+        for (u = 1; u < kUnroll; ++u) {
+            HUF_encodeSymbol(bitC, ip[n - kUnroll - u], ct, /* idx */ 1, /* fast */ 1);
+        }
+        HUF_encodeSymbol(bitC, ip[n - kUnroll - kUnroll], ct, /* idx */ 1, /* fast */ kLastFast);
+        /* Merge bitstream @ index 1 into the bitstream @ index 0 */
+        HUF_mergeIndex1(bitC);
+        HUF_flushBits(bitC, kFastFlush);
+    }
+    assert(n == 0);
+
+}
+
+/**
+ * Returns a tight upper bound on the output space needed by Huffman
+ * with 8 bytes buffer to handle over-writes. If the output is at least
+ * this large we don't need to do bounds checks during Huffman encoding.
+ */
+static size_t HUF_tightCompressBound(size_t srcSize, size_t tableLog)
+{
+    return ((srcSize * tableLog) >> 3) + 8;
+}
 
-#define HUF_FLUSHBITS_2(stream) \
-    if (sizeof((stream)->bitContainer)*8 < HUF_TABLELOG_MAX*4+7) HUF_FLUSHBITS(stream)
 
 FORCE_INLINE_TEMPLATE size_t
 HUF_compress1X_usingCTable_internal_body(void* dst, size_t dstSize,
                                    const void* src, size_t srcSize,
                                    const HUF_CElt* CTable)
 {
+    U32 const tableLog = (U32)CTable[0];
+    HUF_CElt const* ct = CTable + 1;
     const BYTE* ip = (const BYTE*) src;
     BYTE* const ostart = (BYTE*)dst;
     BYTE* const oend = ostart + dstSize;
     BYTE* op = ostart;
-    size_t n;
-    BIT_CStream_t bitC;
+    HUF_CStream_t bitC;
 
     /* init */
     if (dstSize < 8) return 0;   /* not enough space to compress */
-    { size_t const initErr = BIT_initCStream(&bitC, op, (size_t)(oend-op));
+    { size_t const initErr = HUF_initCStream(&bitC, op, (size_t)(oend-op));
       if (HUF_isError(initErr)) return 0; }
 
-    n = srcSize & ~3;  /* join to mod 4 */
-    switch (srcSize & 3)
-    {
-        case 3 : HUF_encodeSymbol(&bitC, ip[n+ 2], CTable);
-                 HUF_FLUSHBITS_2(&bitC);
-		 /* fall-through */
-        case 2 : HUF_encodeSymbol(&bitC, ip[n+ 1], CTable);
-                 HUF_FLUSHBITS_1(&bitC);
-		 /* fall-through */
-        case 1 : HUF_encodeSymbol(&bitC, ip[n+ 0], CTable);
-                 HUF_FLUSHBITS(&bitC);
-		 /* fall-through */
-        case 0 : /* fall-through */
-        default: break;
-    }
-
-    for (; n>0; n-=4) {  /* note : n&3==0 at this stage */
-        HUF_encodeSymbol(&bitC, ip[n- 1], CTable);
-        HUF_FLUSHBITS_1(&bitC);
-        HUF_encodeSymbol(&bitC, ip[n- 2], CTable);
-        HUF_FLUSHBITS_2(&bitC);
-        HUF_encodeSymbol(&bitC, ip[n- 3], CTable);
-        HUF_FLUSHBITS_1(&bitC);
-        HUF_encodeSymbol(&bitC, ip[n- 4], CTable);
-        HUF_FLUSHBITS(&bitC);
+    if (dstSize < HUF_tightCompressBound(srcSize, (size_t)tableLog) || tableLog > 11)
+        HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ MEM_32bits() ? 2 : 4, /* kFast */ 0, /* kLastFast */ 0);
+    else {
+        if (MEM_32bits()) {
+            switch (tableLog) {
+            case 11:
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 2, /* kFastFlush */ 1, /* kLastFast */ 0);
+                break;
+            case 10: ZSTD_FALLTHROUGH;
+            case 9: ZSTD_FALLTHROUGH;
+            case 8:
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 2, /* kFastFlush */ 1, /* kLastFast */ 1);
+                break;
+            case 7: ZSTD_FALLTHROUGH;
+            default:
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 3, /* kFastFlush */ 1, /* kLastFast */ 1);
+                break;
+            }
+        } else {
+            switch (tableLog) {
+            case 11:
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 5, /* kFastFlush */ 1, /* kLastFast */ 0);
+                break;
+            case 10:
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 5, /* kFastFlush */ 1, /* kLastFast */ 1);
+                break;
+            case 9:
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 6, /* kFastFlush */ 1, /* kLastFast */ 0);
+                break;
+            case 8:
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 7, /* kFastFlush */ 1, /* kLastFast */ 0);
+                break;
+            case 7:
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 8, /* kFastFlush */ 1, /* kLastFast */ 0);
+                break;
+            case 6: ZSTD_FALLTHROUGH;
+            default:
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 9, /* kFastFlush */ 1, /* kLastFast */ 1);
+                break;
+            }
+        }
     }
+    assert(bitC.ptr <= bitC.endPtr);
 
-    return BIT_closeCStream(&bitC);
+    return HUF_closeCStream(&bitC);
 }
 
 #if DYNAMIC_BMI2
 
-static TARGET_ATTRIBUTE("bmi2") size_t
+static BMI2_TARGET_ATTRIBUTE size_t
 HUF_compress1X_usingCTable_internal_bmi2(void* dst, size_t dstSize,
                                    const void* src, size_t srcSize,
                                    const HUF_CElt* CTable)
@@ -522,9 +1105,9 @@ HUF_compress1X_usingCTable_internal_default(void* dst, size_t dstSize,
 static size_t
 HUF_compress1X_usingCTable_internal(void* dst, size_t dstSize,
                               const void* src, size_t srcSize,
-                              const HUF_CElt* CTable, const int bmi2)
+                              const HUF_CElt* CTable, const int flags)
 {
-    if (bmi2) {
+    if (flags & HUF_flags_bmi2) {
         return HUF_compress1X_usingCTable_internal_bmi2(dst, dstSize, src, srcSize, CTable);
     }
     return HUF_compress1X_usingCTable_internal_default(dst, dstSize, src, srcSize, CTable);
@@ -535,24 +1118,23 @@ HUF_compress1X_usingCTable_internal(void* dst, size_t dstSize,
 static size_t
 HUF_compress1X_usingCTable_internal(void* dst, size_t dstSize,
                               const void* src, size_t srcSize,
-                              const HUF_CElt* CTable, const int bmi2)
+                              const HUF_CElt* CTable, const int flags)
 {
-    (void)bmi2;
+    (void)flags;
     return HUF_compress1X_usingCTable_internal_body(dst, dstSize, src, srcSize, CTable);
 }
 
 #endif
 
-size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable)
+size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int flags)
 {
-    return HUF_compress1X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, /* bmi2 */ 0);
+    return HUF_compress1X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, flags);
 }
 
-
 static size_t
 HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize,
                               const void* src, size_t srcSize,
-                              const HUF_CElt* CTable, int bmi2)
+                              const HUF_CElt* CTable, int flags)
 {
     size_t const segmentSize = (srcSize+3)/4;   /* first 3 segments */
     const BYTE* ip = (const BYTE*) src;
@@ -566,27 +1148,24 @@ HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize,
     op += 6;   /* jumpTable */
 
     assert(op <= oend);
-    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, bmi2) );
-        if (cSize==0) return 0;
-        assert(cSize <= 65535);
+    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, flags) );
+        if (cSize == 0 || cSize > 65535) return 0;
         MEM_writeLE16(ostart, (U16)cSize);
         op += cSize;
     }
 
     ip += segmentSize;
     assert(op <= oend);
-    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, bmi2) );
-        if (cSize==0) return 0;
-        assert(cSize <= 65535);
+    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, flags) );
+        if (cSize == 0 || cSize > 65535) return 0;
         MEM_writeLE16(ostart+2, (U16)cSize);
         op += cSize;
     }
 
     ip += segmentSize;
     assert(op <= oend);
-    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, bmi2) );
-        if (cSize==0) return 0;
-        assert(cSize <= 65535);
+    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, flags) );
+        if (cSize == 0 || cSize > 65535) return 0;
         MEM_writeLE16(ostart+4, (U16)cSize);
         op += cSize;
     }
@@ -594,17 +1173,17 @@ HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize,
     ip += segmentSize;
     assert(op <= oend);
     assert(ip <= iend);
-    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, (size_t)(iend-ip), CTable, bmi2) );
-        if (cSize==0) return 0;
+    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, (size_t)(iend-ip), CTable, flags) );
+        if (cSize == 0 || cSize > 65535) return 0;
         op += cSize;
     }
 
     return (size_t)(op-ostart);
 }
 
-size_t HUF_compress4X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable)
+size_t HUF_compress4X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int flags)
 {
-    return HUF_compress4X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, /* bmi2 */ 0);
+    return HUF_compress4X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, flags);
 }
 
 typedef enum { HUF_singleStream, HUF_fourStreams } HUF_nbStreams_e;
@@ -612,11 +1191,11 @@ typedef enum { HUF_singleStream, HUF_fourStreams } HUF_nbStreams_e;
 static size_t HUF_compressCTable_internal(
                 BYTE* const ostart, BYTE* op, BYTE* const oend,
                 const void* src, size_t srcSize,
-                HUF_nbStreams_e nbStreams, const HUF_CElt* CTable, const int bmi2)
+                HUF_nbStreams_e nbStreams, const HUF_CElt* CTable, const int flags)
 {
     size_t const cSize = (nbStreams==HUF_singleStream) ?
-                         HUF_compress1X_usingCTable_internal(op, (size_t)(oend - op), src, srcSize, CTable, bmi2) :
-                         HUF_compress4X_usingCTable_internal(op, (size_t)(oend - op), src, srcSize, CTable, bmi2);
+                         HUF_compress1X_usingCTable_internal(op, (size_t)(oend - op), src, srcSize, CTable, flags) :
+                         HUF_compress4X_usingCTable_internal(op, (size_t)(oend - op), src, srcSize, CTable, flags);
     if (HUF_isError(cSize)) { return cSize; }
     if (cSize==0) { return 0; }   /* uncompressible */
     op += cSize;
@@ -628,31 +1207,111 @@ static size_t HUF_compressCTable_internal(
 
 typedef struct {
     unsigned count[HUF_SYMBOLVALUE_MAX + 1];
-    HUF_CElt CTable[HUF_SYMBOLVALUE_MAX + 1];
-    HUF_buildCTable_wksp_tables buildCTable_wksp;
+    HUF_CElt CTable[HUF_CTABLE_SIZE_ST(HUF_SYMBOLVALUE_MAX)];
+    union {
+        HUF_buildCTable_wksp_tables buildCTable_wksp;
+        HUF_WriteCTableWksp writeCTable_wksp;
+        U32 hist_wksp[HIST_WKSP_SIZE_U32];
+    } wksps;
 } HUF_compress_tables_t;
 
+#define SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE 4096
+#define SUSPECT_INCOMPRESSIBLE_SAMPLE_RATIO 10  /* Must be >= 2 */
+
+unsigned HUF_cardinality(const unsigned* count, unsigned maxSymbolValue)
+{
+    unsigned cardinality = 0;
+    unsigned i;
+
+    for (i = 0; i < maxSymbolValue + 1; i++) {
+        if (count[i] != 0) cardinality += 1;
+    }
+
+    return cardinality;
+}
+
+unsigned HUF_minTableLog(unsigned symbolCardinality)
+{
+    U32 minBitsSymbols = ZSTD_highbit32(symbolCardinality) + 1;
+    return minBitsSymbols;
+}
+
+unsigned HUF_optimalTableLog(
+            unsigned maxTableLog,
+            size_t srcSize,
+            unsigned maxSymbolValue,
+            void* workSpace, size_t wkspSize,
+            HUF_CElt* table,
+      const unsigned* count,
+            int flags)
+{
+    assert(srcSize > 1); /* Not supported, RLE should be used instead */
+    assert(wkspSize >= sizeof(HUF_buildCTable_wksp_tables));
+
+    if (!(flags & HUF_flags_optimalDepth)) {
+        /* cheap evaluation, based on FSE */
+        return FSE_optimalTableLog_internal(maxTableLog, srcSize, maxSymbolValue, 1);
+    }
+
+    {   BYTE* dst = (BYTE*)workSpace + sizeof(HUF_WriteCTableWksp);
+        size_t dstSize = wkspSize - sizeof(HUF_WriteCTableWksp);
+        size_t maxBits, hSize, newSize;
+        const unsigned symbolCardinality = HUF_cardinality(count, maxSymbolValue);
+        const unsigned minTableLog = HUF_minTableLog(symbolCardinality);
+        size_t optSize = ((size_t) ~0) - 1;
+        unsigned optLog = maxTableLog, optLogGuess;
+
+        DEBUGLOG(6, "HUF_optimalTableLog: probing huf depth (srcSize=%zu)", srcSize);
+
+        /* Search until size increases */
+        for (optLogGuess = minTableLog; optLogGuess <= maxTableLog; optLogGuess++) {
+            DEBUGLOG(7, "checking for huffLog=%u", optLogGuess);
+            maxBits = HUF_buildCTable_wksp(table, count, maxSymbolValue, optLogGuess, workSpace, wkspSize);
+            if (ERR_isError(maxBits)) continue;
+
+            if (maxBits < optLogGuess && optLogGuess > minTableLog) break;
+
+            hSize = HUF_writeCTable_wksp(dst, dstSize, table, maxSymbolValue, (U32)maxBits, workSpace, wkspSize);
+
+            if (ERR_isError(hSize)) continue;
+
+            newSize = HUF_estimateCompressedSize(table, count, maxSymbolValue) + hSize;
+
+            if (newSize > optSize + 1) {
+                break;
+            }
+
+            if (newSize < optSize) {
+                optSize = newSize;
+                optLog = optLogGuess;
+            }
+        }
+        assert(optLog <= HUF_TABLELOG_MAX);
+        return optLog;
+    }
+}
+
 /* HUF_compress_internal() :
- * `workSpace` must a table of at least HUF_WORKSPACE_SIZE_U32 unsigned */
+ * `workSpace_align4` must be aligned on 4-bytes boundaries,
+ * and occupies the same space as a table of HUF_WORKSPACE_SIZE_U64 unsigned */
 static size_t
 HUF_compress_internal (void* dst, size_t dstSize,
                  const void* src, size_t srcSize,
                        unsigned maxSymbolValue, unsigned huffLog,
                        HUF_nbStreams_e nbStreams,
                        void* workSpace, size_t wkspSize,
-                       HUF_CElt* oldHufTable, HUF_repeat* repeat, int preferRepeat,
-                 const int bmi2)
+                       HUF_CElt* oldHufTable, HUF_repeat* repeat, int flags)
 {
-    HUF_compress_tables_t* const table = (HUF_compress_tables_t*)workSpace;
+    HUF_compress_tables_t* const table = (HUF_compress_tables_t*)HUF_alignUpWorkspace(workSpace, &wkspSize, ZSTD_ALIGNOF(size_t));
     BYTE* const ostart = (BYTE*)dst;
     BYTE* const oend = ostart + dstSize;
     BYTE* op = ostart;
 
-    HUF_STATIC_ASSERT(sizeof(*table) <= HUF_WORKSPACE_SIZE);
+    DEBUGLOG(5, "HUF_compress_internal (srcSize=%zu)", srcSize);
+    HUF_STATIC_ASSERT(sizeof(*table) + HUF_WORKSPACE_MAX_ALIGNMENT <= HUF_WORKSPACE_SIZE);
 
     /* checks & inits */
-    if (((size_t)workSpace & 3) != 0) return ERROR(GENERIC);  /* must be aligned on 4-bytes boundaries */
-    if (wkspSize < HUF_WORKSPACE_SIZE) return ERROR(workSpace_tooSmall);
+    if (wkspSize < sizeof(*table)) return ERROR(workSpace_tooSmall);
     if (!srcSize) return 0;  /* Uncompressed */
     if (!dstSize) return 0;  /* cannot fit anything within dst budget */
     if (srcSize > HUF_BLOCKSIZE_MAX) return ERROR(srcSize_wrong);   /* current block size limit */
@@ -662,17 +1321,34 @@ HUF_compress_internal (void* dst, size_t dstSize,
     if (!huffLog) huffLog = HUF_TABLELOG_DEFAULT;
 
     /* Heuristic : If old table is valid, use it for small inputs */
-    if (preferRepeat && repeat && *repeat == HUF_repeat_valid) {
+    if ((flags & HUF_flags_preferRepeat) && repeat && *repeat == HUF_repeat_valid) {
         return HUF_compressCTable_internal(ostart, op, oend,
                                            src, srcSize,
-                                           nbStreams, oldHufTable, bmi2);
+                                           nbStreams, oldHufTable, flags);
+    }
+
+    /* If uncompressible data is suspected, do a smaller sampling first */
+    DEBUG_STATIC_ASSERT(SUSPECT_INCOMPRESSIBLE_SAMPLE_RATIO >= 2);
+    if ((flags & HUF_flags_suspectUncompressible) && srcSize >= (SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE * SUSPECT_INCOMPRESSIBLE_SAMPLE_RATIO)) {
+        size_t largestTotal = 0;
+        DEBUGLOG(5, "input suspected incompressible : sampling to check");
+        {   unsigned maxSymbolValueBegin = maxSymbolValue;
+            CHECK_V_F(largestBegin, HIST_count_simple (table->count, &maxSymbolValueBegin, (const BYTE*)src, SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE) );
+            largestTotal += largestBegin;
+        }
+        {   unsigned maxSymbolValueEnd = maxSymbolValue;
+            CHECK_V_F(largestEnd, HIST_count_simple (table->count, &maxSymbolValueEnd, (const BYTE*)src + srcSize - SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE, SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE) );
+            largestTotal += largestEnd;
+        }
+        if (largestTotal <= ((2 * SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE) >> 7)+4) return 0;   /* heuristic : probably not compressible enough */
     }
 
     /* Scan input and build symbol stats */
-    {   CHECK_V_F(largest, HIST_count_wksp (table->count, &maxSymbolValue, (const BYTE*)src, srcSize, workSpace, wkspSize) );
+    {   CHECK_V_F(largest, HIST_count_wksp (table->count, &maxSymbolValue, (const BYTE*)src, srcSize, table->wksps.hist_wksp, sizeof(table->wksps.hist_wksp)) );
         if (largest == srcSize) { *ostart = ((const BYTE*)src)[0]; return 1; }   /* single symbol, rle */
         if (largest <= (srcSize >> 7)+4) return 0;   /* heuristic : probably not compressible enough */
     }
+    DEBUGLOG(6, "histogram detail completed (%zu symbols)", showU32(table->count, maxSymbolValue+1));
 
     /* Check validity of previous table */
     if ( repeat
@@ -681,26 +1357,31 @@ HUF_compress_internal (void* dst, size_t dstSize,
         *repeat = HUF_repeat_none;
     }
     /* Heuristic : use existing table for small inputs */
-    if (preferRepeat && repeat && *repeat != HUF_repeat_none) {
+    if ((flags & HUF_flags_preferRepeat) && repeat && *repeat != HUF_repeat_none) {
         return HUF_compressCTable_internal(ostart, op, oend,
                                            src, srcSize,
-                                           nbStreams, oldHufTable, bmi2);
+                                           nbStreams, oldHufTable, flags);
     }
 
     /* Build Huffman Tree */
-    huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue);
+    huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue, &table->wksps, sizeof(table->wksps), table->CTable, table->count, flags);
     {   size_t const maxBits = HUF_buildCTable_wksp(table->CTable, table->count,
                                             maxSymbolValue, huffLog,
-                                            &table->buildCTable_wksp, sizeof(table->buildCTable_wksp));
+                                            &table->wksps.buildCTable_wksp, sizeof(table->wksps.buildCTable_wksp));
         CHECK_F(maxBits);
         huffLog = (U32)maxBits;
-        /* Zero unused symbols in CTable, so we can check it for validity */
-        memset(table->CTable + (maxSymbolValue + 1), 0,
-               sizeof(table->CTable) - ((maxSymbolValue + 1) * sizeof(HUF_CElt)));
+        DEBUGLOG(6, "bit distribution completed (%zu symbols)", showCTableBits(table->CTable + 1, maxSymbolValue+1));
+    }
+    /* Zero unused symbols in CTable, so we can check it for validity */
+    {
+        size_t const ctableSize = HUF_CTABLE_SIZE_ST(maxSymbolValue);
+        size_t const unusedSize = sizeof(table->CTable) - ctableSize * sizeof(HUF_CElt);
+        ZSTD_memset(table->CTable + ctableSize, 0, unusedSize);
     }
 
     /* Write table description header */
-    {   CHECK_V_F(hSize, HUF_writeCTable (op, dstSize, table->CTable, maxSymbolValue, huffLog) );
+    {   CHECK_V_F(hSize, HUF_writeCTable_wksp(op, dstSize, table->CTable, maxSymbolValue, huffLog,
+                                              &table->wksps.writeCTable_wksp, sizeof(table->wksps.writeCTable_wksp)) );
         /* Check if using previous huffman table is beneficial */
         if (repeat && *repeat != HUF_repeat_none) {
             size_t const oldSize = HUF_estimateCompressedSize(oldHufTable, table->count, maxSymbolValue);
@@ -708,7 +1389,7 @@ HUF_compress_internal (void* dst, size_t dstSize,
             if (oldSize <= hSize + newSize || hSize + 12 >= srcSize) {
                 return HUF_compressCTable_internal(ostart, op, oend,
                                                    src, srcSize,
-                                                   nbStreams, oldHufTable, bmi2);
+                                                   nbStreams, oldHufTable, flags);
         }   }
 
         /* Use the new huffman table */
@@ -716,83 +1397,39 @@ HUF_compress_internal (void* dst, size_t dstSize,
         op += hSize;
         if (repeat) { *repeat = HUF_repeat_none; }
         if (oldHufTable)
-            memcpy(oldHufTable, table->CTable, sizeof(table->CTable));  /* Save new table */
+            ZSTD_memcpy(oldHufTable, table->CTable, sizeof(table->CTable));  /* Save new table */
     }
     return HUF_compressCTable_internal(ostart, op, oend,
                                        src, srcSize,
-                                       nbStreams, table->CTable, bmi2);
-}
-
-
-size_t HUF_compress1X_wksp (void* dst, size_t dstSize,
-                      const void* src, size_t srcSize,
-                      unsigned maxSymbolValue, unsigned huffLog,
-                      void* workSpace, size_t wkspSize)
-{
-    return HUF_compress_internal(dst, dstSize, src, srcSize,
-                                 maxSymbolValue, huffLog, HUF_singleStream,
-                                 workSpace, wkspSize,
-                                 NULL, NULL, 0, 0 /*bmi2*/);
+                                       nbStreams, table->CTable, flags);
 }
 
 size_t HUF_compress1X_repeat (void* dst, size_t dstSize,
                       const void* src, size_t srcSize,
                       unsigned maxSymbolValue, unsigned huffLog,
                       void* workSpace, size_t wkspSize,
-                      HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2)
+                      HUF_CElt* hufTable, HUF_repeat* repeat, int flags)
 {
+    DEBUGLOG(5, "HUF_compress1X_repeat (srcSize = %zu)", srcSize);
     return HUF_compress_internal(dst, dstSize, src, srcSize,
                                  maxSymbolValue, huffLog, HUF_singleStream,
                                  workSpace, wkspSize, hufTable,
-                                 repeat, preferRepeat, bmi2);
-}
-
-size_t HUF_compress1X (void* dst, size_t dstSize,
-                 const void* src, size_t srcSize,
-                 unsigned maxSymbolValue, unsigned huffLog)
-{
-    unsigned workSpace[HUF_WORKSPACE_SIZE_U32];
-    return HUF_compress1X_wksp(dst, dstSize, src, srcSize, maxSymbolValue, huffLog, workSpace, sizeof(workSpace));
-}
-
-/* HUF_compress4X_repeat():
- * compress input using 4 streams.
- * provide workspace to generate compression tables */
-size_t HUF_compress4X_wksp (void* dst, size_t dstSize,
-                      const void* src, size_t srcSize,
-                      unsigned maxSymbolValue, unsigned huffLog,
-                      void* workSpace, size_t wkspSize)
-{
-    return HUF_compress_internal(dst, dstSize, src, srcSize,
-                                 maxSymbolValue, huffLog, HUF_fourStreams,
-                                 workSpace, wkspSize,
-                                 NULL, NULL, 0, 0 /*bmi2*/);
+                                 repeat, flags);
 }
 
 /* HUF_compress4X_repeat():
  * compress input using 4 streams.
+ * consider skipping quickly
  * re-use an existing huffman compression table */
 size_t HUF_compress4X_repeat (void* dst, size_t dstSize,
                       const void* src, size_t srcSize,
                       unsigned maxSymbolValue, unsigned huffLog,
                       void* workSpace, size_t wkspSize,
-                      HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2)
+                      HUF_CElt* hufTable, HUF_repeat* repeat, int flags)
 {
+    DEBUGLOG(5, "HUF_compress4X_repeat (srcSize = %zu)", srcSize);
     return HUF_compress_internal(dst, dstSize, src, srcSize,
                                  maxSymbolValue, huffLog, HUF_fourStreams,
                                  workSpace, wkspSize,
-                                 hufTable, repeat, preferRepeat, bmi2);
-}
-
-size_t HUF_compress2 (void* dst, size_t dstSize,
-                const void* src, size_t srcSize,
-                unsigned maxSymbolValue, unsigned huffLog)
-{
-    unsigned workSpace[HUF_WORKSPACE_SIZE_U32];
-    return HUF_compress4X_wksp(dst, dstSize, src, srcSize, maxSymbolValue, huffLog, workSpace, sizeof(workSpace));
-}
-
-size_t HUF_compress (void* dst, size_t maxDstSize, const void* src, size_t srcSize)
-{
-    return HUF_compress2(dst, maxDstSize, src, srcSize, 255, HUF_TABLELOG_DEFAULT);
+                                 hufTable, repeat, flags);
 }
diff --git a/contrib/zstd/huf_decompress.c b/contrib/zstd/huf_decompress.c
index ab5db481c..ea72fe4bd 100644
--- a/contrib/zstd/huf_decompress.c
+++ b/contrib/zstd/huf_decompress.c
@@ -1,7 +1,7 @@
 /* ******************************************************************
  * huff0 huffman decoder,
  * part of Finite State Entropy library
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  *  You can contact the author at :
  *  - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -15,13 +15,20 @@
 /* **************************************************************
 *  Dependencies
 ****************************************************************/
-#include <string.h>     /* memcpy, memset */
+#include "zstd_deps.h"  /* ZSTD_memcpy, ZSTD_memset */
 #include "compiler.h"
 #include "bitstream.h"  /* BIT_* */
 #include "fse.h"        /* to compress headers */
-#define HUF_STATIC_LINKING_ONLY
 #include "huf.h"
 #include "error_private.h"
+#include "zstd_internal.h"
+#include "bits.h"       /* ZSTD_highbit32, ZSTD_countTrailingZeros64 */
+
+/* **************************************************************
+*  Constants
+****************************************************************/
+
+#define HUF_DECODER_FAST_TABLELOG 11
 
 /* **************************************************************
 *  Macros
@@ -36,6 +43,28 @@
 #error "Cannot force the use of the X1 and X2 decoders at the same time!"
 #endif
 
+/* When DYNAMIC_BMI2 is enabled, fast decoders are only called when bmi2 is
+ * supported at runtime, so we can add the BMI2 target attribute.
+ * When it is disabled, we will still get BMI2 if it is enabled statically.
+ */
+#if DYNAMIC_BMI2
+# define HUF_FAST_BMI2_ATTRS BMI2_TARGET_ATTRIBUTE
+#else
+# define HUF_FAST_BMI2_ATTRS
+#endif
+
+#ifdef __cplusplus
+# define HUF_EXTERN_C extern "C"
+#else
+# define HUF_EXTERN_C
+#endif
+#define HUF_ASM_DECL HUF_EXTERN_C
+
+#if DYNAMIC_BMI2
+# define HUF_NEED_BMI2_FUNCTION 1
+#else
+# define HUF_NEED_BMI2_FUNCTION 0
+#endif
 
 /* **************************************************************
 *  Error Management
@@ -53,6 +82,11 @@
 /* **************************************************************
 *  BMI2 Variant Wrappers
 ****************************************************************/
+typedef size_t (*HUF_DecompressUsingDTableFn)(void *dst, size_t dstSize,
+                                              const void *cSrc,
+                                              size_t cSrcSize,
+                                              const HUF_DTable *DTable);
+
 #if DYNAMIC_BMI2
 
 #define HUF_DGEN(fn)                                                        \
@@ -65,7 +99,7 @@
         return fn##_body(dst, dstSize, cSrc, cSrcSize, DTable);             \
     }                                                                       \
                                                                             \
-    static TARGET_ATTRIBUTE("bmi2") size_t fn##_bmi2(                       \
+    static BMI2_TARGET_ATTRIBUTE size_t fn##_bmi2(                          \
                   void* dst,  size_t dstSize,                               \
             const void* cSrc, size_t cSrcSize,                              \
             const HUF_DTable* DTable)                                       \
@@ -74,9 +108,9 @@
     }                                                                       \
                                                                             \
     static size_t fn(void* dst, size_t dstSize, void const* cSrc,           \
-                     size_t cSrcSize, HUF_DTable const* DTable, int bmi2)   \
+                     size_t cSrcSize, HUF_DTable const* DTable, int flags)  \
     {                                                                       \
-        if (bmi2) {                                                         \
+        if (flags & HUF_flags_bmi2) {                                       \
             return fn##_bmi2(dst, dstSize, cSrc, cSrcSize, DTable);         \
         }                                                                   \
         return fn##_default(dst, dstSize, cSrc, cSrcSize, DTable);          \
@@ -86,9 +120,9 @@
 
 #define HUF_DGEN(fn)                                                        \
     static size_t fn(void* dst, size_t dstSize, void const* cSrc,           \
-                     size_t cSrcSize, HUF_DTable const* DTable, int bmi2)   \
+                     size_t cSrcSize, HUF_DTable const* DTable, int flags)  \
     {                                                                       \
-        (void)bmi2;                                                         \
+        (void)flags;                                                        \
         return fn##_body(dst, dstSize, cSrc, cSrcSize, DTable);             \
     }
 
@@ -103,92 +137,357 @@ typedef struct { BYTE maxTableLog; BYTE tableType; BYTE tableLog; BYTE reserved;
 static DTableDesc HUF_getDTableDesc(const HUF_DTable* table)
 {
     DTableDesc dtd;
-    memcpy(&dtd, table, sizeof(dtd));
+    ZSTD_memcpy(&dtd, table, sizeof(dtd));
     return dtd;
 }
 
+static size_t HUF_initFastDStream(BYTE const* ip) {
+    BYTE const lastByte = ip[7];
+    size_t const bitsConsumed = lastByte ? 8 - ZSTD_highbit32(lastByte) : 0;
+    size_t const value = MEM_readLEST(ip) | 1;
+    assert(bitsConsumed <= 8);
+    assert(sizeof(size_t) == 8);
+    return value << bitsConsumed;
+}
+
+
+/**
+ * The input/output arguments to the Huffman fast decoding loop:
+ *
+ * ip [in/out] - The input pointers, must be updated to reflect what is consumed.
+ * op [in/out] - The output pointers, must be updated to reflect what is written.
+ * bits [in/out] - The bitstream containers, must be updated to reflect the current state.
+ * dt [in] - The decoding table.
+ * ilimit [in] - The input limit, stop when any input pointer is below ilimit.
+ * oend [in] - The end of the output stream. op[3] must not cross oend.
+ * iend [in] - The end of each input stream. ip[i] may cross iend[i],
+ *             as long as it is above ilimit, but that indicates corruption.
+ */
+typedef struct {
+    BYTE const* ip[4];
+    BYTE* op[4];
+    U64 bits[4];
+    void const* dt;
+    BYTE const* ilimit;
+    BYTE* oend;
+    BYTE const* iend[4];
+} HUF_DecompressFastArgs;
+
+typedef void (*HUF_DecompressFastLoopFn)(HUF_DecompressFastArgs*);
+
+/**
+ * Initializes args for the fast decoding loop.
+ * @returns 1 on success
+ *          0 if the fallback implementation should be used.
+ *          Or an error code on failure.
+ */
+static size_t HUF_DecompressFastArgs_init(HUF_DecompressFastArgs* args, void* dst, size_t dstSize, void const* src, size_t srcSize, const HUF_DTable* DTable)
+{
+    void const* dt = DTable + 1;
+    U32 const dtLog = HUF_getDTableDesc(DTable).tableLog;
+
+    const BYTE* const ilimit = (const BYTE*)src + 6 + 8;
+
+    BYTE* const oend = (BYTE*)dst + dstSize;
+
+    /* The fast decoding loop assumes 64-bit little-endian.
+     * This condition is false on x32.
+     */
+    if (!MEM_isLittleEndian() || MEM_32bits())
+        return 0;
+
+    /* strict minimum : jump table + 1 byte per stream */
+    if (srcSize < 10)
+        return ERROR(corruption_detected);
+
+    /* Must have at least 8 bytes per stream because we don't handle initializing smaller bit containers.
+     * If table log is not correct at this point, fallback to the old decoder.
+     * On small inputs we don't have enough data to trigger the fast loop, so use the old decoder.
+     */
+    if (dtLog != HUF_DECODER_FAST_TABLELOG)
+        return 0;
+
+    /* Read the jump table. */
+    {
+        const BYTE* const istart = (const BYTE*)src;
+        size_t const length1 = MEM_readLE16(istart);
+        size_t const length2 = MEM_readLE16(istart+2);
+        size_t const length3 = MEM_readLE16(istart+4);
+        size_t const length4 = srcSize - (length1 + length2 + length3 + 6);
+        args->iend[0] = istart + 6;  /* jumpTable */
+        args->iend[1] = args->iend[0] + length1;
+        args->iend[2] = args->iend[1] + length2;
+        args->iend[3] = args->iend[2] + length3;
+
+        /* HUF_initFastDStream() requires this, and this small of an input
+         * won't benefit from the ASM loop anyways.
+         * length1 must be >= 16 so that ip[0] >= ilimit before the loop
+         * starts.
+         */
+        if (length1 < 16 || length2 < 8 || length3 < 8 || length4 < 8)
+            return 0;
+        if (length4 > srcSize) return ERROR(corruption_detected);   /* overflow */
+    }
+    /* ip[] contains the position that is currently loaded into bits[]. */
+    args->ip[0] = args->iend[1] - sizeof(U64);
+    args->ip[1] = args->iend[2] - sizeof(U64);
+    args->ip[2] = args->iend[3] - sizeof(U64);
+    args->ip[3] = (BYTE const*)src + srcSize - sizeof(U64);
+
+    /* op[] contains the output pointers. */
+    args->op[0] = (BYTE*)dst;
+    args->op[1] = args->op[0] + (dstSize+3)/4;
+    args->op[2] = args->op[1] + (dstSize+3)/4;
+    args->op[3] = args->op[2] + (dstSize+3)/4;
+
+    /* No point to call the ASM loop for tiny outputs. */
+    if (args->op[3] >= oend)
+        return 0;
+
+    /* bits[] is the bit container.
+        * It is read from the MSB down to the LSB.
+        * It is shifted left as it is read, and zeros are
+        * shifted in. After the lowest valid bit a 1 is
+        * set, so that CountTrailingZeros(bits[]) can be used
+        * to count how many bits we've consumed.
+        */
+    args->bits[0] = HUF_initFastDStream(args->ip[0]);
+    args->bits[1] = HUF_initFastDStream(args->ip[1]);
+    args->bits[2] = HUF_initFastDStream(args->ip[2]);
+    args->bits[3] = HUF_initFastDStream(args->ip[3]);
+
+    /* If ip[] >= ilimit, it is guaranteed to be safe to
+        * reload bits[]. It may be beyond its section, but is
+        * guaranteed to be valid (>= istart).
+        */
+    args->ilimit = ilimit;
+
+    args->oend = oend;
+    args->dt = dt;
+
+    return 1;
+}
+
+static size_t HUF_initRemainingDStream(BIT_DStream_t* bit, HUF_DecompressFastArgs const* args, int stream, BYTE* segmentEnd)
+{
+    /* Validate that we haven't overwritten. */
+    if (args->op[stream] > segmentEnd)
+        return ERROR(corruption_detected);
+    /* Validate that we haven't read beyond iend[].
+        * Note that ip[] may be < iend[] because the MSB is
+        * the next bit to read, and we may have consumed 100%
+        * of the stream, so down to iend[i] - 8 is valid.
+        */
+    if (args->ip[stream] < args->iend[stream] - 8)
+        return ERROR(corruption_detected);
+
+    /* Construct the BIT_DStream_t. */
+    assert(sizeof(size_t) == 8);
+    bit->bitContainer = MEM_readLEST(args->ip[stream]);
+    bit->bitsConsumed = ZSTD_countTrailingZeros64(args->bits[stream]);
+    bit->start = (const char*)args->iend[0];
+    bit->limitPtr = bit->start + sizeof(size_t);
+    bit->ptr = (const char*)args->ip[stream];
+
+    return 0;
+}
+
 
 #ifndef HUF_FORCE_DECOMPRESS_X2
 
 /*-***************************/
 /*  single-symbol decoding   */
 /*-***************************/
-typedef struct { BYTE byte; BYTE nbBits; } HUF_DEltX1;   /* single-symbol decoding */
+typedef struct { BYTE nbBits; BYTE byte; } HUF_DEltX1;   /* single-symbol decoding */
 
-size_t HUF_readDTableX1_wksp(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize)
+/**
+ * Packs 4 HUF_DEltX1 structs into a U64. This is used to lay down 4 entries at
+ * a time.
+ */
+static U64 HUF_DEltX1_set4(BYTE symbol, BYTE nbBits) {
+    U64 D4;
+    if (MEM_isLittleEndian()) {
+        D4 = (U64)((symbol << 8) + nbBits);
+    } else {
+        D4 = (U64)(symbol + (nbBits << 8));
+    }
+    assert(D4 < (1U << 16));
+    D4 *= 0x0001000100010001ULL;
+    return D4;
+}
+
+/**
+ * Increase the tableLog to targetTableLog and rescales the stats.
+ * If tableLog > targetTableLog this is a no-op.
+ * @returns New tableLog
+ */
+static U32 HUF_rescaleStats(BYTE* huffWeight, U32* rankVal, U32 nbSymbols, U32 tableLog, U32 targetTableLog)
+{
+    if (tableLog > targetTableLog)
+        return tableLog;
+    if (tableLog < targetTableLog) {
+        U32 const scale = targetTableLog - tableLog;
+        U32 s;
+        /* Increase the weight for all non-zero probability symbols by scale. */
+        for (s = 0; s < nbSymbols; ++s) {
+            huffWeight[s] += (BYTE)((huffWeight[s] == 0) ? 0 : scale);
+        }
+        /* Update rankVal to reflect the new weights.
+         * All weights except 0 get moved to weight + scale.
+         * Weights [1, scale] are empty.
+         */
+        for (s = targetTableLog; s > scale; --s) {
+            rankVal[s] = rankVal[s - scale];
+        }
+        for (s = scale; s > 0; --s) {
+            rankVal[s] = 0;
+        }
+    }
+    return targetTableLog;
+}
+
+typedef struct {
+        U32 rankVal[HUF_TABLELOG_ABSOLUTEMAX + 1];
+        U32 rankStart[HUF_TABLELOG_ABSOLUTEMAX + 1];
+        U32 statsWksp[HUF_READ_STATS_WORKSPACE_SIZE_U32];
+        BYTE symbols[HUF_SYMBOLVALUE_MAX + 1];
+        BYTE huffWeight[HUF_SYMBOLVALUE_MAX + 1];
+} HUF_ReadDTableX1_Workspace;
+
+size_t HUF_readDTableX1_wksp(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize, int flags)
 {
     U32 tableLog = 0;
     U32 nbSymbols = 0;
     size_t iSize;
     void* const dtPtr = DTable + 1;
     HUF_DEltX1* const dt = (HUF_DEltX1*)dtPtr;
+    HUF_ReadDTableX1_Workspace* wksp = (HUF_ReadDTableX1_Workspace*)workSpace;
 
-    U32* rankVal;
-    BYTE* huffWeight;
-    size_t spaceUsed32 = 0;
-
-    rankVal = (U32 *)workSpace + spaceUsed32;
-    spaceUsed32 += HUF_TABLELOG_ABSOLUTEMAX + 1;
-    huffWeight = (BYTE *)((U32 *)workSpace + spaceUsed32);
-    spaceUsed32 += HUF_ALIGN(HUF_SYMBOLVALUE_MAX + 1, sizeof(U32)) >> 2;
-
-    if ((spaceUsed32 << 2) > wkspSize) return ERROR(tableLog_tooLarge);
+    DEBUG_STATIC_ASSERT(HUF_DECOMPRESS_WORKSPACE_SIZE >= sizeof(*wksp));
+    if (sizeof(*wksp) > wkspSize) return ERROR(tableLog_tooLarge);
 
     DEBUG_STATIC_ASSERT(sizeof(DTableDesc) == sizeof(HUF_DTable));
-    /* memset(huffWeight, 0, sizeof(huffWeight)); */   /* is not necessary, even though some analyzer complain ... */
+    /* ZSTD_memset(huffWeight, 0, sizeof(huffWeight)); */   /* is not necessary, even though some analyzer complain ... */
 
-    iSize = HUF_readStats(huffWeight, HUF_SYMBOLVALUE_MAX + 1, rankVal, &nbSymbols, &tableLog, src, srcSize);
+    iSize = HUF_readStats_wksp(wksp->huffWeight, HUF_SYMBOLVALUE_MAX + 1, wksp->rankVal, &nbSymbols, &tableLog, src, srcSize, wksp->statsWksp, sizeof(wksp->statsWksp), flags);
     if (HUF_isError(iSize)) return iSize;
 
+
     /* Table header */
     {   DTableDesc dtd = HUF_getDTableDesc(DTable);
+        U32 const maxTableLog = dtd.maxTableLog + 1;
+        U32 const targetTableLog = MIN(maxTableLog, HUF_DECODER_FAST_TABLELOG);
+        tableLog = HUF_rescaleStats(wksp->huffWeight, wksp->rankVal, nbSymbols, tableLog, targetTableLog);
         if (tableLog > (U32)(dtd.maxTableLog+1)) return ERROR(tableLog_tooLarge);   /* DTable too small, Huffman tree cannot fit in */
         dtd.tableType = 0;
         dtd.tableLog = (BYTE)tableLog;
-        memcpy(DTable, &dtd, sizeof(dtd));
+        ZSTD_memcpy(DTable, &dtd, sizeof(dtd));
     }
 
-    /* Calculate starting value for each rank */
-    {   U32 n, nextRankStart = 0;
-        for (n=1; n<tableLog+1; n++) {
-            U32 const current = nextRankStart;
-            nextRankStart += (rankVal[n] << (n-1));
-            rankVal[n] = current;
-    }   }
-
-    /* fill DTable */
-    {   U32 n;
-        size_t const nEnd = nbSymbols;
-        for (n=0; n<nEnd; n++) {
-            size_t const w = huffWeight[n];
-            size_t const length = (1 << w) >> 1;
-            size_t const uStart = rankVal[w];
-            size_t const uEnd = uStart + length;
-            size_t u;
-            HUF_DEltX1 D;
-            D.byte = (BYTE)n;
-            D.nbBits = (BYTE)(tableLog + 1 - w);
-            rankVal[w] = (U32)uEnd;
-            if (length < 4) {
-                /* Use length in the loop bound so the compiler knows it is short. */
-                for (u = 0; u < length; ++u)
-                    dt[uStart + u] = D;
-            } else {
-                /* Unroll the loop 4 times, we know it is a power of 2. */
-                for (u = uStart; u < uEnd; u += 4) {
-                    dt[u + 0] = D;
-                    dt[u + 1] = D;
-                    dt[u + 2] = D;
-                    dt[u + 3] = D;
-    }   }   }   }
-    return iSize;
-}
+    /* Compute symbols and rankStart given rankVal:
+     *
+     * rankVal already contains the number of values of each weight.
+     *
+     * symbols contains the symbols ordered by weight. First are the rankVal[0]
+     * weight 0 symbols, followed by the rankVal[1] weight 1 symbols, and so on.
+     * symbols[0] is filled (but unused) to avoid a branch.
+     *
+     * rankStart contains the offset where each rank belongs in the DTable.
+     * rankStart[0] is not filled because there are no entries in the table for
+     * weight 0.
+     */
+    {   int n;
+        U32 nextRankStart = 0;
+        int const unroll = 4;
+        int const nLimit = (int)nbSymbols - unroll + 1;
+        for (n=0; n<(int)tableLog+1; n++) {
+            U32 const curr = nextRankStart;
+            nextRankStart += wksp->rankVal[n];
+            wksp->rankStart[n] = curr;
+        }
+        for (n=0; n < nLimit; n += unroll) {
+            int u;
+            for (u=0; u < unroll; ++u) {
+                size_t const w = wksp->huffWeight[n+u];
+                wksp->symbols[wksp->rankStart[w]++] = (BYTE)(n+u);
+            }
+        }
+        for (; n < (int)nbSymbols; ++n) {
+            size_t const w = wksp->huffWeight[n];
+            wksp->symbols[wksp->rankStart[w]++] = (BYTE)n;
+        }
+    }
 
-size_t HUF_readDTableX1(HUF_DTable* DTable, const void* src, size_t srcSize)
-{
-    U32 workSpace[HUF_DECOMPRESS_WORKSPACE_SIZE_U32];
-    return HUF_readDTableX1_wksp(DTable, src, srcSize,
-                                 workSpace, sizeof(workSpace));
+    /* fill DTable
+     * We fill all entries of each weight in order.
+     * That way length is a constant for each iteration of the outer loop.
+     * We can switch based on the length to a different inner loop which is
+     * optimized for that particular case.
+     */
+    {   U32 w;
+        int symbol = wksp->rankVal[0];
+        int rankStart = 0;
+        for (w=1; w<tableLog+1; ++w) {
+            int const symbolCount = wksp->rankVal[w];
+            int const length = (1 << w) >> 1;
+            int uStart = rankStart;
+            BYTE const nbBits = (BYTE)(tableLog + 1 - w);
+            int s;
+            int u;
+            switch (length) {
+            case 1:
+                for (s=0; s<symbolCount; ++s) {
+                    HUF_DEltX1 D;
+                    D.byte = wksp->symbols[symbol + s];
+                    D.nbBits = nbBits;
+                    dt[uStart] = D;
+                    uStart += 1;
+                }
+                break;
+            case 2:
+                for (s=0; s<symbolCount; ++s) {
+                    HUF_DEltX1 D;
+                    D.byte = wksp->symbols[symbol + s];
+                    D.nbBits = nbBits;
+                    dt[uStart+0] = D;
+                    dt[uStart+1] = D;
+                    uStart += 2;
+                }
+                break;
+            case 4:
+                for (s=0; s<symbolCount; ++s) {
+                    U64 const D4 = HUF_DEltX1_set4(wksp->symbols[symbol + s], nbBits);
+                    MEM_write64(dt + uStart, D4);
+                    uStart += 4;
+                }
+                break;
+            case 8:
+                for (s=0; s<symbolCount; ++s) {
+                    U64 const D4 = HUF_DEltX1_set4(wksp->symbols[symbol + s], nbBits);
+                    MEM_write64(dt + uStart, D4);
+                    MEM_write64(dt + uStart + 4, D4);
+                    uStart += 8;
+                }
+                break;
+            default:
+                for (s=0; s<symbolCount; ++s) {
+                    U64 const D4 = HUF_DEltX1_set4(wksp->symbols[symbol + s], nbBits);
+                    for (u=0; u < length; u += 16) {
+                        MEM_write64(dt + uStart + u + 0, D4);
+                        MEM_write64(dt + uStart + u + 4, D4);
+                        MEM_write64(dt + uStart + u + 8, D4);
+                        MEM_write64(dt + uStart + u + 12, D4);
+                    }
+                    assert(u == length);
+                    uStart += length;
+                }
+                break;
+            }
+            symbol += symbolCount;
+            rankStart += symbolCount * length;
+        }
+    }
+    return iSize;
 }
 
 FORCE_INLINE_TEMPLATE BYTE
@@ -217,11 +516,15 @@ HUF_decodeStreamX1(BYTE* p, BIT_DStream_t* const bitDPtr, BYTE* const pEnd, cons
     BYTE* const pStart = p;
 
     /* up to 4 symbols at a time */
-    while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) & (p < pEnd-3)) {
-        HUF_DECODE_SYMBOLX1_2(p, bitDPtr);
-        HUF_DECODE_SYMBOLX1_1(p, bitDPtr);
-        HUF_DECODE_SYMBOLX1_2(p, bitDPtr);
-        HUF_DECODE_SYMBOLX1_0(p, bitDPtr);
+    if ((pEnd - p) > 3) {
+        while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) & (p < pEnd-3)) {
+            HUF_DECODE_SYMBOLX1_2(p, bitDPtr);
+            HUF_DECODE_SYMBOLX1_1(p, bitDPtr);
+            HUF_DECODE_SYMBOLX1_2(p, bitDPtr);
+            HUF_DECODE_SYMBOLX1_0(p, bitDPtr);
+        }
+    } else {
+        BIT_reloadDStream(bitDPtr);
     }
 
     /* [0-3] symbols remaining */
@@ -233,7 +536,7 @@ HUF_decodeStreamX1(BYTE* p, BIT_DStream_t* const bitDPtr, BYTE* const pEnd, cons
     while (p < pEnd)
         HUF_DECODE_SYMBOLX1_0(p, bitDPtr);
 
-    return pEnd-pStart;
+    return (size_t)(pEnd-pStart);
 }
 
 FORCE_INLINE_TEMPLATE size_t
@@ -259,6 +562,10 @@ HUF_decompress1X1_usingDTable_internal_body(
     return dstSize;
 }
 
+/* HUF_decompress4X1_usingDTable_internal_body():
+ * Conditions :
+ * @dstSize >= 6
+ */
 FORCE_INLINE_TEMPLATE size_t
 HUF_decompress4X1_usingDTable_internal_body(
           void* dst,  size_t dstSize,
@@ -301,33 +608,37 @@ HUF_decompress4X1_usingDTable_internal_body(
         U32 endSignal = 1;
 
         if (length4 > cSrcSize) return ERROR(corruption_detected);   /* overflow */
+        if (opStart4 > oend) return ERROR(corruption_detected);      /* overflow */
+        if (dstSize < 6) return ERROR(corruption_detected);         /* stream 4-split doesn't work */
         CHECK_F( BIT_initDStream(&bitD1, istart1, length1) );
         CHECK_F( BIT_initDStream(&bitD2, istart2, length2) );
         CHECK_F( BIT_initDStream(&bitD3, istart3, length3) );
         CHECK_F( BIT_initDStream(&bitD4, istart4, length4) );
 
         /* up to 16 symbols per loop (4 symbols per stream) in 64-bit mode */
-        for ( ; (endSignal) & (op4 < olimit) ; ) {
-            HUF_DECODE_SYMBOLX1_2(op1, &bitD1);
-            HUF_DECODE_SYMBOLX1_2(op2, &bitD2);
-            HUF_DECODE_SYMBOLX1_2(op3, &bitD3);
-            HUF_DECODE_SYMBOLX1_2(op4, &bitD4);
-            HUF_DECODE_SYMBOLX1_1(op1, &bitD1);
-            HUF_DECODE_SYMBOLX1_1(op2, &bitD2);
-            HUF_DECODE_SYMBOLX1_1(op3, &bitD3);
-            HUF_DECODE_SYMBOLX1_1(op4, &bitD4);
-            HUF_DECODE_SYMBOLX1_2(op1, &bitD1);
-            HUF_DECODE_SYMBOLX1_2(op2, &bitD2);
-            HUF_DECODE_SYMBOLX1_2(op3, &bitD3);
-            HUF_DECODE_SYMBOLX1_2(op4, &bitD4);
-            HUF_DECODE_SYMBOLX1_0(op1, &bitD1);
-            HUF_DECODE_SYMBOLX1_0(op2, &bitD2);
-            HUF_DECODE_SYMBOLX1_0(op3, &bitD3);
-            HUF_DECODE_SYMBOLX1_0(op4, &bitD4);
-            endSignal &= BIT_reloadDStreamFast(&bitD1) == BIT_DStream_unfinished;
-            endSignal &= BIT_reloadDStreamFast(&bitD2) == BIT_DStream_unfinished;
-            endSignal &= BIT_reloadDStreamFast(&bitD3) == BIT_DStream_unfinished;
-            endSignal &= BIT_reloadDStreamFast(&bitD4) == BIT_DStream_unfinished;
+        if ((size_t)(oend - op4) >= sizeof(size_t)) {
+            for ( ; (endSignal) & (op4 < olimit) ; ) {
+                HUF_DECODE_SYMBOLX1_2(op1, &bitD1);
+                HUF_DECODE_SYMBOLX1_2(op2, &bitD2);
+                HUF_DECODE_SYMBOLX1_2(op3, &bitD3);
+                HUF_DECODE_SYMBOLX1_2(op4, &bitD4);
+                HUF_DECODE_SYMBOLX1_1(op1, &bitD1);
+                HUF_DECODE_SYMBOLX1_1(op2, &bitD2);
+                HUF_DECODE_SYMBOLX1_1(op3, &bitD3);
+                HUF_DECODE_SYMBOLX1_1(op4, &bitD4);
+                HUF_DECODE_SYMBOLX1_2(op1, &bitD1);
+                HUF_DECODE_SYMBOLX1_2(op2, &bitD2);
+                HUF_DECODE_SYMBOLX1_2(op3, &bitD3);
+                HUF_DECODE_SYMBOLX1_2(op4, &bitD4);
+                HUF_DECODE_SYMBOLX1_0(op1, &bitD1);
+                HUF_DECODE_SYMBOLX1_0(op2, &bitD2);
+                HUF_DECODE_SYMBOLX1_0(op3, &bitD3);
+                HUF_DECODE_SYMBOLX1_0(op4, &bitD4);
+                endSignal &= BIT_reloadDStreamFast(&bitD1) == BIT_DStream_unfinished;
+                endSignal &= BIT_reloadDStreamFast(&bitD2) == BIT_DStream_unfinished;
+                endSignal &= BIT_reloadDStreamFast(&bitD3) == BIT_DStream_unfinished;
+                endSignal &= BIT_reloadDStreamFast(&bitD4) == BIT_DStream_unfinished;
+            }
         }
 
         /* check corruption */
@@ -353,99 +664,230 @@ HUF_decompress4X1_usingDTable_internal_body(
     }
 }
 
+#if HUF_NEED_BMI2_FUNCTION
+static BMI2_TARGET_ATTRIBUTE
+size_t HUF_decompress4X1_usingDTable_internal_bmi2(void* dst, size_t dstSize, void const* cSrc,
+                    size_t cSrcSize, HUF_DTable const* DTable) {
+    return HUF_decompress4X1_usingDTable_internal_body(dst, dstSize, cSrc, cSrcSize, DTable);
+}
+#endif
 
-typedef size_t (*HUF_decompress_usingDTable_t)(void *dst, size_t dstSize,
-                                               const void *cSrc,
-                                               size_t cSrcSize,
-                                               const HUF_DTable *DTable);
+static
+size_t HUF_decompress4X1_usingDTable_internal_default(void* dst, size_t dstSize, void const* cSrc,
+                    size_t cSrcSize, HUF_DTable const* DTable) {
+    return HUF_decompress4X1_usingDTable_internal_body(dst, dstSize, cSrc, cSrcSize, DTable);
+}
 
-HUF_DGEN(HUF_decompress1X1_usingDTable_internal)
-HUF_DGEN(HUF_decompress4X1_usingDTable_internal)
+#if ZSTD_ENABLE_ASM_X86_64_BMI2
 
+HUF_ASM_DECL void HUF_decompress4X1_usingDTable_internal_fast_asm_loop(HUF_DecompressFastArgs* args) ZSTDLIB_HIDDEN;
 
+#endif
 
-size_t HUF_decompress1X1_usingDTable(
-          void* dst,  size_t dstSize,
-    const void* cSrc, size_t cSrcSize,
-    const HUF_DTable* DTable)
+static HUF_FAST_BMI2_ATTRS
+void HUF_decompress4X1_usingDTable_internal_fast_c_loop(HUF_DecompressFastArgs* args)
 {
-    DTableDesc dtd = HUF_getDTableDesc(DTable);
-    if (dtd.tableType != 0) return ERROR(GENERIC);
-    return HUF_decompress1X1_usingDTable_internal(dst, dstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
+    U64 bits[4];
+    BYTE const* ip[4];
+    BYTE* op[4];
+    U16 const* const dtable = (U16 const*)args->dt;
+    BYTE* const oend = args->oend;
+    BYTE const* const ilimit = args->ilimit;
+
+    /* Copy the arguments to local variables */
+    ZSTD_memcpy(&bits, &args->bits, sizeof(bits));
+    ZSTD_memcpy(&ip, &args->ip, sizeof(ip));
+    ZSTD_memcpy(&op, &args->op, sizeof(op));
+
+    assert(MEM_isLittleEndian());
+    assert(!MEM_32bits());
+
+    for (;;) {
+        BYTE* olimit;
+        int stream;
+        int symbol;
+
+        /* Assert loop preconditions */
+#ifndef NDEBUG
+        for (stream = 0; stream < 4; ++stream) {
+            assert(op[stream] <= (stream == 3 ? oend : op[stream + 1]));
+            assert(ip[stream] >= ilimit);
+        }
+#endif
+        /* Compute olimit */
+        {
+            /* Each iteration produces 5 output symbols per stream */
+            size_t const oiters = (size_t)(oend - op[3]) / 5;
+            /* Each iteration consumes up to 11 bits * 5 = 55 bits < 7 bytes
+             * per stream.
+             */
+            size_t const iiters = (size_t)(ip[0] - ilimit) / 7;
+            /* We can safely run iters iterations before running bounds checks */
+            size_t const iters = MIN(oiters, iiters);
+            size_t const symbols = iters * 5;
+
+            /* We can simply check that op[3] < olimit, instead of checking all
+             * of our bounds, since we can't hit the other bounds until we've run
+             * iters iterations, which only happens when op[3] == olimit.
+             */
+            olimit = op[3] + symbols;
+
+            /* Exit fast decoding loop once we get close to the end. */
+            if (op[3] + 20 > olimit)
+                break;
+
+            /* Exit the decoding loop if any input pointer has crossed the
+             * previous one. This indicates corruption, and a precondition
+             * to our loop is that ip[i] >= ip[0].
+             */
+            for (stream = 1; stream < 4; ++stream) {
+                if (ip[stream] < ip[stream - 1])
+                    goto _out;
+            }
+        }
+
+#ifndef NDEBUG
+        for (stream = 1; stream < 4; ++stream) {
+            assert(ip[stream] >= ip[stream - 1]);
+        }
+#endif
+
+        do {
+            /* Decode 5 symbols in each of the 4 streams */
+            for (symbol = 0; symbol < 5; ++symbol) {
+                for (stream = 0; stream < 4; ++stream) {
+                    int const index = (int)(bits[stream] >> 53);
+                    int const entry = (int)dtable[index];
+                    bits[stream] <<= (entry & 63);
+                    op[stream][symbol] = (BYTE)((entry >> 8) & 0xFF);
+                }
+            }
+            /* Reload the bitstreams */
+            for (stream = 0; stream < 4; ++stream) {
+                int const ctz = ZSTD_countTrailingZeros64(bits[stream]);
+                int const nbBits = ctz & 7;
+                int const nbBytes = ctz >> 3;
+                op[stream] += 5;
+                ip[stream] -= nbBytes;
+                bits[stream] = MEM_read64(ip[stream]) | 1;
+                bits[stream] <<= nbBits;
+            }
+        } while (op[3] < olimit);
+    }
+
+_out:
+
+    /* Save the final values of each of the state variables back to args. */
+    ZSTD_memcpy(&args->bits, &bits, sizeof(bits));
+    ZSTD_memcpy(&args->ip, &ip, sizeof(ip));
+    ZSTD_memcpy(&args->op, &op, sizeof(op));
 }
 
-size_t HUF_decompress1X1_DCtx_wksp(HUF_DTable* DCtx, void* dst, size_t dstSize,
-                                   const void* cSrc, size_t cSrcSize,
-                                   void* workSpace, size_t wkspSize)
+/**
+ * @returns @p dstSize on success (>= 6)
+ *          0 if the fallback implementation should be used
+ *          An error if an error occurred
+ */
+static HUF_FAST_BMI2_ATTRS
+size_t
+HUF_decompress4X1_usingDTable_internal_fast(
+          void* dst,  size_t dstSize,
+    const void* cSrc, size_t cSrcSize,
+    const HUF_DTable* DTable,
+    HUF_DecompressFastLoopFn loopFn)
 {
-    const BYTE* ip = (const BYTE*) cSrc;
+    void const* dt = DTable + 1;
+    const BYTE* const iend = (const BYTE*)cSrc + 6;
+    BYTE* const oend = (BYTE*)dst + dstSize;
+    HUF_DecompressFastArgs args;
+    {   size_t const ret = HUF_DecompressFastArgs_init(&args, dst, dstSize, cSrc, cSrcSize, DTable);
+        FORWARD_IF_ERROR(ret, "Failed to init fast loop args");
+        if (ret == 0)
+            return 0;
+    }
 
-    size_t const hSize = HUF_readDTableX1_wksp(DCtx, cSrc, cSrcSize, workSpace, wkspSize);
-    if (HUF_isError(hSize)) return hSize;
-    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
-    ip += hSize; cSrcSize -= hSize;
+    assert(args.ip[0] >= args.ilimit);
+    loopFn(&args);
+
+    /* Our loop guarantees that ip[] >= ilimit and that we haven't
+    * overwritten any op[].
+    */
+    assert(args.ip[0] >= iend);
+    assert(args.ip[1] >= iend);
+    assert(args.ip[2] >= iend);
+    assert(args.ip[3] >= iend);
+    assert(args.op[3] <= oend);
+    (void)iend;
+
+    /* finish bit streams one by one. */
+    {   size_t const segmentSize = (dstSize+3) / 4;
+        BYTE* segmentEnd = (BYTE*)dst;
+        int i;
+        for (i = 0; i < 4; ++i) {
+            BIT_DStream_t bit;
+            if (segmentSize <= (size_t)(oend - segmentEnd))
+                segmentEnd += segmentSize;
+            else
+                segmentEnd = oend;
+            FORWARD_IF_ERROR(HUF_initRemainingDStream(&bit, &args, i, segmentEnd), "corruption");
+            /* Decompress and validate that we've produced exactly the expected length. */
+            args.op[i] += HUF_decodeStreamX1(args.op[i], &bit, segmentEnd, (HUF_DEltX1 const*)dt, HUF_DECODER_FAST_TABLELOG);
+            if (args.op[i] != segmentEnd) return ERROR(corruption_detected);
+        }
+    }
 
-    return HUF_decompress1X1_usingDTable_internal(dst, dstSize, ip, cSrcSize, DCtx, /* bmi2 */ 0);
+    /* decoded size */
+    assert(dstSize != 0);
+    return dstSize;
 }
 
+HUF_DGEN(HUF_decompress1X1_usingDTable_internal)
 
-size_t HUF_decompress1X1_DCtx(HUF_DTable* DCtx, void* dst, size_t dstSize,
-                              const void* cSrc, size_t cSrcSize)
+static size_t HUF_decompress4X1_usingDTable_internal(void* dst, size_t dstSize, void const* cSrc,
+                    size_t cSrcSize, HUF_DTable const* DTable, int flags)
 {
-    U32 workSpace[HUF_DECOMPRESS_WORKSPACE_SIZE_U32];
-    return HUF_decompress1X1_DCtx_wksp(DCtx, dst, dstSize, cSrc, cSrcSize,
-                                       workSpace, sizeof(workSpace));
-}
+    HUF_DecompressUsingDTableFn fallbackFn = HUF_decompress4X1_usingDTable_internal_default;
+    HUF_DecompressFastLoopFn loopFn = HUF_decompress4X1_usingDTable_internal_fast_c_loop;
 
-size_t HUF_decompress1X1 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
-{
-    HUF_CREATE_STATIC_DTABLEX1(DTable, HUF_TABLELOG_MAX);
-    return HUF_decompress1X1_DCtx (DTable, dst, dstSize, cSrc, cSrcSize);
-}
+#if DYNAMIC_BMI2
+    if (flags & HUF_flags_bmi2) {
+        fallbackFn = HUF_decompress4X1_usingDTable_internal_bmi2;
+# if ZSTD_ENABLE_ASM_X86_64_BMI2
+        if (!(flags & HUF_flags_disableAsm)) {
+            loopFn = HUF_decompress4X1_usingDTable_internal_fast_asm_loop;
+        }
+# endif
+    } else {
+        return fallbackFn(dst, dstSize, cSrc, cSrcSize, DTable);
+    }
+#endif
 
-size_t HUF_decompress4X1_usingDTable(
-          void* dst,  size_t dstSize,
-    const void* cSrc, size_t cSrcSize,
-    const HUF_DTable* DTable)
-{
-    DTableDesc dtd = HUF_getDTableDesc(DTable);
-    if (dtd.tableType != 0) return ERROR(GENERIC);
-    return HUF_decompress4X1_usingDTable_internal(dst, dstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
+#if ZSTD_ENABLE_ASM_X86_64_BMI2 && defined(__BMI2__)
+    if (!(flags & HUF_flags_disableAsm)) {
+        loopFn = HUF_decompress4X1_usingDTable_internal_fast_asm_loop;
+    }
+#endif
+
+    if (!(flags & HUF_flags_disableFast)) {
+        size_t const ret = HUF_decompress4X1_usingDTable_internal_fast(dst, dstSize, cSrc, cSrcSize, DTable, loopFn);
+        if (ret != 0)
+            return ret;
+    }
+    return fallbackFn(dst, dstSize, cSrc, cSrcSize, DTable);
 }
 
-static size_t HUF_decompress4X1_DCtx_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize,
+static size_t HUF_decompress4X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
                                    const void* cSrc, size_t cSrcSize,
-                                   void* workSpace, size_t wkspSize, int bmi2)
+                                   void* workSpace, size_t wkspSize, int flags)
 {
     const BYTE* ip = (const BYTE*) cSrc;
 
-    size_t const hSize = HUF_readDTableX1_wksp (dctx, cSrc, cSrcSize,
-                                                workSpace, wkspSize);
+    size_t const hSize = HUF_readDTableX1_wksp(dctx, cSrc, cSrcSize, workSpace, wkspSize, flags);
     if (HUF_isError(hSize)) return hSize;
     if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
     ip += hSize; cSrcSize -= hSize;
 
-    return HUF_decompress4X1_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx, bmi2);
-}
-
-size_t HUF_decompress4X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
-                                   const void* cSrc, size_t cSrcSize,
-                                   void* workSpace, size_t wkspSize)
-{
-    return HUF_decompress4X1_DCtx_wksp_bmi2(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, 0);
-}
-
-
-size_t HUF_decompress4X1_DCtx (HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
-{
-    U32 workSpace[HUF_DECOMPRESS_WORKSPACE_SIZE_U32];
-    return HUF_decompress4X1_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize,
-                                       workSpace, sizeof(workSpace));
-}
-size_t HUF_decompress4X1 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
-{
-    HUF_CREATE_STATIC_DTABLEX1(DTable, HUF_TABLELOG_MAX);
-    return HUF_decompress4X1_DCtx(DTable, dst, dstSize, cSrc, cSrcSize);
+    return HUF_decompress4X1_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx, flags);
 }
 
 #endif /* HUF_FORCE_DECOMPRESS_X2 */
@@ -458,209 +900,322 @@ size_t HUF_decompress4X1 (void* dst, size_t dstSize, const void* cSrc, size_t cS
 /* *************************/
 
 typedef struct { U16 sequence; BYTE nbBits; BYTE length; } HUF_DEltX2;  /* double-symbols decoding */
-typedef struct { BYTE symbol; BYTE weight; } sortedSymbol_t;
+typedef struct { BYTE symbol; } sortedSymbol_t;
 typedef U32 rankValCol_t[HUF_TABLELOG_MAX + 1];
 typedef rankValCol_t rankVal_t[HUF_TABLELOG_MAX];
 
+/**
+ * Constructs a HUF_DEltX2 in a U32.
+ */
+static U32 HUF_buildDEltX2U32(U32 symbol, U32 nbBits, U32 baseSeq, int level)
+{
+    U32 seq;
+    DEBUG_STATIC_ASSERT(offsetof(HUF_DEltX2, sequence) == 0);
+    DEBUG_STATIC_ASSERT(offsetof(HUF_DEltX2, nbBits) == 2);
+    DEBUG_STATIC_ASSERT(offsetof(HUF_DEltX2, length) == 3);
+    DEBUG_STATIC_ASSERT(sizeof(HUF_DEltX2) == sizeof(U32));
+    if (MEM_isLittleEndian()) {
+        seq = level == 1 ? symbol : (baseSeq + (symbol << 8));
+        return seq + (nbBits << 16) + ((U32)level << 24);
+    } else {
+        seq = level == 1 ? (symbol << 8) : ((baseSeq << 8) + symbol);
+        return (seq << 16) + (nbBits << 8) + (U32)level;
+    }
+}
 
-/* HUF_fillDTableX2Level2() :
- * `rankValOrigin` must be a table of at least (HUF_TABLELOG_MAX + 1) U32 */
-static void HUF_fillDTableX2Level2(HUF_DEltX2* DTable, U32 sizeLog, const U32 consumed,
-                           const U32* rankValOrigin, const int minWeight,
-                           const sortedSymbol_t* sortedSymbols, const U32 sortedListSize,
-                           U32 nbBitsBaseline, U16 baseSeq)
+/**
+ * Constructs a HUF_DEltX2.
+ */
+static HUF_DEltX2 HUF_buildDEltX2(U32 symbol, U32 nbBits, U32 baseSeq, int level)
 {
     HUF_DEltX2 DElt;
-    U32 rankVal[HUF_TABLELOG_MAX + 1];
+    U32 const val = HUF_buildDEltX2U32(symbol, nbBits, baseSeq, level);
+    DEBUG_STATIC_ASSERT(sizeof(DElt) == sizeof(val));
+    ZSTD_memcpy(&DElt, &val, sizeof(val));
+    return DElt;
+}
+
+/**
+ * Constructs 2 HUF_DEltX2s and packs them into a U64.
+ */
+static U64 HUF_buildDEltX2U64(U32 symbol, U32 nbBits, U16 baseSeq, int level)
+{
+    U32 DElt = HUF_buildDEltX2U32(symbol, nbBits, baseSeq, level);
+    return (U64)DElt + ((U64)DElt << 32);
+}
 
-    /* get pre-calculated rankVal */
-    memcpy(rankVal, rankValOrigin, sizeof(rankVal));
+/**
+ * Fills the DTable rank with all the symbols from [begin, end) that are each
+ * nbBits long.
+ *
+ * @param DTableRank The start of the rank in the DTable.
+ * @param begin The first symbol to fill (inclusive).
+ * @param end The last symbol to fill (exclusive).
+ * @param nbBits Each symbol is nbBits long.
+ * @param tableLog The table log.
+ * @param baseSeq If level == 1 { 0 } else { the first level symbol }
+ * @param level The level in the table. Must be 1 or 2.
+ */
+static void HUF_fillDTableX2ForWeight(
+    HUF_DEltX2* DTableRank,
+    sortedSymbol_t const* begin, sortedSymbol_t const* end,
+    U32 nbBits, U32 tableLog,
+    U16 baseSeq, int const level)
+{
+    U32 const length = 1U << ((tableLog - nbBits) & 0x1F /* quiet static-analyzer */);
+    const sortedSymbol_t* ptr;
+    assert(level >= 1 && level <= 2);
+    switch (length) {
+    case 1:
+        for (ptr = begin; ptr != end; ++ptr) {
+            HUF_DEltX2 const DElt = HUF_buildDEltX2(ptr->symbol, nbBits, baseSeq, level);
+            *DTableRank++ = DElt;
+        }
+        break;
+    case 2:
+        for (ptr = begin; ptr != end; ++ptr) {
+            HUF_DEltX2 const DElt = HUF_buildDEltX2(ptr->symbol, nbBits, baseSeq, level);
+            DTableRank[0] = DElt;
+            DTableRank[1] = DElt;
+            DTableRank += 2;
+        }
+        break;
+    case 4:
+        for (ptr = begin; ptr != end; ++ptr) {
+            U64 const DEltX2 = HUF_buildDEltX2U64(ptr->symbol, nbBits, baseSeq, level);
+            ZSTD_memcpy(DTableRank + 0, &DEltX2, sizeof(DEltX2));
+            ZSTD_memcpy(DTableRank + 2, &DEltX2, sizeof(DEltX2));
+            DTableRank += 4;
+        }
+        break;
+    case 8:
+        for (ptr = begin; ptr != end; ++ptr) {
+            U64 const DEltX2 = HUF_buildDEltX2U64(ptr->symbol, nbBits, baseSeq, level);
+            ZSTD_memcpy(DTableRank + 0, &DEltX2, sizeof(DEltX2));
+            ZSTD_memcpy(DTableRank + 2, &DEltX2, sizeof(DEltX2));
+            ZSTD_memcpy(DTableRank + 4, &DEltX2, sizeof(DEltX2));
+            ZSTD_memcpy(DTableRank + 6, &DEltX2, sizeof(DEltX2));
+            DTableRank += 8;
+        }
+        break;
+    default:
+        for (ptr = begin; ptr != end; ++ptr) {
+            U64 const DEltX2 = HUF_buildDEltX2U64(ptr->symbol, nbBits, baseSeq, level);
+            HUF_DEltX2* const DTableRankEnd = DTableRank + length;
+            for (; DTableRank != DTableRankEnd; DTableRank += 8) {
+                ZSTD_memcpy(DTableRank + 0, &DEltX2, sizeof(DEltX2));
+                ZSTD_memcpy(DTableRank + 2, &DEltX2, sizeof(DEltX2));
+                ZSTD_memcpy(DTableRank + 4, &DEltX2, sizeof(DEltX2));
+                ZSTD_memcpy(DTableRank + 6, &DEltX2, sizeof(DEltX2));
+            }
+        }
+        break;
+    }
+}
 
-    /* fill skipped values */
+/* HUF_fillDTableX2Level2() :
+ * `rankValOrigin` must be a table of at least (HUF_TABLELOG_MAX + 1) U32 */
+static void HUF_fillDTableX2Level2(HUF_DEltX2* DTable, U32 targetLog, const U32 consumedBits,
+                           const U32* rankVal, const int minWeight, const int maxWeight1,
+                           const sortedSymbol_t* sortedSymbols, U32 const* rankStart,
+                           U32 nbBitsBaseline, U16 baseSeq)
+{
+    /* Fill skipped values (all positions up to rankVal[minWeight]).
+     * These are positions only get a single symbol because the combined weight
+     * is too large.
+     */
     if (minWeight>1) {
-        U32 i, skipSize = rankVal[minWeight];
-        MEM_writeLE16(&(DElt.sequence), baseSeq);
-        DElt.nbBits   = (BYTE)(consumed);
-        DElt.length   = 1;
-        for (i = 0; i < skipSize; i++)
-            DTable[i] = DElt;
+        U32 const length = 1U << ((targetLog - consumedBits) & 0x1F /* quiet static-analyzer */);
+        U64 const DEltX2 = HUF_buildDEltX2U64(baseSeq, consumedBits, /* baseSeq */ 0, /* level */ 1);
+        int const skipSize = rankVal[minWeight];
+        assert(length > 1);
+        assert((U32)skipSize < length);
+        switch (length) {
+        case 2:
+            assert(skipSize == 1);
+            ZSTD_memcpy(DTable, &DEltX2, sizeof(DEltX2));
+            break;
+        case 4:
+            assert(skipSize <= 4);
+            ZSTD_memcpy(DTable + 0, &DEltX2, sizeof(DEltX2));
+            ZSTD_memcpy(DTable + 2, &DEltX2, sizeof(DEltX2));
+            break;
+        default:
+            {
+                int i;
+                for (i = 0; i < skipSize; i += 8) {
+                    ZSTD_memcpy(DTable + i + 0, &DEltX2, sizeof(DEltX2));
+                    ZSTD_memcpy(DTable + i + 2, &DEltX2, sizeof(DEltX2));
+                    ZSTD_memcpy(DTable + i + 4, &DEltX2, sizeof(DEltX2));
+                    ZSTD_memcpy(DTable + i + 6, &DEltX2, sizeof(DEltX2));
+                }
+            }
+        }
     }
 
-    /* fill DTable */
-    {   U32 s; for (s=0; s<sortedListSize; s++) {   /* note : sortedSymbols already skipped */
-            const U32 symbol = sortedSymbols[s].symbol;
-            const U32 weight = sortedSymbols[s].weight;
-            const U32 nbBits = nbBitsBaseline - weight;
-            const U32 length = 1 << (sizeLog-nbBits);
-            const U32 start = rankVal[weight];
-            U32 i = start;
-            const U32 end = start + length;
-
-            MEM_writeLE16(&(DElt.sequence), (U16)(baseSeq + (symbol << 8)));
-            DElt.nbBits = (BYTE)(nbBits + consumed);
-            DElt.length = 2;
-            do { DTable[i++] = DElt; } while (i<end);   /* since length >= 1 */
-
-            rankVal[weight] += length;
-    }   }
+    /* Fill each of the second level symbols by weight. */
+    {
+        int w;
+        for (w = minWeight; w < maxWeight1; ++w) {
+            int const begin = rankStart[w];
+            int const end = rankStart[w+1];
+            U32 const nbBits = nbBitsBaseline - w;
+            U32 const totalBits = nbBits + consumedBits;
+            HUF_fillDTableX2ForWeight(
+                DTable + rankVal[w],
+                sortedSymbols + begin, sortedSymbols + end,
+                totalBits, targetLog,
+                baseSeq, /* level */ 2);
+        }
+    }
 }
 
-
 static void HUF_fillDTableX2(HUF_DEltX2* DTable, const U32 targetLog,
-                           const sortedSymbol_t* sortedList, const U32 sortedListSize,
-                           const U32* rankStart, rankVal_t rankValOrigin, const U32 maxWeight,
+                           const sortedSymbol_t* sortedList,
+                           const U32* rankStart, rankValCol_t* rankValOrigin, const U32 maxWeight,
                            const U32 nbBitsBaseline)
 {
-    U32 rankVal[HUF_TABLELOG_MAX + 1];
+    U32* const rankVal = rankValOrigin[0];
     const int scaleLog = nbBitsBaseline - targetLog;   /* note : targetLog >= srcLog, hence scaleLog <= 1 */
     const U32 minBits  = nbBitsBaseline - maxWeight;
-    U32 s;
-
-    memcpy(rankVal, rankValOrigin, sizeof(rankVal));
-
-    /* fill DTable */
-    for (s=0; s<sortedListSize; s++) {
-        const U16 symbol = sortedList[s].symbol;
-        const U32 weight = sortedList[s].weight;
-        const U32 nbBits = nbBitsBaseline - weight;
-        const U32 start = rankVal[weight];
-        const U32 length = 1 << (targetLog-nbBits);
-
-        if (targetLog-nbBits >= minBits) {   /* enough room for a second symbol */
-            U32 sortedRank;
+    int w;
+    int const wEnd = (int)maxWeight + 1;
+
+    /* Fill DTable in order of weight. */
+    for (w = 1; w < wEnd; ++w) {
+        int const begin = (int)rankStart[w];
+        int const end = (int)rankStart[w+1];
+        U32 const nbBits = nbBitsBaseline - w;
+
+        if (targetLog-nbBits >= minBits) {
+            /* Enough room for a second symbol. */
+            int start = rankVal[w];
+            U32 const length = 1U << ((targetLog - nbBits) & 0x1F /* quiet static-analyzer */);
             int minWeight = nbBits + scaleLog;
+            int s;
             if (minWeight < 1) minWeight = 1;
-            sortedRank = rankStart[minWeight];
-            HUF_fillDTableX2Level2(DTable+start, targetLog-nbBits, nbBits,
-                           rankValOrigin[nbBits], minWeight,
-                           sortedList+sortedRank, sortedListSize-sortedRank,
-                           nbBitsBaseline, symbol);
+            /* Fill the DTable for every symbol of weight w.
+             * These symbols get at least 1 second symbol.
+             */
+            for (s = begin; s != end; ++s) {
+                HUF_fillDTableX2Level2(
+                    DTable + start, targetLog, nbBits,
+                    rankValOrigin[nbBits], minWeight, wEnd,
+                    sortedList, rankStart,
+                    nbBitsBaseline, sortedList[s].symbol);
+                start += length;
+            }
         } else {
-            HUF_DEltX2 DElt;
-            MEM_writeLE16(&(DElt.sequence), symbol);
-            DElt.nbBits = (BYTE)(nbBits);
-            DElt.length = 1;
-            {   U32 const end = start + length;
-                U32 u;
-                for (u = start; u < end; u++) DTable[u] = DElt;
-        }   }
-        rankVal[weight] += length;
+            /* Only a single symbol. */
+            HUF_fillDTableX2ForWeight(
+                DTable + rankVal[w],
+                sortedList + begin, sortedList + end,
+                nbBits, targetLog,
+                /* baseSeq */ 0, /* level */ 1);
+        }
     }
 }
 
+typedef struct {
+    rankValCol_t rankVal[HUF_TABLELOG_MAX];
+    U32 rankStats[HUF_TABLELOG_MAX + 1];
+    U32 rankStart0[HUF_TABLELOG_MAX + 3];
+    sortedSymbol_t sortedSymbol[HUF_SYMBOLVALUE_MAX + 1];
+    BYTE weightList[HUF_SYMBOLVALUE_MAX + 1];
+    U32 calleeWksp[HUF_READ_STATS_WORKSPACE_SIZE_U32];
+} HUF_ReadDTableX2_Workspace;
+
 size_t HUF_readDTableX2_wksp(HUF_DTable* DTable,
                        const void* src, size_t srcSize,
-                             void* workSpace, size_t wkspSize)
+                             void* workSpace, size_t wkspSize, int flags)
 {
-    U32 tableLog, maxW, sizeOfSort, nbSymbols;
+    U32 tableLog, maxW, nbSymbols;
     DTableDesc dtd = HUF_getDTableDesc(DTable);
-    U32 const maxTableLog = dtd.maxTableLog;
+    U32 maxTableLog = dtd.maxTableLog;
     size_t iSize;
     void* dtPtr = DTable+1;   /* force compiler to avoid strict-aliasing */
     HUF_DEltX2* const dt = (HUF_DEltX2*)dtPtr;
     U32 *rankStart;
 
-    rankValCol_t* rankVal;
-    U32* rankStats;
-    U32* rankStart0;
-    sortedSymbol_t* sortedSymbol;
-    BYTE* weightList;
-    size_t spaceUsed32 = 0;
-
-    rankVal = (rankValCol_t *)((U32 *)workSpace + spaceUsed32);
-    spaceUsed32 += (sizeof(rankValCol_t) * HUF_TABLELOG_MAX) >> 2;
-    rankStats = (U32 *)workSpace + spaceUsed32;
-    spaceUsed32 += HUF_TABLELOG_MAX + 1;
-    rankStart0 = (U32 *)workSpace + spaceUsed32;
-    spaceUsed32 += HUF_TABLELOG_MAX + 2;
-    sortedSymbol = (sortedSymbol_t *)workSpace + (spaceUsed32 * sizeof(U32)) / sizeof(sortedSymbol_t);
-    spaceUsed32 += HUF_ALIGN(sizeof(sortedSymbol_t) * (HUF_SYMBOLVALUE_MAX + 1), sizeof(U32)) >> 2;
-    weightList = (BYTE *)((U32 *)workSpace + spaceUsed32);
-    spaceUsed32 += HUF_ALIGN(HUF_SYMBOLVALUE_MAX + 1, sizeof(U32)) >> 2;
-
-    if ((spaceUsed32 << 2) > wkspSize) return ERROR(tableLog_tooLarge);
-
-    rankStart = rankStart0 + 1;
-    memset(rankStats, 0, sizeof(U32) * (2 * HUF_TABLELOG_MAX + 2 + 1));
+    HUF_ReadDTableX2_Workspace* const wksp = (HUF_ReadDTableX2_Workspace*)workSpace;
+
+    if (sizeof(*wksp) > wkspSize) return ERROR(GENERIC);
+
+    rankStart = wksp->rankStart0 + 1;
+    ZSTD_memset(wksp->rankStats, 0, sizeof(wksp->rankStats));
+    ZSTD_memset(wksp->rankStart0, 0, sizeof(wksp->rankStart0));
 
     DEBUG_STATIC_ASSERT(sizeof(HUF_DEltX2) == sizeof(HUF_DTable));   /* if compiler fails here, assertion is wrong */
     if (maxTableLog > HUF_TABLELOG_MAX) return ERROR(tableLog_tooLarge);
-    /* memset(weightList, 0, sizeof(weightList)); */  /* is not necessary, even though some analyzer complain ... */
+    /* ZSTD_memset(weightList, 0, sizeof(weightList)); */  /* is not necessary, even though some analyzer complain ... */
 
-    iSize = HUF_readStats(weightList, HUF_SYMBOLVALUE_MAX + 1, rankStats, &nbSymbols, &tableLog, src, srcSize);
+    iSize = HUF_readStats_wksp(wksp->weightList, HUF_SYMBOLVALUE_MAX + 1, wksp->rankStats, &nbSymbols, &tableLog, src, srcSize, wksp->calleeWksp, sizeof(wksp->calleeWksp), flags);
     if (HUF_isError(iSize)) return iSize;
 
     /* check result */
     if (tableLog > maxTableLog) return ERROR(tableLog_tooLarge);   /* DTable can't fit code depth */
+    if (tableLog <= HUF_DECODER_FAST_TABLELOG && maxTableLog > HUF_DECODER_FAST_TABLELOG) maxTableLog = HUF_DECODER_FAST_TABLELOG;
 
     /* find maxWeight */
-    for (maxW = tableLog; rankStats[maxW]==0; maxW--) {}  /* necessarily finds a solution before 0 */
+    for (maxW = tableLog; wksp->rankStats[maxW]==0; maxW--) {}  /* necessarily finds a solution before 0 */
 
     /* Get start index of each weight */
     {   U32 w, nextRankStart = 0;
         for (w=1; w<maxW+1; w++) {
-            U32 current = nextRankStart;
-            nextRankStart += rankStats[w];
-            rankStart[w] = current;
+            U32 curr = nextRankStart;
+            nextRankStart += wksp->rankStats[w];
+            rankStart[w] = curr;
         }
         rankStart[0] = nextRankStart;   /* put all 0w symbols at the end of sorted list*/
-        sizeOfSort = nextRankStart;
+        rankStart[maxW+1] = nextRankStart;
     }
 
     /* sort symbols by weight */
     {   U32 s;
         for (s=0; s<nbSymbols; s++) {
-            U32 const w = weightList[s];
+            U32 const w = wksp->weightList[s];
             U32 const r = rankStart[w]++;
-            sortedSymbol[r].symbol = (BYTE)s;
-            sortedSymbol[r].weight = (BYTE)w;
+            wksp->sortedSymbol[r].symbol = (BYTE)s;
         }
         rankStart[0] = 0;   /* forget 0w symbols; this is beginning of weight(1) */
     }
 
     /* Build rankVal */
-    {   U32* const rankVal0 = rankVal[0];
+    {   U32* const rankVal0 = wksp->rankVal[0];
         {   int const rescale = (maxTableLog-tableLog) - 1;   /* tableLog <= maxTableLog */
             U32 nextRankVal = 0;
             U32 w;
             for (w=1; w<maxW+1; w++) {
-                U32 current = nextRankVal;
-                nextRankVal += rankStats[w] << (w+rescale);
-                rankVal0[w] = current;
+                U32 curr = nextRankVal;
+                nextRankVal += wksp->rankStats[w] << (w+rescale);
+                rankVal0[w] = curr;
         }   }
         {   U32 const minBits = tableLog+1 - maxW;
             U32 consumed;
             for (consumed = minBits; consumed < maxTableLog - minBits + 1; consumed++) {
-                U32* const rankValPtr = rankVal[consumed];
+                U32* const rankValPtr = wksp->rankVal[consumed];
                 U32 w;
                 for (w = 1; w < maxW+1; w++) {
                     rankValPtr[w] = rankVal0[w] >> consumed;
     }   }   }   }
 
     HUF_fillDTableX2(dt, maxTableLog,
-                   sortedSymbol, sizeOfSort,
-                   rankStart0, rankVal, maxW,
+                   wksp->sortedSymbol,
+                   wksp->rankStart0, wksp->rankVal, maxW,
                    tableLog+1);
 
     dtd.tableLog = (BYTE)maxTableLog;
     dtd.tableType = 1;
-    memcpy(DTable, &dtd, sizeof(dtd));
+    ZSTD_memcpy(DTable, &dtd, sizeof(dtd));
     return iSize;
 }
 
-size_t HUF_readDTableX2(HUF_DTable* DTable, const void* src, size_t srcSize)
-{
-  U32 workSpace[HUF_DECOMPRESS_WORKSPACE_SIZE_U32];
-  return HUF_readDTableX2_wksp(DTable, src, srcSize,
-                               workSpace, sizeof(workSpace));
-}
-
 
 FORCE_INLINE_TEMPLATE U32
 HUF_decodeSymbolX2(void* op, BIT_DStream_t* DStream, const HUF_DEltX2* dt, const U32 dtLog)
 {
     size_t const val = BIT_lookBitsFast(DStream, dtLog);   /* note : dtLog >= 1 */
-    memcpy(op, dt+val, 2);
+    ZSTD_memcpy(op, &dt[val].sequence, 2);
     BIT_skipBits(DStream, dt[val].nbBits);
     return dt[val].length;
 }
@@ -669,15 +1224,17 @@ FORCE_INLINE_TEMPLATE U32
 HUF_decodeLastSymbolX2(void* op, BIT_DStream_t* DStream, const HUF_DEltX2* dt, const U32 dtLog)
 {
     size_t const val = BIT_lookBitsFast(DStream, dtLog);   /* note : dtLog >= 1 */
-    memcpy(op, dt+val, 1);
-    if (dt[val].length==1) BIT_skipBits(DStream, dt[val].nbBits);
-    else {
+    ZSTD_memcpy(op, &dt[val].sequence, 1);
+    if (dt[val].length==1) {
+        BIT_skipBits(DStream, dt[val].nbBits);
+    } else {
         if (DStream->bitsConsumed < (sizeof(DStream->bitContainer)*8)) {
             BIT_skipBits(DStream, dt[val].nbBits);
             if (DStream->bitsConsumed > (sizeof(DStream->bitContainer)*8))
                 /* ugly hack; works only because it's the last symbol. Note : can't easily extract nbBits from just this symbol */
                 DStream->bitsConsumed = (sizeof(DStream->bitContainer)*8);
-    }   }
+        }
+    }
     return 1;
 }
 
@@ -699,19 +1256,37 @@ HUF_decodeStreamX2(BYTE* p, BIT_DStream_t* bitDPtr, BYTE* const pEnd,
     BYTE* const pStart = p;
 
     /* up to 8 symbols at a time */
-    while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) & (p < pEnd-(sizeof(bitDPtr->bitContainer)-1))) {
-        HUF_DECODE_SYMBOLX2_2(p, bitDPtr);
-        HUF_DECODE_SYMBOLX2_1(p, bitDPtr);
-        HUF_DECODE_SYMBOLX2_2(p, bitDPtr);
-        HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
+    if ((size_t)(pEnd - p) >= sizeof(bitDPtr->bitContainer)) {
+        if (dtLog <= 11 && MEM_64bits()) {
+            /* up to 10 symbols at a time */
+            while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) & (p < pEnd-9)) {
+                HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
+                HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
+                HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
+                HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
+                HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
+            }
+        } else {
+            /* up to 8 symbols at a time */
+            while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) & (p < pEnd-(sizeof(bitDPtr->bitContainer)-1))) {
+                HUF_DECODE_SYMBOLX2_2(p, bitDPtr);
+                HUF_DECODE_SYMBOLX2_1(p, bitDPtr);
+                HUF_DECODE_SYMBOLX2_2(p, bitDPtr);
+                HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
+            }
+        }
+    } else {
+        BIT_reloadDStream(bitDPtr);
     }
 
     /* closer to end : up to 2 symbols at a time */
-    while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) & (p <= pEnd-2))
-        HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
+    if ((size_t)(pEnd - p) >= 2) {
+        while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) & (p <= pEnd-2))
+            HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
 
-    while (p <= pEnd-2)
-        HUF_DECODE_SYMBOLX2_0(p, bitDPtr);   /* no need to reload : reached the end of DStream */
+        while (p <= pEnd-2)
+            HUF_DECODE_SYMBOLX2_0(p, bitDPtr);   /* no need to reload : reached the end of DStream */
+    }
 
     if (p < pEnd)
         p += HUF_decodeLastSymbolX2(p, bitDPtr, dt, dtLog);
@@ -746,6 +1321,10 @@ HUF_decompress1X2_usingDTable_internal_body(
     return dstSize;
 }
 
+/* HUF_decompress4X2_usingDTable_internal_body():
+ * Conditions:
+ * @dstSize >= 6
+ */
 FORCE_INLINE_TEMPLATE size_t
 HUF_decompress4X2_usingDTable_internal_body(
           void* dst,  size_t dstSize,
@@ -786,58 +1365,62 @@ HUF_decompress4X2_usingDTable_internal_body(
         DTableDesc const dtd = HUF_getDTableDesc(DTable);
         U32 const dtLog = dtd.tableLog;
 
-        if (length4 > cSrcSize) return ERROR(corruption_detected);   /* overflow */
+        if (length4 > cSrcSize) return ERROR(corruption_detected);  /* overflow */
+        if (opStart4 > oend) return ERROR(corruption_detected);     /* overflow */
+        if (dstSize < 6) return ERROR(corruption_detected);         /* stream 4-split doesn't work */
         CHECK_F( BIT_initDStream(&bitD1, istart1, length1) );
         CHECK_F( BIT_initDStream(&bitD2, istart2, length2) );
         CHECK_F( BIT_initDStream(&bitD3, istart3, length3) );
         CHECK_F( BIT_initDStream(&bitD4, istart4, length4) );
 
         /* 16-32 symbols per loop (4-8 symbols per stream) */
-        for ( ; (endSignal) & (op4 < olimit); ) {
+        if ((size_t)(oend - op4) >= sizeof(size_t)) {
+            for ( ; (endSignal) & (op4 < olimit); ) {
 #if defined(__clang__) && (defined(__x86_64__) || defined(__i386__))
-            HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
-            HUF_DECODE_SYMBOLX2_1(op1, &bitD1);
-            HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
-            HUF_DECODE_SYMBOLX2_0(op1, &bitD1);
-            HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
-            HUF_DECODE_SYMBOLX2_1(op2, &bitD2);
-            HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
-            HUF_DECODE_SYMBOLX2_0(op2, &bitD2);
-            endSignal &= BIT_reloadDStreamFast(&bitD1) == BIT_DStream_unfinished;
-            endSignal &= BIT_reloadDStreamFast(&bitD2) == BIT_DStream_unfinished;
-            HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
-            HUF_DECODE_SYMBOLX2_1(op3, &bitD3);
-            HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
-            HUF_DECODE_SYMBOLX2_0(op3, &bitD3);
-            HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
-            HUF_DECODE_SYMBOLX2_1(op4, &bitD4);
-            HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
-            HUF_DECODE_SYMBOLX2_0(op4, &bitD4);
-            endSignal &= BIT_reloadDStreamFast(&bitD3) == BIT_DStream_unfinished;
-            endSignal &= BIT_reloadDStreamFast(&bitD4) == BIT_DStream_unfinished;
+                HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
+                HUF_DECODE_SYMBOLX2_1(op1, &bitD1);
+                HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
+                HUF_DECODE_SYMBOLX2_0(op1, &bitD1);
+                HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
+                HUF_DECODE_SYMBOLX2_1(op2, &bitD2);
+                HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
+                HUF_DECODE_SYMBOLX2_0(op2, &bitD2);
+                endSignal &= BIT_reloadDStreamFast(&bitD1) == BIT_DStream_unfinished;
+                endSignal &= BIT_reloadDStreamFast(&bitD2) == BIT_DStream_unfinished;
+                HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
+                HUF_DECODE_SYMBOLX2_1(op3, &bitD3);
+                HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
+                HUF_DECODE_SYMBOLX2_0(op3, &bitD3);
+                HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
+                HUF_DECODE_SYMBOLX2_1(op4, &bitD4);
+                HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
+                HUF_DECODE_SYMBOLX2_0(op4, &bitD4);
+                endSignal &= BIT_reloadDStreamFast(&bitD3) == BIT_DStream_unfinished;
+                endSignal &= BIT_reloadDStreamFast(&bitD4) == BIT_DStream_unfinished;
 #else
-            HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
-            HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
-            HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
-            HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
-            HUF_DECODE_SYMBOLX2_1(op1, &bitD1);
-            HUF_DECODE_SYMBOLX2_1(op2, &bitD2);
-            HUF_DECODE_SYMBOLX2_1(op3, &bitD3);
-            HUF_DECODE_SYMBOLX2_1(op4, &bitD4);
-            HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
-            HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
-            HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
-            HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
-            HUF_DECODE_SYMBOLX2_0(op1, &bitD1);
-            HUF_DECODE_SYMBOLX2_0(op2, &bitD2);
-            HUF_DECODE_SYMBOLX2_0(op3, &bitD3);
-            HUF_DECODE_SYMBOLX2_0(op4, &bitD4);
-            endSignal = (U32)LIKELY(
-                        (BIT_reloadDStreamFast(&bitD1) == BIT_DStream_unfinished)
-                      & (BIT_reloadDStreamFast(&bitD2) == BIT_DStream_unfinished)
-                      & (BIT_reloadDStreamFast(&bitD3) == BIT_DStream_unfinished)
-                      & (BIT_reloadDStreamFast(&bitD4) == BIT_DStream_unfinished));
+                HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
+                HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
+                HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
+                HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
+                HUF_DECODE_SYMBOLX2_1(op1, &bitD1);
+                HUF_DECODE_SYMBOLX2_1(op2, &bitD2);
+                HUF_DECODE_SYMBOLX2_1(op3, &bitD3);
+                HUF_DECODE_SYMBOLX2_1(op4, &bitD4);
+                HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
+                HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
+                HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
+                HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
+                HUF_DECODE_SYMBOLX2_0(op1, &bitD1);
+                HUF_DECODE_SYMBOLX2_0(op2, &bitD2);
+                HUF_DECODE_SYMBOLX2_0(op3, &bitD3);
+                HUF_DECODE_SYMBOLX2_0(op4, &bitD4);
+                endSignal = (U32)LIKELY((U32)
+                            (BIT_reloadDStreamFast(&bitD1) == BIT_DStream_unfinished)
+                        & (BIT_reloadDStreamFast(&bitD2) == BIT_DStream_unfinished)
+                        & (BIT_reloadDStreamFast(&bitD3) == BIT_DStream_unfinished)
+                        & (BIT_reloadDStreamFast(&bitD4) == BIT_DStream_unfinished));
 #endif
+            }
         }
 
         /* check corruption */
@@ -861,94 +1444,279 @@ HUF_decompress4X2_usingDTable_internal_body(
     }
 }
 
-HUF_DGEN(HUF_decompress1X2_usingDTable_internal)
-HUF_DGEN(HUF_decompress4X2_usingDTable_internal)
+#if HUF_NEED_BMI2_FUNCTION
+static BMI2_TARGET_ATTRIBUTE
+size_t HUF_decompress4X2_usingDTable_internal_bmi2(void* dst, size_t dstSize, void const* cSrc,
+                    size_t cSrcSize, HUF_DTable const* DTable) {
+    return HUF_decompress4X2_usingDTable_internal_body(dst, dstSize, cSrc, cSrcSize, DTable);
+}
+#endif
 
-size_t HUF_decompress1X2_usingDTable(
-          void* dst,  size_t dstSize,
-    const void* cSrc, size_t cSrcSize,
-    const HUF_DTable* DTable)
-{
-    DTableDesc dtd = HUF_getDTableDesc(DTable);
-    if (dtd.tableType != 1) return ERROR(GENERIC);
-    return HUF_decompress1X2_usingDTable_internal(dst, dstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
+static
+size_t HUF_decompress4X2_usingDTable_internal_default(void* dst, size_t dstSize, void const* cSrc,
+                    size_t cSrcSize, HUF_DTable const* DTable) {
+    return HUF_decompress4X2_usingDTable_internal_body(dst, dstSize, cSrc, cSrcSize, DTable);
 }
 
-size_t HUF_decompress1X2_DCtx_wksp(HUF_DTable* DCtx, void* dst, size_t dstSize,
-                                   const void* cSrc, size_t cSrcSize,
-                                   void* workSpace, size_t wkspSize)
+#if ZSTD_ENABLE_ASM_X86_64_BMI2
+
+HUF_ASM_DECL void HUF_decompress4X2_usingDTable_internal_fast_asm_loop(HUF_DecompressFastArgs* args) ZSTDLIB_HIDDEN;
+
+#endif
+
+static HUF_FAST_BMI2_ATTRS
+void HUF_decompress4X2_usingDTable_internal_fast_c_loop(HUF_DecompressFastArgs* args)
 {
-    const BYTE* ip = (const BYTE*) cSrc;
+    U64 bits[4];
+    BYTE const* ip[4];
+    BYTE* op[4];
+    BYTE* oend[4];
+    HUF_DEltX2 const* const dtable = (HUF_DEltX2 const*)args->dt;
+    BYTE const* const ilimit = args->ilimit;
+
+    /* Copy the arguments to local registers. */
+    ZSTD_memcpy(&bits, &args->bits, sizeof(bits));
+    ZSTD_memcpy(&ip, &args->ip, sizeof(ip));
+    ZSTD_memcpy(&op, &args->op, sizeof(op));
+
+    oend[0] = op[1];
+    oend[1] = op[2];
+    oend[2] = op[3];
+    oend[3] = args->oend;
+
+    assert(MEM_isLittleEndian());
+    assert(!MEM_32bits());
+
+    for (;;) {
+        BYTE* olimit;
+        int stream;
+        int symbol;
+
+        /* Assert loop preconditions */
+#ifndef NDEBUG
+        for (stream = 0; stream < 4; ++stream) {
+            assert(op[stream] <= oend[stream]);
+            assert(ip[stream] >= ilimit);
+        }
+#endif
+        /* Compute olimit */
+        {
+            /* Each loop does 5 table lookups for each of the 4 streams.
+             * Each table lookup consumes up to 11 bits of input, and produces
+             * up to 2 bytes of output.
+             */
+            /* We can consume up to 7 bytes of input per iteration per stream.
+             * We also know that each input pointer is >= ip[0]. So we can run
+             * iters loops before running out of input.
+             */
+            size_t iters = (size_t)(ip[0] - ilimit) / 7;
+            /* Each iteration can produce up to 10 bytes of output per stream.
+             * Each output stream my advance at different rates. So take the
+             * minimum number of safe iterations among all the output streams.
+             */
+            for (stream = 0; stream < 4; ++stream) {
+                size_t const oiters = (size_t)(oend[stream] - op[stream]) / 10;
+                iters = MIN(iters, oiters);
+            }
+
+            /* Each iteration produces at least 5 output symbols. So until
+             * op[3] crosses olimit, we know we haven't executed iters
+             * iterations yet. This saves us maintaining an iters counter,
+             * at the expense of computing the remaining # of iterations
+             * more frequently.
+             */
+            olimit = op[3] + (iters * 5);
+
+            /* Exit the fast decoding loop if we are too close to the end. */
+            if (op[3] + 10 > olimit)
+                break;
+
+            /* Exit the decoding loop if any input pointer has crossed the
+             * previous one. This indicates corruption, and a precondition
+             * to our loop is that ip[i] >= ip[0].
+             */
+            for (stream = 1; stream < 4; ++stream) {
+                if (ip[stream] < ip[stream - 1])
+                    goto _out;
+            }
+        }
 
-    size_t const hSize = HUF_readDTableX2_wksp(DCtx, cSrc, cSrcSize,
-                                               workSpace, wkspSize);
-    if (HUF_isError(hSize)) return hSize;
-    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
-    ip += hSize; cSrcSize -= hSize;
+#ifndef NDEBUG
+        for (stream = 1; stream < 4; ++stream) {
+            assert(ip[stream] >= ip[stream - 1]);
+        }
+#endif
 
-    return HUF_decompress1X2_usingDTable_internal(dst, dstSize, ip, cSrcSize, DCtx, /* bmi2 */ 0);
-}
+        do {
+            /* Do 5 table lookups for each of the first 3 streams */
+            for (symbol = 0; symbol < 5; ++symbol) {
+                for (stream = 0; stream < 3; ++stream) {
+                    int const index = (int)(bits[stream] >> 53);
+                    HUF_DEltX2 const entry = dtable[index];
+                    MEM_write16(op[stream], entry.sequence);
+                    bits[stream] <<= (entry.nbBits);
+                    op[stream] += (entry.length);
+                }
+            }
+            /* Do 1 table lookup from the final stream */
+            {
+                int const index = (int)(bits[3] >> 53);
+                HUF_DEltX2 const entry = dtable[index];
+                MEM_write16(op[3], entry.sequence);
+                bits[3] <<= (entry.nbBits);
+                op[3] += (entry.length);
+            }
+            /* Do 4 table lookups from the final stream & reload bitstreams */
+            for (stream = 0; stream < 4; ++stream) {
+                /* Do a table lookup from the final stream.
+                 * This is interleaved with the reloading to reduce register
+                 * pressure. This shouldn't be necessary, but compilers can
+                 * struggle with codegen with high register pressure.
+                 */
+                {
+                    int const index = (int)(bits[3] >> 53);
+                    HUF_DEltX2 const entry = dtable[index];
+                    MEM_write16(op[3], entry.sequence);
+                    bits[3] <<= (entry.nbBits);
+                    op[3] += (entry.length);
+                }
+                /* Reload the bistreams. The final bitstream must be reloaded
+                 * after the 5th symbol was decoded.
+                 */
+                {
+                    int const ctz = ZSTD_countTrailingZeros64(bits[stream]);
+                    int const nbBits = ctz & 7;
+                    int const nbBytes = ctz >> 3;
+                    ip[stream] -= nbBytes;
+                    bits[stream] = MEM_read64(ip[stream]) | 1;
+                    bits[stream] <<= nbBits;
+                }
+            }
+        } while (op[3] < olimit);
+    }
 
+_out:
 
-size_t HUF_decompress1X2_DCtx(HUF_DTable* DCtx, void* dst, size_t dstSize,
-                              const void* cSrc, size_t cSrcSize)
-{
-    U32 workSpace[HUF_DECOMPRESS_WORKSPACE_SIZE_U32];
-    return HUF_decompress1X2_DCtx_wksp(DCtx, dst, dstSize, cSrc, cSrcSize,
-                                       workSpace, sizeof(workSpace));
+    /* Save the final values of each of the state variables back to args. */
+    ZSTD_memcpy(&args->bits, &bits, sizeof(bits));
+    ZSTD_memcpy(&args->ip, &ip, sizeof(ip));
+    ZSTD_memcpy(&args->op, &op, sizeof(op));
 }
 
-size_t HUF_decompress1X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
-{
-    HUF_CREATE_STATIC_DTABLEX2(DTable, HUF_TABLELOG_MAX);
-    return HUF_decompress1X2_DCtx(DTable, dst, dstSize, cSrc, cSrcSize);
-}
 
-size_t HUF_decompress4X2_usingDTable(
+static HUF_FAST_BMI2_ATTRS size_t
+HUF_decompress4X2_usingDTable_internal_fast(
           void* dst,  size_t dstSize,
     const void* cSrc, size_t cSrcSize,
-    const HUF_DTable* DTable)
+    const HUF_DTable* DTable,
+    HUF_DecompressFastLoopFn loopFn) {
+    void const* dt = DTable + 1;
+    const BYTE* const iend = (const BYTE*)cSrc + 6;
+    BYTE* const oend = (BYTE*)dst + dstSize;
+    HUF_DecompressFastArgs args;
+    {
+        size_t const ret = HUF_DecompressFastArgs_init(&args, dst, dstSize, cSrc, cSrcSize, DTable);
+        FORWARD_IF_ERROR(ret, "Failed to init asm args");
+        if (ret == 0)
+            return 0;
+    }
+
+    assert(args.ip[0] >= args.ilimit);
+    loopFn(&args);
+
+    /* note : op4 already verified within main loop */
+    assert(args.ip[0] >= iend);
+    assert(args.ip[1] >= iend);
+    assert(args.ip[2] >= iend);
+    assert(args.ip[3] >= iend);
+    assert(args.op[3] <= oend);
+    (void)iend;
+
+    /* finish bitStreams one by one */
+    {
+        size_t const segmentSize = (dstSize+3) / 4;
+        BYTE* segmentEnd = (BYTE*)dst;
+        int i;
+        for (i = 0; i < 4; ++i) {
+            BIT_DStream_t bit;
+            if (segmentSize <= (size_t)(oend - segmentEnd))
+                segmentEnd += segmentSize;
+            else
+                segmentEnd = oend;
+            FORWARD_IF_ERROR(HUF_initRemainingDStream(&bit, &args, i, segmentEnd), "corruption");
+            args.op[i] += HUF_decodeStreamX2(args.op[i], &bit, segmentEnd, (HUF_DEltX2 const*)dt, HUF_DECODER_FAST_TABLELOG);
+            if (args.op[i] != segmentEnd)
+                return ERROR(corruption_detected);
+        }
+    }
+
+    /* decoded size */
+    return dstSize;
+}
+
+static size_t HUF_decompress4X2_usingDTable_internal(void* dst, size_t dstSize, void const* cSrc,
+                    size_t cSrcSize, HUF_DTable const* DTable, int flags)
 {
-    DTableDesc dtd = HUF_getDTableDesc(DTable);
-    if (dtd.tableType != 1) return ERROR(GENERIC);
-    return HUF_decompress4X2_usingDTable_internal(dst, dstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
+    HUF_DecompressUsingDTableFn fallbackFn = HUF_decompress4X2_usingDTable_internal_default;
+    HUF_DecompressFastLoopFn loopFn = HUF_decompress4X2_usingDTable_internal_fast_c_loop;
+
+#if DYNAMIC_BMI2
+    if (flags & HUF_flags_bmi2) {
+        fallbackFn = HUF_decompress4X2_usingDTable_internal_bmi2;
+# if ZSTD_ENABLE_ASM_X86_64_BMI2
+        if (!(flags & HUF_flags_disableAsm)) {
+            loopFn = HUF_decompress4X2_usingDTable_internal_fast_asm_loop;
+        }
+# endif
+    } else {
+        return fallbackFn(dst, dstSize, cSrc, cSrcSize, DTable);
+    }
+#endif
+
+#if ZSTD_ENABLE_ASM_X86_64_BMI2 && defined(__BMI2__)
+    if (!(flags & HUF_flags_disableAsm)) {
+        loopFn = HUF_decompress4X2_usingDTable_internal_fast_asm_loop;
+    }
+#endif
+
+    if (!(flags & HUF_flags_disableFast)) {
+        size_t const ret = HUF_decompress4X2_usingDTable_internal_fast(dst, dstSize, cSrc, cSrcSize, DTable, loopFn);
+        if (ret != 0)
+            return ret;
+    }
+    return fallbackFn(dst, dstSize, cSrc, cSrcSize, DTable);
 }
 
-static size_t HUF_decompress4X2_DCtx_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize,
+HUF_DGEN(HUF_decompress1X2_usingDTable_internal)
+
+size_t HUF_decompress1X2_DCtx_wksp(HUF_DTable* DCtx, void* dst, size_t dstSize,
                                    const void* cSrc, size_t cSrcSize,
-                                   void* workSpace, size_t wkspSize, int bmi2)
+                                   void* workSpace, size_t wkspSize, int flags)
 {
     const BYTE* ip = (const BYTE*) cSrc;
 
-    size_t hSize = HUF_readDTableX2_wksp(dctx, cSrc, cSrcSize,
-                                         workSpace, wkspSize);
+    size_t const hSize = HUF_readDTableX2_wksp(DCtx, cSrc, cSrcSize,
+                                               workSpace, wkspSize, flags);
     if (HUF_isError(hSize)) return hSize;
     if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
     ip += hSize; cSrcSize -= hSize;
 
-    return HUF_decompress4X2_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx, bmi2);
+    return HUF_decompress1X2_usingDTable_internal(dst, dstSize, ip, cSrcSize, DCtx, flags);
 }
 
-size_t HUF_decompress4X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
+static size_t HUF_decompress4X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
                                    const void* cSrc, size_t cSrcSize,
-                                   void* workSpace, size_t wkspSize)
+                                   void* workSpace, size_t wkspSize, int flags)
 {
-    return HUF_decompress4X2_DCtx_wksp_bmi2(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, /* bmi2 */ 0);
-}
-
+    const BYTE* ip = (const BYTE*) cSrc;
 
-size_t HUF_decompress4X2_DCtx(HUF_DTable* dctx, void* dst, size_t dstSize,
-                              const void* cSrc, size_t cSrcSize)
-{
-    U32 workSpace[HUF_DECOMPRESS_WORKSPACE_SIZE_U32];
-    return HUF_decompress4X2_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize,
-                                       workSpace, sizeof(workSpace));
-}
+    size_t hSize = HUF_readDTableX2_wksp(dctx, cSrc, cSrcSize,
+                                         workSpace, wkspSize, flags);
+    if (HUF_isError(hSize)) return hSize;
+    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
+    ip += hSize; cSrcSize -= hSize;
 
-size_t HUF_decompress4X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
-{
-    HUF_CREATE_STATIC_DTABLEX2(DTable, HUF_TABLELOG_MAX);
-    return HUF_decompress4X2_DCtx(DTable, dst, dstSize, cSrc, cSrcSize);
+    return HUF_decompress4X2_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx, flags);
 }
 
 #endif /* HUF_FORCE_DECOMPRESS_X1 */
@@ -958,66 +1726,28 @@ size_t HUF_decompress4X2 (void* dst, size_t dstSize, const void* cSrc, size_t cS
 /* Universal decompression selectors */
 /* ***********************************/
 
-size_t HUF_decompress1X_usingDTable(void* dst, size_t maxDstSize,
-                                    const void* cSrc, size_t cSrcSize,
-                                    const HUF_DTable* DTable)
-{
-    DTableDesc const dtd = HUF_getDTableDesc(DTable);
-#if defined(HUF_FORCE_DECOMPRESS_X1)
-    (void)dtd;
-    assert(dtd.tableType == 0);
-    return HUF_decompress1X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
-#elif defined(HUF_FORCE_DECOMPRESS_X2)
-    (void)dtd;
-    assert(dtd.tableType == 1);
-    return HUF_decompress1X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
-#else
-    return dtd.tableType ? HUF_decompress1X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0) :
-                           HUF_decompress1X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
-#endif
-}
-
-size_t HUF_decompress4X_usingDTable(void* dst, size_t maxDstSize,
-                                    const void* cSrc, size_t cSrcSize,
-                                    const HUF_DTable* DTable)
-{
-    DTableDesc const dtd = HUF_getDTableDesc(DTable);
-#if defined(HUF_FORCE_DECOMPRESS_X1)
-    (void)dtd;
-    assert(dtd.tableType == 0);
-    return HUF_decompress4X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
-#elif defined(HUF_FORCE_DECOMPRESS_X2)
-    (void)dtd;
-    assert(dtd.tableType == 1);
-    return HUF_decompress4X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
-#else
-    return dtd.tableType ? HUF_decompress4X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0) :
-                           HUF_decompress4X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, /* bmi2 */ 0);
-#endif
-}
-
 
 #if !defined(HUF_FORCE_DECOMPRESS_X1) && !defined(HUF_FORCE_DECOMPRESS_X2)
 typedef struct { U32 tableTime; U32 decode256Time; } algo_time_t;
-static const algo_time_t algoTime[16 /* Quantization */][3 /* single, double, quad */] =
+static const algo_time_t algoTime[16 /* Quantization */][2 /* single, double */] =
 {
     /* single, double, quad */
-    {{0,0}, {1,1}, {2,2}},  /* Q==0 : impossible */
-    {{0,0}, {1,1}, {2,2}},  /* Q==1 : impossible */
-    {{  38,130}, {1313, 74}, {2151, 38}},   /* Q == 2 : 12-18% */
-    {{ 448,128}, {1353, 74}, {2238, 41}},   /* Q == 3 : 18-25% */
-    {{ 556,128}, {1353, 74}, {2238, 47}},   /* Q == 4 : 25-32% */
-    {{ 714,128}, {1418, 74}, {2436, 53}},   /* Q == 5 : 32-38% */
-    {{ 883,128}, {1437, 74}, {2464, 61}},   /* Q == 6 : 38-44% */
-    {{ 897,128}, {1515, 75}, {2622, 68}},   /* Q == 7 : 44-50% */
-    {{ 926,128}, {1613, 75}, {2730, 75}},   /* Q == 8 : 50-56% */
-    {{ 947,128}, {1729, 77}, {3359, 77}},   /* Q == 9 : 56-62% */
-    {{1107,128}, {2083, 81}, {4006, 84}},   /* Q ==10 : 62-69% */
-    {{1177,128}, {2379, 87}, {4785, 88}},   /* Q ==11 : 69-75% */
-    {{1242,128}, {2415, 93}, {5155, 84}},   /* Q ==12 : 75-81% */
-    {{1349,128}, {2644,106}, {5260,106}},   /* Q ==13 : 81-87% */
-    {{1455,128}, {2422,124}, {4174,124}},   /* Q ==14 : 87-93% */
-    {{ 722,128}, {1891,145}, {1936,146}},   /* Q ==15 : 93-99% */
+    {{0,0}, {1,1}},  /* Q==0 : impossible */
+    {{0,0}, {1,1}},  /* Q==1 : impossible */
+    {{ 150,216}, { 381,119}},   /* Q == 2 : 12-18% */
+    {{ 170,205}, { 514,112}},   /* Q == 3 : 18-25% */
+    {{ 177,199}, { 539,110}},   /* Q == 4 : 25-32% */
+    {{ 197,194}, { 644,107}},   /* Q == 5 : 32-38% */
+    {{ 221,192}, { 735,107}},   /* Q == 6 : 38-44% */
+    {{ 256,189}, { 881,106}},   /* Q == 7 : 44-50% */
+    {{ 359,188}, {1167,109}},   /* Q == 8 : 50-56% */
+    {{ 582,187}, {1570,114}},   /* Q == 9 : 56-62% */
+    {{ 688,187}, {1712,122}},   /* Q ==10 : 62-69% */
+    {{ 825,186}, {1965,136}},   /* Q ==11 : 69-75% */
+    {{ 976,185}, {2131,150}},   /* Q ==12 : 75-81% */
+    {{1180,186}, {2070,175}},   /* Q ==13 : 81-87% */
+    {{1377,185}, {1731,202}},   /* Q ==14 : 87-93% */
+    {{1412,185}, {1695,202}},   /* Q ==15 : 93-99% */
 };
 #endif
 
@@ -1044,188 +1774,92 @@ U32 HUF_selectDecoder (size_t dstSize, size_t cSrcSize)
         U32 const D256 = (U32)(dstSize >> 8);
         U32 const DTime0 = algoTime[Q][0].tableTime + (algoTime[Q][0].decode256Time * D256);
         U32 DTime1 = algoTime[Q][1].tableTime + (algoTime[Q][1].decode256Time * D256);
-        DTime1 += DTime1 >> 3;  /* advantage to algorithm using less memory, to reduce cache eviction */
+        DTime1 += DTime1 >> 5;  /* small advantage to algorithm using less memory, to reduce cache eviction */
         return DTime1 < DTime0;
     }
 #endif
 }
 
-
-typedef size_t (*decompressionAlgo)(void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);
-
-size_t HUF_decompress (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
-{
-#if !defined(HUF_FORCE_DECOMPRESS_X1) && !defined(HUF_FORCE_DECOMPRESS_X2)
-    static const decompressionAlgo decompress[2] = { HUF_decompress4X1, HUF_decompress4X2 };
-#endif
-
-    /* validation checks */
-    if (dstSize == 0) return ERROR(dstSize_tooSmall);
-    if (cSrcSize > dstSize) return ERROR(corruption_detected);   /* invalid */
-    if (cSrcSize == dstSize) { memcpy(dst, cSrc, dstSize); return dstSize; }   /* not compressed */
-    if (cSrcSize == 1) { memset(dst, *(const BYTE*)cSrc, dstSize); return dstSize; }   /* RLE */
-
-    {   U32 const algoNb = HUF_selectDecoder(dstSize, cSrcSize);
-#if defined(HUF_FORCE_DECOMPRESS_X1)
-        (void)algoNb;
-        assert(algoNb == 0);
-        return HUF_decompress4X1(dst, dstSize, cSrc, cSrcSize);
-#elif defined(HUF_FORCE_DECOMPRESS_X2)
-        (void)algoNb;
-        assert(algoNb == 1);
-        return HUF_decompress4X2(dst, dstSize, cSrc, cSrcSize);
-#else
-        return decompress[algoNb](dst, dstSize, cSrc, cSrcSize);
-#endif
-    }
-}
-
-size_t HUF_decompress4X_DCtx (HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
-{
-    /* validation checks */
-    if (dstSize == 0) return ERROR(dstSize_tooSmall);
-    if (cSrcSize > dstSize) return ERROR(corruption_detected);   /* invalid */
-    if (cSrcSize == dstSize) { memcpy(dst, cSrc, dstSize); return dstSize; }   /* not compressed */
-    if (cSrcSize == 1) { memset(dst, *(const BYTE*)cSrc, dstSize); return dstSize; }   /* RLE */
-
-    {   U32 const algoNb = HUF_selectDecoder(dstSize, cSrcSize);
-#if defined(HUF_FORCE_DECOMPRESS_X1)
-        (void)algoNb;
-        assert(algoNb == 0);
-        return HUF_decompress4X1_DCtx(dctx, dst, dstSize, cSrc, cSrcSize);
-#elif defined(HUF_FORCE_DECOMPRESS_X2)
-        (void)algoNb;
-        assert(algoNb == 1);
-        return HUF_decompress4X2_DCtx(dctx, dst, dstSize, cSrc, cSrcSize);
-#else
-        return algoNb ? HUF_decompress4X2_DCtx(dctx, dst, dstSize, cSrc, cSrcSize) :
-                        HUF_decompress4X1_DCtx(dctx, dst, dstSize, cSrc, cSrcSize) ;
-#endif
-    }
-}
-
-size_t HUF_decompress4X_hufOnly(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize)
-{
-    U32 workSpace[HUF_DECOMPRESS_WORKSPACE_SIZE_U32];
-    return HUF_decompress4X_hufOnly_wksp(dctx, dst, dstSize, cSrc, cSrcSize,
-                                         workSpace, sizeof(workSpace));
-}
-
-
-size_t HUF_decompress4X_hufOnly_wksp(HUF_DTable* dctx, void* dst,
-                                     size_t dstSize, const void* cSrc,
-                                     size_t cSrcSize, void* workSpace,
-                                     size_t wkspSize)
-{
-    /* validation checks */
-    if (dstSize == 0) return ERROR(dstSize_tooSmall);
-    if (cSrcSize == 0) return ERROR(corruption_detected);
-
-    {   U32 const algoNb = HUF_selectDecoder(dstSize, cSrcSize);
-#if defined(HUF_FORCE_DECOMPRESS_X1)
-        (void)algoNb;
-        assert(algoNb == 0);
-        return HUF_decompress4X1_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize);
-#elif defined(HUF_FORCE_DECOMPRESS_X2)
-        (void)algoNb;
-        assert(algoNb == 1);
-        return HUF_decompress4X2_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize);
-#else
-        return algoNb ? HUF_decompress4X2_DCtx_wksp(dctx, dst, dstSize, cSrc,
-                            cSrcSize, workSpace, wkspSize):
-                        HUF_decompress4X1_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize);
-#endif
-    }
-}
-
 size_t HUF_decompress1X_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
                                   const void* cSrc, size_t cSrcSize,
-                                  void* workSpace, size_t wkspSize)
+                                  void* workSpace, size_t wkspSize, int flags)
 {
     /* validation checks */
     if (dstSize == 0) return ERROR(dstSize_tooSmall);
     if (cSrcSize > dstSize) return ERROR(corruption_detected);   /* invalid */
-    if (cSrcSize == dstSize) { memcpy(dst, cSrc, dstSize); return dstSize; }   /* not compressed */
-    if (cSrcSize == 1) { memset(dst, *(const BYTE*)cSrc, dstSize); return dstSize; }   /* RLE */
+    if (cSrcSize == dstSize) { ZSTD_memcpy(dst, cSrc, dstSize); return dstSize; }   /* not compressed */
+    if (cSrcSize == 1) { ZSTD_memset(dst, *(const BYTE*)cSrc, dstSize); return dstSize; }   /* RLE */
 
     {   U32 const algoNb = HUF_selectDecoder(dstSize, cSrcSize);
 #if defined(HUF_FORCE_DECOMPRESS_X1)
         (void)algoNb;
         assert(algoNb == 0);
         return HUF_decompress1X1_DCtx_wksp(dctx, dst, dstSize, cSrc,
-                                cSrcSize, workSpace, wkspSize);
+                                cSrcSize, workSpace, wkspSize, flags);
 #elif defined(HUF_FORCE_DECOMPRESS_X2)
         (void)algoNb;
         assert(algoNb == 1);
         return HUF_decompress1X2_DCtx_wksp(dctx, dst, dstSize, cSrc,
-                                cSrcSize, workSpace, wkspSize);
+                                cSrcSize, workSpace, wkspSize, flags);
 #else
         return algoNb ? HUF_decompress1X2_DCtx_wksp(dctx, dst, dstSize, cSrc,
-                                cSrcSize, workSpace, wkspSize):
+                                cSrcSize, workSpace, wkspSize, flags):
                         HUF_decompress1X1_DCtx_wksp(dctx, dst, dstSize, cSrc,
-                                cSrcSize, workSpace, wkspSize);
+                                cSrcSize, workSpace, wkspSize, flags);
 #endif
     }
 }
 
-size_t HUF_decompress1X_DCtx(HUF_DTable* dctx, void* dst, size_t dstSize,
-                             const void* cSrc, size_t cSrcSize)
-{
-    U32 workSpace[HUF_DECOMPRESS_WORKSPACE_SIZE_U32];
-    return HUF_decompress1X_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize,
-                                      workSpace, sizeof(workSpace));
-}
-
 
-size_t HUF_decompress1X_usingDTable_bmi2(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int bmi2)
+size_t HUF_decompress1X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int flags)
 {
     DTableDesc const dtd = HUF_getDTableDesc(DTable);
 #if defined(HUF_FORCE_DECOMPRESS_X1)
     (void)dtd;
     assert(dtd.tableType == 0);
-    return HUF_decompress1X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, bmi2);
+    return HUF_decompress1X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
 #elif defined(HUF_FORCE_DECOMPRESS_X2)
     (void)dtd;
     assert(dtd.tableType == 1);
-    return HUF_decompress1X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, bmi2);
+    return HUF_decompress1X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
 #else
-    return dtd.tableType ? HUF_decompress1X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, bmi2) :
-                           HUF_decompress1X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, bmi2);
+    return dtd.tableType ? HUF_decompress1X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags) :
+                           HUF_decompress1X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
 #endif
 }
 
 #ifndef HUF_FORCE_DECOMPRESS_X2
-size_t HUF_decompress1X1_DCtx_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int bmi2)
+size_t HUF_decompress1X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags)
 {
     const BYTE* ip = (const BYTE*) cSrc;
 
-    size_t const hSize = HUF_readDTableX1_wksp(dctx, cSrc, cSrcSize, workSpace, wkspSize);
+    size_t const hSize = HUF_readDTableX1_wksp(dctx, cSrc, cSrcSize, workSpace, wkspSize, flags);
     if (HUF_isError(hSize)) return hSize;
     if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
     ip += hSize; cSrcSize -= hSize;
 
-    return HUF_decompress1X1_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx, bmi2);
+    return HUF_decompress1X1_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx, flags);
 }
 #endif
 
-size_t HUF_decompress4X_usingDTable_bmi2(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int bmi2)
+size_t HUF_decompress4X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int flags)
 {
     DTableDesc const dtd = HUF_getDTableDesc(DTable);
 #if defined(HUF_FORCE_DECOMPRESS_X1)
     (void)dtd;
     assert(dtd.tableType == 0);
-    return HUF_decompress4X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, bmi2);
+    return HUF_decompress4X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
 #elif defined(HUF_FORCE_DECOMPRESS_X2)
     (void)dtd;
     assert(dtd.tableType == 1);
-    return HUF_decompress4X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, bmi2);
+    return HUF_decompress4X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
 #else
-    return dtd.tableType ? HUF_decompress4X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, bmi2) :
-                           HUF_decompress4X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, bmi2);
+    return dtd.tableType ? HUF_decompress4X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags) :
+                           HUF_decompress4X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
 #endif
 }
 
-size_t HUF_decompress4X_hufOnly_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int bmi2)
+size_t HUF_decompress4X_hufOnly_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags)
 {
     /* validation checks */
     if (dstSize == 0) return ERROR(dstSize_tooSmall);
@@ -1235,14 +1869,14 @@ size_t HUF_decompress4X_hufOnly_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t ds
 #if defined(HUF_FORCE_DECOMPRESS_X1)
         (void)algoNb;
         assert(algoNb == 0);
-        return HUF_decompress4X1_DCtx_wksp_bmi2(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, bmi2);
+        return HUF_decompress4X1_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, flags);
 #elif defined(HUF_FORCE_DECOMPRESS_X2)
         (void)algoNb;
         assert(algoNb == 1);
-        return HUF_decompress4X2_DCtx_wksp_bmi2(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, bmi2);
+        return HUF_decompress4X2_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, flags);
 #else
-        return algoNb ? HUF_decompress4X2_DCtx_wksp_bmi2(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, bmi2) :
-                        HUF_decompress4X1_DCtx_wksp_bmi2(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, bmi2);
+        return algoNb ? HUF_decompress4X2_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, flags) :
+                        HUF_decompress4X1_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, flags);
 #endif
     }
 }
diff --git a/contrib/zstd/mem.h b/contrib/zstd/mem.h
index 89c8aea7d..98dd47a04 100644
--- a/contrib/zstd/mem.h
+++ b/contrib/zstd/mem.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -18,8 +18,10 @@ extern "C" {
 /*-****************************************
 *  Dependencies
 ******************************************/
-#include <stddef.h>     /* size_t, ptrdiff_t */
-#include <string.h>     /* memcpy */
+#include <stddef.h>  /* size_t, ptrdiff_t */
+#include "compiler.h"  /* __has_builtin */
+#include "debug.h"  /* DEBUG_STATIC_ASSERT */
+#include "zstd_deps.h"  /* ZSTD_memcpy */
 
 
 /*-****************************************
@@ -39,94 +41,18 @@ extern "C" {
 #  define MEM_STATIC static  /* this version may generate warnings for unused static functions; disable the relevant warning */
 #endif
 
-#ifndef __has_builtin
-#  define __has_builtin(x) 0  /* compat. with non-clang compilers */
-#endif
-
-/* code only tested on 32 and 64 bits systems */
-#define MEM_STATIC_ASSERT(c)   { enum { MEM_static_assert = 1/(int)(!!(c)) }; }
-MEM_STATIC void MEM_check(void) { MEM_STATIC_ASSERT((sizeof(size_t)==4) || (sizeof(size_t)==8)); }
-
-/* detects whether we are being compiled under msan */
-#if defined (__has_feature)
-#  if __has_feature(memory_sanitizer)
-#    define MEMORY_SANITIZER 1
-#  endif
-#endif
-
-#if defined (MEMORY_SANITIZER)
-/* Not all platforms that support msan provide sanitizers/msan_interface.h.
- * We therefore declare the functions we need ourselves, rather than trying to
- * include the header file... */
-
-#include <stdint.h> /* intptr_t */
-
-/* Make memory region fully initialized (without changing its contents). */
-void __msan_unpoison(const volatile void *a, size_t size);
-
-/* Make memory region fully uninitialized (without changing its contents).
-   This is a legacy interface that does not update origin information. Use
-   __msan_allocated_memory() instead. */
-void __msan_poison(const volatile void *a, size_t size);
-
-/* Returns the offset of the first (at least partially) poisoned byte in the
-   memory range, or -1 if the whole range is good. */
-intptr_t __msan_test_shadow(const volatile void *x, size_t size);
-#endif
-
-/* detects whether we are being compiled under asan */
-#if defined (__has_feature)
-#  if __has_feature(address_sanitizer)
-#    define ADDRESS_SANITIZER 1
-#  endif
-#elif defined(__SANITIZE_ADDRESS__)
-#  define ADDRESS_SANITIZER 1
-#endif
-
-#if defined (ADDRESS_SANITIZER)
-/* Not all platforms that support asan provide sanitizers/asan_interface.h.
- * We therefore declare the functions we need ourselves, rather than trying to
- * include the header file... */
-
-/**
- * Marks a memory region (<c>[addr, addr+size)</c>) as unaddressable.
- *
- * This memory must be previously allocated by your program. Instrumented
- * code is forbidden from accessing addresses in this region until it is
- * unpoisoned. This function is not guaranteed to poison the entire region -
- * it could poison only a subregion of <c>[addr, addr+size)</c> due to ASan
- * alignment restrictions.
- *
- * \note This function is not thread-safe because no two threads can poison or
- * unpoison memory in the same memory region simultaneously.
- *
- * \param addr Start of memory region.
- * \param size Size of memory region. */
-void __asan_poison_memory_region(void const volatile *addr, size_t size);
-
-/**
- * Marks a memory region (<c>[addr, addr+size)</c>) as addressable.
- *
- * This memory must be previously allocated by your program. Accessing
- * addresses in this region is allowed until this region is poisoned again.
- * This function could unpoison a super-region of <c>[addr, addr+size)</c> due
- * to ASan alignment restrictions.
- *
- * \note This function is not thread-safe because no two threads can
- * poison or unpoison memory in the same memory region simultaneously.
- *
- * \param addr Start of memory region.
- * \param size Size of memory region. */
-void __asan_unpoison_memory_region(void const volatile *addr, size_t size);
-#endif
-
-
 /*-**************************************************************
 *  Basic Types
 *****************************************************************/
 #if  !defined (__VMS) && (defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
-# include <stdint.h>
+#  if defined(_AIX)
+#    include <inttypes.h>
+#  else
+#    include <stdint.h> /* intptr_t */
+#  endif
   typedef   uint8_t BYTE;
+  typedef   uint8_t U8;
+  typedef    int8_t S8;
   typedef  uint16_t U16;
   typedef   int16_t S16;
   typedef  uint32_t U32;
@@ -139,6 +65,8 @@ void __asan_unpoison_memory_region(void const volatile *addr, size_t size);
 #  error "this implementation requires char to be exactly 8-bit type"
 #endif
   typedef unsigned char      BYTE;
+  typedef unsigned char      U8;
+  typedef   signed char      S8;
 #if USHRT_MAX != 65535
 #  error "this implementation requires short to be exactly 16-bit type"
 #endif
@@ -157,25 +85,63 @@ void __asan_unpoison_memory_region(void const volatile *addr, size_t size);
 
 
 /*-**************************************************************
-*  Memory I/O
+*  Memory I/O API
+*****************************************************************/
+/*=== Static platform detection ===*/
+MEM_STATIC unsigned MEM_32bits(void);
+MEM_STATIC unsigned MEM_64bits(void);
+MEM_STATIC unsigned MEM_isLittleEndian(void);
+
+/*=== Native unaligned read/write ===*/
+MEM_STATIC U16 MEM_read16(const void* memPtr);
+MEM_STATIC U32 MEM_read32(const void* memPtr);
+MEM_STATIC U64 MEM_read64(const void* memPtr);
+MEM_STATIC size_t MEM_readST(const void* memPtr);
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value);
+MEM_STATIC void MEM_write32(void* memPtr, U32 value);
+MEM_STATIC void MEM_write64(void* memPtr, U64 value);
+
+/*=== Little endian unaligned read/write ===*/
+MEM_STATIC U16 MEM_readLE16(const void* memPtr);
+MEM_STATIC U32 MEM_readLE24(const void* memPtr);
+MEM_STATIC U32 MEM_readLE32(const void* memPtr);
+MEM_STATIC U64 MEM_readLE64(const void* memPtr);
+MEM_STATIC size_t MEM_readLEST(const void* memPtr);
+
+MEM_STATIC void MEM_writeLE16(void* memPtr, U16 val);
+MEM_STATIC void MEM_writeLE24(void* memPtr, U32 val);
+MEM_STATIC void MEM_writeLE32(void* memPtr, U32 val32);
+MEM_STATIC void MEM_writeLE64(void* memPtr, U64 val64);
+MEM_STATIC void MEM_writeLEST(void* memPtr, size_t val);
+
+/*=== Big endian unaligned read/write ===*/
+MEM_STATIC U32 MEM_readBE32(const void* memPtr);
+MEM_STATIC U64 MEM_readBE64(const void* memPtr);
+MEM_STATIC size_t MEM_readBEST(const void* memPtr);
+
+MEM_STATIC void MEM_writeBE32(void* memPtr, U32 val32);
+MEM_STATIC void MEM_writeBE64(void* memPtr, U64 val64);
+MEM_STATIC void MEM_writeBEST(void* memPtr, size_t val);
+
+/*=== Byteswap ===*/
+MEM_STATIC U32 MEM_swap32(U32 in);
+MEM_STATIC U64 MEM_swap64(U64 in);
+MEM_STATIC size_t MEM_swapST(size_t in);
+
+
+/*-**************************************************************
+*  Memory I/O Implementation
 *****************************************************************/
-/* MEM_FORCE_MEMORY_ACCESS :
- * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
- * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
- * The below switch allow to select different access method for improved performance.
- * Method 0 (default) : use `memcpy()`. Safe and portable.
- * Method 1 : `__packed` statement. It depends on compiler extension (i.e., not portable).
- *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
+/* MEM_FORCE_MEMORY_ACCESS : For accessing unaligned memory:
+ * Method 0 : always use `memcpy()`. Safe and portable.
+ * Method 1 : Use compiler extension to set unaligned access.
  * Method 2 : direct access. This method is portable but violate C standard.
  *            It can generate buggy code on targets depending on alignment.
- *            In some circumstances, it's the only known way to get the most performance (i.e. GCC + ARMv6)
- * See http://fastcompression.blogspot.fr/2015/08/accessing-unaligned-memory.html for details.
- * Prefer these methods in priority order (0 > 1 > 2)
+ * Default  : method 1 if supported, else method 0
  */
 #ifndef MEM_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
-#  if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
-#    define MEM_FORCE_MEMORY_ACCESS 2
-#  elif defined(__INTEL_COMPILER) || defined(__GNUC__) || defined(__ICCARM__)
+#  ifdef __GNUC__
 #    define MEM_FORCE_MEMORY_ACCESS 1
 #  endif
 #endif
@@ -185,8 +151,22 @@ MEM_STATIC unsigned MEM_64bits(void) { return sizeof(size_t)==8; }
 
 MEM_STATIC unsigned MEM_isLittleEndian(void)
 {
+#if defined(__BYTE_ORDER__) && defined(__ORDER_LITTLE_ENDIAN__) && (__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
+    return 1;
+#elif defined(__BYTE_ORDER__) && defined(__ORDER_BIG_ENDIAN__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+    return 0;
+#elif defined(__clang__) && __LITTLE_ENDIAN__
+    return 1;
+#elif defined(__clang__) && __BIG_ENDIAN__
+    return 0;
+#elif defined(_MSC_VER) && (_M_AMD64 || _M_IX86)
+    return 1;
+#elif defined(__DMC__) && defined(_M_IX86)
+    return 1;
+#else
     const union { U32 u; BYTE c[4]; } one = { 1 };   /* don't use static : performance detrimental  */
     return one.c[0];
+#endif
 }
 
 #if defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==2)
@@ -204,30 +184,19 @@ MEM_STATIC void MEM_write64(void* memPtr, U64 value) { *(U64*)memPtr = value; }
 
 #elif defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==1)
 
-/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
-/* currently only defined for gcc and icc */
-#if defined(_MSC_VER) || (defined(__INTEL_COMPILER) && defined(WIN32))
-    __pragma( pack(push, 1) )
-    typedef struct { U16 v; } unalign16;
-    typedef struct { U32 v; } unalign32;
-    typedef struct { U64 v; } unalign64;
-    typedef struct { size_t v; } unalignArch;
-    __pragma( pack(pop) )
-#else
-    typedef struct { U16 v; } __attribute__((packed)) unalign16;
-    typedef struct { U32 v; } __attribute__((packed)) unalign32;
-    typedef struct { U64 v; } __attribute__((packed)) unalign64;
-    typedef struct { size_t v; } __attribute__((packed)) unalignArch;
-#endif
+typedef __attribute__((aligned(1))) U16 unalign16;
+typedef __attribute__((aligned(1))) U32 unalign32;
+typedef __attribute__((aligned(1))) U64 unalign64;
+typedef __attribute__((aligned(1))) size_t unalignArch;
 
-MEM_STATIC U16 MEM_read16(const void* ptr) { return ((const unalign16*)ptr)->v; }
-MEM_STATIC U32 MEM_read32(const void* ptr) { return ((const unalign32*)ptr)->v; }
-MEM_STATIC U64 MEM_read64(const void* ptr) { return ((const unalign64*)ptr)->v; }
-MEM_STATIC size_t MEM_readST(const void* ptr) { return ((const unalignArch*)ptr)->v; }
+MEM_STATIC U16 MEM_read16(const void* ptr) { return *(const unalign16*)ptr; }
+MEM_STATIC U32 MEM_read32(const void* ptr) { return *(const unalign32*)ptr; }
+MEM_STATIC U64 MEM_read64(const void* ptr) { return *(const unalign64*)ptr; }
+MEM_STATIC size_t MEM_readST(const void* ptr) { return *(const unalignArch*)ptr; }
 
-MEM_STATIC void MEM_write16(void* memPtr, U16 value) { ((unalign16*)memPtr)->v = value; }
-MEM_STATIC void MEM_write32(void* memPtr, U32 value) { ((unalign32*)memPtr)->v = value; }
-MEM_STATIC void MEM_write64(void* memPtr, U64 value) { ((unalign64*)memPtr)->v = value; }
+MEM_STATIC void MEM_write16(void* memPtr, U16 value) { *(unalign16*)memPtr = value; }
+MEM_STATIC void MEM_write32(void* memPtr, U32 value) { *(unalign32*)memPtr = value; }
+MEM_STATIC void MEM_write64(void* memPtr, U64 value) { *(unalign64*)memPtr = value; }
 
 #else
 
@@ -236,41 +205,49 @@ MEM_STATIC void MEM_write64(void* memPtr, U64 value) { ((unalign64*)memPtr)->v =
 
 MEM_STATIC U16 MEM_read16(const void* memPtr)
 {
-    U16 val; memcpy(&val, memPtr, sizeof(val)); return val;
+    U16 val; ZSTD_memcpy(&val, memPtr, sizeof(val)); return val;
 }
 
 MEM_STATIC U32 MEM_read32(const void* memPtr)
 {
-    U32 val; memcpy(&val, memPtr, sizeof(val)); return val;
+    U32 val; ZSTD_memcpy(&val, memPtr, sizeof(val)); return val;
 }
 
 MEM_STATIC U64 MEM_read64(const void* memPtr)
 {
-    U64 val; memcpy(&val, memPtr, sizeof(val)); return val;
+    U64 val; ZSTD_memcpy(&val, memPtr, sizeof(val)); return val;
 }
 
 MEM_STATIC size_t MEM_readST(const void* memPtr)
 {
-    size_t val; memcpy(&val, memPtr, sizeof(val)); return val;
+    size_t val; ZSTD_memcpy(&val, memPtr, sizeof(val)); return val;
 }
 
 MEM_STATIC void MEM_write16(void* memPtr, U16 value)
 {
-    memcpy(memPtr, &value, sizeof(value));
+    ZSTD_memcpy(memPtr, &value, sizeof(value));
 }
 
 MEM_STATIC void MEM_write32(void* memPtr, U32 value)
 {
-    memcpy(memPtr, &value, sizeof(value));
+    ZSTD_memcpy(memPtr, &value, sizeof(value));
 }
 
 MEM_STATIC void MEM_write64(void* memPtr, U64 value)
 {
-    memcpy(memPtr, &value, sizeof(value));
+    ZSTD_memcpy(memPtr, &value, sizeof(value));
 }
 
 #endif /* MEM_FORCE_MEMORY_ACCESS */
 
+MEM_STATIC U32 MEM_swap32_fallback(U32 in)
+{
+    return  ((in << 24) & 0xff000000 ) |
+            ((in <<  8) & 0x00ff0000 ) |
+            ((in >>  8) & 0x0000ff00 ) |
+            ((in >> 24) & 0x000000ff );
+}
+
 MEM_STATIC U32 MEM_swap32(U32 in)
 {
 #if defined(_MSC_VER)     /* Visual Studio */
@@ -279,22 +256,13 @@ MEM_STATIC U32 MEM_swap32(U32 in)
   || (defined(__clang__) && __has_builtin(__builtin_bswap32))
     return __builtin_bswap32(in);
 #else
-    return  ((in << 24) & 0xff000000 ) |
-            ((in <<  8) & 0x00ff0000 ) |
-            ((in >>  8) & 0x0000ff00 ) |
-            ((in >> 24) & 0x000000ff );
+    return MEM_swap32_fallback(in);
 #endif
 }
 
-MEM_STATIC U64 MEM_swap64(U64 in)
+MEM_STATIC U64 MEM_swap64_fallback(U64 in)
 {
-#if defined(_MSC_VER)     /* Visual Studio */
-    return _byteswap_uint64(in);
-#elif (defined (__GNUC__) && (__GNUC__ * 100 + __GNUC_MINOR__ >= 403)) \
-  || (defined(__clang__) && __has_builtin(__builtin_bswap64))
-    return __builtin_bswap64(in);
-#else
-    return  ((in << 56) & 0xff00000000000000ULL) |
+     return  ((in << 56) & 0xff00000000000000ULL) |
             ((in << 40) & 0x00ff000000000000ULL) |
             ((in << 24) & 0x0000ff0000000000ULL) |
             ((in << 8)  & 0x000000ff00000000ULL) |
@@ -302,6 +270,17 @@ MEM_STATIC U64 MEM_swap64(U64 in)
             ((in >> 24) & 0x0000000000ff0000ULL) |
             ((in >> 40) & 0x000000000000ff00ULL) |
             ((in >> 56) & 0x00000000000000ffULL);
+}
+
+MEM_STATIC U64 MEM_swap64(U64 in)
+{
+#if defined(_MSC_VER)     /* Visual Studio */
+    return _byteswap_uint64(in);
+#elif (defined (__GNUC__) && (__GNUC__ * 100 + __GNUC_MINOR__ >= 403)) \
+  || (defined(__clang__) && __has_builtin(__builtin_bswap64))
+    return __builtin_bswap64(in);
+#else
+    return MEM_swap64_fallback(in);
 #endif
 }
 
@@ -338,7 +317,7 @@ MEM_STATIC void MEM_writeLE16(void* memPtr, U16 val)
 
 MEM_STATIC U32 MEM_readLE24(const void* memPtr)
 {
-    return MEM_readLE16(memPtr) + (((const BYTE*)memPtr)[2] << 16);
+    return (U32)MEM_readLE16(memPtr) + ((U32)(((const BYTE*)memPtr)[2]) << 16);
 }
 
 MEM_STATIC void MEM_writeLE24(void* memPtr, U32 val)
@@ -445,6 +424,9 @@ MEM_STATIC void MEM_writeBEST(void* memPtr, size_t val)
         MEM_writeBE64(memPtr, (U64)val);
 }
 
+/* code only tested on 32 and 64 bits systems */
+MEM_STATIC void MEM_check(void) { DEBUG_STATIC_ASSERT((sizeof(size_t)==4) || (sizeof(size_t)==8)); }
+
 
 #if defined (__cplusplus)
 }
diff --git a/contrib/zstd/pool.c b/contrib/zstd/pool.c
index aa4b4de0d..f3d9d0854 100644
--- a/contrib/zstd/pool.c
+++ b/contrib/zstd/pool.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -10,9 +10,9 @@
 
 
 /* ======   Dependencies   ======= */
-#include <stddef.h>    /* size_t */
+#include "zstd_deps.h" /* size_t */
 #include "debug.h"     /* assert */
-#include "zstd_internal.h"  /* ZSTD_malloc, ZSTD_free */
+#include "zstd_internal.h"  /* ZSTD_customCalloc, ZSTD_customFree */
 #include "pool.h"
 
 /* ======   Compiler specifics   ====== */
@@ -86,7 +86,7 @@ static void* POOL_thread(void* opaque) {
         {   POOL_job const job = ctx->queue[ctx->queueHead];
             ctx->queueHead = (ctx->queueHead + 1) % ctx->queueSize;
             ctx->numThreadsBusy++;
-            ctx->queueEmpty = ctx->queueHead == ctx->queueTail;
+            ctx->queueEmpty = (ctx->queueHead == ctx->queueTail);
             /* Unlock the mutex, signal a pusher, and run the job */
             ZSTD_pthread_cond_signal(&ctx->queuePushCond);
             ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
@@ -96,33 +96,37 @@ static void* POOL_thread(void* opaque) {
             /* If the intended queue size was 0, signal after finishing job */
             ZSTD_pthread_mutex_lock(&ctx->queueMutex);
             ctx->numThreadsBusy--;
-            if (ctx->queueSize == 1) {
-                ZSTD_pthread_cond_signal(&ctx->queuePushCond);
-            }
+            ZSTD_pthread_cond_signal(&ctx->queuePushCond);
             ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
         }
     }  /* for (;;) */
     assert(0);  /* Unreachable */
 }
 
+/* ZSTD_createThreadPool() : public access point */
+POOL_ctx* ZSTD_createThreadPool(size_t numThreads) {
+    return POOL_create (numThreads, 0);
+}
+
 POOL_ctx* POOL_create(size_t numThreads, size_t queueSize) {
     return POOL_create_advanced(numThreads, queueSize, ZSTD_defaultCMem);
 }
 
 POOL_ctx* POOL_create_advanced(size_t numThreads, size_t queueSize,
-                               ZSTD_customMem customMem) {
+                               ZSTD_customMem customMem)
+{
     POOL_ctx* ctx;
     /* Check parameters */
     if (!numThreads) { return NULL; }
     /* Allocate the context and zero initialize */
-    ctx = (POOL_ctx*)ZSTD_calloc(sizeof(POOL_ctx), customMem);
+    ctx = (POOL_ctx*)ZSTD_customCalloc(sizeof(POOL_ctx), customMem);
     if (!ctx) { return NULL; }
     /* Initialize the job queue.
      * It needs one extra space since one space is wasted to differentiate
      * empty and full queues.
      */
     ctx->queueSize = queueSize + 1;
-    ctx->queue = (POOL_job*)ZSTD_malloc(ctx->queueSize * sizeof(POOL_job), customMem);
+    ctx->queue = (POOL_job*)ZSTD_customCalloc(ctx->queueSize * sizeof(POOL_job), customMem);
     ctx->queueHead = 0;
     ctx->queueTail = 0;
     ctx->numThreadsBusy = 0;
@@ -136,7 +140,7 @@ POOL_ctx* POOL_create_advanced(size_t numThreads, size_t queueSize,
     }
     ctx->shutdown = 0;
     /* Allocate space for the thread handles */
-    ctx->threads = (ZSTD_pthread_t*)ZSTD_malloc(numThreads * sizeof(ZSTD_pthread_t), customMem);
+    ctx->threads = (ZSTD_pthread_t*)ZSTD_customCalloc(numThreads * sizeof(ZSTD_pthread_t), customMem);
     ctx->threadCapacity = 0;
     ctx->customMem = customMem;
     /* Check for errors */
@@ -169,7 +173,7 @@ static void POOL_join(POOL_ctx* ctx) {
     /* Join all of the threads */
     {   size_t i;
         for (i = 0; i < ctx->threadCapacity; ++i) {
-            ZSTD_pthread_join(ctx->threads[i], NULL);  /* note : could fail */
+            ZSTD_pthread_join(ctx->threads[i]);  /* note : could fail */
     }   }
 }
 
@@ -179,14 +183,27 @@ void POOL_free(POOL_ctx *ctx) {
     ZSTD_pthread_mutex_destroy(&ctx->queueMutex);
     ZSTD_pthread_cond_destroy(&ctx->queuePushCond);
     ZSTD_pthread_cond_destroy(&ctx->queuePopCond);
-    ZSTD_free(ctx->queue, ctx->customMem);
-    ZSTD_free(ctx->threads, ctx->customMem);
-    ZSTD_free(ctx, ctx->customMem);
+    ZSTD_customFree(ctx->queue, ctx->customMem);
+    ZSTD_customFree(ctx->threads, ctx->customMem);
+    ZSTD_customFree(ctx, ctx->customMem);
 }
 
+/*! POOL_joinJobs() :
+ *  Waits for all queued jobs to finish executing.
+ */
+void POOL_joinJobs(POOL_ctx* ctx) {
+    ZSTD_pthread_mutex_lock(&ctx->queueMutex);
+    while(!ctx->queueEmpty || ctx->numThreadsBusy > 0) {
+        ZSTD_pthread_cond_wait(&ctx->queuePushCond, &ctx->queueMutex);
+    }
+    ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
+}
 
+void ZSTD_freeThreadPool (ZSTD_threadPool* pool) {
+  POOL_free (pool);
+}
 
-size_t POOL_sizeof(POOL_ctx *ctx) {
+size_t POOL_sizeof(const POOL_ctx* ctx) {
     if (ctx==NULL) return 0;  /* supports sizeof NULL */
     return sizeof(*ctx)
         + ctx->queueSize * sizeof(POOL_job)
@@ -203,11 +220,11 @@ static int POOL_resize_internal(POOL_ctx* ctx, size_t numThreads)
         return 0;
     }
     /* numThreads > threadCapacity */
-    {   ZSTD_pthread_t* const threadPool = (ZSTD_pthread_t*)ZSTD_malloc(numThreads * sizeof(ZSTD_pthread_t), ctx->customMem);
+    {   ZSTD_pthread_t* const threadPool = (ZSTD_pthread_t*)ZSTD_customCalloc(numThreads * sizeof(ZSTD_pthread_t), ctx->customMem);
         if (!threadPool) return 1;
         /* replace existing thread pool */
-        memcpy(threadPool, ctx->threads, ctx->threadCapacity * sizeof(*threadPool));
-        ZSTD_free(ctx->threads, ctx->customMem);
+        ZSTD_memcpy(threadPool, ctx->threads, ctx->threadCapacity * sizeof(*threadPool));
+        ZSTD_customFree(ctx->threads, ctx->customMem);
         ctx->threads = threadPool;
         /* Initialize additional threads */
         {   size_t threadId;
@@ -251,9 +268,12 @@ static int isQueueFull(POOL_ctx const* ctx) {
 }
 
 
-static void POOL_add_internal(POOL_ctx* ctx, POOL_function function, void *opaque)
+static void
+POOL_add_internal(POOL_ctx* ctx, POOL_function function, void *opaque)
 {
-    POOL_job const job = {function, opaque};
+    POOL_job job;
+    job.function = function;
+    job.opaque = opaque;
     assert(ctx != NULL);
     if (ctx->shutdown) return;
 
@@ -301,21 +321,28 @@ int POOL_tryAdd(POOL_ctx* ctx, POOL_function function, void* opaque)
 struct POOL_ctx_s {
     int dummy;
 };
-static POOL_ctx g_ctx;
+static POOL_ctx g_poolCtx;
 
 POOL_ctx* POOL_create(size_t numThreads, size_t queueSize) {
     return POOL_create_advanced(numThreads, queueSize, ZSTD_defaultCMem);
 }
 
-POOL_ctx* POOL_create_advanced(size_t numThreads, size_t queueSize, ZSTD_customMem customMem) {
+POOL_ctx*
+POOL_create_advanced(size_t numThreads, size_t queueSize, ZSTD_customMem customMem)
+{
     (void)numThreads;
     (void)queueSize;
     (void)customMem;
-    return &g_ctx;
+    return &g_poolCtx;
 }
 
 void POOL_free(POOL_ctx* ctx) {
-    assert(!ctx || ctx == &g_ctx);
+    assert(!ctx || ctx == &g_poolCtx);
+    (void)ctx;
+}
+
+void POOL_joinJobs(POOL_ctx* ctx){
+    assert(!ctx || ctx == &g_poolCtx);
     (void)ctx;
 }
 
@@ -335,9 +362,9 @@ int POOL_tryAdd(POOL_ctx* ctx, POOL_function function, void* opaque) {
     return 1;
 }
 
-size_t POOL_sizeof(POOL_ctx* ctx) {
+size_t POOL_sizeof(const POOL_ctx* ctx) {
     if (ctx==NULL) return 0;  /* supports sizeof NULL */
-    assert(ctx == &g_ctx);
+    assert(ctx == &g_poolCtx);
     return sizeof(*ctx);
 }
 
diff --git a/contrib/zstd/pool.h b/contrib/zstd/pool.h
index 7dcaf010d..14e883feb 100644
--- a/contrib/zstd/pool.h
+++ b/contrib/zstd/pool.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -16,11 +16,8 @@ extern "C" {
 #endif
 
 
-#include <stddef.h>   /* size_t */
-
-#ifndef ZSTD_STATIC_LINKING_ONLY
+#include "zstd_deps.h"
 #define ZSTD_STATIC_LINKING_ONLY   /* ZSTD_customMem */
-#endif
 #include "zstd.h"
 
 typedef struct POOL_ctx_s POOL_ctx;
@@ -41,6 +38,12 @@ POOL_ctx* POOL_create_advanced(size_t numThreads, size_t queueSize,
  */
 void POOL_free(POOL_ctx* ctx);
 
+
+/*! POOL_joinJobs() :
+ *  Waits for all queued jobs to finish executing.
+ */
+void POOL_joinJobs(POOL_ctx* ctx);
+
 /*! POOL_resize() :
  *  Expands or shrinks pool's number of threads.
  *  This is more efficient than releasing + creating a new context,
@@ -56,7 +59,7 @@ int POOL_resize(POOL_ctx* ctx, size_t numThreads);
  * @return threadpool memory usage
  *  note : compatible with NULL (returns 0 in this case)
  */
-size_t POOL_sizeof(POOL_ctx* ctx);
+size_t POOL_sizeof(const POOL_ctx* ctx);
 
 /*! POOL_function :
  *  The function type that can be added to a thread pool.
@@ -73,7 +76,7 @@ void POOL_add(POOL_ctx* ctx, POOL_function function, void* opaque);
 
 
 /*! POOL_tryAdd() :
- *  Add the job `function(opaque)` to thread pool _if_ a worker is available.
+ *  Add the job `function(opaque)` to thread pool _if_ a queue slot is available.
  *  Returns immediately even if not (does not block).
  * @return : 1 if successful, 0 if not.
  */
diff --git a/contrib/zstd/portability_macros.h b/contrib/zstd/portability_macros.h
new file mode 100644
index 000000000..8fd6ea82d
--- /dev/null
+++ b/contrib/zstd/portability_macros.h
@@ -0,0 +1,156 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_PORTABILITY_MACROS_H
+#define ZSTD_PORTABILITY_MACROS_H
+
+/**
+ * This header file contains macro definitions to support portability.
+ * This header is shared between C and ASM code, so it MUST only
+ * contain macro definitions. It MUST not contain any C code.
+ *
+ * This header ONLY defines macros to detect platforms/feature support.
+ *
+ */
+
+
+/* compat. with non-clang compilers */
+#ifndef __has_attribute
+  #define __has_attribute(x) 0
+#endif
+
+/* compat. with non-clang compilers */
+#ifndef __has_builtin
+#  define __has_builtin(x) 0
+#endif
+
+/* compat. with non-clang compilers */
+#ifndef __has_feature
+#  define __has_feature(x) 0
+#endif
+
+/* detects whether we are being compiled under msan */
+#ifndef ZSTD_MEMORY_SANITIZER
+#  if __has_feature(memory_sanitizer)
+#    define ZSTD_MEMORY_SANITIZER 1
+#  else
+#    define ZSTD_MEMORY_SANITIZER 0
+#  endif
+#endif
+
+/* detects whether we are being compiled under asan */
+#ifndef ZSTD_ADDRESS_SANITIZER
+#  if __has_feature(address_sanitizer)
+#    define ZSTD_ADDRESS_SANITIZER 1
+#  elif defined(__SANITIZE_ADDRESS__)
+#    define ZSTD_ADDRESS_SANITIZER 1
+#  else
+#    define ZSTD_ADDRESS_SANITIZER 0
+#  endif
+#endif
+
+/* detects whether we are being compiled under dfsan */
+#ifndef ZSTD_DATAFLOW_SANITIZER
+# if __has_feature(dataflow_sanitizer)
+#  define ZSTD_DATAFLOW_SANITIZER 1
+# else
+#  define ZSTD_DATAFLOW_SANITIZER 0
+# endif
+#endif
+
+/* Mark the internal assembly functions as hidden  */
+#ifdef __ELF__
+# define ZSTD_HIDE_ASM_FUNCTION(func) .hidden func
+#else
+# define ZSTD_HIDE_ASM_FUNCTION(func)
+#endif
+
+/* Enable runtime BMI2 dispatch based on the CPU.
+ * Enabled for clang & gcc >=4.8 on x86 when BMI2 isn't enabled by default.
+ */
+#ifndef DYNAMIC_BMI2
+  #if ((defined(__clang__) && __has_attribute(__target__)) \
+      || (defined(__GNUC__) \
+          && (__GNUC__ >= 5 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8)))) \
+      && (defined(__x86_64__) || defined(_M_X64)) \
+      && !defined(__BMI2__)
+  #  define DYNAMIC_BMI2 1
+  #else
+  #  define DYNAMIC_BMI2 0
+  #endif
+#endif
+
+/**
+ * Only enable assembly for GNUC compatible compilers,
+ * because other platforms may not support GAS assembly syntax.
+ *
+ * Only enable assembly for Linux / MacOS, other platforms may
+ * work, but they haven't been tested. This could likely be
+ * extended to BSD systems.
+ *
+ * Disable assembly when MSAN is enabled, because MSAN requires
+ * 100% of code to be instrumented to work.
+ */
+#if defined(__GNUC__)
+#  if defined(__linux__) || defined(__linux) || defined(__APPLE__)
+#    if ZSTD_MEMORY_SANITIZER
+#      define ZSTD_ASM_SUPPORTED 0
+#    elif ZSTD_DATAFLOW_SANITIZER
+#      define ZSTD_ASM_SUPPORTED 0
+#    else
+#      define ZSTD_ASM_SUPPORTED 1
+#    endif
+#  else
+#    define ZSTD_ASM_SUPPORTED 0
+#  endif
+#else
+#  define ZSTD_ASM_SUPPORTED 0
+#endif
+
+/**
+ * Determines whether we should enable assembly for x86-64
+ * with BMI2.
+ *
+ * Enable if all of the following conditions hold:
+ * - ASM hasn't been explicitly disabled by defining ZSTD_DISABLE_ASM
+ * - Assembly is supported
+ * - We are compiling for x86-64 and either:
+ *   - DYNAMIC_BMI2 is enabled
+ *   - BMI2 is supported at compile time
+ */
+#if !defined(ZSTD_DISABLE_ASM) &&                                 \
+    ZSTD_ASM_SUPPORTED &&                                         \
+    defined(__x86_64__) &&                                        \
+    (DYNAMIC_BMI2 || defined(__BMI2__))
+# define ZSTD_ENABLE_ASM_X86_64_BMI2 1
+#else
+# define ZSTD_ENABLE_ASM_X86_64_BMI2 0
+#endif
+
+/*
+ * For x86 ELF targets, add .note.gnu.property section for Intel CET in
+ * assembly sources when CET is enabled.
+ *
+ * Additionally, any function that may be called indirectly must begin
+ * with ZSTD_CET_ENDBRANCH.
+ */
+#if defined(__ELF__) && (defined(__x86_64__) || defined(__i386__)) \
+    && defined(__has_include)
+# if __has_include(<cet.h>)
+#  include <cet.h>
+#  define ZSTD_CET_ENDBRANCH _CET_ENDBR
+# endif
+#endif
+
+#ifndef ZSTD_CET_ENDBRANCH
+# define ZSTD_CET_ENDBRANCH
+#endif
+
+#endif /* ZSTD_PORTABILITY_MACROS_H */
diff --git a/contrib/zstd/threading.c b/contrib/zstd/threading.c
deleted file mode 100644
index e2edb313e..000000000
--- a/contrib/zstd/threading.c
+++ /dev/null
@@ -1,121 +0,0 @@
-/**
- * Copyright (c) 2016 Tino Reichardt
- * All rights reserved.
- *
- * You can contact the author at:
- * - zstdmt source repository: https://github.com/mcmilk/zstdmt
- *
- * This source code is licensed under both the BSD-style license (found in the
- * LICENSE file in the root directory of this source tree) and the GPLv2 (found
- * in the COPYING file in the root directory of this source tree).
- * You may select, at your option, one of the above-listed licenses.
- */
-
-/**
- * This file will hold wrapper for systems, which do not support pthreads
- */
-
-#include "threading.h"
-
-/* create fake symbol to avoid empty translation unit warning */
-int g_ZSTD_threading_useless_symbol;
-
-#if defined(ZSTD_MULTITHREAD) && defined(_WIN32)
-
-/**
- * Windows minimalist Pthread Wrapper, based on :
- * http://www.cse.wustl.edu/~schmidt/win32-cv-1.html
- */
-
-
-/* ===  Dependencies  === */
-#include <process.h>
-#include <errno.h>
-
-
-/* ===  Implementation  === */
-
-static unsigned __stdcall worker(void *arg)
-{
-    ZSTD_pthread_t* const thread = (ZSTD_pthread_t*) arg;
-    thread->arg = thread->start_routine(thread->arg);
-    return 0;
-}
-
-int ZSTD_pthread_create(ZSTD_pthread_t* thread, const void* unused,
-            void* (*start_routine) (void*), void* arg)
-{
-    (void)unused;
-    thread->arg = arg;
-    thread->start_routine = start_routine;
-    thread->handle = (HANDLE) _beginthreadex(NULL, 0, worker, thread, 0, NULL);
-
-    if (!thread->handle)
-        return errno;
-    else
-        return 0;
-}
-
-int ZSTD_pthread_join(ZSTD_pthread_t thread, void **value_ptr)
-{
-    DWORD result;
-
-    if (!thread.handle) return 0;
-
-    result = WaitForSingleObject(thread.handle, INFINITE);
-    switch (result) {
-    case WAIT_OBJECT_0:
-        if (value_ptr) *value_ptr = thread.arg;
-        return 0;
-    case WAIT_ABANDONED:
-        return EINVAL;
-    default:
-        return GetLastError();
-    }
-}
-
-#endif   /* ZSTD_MULTITHREAD */
-
-#if defined(ZSTD_MULTITHREAD) && DEBUGLEVEL >= 1 && !defined(_WIN32)
-
-#include <stdlib.h>
-
-int ZSTD_pthread_mutex_init(ZSTD_pthread_mutex_t* mutex, pthread_mutexattr_t const* attr)
-{
-    *mutex = (pthread_mutex_t*)malloc(sizeof(pthread_mutex_t));
-    if (!*mutex)
-        return 1;
-    return pthread_mutex_init(*mutex, attr);
-}
-
-int ZSTD_pthread_mutex_destroy(ZSTD_pthread_mutex_t* mutex)
-{
-    if (!*mutex)
-        return 0;
-    {
-        int const ret = pthread_mutex_destroy(*mutex);
-        free(*mutex);
-        return ret;
-    }
-}
-
-int ZSTD_pthread_cond_init(ZSTD_pthread_cond_t* cond, pthread_condattr_t const* attr)
-{
-    *cond = (pthread_cond_t*)malloc(sizeof(pthread_cond_t));
-    if (!*cond)
-        return 1;
-    return pthread_cond_init(*cond, attr);
-}
-
-int ZSTD_pthread_cond_destroy(ZSTD_pthread_cond_t* cond)
-{
-    if (!*cond)
-        return 0;
-    {
-        int const ret = pthread_cond_destroy(*cond);
-        free(*cond);
-        return ret;
-    }
-}
-
-#endif
diff --git a/contrib/zstd/threading.h b/contrib/zstd/threading.h
deleted file mode 100644
index fd0060d5a..000000000
--- a/contrib/zstd/threading.h
+++ /dev/null
@@ -1,155 +0,0 @@
-/**
- * Copyright (c) 2016 Tino Reichardt
- * All rights reserved.
- *
- * You can contact the author at:
- * - zstdmt source repository: https://github.com/mcmilk/zstdmt
- *
- * This source code is licensed under both the BSD-style license (found in the
- * LICENSE file in the root directory of this source tree) and the GPLv2 (found
- * in the COPYING file in the root directory of this source tree).
- * You may select, at your option, one of the above-listed licenses.
- */
-
-#ifndef THREADING_H_938743
-#define THREADING_H_938743
-
-#include "debug.h"
-
-#if defined (__cplusplus)
-extern "C" {
-#endif
-
-#if defined(ZSTD_MULTITHREAD) && defined(_WIN32)
-
-/**
- * Windows minimalist Pthread Wrapper, based on :
- * http://www.cse.wustl.edu/~schmidt/win32-cv-1.html
- */
-#ifdef WINVER
-#  undef WINVER
-#endif
-#define WINVER       0x0600
-
-#ifdef _WIN32_WINNT
-#  undef _WIN32_WINNT
-#endif
-#define _WIN32_WINNT 0x0600
-
-#ifndef WIN32_LEAN_AND_MEAN
-#  define WIN32_LEAN_AND_MEAN
-#endif
-
-#undef ERROR   /* reported already defined on VS 2015 (Rich Geldreich) */
-#include <windows.h>
-#undef ERROR
-#define ERROR(name) ZSTD_ERROR(name)
-
-
-/* mutex */
-#define ZSTD_pthread_mutex_t           CRITICAL_SECTION
-#define ZSTD_pthread_mutex_init(a, b)  ((void)(b), InitializeCriticalSection((a)), 0)
-#define ZSTD_pthread_mutex_destroy(a)  DeleteCriticalSection((a))
-#define ZSTD_pthread_mutex_lock(a)     EnterCriticalSection((a))
-#define ZSTD_pthread_mutex_unlock(a)   LeaveCriticalSection((a))
-
-/* condition variable */
-#define ZSTD_pthread_cond_t             CONDITION_VARIABLE
-#define ZSTD_pthread_cond_init(a, b)    ((void)(b), InitializeConditionVariable((a)), 0)
-#define ZSTD_pthread_cond_destroy(a)    ((void)(a))
-#define ZSTD_pthread_cond_wait(a, b)    SleepConditionVariableCS((a), (b), INFINITE)
-#define ZSTD_pthread_cond_signal(a)     WakeConditionVariable((a))
-#define ZSTD_pthread_cond_broadcast(a)  WakeAllConditionVariable((a))
-
-/* ZSTD_pthread_create() and ZSTD_pthread_join() */
-typedef struct {
-    HANDLE handle;
-    void* (*start_routine)(void*);
-    void* arg;
-} ZSTD_pthread_t;
-
-int ZSTD_pthread_create(ZSTD_pthread_t* thread, const void* unused,
-                   void* (*start_routine) (void*), void* arg);
-
-int ZSTD_pthread_join(ZSTD_pthread_t thread, void** value_ptr);
-
-/**
- * add here more wrappers as required
- */
-
-
-#elif defined(ZSTD_MULTITHREAD)    /* posix assumed ; need a better detection method */
-/* ===   POSIX Systems   === */
-#  include <pthread.h>
-
-#if DEBUGLEVEL < 1
-
-#define ZSTD_pthread_mutex_t            pthread_mutex_t
-#define ZSTD_pthread_mutex_init(a, b)   pthread_mutex_init((a), (b))
-#define ZSTD_pthread_mutex_destroy(a)   pthread_mutex_destroy((a))
-#define ZSTD_pthread_mutex_lock(a)      pthread_mutex_lock((a))
-#define ZSTD_pthread_mutex_unlock(a)    pthread_mutex_unlock((a))
-
-#define ZSTD_pthread_cond_t             pthread_cond_t
-#define ZSTD_pthread_cond_init(a, b)    pthread_cond_init((a), (b))
-#define ZSTD_pthread_cond_destroy(a)    pthread_cond_destroy((a))
-#define ZSTD_pthread_cond_wait(a, b)    pthread_cond_wait((a), (b))
-#define ZSTD_pthread_cond_signal(a)     pthread_cond_signal((a))
-#define ZSTD_pthread_cond_broadcast(a)  pthread_cond_broadcast((a))
-
-#define ZSTD_pthread_t                  pthread_t
-#define ZSTD_pthread_create(a, b, c, d) pthread_create((a), (b), (c), (d))
-#define ZSTD_pthread_join(a, b)         pthread_join((a),(b))
-
-#else /* DEBUGLEVEL >= 1 */
-
-/* Debug implementation of threading.
- * In this implementation we use pointers for mutexes and condition variables.
- * This way, if we forget to init/destroy them the program will crash or ASAN
- * will report leaks.
- */
-
-#define ZSTD_pthread_mutex_t            pthread_mutex_t*
-int ZSTD_pthread_mutex_init(ZSTD_pthread_mutex_t* mutex, pthread_mutexattr_t const* attr);
-int ZSTD_pthread_mutex_destroy(ZSTD_pthread_mutex_t* mutex);
-#define ZSTD_pthread_mutex_lock(a)      pthread_mutex_lock(*(a))
-#define ZSTD_pthread_mutex_unlock(a)    pthread_mutex_unlock(*(a))
-
-#define ZSTD_pthread_cond_t             pthread_cond_t*
-int ZSTD_pthread_cond_init(ZSTD_pthread_cond_t* cond, pthread_condattr_t const* attr);
-int ZSTD_pthread_cond_destroy(ZSTD_pthread_cond_t* cond);
-#define ZSTD_pthread_cond_wait(a, b)    pthread_cond_wait(*(a), *(b))
-#define ZSTD_pthread_cond_signal(a)     pthread_cond_signal(*(a))
-#define ZSTD_pthread_cond_broadcast(a)  pthread_cond_broadcast(*(a))
-
-#define ZSTD_pthread_t                  pthread_t
-#define ZSTD_pthread_create(a, b, c, d) pthread_create((a), (b), (c), (d))
-#define ZSTD_pthread_join(a, b)         pthread_join((a),(b))
-
-#endif
-
-#else  /* ZSTD_MULTITHREAD not defined */
-/* No multithreading support */
-
-typedef int ZSTD_pthread_mutex_t;
-#define ZSTD_pthread_mutex_init(a, b)   ((void)(a), (void)(b), 0)
-#define ZSTD_pthread_mutex_destroy(a)   ((void)(a))
-#define ZSTD_pthread_mutex_lock(a)      ((void)(a))
-#define ZSTD_pthread_mutex_unlock(a)    ((void)(a))
-
-typedef int ZSTD_pthread_cond_t;
-#define ZSTD_pthread_cond_init(a, b)    ((void)(a), (void)(b), 0)
-#define ZSTD_pthread_cond_destroy(a)    ((void)(a))
-#define ZSTD_pthread_cond_wait(a, b)    ((void)(a), (void)(b))
-#define ZSTD_pthread_cond_signal(a)     ((void)(a))
-#define ZSTD_pthread_cond_broadcast(a)  ((void)(a))
-
-/* do not use ZSTD_pthread_t */
-
-#endif /* ZSTD_MULTITHREAD */
-
-#if defined (__cplusplus)
-}
-#endif
-
-#endif /* THREADING_H_938743 */
diff --git a/contrib/zstd/zdict.h b/contrib/zstd/zdict.h
new file mode 100644
index 000000000..2268f948a
--- /dev/null
+++ b/contrib/zstd/zdict.h
@@ -0,0 +1,474 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+#ifndef ZSTD_ZDICT_H
+#define ZSTD_ZDICT_H
+
+/*======  Dependencies  ======*/
+#include <stddef.h>  /* size_t */
+
+
+/* =====   ZDICTLIB_API : control library symbols visibility   ===== */
+#ifndef ZDICTLIB_VISIBLE
+   /* Backwards compatibility with old macro name */
+#  ifdef ZDICTLIB_VISIBILITY
+#    define ZDICTLIB_VISIBLE ZDICTLIB_VISIBILITY
+#  elif defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
+#    define ZDICTLIB_VISIBLE __attribute__ ((visibility ("default")))
+#  else
+#    define ZDICTLIB_VISIBLE
+#  endif
+#endif
+
+#ifndef ZDICTLIB_HIDDEN
+#  if defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
+#    define ZDICTLIB_HIDDEN __attribute__ ((visibility ("hidden")))
+#  else
+#    define ZDICTLIB_HIDDEN
+#  endif
+#endif
+
+#if defined(ZSTD_DLL_EXPORT) && (ZSTD_DLL_EXPORT==1)
+#  define ZDICTLIB_API __declspec(dllexport) ZDICTLIB_VISIBLE
+#elif defined(ZSTD_DLL_IMPORT) && (ZSTD_DLL_IMPORT==1)
+#  define ZDICTLIB_API __declspec(dllimport) ZDICTLIB_VISIBLE /* It isn't required but allows to generate better code, saving a function pointer load from the IAT and an indirect jump.*/
+#else
+#  define ZDICTLIB_API ZDICTLIB_VISIBLE
+#endif
+
+/*******************************************************************************
+ * Zstd dictionary builder
+ *
+ * FAQ
+ * ===
+ * Why should I use a dictionary?
+ * ------------------------------
+ *
+ * Zstd can use dictionaries to improve compression ratio of small data.
+ * Traditionally small files don't compress well because there is very little
+ * repetition in a single sample, since it is small. But, if you are compressing
+ * many similar files, like a bunch of JSON records that share the same
+ * structure, you can train a dictionary on ahead of time on some samples of
+ * these files. Then, zstd can use the dictionary to find repetitions that are
+ * present across samples. This can vastly improve compression ratio.
+ *
+ * When is a dictionary useful?
+ * ----------------------------
+ *
+ * Dictionaries are useful when compressing many small files that are similar.
+ * The larger a file is, the less benefit a dictionary will have. Generally,
+ * we don't expect dictionary compression to be effective past 100KB. And the
+ * smaller a file is, the more we would expect the dictionary to help.
+ *
+ * How do I use a dictionary?
+ * --------------------------
+ *
+ * Simply pass the dictionary to the zstd compressor with
+ * `ZSTD_CCtx_loadDictionary()`. The same dictionary must then be passed to
+ * the decompressor, using `ZSTD_DCtx_loadDictionary()`. There are other
+ * more advanced functions that allow selecting some options, see zstd.h for
+ * complete documentation.
+ *
+ * What is a zstd dictionary?
+ * --------------------------
+ *
+ * A zstd dictionary has two pieces: Its header, and its content. The header
+ * contains a magic number, the dictionary ID, and entropy tables. These
+ * entropy tables allow zstd to save on header costs in the compressed file,
+ * which really matters for small data. The content is just bytes, which are
+ * repeated content that is common across many samples.
+ *
+ * What is a raw content dictionary?
+ * ---------------------------------
+ *
+ * A raw content dictionary is just bytes. It doesn't have a zstd dictionary
+ * header, a dictionary ID, or entropy tables. Any buffer is a valid raw
+ * content dictionary.
+ *
+ * How do I train a dictionary?
+ * ----------------------------
+ *
+ * Gather samples from your use case. These samples should be similar to each
+ * other. If you have several use cases, you could try to train one dictionary
+ * per use case.
+ *
+ * Pass those samples to `ZDICT_trainFromBuffer()` and that will train your
+ * dictionary. There are a few advanced versions of this function, but this
+ * is a great starting point. If you want to further tune your dictionary
+ * you could try `ZDICT_optimizeTrainFromBuffer_cover()`. If that is too slow
+ * you can try `ZDICT_optimizeTrainFromBuffer_fastCover()`.
+ *
+ * If the dictionary training function fails, that is likely because you
+ * either passed too few samples, or a dictionary would not be effective
+ * for your data. Look at the messages that the dictionary trainer printed,
+ * if it doesn't say too few samples, then a dictionary would not be effective.
+ *
+ * How large should my dictionary be?
+ * ----------------------------------
+ *
+ * A reasonable dictionary size, the `dictBufferCapacity`, is about 100KB.
+ * The zstd CLI defaults to a 110KB dictionary. You likely don't need a
+ * dictionary larger than that. But, most use cases can get away with a
+ * smaller dictionary. The advanced dictionary builders can automatically
+ * shrink the dictionary for you, and select the smallest size that doesn't
+ * hurt compression ratio too much. See the `shrinkDict` parameter.
+ * A smaller dictionary can save memory, and potentially speed up
+ * compression.
+ *
+ * How many samples should I provide to the dictionary builder?
+ * ------------------------------------------------------------
+ *
+ * We generally recommend passing ~100x the size of the dictionary
+ * in samples. A few thousand should suffice. Having too few samples
+ * can hurt the dictionaries effectiveness. Having more samples will
+ * only improve the dictionaries effectiveness. But having too many
+ * samples can slow down the dictionary builder.
+ *
+ * How do I determine if a dictionary will be effective?
+ * -----------------------------------------------------
+ *
+ * Simply train a dictionary and try it out. You can use zstd's built in
+ * benchmarking tool to test the dictionary effectiveness.
+ *
+ *   # Benchmark levels 1-3 without a dictionary
+ *   zstd -b1e3 -r /path/to/my/files
+ *   # Benchmark levels 1-3 with a dictionary
+ *   zstd -b1e3 -r /path/to/my/files -D /path/to/my/dictionary
+ *
+ * When should I retrain a dictionary?
+ * -----------------------------------
+ *
+ * You should retrain a dictionary when its effectiveness drops. Dictionary
+ * effectiveness drops as the data you are compressing changes. Generally, we do
+ * expect dictionaries to "decay" over time, as your data changes, but the rate
+ * at which they decay depends on your use case. Internally, we regularly
+ * retrain dictionaries, and if the new dictionary performs significantly
+ * better than the old dictionary, we will ship the new dictionary.
+ *
+ * I have a raw content dictionary, how do I turn it into a zstd dictionary?
+ * -------------------------------------------------------------------------
+ *
+ * If you have a raw content dictionary, e.g. by manually constructing it, or
+ * using a third-party dictionary builder, you can turn it into a zstd
+ * dictionary by using `ZDICT_finalizeDictionary()`. You'll also have to
+ * provide some samples of the data. It will add the zstd header to the
+ * raw content, which contains a dictionary ID and entropy tables, which
+ * will improve compression ratio, and allow zstd to write the dictionary ID
+ * into the frame, if you so choose.
+ *
+ * Do I have to use zstd's dictionary builder?
+ * -------------------------------------------
+ *
+ * No! You can construct dictionary content however you please, it is just
+ * bytes. It will always be valid as a raw content dictionary. If you want
+ * a zstd dictionary, which can improve compression ratio, use
+ * `ZDICT_finalizeDictionary()`.
+ *
+ * What is the attack surface of a zstd dictionary?
+ * ------------------------------------------------
+ *
+ * Zstd is heavily fuzz tested, including loading fuzzed dictionaries, so
+ * zstd should never crash, or access out-of-bounds memory no matter what
+ * the dictionary is. However, if an attacker can control the dictionary
+ * during decompression, they can cause zstd to generate arbitrary bytes,
+ * just like if they controlled the compressed data.
+ *
+ ******************************************************************************/
+
+
+/*! ZDICT_trainFromBuffer():
+ *  Train a dictionary from an array of samples.
+ *  Redirect towards ZDICT_optimizeTrainFromBuffer_fastCover() single-threaded, with d=8, steps=4,
+ *  f=20, and accel=1.
+ *  Samples must be stored concatenated in a single flat buffer `samplesBuffer`,
+ *  supplied with an array of sizes `samplesSizes`, providing the size of each sample, in order.
+ *  The resulting dictionary will be saved into `dictBuffer`.
+ * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`)
+ *          or an error code, which can be tested with ZDICT_isError().
+ *  Note:  Dictionary training will fail if there are not enough samples to construct a
+ *         dictionary, or if most of the samples are too small (< 8 bytes being the lower limit).
+ *         If dictionary training fails, you should use zstd without a dictionary, as the dictionary
+ *         would've been ineffective anyways. If you believe your samples would benefit from a dictionary
+ *         please open an issue with details, and we can look into it.
+ *  Note: ZDICT_trainFromBuffer()'s memory usage is about 6 MB.
+ *  Tips: In general, a reasonable dictionary has a size of ~ 100 KB.
+ *        It's possible to select smaller or larger size, just by specifying `dictBufferCapacity`.
+ *        In general, it's recommended to provide a few thousands samples, though this can vary a lot.
+ *        It's recommended that total size of all samples be about ~x100 times the target size of dictionary.
+ */
+ZDICTLIB_API size_t ZDICT_trainFromBuffer(void* dictBuffer, size_t dictBufferCapacity,
+                                    const void* samplesBuffer,
+                                    const size_t* samplesSizes, unsigned nbSamples);
+
+typedef struct {
+    int      compressionLevel;   /**< optimize for a specific zstd compression level; 0 means default */
+    unsigned notificationLevel;  /**< Write log to stderr; 0 = none (default); 1 = errors; 2 = progression; 3 = details; 4 = debug; */
+    unsigned dictID;             /**< force dictID value; 0 means auto mode (32-bits random value)
+                                  *   NOTE: The zstd format reserves some dictionary IDs for future use.
+                                  *         You may use them in private settings, but be warned that they
+                                  *         may be used by zstd in a public dictionary registry in the future.
+                                  *         These dictionary IDs are:
+                                  *           - low range  : <= 32767
+                                  *           - high range : >= (2^31)
+                                  */
+} ZDICT_params_t;
+
+/*! ZDICT_finalizeDictionary():
+ * Given a custom content as a basis for dictionary, and a set of samples,
+ * finalize dictionary by adding headers and statistics according to the zstd
+ * dictionary format.
+ *
+ * Samples must be stored concatenated in a flat buffer `samplesBuffer`,
+ * supplied with an array of sizes `samplesSizes`, providing the size of each
+ * sample in order. The samples are used to construct the statistics, so they
+ * should be representative of what you will compress with this dictionary.
+ *
+ * The compression level can be set in `parameters`. You should pass the
+ * compression level you expect to use in production. The statistics for each
+ * compression level differ, so tuning the dictionary for the compression level
+ * can help quite a bit.
+ *
+ * You can set an explicit dictionary ID in `parameters`, or allow us to pick
+ * a random dictionary ID for you, but we can't guarantee no collisions.
+ *
+ * The dstDictBuffer and the dictContent may overlap, and the content will be
+ * appended to the end of the header. If the header + the content doesn't fit in
+ * maxDictSize the beginning of the content is truncated to make room, since it
+ * is presumed that the most profitable content is at the end of the dictionary,
+ * since that is the cheapest to reference.
+ *
+ * `maxDictSize` must be >= max(dictContentSize, ZSTD_DICTSIZE_MIN).
+ *
+ * @return: size of dictionary stored into `dstDictBuffer` (<= `maxDictSize`),
+ *          or an error code, which can be tested by ZDICT_isError().
+ * Note: ZDICT_finalizeDictionary() will push notifications into stderr if
+ *       instructed to, using notificationLevel>0.
+ * NOTE: This function currently may fail in several edge cases including:
+ *         * Not enough samples
+ *         * Samples are uncompressible
+ *         * Samples are all exactly the same
+ */
+ZDICTLIB_API size_t ZDICT_finalizeDictionary(void* dstDictBuffer, size_t maxDictSize,
+                                const void* dictContent, size_t dictContentSize,
+                                const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
+                                ZDICT_params_t parameters);
+
+
+/*======   Helper functions   ======*/
+ZDICTLIB_API unsigned ZDICT_getDictID(const void* dictBuffer, size_t dictSize);  /**< extracts dictID; @return zero if error (not a valid dictionary) */
+ZDICTLIB_API size_t ZDICT_getDictHeaderSize(const void* dictBuffer, size_t dictSize);  /* returns dict header size; returns a ZSTD error code on failure */
+ZDICTLIB_API unsigned ZDICT_isError(size_t errorCode);
+ZDICTLIB_API const char* ZDICT_getErrorName(size_t errorCode);
+
+#endif   /* ZSTD_ZDICT_H */
+
+#if defined(ZDICT_STATIC_LINKING_ONLY) && !defined(ZSTD_ZDICT_H_STATIC)
+#define ZSTD_ZDICT_H_STATIC
+
+/* This can be overridden externally to hide static symbols. */
+#ifndef ZDICTLIB_STATIC_API
+#  if defined(ZSTD_DLL_EXPORT) && (ZSTD_DLL_EXPORT==1)
+#    define ZDICTLIB_STATIC_API __declspec(dllexport) ZDICTLIB_VISIBLE
+#  elif defined(ZSTD_DLL_IMPORT) && (ZSTD_DLL_IMPORT==1)
+#    define ZDICTLIB_STATIC_API __declspec(dllimport) ZDICTLIB_VISIBLE
+#  else
+#    define ZDICTLIB_STATIC_API ZDICTLIB_VISIBLE
+#  endif
+#endif
+
+/* ====================================================================================
+ * The definitions in this section are considered experimental.
+ * They should never be used with a dynamic library, as they may change in the future.
+ * They are provided for advanced usages.
+ * Use them only in association with static linking.
+ * ==================================================================================== */
+
+#define ZDICT_DICTSIZE_MIN    256
+/* Deprecated: Remove in v1.6.0 */
+#define ZDICT_CONTENTSIZE_MIN 128
+
+/*! ZDICT_cover_params_t:
+ *  k and d are the only required parameters.
+ *  For others, value 0 means default.
+ */
+typedef struct {
+    unsigned k;                  /* Segment size : constraint: 0 < k : Reasonable range [16, 2048+] */
+    unsigned d;                  /* dmer size : constraint: 0 < d <= k : Reasonable range [6, 16] */
+    unsigned steps;              /* Number of steps : Only used for optimization : 0 means default (40) : Higher means more parameters checked */
+    unsigned nbThreads;          /* Number of threads : constraint: 0 < nbThreads : 1 means single-threaded : Only used for optimization : Ignored if ZSTD_MULTITHREAD is not defined */
+    double splitPoint;           /* Percentage of samples used for training: Only used for optimization : the first nbSamples * splitPoint samples will be used to training, the last nbSamples * (1 - splitPoint) samples will be used for testing, 0 means default (1.0), 1.0 when all samples are used for both training and testing */
+    unsigned shrinkDict;         /* Train dictionaries to shrink in size starting from the minimum size and selects the smallest dictionary that is shrinkDictMaxRegression% worse than the largest dictionary. 0 means no shrinking and 1 means shrinking  */
+    unsigned shrinkDictMaxRegression; /* Sets shrinkDictMaxRegression so that a smaller dictionary can be at worse shrinkDictMaxRegression% worse than the max dict size dictionary. */
+    ZDICT_params_t zParams;
+} ZDICT_cover_params_t;
+
+typedef struct {
+    unsigned k;                  /* Segment size : constraint: 0 < k : Reasonable range [16, 2048+] */
+    unsigned d;                  /* dmer size : constraint: 0 < d <= k : Reasonable range [6, 16] */
+    unsigned f;                  /* log of size of frequency array : constraint: 0 < f <= 31 : 1 means default(20)*/
+    unsigned steps;              /* Number of steps : Only used for optimization : 0 means default (40) : Higher means more parameters checked */
+    unsigned nbThreads;          /* Number of threads : constraint: 0 < nbThreads : 1 means single-threaded : Only used for optimization : Ignored if ZSTD_MULTITHREAD is not defined */
+    double splitPoint;           /* Percentage of samples used for training: Only used for optimization : the first nbSamples * splitPoint samples will be used to training, the last nbSamples * (1 - splitPoint) samples will be used for testing, 0 means default (0.75), 1.0 when all samples are used for both training and testing */
+    unsigned accel;              /* Acceleration level: constraint: 0 < accel <= 10, higher means faster and less accurate, 0 means default(1) */
+    unsigned shrinkDict;         /* Train dictionaries to shrink in size starting from the minimum size and selects the smallest dictionary that is shrinkDictMaxRegression% worse than the largest dictionary. 0 means no shrinking and 1 means shrinking  */
+    unsigned shrinkDictMaxRegression; /* Sets shrinkDictMaxRegression so that a smaller dictionary can be at worse shrinkDictMaxRegression% worse than the max dict size dictionary. */
+
+    ZDICT_params_t zParams;
+} ZDICT_fastCover_params_t;
+
+/*! ZDICT_trainFromBuffer_cover():
+ *  Train a dictionary from an array of samples using the COVER algorithm.
+ *  Samples must be stored concatenated in a single flat buffer `samplesBuffer`,
+ *  supplied with an array of sizes `samplesSizes`, providing the size of each sample, in order.
+ *  The resulting dictionary will be saved into `dictBuffer`.
+ * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`)
+ *          or an error code, which can be tested with ZDICT_isError().
+ *          See ZDICT_trainFromBuffer() for details on failure modes.
+ *  Note: ZDICT_trainFromBuffer_cover() requires about 9 bytes of memory for each input byte.
+ *  Tips: In general, a reasonable dictionary has a size of ~ 100 KB.
+ *        It's possible to select smaller or larger size, just by specifying `dictBufferCapacity`.
+ *        In general, it's recommended to provide a few thousands samples, though this can vary a lot.
+ *        It's recommended that total size of all samples be about ~x100 times the target size of dictionary.
+ */
+ZDICTLIB_STATIC_API size_t ZDICT_trainFromBuffer_cover(
+          void *dictBuffer, size_t dictBufferCapacity,
+    const void *samplesBuffer, const size_t *samplesSizes, unsigned nbSamples,
+          ZDICT_cover_params_t parameters);
+
+/*! ZDICT_optimizeTrainFromBuffer_cover():
+ * The same requirements as above hold for all the parameters except `parameters`.
+ * This function tries many parameter combinations and picks the best parameters.
+ * `*parameters` is filled with the best parameters found,
+ * dictionary constructed with those parameters is stored in `dictBuffer`.
+ *
+ * All of the parameters d, k, steps are optional.
+ * If d is non-zero then we don't check multiple values of d, otherwise we check d = {6, 8}.
+ * if steps is zero it defaults to its default value.
+ * If k is non-zero then we don't check multiple values of k, otherwise we check steps values in [50, 2000].
+ *
+ * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`)
+ *          or an error code, which can be tested with ZDICT_isError().
+ *          On success `*parameters` contains the parameters selected.
+ *          See ZDICT_trainFromBuffer() for details on failure modes.
+ * Note: ZDICT_optimizeTrainFromBuffer_cover() requires about 8 bytes of memory for each input byte and additionally another 5 bytes of memory for each byte of memory for each thread.
+ */
+ZDICTLIB_STATIC_API size_t ZDICT_optimizeTrainFromBuffer_cover(
+          void* dictBuffer, size_t dictBufferCapacity,
+    const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
+          ZDICT_cover_params_t* parameters);
+
+/*! ZDICT_trainFromBuffer_fastCover():
+ *  Train a dictionary from an array of samples using a modified version of COVER algorithm.
+ *  Samples must be stored concatenated in a single flat buffer `samplesBuffer`,
+ *  supplied with an array of sizes `samplesSizes`, providing the size of each sample, in order.
+ *  d and k are required.
+ *  All other parameters are optional, will use default values if not provided
+ *  The resulting dictionary will be saved into `dictBuffer`.
+ * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`)
+ *          or an error code, which can be tested with ZDICT_isError().
+ *          See ZDICT_trainFromBuffer() for details on failure modes.
+ *  Note: ZDICT_trainFromBuffer_fastCover() requires 6 * 2^f bytes of memory.
+ *  Tips: In general, a reasonable dictionary has a size of ~ 100 KB.
+ *        It's possible to select smaller or larger size, just by specifying `dictBufferCapacity`.
+ *        In general, it's recommended to provide a few thousands samples, though this can vary a lot.
+ *        It's recommended that total size of all samples be about ~x100 times the target size of dictionary.
+ */
+ZDICTLIB_STATIC_API size_t ZDICT_trainFromBuffer_fastCover(void *dictBuffer,
+                    size_t dictBufferCapacity, const void *samplesBuffer,
+                    const size_t *samplesSizes, unsigned nbSamples,
+                    ZDICT_fastCover_params_t parameters);
+
+/*! ZDICT_optimizeTrainFromBuffer_fastCover():
+ * The same requirements as above hold for all the parameters except `parameters`.
+ * This function tries many parameter combinations (specifically, k and d combinations)
+ * and picks the best parameters. `*parameters` is filled with the best parameters found,
+ * dictionary constructed with those parameters is stored in `dictBuffer`.
+ * All of the parameters d, k, steps, f, and accel are optional.
+ * If d is non-zero then we don't check multiple values of d, otherwise we check d = {6, 8}.
+ * if steps is zero it defaults to its default value.
+ * If k is non-zero then we don't check multiple values of k, otherwise we check steps values in [50, 2000].
+ * If f is zero, default value of 20 is used.
+ * If accel is zero, default value of 1 is used.
+ *
+ * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`)
+ *          or an error code, which can be tested with ZDICT_isError().
+ *          On success `*parameters` contains the parameters selected.
+ *          See ZDICT_trainFromBuffer() for details on failure modes.
+ * Note: ZDICT_optimizeTrainFromBuffer_fastCover() requires about 6 * 2^f bytes of memory for each thread.
+ */
+ZDICTLIB_STATIC_API size_t ZDICT_optimizeTrainFromBuffer_fastCover(void* dictBuffer,
+                    size_t dictBufferCapacity, const void* samplesBuffer,
+                    const size_t* samplesSizes, unsigned nbSamples,
+                    ZDICT_fastCover_params_t* parameters);
+
+typedef struct {
+    unsigned selectivityLevel;   /* 0 means default; larger => select more => larger dictionary */
+    ZDICT_params_t zParams;
+} ZDICT_legacy_params_t;
+
+/*! ZDICT_trainFromBuffer_legacy():
+ *  Train a dictionary from an array of samples.
+ *  Samples must be stored concatenated in a single flat buffer `samplesBuffer`,
+ *  supplied with an array of sizes `samplesSizes`, providing the size of each sample, in order.
+ *  The resulting dictionary will be saved into `dictBuffer`.
+ * `parameters` is optional and can be provided with values set to 0 to mean "default".
+ * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`)
+ *          or an error code, which can be tested with ZDICT_isError().
+ *          See ZDICT_trainFromBuffer() for details on failure modes.
+ *  Tips: In general, a reasonable dictionary has a size of ~ 100 KB.
+ *        It's possible to select smaller or larger size, just by specifying `dictBufferCapacity`.
+ *        In general, it's recommended to provide a few thousands samples, though this can vary a lot.
+ *        It's recommended that total size of all samples be about ~x100 times the target size of dictionary.
+ *  Note: ZDICT_trainFromBuffer_legacy() will send notifications into stderr if instructed to, using notificationLevel>0.
+ */
+ZDICTLIB_STATIC_API size_t ZDICT_trainFromBuffer_legacy(
+    void* dictBuffer, size_t dictBufferCapacity,
+    const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
+    ZDICT_legacy_params_t parameters);
+
+
+/* Deprecation warnings */
+/* It is generally possible to disable deprecation warnings from compiler,
+   for example with -Wno-deprecated-declarations for gcc
+   or _CRT_SECURE_NO_WARNINGS in Visual.
+   Otherwise, it's also possible to manually define ZDICT_DISABLE_DEPRECATE_WARNINGS */
+#ifdef ZDICT_DISABLE_DEPRECATE_WARNINGS
+#  define ZDICT_DEPRECATED(message) /* disable deprecation warnings */
+#else
+#  define ZDICT_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
+#  if defined (__cplusplus) && (__cplusplus >= 201402) /* C++14 or greater */
+#    define ZDICT_DEPRECATED(message) [[deprecated(message)]]
+#  elif defined(__clang__) || (ZDICT_GCC_VERSION >= 405)
+#    define ZDICT_DEPRECATED(message) __attribute__((deprecated(message)))
+#  elif (ZDICT_GCC_VERSION >= 301)
+#    define ZDICT_DEPRECATED(message) __attribute__((deprecated))
+#  elif defined(_MSC_VER)
+#    define ZDICT_DEPRECATED(message) __declspec(deprecated(message))
+#  else
+#    pragma message("WARNING: You need to implement ZDICT_DEPRECATED for this compiler")
+#    define ZDICT_DEPRECATED(message)
+#  endif
+#endif /* ZDICT_DISABLE_DEPRECATE_WARNINGS */
+
+ZDICT_DEPRECATED("use ZDICT_finalizeDictionary() instead")
+ZDICTLIB_STATIC_API
+size_t ZDICT_addEntropyTablesFromBuffer(void* dictBuffer, size_t dictContentSize, size_t dictBufferCapacity,
+                                  const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples);
+
+
+#endif   /* ZSTD_ZDICT_H_STATIC */
+
+#if defined (__cplusplus)
+}
+#endif
diff --git a/contrib/zstd/zstd.h b/contrib/zstd/zstd.h
index 8c6fc6ae9..95aac0737 100644
--- a/contrib/zstd/zstd.h
+++ b/contrib/zstd/zstd.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -14,27 +14,61 @@ extern "C" {
 #ifndef ZSTD_H_235446
 #define ZSTD_H_235446
 
-/* ======   Dependency   ======*/
+/* ======   Dependencies   ======*/
 #include <limits.h>   /* INT_MAX */
 #include <stddef.h>   /* size_t */
 
 
 /* =====   ZSTDLIB_API : control library symbols visibility   ===== */
-#ifndef ZSTDLIB_VISIBILITY
-#  if defined(__GNUC__) && (__GNUC__ >= 4)
-#    define ZSTDLIB_VISIBILITY __attribute__ ((visibility ("default")))
+#ifndef ZSTDLIB_VISIBLE
+   /* Backwards compatibility with old macro name */
+#  ifdef ZSTDLIB_VISIBILITY
+#    define ZSTDLIB_VISIBLE ZSTDLIB_VISIBILITY
+#  elif defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
+#    define ZSTDLIB_VISIBLE __attribute__ ((visibility ("default")))
 #  else
-#    define ZSTDLIB_VISIBILITY
+#    define ZSTDLIB_VISIBLE
 #  endif
 #endif
+
+#ifndef ZSTDLIB_HIDDEN
+#  if defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
+#    define ZSTDLIB_HIDDEN __attribute__ ((visibility ("hidden")))
+#  else
+#    define ZSTDLIB_HIDDEN
+#  endif
+#endif
+
 #if defined(ZSTD_DLL_EXPORT) && (ZSTD_DLL_EXPORT==1)
-#  define ZSTDLIB_API __declspec(dllexport) ZSTDLIB_VISIBILITY
+#  define ZSTDLIB_API __declspec(dllexport) ZSTDLIB_VISIBLE
 #elif defined(ZSTD_DLL_IMPORT) && (ZSTD_DLL_IMPORT==1)
-#  define ZSTDLIB_API __declspec(dllimport) ZSTDLIB_VISIBILITY /* It isn't required but allows to generate better code, saving a function pointer load from the IAT and an indirect jump.*/
+#  define ZSTDLIB_API __declspec(dllimport) ZSTDLIB_VISIBLE /* It isn't required but allows to generate better code, saving a function pointer load from the IAT and an indirect jump.*/
 #else
-#  define ZSTDLIB_API ZSTDLIB_VISIBILITY
+#  define ZSTDLIB_API ZSTDLIB_VISIBLE
 #endif
 
+/* Deprecation warnings :
+ * Should these warnings be a problem, it is generally possible to disable them,
+ * typically with -Wno-deprecated-declarations for gcc or _CRT_SECURE_NO_WARNINGS in Visual.
+ * Otherwise, it's also possible to define ZSTD_DISABLE_DEPRECATE_WARNINGS.
+ */
+#ifdef ZSTD_DISABLE_DEPRECATE_WARNINGS
+#  define ZSTD_DEPRECATED(message) /* disable deprecation warnings */
+#else
+#  if defined (__cplusplus) && (__cplusplus >= 201402) /* C++14 or greater */
+#    define ZSTD_DEPRECATED(message) [[deprecated(message)]]
+#  elif (defined(GNUC) && (GNUC > 4 || (GNUC == 4 && GNUC_MINOR >= 5))) || defined(__clang__)
+#    define ZSTD_DEPRECATED(message) __attribute__((deprecated(message)))
+#  elif defined(__GNUC__) && (__GNUC__ >= 3)
+#    define ZSTD_DEPRECATED(message) __attribute__((deprecated))
+#  elif defined(_MSC_VER)
+#    define ZSTD_DEPRECATED(message) __declspec(deprecated(message))
+#  else
+#    pragma message("WARNING: You need to implement ZSTD_DEPRECATED for this compiler")
+#    define ZSTD_DEPRECATED(message)
+#  endif
+#endif /* ZSTD_DISABLE_DEPRECATE_WARNINGS */
+
 
 /*******************************************************************************
   Introduction
@@ -71,17 +105,22 @@ extern "C" {
 
 /*------   Version   ------*/
 #define ZSTD_VERSION_MAJOR    1
-#define ZSTD_VERSION_MINOR    4
-#define ZSTD_VERSION_RELEASE  5
-
+#define ZSTD_VERSION_MINOR    5
+#define ZSTD_VERSION_RELEASE  4
 #define ZSTD_VERSION_NUMBER  (ZSTD_VERSION_MAJOR *100*100 + ZSTD_VERSION_MINOR *100 + ZSTD_VERSION_RELEASE)
-ZSTDLIB_API unsigned ZSTD_versionNumber(void);   /**< to check runtime library version */
+
+/*! ZSTD_versionNumber() :
+ *  Return runtime library version, the value is (MAJOR*100*100 + MINOR*100 + RELEASE). */
+ZSTDLIB_API unsigned ZSTD_versionNumber(void);
 
 #define ZSTD_LIB_VERSION ZSTD_VERSION_MAJOR.ZSTD_VERSION_MINOR.ZSTD_VERSION_RELEASE
 #define ZSTD_QUOTE(str) #str
 #define ZSTD_EXPAND_AND_QUOTE(str) ZSTD_QUOTE(str)
 #define ZSTD_VERSION_STRING ZSTD_EXPAND_AND_QUOTE(ZSTD_LIB_VERSION)
-ZSTDLIB_API const char* ZSTD_versionString(void);   /* requires v1.3.0+ */
+
+/*! ZSTD_versionString() :
+ *  Return runtime library version, like "1.4.5". Requires v1.3.0+. */
+ZSTDLIB_API const char* ZSTD_versionString(void);
 
 /* *************************************
  *  Default constant
@@ -104,7 +143,6 @@ ZSTDLIB_API const char* ZSTD_versionString(void);   /* requires v1.3.0+ */
 #define ZSTD_BLOCKSIZE_MAX     (1<<ZSTD_BLOCKSIZELOG_MAX)
 
 
-
 /***************************************
 *  Simple API
 ***************************************/
@@ -159,9 +197,11 @@ ZSTDLIB_API unsigned long long ZSTD_getFrameContentSize(const void *src, size_t
  *  "empty", "unknown" and "error" results to the same return value (0),
  *  while ZSTD_getFrameContentSize() gives them separate return values.
  * @return : decompressed size of `src` frame content _if known and not empty_, 0 otherwise. */
-ZSTDLIB_API unsigned long long ZSTD_getDecompressedSize(const void* src, size_t srcSize);
+ZSTD_DEPRECATED("Replaced by ZSTD_getFrameContentSize")
+ZSTDLIB_API
+unsigned long long ZSTD_getDecompressedSize(const void* src, size_t srcSize);
 
-/*! ZSTD_findFrameCompressedSize() :
+/*! ZSTD_findFrameCompressedSize() : Requires v1.4.0+
  * `src` should point to the start of a ZSTD frame or skippable frame.
  * `srcSize` must be >= first frame size
  * @return : the compressed size of the first frame starting at `src`,
@@ -171,12 +211,35 @@ ZSTDLIB_API size_t ZSTD_findFrameCompressedSize(const void* src, size_t srcSize)
 
 
 /*======  Helper functions  ======*/
-#define ZSTD_COMPRESSBOUND(srcSize)   ((srcSize) + ((srcSize)>>8) + (((srcSize) < (128<<10)) ? (((128<<10) - (srcSize)) >> 11) /* margin, from 64 to 0 */ : 0))  /* this formula ensures that bound(A) + bound(B) <= bound(A+B) as long as A and B >= 128 KB */
-ZSTDLIB_API size_t      ZSTD_compressBound(size_t srcSize); /*!< maximum compressed size in worst case single-pass scenario */
+/* ZSTD_compressBound() :
+ * maximum compressed size in worst case single-pass scenario.
+ * When invoking `ZSTD_compress()` or any other one-pass compression function,
+ * it's recommended to provide @dstCapacity >= ZSTD_compressBound(srcSize)
+ * as it eliminates one potential failure scenario,
+ * aka not enough room in dst buffer to write the compressed frame.
+ * Note : ZSTD_compressBound() itself can fail, if @srcSize > ZSTD_MAX_INPUT_SIZE .
+ *        In which case, ZSTD_compressBound() will return an error code
+ *        which can be tested using ZSTD_isError().
+ *
+ * ZSTD_COMPRESSBOUND() :
+ * same as ZSTD_compressBound(), but as a macro.
+ * It can be used to produce constants, which can be useful for static allocation,
+ * for example to size a static array on stack.
+ * Will produce constant value 0 if srcSize too large.
+ */
+#define ZSTD_MAX_INPUT_SIZE ((sizeof(size_t)==8) ? 0xFF00FF00FF00FF00LLU : 0xFF00FF00U)
+#define ZSTD_COMPRESSBOUND(srcSize)   (((size_t)(srcSize) >= ZSTD_MAX_INPUT_SIZE) ? 0 : (srcSize) + ((srcSize)>>8) + (((srcSize) < (128<<10)) ? (((128<<10) - (srcSize)) >> 11) /* margin, from 64 to 0 */ : 0))  /* this formula ensures that bound(A) + bound(B) <= bound(A+B) as long as A and B >= 128 KB */
+ZSTDLIB_API size_t ZSTD_compressBound(size_t srcSize); /*!< maximum compressed size in worst case single-pass scenario */
+/* ZSTD_isError() :
+ * Most ZSTD_* functions returning a size_t value can be tested for error,
+ * using ZSTD_isError().
+ * @return 1 if error, 0 otherwise
+ */
 ZSTDLIB_API unsigned    ZSTD_isError(size_t code);          /*!< tells if a `size_t` function result is an error code */
 ZSTDLIB_API const char* ZSTD_getErrorName(size_t code);     /*!< provides readable string from an error code */
-ZSTDLIB_API int         ZSTD_minCLevel(void);               /*!< minimum negative compression level allowed */
+ZSTDLIB_API int         ZSTD_minCLevel(void);               /*!< minimum negative compression level allowed, requires v1.4.0+ */
 ZSTDLIB_API int         ZSTD_maxCLevel(void);               /*!< maximum compression level available */
+ZSTDLIB_API int         ZSTD_defaultCLevel(void);           /*!< default compression level, specified by ZSTD_CLEVEL_DEFAULT, requires v1.5.0+ */
 
 
 /***************************************
@@ -194,7 +257,7 @@ ZSTDLIB_API int         ZSTD_maxCLevel(void);               /*!< maximum compres
  */
 typedef struct ZSTD_CCtx_s ZSTD_CCtx;
 ZSTDLIB_API ZSTD_CCtx* ZSTD_createCCtx(void);
-ZSTDLIB_API size_t     ZSTD_freeCCtx(ZSTD_CCtx* cctx);
+ZSTDLIB_API size_t     ZSTD_freeCCtx(ZSTD_CCtx* cctx);  /* accept NULL pointer */
 
 /*! ZSTD_compressCCtx() :
  *  Same as ZSTD_compress(), using an explicit ZSTD_CCtx.
@@ -217,7 +280,7 @@ ZSTDLIB_API size_t ZSTD_compressCCtx(ZSTD_CCtx* cctx,
  *  Use one context per thread for parallel execution. */
 typedef struct ZSTD_DCtx_s ZSTD_DCtx;
 ZSTDLIB_API ZSTD_DCtx* ZSTD_createDCtx(void);
-ZSTDLIB_API size_t     ZSTD_freeDCtx(ZSTD_DCtx* dctx);
+ZSTDLIB_API size_t     ZSTD_freeDCtx(ZSTD_DCtx* dctx);  /* accept NULL pointer */
 
 /*! ZSTD_decompressDCtx() :
  *  Same as ZSTD_decompress(),
@@ -229,9 +292,9 @@ ZSTDLIB_API size_t ZSTD_decompressDCtx(ZSTD_DCtx* dctx,
                                  const void* src, size_t srcSize);
 
 
-/***************************************
-*  Advanced compression API
-***************************************/
+/*********************************************
+*  Advanced compression API (Requires v1.4.0+)
+**********************************************/
 
 /* API design :
  *   Parameters are pushed one by one into an existing context,
@@ -242,7 +305,7 @@ ZSTDLIB_API size_t ZSTD_decompressDCtx(ZSTD_DCtx* dctx,
  *
  *   It's possible to reset all parameters to "default" using ZSTD_CCtx_reset().
  *
- *   This API supercedes all other "advanced" API entry points in the experimental section.
+ *   This API supersedes all other "advanced" API entry points in the experimental section.
  *   In the future, we expect to remove from experimental API entry points which are redundant with this API.
  */
 
@@ -261,7 +324,6 @@ typedef enum { ZSTD_fast=1,
                          Only the order (from fast to strong) is guaranteed */
 } ZSTD_strategy;
 
-
 typedef enum {
 
     /* compression parameters
@@ -327,14 +389,15 @@ typedef enum {
                               * The higher the value of selected strategy, the more complex it is,
                               * resulting in stronger and slower compression.
                               * Special: value 0 means "use default strategy". */
-
     /* LDM mode parameters */
     ZSTD_c_enableLongDistanceMatching=160, /* Enable long distance matching.
                                      * This parameter is designed to improve compression ratio
                                      * for large inputs, by finding large matches at long distance.
                                      * It increases memory usage and window size.
                                      * Note: enabling this parameter increases default ZSTD_c_windowLog to 128 MB
-                                     * except when expressly set to a different value. */
+                                     * except when expressly set to a different value.
+                                     * Note: will be enabled by default if ZSTD_c_windowLog >= 128 MB and
+                                     * compression strategy >= ZSTD_btopt (== compression level 16+) */
     ZSTD_c_ldmHashLog=161,   /* Size of the table for long distance matching, as a power of 2.
                               * Larger values increase memory usage and compression ratio,
                               * but decrease compression speed.
@@ -365,20 +428,24 @@ typedef enum {
     ZSTD_c_dictIDFlag=202,   /* When applicable, dictionary's ID is written into frame header (default:1) */
 
     /* multi-threading parameters */
-    /* These parameters are only useful if multi-threading is enabled (compiled with build macro ZSTD_MULTITHREAD).
-     * They return an error otherwise. */
+    /* These parameters are only active if multi-threading is enabled (compiled with build macro ZSTD_MULTITHREAD).
+     * Otherwise, trying to set any other value than default (0) will be a no-op and return an error.
+     * In a situation where it's unknown if the linked library supports multi-threading or not,
+     * setting ZSTD_c_nbWorkers to any value >= 1 and consulting the return value provides a quick way to check this property.
+     */
     ZSTD_c_nbWorkers=400,    /* Select how many threads will be spawned to compress in parallel.
-                              * When nbWorkers >= 1, triggers asynchronous mode when used with ZSTD_compressStream*() :
+                              * When nbWorkers >= 1, triggers asynchronous mode when invoking ZSTD_compressStream*() :
                               * ZSTD_compressStream*() consumes input and flush output if possible, but immediately gives back control to caller,
-                              * while compression work is performed in parallel, within worker threads.
+                              * while compression is performed in parallel, within worker thread(s).
                               * (note : a strong exception to this rule is when first invocation of ZSTD_compressStream2() sets ZSTD_e_end :
                               *  in which case, ZSTD_compressStream2() delegates to ZSTD_compress2(), which is always a blocking call).
                               * More workers improve speed, but also increase memory usage.
-                              * Default value is `0`, aka "single-threaded mode" : no worker is spawned, compression is performed inside Caller's thread, all invocations are blocking */
+                              * Default value is `0`, aka "single-threaded mode" : no worker is spawned,
+                              * compression is performed inside Caller's thread, and all invocations are blocking */
     ZSTD_c_jobSize=401,      /* Size of a compression job. This value is enforced only when nbWorkers >= 1.
                               * Each compression job is completed in parallel, so this value can indirectly impact the nb of active threads.
                               * 0 means default, which is dynamically determined based on compression parameters.
-                              * Job size must be a minimum of overlap size, or 1 MB, whichever is largest.
+                              * Job size must be a minimum of overlap size, or ZSTDMT_JOBSIZE_MIN (= 512 KB), whichever is largest.
                               * The minimum size is automatically and transparently enforced. */
     ZSTD_c_overlapLog=402,   /* Control the overlap size, as a fraction of window size.
                               * The overlap size is an amount of data reloaded from previous job at the beginning of a new job.
@@ -403,6 +470,16 @@ typedef enum {
      * ZSTD_c_literalCompressionMode
      * ZSTD_c_targetCBlockSize
      * ZSTD_c_srcSizeHint
+     * ZSTD_c_enableDedicatedDictSearch
+     * ZSTD_c_stableInBuffer
+     * ZSTD_c_stableOutBuffer
+     * ZSTD_c_blockDelimiters
+     * ZSTD_c_validateSequences
+     * ZSTD_c_useBlockSplitter
+     * ZSTD_c_useRowMatchFinder
+     * ZSTD_c_prefetchCDictTables
+     * ZSTD_c_enableSeqProducerFallback
+     * ZSTD_c_maxBlockSize
      * Because they are not stable, it's necessary to define ZSTD_STATIC_LINKING_ONLY to access them.
      * note : never ever use experimentalParam? names directly;
      *        also, the enums values themselves are unstable and can still change.
@@ -413,7 +490,19 @@ typedef enum {
      ZSTD_c_experimentalParam4=1001,
      ZSTD_c_experimentalParam5=1002,
      ZSTD_c_experimentalParam6=1003,
-     ZSTD_c_experimentalParam7=1004
+     ZSTD_c_experimentalParam7=1004,
+     ZSTD_c_experimentalParam8=1005,
+     ZSTD_c_experimentalParam9=1006,
+     ZSTD_c_experimentalParam10=1007,
+     ZSTD_c_experimentalParam11=1008,
+     ZSTD_c_experimentalParam12=1009,
+     ZSTD_c_experimentalParam13=1010,
+     ZSTD_c_experimentalParam14=1011,
+     ZSTD_c_experimentalParam15=1012,
+     ZSTD_c_experimentalParam16=1013,
+     ZSTD_c_experimentalParam17=1014,
+     ZSTD_c_experimentalParam18=1015,
+     ZSTD_c_experimentalParam19=1016
 } ZSTD_cParameter;
 
 typedef struct {
@@ -476,7 +565,7 @@ typedef enum {
  *                  They will be used to compress next frame.
  *                  Resetting session never fails.
  *  - The parameters : changes all parameters back to "default".
- *                  This removes any reference to any dictionary too.
+ *                  This also removes any reference to any dictionary or external sequence producer.
  *                  Parameters can only be changed between 2 sessions (i.e. no compression is currently ongoing)
  *                  otherwise the reset fails, and function returns an error value (which can be tested using ZSTD_isError())
  *  - Both : similar to resetting the session, followed by resetting parameters.
@@ -498,9 +587,9 @@ ZSTDLIB_API size_t ZSTD_compress2( ZSTD_CCtx* cctx,
                              const void* src, size_t srcSize);
 
 
-/***************************************
-*  Advanced decompression API
-***************************************/
+/***********************************************
+*  Advanced decompression API (Requires v1.4.0+)
+************************************************/
 
 /* The advanced API pushes parameters one by one into an existing DCtx context.
  * Parameters are sticky, and remain valid for all following frames
@@ -524,11 +613,17 @@ typedef enum {
      * At the time of this writing, they include :
      * ZSTD_d_format
      * ZSTD_d_stableOutBuffer
+     * ZSTD_d_forceIgnoreChecksum
+     * ZSTD_d_refMultipleDDicts
+     * ZSTD_d_disableHuffmanAssembly
      * Because they are not stable, it's necessary to define ZSTD_STATIC_LINKING_ONLY to access them.
      * note : never ever use experimentalParam? names directly
      */
      ZSTD_d_experimentalParam1=1000,
-     ZSTD_d_experimentalParam2=1001
+     ZSTD_d_experimentalParam2=1001,
+     ZSTD_d_experimentalParam3=1002,
+     ZSTD_d_experimentalParam4=1003,
+     ZSTD_d_experimentalParam5=1004
 
 } ZSTD_dParameter;
 
@@ -642,7 +737,7 @@ typedef ZSTD_CCtx ZSTD_CStream;  /**< CCtx and CStream are now effectively same
                                  /* Continue to distinguish them for compatibility with older versions <= v1.2.0 */
 /*===== ZSTD_CStream management functions =====*/
 ZSTDLIB_API ZSTD_CStream* ZSTD_createCStream(void);
-ZSTDLIB_API size_t ZSTD_freeCStream(ZSTD_CStream* zcs);
+ZSTDLIB_API size_t ZSTD_freeCStream(ZSTD_CStream* zcs);  /* accept NULL pointer */
 
 /*===== Streaming compression functions =====*/
 typedef enum {
@@ -658,14 +753,15 @@ typedef enum {
                         : note : multithreaded compression will block to flush as much output as possible. */
 } ZSTD_EndDirective;
 
-/*! ZSTD_compressStream2() :
+/*! ZSTD_compressStream2() : Requires v1.4.0+
  *  Behaves about the same as ZSTD_compressStream, with additional control on end directive.
  *  - Compression parameters are pushed into CCtx before starting compression, using ZSTD_CCtx_set*()
  *  - Compression parameters cannot be changed once compression is started (save a list of exceptions in multi-threading mode)
  *  - output->pos must be <= dstCapacity, input->pos must be <= srcSize
  *  - output->pos and input->pos will be updated. They are guaranteed to remain below their respective limit.
+ *  - endOp must be a valid directive
  *  - When nbWorkers==0 (default), function is blocking : it completes its job before returning to caller.
- *  - When nbWorkers>=1, function is non-blocking : it just acquires a copy of input, and distributes jobs to internal worker threads, flush whatever is available,
+ *  - When nbWorkers>=1, function is non-blocking : it copies a portion of input, distributes jobs to internal worker threads, flush to output whatever is available,
  *                                                  and then immediately returns, just indicating that there is some data remaining to be flushed.
  *                                                  The function nonetheless guarantees forward progress : it will return only after it reads or write at least 1+ byte.
  *  - Exception : if the first call requests a ZSTD_e_end directive and provides enough dstCapacity, the function delegates to ZSTD_compress2() which is always blocking.
@@ -703,11 +799,9 @@ ZSTDLIB_API size_t ZSTD_CStreamOutSize(void);   /**< recommended size for output
 
 
 /* *****************************************************************************
- * This following is a legacy streaming API.
+ * This following is a legacy streaming API, available since v1.0+ .
  * It can be replaced by ZSTD_CCtx_reset() and ZSTD_compressStream2().
  * It is redundant, but remains fully supported.
- * Advanced parameters and dictionary compression can only be used through the
- * new API.
  ******************************************************************************/
 
 /*!
@@ -716,6 +810,9 @@ ZSTDLIB_API size_t ZSTD_CStreamOutSize(void);   /**< recommended size for output
  *     ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only);
  *     ZSTD_CCtx_refCDict(zcs, NULL); // clear the dictionary (if any)
  *     ZSTD_CCtx_setParameter(zcs, ZSTD_c_compressionLevel, compressionLevel);
+ *
+ * Note that ZSTD_initCStream() clears any previously set dictionary. Use the new API
+ * to compress with a dictionary.
  */
 ZSTDLIB_API size_t ZSTD_initCStream(ZSTD_CStream* zcs, int compressionLevel);
 /*!
@@ -762,17 +859,35 @@ typedef ZSTD_DCtx ZSTD_DStream;  /**< DCtx and DStream are now effectively same
                                  /* For compatibility with versions <= v1.2.0, prefer differentiating them. */
 /*===== ZSTD_DStream management functions =====*/
 ZSTDLIB_API ZSTD_DStream* ZSTD_createDStream(void);
-ZSTDLIB_API size_t ZSTD_freeDStream(ZSTD_DStream* zds);
+ZSTDLIB_API size_t ZSTD_freeDStream(ZSTD_DStream* zds);  /* accept NULL pointer */
 
 /*===== Streaming decompression functions =====*/
 
-/* This function is redundant with the advanced API and equivalent to:
+/*! ZSTD_initDStream() :
+ * Initialize/reset DStream state for new decompression operation.
+ * Call before new decompression operation using same DStream.
  *
+ * Note : This function is redundant with the advanced API and equivalent to:
  *     ZSTD_DCtx_reset(zds, ZSTD_reset_session_only);
  *     ZSTD_DCtx_refDDict(zds, NULL);
  */
 ZSTDLIB_API size_t ZSTD_initDStream(ZSTD_DStream* zds);
 
+/*! ZSTD_decompressStream() :
+ * Streaming decompression function.
+ * Call repetitively to consume full input updating it as necessary.
+ * Function will update both input and output `pos` fields exposing current state via these fields:
+ * - `input.pos < input.size`, some input remaining and caller should provide remaining input
+ *   on the next call.
+ * - `output.pos < output.size`, decoder finished and flushed all remaining buffers.
+ * - `output.pos == output.size`, potentially uncflushed data present in the internal buffers,
+ *   call ZSTD_decompressStream() again to flush remaining data to output.
+ * Note : with no additional input, amount of data flushed <= ZSTD_BLOCKSIZE_MAX.
+ *
+ * @return : 0 when a frame is completely decoded and fully flushed,
+ *           or an error code, which can be tested using ZSTD_isError(),
+ *           or any other value > 0, which means there is some decoding or flushing to do to complete current frame.
+ */
 ZSTDLIB_API size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inBuffer* input);
 
 ZSTDLIB_API size_t ZSTD_DStreamInSize(void);    /*!< recommended size for input buffer */
@@ -785,7 +900,7 @@ ZSTDLIB_API size_t ZSTD_DStreamOutSize(void);   /*!< recommended size for output
 /*! ZSTD_compress_usingDict() :
  *  Compression at an explicit compression level using a Dictionary.
  *  A dictionary can be any arbitrary data segment (also called a prefix),
- *  or a buffer with specified information (see dictBuilder/zdict.h).
+ *  or a buffer with specified information (see zdict.h).
  *  Note : This function loads the dictionary, resulting in significant startup delay.
  *         It's intended for a dictionary used only once.
  *  Note 2 : When `dict == NULL || dictSize < 8` no dictionary is used. */
@@ -828,7 +943,8 @@ ZSTDLIB_API ZSTD_CDict* ZSTD_createCDict(const void* dictBuffer, size_t dictSize
                                          int compressionLevel);
 
 /*! ZSTD_freeCDict() :
- *  Function frees memory allocated by ZSTD_createCDict(). */
+ *  Function frees memory allocated by ZSTD_createCDict().
+ *  If a NULL pointer is passed, no operation is performed. */
 ZSTDLIB_API size_t      ZSTD_freeCDict(ZSTD_CDict* CDict);
 
 /*! ZSTD_compress_usingCDict() :
@@ -850,7 +966,8 @@ typedef struct ZSTD_DDict_s ZSTD_DDict;
 ZSTDLIB_API ZSTD_DDict* ZSTD_createDDict(const void* dictBuffer, size_t dictSize);
 
 /*! ZSTD_freeDDict() :
- *  Function frees memory allocated with ZSTD_createDDict() */
+ *  Function frees memory allocated with ZSTD_createDDict()
+ *  If a NULL pointer is passed, no operation is performed. */
 ZSTDLIB_API size_t      ZSTD_freeDDict(ZSTD_DDict* ddict);
 
 /*! ZSTD_decompress_usingDDict() :
@@ -866,24 +983,30 @@ ZSTDLIB_API size_t ZSTD_decompress_usingDDict(ZSTD_DCtx* dctx,
  *  Dictionary helper functions
  *******************************/
 
-/*! ZSTD_getDictID_fromDict() :
+/*! ZSTD_getDictID_fromDict() : Requires v1.4.0+
  *  Provides the dictID stored within dictionary.
  *  if @return == 0, the dictionary is not conformant with Zstandard specification.
  *  It can still be loaded, but as a content-only dictionary. */
 ZSTDLIB_API unsigned ZSTD_getDictID_fromDict(const void* dict, size_t dictSize);
 
-/*! ZSTD_getDictID_fromDDict() :
+/*! ZSTD_getDictID_fromCDict() : Requires v1.5.0+
+ *  Provides the dictID of the dictionary loaded into `cdict`.
+ *  If @return == 0, the dictionary is not conformant to Zstandard specification, or empty.
+ *  Non-conformant dictionaries can still be loaded, but as content-only dictionaries. */
+ZSTDLIB_API unsigned ZSTD_getDictID_fromCDict(const ZSTD_CDict* cdict);
+
+/*! ZSTD_getDictID_fromDDict() : Requires v1.4.0+
  *  Provides the dictID of the dictionary loaded into `ddict`.
  *  If @return == 0, the dictionary is not conformant to Zstandard specification, or empty.
  *  Non-conformant dictionaries can still be loaded, but as content-only dictionaries. */
 ZSTDLIB_API unsigned ZSTD_getDictID_fromDDict(const ZSTD_DDict* ddict);
 
-/*! ZSTD_getDictID_fromFrame() :
+/*! ZSTD_getDictID_fromFrame() : Requires v1.4.0+
  *  Provides the dictID required to decompressed the frame stored within `src`.
  *  If @return == 0, the dictID could not be decoded.
  *  This could for one of the following reasons :
  *  - The frame does not require a dictionary to be decoded (most common case).
- *  - The frame was built with dictID intentionally removed. Whatever dictionary is necessary is a hidden information.
+ *  - The frame was built with dictID intentionally removed. Whatever dictionary is necessary is a hidden piece of information.
  *    Note : this use case also happens when using a non-conformant dictionary.
  *  - `srcSize` is too small, and as a result, the frame header could not be decoded (only possible if `srcSize < ZSTD_FRAMEHEADERSIZE_MAX`).
  *  - This is not a Zstandard frame.
@@ -892,23 +1015,24 @@ ZSTDLIB_API unsigned ZSTD_getDictID_fromFrame(const void* src, size_t srcSize);
 
 
 /*******************************************************************************
- * Advanced dictionary and prefix API
+ * Advanced dictionary and prefix API (Requires v1.4.0+)
  *
  * This API allows dictionaries to be used with ZSTD_compress2(),
- * ZSTD_compressStream2(), and ZSTD_decompress(). Dictionaries are sticky, and
+ * ZSTD_compressStream2(), and ZSTD_decompressDCtx(). Dictionaries are sticky, and
  * only reset with the context is reset with ZSTD_reset_parameters or
  * ZSTD_reset_session_and_parameters. Prefixes are single-use.
  ******************************************************************************/
 
 
-/*! ZSTD_CCtx_loadDictionary() :
+/*! ZSTD_CCtx_loadDictionary() : Requires v1.4.0+
  *  Create an internal CDict from `dict` buffer.
  *  Decompression will have to use same dictionary.
  * @result : 0, or an error code (which can be tested with ZSTD_isError()).
  *  Special: Loading a NULL (or 0-size) dictionary invalidates previous dictionary,
  *           meaning "return to no-dictionary mode".
- *  Note 1 : Dictionary is sticky, it will be used for all future compressed frames.
- *           To return to "no-dictionary" situation, load a NULL dictionary (or reset parameters).
+ *  Note 1 : Dictionary is sticky, it will be used for all future compressed frames,
+ *           until parameters are reset, a new dictionary is loaded, or the dictionary
+ *           is explicitly invalidated by loading a NULL dictionary.
  *  Note 2 : Loading a dictionary involves building tables.
  *           It's also a CPU consuming operation, with non-negligible impact on latency.
  *           Tables are dependent on compression parameters, and for this reason,
@@ -920,11 +1044,11 @@ ZSTDLIB_API unsigned ZSTD_getDictID_fromFrame(const void* src, size_t srcSize);
  *           to precisely select how dictionary content must be interpreted. */
 ZSTDLIB_API size_t ZSTD_CCtx_loadDictionary(ZSTD_CCtx* cctx, const void* dict, size_t dictSize);
 
-/*! ZSTD_CCtx_refCDict() :
- *  Reference a prepared dictionary, to be used for all next compressed frames.
+/*! ZSTD_CCtx_refCDict() : Requires v1.4.0+
+ *  Reference a prepared dictionary, to be used for all future compressed frames.
  *  Note that compression parameters are enforced from within CDict,
  *  and supersede any compression parameter previously set within CCtx.
- *  The parameters ignored are labled as "superseded-by-cdict" in the ZSTD_cParameter enum docs.
+ *  The parameters ignored are labelled as "superseded-by-cdict" in the ZSTD_cParameter enum docs.
  *  The ignored parameters will be used again if the CCtx is returned to no-dictionary mode.
  *  The dictionary will remain valid for future compressed frames using same CCtx.
  * @result : 0, or an error code (which can be tested with ZSTD_isError()).
@@ -934,7 +1058,7 @@ ZSTDLIB_API size_t ZSTD_CCtx_loadDictionary(ZSTD_CCtx* cctx, const void* dict, s
  *  Note 2 : CDict is just referenced, its lifetime must outlive its usage within CCtx. */
 ZSTDLIB_API size_t ZSTD_CCtx_refCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict);
 
-/*! ZSTD_CCtx_refPrefix() :
+/*! ZSTD_CCtx_refPrefix() : Requires v1.4.0+
  *  Reference a prefix (single-usage dictionary) for next compressed frame.
  *  A prefix is **only used once**. Tables are discarded at end of frame (ZSTD_e_end).
  *  Decompression will need same prefix to properly regenerate data.
@@ -955,10 +1079,10 @@ ZSTDLIB_API size_t ZSTD_CCtx_refCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict);
 ZSTDLIB_API size_t ZSTD_CCtx_refPrefix(ZSTD_CCtx* cctx,
                                  const void* prefix, size_t prefixSize);
 
-/*! ZSTD_DCtx_loadDictionary() :
- *  Create an internal DDict from dict buffer,
- *  to be used to decompress next frames.
- *  The dictionary remains valid for all future frames, until explicitly invalidated.
+/*! ZSTD_DCtx_loadDictionary() : Requires v1.4.0+
+ *  Create an internal DDict from dict buffer, to be used to decompress all future frames.
+ *  The dictionary remains valid for all future frames, until explicitly invalidated, or
+ *  a new dictionary is loaded.
  * @result : 0, or an error code (which can be tested with ZSTD_isError()).
  *  Special : Adding a NULL (or 0-size) dictionary invalidates any previous dictionary,
  *            meaning "return to no-dictionary mode".
@@ -972,18 +1096,26 @@ ZSTDLIB_API size_t ZSTD_CCtx_refPrefix(ZSTD_CCtx* cctx,
  */
 ZSTDLIB_API size_t ZSTD_DCtx_loadDictionary(ZSTD_DCtx* dctx, const void* dict, size_t dictSize);
 
-/*! ZSTD_DCtx_refDDict() :
+/*! ZSTD_DCtx_refDDict() : Requires v1.4.0+
  *  Reference a prepared dictionary, to be used to decompress next frames.
  *  The dictionary remains active for decompression of future frames using same DCtx.
+ *
+ *  If called with ZSTD_d_refMultipleDDicts enabled, repeated calls of this function
+ *  will store the DDict references in a table, and the DDict used for decompression
+ *  will be determined at decompression time, as per the dict ID in the frame.
+ *  The memory for the table is allocated on the first call to refDDict, and can be
+ *  freed with ZSTD_freeDCtx().
+ *
+ *  If called with ZSTD_d_refMultipleDDicts disabled (the default), only one dictionary
+ *  will be managed, and referencing a dictionary effectively "discards" any previous one.
+ *
  * @result : 0, or an error code (which can be tested with ZSTD_isError()).
- *  Note 1 : Currently, only one dictionary can be managed.
- *           Referencing a new dictionary effectively "discards" any previous one.
  *  Special: referencing a NULL DDict means "return to no-dictionary mode".
  *  Note 2 : DDict is just referenced, its lifetime must outlive its usage from DCtx.
  */
 ZSTDLIB_API size_t ZSTD_DCtx_refDDict(ZSTD_DCtx* dctx, const ZSTD_DDict* ddict);
 
-/*! ZSTD_DCtx_refPrefix() :
+/*! ZSTD_DCtx_refPrefix() : Requires v1.4.0+
  *  Reference a prefix (single-usage dictionary) to decompress next frame.
  *  This is the reverse operation of ZSTD_CCtx_refPrefix(),
  *  and must use the same prefix as the one used during compression.
@@ -1004,7 +1136,7 @@ ZSTDLIB_API size_t ZSTD_DCtx_refPrefix(ZSTD_DCtx* dctx,
 
 /* ===   Memory management   === */
 
-/*! ZSTD_sizeof_*() :
+/*! ZSTD_sizeof_*() : Requires v1.4.0+
  *  These functions give the _current_ memory usage of selected object.
  *  Note that object memory usage can evolve (increase or decrease) over time. */
 ZSTDLIB_API size_t ZSTD_sizeof_CCtx(const ZSTD_CCtx* cctx);
@@ -1029,6 +1161,17 @@ ZSTDLIB_API size_t ZSTD_sizeof_DDict(const ZSTD_DDict* ddict);
 #if defined(ZSTD_STATIC_LINKING_ONLY) && !defined(ZSTD_H_ZSTD_STATIC_LINKING_ONLY)
 #define ZSTD_H_ZSTD_STATIC_LINKING_ONLY
 
+/* This can be overridden externally to hide static symbols. */
+#ifndef ZSTDLIB_STATIC_API
+#  if defined(ZSTD_DLL_EXPORT) && (ZSTD_DLL_EXPORT==1)
+#    define ZSTDLIB_STATIC_API __declspec(dllexport) ZSTDLIB_VISIBLE
+#  elif defined(ZSTD_DLL_IMPORT) && (ZSTD_DLL_IMPORT==1)
+#    define ZSTDLIB_STATIC_API __declspec(dllimport) ZSTDLIB_VISIBLE
+#  else
+#    define ZSTDLIB_STATIC_API ZSTDLIB_VISIBLE
+#  endif
+#endif
+
 /****************************************************************************************
  *   experimental API (static linking only)
  ****************************************************************************************
@@ -1063,6 +1206,7 @@ ZSTDLIB_API size_t ZSTD_sizeof_DDict(const ZSTD_DDict* ddict);
 #define ZSTD_TARGETLENGTH_MIN     0   /* note : comparing this constant to an unsigned results in a tautological test */
 #define ZSTD_STRATEGY_MIN        ZSTD_fast
 #define ZSTD_STRATEGY_MAX        ZSTD_btultra2
+#define ZSTD_BLOCKSIZE_MAX_MIN (1 << 10) /* The minimum valid max blocksize. Maximum blocksizes smaller than this make compressBound() inaccurate. */
 
 
 #define ZSTD_OVERLAPLOG_MIN       0
@@ -1091,30 +1235,46 @@ ZSTDLIB_API size_t ZSTD_sizeof_DDict(const ZSTD_DDict* ddict);
 #define ZSTD_SRCSIZEHINT_MIN        0
 #define ZSTD_SRCSIZEHINT_MAX        INT_MAX
 
-/* internal */
-#define ZSTD_HASHLOG3_MAX           17
-
 
 /* ---  Advanced types  --- */
 
 typedef struct ZSTD_CCtx_params_s ZSTD_CCtx_params;
 
 typedef struct {
-    unsigned int matchPos; /* Match pos in dst */
-    /* If seqDef.offset > 3, then this is seqDef.offset - 3
-     * If seqDef.offset < 3, then this is the corresponding repeat offset
-     * But if seqDef.offset < 3 and litLength == 0, this is the
-     *   repeat offset before the corresponding repeat offset
-     * And if seqDef.offset == 3 and litLength == 0, this is the
-     *   most recent repeat offset - 1
-     */
-    unsigned int offset;
-    unsigned int litLength; /* Literal length */
-    unsigned int matchLength; /* Match length */
-    /* 0 when seq not rep and seqDef.offset otherwise
-     * when litLength == 0 this will be <= 4, otherwise <= 3 like normal
-     */
-    unsigned int rep;
+    unsigned int offset;      /* The offset of the match. (NOT the same as the offset code)
+                               * If offset == 0 and matchLength == 0, this sequence represents the last
+                               * literals in the block of litLength size.
+                               */
+
+    unsigned int litLength;   /* Literal length of the sequence. */
+    unsigned int matchLength; /* Match length of the sequence. */
+
+                              /* Note: Users of this API may provide a sequence with matchLength == litLength == offset == 0.
+                               * In this case, we will treat the sequence as a marker for a block boundary.
+                               */
+
+    unsigned int rep;         /* Represents which repeat offset is represented by the field 'offset'.
+                               * Ranges from [0, 3].
+                               *
+                               * Repeat offsets are essentially previous offsets from previous sequences sorted in
+                               * recency order. For more detail, see doc/zstd_compression_format.md
+                               *
+                               * If rep == 0, then 'offset' does not contain a repeat offset.
+                               * If rep > 0:
+                               *  If litLength != 0:
+                               *      rep == 1 --> offset == repeat_offset_1
+                               *      rep == 2 --> offset == repeat_offset_2
+                               *      rep == 3 --> offset == repeat_offset_3
+                               *  If litLength == 0:
+                               *      rep == 1 --> offset == repeat_offset_2
+                               *      rep == 2 --> offset == repeat_offset_3
+                               *      rep == 3 --> offset == repeat_offset_1 - 1
+                               *
+                               * Note: This field is optional. ZSTD_generateSequences() will calculate the value of
+                               * 'rep', but repeat offsets do not necessarily need to be calculated from an external
+                               * sequence provider's perspective. For example, ZSTD_compressSequences() does not
+                               * use this 'rep' field at all (as of now).
+                               */
 } ZSTD_Sequence;
 
 typedef struct {
@@ -1157,6 +1317,18 @@ typedef enum {
 } ZSTD_format_e;
 
 typedef enum {
+    /* Note: this enum controls ZSTD_d_forceIgnoreChecksum */
+    ZSTD_d_validateChecksum = 0,
+    ZSTD_d_ignoreChecksum = 1
+} ZSTD_forceIgnoreChecksum_e;
+
+typedef enum {
+    /* Note: this enum controls ZSTD_d_refMultipleDDicts */
+    ZSTD_rmd_refSingleDDict = 0,
+    ZSTD_rmd_refMultipleDDicts = 1
+} ZSTD_refMultipleDDicts_e;
+
+typedef enum {
     /* Note: this enum and the behavior it controls are effectively internal
      * implementation details of the compressor. They are expected to continue
      * to evolve and should be considered only in the context of extremely
@@ -1204,6 +1376,15 @@ typedef enum {
   ZSTD_lcm_uncompressed = 2   /**< Always emit uncompressed literals. */
 } ZSTD_literalCompressionMode_e;
 
+typedef enum {
+  /* Note: This enum controls features which are conditionally beneficial. Zstd typically will make a final
+   * decision on whether or not to enable the feature (ZSTD_ps_auto), but setting the switch to ZSTD_ps_enable
+   * or ZSTD_ps_disable allow for a force enable/disable the feature.
+   */
+  ZSTD_ps_auto = 0,         /* Let the library automatically determine whether the feature shall be enabled */
+  ZSTD_ps_enable = 1,       /* Force-enable the feature */
+  ZSTD_ps_disable = 2       /* Do not use the feature */
+} ZSTD_paramSwitch_e;
 
 /***************************************
 *  Frame size functions
@@ -1230,14 +1411,14 @@ typedef enum {
  *   note 5 : ZSTD_findDecompressedSize handles multiple frames, and so it must traverse the input to
  *            read each contained frame header.  This is fast as most of the data is skipped,
  *            however it does mean that all frame data must be present and valid. */
-ZSTDLIB_API unsigned long long ZSTD_findDecompressedSize(const void* src, size_t srcSize);
+ZSTDLIB_STATIC_API unsigned long long ZSTD_findDecompressedSize(const void* src, size_t srcSize);
 
 /*! ZSTD_decompressBound() :
  *  `src` should point to the start of a series of ZSTD encoded and/or skippable frames
  *  `srcSize` must be the _exact_ size of this series
  *       (i.e. there should be a frame boundary at `src + srcSize`)
  *  @return : - upper-bound for the decompressed size of all data in all successive frames
- *            - if an error occured: ZSTD_CONTENTSIZE_ERROR
+ *            - if an error occurred: ZSTD_CONTENTSIZE_ERROR
  *
  *  note 1  : an error can occur if `src` contains an invalid or incorrectly formatted frame.
  *  note 2  : the upper-bound is exact when the decompressed size field is available in every ZSTD encoded frame of `src`.
@@ -1245,22 +1426,178 @@ ZSTDLIB_API unsigned long long ZSTD_findDecompressedSize(const void* src, size_t
  *  note 3  : when the decompressed size field isn't available, the upper-bound for that frame is calculated by:
  *              upper-bound = # blocks * min(128 KB, Window_Size)
  */
-ZSTDLIB_API unsigned long long ZSTD_decompressBound(const void* src, size_t srcSize);
+ZSTDLIB_STATIC_API unsigned long long ZSTD_decompressBound(const void* src, size_t srcSize);
 
 /*! ZSTD_frameHeaderSize() :
  *  srcSize must be >= ZSTD_FRAMEHEADERSIZE_PREFIX.
  * @return : size of the Frame Header,
  *           or an error code (if srcSize is too small) */
-ZSTDLIB_API size_t ZSTD_frameHeaderSize(const void* src, size_t srcSize);
+ZSTDLIB_STATIC_API size_t ZSTD_frameHeaderSize(const void* src, size_t srcSize);
+
+/*! ZSTD_decompressionMargin() :
+ * Zstd supports in-place decompression, where the input and output buffers overlap.
+ * In this case, the output buffer must be at least (Margin + Output_Size) bytes large,
+ * and the input buffer must be at the end of the output buffer.
+ *
+ *  _______________________ Output Buffer ________________________
+ * |                                                              |
+ * |                                        ____ Input Buffer ____|
+ * |                                       |                      |
+ * v                                       v                      v
+ * |---------------------------------------|-----------|----------|
+ * ^                                                   ^          ^
+ * |___________________ Output_Size ___________________|_ Margin _|
+ *
+ * NOTE: See also ZSTD_DECOMPRESSION_MARGIN().
+ * NOTE: This applies only to single-pass decompression through ZSTD_decompress() or
+ * ZSTD_decompressDCtx().
+ * NOTE: This function supports multi-frame input.
+ *
+ * @param src The compressed frame(s)
+ * @param srcSize The size of the compressed frame(s)
+ * @returns The decompression margin or an error that can be checked with ZSTD_isError().
+ */
+ZSTDLIB_STATIC_API size_t ZSTD_decompressionMargin(const void* src, size_t srcSize);
+
+/*! ZSTD_DECOMPRESS_MARGIN() :
+ * Similar to ZSTD_decompressionMargin(), but instead of computing the margin from
+ * the compressed frame, compute it from the original size and the blockSizeLog.
+ * See ZSTD_decompressionMargin() for details.
+ *
+ * WARNING: This macro does not support multi-frame input, the input must be a single
+ * zstd frame. If you need that support use the function, or implement it yourself.
+ *
+ * @param originalSize The original uncompressed size of the data.
+ * @param blockSize    The block size == MIN(windowSize, ZSTD_BLOCKSIZE_MAX).
+ *                     Unless you explicitly set the windowLog smaller than
+ *                     ZSTD_BLOCKSIZELOG_MAX you can just use ZSTD_BLOCKSIZE_MAX.
+ */
+#define ZSTD_DECOMPRESSION_MARGIN(originalSize, blockSize) ((size_t)(                                              \
+        ZSTD_FRAMEHEADERSIZE_MAX                                                              /* Frame header */ + \
+        4                                                                                         /* checksum */ + \
+        ((originalSize) == 0 ? 0 : 3 * (((originalSize) + (blockSize) - 1) / blockSize)) /* 3 bytes per block */ + \
+        (blockSize)                                                                    /* One block of margin */   \
+    ))
+
+typedef enum {
+  ZSTD_sf_noBlockDelimiters = 0,         /* Representation of ZSTD_Sequence has no block delimiters, sequences only */
+  ZSTD_sf_explicitBlockDelimiters = 1    /* Representation of ZSTD_Sequence contains explicit block delimiters */
+} ZSTD_sequenceFormat_e;
+
+/*! ZSTD_sequenceBound() :
+ * `srcSize` : size of the input buffer
+ *  @return : upper-bound for the number of sequences that can be generated
+ *            from a buffer of srcSize bytes
+ *
+ *  note : returns number of sequences - to get bytes, multiply by sizeof(ZSTD_Sequence).
+ */
+ZSTDLIB_STATIC_API size_t ZSTD_sequenceBound(size_t srcSize);
+
+/*! ZSTD_generateSequences() :
+ * Generate sequences using ZSTD_compress2(), given a source buffer.
+ *
+ * Each block will end with a dummy sequence
+ * with offset == 0, matchLength == 0, and litLength == length of last literals.
+ * litLength may be == 0, and if so, then the sequence of (of: 0 ml: 0 ll: 0)
+ * simply acts as a block delimiter.
+ *
+ * @zc can be used to insert custom compression params.
+ * This function invokes ZSTD_compress2().
+ *
+ * The output of this function can be fed into ZSTD_compressSequences() with CCtx
+ * setting of ZSTD_c_blockDelimiters as ZSTD_sf_explicitBlockDelimiters
+ * @return : number of sequences generated
+ */
+
+ZSTDLIB_STATIC_API size_t
+ZSTD_generateSequences( ZSTD_CCtx* zc,
+                        ZSTD_Sequence* outSeqs, size_t outSeqsSize,
+                        const void* src, size_t srcSize);
+
+/*! ZSTD_mergeBlockDelimiters() :
+ * Given an array of ZSTD_Sequence, remove all sequences that represent block delimiters/last literals
+ * by merging them into the literals of the next sequence.
+ *
+ * As such, the final generated result has no explicit representation of block boundaries,
+ * and the final last literals segment is not represented in the sequences.
+ *
+ * The output of this function can be fed into ZSTD_compressSequences() with CCtx
+ * setting of ZSTD_c_blockDelimiters as ZSTD_sf_noBlockDelimiters
+ * @return : number of sequences left after merging
+ */
+ZSTDLIB_STATIC_API size_t ZSTD_mergeBlockDelimiters(ZSTD_Sequence* sequences, size_t seqsSize);
+
+/*! ZSTD_compressSequences() :
+ * Compress an array of ZSTD_Sequence, associated with @src buffer, into dst.
+ * @src contains the entire input (not just the literals).
+ * If @srcSize > sum(sequence.length), the remaining bytes are considered all literals
+ * If a dictionary is included, then the cctx should reference the dict. (see: ZSTD_CCtx_refCDict(), ZSTD_CCtx_loadDictionary(), etc.)
+ * The entire source is compressed into a single frame.
+ *
+ * The compression behavior changes based on cctx params. In particular:
+ *    If ZSTD_c_blockDelimiters == ZSTD_sf_noBlockDelimiters, the array of ZSTD_Sequence is expected to contain
+ *    no block delimiters (defined in ZSTD_Sequence). Block boundaries are roughly determined based on
+ *    the block size derived from the cctx, and sequences may be split. This is the default setting.
+ *
+ *    If ZSTD_c_blockDelimiters == ZSTD_sf_explicitBlockDelimiters, the array of ZSTD_Sequence is expected to contain
+ *    block delimiters (defined in ZSTD_Sequence). Behavior is undefined if no block delimiters are provided.
+ *
+ *    If ZSTD_c_validateSequences == 0, this function will blindly accept the sequences provided. Invalid sequences cause undefined
+ *    behavior. If ZSTD_c_validateSequences == 1, then if sequence is invalid (see doc/zstd_compression_format.md for
+ *    specifics regarding offset/matchlength requirements) then the function will bail out and return an error.
+ *
+ *    In addition to the two adjustable experimental params, there are other important cctx params.
+ *    - ZSTD_c_minMatch MUST be set as less than or equal to the smallest match generated by the match finder. It has a minimum value of ZSTD_MINMATCH_MIN.
+ *    - ZSTD_c_compressionLevel accordingly adjusts the strength of the entropy coder, as it would in typical compression.
+ *    - ZSTD_c_windowLog affects offset validation: this function will return an error at higher debug levels if a provided offset
+ *      is larger than what the spec allows for a given window log and dictionary (if present). See: doc/zstd_compression_format.md
+ *
+ * Note: Repcodes are, as of now, always re-calculated within this function, so ZSTD_Sequence::rep is unused.
+ * Note 2: Once we integrate ability to ingest repcodes, the explicit block delims mode must respect those repcodes exactly,
+ *         and cannot emit an RLE block that disagrees with the repcode history
+ * @return : final compressed size, or a ZSTD error code.
+ */
+ZSTDLIB_STATIC_API size_t
+ZSTD_compressSequences( ZSTD_CCtx* cctx, void* dst, size_t dstSize,
+                        const ZSTD_Sequence* inSeqs, size_t inSeqsSize,
+                        const void* src, size_t srcSize);
 
-/*! ZSTD_getSequences() :
- * Extract sequences from the sequence store
- * zc can be used to insert custom compression params.
- * This function invokes ZSTD_compress2
- * @return : number of sequences extracted
+
+/*! ZSTD_writeSkippableFrame() :
+ * Generates a zstd skippable frame containing data given by src, and writes it to dst buffer.
+ *
+ * Skippable frames begin with a 4-byte magic number. There are 16 possible choices of magic number,
+ * ranging from ZSTD_MAGIC_SKIPPABLE_START to ZSTD_MAGIC_SKIPPABLE_START+15.
+ * As such, the parameter magicVariant controls the exact skippable frame magic number variant used, so
+ * the magic number used will be ZSTD_MAGIC_SKIPPABLE_START + magicVariant.
+ *
+ * Returns an error if destination buffer is not large enough, if the source size is not representable
+ * with a 4-byte unsigned int, or if the parameter magicVariant is greater than 15 (and therefore invalid).
+ *
+ * @return : number of bytes written or a ZSTD error.
  */
-ZSTDLIB_API size_t ZSTD_getSequences(ZSTD_CCtx* zc, ZSTD_Sequence* outSeqs,
-    size_t outSeqsSize, const void* src, size_t srcSize);
+ZSTDLIB_STATIC_API size_t ZSTD_writeSkippableFrame(void* dst, size_t dstCapacity,
+                                            const void* src, size_t srcSize, unsigned magicVariant);
+
+/*! ZSTD_readSkippableFrame() :
+ * Retrieves a zstd skippable frame containing data given by src, and writes it to dst buffer.
+ *
+ * The parameter magicVariant will receive the magicVariant that was supplied when the frame was written,
+ * i.e. magicNumber - ZSTD_MAGIC_SKIPPABLE_START.  This can be NULL if the caller is not interested
+ * in the magicVariant.
+ *
+ * Returns an error if destination buffer is not large enough, or if the frame is not skippable.
+ *
+ * @return : number of bytes written or a ZSTD error.
+ */
+ZSTDLIB_API size_t ZSTD_readSkippableFrame(void* dst, size_t dstCapacity, unsigned* magicVariant,
+                                            const void* src, size_t srcSize);
+
+/*! ZSTD_isSkippableFrame() :
+ *  Tells if the content of `buffer` starts with a valid Frame Identifier for a skippable frame.
+ */
+ZSTDLIB_API unsigned ZSTD_isSkippableFrame(const void* buffer, size_t size);
+
 
 
 /***************************************
@@ -1287,13 +1624,16 @@ ZSTDLIB_API size_t ZSTD_getSequences(ZSTD_CCtx* zc, ZSTD_Sequence* outSeqs,
  *  and ZSTD_estimateCCtxSize_usingCCtxParams(), which can be used in tandem with ZSTD_CCtxParams_setParameter().
  *  Both can be used to estimate memory using custom compression parameters and arbitrary srcSize limits.
  *
- *  Note 2 : only single-threaded compression is supported.
+ *  Note : only single-threaded compression is supported.
  *  ZSTD_estimateCCtxSize_usingCCtxParams() will return an error code if ZSTD_c_nbWorkers is >= 1.
+ *
+ *  Note 2 : ZSTD_estimateCCtxSize* functions are not compatible with the Block-Level Sequence Producer API at this time.
+ *  Size estimates assume that no external sequence producer is registered.
  */
-ZSTDLIB_API size_t ZSTD_estimateCCtxSize(int compressionLevel);
-ZSTDLIB_API size_t ZSTD_estimateCCtxSize_usingCParams(ZSTD_compressionParameters cParams);
-ZSTDLIB_API size_t ZSTD_estimateCCtxSize_usingCCtxParams(const ZSTD_CCtx_params* params);
-ZSTDLIB_API size_t ZSTD_estimateDCtxSize(void);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateCCtxSize(int compressionLevel);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateCCtxSize_usingCParams(ZSTD_compressionParameters cParams);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateCCtxSize_usingCCtxParams(const ZSTD_CCtx_params* params);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateDCtxSize(void);
 
 /*! ZSTD_estimateCStreamSize() :
  *  ZSTD_estimateCStreamSize() will provide a budget large enough for any compression level up to selected one.
@@ -1307,21 +1647,26 @@ ZSTDLIB_API size_t ZSTD_estimateDCtxSize(void);
  *  or deducted from a valid frame Header, using ZSTD_estimateDStreamSize_fromFrame();
  *  Note : if streaming is init with function ZSTD_init?Stream_usingDict(),
  *         an internal ?Dict will be created, which additional size is not estimated here.
- *         In this case, get total size by adding ZSTD_estimate?DictSize */
-ZSTDLIB_API size_t ZSTD_estimateCStreamSize(int compressionLevel);
-ZSTDLIB_API size_t ZSTD_estimateCStreamSize_usingCParams(ZSTD_compressionParameters cParams);
-ZSTDLIB_API size_t ZSTD_estimateCStreamSize_usingCCtxParams(const ZSTD_CCtx_params* params);
-ZSTDLIB_API size_t ZSTD_estimateDStreamSize(size_t windowSize);
-ZSTDLIB_API size_t ZSTD_estimateDStreamSize_fromFrame(const void* src, size_t srcSize);
+ *         In this case, get total size by adding ZSTD_estimate?DictSize
+ *  Note 2 : only single-threaded compression is supported.
+ *  ZSTD_estimateCStreamSize_usingCCtxParams() will return an error code if ZSTD_c_nbWorkers is >= 1.
+ *  Note 3 : ZSTD_estimateCStreamSize* functions are not compatible with the Block-Level Sequence Producer API at this time.
+ *  Size estimates assume that no external sequence producer is registered.
+ */
+ZSTDLIB_STATIC_API size_t ZSTD_estimateCStreamSize(int compressionLevel);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateCStreamSize_usingCParams(ZSTD_compressionParameters cParams);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateCStreamSize_usingCCtxParams(const ZSTD_CCtx_params* params);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateDStreamSize(size_t windowSize);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateDStreamSize_fromFrame(const void* src, size_t srcSize);
 
 /*! ZSTD_estimate?DictSize() :
  *  ZSTD_estimateCDictSize() will bet that src size is relatively "small", and content is copied, like ZSTD_createCDict().
  *  ZSTD_estimateCDictSize_advanced() makes it possible to control compression parameters precisely, like ZSTD_createCDict_advanced().
  *  Note : dictionaries created by reference (`ZSTD_dlm_byRef`) are logically smaller.
  */
-ZSTDLIB_API size_t ZSTD_estimateCDictSize(size_t dictSize, int compressionLevel);
-ZSTDLIB_API size_t ZSTD_estimateCDictSize_advanced(size_t dictSize, ZSTD_compressionParameters cParams, ZSTD_dictLoadMethod_e dictLoadMethod);
-ZSTDLIB_API size_t ZSTD_estimateDDictSize(size_t dictSize, ZSTD_dictLoadMethod_e dictLoadMethod);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateCDictSize(size_t dictSize, int compressionLevel);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateCDictSize_advanced(size_t dictSize, ZSTD_compressionParameters cParams, ZSTD_dictLoadMethod_e dictLoadMethod);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateDDictSize(size_t dictSize, ZSTD_dictLoadMethod_e dictLoadMethod);
 
 /*! ZSTD_initStatic*() :
  *  Initialize an object using a pre-allocated fixed-size buffer.
@@ -1344,20 +1689,20 @@ ZSTDLIB_API size_t ZSTD_estimateDDictSize(size_t dictSize, ZSTD_dictLoadMethod_e
  *  Limitation 2 : static cctx currently not compatible with multi-threading.
  *  Limitation 3 : static dctx is incompatible with legacy support.
  */
-ZSTDLIB_API ZSTD_CCtx*    ZSTD_initStaticCCtx(void* workspace, size_t workspaceSize);
-ZSTDLIB_API ZSTD_CStream* ZSTD_initStaticCStream(void* workspace, size_t workspaceSize);    /**< same as ZSTD_initStaticCCtx() */
+ZSTDLIB_STATIC_API ZSTD_CCtx*    ZSTD_initStaticCCtx(void* workspace, size_t workspaceSize);
+ZSTDLIB_STATIC_API ZSTD_CStream* ZSTD_initStaticCStream(void* workspace, size_t workspaceSize);    /**< same as ZSTD_initStaticCCtx() */
 
-ZSTDLIB_API ZSTD_DCtx*    ZSTD_initStaticDCtx(void* workspace, size_t workspaceSize);
-ZSTDLIB_API ZSTD_DStream* ZSTD_initStaticDStream(void* workspace, size_t workspaceSize);    /**< same as ZSTD_initStaticDCtx() */
+ZSTDLIB_STATIC_API ZSTD_DCtx*    ZSTD_initStaticDCtx(void* workspace, size_t workspaceSize);
+ZSTDLIB_STATIC_API ZSTD_DStream* ZSTD_initStaticDStream(void* workspace, size_t workspaceSize);    /**< same as ZSTD_initStaticDCtx() */
 
-ZSTDLIB_API const ZSTD_CDict* ZSTD_initStaticCDict(
+ZSTDLIB_STATIC_API const ZSTD_CDict* ZSTD_initStaticCDict(
                                         void* workspace, size_t workspaceSize,
                                         const void* dict, size_t dictSize,
                                         ZSTD_dictLoadMethod_e dictLoadMethod,
                                         ZSTD_dictContentType_e dictContentType,
                                         ZSTD_compressionParameters cParams);
 
-ZSTDLIB_API const ZSTD_DDict* ZSTD_initStaticDDict(
+ZSTDLIB_STATIC_API const ZSTD_DDict* ZSTD_initStaticDDict(
                                         void* workspace, size_t workspaceSize,
                                         const void* dict, size_t dictSize,
                                         ZSTD_dictLoadMethod_e dictLoadMethod,
@@ -1372,25 +1717,55 @@ ZSTDLIB_API const ZSTD_DDict* ZSTD_initStaticDDict(
 typedef void* (*ZSTD_allocFunction) (void* opaque, size_t size);
 typedef void  (*ZSTD_freeFunction) (void* opaque, void* address);
 typedef struct { ZSTD_allocFunction customAlloc; ZSTD_freeFunction customFree; void* opaque; } ZSTD_customMem;
-static ZSTD_customMem const ZSTD_defaultCMem = { NULL, NULL, NULL };  /**< this constant defers to stdlib's functions */
+static
+#ifdef __GNUC__
+__attribute__((__unused__))
+#endif
+ZSTD_customMem const ZSTD_defaultCMem = { NULL, NULL, NULL };  /**< this constant defers to stdlib's functions */
 
-ZSTDLIB_API ZSTD_CCtx*    ZSTD_createCCtx_advanced(ZSTD_customMem customMem);
-ZSTDLIB_API ZSTD_CStream* ZSTD_createCStream_advanced(ZSTD_customMem customMem);
-ZSTDLIB_API ZSTD_DCtx*    ZSTD_createDCtx_advanced(ZSTD_customMem customMem);
-ZSTDLIB_API ZSTD_DStream* ZSTD_createDStream_advanced(ZSTD_customMem customMem);
+ZSTDLIB_STATIC_API ZSTD_CCtx*    ZSTD_createCCtx_advanced(ZSTD_customMem customMem);
+ZSTDLIB_STATIC_API ZSTD_CStream* ZSTD_createCStream_advanced(ZSTD_customMem customMem);
+ZSTDLIB_STATIC_API ZSTD_DCtx*    ZSTD_createDCtx_advanced(ZSTD_customMem customMem);
+ZSTDLIB_STATIC_API ZSTD_DStream* ZSTD_createDStream_advanced(ZSTD_customMem customMem);
 
-ZSTDLIB_API ZSTD_CDict* ZSTD_createCDict_advanced(const void* dict, size_t dictSize,
+ZSTDLIB_STATIC_API ZSTD_CDict* ZSTD_createCDict_advanced(const void* dict, size_t dictSize,
                                                   ZSTD_dictLoadMethod_e dictLoadMethod,
                                                   ZSTD_dictContentType_e dictContentType,
                                                   ZSTD_compressionParameters cParams,
                                                   ZSTD_customMem customMem);
 
-ZSTDLIB_API ZSTD_DDict* ZSTD_createDDict_advanced(const void* dict, size_t dictSize,
-                                                  ZSTD_dictLoadMethod_e dictLoadMethod,
-                                                  ZSTD_dictContentType_e dictContentType,
-                                                  ZSTD_customMem customMem);
+/*! Thread pool :
+ *  These prototypes make it possible to share a thread pool among multiple compression contexts.
+ *  This can limit resources for applications with multiple threads where each one uses
+ *  a threaded compression mode (via ZSTD_c_nbWorkers parameter).
+ *  ZSTD_createThreadPool creates a new thread pool with a given number of threads.
+ *  Note that the lifetime of such pool must exist while being used.
+ *  ZSTD_CCtx_refThreadPool assigns a thread pool to a context (use NULL argument value
+ *  to use an internal thread pool).
+ *  ZSTD_freeThreadPool frees a thread pool, accepts NULL pointer.
+ */
+typedef struct POOL_ctx_s ZSTD_threadPool;
+ZSTDLIB_STATIC_API ZSTD_threadPool* ZSTD_createThreadPool(size_t numThreads);
+ZSTDLIB_STATIC_API void ZSTD_freeThreadPool (ZSTD_threadPool* pool);  /* accept NULL pointer */
+ZSTDLIB_STATIC_API size_t ZSTD_CCtx_refThreadPool(ZSTD_CCtx* cctx, ZSTD_threadPool* pool);
 
 
+/*
+ * This API is temporary and is expected to change or disappear in the future!
+ */
+ZSTDLIB_STATIC_API ZSTD_CDict* ZSTD_createCDict_advanced2(
+    const void* dict, size_t dictSize,
+    ZSTD_dictLoadMethod_e dictLoadMethod,
+    ZSTD_dictContentType_e dictContentType,
+    const ZSTD_CCtx_params* cctxParams,
+    ZSTD_customMem customMem);
+
+ZSTDLIB_STATIC_API ZSTD_DDict* ZSTD_createDDict_advanced(
+    const void* dict, size_t dictSize,
+    ZSTD_dictLoadMethod_e dictLoadMethod,
+    ZSTD_dictContentType_e dictContentType,
+    ZSTD_customMem customMem);
+
 
 /***************************************
 *  Advanced compression functions
@@ -1402,22 +1777,22 @@ ZSTDLIB_API ZSTD_DDict* ZSTD_createDDict_advanced(const void* dict, size_t dictS
  *  As a consequence, `dictBuffer` **must** outlive CDict,
  *  and its content must remain unmodified throughout the lifetime of CDict.
  *  note: equivalent to ZSTD_createCDict_advanced(), with dictLoadMethod==ZSTD_dlm_byRef */
-ZSTDLIB_API ZSTD_CDict* ZSTD_createCDict_byReference(const void* dictBuffer, size_t dictSize, int compressionLevel);
+ZSTDLIB_STATIC_API ZSTD_CDict* ZSTD_createCDict_byReference(const void* dictBuffer, size_t dictSize, int compressionLevel);
 
 /*! ZSTD_getCParams() :
  * @return ZSTD_compressionParameters structure for a selected compression level and estimated srcSize.
  * `estimatedSrcSize` value is optional, select 0 if not known */
-ZSTDLIB_API ZSTD_compressionParameters ZSTD_getCParams(int compressionLevel, unsigned long long estimatedSrcSize, size_t dictSize);
+ZSTDLIB_STATIC_API ZSTD_compressionParameters ZSTD_getCParams(int compressionLevel, unsigned long long estimatedSrcSize, size_t dictSize);
 
 /*! ZSTD_getParams() :
  *  same as ZSTD_getCParams(), but @return a full `ZSTD_parameters` object instead of sub-component `ZSTD_compressionParameters`.
  *  All fields of `ZSTD_frameParameters` are set to default : contentSize=1, checksum=0, noDictID=0 */
-ZSTDLIB_API ZSTD_parameters ZSTD_getParams(int compressionLevel, unsigned long long estimatedSrcSize, size_t dictSize);
+ZSTDLIB_STATIC_API ZSTD_parameters ZSTD_getParams(int compressionLevel, unsigned long long estimatedSrcSize, size_t dictSize);
 
 /*! ZSTD_checkCParams() :
  *  Ensure param values remain within authorized range.
  * @return 0 on success, or an error code (can be checked with ZSTD_isError()) */
-ZSTDLIB_API size_t ZSTD_checkCParams(ZSTD_compressionParameters params);
+ZSTDLIB_STATIC_API size_t ZSTD_checkCParams(ZSTD_compressionParameters params);
 
 /*! ZSTD_adjustCParams() :
  *  optimize params for a given `srcSize` and `dictSize`.
@@ -1425,23 +1800,34 @@ ZSTDLIB_API size_t ZSTD_checkCParams(ZSTD_compressionParameters params);
  * `dictSize` must be `0` when there is no dictionary.
  *  cPar can be invalid : all parameters will be clamped within valid range in the @return struct.
  *  This function never fails (wide contract) */
-ZSTDLIB_API ZSTD_compressionParameters ZSTD_adjustCParams(ZSTD_compressionParameters cPar, unsigned long long srcSize, size_t dictSize);
+ZSTDLIB_STATIC_API ZSTD_compressionParameters ZSTD_adjustCParams(ZSTD_compressionParameters cPar, unsigned long long srcSize, size_t dictSize);
+
+/*! ZSTD_CCtx_setCParams() :
+ *  Set all parameters provided within @cparams into the working @cctx.
+ *  Note : if modifying parameters during compression (MT mode only),
+ *         note that changes to the .windowLog parameter will be ignored.
+ * @return 0 on success, or an error code (can be checked with ZSTD_isError()) */
+ZSTDLIB_STATIC_API size_t ZSTD_CCtx_setCParams(ZSTD_CCtx* cctx, ZSTD_compressionParameters cparams);
 
 /*! ZSTD_compress_advanced() :
  *  Note : this function is now DEPRECATED.
  *         It can be replaced by ZSTD_compress2(), in combination with ZSTD_CCtx_setParameter() and other parameter setters.
- *  This prototype will be marked as deprecated and generate compilation warning on reaching v1.5.x */
-ZSTDLIB_API size_t ZSTD_compress_advanced(ZSTD_CCtx* cctx,
-                                          void* dst, size_t dstCapacity,
-                                    const void* src, size_t srcSize,
-                                    const void* dict,size_t dictSize,
-                                          ZSTD_parameters params);
+ *  This prototype will generate compilation warnings. */
+ZSTD_DEPRECATED("use ZSTD_compress2")
+ZSTDLIB_STATIC_API
+size_t ZSTD_compress_advanced(ZSTD_CCtx* cctx,
+                              void* dst, size_t dstCapacity,
+                        const void* src, size_t srcSize,
+                        const void* dict,size_t dictSize,
+                              ZSTD_parameters params);
 
 /*! ZSTD_compress_usingCDict_advanced() :
- *  Note : this function is now REDUNDANT.
+ *  Note : this function is now DEPRECATED.
  *         It can be replaced by ZSTD_compress2(), in combination with ZSTD_CCtx_loadDictionary() and other parameter setters.
- *  This prototype will be marked as deprecated and generate compilation warning in some future version */
-ZSTDLIB_API size_t ZSTD_compress_usingCDict_advanced(ZSTD_CCtx* cctx,
+ *  This prototype will generate compilation warnings. */
+ZSTD_DEPRECATED("use ZSTD_compress2 with ZSTD_CCtx_loadDictionary")
+ZSTDLIB_STATIC_API
+size_t ZSTD_compress_usingCDict_advanced(ZSTD_CCtx* cctx,
                                               void* dst, size_t dstCapacity,
                                         const void* src, size_t srcSize,
                                         const ZSTD_CDict* cdict,
@@ -1451,18 +1837,18 @@ ZSTDLIB_API size_t ZSTD_compress_usingCDict_advanced(ZSTD_CCtx* cctx,
 /*! ZSTD_CCtx_loadDictionary_byReference() :
  *  Same as ZSTD_CCtx_loadDictionary(), but dictionary content is referenced, instead of being copied into CCtx.
  *  It saves some memory, but also requires that `dict` outlives its usage within `cctx` */
-ZSTDLIB_API size_t ZSTD_CCtx_loadDictionary_byReference(ZSTD_CCtx* cctx, const void* dict, size_t dictSize);
+ZSTDLIB_STATIC_API size_t ZSTD_CCtx_loadDictionary_byReference(ZSTD_CCtx* cctx, const void* dict, size_t dictSize);
 
 /*! ZSTD_CCtx_loadDictionary_advanced() :
  *  Same as ZSTD_CCtx_loadDictionary(), but gives finer control over
  *  how to load the dictionary (by copy ? by reference ?)
  *  and how to interpret it (automatic ? force raw mode ? full mode only ?) */
-ZSTDLIB_API size_t ZSTD_CCtx_loadDictionary_advanced(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, ZSTD_dictLoadMethod_e dictLoadMethod, ZSTD_dictContentType_e dictContentType);
+ZSTDLIB_STATIC_API size_t ZSTD_CCtx_loadDictionary_advanced(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, ZSTD_dictLoadMethod_e dictLoadMethod, ZSTD_dictContentType_e dictContentType);
 
 /*! ZSTD_CCtx_refPrefix_advanced() :
  *  Same as ZSTD_CCtx_refPrefix(), but gives finer control over
  *  how to interpret prefix content (automatic ? force raw mode (default) ? full mode only ?) */
-ZSTDLIB_API size_t ZSTD_CCtx_refPrefix_advanced(ZSTD_CCtx* cctx, const void* prefix, size_t prefixSize, ZSTD_dictContentType_e dictContentType);
+ZSTDLIB_STATIC_API size_t ZSTD_CCtx_refPrefix_advanced(ZSTD_CCtx* cctx, const void* prefix, size_t prefixSize, ZSTD_dictContentType_e dictContentType);
 
 /* ===   experimental parameters   === */
 /* these parameters can be used with ZSTD_setParameter()
@@ -1501,9 +1887,15 @@ ZSTDLIB_API size_t ZSTD_CCtx_refPrefix_advanced(ZSTD_CCtx* cctx, const void* pre
  * See the comments on that enum for an explanation of the feature. */
 #define ZSTD_c_forceAttachDict ZSTD_c_experimentalParam4
 
-/* Controls how the literals are compressed (default is auto).
- * The value must be of type ZSTD_literalCompressionMode_e.
- * See ZSTD_literalCompressionMode_t enum definition for details.
+/* Controlled with ZSTD_paramSwitch_e enum.
+ * Default is ZSTD_ps_auto.
+ * Set to ZSTD_ps_disable to never compress literals.
+ * Set to ZSTD_ps_enable to always compress literals. (Note: uncompressed literals
+ * may still be emitted if huffman is not beneficial to use.)
+ *
+ * By default, in ZSTD_ps_auto, the library will decide at runtime whether to use
+ * literals compression based on the compression parameters - specifically,
+ * negative compression levels do not use literal compression.
  */
 #define ZSTD_c_literalCompressionMode ZSTD_c_experimentalParam5
 
@@ -1518,12 +1910,265 @@ ZSTDLIB_API size_t ZSTD_CCtx_refPrefix_advanced(ZSTD_CCtx* cctx, const void* pre
  * but compression ratio may regress significantly if guess considerably underestimates */
 #define ZSTD_c_srcSizeHint ZSTD_c_experimentalParam7
 
+/* Controls whether the new and experimental "dedicated dictionary search
+ * structure" can be used. This feature is still rough around the edges, be
+ * prepared for surprising behavior!
+ *
+ * How to use it:
+ *
+ * When using a CDict, whether to use this feature or not is controlled at
+ * CDict creation, and it must be set in a CCtxParams set passed into that
+ * construction (via ZSTD_createCDict_advanced2()). A compression will then
+ * use the feature or not based on how the CDict was constructed; the value of
+ * this param, set in the CCtx, will have no effect.
+ *
+ * However, when a dictionary buffer is passed into a CCtx, such as via
+ * ZSTD_CCtx_loadDictionary(), this param can be set on the CCtx to control
+ * whether the CDict that is created internally can use the feature or not.
+ *
+ * What it does:
+ *
+ * Normally, the internal data structures of the CDict are analogous to what
+ * would be stored in a CCtx after compressing the contents of a dictionary.
+ * To an approximation, a compression using a dictionary can then use those
+ * data structures to simply continue what is effectively a streaming
+ * compression where the simulated compression of the dictionary left off.
+ * Which is to say, the search structures in the CDict are normally the same
+ * format as in the CCtx.
+ *
+ * It is possible to do better, since the CDict is not like a CCtx: the search
+ * structures are written once during CDict creation, and then are only read
+ * after that, while the search structures in the CCtx are both read and
+ * written as the compression goes along. This means we can choose a search
+ * structure for the dictionary that is read-optimized.
+ *
+ * This feature enables the use of that different structure.
+ *
+ * Note that some of the members of the ZSTD_compressionParameters struct have
+ * different semantics and constraints in the dedicated search structure. It is
+ * highly recommended that you simply set a compression level in the CCtxParams
+ * you pass into the CDict creation call, and avoid messing with the cParams
+ * directly.
+ *
+ * Effects:
+ *
+ * This will only have any effect when the selected ZSTD_strategy
+ * implementation supports this feature. Currently, that's limited to
+ * ZSTD_greedy, ZSTD_lazy, and ZSTD_lazy2.
+ *
+ * Note that this means that the CDict tables can no longer be copied into the
+ * CCtx, so the dict attachment mode ZSTD_dictForceCopy will no longer be
+ * usable. The dictionary can only be attached or reloaded.
+ *
+ * In general, you should expect compression to be faster--sometimes very much
+ * so--and CDict creation to be slightly slower. Eventually, we will probably
+ * make this mode the default.
+ */
+#define ZSTD_c_enableDedicatedDictSearch ZSTD_c_experimentalParam8
+
+/* ZSTD_c_stableInBuffer
+ * Experimental parameter.
+ * Default is 0 == disabled. Set to 1 to enable.
+ *
+ * Tells the compressor that input data presented with ZSTD_inBuffer
+ * will ALWAYS be the same between calls.
+ * Technically, the @src pointer must never be changed,
+ * and the @pos field can only be updated by zstd.
+ * However, it's possible to increase the @size field,
+ * allowing scenarios where more data can be appended after compressions starts.
+ * These conditions are checked by the compressor,
+ * and compression will fail if they are not respected.
+ * Also, data in the ZSTD_inBuffer within the range [src, src + pos)
+ * MUST not be modified during compression or it will result in data corruption.
+ *
+ * When this flag is enabled zstd won't allocate an input window buffer,
+ * because the user guarantees it can reference the ZSTD_inBuffer until
+ * the frame is complete. But, it will still allocate an output buffer
+ * large enough to fit a block (see ZSTD_c_stableOutBuffer). This will also
+ * avoid the memcpy() from the input buffer to the input window buffer.
+ *
+ * NOTE: So long as the ZSTD_inBuffer always points to valid memory, using
+ * this flag is ALWAYS memory safe, and will never access out-of-bounds
+ * memory. However, compression WILL fail if conditions are not respected.
+ *
+ * WARNING: The data in the ZSTD_inBuffer in the range [src, src + pos) MUST
+ * not be modified during compression or it will result in data corruption.
+ * This is because zstd needs to reference data in the ZSTD_inBuffer to find
+ * matches. Normally zstd maintains its own window buffer for this purpose,
+ * but passing this flag tells zstd to rely on user provided buffer instead.
+ */
+#define ZSTD_c_stableInBuffer ZSTD_c_experimentalParam9
+
+/* ZSTD_c_stableOutBuffer
+ * Experimental parameter.
+ * Default is 0 == disabled. Set to 1 to enable.
+ *
+ * Tells he compressor that the ZSTD_outBuffer will not be resized between
+ * calls. Specifically: (out.size - out.pos) will never grow. This gives the
+ * compressor the freedom to say: If the compressed data doesn't fit in the
+ * output buffer then return ZSTD_error_dstSizeTooSmall. This allows us to
+ * always decompress directly into the output buffer, instead of decompressing
+ * into an internal buffer and copying to the output buffer.
+ *
+ * When this flag is enabled zstd won't allocate an output buffer, because
+ * it can write directly to the ZSTD_outBuffer. It will still allocate the
+ * input window buffer (see ZSTD_c_stableInBuffer).
+ *
+ * Zstd will check that (out.size - out.pos) never grows and return an error
+ * if it does. While not strictly necessary, this should prevent surprises.
+ */
+#define ZSTD_c_stableOutBuffer ZSTD_c_experimentalParam10
+
+/* ZSTD_c_blockDelimiters
+ * Default is 0 == ZSTD_sf_noBlockDelimiters.
+ *
+ * For use with sequence compression API: ZSTD_compressSequences().
+ *
+ * Designates whether or not the given array of ZSTD_Sequence contains block delimiters
+ * and last literals, which are defined as sequences with offset == 0 and matchLength == 0.
+ * See the definition of ZSTD_Sequence for more specifics.
+ */
+#define ZSTD_c_blockDelimiters ZSTD_c_experimentalParam11
+
+/* ZSTD_c_validateSequences
+ * Default is 0 == disabled. Set to 1 to enable sequence validation.
+ *
+ * For use with sequence compression API: ZSTD_compressSequences().
+ * Designates whether or not we validate sequences provided to ZSTD_compressSequences()
+ * during function execution.
+ *
+ * Without validation, providing a sequence that does not conform to the zstd spec will cause
+ * undefined behavior, and may produce a corrupted block.
+ *
+ * With validation enabled, if sequence is invalid (see doc/zstd_compression_format.md for
+ * specifics regarding offset/matchlength requirements) then the function will bail out and
+ * return an error.
+ *
+ */
+#define ZSTD_c_validateSequences ZSTD_c_experimentalParam12
+
+/* ZSTD_c_useBlockSplitter
+ * Controlled with ZSTD_paramSwitch_e enum.
+ * Default is ZSTD_ps_auto.
+ * Set to ZSTD_ps_disable to never use block splitter.
+ * Set to ZSTD_ps_enable to always use block splitter.
+ *
+ * By default, in ZSTD_ps_auto, the library will decide at runtime whether to use
+ * block splitting based on the compression parameters.
+ */
+#define ZSTD_c_useBlockSplitter ZSTD_c_experimentalParam13
+
+/* ZSTD_c_useRowMatchFinder
+ * Controlled with ZSTD_paramSwitch_e enum.
+ * Default is ZSTD_ps_auto.
+ * Set to ZSTD_ps_disable to never use row-based matchfinder.
+ * Set to ZSTD_ps_enable to force usage of row-based matchfinder.
+ *
+ * By default, in ZSTD_ps_auto, the library will decide at runtime whether to use
+ * the row-based matchfinder based on support for SIMD instructions and the window log.
+ * Note that this only pertains to compression strategies: greedy, lazy, and lazy2
+ */
+#define ZSTD_c_useRowMatchFinder ZSTD_c_experimentalParam14
+
+/* ZSTD_c_deterministicRefPrefix
+ * Default is 0 == disabled. Set to 1 to enable.
+ *
+ * Zstd produces different results for prefix compression when the prefix is
+ * directly adjacent to the data about to be compressed vs. when it isn't.
+ * This is because zstd detects that the two buffers are contiguous and it can
+ * use a more efficient match finding algorithm. However, this produces different
+ * results than when the two buffers are non-contiguous. This flag forces zstd
+ * to always load the prefix in non-contiguous mode, even if it happens to be
+ * adjacent to the data, to guarantee determinism.
+ *
+ * If you really care about determinism when using a dictionary or prefix,
+ * like when doing delta compression, you should select this option. It comes
+ * at a speed penalty of about ~2.5% if the dictionary and data happened to be
+ * contiguous, and is free if they weren't contiguous. We don't expect that
+ * intentionally making the dictionary and data contiguous will be worth the
+ * cost to memcpy() the data.
+ */
+#define ZSTD_c_deterministicRefPrefix ZSTD_c_experimentalParam15
+
+/* ZSTD_c_prefetchCDictTables
+ * Controlled with ZSTD_paramSwitch_e enum. Default is ZSTD_ps_auto.
+ *
+ * In some situations, zstd uses CDict tables in-place rather than copying them
+ * into the working context. (See docs on ZSTD_dictAttachPref_e above for details).
+ * In such situations, compression speed is seriously impacted when CDict tables are
+ * "cold" (outside CPU cache). This parameter instructs zstd to prefetch CDict tables
+ * when they are used in-place.
+ *
+ * For sufficiently small inputs, the cost of the prefetch will outweigh the benefit.
+ * For sufficiently large inputs, zstd will by default memcpy() CDict tables
+ * into the working context, so there is no need to prefetch. This parameter is
+ * targeted at a middle range of input sizes, where a prefetch is cheap enough to be
+ * useful but memcpy() is too expensive. The exact range of input sizes where this
+ * makes sense is best determined by careful experimentation.
+ *
+ * Note: for this parameter, ZSTD_ps_auto is currently equivalent to ZSTD_ps_disable,
+ * but in the future zstd may conditionally enable this feature via an auto-detection
+ * heuristic for cold CDicts.
+ * Use ZSTD_ps_disable to opt out of prefetching under any circumstances.
+ */
+#define ZSTD_c_prefetchCDictTables ZSTD_c_experimentalParam16
+
+/* ZSTD_c_enableSeqProducerFallback
+ * Allowed values are 0 (disable) and 1 (enable). The default setting is 0.
+ *
+ * Controls whether zstd will fall back to an internal sequence producer if an
+ * external sequence producer is registered and returns an error code. This fallback
+ * is block-by-block: the internal sequence producer will only be called for blocks
+ * where the external sequence producer returns an error code. Fallback parsing will
+ * follow any other cParam settings, such as compression level, the same as in a
+ * normal (fully-internal) compression operation.
+ *
+ * The user is strongly encouraged to read the full Block-Level Sequence Producer API
+ * documentation (below) before setting this parameter. */
+#define ZSTD_c_enableSeqProducerFallback ZSTD_c_experimentalParam17
+
+/* ZSTD_c_maxBlockSize
+ * Allowed values are between 1KB and ZSTD_BLOCKSIZE_MAX (128KB).
+ * The default is ZSTD_BLOCKSIZE_MAX, and setting to 0 will set to the default.
+ *
+ * This parameter can be used to set an upper bound on the blocksize
+ * that overrides the default ZSTD_BLOCKSIZE_MAX. It cannot be used to set upper
+ * bounds greater than ZSTD_BLOCKSIZE_MAX or bounds lower than 1KB (will make
+ * compressBound() innacurate). Only currently meant to be used for testing.
+ *
+ */
+#define ZSTD_c_maxBlockSize ZSTD_c_experimentalParam18
+
+/* ZSTD_c_searchForExternalRepcodes
+ * This parameter affects how zstd parses external sequences, such as sequences
+ * provided through the compressSequences() API or from an external block-level
+ * sequence producer.
+ *
+ * If set to ZSTD_ps_enable, the library will check for repeated offsets in
+ * external sequences, even if those repcodes are not explicitly indicated in
+ * the "rep" field. Note that this is the only way to exploit repcode matches
+ * while using compressSequences() or an external sequence producer, since zstd
+ * currently ignores the "rep" field of external sequences.
+ *
+ * If set to ZSTD_ps_disable, the library will not exploit repeated offsets in
+ * external sequences, regardless of whether the "rep" field has been set. This
+ * reduces sequence compression overhead by about 25% while sacrificing some
+ * compression ratio.
+ *
+ * The default value is ZSTD_ps_auto, for which the library will enable/disable
+ * based on compression level.
+ *
+ * Note: for now, this param only has an effect if ZSTD_c_blockDelimiters is
+ * set to ZSTD_sf_explicitBlockDelimiters. That may change in the future.
+ */
+#define ZSTD_c_searchForExternalRepcodes ZSTD_c_experimentalParam19
+
 /*! ZSTD_CCtx_getParameter() :
  *  Get the requested compression parameter value, selected by enum ZSTD_cParameter,
  *  and store it into int* value.
  * @return : 0, or an error code (which can be tested with ZSTD_isError()).
  */
-ZSTDLIB_API size_t ZSTD_CCtx_getParameter(ZSTD_CCtx* cctx, ZSTD_cParameter param, int* value);
+ZSTDLIB_STATIC_API size_t ZSTD_CCtx_getParameter(const ZSTD_CCtx* cctx, ZSTD_cParameter param, int* value);
 
 
 /*! ZSTD_CCtx_params :
@@ -1538,45 +2183,47 @@ ZSTDLIB_API size_t ZSTD_CCtx_getParameter(ZSTD_CCtx* cctx, ZSTD_cParameter param
  *                                    These parameters will be applied to
  *                                    all subsequent frames.
  *  - ZSTD_compressStream2() : Do compression using the CCtx.
- *  - ZSTD_freeCCtxParams() : Free the memory.
+ *  - ZSTD_freeCCtxParams() : Free the memory, accept NULL pointer.
  *
  *  This can be used with ZSTD_estimateCCtxSize_advanced_usingCCtxParams()
  *  for static allocation of CCtx for single-threaded compression.
  */
-ZSTDLIB_API ZSTD_CCtx_params* ZSTD_createCCtxParams(void);
-ZSTDLIB_API size_t ZSTD_freeCCtxParams(ZSTD_CCtx_params* params);
+ZSTDLIB_STATIC_API ZSTD_CCtx_params* ZSTD_createCCtxParams(void);
+ZSTDLIB_STATIC_API size_t ZSTD_freeCCtxParams(ZSTD_CCtx_params* params);  /* accept NULL pointer */
 
 /*! ZSTD_CCtxParams_reset() :
  *  Reset params to default values.
  */
-ZSTDLIB_API size_t ZSTD_CCtxParams_reset(ZSTD_CCtx_params* params);
+ZSTDLIB_STATIC_API size_t ZSTD_CCtxParams_reset(ZSTD_CCtx_params* params);
 
 /*! ZSTD_CCtxParams_init() :
  *  Initializes the compression parameters of cctxParams according to
  *  compression level. All other parameters are reset to their default values.
  */
-ZSTDLIB_API size_t ZSTD_CCtxParams_init(ZSTD_CCtx_params* cctxParams, int compressionLevel);
+ZSTDLIB_STATIC_API size_t ZSTD_CCtxParams_init(ZSTD_CCtx_params* cctxParams, int compressionLevel);
 
 /*! ZSTD_CCtxParams_init_advanced() :
  *  Initializes the compression and frame parameters of cctxParams according to
  *  params. All other parameters are reset to their default values.
  */
-ZSTDLIB_API size_t ZSTD_CCtxParams_init_advanced(ZSTD_CCtx_params* cctxParams, ZSTD_parameters params);
+ZSTDLIB_STATIC_API size_t ZSTD_CCtxParams_init_advanced(ZSTD_CCtx_params* cctxParams, ZSTD_parameters params);
 
-/*! ZSTD_CCtxParams_setParameter() :
+/*! ZSTD_CCtxParams_setParameter() : Requires v1.4.0+
  *  Similar to ZSTD_CCtx_setParameter.
  *  Set one compression parameter, selected by enum ZSTD_cParameter.
- *  Parameters must be applied to a ZSTD_CCtx using ZSTD_CCtx_setParametersUsingCCtxParams().
- * @result : 0, or an error code (which can be tested with ZSTD_isError()).
+ *  Parameters must be applied to a ZSTD_CCtx using
+ *  ZSTD_CCtx_setParametersUsingCCtxParams().
+ * @result : a code representing success or failure (which can be tested with
+ *           ZSTD_isError()).
  */
-ZSTDLIB_API size_t ZSTD_CCtxParams_setParameter(ZSTD_CCtx_params* params, ZSTD_cParameter param, int value);
+ZSTDLIB_STATIC_API size_t ZSTD_CCtxParams_setParameter(ZSTD_CCtx_params* params, ZSTD_cParameter param, int value);
 
 /*! ZSTD_CCtxParams_getParameter() :
  * Similar to ZSTD_CCtx_getParameter.
  * Get the requested value of one compression parameter, selected by enum ZSTD_cParameter.
  * @result : 0, or an error code (which can be tested with ZSTD_isError()).
  */
-ZSTDLIB_API size_t ZSTD_CCtxParams_getParameter(ZSTD_CCtx_params* params, ZSTD_cParameter param, int* value);
+ZSTDLIB_STATIC_API size_t ZSTD_CCtxParams_getParameter(const ZSTD_CCtx_params* params, ZSTD_cParameter param, int* value);
 
 /*! ZSTD_CCtx_setParametersUsingCCtxParams() :
  *  Apply a set of ZSTD_CCtx_params to the compression context.
@@ -1585,7 +2232,7 @@ ZSTDLIB_API size_t ZSTD_CCtxParams_getParameter(ZSTD_CCtx_params* params, ZSTD_c
  *    if nbWorkers>=1, new parameters will be picked up at next job,
  *       with a few restrictions (windowLog, pledgedSrcSize, nbWorkers, jobSize, and overlapLog are not updated).
  */
-ZSTDLIB_API size_t ZSTD_CCtx_setParametersUsingCCtxParams(
+ZSTDLIB_STATIC_API size_t ZSTD_CCtx_setParametersUsingCCtxParams(
         ZSTD_CCtx* cctx, const ZSTD_CCtx_params* params);
 
 /*! ZSTD_compressStream2_simpleArgs() :
@@ -1594,7 +2241,7 @@ ZSTDLIB_API size_t ZSTD_CCtx_setParametersUsingCCtxParams(
  *  This variant might be helpful for binders from dynamic languages
  *  which have troubles handling structures containing memory pointers.
  */
-ZSTDLIB_API size_t ZSTD_compressStream2_simpleArgs (
+ZSTDLIB_STATIC_API size_t ZSTD_compressStream2_simpleArgs (
                             ZSTD_CCtx* cctx,
                             void* dst, size_t dstCapacity, size_t* dstPos,
                       const void* src, size_t srcSize, size_t* srcPos,
@@ -1610,33 +2257,33 @@ ZSTDLIB_API size_t ZSTD_compressStream2_simpleArgs (
  *  Note : Frame Identifier is 4 bytes. If `size < 4`, @return will always be 0.
  *  Note 2 : Legacy Frame Identifiers are considered valid only if Legacy Support is enabled.
  *  Note 3 : Skippable Frame Identifiers are considered valid. */
-ZSTDLIB_API unsigned ZSTD_isFrame(const void* buffer, size_t size);
+ZSTDLIB_STATIC_API unsigned ZSTD_isFrame(const void* buffer, size_t size);
 
 /*! ZSTD_createDDict_byReference() :
  *  Create a digested dictionary, ready to start decompression operation without startup delay.
  *  Dictionary content is referenced, and therefore stays in dictBuffer.
  *  It is important that dictBuffer outlives DDict,
  *  it must remain read accessible throughout the lifetime of DDict */
-ZSTDLIB_API ZSTD_DDict* ZSTD_createDDict_byReference(const void* dictBuffer, size_t dictSize);
+ZSTDLIB_STATIC_API ZSTD_DDict* ZSTD_createDDict_byReference(const void* dictBuffer, size_t dictSize);
 
 /*! ZSTD_DCtx_loadDictionary_byReference() :
  *  Same as ZSTD_DCtx_loadDictionary(),
  *  but references `dict` content instead of copying it into `dctx`.
  *  This saves memory if `dict` remains around.,
  *  However, it's imperative that `dict` remains accessible (and unmodified) while being used, so it must outlive decompression. */
-ZSTDLIB_API size_t ZSTD_DCtx_loadDictionary_byReference(ZSTD_DCtx* dctx, const void* dict, size_t dictSize);
+ZSTDLIB_STATIC_API size_t ZSTD_DCtx_loadDictionary_byReference(ZSTD_DCtx* dctx, const void* dict, size_t dictSize);
 
 /*! ZSTD_DCtx_loadDictionary_advanced() :
  *  Same as ZSTD_DCtx_loadDictionary(),
  *  but gives direct control over
  *  how to load the dictionary (by copy ? by reference ?)
  *  and how to interpret it (automatic ? force raw mode ? full mode only ?). */
-ZSTDLIB_API size_t ZSTD_DCtx_loadDictionary_advanced(ZSTD_DCtx* dctx, const void* dict, size_t dictSize, ZSTD_dictLoadMethod_e dictLoadMethod, ZSTD_dictContentType_e dictContentType);
+ZSTDLIB_STATIC_API size_t ZSTD_DCtx_loadDictionary_advanced(ZSTD_DCtx* dctx, const void* dict, size_t dictSize, ZSTD_dictLoadMethod_e dictLoadMethod, ZSTD_dictContentType_e dictContentType);
 
 /*! ZSTD_DCtx_refPrefix_advanced() :
  *  Same as ZSTD_DCtx_refPrefix(), but gives finer control over
  *  how to interpret prefix content (automatic ? force raw mode (default) ? full mode only ?) */
-ZSTDLIB_API size_t ZSTD_DCtx_refPrefix_advanced(ZSTD_DCtx* dctx, const void* prefix, size_t prefixSize, ZSTD_dictContentType_e dictContentType);
+ZSTDLIB_STATIC_API size_t ZSTD_DCtx_refPrefix_advanced(ZSTD_DCtx* dctx, const void* prefix, size_t prefixSize, ZSTD_dictContentType_e dictContentType);
 
 /*! ZSTD_DCtx_setMaxWindowSize() :
  *  Refuses allocating internal buffers for frames requiring a window size larger than provided limit.
@@ -1645,7 +2292,14 @@ ZSTDLIB_API size_t ZSTD_DCtx_refPrefix_advanced(ZSTD_DCtx* dctx, const void* pre
  *  By default, a decompression context accepts all window sizes <= (1 << ZSTD_WINDOWLOG_LIMIT_DEFAULT)
  * @return : 0, or an error code (which can be tested using ZSTD_isError()).
  */
-ZSTDLIB_API size_t ZSTD_DCtx_setMaxWindowSize(ZSTD_DCtx* dctx, size_t maxWindowSize);
+ZSTDLIB_STATIC_API size_t ZSTD_DCtx_setMaxWindowSize(ZSTD_DCtx* dctx, size_t maxWindowSize);
+
+/*! ZSTD_DCtx_getParameter() :
+ *  Get the requested decompression parameter value, selected by enum ZSTD_dParameter,
+ *  and store it into int* value.
+ * @return : 0, or an error code (which can be tested with ZSTD_isError()).
+ */
+ZSTDLIB_STATIC_API size_t ZSTD_DCtx_getParameter(ZSTD_DCtx* dctx, ZSTD_dParameter param, int* value);
 
 /* ZSTD_d_format
  * experimental parameter,
@@ -1665,7 +2319,7 @@ ZSTDLIB_API size_t ZSTD_DCtx_setMaxWindowSize(ZSTD_DCtx* dctx, size_t maxWindowS
  * in the range [dst, dst + pos) MUST not be modified during decompression
  * or you will get data corruption.
  *
- * When this flags is enabled zstd won't allocate an output buffer, because
+ * When this flag is enabled zstd won't allocate an output buffer, because
  * it can write directly to the ZSTD_outBuffer, but it will still allocate
  * an input buffer large enough to fit any compressed block. This will also
  * avoid the memcpy() from the internal output buffer to the ZSTD_outBuffer.
@@ -1684,12 +2338,61 @@ ZSTDLIB_API size_t ZSTD_DCtx_setMaxWindowSize(ZSTD_DCtx* dctx, size_t maxWindowS
  */
 #define ZSTD_d_stableOutBuffer ZSTD_d_experimentalParam2
 
+/* ZSTD_d_forceIgnoreChecksum
+ * Experimental parameter.
+ * Default is 0 == disabled. Set to 1 to enable
+ *
+ * Tells the decompressor to skip checksum validation during decompression, regardless
+ * of whether checksumming was specified during compression. This offers some
+ * slight performance benefits, and may be useful for debugging.
+ * Param has values of type ZSTD_forceIgnoreChecksum_e
+ */
+#define ZSTD_d_forceIgnoreChecksum ZSTD_d_experimentalParam3
+
+/* ZSTD_d_refMultipleDDicts
+ * Experimental parameter.
+ * Default is 0 == disabled. Set to 1 to enable
+ *
+ * If enabled and dctx is allocated on the heap, then additional memory will be allocated
+ * to store references to multiple ZSTD_DDict. That is, multiple calls of ZSTD_refDDict()
+ * using a given ZSTD_DCtx, rather than overwriting the previous DDict reference, will instead
+ * store all references. At decompression time, the appropriate dictID is selected
+ * from the set of DDicts based on the dictID in the frame.
+ *
+ * Usage is simply calling ZSTD_refDDict() on multiple dict buffers.
+ *
+ * Param has values of byte ZSTD_refMultipleDDicts_e
+ *
+ * WARNING: Enabling this parameter and calling ZSTD_DCtx_refDDict(), will trigger memory
+ * allocation for the hash table. ZSTD_freeDCtx() also frees this memory.
+ * Memory is allocated as per ZSTD_DCtx::customMem.
+ *
+ * Although this function allocates memory for the table, the user is still responsible for
+ * memory management of the underlying ZSTD_DDict* themselves.
+ */
+#define ZSTD_d_refMultipleDDicts ZSTD_d_experimentalParam4
+
+/* ZSTD_d_disableHuffmanAssembly
+ * Set to 1 to disable the Huffman assembly implementation.
+ * The default value is 0, which allows zstd to use the Huffman assembly
+ * implementation if available.
+ *
+ * This parameter can be used to disable Huffman assembly at runtime.
+ * If you want to disable it at compile time you can define the macro
+ * ZSTD_DISABLE_ASM.
+ */
+#define ZSTD_d_disableHuffmanAssembly ZSTD_d_experimentalParam5
+
+
 /*! ZSTD_DCtx_setFormat() :
+ *  This function is REDUNDANT. Prefer ZSTD_DCtx_setParameter().
  *  Instruct the decoder context about what kind of data to decode next.
  *  This instruction is mandatory to decode data without a fully-formed header,
  *  such ZSTD_f_zstd1_magicless for example.
  * @return : 0, or an error code (which can be tested using ZSTD_isError()). */
-ZSTDLIB_API size_t ZSTD_DCtx_setFormat(ZSTD_DCtx* dctx, ZSTD_format_e format);
+ZSTD_DEPRECATED("use ZSTD_DCtx_setParameter() instead")
+ZSTDLIB_STATIC_API
+size_t ZSTD_DCtx_setFormat(ZSTD_DCtx* dctx, ZSTD_format_e format);
 
 /*! ZSTD_decompressStream_simpleArgs() :
  *  Same as ZSTD_decompressStream(),
@@ -1697,7 +2400,7 @@ ZSTDLIB_API size_t ZSTD_DCtx_setFormat(ZSTD_DCtx* dctx, ZSTD_format_e format);
  *  This can be helpful for binders from dynamic languages
  *  which have troubles handling structures containing memory pointers.
  */
-ZSTDLIB_API size_t ZSTD_decompressStream_simpleArgs (
+ZSTDLIB_STATIC_API size_t ZSTD_decompressStream_simpleArgs (
                             ZSTD_DCtx* dctx,
                             void* dst, size_t dstCapacity, size_t* dstPos,
                       const void* src, size_t srcSize, size_t* srcPos);
@@ -1711,8 +2414,9 @@ ZSTDLIB_API size_t ZSTD_decompressStream_simpleArgs (
 ********************************************************************/
 
 /*=====   Advanced Streaming compression functions  =====*/
-/**! ZSTD_initCStream_srcSize() :
- * This function is deprecated, and equivalent to:
+
+/*! ZSTD_initCStream_srcSize() :
+ * This function is DEPRECATED, and equivalent to:
  *     ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only);
  *     ZSTD_CCtx_refCDict(zcs, NULL); // clear the dictionary (if any)
  *     ZSTD_CCtx_setParameter(zcs, ZSTD_c_compressionLevel, compressionLevel);
@@ -1721,15 +2425,16 @@ ZSTDLIB_API size_t ZSTD_decompressStream_simpleArgs (
  * pledgedSrcSize must be correct. If it is not known at init time, use
  * ZSTD_CONTENTSIZE_UNKNOWN. Note that, for compatibility with older programs,
  * "0" also disables frame content size field. It may be enabled in the future.
- * Note : this prototype will be marked as deprecated and generate compilation warnings on reaching v1.5.x
+ * This prototype will generate compilation warnings.
  */
-ZSTDLIB_API size_t
-ZSTD_initCStream_srcSize(ZSTD_CStream* zcs,
+ZSTD_DEPRECATED("use ZSTD_CCtx_reset, see zstd.h for detailed instructions")
+ZSTDLIB_STATIC_API
+size_t ZSTD_initCStream_srcSize(ZSTD_CStream* zcs,
                          int compressionLevel,
                          unsigned long long pledgedSrcSize);
 
-/**! ZSTD_initCStream_usingDict() :
- * This function is deprecated, and is equivalent to:
+/*! ZSTD_initCStream_usingDict() :
+ * This function is DEPRECATED, and is equivalent to:
  *     ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only);
  *     ZSTD_CCtx_setParameter(zcs, ZSTD_c_compressionLevel, compressionLevel);
  *     ZSTD_CCtx_loadDictionary(zcs, dict, dictSize);
@@ -1738,15 +2443,16 @@ ZSTD_initCStream_srcSize(ZSTD_CStream* zcs,
  * dict == NULL or dictSize < 8, in which case no dict is used.
  * Note: dict is loaded with ZSTD_dct_auto (treated as a full zstd dictionary if
  * it begins with ZSTD_MAGIC_DICTIONARY, else as raw content) and ZSTD_dlm_byCopy.
- * Note : this prototype will be marked as deprecated and generate compilation warnings on reaching v1.5.x
+ * This prototype will generate compilation warnings.
  */
-ZSTDLIB_API size_t
-ZSTD_initCStream_usingDict(ZSTD_CStream* zcs,
+ZSTD_DEPRECATED("use ZSTD_CCtx_reset, see zstd.h for detailed instructions")
+ZSTDLIB_STATIC_API
+size_t ZSTD_initCStream_usingDict(ZSTD_CStream* zcs,
                      const void* dict, size_t dictSize,
                            int compressionLevel);
 
-/**! ZSTD_initCStream_advanced() :
- * This function is deprecated, and is approximately equivalent to:
+/*! ZSTD_initCStream_advanced() :
+ * This function is DEPRECATED, and is approximately equivalent to:
  *     ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only);
  *     // Pseudocode: Set each zstd parameter and leave the rest as-is.
  *     for ((param, value) : params) {
@@ -1758,25 +2464,28 @@ ZSTD_initCStream_usingDict(ZSTD_CStream* zcs,
  * dict is loaded with ZSTD_dct_auto and ZSTD_dlm_byCopy.
  * pledgedSrcSize must be correct.
  * If srcSize is not known at init time, use value ZSTD_CONTENTSIZE_UNKNOWN.
- * Note : this prototype will be marked as deprecated and generate compilation warnings on reaching v1.5.x
+ * This prototype will generate compilation warnings.
  */
-ZSTDLIB_API size_t
-ZSTD_initCStream_advanced(ZSTD_CStream* zcs,
+ZSTD_DEPRECATED("use ZSTD_CCtx_reset, see zstd.h for detailed instructions")
+ZSTDLIB_STATIC_API
+size_t ZSTD_initCStream_advanced(ZSTD_CStream* zcs,
                     const void* dict, size_t dictSize,
                           ZSTD_parameters params,
                           unsigned long long pledgedSrcSize);
 
-/**! ZSTD_initCStream_usingCDict() :
- * This function is deprecated, and equivalent to:
+/*! ZSTD_initCStream_usingCDict() :
+ * This function is DEPRECATED, and equivalent to:
  *     ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only);
  *     ZSTD_CCtx_refCDict(zcs, cdict);
  *
  * note : cdict will just be referenced, and must outlive compression session
- * Note : this prototype will be marked as deprecated and generate compilation warnings on reaching v1.5.x
+ * This prototype will generate compilation warnings.
  */
-ZSTDLIB_API size_t ZSTD_initCStream_usingCDict(ZSTD_CStream* zcs, const ZSTD_CDict* cdict);
+ZSTD_DEPRECATED("use ZSTD_CCtx_reset and ZSTD_CCtx_refCDict, see zstd.h for detailed instructions")
+ZSTDLIB_STATIC_API
+size_t ZSTD_initCStream_usingCDict(ZSTD_CStream* zcs, const ZSTD_CDict* cdict);
 
-/**! ZSTD_initCStream_usingCDict_advanced() :
+/*! ZSTD_initCStream_usingCDict_advanced() :
  *   This function is DEPRECATED, and is approximately equivalent to:
  *     ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only);
  *     // Pseudocode: Set each zstd frame parameter and leave the rest as-is.
@@ -1789,18 +2498,22 @@ ZSTDLIB_API size_t ZSTD_initCStream_usingCDict(ZSTD_CStream* zcs, const ZSTD_CDi
  * same as ZSTD_initCStream_usingCDict(), with control over frame parameters.
  * pledgedSrcSize must be correct. If srcSize is not known at init time, use
  * value ZSTD_CONTENTSIZE_UNKNOWN.
- * Note : this prototype will be marked as deprecated and generate compilation warnings on reaching v1.5.x
+ * This prototype will generate compilation warnings.
  */
-ZSTDLIB_API size_t
-ZSTD_initCStream_usingCDict_advanced(ZSTD_CStream* zcs,
+ZSTD_DEPRECATED("use ZSTD_CCtx_reset and ZSTD_CCtx_refCDict, see zstd.h for detailed instructions")
+ZSTDLIB_STATIC_API
+size_t ZSTD_initCStream_usingCDict_advanced(ZSTD_CStream* zcs,
                                const ZSTD_CDict* cdict,
                                      ZSTD_frameParameters fParams,
                                      unsigned long long pledgedSrcSize);
 
 /*! ZSTD_resetCStream() :
- * This function is deprecated, and is equivalent to:
+ * This function is DEPRECATED, and is equivalent to:
  *     ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only);
  *     ZSTD_CCtx_setPledgedSrcSize(zcs, pledgedSrcSize);
+ * Note: ZSTD_resetCStream() interprets pledgedSrcSize == 0 as ZSTD_CONTENTSIZE_UNKNOWN, but
+ *       ZSTD_CCtx_setPledgedSrcSize() does not do the same, so ZSTD_CONTENTSIZE_UNKNOWN must be
+ *       explicitly specified.
  *
  *  start a new frame, using same parameters from previous frame.
  *  This is typically useful to skip dictionary loading stage, since it will re-use it in-place.
@@ -1810,9 +2523,11 @@ ZSTD_initCStream_usingCDict_advanced(ZSTD_CStream* zcs,
  *  For the time being, pledgedSrcSize==0 is interpreted as "srcSize unknown" for compatibility with older programs,
  *  but it will change to mean "empty" in future version, so use macro ZSTD_CONTENTSIZE_UNKNOWN instead.
  * @return : 0, or an error code (which can be tested using ZSTD_isError())
- *  Note : this prototype will be marked as deprecated and generate compilation warnings on reaching v1.5.x
+ *  This prototype will generate compilation warnings.
  */
-ZSTDLIB_API size_t ZSTD_resetCStream(ZSTD_CStream* zcs, unsigned long long pledgedSrcSize);
+ZSTD_DEPRECATED("use ZSTD_CCtx_reset, see zstd.h for detailed instructions")
+ZSTDLIB_STATIC_API
+size_t ZSTD_resetCStream(ZSTD_CStream* zcs, unsigned long long pledgedSrcSize);
 
 
 typedef struct {
@@ -1830,7 +2545,7 @@ typedef struct {
  * Note : (ingested - consumed) is amount of input data buffered internally, not yet compressed.
  * Aggregates progression inside active worker threads.
  */
-ZSTDLIB_API ZSTD_frameProgression ZSTD_getFrameProgression(const ZSTD_CCtx* cctx);
+ZSTDLIB_STATIC_API ZSTD_frameProgression ZSTD_getFrameProgression(const ZSTD_CCtx* cctx);
 
 /*! ZSTD_toFlushNow() :
  *  Tell how many bytes are ready to be flushed immediately.
@@ -1845,41 +2560,42 @@ ZSTDLIB_API ZSTD_frameProgression ZSTD_getFrameProgression(const ZSTD_CCtx* cctx
  *    therefore flush speed is limited by production speed of oldest job
  *    irrespective of the speed of concurrent (and newer) jobs.
  */
-ZSTDLIB_API size_t ZSTD_toFlushNow(ZSTD_CCtx* cctx);
+ZSTDLIB_STATIC_API size_t ZSTD_toFlushNow(ZSTD_CCtx* cctx);
 
 
 /*=====   Advanced Streaming decompression functions  =====*/
-/**
+
+/*!
  * This function is deprecated, and is equivalent to:
  *
  *     ZSTD_DCtx_reset(zds, ZSTD_reset_session_only);
  *     ZSTD_DCtx_loadDictionary(zds, dict, dictSize);
  *
  * note: no dictionary will be used if dict == NULL or dictSize < 8
- * Note : this prototype will be marked as deprecated and generate compilation warnings on reaching v1.5.x
  */
-ZSTDLIB_API size_t ZSTD_initDStream_usingDict(ZSTD_DStream* zds, const void* dict, size_t dictSize);
+ZSTD_DEPRECATED("use ZSTD_DCtx_reset + ZSTD_DCtx_loadDictionary, see zstd.h for detailed instructions")
+ZSTDLIB_STATIC_API size_t ZSTD_initDStream_usingDict(ZSTD_DStream* zds, const void* dict, size_t dictSize);
 
-/**
+/*!
  * This function is deprecated, and is equivalent to:
  *
  *     ZSTD_DCtx_reset(zds, ZSTD_reset_session_only);
  *     ZSTD_DCtx_refDDict(zds, ddict);
  *
  * note : ddict is referenced, it must outlive decompression session
- * Note : this prototype will be marked as deprecated and generate compilation warnings on reaching v1.5.x
  */
-ZSTDLIB_API size_t ZSTD_initDStream_usingDDict(ZSTD_DStream* zds, const ZSTD_DDict* ddict);
+ZSTD_DEPRECATED("use ZSTD_DCtx_reset + ZSTD_DCtx_refDDict, see zstd.h for detailed instructions")
+ZSTDLIB_STATIC_API size_t ZSTD_initDStream_usingDDict(ZSTD_DStream* zds, const ZSTD_DDict* ddict);
 
-/**
+/*!
  * This function is deprecated, and is equivalent to:
  *
  *     ZSTD_DCtx_reset(zds, ZSTD_reset_session_only);
  *
  * re-use decompression parameters from previous init; saves dictionary loading
- * Note : this prototype will be marked as deprecated and generate compilation warnings on reaching v1.5.x
  */
-ZSTDLIB_API size_t ZSTD_resetDStream(ZSTD_DStream* zds);
+ZSTD_DEPRECATED("use ZSTD_DCtx_reset, see zstd.h for detailed instructions")
+ZSTDLIB_STATIC_API size_t ZSTD_resetDStream(ZSTD_DStream* zds);
 
 
 /*********************************************************************
@@ -1898,9 +2614,7 @@ ZSTDLIB_API size_t ZSTD_resetDStream(ZSTD_DStream* zds);
   ZSTD_CCtx object can be re-used multiple times within successive compression operations.
 
   Start by initializing a context.
-  Use ZSTD_compressBegin(), or ZSTD_compressBegin_usingDict() for dictionary compression,
-  or ZSTD_compressBegin_advanced(), for finer parameter control.
-  It's also possible to duplicate a reference context which has already been initialized, using ZSTD_copyCCtx()
+  Use ZSTD_compressBegin(), or ZSTD_compressBegin_usingDict() for dictionary compression.
 
   Then, consume your input using ZSTD_compressContinue().
   There are some important considerations to keep in mind when using this advanced function :
@@ -1922,18 +2636,25 @@ ZSTDLIB_API size_t ZSTD_resetDStream(ZSTD_DStream* zds);
 */
 
 /*=====   Buffer-less streaming compression functions  =====*/
-ZSTDLIB_API size_t ZSTD_compressBegin(ZSTD_CCtx* cctx, int compressionLevel);
-ZSTDLIB_API size_t ZSTD_compressBegin_usingDict(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, int compressionLevel);
-ZSTDLIB_API size_t ZSTD_compressBegin_advanced(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, ZSTD_parameters params, unsigned long long pledgedSrcSize); /**< pledgedSrcSize : If srcSize is not known at init time, use ZSTD_CONTENTSIZE_UNKNOWN */
-ZSTDLIB_API size_t ZSTD_compressBegin_usingCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict); /**< note: fails if cdict==NULL */
-ZSTDLIB_API size_t ZSTD_compressBegin_usingCDict_advanced(ZSTD_CCtx* const cctx, const ZSTD_CDict* const cdict, ZSTD_frameParameters const fParams, unsigned long long const pledgedSrcSize);   /* compression parameters are already set within cdict. pledgedSrcSize must be correct. If srcSize is not known, use macro ZSTD_CONTENTSIZE_UNKNOWN */
-ZSTDLIB_API size_t ZSTD_copyCCtx(ZSTD_CCtx* cctx, const ZSTD_CCtx* preparedCCtx, unsigned long long pledgedSrcSize); /**<  note: if pledgedSrcSize is not known, use ZSTD_CONTENTSIZE_UNKNOWN */
-
-ZSTDLIB_API size_t ZSTD_compressContinue(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
-ZSTDLIB_API size_t ZSTD_compressEnd(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
-
-
-/*-
+ZSTDLIB_STATIC_API size_t ZSTD_compressBegin(ZSTD_CCtx* cctx, int compressionLevel);
+ZSTDLIB_STATIC_API size_t ZSTD_compressBegin_usingDict(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, int compressionLevel);
+ZSTDLIB_STATIC_API size_t ZSTD_compressBegin_usingCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict); /**< note: fails if cdict==NULL */
+
+ZSTD_DEPRECATED("This function will likely be removed in a future release. It is misleading and has very limited utility.")
+ZSTDLIB_STATIC_API
+size_t ZSTD_copyCCtx(ZSTD_CCtx* cctx, const ZSTD_CCtx* preparedCCtx, unsigned long long pledgedSrcSize); /**<  note: if pledgedSrcSize is not known, use ZSTD_CONTENTSIZE_UNKNOWN */
+
+ZSTDLIB_STATIC_API size_t ZSTD_compressContinue(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+ZSTDLIB_STATIC_API size_t ZSTD_compressEnd(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+
+/* The ZSTD_compressBegin_advanced() and ZSTD_compressBegin_usingCDict_advanced() are now DEPRECATED and will generate a compiler warning */
+ZSTD_DEPRECATED("use advanced API to access custom parameters")
+ZSTDLIB_STATIC_API
+size_t ZSTD_compressBegin_advanced(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, ZSTD_parameters params, unsigned long long pledgedSrcSize); /**< pledgedSrcSize : If srcSize is not known at init time, use ZSTD_CONTENTSIZE_UNKNOWN */
+ZSTD_DEPRECATED("use advanced API to access custom parameters")
+ZSTDLIB_STATIC_API
+size_t ZSTD_compressBegin_usingCDict_advanced(ZSTD_CCtx* const cctx, const ZSTD_CDict* const cdict, ZSTD_frameParameters const fParams, unsigned long long const pledgedSrcSize);   /* compression parameters are already set within cdict. pledgedSrcSize must be correct. If srcSize is not known, use macro ZSTD_CONTENTSIZE_UNKNOWN */
+/**
   Buffer-less streaming decompression (synchronous mode)
 
   A ZSTD_DCtx object is required to track streaming operations.
@@ -1944,8 +2665,8 @@ ZSTDLIB_API size_t ZSTD_compressEnd(ZSTD_CCtx* cctx, void* dst, size_t dstCapaci
   Frame header is extracted from the beginning of compressed frame, so providing only the frame's beginning is enough.
   Data fragment must be large enough to ensure successful decoding.
  `ZSTD_frameHeaderSize_max` bytes is guaranteed to always be large enough.
-  @result : 0 : successful decoding, the `ZSTD_frameHeader` structure is correctly filled.
-           >0 : `srcSize` is too small, please provide at least @result bytes on next attempt.
+  result  : 0 : successful decoding, the `ZSTD_frameHeader` structure is correctly filled.
+           >0 : `srcSize` is too small, please provide at least result bytes on next attempt.
            errorCode, which can be tested using ZSTD_isError().
 
   It fills a ZSTD_frameHeader structure with important information to correctly decode the frame,
@@ -1964,7 +2685,7 @@ ZSTDLIB_API size_t ZSTD_compressEnd(ZSTD_CCtx* cctx, void* dst, size_t dstCapaci
 
   The most memory efficient way is to use a round buffer of sufficient size.
   Sufficient size is determined by invoking ZSTD_decodingBufferSize_min(),
-  which can @return an error code if required value is too large for current system (in 32-bits mode).
+  which can return an error code if required value is too large for current system (in 32-bits mode).
   In a round buffer methodology, ZSTD_decompressContinue() decompresses each block next to previous one,
   up to the moment there is not enough room left in the buffer to guarantee decoding another full block,
   which maximum size is provided in `ZSTD_frameHeader` structure, field `blockSizeMax`.
@@ -1984,7 +2705,7 @@ ZSTDLIB_API size_t ZSTD_compressEnd(ZSTD_CCtx* cctx, void* dst, size_t dstCapaci
   ZSTD_nextSrcSizeToDecompress() tells how many bytes to provide as 'srcSize' to ZSTD_decompressContinue().
   ZSTD_decompressContinue() requires this _exact_ amount of bytes, or it will fail.
 
- @result of ZSTD_decompressContinue() is the number of bytes regenerated within 'dst' (necessarily <= dstCapacity).
+  result of ZSTD_decompressContinue() is the number of bytes regenerated within 'dst' (necessarily <= dstCapacity).
   It can be zero : it just means ZSTD_decompressContinue() has decoded some metadata item.
   It can also be an error code, which can be tested with ZSTD_isError().
 
@@ -2016,6 +2737,8 @@ typedef struct {
     unsigned headerSize;
     unsigned dictID;
     unsigned checksumFlag;
+    unsigned _reserved1;
+    unsigned _reserved2;
 } ZSTD_frameHeader;
 
 /*! ZSTD_getFrameHeader() :
@@ -2023,24 +2746,25 @@ typedef struct {
  * @return : 0, `zfhPtr` is correctly filled,
  *          >0, `srcSize` is too small, value is wanted `srcSize` amount,
  *           or an error code, which can be tested using ZSTD_isError() */
-ZSTDLIB_API size_t ZSTD_getFrameHeader(ZSTD_frameHeader* zfhPtr, const void* src, size_t srcSize);   /**< doesn't consume input */
+ZSTDLIB_STATIC_API size_t ZSTD_getFrameHeader(ZSTD_frameHeader* zfhPtr, const void* src, size_t srcSize);   /**< doesn't consume input */
 /*! ZSTD_getFrameHeader_advanced() :
  *  same as ZSTD_getFrameHeader(),
  *  with added capability to select a format (like ZSTD_f_zstd1_magicless) */
-ZSTDLIB_API size_t ZSTD_getFrameHeader_advanced(ZSTD_frameHeader* zfhPtr, const void* src, size_t srcSize, ZSTD_format_e format);
-ZSTDLIB_API size_t ZSTD_decodingBufferSize_min(unsigned long long windowSize, unsigned long long frameContentSize);  /**< when frame content size is not known, pass in frameContentSize == ZSTD_CONTENTSIZE_UNKNOWN */
+ZSTDLIB_STATIC_API size_t ZSTD_getFrameHeader_advanced(ZSTD_frameHeader* zfhPtr, const void* src, size_t srcSize, ZSTD_format_e format);
+ZSTDLIB_STATIC_API size_t ZSTD_decodingBufferSize_min(unsigned long long windowSize, unsigned long long frameContentSize);  /**< when frame content size is not known, pass in frameContentSize == ZSTD_CONTENTSIZE_UNKNOWN */
 
-ZSTDLIB_API size_t ZSTD_decompressBegin(ZSTD_DCtx* dctx);
-ZSTDLIB_API size_t ZSTD_decompressBegin_usingDict(ZSTD_DCtx* dctx, const void* dict, size_t dictSize);
-ZSTDLIB_API size_t ZSTD_decompressBegin_usingDDict(ZSTD_DCtx* dctx, const ZSTD_DDict* ddict);
+ZSTDLIB_STATIC_API size_t ZSTD_decompressBegin(ZSTD_DCtx* dctx);
+ZSTDLIB_STATIC_API size_t ZSTD_decompressBegin_usingDict(ZSTD_DCtx* dctx, const void* dict, size_t dictSize);
+ZSTDLIB_STATIC_API size_t ZSTD_decompressBegin_usingDDict(ZSTD_DCtx* dctx, const ZSTD_DDict* ddict);
 
-ZSTDLIB_API size_t ZSTD_nextSrcSizeToDecompress(ZSTD_DCtx* dctx);
-ZSTDLIB_API size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+ZSTDLIB_STATIC_API size_t ZSTD_nextSrcSizeToDecompress(ZSTD_DCtx* dctx);
+ZSTDLIB_STATIC_API size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
 
 /* misc */
-ZSTDLIB_API void   ZSTD_copyDCtx(ZSTD_DCtx* dctx, const ZSTD_DCtx* preparedDCtx);
+ZSTD_DEPRECATED("This function will likely be removed in the next minor release. It is misleading and has very limited utility.")
+ZSTDLIB_STATIC_API void   ZSTD_copyDCtx(ZSTD_DCtx* dctx, const ZSTD_DCtx* preparedDCtx);
 typedef enum { ZSTDnit_frameHeader, ZSTDnit_blockHeader, ZSTDnit_block, ZSTDnit_lastBlock, ZSTDnit_checksum, ZSTDnit_skippableFrame } ZSTD_nextInputType_e;
-ZSTDLIB_API ZSTD_nextInputType_e ZSTD_nextInputType(ZSTD_DCtx* dctx);
+ZSTDLIB_STATIC_API ZSTD_nextInputType_e ZSTD_nextInputType(ZSTD_DCtx* dctx);
 
 
 
@@ -2060,7 +2784,6 @@ ZSTDLIB_API ZSTD_nextInputType_e ZSTD_nextInputType(ZSTD_DCtx* dctx);
     - It is necessary to init context before starting
       + compression : any ZSTD_compressBegin*() variant, including with dictionary
       + decompression : any ZSTD_decompressBegin*() variant, including with dictionary
-      + copyCCtx() and copyDCtx() can be used too
     - Block size is limited, it must be <= ZSTD_getBlockSize() <= ZSTD_BLOCKSIZE_MAX == 128 KB
       + If input is larger than a block size, it's necessary to split input data into multiple blocks
       + For inputs larger than a single block, consider using regular ZSTD_compress() instead.
@@ -2077,12 +2800,173 @@ ZSTDLIB_API ZSTD_nextInputType_e ZSTD_nextInputType(ZSTD_DCtx* dctx);
 */
 
 /*=====   Raw zstd block functions  =====*/
-ZSTDLIB_API size_t ZSTD_getBlockSize   (const ZSTD_CCtx* cctx);
-ZSTDLIB_API size_t ZSTD_compressBlock  (ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
-ZSTDLIB_API size_t ZSTD_decompressBlock(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
-ZSTDLIB_API size_t ZSTD_insertBlock    (ZSTD_DCtx* dctx, const void* blockStart, size_t blockSize);  /**< insert uncompressed block into `dctx` history. Useful for multi-blocks decompression. */
+ZSTDLIB_STATIC_API size_t ZSTD_getBlockSize   (const ZSTD_CCtx* cctx);
+ZSTDLIB_STATIC_API size_t ZSTD_compressBlock  (ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+ZSTDLIB_STATIC_API size_t ZSTD_decompressBlock(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+ZSTDLIB_STATIC_API size_t ZSTD_insertBlock    (ZSTD_DCtx* dctx, const void* blockStart, size_t blockSize);  /**< insert uncompressed block into `dctx` history. Useful for multi-blocks decompression. */
 
 
+/* ********************* BLOCK-LEVEL SEQUENCE PRODUCER API *********************
+ *
+ * *** OVERVIEW ***
+ * The Block-Level Sequence Producer API allows users to provide their own custom
+ * sequence producer which libzstd invokes to process each block. The produced list
+ * of sequences (literals and matches) is then post-processed by libzstd to produce
+ * valid compressed blocks.
+ *
+ * This block-level offload API is a more granular complement of the existing
+ * frame-level offload API compressSequences() (introduced in v1.5.1). It offers
+ * an easier migration story for applications already integrated with libzstd: the
+ * user application continues to invoke the same compression functions
+ * ZSTD_compress2() or ZSTD_compressStream2() as usual, and transparently benefits
+ * from the specific advantages of the external sequence producer. For example,
+ * the sequence producer could be tuned to take advantage of known characteristics
+ * of the input, to offer better speed / ratio, or could leverage hardware
+ * acceleration not available within libzstd itself.
+ *
+ * See contrib/externalSequenceProducer for an example program employing the
+ * Block-Level Sequence Producer API.
+ *
+ * *** USAGE ***
+ * The user is responsible for implementing a function of type
+ * ZSTD_sequenceProducer_F. For each block, zstd will pass the following
+ * arguments to the user-provided function:
+ *
+ *   - sequenceProducerState: a pointer to a user-managed state for the sequence
+ *     producer.
+ *
+ *   - outSeqs, outSeqsCapacity: an output buffer for the sequence producer.
+ *     outSeqsCapacity is guaranteed >= ZSTD_sequenceBound(srcSize). The memory
+ *     backing outSeqs is managed by the CCtx.
+ *
+ *   - src, srcSize: an input buffer for the sequence producer to parse.
+ *     srcSize is guaranteed to be <= ZSTD_BLOCKSIZE_MAX.
+ *
+ *   - dict, dictSize: a history buffer, which may be empty, which the sequence
+ *     producer may reference as it parses the src buffer. Currently, zstd will
+ *     always pass dictSize == 0 into external sequence producers, but this will
+ *     change in the future.
+ *
+ *   - compressionLevel: a signed integer representing the zstd compression level
+ *     set by the user for the current operation. The sequence producer may choose
+ *     to use this information to change its compression strategy and speed/ratio
+ *     tradeoff. Note: the compression level does not reflect zstd parameters set
+ *     through the advanced API.
+ *
+ *   - windowSize: a size_t representing the maximum allowed offset for external
+ *     sequences. Note that sequence offsets are sometimes allowed to exceed the
+ *     windowSize if a dictionary is present, see doc/zstd_compression_format.md
+ *     for details.
+ *
+ * The user-provided function shall return a size_t representing the number of
+ * sequences written to outSeqs. This return value will be treated as an error
+ * code if it is greater than outSeqsCapacity. The return value must be non-zero
+ * if srcSize is non-zero. The ZSTD_SEQUENCE_PRODUCER_ERROR macro is provided
+ * for convenience, but any value greater than outSeqsCapacity will be treated as
+ * an error code.
+ *
+ * If the user-provided function does not return an error code, the sequences
+ * written to outSeqs must be a valid parse of the src buffer. Data corruption may
+ * occur if the parse is not valid. A parse is defined to be valid if the
+ * following conditions hold:
+ *   - The sum of matchLengths and literalLengths must equal srcSize.
+ *   - All sequences in the parse, except for the final sequence, must have
+ *     matchLength >= ZSTD_MINMATCH_MIN. The final sequence must have
+ *     matchLength >= ZSTD_MINMATCH_MIN or matchLength == 0.
+ *   - All offsets must respect the windowSize parameter as specified in
+ *     doc/zstd_compression_format.md.
+ *   - If the final sequence has matchLength == 0, it must also have offset == 0.
+ *
+ * zstd will only validate these conditions (and fail compression if they do not
+ * hold) if the ZSTD_c_validateSequences cParam is enabled. Note that sequence
+ * validation has a performance cost.
+ *
+ * If the user-provided function returns an error, zstd will either fall back
+ * to an internal sequence producer or fail the compression operation. The user can
+ * choose between the two behaviors by setting the ZSTD_c_enableSeqProducerFallback
+ * cParam. Fallback compression will follow any other cParam settings, such as
+ * compression level, the same as in a normal compression operation.
+ *
+ * The user shall instruct zstd to use a particular ZSTD_sequenceProducer_F
+ * function by calling
+ *         ZSTD_registerSequenceProducer(cctx,
+ *                                       sequenceProducerState,
+ *                                       sequenceProducer)
+ * This setting will persist until the next parameter reset of the CCtx.
+ *
+ * The sequenceProducerState must be initialized by the user before calling
+ * ZSTD_registerSequenceProducer(). The user is responsible for destroying the
+ * sequenceProducerState.
+ *
+ * *** LIMITATIONS ***
+ * This API is compatible with all zstd compression APIs which respect advanced parameters.
+ * However, there are three limitations:
+ *
+ * First, the ZSTD_c_enableLongDistanceMatching cParam is not currently supported.
+ * COMPRESSION WILL FAIL if it is enabled and the user tries to compress with a block-level
+ * external sequence producer.
+ *   - Note that ZSTD_c_enableLongDistanceMatching is auto-enabled by default in some
+ *     cases (see its documentation for details). Users must explicitly set
+ *     ZSTD_c_enableLongDistanceMatching to ZSTD_ps_disable in such cases if an external
+ *     sequence producer is registered.
+ *   - As of this writing, ZSTD_c_enableLongDistanceMatching is disabled by default
+ *     whenever ZSTD_c_windowLog < 128MB, but that's subject to change. Users should
+ *     check the docs on ZSTD_c_enableLongDistanceMatching whenever the Block-Level Sequence
+ *     Producer API is used in conjunction with advanced settings (like ZSTD_c_windowLog).
+ *
+ * Second, history buffers are not currently supported. Concretely, zstd will always pass
+ * dictSize == 0 to the external sequence producer (for now). This has two implications:
+ *   - Dictionaries are not currently supported. Compression will *not* fail if the user
+ *     references a dictionary, but the dictionary won't have any effect.
+ *   - Stream history is not currently supported. All advanced compression APIs, including
+ *     streaming APIs, work with external sequence producers, but each block is treated as
+ *     an independent chunk without history from previous blocks.
+ *
+ * Third, multi-threading within a single compression is not currently supported. In other words,
+ * COMPRESSION WILL FAIL if ZSTD_c_nbWorkers > 0 and an external sequence producer is registered.
+ * Multi-threading across compressions is fine: simply create one CCtx per thread.
+ *
+ * Long-term, we plan to overcome all three limitations. There is no technical blocker to
+ * overcoming them. It is purely a question of engineering effort.
+ */
+
+#define ZSTD_SEQUENCE_PRODUCER_ERROR ((size_t)(-1))
+
+typedef size_t ZSTD_sequenceProducer_F (
+  void* sequenceProducerState,
+  ZSTD_Sequence* outSeqs, size_t outSeqsCapacity,
+  const void* src, size_t srcSize,
+  const void* dict, size_t dictSize,
+  int compressionLevel,
+  size_t windowSize
+);
+
+/*! ZSTD_registerSequenceProducer() :
+ * Instruct zstd to use a block-level external sequence producer function.
+ *
+ * The sequenceProducerState must be initialized by the caller, and the caller is
+ * responsible for managing its lifetime. This parameter is sticky across
+ * compressions. It will remain set until the user explicitly resets compression
+ * parameters.
+ *
+ * Sequence producer registration is considered to be an "advanced parameter",
+ * part of the "advanced API". This means it will only have an effect on compression
+ * APIs which respect advanced parameters, such as compress2() and compressStream2().
+ * Older compression APIs such as compressCCtx(), which predate the introduction of
+ * "advanced parameters", will ignore any external sequence producer setting.
+ *
+ * The sequence producer can be "cleared" by registering a NULL function pointer. This
+ * removes all limitations described above in the "LIMITATIONS" section of the API docs.
+ *
+ * The user is strongly encouraged to read the full API documentation (above) before
+ * calling this function. */
+ZSTDLIB_STATIC_API void
+ZSTD_registerSequenceProducer(
+  ZSTD_CCtx* cctx,
+  void* sequenceProducerState,
+  ZSTD_sequenceProducer_F* sequenceProducer
+);
+
 #endif   /* ZSTD_H_ZSTD_STATIC_LINKING_ONLY */
 
 #if defined (__cplusplus)
diff --git a/contrib/zstd/zstd_common.c b/contrib/zstd/zstd_common.c
index 91fe3323a..320855247 100644
--- a/contrib/zstd/zstd_common.c
+++ b/contrib/zstd/zstd_common.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -13,8 +13,8 @@
 /*-*************************************
 *  Dependencies
 ***************************************/
-#include <stdlib.h>      /* malloc, calloc, free */
-#include <string.h>      /* memset */
+#define ZSTD_DEPS_NEED_MALLOC
+#include "zstd_deps.h"   /* ZSTD_malloc, ZSTD_calloc, ZSTD_free, ZSTD_memset */
 #include "error_private.h"
 #include "zstd_internal.h"
 
@@ -53,31 +53,31 @@ const char* ZSTD_getErrorString(ZSTD_ErrorCode code) { return ERR_getErrorString
 /*=**************************************************************
 *  Custom allocator
 ****************************************************************/
-void* ZSTD_malloc(size_t size, ZSTD_customMem customMem)
+void* ZSTD_customMalloc(size_t size, ZSTD_customMem customMem)
 {
     if (customMem.customAlloc)
         return customMem.customAlloc(customMem.opaque, size);
-    return malloc(size);
+    return ZSTD_malloc(size);
 }
 
-void* ZSTD_calloc(size_t size, ZSTD_customMem customMem)
+void* ZSTD_customCalloc(size_t size, ZSTD_customMem customMem)
 {
     if (customMem.customAlloc) {
         /* calloc implemented as malloc+memset;
          * not as efficient as calloc, but next best guess for custom malloc */
         void* const ptr = customMem.customAlloc(customMem.opaque, size);
-        memset(ptr, 0, size);
+        ZSTD_memset(ptr, 0, size);
         return ptr;
     }
-    return calloc(1, size);
+    return ZSTD_calloc(1, size);
 }
 
-void ZSTD_free(void* ptr, ZSTD_customMem customMem)
+void ZSTD_customFree(void* ptr, ZSTD_customMem customMem)
 {
     if (ptr!=NULL) {
         if (customMem.customFree)
             customMem.customFree(customMem.opaque, ptr);
         else
-            free(ptr);
+            ZSTD_free(ptr);
     }
 }
diff --git a/contrib/zstd/zstd_compress.c b/contrib/zstd/zstd_compress.c
index 526cb21f6..e11252c6a 100644
--- a/contrib/zstd/zstd_compress.c
+++ b/contrib/zstd/zstd_compress.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -11,14 +11,11 @@
 /*-*************************************
 *  Dependencies
 ***************************************/
-#include <limits.h>         /* INT_MAX */
-#include <string.h>         /* memset */
-#include "cpu.h"
+#include "zstd_deps.h"  /* INT_MAX, ZSTD_memset, ZSTD_memcpy */
 #include "mem.h"
 #include "hist.h"           /* HIST_countFast_wksp */
 #define FSE_STATIC_LINKING_ONLY   /* FSE_encodeSymbol */
 #include "fse.h"
-#define HUF_STATIC_LINKING_ONLY
 #include "huf.h"
 #include "zstd_compress_internal.h"
 #include "zstd_compress_sequences.h"
@@ -29,20 +26,49 @@
 #include "zstd_opt.h"
 #include "zstd_ldm.h"
 #include "zstd_compress_superblock.h"
+#include  "bits.h"      /* ZSTD_highbit32 */
+
+/* ***************************************************************
+*  Tuning parameters
+*****************************************************************/
+/*!
+ * COMPRESS_HEAPMODE :
+ * Select how default decompression function ZSTD_compress() allocates its context,
+ * on stack (0, default), or into heap (1).
+ * Note that functions with explicit context such as ZSTD_compressCCtx() are unaffected.
+ */
+#ifndef ZSTD_COMPRESS_HEAPMODE
+#  define ZSTD_COMPRESS_HEAPMODE 0
+#endif
 
+/*!
+ * ZSTD_HASHLOG3_MAX :
+ * Maximum size of the hash table dedicated to find 3-bytes matches,
+ * in log format, aka 17 => 1 << 17 == 128Ki positions.
+ * This structure is only used in zstd_opt.
+ * Since allocation is centralized for all strategies, it has to be known here.
+ * The actual (selected) size of the hash table is then stored in ZSTD_matchState_t.hashLog3,
+ * so that zstd_opt.c doesn't need to know about this constant.
+ */
+#ifndef ZSTD_HASHLOG3_MAX
+#  define ZSTD_HASHLOG3_MAX 17
+#endif
 
 /*-*************************************
 *  Helper functions
 ***************************************/
 /* ZSTD_compressBound()
- * Note that the result from this function is only compatible with the "normal"
- * full-block strategy.
- * When there are a lot of small blocks due to frequent flush in streaming mode
- * the overhead of headers can make the compressed data to be larger than the
- * return value of ZSTD_compressBound().
+ * Note that the result from this function is only valid for
+ * the one-pass compression functions.
+ * When employing the streaming mode,
+ * if flushes are frequently altering the size of blocks,
+ * the overhead from block headers can make the compressed data larger
+ * than the return value of ZSTD_compressBound().
  */
 size_t ZSTD_compressBound(size_t srcSize) {
-    return ZSTD_COMPRESSBOUND(srcSize);
+    size_t const r = ZSTD_COMPRESSBOUND(srcSize);
+    if (r==0) return ERROR(srcSize_wrong);
+    return r;
 }
 
 
@@ -52,6 +78,7 @@ size_t ZSTD_compressBound(size_t srcSize) {
 struct ZSTD_CDict_s {
     const void* dictContent;
     size_t dictContentSize;
+    ZSTD_dictContentType_e dictContentType; /* The dictContentType the CDict was created with */
     U32* entropyWorkspace; /* entropy workspace of HUF_WORKSPACE_SIZE bytes */
     ZSTD_cwksp workspace;
     ZSTD_matchState_t matchState;
@@ -59,6 +86,10 @@ struct ZSTD_CDict_s {
     ZSTD_customMem customMem;
     U32 dictID;
     int compressionLevel; /* 0 indicates that advanced API was used to select CDict params */
+    ZSTD_paramSwitch_e useRowMatchFinder; /* Indicates whether the CDict was created with params that would use
+                                           * row-based matchfinder. Unless the cdict is reloaded, we will use
+                                           * the same greedy/lazy matchfinder at compression time.
+                                           */
 };  /* typedef'd to ZSTD_CDict within "zstd.h" */
 
 ZSTD_CCtx* ZSTD_createCCtx(void)
@@ -69,9 +100,9 @@ ZSTD_CCtx* ZSTD_createCCtx(void)
 static void ZSTD_initCCtx(ZSTD_CCtx* cctx, ZSTD_customMem memManager)
 {
     assert(cctx != NULL);
-    memset(cctx, 0, sizeof(*cctx));
+    ZSTD_memset(cctx, 0, sizeof(*cctx));
     cctx->customMem = memManager;
-    cctx->bmi2 = ZSTD_cpuid_bmi2(ZSTD_cpuid());
+    cctx->bmi2 = ZSTD_cpuSupportsBmi2();
     {   size_t const err = ZSTD_CCtx_reset(cctx, ZSTD_reset_parameters);
         assert(!ZSTD_isError(err));
         (void)err;
@@ -82,8 +113,8 @@ ZSTD_CCtx* ZSTD_createCCtx_advanced(ZSTD_customMem customMem)
 {
     ZSTD_STATIC_ASSERT(zcss_init==0);
     ZSTD_STATIC_ASSERT(ZSTD_CONTENTSIZE_UNKNOWN==(0ULL - 1));
-    if (!customMem.customAlloc ^ !customMem.customFree) return NULL;
-    {   ZSTD_CCtx* const cctx = (ZSTD_CCtx*)ZSTD_malloc(sizeof(ZSTD_CCtx), customMem);
+    if ((!customMem.customAlloc) ^ (!customMem.customFree)) return NULL;
+    {   ZSTD_CCtx* const cctx = (ZSTD_CCtx*)ZSTD_customMalloc(sizeof(ZSTD_CCtx), customMem);
         if (!cctx) return NULL;
         ZSTD_initCCtx(cctx, customMem);
         return cctx;
@@ -96,20 +127,20 @@ ZSTD_CCtx* ZSTD_initStaticCCtx(void* workspace, size_t workspaceSize)
     ZSTD_CCtx* cctx;
     if (workspaceSize <= sizeof(ZSTD_CCtx)) return NULL;  /* minimum size */
     if ((size_t)workspace & 7) return NULL;  /* must be 8-aligned */
-    ZSTD_cwksp_init(&ws, workspace, workspaceSize);
+    ZSTD_cwksp_init(&ws, workspace, workspaceSize, ZSTD_cwksp_static_alloc);
 
     cctx = (ZSTD_CCtx*)ZSTD_cwksp_reserve_object(&ws, sizeof(ZSTD_CCtx));
     if (cctx == NULL) return NULL;
 
-    memset(cctx, 0, sizeof(ZSTD_CCtx));
+    ZSTD_memset(cctx, 0, sizeof(ZSTD_CCtx));
     ZSTD_cwksp_move(&cctx->workspace, &ws);
     cctx->staticSize = workspaceSize;
 
     /* statically sized space. entropyWorkspace never moves (but prev/next block swap places) */
-    if (!ZSTD_cwksp_check_available(&cctx->workspace, HUF_WORKSPACE_SIZE + 2 * sizeof(ZSTD_compressedBlockState_t))) return NULL;
+    if (!ZSTD_cwksp_check_available(&cctx->workspace, ENTROPY_WORKSPACE_SIZE + 2 * sizeof(ZSTD_compressedBlockState_t))) return NULL;
     cctx->blockState.prevCBlock = (ZSTD_compressedBlockState_t*)ZSTD_cwksp_reserve_object(&cctx->workspace, sizeof(ZSTD_compressedBlockState_t));
     cctx->blockState.nextCBlock = (ZSTD_compressedBlockState_t*)ZSTD_cwksp_reserve_object(&cctx->workspace, sizeof(ZSTD_compressedBlockState_t));
-    cctx->entropyWorkspace = (U32*)ZSTD_cwksp_reserve_object(&cctx->workspace, HUF_WORKSPACE_SIZE);
+    cctx->entropyWorkspace = (U32*)ZSTD_cwksp_reserve_object(&cctx->workspace, ENTROPY_WORKSPACE_SIZE);
     cctx->bmi2 = ZSTD_cpuid_bmi2(ZSTD_cpuid());
     return cctx;
 }
@@ -119,10 +150,10 @@ ZSTD_CCtx* ZSTD_initStaticCCtx(void* workspace, size_t workspaceSize)
  */
 static void ZSTD_clearAllDicts(ZSTD_CCtx* cctx)
 {
-    ZSTD_free(cctx->localDict.dictBuffer, cctx->customMem);
+    ZSTD_customFree(cctx->localDict.dictBuffer, cctx->customMem);
     ZSTD_freeCDict(cctx->localDict.cdict);
-    memset(&cctx->localDict, 0, sizeof(cctx->localDict));
-    memset(&cctx->prefixDict, 0, sizeof(cctx->prefixDict));
+    ZSTD_memset(&cctx->localDict, 0, sizeof(cctx->localDict));
+    ZSTD_memset(&cctx->prefixDict, 0, sizeof(cctx->prefixDict));
     cctx->cdict = NULL;
 }
 
@@ -149,12 +180,9 @@ size_t ZSTD_freeCCtx(ZSTD_CCtx* cctx)
     if (cctx==NULL) return 0;   /* support free on NULL */
     RETURN_ERROR_IF(cctx->staticSize, memory_allocation,
                     "not compatible with static CCtx");
-    {
-        int cctxInWorkspace = ZSTD_cwksp_owns_buffer(&cctx->workspace, cctx);
+    {   int cctxInWorkspace = ZSTD_cwksp_owns_buffer(&cctx->workspace, cctx);
         ZSTD_freeCCtxContent(cctx);
-        if (!cctxInWorkspace) {
-            ZSTD_free(cctx, cctx->customMem);
-        }
+        if (!cctxInWorkspace) ZSTD_customFree(cctx, cctx->customMem);
     }
     return 0;
 }
@@ -189,15 +217,116 @@ size_t ZSTD_sizeof_CStream(const ZSTD_CStream* zcs)
 /* private API call, for dictBuilder only */
 const seqStore_t* ZSTD_getSeqStore(const ZSTD_CCtx* ctx) { return &(ctx->seqStore); }
 
+/* Returns true if the strategy supports using a row based matchfinder */
+static int ZSTD_rowMatchFinderSupported(const ZSTD_strategy strategy) {
+    return (strategy >= ZSTD_greedy && strategy <= ZSTD_lazy2);
+}
+
+/* Returns true if the strategy and useRowMatchFinder mode indicate that we will use the row based matchfinder
+ * for this compression.
+ */
+static int ZSTD_rowMatchFinderUsed(const ZSTD_strategy strategy, const ZSTD_paramSwitch_e mode) {
+    assert(mode != ZSTD_ps_auto);
+    return ZSTD_rowMatchFinderSupported(strategy) && (mode == ZSTD_ps_enable);
+}
+
+/* Returns row matchfinder usage given an initial mode and cParams */
+static ZSTD_paramSwitch_e ZSTD_resolveRowMatchFinderMode(ZSTD_paramSwitch_e mode,
+                                                         const ZSTD_compressionParameters* const cParams) {
+#if defined(ZSTD_ARCH_X86_SSE2) || defined(ZSTD_ARCH_ARM_NEON)
+    int const kHasSIMD128 = 1;
+#else
+    int const kHasSIMD128 = 0;
+#endif
+    if (mode != ZSTD_ps_auto) return mode; /* if requested enabled, but no SIMD, we still will use row matchfinder */
+    mode = ZSTD_ps_disable;
+    if (!ZSTD_rowMatchFinderSupported(cParams->strategy)) return mode;
+    if (kHasSIMD128) {
+        if (cParams->windowLog > 14) mode = ZSTD_ps_enable;
+    } else {
+        if (cParams->windowLog > 17) mode = ZSTD_ps_enable;
+    }
+    return mode;
+}
+
+/* Returns block splitter usage (generally speaking, when using slower/stronger compression modes) */
+static ZSTD_paramSwitch_e ZSTD_resolveBlockSplitterMode(ZSTD_paramSwitch_e mode,
+                                                        const ZSTD_compressionParameters* const cParams) {
+    if (mode != ZSTD_ps_auto) return mode;
+    return (cParams->strategy >= ZSTD_btopt && cParams->windowLog >= 17) ? ZSTD_ps_enable : ZSTD_ps_disable;
+}
+
+/* Returns 1 if the arguments indicate that we should allocate a chainTable, 0 otherwise */
+static int ZSTD_allocateChainTable(const ZSTD_strategy strategy,
+                                   const ZSTD_paramSwitch_e useRowMatchFinder,
+                                   const U32 forDDSDict) {
+    assert(useRowMatchFinder != ZSTD_ps_auto);
+    /* We always should allocate a chaintable if we are allocating a matchstate for a DDS dictionary matchstate.
+     * We do not allocate a chaintable if we are using ZSTD_fast, or are using the row-based matchfinder.
+     */
+    return forDDSDict || ((strategy != ZSTD_fast) && !ZSTD_rowMatchFinderUsed(strategy, useRowMatchFinder));
+}
+
+/* Returns ZSTD_ps_enable if compression parameters are such that we should
+ * enable long distance matching (wlog >= 27, strategy >= btopt).
+ * Returns ZSTD_ps_disable otherwise.
+ */
+static ZSTD_paramSwitch_e ZSTD_resolveEnableLdm(ZSTD_paramSwitch_e mode,
+                                 const ZSTD_compressionParameters* const cParams) {
+    if (mode != ZSTD_ps_auto) return mode;
+    return (cParams->strategy >= ZSTD_btopt && cParams->windowLog >= 27) ? ZSTD_ps_enable : ZSTD_ps_disable;
+}
+
+static int ZSTD_resolveExternalSequenceValidation(int mode) {
+    return mode;
+}
+
+/* Resolves maxBlockSize to the default if no value is present. */
+static size_t ZSTD_resolveMaxBlockSize(size_t maxBlockSize) {
+    if (maxBlockSize == 0) {
+        return ZSTD_BLOCKSIZE_MAX;
+    } else {
+        return maxBlockSize;
+    }
+}
+
+static ZSTD_paramSwitch_e ZSTD_resolveExternalRepcodeSearch(ZSTD_paramSwitch_e value, int cLevel) {
+    if (value != ZSTD_ps_auto) return value;
+    if (cLevel < 10) {
+        return ZSTD_ps_disable;
+    } else {
+        return ZSTD_ps_enable;
+    }
+}
+
+/* Returns 1 if compression parameters are such that CDict hashtable and chaintable indices are tagged.
+ * If so, the tags need to be removed in ZSTD_resetCCtx_byCopyingCDict. */
+static int ZSTD_CDictIndicesAreTagged(const ZSTD_compressionParameters* const cParams) {
+    return cParams->strategy == ZSTD_fast || cParams->strategy == ZSTD_dfast;
+}
+
 static ZSTD_CCtx_params ZSTD_makeCCtxParamsFromCParams(
         ZSTD_compressionParameters cParams)
 {
     ZSTD_CCtx_params cctxParams;
-    memset(&cctxParams, 0, sizeof(cctxParams));
+    /* should not matter, as all cParams are presumed properly defined */
+    ZSTD_CCtxParams_init(&cctxParams, ZSTD_CLEVEL_DEFAULT);
     cctxParams.cParams = cParams;
-    cctxParams.compressionLevel = ZSTD_CLEVEL_DEFAULT;  /* should not matter, as all cParams are presumed properly defined */
+
+    /* Adjust advanced params according to cParams */
+    cctxParams.ldmParams.enableLdm = ZSTD_resolveEnableLdm(cctxParams.ldmParams.enableLdm, &cParams);
+    if (cctxParams.ldmParams.enableLdm == ZSTD_ps_enable) {
+        ZSTD_ldm_adjustParameters(&cctxParams.ldmParams, &cParams);
+        assert(cctxParams.ldmParams.hashLog >= cctxParams.ldmParams.bucketSizeLog);
+        assert(cctxParams.ldmParams.hashRateLog < 32);
+    }
+    cctxParams.useBlockSplitter = ZSTD_resolveBlockSplitterMode(cctxParams.useBlockSplitter, &cParams);
+    cctxParams.useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(cctxParams.useRowMatchFinder, &cParams);
+    cctxParams.validateSequences = ZSTD_resolveExternalSequenceValidation(cctxParams.validateSequences);
+    cctxParams.maxBlockSize = ZSTD_resolveMaxBlockSize(cctxParams.maxBlockSize);
+    cctxParams.searchForExternalRepcodes = ZSTD_resolveExternalRepcodeSearch(cctxParams.searchForExternalRepcodes,
+                                                                             cctxParams.compressionLevel);
     assert(!ZSTD_checkCParams(cParams));
-    cctxParams.fParams.contentSizeFlag = 1;
     return cctxParams;
 }
 
@@ -205,13 +334,12 @@ static ZSTD_CCtx_params* ZSTD_createCCtxParams_advanced(
         ZSTD_customMem customMem)
 {
     ZSTD_CCtx_params* params;
-    if (!customMem.customAlloc ^ !customMem.customFree) return NULL;
-    params = (ZSTD_CCtx_params*)ZSTD_calloc(
+    if ((!customMem.customAlloc) ^ (!customMem.customFree)) return NULL;
+    params = (ZSTD_CCtx_params*)ZSTD_customCalloc(
             sizeof(ZSTD_CCtx_params), customMem);
     if (!params) { return NULL; }
+    ZSTD_CCtxParams_init(params, ZSTD_CLEVEL_DEFAULT);
     params->customMem = customMem;
-    params->compressionLevel = ZSTD_CLEVEL_DEFAULT;
-    params->fParams.contentSizeFlag = 1;
     return params;
 }
 
@@ -223,7 +351,7 @@ ZSTD_CCtx_params* ZSTD_createCCtxParams(void)
 size_t ZSTD_freeCCtxParams(ZSTD_CCtx_params* params)
 {
     if (params == NULL) { return 0; }
-    ZSTD_free(params, params->customMem);
+    ZSTD_customFree(params, params->customMem);
     return 0;
 }
 
@@ -234,35 +362,63 @@ size_t ZSTD_CCtxParams_reset(ZSTD_CCtx_params* params)
 
 size_t ZSTD_CCtxParams_init(ZSTD_CCtx_params* cctxParams, int compressionLevel) {
     RETURN_ERROR_IF(!cctxParams, GENERIC, "NULL pointer!");
-    memset(cctxParams, 0, sizeof(*cctxParams));
+    ZSTD_memset(cctxParams, 0, sizeof(*cctxParams));
     cctxParams->compressionLevel = compressionLevel;
     cctxParams->fParams.contentSizeFlag = 1;
     return 0;
 }
 
+#define ZSTD_NO_CLEVEL 0
+
+/**
+ * Initializes `cctxParams` from `params` and `compressionLevel`.
+ * @param compressionLevel If params are derived from a compression level then that compression level, otherwise ZSTD_NO_CLEVEL.
+ */
+static void
+ZSTD_CCtxParams_init_internal(ZSTD_CCtx_params* cctxParams,
+                        const ZSTD_parameters* params,
+                              int compressionLevel)
+{
+    assert(!ZSTD_checkCParams(params->cParams));
+    ZSTD_memset(cctxParams, 0, sizeof(*cctxParams));
+    cctxParams->cParams = params->cParams;
+    cctxParams->fParams = params->fParams;
+    /* Should not matter, as all cParams are presumed properly defined.
+     * But, set it for tracing anyway.
+     */
+    cctxParams->compressionLevel = compressionLevel;
+    cctxParams->useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(cctxParams->useRowMatchFinder, &params->cParams);
+    cctxParams->useBlockSplitter = ZSTD_resolveBlockSplitterMode(cctxParams->useBlockSplitter, &params->cParams);
+    cctxParams->ldmParams.enableLdm = ZSTD_resolveEnableLdm(cctxParams->ldmParams.enableLdm, &params->cParams);
+    cctxParams->validateSequences = ZSTD_resolveExternalSequenceValidation(cctxParams->validateSequences);
+    cctxParams->maxBlockSize = ZSTD_resolveMaxBlockSize(cctxParams->maxBlockSize);
+    cctxParams->searchForExternalRepcodes = ZSTD_resolveExternalRepcodeSearch(cctxParams->searchForExternalRepcodes, compressionLevel);
+    DEBUGLOG(4, "ZSTD_CCtxParams_init_internal: useRowMatchFinder=%d, useBlockSplitter=%d ldm=%d",
+                cctxParams->useRowMatchFinder, cctxParams->useBlockSplitter, cctxParams->ldmParams.enableLdm);
+}
+
 size_t ZSTD_CCtxParams_init_advanced(ZSTD_CCtx_params* cctxParams, ZSTD_parameters params)
 {
     RETURN_ERROR_IF(!cctxParams, GENERIC, "NULL pointer!");
     FORWARD_IF_ERROR( ZSTD_checkCParams(params.cParams) , "");
-    memset(cctxParams, 0, sizeof(*cctxParams));
-    assert(!ZSTD_checkCParams(params.cParams));
-    cctxParams->cParams = params.cParams;
-    cctxParams->fParams = params.fParams;
-    cctxParams->compressionLevel = ZSTD_CLEVEL_DEFAULT;   /* should not matter, as all cParams are presumed properly defined */
+    ZSTD_CCtxParams_init_internal(cctxParams, &params, ZSTD_NO_CLEVEL);
     return 0;
 }
 
-/* ZSTD_assignParamsToCCtxParams() :
- * params is presumed valid at this stage */
-static ZSTD_CCtx_params ZSTD_assignParamsToCCtxParams(
-        const ZSTD_CCtx_params* cctxParams, const ZSTD_parameters* params)
+/**
+ * Sets cctxParams' cParams and fParams from params, but otherwise leaves them alone.
+ * @param params Validated zstd parameters.
+ */
+static void ZSTD_CCtxParams_setZstdParams(
+        ZSTD_CCtx_params* cctxParams, const ZSTD_parameters* params)
 {
-    ZSTD_CCtx_params ret = *cctxParams;
     assert(!ZSTD_checkCParams(params->cParams));
-    ret.cParams = params->cParams;
-    ret.fParams = params->fParams;
-    ret.compressionLevel = ZSTD_CLEVEL_DEFAULT;   /* should not matter, as all cParams are presumed properly defined */
-    return ret;
+    cctxParams->cParams = params->cParams;
+    cctxParams->fParams = params->fParams;
+    /* Should not matter, as all cParams are presumed properly defined.
+     * But, set it for tracing anyway.
+     */
+    cctxParams->compressionLevel = ZSTD_NO_CLEVEL;
 }
 
 ZSTD_bounds ZSTD_cParam_getBounds(ZSTD_cParameter param)
@@ -354,11 +510,16 @@ ZSTD_bounds ZSTD_cParam_getBounds(ZSTD_cParameter param)
 #endif
         return bounds;
 
-    case ZSTD_c_enableLongDistanceMatching:
+    case ZSTD_c_enableDedicatedDictSearch:
         bounds.lowerBound = 0;
         bounds.upperBound = 1;
         return bounds;
 
+    case ZSTD_c_enableLongDistanceMatching:
+        bounds.lowerBound = (int)ZSTD_ps_auto;
+        bounds.upperBound = (int)ZSTD_ps_disable;
+        return bounds;
+
     case ZSTD_c_ldmHashLog:
         bounds.lowerBound = ZSTD_LDM_HASHLOG_MIN;
         bounds.upperBound = ZSTD_LDM_HASHLOG_MAX;
@@ -397,15 +558,15 @@ ZSTD_bounds ZSTD_cParam_getBounds(ZSTD_cParameter param)
         return bounds;
 
     case ZSTD_c_forceAttachDict:
-        ZSTD_STATIC_ASSERT(ZSTD_dictDefaultAttach < ZSTD_dictForceCopy);
+        ZSTD_STATIC_ASSERT(ZSTD_dictDefaultAttach < ZSTD_dictForceLoad);
         bounds.lowerBound = ZSTD_dictDefaultAttach;
         bounds.upperBound = ZSTD_dictForceLoad;       /* note : how to ensure at compile time that this is the highest value enum ? */
         return bounds;
 
     case ZSTD_c_literalCompressionMode:
-        ZSTD_STATIC_ASSERT(ZSTD_lcm_auto < ZSTD_lcm_huffman && ZSTD_lcm_huffman < ZSTD_lcm_uncompressed);
-        bounds.lowerBound = ZSTD_lcm_auto;
-        bounds.upperBound = ZSTD_lcm_uncompressed;
+        ZSTD_STATIC_ASSERT(ZSTD_ps_auto < ZSTD_ps_enable && ZSTD_ps_enable < ZSTD_ps_disable);
+        bounds.lowerBound = (int)ZSTD_ps_auto;
+        bounds.upperBound = (int)ZSTD_ps_disable;
         return bounds;
 
     case ZSTD_c_targetCBlockSize:
@@ -418,6 +579,57 @@ ZSTD_bounds ZSTD_cParam_getBounds(ZSTD_cParameter param)
         bounds.upperBound = ZSTD_SRCSIZEHINT_MAX;
         return bounds;
 
+    case ZSTD_c_stableInBuffer:
+    case ZSTD_c_stableOutBuffer:
+        bounds.lowerBound = (int)ZSTD_bm_buffered;
+        bounds.upperBound = (int)ZSTD_bm_stable;
+        return bounds;
+
+    case ZSTD_c_blockDelimiters:
+        bounds.lowerBound = (int)ZSTD_sf_noBlockDelimiters;
+        bounds.upperBound = (int)ZSTD_sf_explicitBlockDelimiters;
+        return bounds;
+
+    case ZSTD_c_validateSequences:
+        bounds.lowerBound = 0;
+        bounds.upperBound = 1;
+        return bounds;
+
+    case ZSTD_c_useBlockSplitter:
+        bounds.lowerBound = (int)ZSTD_ps_auto;
+        bounds.upperBound = (int)ZSTD_ps_disable;
+        return bounds;
+
+    case ZSTD_c_useRowMatchFinder:
+        bounds.lowerBound = (int)ZSTD_ps_auto;
+        bounds.upperBound = (int)ZSTD_ps_disable;
+        return bounds;
+
+    case ZSTD_c_deterministicRefPrefix:
+        bounds.lowerBound = 0;
+        bounds.upperBound = 1;
+        return bounds;
+
+    case ZSTD_c_prefetchCDictTables:
+        bounds.lowerBound = (int)ZSTD_ps_auto;
+        bounds.upperBound = (int)ZSTD_ps_disable;
+        return bounds;
+
+    case ZSTD_c_enableSeqProducerFallback:
+        bounds.lowerBound = 0;
+        bounds.upperBound = 1;
+        return bounds;
+
+    case ZSTD_c_maxBlockSize:
+        bounds.lowerBound = ZSTD_BLOCKSIZE_MAX_MIN;
+        bounds.upperBound = ZSTD_BLOCKSIZE_MAX;
+        return bounds;
+
+    case ZSTD_c_searchForExternalRepcodes:
+        bounds.lowerBound = (int)ZSTD_ps_auto;
+        bounds.upperBound = (int)ZSTD_ps_disable;
+        return bounds;
+
     default:
         bounds.error = ERROR(parameter_unsupported);
         return bounds;
@@ -465,6 +677,7 @@ static int ZSTD_isUpdateAuthorized(ZSTD_cParameter param)
     case ZSTD_c_jobSize:
     case ZSTD_c_overlapLog:
     case ZSTD_c_rsyncable:
+    case ZSTD_c_enableDedicatedDictSearch:
     case ZSTD_c_enableLongDistanceMatching:
     case ZSTD_c_ldmHashLog:
     case ZSTD_c_ldmMinMatch:
@@ -474,6 +687,17 @@ static int ZSTD_isUpdateAuthorized(ZSTD_cParameter param)
     case ZSTD_c_literalCompressionMode:
     case ZSTD_c_targetCBlockSize:
     case ZSTD_c_srcSizeHint:
+    case ZSTD_c_stableInBuffer:
+    case ZSTD_c_stableOutBuffer:
+    case ZSTD_c_blockDelimiters:
+    case ZSTD_c_validateSequences:
+    case ZSTD_c_useBlockSplitter:
+    case ZSTD_c_useRowMatchFinder:
+    case ZSTD_c_deterministicRefPrefix:
+    case ZSTD_c_prefetchCDictTables:
+    case ZSTD_c_enableSeqProducerFallback:
+    case ZSTD_c_maxBlockSize:
+    case ZSTD_c_searchForExternalRepcodes:
     default:
         return 0;
     }
@@ -486,7 +710,7 @@ size_t ZSTD_CCtx_setParameter(ZSTD_CCtx* cctx, ZSTD_cParameter param, int value)
         if (ZSTD_isUpdateAuthorized(param)) {
             cctx->cParamsChanged = 1;
         } else {
-            RETURN_ERROR(stage_wrong, "can only set params in ctx init stage");
+            RETURN_ERROR(stage_wrong, "can only set params in cctx init stage");
     }   }
 
     switch(param)
@@ -515,12 +739,24 @@ size_t ZSTD_CCtx_setParameter(ZSTD_CCtx* cctx, ZSTD_cParameter param, int value)
     case ZSTD_c_jobSize:
     case ZSTD_c_overlapLog:
     case ZSTD_c_rsyncable:
+    case ZSTD_c_enableDedicatedDictSearch:
     case ZSTD_c_enableLongDistanceMatching:
     case ZSTD_c_ldmHashLog:
     case ZSTD_c_ldmMinMatch:
     case ZSTD_c_ldmBucketSizeLog:
     case ZSTD_c_targetCBlockSize:
     case ZSTD_c_srcSizeHint:
+    case ZSTD_c_stableInBuffer:
+    case ZSTD_c_stableOutBuffer:
+    case ZSTD_c_blockDelimiters:
+    case ZSTD_c_validateSequences:
+    case ZSTD_c_useBlockSplitter:
+    case ZSTD_c_useRowMatchFinder:
+    case ZSTD_c_deterministicRefPrefix:
+    case ZSTD_c_prefetchCDictTables:
+    case ZSTD_c_enableSeqProducerFallback:
+    case ZSTD_c_maxBlockSize:
+    case ZSTD_c_searchForExternalRepcodes:
         break;
 
     default: RETURN_ERROR(parameter_unsupported, "unknown parameter");
@@ -541,9 +777,10 @@ size_t ZSTD_CCtxParams_setParameter(ZSTD_CCtx_params* CCtxParams,
 
     case ZSTD_c_compressionLevel : {
         FORWARD_IF_ERROR(ZSTD_cParam_clampBounds(param, &value), "");
-        if (value) {  /* 0 : does not change current level */
+        if (value == 0)
+            CCtxParams->compressionLevel = ZSTD_CLEVEL_DEFAULT; /* 0 == default */
+        else
             CCtxParams->compressionLevel = value;
-        }
         if (CCtxParams->compressionLevel >= 0) return (size_t)CCtxParams->compressionLevel;
         return 0;  /* return type (size_t) cannot represent negative values */
     }
@@ -575,12 +812,12 @@ size_t ZSTD_CCtxParams_setParameter(ZSTD_CCtx_params* CCtxParams,
     case ZSTD_c_minMatch :
         if (value!=0)   /* 0 => use default */
             BOUNDCHECK(ZSTD_c_minMatch, value);
-        CCtxParams->cParams.minMatch = value;
+        CCtxParams->cParams.minMatch = (U32)value;
         return CCtxParams->cParams.minMatch;
 
     case ZSTD_c_targetLength :
         BOUNDCHECK(ZSTD_c_targetLength, value);
-        CCtxParams->cParams.targetLength = value;
+        CCtxParams->cParams.targetLength = (U32)value;
         return CCtxParams->cParams.targetLength;
 
     case ZSTD_c_strategy :
@@ -593,12 +830,12 @@ size_t ZSTD_CCtxParams_setParameter(ZSTD_CCtx_params* CCtxParams,
         /* Content size written in frame header _when known_ (default:1) */
         DEBUGLOG(4, "set content size flag = %u", (value!=0));
         CCtxParams->fParams.contentSizeFlag = value != 0;
-        return CCtxParams->fParams.contentSizeFlag;
+        return (size_t)CCtxParams->fParams.contentSizeFlag;
 
     case ZSTD_c_checksumFlag :
         /* A 32-bits content checksum will be calculated and written at end of frame (default:0) */
         CCtxParams->fParams.checksumFlag = value != 0;
-        return CCtxParams->fParams.checksumFlag;
+        return (size_t)CCtxParams->fParams.checksumFlag;
 
     case ZSTD_c_dictIDFlag : /* When applicable, dictionary's dictID is provided in frame header (default:1) */
         DEBUGLOG(4, "set dictIDFlag = %u", (value!=0));
@@ -607,18 +844,18 @@ size_t ZSTD_CCtxParams_setParameter(ZSTD_CCtx_params* CCtxParams,
 
     case ZSTD_c_forceMaxWindow :
         CCtxParams->forceWindow = (value != 0);
-        return CCtxParams->forceWindow;
+        return (size_t)CCtxParams->forceWindow;
 
     case ZSTD_c_forceAttachDict : {
         const ZSTD_dictAttachPref_e pref = (ZSTD_dictAttachPref_e)value;
-        BOUNDCHECK(ZSTD_c_forceAttachDict, pref);
+        BOUNDCHECK(ZSTD_c_forceAttachDict, (int)pref);
         CCtxParams->attachDictPref = pref;
         return CCtxParams->attachDictPref;
     }
 
     case ZSTD_c_literalCompressionMode : {
-        const ZSTD_literalCompressionMode_e lcm = (ZSTD_literalCompressionMode_e)value;
-        BOUNDCHECK(ZSTD_c_literalCompressionMode, lcm);
+        const ZSTD_paramSwitch_e lcm = (ZSTD_paramSwitch_e)value;
+        BOUNDCHECK(ZSTD_c_literalCompressionMode, (int)lcm);
         CCtxParams->literalCompressionMode = lcm;
         return CCtxParams->literalCompressionMode;
     }
@@ -667,57 +904,118 @@ size_t ZSTD_CCtxParams_setParameter(ZSTD_CCtx_params* CCtxParams,
         return CCtxParams->rsyncable;
 #endif
 
+    case ZSTD_c_enableDedicatedDictSearch :
+        CCtxParams->enableDedicatedDictSearch = (value!=0);
+        return (size_t)CCtxParams->enableDedicatedDictSearch;
+
     case ZSTD_c_enableLongDistanceMatching :
-        CCtxParams->ldmParams.enableLdm = (value!=0);
+        BOUNDCHECK(ZSTD_c_enableLongDistanceMatching, value);
+        CCtxParams->ldmParams.enableLdm = (ZSTD_paramSwitch_e)value;
         return CCtxParams->ldmParams.enableLdm;
 
     case ZSTD_c_ldmHashLog :
         if (value!=0)   /* 0 ==> auto */
             BOUNDCHECK(ZSTD_c_ldmHashLog, value);
-        CCtxParams->ldmParams.hashLog = value;
+        CCtxParams->ldmParams.hashLog = (U32)value;
         return CCtxParams->ldmParams.hashLog;
 
     case ZSTD_c_ldmMinMatch :
         if (value!=0)   /* 0 ==> default */
             BOUNDCHECK(ZSTD_c_ldmMinMatch, value);
-        CCtxParams->ldmParams.minMatchLength = value;
+        CCtxParams->ldmParams.minMatchLength = (U32)value;
         return CCtxParams->ldmParams.minMatchLength;
 
     case ZSTD_c_ldmBucketSizeLog :
         if (value!=0)   /* 0 ==> default */
             BOUNDCHECK(ZSTD_c_ldmBucketSizeLog, value);
-        CCtxParams->ldmParams.bucketSizeLog = value;
+        CCtxParams->ldmParams.bucketSizeLog = (U32)value;
         return CCtxParams->ldmParams.bucketSizeLog;
 
     case ZSTD_c_ldmHashRateLog :
-        RETURN_ERROR_IF(value > ZSTD_WINDOWLOG_MAX - ZSTD_HASHLOG_MIN,
-                        parameter_outOfBound, "Param out of bounds!");
-        CCtxParams->ldmParams.hashRateLog = value;
+        if (value!=0)   /* 0 ==> default */
+            BOUNDCHECK(ZSTD_c_ldmHashRateLog, value);
+        CCtxParams->ldmParams.hashRateLog = (U32)value;
         return CCtxParams->ldmParams.hashRateLog;
 
     case ZSTD_c_targetCBlockSize :
         if (value!=0)   /* 0 ==> default */
             BOUNDCHECK(ZSTD_c_targetCBlockSize, value);
-        CCtxParams->targetCBlockSize = value;
+        CCtxParams->targetCBlockSize = (U32)value;
         return CCtxParams->targetCBlockSize;
 
     case ZSTD_c_srcSizeHint :
         if (value!=0)    /* 0 ==> default */
             BOUNDCHECK(ZSTD_c_srcSizeHint, value);
         CCtxParams->srcSizeHint = value;
-        return CCtxParams->srcSizeHint;
+        return (size_t)CCtxParams->srcSizeHint;
+
+    case ZSTD_c_stableInBuffer:
+        BOUNDCHECK(ZSTD_c_stableInBuffer, value);
+        CCtxParams->inBufferMode = (ZSTD_bufferMode_e)value;
+        return CCtxParams->inBufferMode;
+
+    case ZSTD_c_stableOutBuffer:
+        BOUNDCHECK(ZSTD_c_stableOutBuffer, value);
+        CCtxParams->outBufferMode = (ZSTD_bufferMode_e)value;
+        return CCtxParams->outBufferMode;
+
+    case ZSTD_c_blockDelimiters:
+        BOUNDCHECK(ZSTD_c_blockDelimiters, value);
+        CCtxParams->blockDelimiters = (ZSTD_sequenceFormat_e)value;
+        return CCtxParams->blockDelimiters;
+
+    case ZSTD_c_validateSequences:
+        BOUNDCHECK(ZSTD_c_validateSequences, value);
+        CCtxParams->validateSequences = value;
+        return CCtxParams->validateSequences;
+
+    case ZSTD_c_useBlockSplitter:
+        BOUNDCHECK(ZSTD_c_useBlockSplitter, value);
+        CCtxParams->useBlockSplitter = (ZSTD_paramSwitch_e)value;
+        return CCtxParams->useBlockSplitter;
+
+    case ZSTD_c_useRowMatchFinder:
+        BOUNDCHECK(ZSTD_c_useRowMatchFinder, value);
+        CCtxParams->useRowMatchFinder = (ZSTD_paramSwitch_e)value;
+        return CCtxParams->useRowMatchFinder;
+
+    case ZSTD_c_deterministicRefPrefix:
+        BOUNDCHECK(ZSTD_c_deterministicRefPrefix, value);
+        CCtxParams->deterministicRefPrefix = !!value;
+        return CCtxParams->deterministicRefPrefix;
+
+    case ZSTD_c_prefetchCDictTables:
+        BOUNDCHECK(ZSTD_c_prefetchCDictTables, value);
+        CCtxParams->prefetchCDictTables = (ZSTD_paramSwitch_e)value;
+        return CCtxParams->prefetchCDictTables;
+
+    case ZSTD_c_enableSeqProducerFallback:
+        BOUNDCHECK(ZSTD_c_enableSeqProducerFallback, value);
+        CCtxParams->enableMatchFinderFallback = value;
+        return CCtxParams->enableMatchFinderFallback;
+
+    case ZSTD_c_maxBlockSize:
+        if (value!=0)    /* 0 ==> default */
+            BOUNDCHECK(ZSTD_c_maxBlockSize, value);
+        CCtxParams->maxBlockSize = value;
+        return CCtxParams->maxBlockSize;
+
+    case ZSTD_c_searchForExternalRepcodes:
+        BOUNDCHECK(ZSTD_c_searchForExternalRepcodes, value);
+        CCtxParams->searchForExternalRepcodes = (ZSTD_paramSwitch_e)value;
+        return CCtxParams->searchForExternalRepcodes;
 
     default: RETURN_ERROR(parameter_unsupported, "unknown parameter");
     }
 }
 
-size_t ZSTD_CCtx_getParameter(ZSTD_CCtx* cctx, ZSTD_cParameter param, int* value)
+size_t ZSTD_CCtx_getParameter(ZSTD_CCtx const* cctx, ZSTD_cParameter param, int* value)
 {
     return ZSTD_CCtxParams_getParameter(&cctx->requestedParams, param, value);
 }
 
 size_t ZSTD_CCtxParams_getParameter(
-        ZSTD_CCtx_params* CCtxParams, ZSTD_cParameter param, int* value)
+        ZSTD_CCtx_params const* CCtxParams, ZSTD_cParameter param, int* value)
 {
     switch(param)
     {
@@ -794,6 +1092,9 @@ size_t ZSTD_CCtxParams_getParameter(
         *value = CCtxParams->rsyncable;
         break;
 #endif
+    case ZSTD_c_enableDedicatedDictSearch :
+        *value = CCtxParams->enableDedicatedDictSearch;
+        break;
     case ZSTD_c_enableLongDistanceMatching :
         *value = CCtxParams->ldmParams.enableLdm;
         break;
@@ -815,6 +1116,39 @@ size_t ZSTD_CCtxParams_getParameter(
     case ZSTD_c_srcSizeHint :
         *value = (int)CCtxParams->srcSizeHint;
         break;
+    case ZSTD_c_stableInBuffer :
+        *value = (int)CCtxParams->inBufferMode;
+        break;
+    case ZSTD_c_stableOutBuffer :
+        *value = (int)CCtxParams->outBufferMode;
+        break;
+    case ZSTD_c_blockDelimiters :
+        *value = (int)CCtxParams->blockDelimiters;
+        break;
+    case ZSTD_c_validateSequences :
+        *value = (int)CCtxParams->validateSequences;
+        break;
+    case ZSTD_c_useBlockSplitter :
+        *value = (int)CCtxParams->useBlockSplitter;
+        break;
+    case ZSTD_c_useRowMatchFinder :
+        *value = (int)CCtxParams->useRowMatchFinder;
+        break;
+    case ZSTD_c_deterministicRefPrefix:
+        *value = (int)CCtxParams->deterministicRefPrefix;
+        break;
+    case ZSTD_c_prefetchCDictTables:
+        *value = (int)CCtxParams->prefetchCDictTables;
+        break;
+    case ZSTD_c_enableSeqProducerFallback:
+        *value = CCtxParams->enableMatchFinderFallback;
+        break;
+    case ZSTD_c_maxBlockSize:
+        *value = (int)CCtxParams->maxBlockSize;
+        break;
+    case ZSTD_c_searchForExternalRepcodes:
+        *value = (int)CCtxParams->searchForExternalRepcodes;
+        break;
     default: RETURN_ERROR(parameter_unsupported, "unknown parameter");
     }
     return 0;
@@ -841,15 +1175,38 @@ size_t ZSTD_CCtx_setParametersUsingCCtxParams(
     return 0;
 }
 
-ZSTDLIB_API size_t ZSTD_CCtx_setPledgedSrcSize(ZSTD_CCtx* cctx, unsigned long long pledgedSrcSize)
+size_t ZSTD_CCtx_setCParams(ZSTD_CCtx* cctx, ZSTD_compressionParameters cparams)
 {
-    DEBUGLOG(4, "ZSTD_CCtx_setPledgedSrcSize to %u bytes", (U32)pledgedSrcSize);
+    DEBUGLOG(4, "ZSTD_CCtx_setCParams");
+    assert(cctx != NULL);
+    if (cctx->streamStage != zcss_init) {
+        /* All parameters in @cparams are allowed to be updated during MT compression.
+         * This must be signaled, so that MT compression picks up the changes */
+        cctx->cParamsChanged = 1;
+    }
+    /* only update if parameters are valid */
+    FORWARD_IF_ERROR(ZSTD_checkCParams(cparams), "");
+    cctx->requestedParams.cParams = cparams;
+    return 0;
+}
+
+size_t ZSTD_CCtx_setPledgedSrcSize(ZSTD_CCtx* cctx, unsigned long long pledgedSrcSize)
+{
+    DEBUGLOG(4, "ZSTD_CCtx_setPledgedSrcSize to %llu bytes", pledgedSrcSize);
     RETURN_ERROR_IF(cctx->streamStage != zcss_init, stage_wrong,
                     "Can't set pledgedSrcSize when not in init stage.");
     cctx->pledgedSrcSizePlusOne = pledgedSrcSize+1;
     return 0;
 }
 
+static ZSTD_compressionParameters ZSTD_dedicatedDictSearch_getCParams(
+        int const compressionLevel,
+        size_t const dictSize);
+static int ZSTD_dedicatedDictSearch_isSupported(
+        const ZSTD_compressionParameters* cParams);
+static void ZSTD_dedicatedDictSearch_revertCParams(
+        ZSTD_compressionParameters* cParams);
+
 /**
  * Initializes the local dict using the requested parameters.
  * NOTE: This does not use the pledged src size, because it may be used for more
@@ -858,8 +1215,6 @@ ZSTDLIB_API size_t ZSTD_CCtx_setPledgedSrcSize(ZSTD_CCtx* cctx, unsigned long lo
 static size_t ZSTD_initLocalDict(ZSTD_CCtx* cctx)
 {
     ZSTD_localDict* const dl = &cctx->localDict;
-    ZSTD_compressionParameters const cParams = ZSTD_getCParamsFromCCtxParams(
-            &cctx->requestedParams, ZSTD_CONTENTSIZE_UNKNOWN, dl->dictSize);
     if (dl->dict == NULL) {
         /* No local dictionary. */
         assert(dl->dictBuffer == NULL);
@@ -876,12 +1231,12 @@ static size_t ZSTD_initLocalDict(ZSTD_CCtx* cctx)
     assert(cctx->cdict == NULL);
     assert(cctx->prefixDict.dict == NULL);
 
-    dl->cdict = ZSTD_createCDict_advanced(
+    dl->cdict = ZSTD_createCDict_advanced2(
             dl->dict,
             dl->dictSize,
             ZSTD_dlm_byRef,
             dl->dictContentType,
-            cParams,
+            &cctx->requestedParams,
             cctx->customMem);
     RETURN_ERROR_IF(!dl->cdict, memory_allocation, "ZSTD_createCDict_advanced failed");
     cctx->cdict = dl->cdict;
@@ -894,8 +1249,6 @@ size_t ZSTD_CCtx_loadDictionary_advanced(
 {
     RETURN_ERROR_IF(cctx->streamStage != zcss_init, stage_wrong,
                     "Can't load a dictionary when ctx is not in init stage.");
-    RETURN_ERROR_IF(cctx->staticSize, memory_allocation,
-                    "no malloc for static CCtx");
     DEBUGLOG(4, "ZSTD_CCtx_loadDictionary_advanced (size: %u)", (U32)dictSize);
     ZSTD_clearAllDicts(cctx);  /* in case one already exists */
     if (dict == NULL || dictSize == 0)  /* no dictionary mode */
@@ -903,9 +1256,12 @@ size_t ZSTD_CCtx_loadDictionary_advanced(
     if (dictLoadMethod == ZSTD_dlm_byRef) {
         cctx->localDict.dict = dict;
     } else {
-        void* dictBuffer = ZSTD_malloc(dictSize, cctx->customMem);
+        void* dictBuffer;
+        RETURN_ERROR_IF(cctx->staticSize, memory_allocation,
+                        "no malloc for static CCtx");
+        dictBuffer = ZSTD_customMalloc(dictSize, cctx->customMem);
         RETURN_ERROR_IF(!dictBuffer, memory_allocation, "NULL pointer!");
-        memcpy(dictBuffer, dict, dictSize);
+        ZSTD_memcpy(dictBuffer, dict, dictSize);
         cctx->localDict.dictBuffer = dictBuffer;
         cctx->localDict.dict = dictBuffer;
     }
@@ -914,14 +1270,14 @@ size_t ZSTD_CCtx_loadDictionary_advanced(
     return 0;
 }
 
-ZSTDLIB_API size_t ZSTD_CCtx_loadDictionary_byReference(
+size_t ZSTD_CCtx_loadDictionary_byReference(
       ZSTD_CCtx* cctx, const void* dict, size_t dictSize)
 {
     return ZSTD_CCtx_loadDictionary_advanced(
             cctx, dict, dictSize, ZSTD_dlm_byRef, ZSTD_dct_auto);
 }
 
-ZSTDLIB_API size_t ZSTD_CCtx_loadDictionary(ZSTD_CCtx* cctx, const void* dict, size_t dictSize)
+size_t ZSTD_CCtx_loadDictionary(ZSTD_CCtx* cctx, const void* dict, size_t dictSize)
 {
     return ZSTD_CCtx_loadDictionary_advanced(
             cctx, dict, dictSize, ZSTD_dlm_byCopy, ZSTD_dct_auto);
@@ -938,6 +1294,14 @@ size_t ZSTD_CCtx_refCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict)
     return 0;
 }
 
+size_t ZSTD_CCtx_refThreadPool(ZSTD_CCtx* cctx, ZSTD_threadPool* pool)
+{
+    RETURN_ERROR_IF(cctx->streamStage != zcss_init, stage_wrong,
+                    "Can't ref a pool when ctx not in init stage.");
+    cctx->pool = pool;
+    return 0;
+}
+
 size_t ZSTD_CCtx_refPrefix(ZSTD_CCtx* cctx, const void* prefix, size_t prefixSize)
 {
     return ZSTD_CCtx_refPrefix_advanced(cctx, prefix, prefixSize, ZSTD_dct_rawContent);
@@ -971,6 +1335,7 @@ size_t ZSTD_CCtx_reset(ZSTD_CCtx* cctx, ZSTD_ResetDirective reset)
         RETURN_ERROR_IF(cctx->streamStage != zcss_init, stage_wrong,
                         "Can't reset parameters only when not in init stage.");
         ZSTD_clearAllDicts(cctx);
+        ZSTD_memset(&cctx->externalMatchCtx, 0, sizeof(cctx->externalMatchCtx));
         return ZSTD_CCtxParams_reset(&cctx->requestedParams);
     }
     return 0;
@@ -1022,42 +1387,141 @@ U32 ZSTD_cycleLog(U32 hashLog, ZSTD_strategy strat)
     return hashLog - btScale;
 }
 
+/** ZSTD_dictAndWindowLog() :
+ * Returns an adjusted window log that is large enough to fit the source and the dictionary.
+ * The zstd format says that the entire dictionary is valid if one byte of the dictionary
+ * is within the window. So the hashLog and chainLog should be large enough to reference both
+ * the dictionary and the window. So we must use this adjusted dictAndWindowLog when downsizing
+ * the hashLog and windowLog.
+ * NOTE: srcSize must not be ZSTD_CONTENTSIZE_UNKNOWN.
+ */
+static U32 ZSTD_dictAndWindowLog(U32 windowLog, U64 srcSize, U64 dictSize)
+{
+    const U64 maxWindowSize = 1ULL << ZSTD_WINDOWLOG_MAX;
+    /* No dictionary ==> No change */
+    if (dictSize == 0) {
+        return windowLog;
+    }
+    assert(windowLog <= ZSTD_WINDOWLOG_MAX);
+    assert(srcSize != ZSTD_CONTENTSIZE_UNKNOWN); /* Handled in ZSTD_adjustCParams_internal() */
+    {
+        U64 const windowSize = 1ULL << windowLog;
+        U64 const dictAndWindowSize = dictSize + windowSize;
+        /* If the window size is already large enough to fit both the source and the dictionary
+         * then just use the window size. Otherwise adjust so that it fits the dictionary and
+         * the window.
+         */
+        if (windowSize >= dictSize + srcSize) {
+            return windowLog; /* Window size large enough already */
+        } else if (dictAndWindowSize >= maxWindowSize) {
+            return ZSTD_WINDOWLOG_MAX; /* Larger than max window log */
+        } else  {
+            return ZSTD_highbit32((U32)dictAndWindowSize - 1) + 1;
+        }
+    }
+}
+
 /** ZSTD_adjustCParams_internal() :
  *  optimize `cPar` for a specified input (`srcSize` and `dictSize`).
  *  mostly downsize to reduce memory consumption and initialization latency.
  * `srcSize` can be ZSTD_CONTENTSIZE_UNKNOWN when not known.
+ * `mode` is the mode for parameter adjustment. See docs for `ZSTD_cParamMode_e`.
  *  note : `srcSize==0` means 0!
  *  condition : cPar is presumed validated (can be checked using ZSTD_checkCParams()). */
 static ZSTD_compressionParameters
 ZSTD_adjustCParams_internal(ZSTD_compressionParameters cPar,
                             unsigned long long srcSize,
-                            size_t dictSize)
+                            size_t dictSize,
+                            ZSTD_cParamMode_e mode,
+                            ZSTD_paramSwitch_e useRowMatchFinder)
 {
-    static const U64 minSrcSize = 513; /* (1<<9) + 1 */
-    static const U64 maxWindowResize = 1ULL << (ZSTD_WINDOWLOG_MAX-1);
+    const U64 minSrcSize = 513; /* (1<<9) + 1 */
+    const U64 maxWindowResize = 1ULL << (ZSTD_WINDOWLOG_MAX-1);
     assert(ZSTD_checkCParams(cPar)==0);
 
-    if (dictSize && srcSize == ZSTD_CONTENTSIZE_UNKNOWN)
-        srcSize = minSrcSize;
+    switch (mode) {
+    case ZSTD_cpm_unknown:
+    case ZSTD_cpm_noAttachDict:
+        /* If we don't know the source size, don't make any
+         * assumptions about it. We will already have selected
+         * smaller parameters if a dictionary is in use.
+         */
+        break;
+    case ZSTD_cpm_createCDict:
+        /* Assume a small source size when creating a dictionary
+         * with an unknown source size.
+         */
+        if (dictSize && srcSize == ZSTD_CONTENTSIZE_UNKNOWN)
+            srcSize = minSrcSize;
+        break;
+    case ZSTD_cpm_attachDict:
+        /* Dictionary has its own dedicated parameters which have
+         * already been selected. We are selecting parameters
+         * for only the source.
+         */
+        dictSize = 0;
+        break;
+    default:
+        assert(0);
+        break;
+    }
 
     /* resize windowLog if input is small enough, to use less memory */
-    if ( (srcSize < maxWindowResize)
-      && (dictSize < maxWindowResize) )  {
+    if ( (srcSize <= maxWindowResize)
+      && (dictSize <= maxWindowResize) )  {
         U32 const tSize = (U32)(srcSize + dictSize);
         static U32 const hashSizeMin = 1 << ZSTD_HASHLOG_MIN;
         U32 const srcLog = (tSize < hashSizeMin) ? ZSTD_HASHLOG_MIN :
                             ZSTD_highbit32(tSize-1) + 1;
         if (cPar.windowLog > srcLog) cPar.windowLog = srcLog;
     }
-    if (cPar.hashLog > cPar.windowLog+1) cPar.hashLog = cPar.windowLog+1;
-    {   U32 const cycleLog = ZSTD_cycleLog(cPar.chainLog, cPar.strategy);
-        if (cycleLog > cPar.windowLog)
-            cPar.chainLog -= (cycleLog - cPar.windowLog);
+    if (srcSize != ZSTD_CONTENTSIZE_UNKNOWN) {
+        U32 const dictAndWindowLog = ZSTD_dictAndWindowLog(cPar.windowLog, (U64)srcSize, (U64)dictSize);
+        U32 const cycleLog = ZSTD_cycleLog(cPar.chainLog, cPar.strategy);
+        if (cPar.hashLog > dictAndWindowLog+1) cPar.hashLog = dictAndWindowLog+1;
+        if (cycleLog > dictAndWindowLog)
+            cPar.chainLog -= (cycleLog - dictAndWindowLog);
     }
 
     if (cPar.windowLog < ZSTD_WINDOWLOG_ABSOLUTEMIN)
         cPar.windowLog = ZSTD_WINDOWLOG_ABSOLUTEMIN;  /* minimum wlog required for valid frame header */
 
+    /* We can't use more than 32 bits of hash in total, so that means that we require:
+     * (hashLog + 8) <= 32 && (chainLog + 8) <= 32
+     */
+    if (mode == ZSTD_cpm_createCDict && ZSTD_CDictIndicesAreTagged(&cPar)) {
+        U32 const maxShortCacheHashLog = 32 - ZSTD_SHORT_CACHE_TAG_BITS;
+        if (cPar.hashLog > maxShortCacheHashLog) {
+            cPar.hashLog = maxShortCacheHashLog;
+        }
+        if (cPar.chainLog > maxShortCacheHashLog) {
+            cPar.chainLog = maxShortCacheHashLog;
+        }
+    }
+
+
+    /* At this point, we aren't 100% sure if we are using the row match finder.
+     * Unless it is explicitly disabled, conservatively assume that it is enabled.
+     * In this case it will only be disabled for small sources, so shrinking the
+     * hash log a little bit shouldn't result in any ratio loss.
+     */
+    if (useRowMatchFinder == ZSTD_ps_auto)
+        useRowMatchFinder = ZSTD_ps_enable;
+
+    /* We can't hash more than 32-bits in total. So that means that we require:
+     * (hashLog - rowLog + 8) <= 32
+     */
+    if (ZSTD_rowMatchFinderUsed(cPar.strategy, useRowMatchFinder)) {
+        /* Switch to 32-entry rows if searchLog is 5 (or more) */
+        U32 const rowLog = BOUNDED(4, cPar.searchLog, 6);
+        U32 const maxRowHashLog = 32 - ZSTD_ROW_HASH_TAG_BITS;
+        U32 const maxHashLog = maxRowHashLog + rowLog;
+        assert(cPar.hashLog >= rowLog);
+        if (cPar.hashLog > maxHashLog) {
+            cPar.hashLog = maxHashLog;
+        }
+    }
+
     return cPar;
 }
 
@@ -1068,38 +1532,50 @@ ZSTD_adjustCParams(ZSTD_compressionParameters cPar,
 {
     cPar = ZSTD_clampCParams(cPar);   /* resulting cPar is necessarily valid (all parameters within range) */
     if (srcSize == 0) srcSize = ZSTD_CONTENTSIZE_UNKNOWN;
-    return ZSTD_adjustCParams_internal(cPar, srcSize, dictSize);
+    return ZSTD_adjustCParams_internal(cPar, srcSize, dictSize, ZSTD_cpm_unknown, ZSTD_ps_auto);
 }
 
-static ZSTD_compressionParameters ZSTD_getCParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize);
-static ZSTD_parameters ZSTD_getParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize);
+static ZSTD_compressionParameters ZSTD_getCParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode);
+static ZSTD_parameters ZSTD_getParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode);
+
+static void ZSTD_overrideCParams(
+              ZSTD_compressionParameters* cParams,
+        const ZSTD_compressionParameters* overrides)
+{
+    if (overrides->windowLog)    cParams->windowLog    = overrides->windowLog;
+    if (overrides->hashLog)      cParams->hashLog      = overrides->hashLog;
+    if (overrides->chainLog)     cParams->chainLog     = overrides->chainLog;
+    if (overrides->searchLog)    cParams->searchLog    = overrides->searchLog;
+    if (overrides->minMatch)     cParams->minMatch     = overrides->minMatch;
+    if (overrides->targetLength) cParams->targetLength = overrides->targetLength;
+    if (overrides->strategy)     cParams->strategy     = overrides->strategy;
+}
 
 ZSTD_compressionParameters ZSTD_getCParamsFromCCtxParams(
-        const ZSTD_CCtx_params* CCtxParams, U64 srcSizeHint, size_t dictSize)
+        const ZSTD_CCtx_params* CCtxParams, U64 srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode)
 {
     ZSTD_compressionParameters cParams;
     if (srcSizeHint == ZSTD_CONTENTSIZE_UNKNOWN && CCtxParams->srcSizeHint > 0) {
       srcSizeHint = CCtxParams->srcSizeHint;
     }
-    cParams = ZSTD_getCParams_internal(CCtxParams->compressionLevel, srcSizeHint, dictSize);
-    if (CCtxParams->ldmParams.enableLdm) cParams.windowLog = ZSTD_LDM_DEFAULT_WINDOW_LOG;
-    if (CCtxParams->cParams.windowLog) cParams.windowLog = CCtxParams->cParams.windowLog;
-    if (CCtxParams->cParams.hashLog) cParams.hashLog = CCtxParams->cParams.hashLog;
-    if (CCtxParams->cParams.chainLog) cParams.chainLog = CCtxParams->cParams.chainLog;
-    if (CCtxParams->cParams.searchLog) cParams.searchLog = CCtxParams->cParams.searchLog;
-    if (CCtxParams->cParams.minMatch) cParams.minMatch = CCtxParams->cParams.minMatch;
-    if (CCtxParams->cParams.targetLength) cParams.targetLength = CCtxParams->cParams.targetLength;
-    if (CCtxParams->cParams.strategy) cParams.strategy = CCtxParams->cParams.strategy;
+    cParams = ZSTD_getCParams_internal(CCtxParams->compressionLevel, srcSizeHint, dictSize, mode);
+    if (CCtxParams->ldmParams.enableLdm == ZSTD_ps_enable) cParams.windowLog = ZSTD_LDM_DEFAULT_WINDOW_LOG;
+    ZSTD_overrideCParams(&cParams, &CCtxParams->cParams);
     assert(!ZSTD_checkCParams(cParams));
     /* srcSizeHint == 0 means 0 */
-    return ZSTD_adjustCParams_internal(cParams, srcSizeHint, dictSize);
+    return ZSTD_adjustCParams_internal(cParams, srcSizeHint, dictSize, mode, CCtxParams->useRowMatchFinder);
 }
 
 static size_t
 ZSTD_sizeof_matchState(const ZSTD_compressionParameters* const cParams,
+                       const ZSTD_paramSwitch_e useRowMatchFinder,
+                       const U32 enableDedicatedDictSearch,
                        const U32 forCCtx)
 {
-    size_t const chainSize = (cParams->strategy == ZSTD_fast) ? 0 : ((size_t)1 << cParams->chainLog);
+    /* chain table size should be 0 for fast or row-hash strategies */
+    size_t const chainSize = ZSTD_allocateChainTable(cParams->strategy, useRowMatchFinder, enableDedicatedDictSearch && !forCCtx)
+                                ? ((size_t)1 << cParams->chainLog)
+                                : 0;
     size_t const hSize = ((size_t)1) << cParams->hashLog;
     U32    const hashLog3 = (forCCtx && cParams->minMatch==3) ? MIN(ZSTD_HASHLOG3_MAX, cParams->windowLog) : 0;
     size_t const h3Size = hashLog3 ? ((size_t)1) << hashLog3 : 0;
@@ -1109,71 +1585,131 @@ ZSTD_sizeof_matchState(const ZSTD_compressionParameters* const cParams,
                             + hSize * sizeof(U32)
                             + h3Size * sizeof(U32);
     size_t const optPotentialSpace =
-        ZSTD_cwksp_alloc_size((MaxML+1) * sizeof(U32))
-      + ZSTD_cwksp_alloc_size((MaxLL+1) * sizeof(U32))
-      + ZSTD_cwksp_alloc_size((MaxOff+1) * sizeof(U32))
-      + ZSTD_cwksp_alloc_size((1<<Litbits) * sizeof(U32))
-      + ZSTD_cwksp_alloc_size((ZSTD_OPT_NUM+1) * sizeof(ZSTD_match_t))
-      + ZSTD_cwksp_alloc_size((ZSTD_OPT_NUM+1) * sizeof(ZSTD_optimal_t));
+        ZSTD_cwksp_aligned_alloc_size((MaxML+1) * sizeof(U32))
+      + ZSTD_cwksp_aligned_alloc_size((MaxLL+1) * sizeof(U32))
+      + ZSTD_cwksp_aligned_alloc_size((MaxOff+1) * sizeof(U32))
+      + ZSTD_cwksp_aligned_alloc_size((1<<Litbits) * sizeof(U32))
+      + ZSTD_cwksp_aligned_alloc_size((ZSTD_OPT_NUM+1) * sizeof(ZSTD_match_t))
+      + ZSTD_cwksp_aligned_alloc_size((ZSTD_OPT_NUM+1) * sizeof(ZSTD_optimal_t));
+    size_t const lazyAdditionalSpace = ZSTD_rowMatchFinderUsed(cParams->strategy, useRowMatchFinder)
+                                            ? ZSTD_cwksp_aligned_alloc_size(hSize*sizeof(U16))
+                                            : 0;
     size_t const optSpace = (forCCtx && (cParams->strategy >= ZSTD_btopt))
                                 ? optPotentialSpace
                                 : 0;
+    size_t const slackSpace = ZSTD_cwksp_slack_space_required();
+
+    /* tables are guaranteed to be sized in multiples of 64 bytes (or 16 uint32_t) */
+    ZSTD_STATIC_ASSERT(ZSTD_HASHLOG_MIN >= 4 && ZSTD_WINDOWLOG_MIN >= 4 && ZSTD_CHAINLOG_MIN >= 4);
+    assert(useRowMatchFinder != ZSTD_ps_auto);
+
     DEBUGLOG(4, "chainSize: %u - hSize: %u - h3Size: %u",
                 (U32)chainSize, (U32)hSize, (U32)h3Size);
-    return tableSpace + optSpace;
+    return tableSpace + optSpace + slackSpace + lazyAdditionalSpace;
+}
+
+/* Helper function for calculating memory requirements.
+ * Gives a tighter bound than ZSTD_sequenceBound() by taking minMatch into account. */
+static size_t ZSTD_maxNbSeq(size_t blockSize, unsigned minMatch, int useSequenceProducer) {
+    U32 const divider = (minMatch==3 || useSequenceProducer) ? 3 : 4;
+    return blockSize / divider;
+}
+
+static size_t ZSTD_estimateCCtxSize_usingCCtxParams_internal(
+        const ZSTD_compressionParameters* cParams,
+        const ldmParams_t* ldmParams,
+        const int isStatic,
+        const ZSTD_paramSwitch_e useRowMatchFinder,
+        const size_t buffInSize,
+        const size_t buffOutSize,
+        const U64 pledgedSrcSize,
+        int useSequenceProducer,
+        size_t maxBlockSize)
+{
+    size_t const windowSize = (size_t) BOUNDED(1ULL, 1ULL << cParams->windowLog, pledgedSrcSize);
+    size_t const blockSize = MIN(ZSTD_resolveMaxBlockSize(maxBlockSize), windowSize);
+    size_t const maxNbSeq = ZSTD_maxNbSeq(blockSize, cParams->minMatch, useSequenceProducer);
+    size_t const tokenSpace = ZSTD_cwksp_alloc_size(WILDCOPY_OVERLENGTH + blockSize)
+                            + ZSTD_cwksp_aligned_alloc_size(maxNbSeq * sizeof(seqDef))
+                            + 3 * ZSTD_cwksp_alloc_size(maxNbSeq * sizeof(BYTE));
+    size_t const entropySpace = ZSTD_cwksp_alloc_size(ENTROPY_WORKSPACE_SIZE);
+    size_t const blockStateSpace = 2 * ZSTD_cwksp_alloc_size(sizeof(ZSTD_compressedBlockState_t));
+    size_t const matchStateSize = ZSTD_sizeof_matchState(cParams, useRowMatchFinder, /* enableDedicatedDictSearch */ 0, /* forCCtx */ 1);
+
+    size_t const ldmSpace = ZSTD_ldm_getTableSize(*ldmParams);
+    size_t const maxNbLdmSeq = ZSTD_ldm_getMaxNbSeq(*ldmParams, blockSize);
+    size_t const ldmSeqSpace = ldmParams->enableLdm == ZSTD_ps_enable ?
+        ZSTD_cwksp_aligned_alloc_size(maxNbLdmSeq * sizeof(rawSeq)) : 0;
+
+
+    size_t const bufferSpace = ZSTD_cwksp_alloc_size(buffInSize)
+                             + ZSTD_cwksp_alloc_size(buffOutSize);
+
+    size_t const cctxSpace = isStatic ? ZSTD_cwksp_alloc_size(sizeof(ZSTD_CCtx)) : 0;
+
+    size_t const maxNbExternalSeq = ZSTD_sequenceBound(blockSize);
+    size_t const externalSeqSpace = useSequenceProducer
+        ? ZSTD_cwksp_aligned_alloc_size(maxNbExternalSeq * sizeof(ZSTD_Sequence))
+        : 0;
+
+    size_t const neededSpace =
+        cctxSpace +
+        entropySpace +
+        blockStateSpace +
+        ldmSpace +
+        ldmSeqSpace +
+        matchStateSize +
+        tokenSpace +
+        bufferSpace +
+        externalSeqSpace;
+
+    DEBUGLOG(5, "estimate workspace : %u", (U32)neededSpace);
+    return neededSpace;
 }
 
 size_t ZSTD_estimateCCtxSize_usingCCtxParams(const ZSTD_CCtx_params* params)
 {
-    RETURN_ERROR_IF(params->nbWorkers > 0, GENERIC, "Estimate CCtx size is supported for single-threaded compression only.");
-    {   ZSTD_compressionParameters const cParams =
-                ZSTD_getCParamsFromCCtxParams(params, ZSTD_CONTENTSIZE_UNKNOWN, 0);
-        size_t const blockSize = MIN(ZSTD_BLOCKSIZE_MAX, (size_t)1 << cParams.windowLog);
-        U32    const divider = (cParams.minMatch==3) ? 3 : 4;
-        size_t const maxNbSeq = blockSize / divider;
-        size_t const tokenSpace = ZSTD_cwksp_alloc_size(WILDCOPY_OVERLENGTH + blockSize)
-                                + ZSTD_cwksp_alloc_size(maxNbSeq * sizeof(seqDef))
-                                + 3 * ZSTD_cwksp_alloc_size(maxNbSeq * sizeof(BYTE));
-        size_t const entropySpace = ZSTD_cwksp_alloc_size(HUF_WORKSPACE_SIZE);
-        size_t const blockStateSpace = 2 * ZSTD_cwksp_alloc_size(sizeof(ZSTD_compressedBlockState_t));
-        size_t const matchStateSize = ZSTD_sizeof_matchState(&cParams, /* forCCtx */ 1);
-
-        size_t const ldmSpace = ZSTD_ldm_getTableSize(params->ldmParams);
-        size_t const ldmSeqSpace = ZSTD_cwksp_alloc_size(ZSTD_ldm_getMaxNbSeq(params->ldmParams, blockSize) * sizeof(rawSeq));
-
-        /* estimateCCtxSize is for one-shot compression. So no buffers should
-         * be needed. However, we still allocate two 0-sized buffers, which can
-         * take space under ASAN. */
-        size_t const bufferSpace = ZSTD_cwksp_alloc_size(0)
-                                 + ZSTD_cwksp_alloc_size(0);
-
-        size_t const cctxSpace = ZSTD_cwksp_alloc_size(sizeof(ZSTD_CCtx));
+    ZSTD_compressionParameters const cParams =
+                ZSTD_getCParamsFromCCtxParams(params, ZSTD_CONTENTSIZE_UNKNOWN, 0, ZSTD_cpm_noAttachDict);
+    ZSTD_paramSwitch_e const useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(params->useRowMatchFinder,
+                                                                               &cParams);
 
-        size_t const neededSpace =
-            cctxSpace +
-            entropySpace +
-            blockStateSpace +
-            ldmSpace +
-            ldmSeqSpace +
-            matchStateSize +
-            tokenSpace +
-            bufferSpace;
-
-        DEBUGLOG(5, "estimate workspace : %u", (U32)neededSpace);
-        return neededSpace;
-    }
+    RETURN_ERROR_IF(params->nbWorkers > 0, GENERIC, "Estimate CCtx size is supported for single-threaded compression only.");
+    /* estimateCCtxSize is for one-shot compression. So no buffers should
+     * be needed. However, we still allocate two 0-sized buffers, which can
+     * take space under ASAN. */
+    return ZSTD_estimateCCtxSize_usingCCtxParams_internal(
+        &cParams, &params->ldmParams, 1, useRowMatchFinder, 0, 0, ZSTD_CONTENTSIZE_UNKNOWN, params->useSequenceProducer, params->maxBlockSize);
 }
 
 size_t ZSTD_estimateCCtxSize_usingCParams(ZSTD_compressionParameters cParams)
 {
-    ZSTD_CCtx_params const params = ZSTD_makeCCtxParamsFromCParams(cParams);
-    return ZSTD_estimateCCtxSize_usingCCtxParams(&params);
+    ZSTD_CCtx_params initialParams = ZSTD_makeCCtxParamsFromCParams(cParams);
+    if (ZSTD_rowMatchFinderSupported(cParams.strategy)) {
+        /* Pick bigger of not using and using row-based matchfinder for greedy and lazy strategies */
+        size_t noRowCCtxSize;
+        size_t rowCCtxSize;
+        initialParams.useRowMatchFinder = ZSTD_ps_disable;
+        noRowCCtxSize = ZSTD_estimateCCtxSize_usingCCtxParams(&initialParams);
+        initialParams.useRowMatchFinder = ZSTD_ps_enable;
+        rowCCtxSize = ZSTD_estimateCCtxSize_usingCCtxParams(&initialParams);
+        return MAX(noRowCCtxSize, rowCCtxSize);
+    } else {
+        return ZSTD_estimateCCtxSize_usingCCtxParams(&initialParams);
+    }
 }
 
 static size_t ZSTD_estimateCCtxSize_internal(int compressionLevel)
 {
-    ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, 0);
-    return ZSTD_estimateCCtxSize_usingCParams(cParams);
+    int tier = 0;
+    size_t largestSize = 0;
+    static const unsigned long long srcSizeTiers[4] = {16 KB, 128 KB, 256 KB, ZSTD_CONTENTSIZE_UNKNOWN};
+    for (; tier < 4; ++tier) {
+        /* Choose the set of cParams for a given level across all srcSizes that give the largest cctxSize */
+        ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, srcSizeTiers[tier], 0, ZSTD_cpm_noAttachDict);
+        largestSize = MAX(ZSTD_estimateCCtxSize_usingCParams(cParams), largestSize);
+    }
+    return largestSize;
 }
 
 size_t ZSTD_estimateCCtxSize(int compressionLevel)
@@ -1181,6 +1717,7 @@ size_t ZSTD_estimateCCtxSize(int compressionLevel)
     int level;
     size_t memBudget = 0;
     for (level=MIN(compressionLevel, 1); level<=compressionLevel; level++) {
+        /* Ensure monotonically increasing memory usage as compression level increases */
         size_t const newMB = ZSTD_estimateCCtxSize_internal(level);
         if (newMB > memBudget) memBudget = newMB;
     }
@@ -1191,27 +1728,42 @@ size_t ZSTD_estimateCStreamSize_usingCCtxParams(const ZSTD_CCtx_params* params)
 {
     RETURN_ERROR_IF(params->nbWorkers > 0, GENERIC, "Estimate CCtx size is supported for single-threaded compression only.");
     {   ZSTD_compressionParameters const cParams =
-                ZSTD_getCParamsFromCCtxParams(params, ZSTD_CONTENTSIZE_UNKNOWN, 0);
-        size_t const CCtxSize = ZSTD_estimateCCtxSize_usingCCtxParams(params);
-        size_t const blockSize = MIN(ZSTD_BLOCKSIZE_MAX, (size_t)1 << cParams.windowLog);
-        size_t const inBuffSize = ((size_t)1 << cParams.windowLog) + blockSize;
-        size_t const outBuffSize = ZSTD_compressBound(blockSize) + 1;
-        size_t const streamingSize = ZSTD_cwksp_alloc_size(inBuffSize)
-                                   + ZSTD_cwksp_alloc_size(outBuffSize);
-
-        return CCtxSize + streamingSize;
+                ZSTD_getCParamsFromCCtxParams(params, ZSTD_CONTENTSIZE_UNKNOWN, 0, ZSTD_cpm_noAttachDict);
+        size_t const blockSize = MIN(ZSTD_resolveMaxBlockSize(params->maxBlockSize), (size_t)1 << cParams.windowLog);
+        size_t const inBuffSize = (params->inBufferMode == ZSTD_bm_buffered)
+                ? ((size_t)1 << cParams.windowLog) + blockSize
+                : 0;
+        size_t const outBuffSize = (params->outBufferMode == ZSTD_bm_buffered)
+                ? ZSTD_compressBound(blockSize) + 1
+                : 0;
+        ZSTD_paramSwitch_e const useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(params->useRowMatchFinder, &params->cParams);
+
+        return ZSTD_estimateCCtxSize_usingCCtxParams_internal(
+            &cParams, &params->ldmParams, 1, useRowMatchFinder, inBuffSize, outBuffSize,
+            ZSTD_CONTENTSIZE_UNKNOWN, params->useSequenceProducer, params->maxBlockSize);
     }
 }
 
 size_t ZSTD_estimateCStreamSize_usingCParams(ZSTD_compressionParameters cParams)
 {
-    ZSTD_CCtx_params const params = ZSTD_makeCCtxParamsFromCParams(cParams);
-    return ZSTD_estimateCStreamSize_usingCCtxParams(&params);
+    ZSTD_CCtx_params initialParams = ZSTD_makeCCtxParamsFromCParams(cParams);
+    if (ZSTD_rowMatchFinderSupported(cParams.strategy)) {
+        /* Pick bigger of not using and using row-based matchfinder for greedy and lazy strategies */
+        size_t noRowCCtxSize;
+        size_t rowCCtxSize;
+        initialParams.useRowMatchFinder = ZSTD_ps_disable;
+        noRowCCtxSize = ZSTD_estimateCStreamSize_usingCCtxParams(&initialParams);
+        initialParams.useRowMatchFinder = ZSTD_ps_enable;
+        rowCCtxSize = ZSTD_estimateCStreamSize_usingCCtxParams(&initialParams);
+        return MAX(noRowCCtxSize, rowCCtxSize);
+    } else {
+        return ZSTD_estimateCStreamSize_usingCCtxParams(&initialParams);
+    }
 }
 
 static size_t ZSTD_estimateCStreamSize_internal(int compressionLevel)
 {
-    ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, 0);
+    ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, 0, ZSTD_cpm_noAttachDict);
     return ZSTD_estimateCStreamSize_usingCParams(cParams);
 }
 
@@ -1305,16 +1857,6 @@ static void ZSTD_invalidateMatchState(ZSTD_matchState_t* ms)
 }
 
 /**
- * Indicates whether this compression proceeds directly from user-provided
- * source buffer to user-provided destination buffer (ZSTDb_not_buffered), or
- * whether the context needs to buffer the input/output (ZSTDb_buffered).
- */
-typedef enum {
-    ZSTDb_not_buffered,
-    ZSTDb_buffered
-} ZSTD_buffered_policy_e;
-
-/**
  * Controls, for this matchState reset, whether the tables need to be cleared /
  * prepared for the coming compression (ZSTDcrp_makeClean), or whether the
  * tables can be left unclean (ZSTDcrp_leaveDirty), because we know that a
@@ -1341,20 +1883,27 @@ typedef enum {
     ZSTD_resetTarget_CCtx
 } ZSTD_resetTarget_e;
 
+
 static size_t
 ZSTD_reset_matchState(ZSTD_matchState_t* ms,
                       ZSTD_cwksp* ws,
                 const ZSTD_compressionParameters* cParams,
+                const ZSTD_paramSwitch_e useRowMatchFinder,
                 const ZSTD_compResetPolicy_e crp,
                 const ZSTD_indexResetPolicy_e forceResetIndex,
                 const ZSTD_resetTarget_e forWho)
 {
-    size_t const chainSize = (cParams->strategy == ZSTD_fast) ? 0 : ((size_t)1 << cParams->chainLog);
+    /* disable chain table allocation for fast or row-based strategies */
+    size_t const chainSize = ZSTD_allocateChainTable(cParams->strategy, useRowMatchFinder,
+                                                     ms->dedicatedDictSearch && (forWho == ZSTD_resetTarget_CDict))
+                                ? ((size_t)1 << cParams->chainLog)
+                                : 0;
     size_t const hSize = ((size_t)1) << cParams->hashLog;
     U32    const hashLog3 = ((forWho == ZSTD_resetTarget_CCtx) && cParams->minMatch==3) ? MIN(ZSTD_HASHLOG3_MAX, cParams->windowLog) : 0;
     size_t const h3Size = hashLog3 ? ((size_t)1) << hashLog3 : 0;
 
     DEBUGLOG(4, "reset indices : %u", forceResetIndex == ZSTDirp_reset);
+    assert(useRowMatchFinder != ZSTD_ps_auto);
     if (forceResetIndex == ZSTDirp_reset) {
         ZSTD_window_init(&ms->window);
         ZSTD_cwksp_mark_tables_dirty(ws);
@@ -1393,11 +1942,23 @@ ZSTD_reset_matchState(ZSTD_matchState_t* ms,
         ms->opt.priceTable = (ZSTD_optimal_t*)ZSTD_cwksp_reserve_aligned(ws, (ZSTD_OPT_NUM+1) * sizeof(ZSTD_optimal_t));
     }
 
+    if (ZSTD_rowMatchFinderUsed(cParams->strategy, useRowMatchFinder)) {
+        {   /* Row match finder needs an additional table of hashes ("tags") */
+            size_t const tagTableSize = hSize*sizeof(U16);
+            ms->tagTable = (U16*)ZSTD_cwksp_reserve_aligned(ws, tagTableSize);
+            if (ms->tagTable) ZSTD_memset(ms->tagTable, 0, tagTableSize);
+        }
+        {   /* Switch to 32-entry rows if searchLog is 5 (or more) */
+            U32 const rowLog = BOUNDED(4, cParams->searchLog, 6);
+            assert(cParams->hashLog >= rowLog);
+            ms->rowHashLog = cParams->hashLog - rowLog;
+        }
+    }
+
     ms->cParams = *cParams;
 
     RETURN_ERROR_IF(ZSTD_cwksp_reserve_failed(ws), memory_allocation,
                     "failed a workspace allocation in ZSTD_reset_matchState");
-
     return 0;
 }
 
@@ -1414,75 +1975,87 @@ static int ZSTD_indexTooCloseToMax(ZSTD_window_t w)
     return (size_t)(w.nextSrc - w.base) > (ZSTD_CURRENT_MAX - ZSTD_INDEXOVERFLOW_MARGIN);
 }
 
+/** ZSTD_dictTooBig():
+ * When dictionaries are larger than ZSTD_CHUNKSIZE_MAX they can't be loaded in
+ * one go generically. So we ensure that in that case we reset the tables to zero,
+ * so that we can load as much of the dictionary as possible.
+ */
+static int ZSTD_dictTooBig(size_t const loadedDictSize)
+{
+    return loadedDictSize > ZSTD_CHUNKSIZE_MAX;
+}
+
 /*! ZSTD_resetCCtx_internal() :
-    note : `params` are assumed fully validated at this stage */
+ * @param loadedDictSize The size of the dictionary to be loaded
+ * into the context, if any. If no dictionary is used, or the
+ * dictionary is being attached / copied, then pass 0.
+ * note : `params` are assumed fully validated at this stage.
+ */
 static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc,
-                                      ZSTD_CCtx_params params,
+                                      ZSTD_CCtx_params const* params,
                                       U64 const pledgedSrcSize,
+                                      size_t const loadedDictSize,
                                       ZSTD_compResetPolicy_e const crp,
                                       ZSTD_buffered_policy_e const zbuff)
 {
     ZSTD_cwksp* const ws = &zc->workspace;
-    DEBUGLOG(4, "ZSTD_resetCCtx_internal: pledgedSrcSize=%u, wlog=%u",
-                (U32)pledgedSrcSize, params.cParams.windowLog);
-    assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams)));
+    DEBUGLOG(4, "ZSTD_resetCCtx_internal: pledgedSrcSize=%u, wlog=%u, useRowMatchFinder=%d useBlockSplitter=%d",
+                (U32)pledgedSrcSize, params->cParams.windowLog, (int)params->useRowMatchFinder, (int)params->useBlockSplitter);
+    assert(!ZSTD_isError(ZSTD_checkCParams(params->cParams)));
 
     zc->isFirstBlock = 1;
 
-    if (params.ldmParams.enableLdm) {
+    /* Set applied params early so we can modify them for LDM,
+     * and point params at the applied params.
+     */
+    zc->appliedParams = *params;
+    params = &zc->appliedParams;
+
+    assert(params->useRowMatchFinder != ZSTD_ps_auto);
+    assert(params->useBlockSplitter != ZSTD_ps_auto);
+    assert(params->ldmParams.enableLdm != ZSTD_ps_auto);
+    assert(params->maxBlockSize != 0);
+    if (params->ldmParams.enableLdm == ZSTD_ps_enable) {
         /* Adjust long distance matching parameters */
-        ZSTD_ldm_adjustParameters(&params.ldmParams, &params.cParams);
-        assert(params.ldmParams.hashLog >= params.ldmParams.bucketSizeLog);
-        assert(params.ldmParams.hashRateLog < 32);
-        zc->ldmState.hashPower = ZSTD_rollingHash_primePower(params.ldmParams.minMatchLength);
-    }
-
-    {   size_t const windowSize = MAX(1, (size_t)MIN(((U64)1 << params.cParams.windowLog), pledgedSrcSize));
-        size_t const blockSize = MIN(ZSTD_BLOCKSIZE_MAX, windowSize);
-        U32    const divider = (params.cParams.minMatch==3) ? 3 : 4;
-        size_t const maxNbSeq = blockSize / divider;
-        size_t const tokenSpace = ZSTD_cwksp_alloc_size(WILDCOPY_OVERLENGTH + blockSize)
-                                + ZSTD_cwksp_alloc_size(maxNbSeq * sizeof(seqDef))
-                                + 3 * ZSTD_cwksp_alloc_size(maxNbSeq * sizeof(BYTE));
-        size_t const buffOutSize = (zbuff==ZSTDb_buffered) ? ZSTD_compressBound(blockSize)+1 : 0;
-        size_t const buffInSize = (zbuff==ZSTDb_buffered) ? windowSize + blockSize : 0;
-        size_t const matchStateSize = ZSTD_sizeof_matchState(&params.cParams, /* forCCtx */ 1);
-        size_t const maxNbLdmSeq = ZSTD_ldm_getMaxNbSeq(params.ldmParams, blockSize);
-
-        ZSTD_indexResetPolicy_e needsIndexReset = zc->initialized ? ZSTDirp_continue : ZSTDirp_reset;
-
-        if (ZSTD_indexTooCloseToMax(zc->blockState.matchState.window)) {
-            needsIndexReset = ZSTDirp_reset;
-        }
+        ZSTD_ldm_adjustParameters(&zc->appliedParams.ldmParams, &params->cParams);
+        assert(params->ldmParams.hashLog >= params->ldmParams.bucketSizeLog);
+        assert(params->ldmParams.hashRateLog < 32);
+    }
 
-        if (!zc->staticSize) ZSTD_cwksp_bump_oversized_duration(ws, 0);
+    {   size_t const windowSize = MAX(1, (size_t)MIN(((U64)1 << params->cParams.windowLog), pledgedSrcSize));
+        size_t const blockSize = MIN(params->maxBlockSize, windowSize);
+        size_t const maxNbSeq = ZSTD_maxNbSeq(blockSize, params->cParams.minMatch, params->useSequenceProducer);
+        size_t const buffOutSize = (zbuff == ZSTDb_buffered && params->outBufferMode == ZSTD_bm_buffered)
+                ? ZSTD_compressBound(blockSize) + 1
+                : 0;
+        size_t const buffInSize = (zbuff == ZSTDb_buffered && params->inBufferMode == ZSTD_bm_buffered)
+                ? windowSize + blockSize
+                : 0;
+        size_t const maxNbLdmSeq = ZSTD_ldm_getMaxNbSeq(params->ldmParams, blockSize);
+
+        int const indexTooClose = ZSTD_indexTooCloseToMax(zc->blockState.matchState.window);
+        int const dictTooBig = ZSTD_dictTooBig(loadedDictSize);
+        ZSTD_indexResetPolicy_e needsIndexReset =
+            (indexTooClose || dictTooBig || !zc->initialized) ? ZSTDirp_reset : ZSTDirp_continue;
+
+        size_t const neededSpace =
+            ZSTD_estimateCCtxSize_usingCCtxParams_internal(
+                &params->cParams, &params->ldmParams, zc->staticSize != 0, params->useRowMatchFinder,
+                buffInSize, buffOutSize, pledgedSrcSize, params->useSequenceProducer, params->maxBlockSize);
+        int resizeWorkspace;
+
+        FORWARD_IF_ERROR(neededSpace, "cctx size estimate failed!");
 
-        /* Check if workspace is large enough, alloc a new one if needed */
-        {   size_t const cctxSpace = zc->staticSize ? ZSTD_cwksp_alloc_size(sizeof(ZSTD_CCtx)) : 0;
-            size_t const entropySpace = ZSTD_cwksp_alloc_size(HUF_WORKSPACE_SIZE);
-            size_t const blockStateSpace = 2 * ZSTD_cwksp_alloc_size(sizeof(ZSTD_compressedBlockState_t));
-            size_t const bufferSpace = ZSTD_cwksp_alloc_size(buffInSize) + ZSTD_cwksp_alloc_size(buffOutSize);
-            size_t const ldmSpace = ZSTD_ldm_getTableSize(params.ldmParams);
-            size_t const ldmSeqSpace = ZSTD_cwksp_alloc_size(maxNbLdmSeq * sizeof(rawSeq));
-
-            size_t const neededSpace =
-                cctxSpace +
-                entropySpace +
-                blockStateSpace +
-                ldmSpace +
-                ldmSeqSpace +
-                matchStateSize +
-                tokenSpace +
-                bufferSpace;
+        if (!zc->staticSize) ZSTD_cwksp_bump_oversized_duration(ws, 0);
 
+        {   /* Check if workspace is large enough, alloc a new one if needed */
             int const workspaceTooSmall = ZSTD_cwksp_sizeof(ws) < neededSpace;
             int const workspaceWasteful = ZSTD_cwksp_check_wasteful(ws, neededSpace);
-
-            DEBUGLOG(4, "Need %zuKB workspace, including %zuKB for match state, and %zuKB for buffers",
-                        neededSpace>>10, matchStateSize>>10, bufferSpace>>10);
+            resizeWorkspace = workspaceTooSmall || workspaceWasteful;
+            DEBUGLOG(4, "Need %zu B workspace", neededSpace);
             DEBUGLOG(4, "windowSize: %zu - blockSize: %zu", windowSize, blockSize);
 
-            if (workspaceTooSmall || workspaceWasteful) {
+            if (resizeWorkspace) {
                 DEBUGLOG(4, "Resize workspaceSize from %zuKB to %zuKB",
                             ZSTD_cwksp_sizeof(ws) >> 10,
                             neededSpace >> 10);
@@ -1503,15 +2076,15 @@ static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc,
                 RETURN_ERROR_IF(zc->blockState.prevCBlock == NULL, memory_allocation, "couldn't allocate prevCBlock");
                 zc->blockState.nextCBlock = (ZSTD_compressedBlockState_t*) ZSTD_cwksp_reserve_object(ws, sizeof(ZSTD_compressedBlockState_t));
                 RETURN_ERROR_IF(zc->blockState.nextCBlock == NULL, memory_allocation, "couldn't allocate nextCBlock");
-                zc->entropyWorkspace = (U32*) ZSTD_cwksp_reserve_object(ws, HUF_WORKSPACE_SIZE);
-                RETURN_ERROR_IF(zc->blockState.nextCBlock == NULL, memory_allocation, "couldn't allocate entropyWorkspace");
+                zc->entropyWorkspace = (U32*) ZSTD_cwksp_reserve_object(ws, ENTROPY_WORKSPACE_SIZE);
+                RETURN_ERROR_IF(zc->entropyWorkspace == NULL, memory_allocation, "couldn't allocate entropyWorkspace");
         }   }
 
         ZSTD_cwksp_clear(ws);
 
         /* init params */
-        zc->appliedParams = params;
-        zc->blockState.matchState.cParams = params.cParams;
+        zc->blockState.matchState.cParams = params->cParams;
+        zc->blockState.matchState.prefetchCDictTables = params->prefetchCDictTables == ZSTD_ps_enable;
         zc->pledgedSrcSizePlusOne = pledgedSrcSize+1;
         zc->consumedSrcSize = 0;
         zc->producedCSize = 0;
@@ -1524,6 +2097,7 @@ static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc,
         XXH64_reset(&zc->xxhState, 0);
         zc->stage = ZSTDcs_init;
         zc->dictID = 0;
+        zc->dictContentSize = 0;
 
         ZSTD_reset_compressedBlockState(zc->blockState.prevCBlock);
 
@@ -1534,19 +2108,20 @@ static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc,
         zc->seqStore.maxNbLit = blockSize;
 
         /* buffers */
+        zc->bufferedPolicy = zbuff;
         zc->inBuffSize = buffInSize;
         zc->inBuff = (char*)ZSTD_cwksp_reserve_buffer(ws, buffInSize);
         zc->outBuffSize = buffOutSize;
         zc->outBuff = (char*)ZSTD_cwksp_reserve_buffer(ws, buffOutSize);
 
         /* ldm bucketOffsets table */
-        if (params.ldmParams.enableLdm) {
+        if (params->ldmParams.enableLdm == ZSTD_ps_enable) {
             /* TODO: avoid memset? */
-            size_t const ldmBucketSize =
-                  ((size_t)1) << (params.ldmParams.hashLog -
-                                  params.ldmParams.bucketSizeLog);
-            zc->ldmState.bucketOffsets = ZSTD_cwksp_reserve_buffer(ws, ldmBucketSize);
-            memset(zc->ldmState.bucketOffsets, 0, ldmBucketSize);
+            size_t const numBuckets =
+                  ((size_t)1) << (params->ldmParams.hashLog -
+                                  params->ldmParams.bucketSizeLog);
+            zc->ldmState.bucketOffsets = ZSTD_cwksp_reserve_buffer(ws, numBuckets);
+            ZSTD_memset(zc->ldmState.bucketOffsets, 0, numBuckets);
         }
 
         /* sequences storage */
@@ -1560,26 +2135,36 @@ static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc,
         FORWARD_IF_ERROR(ZSTD_reset_matchState(
             &zc->blockState.matchState,
             ws,
-            &params.cParams,
+            &params->cParams,
+            params->useRowMatchFinder,
             crp,
             needsIndexReset,
             ZSTD_resetTarget_CCtx), "");
 
         /* ldm hash table */
-        if (params.ldmParams.enableLdm) {
+        if (params->ldmParams.enableLdm == ZSTD_ps_enable) {
             /* TODO: avoid memset? */
-            size_t const ldmHSize = ((size_t)1) << params.ldmParams.hashLog;
+            size_t const ldmHSize = ((size_t)1) << params->ldmParams.hashLog;
             zc->ldmState.hashTable = (ldmEntry_t*)ZSTD_cwksp_reserve_aligned(ws, ldmHSize * sizeof(ldmEntry_t));
-            memset(zc->ldmState.hashTable, 0, ldmHSize * sizeof(ldmEntry_t));
+            ZSTD_memset(zc->ldmState.hashTable, 0, ldmHSize * sizeof(ldmEntry_t));
             zc->ldmSequences = (rawSeq*)ZSTD_cwksp_reserve_aligned(ws, maxNbLdmSeq * sizeof(rawSeq));
             zc->maxNbLdmSequences = maxNbLdmSeq;
 
             ZSTD_window_init(&zc->ldmState.window);
-            ZSTD_window_clear(&zc->ldmState.window);
             zc->ldmState.loadedDictEnd = 0;
         }
 
+        /* reserve space for block-level external sequences */
+        if (params->useSequenceProducer) {
+            size_t const maxNbExternalSeq = ZSTD_sequenceBound(blockSize);
+            zc->externalMatchCtx.seqBufferCapacity = maxNbExternalSeq;
+            zc->externalMatchCtx.seqBuffer =
+                (ZSTD_Sequence*)ZSTD_cwksp_reserve_aligned(ws, maxNbExternalSeq * sizeof(ZSTD_Sequence));
+        }
+
         DEBUGLOG(3, "wksp: finished allocating, %zd bytes remain available", ZSTD_cwksp_available_space(ws));
+        assert(ZSTD_cwksp_estimated_space_within_bounds(ws, neededSpace, resizeWorkspace));
+
         zc->initialized = 1;
 
         return 0;
@@ -1618,12 +2203,14 @@ static int ZSTD_shouldAttachDict(const ZSTD_CDict* cdict,
                                  U64 pledgedSrcSize)
 {
     size_t cutoff = attachDictSizeCutoffs[cdict->matchState.cParams.strategy];
-    return ( pledgedSrcSize <= cutoff
-          || pledgedSrcSize == ZSTD_CONTENTSIZE_UNKNOWN
-          || params->attachDictPref == ZSTD_dictForceAttach )
-        && params->attachDictPref != ZSTD_dictForceCopy
-        && !params->forceWindow; /* dictMatchState isn't correctly
-                                 * handled in _enforceMaxDist */
+    int const dedicatedDictSearch = cdict->matchState.dedicatedDictSearch;
+    return dedicatedDictSearch
+        || ( ( pledgedSrcSize <= cutoff
+            || pledgedSrcSize == ZSTD_CONTENTSIZE_UNKNOWN
+            || params->attachDictPref == ZSTD_dictForceAttach )
+          && params->attachDictPref != ZSTD_dictForceCopy
+          && !params->forceWindow ); /* dictMatchState isn't correctly
+                                      * handled in _enforceMaxDist */
 }
 
 static size_t
@@ -1633,17 +2220,29 @@ ZSTD_resetCCtx_byAttachingCDict(ZSTD_CCtx* cctx,
                         U64 pledgedSrcSize,
                         ZSTD_buffered_policy_e zbuff)
 {
-    {   const ZSTD_compressionParameters* const cdict_cParams = &cdict->matchState.cParams;
+    DEBUGLOG(4, "ZSTD_resetCCtx_byAttachingCDict() pledgedSrcSize=%llu",
+                (unsigned long long)pledgedSrcSize);
+    {
+        ZSTD_compressionParameters adjusted_cdict_cParams = cdict->matchState.cParams;
         unsigned const windowLog = params.cParams.windowLog;
         assert(windowLog != 0);
         /* Resize working context table params for input only, since the dict
          * has its own tables. */
-        /* pledgeSrcSize == 0 means 0! */
-        params.cParams = ZSTD_adjustCParams_internal(*cdict_cParams, pledgedSrcSize, 0);
+        /* pledgedSrcSize == 0 means 0! */
+
+        if (cdict->matchState.dedicatedDictSearch) {
+            ZSTD_dedicatedDictSearch_revertCParams(&adjusted_cdict_cParams);
+        }
+
+        params.cParams = ZSTD_adjustCParams_internal(adjusted_cdict_cParams, pledgedSrcSize,
+                                                     cdict->dictContentSize, ZSTD_cpm_attachDict,
+                                                     params.useRowMatchFinder);
         params.cParams.windowLog = windowLog;
-        FORWARD_IF_ERROR(ZSTD_resetCCtx_internal(cctx, params, pledgedSrcSize,
+        params.useRowMatchFinder = cdict->useRowMatchFinder;    /* cdict overrides */
+        FORWARD_IF_ERROR(ZSTD_resetCCtx_internal(cctx, &params, pledgedSrcSize,
+                                                 /* loadedDictSize */ 0,
                                                  ZSTDcrp_makeClean, zbuff), "");
-        assert(cctx->appliedParams.cParams.strategy == cdict_cParams->strategy);
+        assert(cctx->appliedParams.cParams.strategy == adjusted_cdict_cParams.strategy);
     }
 
     {   const U32 cdictEnd = (U32)( cdict->matchState.window.nextSrc
@@ -1668,13 +2267,30 @@ ZSTD_resetCCtx_byAttachingCDict(ZSTD_CCtx* cctx,
     }   }
 
     cctx->dictID = cdict->dictID;
+    cctx->dictContentSize = cdict->dictContentSize;
 
     /* copy block state */
-    memcpy(cctx->blockState.prevCBlock, &cdict->cBlockState, sizeof(cdict->cBlockState));
+    ZSTD_memcpy(cctx->blockState.prevCBlock, &cdict->cBlockState, sizeof(cdict->cBlockState));
 
     return 0;
 }
 
+static void ZSTD_copyCDictTableIntoCCtx(U32* dst, U32 const* src, size_t tableSize,
+                                        ZSTD_compressionParameters const* cParams) {
+    if (ZSTD_CDictIndicesAreTagged(cParams)){
+        /* Remove tags from the CDict table if they are present.
+         * See docs on "short cache" in zstd_compress_internal.h for context. */
+        size_t i;
+        for (i = 0; i < tableSize; i++) {
+            U32 const taggedIndex = src[i];
+            U32 const index = taggedIndex >> ZSTD_SHORT_CACHE_TAG_BITS;
+            dst[i] = index;
+        }
+    } else {
+        ZSTD_memcpy(dst, src, tableSize * sizeof(U32));
+    }
+}
+
 static size_t ZSTD_resetCCtx_byCopyingCDict(ZSTD_CCtx* cctx,
                             const ZSTD_CDict* cdict,
                             ZSTD_CCtx_params params,
@@ -1683,14 +2299,18 @@ static size_t ZSTD_resetCCtx_byCopyingCDict(ZSTD_CCtx* cctx,
 {
     const ZSTD_compressionParameters *cdict_cParams = &cdict->matchState.cParams;
 
-    DEBUGLOG(4, "copying dictionary into context");
+    assert(!cdict->matchState.dedicatedDictSearch);
+    DEBUGLOG(4, "ZSTD_resetCCtx_byCopyingCDict() pledgedSrcSize=%llu",
+                (unsigned long long)pledgedSrcSize);
 
     {   unsigned const windowLog = params.cParams.windowLog;
         assert(windowLog != 0);
         /* Copy only compression parameters related to tables. */
         params.cParams = *cdict_cParams;
         params.cParams.windowLog = windowLog;
-        FORWARD_IF_ERROR(ZSTD_resetCCtx_internal(cctx, params, pledgedSrcSize,
+        params.useRowMatchFinder = cdict->useRowMatchFinder;
+        FORWARD_IF_ERROR(ZSTD_resetCCtx_internal(cctx, &params, pledgedSrcSize,
+                                                 /* loadedDictSize */ 0,
                                                  ZSTDcrp_leaveDirty, zbuff), "");
         assert(cctx->appliedParams.cParams.strategy == cdict_cParams->strategy);
         assert(cctx->appliedParams.cParams.hashLog == cdict_cParams->hashLog);
@@ -1698,24 +2318,38 @@ static size_t ZSTD_resetCCtx_byCopyingCDict(ZSTD_CCtx* cctx,
     }
 
     ZSTD_cwksp_mark_tables_dirty(&cctx->workspace);
+    assert(params.useRowMatchFinder != ZSTD_ps_auto);
 
     /* copy tables */
-    {   size_t const chainSize = (cdict_cParams->strategy == ZSTD_fast) ? 0 : ((size_t)1 << cdict_cParams->chainLog);
+    {   size_t const chainSize = ZSTD_allocateChainTable(cdict_cParams->strategy, cdict->useRowMatchFinder, 0 /* DDS guaranteed disabled */)
+                                                            ? ((size_t)1 << cdict_cParams->chainLog)
+                                                            : 0;
         size_t const hSize =  (size_t)1 << cdict_cParams->hashLog;
 
-        memcpy(cctx->blockState.matchState.hashTable,
-               cdict->matchState.hashTable,
-               hSize * sizeof(U32));
-        memcpy(cctx->blockState.matchState.chainTable,
-               cdict->matchState.chainTable,
-               chainSize * sizeof(U32));
+        ZSTD_copyCDictTableIntoCCtx(cctx->blockState.matchState.hashTable,
+                                cdict->matchState.hashTable,
+                                hSize, cdict_cParams);
+
+        /* Do not copy cdict's chainTable if cctx has parameters such that it would not use chainTable */
+        if (ZSTD_allocateChainTable(cctx->appliedParams.cParams.strategy, cctx->appliedParams.useRowMatchFinder, 0 /* forDDSDict */)) {
+            ZSTD_copyCDictTableIntoCCtx(cctx->blockState.matchState.chainTable,
+                                    cdict->matchState.chainTable,
+                                    chainSize, cdict_cParams);
+        }
+        /* copy tag table */
+        if (ZSTD_rowMatchFinderUsed(cdict_cParams->strategy, cdict->useRowMatchFinder)) {
+            size_t const tagTableSize = hSize*sizeof(U16);
+            ZSTD_memcpy(cctx->blockState.matchState.tagTable,
+                cdict->matchState.tagTable,
+                tagTableSize);
+        }
     }
 
     /* Zero the hashTable3, since the cdict never fills it */
     {   int const h3log = cctx->blockState.matchState.hashLog3;
         size_t const h3Size = h3log ? ((size_t)1 << h3log) : 0;
         assert(cdict->matchState.hashLog3 == 0);
-        memset(cctx->blockState.matchState.hashTable3, 0, h3Size * sizeof(U32));
+        ZSTD_memset(cctx->blockState.matchState.hashTable3, 0, h3Size * sizeof(U32));
     }
 
     ZSTD_cwksp_mark_tables_clean(&cctx->workspace);
@@ -1729,9 +2363,10 @@ static size_t ZSTD_resetCCtx_byCopyingCDict(ZSTD_CCtx* cctx,
     }
 
     cctx->dictID = cdict->dictID;
+    cctx->dictContentSize = cdict->dictContentSize;
 
     /* copy block state */
-    memcpy(cctx->blockState.prevCBlock, &cdict->cBlockState, sizeof(cdict->cBlockState));
+    ZSTD_memcpy(cctx->blockState.prevCBlock, &cdict->cBlockState, sizeof(cdict->cBlockState));
 
     return 0;
 }
@@ -1771,16 +2406,23 @@ static size_t ZSTD_copyCCtx_internal(ZSTD_CCtx* dstCCtx,
                             U64 pledgedSrcSize,
                             ZSTD_buffered_policy_e zbuff)
 {
-    DEBUGLOG(5, "ZSTD_copyCCtx_internal");
     RETURN_ERROR_IF(srcCCtx->stage!=ZSTDcs_init, stage_wrong,
                     "Can't copy a ctx that's not in init stage.");
-
-    memcpy(&dstCCtx->customMem, &srcCCtx->customMem, sizeof(ZSTD_customMem));
+    DEBUGLOG(5, "ZSTD_copyCCtx_internal");
+    ZSTD_memcpy(&dstCCtx->customMem, &srcCCtx->customMem, sizeof(ZSTD_customMem));
     {   ZSTD_CCtx_params params = dstCCtx->requestedParams;
         /* Copy only compression parameters related to tables. */
         params.cParams = srcCCtx->appliedParams.cParams;
+        assert(srcCCtx->appliedParams.useRowMatchFinder != ZSTD_ps_auto);
+        assert(srcCCtx->appliedParams.useBlockSplitter != ZSTD_ps_auto);
+        assert(srcCCtx->appliedParams.ldmParams.enableLdm != ZSTD_ps_auto);
+        params.useRowMatchFinder = srcCCtx->appliedParams.useRowMatchFinder;
+        params.useBlockSplitter = srcCCtx->appliedParams.useBlockSplitter;
+        params.ldmParams = srcCCtx->appliedParams.ldmParams;
         params.fParams = fParams;
-        ZSTD_resetCCtx_internal(dstCCtx, params, pledgedSrcSize,
+        params.maxBlockSize = srcCCtx->appliedParams.maxBlockSize;
+        ZSTD_resetCCtx_internal(dstCCtx, &params, pledgedSrcSize,
+                                /* loadedDictSize */ 0,
                                 ZSTDcrp_leaveDirty, zbuff);
         assert(dstCCtx->appliedParams.cParams.windowLog == srcCCtx->appliedParams.cParams.windowLog);
         assert(dstCCtx->appliedParams.cParams.strategy == srcCCtx->appliedParams.cParams.strategy);
@@ -1792,18 +2434,22 @@ static size_t ZSTD_copyCCtx_internal(ZSTD_CCtx* dstCCtx,
     ZSTD_cwksp_mark_tables_dirty(&dstCCtx->workspace);
 
     /* copy tables */
-    {   size_t const chainSize = (srcCCtx->appliedParams.cParams.strategy == ZSTD_fast) ? 0 : ((size_t)1 << srcCCtx->appliedParams.cParams.chainLog);
+    {   size_t const chainSize = ZSTD_allocateChainTable(srcCCtx->appliedParams.cParams.strategy,
+                                                         srcCCtx->appliedParams.useRowMatchFinder,
+                                                         0 /* forDDSDict */)
+                                    ? ((size_t)1 << srcCCtx->appliedParams.cParams.chainLog)
+                                    : 0;
         size_t const hSize =  (size_t)1 << srcCCtx->appliedParams.cParams.hashLog;
         int const h3log = srcCCtx->blockState.matchState.hashLog3;
         size_t const h3Size = h3log ? ((size_t)1 << h3log) : 0;
 
-        memcpy(dstCCtx->blockState.matchState.hashTable,
+        ZSTD_memcpy(dstCCtx->blockState.matchState.hashTable,
                srcCCtx->blockState.matchState.hashTable,
                hSize * sizeof(U32));
-        memcpy(dstCCtx->blockState.matchState.chainTable,
+        ZSTD_memcpy(dstCCtx->blockState.matchState.chainTable,
                srcCCtx->blockState.matchState.chainTable,
                chainSize * sizeof(U32));
-        memcpy(dstCCtx->blockState.matchState.hashTable3,
+        ZSTD_memcpy(dstCCtx->blockState.matchState.hashTable3,
                srcCCtx->blockState.matchState.hashTable3,
                h3Size * sizeof(U32));
     }
@@ -1819,9 +2465,10 @@ static size_t ZSTD_copyCCtx_internal(ZSTD_CCtx* dstCCtx,
         dstMatchState->loadedDictEnd= srcMatchState->loadedDictEnd;
     }
     dstCCtx->dictID = srcCCtx->dictID;
+    dstCCtx->dictContentSize = srcCCtx->dictContentSize;
 
     /* copy block state */
-    memcpy(dstCCtx->blockState.prevCBlock, srcCCtx->blockState.prevCBlock, sizeof(*srcCCtx->blockState.prevCBlock));
+    ZSTD_memcpy(dstCCtx->blockState.prevCBlock, srcCCtx->blockState.prevCBlock, sizeof(*srcCCtx->blockState.prevCBlock));
 
     return 0;
 }
@@ -1834,7 +2481,7 @@ static size_t ZSTD_copyCCtx_internal(ZSTD_CCtx* dstCCtx,
 size_t ZSTD_copyCCtx(ZSTD_CCtx* dstCCtx, const ZSTD_CCtx* srcCCtx, unsigned long long pledgedSrcSize)
 {
     ZSTD_frameParameters fParams = { 1 /*content*/, 0 /*checksum*/, 0 /*noDictID*/ };
-    ZSTD_buffered_policy_e const zbuff = (ZSTD_buffered_policy_e)(srcCCtx->inBuffSize>0);
+    ZSTD_buffered_policy_e const zbuff = srcCCtx->bufferedPolicy;
     ZSTD_STATIC_ASSERT((U32)ZSTDb_buffered==1);
     if (pledgedSrcSize==0) pledgedSrcSize = ZSTD_CONTENTSIZE_UNKNOWN;
     fParams.contentSizeFlag = (pledgedSrcSize != ZSTD_CONTENTSIZE_UNKNOWN);
@@ -1858,10 +2505,12 @@ ZSTD_reduceTable_internal (U32* const table, U32 const size, U32 const reducerVa
     int const nbRows = (int)size / ZSTD_ROWSIZE;
     int cellNb = 0;
     int rowNb;
+    /* Protect special index values < ZSTD_WINDOW_START_INDEX. */
+    U32 const reducerThreshold = reducerValue + ZSTD_WINDOW_START_INDEX;
     assert((size & (ZSTD_ROWSIZE-1)) == 0);  /* multiple of ZSTD_ROWSIZE */
     assert(size < (1U<<31));   /* can be casted to int */
 
-#if defined (MEMORY_SANITIZER) && !defined (ZSTD_MSAN_DONT_POISON_WORKSPACE)
+#if ZSTD_MEMORY_SANITIZER && !defined (ZSTD_MSAN_DONT_POISON_WORKSPACE)
     /* To validate that the table re-use logic is sound, and that we don't
      * access table space that we haven't cleaned, we re-"poison" the table
      * space every time we mark it dirty.
@@ -1877,12 +2526,17 @@ ZSTD_reduceTable_internal (U32* const table, U32 const size, U32 const reducerVa
     for (rowNb=0 ; rowNb < nbRows ; rowNb++) {
         int column;
         for (column=0; column<ZSTD_ROWSIZE; column++) {
-            if (preserveMark) {
-                U32 const adder = (table[cellNb] == ZSTD_DUBT_UNSORTED_MARK) ? reducerValue : 0;
-                table[cellNb] += adder;
+            U32 newVal;
+            if (preserveMark && table[cellNb] == ZSTD_DUBT_UNSORTED_MARK) {
+                /* This write is pointless, but is required(?) for the compiler
+                 * to auto-vectorize the loop. */
+                newVal = ZSTD_DUBT_UNSORTED_MARK;
+            } else if (table[cellNb] < reducerThreshold) {
+                newVal = 0;
+            } else {
+                newVal = table[cellNb] - reducerValue;
             }
-            if (table[cellNb] < reducerValue) table[cellNb] = 0;
-            else table[cellNb] -= reducerValue;
+            table[cellNb] = newVal;
             cellNb++;
     }   }
 }
@@ -1905,7 +2559,7 @@ static void ZSTD_reduceIndex (ZSTD_matchState_t* ms, ZSTD_CCtx_params const* par
         ZSTD_reduceTable(ms->hashTable, hSize, reducerValue);
     }
 
-    if (params->cParams.strategy != ZSTD_fast) {
+    if (ZSTD_allocateChainTable(params->cParams.strategy, params->useRowMatchFinder, (U32)ms->dedicatedDictSearch)) {
         U32 const chainSize = (U32)1 << params->cParams.chainLog;
         if (params->cParams.strategy == ZSTD_btlazy2)
             ZSTD_reduceTable_btlazy2(ms->chainTable, chainSize, reducerValue);
@@ -1926,7 +2580,7 @@ static void ZSTD_reduceIndex (ZSTD_matchState_t* ms, ZSTD_CCtx_params const* par
 
 /* See doc/zstd_compression_format.md for detailed format description */
 
-void ZSTD_seqToCodes(const seqStore_t* seqStorePtr)
+int ZSTD_seqToCodes(const seqStore_t* seqStorePtr)
 {
     const seqDef* const sequences = seqStorePtr->sequencesStart;
     BYTE* const llCodeTable = seqStorePtr->llCode;
@@ -1934,18 +2588,24 @@ void ZSTD_seqToCodes(const seqStore_t* seqStorePtr)
     BYTE* const mlCodeTable = seqStorePtr->mlCode;
     U32 const nbSeq = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
     U32 u;
+    int longOffsets = 0;
     assert(nbSeq <= seqStorePtr->maxNbSeq);
     for (u=0; u<nbSeq; u++) {
         U32 const llv = sequences[u].litLength;
-        U32 const mlv = sequences[u].matchLength;
+        U32 const ofCode = ZSTD_highbit32(sequences[u].offBase);
+        U32 const mlv = sequences[u].mlBase;
         llCodeTable[u] = (BYTE)ZSTD_LLcode(llv);
-        ofCodeTable[u] = (BYTE)ZSTD_highbit32(sequences[u].offset);
+        ofCodeTable[u] = (BYTE)ofCode;
         mlCodeTable[u] = (BYTE)ZSTD_MLcode(mlv);
+        assert(!(MEM_64bits() && ofCode >= STREAM_ACCUMULATOR_MIN));
+        if (MEM_32bits() && ofCode >= STREAM_ACCUMULATOR_MIN)
+            longOffsets = 1;
     }
-    if (seqStorePtr->longLengthID==1)
+    if (seqStorePtr->longLengthType==ZSTD_llt_literalLength)
         llCodeTable[seqStorePtr->longLengthPos] = MaxLL;
-    if (seqStorePtr->longLengthID==2)
+    if (seqStorePtr->longLengthType==ZSTD_llt_matchLength)
         mlCodeTable[seqStorePtr->longLengthPos] = MaxML;
+    return longOffsets;
 }
 
 /* ZSTD_useTargetCBlockSize():
@@ -1958,163 +2618,250 @@ static int ZSTD_useTargetCBlockSize(const ZSTD_CCtx_params* cctxParams)
     return (cctxParams->targetCBlockSize != 0);
 }
 
-/* ZSTD_compressSequences_internal():
- * actually compresses both literals and sequences */
-MEM_STATIC size_t
-ZSTD_compressSequences_internal(seqStore_t* seqStorePtr,
-                          const ZSTD_entropyCTables_t* prevEntropy,
-                                ZSTD_entropyCTables_t* nextEntropy,
-                          const ZSTD_CCtx_params* cctxParams,
-                                void* dst, size_t dstCapacity,
-                                void* entropyWorkspace, size_t entropyWkspSize,
-                          const int bmi2)
+/* ZSTD_blockSplitterEnabled():
+ * Returns if block splitting param is being used
+ * If used, compression will do best effort to split a block in order to improve compression ratio.
+ * At the time this function is called, the parameter must be finalized.
+ * Returns 1 if true, 0 otherwise. */
+static int ZSTD_blockSplitterEnabled(ZSTD_CCtx_params* cctxParams)
 {
-    const int longOffsets = cctxParams->cParams.windowLog > STREAM_ACCUMULATOR_MIN;
-    ZSTD_strategy const strategy = cctxParams->cParams.strategy;
-    unsigned count[MaxSeq+1];
-    FSE_CTable* CTable_LitLength = nextEntropy->fse.litlengthCTable;
-    FSE_CTable* CTable_OffsetBits = nextEntropy->fse.offcodeCTable;
-    FSE_CTable* CTable_MatchLength = nextEntropy->fse.matchlengthCTable;
-    U32 LLtype, Offtype, MLtype;   /* compressed, raw or rle */
-    const seqDef* const sequences = seqStorePtr->sequencesStart;
+    DEBUGLOG(5, "ZSTD_blockSplitterEnabled (useBlockSplitter=%d)", cctxParams->useBlockSplitter);
+    assert(cctxParams->useBlockSplitter != ZSTD_ps_auto);
+    return (cctxParams->useBlockSplitter == ZSTD_ps_enable);
+}
+
+/* Type returned by ZSTD_buildSequencesStatistics containing finalized symbol encoding types
+ * and size of the sequences statistics
+ */
+typedef struct {
+    U32 LLtype;
+    U32 Offtype;
+    U32 MLtype;
+    size_t size;
+    size_t lastCountSize; /* Accounts for bug in 1.3.4. More detail in ZSTD_entropyCompressSeqStore_internal() */
+    int longOffsets;
+} ZSTD_symbolEncodingTypeStats_t;
+
+/* ZSTD_buildSequencesStatistics():
+ * Returns a ZSTD_symbolEncodingTypeStats_t, or a zstd error code in the `size` field.
+ * Modifies `nextEntropy` to have the appropriate values as a side effect.
+ * nbSeq must be greater than 0.
+ *
+ * entropyWkspSize must be of size at least ENTROPY_WORKSPACE_SIZE - (MaxSeq + 1)*sizeof(U32)
+ */
+static ZSTD_symbolEncodingTypeStats_t
+ZSTD_buildSequencesStatistics(
+                const seqStore_t* seqStorePtr, size_t nbSeq,
+                const ZSTD_fseCTables_t* prevEntropy, ZSTD_fseCTables_t* nextEntropy,
+                      BYTE* dst, const BYTE* const dstEnd,
+                      ZSTD_strategy strategy, unsigned* countWorkspace,
+                      void* entropyWorkspace, size_t entropyWkspSize)
+{
+    BYTE* const ostart = dst;
+    const BYTE* const oend = dstEnd;
+    BYTE* op = ostart;
+    FSE_CTable* CTable_LitLength = nextEntropy->litlengthCTable;
+    FSE_CTable* CTable_OffsetBits = nextEntropy->offcodeCTable;
+    FSE_CTable* CTable_MatchLength = nextEntropy->matchlengthCTable;
     const BYTE* const ofCodeTable = seqStorePtr->ofCode;
     const BYTE* const llCodeTable = seqStorePtr->llCode;
     const BYTE* const mlCodeTable = seqStorePtr->mlCode;
-    BYTE* const ostart = (BYTE*)dst;
-    BYTE* const oend = ostart + dstCapacity;
-    BYTE* op = ostart;
-    size_t const nbSeq = (size_t)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
-    BYTE* seqHead;
-    BYTE* lastNCount = NULL;
-
-    DEBUGLOG(5, "ZSTD_compressSequences_internal (nbSeq=%zu)", nbSeq);
-    ZSTD_STATIC_ASSERT(HUF_WORKSPACE_SIZE >= (1<<MAX(MLFSELog,LLFSELog)));
-
-    /* Compress literals */
-    {   const BYTE* const literals = seqStorePtr->litStart;
-        size_t const litSize = (size_t)(seqStorePtr->lit - literals);
-        size_t const cSize = ZSTD_compressLiterals(
-                                    &prevEntropy->huf, &nextEntropy->huf,
-                                    cctxParams->cParams.strategy,
-                                    ZSTD_disableLiteralsCompression(cctxParams),
-                                    op, dstCapacity,
-                                    literals, litSize,
-                                    entropyWorkspace, entropyWkspSize,
-                                    bmi2);
-        FORWARD_IF_ERROR(cSize, "ZSTD_compressLiterals failed");
-        assert(cSize <= dstCapacity);
-        op += cSize;
-    }
-
-    /* Sequences Header */
-    RETURN_ERROR_IF((oend-op) < 3 /*max nbSeq Size*/ + 1 /*seqHead*/,
-                    dstSize_tooSmall, "Can't fit seq hdr in output buf!");
-    if (nbSeq < 128) {
-        *op++ = (BYTE)nbSeq;
-    } else if (nbSeq < LONGNBSEQ) {
-        op[0] = (BYTE)((nbSeq>>8) + 0x80);
-        op[1] = (BYTE)nbSeq;
-        op+=2;
-    } else {
-        op[0]=0xFF;
-        MEM_writeLE16(op+1, (U16)(nbSeq - LONGNBSEQ));
-        op+=3;
-    }
-    assert(op <= oend);
-    if (nbSeq==0) {
-        /* Copy the old tables over as if we repeated them */
-        memcpy(&nextEntropy->fse, &prevEntropy->fse, sizeof(prevEntropy->fse));
-        return (size_t)(op - ostart);
-    }
-
-    /* seqHead : flags for FSE encoding type */
-    seqHead = op++;
-    assert(op <= oend);
+    ZSTD_symbolEncodingTypeStats_t stats;
 
+    stats.lastCountSize = 0;
     /* convert length/distances into codes */
-    ZSTD_seqToCodes(seqStorePtr);
+    stats.longOffsets = ZSTD_seqToCodes(seqStorePtr);
+    assert(op <= oend);
+    assert(nbSeq != 0); /* ZSTD_selectEncodingType() divides by nbSeq */
     /* build CTable for Literal Lengths */
     {   unsigned max = MaxLL;
-        size_t const mostFrequent = HIST_countFast_wksp(count, &max, llCodeTable, nbSeq, entropyWorkspace, entropyWkspSize);   /* can't fail */
+        size_t const mostFrequent = HIST_countFast_wksp(countWorkspace, &max, llCodeTable, nbSeq, entropyWorkspace, entropyWkspSize);   /* can't fail */
         DEBUGLOG(5, "Building LL table");
-        nextEntropy->fse.litlength_repeatMode = prevEntropy->fse.litlength_repeatMode;
-        LLtype = ZSTD_selectEncodingType(&nextEntropy->fse.litlength_repeatMode,
-                                        count, max, mostFrequent, nbSeq,
-                                        LLFSELog, prevEntropy->fse.litlengthCTable,
+        nextEntropy->litlength_repeatMode = prevEntropy->litlength_repeatMode;
+        stats.LLtype = ZSTD_selectEncodingType(&nextEntropy->litlength_repeatMode,
+                                        countWorkspace, max, mostFrequent, nbSeq,
+                                        LLFSELog, prevEntropy->litlengthCTable,
                                         LL_defaultNorm, LL_defaultNormLog,
                                         ZSTD_defaultAllowed, strategy);
         assert(set_basic < set_compressed && set_rle < set_compressed);
-        assert(!(LLtype < set_compressed && nextEntropy->fse.litlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */
+        assert(!(stats.LLtype < set_compressed && nextEntropy->litlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */
         {   size_t const countSize = ZSTD_buildCTable(
                 op, (size_t)(oend - op),
-                CTable_LitLength, LLFSELog, (symbolEncodingType_e)LLtype,
-                count, max, llCodeTable, nbSeq,
+                CTable_LitLength, LLFSELog, (symbolEncodingType_e)stats.LLtype,
+                countWorkspace, max, llCodeTable, nbSeq,
                 LL_defaultNorm, LL_defaultNormLog, MaxLL,
-                prevEntropy->fse.litlengthCTable,
-                sizeof(prevEntropy->fse.litlengthCTable),
+                prevEntropy->litlengthCTable,
+                sizeof(prevEntropy->litlengthCTable),
                 entropyWorkspace, entropyWkspSize);
-            FORWARD_IF_ERROR(countSize, "ZSTD_buildCTable for LitLens failed");
-            if (LLtype == set_compressed)
-                lastNCount = op;
+            if (ZSTD_isError(countSize)) {
+                DEBUGLOG(3, "ZSTD_buildCTable for LitLens failed");
+                stats.size = countSize;
+                return stats;
+            }
+            if (stats.LLtype == set_compressed)
+                stats.lastCountSize = countSize;
             op += countSize;
             assert(op <= oend);
     }   }
     /* build CTable for Offsets */
     {   unsigned max = MaxOff;
         size_t const mostFrequent = HIST_countFast_wksp(
-            count, &max, ofCodeTable, nbSeq, entropyWorkspace, entropyWkspSize);  /* can't fail */
+            countWorkspace, &max, ofCodeTable, nbSeq, entropyWorkspace, entropyWkspSize);  /* can't fail */
         /* We can only use the basic table if max <= DefaultMaxOff, otherwise the offsets are too large */
         ZSTD_defaultPolicy_e const defaultPolicy = (max <= DefaultMaxOff) ? ZSTD_defaultAllowed : ZSTD_defaultDisallowed;
         DEBUGLOG(5, "Building OF table");
-        nextEntropy->fse.offcode_repeatMode = prevEntropy->fse.offcode_repeatMode;
-        Offtype = ZSTD_selectEncodingType(&nextEntropy->fse.offcode_repeatMode,
-                                        count, max, mostFrequent, nbSeq,
-                                        OffFSELog, prevEntropy->fse.offcodeCTable,
+        nextEntropy->offcode_repeatMode = prevEntropy->offcode_repeatMode;
+        stats.Offtype = ZSTD_selectEncodingType(&nextEntropy->offcode_repeatMode,
+                                        countWorkspace, max, mostFrequent, nbSeq,
+                                        OffFSELog, prevEntropy->offcodeCTable,
                                         OF_defaultNorm, OF_defaultNormLog,
                                         defaultPolicy, strategy);
-        assert(!(Offtype < set_compressed && nextEntropy->fse.offcode_repeatMode != FSE_repeat_none)); /* We don't copy tables */
+        assert(!(stats.Offtype < set_compressed && nextEntropy->offcode_repeatMode != FSE_repeat_none)); /* We don't copy tables */
         {   size_t const countSize = ZSTD_buildCTable(
                 op, (size_t)(oend - op),
-                CTable_OffsetBits, OffFSELog, (symbolEncodingType_e)Offtype,
-                count, max, ofCodeTable, nbSeq,
+                CTable_OffsetBits, OffFSELog, (symbolEncodingType_e)stats.Offtype,
+                countWorkspace, max, ofCodeTable, nbSeq,
                 OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff,
-                prevEntropy->fse.offcodeCTable,
-                sizeof(prevEntropy->fse.offcodeCTable),
+                prevEntropy->offcodeCTable,
+                sizeof(prevEntropy->offcodeCTable),
                 entropyWorkspace, entropyWkspSize);
-            FORWARD_IF_ERROR(countSize, "ZSTD_buildCTable for Offsets failed");
-            if (Offtype == set_compressed)
-                lastNCount = op;
+            if (ZSTD_isError(countSize)) {
+                DEBUGLOG(3, "ZSTD_buildCTable for Offsets failed");
+                stats.size = countSize;
+                return stats;
+            }
+            if (stats.Offtype == set_compressed)
+                stats.lastCountSize = countSize;
             op += countSize;
             assert(op <= oend);
     }   }
     /* build CTable for MatchLengths */
     {   unsigned max = MaxML;
         size_t const mostFrequent = HIST_countFast_wksp(
-            count, &max, mlCodeTable, nbSeq, entropyWorkspace, entropyWkspSize);   /* can't fail */
+            countWorkspace, &max, mlCodeTable, nbSeq, entropyWorkspace, entropyWkspSize);   /* can't fail */
         DEBUGLOG(5, "Building ML table (remaining space : %i)", (int)(oend-op));
-        nextEntropy->fse.matchlength_repeatMode = prevEntropy->fse.matchlength_repeatMode;
-        MLtype = ZSTD_selectEncodingType(&nextEntropy->fse.matchlength_repeatMode,
-                                        count, max, mostFrequent, nbSeq,
-                                        MLFSELog, prevEntropy->fse.matchlengthCTable,
+        nextEntropy->matchlength_repeatMode = prevEntropy->matchlength_repeatMode;
+        stats.MLtype = ZSTD_selectEncodingType(&nextEntropy->matchlength_repeatMode,
+                                        countWorkspace, max, mostFrequent, nbSeq,
+                                        MLFSELog, prevEntropy->matchlengthCTable,
                                         ML_defaultNorm, ML_defaultNormLog,
                                         ZSTD_defaultAllowed, strategy);
-        assert(!(MLtype < set_compressed && nextEntropy->fse.matchlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */
+        assert(!(stats.MLtype < set_compressed && nextEntropy->matchlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */
         {   size_t const countSize = ZSTD_buildCTable(
                 op, (size_t)(oend - op),
-                CTable_MatchLength, MLFSELog, (symbolEncodingType_e)MLtype,
-                count, max, mlCodeTable, nbSeq,
+                CTable_MatchLength, MLFSELog, (symbolEncodingType_e)stats.MLtype,
+                countWorkspace, max, mlCodeTable, nbSeq,
                 ML_defaultNorm, ML_defaultNormLog, MaxML,
-                prevEntropy->fse.matchlengthCTable,
-                sizeof(prevEntropy->fse.matchlengthCTable),
+                prevEntropy->matchlengthCTable,
+                sizeof(prevEntropy->matchlengthCTable),
                 entropyWorkspace, entropyWkspSize);
-            FORWARD_IF_ERROR(countSize, "ZSTD_buildCTable for MatchLengths failed");
-            if (MLtype == set_compressed)
-                lastNCount = op;
+            if (ZSTD_isError(countSize)) {
+                DEBUGLOG(3, "ZSTD_buildCTable for MatchLengths failed");
+                stats.size = countSize;
+                return stats;
+            }
+            if (stats.MLtype == set_compressed)
+                stats.lastCountSize = countSize;
             op += countSize;
             assert(op <= oend);
     }   }
+    stats.size = (size_t)(op-ostart);
+    return stats;
+}
+
+/* ZSTD_entropyCompressSeqStore_internal():
+ * compresses both literals and sequences
+ * Returns compressed size of block, or a zstd error.
+ */
+#define SUSPECT_UNCOMPRESSIBLE_LITERAL_RATIO 20
+MEM_STATIC size_t
+ZSTD_entropyCompressSeqStore_internal(
+                        const seqStore_t* seqStorePtr,
+                        const ZSTD_entropyCTables_t* prevEntropy,
+                              ZSTD_entropyCTables_t* nextEntropy,
+                        const ZSTD_CCtx_params* cctxParams,
+                              void* dst, size_t dstCapacity,
+                              void* entropyWorkspace, size_t entropyWkspSize,
+                        const int bmi2)
+{
+    ZSTD_strategy const strategy = cctxParams->cParams.strategy;
+    unsigned* count = (unsigned*)entropyWorkspace;
+    FSE_CTable* CTable_LitLength = nextEntropy->fse.litlengthCTable;
+    FSE_CTable* CTable_OffsetBits = nextEntropy->fse.offcodeCTable;
+    FSE_CTable* CTable_MatchLength = nextEntropy->fse.matchlengthCTable;
+    const seqDef* const sequences = seqStorePtr->sequencesStart;
+    const size_t nbSeq = (size_t)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
+    const BYTE* const ofCodeTable = seqStorePtr->ofCode;
+    const BYTE* const llCodeTable = seqStorePtr->llCode;
+    const BYTE* const mlCodeTable = seqStorePtr->mlCode;
+    BYTE* const ostart = (BYTE*)dst;
+    BYTE* const oend = ostart + dstCapacity;
+    BYTE* op = ostart;
+    size_t lastCountSize;
+    int longOffsets = 0;
+
+    entropyWorkspace = count + (MaxSeq + 1);
+    entropyWkspSize -= (MaxSeq + 1) * sizeof(*count);
+
+    DEBUGLOG(5, "ZSTD_entropyCompressSeqStore_internal (nbSeq=%zu, dstCapacity=%zu)", nbSeq, dstCapacity);
+    ZSTD_STATIC_ASSERT(HUF_WORKSPACE_SIZE >= (1<<MAX(MLFSELog,LLFSELog)));
+    assert(entropyWkspSize >= HUF_WORKSPACE_SIZE);
+
+    /* Compress literals */
+    {   const BYTE* const literals = seqStorePtr->litStart;
+        size_t const numSequences = (size_t)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
+        size_t const numLiterals = (size_t)(seqStorePtr->lit - seqStorePtr->litStart);
+        /* Base suspicion of uncompressibility on ratio of literals to sequences */
+        unsigned const suspectUncompressible = (numSequences == 0) || (numLiterals / numSequences >= SUSPECT_UNCOMPRESSIBLE_LITERAL_RATIO);
+        size_t const litSize = (size_t)(seqStorePtr->lit - literals);
+
+        size_t const cSize = ZSTD_compressLiterals(
+                                    op, dstCapacity,
+                                    literals, litSize,
+                                    entropyWorkspace, entropyWkspSize,
+                                    &prevEntropy->huf, &nextEntropy->huf,
+                                    cctxParams->cParams.strategy,
+                                    ZSTD_literalsCompressionIsDisabled(cctxParams),
+                                    suspectUncompressible, bmi2);
+        FORWARD_IF_ERROR(cSize, "ZSTD_compressLiterals failed");
+        assert(cSize <= dstCapacity);
+        op += cSize;
+    }
 
-    *seqHead = (BYTE)((LLtype<<6) + (Offtype<<4) + (MLtype<<2));
+    /* Sequences Header */
+    RETURN_ERROR_IF((oend-op) < 3 /*max nbSeq Size*/ + 1 /*seqHead*/,
+                    dstSize_tooSmall, "Can't fit seq hdr in output buf!");
+    if (nbSeq < 128) {
+        *op++ = (BYTE)nbSeq;
+    } else if (nbSeq < LONGNBSEQ) {
+        op[0] = (BYTE)((nbSeq>>8) + 0x80);
+        op[1] = (BYTE)nbSeq;
+        op+=2;
+    } else {
+        op[0]=0xFF;
+        MEM_writeLE16(op+1, (U16)(nbSeq - LONGNBSEQ));
+        op+=3;
+    }
+    assert(op <= oend);
+    if (nbSeq==0) {
+        /* Copy the old tables over as if we repeated them */
+        ZSTD_memcpy(&nextEntropy->fse, &prevEntropy->fse, sizeof(prevEntropy->fse));
+        return (size_t)(op - ostart);
+    }
+    {   BYTE* const seqHead = op++;
+        /* build stats for sequences */
+        const ZSTD_symbolEncodingTypeStats_t stats =
+                ZSTD_buildSequencesStatistics(seqStorePtr, nbSeq,
+                                             &prevEntropy->fse, &nextEntropy->fse,
+                                              op, oend,
+                                              strategy, count,
+                                              entropyWorkspace, entropyWkspSize);
+        FORWARD_IF_ERROR(stats.size, "ZSTD_buildSequencesStatistics failed!");
+        *seqHead = (BYTE)((stats.LLtype<<6) + (stats.Offtype<<4) + (stats.MLtype<<2));
+        lastCountSize = stats.lastCountSize;
+        op += stats.size;
+        longOffsets = stats.longOffsets;
+    }
 
     {   size_t const bitstreamSize = ZSTD_encodeSequences(
                                         op, (size_t)(oend - op),
@@ -2134,9 +2881,9 @@ ZSTD_compressSequences_internal(seqStore_t* seqStorePtr,
          * In this exceedingly rare case, we will simply emit an uncompressed
          * block, since it isn't worth optimizing.
          */
-        if (lastNCount && (op - lastNCount) < 4) {
-            /* NCountSize >= 2 && bitstreamSize > 0 ==> lastCountSize == 3 */
-            assert(op - lastNCount == 3);
+        if (lastCountSize && (lastCountSize + bitstreamSize) < 4) {
+            /* lastCountSize >= 2 && bitstreamSize > 0 ==> lastCountSize == 3 */
+            assert(lastCountSize + bitstreamSize == 3);
             DEBUGLOG(5, "Avoiding bug in zstd decoder in versions <= 1.3.4 by "
                         "emitting an uncompressed block.");
             return 0;
@@ -2148,16 +2895,17 @@ ZSTD_compressSequences_internal(seqStore_t* seqStorePtr,
 }
 
 MEM_STATIC size_t
-ZSTD_compressSequences(seqStore_t* seqStorePtr,
-                       const ZSTD_entropyCTables_t* prevEntropy,
-                             ZSTD_entropyCTables_t* nextEntropy,
-                       const ZSTD_CCtx_params* cctxParams,
-                             void* dst, size_t dstCapacity,
-                             size_t srcSize,
-                             void* entropyWorkspace, size_t entropyWkspSize,
-                             int bmi2)
-{
-    size_t const cSize = ZSTD_compressSequences_internal(
+ZSTD_entropyCompressSeqStore(
+                    const seqStore_t* seqStorePtr,
+                    const ZSTD_entropyCTables_t* prevEntropy,
+                          ZSTD_entropyCTables_t* nextEntropy,
+                    const ZSTD_CCtx_params* cctxParams,
+                          void* dst, size_t dstCapacity,
+                          size_t srcSize,
+                          void* entropyWorkspace, size_t entropyWkspSize,
+                          int bmi2)
+{
+    size_t const cSize = ZSTD_entropyCompressSeqStore_internal(
                             seqStorePtr, prevEntropy, nextEntropy, cctxParams,
                             dst, dstCapacity,
                             entropyWorkspace, entropyWkspSize, bmi2);
@@ -2165,24 +2913,30 @@ ZSTD_compressSequences(seqStore_t* seqStorePtr,
     /* When srcSize <= dstCapacity, there is enough space to write a raw uncompressed block.
      * Since we ran out of space, block must be not compressible, so fall back to raw uncompressed block.
      */
-    if ((cSize == ERROR(dstSize_tooSmall)) & (srcSize <= dstCapacity))
+    if ((cSize == ERROR(dstSize_tooSmall)) & (srcSize <= dstCapacity)) {
+        DEBUGLOG(4, "not enough dstCapacity (%zu) for ZSTD_entropyCompressSeqStore_internal()=> do not compress block", dstCapacity);
         return 0;  /* block not compressed */
-    FORWARD_IF_ERROR(cSize, "ZSTD_compressSequences_internal failed");
+    }
+    FORWARD_IF_ERROR(cSize, "ZSTD_entropyCompressSeqStore_internal failed");
 
     /* Check compressibility */
     {   size_t const maxCSize = srcSize - ZSTD_minGain(srcSize, cctxParams->cParams.strategy);
         if (cSize >= maxCSize) return 0;  /* block not compressed */
     }
-
+    DEBUGLOG(5, "ZSTD_entropyCompressSeqStore() cSize: %zu", cSize);
+    /* libzstd decoder before  > v1.5.4 is not compatible with compressed blocks of size ZSTD_BLOCKSIZE_MAX exactly.
+     * This restriction is indirectly already fulfilled by respecting ZSTD_minGain() condition above.
+     */
+    assert(cSize < ZSTD_BLOCKSIZE_MAX);
     return cSize;
 }
 
 /* ZSTD_selectBlockCompressor() :
  * Not static, but internal use only (used by long distance matcher)
  * assumption : strat is a valid strategy */
-ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_dictMode_e dictMode)
+ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_paramSwitch_e useRowMatchFinder, ZSTD_dictMode_e dictMode)
 {
-    static const ZSTD_blockCompressor blockCompressor[3][ZSTD_STRATEGY_MAX+1] = {
+    static const ZSTD_blockCompressor blockCompressor[4][ZSTD_STRATEGY_MAX+1] = {
         { ZSTD_compressBlock_fast  /* default for 0 */,
           ZSTD_compressBlock_fast,
           ZSTD_compressBlock_doubleFast,
@@ -2212,13 +2966,44 @@ ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_dictMo
           ZSTD_compressBlock_btlazy2_dictMatchState,
           ZSTD_compressBlock_btopt_dictMatchState,
           ZSTD_compressBlock_btultra_dictMatchState,
-          ZSTD_compressBlock_btultra_dictMatchState }
+          ZSTD_compressBlock_btultra_dictMatchState },
+        { NULL  /* default for 0 */,
+          NULL,
+          NULL,
+          ZSTD_compressBlock_greedy_dedicatedDictSearch,
+          ZSTD_compressBlock_lazy_dedicatedDictSearch,
+          ZSTD_compressBlock_lazy2_dedicatedDictSearch,
+          NULL,
+          NULL,
+          NULL,
+          NULL }
     };
     ZSTD_blockCompressor selectedCompressor;
     ZSTD_STATIC_ASSERT((unsigned)ZSTD_fast == 1);
 
     assert(ZSTD_cParam_withinBounds(ZSTD_c_strategy, strat));
-    selectedCompressor = blockCompressor[(int)dictMode][(int)strat];
+    DEBUGLOG(4, "Selected block compressor: dictMode=%d strat=%d rowMatchfinder=%d", (int)dictMode, (int)strat, (int)useRowMatchFinder);
+    if (ZSTD_rowMatchFinderUsed(strat, useRowMatchFinder)) {
+        static const ZSTD_blockCompressor rowBasedBlockCompressors[4][3] = {
+            { ZSTD_compressBlock_greedy_row,
+            ZSTD_compressBlock_lazy_row,
+            ZSTD_compressBlock_lazy2_row },
+            { ZSTD_compressBlock_greedy_extDict_row,
+            ZSTD_compressBlock_lazy_extDict_row,
+            ZSTD_compressBlock_lazy2_extDict_row },
+            { ZSTD_compressBlock_greedy_dictMatchState_row,
+            ZSTD_compressBlock_lazy_dictMatchState_row,
+            ZSTD_compressBlock_lazy2_dictMatchState_row },
+            { ZSTD_compressBlock_greedy_dedicatedDictSearch_row,
+            ZSTD_compressBlock_lazy_dedicatedDictSearch_row,
+            ZSTD_compressBlock_lazy2_dedicatedDictSearch_row }
+        };
+        DEBUGLOG(4, "Selecting a row-based matchfinder");
+        assert(useRowMatchFinder != ZSTD_ps_auto);
+        selectedCompressor = rowBasedBlockCompressors[(int)dictMode][(int)strat - (int)ZSTD_greedy];
+    } else {
+        selectedCompressor = blockCompressor[(int)dictMode][(int)strat];
+    }
     assert(selectedCompressor != NULL);
     return selectedCompressor;
 }
@@ -2226,7 +3011,7 @@ ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_dictMo
 static void ZSTD_storeLastLiterals(seqStore_t* seqStorePtr,
                                    const BYTE* anchor, size_t lastLLSize)
 {
-    memcpy(seqStorePtr->lit, anchor, lastLLSize);
+    ZSTD_memcpy(seqStorePtr->lit, anchor, lastLLSize);
     seqStorePtr->lit += lastLLSize;
 }
 
@@ -2234,7 +3019,73 @@ void ZSTD_resetSeqStore(seqStore_t* ssPtr)
 {
     ssPtr->lit = ssPtr->litStart;
     ssPtr->sequences = ssPtr->sequencesStart;
-    ssPtr->longLengthID = 0;
+    ssPtr->longLengthType = ZSTD_llt_none;
+}
+
+/* ZSTD_postProcessSequenceProducerResult() :
+ * Validates and post-processes sequences obtained through the external matchfinder API:
+ *   - Checks whether nbExternalSeqs represents an error condition.
+ *   - Appends a block delimiter to outSeqs if one is not already present.
+ *     See zstd.h for context regarding block delimiters.
+ * Returns the number of sequences after post-processing, or an error code. */
+static size_t ZSTD_postProcessSequenceProducerResult(
+    ZSTD_Sequence* outSeqs, size_t nbExternalSeqs, size_t outSeqsCapacity, size_t srcSize
+) {
+    RETURN_ERROR_IF(
+        nbExternalSeqs > outSeqsCapacity,
+        sequenceProducer_failed,
+        "External sequence producer returned error code %lu",
+        (unsigned long)nbExternalSeqs
+    );
+
+    RETURN_ERROR_IF(
+        nbExternalSeqs == 0 && srcSize > 0,
+        sequenceProducer_failed,
+        "Got zero sequences from external sequence producer for a non-empty src buffer!"
+    );
+
+    if (srcSize == 0) {
+        ZSTD_memset(&outSeqs[0], 0, sizeof(ZSTD_Sequence));
+        return 1;
+    }
+
+    {
+        ZSTD_Sequence const lastSeq = outSeqs[nbExternalSeqs - 1];
+
+        /* We can return early if lastSeq is already a block delimiter. */
+        if (lastSeq.offset == 0 && lastSeq.matchLength == 0) {
+            return nbExternalSeqs;
+        }
+
+        /* This error condition is only possible if the external matchfinder
+         * produced an invalid parse, by definition of ZSTD_sequenceBound(). */
+        RETURN_ERROR_IF(
+            nbExternalSeqs == outSeqsCapacity,
+            sequenceProducer_failed,
+            "nbExternalSeqs == outSeqsCapacity but lastSeq is not a block delimiter!"
+        );
+
+        /* lastSeq is not a block delimiter, so we need to append one. */
+        ZSTD_memset(&outSeqs[nbExternalSeqs], 0, sizeof(ZSTD_Sequence));
+        return nbExternalSeqs + 1;
+    }
+}
+
+/* ZSTD_fastSequenceLengthSum() :
+ * Returns sum(litLen) + sum(matchLen) + lastLits for *seqBuf*.
+ * Similar to another function in zstd_compress.c (determine_blockSize),
+ * except it doesn't check for a block delimiter to end summation.
+ * Removing the early exit allows the compiler to auto-vectorize (https://godbolt.org/z/cY1cajz9P).
+ * This function can be deleted and replaced by determine_blockSize after we resolve issue #3456. */
+static size_t ZSTD_fastSequenceLengthSum(ZSTD_Sequence const* seqBuf, size_t seqBufSize) {
+    size_t matchLenSum, litLenSum, i;
+    matchLenSum = 0;
+    litLenSum = 0;
+    for (i = 0; i < seqBufSize; i++) {
+        litLenSum += seqBuf[i].litLength;
+        matchLenSum += seqBuf[i].matchLength;
+    }
+    return litLenSum + matchLenSum;
 }
 
 typedef enum { ZSTDbss_compress, ZSTDbss_noCompress } ZSTD_buildSeqStore_e;
@@ -2246,8 +3097,14 @@ static size_t ZSTD_buildSeqStore(ZSTD_CCtx* zc, const void* src, size_t srcSize)
     assert(srcSize <= ZSTD_BLOCKSIZE_MAX);
     /* Assert that we have correctly flushed the ctx params into the ms's copy */
     ZSTD_assertEqualCParams(zc->appliedParams.cParams, ms->cParams);
-    if (srcSize < MIN_CBLOCK_SIZE+ZSTD_blockHeaderSize+1) {
-        ZSTD_ldm_skipSequences(&zc->externSeqStore, srcSize, zc->appliedParams.cParams.minMatch);
+    /* TODO: See 3090. We reduced MIN_CBLOCK_SIZE from 3 to 2 so to compensate we are adding
+     * additional 1. We need to revisit and change this logic to be more consistent */
+    if (srcSize < MIN_CBLOCK_SIZE+ZSTD_blockHeaderSize+1+1) {
+        if (zc->appliedParams.cParams.strategy >= ZSTD_btopt) {
+            ZSTD_ldm_skipRawSeqStoreBytes(&zc->externSeqStore, srcSize);
+        } else {
+            ZSTD_ldm_skipSequences(&zc->externSeqStore, srcSize, zc->appliedParams.cParams.minMatch);
+        }
         return ZSTDbss_noCompress; /* don't even attempt compression below a certain srcSize */
     }
     ZSTD_resetSeqStore(&(zc->seqStore));
@@ -2263,10 +3120,10 @@ static size_t ZSTD_buildSeqStore(ZSTD_CCtx* zc, const void* src, size_t srcSize)
     /* limited update after a very long match */
     {   const BYTE* const base = ms->window.base;
         const BYTE* const istart = (const BYTE*)src;
-        const U32 current = (U32)(istart-base);
+        const U32 curr = (U32)(istart-base);
         if (sizeof(ptrdiff_t)==8) assert(istart - base < (ptrdiff_t)(U32)(-1));   /* ensure no overflow */
-        if (current > ms->nextToUpdate + 384)
-            ms->nextToUpdate = current - MIN(192, (U32)(current - ms->nextToUpdate - 384));
+        if (curr > ms->nextToUpdate + 384)
+            ms->nextToUpdate = curr - MIN(192, (U32)(curr - ms->nextToUpdate - 384));
     }
 
     /* select and store sequences */
@@ -2277,16 +3134,34 @@ static size_t ZSTD_buildSeqStore(ZSTD_CCtx* zc, const void* src, size_t srcSize)
                 zc->blockState.nextCBlock->rep[i] = zc->blockState.prevCBlock->rep[i];
         }
         if (zc->externSeqStore.pos < zc->externSeqStore.size) {
-            assert(!zc->appliedParams.ldmParams.enableLdm);
+            assert(zc->appliedParams.ldmParams.enableLdm == ZSTD_ps_disable);
+
+            /* External matchfinder + LDM is technically possible, just not implemented yet.
+             * We need to revisit soon and implement it. */
+            RETURN_ERROR_IF(
+                zc->appliedParams.useSequenceProducer,
+                parameter_combination_unsupported,
+                "Long-distance matching with external sequence producer enabled is not currently supported."
+            );
+
             /* Updates ldmSeqStore.pos */
             lastLLSize =
                 ZSTD_ldm_blockCompress(&zc->externSeqStore,
                                        ms, &zc->seqStore,
                                        zc->blockState.nextCBlock->rep,
+                                       zc->appliedParams.useRowMatchFinder,
                                        src, srcSize);
             assert(zc->externSeqStore.pos <= zc->externSeqStore.size);
-        } else if (zc->appliedParams.ldmParams.enableLdm) {
-            rawSeqStore_t ldmSeqStore = {NULL, 0, 0, 0};
+        } else if (zc->appliedParams.ldmParams.enableLdm == ZSTD_ps_enable) {
+            rawSeqStore_t ldmSeqStore = kNullRawSeqStore;
+
+            /* External matchfinder + LDM is technically possible, just not implemented yet.
+             * We need to revisit soon and implement it. */
+            RETURN_ERROR_IF(
+                zc->appliedParams.useSequenceProducer,
+                parameter_combination_unsupported,
+                "Long-distance matching with external sequence producer enabled is not currently supported."
+            );
 
             ldmSeqStore.seq = zc->ldmSequences;
             ldmSeqStore.capacity = zc->maxNbLdmSequences;
@@ -2299,10 +3174,75 @@ static size_t ZSTD_buildSeqStore(ZSTD_CCtx* zc, const void* src, size_t srcSize)
                 ZSTD_ldm_blockCompress(&ldmSeqStore,
                                        ms, &zc->seqStore,
                                        zc->blockState.nextCBlock->rep,
+                                       zc->appliedParams.useRowMatchFinder,
                                        src, srcSize);
             assert(ldmSeqStore.pos == ldmSeqStore.size);
-        } else {   /* not long range mode */
-            ZSTD_blockCompressor const blockCompressor = ZSTD_selectBlockCompressor(zc->appliedParams.cParams.strategy, dictMode);
+        } else if (zc->appliedParams.useSequenceProducer) {
+            assert(
+                zc->externalMatchCtx.seqBufferCapacity >= ZSTD_sequenceBound(srcSize)
+            );
+            assert(zc->externalMatchCtx.mFinder != NULL);
+
+            {   U32 const windowSize = (U32)1 << zc->appliedParams.cParams.windowLog;
+
+                size_t const nbExternalSeqs = (zc->externalMatchCtx.mFinder)(
+                    zc->externalMatchCtx.mState,
+                    zc->externalMatchCtx.seqBuffer,
+                    zc->externalMatchCtx.seqBufferCapacity,
+                    src, srcSize,
+                    NULL, 0,  /* dict and dictSize, currently not supported */
+                    zc->appliedParams.compressionLevel,
+                    windowSize
+                );
+
+                size_t const nbPostProcessedSeqs = ZSTD_postProcessSequenceProducerResult(
+                    zc->externalMatchCtx.seqBuffer,
+                    nbExternalSeqs,
+                    zc->externalMatchCtx.seqBufferCapacity,
+                    srcSize
+                );
+
+                /* Return early if there is no error, since we don't need to worry about last literals */
+                if (!ZSTD_isError(nbPostProcessedSeqs)) {
+                    ZSTD_sequencePosition seqPos = {0,0,0};
+                    size_t const seqLenSum = ZSTD_fastSequenceLengthSum(zc->externalMatchCtx.seqBuffer, nbPostProcessedSeqs);
+                    RETURN_ERROR_IF(seqLenSum > srcSize, externalSequences_invalid, "External sequences imply too large a block!");
+                    FORWARD_IF_ERROR(
+                        ZSTD_copySequencesToSeqStoreExplicitBlockDelim(
+                            zc, &seqPos,
+                            zc->externalMatchCtx.seqBuffer, nbPostProcessedSeqs,
+                            src, srcSize,
+                            zc->appliedParams.searchForExternalRepcodes
+                        ),
+                        "Failed to copy external sequences to seqStore!"
+                    );
+                    ms->ldmSeqStore = NULL;
+                    DEBUGLOG(5, "Copied %lu sequences from external sequence producer to internal seqStore.", (unsigned long)nbExternalSeqs);
+                    return ZSTDbss_compress;
+                }
+
+                /* Propagate the error if fallback is disabled */
+                if (!zc->appliedParams.enableMatchFinderFallback) {
+                    return nbPostProcessedSeqs;
+                }
+
+                /* Fallback to software matchfinder */
+                {   ZSTD_blockCompressor const blockCompressor = ZSTD_selectBlockCompressor(zc->appliedParams.cParams.strategy,
+                                                                                            zc->appliedParams.useRowMatchFinder,
+                                                                                            dictMode);
+                    ms->ldmSeqStore = NULL;
+                    DEBUGLOG(
+                        5,
+                        "External sequence producer returned error code %lu. Falling back to internal parser.",
+                        (unsigned long)nbExternalSeqs
+                    );
+                    lastLLSize = blockCompressor(ms, &zc->seqStore, zc->blockState.nextCBlock->rep, src, srcSize);
+            }   }
+        } else {   /* not long range mode and no external matchfinder */
+            ZSTD_blockCompressor const blockCompressor = ZSTD_selectBlockCompressor(zc->appliedParams.cParams.strategy,
+                                                                                    zc->appliedParams.useRowMatchFinder,
+                                                                                    dictMode);
+            ms->ldmSeqStore = NULL;
             lastLLSize = blockCompressor(ms, &zc->seqStore, zc->blockState.nextCBlock->rep, src, srcSize);
         }
         {   const BYTE* const lastLiterals = (const BYTE*)src + srcSize - lastLLSize;
@@ -2314,59 +3254,76 @@ static size_t ZSTD_buildSeqStore(ZSTD_CCtx* zc, const void* src, size_t srcSize)
 static void ZSTD_copyBlockSequences(ZSTD_CCtx* zc)
 {
     const seqStore_t* seqStore = ZSTD_getSeqStore(zc);
-    const seqDef* seqs = seqStore->sequencesStart;
-    size_t seqsSize = seqStore->sequences - seqs;
+    const seqDef* seqStoreSeqs = seqStore->sequencesStart;
+    size_t seqStoreSeqSize = seqStore->sequences - seqStoreSeqs;
+    size_t seqStoreLiteralsSize = (size_t)(seqStore->lit - seqStore->litStart);
+    size_t literalsRead = 0;
+    size_t lastLLSize;
 
     ZSTD_Sequence* outSeqs = &zc->seqCollector.seqStart[zc->seqCollector.seqIndex];
-    size_t i; size_t position; int repIdx;
+    size_t i;
+    repcodes_t updatedRepcodes;
 
     assert(zc->seqCollector.seqIndex + 1 < zc->seqCollector.maxSequences);
-    for (i = 0, position = 0; i < seqsSize; ++i) {
-        outSeqs[i].offset = seqs[i].offset;
-        outSeqs[i].litLength = seqs[i].litLength;
-        outSeqs[i].matchLength = seqs[i].matchLength + MINMATCH;
+    /* Ensure we have enough space for last literals "sequence" */
+    assert(zc->seqCollector.maxSequences >= seqStoreSeqSize + 1);
+    ZSTD_memcpy(updatedRepcodes.rep, zc->blockState.prevCBlock->rep, sizeof(repcodes_t));
+    for (i = 0; i < seqStoreSeqSize; ++i) {
+        U32 rawOffset = seqStoreSeqs[i].offBase - ZSTD_REP_NUM;
+        outSeqs[i].litLength = seqStoreSeqs[i].litLength;
+        outSeqs[i].matchLength = seqStoreSeqs[i].mlBase + MINMATCH;
+        outSeqs[i].rep = 0;
 
         if (i == seqStore->longLengthPos) {
-            if (seqStore->longLengthID == 1) {
+            if (seqStore->longLengthType == ZSTD_llt_literalLength) {
                 outSeqs[i].litLength += 0x10000;
-            } else if (seqStore->longLengthID == 2) {
+            } else if (seqStore->longLengthType == ZSTD_llt_matchLength) {
                 outSeqs[i].matchLength += 0x10000;
             }
         }
 
-        if (outSeqs[i].offset <= ZSTD_REP_NUM) {
-            outSeqs[i].rep = outSeqs[i].offset;
-            repIdx = (unsigned int)i - outSeqs[i].offset;
-
-            if (outSeqs[i].litLength == 0) {
-                if (outSeqs[i].offset < 3) {
-                    --repIdx;
+        if (seqStoreSeqs[i].offBase <= ZSTD_REP_NUM) {
+            /* Derive the correct offset corresponding to a repcode */
+            outSeqs[i].rep = seqStoreSeqs[i].offBase;
+            if (outSeqs[i].litLength != 0) {
+                rawOffset = updatedRepcodes.rep[outSeqs[i].rep - 1];
+            } else {
+                if (outSeqs[i].rep == 3) {
+                    rawOffset = updatedRepcodes.rep[0] - 1;
                 } else {
-                    repIdx = (unsigned int)i - 1;
+                    rawOffset = updatedRepcodes.rep[outSeqs[i].rep];
                 }
-                ++outSeqs[i].rep;
             }
-            assert(repIdx >= -3);
-            outSeqs[i].offset = repIdx >= 0 ? outSeqs[repIdx].offset : repStartValue[-repIdx - 1];
-            if (outSeqs[i].rep == 4) {
-                --outSeqs[i].offset;
-            }
-        } else {
-            outSeqs[i].offset -= ZSTD_REP_NUM;
         }
-
-        position += outSeqs[i].litLength;
-        outSeqs[i].matchPos = (unsigned int)position;
-        position += outSeqs[i].matchLength;
+        outSeqs[i].offset = rawOffset;
+        /* seqStoreSeqs[i].offset == offCode+1, and ZSTD_updateRep() expects offCode
+           so we provide seqStoreSeqs[i].offset - 1 */
+        ZSTD_updateRep(updatedRepcodes.rep,
+                       seqStoreSeqs[i].offBase,
+                       seqStoreSeqs[i].litLength == 0);
+        literalsRead += outSeqs[i].litLength;
     }
-    zc->seqCollector.seqIndex += seqsSize;
+    /* Insert last literals (if any exist) in the block as a sequence with ml == off == 0.
+     * If there are no last literals, then we'll emit (of: 0, ml: 0, ll: 0), which is a marker
+     * for the block boundary, according to the API.
+     */
+    assert(seqStoreLiteralsSize >= literalsRead);
+    lastLLSize = seqStoreLiteralsSize - literalsRead;
+    outSeqs[i].litLength = (U32)lastLLSize;
+    outSeqs[i].matchLength = outSeqs[i].offset = outSeqs[i].rep = 0;
+    seqStoreSeqSize++;
+    zc->seqCollector.seqIndex += seqStoreSeqSize;
 }
 
-size_t ZSTD_getSequences(ZSTD_CCtx* zc, ZSTD_Sequence* outSeqs,
-    size_t outSeqsSize, const void* src, size_t srcSize)
+size_t ZSTD_sequenceBound(size_t srcSize) {
+    return (srcSize / ZSTD_MINMATCH_MIN) + 1;
+}
+
+size_t ZSTD_generateSequences(ZSTD_CCtx* zc, ZSTD_Sequence* outSeqs,
+                              size_t outSeqsSize, const void* src, size_t srcSize)
 {
     const size_t dstCapacity = ZSTD_compressBound(srcSize);
-    void* dst = ZSTD_malloc(dstCapacity, ZSTD_defaultCMem);
+    void* dst = ZSTD_customMalloc(dstCapacity, ZSTD_defaultCMem);
     SeqCollector seqCollector;
 
     RETURN_ERROR_IF(dst == NULL, memory_allocation, "NULL pointer!");
@@ -2378,17 +3335,46 @@ size_t ZSTD_getSequences(ZSTD_CCtx* zc, ZSTD_Sequence* outSeqs,
     zc->seqCollector = seqCollector;
 
     ZSTD_compress2(zc, dst, dstCapacity, src, srcSize);
-    ZSTD_free(dst, ZSTD_defaultCMem);
+    ZSTD_customFree(dst, ZSTD_defaultCMem);
     return zc->seqCollector.seqIndex;
 }
 
-/* Returns true if the given block is a RLE block */
-static int ZSTD_isRLE(const BYTE *ip, size_t length) {
+size_t ZSTD_mergeBlockDelimiters(ZSTD_Sequence* sequences, size_t seqsSize) {
+    size_t in = 0;
+    size_t out = 0;
+    for (; in < seqsSize; ++in) {
+        if (sequences[in].offset == 0 && sequences[in].matchLength == 0) {
+            if (in != seqsSize - 1) {
+                sequences[in+1].litLength += sequences[in].litLength;
+            }
+        } else {
+            sequences[out] = sequences[in];
+            ++out;
+        }
+    }
+    return out;
+}
+
+/* Unrolled loop to read four size_ts of input at a time. Returns 1 if is RLE, 0 if not. */
+static int ZSTD_isRLE(const BYTE* src, size_t length) {
+    const BYTE* ip = src;
+    const BYTE value = ip[0];
+    const size_t valueST = (size_t)((U64)value * 0x0101010101010101ULL);
+    const size_t unrollSize = sizeof(size_t) * 4;
+    const size_t unrollMask = unrollSize - 1;
+    const size_t prefixLength = length & unrollMask;
     size_t i;
-    if (length < 2) return 1;
-    for (i = 1; i < length; ++i) {
-        if (ip[0] != ip[i]) return 0;
+    if (length == 1) return 1;
+    /* Check if prefix is RLE first before using unrolled loop */
+    if (prefixLength && ZSTD_count(ip+1, ip, ip+prefixLength) != prefixLength-1) {
+        return 0;
     }
+    for (i = prefixLength; i != length; i += unrollSize) {
+        size_t u;
+        for (u = 0; u < unrollSize; u += sizeof(size_t)) {
+            if (MEM_readST(ip + i + u) != valueST) {
+                return 0;
+    }   }   }
     return 1;
 }
 
@@ -2404,20 +3390,787 @@ static int ZSTD_maybeRLE(seqStore_t const* seqStore)
     return nbSeqs < 4 && nbLits < 10;
 }
 
-static void ZSTD_confirmRepcodesAndEntropyTables(ZSTD_CCtx* zc)
+static void
+ZSTD_blockState_confirmRepcodesAndEntropyTables(ZSTD_blockState_t* const bs)
+{
+    ZSTD_compressedBlockState_t* const tmp = bs->prevCBlock;
+    bs->prevCBlock = bs->nextCBlock;
+    bs->nextCBlock = tmp;
+}
+
+/* Writes the block header */
+static void
+writeBlockHeader(void* op, size_t cSize, size_t blockSize, U32 lastBlock)
+{
+    U32 const cBlockHeader = cSize == 1 ?
+                        lastBlock + (((U32)bt_rle)<<1) + (U32)(blockSize << 3) :
+                        lastBlock + (((U32)bt_compressed)<<1) + (U32)(cSize << 3);
+    MEM_writeLE24(op, cBlockHeader);
+    DEBUGLOG(3, "writeBlockHeader: cSize: %zu blockSize: %zu lastBlock: %u", cSize, blockSize, lastBlock);
+}
+
+/** ZSTD_buildBlockEntropyStats_literals() :
+ *  Builds entropy for the literals.
+ *  Stores literals block type (raw, rle, compressed, repeat) and
+ *  huffman description table to hufMetadata.
+ *  Requires ENTROPY_WORKSPACE_SIZE workspace
+ * @return : size of huffman description table, or an error code
+ */
+static size_t
+ZSTD_buildBlockEntropyStats_literals(void* const src, size_t srcSize,
+                               const ZSTD_hufCTables_t* prevHuf,
+                                     ZSTD_hufCTables_t* nextHuf,
+                                     ZSTD_hufCTablesMetadata_t* hufMetadata,
+                               const int literalsCompressionIsDisabled,
+                                     void* workspace, size_t wkspSize,
+                                     int hufFlags)
+{
+    BYTE* const wkspStart = (BYTE*)workspace;
+    BYTE* const wkspEnd = wkspStart + wkspSize;
+    BYTE* const countWkspStart = wkspStart;
+    unsigned* const countWksp = (unsigned*)workspace;
+    const size_t countWkspSize = (HUF_SYMBOLVALUE_MAX + 1) * sizeof(unsigned);
+    BYTE* const nodeWksp = countWkspStart + countWkspSize;
+    const size_t nodeWkspSize = (size_t)(wkspEnd - nodeWksp);
+    unsigned maxSymbolValue = HUF_SYMBOLVALUE_MAX;
+    unsigned huffLog = LitHufLog;
+    HUF_repeat repeat = prevHuf->repeatMode;
+    DEBUGLOG(5, "ZSTD_buildBlockEntropyStats_literals (srcSize=%zu)", srcSize);
+
+    /* Prepare nextEntropy assuming reusing the existing table */
+    ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
+
+    if (literalsCompressionIsDisabled) {
+        DEBUGLOG(5, "set_basic - disabled");
+        hufMetadata->hType = set_basic;
+        return 0;
+    }
+
+    /* small ? don't even attempt compression (speed opt) */
+#ifndef COMPRESS_LITERALS_SIZE_MIN
+# define COMPRESS_LITERALS_SIZE_MIN 63  /* heuristic */
+#endif
+    {   size_t const minLitSize = (prevHuf->repeatMode == HUF_repeat_valid) ? 6 : COMPRESS_LITERALS_SIZE_MIN;
+        if (srcSize <= minLitSize) {
+            DEBUGLOG(5, "set_basic - too small");
+            hufMetadata->hType = set_basic;
+            return 0;
+    }   }
+
+    /* Scan input and build symbol stats */
+    {   size_t const largest =
+            HIST_count_wksp (countWksp, &maxSymbolValue,
+                            (const BYTE*)src, srcSize,
+                            workspace, wkspSize);
+        FORWARD_IF_ERROR(largest, "HIST_count_wksp failed");
+        if (largest == srcSize) {
+            /* only one literal symbol */
+            DEBUGLOG(5, "set_rle");
+            hufMetadata->hType = set_rle;
+            return 0;
+        }
+        if (largest <= (srcSize >> 7)+4) {
+            /* heuristic: likely not compressible */
+            DEBUGLOG(5, "set_basic - no gain");
+            hufMetadata->hType = set_basic;
+            return 0;
+    }   }
+
+    /* Validate the previous Huffman table */
+    if (repeat == HUF_repeat_check
+      && !HUF_validateCTable((HUF_CElt const*)prevHuf->CTable, countWksp, maxSymbolValue)) {
+        repeat = HUF_repeat_none;
+    }
+
+    /* Build Huffman Tree */
+    ZSTD_memset(nextHuf->CTable, 0, sizeof(nextHuf->CTable));
+    huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue, nodeWksp, nodeWkspSize, nextHuf->CTable, countWksp, hufFlags);
+    assert(huffLog <= LitHufLog);
+    {   size_t const maxBits = HUF_buildCTable_wksp((HUF_CElt*)nextHuf->CTable, countWksp,
+                                                    maxSymbolValue, huffLog,
+                                                    nodeWksp, nodeWkspSize);
+        FORWARD_IF_ERROR(maxBits, "HUF_buildCTable_wksp");
+        huffLog = (U32)maxBits;
+    }
+    {   /* Build and write the CTable */
+        size_t const newCSize = HUF_estimateCompressedSize(
+                (HUF_CElt*)nextHuf->CTable, countWksp, maxSymbolValue);
+        size_t const hSize = HUF_writeCTable_wksp(
+                hufMetadata->hufDesBuffer, sizeof(hufMetadata->hufDesBuffer),
+                (HUF_CElt*)nextHuf->CTable, maxSymbolValue, huffLog,
+                nodeWksp, nodeWkspSize);
+        /* Check against repeating the previous CTable */
+        if (repeat != HUF_repeat_none) {
+            size_t const oldCSize = HUF_estimateCompressedSize(
+                    (HUF_CElt const*)prevHuf->CTable, countWksp, maxSymbolValue);
+            if (oldCSize < srcSize && (oldCSize <= hSize + newCSize || hSize + 12 >= srcSize)) {
+                DEBUGLOG(5, "set_repeat - smaller");
+                ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
+                hufMetadata->hType = set_repeat;
+                return 0;
+        }   }
+        if (newCSize + hSize >= srcSize) {
+            DEBUGLOG(5, "set_basic - no gains");
+            ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
+            hufMetadata->hType = set_basic;
+            return 0;
+        }
+        DEBUGLOG(5, "set_compressed (hSize=%u)", (U32)hSize);
+        hufMetadata->hType = set_compressed;
+        nextHuf->repeatMode = HUF_repeat_check;
+        return hSize;
+    }
+}
+
+
+/* ZSTD_buildDummySequencesStatistics():
+ * Returns a ZSTD_symbolEncodingTypeStats_t with all encoding types as set_basic,
+ * and updates nextEntropy to the appropriate repeatMode.
+ */
+static ZSTD_symbolEncodingTypeStats_t
+ZSTD_buildDummySequencesStatistics(ZSTD_fseCTables_t* nextEntropy)
+{
+    ZSTD_symbolEncodingTypeStats_t stats = {set_basic, set_basic, set_basic, 0, 0, 0};
+    nextEntropy->litlength_repeatMode = FSE_repeat_none;
+    nextEntropy->offcode_repeatMode = FSE_repeat_none;
+    nextEntropy->matchlength_repeatMode = FSE_repeat_none;
+    return stats;
+}
+
+/** ZSTD_buildBlockEntropyStats_sequences() :
+ *  Builds entropy for the sequences.
+ *  Stores symbol compression modes and fse table to fseMetadata.
+ *  Requires ENTROPY_WORKSPACE_SIZE wksp.
+ * @return : size of fse tables or error code */
+static size_t
+ZSTD_buildBlockEntropyStats_sequences(
+                const seqStore_t* seqStorePtr,
+                const ZSTD_fseCTables_t* prevEntropy,
+                      ZSTD_fseCTables_t* nextEntropy,
+                const ZSTD_CCtx_params* cctxParams,
+                      ZSTD_fseCTablesMetadata_t* fseMetadata,
+                      void* workspace, size_t wkspSize)
+{
+    ZSTD_strategy const strategy = cctxParams->cParams.strategy;
+    size_t const nbSeq = (size_t)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
+    BYTE* const ostart = fseMetadata->fseTablesBuffer;
+    BYTE* const oend = ostart + sizeof(fseMetadata->fseTablesBuffer);
+    BYTE* op = ostart;
+    unsigned* countWorkspace = (unsigned*)workspace;
+    unsigned* entropyWorkspace = countWorkspace + (MaxSeq + 1);
+    size_t entropyWorkspaceSize = wkspSize - (MaxSeq + 1) * sizeof(*countWorkspace);
+    ZSTD_symbolEncodingTypeStats_t stats;
+
+    DEBUGLOG(5, "ZSTD_buildBlockEntropyStats_sequences (nbSeq=%zu)", nbSeq);
+    stats = nbSeq != 0 ? ZSTD_buildSequencesStatistics(seqStorePtr, nbSeq,
+                                          prevEntropy, nextEntropy, op, oend,
+                                          strategy, countWorkspace,
+                                          entropyWorkspace, entropyWorkspaceSize)
+                       : ZSTD_buildDummySequencesStatistics(nextEntropy);
+    FORWARD_IF_ERROR(stats.size, "ZSTD_buildSequencesStatistics failed!");
+    fseMetadata->llType = (symbolEncodingType_e) stats.LLtype;
+    fseMetadata->ofType = (symbolEncodingType_e) stats.Offtype;
+    fseMetadata->mlType = (symbolEncodingType_e) stats.MLtype;
+    fseMetadata->lastCountSize = stats.lastCountSize;
+    return stats.size;
+}
+
+
+/** ZSTD_buildBlockEntropyStats() :
+ *  Builds entropy for the block.
+ *  Requires workspace size ENTROPY_WORKSPACE_SIZE
+ * @return : 0 on success, or an error code
+ *  Note : also employed in superblock
+ */
+size_t ZSTD_buildBlockEntropyStats(
+            const seqStore_t* seqStorePtr,
+            const ZSTD_entropyCTables_t* prevEntropy,
+                  ZSTD_entropyCTables_t* nextEntropy,
+            const ZSTD_CCtx_params* cctxParams,
+                  ZSTD_entropyCTablesMetadata_t* entropyMetadata,
+                  void* workspace, size_t wkspSize)
+{
+    size_t const litSize = (size_t)(seqStorePtr->lit - seqStorePtr->litStart);
+    int const huf_useOptDepth = (cctxParams->cParams.strategy >= HUF_OPTIMAL_DEPTH_THRESHOLD);
+    int const hufFlags = huf_useOptDepth ? HUF_flags_optimalDepth : 0;
+
+    entropyMetadata->hufMetadata.hufDesSize =
+        ZSTD_buildBlockEntropyStats_literals(seqStorePtr->litStart, litSize,
+                                            &prevEntropy->huf, &nextEntropy->huf,
+                                            &entropyMetadata->hufMetadata,
+                                            ZSTD_literalsCompressionIsDisabled(cctxParams),
+                                            workspace, wkspSize, hufFlags);
+
+    FORWARD_IF_ERROR(entropyMetadata->hufMetadata.hufDesSize, "ZSTD_buildBlockEntropyStats_literals failed");
+    entropyMetadata->fseMetadata.fseTablesSize =
+        ZSTD_buildBlockEntropyStats_sequences(seqStorePtr,
+                                              &prevEntropy->fse, &nextEntropy->fse,
+                                              cctxParams,
+                                              &entropyMetadata->fseMetadata,
+                                              workspace, wkspSize);
+    FORWARD_IF_ERROR(entropyMetadata->fseMetadata.fseTablesSize, "ZSTD_buildBlockEntropyStats_sequences failed");
+    return 0;
+}
+
+/* Returns the size estimate for the literals section (header + content) of a block */
+static size_t
+ZSTD_estimateBlockSize_literal(const BYTE* literals, size_t litSize,
+                               const ZSTD_hufCTables_t* huf,
+                               const ZSTD_hufCTablesMetadata_t* hufMetadata,
+                               void* workspace, size_t wkspSize,
+                               int writeEntropy)
+{
+    unsigned* const countWksp = (unsigned*)workspace;
+    unsigned maxSymbolValue = HUF_SYMBOLVALUE_MAX;
+    size_t literalSectionHeaderSize = 3 + (litSize >= 1 KB) + (litSize >= 16 KB);
+    U32 singleStream = litSize < 256;
+
+    if (hufMetadata->hType == set_basic) return litSize;
+    else if (hufMetadata->hType == set_rle) return 1;
+    else if (hufMetadata->hType == set_compressed || hufMetadata->hType == set_repeat) {
+        size_t const largest = HIST_count_wksp (countWksp, &maxSymbolValue, (const BYTE*)literals, litSize, workspace, wkspSize);
+        if (ZSTD_isError(largest)) return litSize;
+        {   size_t cLitSizeEstimate = HUF_estimateCompressedSize((const HUF_CElt*)huf->CTable, countWksp, maxSymbolValue);
+            if (writeEntropy) cLitSizeEstimate += hufMetadata->hufDesSize;
+            if (!singleStream) cLitSizeEstimate += 6; /* multi-stream huffman uses 6-byte jump table */
+            return cLitSizeEstimate + literalSectionHeaderSize;
+    }   }
+    assert(0); /* impossible */
+    return 0;
+}
+
+/* Returns the size estimate for the FSE-compressed symbols (of, ml, ll) of a block */
+static size_t
+ZSTD_estimateBlockSize_symbolType(symbolEncodingType_e type,
+                    const BYTE* codeTable, size_t nbSeq, unsigned maxCode,
+                    const FSE_CTable* fseCTable,
+                    const U8* additionalBits,
+                    short const* defaultNorm, U32 defaultNormLog, U32 defaultMax,
+                    void* workspace, size_t wkspSize)
+{
+    unsigned* const countWksp = (unsigned*)workspace;
+    const BYTE* ctp = codeTable;
+    const BYTE* const ctStart = ctp;
+    const BYTE* const ctEnd = ctStart + nbSeq;
+    size_t cSymbolTypeSizeEstimateInBits = 0;
+    unsigned max = maxCode;
+
+    HIST_countFast_wksp(countWksp, &max, codeTable, nbSeq, workspace, wkspSize);  /* can't fail */
+    if (type == set_basic) {
+        /* We selected this encoding type, so it must be valid. */
+        assert(max <= defaultMax);
+        (void)defaultMax;
+        cSymbolTypeSizeEstimateInBits = ZSTD_crossEntropyCost(defaultNorm, defaultNormLog, countWksp, max);
+    } else if (type == set_rle) {
+        cSymbolTypeSizeEstimateInBits = 0;
+    } else if (type == set_compressed || type == set_repeat) {
+        cSymbolTypeSizeEstimateInBits = ZSTD_fseBitCost(fseCTable, countWksp, max);
+    }
+    if (ZSTD_isError(cSymbolTypeSizeEstimateInBits)) {
+        return nbSeq * 10;
+    }
+    while (ctp < ctEnd) {
+        if (additionalBits) cSymbolTypeSizeEstimateInBits += additionalBits[*ctp];
+        else cSymbolTypeSizeEstimateInBits += *ctp; /* for offset, offset code is also the number of additional bits */
+        ctp++;
+    }
+    return cSymbolTypeSizeEstimateInBits >> 3;
+}
+
+/* Returns the size estimate for the sequences section (header + content) of a block */
+static size_t
+ZSTD_estimateBlockSize_sequences(const BYTE* ofCodeTable,
+                                 const BYTE* llCodeTable,
+                                 const BYTE* mlCodeTable,
+                                 size_t nbSeq,
+                                 const ZSTD_fseCTables_t* fseTables,
+                                 const ZSTD_fseCTablesMetadata_t* fseMetadata,
+                                 void* workspace, size_t wkspSize,
+                                 int writeEntropy)
+{
+    size_t sequencesSectionHeaderSize = 1 /* seqHead */ + 1 /* min seqSize size */ + (nbSeq >= 128) + (nbSeq >= LONGNBSEQ);
+    size_t cSeqSizeEstimate = 0;
+    cSeqSizeEstimate += ZSTD_estimateBlockSize_symbolType(fseMetadata->ofType, ofCodeTable, nbSeq, MaxOff,
+                                    fseTables->offcodeCTable, NULL,
+                                    OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff,
+                                    workspace, wkspSize);
+    cSeqSizeEstimate += ZSTD_estimateBlockSize_symbolType(fseMetadata->llType, llCodeTable, nbSeq, MaxLL,
+                                    fseTables->litlengthCTable, LL_bits,
+                                    LL_defaultNorm, LL_defaultNormLog, MaxLL,
+                                    workspace, wkspSize);
+    cSeqSizeEstimate += ZSTD_estimateBlockSize_symbolType(fseMetadata->mlType, mlCodeTable, nbSeq, MaxML,
+                                    fseTables->matchlengthCTable, ML_bits,
+                                    ML_defaultNorm, ML_defaultNormLog, MaxML,
+                                    workspace, wkspSize);
+    if (writeEntropy) cSeqSizeEstimate += fseMetadata->fseTablesSize;
+    return cSeqSizeEstimate + sequencesSectionHeaderSize;
+}
+
+/* Returns the size estimate for a given stream of literals, of, ll, ml */
+static size_t
+ZSTD_estimateBlockSize(const BYTE* literals, size_t litSize,
+                       const BYTE* ofCodeTable,
+                       const BYTE* llCodeTable,
+                       const BYTE* mlCodeTable,
+                       size_t nbSeq,
+                       const ZSTD_entropyCTables_t* entropy,
+                       const ZSTD_entropyCTablesMetadata_t* entropyMetadata,
+                       void* workspace, size_t wkspSize,
+                       int writeLitEntropy, int writeSeqEntropy)
+{
+    size_t const literalsSize = ZSTD_estimateBlockSize_literal(literals, litSize,
+                                    &entropy->huf, &entropyMetadata->hufMetadata,
+                                    workspace, wkspSize, writeLitEntropy);
+    size_t const seqSize = ZSTD_estimateBlockSize_sequences(ofCodeTable, llCodeTable, mlCodeTable,
+                                    nbSeq, &entropy->fse, &entropyMetadata->fseMetadata,
+                                    workspace, wkspSize, writeSeqEntropy);
+    return seqSize + literalsSize + ZSTD_blockHeaderSize;
+}
+
+/* Builds entropy statistics and uses them for blocksize estimation.
+ *
+ * @return: estimated compressed size of the seqStore, or a zstd error.
+ */
+static size_t
+ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(seqStore_t* seqStore, ZSTD_CCtx* zc)
+{
+    ZSTD_entropyCTablesMetadata_t* const entropyMetadata = &zc->blockSplitCtx.entropyMetadata;
+    DEBUGLOG(6, "ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize()");
+    FORWARD_IF_ERROR(ZSTD_buildBlockEntropyStats(seqStore,
+                    &zc->blockState.prevCBlock->entropy,
+                    &zc->blockState.nextCBlock->entropy,
+                    &zc->appliedParams,
+                    entropyMetadata,
+                    zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE), "");
+    return ZSTD_estimateBlockSize(
+                    seqStore->litStart, (size_t)(seqStore->lit - seqStore->litStart),
+                    seqStore->ofCode, seqStore->llCode, seqStore->mlCode,
+                    (size_t)(seqStore->sequences - seqStore->sequencesStart),
+                    &zc->blockState.nextCBlock->entropy,
+                    entropyMetadata,
+                    zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE,
+                    (int)(entropyMetadata->hufMetadata.hType == set_compressed), 1);
+}
+
+/* Returns literals bytes represented in a seqStore */
+static size_t ZSTD_countSeqStoreLiteralsBytes(const seqStore_t* const seqStore)
+{
+    size_t literalsBytes = 0;
+    size_t const nbSeqs = (size_t)(seqStore->sequences - seqStore->sequencesStart);
+    size_t i;
+    for (i = 0; i < nbSeqs; ++i) {
+        seqDef const seq = seqStore->sequencesStart[i];
+        literalsBytes += seq.litLength;
+        if (i == seqStore->longLengthPos && seqStore->longLengthType == ZSTD_llt_literalLength) {
+            literalsBytes += 0x10000;
+    }   }
+    return literalsBytes;
+}
+
+/* Returns match bytes represented in a seqStore */
+static size_t ZSTD_countSeqStoreMatchBytes(const seqStore_t* const seqStore)
+{
+    size_t matchBytes = 0;
+    size_t const nbSeqs = (size_t)(seqStore->sequences - seqStore->sequencesStart);
+    size_t i;
+    for (i = 0; i < nbSeqs; ++i) {
+        seqDef seq = seqStore->sequencesStart[i];
+        matchBytes += seq.mlBase + MINMATCH;
+        if (i == seqStore->longLengthPos && seqStore->longLengthType == ZSTD_llt_matchLength) {
+            matchBytes += 0x10000;
+    }   }
+    return matchBytes;
+}
+
+/* Derives the seqStore that is a chunk of the originalSeqStore from [startIdx, endIdx).
+ * Stores the result in resultSeqStore.
+ */
+static void ZSTD_deriveSeqStoreChunk(seqStore_t* resultSeqStore,
+                               const seqStore_t* originalSeqStore,
+                                     size_t startIdx, size_t endIdx)
+{
+    *resultSeqStore = *originalSeqStore;
+    if (startIdx > 0) {
+        resultSeqStore->sequences = originalSeqStore->sequencesStart + startIdx;
+        resultSeqStore->litStart += ZSTD_countSeqStoreLiteralsBytes(resultSeqStore);
+    }
+
+    /* Move longLengthPos into the correct position if necessary */
+    if (originalSeqStore->longLengthType != ZSTD_llt_none) {
+        if (originalSeqStore->longLengthPos < startIdx || originalSeqStore->longLengthPos > endIdx) {
+            resultSeqStore->longLengthType = ZSTD_llt_none;
+        } else {
+            resultSeqStore->longLengthPos -= (U32)startIdx;
+        }
+    }
+    resultSeqStore->sequencesStart = originalSeqStore->sequencesStart + startIdx;
+    resultSeqStore->sequences = originalSeqStore->sequencesStart + endIdx;
+    if (endIdx == (size_t)(originalSeqStore->sequences - originalSeqStore->sequencesStart)) {
+        /* This accounts for possible last literals if the derived chunk reaches the end of the block */
+        assert(resultSeqStore->lit == originalSeqStore->lit);
+    } else {
+        size_t const literalsBytes = ZSTD_countSeqStoreLiteralsBytes(resultSeqStore);
+        resultSeqStore->lit = resultSeqStore->litStart + literalsBytes;
+    }
+    resultSeqStore->llCode += startIdx;
+    resultSeqStore->mlCode += startIdx;
+    resultSeqStore->ofCode += startIdx;
+}
+
+/**
+ * Returns the raw offset represented by the combination of offBase, ll0, and repcode history.
+ * offBase must represent a repcode in the numeric representation of ZSTD_storeSeq().
+ */
+static U32
+ZSTD_resolveRepcodeToRawOffset(const U32 rep[ZSTD_REP_NUM], const U32 offBase, const U32 ll0)
+{
+    U32 const adjustedRepCode = OFFBASE_TO_REPCODE(offBase) - 1 + ll0;  /* [ 0 - 3 ] */
+    assert(OFFBASE_IS_REPCODE(offBase));
+    if (adjustedRepCode == ZSTD_REP_NUM) {
+        assert(ll0);
+        /* litlength == 0 and offCode == 2 implies selection of first repcode - 1
+         * This is only valid if it results in a valid offset value, aka > 0.
+         * Note : it may happen that `rep[0]==1` in exceptional circumstances.
+         * In which case this function will return 0, which is an invalid offset.
+         * It's not an issue though, since this value will be
+         * compared and discarded within ZSTD_seqStore_resolveOffCodes().
+         */
+        return rep[0] - 1;
+    }
+    return rep[adjustedRepCode];
+}
+
+/**
+ * ZSTD_seqStore_resolveOffCodes() reconciles any possible divergences in offset history that may arise
+ * due to emission of RLE/raw blocks that disturb the offset history,
+ * and replaces any repcodes within the seqStore that may be invalid.
+ *
+ * dRepcodes are updated as would be on the decompression side.
+ * cRepcodes are updated exactly in accordance with the seqStore.
+ *
+ * Note : this function assumes seq->offBase respects the following numbering scheme :
+ *        0 : invalid
+ *        1-3 : repcode 1-3
+ *        4+ : real_offset+3
+ */
+static void
+ZSTD_seqStore_resolveOffCodes(repcodes_t* const dRepcodes, repcodes_t* const cRepcodes,
+                        const seqStore_t* const seqStore, U32 const nbSeq)
+{
+    U32 idx = 0;
+    for (; idx < nbSeq; ++idx) {
+        seqDef* const seq = seqStore->sequencesStart + idx;
+        U32 const ll0 = (seq->litLength == 0);
+        U32 const offBase = seq->offBase;
+        assert(offBase > 0);
+        if (OFFBASE_IS_REPCODE(offBase)) {
+            U32 const dRawOffset = ZSTD_resolveRepcodeToRawOffset(dRepcodes->rep, offBase, ll0);
+            U32 const cRawOffset = ZSTD_resolveRepcodeToRawOffset(cRepcodes->rep, offBase, ll0);
+            /* Adjust simulated decompression repcode history if we come across a mismatch. Replace
+             * the repcode with the offset it actually references, determined by the compression
+             * repcode history.
+             */
+            if (dRawOffset != cRawOffset) {
+                seq->offBase = OFFSET_TO_OFFBASE(cRawOffset);
+            }
+        }
+        /* Compression repcode history is always updated with values directly from the unmodified seqStore.
+         * Decompression repcode history may use modified seq->offset value taken from compression repcode history.
+         */
+        ZSTD_updateRep(dRepcodes->rep, seq->offBase, ll0);
+        ZSTD_updateRep(cRepcodes->rep, offBase, ll0);
+    }
+}
+
+/* ZSTD_compressSeqStore_singleBlock():
+ * Compresses a seqStore into a block with a block header, into the buffer dst.
+ *
+ * Returns the total size of that block (including header) or a ZSTD error code.
+ */
+static size_t
+ZSTD_compressSeqStore_singleBlock(ZSTD_CCtx* zc,
+                            const seqStore_t* const seqStore,
+                                  repcodes_t* const dRep, repcodes_t* const cRep,
+                                  void* dst, size_t dstCapacity,
+                            const void* src, size_t srcSize,
+                                  U32 lastBlock, U32 isPartition)
+{
+    const U32 rleMaxLength = 25;
+    BYTE* op = (BYTE*)dst;
+    const BYTE* ip = (const BYTE*)src;
+    size_t cSize;
+    size_t cSeqsSize;
+
+    /* In case of an RLE or raw block, the simulated decompression repcode history must be reset */
+    repcodes_t const dRepOriginal = *dRep;
+    DEBUGLOG(5, "ZSTD_compressSeqStore_singleBlock");
+    if (isPartition)
+        ZSTD_seqStore_resolveOffCodes(dRep, cRep, seqStore, (U32)(seqStore->sequences - seqStore->sequencesStart));
+
+    RETURN_ERROR_IF(dstCapacity < ZSTD_blockHeaderSize, dstSize_tooSmall, "Block header doesn't fit");
+    cSeqsSize = ZSTD_entropyCompressSeqStore(seqStore,
+                &zc->blockState.prevCBlock->entropy, &zc->blockState.nextCBlock->entropy,
+                &zc->appliedParams,
+                op + ZSTD_blockHeaderSize, dstCapacity - ZSTD_blockHeaderSize,
+                srcSize,
+                zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */,
+                zc->bmi2);
+    FORWARD_IF_ERROR(cSeqsSize, "ZSTD_entropyCompressSeqStore failed!");
+
+    if (!zc->isFirstBlock &&
+        cSeqsSize < rleMaxLength &&
+        ZSTD_isRLE((BYTE const*)src, srcSize)) {
+        /* We don't want to emit our first block as a RLE even if it qualifies because
+        * doing so will cause the decoder (cli only) to throw a "should consume all input error."
+        * This is only an issue for zstd <= v1.4.3
+        */
+        cSeqsSize = 1;
+    }
+
+    if (zc->seqCollector.collectSequences) {
+        ZSTD_copyBlockSequences(zc);
+        ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState);
+        return 0;
+    }
+
+    if (cSeqsSize == 0) {
+        cSize = ZSTD_noCompressBlock(op, dstCapacity, ip, srcSize, lastBlock);
+        FORWARD_IF_ERROR(cSize, "Nocompress block failed");
+        DEBUGLOG(4, "Writing out nocompress block, size: %zu", cSize);
+        *dRep = dRepOriginal; /* reset simulated decompression repcode history */
+    } else if (cSeqsSize == 1) {
+        cSize = ZSTD_rleCompressBlock(op, dstCapacity, *ip, srcSize, lastBlock);
+        FORWARD_IF_ERROR(cSize, "RLE compress block failed");
+        DEBUGLOG(4, "Writing out RLE block, size: %zu", cSize);
+        *dRep = dRepOriginal; /* reset simulated decompression repcode history */
+    } else {
+        ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState);
+        writeBlockHeader(op, cSeqsSize, srcSize, lastBlock);
+        cSize = ZSTD_blockHeaderSize + cSeqsSize;
+        DEBUGLOG(4, "Writing out compressed block, size: %zu", cSize);
+    }
+
+    if (zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode == FSE_repeat_valid)
+        zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode = FSE_repeat_check;
+
+    return cSize;
+}
+
+/* Struct to keep track of where we are in our recursive calls. */
+typedef struct {
+    U32* splitLocations;    /* Array of split indices */
+    size_t idx;             /* The current index within splitLocations being worked on */
+} seqStoreSplits;
+
+#define MIN_SEQUENCES_BLOCK_SPLITTING 300
+
+/* Helper function to perform the recursive search for block splits.
+ * Estimates the cost of seqStore prior to split, and estimates the cost of splitting the sequences in half.
+ * If advantageous to split, then we recurse down the two sub-blocks.
+ * If not, or if an error occurred in estimation, then we do not recurse.
+ *
+ * Note: The recursion depth is capped by a heuristic minimum number of sequences,
+ * defined by MIN_SEQUENCES_BLOCK_SPLITTING.
+ * In theory, this means the absolute largest recursion depth is 10 == log2(maxNbSeqInBlock/MIN_SEQUENCES_BLOCK_SPLITTING).
+ * In practice, recursion depth usually doesn't go beyond 4.
+ *
+ * Furthermore, the number of splits is capped by ZSTD_MAX_NB_BLOCK_SPLITS.
+ * At ZSTD_MAX_NB_BLOCK_SPLITS == 196 with the current existing blockSize
+ * maximum of 128 KB, this value is actually impossible to reach.
+ */
+static void
+ZSTD_deriveBlockSplitsHelper(seqStoreSplits* splits, size_t startIdx, size_t endIdx,
+                             ZSTD_CCtx* zc, const seqStore_t* origSeqStore)
+{
+    seqStore_t* const fullSeqStoreChunk = &zc->blockSplitCtx.fullSeqStoreChunk;
+    seqStore_t* const firstHalfSeqStore = &zc->blockSplitCtx.firstHalfSeqStore;
+    seqStore_t* const secondHalfSeqStore = &zc->blockSplitCtx.secondHalfSeqStore;
+    size_t estimatedOriginalSize;
+    size_t estimatedFirstHalfSize;
+    size_t estimatedSecondHalfSize;
+    size_t midIdx = (startIdx + endIdx)/2;
+
+    DEBUGLOG(5, "ZSTD_deriveBlockSplitsHelper: startIdx=%zu endIdx=%zu", startIdx, endIdx);
+    assert(endIdx >= startIdx);
+    if (endIdx - startIdx < MIN_SEQUENCES_BLOCK_SPLITTING || splits->idx >= ZSTD_MAX_NB_BLOCK_SPLITS) {
+        DEBUGLOG(6, "ZSTD_deriveBlockSplitsHelper: Too few sequences (%zu)", endIdx - startIdx);
+        return;
+    }
+    ZSTD_deriveSeqStoreChunk(fullSeqStoreChunk, origSeqStore, startIdx, endIdx);
+    ZSTD_deriveSeqStoreChunk(firstHalfSeqStore, origSeqStore, startIdx, midIdx);
+    ZSTD_deriveSeqStoreChunk(secondHalfSeqStore, origSeqStore, midIdx, endIdx);
+    estimatedOriginalSize = ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(fullSeqStoreChunk, zc);
+    estimatedFirstHalfSize = ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(firstHalfSeqStore, zc);
+    estimatedSecondHalfSize = ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(secondHalfSeqStore, zc);
+    DEBUGLOG(5, "Estimated original block size: %zu -- First half split: %zu -- Second half split: %zu",
+             estimatedOriginalSize, estimatedFirstHalfSize, estimatedSecondHalfSize);
+    if (ZSTD_isError(estimatedOriginalSize) || ZSTD_isError(estimatedFirstHalfSize) || ZSTD_isError(estimatedSecondHalfSize)) {
+        return;
+    }
+    if (estimatedFirstHalfSize + estimatedSecondHalfSize < estimatedOriginalSize) {
+        DEBUGLOG(5, "split decided at seqNb:%zu", midIdx);
+        ZSTD_deriveBlockSplitsHelper(splits, startIdx, midIdx, zc, origSeqStore);
+        splits->splitLocations[splits->idx] = (U32)midIdx;
+        splits->idx++;
+        ZSTD_deriveBlockSplitsHelper(splits, midIdx, endIdx, zc, origSeqStore);
+    }
+}
+
+/* Base recursive function.
+ * Populates a table with intra-block partition indices that can improve compression ratio.
+ *
+ * @return: number of splits made (which equals the size of the partition table - 1).
+ */
+static size_t ZSTD_deriveBlockSplits(ZSTD_CCtx* zc, U32 partitions[], U32 nbSeq)
+{
+    seqStoreSplits splits;
+    splits.splitLocations = partitions;
+    splits.idx = 0;
+    if (nbSeq <= 4) {
+        DEBUGLOG(5, "ZSTD_deriveBlockSplits: Too few sequences to split (%u <= 4)", nbSeq);
+        /* Refuse to try and split anything with less than 4 sequences */
+        return 0;
+    }
+    ZSTD_deriveBlockSplitsHelper(&splits, 0, nbSeq, zc, &zc->seqStore);
+    splits.splitLocations[splits.idx] = nbSeq;
+    DEBUGLOG(5, "ZSTD_deriveBlockSplits: final nb partitions: %zu", splits.idx+1);
+    return splits.idx;
+}
+
+/* ZSTD_compressBlock_splitBlock():
+ * Attempts to split a given block into multiple blocks to improve compression ratio.
+ *
+ * Returns combined size of all blocks (which includes headers), or a ZSTD error code.
+ */
+static size_t
+ZSTD_compressBlock_splitBlock_internal(ZSTD_CCtx* zc,
+                                    void* dst, size_t dstCapacity,
+                              const void* src, size_t blockSize,
+                                    U32 lastBlock, U32 nbSeq)
+{
+    size_t cSize = 0;
+    const BYTE* ip = (const BYTE*)src;
+    BYTE* op = (BYTE*)dst;
+    size_t i = 0;
+    size_t srcBytesTotal = 0;
+    U32* const partitions = zc->blockSplitCtx.partitions; /* size == ZSTD_MAX_NB_BLOCK_SPLITS */
+    seqStore_t* const nextSeqStore = &zc->blockSplitCtx.nextSeqStore;
+    seqStore_t* const currSeqStore = &zc->blockSplitCtx.currSeqStore;
+    size_t const numSplits = ZSTD_deriveBlockSplits(zc, partitions, nbSeq);
+
+    /* If a block is split and some partitions are emitted as RLE/uncompressed, then repcode history
+     * may become invalid. In order to reconcile potentially invalid repcodes, we keep track of two
+     * separate repcode histories that simulate repcode history on compression and decompression side,
+     * and use the histories to determine whether we must replace a particular repcode with its raw offset.
+     *
+     * 1) cRep gets updated for each partition, regardless of whether the block was emitted as uncompressed
+     *    or RLE. This allows us to retrieve the offset value that an invalid repcode references within
+     *    a nocompress/RLE block.
+     * 2) dRep gets updated only for compressed partitions, and when a repcode gets replaced, will use
+     *    the replacement offset value rather than the original repcode to update the repcode history.
+     *    dRep also will be the final repcode history sent to the next block.
+     *
+     * See ZSTD_seqStore_resolveOffCodes() for more details.
+     */
+    repcodes_t dRep;
+    repcodes_t cRep;
+    ZSTD_memcpy(dRep.rep, zc->blockState.prevCBlock->rep, sizeof(repcodes_t));
+    ZSTD_memcpy(cRep.rep, zc->blockState.prevCBlock->rep, sizeof(repcodes_t));
+    ZSTD_memset(nextSeqStore, 0, sizeof(seqStore_t));
+
+    DEBUGLOG(5, "ZSTD_compressBlock_splitBlock_internal (dstCapacity=%u, dictLimit=%u, nextToUpdate=%u)",
+                (unsigned)dstCapacity, (unsigned)zc->blockState.matchState.window.dictLimit,
+                (unsigned)zc->blockState.matchState.nextToUpdate);
+
+    if (numSplits == 0) {
+        size_t cSizeSingleBlock =
+            ZSTD_compressSeqStore_singleBlock(zc, &zc->seqStore,
+                                            &dRep, &cRep,
+                                            op, dstCapacity,
+                                            ip, blockSize,
+                                            lastBlock, 0 /* isPartition */);
+        FORWARD_IF_ERROR(cSizeSingleBlock, "Compressing single block from splitBlock_internal() failed!");
+        DEBUGLOG(5, "ZSTD_compressBlock_splitBlock_internal: No splits");
+        assert(zc->blockSize <= ZSTD_BLOCKSIZE_MAX);
+        assert(cSizeSingleBlock <= zc->blockSize + ZSTD_blockHeaderSize);
+        return cSizeSingleBlock;
+    }
+
+    ZSTD_deriveSeqStoreChunk(currSeqStore, &zc->seqStore, 0, partitions[0]);
+    for (i = 0; i <= numSplits; ++i) {
+        size_t cSizeChunk;
+        U32 const lastPartition = (i == numSplits);
+        U32 lastBlockEntireSrc = 0;
+
+        size_t srcBytes = ZSTD_countSeqStoreLiteralsBytes(currSeqStore) + ZSTD_countSeqStoreMatchBytes(currSeqStore);
+        srcBytesTotal += srcBytes;
+        if (lastPartition) {
+            /* This is the final partition, need to account for possible last literals */
+            srcBytes += blockSize - srcBytesTotal;
+            lastBlockEntireSrc = lastBlock;
+        } else {
+            ZSTD_deriveSeqStoreChunk(nextSeqStore, &zc->seqStore, partitions[i], partitions[i+1]);
+        }
+
+        cSizeChunk = ZSTD_compressSeqStore_singleBlock(zc, currSeqStore,
+                                                      &dRep, &cRep,
+                                                       op, dstCapacity,
+                                                       ip, srcBytes,
+                                                       lastBlockEntireSrc, 1 /* isPartition */);
+        DEBUGLOG(5, "Estimated size: %zu vs %zu : actual size",
+                    ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(currSeqStore, zc), cSizeChunk);
+        FORWARD_IF_ERROR(cSizeChunk, "Compressing chunk failed!");
+
+        ip += srcBytes;
+        op += cSizeChunk;
+        dstCapacity -= cSizeChunk;
+        cSize += cSizeChunk;
+        *currSeqStore = *nextSeqStore;
+        assert(cSizeChunk <= zc->blockSize + ZSTD_blockHeaderSize);
+    }
+    /* cRep and dRep may have diverged during the compression.
+     * If so, we use the dRep repcodes for the next block.
+     */
+    ZSTD_memcpy(zc->blockState.prevCBlock->rep, dRep.rep, sizeof(repcodes_t));
+    return cSize;
+}
+
+static size_t
+ZSTD_compressBlock_splitBlock(ZSTD_CCtx* zc,
+                              void* dst, size_t dstCapacity,
+                              const void* src, size_t srcSize, U32 lastBlock)
 {
-    ZSTD_compressedBlockState_t* const tmp = zc->blockState.prevCBlock;
-    zc->blockState.prevCBlock = zc->blockState.nextCBlock;
-    zc->blockState.nextCBlock = tmp;
+    U32 nbSeq;
+    size_t cSize;
+    DEBUGLOG(4, "ZSTD_compressBlock_splitBlock");
+    assert(zc->appliedParams.useBlockSplitter == ZSTD_ps_enable);
+
+    {   const size_t bss = ZSTD_buildSeqStore(zc, src, srcSize);
+        FORWARD_IF_ERROR(bss, "ZSTD_buildSeqStore failed");
+        if (bss == ZSTDbss_noCompress) {
+            if (zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode == FSE_repeat_valid)
+                zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode = FSE_repeat_check;
+            cSize = ZSTD_noCompressBlock(dst, dstCapacity, src, srcSize, lastBlock);
+            FORWARD_IF_ERROR(cSize, "ZSTD_noCompressBlock failed");
+            DEBUGLOG(4, "ZSTD_compressBlock_splitBlock: Nocompress block");
+            return cSize;
+        }
+        nbSeq = (U32)(zc->seqStore.sequences - zc->seqStore.sequencesStart);
+    }
+
+    cSize = ZSTD_compressBlock_splitBlock_internal(zc, dst, dstCapacity, src, srcSize, lastBlock, nbSeq);
+    FORWARD_IF_ERROR(cSize, "Splitting blocks failed!");
+    return cSize;
 }
 
-static size_t ZSTD_compressBlock_internal(ZSTD_CCtx* zc,
-                                        void* dst, size_t dstCapacity,
-                                        const void* src, size_t srcSize, U32 frame)
+static size_t
+ZSTD_compressBlock_internal(ZSTD_CCtx* zc,
+                            void* dst, size_t dstCapacity,
+                            const void* src, size_t srcSize, U32 frame)
 {
-    /* This the upper bound for the length of an rle block.
-     * This isn't the actual upper bound. Finding the real threshold
-     * needs further investigation.
+    /* This is an estimated upper bound for the length of an rle block.
+     * This isn't the actual upper bound.
+     * Finding the real threshold needs further investigation.
      */
     const U32 rleMaxLength = 25;
     size_t cSize;
@@ -2434,16 +4187,17 @@ static size_t ZSTD_compressBlock_internal(ZSTD_CCtx* zc,
 
     if (zc->seqCollector.collectSequences) {
         ZSTD_copyBlockSequences(zc);
+        ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState);
         return 0;
     }
 
     /* encode sequences and literals */
-    cSize = ZSTD_compressSequences(&zc->seqStore,
+    cSize = ZSTD_entropyCompressSeqStore(&zc->seqStore,
             &zc->blockState.prevCBlock->entropy, &zc->blockState.nextCBlock->entropy,
             &zc->appliedParams,
             dst, dstCapacity,
             srcSize,
-            zc->entropyWorkspace, HUF_WORKSPACE_SIZE /* statically allocated in resetCCtx */,
+            zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */,
             zc->bmi2);
 
     if (frame &&
@@ -2461,7 +4215,7 @@ static size_t ZSTD_compressBlock_internal(ZSTD_CCtx* zc,
 
 out:
     if (!ZSTD_isError(cSize) && cSize > 1) {
-        ZSTD_confirmRepcodesAndEntropyTables(zc);
+        ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState);
     }
     /* We check that dictionaries have offset codes available for the first
      * block. After the first block, the offcode table might not have large
@@ -2508,18 +4262,19 @@ static size_t ZSTD_compressBlock_targetCBlockSize_body(ZSTD_CCtx* zc,
          *   * cSize >= blockBound(srcSize): We have expanded the block too much so
          *     emit an uncompressed block.
          */
-        {
-            size_t const cSize = ZSTD_compressSuperBlock(zc, dst, dstCapacity, src, srcSize, lastBlock);
+        {   size_t const cSize =
+                ZSTD_compressSuperBlock(zc, dst, dstCapacity, src, srcSize, lastBlock);
             if (cSize != ERROR(dstSize_tooSmall)) {
-                size_t const maxCSize = srcSize - ZSTD_minGain(srcSize, zc->appliedParams.cParams.strategy);
+                size_t const maxCSize =
+                    srcSize - ZSTD_minGain(srcSize, zc->appliedParams.cParams.strategy);
                 FORWARD_IF_ERROR(cSize, "ZSTD_compressSuperBlock failed");
                 if (cSize != 0 && cSize < maxCSize + ZSTD_blockHeaderSize) {
-                    ZSTD_confirmRepcodesAndEntropyTables(zc);
+                    ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState);
                     return cSize;
                 }
             }
         }
-    }
+    } /* if (bss == ZSTDbss_compress)*/
 
     DEBUGLOG(6, "Resorting to ZSTD_noCompressBlock()");
     /* Superblock compression failed, attempt to emit a single no compress block.
@@ -2554,9 +4309,9 @@ static void ZSTD_overflowCorrectIfNeeded(ZSTD_matchState_t* ms,
                                          void const* ip,
                                          void const* iend)
 {
-    if (ZSTD_window_needOverflowCorrection(ms->window, iend)) {
-        U32 const maxDist = (U32)1 << params->cParams.windowLog;
-        U32 const cycleLog = ZSTD_cycleLog(params->cParams.chainLog, params->cParams.strategy);
+    U32 const cycleLog = ZSTD_cycleLog(params->cParams.chainLog, params->cParams.strategy);
+    U32 const maxDist = (U32)1 << params->cParams.windowLog;
+    if (ZSTD_window_needOverflowCorrection(ms->window, cycleLog, maxDist, ms->loadedDictEnd, ip, iend)) {
         U32 const correction = ZSTD_window_correctOverflow(&ms->window, cycleLog, maxDist, ip);
         ZSTD_STATIC_ASSERT(ZSTD_CHAINLOG_MAX <= 30);
         ZSTD_STATIC_ASSERT(ZSTD_WINDOWLOG_MAX_32 <= 30);
@@ -2577,9 +4332,9 @@ static void ZSTD_overflowCorrectIfNeeded(ZSTD_matchState_t* ms,
 *   All blocks will be terminated, all input will be consumed.
 *   Function will issue an error if there is not enough `dstCapacity` to hold the compressed content.
 *   Frame is supposed already started (header already produced)
-*   @return : compressed size, or an error code
+*  @return : compressed size, or an error code
 */
-static size_t ZSTD_compress_frameChunk (ZSTD_CCtx* cctx,
+static size_t ZSTD_compress_frameChunk(ZSTD_CCtx* cctx,
                                      void* dst, size_t dstCapacity,
                                const void* src, size_t srcSize,
                                      U32 lastFrameChunk)
@@ -2593,7 +4348,7 @@ static size_t ZSTD_compress_frameChunk (ZSTD_CCtx* cctx,
 
     assert(cctx->appliedParams.cParams.windowLog <= ZSTD_WINDOWLOG_MAX);
 
-    DEBUGLOG(5, "ZSTD_compress_frameChunk (blockSize=%u)", (unsigned)blockSize);
+    DEBUGLOG(4, "ZSTD_compress_frameChunk (blockSize=%u)", (unsigned)blockSize);
     if (cctx->appliedParams.fParams.checksumFlag && srcSize)
         XXH64_update(&cctx->xxhState, src, srcSize);
 
@@ -2601,7 +4356,9 @@ static size_t ZSTD_compress_frameChunk (ZSTD_CCtx* cctx,
         ZSTD_matchState_t* const ms = &cctx->blockState.matchState;
         U32 const lastBlock = lastFrameChunk & (blockSize >= remaining);
 
-        RETURN_ERROR_IF(dstCapacity < ZSTD_blockHeaderSize + MIN_CBLOCK_SIZE,
+        /* TODO: See 3090. We reduced MIN_CBLOCK_SIZE from 3 to 2 so to compensate we are adding
+         * additional 1. We need to revisit and change this logic to be more consistent */
+        RETURN_ERROR_IF(dstCapacity < ZSTD_blockHeaderSize + MIN_CBLOCK_SIZE + 1,
                         dstSize_tooSmall,
                         "not enough space to store compressed block");
         if (remaining < blockSize) blockSize = remaining;
@@ -2609,6 +4366,7 @@ static size_t ZSTD_compress_frameChunk (ZSTD_CCtx* cctx,
         ZSTD_overflowCorrectIfNeeded(
             ms, &cctx->workspace, &cctx->appliedParams, ip, ip + blockSize);
         ZSTD_checkDictValidity(&ms->window, ip + blockSize, maxDist, &ms->loadedDictEnd, &ms->dictMatchState);
+        ZSTD_window_enforceMaxDist(&ms->window, ip, maxDist, &ms->loadedDictEnd, &ms->dictMatchState);
 
         /* Ensure hash/chain table insertion resumes no sooner than lowlimit */
         if (ms->nextToUpdate < ms->window.lowLimit) ms->nextToUpdate = ms->window.lowLimit;
@@ -2619,6 +4377,10 @@ static size_t ZSTD_compress_frameChunk (ZSTD_CCtx* cctx,
                 FORWARD_IF_ERROR(cSize, "ZSTD_compressBlock_targetCBlockSize failed");
                 assert(cSize > 0);
                 assert(cSize <= blockSize + ZSTD_blockHeaderSize);
+            } else if (ZSTD_blockSplitterEnabled(&cctx->appliedParams)) {
+                cSize = ZSTD_compressBlock_splitBlock(cctx, op, dstCapacity, ip, blockSize, lastBlock);
+                FORWARD_IF_ERROR(cSize, "ZSTD_compressBlock_splitBlock failed");
+                assert(cSize > 0 || cctx->seqCollector.collectSequences == 1);
             } else {
                 cSize = ZSTD_compressBlock_internal(cctx,
                                         op+ZSTD_blockHeaderSize, dstCapacity-ZSTD_blockHeaderSize,
@@ -2635,7 +4397,7 @@ static size_t ZSTD_compress_frameChunk (ZSTD_CCtx* cctx,
                     MEM_writeLE24(op, cBlockHeader);
                     cSize += ZSTD_blockHeaderSize;
                 }
-            }
+            }  /* if (ZSTD_useTargetCBlockSize(&cctx->appliedParams))*/
 
 
             ip += blockSize;
@@ -2673,7 +4435,6 @@ static size_t ZSTD_writeFrameHeader(void* dst, size_t dstCapacity,
                     "dst buf is too small to fit worst-case frame header size.");
     DEBUGLOG(4, "ZSTD_writeFrameHeader : dictIDFlag : %u ; dictID : %u ; dictIDSizeCode : %u",
                 !params->fParams.noDictIDFlag, (unsigned)dictID, (unsigned)dictIDSizeCode);
-
     if (params->format == ZSTD_f_zstd1) {
         MEM_writeLE32(dst, ZSTD_MAGICNUMBER);
         pos = 4;
@@ -2682,7 +4443,9 @@ static size_t ZSTD_writeFrameHeader(void* dst, size_t dstCapacity,
     if (!singleSegment) op[pos++] = windowLogByte;
     switch(dictIDSizeCode)
     {
-        default:  assert(0); /* impossible */
+        default:
+            assert(0); /* impossible */
+            ZSTD_FALLTHROUGH;
         case 0 : break;
         case 1 : op[pos] = (BYTE)(dictID); pos++; break;
         case 2 : MEM_writeLE16(op+pos, (U16)dictID); pos+=2; break;
@@ -2690,7 +4453,9 @@ static size_t ZSTD_writeFrameHeader(void* dst, size_t dstCapacity,
     }
     switch(fcsCode)
     {
-        default:  assert(0); /* impossible */
+        default:
+            assert(0); /* impossible */
+            ZSTD_FALLTHROUGH;
         case 0 : if (singleSegment) op[pos++] = (BYTE)(pledgedSrcSize); break;
         case 1 : MEM_writeLE16(op+pos, (U16)(pledgedSrcSize-256)); pos+=2; break;
         case 2 : MEM_writeLE32(op+pos, (U32)(pledgedSrcSize)); pos+=4; break;
@@ -2699,6 +4464,26 @@ static size_t ZSTD_writeFrameHeader(void* dst, size_t dstCapacity,
     return pos;
 }
 
+/* ZSTD_writeSkippableFrame_advanced() :
+ * Writes out a skippable frame with the specified magic number variant (16 are supported),
+ * from ZSTD_MAGIC_SKIPPABLE_START to ZSTD_MAGIC_SKIPPABLE_START+15, and the desired source data.
+ *
+ * Returns the total number of bytes written, or a ZSTD error code.
+ */
+size_t ZSTD_writeSkippableFrame(void* dst, size_t dstCapacity,
+                                const void* src, size_t srcSize, unsigned magicVariant) {
+    BYTE* op = (BYTE*)dst;
+    RETURN_ERROR_IF(dstCapacity < srcSize + ZSTD_SKIPPABLEHEADERSIZE /* Skippable frame overhead */,
+                    dstSize_tooSmall, "Not enough room for skippable frame");
+    RETURN_ERROR_IF(srcSize > (unsigned)0xFFFFFFFF, srcSize_wrong, "Src size too large for skippable frame");
+    RETURN_ERROR_IF(magicVariant > 15, parameter_outOfBound, "Skippable frame magic number variant not supported");
+
+    MEM_writeLE32(op, (U32)(ZSTD_MAGIC_SKIPPABLE_START + magicVariant));
+    MEM_writeLE32(op+4, (U32)srcSize);
+    ZSTD_memcpy(op+8, src, srcSize);
+    return srcSize + ZSTD_SKIPPABLEHEADERSIZE;
+}
+
 /* ZSTD_writeLastEmptyBlock() :
  * output an empty Block with end-of-frame mark to complete a frame
  * @return : size of data written into `dst` (== ZSTD_blockHeaderSize (defined in zstd_internal.h))
@@ -2718,13 +4503,14 @@ size_t ZSTD_referenceExternalSequences(ZSTD_CCtx* cctx, rawSeq* seq, size_t nbSe
 {
     RETURN_ERROR_IF(cctx->stage != ZSTDcs_init, stage_wrong,
                     "wrong cctx stage");
-    RETURN_ERROR_IF(cctx->appliedParams.ldmParams.enableLdm,
+    RETURN_ERROR_IF(cctx->appliedParams.ldmParams.enableLdm == ZSTD_ps_enable,
                     parameter_unsupported,
                     "incompatible with ldm");
     cctx->externSeqStore.seq = seq;
     cctx->externSeqStore.size = nbSeq;
     cctx->externSeqStore.capacity = nbSeq;
     cctx->externSeqStore.pos = 0;
+    cctx->externSeqStore.posInSequence = 0;
     return 0;
 }
 
@@ -2754,11 +4540,12 @@ static size_t ZSTD_compressContinue_internal (ZSTD_CCtx* cctx,
 
     if (!srcSize) return fhSize;  /* do not generate an empty block if no input */
 
-    if (!ZSTD_window_update(&ms->window, src, srcSize)) {
+    if (!ZSTD_window_update(&ms->window, src, srcSize, ms->forceNonContiguous)) {
+        ms->forceNonContiguous = 0;
         ms->nextToUpdate = ms->window.dictLimit;
     }
-    if (cctx->appliedParams.ldmParams.enableLdm) {
-        ZSTD_window_update(&cctx->ldmState.window, src, srcSize);
+    if (cctx->appliedParams.ldmParams.enableLdm == ZSTD_ps_enable) {
+        ZSTD_window_update(&cctx->ldmState.window, src, srcSize, /* forceNonContiguous */ 0);
     }
 
     if (!frame) {
@@ -2802,7 +4589,7 @@ size_t ZSTD_getBlockSize(const ZSTD_CCtx* cctx)
 {
     ZSTD_compressionParameters const cParams = cctx->appliedParams.cParams;
     assert(!ZSTD_checkCParams(cParams));
-    return MIN (ZSTD_BLOCKSIZE_MAX, (U32)1 << cParams.windowLog);
+    return MIN(cctx->appliedParams.maxBlockSize, (size_t)1 << cParams.windowLog);
 }
 
 size_t ZSTD_compressBlock(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize)
@@ -2822,63 +4609,106 @@ static size_t ZSTD_loadDictionaryContent(ZSTD_matchState_t* ms,
                                          ZSTD_cwksp* ws,
                                          ZSTD_CCtx_params const* params,
                                          const void* src, size_t srcSize,
-                                         ZSTD_dictTableLoadMethod_e dtlm)
+                                         ZSTD_dictTableLoadMethod_e dtlm,
+                                         ZSTD_tableFillPurpose_e tfp)
 {
     const BYTE* ip = (const BYTE*) src;
     const BYTE* const iend = ip + srcSize;
+    int const loadLdmDict = params->ldmParams.enableLdm == ZSTD_ps_enable && ls != NULL;
 
-    ZSTD_window_update(&ms->window, src, srcSize);
-    ms->loadedDictEnd = params->forceWindow ? 0 : (U32)(iend - ms->window.base);
+    /* Assert that the ms params match the params we're being given */
+    ZSTD_assertEqualCParams(params->cParams, ms->cParams);
 
-    if (params->ldmParams.enableLdm && ls != NULL) {
-        ZSTD_window_update(&ls->window, src, srcSize);
-        ls->loadedDictEnd = params->forceWindow ? 0 : (U32)(iend - ls->window.base);
-    }
+    {   /* Ensure large dictionaries can't cause index overflow */
 
-    /* Assert that we the ms params match the params we're being given */
-    ZSTD_assertEqualCParams(params->cParams, ms->cParams);
+        /* Allow the dictionary to set indices up to exactly ZSTD_CURRENT_MAX.
+         * Dictionaries right at the edge will immediately trigger overflow
+         * correction, but I don't want to insert extra constraints here.
+         */
+        U32 maxDictSize = ZSTD_CURRENT_MAX - ZSTD_WINDOW_START_INDEX;
+
+        int const CDictTaggedIndices = ZSTD_CDictIndicesAreTagged(&params->cParams);
+        if (CDictTaggedIndices && tfp == ZSTD_tfp_forCDict) {
+            /* Some dictionary matchfinders in zstd use "short cache",
+             * which treats the lower ZSTD_SHORT_CACHE_TAG_BITS of each
+             * CDict hashtable entry as a tag rather than as part of an index.
+             * When short cache is used, we need to truncate the dictionary
+             * so that its indices don't overlap with the tag. */
+            U32 const shortCacheMaxDictSize = (1u << (32 - ZSTD_SHORT_CACHE_TAG_BITS)) - ZSTD_WINDOW_START_INDEX;
+            maxDictSize = MIN(maxDictSize, shortCacheMaxDictSize);
+            assert(!loadLdmDict);
+        }
 
-    if (srcSize <= HASH_READ_SIZE) return 0;
+        /* If the dictionary is too large, only load the suffix of the dictionary. */
+        if (srcSize > maxDictSize) {
+            ip = iend - maxDictSize;
+            src = ip;
+            srcSize = maxDictSize;
+    }   }
 
-    while (iend - ip > HASH_READ_SIZE) {
-        size_t const remaining = (size_t)(iend - ip);
-        size_t const chunk = MIN(remaining, ZSTD_CHUNKSIZE_MAX);
-        const BYTE* const ichunk = ip + chunk;
+    if (srcSize > ZSTD_CHUNKSIZE_MAX) {
+        /* We must have cleared our windows when our source is this large. */
+        assert(ZSTD_window_isEmpty(ms->window));
+        if (loadLdmDict) assert(ZSTD_window_isEmpty(ls->window));
+    }
 
-        ZSTD_overflowCorrectIfNeeded(ms, ws, params, ip, ichunk);
+    DEBUGLOG(4, "ZSTD_loadDictionaryContent(): useRowMatchFinder=%d", (int)params->useRowMatchFinder);
+    ZSTD_window_update(&ms->window, src, srcSize, /* forceNonContiguous */ 0);
+    ms->loadedDictEnd = params->forceWindow ? 0 : (U32)(iend - ms->window.base);
+    ms->forceNonContiguous = params->deterministicRefPrefix;
 
-        if (params->ldmParams.enableLdm && ls != NULL)
-            ZSTD_ldm_fillHashTable(ls, (const BYTE*)src, (const BYTE*)src + srcSize, &params->ldmParams);
+    if (loadLdmDict) {
+        ZSTD_window_update(&ls->window, src, srcSize, /* forceNonContiguous */ 0);
+        ls->loadedDictEnd = params->forceWindow ? 0 : (U32)(iend - ls->window.base);
+    }
 
-        switch(params->cParams.strategy)
-        {
-        case ZSTD_fast:
-            ZSTD_fillHashTable(ms, ichunk, dtlm);
-            break;
-        case ZSTD_dfast:
-            ZSTD_fillDoubleHashTable(ms, ichunk, dtlm);
-            break;
+    if (srcSize <= HASH_READ_SIZE) return 0;
 
-        case ZSTD_greedy:
-        case ZSTD_lazy:
-        case ZSTD_lazy2:
-            if (chunk >= HASH_READ_SIZE)
-                ZSTD_insertAndFindFirstIndex(ms, ichunk-HASH_READ_SIZE);
-            break;
+    ZSTD_overflowCorrectIfNeeded(ms, ws, params, ip, iend);
 
-        case ZSTD_btlazy2:   /* we want the dictionary table fully sorted */
-        case ZSTD_btopt:
-        case ZSTD_btultra:
-        case ZSTD_btultra2:
-            if (chunk >= HASH_READ_SIZE)
-                ZSTD_updateTree(ms, ichunk-HASH_READ_SIZE, ichunk);
-            break;
+    if (loadLdmDict)
+        ZSTD_ldm_fillHashTable(ls, ip, iend, &params->ldmParams);
 
-        default:
-            assert(0);  /* not possible : not a valid strategy id */
+    switch(params->cParams.strategy)
+    {
+    case ZSTD_fast:
+        ZSTD_fillHashTable(ms, iend, dtlm, tfp);
+        break;
+    case ZSTD_dfast:
+        ZSTD_fillDoubleHashTable(ms, iend, dtlm, tfp);
+        break;
+
+    case ZSTD_greedy:
+    case ZSTD_lazy:
+    case ZSTD_lazy2:
+        assert(srcSize >= HASH_READ_SIZE);
+        if (ms->dedicatedDictSearch) {
+            assert(ms->chainTable != NULL);
+            ZSTD_dedicatedDictSearch_lazy_loadDictionary(ms, iend-HASH_READ_SIZE);
+        } else {
+            assert(params->useRowMatchFinder != ZSTD_ps_auto);
+            if (params->useRowMatchFinder == ZSTD_ps_enable) {
+                size_t const tagTableSize = ((size_t)1 << params->cParams.hashLog) * sizeof(U16);
+                ZSTD_memset(ms->tagTable, 0, tagTableSize);
+                ZSTD_row_update(ms, iend-HASH_READ_SIZE);
+                DEBUGLOG(4, "Using row-based hash table for lazy dict");
+            } else {
+                ZSTD_insertAndFindFirstIndex(ms, iend-HASH_READ_SIZE);
+                DEBUGLOG(4, "Using chain-based hash table for lazy dict");
+            }
         }
+        break;
 
-        ip = ichunk;
+    case ZSTD_btlazy2:   /* we want the dictionary table fully sorted */
+    case ZSTD_btopt:
+    case ZSTD_btultra:
+    case ZSTD_btultra2:
+        assert(srcSize >= HASH_READ_SIZE);
+        ZSTD_updateTree(ms, iend-HASH_READ_SIZE, iend);
+        break;
+
+    default:
+        assert(0);  /* not possible : not a valid strategy id */
     }
 
     ms->nextToUpdate = (U32)(iend - ms->window.base);
@@ -2887,22 +4717,28 @@ static size_t ZSTD_loadDictionaryContent(ZSTD_matchState_t* ms,
 
 
 /* Dictionaries that assign zero probability to symbols that show up causes problems
-   when FSE encoding.  Refuse dictionaries that assign zero probability to symbols
-   that we may encounter during compression.
-   NOTE: This behavior is not standard and could be improved in the future. */
-static size_t ZSTD_checkDictNCount(short* normalizedCounter, unsigned dictMaxSymbolValue, unsigned maxSymbolValue) {
+ * when FSE encoding. Mark dictionaries with zero probability symbols as FSE_repeat_check
+ * and only dictionaries with 100% valid symbols can be assumed valid.
+ */
+static FSE_repeat ZSTD_dictNCountRepeat(short* normalizedCounter, unsigned dictMaxSymbolValue, unsigned maxSymbolValue)
+{
     U32 s;
-    RETURN_ERROR_IF(dictMaxSymbolValue < maxSymbolValue, dictionary_corrupted, "dict fse tables don't have all symbols");
+    if (dictMaxSymbolValue < maxSymbolValue) {
+        return FSE_repeat_check;
+    }
     for (s = 0; s <= maxSymbolValue; ++s) {
-        RETURN_ERROR_IF(normalizedCounter[s] == 0, dictionary_corrupted, "dict fse tables don't have all symbols");
+        if (normalizedCounter[s] == 0) {
+            return FSE_repeat_check;
+        }
     }
-    return 0;
+    return FSE_repeat_valid;
 }
 
 size_t ZSTD_loadCEntropy(ZSTD_compressedBlockState_t* bs, void* workspace,
-                         short* offcodeNCount, unsigned* offcodeMaxValue,
                          const void* const dict, size_t dictSize)
 {
+    short offcodeNCount[MaxOff+1];
+    unsigned offcodeMaxValue = MaxOff;
     const BYTE* dictPtr = (const BYTE*)dict;    /* skip magic num and dict ID */
     const BYTE* const dictEnd = dictPtr + dictSize;
     dictPtr += 8;
@@ -2924,16 +4760,16 @@ size_t ZSTD_loadCEntropy(ZSTD_compressedBlockState_t* bs, void* workspace,
     }
 
     {   unsigned offcodeLog;
-        size_t const offcodeHeaderSize = FSE_readNCount(offcodeNCount, offcodeMaxValue, &offcodeLog, dictPtr, dictEnd-dictPtr);
+        size_t const offcodeHeaderSize = FSE_readNCount(offcodeNCount, &offcodeMaxValue, &offcodeLog, dictPtr, dictEnd-dictPtr);
         RETURN_ERROR_IF(FSE_isError(offcodeHeaderSize), dictionary_corrupted, "");
         RETURN_ERROR_IF(offcodeLog > OffFSELog, dictionary_corrupted, "");
-        /* Defer checking offcodeMaxValue because we need to know the size of the dictionary content */
         /* fill all offset symbols to avoid garbage at end of table */
         RETURN_ERROR_IF(FSE_isError(FSE_buildCTable_wksp(
                 bs->entropy.fse.offcodeCTable,
                 offcodeNCount, MaxOff, offcodeLog,
                 workspace, HUF_WORKSPACE_SIZE)),
             dictionary_corrupted, "");
+        /* Defer checking offcodeMaxValue because we need to know the size of the dictionary content */
         dictPtr += offcodeHeaderSize;
     }
 
@@ -2942,13 +4778,12 @@ size_t ZSTD_loadCEntropy(ZSTD_compressedBlockState_t* bs, void* workspace,
         size_t const matchlengthHeaderSize = FSE_readNCount(matchlengthNCount, &matchlengthMaxValue, &matchlengthLog, dictPtr, dictEnd-dictPtr);
         RETURN_ERROR_IF(FSE_isError(matchlengthHeaderSize), dictionary_corrupted, "");
         RETURN_ERROR_IF(matchlengthLog > MLFSELog, dictionary_corrupted, "");
-        /* Every match length code must have non-zero probability */
-        FORWARD_IF_ERROR( ZSTD_checkDictNCount(matchlengthNCount, matchlengthMaxValue, MaxML), "");
         RETURN_ERROR_IF(FSE_isError(FSE_buildCTable_wksp(
                 bs->entropy.fse.matchlengthCTable,
                 matchlengthNCount, matchlengthMaxValue, matchlengthLog,
                 workspace, HUF_WORKSPACE_SIZE)),
             dictionary_corrupted, "");
+        bs->entropy.fse.matchlength_repeatMode = ZSTD_dictNCountRepeat(matchlengthNCount, matchlengthMaxValue, MaxML);
         dictPtr += matchlengthHeaderSize;
     }
 
@@ -2957,13 +4792,12 @@ size_t ZSTD_loadCEntropy(ZSTD_compressedBlockState_t* bs, void* workspace,
         size_t const litlengthHeaderSize = FSE_readNCount(litlengthNCount, &litlengthMaxValue, &litlengthLog, dictPtr, dictEnd-dictPtr);
         RETURN_ERROR_IF(FSE_isError(litlengthHeaderSize), dictionary_corrupted, "");
         RETURN_ERROR_IF(litlengthLog > LLFSELog, dictionary_corrupted, "");
-        /* Every literal length code must have non-zero probability */
-        FORWARD_IF_ERROR( ZSTD_checkDictNCount(litlengthNCount, litlengthMaxValue, MaxLL), "");
         RETURN_ERROR_IF(FSE_isError(FSE_buildCTable_wksp(
                 bs->entropy.fse.litlengthCTable,
                 litlengthNCount, litlengthMaxValue, litlengthLog,
                 workspace, HUF_WORKSPACE_SIZE)),
             dictionary_corrupted, "");
+        bs->entropy.fse.litlength_repeatMode = ZSTD_dictNCountRepeat(litlengthNCount, litlengthMaxValue, MaxLL);
         dictPtr += litlengthHeaderSize;
     }
 
@@ -2973,12 +4807,28 @@ size_t ZSTD_loadCEntropy(ZSTD_compressedBlockState_t* bs, void* workspace,
     bs->rep[2] = MEM_readLE32(dictPtr+8);
     dictPtr += 12;
 
+    {   size_t const dictContentSize = (size_t)(dictEnd - dictPtr);
+        U32 offcodeMax = MaxOff;
+        if (dictContentSize <= ((U32)-1) - 128 KB) {
+            U32 const maxOffset = (U32)dictContentSize + 128 KB; /* The maximum offset that must be supported */
+            offcodeMax = ZSTD_highbit32(maxOffset); /* Calculate minimum offset code required to represent maxOffset */
+        }
+        /* All offset values <= dictContentSize + 128 KB must be representable for a valid table */
+        bs->entropy.fse.offcode_repeatMode = ZSTD_dictNCountRepeat(offcodeNCount, offcodeMaxValue, MIN(offcodeMax, MaxOff));
+
+        /* All repCodes must be <= dictContentSize and != 0 */
+        {   U32 u;
+            for (u=0; u<3; u++) {
+                RETURN_ERROR_IF(bs->rep[u] == 0, dictionary_corrupted, "");
+                RETURN_ERROR_IF(bs->rep[u] > dictContentSize, dictionary_corrupted, "");
+    }   }   }
+
     return dictPtr - (const BYTE*)dict;
 }
 
 /* Dictionary format :
  * See :
- * https://github.com/facebook/zstd/blob/master/doc/zstd_compression_format.md#dictionary-format
+ * https://github.com/facebook/zstd/blob/release/doc/zstd_compression_format.md#dictionary-format
  */
 /*! ZSTD_loadZstdDictionary() :
  * @return : dictID, or an error code
@@ -2991,46 +4841,28 @@ static size_t ZSTD_loadZstdDictionary(ZSTD_compressedBlockState_t* bs,
                                       ZSTD_CCtx_params const* params,
                                       const void* dict, size_t dictSize,
                                       ZSTD_dictTableLoadMethod_e dtlm,
+                                      ZSTD_tableFillPurpose_e tfp,
                                       void* workspace)
 {
     const BYTE* dictPtr = (const BYTE*)dict;
     const BYTE* const dictEnd = dictPtr + dictSize;
-    short offcodeNCount[MaxOff+1];
-    unsigned offcodeMaxValue = MaxOff;
     size_t dictID;
     size_t eSize;
-
     ZSTD_STATIC_ASSERT(HUF_WORKSPACE_SIZE >= (1<<MAX(MLFSELog,LLFSELog)));
     assert(dictSize >= 8);
     assert(MEM_readLE32(dictPtr) == ZSTD_MAGIC_DICTIONARY);
 
     dictID = params->fParams.noDictIDFlag ? 0 :  MEM_readLE32(dictPtr + 4 /* skip magic number */ );
-    eSize = ZSTD_loadCEntropy(bs, workspace, offcodeNCount, &offcodeMaxValue, dict, dictSize);
+    eSize = ZSTD_loadCEntropy(bs, workspace, dict, dictSize);
     FORWARD_IF_ERROR(eSize, "ZSTD_loadCEntropy failed");
     dictPtr += eSize;
 
-    {   size_t const dictContentSize = (size_t)(dictEnd - dictPtr);
-        U32 offcodeMax = MaxOff;
-        if (dictContentSize <= ((U32)-1) - 128 KB) {
-            U32 const maxOffset = (U32)dictContentSize + 128 KB; /* The maximum offset that must be supported */
-            offcodeMax = ZSTD_highbit32(maxOffset); /* Calculate minimum offset code required to represent maxOffset */
-        }
-        /* All offset values <= dictContentSize + 128 KB must be representable */
-        FORWARD_IF_ERROR(ZSTD_checkDictNCount(offcodeNCount, offcodeMaxValue, MIN(offcodeMax, MaxOff)), "");
-        /* All repCodes must be <= dictContentSize and != 0*/
-        {   U32 u;
-            for (u=0; u<3; u++) {
-                RETURN_ERROR_IF(bs->rep[u] == 0, dictionary_corrupted, "");
-                RETURN_ERROR_IF(bs->rep[u] > dictContentSize, dictionary_corrupted, "");
-        }   }
-
-        bs->entropy.fse.offcode_repeatMode = FSE_repeat_valid;
-        bs->entropy.fse.matchlength_repeatMode = FSE_repeat_valid;
-        bs->entropy.fse.litlength_repeatMode = FSE_repeat_valid;
+    {
+        size_t const dictContentSize = (size_t)(dictEnd - dictPtr);
         FORWARD_IF_ERROR(ZSTD_loadDictionaryContent(
-            ms, NULL, ws, params, dictPtr, dictContentSize, dtlm), "");
-        return dictID;
+            ms, NULL, ws, params, dictPtr, dictContentSize, dtlm, tfp), "");
     }
+    return dictID;
 }
 
 /** ZSTD_compress_insertDictionary() :
@@ -3044,6 +4876,7 @@ ZSTD_compress_insertDictionary(ZSTD_compressedBlockState_t* bs,
                          const void* dict, size_t dictSize,
                                ZSTD_dictContentType_e dictContentType,
                                ZSTD_dictTableLoadMethod_e dtlm,
+                               ZSTD_tableFillPurpose_e tfp,
                                void* workspace)
 {
     DEBUGLOG(4, "ZSTD_compress_insertDictionary (dictSize=%u)", (U32)dictSize);
@@ -3056,13 +4889,13 @@ ZSTD_compress_insertDictionary(ZSTD_compressedBlockState_t* bs,
 
     /* dict restricted modes */
     if (dictContentType == ZSTD_dct_rawContent)
-        return ZSTD_loadDictionaryContent(ms, ls, ws, params, dict, dictSize, dtlm);
+        return ZSTD_loadDictionaryContent(ms, ls, ws, params, dict, dictSize, dtlm, tfp);
 
     if (MEM_readLE32(dict) != ZSTD_MAGIC_DICTIONARY) {
         if (dictContentType == ZSTD_dct_auto) {
             DEBUGLOG(4, "raw content dictionary detected");
             return ZSTD_loadDictionaryContent(
-                ms, ls, ws, params, dict, dictSize, dtlm);
+                ms, ls, ws, params, dict, dictSize, dtlm, tfp);
         }
         RETURN_ERROR_IF(dictContentType == ZSTD_dct_fullDict, dictionary_wrong, "");
         assert(0);   /* impossible */
@@ -3070,13 +4903,14 @@ ZSTD_compress_insertDictionary(ZSTD_compressedBlockState_t* bs,
 
     /* dict as full zstd dictionary */
     return ZSTD_loadZstdDictionary(
-        bs, ms, ws, params, dict, dictSize, dtlm, workspace);
+        bs, ms, ws, params, dict, dictSize, dtlm, tfp, workspace);
 }
 
 #define ZSTD_USE_CDICT_PARAMS_SRCSIZE_CUTOFF (128 KB)
-#define ZSTD_USE_CDICT_PARAMS_DICTSIZE_MULTIPLIER (6)
+#define ZSTD_USE_CDICT_PARAMS_DICTSIZE_MULTIPLIER (6ULL)
 
 /*! ZSTD_compressBegin_internal() :
+ * Assumption : either @dict OR @cdict (or none) is non-NULL, never both
  * @return : 0, or an error code */
 static size_t ZSTD_compressBegin_internal(ZSTD_CCtx* cctx,
                                     const void* dict, size_t dictSize,
@@ -3086,6 +4920,10 @@ static size_t ZSTD_compressBegin_internal(ZSTD_CCtx* cctx,
                                     const ZSTD_CCtx_params* params, U64 pledgedSrcSize,
                                     ZSTD_buffered_policy_e zbuff)
 {
+    size_t const dictContentSize = cdict ? cdict->dictContentSize : dictSize;
+#if ZSTD_TRACE
+    cctx->traceCtx = (ZSTD_trace_compress_begin != NULL) ? ZSTD_trace_compress_begin(cctx) : 0;
+#endif
     DEBUGLOG(4, "ZSTD_compressBegin_internal: wlog=%u", params->cParams.windowLog);
     /* params are supposed to be fully validated at this point */
     assert(!ZSTD_isError(ZSTD_checkCParams(params->cParams)));
@@ -3100,21 +4938,23 @@ static size_t ZSTD_compressBegin_internal(ZSTD_CCtx* cctx,
         return ZSTD_resetCCtx_usingCDict(cctx, cdict, params, pledgedSrcSize, zbuff);
     }
 
-    FORWARD_IF_ERROR( ZSTD_resetCCtx_internal(cctx, *params, pledgedSrcSize,
+    FORWARD_IF_ERROR( ZSTD_resetCCtx_internal(cctx, params, pledgedSrcSize,
+                                     dictContentSize,
                                      ZSTDcrp_makeClean, zbuff) , "");
     {   size_t const dictID = cdict ?
                 ZSTD_compress_insertDictionary(
                         cctx->blockState.prevCBlock, &cctx->blockState.matchState,
                         &cctx->ldmState, &cctx->workspace, &cctx->appliedParams, cdict->dictContent,
-                        cdict->dictContentSize, dictContentType, dtlm,
-                        cctx->entropyWorkspace)
+                        cdict->dictContentSize, cdict->dictContentType, dtlm,
+                        ZSTD_tfp_forCCtx, cctx->entropyWorkspace)
               : ZSTD_compress_insertDictionary(
                         cctx->blockState.prevCBlock, &cctx->blockState.matchState,
                         &cctx->ldmState, &cctx->workspace, &cctx->appliedParams, dict, dictSize,
-                        dictContentType, dtlm, cctx->entropyWorkspace);
+                        dictContentType, dtlm, ZSTD_tfp_forCCtx, cctx->entropyWorkspace);
         FORWARD_IF_ERROR(dictID, "ZSTD_compress_insertDictionary failed");
         assert(dictID <= UINT_MAX);
         cctx->dictID = (U32)dictID;
+        cctx->dictContentSize = dictContentSize;
     }
     return 0;
 }
@@ -3143,19 +4983,21 @@ size_t ZSTD_compressBegin_advanced(ZSTD_CCtx* cctx,
                              const void* dict, size_t dictSize,
                                    ZSTD_parameters params, unsigned long long pledgedSrcSize)
 {
-    ZSTD_CCtx_params const cctxParams =
-            ZSTD_assignParamsToCCtxParams(&cctx->requestedParams, &params);
+    ZSTD_CCtx_params cctxParams;
+    ZSTD_CCtxParams_init_internal(&cctxParams, &params, ZSTD_NO_CLEVEL);
     return ZSTD_compressBegin_advanced_internal(cctx,
                                             dict, dictSize, ZSTD_dct_auto, ZSTD_dtlm_fast,
                                             NULL /*cdict*/,
                                             &cctxParams, pledgedSrcSize);
 }
 
-size_t ZSTD_compressBegin_usingDict(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, int compressionLevel)
+size_t
+ZSTD_compressBegin_usingDict(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, int compressionLevel)
 {
-    ZSTD_parameters const params = ZSTD_getParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize);
-    ZSTD_CCtx_params const cctxParams =
-            ZSTD_assignParamsToCCtxParams(&cctx->requestedParams, &params);
+    ZSTD_CCtx_params cctxParams;
+    {   ZSTD_parameters const params = ZSTD_getParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_noAttachDict);
+        ZSTD_CCtxParams_init_internal(&cctxParams, &params, (compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT : compressionLevel);
+    }
     DEBUGLOG(4, "ZSTD_compressBegin_usingDict (dictSize=%u)", (unsigned)dictSize);
     return ZSTD_compressBegin_internal(cctx, dict, dictSize, ZSTD_dct_auto, ZSTD_dtlm_fast, NULL,
                                        &cctxParams, ZSTD_CONTENTSIZE_UNKNOWN, ZSTDb_not_buffered);
@@ -3209,6 +5051,30 @@ static size_t ZSTD_writeEpilogue(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity)
     return op-ostart;
 }
 
+void ZSTD_CCtx_trace(ZSTD_CCtx* cctx, size_t extraCSize)
+{
+#if ZSTD_TRACE
+    if (cctx->traceCtx && ZSTD_trace_compress_end != NULL) {
+        int const streaming = cctx->inBuffSize > 0 || cctx->outBuffSize > 0 || cctx->appliedParams.nbWorkers > 0;
+        ZSTD_Trace trace;
+        ZSTD_memset(&trace, 0, sizeof(trace));
+        trace.version = ZSTD_VERSION_NUMBER;
+        trace.streaming = streaming;
+        trace.dictionaryID = cctx->dictID;
+        trace.dictionarySize = cctx->dictContentSize;
+        trace.uncompressedSize = cctx->consumedSrcSize;
+        trace.compressedSize = cctx->producedCSize + extraCSize;
+        trace.params = &cctx->appliedParams;
+        trace.cctx = cctx;
+        ZSTD_trace_compress_end(cctx->traceCtx, &trace);
+    }
+    cctx->traceCtx = 0;
+#else
+    (void)cctx;
+    (void)extraCSize;
+#endif
+}
+
 size_t ZSTD_compressEnd (ZSTD_CCtx* cctx,
                          void* dst, size_t dstCapacity,
                    const void* src, size_t srcSize)
@@ -3231,26 +5097,10 @@ size_t ZSTD_compressEnd (ZSTD_CCtx* cctx,
             (unsigned)cctx->pledgedSrcSizePlusOne-1,
             (unsigned)cctx->consumedSrcSize);
     }
+    ZSTD_CCtx_trace(cctx, endResult);
     return cSize + endResult;
 }
 
-
-static size_t ZSTD_compress_internal (ZSTD_CCtx* cctx,
-                                      void* dst, size_t dstCapacity,
-                                const void* src, size_t srcSize,
-                                const void* dict,size_t dictSize,
-                                const ZSTD_parameters* params)
-{
-    ZSTD_CCtx_params const cctxParams =
-            ZSTD_assignParamsToCCtxParams(&cctx->requestedParams, params);
-    DEBUGLOG(4, "ZSTD_compress_internal");
-    return ZSTD_compress_advanced_internal(cctx,
-                                           dst, dstCapacity,
-                                           src, srcSize,
-                                           dict, dictSize,
-                                           &cctxParams);
-}
-
 size_t ZSTD_compress_advanced (ZSTD_CCtx* cctx,
                                void* dst, size_t dstCapacity,
                          const void* src, size_t srcSize,
@@ -3259,11 +5109,12 @@ size_t ZSTD_compress_advanced (ZSTD_CCtx* cctx,
 {
     DEBUGLOG(4, "ZSTD_compress_advanced");
     FORWARD_IF_ERROR(ZSTD_checkCParams(params.cParams), "");
-    return ZSTD_compress_internal(cctx,
-                                  dst, dstCapacity,
-                                  src, srcSize,
-                                  dict, dictSize,
-                                  &params);
+    ZSTD_CCtxParams_init_internal(&cctx->simpleApiParams, &params, ZSTD_NO_CLEVEL);
+    return ZSTD_compress_advanced_internal(cctx,
+                                           dst, dstCapacity,
+                                           src, srcSize,
+                                           dict, dictSize,
+                                           &cctx->simpleApiParams);
 }
 
 /* Internal */
@@ -3287,11 +5138,13 @@ size_t ZSTD_compress_usingDict(ZSTD_CCtx* cctx,
                          const void* dict, size_t dictSize,
                                int compressionLevel)
 {
-    ZSTD_parameters const params = ZSTD_getParams_internal(compressionLevel, srcSize, dict ? dictSize : 0);
-    ZSTD_CCtx_params cctxParams = ZSTD_assignParamsToCCtxParams(&cctx->requestedParams, &params);
+    {
+        ZSTD_parameters const params = ZSTD_getParams_internal(compressionLevel, srcSize, dict ? dictSize : 0, ZSTD_cpm_noAttachDict);
+        assert(params.fParams.contentSizeFlag == 1);
+        ZSTD_CCtxParams_init_internal(&cctx->simpleApiParams, &params, (compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT: compressionLevel);
+    }
     DEBUGLOG(4, "ZSTD_compress_usingDict (srcSize=%u)", (unsigned)srcSize);
-    assert(params.fParams.contentSizeFlag == 1);
-    return ZSTD_compress_advanced_internal(cctx, dst, dstCapacity, src, srcSize, dict, dictSize, &cctxParams);
+    return ZSTD_compress_advanced_internal(cctx, dst, dstCapacity, src, srcSize, dict, dictSize, &cctx->simpleApiParams);
 }
 
 size_t ZSTD_compressCCtx(ZSTD_CCtx* cctx,
@@ -3309,10 +5162,17 @@ size_t ZSTD_compress(void* dst, size_t dstCapacity,
                      int compressionLevel)
 {
     size_t result;
+#if ZSTD_COMPRESS_HEAPMODE
+    ZSTD_CCtx* cctx = ZSTD_createCCtx();
+    RETURN_ERROR_IF(!cctx, memory_allocation, "ZSTD_createCCtx failed");
+    result = ZSTD_compressCCtx(cctx, dst, dstCapacity, src, srcSize, compressionLevel);
+    ZSTD_freeCCtx(cctx);
+#else
     ZSTD_CCtx ctxBody;
     ZSTD_initCCtx(&ctxBody, ZSTD_defaultCMem);
     result = ZSTD_compressCCtx(&ctxBody, dst, dstCapacity, src, srcSize, compressionLevel);
     ZSTD_freeCCtxContent(&ctxBody);   /* can't free ctxBody itself, as it's on stack; free only heap content */
+#endif
     return result;
 }
 
@@ -3328,14 +5188,17 @@ size_t ZSTD_estimateCDictSize_advanced(
     DEBUGLOG(5, "sizeof(ZSTD_CDict) : %u", (unsigned)sizeof(ZSTD_CDict));
     return ZSTD_cwksp_alloc_size(sizeof(ZSTD_CDict))
          + ZSTD_cwksp_alloc_size(HUF_WORKSPACE_SIZE)
-         + ZSTD_sizeof_matchState(&cParams, /* forCCtx */ 0)
+         /* enableDedicatedDictSearch == 1 ensures that CDict estimation will not be too small
+          * in case we are using DDS with row-hash. */
+         + ZSTD_sizeof_matchState(&cParams, ZSTD_resolveRowMatchFinderMode(ZSTD_ps_auto, &cParams),
+                                  /* enableDedicatedDictSearch */ 1, /* forCCtx */ 0)
          + (dictLoadMethod == ZSTD_dlm_byRef ? 0
             : ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(dictSize, sizeof(void *))));
 }
 
 size_t ZSTD_estimateCDictSize(size_t dictSize, int compressionLevel)
 {
-    ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize);
+    ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_createCDict);
     return ZSTD_estimateCDictSize_advanced(dictSize, cParams, ZSTD_dlm_byCopy);
 }
 
@@ -3353,20 +5216,22 @@ static size_t ZSTD_initCDict_internal(
               const void* dictBuffer, size_t dictSize,
                     ZSTD_dictLoadMethod_e dictLoadMethod,
                     ZSTD_dictContentType_e dictContentType,
-                    ZSTD_compressionParameters cParams)
+                    ZSTD_CCtx_params params)
 {
     DEBUGLOG(3, "ZSTD_initCDict_internal (dictContentType:%u)", (unsigned)dictContentType);
-    assert(!ZSTD_checkCParams(cParams));
-    cdict->matchState.cParams = cParams;
+    assert(!ZSTD_checkCParams(params.cParams));
+    cdict->matchState.cParams = params.cParams;
+    cdict->matchState.dedicatedDictSearch = params.enableDedicatedDictSearch;
     if ((dictLoadMethod == ZSTD_dlm_byRef) || (!dictBuffer) || (!dictSize)) {
         cdict->dictContent = dictBuffer;
     } else {
          void *internalBuffer = ZSTD_cwksp_reserve_object(&cdict->workspace, ZSTD_cwksp_align(dictSize, sizeof(void*)));
         RETURN_ERROR_IF(!internalBuffer, memory_allocation, "NULL pointer!");
         cdict->dictContent = internalBuffer;
-        memcpy(internalBuffer, dictBuffer, dictSize);
+        ZSTD_memcpy(internalBuffer, dictBuffer, dictSize);
     }
     cdict->dictContentSize = dictSize;
+    cdict->dictContentType = dictContentType;
 
     cdict->entropyWorkspace = (U32*)ZSTD_cwksp_reserve_object(&cdict->workspace, HUF_WORKSPACE_SIZE);
 
@@ -3376,22 +5241,20 @@ static size_t ZSTD_initCDict_internal(
     FORWARD_IF_ERROR(ZSTD_reset_matchState(
         &cdict->matchState,
         &cdict->workspace,
-        &cParams,
+        &params.cParams,
+        params.useRowMatchFinder,
         ZSTDcrp_makeClean,
         ZSTDirp_reset,
         ZSTD_resetTarget_CDict), "");
     /* (Maybe) load the dictionary
      * Skips loading the dictionary if it is < 8 bytes.
      */
-    {   ZSTD_CCtx_params params;
-        memset(&params, 0, sizeof(params));
-        params.compressionLevel = ZSTD_CLEVEL_DEFAULT;
+    {   params.compressionLevel = ZSTD_CLEVEL_DEFAULT;
         params.fParams.contentSizeFlag = 1;
-        params.cParams = cParams;
         {   size_t const dictID = ZSTD_compress_insertDictionary(
                     &cdict->cBlockState, &cdict->matchState, NULL, &cdict->workspace,
                     &params, cdict->dictContent, cdict->dictContentSize,
-                    dictContentType, ZSTD_dtlm_full, cdict->entropyWorkspace);
+                    dictContentType, ZSTD_dtlm_full, ZSTD_tfp_forCDict, cdict->entropyWorkspace);
             FORWARD_IF_ERROR(dictID, "ZSTD_compress_insertDictionary failed");
             assert(dictID <= (size_t)(U32)-1);
             cdict->dictID = (U32)dictID;
@@ -3401,66 +5264,129 @@ static size_t ZSTD_initCDict_internal(
     return 0;
 }
 
-ZSTD_CDict* ZSTD_createCDict_advanced(const void* dictBuffer, size_t dictSize,
+static ZSTD_CDict* ZSTD_createCDict_advanced_internal(size_t dictSize,
                                       ZSTD_dictLoadMethod_e dictLoadMethod,
-                                      ZSTD_dictContentType_e dictContentType,
-                                      ZSTD_compressionParameters cParams, ZSTD_customMem customMem)
+                                      ZSTD_compressionParameters cParams,
+                                      ZSTD_paramSwitch_e useRowMatchFinder,
+                                      U32 enableDedicatedDictSearch,
+                                      ZSTD_customMem customMem)
 {
-    DEBUGLOG(3, "ZSTD_createCDict_advanced, mode %u", (unsigned)dictContentType);
-    if (!customMem.customAlloc ^ !customMem.customFree) return NULL;
+    if ((!customMem.customAlloc) ^ (!customMem.customFree)) return NULL;
 
     {   size_t const workspaceSize =
             ZSTD_cwksp_alloc_size(sizeof(ZSTD_CDict)) +
             ZSTD_cwksp_alloc_size(HUF_WORKSPACE_SIZE) +
-            ZSTD_sizeof_matchState(&cParams, /* forCCtx */ 0) +
+            ZSTD_sizeof_matchState(&cParams, useRowMatchFinder, enableDedicatedDictSearch, /* forCCtx */ 0) +
             (dictLoadMethod == ZSTD_dlm_byRef ? 0
              : ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(dictSize, sizeof(void*))));
-        void* const workspace = ZSTD_malloc(workspaceSize, customMem);
+        void* const workspace = ZSTD_customMalloc(workspaceSize, customMem);
         ZSTD_cwksp ws;
         ZSTD_CDict* cdict;
 
         if (!workspace) {
-            ZSTD_free(workspace, customMem);
+            ZSTD_customFree(workspace, customMem);
             return NULL;
         }
 
-        ZSTD_cwksp_init(&ws, workspace, workspaceSize);
+        ZSTD_cwksp_init(&ws, workspace, workspaceSize, ZSTD_cwksp_dynamic_alloc);
 
         cdict = (ZSTD_CDict*)ZSTD_cwksp_reserve_object(&ws, sizeof(ZSTD_CDict));
         assert(cdict != NULL);
         ZSTD_cwksp_move(&cdict->workspace, &ws);
         cdict->customMem = customMem;
-        cdict->compressionLevel = 0; /* signals advanced API usage */
+        cdict->compressionLevel = ZSTD_NO_CLEVEL; /* signals advanced API usage */
+        cdict->useRowMatchFinder = useRowMatchFinder;
+        return cdict;
+    }
+}
 
-        if (ZSTD_isError( ZSTD_initCDict_internal(cdict,
-                                        dictBuffer, dictSize,
-                                        dictLoadMethod, dictContentType,
-                                        cParams) )) {
-            ZSTD_freeCDict(cdict);
-            return NULL;
-        }
+ZSTD_CDict* ZSTD_createCDict_advanced(const void* dictBuffer, size_t dictSize,
+                                      ZSTD_dictLoadMethod_e dictLoadMethod,
+                                      ZSTD_dictContentType_e dictContentType,
+                                      ZSTD_compressionParameters cParams,
+                                      ZSTD_customMem customMem)
+{
+    ZSTD_CCtx_params cctxParams;
+    ZSTD_memset(&cctxParams, 0, sizeof(cctxParams));
+    ZSTD_CCtxParams_init(&cctxParams, 0);
+    cctxParams.cParams = cParams;
+    cctxParams.customMem = customMem;
+    return ZSTD_createCDict_advanced2(
+        dictBuffer, dictSize,
+        dictLoadMethod, dictContentType,
+        &cctxParams, customMem);
+}
+
+ZSTD_CDict* ZSTD_createCDict_advanced2(
+        const void* dict, size_t dictSize,
+        ZSTD_dictLoadMethod_e dictLoadMethod,
+        ZSTD_dictContentType_e dictContentType,
+        const ZSTD_CCtx_params* originalCctxParams,
+        ZSTD_customMem customMem)
+{
+    ZSTD_CCtx_params cctxParams = *originalCctxParams;
+    ZSTD_compressionParameters cParams;
+    ZSTD_CDict* cdict;
 
-        return cdict;
+    DEBUGLOG(3, "ZSTD_createCDict_advanced2, mode %u", (unsigned)dictContentType);
+    if (!customMem.customAlloc ^ !customMem.customFree) return NULL;
+
+    if (cctxParams.enableDedicatedDictSearch) {
+        cParams = ZSTD_dedicatedDictSearch_getCParams(
+            cctxParams.compressionLevel, dictSize);
+        ZSTD_overrideCParams(&cParams, &cctxParams.cParams);
+    } else {
+        cParams = ZSTD_getCParamsFromCCtxParams(
+            &cctxParams, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_createCDict);
+    }
+
+    if (!ZSTD_dedicatedDictSearch_isSupported(&cParams)) {
+        /* Fall back to non-DDSS params */
+        cctxParams.enableDedicatedDictSearch = 0;
+        cParams = ZSTD_getCParamsFromCCtxParams(
+            &cctxParams, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_createCDict);
     }
+
+    DEBUGLOG(3, "ZSTD_createCDict_advanced2: DDS: %u", cctxParams.enableDedicatedDictSearch);
+    cctxParams.cParams = cParams;
+    cctxParams.useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(cctxParams.useRowMatchFinder, &cParams);
+
+    cdict = ZSTD_createCDict_advanced_internal(dictSize,
+                        dictLoadMethod, cctxParams.cParams,
+                        cctxParams.useRowMatchFinder, cctxParams.enableDedicatedDictSearch,
+                        customMem);
+
+    if (ZSTD_isError( ZSTD_initCDict_internal(cdict,
+                                    dict, dictSize,
+                                    dictLoadMethod, dictContentType,
+                                    cctxParams) )) {
+        ZSTD_freeCDict(cdict);
+        return NULL;
+    }
+
+    return cdict;
 }
 
 ZSTD_CDict* ZSTD_createCDict(const void* dict, size_t dictSize, int compressionLevel)
 {
-    ZSTD_compressionParameters cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize);
-    ZSTD_CDict* cdict = ZSTD_createCDict_advanced(dict, dictSize,
+    ZSTD_compressionParameters cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_createCDict);
+    ZSTD_CDict* const cdict = ZSTD_createCDict_advanced(dict, dictSize,
                                                   ZSTD_dlm_byCopy, ZSTD_dct_auto,
                                                   cParams, ZSTD_defaultCMem);
     if (cdict)
-        cdict->compressionLevel = compressionLevel == 0 ? ZSTD_CLEVEL_DEFAULT : compressionLevel;
+        cdict->compressionLevel = (compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT : compressionLevel;
     return cdict;
 }
 
 ZSTD_CDict* ZSTD_createCDict_byReference(const void* dict, size_t dictSize, int compressionLevel)
 {
-    ZSTD_compressionParameters cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize);
-    return ZSTD_createCDict_advanced(dict, dictSize,
+    ZSTD_compressionParameters cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_createCDict);
+    ZSTD_CDict* const cdict = ZSTD_createCDict_advanced(dict, dictSize,
                                      ZSTD_dlm_byRef, ZSTD_dct_auto,
                                      cParams, ZSTD_defaultCMem);
+    if (cdict)
+        cdict->compressionLevel = (compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT : compressionLevel;
+    return cdict;
 }
 
 size_t ZSTD_freeCDict(ZSTD_CDict* cdict)
@@ -3470,7 +5396,7 @@ size_t ZSTD_freeCDict(ZSTD_CDict* cdict)
         int cdictInWorkspace = ZSTD_cwksp_owns_buffer(&cdict->workspace, cdict);
         ZSTD_cwksp_free(&cdict->workspace, cMem);
         if (!cdictInWorkspace) {
-            ZSTD_free(cdict, cMem);
+            ZSTD_customFree(cdict, cMem);
         }
         return 0;
     }
@@ -3496,19 +5422,22 @@ const ZSTD_CDict* ZSTD_initStaticCDict(
                                  ZSTD_dictContentType_e dictContentType,
                                  ZSTD_compressionParameters cParams)
 {
-    size_t const matchStateSize = ZSTD_sizeof_matchState(&cParams, /* forCCtx */ 0);
+    ZSTD_paramSwitch_e const useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(ZSTD_ps_auto, &cParams);
+    /* enableDedicatedDictSearch == 1 ensures matchstate is not too small in case this CDict will be used for DDS + row hash */
+    size_t const matchStateSize = ZSTD_sizeof_matchState(&cParams, useRowMatchFinder, /* enableDedicatedDictSearch */ 1, /* forCCtx */ 0);
     size_t const neededSize = ZSTD_cwksp_alloc_size(sizeof(ZSTD_CDict))
                             + (dictLoadMethod == ZSTD_dlm_byRef ? 0
                                : ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(dictSize, sizeof(void*))))
                             + ZSTD_cwksp_alloc_size(HUF_WORKSPACE_SIZE)
                             + matchStateSize;
     ZSTD_CDict* cdict;
+    ZSTD_CCtx_params params;
 
     if ((size_t)workspace & 7) return NULL;  /* 8-aligned */
 
     {
         ZSTD_cwksp ws;
-        ZSTD_cwksp_init(&ws, workspace, workspaceSize);
+        ZSTD_cwksp_init(&ws, workspace, workspaceSize, ZSTD_cwksp_static_alloc);
         cdict = (ZSTD_CDict*)ZSTD_cwksp_reserve_object(&ws, sizeof(ZSTD_CDict));
         if (cdict == NULL) return NULL;
         ZSTD_cwksp_move(&cdict->workspace, &ws);
@@ -3518,10 +5447,15 @@ const ZSTD_CDict* ZSTD_initStaticCDict(
         (unsigned)workspaceSize, (unsigned)neededSize, (unsigned)(workspaceSize < neededSize));
     if (workspaceSize < neededSize) return NULL;
 
+    ZSTD_CCtxParams_init(&params, 0);
+    params.cParams = cParams;
+    params.useRowMatchFinder = useRowMatchFinder;
+    cdict->useRowMatchFinder = useRowMatchFinder;
+
     if (ZSTD_isError( ZSTD_initCDict_internal(cdict,
                                               dict, dictSize,
                                               dictLoadMethod, dictContentType,
-                                              cParams) ))
+                                              params) ))
         return NULL;
 
     return cdict;
@@ -3533,61 +5467,98 @@ ZSTD_compressionParameters ZSTD_getCParamsFromCDict(const ZSTD_CDict* cdict)
     return cdict->matchState.cParams;
 }
 
-/* ZSTD_compressBegin_usingCDict_advanced() :
- * cdict must be != NULL */
-size_t ZSTD_compressBegin_usingCDict_advanced(
+/*! ZSTD_getDictID_fromCDict() :
+ *  Provides the dictID of the dictionary loaded into `cdict`.
+ *  If @return == 0, the dictionary is not conformant to Zstandard specification, or empty.
+ *  Non-conformant dictionaries can still be loaded, but as content-only dictionaries. */
+unsigned ZSTD_getDictID_fromCDict(const ZSTD_CDict* cdict)
+{
+    if (cdict==NULL) return 0;
+    return cdict->dictID;
+}
+
+/* ZSTD_compressBegin_usingCDict_internal() :
+ * Implementation of various ZSTD_compressBegin_usingCDict* functions.
+ */
+static size_t ZSTD_compressBegin_usingCDict_internal(
     ZSTD_CCtx* const cctx, const ZSTD_CDict* const cdict,
     ZSTD_frameParameters const fParams, unsigned long long const pledgedSrcSize)
 {
-    DEBUGLOG(4, "ZSTD_compressBegin_usingCDict_advanced");
+    ZSTD_CCtx_params cctxParams;
+    DEBUGLOG(4, "ZSTD_compressBegin_usingCDict_internal");
     RETURN_ERROR_IF(cdict==NULL, dictionary_wrong, "NULL pointer!");
-    {   ZSTD_CCtx_params params = cctx->requestedParams;
+    /* Initialize the cctxParams from the cdict */
+    {
+        ZSTD_parameters params;
+        params.fParams = fParams;
         params.cParams = ( pledgedSrcSize < ZSTD_USE_CDICT_PARAMS_SRCSIZE_CUTOFF
                         || pledgedSrcSize < cdict->dictContentSize * ZSTD_USE_CDICT_PARAMS_DICTSIZE_MULTIPLIER
                         || pledgedSrcSize == ZSTD_CONTENTSIZE_UNKNOWN
-                        || cdict->compressionLevel == 0 )
-                      && (params.attachDictPref != ZSTD_dictForceLoad) ?
+                        || cdict->compressionLevel == 0 ) ?
                 ZSTD_getCParamsFromCDict(cdict)
               : ZSTD_getCParams(cdict->compressionLevel,
                                 pledgedSrcSize,
                                 cdict->dictContentSize);
-        /* Increase window log to fit the entire dictionary and source if the
-         * source size is known. Limit the increase to 19, which is the
-         * window log for compression level 1 with the largest source size.
-         */
-        if (pledgedSrcSize != ZSTD_CONTENTSIZE_UNKNOWN) {
-            U32 const limitedSrcSize = (U32)MIN(pledgedSrcSize, 1U << 19);
-            U32 const limitedSrcLog = limitedSrcSize > 1 ? ZSTD_highbit32(limitedSrcSize - 1) + 1 : 1;
-            params.cParams.windowLog = MAX(params.cParams.windowLog, limitedSrcLog);
-        }
-        params.fParams = fParams;
-        return ZSTD_compressBegin_internal(cctx,
-                                           NULL, 0, ZSTD_dct_auto, ZSTD_dtlm_fast,
-                                           cdict,
-                                           &params, pledgedSrcSize,
-                                           ZSTDb_not_buffered);
+        ZSTD_CCtxParams_init_internal(&cctxParams, &params, cdict->compressionLevel);
+    }
+    /* Increase window log to fit the entire dictionary and source if the
+     * source size is known. Limit the increase to 19, which is the
+     * window log for compression level 1 with the largest source size.
+     */
+    if (pledgedSrcSize != ZSTD_CONTENTSIZE_UNKNOWN) {
+        U32 const limitedSrcSize = (U32)MIN(pledgedSrcSize, 1U << 19);
+        U32 const limitedSrcLog = limitedSrcSize > 1 ? ZSTD_highbit32(limitedSrcSize - 1) + 1 : 1;
+        cctxParams.cParams.windowLog = MAX(cctxParams.cParams.windowLog, limitedSrcLog);
     }
+    return ZSTD_compressBegin_internal(cctx,
+                                        NULL, 0, ZSTD_dct_auto, ZSTD_dtlm_fast,
+                                        cdict,
+                                        &cctxParams, pledgedSrcSize,
+                                        ZSTDb_not_buffered);
+}
+
+
+/* ZSTD_compressBegin_usingCDict_advanced() :
+ * This function is DEPRECATED.
+ * cdict must be != NULL */
+size_t ZSTD_compressBegin_usingCDict_advanced(
+    ZSTD_CCtx* const cctx, const ZSTD_CDict* const cdict,
+    ZSTD_frameParameters const fParams, unsigned long long const pledgedSrcSize)
+{
+    return ZSTD_compressBegin_usingCDict_internal(cctx, cdict, fParams, pledgedSrcSize);
 }
 
 /* ZSTD_compressBegin_usingCDict() :
- * pledgedSrcSize=0 means "unknown"
- * if pledgedSrcSize>0, it will enable contentSizeFlag */
+ * cdict must be != NULL */
 size_t ZSTD_compressBegin_usingCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict)
 {
     ZSTD_frameParameters const fParams = { 0 /*content*/, 0 /*checksum*/, 0 /*noDictID*/ };
-    DEBUGLOG(4, "ZSTD_compressBegin_usingCDict : dictIDFlag == %u", !fParams.noDictIDFlag);
-    return ZSTD_compressBegin_usingCDict_advanced(cctx, cdict, fParams, ZSTD_CONTENTSIZE_UNKNOWN);
+    return ZSTD_compressBegin_usingCDict_internal(cctx, cdict, fParams, ZSTD_CONTENTSIZE_UNKNOWN);
 }
 
-size_t ZSTD_compress_usingCDict_advanced(ZSTD_CCtx* cctx,
+/*! ZSTD_compress_usingCDict_internal():
+ * Implementation of various ZSTD_compress_usingCDict* functions.
+ */
+static size_t ZSTD_compress_usingCDict_internal(ZSTD_CCtx* cctx,
                                 void* dst, size_t dstCapacity,
                                 const void* src, size_t srcSize,
                                 const ZSTD_CDict* cdict, ZSTD_frameParameters fParams)
 {
-    FORWARD_IF_ERROR(ZSTD_compressBegin_usingCDict_advanced(cctx, cdict, fParams, srcSize), "");   /* will check if cdict != NULL */
+    FORWARD_IF_ERROR(ZSTD_compressBegin_usingCDict_internal(cctx, cdict, fParams, srcSize), ""); /* will check if cdict != NULL */
     return ZSTD_compressEnd(cctx, dst, dstCapacity, src, srcSize);
 }
 
+/*! ZSTD_compress_usingCDict_advanced():
+ * This function is DEPRECATED.
+ */
+size_t ZSTD_compress_usingCDict_advanced(ZSTD_CCtx* cctx,
+                                void* dst, size_t dstCapacity,
+                                const void* src, size_t srcSize,
+                                const ZSTD_CDict* cdict, ZSTD_frameParameters fParams)
+{
+    return ZSTD_compress_usingCDict_internal(cctx, dst, dstCapacity, src, srcSize, cdict, fParams);
+}
+
 /*! ZSTD_compress_usingCDict() :
  *  Compression using a digested Dictionary.
  *  Faster startup than ZSTD_compress_usingDict(), recommended when same dictionary is used multiple times.
@@ -3599,7 +5570,7 @@ size_t ZSTD_compress_usingCDict(ZSTD_CCtx* cctx,
                                 const ZSTD_CDict* cdict)
 {
     ZSTD_frameParameters const fParams = { 1 /*content*/, 0 /*checksum*/, 0 /*noDictID*/ };
-    return ZSTD_compress_usingCDict_advanced(cctx, dst, dstCapacity, src, srcSize, cdict, fParams);
+    return ZSTD_compress_usingCDict_internal(cctx, dst, dstCapacity, src, srcSize, cdict, fParams);
 }
 
 
@@ -3640,32 +5611,12 @@ size_t ZSTD_CStreamOutSize(void)
     return ZSTD_compressBound(ZSTD_BLOCKSIZE_MAX) + ZSTD_blockHeaderSize + 4 /* 32-bits hash */ ;
 }
 
-static size_t ZSTD_resetCStream_internal(ZSTD_CStream* cctx,
-                    const void* const dict, size_t const dictSize, ZSTD_dictContentType_e const dictContentType,
-                    const ZSTD_CDict* const cdict,
-                    ZSTD_CCtx_params params, unsigned long long const pledgedSrcSize)
+static ZSTD_cParamMode_e ZSTD_getCParamMode(ZSTD_CDict const* cdict, ZSTD_CCtx_params const* params, U64 pledgedSrcSize)
 {
-    DEBUGLOG(4, "ZSTD_resetCStream_internal");
-    /* Finalize the compression parameters */
-    params.cParams = ZSTD_getCParamsFromCCtxParams(&params, pledgedSrcSize, dictSize);
-    /* params are supposed to be fully validated at this point */
-    assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams)));
-    assert(!((dict) && (cdict)));  /* either dict or cdict, not both */
-
-    FORWARD_IF_ERROR( ZSTD_compressBegin_internal(cctx,
-                                         dict, dictSize, dictContentType, ZSTD_dtlm_fast,
-                                         cdict,
-                                         &params, pledgedSrcSize,
-                                         ZSTDb_buffered) , "");
-
-    cctx->inToCompress = 0;
-    cctx->inBuffPos = 0;
-    cctx->inBuffTarget = cctx->blockSize
-                      + (cctx->blockSize == pledgedSrcSize);   /* for small input: avoid automatic flush on reaching end of block, since it would require to add a 3-bytes null block to end frame */
-    cctx->outBuffContentSize = cctx->outBuffFlushedSize = 0;
-    cctx->streamStage = zcss_load;
-    cctx->frameEnded = 0;
-    return 0;   /* ready to go */
+    if (cdict != NULL && ZSTD_shouldAttachDict(cdict, params, pledgedSrcSize))
+        return ZSTD_cpm_attachDict;
+    else
+        return ZSTD_cpm_noAttachDict;
 }
 
 /* ZSTD_resetCStream():
@@ -3749,7 +5700,7 @@ size_t ZSTD_initCStream_advanced(ZSTD_CStream* zcs,
     FORWARD_IF_ERROR( ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only) , "");
     FORWARD_IF_ERROR( ZSTD_CCtx_setPledgedSrcSize(zcs, pledgedSrcSize) , "");
     FORWARD_IF_ERROR( ZSTD_checkCParams(params.cParams) , "");
-    zcs->requestedParams = ZSTD_assignParamsToCCtxParams(&zcs->requestedParams, &params);
+    ZSTD_CCtxParams_setZstdParams(&zcs->requestedParams, &params);
     FORWARD_IF_ERROR( ZSTD_CCtx_loadDictionary(zcs, dict, dictSize) , "");
     return 0;
 }
@@ -3791,36 +5742,54 @@ size_t ZSTD_initCStream(ZSTD_CStream* zcs, int compressionLevel)
 
 static size_t ZSTD_nextInputSizeHint(const ZSTD_CCtx* cctx)
 {
-    size_t hintInSize = cctx->inBuffTarget - cctx->inBuffPos;
-    if (hintInSize==0) hintInSize = cctx->blockSize;
-    return hintInSize;
+    if (cctx->appliedParams.inBufferMode == ZSTD_bm_stable) {
+        return cctx->blockSize - cctx->stableIn_notConsumed;
+    }
+    assert(cctx->appliedParams.inBufferMode == ZSTD_bm_buffered);
+    {   size_t hintInSize = cctx->inBuffTarget - cctx->inBuffPos;
+        if (hintInSize==0) hintInSize = cctx->blockSize;
+        return hintInSize;
+    }
 }
 
 /** ZSTD_compressStream_generic():
  *  internal function for all *compressStream*() variants
- *  non-static, because can be called from zstdmt_compress.c
- * @return : hint size for next input */
+ * @return : hint size for next input to complete ongoing block */
 static size_t ZSTD_compressStream_generic(ZSTD_CStream* zcs,
                                           ZSTD_outBuffer* output,
                                           ZSTD_inBuffer* input,
                                           ZSTD_EndDirective const flushMode)
 {
-    const char* const istart = (const char*)input->src;
-    const char* const iend = input->size != 0 ? istart + input->size : istart;
-    const char* ip = input->pos != 0 ? istart + input->pos : istart;
-    char* const ostart = (char*)output->dst;
-    char* const oend = output->size != 0 ? ostart + output->size : ostart;
-    char* op = output->pos != 0 ? ostart + output->pos : ostart;
+    const char* const istart = (assert(input != NULL), (const char*)input->src);
+    const char* const iend = (istart != NULL) ? istart + input->size : istart;
+    const char* ip = (istart != NULL) ? istart + input->pos : istart;
+    char* const ostart = (assert(output != NULL), (char*)output->dst);
+    char* const oend = (ostart != NULL) ? ostart + output->size : ostart;
+    char* op = (ostart != NULL) ? ostart + output->pos : ostart;
     U32 someMoreWork = 1;
 
     /* check expectations */
-    DEBUGLOG(5, "ZSTD_compressStream_generic, flush=%u", (unsigned)flushMode);
-    assert(zcs->inBuff != NULL);
-    assert(zcs->inBuffSize > 0);
-    assert(zcs->outBuff !=  NULL);
-    assert(zcs->outBuffSize > 0);
-    assert(output->pos <= output->size);
+    DEBUGLOG(5, "ZSTD_compressStream_generic, flush=%i, srcSize = %zu", (int)flushMode, input->size - input->pos);
+    assert(zcs != NULL);
+    if (zcs->appliedParams.inBufferMode == ZSTD_bm_stable) {
+        assert(input->pos >= zcs->stableIn_notConsumed);
+        input->pos -= zcs->stableIn_notConsumed;
+        ip -= zcs->stableIn_notConsumed;
+        zcs->stableIn_notConsumed = 0;
+    }
+    if (zcs->appliedParams.inBufferMode == ZSTD_bm_buffered) {
+        assert(zcs->inBuff != NULL);
+        assert(zcs->inBuffSize > 0);
+    }
+    if (zcs->appliedParams.outBufferMode == ZSTD_bm_buffered) {
+        assert(zcs->outBuff !=  NULL);
+        assert(zcs->outBuffSize > 0);
+    }
+    if (input->src == NULL) assert(input->size == 0);
     assert(input->pos <= input->size);
+    if (output->dst == NULL) assert(output->size == 0);
+    assert(output->pos <= output->size);
+    assert((U32)flushMode <= (U32)ZSTD_e_end);
 
     while (someMoreWork) {
         switch(zcs->streamStage)
@@ -3830,7 +5799,8 @@ static size_t ZSTD_compressStream_generic(ZSTD_CStream* zcs,
 
         case zcss_load:
             if ( (flushMode == ZSTD_e_end)
-              && ((size_t)(oend-op) >= ZSTD_compressBound(iend-ip))  /* enough dstCapacity */
+              && ( (size_t)(oend-op) >= ZSTD_compressBound(iend-ip)     /* Enough output space */
+                || zcs->appliedParams.outBufferMode == ZSTD_bm_stable)  /* OR we are allowed to return dstSizeTooSmall */
               && (zcs->inBuffPos == 0) ) {
                 /* shortcut to compression pass directly into output buffer */
                 size_t const cSize = ZSTD_compressEnd(zcs,
@@ -3843,14 +5813,14 @@ static size_t ZSTD_compressStream_generic(ZSTD_CStream* zcs,
                 ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only);
                 someMoreWork = 0; break;
             }
-            /* complete loading into inBuffer */
-            {   size_t const toLoad = zcs->inBuffTarget - zcs->inBuffPos;
+            /* complete loading into inBuffer in buffered mode */
+            if (zcs->appliedParams.inBufferMode == ZSTD_bm_buffered) {
+                size_t const toLoad = zcs->inBuffTarget - zcs->inBuffPos;
                 size_t const loaded = ZSTD_limitCopy(
                                         zcs->inBuff + zcs->inBuffPos, toLoad,
                                         ip, iend-ip);
                 zcs->inBuffPos += loaded;
-                if (loaded != 0)
-                    ip += loaded;
+                if (ip) ip += loaded;
                 if ( (flushMode == ZSTD_e_continue)
                   && (zcs->inBuffPos < zcs->inBuffTarget) ) {
                     /* not enough input to fill full block : stop here */
@@ -3861,34 +5831,62 @@ static size_t ZSTD_compressStream_generic(ZSTD_CStream* zcs,
                     /* empty */
                     someMoreWork = 0; break;
                 }
+            } else {
+                assert(zcs->appliedParams.inBufferMode == ZSTD_bm_stable);
+                if ( (flushMode == ZSTD_e_continue)
+                  && ( (size_t)(iend - ip) < zcs->blockSize) ) {
+                    /* can't compress a full block : stop here */
+                    zcs->stableIn_notConsumed = (size_t)(iend - ip);
+                    ip = iend;  /* pretend to have consumed input */
+                    someMoreWork = 0; break;
+                }
+                if ( (flushMode == ZSTD_e_flush)
+                  && (ip == iend) ) {
+                    /* empty */
+                    someMoreWork = 0; break;
+                }
             }
             /* compress current block (note : this stage cannot be stopped in the middle) */
             DEBUGLOG(5, "stream compression stage (flushMode==%u)", flushMode);
-            {   void* cDst;
+            {   int const inputBuffered = (zcs->appliedParams.inBufferMode == ZSTD_bm_buffered);
+                void* cDst;
                 size_t cSize;
-                size_t const iSize = zcs->inBuffPos - zcs->inToCompress;
                 size_t oSize = oend-op;
-                unsigned const lastBlock = (flushMode == ZSTD_e_end) && (ip==iend);
-                if (oSize >= ZSTD_compressBound(iSize))
+                size_t const iSize = inputBuffered ? zcs->inBuffPos - zcs->inToCompress
+                                                   : MIN((size_t)(iend - ip), zcs->blockSize);
+                if (oSize >= ZSTD_compressBound(iSize) || zcs->appliedParams.outBufferMode == ZSTD_bm_stable)
                     cDst = op;   /* compress into output buffer, to skip flush stage */
                 else
                     cDst = zcs->outBuff, oSize = zcs->outBuffSize;
-                cSize = lastBlock ?
-                        ZSTD_compressEnd(zcs, cDst, oSize,
-                                    zcs->inBuff + zcs->inToCompress, iSize) :
-                        ZSTD_compressContinue(zcs, cDst, oSize,
-                                    zcs->inBuff + zcs->inToCompress, iSize);
-                FORWARD_IF_ERROR(cSize, "%s", lastBlock ? "ZSTD_compressEnd failed" : "ZSTD_compressContinue failed");
-                zcs->frameEnded = lastBlock;
-                /* prepare next block */
-                zcs->inBuffTarget = zcs->inBuffPos + zcs->blockSize;
-                if (zcs->inBuffTarget > zcs->inBuffSize)
-                    zcs->inBuffPos = 0, zcs->inBuffTarget = zcs->blockSize;
-                DEBUGLOG(5, "inBuffTarget:%u / inBuffSize:%u",
-                         (unsigned)zcs->inBuffTarget, (unsigned)zcs->inBuffSize);
-                if (!lastBlock)
-                    assert(zcs->inBuffTarget <= zcs->inBuffSize);
-                zcs->inToCompress = zcs->inBuffPos;
+                if (inputBuffered) {
+                    unsigned const lastBlock = (flushMode == ZSTD_e_end) && (ip==iend);
+                    cSize = lastBlock ?
+                            ZSTD_compressEnd(zcs, cDst, oSize,
+                                        zcs->inBuff + zcs->inToCompress, iSize) :
+                            ZSTD_compressContinue(zcs, cDst, oSize,
+                                        zcs->inBuff + zcs->inToCompress, iSize);
+                    FORWARD_IF_ERROR(cSize, "%s", lastBlock ? "ZSTD_compressEnd failed" : "ZSTD_compressContinue failed");
+                    zcs->frameEnded = lastBlock;
+                    /* prepare next block */
+                    zcs->inBuffTarget = zcs->inBuffPos + zcs->blockSize;
+                    if (zcs->inBuffTarget > zcs->inBuffSize)
+                        zcs->inBuffPos = 0, zcs->inBuffTarget = zcs->blockSize;
+                    DEBUGLOG(5, "inBuffTarget:%u / inBuffSize:%u",
+                            (unsigned)zcs->inBuffTarget, (unsigned)zcs->inBuffSize);
+                    if (!lastBlock)
+                        assert(zcs->inBuffTarget <= zcs->inBuffSize);
+                    zcs->inToCompress = zcs->inBuffPos;
+                } else { /* !inputBuffered, hence ZSTD_bm_stable */
+                    unsigned const lastBlock = (flushMode == ZSTD_e_end) && (ip + iSize == iend);
+                    cSize = lastBlock ?
+                            ZSTD_compressEnd(zcs, cDst, oSize, ip, iSize) :
+                            ZSTD_compressContinue(zcs, cDst, oSize, ip, iSize);
+                    /* Consume the input prior to error checking to mirror buffered mode. */
+                    if (ip) ip += iSize;
+                    FORWARD_IF_ERROR(cSize, "%s", lastBlock ? "ZSTD_compressEnd failed" : "ZSTD_compressContinue failed");
+                    zcs->frameEnded = lastBlock;
+                    if (lastBlock) assert(ip == iend);
+                }
                 if (cDst == op) {  /* no need to flush */
                     op += cSize;
                     if (zcs->frameEnded) {
@@ -3902,9 +5900,10 @@ static size_t ZSTD_compressStream_generic(ZSTD_CStream* zcs,
                 zcs->outBuffFlushedSize = 0;
                 zcs->streamStage = zcss_flush; /* pass-through to flush stage */
             }
-	    /* fall-through */
+	    ZSTD_FALLTHROUGH;
         case zcss_flush:
             DEBUGLOG(5, "flush stage");
+            assert(zcs->appliedParams.outBufferMode == ZSTD_bm_buffered);
             {   size_t const toFlush = zcs->outBuffContentSize - zcs->outBuffFlushedSize;
                 size_t const flushed = ZSTD_limitCopy(op, (size_t)(oend-op),
                             zcs->outBuff + zcs->outBuffFlushedSize, toFlush);
@@ -3959,7 +5958,141 @@ size_t ZSTD_compressStream(ZSTD_CStream* zcs, ZSTD_outBuffer* output, ZSTD_inBuf
     return ZSTD_nextInputSizeHint_MTorST(zcs);
 }
 
+/* After a compression call set the expected input/output buffer.
+ * This is validated at the start of the next compression call.
+ */
+static void
+ZSTD_setBufferExpectations(ZSTD_CCtx* cctx, const ZSTD_outBuffer* output, const ZSTD_inBuffer* input)
+{
+    DEBUGLOG(5, "ZSTD_setBufferExpectations (for advanced stable in/out modes)");
+    if (cctx->appliedParams.inBufferMode == ZSTD_bm_stable) {
+        cctx->expectedInBuffer = *input;
+    }
+    if (cctx->appliedParams.outBufferMode == ZSTD_bm_stable) {
+        cctx->expectedOutBufferSize = output->size - output->pos;
+    }
+}
+
+/* Validate that the input/output buffers match the expectations set by
+ * ZSTD_setBufferExpectations.
+ */
+static size_t ZSTD_checkBufferStability(ZSTD_CCtx const* cctx,
+                                        ZSTD_outBuffer const* output,
+                                        ZSTD_inBuffer const* input,
+                                        ZSTD_EndDirective endOp)
+{
+    if (cctx->appliedParams.inBufferMode == ZSTD_bm_stable) {
+        ZSTD_inBuffer const expect = cctx->expectedInBuffer;
+        if (expect.src != input->src || expect.pos != input->pos)
+            RETURN_ERROR(stabilityCondition_notRespected, "ZSTD_c_stableInBuffer enabled but input differs!");
+    }
+    (void)endOp;
+    if (cctx->appliedParams.outBufferMode == ZSTD_bm_stable) {
+        size_t const outBufferSize = output->size - output->pos;
+        if (cctx->expectedOutBufferSize != outBufferSize)
+            RETURN_ERROR(stabilityCondition_notRespected, "ZSTD_c_stableOutBuffer enabled but output size differs!");
+    }
+    return 0;
+}
+
+static size_t ZSTD_CCtx_init_compressStream2(ZSTD_CCtx* cctx,
+                                             ZSTD_EndDirective endOp,
+                                             size_t inSize)
+{
+    ZSTD_CCtx_params params = cctx->requestedParams;
+    ZSTD_prefixDict const prefixDict = cctx->prefixDict;
+    FORWARD_IF_ERROR( ZSTD_initLocalDict(cctx) , ""); /* Init the local dict if present. */
+    ZSTD_memset(&cctx->prefixDict, 0, sizeof(cctx->prefixDict));   /* single usage */
+    assert(prefixDict.dict==NULL || cctx->cdict==NULL);    /* only one can be set */
+    if (cctx->cdict && !cctx->localDict.cdict) {
+        /* Let the cdict's compression level take priority over the requested params.
+         * But do not take the cdict's compression level if the "cdict" is actually a localDict
+         * generated from ZSTD_initLocalDict().
+         */
+        params.compressionLevel = cctx->cdict->compressionLevel;
+    }
+    DEBUGLOG(4, "ZSTD_compressStream2 : transparent init stage");
+    if (endOp == ZSTD_e_end) cctx->pledgedSrcSizePlusOne = inSize + 1;  /* auto-determine pledgedSrcSize */
+
+    {   size_t const dictSize = prefixDict.dict
+                ? prefixDict.dictSize
+                : (cctx->cdict ? cctx->cdict->dictContentSize : 0);
+        ZSTD_cParamMode_e const mode = ZSTD_getCParamMode(cctx->cdict, &params, cctx->pledgedSrcSizePlusOne - 1);
+        params.cParams = ZSTD_getCParamsFromCCtxParams(
+                &params, cctx->pledgedSrcSizePlusOne-1,
+                dictSize, mode);
+    }
+
+    params.useBlockSplitter = ZSTD_resolveBlockSplitterMode(params.useBlockSplitter, &params.cParams);
+    params.ldmParams.enableLdm = ZSTD_resolveEnableLdm(params.ldmParams.enableLdm, &params.cParams);
+    params.useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(params.useRowMatchFinder, &params.cParams);
+    params.validateSequences = ZSTD_resolveExternalSequenceValidation(params.validateSequences);
+    params.maxBlockSize = ZSTD_resolveMaxBlockSize(params.maxBlockSize);
+    params.searchForExternalRepcodes = ZSTD_resolveExternalRepcodeSearch(params.searchForExternalRepcodes, params.compressionLevel);
+
+#ifdef ZSTD_MULTITHREAD
+    /* If external matchfinder is enabled, make sure to fail before checking job size (for consistency) */
+    RETURN_ERROR_IF(
+        params.useSequenceProducer == 1 && params.nbWorkers >= 1,
+        parameter_combination_unsupported,
+        "External sequence producer isn't supported with nbWorkers >= 1"
+    );
+
+    if ((cctx->pledgedSrcSizePlusOne-1) <= ZSTDMT_JOBSIZE_MIN) {
+        params.nbWorkers = 0; /* do not invoke multi-threading when src size is too small */
+    }
+    if (params.nbWorkers > 0) {
+#if ZSTD_TRACE
+        cctx->traceCtx = (ZSTD_trace_compress_begin != NULL) ? ZSTD_trace_compress_begin(cctx) : 0;
+#endif
+        /* mt context creation */
+        if (cctx->mtctx == NULL) {
+            DEBUGLOG(4, "ZSTD_compressStream2: creating new mtctx for nbWorkers=%u",
+                        params.nbWorkers);
+            cctx->mtctx = ZSTDMT_createCCtx_advanced((U32)params.nbWorkers, cctx->customMem, cctx->pool);
+            RETURN_ERROR_IF(cctx->mtctx == NULL, memory_allocation, "NULL pointer!");
+        }
+        /* mt compression */
+        DEBUGLOG(4, "call ZSTDMT_initCStream_internal as nbWorkers=%u", params.nbWorkers);
+        FORWARD_IF_ERROR( ZSTDMT_initCStream_internal(
+                    cctx->mtctx,
+                    prefixDict.dict, prefixDict.dictSize, prefixDict.dictContentType,
+                    cctx->cdict, params, cctx->pledgedSrcSizePlusOne-1) , "");
+        cctx->dictID = cctx->cdict ? cctx->cdict->dictID : 0;
+        cctx->dictContentSize = cctx->cdict ? cctx->cdict->dictContentSize : prefixDict.dictSize;
+        cctx->consumedSrcSize = 0;
+        cctx->producedCSize = 0;
+        cctx->streamStage = zcss_load;
+        cctx->appliedParams = params;
+    } else
+#endif  /* ZSTD_MULTITHREAD */
+    {   U64 const pledgedSrcSize = cctx->pledgedSrcSizePlusOne - 1;
+        assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams)));
+        FORWARD_IF_ERROR( ZSTD_compressBegin_internal(cctx,
+                prefixDict.dict, prefixDict.dictSize, prefixDict.dictContentType, ZSTD_dtlm_fast,
+                cctx->cdict,
+                &params, pledgedSrcSize,
+                ZSTDb_buffered) , "");
+        assert(cctx->appliedParams.nbWorkers == 0);
+        cctx->inToCompress = 0;
+        cctx->inBuffPos = 0;
+        if (cctx->appliedParams.inBufferMode == ZSTD_bm_buffered) {
+            /* for small input: avoid automatic flush on reaching end of block, since
+            * it would require to add a 3-bytes null block to end frame
+            */
+            cctx->inBuffTarget = cctx->blockSize + (cctx->blockSize == pledgedSrcSize);
+        } else {
+            cctx->inBuffTarget = 0;
+        }
+        cctx->outBuffContentSize = cctx->outBuffFlushedSize = 0;
+        cctx->streamStage = zcss_load;
+        cctx->frameEnded = 0;
+    }
+    return 0;
+}
 
+/* @return provides a minimum amount of data remaining to be flushed from internal buffers
+ */
 size_t ZSTD_compressStream2( ZSTD_CCtx* cctx,
                              ZSTD_outBuffer* output,
                              ZSTD_inBuffer* input,
@@ -3967,82 +6100,95 @@ size_t ZSTD_compressStream2( ZSTD_CCtx* cctx,
 {
     DEBUGLOG(5, "ZSTD_compressStream2, endOp=%u ", (unsigned)endOp);
     /* check conditions */
-    RETURN_ERROR_IF(output->pos > output->size, GENERIC, "invalid buffer");
-    RETURN_ERROR_IF(input->pos  > input->size, GENERIC, "invalid buffer");
-    assert(cctx!=NULL);
+    RETURN_ERROR_IF(output->pos > output->size, dstSize_tooSmall, "invalid output buffer");
+    RETURN_ERROR_IF(input->pos  > input->size, srcSize_wrong, "invalid input buffer");
+    RETURN_ERROR_IF((U32)endOp > (U32)ZSTD_e_end, parameter_outOfBound, "invalid endDirective");
+    assert(cctx != NULL);
 
     /* transparent initialization stage */
     if (cctx->streamStage == zcss_init) {
-        ZSTD_CCtx_params params = cctx->requestedParams;
-        ZSTD_prefixDict const prefixDict = cctx->prefixDict;
-        FORWARD_IF_ERROR( ZSTD_initLocalDict(cctx) , ""); /* Init the local dict if present. */
-        memset(&cctx->prefixDict, 0, sizeof(cctx->prefixDict));   /* single usage */
-        assert(prefixDict.dict==NULL || cctx->cdict==NULL);    /* only one can be set */
-        DEBUGLOG(4, "ZSTD_compressStream2 : transparent init stage");
-        if (endOp == ZSTD_e_end) cctx->pledgedSrcSizePlusOne = input->size + 1;  /* auto-fix pledgedSrcSize */
-        params.cParams = ZSTD_getCParamsFromCCtxParams(
-                &cctx->requestedParams, cctx->pledgedSrcSizePlusOne-1, 0 /*dictSize*/);
-
-
-#ifdef ZSTD_MULTITHREAD
-        if ((cctx->pledgedSrcSizePlusOne-1) <= ZSTDMT_JOBSIZE_MIN) {
-            params.nbWorkers = 0; /* do not invoke multi-threading when src size is too small */
-        }
-        if (params.nbWorkers > 0) {
-            /* mt context creation */
-            if (cctx->mtctx == NULL) {
-                DEBUGLOG(4, "ZSTD_compressStream2: creating new mtctx for nbWorkers=%u",
-                            params.nbWorkers);
-                cctx->mtctx = ZSTDMT_createCCtx_advanced((U32)params.nbWorkers, cctx->customMem);
-                RETURN_ERROR_IF(cctx->mtctx == NULL, memory_allocation, "NULL pointer!");
+        size_t const inputSize = input->size - input->pos;  /* no obligation to start from pos==0 */
+        size_t const totalInputSize = inputSize + cctx->stableIn_notConsumed;
+        if ( (cctx->requestedParams.inBufferMode == ZSTD_bm_stable) /* input is presumed stable, across invocations */
+          && (endOp == ZSTD_e_continue)                             /* no flush requested, more input to come */
+          && (totalInputSize < ZSTD_BLOCKSIZE_MAX) ) {              /* not even reached one block yet */
+            if (cctx->stableIn_notConsumed) {  /* not the first time */
+                /* check stable source guarantees */
+                RETURN_ERROR_IF(input->src != cctx->expectedInBuffer.src, stabilityCondition_notRespected, "stableInBuffer condition not respected: wrong src pointer");
+                RETURN_ERROR_IF(input->pos != cctx->expectedInBuffer.size, stabilityCondition_notRespected, "stableInBuffer condition not respected: externally modified pos");
             }
-            /* mt compression */
-            DEBUGLOG(4, "call ZSTDMT_initCStream_internal as nbWorkers=%u", params.nbWorkers);
-            FORWARD_IF_ERROR( ZSTDMT_initCStream_internal(
-                        cctx->mtctx,
-                        prefixDict.dict, prefixDict.dictSize, prefixDict.dictContentType,
-                        cctx->cdict, params, cctx->pledgedSrcSizePlusOne-1) , "");
-            cctx->streamStage = zcss_load;
-            cctx->appliedParams.nbWorkers = params.nbWorkers;
-        } else
-#endif
-        {   FORWARD_IF_ERROR( ZSTD_resetCStream_internal(cctx,
-                            prefixDict.dict, prefixDict.dictSize, prefixDict.dictContentType,
-                            cctx->cdict,
-                            params, cctx->pledgedSrcSizePlusOne-1) , "");
-            assert(cctx->streamStage == zcss_load);
-            assert(cctx->appliedParams.nbWorkers == 0);
-    }   }
+            /* pretend input was consumed, to give a sense forward progress */
+            input->pos = input->size;
+            /* save stable inBuffer, for later control, and flush/end */
+            cctx->expectedInBuffer = *input;
+            /* but actually input wasn't consumed, so keep track of position from where compression shall resume */
+            cctx->stableIn_notConsumed += inputSize;
+            /* don't initialize yet, wait for the first block of flush() order, for better parameters adaptation */
+            return ZSTD_FRAMEHEADERSIZE_MIN(cctx->requestedParams.format);  /* at least some header to produce */
+        }
+        FORWARD_IF_ERROR(ZSTD_CCtx_init_compressStream2(cctx, endOp, totalInputSize), "compressStream2 initialization failed");
+        ZSTD_setBufferExpectations(cctx, output, input);   /* Set initial buffer expectations now that we've initialized */
+    }
     /* end of transparent initialization stage */
 
+    FORWARD_IF_ERROR(ZSTD_checkBufferStability(cctx, output, input, endOp), "invalid buffers");
     /* compression stage */
 #ifdef ZSTD_MULTITHREAD
     if (cctx->appliedParams.nbWorkers > 0) {
-        int const forceMaxProgress = (endOp == ZSTD_e_flush || endOp == ZSTD_e_end);
         size_t flushMin;
-        assert(forceMaxProgress || endOp == ZSTD_e_continue /* Protection for a new flush type */);
         if (cctx->cParamsChanged) {
             ZSTDMT_updateCParams_whileCompressing(cctx->mtctx, &cctx->requestedParams);
             cctx->cParamsChanged = 0;
         }
-        do {
+        if (cctx->stableIn_notConsumed) {
+            assert(cctx->appliedParams.inBufferMode == ZSTD_bm_stable);
+            /* some early data was skipped - make it available for consumption */
+            assert(input->pos >= cctx->stableIn_notConsumed);
+            input->pos -= cctx->stableIn_notConsumed;
+            cctx->stableIn_notConsumed = 0;
+        }
+        for (;;) {
+            size_t const ipos = input->pos;
+            size_t const opos = output->pos;
             flushMin = ZSTDMT_compressStream_generic(cctx->mtctx, output, input, endOp);
+            cctx->consumedSrcSize += (U64)(input->pos - ipos);
+            cctx->producedCSize += (U64)(output->pos - opos);
             if ( ZSTD_isError(flushMin)
               || (endOp == ZSTD_e_end && flushMin == 0) ) { /* compression completed */
+                if (flushMin == 0)
+                    ZSTD_CCtx_trace(cctx, 0);
                 ZSTD_CCtx_reset(cctx, ZSTD_reset_session_only);
             }
             FORWARD_IF_ERROR(flushMin, "ZSTDMT_compressStream_generic failed");
-        } while (forceMaxProgress && flushMin != 0 && output->pos < output->size);
+
+            if (endOp == ZSTD_e_continue) {
+                /* We only require some progress with ZSTD_e_continue, not maximal progress.
+                 * We're done if we've consumed or produced any bytes, or either buffer is
+                 * full.
+                 */
+                if (input->pos != ipos || output->pos != opos || input->pos == input->size || output->pos == output->size)
+                    break;
+            } else {
+                assert(endOp == ZSTD_e_flush || endOp == ZSTD_e_end);
+                /* We require maximal progress. We're done when the flush is complete or the
+                 * output buffer is full.
+                 */
+                if (flushMin == 0 || output->pos == output->size)
+                    break;
+            }
+        }
         DEBUGLOG(5, "completed ZSTD_compressStream2 delegating to ZSTDMT_compressStream_generic");
         /* Either we don't require maximum forward progress, we've finished the
          * flush, or we are out of output space.
          */
-        assert(!forceMaxProgress || flushMin == 0 || output->pos == output->size);
+        assert(endOp == ZSTD_e_continue || flushMin == 0 || output->pos == output->size);
+        ZSTD_setBufferExpectations(cctx, output, input);
         return flushMin;
     }
-#endif
+#endif /* ZSTD_MULTITHREAD */
     FORWARD_IF_ERROR( ZSTD_compressStream_generic(cctx, output, input, endOp) , "");
     DEBUGLOG(5, "completed ZSTD_compressStream2");
+    ZSTD_setBufferExpectations(cctx, output, input);
     return cctx->outBuffContentSize - cctx->outBuffFlushedSize; /* remaining to flush */
 }
 
@@ -4052,27 +6198,43 @@ size_t ZSTD_compressStream2_simpleArgs (
                       const void* src, size_t srcSize, size_t* srcPos,
                             ZSTD_EndDirective endOp)
 {
-    ZSTD_outBuffer output = { dst, dstCapacity, *dstPos };
-    ZSTD_inBuffer  input  = { src, srcSize, *srcPos };
+    ZSTD_outBuffer output;
+    ZSTD_inBuffer  input;
+    output.dst = dst;
+    output.size = dstCapacity;
+    output.pos = *dstPos;
+    input.src = src;
+    input.size = srcSize;
+    input.pos = *srcPos;
     /* ZSTD_compressStream2() will check validity of dstPos and srcPos */
-    size_t const cErr = ZSTD_compressStream2(cctx, &output, &input, endOp);
-    *dstPos = output.pos;
-    *srcPos = input.pos;
-    return cErr;
+    {   size_t const cErr = ZSTD_compressStream2(cctx, &output, &input, endOp);
+        *dstPos = output.pos;
+        *srcPos = input.pos;
+        return cErr;
+    }
 }
 
 size_t ZSTD_compress2(ZSTD_CCtx* cctx,
                       void* dst, size_t dstCapacity,
                       const void* src, size_t srcSize)
 {
+    ZSTD_bufferMode_e const originalInBufferMode = cctx->requestedParams.inBufferMode;
+    ZSTD_bufferMode_e const originalOutBufferMode = cctx->requestedParams.outBufferMode;
     DEBUGLOG(4, "ZSTD_compress2 (srcSize=%u)", (unsigned)srcSize);
     ZSTD_CCtx_reset(cctx, ZSTD_reset_session_only);
+    /* Enable stable input/output buffers. */
+    cctx->requestedParams.inBufferMode = ZSTD_bm_stable;
+    cctx->requestedParams.outBufferMode = ZSTD_bm_stable;
     {   size_t oPos = 0;
         size_t iPos = 0;
         size_t const result = ZSTD_compressStream2_simpleArgs(cctx,
                                         dst, dstCapacity, &oPos,
                                         src, srcSize, &iPos,
                                         ZSTD_e_end);
+        /* Reset to the original values. */
+        cctx->requestedParams.inBufferMode = originalInBufferMode;
+        cctx->requestedParams.outBufferMode = originalOutBufferMode;
+
         FORWARD_IF_ERROR(result, "ZSTD_compressStream2_simpleArgs failed");
         if (result != 0) {  /* compression not completed, due to lack of output space */
             assert(oPos == dstCapacity);
@@ -4083,22 +6245,512 @@ size_t ZSTD_compress2(ZSTD_CCtx* cctx,
     }
 }
 
+/* ZSTD_validateSequence() :
+ * @offCode : is presumed to follow format required by ZSTD_storeSeq()
+ * @returns a ZSTD error code if sequence is not valid
+ */
+static size_t
+ZSTD_validateSequence(U32 offCode, U32 matchLength, U32 minMatch,
+                      size_t posInSrc, U32 windowLog, size_t dictSize, int useSequenceProducer)
+{
+    U32 const windowSize = 1u << windowLog;
+    /* posInSrc represents the amount of data the decoder would decode up to this point.
+     * As long as the amount of data decoded is less than or equal to window size, offsets may be
+     * larger than the total length of output decoded in order to reference the dict, even larger than
+     * window size. After output surpasses windowSize, we're limited to windowSize offsets again.
+     */
+    size_t const offsetBound = posInSrc > windowSize ? (size_t)windowSize : posInSrc + (size_t)dictSize;
+    size_t const matchLenLowerBound = (minMatch == 3 || useSequenceProducer) ? 3 : 4;
+    RETURN_ERROR_IF(offCode > OFFSET_TO_OFFBASE(offsetBound), externalSequences_invalid, "Offset too large!");
+    /* Validate maxNbSeq is large enough for the given matchLength and minMatch */
+    RETURN_ERROR_IF(matchLength < matchLenLowerBound, externalSequences_invalid, "Matchlength too small for the minMatch");
+    return 0;
+}
+
+/* Returns an offset code, given a sequence's raw offset, the ongoing repcode array, and whether litLength == 0 */
+static U32 ZSTD_finalizeOffBase(U32 rawOffset, const U32 rep[ZSTD_REP_NUM], U32 ll0)
+{
+    U32 offBase = OFFSET_TO_OFFBASE(rawOffset);
+
+    if (!ll0 && rawOffset == rep[0]) {
+        offBase = REPCODE1_TO_OFFBASE;
+    } else if (rawOffset == rep[1]) {
+        offBase = REPCODE_TO_OFFBASE(2 - ll0);
+    } else if (rawOffset == rep[2]) {
+        offBase = REPCODE_TO_OFFBASE(3 - ll0);
+    } else if (ll0 && rawOffset == rep[0] - 1) {
+        offBase = REPCODE3_TO_OFFBASE;
+    }
+    return offBase;
+}
+
+size_t
+ZSTD_copySequencesToSeqStoreExplicitBlockDelim(ZSTD_CCtx* cctx,
+                                              ZSTD_sequencePosition* seqPos,
+                                        const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
+                                        const void* src, size_t blockSize,
+                                        ZSTD_paramSwitch_e externalRepSearch)
+{
+    U32 idx = seqPos->idx;
+    U32 const startIdx = idx;
+    BYTE const* ip = (BYTE const*)(src);
+    const BYTE* const iend = ip + blockSize;
+    repcodes_t updatedRepcodes;
+    U32 dictSize;
+
+    DEBUGLOG(5, "ZSTD_copySequencesToSeqStoreExplicitBlockDelim (blockSize = %zu)", blockSize);
+
+    if (cctx->cdict) {
+        dictSize = (U32)cctx->cdict->dictContentSize;
+    } else if (cctx->prefixDict.dict) {
+        dictSize = (U32)cctx->prefixDict.dictSize;
+    } else {
+        dictSize = 0;
+    }
+    ZSTD_memcpy(updatedRepcodes.rep, cctx->blockState.prevCBlock->rep, sizeof(repcodes_t));
+    for (; idx < inSeqsSize && (inSeqs[idx].matchLength != 0 || inSeqs[idx].offset != 0); ++idx) {
+        U32 const litLength = inSeqs[idx].litLength;
+        U32 const matchLength = inSeqs[idx].matchLength;
+        U32 offBase;
+
+        if (externalRepSearch == ZSTD_ps_disable) {
+            offBase = OFFSET_TO_OFFBASE(inSeqs[idx].offset);
+        } else {
+            U32 const ll0 = (litLength == 0);
+            offBase = ZSTD_finalizeOffBase(inSeqs[idx].offset, updatedRepcodes.rep, ll0);
+            ZSTD_updateRep(updatedRepcodes.rep, offBase, ll0);
+        }
+
+        DEBUGLOG(6, "Storing sequence: (of: %u, ml: %u, ll: %u)", offBase, matchLength, litLength);
+        if (cctx->appliedParams.validateSequences) {
+            seqPos->posInSrc += litLength + matchLength;
+            FORWARD_IF_ERROR(ZSTD_validateSequence(offBase, matchLength, cctx->appliedParams.cParams.minMatch, seqPos->posInSrc,
+                                                cctx->appliedParams.cParams.windowLog, dictSize, cctx->appliedParams.useSequenceProducer),
+                                                "Sequence validation failed");
+        }
+        RETURN_ERROR_IF(idx - seqPos->idx >= cctx->seqStore.maxNbSeq, externalSequences_invalid,
+                        "Not enough memory allocated. Try adjusting ZSTD_c_minMatch.");
+        ZSTD_storeSeq(&cctx->seqStore, litLength, ip, iend, offBase, matchLength);
+        ip += matchLength + litLength;
+    }
+
+    /* If we skipped repcode search while parsing, we need to update repcodes now */
+    assert(externalRepSearch != ZSTD_ps_auto);
+    assert(idx >= startIdx);
+    if (externalRepSearch == ZSTD_ps_disable && idx != startIdx) {
+        U32* const rep = updatedRepcodes.rep;
+        U32 lastSeqIdx = idx - 1; /* index of last non-block-delimiter sequence */
+
+        if (lastSeqIdx >= startIdx + 2) {
+            rep[2] = inSeqs[lastSeqIdx - 2].offset;
+            rep[1] = inSeqs[lastSeqIdx - 1].offset;
+            rep[0] = inSeqs[lastSeqIdx].offset;
+        } else if (lastSeqIdx == startIdx + 1) {
+            rep[2] = rep[0];
+            rep[1] = inSeqs[lastSeqIdx - 1].offset;
+            rep[0] = inSeqs[lastSeqIdx].offset;
+        } else {
+            assert(lastSeqIdx == startIdx);
+            rep[2] = rep[1];
+            rep[1] = rep[0];
+            rep[0] = inSeqs[lastSeqIdx].offset;
+        }
+    }
+
+    ZSTD_memcpy(cctx->blockState.nextCBlock->rep, updatedRepcodes.rep, sizeof(repcodes_t));
+
+    if (inSeqs[idx].litLength) {
+        DEBUGLOG(6, "Storing last literals of size: %u", inSeqs[idx].litLength);
+        ZSTD_storeLastLiterals(&cctx->seqStore, ip, inSeqs[idx].litLength);
+        ip += inSeqs[idx].litLength;
+        seqPos->posInSrc += inSeqs[idx].litLength;
+    }
+    RETURN_ERROR_IF(ip != iend, externalSequences_invalid, "Blocksize doesn't agree with block delimiter!");
+    seqPos->idx = idx+1;
+    return 0;
+}
+
+size_t
+ZSTD_copySequencesToSeqStoreNoBlockDelim(ZSTD_CCtx* cctx, ZSTD_sequencePosition* seqPos,
+                                   const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
+                                   const void* src, size_t blockSize, ZSTD_paramSwitch_e externalRepSearch)
+{
+    U32 idx = seqPos->idx;
+    U32 startPosInSequence = seqPos->posInSequence;
+    U32 endPosInSequence = seqPos->posInSequence + (U32)blockSize;
+    size_t dictSize;
+    BYTE const* ip = (BYTE const*)(src);
+    BYTE const* iend = ip + blockSize;  /* May be adjusted if we decide to process fewer than blockSize bytes */
+    repcodes_t updatedRepcodes;
+    U32 bytesAdjustment = 0;
+    U32 finalMatchSplit = 0;
+
+    /* TODO(embg) support fast parsing mode in noBlockDelim mode */
+    (void)externalRepSearch;
+
+    if (cctx->cdict) {
+        dictSize = cctx->cdict->dictContentSize;
+    } else if (cctx->prefixDict.dict) {
+        dictSize = cctx->prefixDict.dictSize;
+    } else {
+        dictSize = 0;
+    }
+    DEBUGLOG(5, "ZSTD_copySequencesToSeqStoreNoBlockDelim: idx: %u PIS: %u blockSize: %zu", idx, startPosInSequence, blockSize);
+    DEBUGLOG(5, "Start seq: idx: %u (of: %u ml: %u ll: %u)", idx, inSeqs[idx].offset, inSeqs[idx].matchLength, inSeqs[idx].litLength);
+    ZSTD_memcpy(updatedRepcodes.rep, cctx->blockState.prevCBlock->rep, sizeof(repcodes_t));
+    while (endPosInSequence && idx < inSeqsSize && !finalMatchSplit) {
+        const ZSTD_Sequence currSeq = inSeqs[idx];
+        U32 litLength = currSeq.litLength;
+        U32 matchLength = currSeq.matchLength;
+        U32 const rawOffset = currSeq.offset;
+        U32 offBase;
+
+        /* Modify the sequence depending on where endPosInSequence lies */
+        if (endPosInSequence >= currSeq.litLength + currSeq.matchLength) {
+            if (startPosInSequence >= litLength) {
+                startPosInSequence -= litLength;
+                litLength = 0;
+                matchLength -= startPosInSequence;
+            } else {
+                litLength -= startPosInSequence;
+            }
+            /* Move to the next sequence */
+            endPosInSequence -= currSeq.litLength + currSeq.matchLength;
+            startPosInSequence = 0;
+        } else {
+            /* This is the final (partial) sequence we're adding from inSeqs, and endPosInSequence
+               does not reach the end of the match. So, we have to split the sequence */
+            DEBUGLOG(6, "Require a split: diff: %u, idx: %u PIS: %u",
+                     currSeq.litLength + currSeq.matchLength - endPosInSequence, idx, endPosInSequence);
+            if (endPosInSequence > litLength) {
+                U32 firstHalfMatchLength;
+                litLength = startPosInSequence >= litLength ? 0 : litLength - startPosInSequence;
+                firstHalfMatchLength = endPosInSequence - startPosInSequence - litLength;
+                if (matchLength > blockSize && firstHalfMatchLength >= cctx->appliedParams.cParams.minMatch) {
+                    /* Only ever split the match if it is larger than the block size */
+                    U32 secondHalfMatchLength = currSeq.matchLength + currSeq.litLength - endPosInSequence;
+                    if (secondHalfMatchLength < cctx->appliedParams.cParams.minMatch) {
+                        /* Move the endPosInSequence backward so that it creates match of minMatch length */
+                        endPosInSequence -= cctx->appliedParams.cParams.minMatch - secondHalfMatchLength;
+                        bytesAdjustment = cctx->appliedParams.cParams.minMatch - secondHalfMatchLength;
+                        firstHalfMatchLength -= bytesAdjustment;
+                    }
+                    matchLength = firstHalfMatchLength;
+                    /* Flag that we split the last match - after storing the sequence, exit the loop,
+                       but keep the value of endPosInSequence */
+                    finalMatchSplit = 1;
+                } else {
+                    /* Move the position in sequence backwards so that we don't split match, and break to store
+                     * the last literals. We use the original currSeq.litLength as a marker for where endPosInSequence
+                     * should go. We prefer to do this whenever it is not necessary to split the match, or if doing so
+                     * would cause the first half of the match to be too small
+                     */
+                    bytesAdjustment = endPosInSequence - currSeq.litLength;
+                    endPosInSequence = currSeq.litLength;
+                    break;
+                }
+            } else {
+                /* This sequence ends inside the literals, break to store the last literals */
+                break;
+            }
+        }
+        /* Check if this offset can be represented with a repcode */
+        {   U32 const ll0 = (litLength == 0);
+            offBase = ZSTD_finalizeOffBase(rawOffset, updatedRepcodes.rep, ll0);
+            ZSTD_updateRep(updatedRepcodes.rep, offBase, ll0);
+        }
+
+        if (cctx->appliedParams.validateSequences) {
+            seqPos->posInSrc += litLength + matchLength;
+            FORWARD_IF_ERROR(ZSTD_validateSequence(offBase, matchLength, cctx->appliedParams.cParams.minMatch, seqPos->posInSrc,
+                                                   cctx->appliedParams.cParams.windowLog, dictSize, cctx->appliedParams.useSequenceProducer),
+                                                   "Sequence validation failed");
+        }
+        DEBUGLOG(6, "Storing sequence: (of: %u, ml: %u, ll: %u)", offBase, matchLength, litLength);
+        RETURN_ERROR_IF(idx - seqPos->idx >= cctx->seqStore.maxNbSeq, externalSequences_invalid,
+                        "Not enough memory allocated. Try adjusting ZSTD_c_minMatch.");
+        ZSTD_storeSeq(&cctx->seqStore, litLength, ip, iend, offBase, matchLength);
+        ip += matchLength + litLength;
+        if (!finalMatchSplit)
+            idx++; /* Next Sequence */
+    }
+    DEBUGLOG(5, "Ending seq: idx: %u (of: %u ml: %u ll: %u)", idx, inSeqs[idx].offset, inSeqs[idx].matchLength, inSeqs[idx].litLength);
+    assert(idx == inSeqsSize || endPosInSequence <= inSeqs[idx].litLength + inSeqs[idx].matchLength);
+    seqPos->idx = idx;
+    seqPos->posInSequence = endPosInSequence;
+    ZSTD_memcpy(cctx->blockState.nextCBlock->rep, updatedRepcodes.rep, sizeof(repcodes_t));
+
+    iend -= bytesAdjustment;
+    if (ip != iend) {
+        /* Store any last literals */
+        U32 lastLLSize = (U32)(iend - ip);
+        assert(ip <= iend);
+        DEBUGLOG(6, "Storing last literals of size: %u", lastLLSize);
+        ZSTD_storeLastLiterals(&cctx->seqStore, ip, lastLLSize);
+        seqPos->posInSrc += lastLLSize;
+    }
+
+    return bytesAdjustment;
+}
+
+typedef size_t (*ZSTD_sequenceCopier) (ZSTD_CCtx* cctx, ZSTD_sequencePosition* seqPos,
+                                       const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
+                                       const void* src, size_t blockSize, ZSTD_paramSwitch_e externalRepSearch);
+static ZSTD_sequenceCopier ZSTD_selectSequenceCopier(ZSTD_sequenceFormat_e mode)
+{
+    ZSTD_sequenceCopier sequenceCopier = NULL;
+    assert(ZSTD_cParam_withinBounds(ZSTD_c_blockDelimiters, mode));
+    if (mode == ZSTD_sf_explicitBlockDelimiters) {
+        return ZSTD_copySequencesToSeqStoreExplicitBlockDelim;
+    } else if (mode == ZSTD_sf_noBlockDelimiters) {
+        return ZSTD_copySequencesToSeqStoreNoBlockDelim;
+    }
+    assert(sequenceCopier != NULL);
+    return sequenceCopier;
+}
+
+/* Discover the size of next block by searching for the delimiter.
+ * Note that a block delimiter **must** exist in this mode,
+ * otherwise it's an input error.
+ * The block size retrieved will be later compared to ensure it remains within bounds */
+static size_t
+blockSize_explicitDelimiter(const ZSTD_Sequence* inSeqs, size_t inSeqsSize, ZSTD_sequencePosition seqPos)
+{
+    int end = 0;
+    size_t blockSize = 0;
+    size_t spos = seqPos.idx;
+    DEBUGLOG(6, "blockSize_explicitDelimiter : seq %zu / %zu", spos, inSeqsSize);
+    assert(spos <= inSeqsSize);
+    while (spos < inSeqsSize) {
+        end = (inSeqs[spos].offset == 0);
+        blockSize += inSeqs[spos].litLength + inSeqs[spos].matchLength;
+        if (end) {
+            if (inSeqs[spos].matchLength != 0)
+                RETURN_ERROR(externalSequences_invalid, "delimiter format error : both matchlength and offset must be == 0");
+            break;
+        }
+        spos++;
+    }
+    if (!end)
+        RETURN_ERROR(externalSequences_invalid, "Reached end of sequences without finding a block delimiter");
+    return blockSize;
+}
+
+/* More a "target" block size */
+static size_t blockSize_noDelimiter(size_t blockSize, size_t remaining)
+{
+    int const lastBlock = (remaining <= blockSize);
+    return lastBlock ? remaining : blockSize;
+}
+
+static size_t determine_blockSize(ZSTD_sequenceFormat_e mode,
+                           size_t blockSize, size_t remaining,
+                     const ZSTD_Sequence* inSeqs, size_t inSeqsSize, ZSTD_sequencePosition seqPos)
+{
+    DEBUGLOG(6, "determine_blockSize : remainingSize = %zu", remaining);
+    if (mode == ZSTD_sf_noBlockDelimiters)
+        return blockSize_noDelimiter(blockSize, remaining);
+    {   size_t const explicitBlockSize = blockSize_explicitDelimiter(inSeqs, inSeqsSize, seqPos);
+        FORWARD_IF_ERROR(explicitBlockSize, "Error while determining block size with explicit delimiters");
+        if (explicitBlockSize > blockSize)
+            RETURN_ERROR(externalSequences_invalid, "sequences incorrectly define a too large block");
+        if (explicitBlockSize > remaining)
+            RETURN_ERROR(externalSequences_invalid, "sequences define a frame longer than source");
+        return explicitBlockSize;
+    }
+}
+
+/* Compress, block-by-block, all of the sequences given.
+ *
+ * Returns the cumulative size of all compressed blocks (including their headers),
+ * otherwise a ZSTD error.
+ */
+static size_t
+ZSTD_compressSequences_internal(ZSTD_CCtx* cctx,
+                                void* dst, size_t dstCapacity,
+                          const ZSTD_Sequence* inSeqs, size_t inSeqsSize,
+                          const void* src, size_t srcSize)
+{
+    size_t cSize = 0;
+    size_t remaining = srcSize;
+    ZSTD_sequencePosition seqPos = {0, 0, 0};
+
+    BYTE const* ip = (BYTE const*)src;
+    BYTE* op = (BYTE*)dst;
+    ZSTD_sequenceCopier const sequenceCopier = ZSTD_selectSequenceCopier(cctx->appliedParams.blockDelimiters);
+
+    DEBUGLOG(4, "ZSTD_compressSequences_internal srcSize: %zu, inSeqsSize: %zu", srcSize, inSeqsSize);
+    /* Special case: empty frame */
+    if (remaining == 0) {
+        U32 const cBlockHeader24 = 1 /* last block */ + (((U32)bt_raw)<<1);
+        RETURN_ERROR_IF(dstCapacity<4, dstSize_tooSmall, "No room for empty frame block header");
+        MEM_writeLE32(op, cBlockHeader24);
+        op += ZSTD_blockHeaderSize;
+        dstCapacity -= ZSTD_blockHeaderSize;
+        cSize += ZSTD_blockHeaderSize;
+    }
+
+    while (remaining) {
+        size_t compressedSeqsSize;
+        size_t cBlockSize;
+        size_t additionalByteAdjustment;
+        size_t blockSize = determine_blockSize(cctx->appliedParams.blockDelimiters,
+                                        cctx->blockSize, remaining,
+                                        inSeqs, inSeqsSize, seqPos);
+        U32 const lastBlock = (blockSize == remaining);
+        FORWARD_IF_ERROR(blockSize, "Error while trying to determine block size");
+        assert(blockSize <= remaining);
+        ZSTD_resetSeqStore(&cctx->seqStore);
+        DEBUGLOG(5, "Working on new block. Blocksize: %zu (total:%zu)", blockSize, (ip - (const BYTE*)src) + blockSize);
+
+        additionalByteAdjustment = sequenceCopier(cctx, &seqPos, inSeqs, inSeqsSize, ip, blockSize, cctx->appliedParams.searchForExternalRepcodes);
+        FORWARD_IF_ERROR(additionalByteAdjustment, "Bad sequence copy");
+        blockSize -= additionalByteAdjustment;
+
+        /* If blocks are too small, emit as a nocompress block */
+        /* TODO: See 3090. We reduced MIN_CBLOCK_SIZE from 3 to 2 so to compensate we are adding
+         * additional 1. We need to revisit and change this logic to be more consistent */
+        if (blockSize < MIN_CBLOCK_SIZE+ZSTD_blockHeaderSize+1+1) {
+            cBlockSize = ZSTD_noCompressBlock(op, dstCapacity, ip, blockSize, lastBlock);
+            FORWARD_IF_ERROR(cBlockSize, "Nocompress block failed");
+            DEBUGLOG(5, "Block too small, writing out nocompress block: cSize: %zu", cBlockSize);
+            cSize += cBlockSize;
+            ip += blockSize;
+            op += cBlockSize;
+            remaining -= blockSize;
+            dstCapacity -= cBlockSize;
+            continue;
+        }
+
+        RETURN_ERROR_IF(dstCapacity < ZSTD_blockHeaderSize, dstSize_tooSmall, "not enough dstCapacity to write a new compressed block");
+        compressedSeqsSize = ZSTD_entropyCompressSeqStore(&cctx->seqStore,
+                                &cctx->blockState.prevCBlock->entropy, &cctx->blockState.nextCBlock->entropy,
+                                &cctx->appliedParams,
+                                op + ZSTD_blockHeaderSize /* Leave space for block header */, dstCapacity - ZSTD_blockHeaderSize,
+                                blockSize,
+                                cctx->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */,
+                                cctx->bmi2);
+        FORWARD_IF_ERROR(compressedSeqsSize, "Compressing sequences of block failed");
+        DEBUGLOG(5, "Compressed sequences size: %zu", compressedSeqsSize);
+
+        if (!cctx->isFirstBlock &&
+            ZSTD_maybeRLE(&cctx->seqStore) &&
+            ZSTD_isRLE(ip, blockSize)) {
+            /* We don't want to emit our first block as a RLE even if it qualifies because
+            * doing so will cause the decoder (cli only) to throw a "should consume all input error."
+            * This is only an issue for zstd <= v1.4.3
+            */
+            compressedSeqsSize = 1;
+        }
+
+        if (compressedSeqsSize == 0) {
+            /* ZSTD_noCompressBlock writes the block header as well */
+            cBlockSize = ZSTD_noCompressBlock(op, dstCapacity, ip, blockSize, lastBlock);
+            FORWARD_IF_ERROR(cBlockSize, "ZSTD_noCompressBlock failed");
+            DEBUGLOG(5, "Writing out nocompress block, size: %zu", cBlockSize);
+        } else if (compressedSeqsSize == 1) {
+            cBlockSize = ZSTD_rleCompressBlock(op, dstCapacity, *ip, blockSize, lastBlock);
+            FORWARD_IF_ERROR(cBlockSize, "ZSTD_rleCompressBlock failed");
+            DEBUGLOG(5, "Writing out RLE block, size: %zu", cBlockSize);
+        } else {
+            U32 cBlockHeader;
+            /* Error checking and repcodes update */
+            ZSTD_blockState_confirmRepcodesAndEntropyTables(&cctx->blockState);
+            if (cctx->blockState.prevCBlock->entropy.fse.offcode_repeatMode == FSE_repeat_valid)
+                cctx->blockState.prevCBlock->entropy.fse.offcode_repeatMode = FSE_repeat_check;
+
+            /* Write block header into beginning of block*/
+            cBlockHeader = lastBlock + (((U32)bt_compressed)<<1) + (U32)(compressedSeqsSize << 3);
+            MEM_writeLE24(op, cBlockHeader);
+            cBlockSize = ZSTD_blockHeaderSize + compressedSeqsSize;
+            DEBUGLOG(5, "Writing out compressed block, size: %zu", cBlockSize);
+        }
+
+        cSize += cBlockSize;
+
+        if (lastBlock) {
+            break;
+        } else {
+            ip += blockSize;
+            op += cBlockSize;
+            remaining -= blockSize;
+            dstCapacity -= cBlockSize;
+            cctx->isFirstBlock = 0;
+        }
+        DEBUGLOG(5, "cSize running total: %zu (remaining dstCapacity=%zu)", cSize, dstCapacity);
+    }
+
+    DEBUGLOG(4, "cSize final total: %zu", cSize);
+    return cSize;
+}
+
+size_t ZSTD_compressSequences(ZSTD_CCtx* cctx,
+                              void* dst, size_t dstCapacity,
+                              const ZSTD_Sequence* inSeqs, size_t inSeqsSize,
+                              const void* src, size_t srcSize)
+{
+    BYTE* op = (BYTE*)dst;
+    size_t cSize = 0;
+    size_t compressedBlocksSize = 0;
+    size_t frameHeaderSize = 0;
+
+    /* Transparent initialization stage, same as compressStream2() */
+    DEBUGLOG(4, "ZSTD_compressSequences (dstCapacity=%zu)", dstCapacity);
+    assert(cctx != NULL);
+    FORWARD_IF_ERROR(ZSTD_CCtx_init_compressStream2(cctx, ZSTD_e_end, srcSize), "CCtx initialization failed");
+    /* Begin writing output, starting with frame header */
+    frameHeaderSize = ZSTD_writeFrameHeader(op, dstCapacity, &cctx->appliedParams, srcSize, cctx->dictID);
+    op += frameHeaderSize;
+    dstCapacity -= frameHeaderSize;
+    cSize += frameHeaderSize;
+    if (cctx->appliedParams.fParams.checksumFlag && srcSize) {
+        XXH64_update(&cctx->xxhState, src, srcSize);
+    }
+    /* cSize includes block header size and compressed sequences size */
+    compressedBlocksSize = ZSTD_compressSequences_internal(cctx,
+                                                           op, dstCapacity,
+                                                           inSeqs, inSeqsSize,
+                                                           src, srcSize);
+    FORWARD_IF_ERROR(compressedBlocksSize, "Compressing blocks failed!");
+    cSize += compressedBlocksSize;
+    dstCapacity -= compressedBlocksSize;
+
+    if (cctx->appliedParams.fParams.checksumFlag) {
+        U32 const checksum = (U32) XXH64_digest(&cctx->xxhState);
+        RETURN_ERROR_IF(dstCapacity<4, dstSize_tooSmall, "no room for checksum");
+        DEBUGLOG(4, "Write checksum : %08X", (unsigned)checksum);
+        MEM_writeLE32((char*)dst + cSize, checksum);
+        cSize += 4;
+    }
+
+    DEBUGLOG(4, "Final compressed size: %zu", cSize);
+    return cSize;
+}
+
 /*======   Finalize   ======*/
 
+static ZSTD_inBuffer inBuffer_forEndFlush(const ZSTD_CStream* zcs)
+{
+    const ZSTD_inBuffer nullInput = { NULL, 0, 0 };
+    const int stableInput = (zcs->appliedParams.inBufferMode == ZSTD_bm_stable);
+    return stableInput ? zcs->expectedInBuffer : nullInput;
+}
+
 /*! ZSTD_flushStream() :
  * @return : amount of data remaining to flush */
 size_t ZSTD_flushStream(ZSTD_CStream* zcs, ZSTD_outBuffer* output)
 {
-    ZSTD_inBuffer input = { NULL, 0, 0 };
+    ZSTD_inBuffer input = inBuffer_forEndFlush(zcs);
+    input.size = input.pos; /* do not ingest more input during flush */
     return ZSTD_compressStream2(zcs, output, &input, ZSTD_e_flush);
 }
 
 
 size_t ZSTD_endStream(ZSTD_CStream* zcs, ZSTD_outBuffer* output)
 {
-    ZSTD_inBuffer input = { NULL, 0, 0 };
+    ZSTD_inBuffer input = inBuffer_forEndFlush(zcs);
     size_t const remainingToFlush = ZSTD_compressStream2(zcs, output, &input, ZSTD_e_end);
-    FORWARD_IF_ERROR( remainingToFlush , "ZSTD_compressStream2 failed");
+    FORWARD_IF_ERROR(remainingToFlush , "ZSTD_compressStream2(,,ZSTD_e_end) failed");
     if (zcs->appliedParams.nbWorkers > 0) return remainingToFlush;   /* minimal estimation */
     /* single thread mode : attempt to calculate remaining to flush more precisely */
     {   size_t const lastBlockSize = zcs->frameEnded ? 0 : ZSTD_BLOCKHEADERSIZE;
@@ -4111,137 +6763,116 @@ size_t ZSTD_endStream(ZSTD_CStream* zcs, ZSTD_outBuffer* output)
 
 
 /*-=====  Pre-defined compression levels  =====-*/
+#include "clevels.h"
 
-#define ZSTD_MAX_CLEVEL     22
 int ZSTD_maxCLevel(void) { return ZSTD_MAX_CLEVEL; }
 int ZSTD_minCLevel(void) { return (int)-ZSTD_TARGETLENGTH_MAX; }
+int ZSTD_defaultCLevel(void) { return ZSTD_CLEVEL_DEFAULT; }
 
-static const ZSTD_compressionParameters ZSTD_defaultCParameters[4][ZSTD_MAX_CLEVEL+1] = {
-{   /* "default" - for any srcSize > 256 KB */
-    /* W,  C,  H,  S,  L, TL, strat */
-    { 19, 12, 13,  1,  6,  1, ZSTD_fast    },  /* base for negative levels */
-    { 19, 13, 14,  1,  7,  0, ZSTD_fast    },  /* level  1 */
-    { 20, 15, 16,  1,  6,  0, ZSTD_fast    },  /* level  2 */
-    { 21, 16, 17,  1,  5,  0, ZSTD_dfast   },  /* level  3 */
-    { 21, 18, 18,  1,  5,  0, ZSTD_dfast   },  /* level  4 */
-    { 21, 18, 19,  2,  5,  2, ZSTD_greedy  },  /* level  5 */
-    { 21, 19, 19,  3,  5,  4, ZSTD_greedy  },  /* level  6 */
-    { 21, 19, 19,  3,  5,  8, ZSTD_lazy    },  /* level  7 */
-    { 21, 19, 19,  3,  5, 16, ZSTD_lazy2   },  /* level  8 */
-    { 21, 19, 20,  4,  5, 16, ZSTD_lazy2   },  /* level  9 */
-    { 22, 20, 21,  4,  5, 16, ZSTD_lazy2   },  /* level 10 */
-    { 22, 21, 22,  4,  5, 16, ZSTD_lazy2   },  /* level 11 */
-    { 22, 21, 22,  5,  5, 16, ZSTD_lazy2   },  /* level 12 */
-    { 22, 21, 22,  5,  5, 32, ZSTD_btlazy2 },  /* level 13 */
-    { 22, 22, 23,  5,  5, 32, ZSTD_btlazy2 },  /* level 14 */
-    { 22, 23, 23,  6,  5, 32, ZSTD_btlazy2 },  /* level 15 */
-    { 22, 22, 22,  5,  5, 48, ZSTD_btopt   },  /* level 16 */
-    { 23, 23, 22,  5,  4, 64, ZSTD_btopt   },  /* level 17 */
-    { 23, 23, 22,  6,  3, 64, ZSTD_btultra },  /* level 18 */
-    { 23, 24, 22,  7,  3,256, ZSTD_btultra2},  /* level 19 */
-    { 25, 25, 23,  7,  3,256, ZSTD_btultra2},  /* level 20 */
-    { 26, 26, 24,  7,  3,512, ZSTD_btultra2},  /* level 21 */
-    { 27, 27, 25,  9,  3,999, ZSTD_btultra2},  /* level 22 */
-},
-{   /* for srcSize <= 256 KB */
-    /* W,  C,  H,  S,  L,  T, strat */
-    { 18, 12, 13,  1,  5,  1, ZSTD_fast    },  /* base for negative levels */
-    { 18, 13, 14,  1,  6,  0, ZSTD_fast    },  /* level  1 */
-    { 18, 14, 14,  1,  5,  0, ZSTD_dfast   },  /* level  2 */
-    { 18, 16, 16,  1,  4,  0, ZSTD_dfast   },  /* level  3 */
-    { 18, 16, 17,  2,  5,  2, ZSTD_greedy  },  /* level  4.*/
-    { 18, 18, 18,  3,  5,  2, ZSTD_greedy  },  /* level  5.*/
-    { 18, 18, 19,  3,  5,  4, ZSTD_lazy    },  /* level  6.*/
-    { 18, 18, 19,  4,  4,  4, ZSTD_lazy    },  /* level  7 */
-    { 18, 18, 19,  4,  4,  8, ZSTD_lazy2   },  /* level  8 */
-    { 18, 18, 19,  5,  4,  8, ZSTD_lazy2   },  /* level  9 */
-    { 18, 18, 19,  6,  4,  8, ZSTD_lazy2   },  /* level 10 */
-    { 18, 18, 19,  5,  4, 12, ZSTD_btlazy2 },  /* level 11.*/
-    { 18, 19, 19,  7,  4, 12, ZSTD_btlazy2 },  /* level 12.*/
-    { 18, 18, 19,  4,  4, 16, ZSTD_btopt   },  /* level 13 */
-    { 18, 18, 19,  4,  3, 32, ZSTD_btopt   },  /* level 14.*/
-    { 18, 18, 19,  6,  3,128, ZSTD_btopt   },  /* level 15.*/
-    { 18, 19, 19,  6,  3,128, ZSTD_btultra },  /* level 16.*/
-    { 18, 19, 19,  8,  3,256, ZSTD_btultra },  /* level 17.*/
-    { 18, 19, 19,  6,  3,128, ZSTD_btultra2},  /* level 18.*/
-    { 18, 19, 19,  8,  3,256, ZSTD_btultra2},  /* level 19.*/
-    { 18, 19, 19, 10,  3,512, ZSTD_btultra2},  /* level 20.*/
-    { 18, 19, 19, 12,  3,512, ZSTD_btultra2},  /* level 21.*/
-    { 18, 19, 19, 13,  3,999, ZSTD_btultra2},  /* level 22.*/
-},
-{   /* for srcSize <= 128 KB */
-    /* W,  C,  H,  S,  L,  T, strat */
-    { 17, 12, 12,  1,  5,  1, ZSTD_fast    },  /* base for negative levels */
-    { 17, 12, 13,  1,  6,  0, ZSTD_fast    },  /* level  1 */
-    { 17, 13, 15,  1,  5,  0, ZSTD_fast    },  /* level  2 */
-    { 17, 15, 16,  2,  5,  0, ZSTD_dfast   },  /* level  3 */
-    { 17, 17, 17,  2,  4,  0, ZSTD_dfast   },  /* level  4 */
-    { 17, 16, 17,  3,  4,  2, ZSTD_greedy  },  /* level  5 */
-    { 17, 17, 17,  3,  4,  4, ZSTD_lazy    },  /* level  6 */
-    { 17, 17, 17,  3,  4,  8, ZSTD_lazy2   },  /* level  7 */
-    { 17, 17, 17,  4,  4,  8, ZSTD_lazy2   },  /* level  8 */
-    { 17, 17, 17,  5,  4,  8, ZSTD_lazy2   },  /* level  9 */
-    { 17, 17, 17,  6,  4,  8, ZSTD_lazy2   },  /* level 10 */
-    { 17, 17, 17,  5,  4,  8, ZSTD_btlazy2 },  /* level 11 */
-    { 17, 18, 17,  7,  4, 12, ZSTD_btlazy2 },  /* level 12 */
-    { 17, 18, 17,  3,  4, 12, ZSTD_btopt   },  /* level 13.*/
-    { 17, 18, 17,  4,  3, 32, ZSTD_btopt   },  /* level 14.*/
-    { 17, 18, 17,  6,  3,256, ZSTD_btopt   },  /* level 15.*/
-    { 17, 18, 17,  6,  3,128, ZSTD_btultra },  /* level 16.*/
-    { 17, 18, 17,  8,  3,256, ZSTD_btultra },  /* level 17.*/
-    { 17, 18, 17, 10,  3,512, ZSTD_btultra },  /* level 18.*/
-    { 17, 18, 17,  5,  3,256, ZSTD_btultra2},  /* level 19.*/
-    { 17, 18, 17,  7,  3,512, ZSTD_btultra2},  /* level 20.*/
-    { 17, 18, 17,  9,  3,512, ZSTD_btultra2},  /* level 21.*/
-    { 17, 18, 17, 11,  3,999, ZSTD_btultra2},  /* level 22.*/
-},
-{   /* for srcSize <= 16 KB */
-    /* W,  C,  H,  S,  L,  T, strat */
-    { 14, 12, 13,  1,  5,  1, ZSTD_fast    },  /* base for negative levels */
-    { 14, 14, 15,  1,  5,  0, ZSTD_fast    },  /* level  1 */
-    { 14, 14, 15,  1,  4,  0, ZSTD_fast    },  /* level  2 */
-    { 14, 14, 15,  2,  4,  0, ZSTD_dfast   },  /* level  3 */
-    { 14, 14, 14,  4,  4,  2, ZSTD_greedy  },  /* level  4 */
-    { 14, 14, 14,  3,  4,  4, ZSTD_lazy    },  /* level  5.*/
-    { 14, 14, 14,  4,  4,  8, ZSTD_lazy2   },  /* level  6 */
-    { 14, 14, 14,  6,  4,  8, ZSTD_lazy2   },  /* level  7 */
-    { 14, 14, 14,  8,  4,  8, ZSTD_lazy2   },  /* level  8.*/
-    { 14, 15, 14,  5,  4,  8, ZSTD_btlazy2 },  /* level  9.*/
-    { 14, 15, 14,  9,  4,  8, ZSTD_btlazy2 },  /* level 10.*/
-    { 14, 15, 14,  3,  4, 12, ZSTD_btopt   },  /* level 11.*/
-    { 14, 15, 14,  4,  3, 24, ZSTD_btopt   },  /* level 12.*/
-    { 14, 15, 14,  5,  3, 32, ZSTD_btultra },  /* level 13.*/
-    { 14, 15, 15,  6,  3, 64, ZSTD_btultra },  /* level 14.*/
-    { 14, 15, 15,  7,  3,256, ZSTD_btultra },  /* level 15.*/
-    { 14, 15, 15,  5,  3, 48, ZSTD_btultra2},  /* level 16.*/
-    { 14, 15, 15,  6,  3,128, ZSTD_btultra2},  /* level 17.*/
-    { 14, 15, 15,  7,  3,256, ZSTD_btultra2},  /* level 18.*/
-    { 14, 15, 15,  8,  3,256, ZSTD_btultra2},  /* level 19.*/
-    { 14, 15, 15,  8,  3,512, ZSTD_btultra2},  /* level 20.*/
-    { 14, 15, 15,  9,  3,512, ZSTD_btultra2},  /* level 21.*/
-    { 14, 15, 15, 10,  3,999, ZSTD_btultra2},  /* level 22.*/
-},
-};
+static ZSTD_compressionParameters ZSTD_dedicatedDictSearch_getCParams(int const compressionLevel, size_t const dictSize)
+{
+    ZSTD_compressionParameters cParams = ZSTD_getCParams_internal(compressionLevel, 0, dictSize, ZSTD_cpm_createCDict);
+    switch (cParams.strategy) {
+        case ZSTD_fast:
+        case ZSTD_dfast:
+            break;
+        case ZSTD_greedy:
+        case ZSTD_lazy:
+        case ZSTD_lazy2:
+            cParams.hashLog += ZSTD_LAZY_DDSS_BUCKET_LOG;
+            break;
+        case ZSTD_btlazy2:
+        case ZSTD_btopt:
+        case ZSTD_btultra:
+        case ZSTD_btultra2:
+            break;
+    }
+    return cParams;
+}
+
+static int ZSTD_dedicatedDictSearch_isSupported(
+        ZSTD_compressionParameters const* cParams)
+{
+    return (cParams->strategy >= ZSTD_greedy)
+        && (cParams->strategy <= ZSTD_lazy2)
+        && (cParams->hashLog > cParams->chainLog)
+        && (cParams->chainLog <= 24);
+}
+
+/**
+ * Reverses the adjustment applied to cparams when enabling dedicated dict
+ * search. This is used to recover the params set to be used in the working
+ * context. (Otherwise, those tables would also grow.)
+ */
+static void ZSTD_dedicatedDictSearch_revertCParams(
+        ZSTD_compressionParameters* cParams) {
+    switch (cParams->strategy) {
+        case ZSTD_fast:
+        case ZSTD_dfast:
+            break;
+        case ZSTD_greedy:
+        case ZSTD_lazy:
+        case ZSTD_lazy2:
+            cParams->hashLog -= ZSTD_LAZY_DDSS_BUCKET_LOG;
+            if (cParams->hashLog < ZSTD_HASHLOG_MIN) {
+                cParams->hashLog = ZSTD_HASHLOG_MIN;
+            }
+            break;
+        case ZSTD_btlazy2:
+        case ZSTD_btopt:
+        case ZSTD_btultra:
+        case ZSTD_btultra2:
+            break;
+    }
+}
+
+static U64 ZSTD_getCParamRowSize(U64 srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode)
+{
+    switch (mode) {
+    case ZSTD_cpm_unknown:
+    case ZSTD_cpm_noAttachDict:
+    case ZSTD_cpm_createCDict:
+        break;
+    case ZSTD_cpm_attachDict:
+        dictSize = 0;
+        break;
+    default:
+        assert(0);
+        break;
+    }
+    {   int const unknown = srcSizeHint == ZSTD_CONTENTSIZE_UNKNOWN;
+        size_t const addedSize = unknown && dictSize > 0 ? 500 : 0;
+        return unknown && dictSize == 0 ? ZSTD_CONTENTSIZE_UNKNOWN : srcSizeHint+dictSize+addedSize;
+    }
+}
 
 /*! ZSTD_getCParams_internal() :
  * @return ZSTD_compressionParameters structure for a selected compression level, srcSize and dictSize.
  *  Note: srcSizeHint 0 means 0, use ZSTD_CONTENTSIZE_UNKNOWN for unknown.
- *        Use dictSize == 0 for unknown or unused. */
-static ZSTD_compressionParameters ZSTD_getCParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize)
+ *        Use dictSize == 0 for unknown or unused.
+ *  Note: `mode` controls how we treat the `dictSize`. See docs for `ZSTD_cParamMode_e`. */
+static ZSTD_compressionParameters ZSTD_getCParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode)
 {
-    int const unknown = srcSizeHint == ZSTD_CONTENTSIZE_UNKNOWN;
-    size_t const addedSize = unknown && dictSize > 0 ? 500 : 0;
-    U64 const rSize = unknown && dictSize == 0 ? ZSTD_CONTENTSIZE_UNKNOWN : srcSizeHint+dictSize+addedSize;
+    U64 const rSize = ZSTD_getCParamRowSize(srcSizeHint, dictSize, mode);
     U32 const tableID = (rSize <= 256 KB) + (rSize <= 128 KB) + (rSize <= 16 KB);
-    int row = compressionLevel;
+    int row;
     DEBUGLOG(5, "ZSTD_getCParams_internal (cLevel=%i)", compressionLevel);
+
+    /* row */
     if (compressionLevel == 0) row = ZSTD_CLEVEL_DEFAULT;   /* 0 == default */
-    if (compressionLevel < 0) row = 0;   /* entry 0 is baseline for fast mode */
-    if (compressionLevel > ZSTD_MAX_CLEVEL) row = ZSTD_MAX_CLEVEL;
+    else if (compressionLevel < 0) row = 0;   /* entry 0 is baseline for fast mode */
+    else if (compressionLevel > ZSTD_MAX_CLEVEL) row = ZSTD_MAX_CLEVEL;
+    else row = compressionLevel;
+
     {   ZSTD_compressionParameters cp = ZSTD_defaultCParameters[tableID][row];
-        if (compressionLevel < 0) cp.targetLength = (unsigned)(-compressionLevel);   /* acceleration factor */
+        DEBUGLOG(5, "ZSTD_getCParams_internal selected tableID: %u row: %u strat: %u", tableID, row, (U32)cp.strategy);
+        /* acceleration factor */
+        if (compressionLevel < 0) {
+            int const clampedCompressionLevel = MAX(ZSTD_minCLevel(), compressionLevel);
+            cp.targetLength = (unsigned)(-clampedCompressionLevel);
+        }
         /* refine parameters based on srcSize & dictSize */
-        return ZSTD_adjustCParams_internal(cp, srcSizeHint, dictSize);
+        return ZSTD_adjustCParams_internal(cp, srcSizeHint, dictSize, mode, ZSTD_ps_auto);
     }
 }
 
@@ -4251,18 +6882,18 @@ static ZSTD_compressionParameters ZSTD_getCParams_internal(int compressionLevel,
 ZSTD_compressionParameters ZSTD_getCParams(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize)
 {
     if (srcSizeHint == 0) srcSizeHint = ZSTD_CONTENTSIZE_UNKNOWN;
-    return ZSTD_getCParams_internal(compressionLevel, srcSizeHint, dictSize);
+    return ZSTD_getCParams_internal(compressionLevel, srcSizeHint, dictSize, ZSTD_cpm_unknown);
 }
 
 /*! ZSTD_getParams() :
  *  same idea as ZSTD_getCParams()
  * @return a `ZSTD_parameters` structure (instead of `ZSTD_compressionParameters`).
  *  Fields of `ZSTD_frameParameters` are set to default values */
-static ZSTD_parameters ZSTD_getParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize) {
+static ZSTD_parameters ZSTD_getParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode) {
     ZSTD_parameters params;
-    ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, srcSizeHint, dictSize);
+    ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, srcSizeHint, dictSize, mode);
     DEBUGLOG(5, "ZSTD_getParams (cLevel=%i)", compressionLevel);
-    memset(&params, 0, sizeof(params));
+    ZSTD_memset(&params, 0, sizeof(params));
     params.cParams = cParams;
     params.fParams.contentSizeFlag = 1;
     return params;
@@ -4274,5 +6905,23 @@ static ZSTD_parameters ZSTD_getParams_internal(int compressionLevel, unsigned lo
  *  Fields of `ZSTD_frameParameters` are set to default values */
 ZSTD_parameters ZSTD_getParams(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize) {
     if (srcSizeHint == 0) srcSizeHint = ZSTD_CONTENTSIZE_UNKNOWN;
-    return ZSTD_getParams_internal(compressionLevel, srcSizeHint, dictSize);
+    return ZSTD_getParams_internal(compressionLevel, srcSizeHint, dictSize, ZSTD_cpm_unknown);
+}
+
+void ZSTD_registerSequenceProducer(
+    ZSTD_CCtx* zc, void* mState,
+    ZSTD_sequenceProducer_F* mFinder
+) {
+    if (mFinder != NULL) {
+        ZSTD_externalMatchCtx emctx;
+        emctx.mState = mState;
+        emctx.mFinder = mFinder;
+        emctx.seqBuffer = NULL;
+        emctx.seqBufferCapacity = 0;
+        zc->externalMatchCtx = emctx;
+        zc->requestedParams.useSequenceProducer = 1;
+    } else {
+        ZSTD_memset(&zc->externalMatchCtx, 0, sizeof(zc->externalMatchCtx));
+        zc->requestedParams.useSequenceProducer = 0;
+    }
 }
diff --git a/contrib/zstd/zstd_compress_internal.h b/contrib/zstd/zstd_compress_internal.h
index c500687ce..e53b3a53f 100644
--- a/contrib/zstd/zstd_compress_internal.h
+++ b/contrib/zstd/zstd_compress_internal.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -23,12 +23,12 @@
 #ifdef ZSTD_MULTITHREAD
 #  include "zstdmt_compress.h"
 #endif
+#include "bits.h" /* ZSTD_highbit32, ZSTD_NbCommonBytes */
 
 #if defined (__cplusplus)
 extern "C" {
 #endif
 
-
 /*-*************************************
 *  Constants
 ***************************************/
@@ -64,7 +64,7 @@ typedef struct {
 } ZSTD_localDict;
 
 typedef struct {
-    U32 CTable[HUF_CTABLE_SIZE_U32(255)];
+    HUF_CElt CTable[HUF_CTABLE_SIZE_ST(255)];
     HUF_repeat repeatMode;
 } ZSTD_hufCTables_t;
 
@@ -82,12 +82,83 @@ typedef struct {
     ZSTD_fseCTables_t fse;
 } ZSTD_entropyCTables_t;
 
+/***********************************************
+*  Entropy buffer statistics structs and funcs *
+***********************************************/
+/** ZSTD_hufCTablesMetadata_t :
+ *  Stores Literals Block Type for a super-block in hType, and
+ *  huffman tree description in hufDesBuffer.
+ *  hufDesSize refers to the size of huffman tree description in bytes.
+ *  This metadata is populated in ZSTD_buildBlockEntropyStats_literals() */
 typedef struct {
-    U32 off;
-    U32 len;
+    symbolEncodingType_e hType;
+    BYTE hufDesBuffer[ZSTD_MAX_HUF_HEADER_SIZE];
+    size_t hufDesSize;
+} ZSTD_hufCTablesMetadata_t;
+
+/** ZSTD_fseCTablesMetadata_t :
+ *  Stores symbol compression modes for a super-block in {ll, ol, ml}Type, and
+ *  fse tables in fseTablesBuffer.
+ *  fseTablesSize refers to the size of fse tables in bytes.
+ *  This metadata is populated in ZSTD_buildBlockEntropyStats_sequences() */
+typedef struct {
+    symbolEncodingType_e llType;
+    symbolEncodingType_e ofType;
+    symbolEncodingType_e mlType;
+    BYTE fseTablesBuffer[ZSTD_MAX_FSE_HEADERS_SIZE];
+    size_t fseTablesSize;
+    size_t lastCountSize; /* This is to account for bug in 1.3.4. More detail in ZSTD_entropyCompressSeqStore_internal() */
+} ZSTD_fseCTablesMetadata_t;
+
+typedef struct {
+    ZSTD_hufCTablesMetadata_t hufMetadata;
+    ZSTD_fseCTablesMetadata_t fseMetadata;
+} ZSTD_entropyCTablesMetadata_t;
+
+/** ZSTD_buildBlockEntropyStats() :
+ *  Builds entropy for the block.
+ *  @return : 0 on success or error code */
+size_t ZSTD_buildBlockEntropyStats(
+                    const seqStore_t* seqStorePtr,
+                    const ZSTD_entropyCTables_t* prevEntropy,
+                          ZSTD_entropyCTables_t* nextEntropy,
+                    const ZSTD_CCtx_params* cctxParams,
+                          ZSTD_entropyCTablesMetadata_t* entropyMetadata,
+                          void* workspace, size_t wkspSize);
+
+/*********************************
+*  Compression internals structs *
+*********************************/
+
+typedef struct {
+    U32 off;            /* Offset sumtype code for the match, using ZSTD_storeSeq() format */
+    U32 len;            /* Raw length of match */
 } ZSTD_match_t;
 
 typedef struct {
+    U32 offset;         /* Offset of sequence */
+    U32 litLength;      /* Length of literals prior to match */
+    U32 matchLength;    /* Raw length of match */
+} rawSeq;
+
+typedef struct {
+  rawSeq* seq;          /* The start of the sequences */
+  size_t pos;           /* The index in seq where reading stopped. pos <= size. */
+  size_t posInSequence; /* The position within the sequence at seq[pos] where reading
+                           stopped. posInSequence <= seq[pos].litLength + seq[pos].matchLength */
+  size_t size;          /* The number of sequences. <= capacity. */
+  size_t capacity;      /* The capacity starting from `seq` pointer */
+} rawSeqStore_t;
+
+typedef struct {
+    U32 idx;            /* Index in array of ZSTD_Sequence */
+    U32 posInSequence;  /* Position within sequence at idx */
+    size_t posInSrc;    /* Number of bytes given by sequences provided so far */
+} ZSTD_sequencePosition;
+
+UNUSED_ATTR static const rawSeqStore_t kNullRawSeqStore = {NULL, 0, 0, 0, 0};
+
+typedef struct {
     int price;
     U32 off;
     U32 mlen;
@@ -116,7 +187,7 @@ typedef struct {
     U32  offCodeSumBasePrice;    /* to compare to log2(offreq)  */
     ZSTD_OptPrice_e priceType;   /* prices can be determined dynamically, or follow a pre-defined cost structure */
     const ZSTD_entropyCTables_t* symbolCosts;  /* pre-calculated dictionary statistics */
-    ZSTD_literalCompressionMode_e literalCompressionMode;
+    ZSTD_paramSwitch_e literalCompressionMode;
 } optState_t;
 
 typedef struct {
@@ -125,14 +196,23 @@ typedef struct {
 } ZSTD_compressedBlockState_t;
 
 typedef struct {
-    BYTE const* nextSrc;    /* next block here to continue on current prefix */
-    BYTE const* base;       /* All regular indexes relative to this position */
-    BYTE const* dictBase;   /* extDict indexes relative to this position */
-    U32 dictLimit;          /* below that point, need extDict */
-    U32 lowLimit;           /* below that point, no more valid data */
+    BYTE const* nextSrc;       /* next block here to continue on current prefix */
+    BYTE const* base;          /* All regular indexes relative to this position */
+    BYTE const* dictBase;      /* extDict indexes relative to this position */
+    U32 dictLimit;             /* below that point, need extDict */
+    U32 lowLimit;              /* below that point, no more valid data */
+    U32 nbOverflowCorrections; /* Number of times overflow correction has run since
+                                * ZSTD_window_init(). Useful for debugging coredumps
+                                * and for ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY.
+                                */
 } ZSTD_window_t;
 
+#define ZSTD_WINDOW_START_INDEX 2
+
 typedef struct ZSTD_matchState_t ZSTD_matchState_t;
+
+#define ZSTD_ROW_HASH_CACHE_SIZE 8       /* Size of prefetching hash cache for row-based matchfinder */
+
 struct ZSTD_matchState_t {
     ZSTD_window_t window;   /* State for window round buffer management */
     U32 loadedDictEnd;      /* index of end of dictionary, within context's referential.
@@ -144,12 +224,29 @@ struct ZSTD_matchState_t {
                              */
     U32 nextToUpdate;       /* index from which to continue table update */
     U32 hashLog3;           /* dispatch table for matches of len==3 : larger == faster, more memory */
+
+    U32 rowHashLog;                          /* For row-based matchfinder: Hashlog based on nb of rows in the hashTable.*/
+    U16* tagTable;                           /* For row-based matchFinder: A row-based table containing the hashes and head index. */
+    U32 hashCache[ZSTD_ROW_HASH_CACHE_SIZE]; /* For row-based matchFinder: a cache of hashes to improve speed */
+
     U32* hashTable;
     U32* hashTable3;
     U32* chainTable;
+
+    U32 forceNonContiguous; /* Non-zero if we should force non-contiguous load for the next window update. */
+
+    int dedicatedDictSearch;  /* Indicates whether this matchState is using the
+                               * dedicated dictionary search structure.
+                               */
     optState_t opt;         /* optimal parser state */
     const ZSTD_matchState_t* dictMatchState;
     ZSTD_compressionParameters cParams;
+    const rawSeqStore_t* ldmSeqStore;
+
+    /* Controls prefetching in some dictMatchState matchfinders.
+     * This behavior is controlled from the cctx ms.
+     * This parameter has no effect in the cdict ms. */
+    int prefetchCDictTables;
 };
 
 typedef struct {
@@ -164,16 +261,25 @@ typedef struct {
 } ldmEntry_t;
 
 typedef struct {
+    BYTE const* split;
+    U32 hash;
+    U32 checksum;
+    ldmEntry_t* bucket;
+} ldmMatchCandidate_t;
+
+#define LDM_BATCH_SIZE 64
+
+typedef struct {
     ZSTD_window_t window;   /* State for the window round buffer management */
     ldmEntry_t* hashTable;
     U32 loadedDictEnd;
     BYTE* bucketOffsets;    /* Next position in bucket to insert entry */
-    U64 hashPower;          /* Used to compute the rolling hash.
-                             * Depends on ldmParams.minMatchLength */
+    size_t splitIndices[LDM_BATCH_SIZE];
+    ldmMatchCandidate_t matchCandidates[LDM_BATCH_SIZE];
 } ldmState_t;
 
 typedef struct {
-    U32 enableLdm;          /* 1 if enable long distance matching */
+    ZSTD_paramSwitch_e enableLdm; /* ZSTD_ps_enable to enable LDM. ZSTD_ps_auto by default */
     U32 hashLog;            /* Log size of hashTable */
     U32 bucketSizeLog;      /* Log bucket size for collision resolution, at most 8 */
     U32 minMatchLength;     /* Minimum match length */
@@ -182,19 +288,6 @@ typedef struct {
 } ldmParams_t;
 
 typedef struct {
-    U32 offset;
-    U32 litLength;
-    U32 matchLength;
-} rawSeq;
-
-typedef struct {
-  rawSeq* seq;     /* The start of the sequences */
-  size_t pos;      /* The position where reading stopped. <= size. */
-  size_t size;     /* The number of sequences. <= capacity. */
-  size_t capacity; /* The capacity starting from `seq` pointer */
-} rawSeqStore_t;
-
-typedef struct {
     int collectSequences;
     ZSTD_Sequence* seqStart;
     size_t seqIndex;
@@ -217,7 +310,7 @@ struct ZSTD_CCtx_params_s {
                                 * There is no guarantee that hint is close to actual source size */
 
     ZSTD_dictAttachPref_e attachDictPref;
-    ZSTD_literalCompressionMode_e literalCompressionMode;
+    ZSTD_paramSwitch_e literalCompressionMode;
 
     /* Multithreading: used to pass parameters to mtctx */
     int nbWorkers;
@@ -228,17 +321,94 @@ struct ZSTD_CCtx_params_s {
     /* Long distance matching parameters */
     ldmParams_t ldmParams;
 
+    /* Dedicated dict search algorithm trigger */
+    int enableDedicatedDictSearch;
+
+    /* Input/output buffer modes */
+    ZSTD_bufferMode_e inBufferMode;
+    ZSTD_bufferMode_e outBufferMode;
+
+    /* Sequence compression API */
+    ZSTD_sequenceFormat_e blockDelimiters;
+    int validateSequences;
+
+    /* Block splitting */
+    ZSTD_paramSwitch_e useBlockSplitter;
+
+    /* Param for deciding whether to use row-based matchfinder */
+    ZSTD_paramSwitch_e useRowMatchFinder;
+
+    /* Always load a dictionary in ext-dict mode (not prefix mode)? */
+    int deterministicRefPrefix;
+
     /* Internal use, for createCCtxParams() and freeCCtxParams() only */
     ZSTD_customMem customMem;
+
+    /* Controls prefetching in some dictMatchState matchfinders */
+    ZSTD_paramSwitch_e prefetchCDictTables;
+
+    /* Controls whether zstd will fall back to an internal matchfinder
+     * if the external matchfinder returns an error code. */
+    int enableMatchFinderFallback;
+
+    /* Indicates whether an external matchfinder has been referenced.
+     * Users can't set this externally.
+     * It is set internally in ZSTD_registerSequenceProducer(). */
+    int useSequenceProducer;
+
+    /* Adjust the max block size*/
+    size_t maxBlockSize;
+
+    /* Controls repcode search in external sequence parsing */
+    ZSTD_paramSwitch_e searchForExternalRepcodes;
 };  /* typedef'd to ZSTD_CCtx_params within "zstd.h" */
 
+#define COMPRESS_SEQUENCES_WORKSPACE_SIZE (sizeof(unsigned) * (MaxSeq + 2))
+#define ENTROPY_WORKSPACE_SIZE (HUF_WORKSPACE_SIZE + COMPRESS_SEQUENCES_WORKSPACE_SIZE)
+
+/**
+ * Indicates whether this compression proceeds directly from user-provided
+ * source buffer to user-provided destination buffer (ZSTDb_not_buffered), or
+ * whether the context needs to buffer the input/output (ZSTDb_buffered).
+ */
+typedef enum {
+    ZSTDb_not_buffered,
+    ZSTDb_buffered
+} ZSTD_buffered_policy_e;
+
+/**
+ * Struct that contains all elements of block splitter that should be allocated
+ * in a wksp.
+ */
+#define ZSTD_MAX_NB_BLOCK_SPLITS 196
+typedef struct {
+    seqStore_t fullSeqStoreChunk;
+    seqStore_t firstHalfSeqStore;
+    seqStore_t secondHalfSeqStore;
+    seqStore_t currSeqStore;
+    seqStore_t nextSeqStore;
+
+    U32 partitions[ZSTD_MAX_NB_BLOCK_SPLITS];
+    ZSTD_entropyCTablesMetadata_t entropyMetadata;
+} ZSTD_blockSplitCtx;
+
+/* Context for block-level external matchfinder API */
+typedef struct {
+  void* mState;
+  ZSTD_sequenceProducer_F* mFinder;
+  ZSTD_Sequence* seqBuffer;
+  size_t seqBufferCapacity;
+} ZSTD_externalMatchCtx;
+
 struct ZSTD_CCtx_s {
     ZSTD_compressionStage_e stage;
     int cParamsChanged;                  /* == 1 if cParams(except wlog) or compression level are changed in requestedParams. Triggers transmission of new params to ZSTDMT (if available) then reset to 0. */
     int bmi2;                            /* == 1 if the CPU supports BMI2 and 0 otherwise. CPU support is determined dynamically once per context lifetime. */
     ZSTD_CCtx_params requestedParams;
     ZSTD_CCtx_params appliedParams;
+    ZSTD_CCtx_params simpleApiParams;    /* Param storage used by the simple API - not sticky. Must only be used in top-level simple API functions for storage. */
     U32   dictID;
+    size_t dictContentSize;
 
     ZSTD_cwksp workspace; /* manages buffer for dynamic allocations */
     size_t blockSize;
@@ -247,6 +417,7 @@ struct ZSTD_CCtx_s {
     unsigned long long producedCSize;
     XXH64_state_t xxhState;
     ZSTD_customMem customMem;
+    ZSTD_threadPool* pool;
     size_t staticSize;
     SeqCollector seqCollector;
     int isFirstBlock;
@@ -258,7 +429,10 @@ struct ZSTD_CCtx_s {
     size_t maxNbLdmSequences;
     rawSeqStore_t externSeqStore; /* Mutable reference to external sequences */
     ZSTD_blockState_t blockState;
-    U32* entropyWorkspace;  /* entropy workspace of HUF_WORKSPACE_SIZE bytes */
+    U32* entropyWorkspace;  /* entropy workspace of ENTROPY_WORKSPACE_SIZE bytes */
+
+    /* Whether we are streaming or not */
+    ZSTD_buffered_policy_e bufferedPolicy;
 
     /* streaming */
     char*  inBuff;
@@ -273,6 +447,11 @@ struct ZSTD_CCtx_s {
     ZSTD_cStreamStage streamStage;
     U32    frameEnded;
 
+    /* Stable in/out buffer verification */
+    ZSTD_inBuffer expectedInBuffer;
+    size_t stableIn_notConsumed; /* nb bytes within stable input buffer that are said to be consumed but are not */
+    size_t expectedOutBufferSize;
+
     /* Dictionary */
     ZSTD_localDict localDict;
     const ZSTD_CDict* cdict;
@@ -282,17 +461,53 @@ struct ZSTD_CCtx_s {
 #ifdef ZSTD_MULTITHREAD
     ZSTDMT_CCtx* mtctx;
 #endif
-};
 
-typedef enum { ZSTD_dtlm_fast, ZSTD_dtlm_full } ZSTD_dictTableLoadMethod_e;
+    /* Tracing */
+#if ZSTD_TRACE
+    ZSTD_TraceCtx traceCtx;
+#endif
 
-typedef enum { ZSTD_noDict = 0, ZSTD_extDict = 1, ZSTD_dictMatchState = 2 } ZSTD_dictMode_e;
+    /* Workspace for block splitter */
+    ZSTD_blockSplitCtx blockSplitCtx;
+
+    /* Workspace for external matchfinder */
+    ZSTD_externalMatchCtx externalMatchCtx;
+};
 
+typedef enum { ZSTD_dtlm_fast, ZSTD_dtlm_full } ZSTD_dictTableLoadMethod_e;
+typedef enum { ZSTD_tfp_forCCtx, ZSTD_tfp_forCDict } ZSTD_tableFillPurpose_e;
+
+typedef enum {
+    ZSTD_noDict = 0,
+    ZSTD_extDict = 1,
+    ZSTD_dictMatchState = 2,
+    ZSTD_dedicatedDictSearch = 3
+} ZSTD_dictMode_e;
+
+typedef enum {
+    ZSTD_cpm_noAttachDict = 0,  /* Compression with ZSTD_noDict or ZSTD_extDict.
+                                 * In this mode we use both the srcSize and the dictSize
+                                 * when selecting and adjusting parameters.
+                                 */
+    ZSTD_cpm_attachDict = 1,    /* Compression with ZSTD_dictMatchState or ZSTD_dedicatedDictSearch.
+                                 * In this mode we only take the srcSize into account when selecting
+                                 * and adjusting parameters.
+                                 */
+    ZSTD_cpm_createCDict = 2,   /* Creating a CDict.
+                                 * In this mode we take both the source size and the dictionary size
+                                 * into account when selecting and adjusting the parameters.
+                                 */
+    ZSTD_cpm_unknown = 3        /* ZSTD_getCParams, ZSTD_getParams, ZSTD_adjustParams.
+                                 * We don't know what these parameters are for. We default to the legacy
+                                 * behavior of taking both the source size and the dict size into account
+                                 * when selecting and adjusting parameters.
+                                 */
+} ZSTD_cParamMode_e;
 
 typedef size_t (*ZSTD_blockCompressor) (
         ZSTD_matchState_t* bs, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
-ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_dictMode_e dictMode);
+ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_paramSwitch_e rowMatchfinderMode, ZSTD_dictMode_e dictMode);
 
 
 MEM_STATIC U32 ZSTD_LLcode(U32 litLength)
@@ -326,31 +541,6 @@ MEM_STATIC U32 ZSTD_MLcode(U32 mlBase)
     return (mlBase > 127) ? ZSTD_highbit32(mlBase) + ML_deltaCode : ML_Code[mlBase];
 }
 
-typedef struct repcodes_s {
-    U32 rep[3];
-} repcodes_t;
-
-MEM_STATIC repcodes_t ZSTD_updateRep(U32 const rep[3], U32 const offset, U32 const ll0)
-{
-    repcodes_t newReps;
-    if (offset >= ZSTD_REP_NUM) {  /* full offset */
-        newReps.rep[2] = rep[1];
-        newReps.rep[1] = rep[0];
-        newReps.rep[0] = offset - ZSTD_REP_MOVE;
-    } else {   /* repcode */
-        U32 const repCode = offset + ll0;
-        if (repCode > 0) {  /* note : if repCode==0, no change */
-            U32 const currentOffset = (repCode==ZSTD_REP_NUM) ? (rep[0] - 1) : rep[repCode];
-            newReps.rep[2] = (repCode >= 2) ? rep[1] : rep[2];
-            newReps.rep[1] = rep[0];
-            newReps.rep[0] = currentOffset;
-        } else {   /* repCode == 0 */
-            memcpy(&newReps, rep, sizeof(newReps));
-        }
-    }
-    return newReps;
-}
-
 /* ZSTD_cParam_withinBounds:
  * @return 1 if value is within cParam bounds,
  * 0 otherwise */
@@ -366,17 +556,20 @@ MEM_STATIC int ZSTD_cParam_withinBounds(ZSTD_cParameter cParam, int value)
 /* ZSTD_noCompressBlock() :
  * Writes uncompressed block to dst buffer from given src.
  * Returns the size of the block */
-MEM_STATIC size_t ZSTD_noCompressBlock (void* dst, size_t dstCapacity, const void* src, size_t srcSize, U32 lastBlock)
+MEM_STATIC size_t
+ZSTD_noCompressBlock(void* dst, size_t dstCapacity, const void* src, size_t srcSize, U32 lastBlock)
 {
     U32 const cBlockHeader24 = lastBlock + (((U32)bt_raw)<<1) + (U32)(srcSize << 3);
+    DEBUGLOG(5, "ZSTD_noCompressBlock (srcSize=%zu, dstCapacity=%zu)", srcSize, dstCapacity);
     RETURN_ERROR_IF(srcSize + ZSTD_blockHeaderSize > dstCapacity,
                     dstSize_tooSmall, "dst buf too small for uncompressed block");
     MEM_writeLE24(dst, cBlockHeader24);
-    memcpy((BYTE*)dst + ZSTD_blockHeaderSize, src, srcSize);
+    ZSTD_memcpy((BYTE*)dst + ZSTD_blockHeaderSize, src, srcSize);
     return ZSTD_blockHeaderSize + srcSize;
 }
 
-MEM_STATIC size_t ZSTD_rleCompressBlock (void* dst, size_t dstCapacity, BYTE src, size_t srcSize, U32 lastBlock)
+MEM_STATIC size_t
+ZSTD_rleCompressBlock(void* dst, size_t dstCapacity, BYTE src, size_t srcSize, U32 lastBlock)
 {
     BYTE* const op = (BYTE*)dst;
     U32 const cBlockHeader = lastBlock + (((U32)bt_rle)<<1) + (U32)(srcSize << 3);
@@ -395,21 +588,21 @@ MEM_STATIC size_t ZSTD_minGain(size_t srcSize, ZSTD_strategy strat)
 {
     U32 const minlog = (strat>=ZSTD_btultra) ? (U32)(strat) - 1 : 6;
     ZSTD_STATIC_ASSERT(ZSTD_btultra == 8);
-    assert(ZSTD_cParam_withinBounds(ZSTD_c_strategy, strat));
+    assert(ZSTD_cParam_withinBounds(ZSTD_c_strategy, (int)strat));
     return (srcSize >> minlog) + 2;
 }
 
-MEM_STATIC int ZSTD_disableLiteralsCompression(const ZSTD_CCtx_params* cctxParams)
+MEM_STATIC int ZSTD_literalsCompressionIsDisabled(const ZSTD_CCtx_params* cctxParams)
 {
     switch (cctxParams->literalCompressionMode) {
-    case ZSTD_lcm_huffman:
+    case ZSTD_ps_enable:
         return 0;
-    case ZSTD_lcm_uncompressed:
+    case ZSTD_ps_disable:
         return 1;
     default:
         assert(0 /* impossible: pre-validated */);
-        /* fall-through */
-    case ZSTD_lcm_auto:
+        ZSTD_FALLTHROUGH;
+    case ZSTD_ps_auto:
         return (cctxParams->cParams.strategy == ZSTD_fast) && (cctxParams->cParams.targetLength > 0);
     }
 }
@@ -419,7 +612,9 @@ MEM_STATIC int ZSTD_disableLiteralsCompression(const ZSTD_CCtx_params* cctxParam
  *  Only called when the sequence ends past ilimit_w, so it only needs to be optimized for single
  *  large copies.
  */
-static void ZSTD_safecopyLiterals(BYTE* op, BYTE const* ip, BYTE const* const iend, BYTE const* ilimit_w) {
+static void
+ZSTD_safecopyLiterals(BYTE* op, BYTE const* ip, BYTE const* const iend, BYTE const* ilimit_w)
+{
     assert(iend > ilimit_w);
     if (ip <= ilimit_w) {
         ZSTD_wildcopy(op, ip, ilimit_w - ip, ZSTD_no_overlap);
@@ -429,14 +624,28 @@ static void ZSTD_safecopyLiterals(BYTE* op, BYTE const* ip, BYTE const* const ie
     while (ip < iend) *op++ = *ip++;
 }
 
+
+#define REPCODE1_TO_OFFBASE REPCODE_TO_OFFBASE(1)
+#define REPCODE2_TO_OFFBASE REPCODE_TO_OFFBASE(2)
+#define REPCODE3_TO_OFFBASE REPCODE_TO_OFFBASE(3)
+#define REPCODE_TO_OFFBASE(r) (assert((r)>=1), assert((r)<=ZSTD_REP_NUM), (r)) /* accepts IDs 1,2,3 */
+#define OFFSET_TO_OFFBASE(o)  (assert((o)>0), o + ZSTD_REP_NUM)
+#define OFFBASE_IS_OFFSET(o)  ((o) > ZSTD_REP_NUM)
+#define OFFBASE_IS_REPCODE(o) ( 1 <= (o) && (o) <= ZSTD_REP_NUM)
+#define OFFBASE_TO_OFFSET(o)  (assert(OFFBASE_IS_OFFSET(o)), (o) - ZSTD_REP_NUM)
+#define OFFBASE_TO_REPCODE(o) (assert(OFFBASE_IS_REPCODE(o)), (o))  /* returns ID 1,2,3 */
+
 /*! ZSTD_storeSeq() :
- *  Store a sequence (litlen, litPtr, offCode and mlBase) into seqStore_t.
- *  `offCode` : distance to match + ZSTD_REP_MOVE (values <= ZSTD_REP_MOVE are repCodes).
- *  `mlBase` : matchLength - MINMATCH
- *  Allowed to overread literals up to litLimit.
+ *  Store a sequence (litlen, litPtr, offBase and matchLength) into seqStore_t.
+ *  @offBase : Users should employ macros REPCODE_TO_OFFBASE() and OFFSET_TO_OFFBASE().
+ *  @matchLength : must be >= MINMATCH
+ *  Allowed to over-read literals up to litLimit.
 */
-HINT_INLINE UNUSED_ATTR
-void ZSTD_storeSeq(seqStore_t* seqStorePtr, size_t litLength, const BYTE* literals, const BYTE* litLimit, U32 offCode, size_t mlBase)
+HINT_INLINE UNUSED_ATTR void
+ZSTD_storeSeq(seqStore_t* seqStorePtr,
+              size_t litLength, const BYTE* literals, const BYTE* litLimit,
+              U32 offBase,
+              size_t matchLength)
 {
     BYTE const* const litLimit_w = litLimit - WILDCOPY_OVERLENGTH;
     BYTE const* const litEnd = literals + litLength;
@@ -444,8 +653,8 @@ void ZSTD_storeSeq(seqStore_t* seqStorePtr, size_t litLength, const BYTE* litera
     static const BYTE* g_start = NULL;
     if (g_start==NULL) g_start = (const BYTE*)literals;  /* note : index only works for compression within a single segment */
     {   U32 const pos = (U32)((const BYTE*)literals - g_start);
-        DEBUGLOG(6, "Cpos%7u :%3u literals, match%4u bytes at offCode%7u",
-               pos, (U32)litLength, (U32)mlBase+MINMATCH, (U32)offCode);
+        DEBUGLOG(6, "Cpos%7u :%3u literals, match%4u bytes at offBase%7u",
+               pos, (U32)litLength, (U32)matchLength, (U32)offBase);
     }
 #endif
     assert((size_t)(seqStorePtr->sequences - seqStorePtr->sequencesStart) < seqStorePtr->maxNbSeq);
@@ -455,9 +664,9 @@ void ZSTD_storeSeq(seqStore_t* seqStorePtr, size_t litLength, const BYTE* litera
     assert(literals + litLength <= litLimit);
     if (litEnd <= litLimit_w) {
         /* Common case we can use wildcopy.
-	 * First copy 16 bytes, because literals are likely short.
-	 */
-        assert(WILDCOPY_OVERLENGTH >= 16);
+         * First copy 16 bytes, because literals are likely short.
+         */
+        ZSTD_STATIC_ASSERT(WILDCOPY_OVERLENGTH >= 16);
         ZSTD_copy16(seqStorePtr->lit, literals);
         if (litLength > 16) {
             ZSTD_wildcopy(seqStorePtr->lit+16, literals+16, (ptrdiff_t)litLength-16, ZSTD_no_overlap);
@@ -469,95 +678,70 @@ void ZSTD_storeSeq(seqStore_t* seqStorePtr, size_t litLength, const BYTE* litera
 
     /* literal Length */
     if (litLength>0xFFFF) {
-        assert(seqStorePtr->longLengthID == 0); /* there can only be a single long length */
-        seqStorePtr->longLengthID = 1;
+        assert(seqStorePtr->longLengthType == ZSTD_llt_none); /* there can only be a single long length */
+        seqStorePtr->longLengthType = ZSTD_llt_literalLength;
         seqStorePtr->longLengthPos = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
     }
     seqStorePtr->sequences[0].litLength = (U16)litLength;
 
     /* match offset */
-    seqStorePtr->sequences[0].offset = offCode + 1;
+    seqStorePtr->sequences[0].offBase = offBase;
 
     /* match Length */
-    if (mlBase>0xFFFF) {
-        assert(seqStorePtr->longLengthID == 0); /* there can only be a single long length */
-        seqStorePtr->longLengthID = 2;
-        seqStorePtr->longLengthPos = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
+    assert(matchLength >= MINMATCH);
+    {   size_t const mlBase = matchLength - MINMATCH;
+        if (mlBase>0xFFFF) {
+            assert(seqStorePtr->longLengthType == ZSTD_llt_none); /* there can only be a single long length */
+            seqStorePtr->longLengthType = ZSTD_llt_matchLength;
+            seqStorePtr->longLengthPos = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
+        }
+        seqStorePtr->sequences[0].mlBase = (U16)mlBase;
     }
-    seqStorePtr->sequences[0].matchLength = (U16)mlBase;
 
     seqStorePtr->sequences++;
 }
 
-
-/*-*************************************
-*  Match length counter
-***************************************/
-static unsigned ZSTD_NbCommonBytes (size_t val)
+/* ZSTD_updateRep() :
+ * updates in-place @rep (array of repeat offsets)
+ * @offBase : sum-type, using numeric representation of ZSTD_storeSeq()
+ */
+MEM_STATIC void
+ZSTD_updateRep(U32 rep[ZSTD_REP_NUM], U32 const offBase, U32 const ll0)
 {
-    if (MEM_isLittleEndian()) {
-        if (MEM_64bits()) {
-#       if defined(_MSC_VER) && defined(_WIN64)
-            unsigned long r = 0;
-            return _BitScanForward64( &r, (U64)val ) ? (unsigned)(r >> 3) : 0;
-#       elif defined(__GNUC__) && (__GNUC__ >= 4)
-            return (__builtin_ctzll((U64)val) >> 3);
-#       else
-            static const int DeBruijnBytePos[64] = { 0, 0, 0, 0, 0, 1, 1, 2,
-                                                     0, 3, 1, 3, 1, 4, 2, 7,
-                                                     0, 2, 3, 6, 1, 5, 3, 5,
-                                                     1, 3, 4, 4, 2, 5, 6, 7,
-                                                     7, 0, 1, 2, 3, 3, 4, 6,
-                                                     2, 6, 5, 5, 3, 4, 5, 6,
-                                                     7, 1, 2, 4, 6, 4, 4, 5,
-                                                     7, 2, 6, 5, 7, 6, 7, 7 };
-            return DeBruijnBytePos[((U64)((val & -(long long)val) * 0x0218A392CDABBD3FULL)) >> 58];
-#       endif
-        } else { /* 32 bits */
-#       if defined(_MSC_VER)
-            unsigned long r=0;
-            return _BitScanForward( &r, (U32)val ) ? (unsigned)(r >> 3) : 0;
-#       elif defined(__GNUC__) && (__GNUC__ >= 3)
-            return (__builtin_ctz((U32)val) >> 3);
-#       else
-            static const int DeBruijnBytePos[32] = { 0, 0, 3, 0, 3, 1, 3, 0,
-                                                     3, 2, 2, 1, 3, 2, 0, 1,
-                                                     3, 3, 1, 2, 2, 2, 2, 0,
-                                                     3, 1, 2, 0, 1, 0, 1, 1 };
-            return DeBruijnBytePos[((U32)((val & -(S32)val) * 0x077CB531U)) >> 27];
-#       endif
+    if (OFFBASE_IS_OFFSET(offBase)) {  /* full offset */
+        rep[2] = rep[1];
+        rep[1] = rep[0];
+        rep[0] = OFFBASE_TO_OFFSET(offBase);
+    } else {   /* repcode */
+        U32 const repCode = OFFBASE_TO_REPCODE(offBase) - 1 + ll0;
+        if (repCode > 0) {  /* note : if repCode==0, no change */
+            U32 const currentOffset = (repCode==ZSTD_REP_NUM) ? (rep[0] - 1) : rep[repCode];
+            rep[2] = (repCode >= 2) ? rep[1] : rep[2];
+            rep[1] = rep[0];
+            rep[0] = currentOffset;
+        } else {   /* repCode == 0 */
+            /* nothing to do */
         }
-    } else {  /* Big Endian CPU */
-        if (MEM_64bits()) {
-#       if defined(_MSC_VER) && defined(_WIN64)
-            unsigned long r = 0;
-            return _BitScanReverse64( &r, val ) ? (unsigned)(r >> 3) : 0;
-#       elif defined(__GNUC__) && (__GNUC__ >= 4)
-            return (__builtin_clzll(val) >> 3);
-#       else
-            unsigned r;
-            const unsigned n32 = sizeof(size_t)*4;   /* calculate this way due to compiler complaining in 32-bits mode */
-            if (!(val>>n32)) { r=4; } else { r=0; val>>=n32; }
-            if (!(val>>16)) { r+=2; val>>=8; } else { val>>=24; }
-            r += (!val);
-            return r;
-#       endif
-        } else { /* 32 bits */
-#       if defined(_MSC_VER)
-            unsigned long r = 0;
-            return _BitScanReverse( &r, (unsigned long)val ) ? (unsigned)(r >> 3) : 0;
-#       elif defined(__GNUC__) && (__GNUC__ >= 3)
-            return (__builtin_clz((U32)val) >> 3);
-#       else
-            unsigned r;
-            if (!(val>>16)) { r=2; val>>=8; } else { r=0; val>>=24; }
-            r += (!val);
-            return r;
-#       endif
-    }   }
+    }
+}
+
+typedef struct repcodes_s {
+    U32 rep[3];
+} repcodes_t;
+
+MEM_STATIC repcodes_t
+ZSTD_newRep(U32 const rep[ZSTD_REP_NUM], U32 const offBase, U32 const ll0)
+{
+    repcodes_t newReps;
+    ZSTD_memcpy(&newReps, rep, sizeof(newReps));
+    ZSTD_updateRep(newReps.rep, offBase, ll0);
+    return newReps;
 }
 
 
+/*-*************************************
+*  Match length counter
+***************************************/
 MEM_STATIC size_t ZSTD_count(const BYTE* pIn, const BYTE* pMatch, const BYTE* const pInLimit)
 {
     const BYTE* const pStart = pIn;
@@ -603,31 +787,36 @@ ZSTD_count_2segments(const BYTE* ip, const BYTE* match,
  *  Hashes
  ***************************************/
 static const U32 prime3bytes = 506832829U;
-static U32    ZSTD_hash3(U32 u, U32 h) { return ((u << (32-24)) * prime3bytes)  >> (32-h) ; }
+static U32    ZSTD_hash3(U32 u, U32 h) { assert(h <= 32); return ((u << (32-24)) * prime3bytes)  >> (32-h) ; }
 MEM_STATIC size_t ZSTD_hash3Ptr(const void* ptr, U32 h) { return ZSTD_hash3(MEM_readLE32(ptr), h); } /* only in zstd_opt.h */
 
 static const U32 prime4bytes = 2654435761U;
-static U32    ZSTD_hash4(U32 u, U32 h) { return (u * prime4bytes) >> (32-h) ; }
-static size_t ZSTD_hash4Ptr(const void* ptr, U32 h) { return ZSTD_hash4(MEM_read32(ptr), h); }
+static U32    ZSTD_hash4(U32 u, U32 h) { assert(h <= 32); return (u * prime4bytes) >> (32-h) ; }
+static size_t ZSTD_hash4Ptr(const void* ptr, U32 h) { return ZSTD_hash4(MEM_readLE32(ptr), h); }
 
 static const U64 prime5bytes = 889523592379ULL;
-static size_t ZSTD_hash5(U64 u, U32 h) { return (size_t)(((u  << (64-40)) * prime5bytes) >> (64-h)) ; }
+static size_t ZSTD_hash5(U64 u, U32 h) { assert(h <= 64); return (size_t)(((u  << (64-40)) * prime5bytes) >> (64-h)) ; }
 static size_t ZSTD_hash5Ptr(const void* p, U32 h) { return ZSTD_hash5(MEM_readLE64(p), h); }
 
 static const U64 prime6bytes = 227718039650203ULL;
-static size_t ZSTD_hash6(U64 u, U32 h) { return (size_t)(((u  << (64-48)) * prime6bytes) >> (64-h)) ; }
+static size_t ZSTD_hash6(U64 u, U32 h) { assert(h <= 64); return (size_t)(((u  << (64-48)) * prime6bytes) >> (64-h)) ; }
 static size_t ZSTD_hash6Ptr(const void* p, U32 h) { return ZSTD_hash6(MEM_readLE64(p), h); }
 
 static const U64 prime7bytes = 58295818150454627ULL;
-static size_t ZSTD_hash7(U64 u, U32 h) { return (size_t)(((u  << (64-56)) * prime7bytes) >> (64-h)) ; }
+static size_t ZSTD_hash7(U64 u, U32 h) { assert(h <= 64); return (size_t)(((u  << (64-56)) * prime7bytes) >> (64-h)) ; }
 static size_t ZSTD_hash7Ptr(const void* p, U32 h) { return ZSTD_hash7(MEM_readLE64(p), h); }
 
 static const U64 prime8bytes = 0xCF1BBCDCB7A56463ULL;
-static size_t ZSTD_hash8(U64 u, U32 h) { return (size_t)(((u) * prime8bytes) >> (64-h)) ; }
+static size_t ZSTD_hash8(U64 u, U32 h) { assert(h <= 64); return (size_t)(((u) * prime8bytes) >> (64-h)) ; }
 static size_t ZSTD_hash8Ptr(const void* p, U32 h) { return ZSTD_hash8(MEM_readLE64(p), h); }
 
-MEM_STATIC size_t ZSTD_hashPtr(const void* p, U32 hBits, U32 mls)
+MEM_STATIC FORCE_INLINE_ATTR
+size_t ZSTD_hashPtr(const void* p, U32 hBits, U32 mls)
 {
+    /* Although some of these hashes do support hBits up to 64, some do not.
+     * To be on the safe side, always avoid hBits > 32. */
+    assert(hBits <= 32);
+
     switch(mls)
     {
     default:
@@ -723,6 +912,13 @@ MEM_STATIC void ZSTD_window_clear(ZSTD_window_t* window)
     window->dictLimit = end;
 }
 
+MEM_STATIC U32 ZSTD_window_isEmpty(ZSTD_window_t const window)
+{
+    return window.dictLimit == ZSTD_WINDOW_START_INDEX &&
+           window.lowLimit == ZSTD_WINDOW_START_INDEX &&
+           (window.nextSrc - window.base) == ZSTD_WINDOW_START_INDEX;
+}
+
 /**
  * ZSTD_window_hasExtDict():
  * Returns non-zero if the window has a non-empty extDict.
@@ -742,20 +938,76 @@ MEM_STATIC ZSTD_dictMode_e ZSTD_matchState_dictMode(const ZSTD_matchState_t *ms)
     return ZSTD_window_hasExtDict(ms->window) ?
         ZSTD_extDict :
         ms->dictMatchState != NULL ?
-            ZSTD_dictMatchState :
+            (ms->dictMatchState->dedicatedDictSearch ? ZSTD_dedicatedDictSearch : ZSTD_dictMatchState) :
             ZSTD_noDict;
 }
 
+/* Defining this macro to non-zero tells zstd to run the overflow correction
+ * code much more frequently. This is very inefficient, and should only be
+ * used for tests and fuzzers.
+ */
+#ifndef ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY
+#  ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
+#    define ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY 1
+#  else
+#    define ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY 0
+#  endif
+#endif
+
+/**
+ * ZSTD_window_canOverflowCorrect():
+ * Returns non-zero if the indices are large enough for overflow correction
+ * to work correctly without impacting compression ratio.
+ */
+MEM_STATIC U32 ZSTD_window_canOverflowCorrect(ZSTD_window_t const window,
+                                              U32 cycleLog,
+                                              U32 maxDist,
+                                              U32 loadedDictEnd,
+                                              void const* src)
+{
+    U32 const cycleSize = 1u << cycleLog;
+    U32 const curr = (U32)((BYTE const*)src - window.base);
+    U32 const minIndexToOverflowCorrect = cycleSize
+                                        + MAX(maxDist, cycleSize)
+                                        + ZSTD_WINDOW_START_INDEX;
+
+    /* Adjust the min index to backoff the overflow correction frequency,
+     * so we don't waste too much CPU in overflow correction. If this
+     * computation overflows we don't really care, we just need to make
+     * sure it is at least minIndexToOverflowCorrect.
+     */
+    U32 const adjustment = window.nbOverflowCorrections + 1;
+    U32 const adjustedIndex = MAX(minIndexToOverflowCorrect * adjustment,
+                                  minIndexToOverflowCorrect);
+    U32 const indexLargeEnough = curr > adjustedIndex;
+
+    /* Only overflow correct early if the dictionary is invalidated already,
+     * so we don't hurt compression ratio.
+     */
+    U32 const dictionaryInvalidated = curr > maxDist + loadedDictEnd;
+
+    return indexLargeEnough && dictionaryInvalidated;
+}
+
 /**
  * ZSTD_window_needOverflowCorrection():
  * Returns non-zero if the indices are getting too large and need overflow
  * protection.
  */
 MEM_STATIC U32 ZSTD_window_needOverflowCorrection(ZSTD_window_t const window,
+                                                  U32 cycleLog,
+                                                  U32 maxDist,
+                                                  U32 loadedDictEnd,
+                                                  void const* src,
                                                   void const* srcEnd)
 {
-    U32 const current = (U32)((BYTE const*)srcEnd - window.base);
-    return current > ZSTD_CURRENT_MAX;
+    U32 const curr = (U32)((BYTE const*)srcEnd - window.base);
+    if (ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY) {
+        if (ZSTD_window_canOverflowCorrect(window, cycleLog, maxDist, loadedDictEnd, src)) {
+            return 1;
+        }
+    }
+    return curr > ZSTD_CURRENT_MAX;
 }
 
 /**
@@ -766,7 +1018,6 @@ MEM_STATIC U32 ZSTD_window_needOverflowCorrection(ZSTD_window_t const window,
  *
  * The least significant cycleLog bits of the indices must remain the same,
  * which may be 0. Every index up to maxDist in the past must be valid.
- * NOTE: (maxDist & cycleMask) must be zero.
  */
 MEM_STATIC U32 ZSTD_window_correctOverflow(ZSTD_window_t* window, U32 cycleLog,
                                            U32 maxDist, void const* src)
@@ -790,32 +1041,52 @@ MEM_STATIC U32 ZSTD_window_correctOverflow(ZSTD_window_t* window, U32 cycleLog,
      * 3. (cctx->lowLimit + 1<<windowLog) < 1<<32:
      *    windowLog <= 31 ==> 3<<29 + 1<<windowLog < 7<<29 < 1<<32.
      */
-    U32 const cycleMask = (1U << cycleLog) - 1;
-    U32 const current = (U32)((BYTE const*)src - window->base);
-    U32 const currentCycle0 = current & cycleMask;
-    /* Exclude zero so that newCurrent - maxDist >= 1. */
-    U32 const currentCycle1 = currentCycle0 == 0 ? (1U << cycleLog) : currentCycle0;
-    U32 const newCurrent = currentCycle1 + maxDist;
-    U32 const correction = current - newCurrent;
-    assert((maxDist & cycleMask) == 0);
-    assert(current > newCurrent);
-    /* Loose bound, should be around 1<<29 (see above) */
-    assert(correction > 1<<28);
+    U32 const cycleSize = 1u << cycleLog;
+    U32 const cycleMask = cycleSize - 1;
+    U32 const curr = (U32)((BYTE const*)src - window->base);
+    U32 const currentCycle = curr & cycleMask;
+    /* Ensure newCurrent - maxDist >= ZSTD_WINDOW_START_INDEX. */
+    U32 const currentCycleCorrection = currentCycle < ZSTD_WINDOW_START_INDEX
+                                     ? MAX(cycleSize, ZSTD_WINDOW_START_INDEX)
+                                     : 0;
+    U32 const newCurrent = currentCycle
+                         + currentCycleCorrection
+                         + MAX(maxDist, cycleSize);
+    U32 const correction = curr - newCurrent;
+    /* maxDist must be a power of two so that:
+     *   (newCurrent & cycleMask) == (curr & cycleMask)
+     * This is required to not corrupt the chains / binary tree.
+     */
+    assert((maxDist & (maxDist - 1)) == 0);
+    assert((curr & cycleMask) == (newCurrent & cycleMask));
+    assert(curr > newCurrent);
+    if (!ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY) {
+        /* Loose bound, should be around 1<<29 (see above) */
+        assert(correction > 1<<28);
+    }
 
     window->base += correction;
     window->dictBase += correction;
-    if (window->lowLimit <= correction) window->lowLimit = 1;
-    else window->lowLimit -= correction;
-    if (window->dictLimit <= correction) window->dictLimit = 1;
-    else window->dictLimit -= correction;
+    if (window->lowLimit < correction + ZSTD_WINDOW_START_INDEX) {
+        window->lowLimit = ZSTD_WINDOW_START_INDEX;
+    } else {
+        window->lowLimit -= correction;
+    }
+    if (window->dictLimit < correction + ZSTD_WINDOW_START_INDEX) {
+        window->dictLimit = ZSTD_WINDOW_START_INDEX;
+    } else {
+        window->dictLimit -= correction;
+    }
 
     /* Ensure we can still reference the full window. */
     assert(newCurrent >= maxDist);
-    assert(newCurrent - maxDist >= 1);
+    assert(newCurrent - maxDist >= ZSTD_WINDOW_START_INDEX);
     /* Ensure that lowLimit and dictLimit didn't underflow. */
     assert(window->lowLimit <= newCurrent);
     assert(window->dictLimit <= newCurrent);
 
+    ++window->nbOverflowCorrections;
+
     DEBUGLOG(4, "Correction of 0x%x bytes to lowLimit=0x%x", correction,
              window->lowLimit);
     return correction;
@@ -904,10 +1175,15 @@ ZSTD_checkDictValidity(const ZSTD_window_t* window,
                     (unsigned)blockEndIdx, (unsigned)maxDist, (unsigned)loadedDictEnd);
         assert(blockEndIdx >= loadedDictEnd);
 
-        if (blockEndIdx > loadedDictEnd + maxDist) {
+        if (blockEndIdx > loadedDictEnd + maxDist || loadedDictEnd != window->dictLimit) {
             /* On reaching window size, dictionaries are invalidated.
              * For simplification, if window size is reached anywhere within next block,
              * the dictionary is invalidated for the full block.
+             *
+             * We also have to invalidate the dictionary if ZSTD_window_update() has detected
+             * non-contiguous segments, which means that loadedDictEnd != window->dictLimit.
+             * loadedDictEnd may be 0, if forceWindow is true, but in that case we never use
+             * dictMatchState, so setting it to NULL is not a problem.
              */
             DEBUGLOG(6, "invalidating dictionary for current block (distance > windowSize)");
             *loadedDictEndPtr = 0;
@@ -919,12 +1195,14 @@ ZSTD_checkDictValidity(const ZSTD_window_t* window,
 }
 
 MEM_STATIC void ZSTD_window_init(ZSTD_window_t* window) {
-    memset(window, 0, sizeof(*window));
-    window->base = (BYTE const*)"";
-    window->dictBase = (BYTE const*)"";
-    window->dictLimit = 1;    /* start from 1, so that 1st position is valid */
-    window->lowLimit = 1;     /* it ensures first and later CCtx usages compress the same */
-    window->nextSrc = window->base + 1;   /* see issue #1241 */
+    ZSTD_memset(window, 0, sizeof(*window));
+    window->base = (BYTE const*)" ";
+    window->dictBase = (BYTE const*)" ";
+    ZSTD_STATIC_ASSERT(ZSTD_DUBT_UNSORTED_MARK < ZSTD_WINDOW_START_INDEX); /* Start above ZSTD_DUBT_UNSORTED_MARK */
+    window->dictLimit = ZSTD_WINDOW_START_INDEX;    /* start from >0, so that 1st position is valid */
+    window->lowLimit = ZSTD_WINDOW_START_INDEX;     /* it ensures first and later CCtx usages compress the same */
+    window->nextSrc = window->base + ZSTD_WINDOW_START_INDEX;   /* see issue #1241 */
+    window->nbOverflowCorrections = 0;
 }
 
 /**
@@ -935,7 +1213,8 @@ MEM_STATIC void ZSTD_window_init(ZSTD_window_t* window) {
  * Returns non-zero if the segment is contiguous.
  */
 MEM_STATIC U32 ZSTD_window_update(ZSTD_window_t* window,
-                                  void const* src, size_t srcSize)
+                                  void const* src, size_t srcSize,
+                                  int forceNonContiguous)
 {
     BYTE const* const ip = (BYTE const*)src;
     U32 contiguous = 1;
@@ -945,7 +1224,7 @@ MEM_STATIC U32 ZSTD_window_update(ZSTD_window_t* window,
     assert(window->base != NULL);
     assert(window->dictBase != NULL);
     /* Check if blocks follow each other */
-    if (src != window->nextSrc) {
+    if (src != window->nextSrc || forceNonContiguous) {
         /* not contiguous */
         size_t const distanceFromBase = (size_t)(window->nextSrc - window->base);
         DEBUGLOG(5, "Non contiguous blocks, new segment starts at %u", window->dictLimit);
@@ -973,25 +1252,32 @@ MEM_STATIC U32 ZSTD_window_update(ZSTD_window_t* window,
 /**
  * Returns the lowest allowed match index. It may either be in the ext-dict or the prefix.
  */
-MEM_STATIC U32 ZSTD_getLowestMatchIndex(const ZSTD_matchState_t* ms, U32 current, unsigned windowLog)
+MEM_STATIC U32 ZSTD_getLowestMatchIndex(const ZSTD_matchState_t* ms, U32 curr, unsigned windowLog)
 {
-    U32    const maxDistance = 1U << windowLog;
-    U32    const lowestValid = ms->window.lowLimit;
-    U32    const withinWindow = (current - lowestValid > maxDistance) ? current - maxDistance : lowestValid;
-    U32    const isDictionary = (ms->loadedDictEnd != 0);
-    U32    const matchLowest = isDictionary ? lowestValid : withinWindow;
+    U32 const maxDistance = 1U << windowLog;
+    U32 const lowestValid = ms->window.lowLimit;
+    U32 const withinWindow = (curr - lowestValid > maxDistance) ? curr - maxDistance : lowestValid;
+    U32 const isDictionary = (ms->loadedDictEnd != 0);
+    /* When using a dictionary the entire dictionary is valid if a single byte of the dictionary
+     * is within the window. We invalidate the dictionary (and set loadedDictEnd to 0) when it isn't
+     * valid for the entire block. So this check is sufficient to find the lowest valid match index.
+     */
+    U32 const matchLowest = isDictionary ? lowestValid : withinWindow;
     return matchLowest;
 }
 
 /**
  * Returns the lowest allowed match index in the prefix.
  */
-MEM_STATIC U32 ZSTD_getLowestPrefixIndex(const ZSTD_matchState_t* ms, U32 current, unsigned windowLog)
+MEM_STATIC U32 ZSTD_getLowestPrefixIndex(const ZSTD_matchState_t* ms, U32 curr, unsigned windowLog)
 {
     U32    const maxDistance = 1U << windowLog;
     U32    const lowestValid = ms->window.dictLimit;
-    U32    const withinWindow = (current - lowestValid > maxDistance) ? current - maxDistance : lowestValid;
+    U32    const withinWindow = (curr - lowestValid > maxDistance) ? curr - maxDistance : lowestValid;
     U32    const isDictionary = (ms->loadedDictEnd != 0);
+    /* When computing the lowest prefix index we need to take the dictionary into account to handle
+     * the edge case where the dictionary and the source are contiguous in memory.
+     */
     U32    const matchLowest = isDictionary ? lowestValid : withinWindow;
     return matchLowest;
 }
@@ -1029,6 +1315,42 @@ MEM_STATIC void ZSTD_debugTable(const U32* table, U32 max)
 
 #endif
 
+/* Short Cache */
+
+/* Normally, zstd matchfinders follow this flow:
+ *     1. Compute hash at ip
+ *     2. Load index from hashTable[hash]
+ *     3. Check if *ip == *(base + index)
+ * In dictionary compression, loading *(base + index) is often an L2 or even L3 miss.
+ *
+ * Short cache is an optimization which allows us to avoid step 3 most of the time
+ * when the data doesn't actually match. With short cache, the flow becomes:
+ *     1. Compute (hash, currentTag) at ip. currentTag is an 8-bit independent hash at ip.
+ *     2. Load (index, matchTag) from hashTable[hash]. See ZSTD_writeTaggedIndex to understand how this works.
+ *     3. Only if currentTag == matchTag, check *ip == *(base + index). Otherwise, continue.
+ *
+ * Currently, short cache is only implemented in CDict hashtables. Thus, its use is limited to
+ * dictMatchState matchfinders.
+ */
+#define ZSTD_SHORT_CACHE_TAG_BITS 8
+#define ZSTD_SHORT_CACHE_TAG_MASK ((1u << ZSTD_SHORT_CACHE_TAG_BITS) - 1)
+
+/* Helper function for ZSTD_fillHashTable and ZSTD_fillDoubleHashTable.
+ * Unpacks hashAndTag into (hash, tag), then packs (index, tag) into hashTable[hash]. */
+MEM_STATIC void ZSTD_writeTaggedIndex(U32* const hashTable, size_t hashAndTag, U32 index) {
+    size_t const hash = hashAndTag >> ZSTD_SHORT_CACHE_TAG_BITS;
+    U32 const tag = (U32)(hashAndTag & ZSTD_SHORT_CACHE_TAG_MASK);
+    assert(index >> (32 - ZSTD_SHORT_CACHE_TAG_BITS) == 0);
+    hashTable[hash] = (index << ZSTD_SHORT_CACHE_TAG_BITS) | tag;
+}
+
+/* Helper function for short cache matchfinders.
+ * Unpacks tag1 and tag2 from lower bits of packedTag1 and packedTag2, then checks if the tags match. */
+MEM_STATIC int ZSTD_comparePackedTags(size_t packedTag1, size_t packedTag2) {
+    U32 const tag1 = packedTag1 & ZSTD_SHORT_CACHE_TAG_MASK;
+    U32 const tag2 = packedTag2 & ZSTD_SHORT_CACHE_TAG_MASK;
+    return tag1 == tag2;
+}
 
 #if defined (__cplusplus)
 }
@@ -1045,7 +1367,6 @@ MEM_STATIC void ZSTD_debugTable(const U32* table, U32 max)
  * assumptions : magic number supposed already checked
  *               and dictSize >= 8 */
 size_t ZSTD_loadCEntropy(ZSTD_compressedBlockState_t* bs, void* workspace,
-                         short* offcodeNCount, unsigned* offcodeMaxValue,
                          const void* const dict, size_t dictSize);
 
 void ZSTD_reset_compressedBlockState(ZSTD_compressedBlockState_t* bs);
@@ -1061,7 +1382,7 @@ void ZSTD_reset_compressedBlockState(ZSTD_compressedBlockState_t* bs);
  * Note: srcSizeHint == 0 means 0!
  */
 ZSTD_compressionParameters ZSTD_getCParamsFromCCtxParams(
-        const ZSTD_CCtx_params* CCtxParams, U64 srcSizeHint, size_t dictSize);
+        const ZSTD_CCtx_params* CCtxParams, U64 srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode);
 
 /*! ZSTD_initCStream_internal() :
  *  Private use only. Init streaming operation.
@@ -1122,4 +1443,36 @@ size_t ZSTD_referenceExternalSequences(ZSTD_CCtx* cctx, rawSeq* seq, size_t nbSe
  *  condition for correct operation : hashLog > 1 */
 U32 ZSTD_cycleLog(U32 hashLog, ZSTD_strategy strat);
 
+/** ZSTD_CCtx_trace() :
+ *  Trace the end of a compression call.
+ */
+void ZSTD_CCtx_trace(ZSTD_CCtx* cctx, size_t extraCSize);
+
+/* Returns 0 on success, and a ZSTD_error otherwise. This function scans through an array of
+ * ZSTD_Sequence, storing the sequences it finds, until it reaches a block delimiter.
+ * Note that the block delimiter must include the last literals of the block.
+ */
+size_t
+ZSTD_copySequencesToSeqStoreExplicitBlockDelim(ZSTD_CCtx* cctx,
+                                              ZSTD_sequencePosition* seqPos,
+                                        const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
+                                        const void* src, size_t blockSize, ZSTD_paramSwitch_e externalRepSearch);
+
+/* Returns the number of bytes to move the current read position back by.
+ * Only non-zero if we ended up splitting a sequence.
+ * Otherwise, it may return a ZSTD error if something went wrong.
+ *
+ * This function will attempt to scan through blockSize bytes
+ * represented by the sequences in @inSeqs,
+ * storing any (partial) sequences.
+ *
+ * Occasionally, we may want to change the actual number of bytes we consumed from inSeqs to
+ * avoid splitting a match, or to avoid splitting a match such that it would produce a match
+ * smaller than MINMATCH. In this case, we return the number of bytes that we didn't read from this block.
+ */
+size_t
+ZSTD_copySequencesToSeqStoreNoBlockDelim(ZSTD_CCtx* cctx, ZSTD_sequencePosition* seqPos,
+                                   const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
+                                   const void* src, size_t blockSize, ZSTD_paramSwitch_e externalRepSearch);
+
 #endif /* ZSTD_COMPRESS_H */
diff --git a/contrib/zstd/zstd_compress_literals.c b/contrib/zstd/zstd_compress_literals.c
index 17e7168d8..bfd4f11ab 100644
--- a/contrib/zstd/zstd_compress_literals.c
+++ b/contrib/zstd/zstd_compress_literals.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -13,11 +13,36 @@
  ***************************************/
 #include "zstd_compress_literals.h"
 
+
+/* **************************************************************
+*  Debug Traces
+****************************************************************/
+#if DEBUGLEVEL >= 2
+
+static size_t showHexa(const void* src, size_t srcSize)
+{
+    const BYTE* const ip = (const BYTE*)src;
+    size_t u;
+    for (u=0; u<srcSize; u++) {
+        RAWLOG(5, " %02X", ip[u]); (void)ip;
+    }
+    RAWLOG(5, " \n");
+    return srcSize;
+}
+
+#endif
+
+
+/* **************************************************************
+*  Literals compression - special cases
+****************************************************************/
 size_t ZSTD_noCompressLiterals (void* dst, size_t dstCapacity, const void* src, size_t srcSize)
 {
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     U32   const flSize = 1 + (srcSize>31) + (srcSize>4095);
 
+    DEBUGLOG(5, "ZSTD_noCompressLiterals: srcSize=%zu, dstCapacity=%zu", srcSize, dstCapacity);
+
     RETURN_ERROR_IF(srcSize + flSize > dstCapacity, dstSize_tooSmall, "");
 
     switch(flSize)
@@ -35,17 +60,31 @@ size_t ZSTD_noCompressLiterals (void* dst, size_t dstCapacity, const void* src,
             assert(0);
     }
 
-    memcpy(ostart + flSize, src, srcSize);
-    DEBUGLOG(5, "Raw literals: %u -> %u", (U32)srcSize, (U32)(srcSize + flSize));
+    ZSTD_memcpy(ostart + flSize, src, srcSize);
+    DEBUGLOG(5, "Raw (uncompressed) literals: %u -> %u", (U32)srcSize, (U32)(srcSize + flSize));
     return srcSize + flSize;
 }
 
+static int allBytesIdentical(const void* src, size_t srcSize)
+{
+    assert(srcSize >= 1);
+    assert(src != NULL);
+    {   const BYTE b = ((const BYTE*)src)[0];
+        size_t p;
+        for (p=1; p<srcSize; p++) {
+            if (((const BYTE*)src)[p] != b) return 0;
+        }
+        return 1;
+    }
+}
+
 size_t ZSTD_compressRleLiteralsBlock (void* dst, size_t dstCapacity, const void* src, size_t srcSize)
 {
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     U32   const flSize = 1 + (srcSize>31) + (srcSize>4095);
 
-    (void)dstCapacity;  /* dstCapacity already guaranteed to be >=4, hence large enough */
+    assert(dstCapacity >= 4); (void)dstCapacity;
+    assert(allBytesIdentical(src, srcSize));
 
     switch(flSize)
     {
@@ -63,68 +102,103 @@ size_t ZSTD_compressRleLiteralsBlock (void* dst, size_t dstCapacity, const void*
     }
 
     ostart[flSize] = *(const BYTE*)src;
-    DEBUGLOG(5, "RLE literals: %u -> %u", (U32)srcSize, (U32)flSize + 1);
+    DEBUGLOG(5, "RLE : Repeated Literal (%02X: %u times) -> %u bytes encoded", ((const BYTE*)src)[0], (U32)srcSize, (U32)flSize + 1);
     return flSize+1;
 }
 
-size_t ZSTD_compressLiterals (ZSTD_hufCTables_t const* prevHuf,
-                              ZSTD_hufCTables_t* nextHuf,
-                              ZSTD_strategy strategy, int disableLiteralCompression,
-                              void* dst, size_t dstCapacity,
-                        const void* src, size_t srcSize,
-                              void* entropyWorkspace, size_t entropyWorkspaceSize,
-                        const int bmi2)
+/* ZSTD_minLiteralsToCompress() :
+ * returns minimal amount of literals
+ * for literal compression to even be attempted.
+ * Minimum is made tighter as compression strategy increases.
+ */
+static size_t
+ZSTD_minLiteralsToCompress(ZSTD_strategy strategy, HUF_repeat huf_repeat)
+{
+    assert((int)strategy >= 0);
+    assert((int)strategy <= 9);
+    /* btultra2 : min 8 bytes;
+     * then 2x larger for each successive compression strategy
+     * max threshold 64 bytes */
+    {   int const shift = MIN(9-(int)strategy, 3);
+        size_t const mintc = (huf_repeat == HUF_repeat_valid) ? 6 : (size_t)8 << shift;
+        DEBUGLOG(7, "minLiteralsToCompress = %zu", mintc);
+        return mintc;
+    }
+}
+
+size_t ZSTD_compressLiterals (
+                  void* dst, size_t dstCapacity,
+            const void* src, size_t srcSize,
+                  void* entropyWorkspace, size_t entropyWorkspaceSize,
+            const ZSTD_hufCTables_t* prevHuf,
+                  ZSTD_hufCTables_t* nextHuf,
+                  ZSTD_strategy strategy,
+                  int disableLiteralCompression,
+                  int suspectUncompressible,
+                  int bmi2)
 {
-    size_t const minGain = ZSTD_minGain(srcSize, strategy);
     size_t const lhSize = 3 + (srcSize >= 1 KB) + (srcSize >= 16 KB);
     BYTE*  const ostart = (BYTE*)dst;
     U32 singleStream = srcSize < 256;
     symbolEncodingType_e hType = set_compressed;
     size_t cLitSize;
 
-    DEBUGLOG(5,"ZSTD_compressLiterals (disableLiteralCompression=%i srcSize=%u)",
-                disableLiteralCompression, (U32)srcSize);
+    DEBUGLOG(5,"ZSTD_compressLiterals (disableLiteralCompression=%i, srcSize=%u, dstCapacity=%zu)",
+                disableLiteralCompression, (U32)srcSize, dstCapacity);
+
+    DEBUGLOG(6, "Completed literals listing (%zu bytes)", showHexa(src, srcSize));
 
     /* Prepare nextEntropy assuming reusing the existing table */
-    memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
+    ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
 
     if (disableLiteralCompression)
         return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
 
-    /* small ? don't even attempt compression (speed opt) */
-#   define COMPRESS_LITERALS_SIZE_MIN 63
-    {   size_t const minLitSize = (prevHuf->repeatMode == HUF_repeat_valid) ? 6 : COMPRESS_LITERALS_SIZE_MIN;
-        if (srcSize <= minLitSize) return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
-    }
+    /* if too small, don't even attempt compression (speed opt) */
+    if (srcSize < ZSTD_minLiteralsToCompress(strategy, prevHuf->repeatMode))
+        return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
 
     RETURN_ERROR_IF(dstCapacity < lhSize+1, dstSize_tooSmall, "not enough space for compression");
     {   HUF_repeat repeat = prevHuf->repeatMode;
-        int const preferRepeat = strategy < ZSTD_lazy ? srcSize <= 1024 : 0;
+        int const flags = 0
+            | (bmi2 ? HUF_flags_bmi2 : 0)
+            | (strategy < ZSTD_lazy && srcSize <= 1024 ? HUF_flags_preferRepeat : 0)
+            | (strategy >= HUF_OPTIMAL_DEPTH_THRESHOLD ? HUF_flags_optimalDepth : 0)
+            | (suspectUncompressible ? HUF_flags_suspectUncompressible : 0);
+
+        typedef size_t (*huf_compress_f)(void*, size_t, const void*, size_t, unsigned, unsigned, void*, size_t, HUF_CElt*, HUF_repeat*, int);
+        huf_compress_f huf_compress;
         if (repeat == HUF_repeat_valid && lhSize == 3) singleStream = 1;
-        cLitSize = singleStream ?
-            HUF_compress1X_repeat(
-                ostart+lhSize, dstCapacity-lhSize, src, srcSize,
-                HUF_SYMBOLVALUE_MAX, HUF_TABLELOG_DEFAULT, entropyWorkspace, entropyWorkspaceSize,
-                (HUF_CElt*)nextHuf->CTable, &repeat, preferRepeat, bmi2) :
-            HUF_compress4X_repeat(
-                ostart+lhSize, dstCapacity-lhSize, src, srcSize,
-                HUF_SYMBOLVALUE_MAX, HUF_TABLELOG_DEFAULT, entropyWorkspace, entropyWorkspaceSize,
-                (HUF_CElt*)nextHuf->CTable, &repeat, preferRepeat, bmi2);
+        huf_compress = singleStream ? HUF_compress1X_repeat : HUF_compress4X_repeat;
+        cLitSize = huf_compress(ostart+lhSize, dstCapacity-lhSize,
+                                src, srcSize,
+                                HUF_SYMBOLVALUE_MAX, LitHufLog,
+                                entropyWorkspace, entropyWorkspaceSize,
+                                (HUF_CElt*)nextHuf->CTable,
+                                &repeat, flags);
+        DEBUGLOG(5, "%zu literals compressed into %zu bytes (before header)", srcSize, cLitSize);
         if (repeat != HUF_repeat_none) {
             /* reused the existing table */
-            DEBUGLOG(5, "Reusing previous huffman table");
+            DEBUGLOG(5, "reusing statistics from previous huffman block");
             hType = set_repeat;
         }
     }
 
-    if ((cLitSize==0) | (cLitSize >= srcSize - minGain) | ERR_isError(cLitSize)) {
-        memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
-        return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
-    }
+    {   size_t const minGain = ZSTD_minGain(srcSize, strategy);
+        if ((cLitSize==0) || (cLitSize >= srcSize - minGain) || ERR_isError(cLitSize)) {
+            ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
+            return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
+    }   }
     if (cLitSize==1) {
-        memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
-        return ZSTD_compressRleLiteralsBlock(dst, dstCapacity, src, srcSize);
-    }
+        /* A return value of 1 signals that the alphabet consists of a single symbol.
+         * However, in some rare circumstances, it could be the compressed size (a single byte).
+         * For that outcome to have a chance to happen, it's necessary that `srcSize < 8`.
+         * (it's also necessary to not generate statistics).
+         * Therefore, in such a case, actively check that all bytes are identical. */
+        if ((srcSize >= 8) || allBytesIdentical(src, srcSize)) {
+            ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
+            return ZSTD_compressRleLiteralsBlock(dst, dstCapacity, src, srcSize);
+    }   }
 
     if (hType == set_compressed) {
         /* using a newly constructed table */
@@ -135,16 +209,19 @@ size_t ZSTD_compressLiterals (ZSTD_hufCTables_t const* prevHuf,
     switch(lhSize)
     {
     case 3: /* 2 - 2 - 10 - 10 */
-        {   U32 const lhc = hType + ((!singleStream) << 2) + ((U32)srcSize<<4) + ((U32)cLitSize<<14);
+        if (!singleStream) assert(srcSize >= MIN_LITERALS_FOR_4_STREAMS);
+        {   U32 const lhc = hType + ((U32)(!singleStream) << 2) + ((U32)srcSize<<4) + ((U32)cLitSize<<14);
             MEM_writeLE24(ostart, lhc);
             break;
         }
     case 4: /* 2 - 2 - 14 - 14 */
+        assert(srcSize >= MIN_LITERALS_FOR_4_STREAMS);
         {   U32 const lhc = hType + (2 << 2) + ((U32)srcSize<<4) + ((U32)cLitSize<<18);
             MEM_writeLE32(ostart, lhc);
             break;
         }
     case 5: /* 2 - 2 - 18 - 18 */
+        assert(srcSize >= MIN_LITERALS_FOR_4_STREAMS);
         {   U32 const lhc = hType + (3 << 2) + ((U32)srcSize<<4) + ((U32)cLitSize<<22);
             MEM_writeLE32(ostart, lhc);
             ostart[4] = (BYTE)(cLitSize >> 10);
diff --git a/contrib/zstd/zstd_compress_literals.h b/contrib/zstd/zstd_compress_literals.h
index 8b0870574..b060c8ad2 100644
--- a/contrib/zstd/zstd_compress_literals.h
+++ b/contrib/zstd/zstd_compress_literals.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -16,14 +16,24 @@
 
 size_t ZSTD_noCompressLiterals (void* dst, size_t dstCapacity, const void* src, size_t srcSize);
 
+/* ZSTD_compressRleLiteralsBlock() :
+ * Conditions :
+ * - All bytes in @src are identical
+ * - dstCapacity >= 4 */
 size_t ZSTD_compressRleLiteralsBlock (void* dst, size_t dstCapacity, const void* src, size_t srcSize);
 
-size_t ZSTD_compressLiterals (ZSTD_hufCTables_t const* prevHuf,
-                              ZSTD_hufCTables_t* nextHuf,
-                              ZSTD_strategy strategy, int disableLiteralCompression,
-                              void* dst, size_t dstCapacity,
+/* ZSTD_compressLiterals():
+ * @entropyWorkspace: must be aligned on 4-bytes boundaries
+ * @entropyWorkspaceSize : must be >= HUF_WORKSPACE_SIZE
+ * @suspectUncompressible: sampling checks, to potentially skip huffman coding
+ */
+size_t ZSTD_compressLiterals (void* dst, size_t dstCapacity,
                         const void* src, size_t srcSize,
                               void* entropyWorkspace, size_t entropyWorkspaceSize,
-                        const int bmi2);
+                        const ZSTD_hufCTables_t* prevHuf,
+                              ZSTD_hufCTables_t* nextHuf,
+                              ZSTD_strategy strategy, int disableLiteralCompression,
+                              int suspectUncompressible,
+                              int bmi2);
 
 #endif /* ZSTD_COMPRESS_LITERALS_H */
diff --git a/contrib/zstd/zstd_compress_sequences.c b/contrib/zstd/zstd_compress_sequences.c
index f9f8097c8..8872d4d35 100644
--- a/contrib/zstd/zstd_compress_sequences.c
+++ b/contrib/zstd/zstd_compress_sequences.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -51,6 +51,19 @@ static unsigned ZSTD_getFSEMaxSymbolValue(FSE_CTable const* ctable) {
 }
 
 /**
+ * Returns true if we should use ncount=-1 else we should
+ * use ncount=1 for low probability symbols instead.
+ */
+static unsigned ZSTD_useLowProbCount(size_t const nbSeq)
+{
+    /* Heuristic: This should cover most blocks <= 16K and
+     * start to fade out after 16K to about 32K depending on
+     * compressibility.
+     */
+    return nbSeq >= 2048;
+}
+
+/**
  * Returns the cost in bytes of encoding the normalized count header.
  * Returns an error if any of the helper functions return an error.
  */
@@ -60,7 +73,7 @@ static size_t ZSTD_NCountCost(unsigned const* count, unsigned const max,
     BYTE wksp[FSE_NCOUNTBOUND];
     S16 norm[MaxSeq + 1];
     const U32 tableLog = FSE_optimalTableLog(FSELog, nbSeq, max);
-    FORWARD_IF_ERROR(FSE_normalizeCount(norm, tableLog, count, nbSeq, max), "");
+    FORWARD_IF_ERROR(FSE_normalizeCount(norm, tableLog, count, nbSeq, max, ZSTD_useLowProbCount(nbSeq)), "");
     return FSE_writeNCount(wksp, sizeof(wksp), norm, max, tableLog);
 }
 
@@ -72,6 +85,8 @@ static size_t ZSTD_entropyCost(unsigned const* count, unsigned const max, size_t
 {
     unsigned cost = 0;
     unsigned s;
+
+    assert(total > 0);
     for (s = 0; s <= max; ++s) {
         unsigned norm = (unsigned)((256 * count[s]) / total);
         if (count[s] != 0 && norm == 0)
@@ -151,7 +166,7 @@ ZSTD_selectEncodingType(
     if (mostFrequent == nbSeq) {
         *repeatMode = FSE_repeat_none;
         if (isDefaultAllowed && nbSeq <= 2) {
-            /* Prefer set_basic over set_rle when there are 2 or less symbols,
+            /* Prefer set_basic over set_rle when there are 2 or fewer symbols,
              * since RLE uses 1 byte, but set_basic uses 5-6 bits per symbol.
              * If basic encoding isn't possible, always choose RLE.
              */
@@ -219,6 +234,11 @@ ZSTD_selectEncodingType(
     return set_compressed;
 }
 
+typedef struct {
+    S16 norm[MaxSeq + 1];
+    U32 wksp[FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(MaxSeq, MaxFSELog)];
+} ZSTD_BuildCTableWksp;
+
 size_t
 ZSTD_buildCTable(void* dst, size_t dstCapacity,
                 FSE_CTable* nextCTable, U32 FSELog, symbolEncodingType_e type,
@@ -239,13 +259,13 @@ ZSTD_buildCTable(void* dst, size_t dstCapacity,
         *op = codeTable[0];
         return 1;
     case set_repeat:
-        memcpy(nextCTable, prevCTable, prevCTableSize);
+        ZSTD_memcpy(nextCTable, prevCTable, prevCTableSize);
         return 0;
     case set_basic:
         FORWARD_IF_ERROR(FSE_buildCTable_wksp(nextCTable, defaultNorm, defaultMax, defaultNormLog, entropyWorkspace, entropyWorkspaceSize), "");  /* note : could be pre-calculated */
         return 0;
     case set_compressed: {
-        S16 norm[MaxSeq + 1];
+        ZSTD_BuildCTableWksp* wksp = (ZSTD_BuildCTableWksp*)entropyWorkspace;
         size_t nbSeq_1 = nbSeq;
         const U32 tableLog = FSE_optimalTableLog(FSELog, nbSeq, max);
         if (count[codeTable[nbSeq-1]] > 1) {
@@ -253,10 +273,13 @@ ZSTD_buildCTable(void* dst, size_t dstCapacity,
             nbSeq_1--;
         }
         assert(nbSeq_1 > 1);
-        FORWARD_IF_ERROR(FSE_normalizeCount(norm, tableLog, count, nbSeq_1, max), "");
-        {   size_t const NCountSize = FSE_writeNCount(op, oend - op, norm, max, tableLog);   /* overflow protected */
+        assert(entropyWorkspaceSize >= sizeof(ZSTD_BuildCTableWksp));
+        (void)entropyWorkspaceSize;
+        FORWARD_IF_ERROR(FSE_normalizeCount(wksp->norm, tableLog, count, nbSeq_1, max, ZSTD_useLowProbCount(nbSeq_1)), "FSE_normalizeCount failed");
+        assert(oend >= op);
+        {   size_t const NCountSize = FSE_writeNCount(op, (size_t)(oend - op), wksp->norm, max, tableLog);   /* overflow protected */
             FORWARD_IF_ERROR(NCountSize, "FSE_writeNCount failed");
-            FORWARD_IF_ERROR(FSE_buildCTable_wksp(nextCTable, norm, max, tableLog, entropyWorkspace, entropyWorkspaceSize), "");
+            FORWARD_IF_ERROR(FSE_buildCTable_wksp(nextCTable, wksp->norm, max, tableLog, wksp->wksp, sizeof(wksp->wksp)), "FSE_buildCTable_wksp failed");
             return NCountSize;
         }
     }
@@ -290,19 +313,19 @@ ZSTD_encodeSequences_body(
     FSE_initCState2(&stateLitLength,   CTable_LitLength,   llCodeTable[nbSeq-1]);
     BIT_addBits(&blockStream, sequences[nbSeq-1].litLength, LL_bits[llCodeTable[nbSeq-1]]);
     if (MEM_32bits()) BIT_flushBits(&blockStream);
-    BIT_addBits(&blockStream, sequences[nbSeq-1].matchLength, ML_bits[mlCodeTable[nbSeq-1]]);
+    BIT_addBits(&blockStream, sequences[nbSeq-1].mlBase, ML_bits[mlCodeTable[nbSeq-1]]);
     if (MEM_32bits()) BIT_flushBits(&blockStream);
     if (longOffsets) {
         U32 const ofBits = ofCodeTable[nbSeq-1];
         unsigned const extraBits = ofBits - MIN(ofBits, STREAM_ACCUMULATOR_MIN-1);
         if (extraBits) {
-            BIT_addBits(&blockStream, sequences[nbSeq-1].offset, extraBits);
+            BIT_addBits(&blockStream, sequences[nbSeq-1].offBase, extraBits);
             BIT_flushBits(&blockStream);
         }
-        BIT_addBits(&blockStream, sequences[nbSeq-1].offset >> extraBits,
+        BIT_addBits(&blockStream, sequences[nbSeq-1].offBase >> extraBits,
                     ofBits - extraBits);
     } else {
-        BIT_addBits(&blockStream, sequences[nbSeq-1].offset, ofCodeTable[nbSeq-1]);
+        BIT_addBits(&blockStream, sequences[nbSeq-1].offBase, ofCodeTable[nbSeq-1]);
     }
     BIT_flushBits(&blockStream);
 
@@ -316,8 +339,8 @@ ZSTD_encodeSequences_body(
             U32  const mlBits = ML_bits[mlCode];
             DEBUGLOG(6, "encoding: litlen:%2u - matchlen:%2u - offCode:%7u",
                         (unsigned)sequences[n].litLength,
-                        (unsigned)sequences[n].matchLength + MINMATCH,
-                        (unsigned)sequences[n].offset);
+                        (unsigned)sequences[n].mlBase + MINMATCH,
+                        (unsigned)sequences[n].offBase);
                                                                             /* 32b*/  /* 64b*/
                                                                             /* (7)*/  /* (7)*/
             FSE_encodeSymbol(&blockStream, &stateOffsetBits, ofCode);       /* 15 */  /* 15 */
@@ -328,18 +351,18 @@ ZSTD_encodeSequences_body(
                 BIT_flushBits(&blockStream);                                /* (7)*/
             BIT_addBits(&blockStream, sequences[n].litLength, llBits);
             if (MEM_32bits() && ((llBits+mlBits)>24)) BIT_flushBits(&blockStream);
-            BIT_addBits(&blockStream, sequences[n].matchLength, mlBits);
+            BIT_addBits(&blockStream, sequences[n].mlBase, mlBits);
             if (MEM_32bits() || (ofBits+mlBits+llBits > 56)) BIT_flushBits(&blockStream);
             if (longOffsets) {
                 unsigned const extraBits = ofBits - MIN(ofBits, STREAM_ACCUMULATOR_MIN-1);
                 if (extraBits) {
-                    BIT_addBits(&blockStream, sequences[n].offset, extraBits);
+                    BIT_addBits(&blockStream, sequences[n].offBase, extraBits);
                     BIT_flushBits(&blockStream);                            /* (7)*/
                 }
-                BIT_addBits(&blockStream, sequences[n].offset >> extraBits,
+                BIT_addBits(&blockStream, sequences[n].offBase >> extraBits,
                             ofBits - extraBits);                            /* 31 */
             } else {
-                BIT_addBits(&blockStream, sequences[n].offset, ofBits);     /* 31 */
+                BIT_addBits(&blockStream, sequences[n].offBase, ofBits);     /* 31 */
             }
             BIT_flushBits(&blockStream);                                    /* (7)*/
             DEBUGLOG(7, "remaining space : %i", (int)(blockStream.endPtr - blockStream.ptr));
@@ -376,7 +399,7 @@ ZSTD_encodeSequences_default(
 
 #if DYNAMIC_BMI2
 
-static TARGET_ATTRIBUTE("bmi2") size_t
+static BMI2_TARGET_ATTRIBUTE size_t
 ZSTD_encodeSequences_bmi2(
             void* dst, size_t dstCapacity,
             FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,
diff --git a/contrib/zstd/zstd_compress_sequences.h b/contrib/zstd/zstd_compress_sequences.h
index b884e45b4..37848d8db 100644
--- a/contrib/zstd/zstd_compress_sequences.h
+++ b/contrib/zstd/zstd_compress_sequences.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/contrib/zstd/zstd_compress_superblock.c b/contrib/zstd/zstd_compress_superblock.c
index d72a8000a..834908520 100644
--- a/contrib/zstd/zstd_compress_superblock.c
+++ b/contrib/zstd/zstd_compress_superblock.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -15,288 +15,10 @@
 
 #include "zstd_internal.h"  /* ZSTD_getSequenceLength */
 #include "hist.h"                     /* HIST_countFast_wksp */
-#include "zstd_compress_internal.h"
+#include "zstd_compress_internal.h"   /* ZSTD_[huf|fse|entropy]CTablesMetadata_t */
 #include "zstd_compress_sequences.h"
 #include "zstd_compress_literals.h"
 
-/*-*************************************
-*  Superblock entropy buffer structs
-***************************************/
-/** ZSTD_hufCTablesMetadata_t :
- *  Stores Literals Block Type for a super-block in hType, and
- *  huffman tree description in hufDesBuffer.
- *  hufDesSize refers to the size of huffman tree description in bytes.
- *  This metadata is populated in ZSTD_buildSuperBlockEntropy_literal() */
-typedef struct {
-    symbolEncodingType_e hType;
-    BYTE hufDesBuffer[500]; /* TODO give name to this value */
-    size_t hufDesSize;
-} ZSTD_hufCTablesMetadata_t;
-
-/** ZSTD_fseCTablesMetadata_t :
- *  Stores symbol compression modes for a super-block in {ll, ol, ml}Type, and
- *  fse tables in fseTablesBuffer.
- *  fseTablesSize refers to the size of fse tables in bytes.
- *  This metadata is populated in ZSTD_buildSuperBlockEntropy_sequences() */
-typedef struct {
-    symbolEncodingType_e llType;
-    symbolEncodingType_e ofType;
-    symbolEncodingType_e mlType;
-    BYTE fseTablesBuffer[500]; /* TODO give name to this value */
-    size_t fseTablesSize;
-    size_t lastCountSize; /* This is to account for bug in 1.3.4. More detail in ZSTD_compressSubBlock_sequences() */
-} ZSTD_fseCTablesMetadata_t;
-
-typedef struct {
-    ZSTD_hufCTablesMetadata_t hufMetadata;
-    ZSTD_fseCTablesMetadata_t fseMetadata;
-} ZSTD_entropyCTablesMetadata_t;
-
-
-/** ZSTD_buildSuperBlockEntropy_literal() :
- *  Builds entropy for the super-block literals.
- *  Stores literals block type (raw, rle, compressed, repeat) and
- *  huffman description table to hufMetadata.
- *  @return : size of huffman description table or error code */
-static size_t ZSTD_buildSuperBlockEntropy_literal(void* const src, size_t srcSize,
-                                            const ZSTD_hufCTables_t* prevHuf,
-                                                  ZSTD_hufCTables_t* nextHuf,
-                                                  ZSTD_hufCTablesMetadata_t* hufMetadata,
-                                                  const int disableLiteralsCompression,
-                                                  void* workspace, size_t wkspSize)
-{
-    BYTE* const wkspStart = (BYTE*)workspace;
-    BYTE* const wkspEnd = wkspStart + wkspSize;
-    BYTE* const countWkspStart = wkspStart;
-    unsigned* const countWksp = (unsigned*)workspace;
-    const size_t countWkspSize = (HUF_SYMBOLVALUE_MAX + 1) * sizeof(unsigned);
-    BYTE* const nodeWksp = countWkspStart + countWkspSize;
-    const size_t nodeWkspSize = wkspEnd-nodeWksp;
-    unsigned maxSymbolValue = 255;
-    unsigned huffLog = HUF_TABLELOG_DEFAULT;
-    HUF_repeat repeat = prevHuf->repeatMode;
-
-    DEBUGLOG(5, "ZSTD_buildSuperBlockEntropy_literal (srcSize=%zu)", srcSize);
-
-    /* Prepare nextEntropy assuming reusing the existing table */
-    memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
-
-    if (disableLiteralsCompression) {
-        DEBUGLOG(5, "set_basic - disabled");
-        hufMetadata->hType = set_basic;
-        return 0;
-    }
-
-    /* small ? don't even attempt compression (speed opt) */
-#   define COMPRESS_LITERALS_SIZE_MIN 63
-    {   size_t const minLitSize = (prevHuf->repeatMode == HUF_repeat_valid) ? 6 : COMPRESS_LITERALS_SIZE_MIN;
-        if (srcSize <= minLitSize) {
-            DEBUGLOG(5, "set_basic - too small");
-            hufMetadata->hType = set_basic;
-            return 0;
-        }
-    }
-
-    /* Scan input and build symbol stats */
-    {   size_t const largest = HIST_count_wksp (countWksp, &maxSymbolValue, (const BYTE*)src, srcSize, workspace, wkspSize);
-        FORWARD_IF_ERROR(largest, "HIST_count_wksp failed");
-        if (largest == srcSize) {
-            DEBUGLOG(5, "set_rle");
-            hufMetadata->hType = set_rle;
-            return 0;
-        }
-        if (largest <= (srcSize >> 7)+4) {
-            DEBUGLOG(5, "set_basic - no gain");
-            hufMetadata->hType = set_basic;
-            return 0;
-        }
-    }
-
-    /* Validate the previous Huffman table */
-    if (repeat == HUF_repeat_check && !HUF_validateCTable((HUF_CElt const*)prevHuf->CTable, countWksp, maxSymbolValue)) {
-        repeat = HUF_repeat_none;
-    }
-
-    /* Build Huffman Tree */
-    memset(nextHuf->CTable, 0, sizeof(nextHuf->CTable));
-    huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue);
-    {   size_t const maxBits = HUF_buildCTable_wksp((HUF_CElt*)nextHuf->CTable, countWksp,
-                                                    maxSymbolValue, huffLog,
-                                                    nodeWksp, nodeWkspSize);
-        FORWARD_IF_ERROR(maxBits, "HUF_buildCTable_wksp");
-        huffLog = (U32)maxBits;
-        {   /* Build and write the CTable */
-            size_t const newCSize = HUF_estimateCompressedSize(
-                    (HUF_CElt*)nextHuf->CTable, countWksp, maxSymbolValue);
-            size_t const hSize = HUF_writeCTable(
-                    hufMetadata->hufDesBuffer, sizeof(hufMetadata->hufDesBuffer),
-                    (HUF_CElt*)nextHuf->CTable, maxSymbolValue, huffLog);
-            /* Check against repeating the previous CTable */
-            if (repeat != HUF_repeat_none) {
-                size_t const oldCSize = HUF_estimateCompressedSize(
-                        (HUF_CElt const*)prevHuf->CTable, countWksp, maxSymbolValue);
-                if (oldCSize < srcSize && (oldCSize <= hSize + newCSize || hSize + 12 >= srcSize)) {
-                    DEBUGLOG(5, "set_repeat - smaller");
-                    memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
-                    hufMetadata->hType = set_repeat;
-                    return 0;
-                }
-            }
-            if (newCSize + hSize >= srcSize) {
-                DEBUGLOG(5, "set_basic - no gains");
-                memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
-                hufMetadata->hType = set_basic;
-                return 0;
-            }
-            DEBUGLOG(5, "set_compressed (hSize=%u)", (U32)hSize);
-            hufMetadata->hType = set_compressed;
-            nextHuf->repeatMode = HUF_repeat_check;
-            return hSize;
-        }
-    }
-}
-
-/** ZSTD_buildSuperBlockEntropy_sequences() :
- *  Builds entropy for the super-block sequences.
- *  Stores symbol compression modes and fse table to fseMetadata.
- *  @return : size of fse tables or error code */
-static size_t ZSTD_buildSuperBlockEntropy_sequences(seqStore_t* seqStorePtr,
-                                              const ZSTD_fseCTables_t* prevEntropy,
-                                                    ZSTD_fseCTables_t* nextEntropy,
-                                              const ZSTD_CCtx_params* cctxParams,
-                                                    ZSTD_fseCTablesMetadata_t* fseMetadata,
-                                                    void* workspace, size_t wkspSize)
-{
-    BYTE* const wkspStart = (BYTE*)workspace;
-    BYTE* const wkspEnd = wkspStart + wkspSize;
-    BYTE* const countWkspStart = wkspStart;
-    unsigned* const countWksp = (unsigned*)workspace;
-    const size_t countWkspSize = (MaxSeq + 1) * sizeof(unsigned);
-    BYTE* const cTableWksp = countWkspStart + countWkspSize;
-    const size_t cTableWkspSize = wkspEnd-cTableWksp;
-    ZSTD_strategy const strategy = cctxParams->cParams.strategy;
-    FSE_CTable* CTable_LitLength = nextEntropy->litlengthCTable;
-    FSE_CTable* CTable_OffsetBits = nextEntropy->offcodeCTable;
-    FSE_CTable* CTable_MatchLength = nextEntropy->matchlengthCTable;
-    const BYTE* const ofCodeTable = seqStorePtr->ofCode;
-    const BYTE* const llCodeTable = seqStorePtr->llCode;
-    const BYTE* const mlCodeTable = seqStorePtr->mlCode;
-    size_t const nbSeq = seqStorePtr->sequences - seqStorePtr->sequencesStart;
-    BYTE* const ostart = fseMetadata->fseTablesBuffer;
-    BYTE* const oend = ostart + sizeof(fseMetadata->fseTablesBuffer);
-    BYTE* op = ostart;
-
-    assert(cTableWkspSize >= (1 << MaxFSELog) * sizeof(FSE_FUNCTION_TYPE));
-    DEBUGLOG(5, "ZSTD_buildSuperBlockEntropy_sequences (nbSeq=%zu)", nbSeq);
-    memset(workspace, 0, wkspSize);
-
-    fseMetadata->lastCountSize = 0;
-    /* convert length/distances into codes */
-    ZSTD_seqToCodes(seqStorePtr);
-    /* build CTable for Literal Lengths */
-    {   U32 LLtype;
-        unsigned max = MaxLL;
-        size_t const mostFrequent = HIST_countFast_wksp(countWksp, &max, llCodeTable, nbSeq, workspace, wkspSize);  /* can't fail */
-        DEBUGLOG(5, "Building LL table");
-        nextEntropy->litlength_repeatMode = prevEntropy->litlength_repeatMode;
-        LLtype = ZSTD_selectEncodingType(&nextEntropy->litlength_repeatMode,
-                                        countWksp, max, mostFrequent, nbSeq,
-                                        LLFSELog, prevEntropy->litlengthCTable,
-                                        LL_defaultNorm, LL_defaultNormLog,
-                                        ZSTD_defaultAllowed, strategy);
-        assert(set_basic < set_compressed && set_rle < set_compressed);
-        assert(!(LLtype < set_compressed && nextEntropy->litlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */
-        {   size_t const countSize = ZSTD_buildCTable(op, oend - op, CTable_LitLength, LLFSELog, (symbolEncodingType_e)LLtype,
-                                                    countWksp, max, llCodeTable, nbSeq, LL_defaultNorm, LL_defaultNormLog, MaxLL,
-                                                    prevEntropy->litlengthCTable, sizeof(prevEntropy->litlengthCTable),
-                                                    cTableWksp, cTableWkspSize);
-            FORWARD_IF_ERROR(countSize, "ZSTD_buildCTable for LitLens failed");
-            if (LLtype == set_compressed)
-                fseMetadata->lastCountSize = countSize;
-            op += countSize;
-            fseMetadata->llType = (symbolEncodingType_e) LLtype;
-    }   }
-    /* build CTable for Offsets */
-    {   U32 Offtype;
-        unsigned max = MaxOff;
-        size_t const mostFrequent = HIST_countFast_wksp(countWksp, &max, ofCodeTable, nbSeq, workspace, wkspSize);  /* can't fail */
-        /* We can only use the basic table if max <= DefaultMaxOff, otherwise the offsets are too large */
-        ZSTD_defaultPolicy_e const defaultPolicy = (max <= DefaultMaxOff) ? ZSTD_defaultAllowed : ZSTD_defaultDisallowed;
-        DEBUGLOG(5, "Building OF table");
-        nextEntropy->offcode_repeatMode = prevEntropy->offcode_repeatMode;
-        Offtype = ZSTD_selectEncodingType(&nextEntropy->offcode_repeatMode,
-                                        countWksp, max, mostFrequent, nbSeq,
-                                        OffFSELog, prevEntropy->offcodeCTable,
-                                        OF_defaultNorm, OF_defaultNormLog,
-                                        defaultPolicy, strategy);
-        assert(!(Offtype < set_compressed && nextEntropy->offcode_repeatMode != FSE_repeat_none)); /* We don't copy tables */
-        {   size_t const countSize = ZSTD_buildCTable(op, oend - op, CTable_OffsetBits, OffFSELog, (symbolEncodingType_e)Offtype,
-                                                    countWksp, max, ofCodeTable, nbSeq, OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff,
-                                                    prevEntropy->offcodeCTable, sizeof(prevEntropy->offcodeCTable),
-                                                    cTableWksp, cTableWkspSize);
-            FORWARD_IF_ERROR(countSize, "ZSTD_buildCTable for Offsets failed");
-            if (Offtype == set_compressed)
-                fseMetadata->lastCountSize = countSize;
-            op += countSize;
-            fseMetadata->ofType = (symbolEncodingType_e) Offtype;
-    }   }
-    /* build CTable for MatchLengths */
-    {   U32 MLtype;
-        unsigned max = MaxML;
-        size_t const mostFrequent = HIST_countFast_wksp(countWksp, &max, mlCodeTable, nbSeq, workspace, wkspSize);   /* can't fail */
-        DEBUGLOG(5, "Building ML table (remaining space : %i)", (int)(oend-op));
-        nextEntropy->matchlength_repeatMode = prevEntropy->matchlength_repeatMode;
-        MLtype = ZSTD_selectEncodingType(&nextEntropy->matchlength_repeatMode,
-                                        countWksp, max, mostFrequent, nbSeq,
-                                        MLFSELog, prevEntropy->matchlengthCTable,
-                                        ML_defaultNorm, ML_defaultNormLog,
-                                        ZSTD_defaultAllowed, strategy);
-        assert(!(MLtype < set_compressed && nextEntropy->matchlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */
-        {   size_t const countSize = ZSTD_buildCTable(op, oend - op, CTable_MatchLength, MLFSELog, (symbolEncodingType_e)MLtype,
-                                                    countWksp, max, mlCodeTable, nbSeq, ML_defaultNorm, ML_defaultNormLog, MaxML,
-                                                    prevEntropy->matchlengthCTable, sizeof(prevEntropy->matchlengthCTable),
-                                                    cTableWksp, cTableWkspSize);
-            FORWARD_IF_ERROR(countSize, "ZSTD_buildCTable for MatchLengths failed");
-            if (MLtype == set_compressed)
-                fseMetadata->lastCountSize = countSize;
-            op += countSize;
-            fseMetadata->mlType = (symbolEncodingType_e) MLtype;
-    }   }
-    assert((size_t) (op-ostart) <= sizeof(fseMetadata->fseTablesBuffer));
-    return op-ostart;
-}
-
-
-/** ZSTD_buildSuperBlockEntropy() :
- *  Builds entropy for the super-block.
- *  @return : 0 on success or error code */
-static size_t
-ZSTD_buildSuperBlockEntropy(seqStore_t* seqStorePtr,
-                      const ZSTD_entropyCTables_t* prevEntropy,
-                            ZSTD_entropyCTables_t* nextEntropy,
-                      const ZSTD_CCtx_params* cctxParams,
-                            ZSTD_entropyCTablesMetadata_t* entropyMetadata,
-                            void* workspace, size_t wkspSize)
-{
-    size_t const litSize = seqStorePtr->lit - seqStorePtr->litStart;
-    DEBUGLOG(5, "ZSTD_buildSuperBlockEntropy");
-    entropyMetadata->hufMetadata.hufDesSize =
-        ZSTD_buildSuperBlockEntropy_literal(seqStorePtr->litStart, litSize,
-                                            &prevEntropy->huf, &nextEntropy->huf,
-                                            &entropyMetadata->hufMetadata,
-                                            ZSTD_disableLiteralsCompression(cctxParams),
-                                            workspace, wkspSize);
-    FORWARD_IF_ERROR(entropyMetadata->hufMetadata.hufDesSize, "ZSTD_buildSuperBlockEntropy_literal failed");
-    entropyMetadata->fseMetadata.fseTablesSize =
-        ZSTD_buildSuperBlockEntropy_sequences(seqStorePtr,
-                                              &prevEntropy->fse, &nextEntropy->fse,
-                                              cctxParams,
-                                              &entropyMetadata->fseMetadata,
-                                              workspace, wkspSize);
-    FORWARD_IF_ERROR(entropyMetadata->fseMetadata.fseTablesSize, "ZSTD_buildSuperBlockEntropy_sequences failed");
-    return 0;
-}
-
 /** ZSTD_compressSubBlock_literal() :
  *  Compresses literals section for a sub-block.
  *  When we have to write the Huffman table we will sometimes choose a header
@@ -304,7 +26,7 @@ ZSTD_buildSuperBlockEntropy(seqStore_t* seqStorePtr,
  *  before we know the table size + compressed size, so we have a bound on the
  *  table size. If we guessed incorrectly, we fall back to uncompressed literals.
  *
- *  We write the header when writeEntropy=1 and set entropyWrriten=1 when we succeeded
+ *  We write the header when writeEntropy=1 and set entropyWritten=1 when we succeeded
  *  in writing the header, otherwise it is set to 0.
  *
  *  hufMetadata->hType has literals block type info.
@@ -314,13 +36,14 @@ ZSTD_buildSuperBlockEntropy(seqStore_t* seqStorePtr,
  *      If it is set_compressed, first sub-block's literals section will be Treeless_Literals_Block
  *      and the following sub-blocks' literals sections will be Treeless_Literals_Block.
  *  @return : compressed size of literals section of a sub-block
- *            Or 0 if it unable to compress.
+ *            Or 0 if unable to compress.
  *            Or error code */
-static size_t ZSTD_compressSubBlock_literal(const HUF_CElt* hufTable,
-                                    const ZSTD_hufCTablesMetadata_t* hufMetadata,
-                                    const BYTE* literals, size_t litSize,
-                                    void* dst, size_t dstSize,
-                                    const int bmi2, int writeEntropy, int* entropyWritten)
+static size_t
+ZSTD_compressSubBlock_literal(const HUF_CElt* hufTable,
+                              const ZSTD_hufCTablesMetadata_t* hufMetadata,
+                              const BYTE* literals, size_t litSize,
+                              void* dst, size_t dstSize,
+                              const int bmi2, int writeEntropy, int* entropyWritten)
 {
     size_t const header = writeEntropy ? 200 : 0;
     size_t const lhSize = 3 + (litSize >= (1 KB - header)) + (litSize >= (16 KB - header));
@@ -331,8 +54,6 @@ static size_t ZSTD_compressSubBlock_literal(const HUF_CElt* hufTable,
     symbolEncodingType_e hType = writeEntropy ? hufMetadata->hType : set_repeat;
     size_t cLitSize = 0;
 
-    (void)bmi2; /* TODO bmi2... */
-
     DEBUGLOG(5, "ZSTD_compressSubBlock_literal (litSize=%zu, lhSize=%zu, writeEntropy=%d)", litSize, lhSize, writeEntropy);
 
     *entropyWritten = 0;
@@ -348,15 +69,15 @@ static size_t ZSTD_compressSubBlock_literal(const HUF_CElt* hufTable,
     assert(hufMetadata->hType == set_compressed || hufMetadata->hType == set_repeat);
 
     if (writeEntropy && hufMetadata->hType == set_compressed) {
-        memcpy(op, hufMetadata->hufDesBuffer, hufMetadata->hufDesSize);
+        ZSTD_memcpy(op, hufMetadata->hufDesBuffer, hufMetadata->hufDesSize);
         op += hufMetadata->hufDesSize;
         cLitSize += hufMetadata->hufDesSize;
         DEBUGLOG(5, "ZSTD_compressSubBlock_literal (hSize=%zu)", hufMetadata->hufDesSize);
     }
 
-    /* TODO bmi2 */
-    {   const size_t cSize = singleStream ? HUF_compress1X_usingCTable(op, oend-op, literals, litSize, hufTable)
-                                          : HUF_compress4X_usingCTable(op, oend-op, literals, litSize, hufTable);
+    {   int const flags = bmi2 ? HUF_flags_bmi2 : 0;
+        const size_t cSize = singleStream ? HUF_compress1X_usingCTable(op, oend-op, literals, litSize, hufTable, flags)
+                                          : HUF_compress4X_usingCTable(op, oend-op, literals, litSize, hufTable, flags);
         op += cSize;
         cLitSize += cSize;
         if (cSize == 0 || ERR_isError(cSize)) {
@@ -404,12 +125,17 @@ static size_t ZSTD_compressSubBlock_literal(const HUF_CElt* hufTable,
     return op-ostart;
 }
 
-static size_t ZSTD_seqDecompressedSize(seqStore_t const* seqStore, const seqDef* sequences, size_t nbSeq, size_t litSize, int lastSequence) {
+static size_t
+ZSTD_seqDecompressedSize(seqStore_t const* seqStore,
+                   const seqDef* sequences, size_t nbSeq,
+                         size_t litSize, int lastSequence)
+{
     const seqDef* const sstart = sequences;
     const seqDef* const send = sequences + nbSeq;
     const seqDef* sp = sstart;
     size_t matchLengthSum = 0;
     size_t litLengthSum = 0;
+    (void)(litLengthSum); /* suppress unused variable warning on some environments */
     while (send-sp > 0) {
         ZSTD_sequenceLength const seqLen = ZSTD_getSequenceLength(seqStore, sp);
         litLengthSum += seqLen.litLength;
@@ -433,13 +159,14 @@ static size_t ZSTD_seqDecompressedSize(seqStore_t const* seqStore, const seqDef*
  *  @return : compressed size of sequences section of a sub-block
  *            Or 0 if it is unable to compress
  *            Or error code. */
-static size_t ZSTD_compressSubBlock_sequences(const ZSTD_fseCTables_t* fseTables,
-                                              const ZSTD_fseCTablesMetadata_t* fseMetadata,
-                                              const seqDef* sequences, size_t nbSeq,
-                                              const BYTE* llCode, const BYTE* mlCode, const BYTE* ofCode,
-                                              const ZSTD_CCtx_params* cctxParams,
-                                              void* dst, size_t dstCapacity,
-                                              const int bmi2, int writeEntropy, int* entropyWritten)
+static size_t
+ZSTD_compressSubBlock_sequences(const ZSTD_fseCTables_t* fseTables,
+                                const ZSTD_fseCTablesMetadata_t* fseMetadata,
+                                const seqDef* sequences, size_t nbSeq,
+                                const BYTE* llCode, const BYTE* mlCode, const BYTE* ofCode,
+                                const ZSTD_CCtx_params* cctxParams,
+                                void* dst, size_t dstCapacity,
+                                const int bmi2, int writeEntropy, int* entropyWritten)
 {
     const int longOffsets = cctxParams->cParams.windowLog > STREAM_ACCUMULATOR_MIN;
     BYTE* const ostart = (BYTE*)dst;
@@ -474,7 +201,7 @@ static size_t ZSTD_compressSubBlock_sequences(const ZSTD_fseCTables_t* fseTables
         const U32 MLtype = fseMetadata->mlType;
         DEBUGLOG(5, "ZSTD_compressSubBlock_sequences (fseTablesSize=%zu)", fseMetadata->fseTablesSize);
         *seqHead = (BYTE)((LLtype<<6) + (Offtype<<4) + (MLtype<<2));
-        memcpy(op, fseMetadata->fseTablesBuffer, fseMetadata->fseTablesSize);
+        ZSTD_memcpy(op, fseMetadata->fseTablesBuffer, fseMetadata->fseTablesSize);
         op += fseMetadata->fseTablesSize;
     } else {
         const U32 repeat = set_repeat;
@@ -602,8 +329,8 @@ static size_t ZSTD_estimateSubBlockSize_literal(const BYTE* literals, size_t lit
 static size_t ZSTD_estimateSubBlockSize_symbolType(symbolEncodingType_e type,
                         const BYTE* codeTable, unsigned maxCode,
                         size_t nbSeq, const FSE_CTable* fseCTable,
-                        const U32* additionalBits,
-                        short const* defaultNorm, U32 defaultNormLog,
+                        const U8* additionalBits,
+                        short const* defaultNorm, U32 defaultNormLog, U32 defaultMax,
                         void* workspace, size_t wkspSize)
 {
     unsigned* const countWksp = (unsigned*)workspace;
@@ -615,7 +342,11 @@ static size_t ZSTD_estimateSubBlockSize_symbolType(symbolEncodingType_e type,
 
     HIST_countFast_wksp(countWksp, &max, codeTable, nbSeq, workspace, wkspSize);  /* can't fail */
     if (type == set_basic) {
-        cSymbolTypeSizeEstimateInBits = ZSTD_crossEntropyCost(defaultNorm, defaultNormLog, countWksp, max);
+        /* We selected this encoding type, so it must be valid. */
+        assert(max <= defaultMax);
+        cSymbolTypeSizeEstimateInBits = max <= defaultMax
+                ? ZSTD_crossEntropyCost(defaultNorm, defaultNormLog, countWksp, max)
+                : ERROR(GENERIC);
     } else if (type == set_rle) {
         cSymbolTypeSizeEstimateInBits = 0;
     } else if (type == set_compressed || type == set_repeat) {
@@ -639,19 +370,20 @@ static size_t ZSTD_estimateSubBlockSize_sequences(const BYTE* ofCodeTable,
                                                   void* workspace, size_t wkspSize,
                                                   int writeEntropy)
 {
-    size_t sequencesSectionHeaderSize = 3; /* Use hard coded size of 3 bytes */
+    size_t const sequencesSectionHeaderSize = 3; /* Use hard coded size of 3 bytes */
     size_t cSeqSizeEstimate = 0;
+    if (nbSeq == 0) return sequencesSectionHeaderSize;
     cSeqSizeEstimate += ZSTD_estimateSubBlockSize_symbolType(fseMetadata->ofType, ofCodeTable, MaxOff,
                                          nbSeq, fseTables->offcodeCTable, NULL,
-                                         OF_defaultNorm, OF_defaultNormLog,
+                                         OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff,
                                          workspace, wkspSize);
     cSeqSizeEstimate += ZSTD_estimateSubBlockSize_symbolType(fseMetadata->llType, llCodeTable, MaxLL,
                                          nbSeq, fseTables->litlengthCTable, LL_bits,
-                                         LL_defaultNorm, LL_defaultNormLog,
+                                         LL_defaultNorm, LL_defaultNormLog, MaxLL,
                                          workspace, wkspSize);
     cSeqSizeEstimate += ZSTD_estimateSubBlockSize_symbolType(fseMetadata->mlType, mlCodeTable, MaxML,
                                          nbSeq, fseTables->matchlengthCTable, ML_bits,
-                                         ML_defaultNorm, ML_defaultNormLog,
+                                         ML_defaultNorm, ML_defaultNormLog, MaxML,
                                          workspace, wkspSize);
     if (writeEntropy) cSeqSizeEstimate += fseMetadata->fseTablesSize;
     return cSeqSizeEstimate + sequencesSectionHeaderSize;
@@ -747,7 +479,7 @@ static size_t ZSTD_compressSubBlock_multi(const seqStore_t* seqStorePtr,
         /* I think there is an optimization opportunity here.
          * Calling ZSTD_estimateSubBlockSize for every sequence can be wasteful
          * since it recalculates estimate from scratch.
-         * For example, it would recount literal distribution and symbol codes everytime.
+         * For example, it would recount literal distribution and symbol codes every time.
          */
         cBlockSizeEstimate = ZSTD_estimateSubBlockSize(lp, litSize, ofCodePtr, llCodePtr, mlCodePtr, seqCount,
                                                        &nextCBlock->entropy, entropyMetadata,
@@ -790,7 +522,7 @@ static size_t ZSTD_compressSubBlock_multi(const seqStore_t* seqStorePtr,
     } while (!lastSequence);
     if (writeLitEntropy) {
         DEBUGLOG(5, "ZSTD_compressSubBlock_multi has literal entropy tables unwritten");
-        memcpy(&nextCBlock->entropy.huf, &prevCBlock->entropy.huf, sizeof(prevCBlock->entropy.huf));
+        ZSTD_memcpy(&nextCBlock->entropy.huf, &prevCBlock->entropy.huf, sizeof(prevCBlock->entropy.huf));
     }
     if (writeSeqEntropy && ZSTD_needSequenceEntropyTables(&entropyMetadata->fseMetadata)) {
         /* If we haven't written our entropy tables, then we've violated our contract and
@@ -809,11 +541,11 @@ static size_t ZSTD_compressSubBlock_multi(const seqStore_t* seqStorePtr,
         if (sp < send) {
             seqDef const* seq;
             repcodes_t rep;
-            memcpy(&rep, prevCBlock->rep, sizeof(rep));
+            ZSTD_memcpy(&rep, prevCBlock->rep, sizeof(rep));
             for (seq = sstart; seq < sp; ++seq) {
-                rep = ZSTD_updateRep(rep.rep, seq->offset - 1, ZSTD_getSequenceLength(seqStorePtr, seq).litLength == 0);
+                ZSTD_updateRep(rep.rep, seq->offBase, ZSTD_getSequenceLength(seqStorePtr, seq).litLength == 0);
             }
-            memcpy(nextCBlock->rep, &rep, sizeof(rep));
+            ZSTD_memcpy(nextCBlock->rep, &rep, sizeof(rep));
         }
     }
     DEBUGLOG(5, "ZSTD_compressSubBlock_multi compressed");
@@ -826,12 +558,12 @@ size_t ZSTD_compressSuperBlock(ZSTD_CCtx* zc,
                                unsigned lastBlock) {
     ZSTD_entropyCTablesMetadata_t entropyMetadata;
 
-    FORWARD_IF_ERROR(ZSTD_buildSuperBlockEntropy(&zc->seqStore,
+    FORWARD_IF_ERROR(ZSTD_buildBlockEntropyStats(&zc->seqStore,
           &zc->blockState.prevCBlock->entropy,
           &zc->blockState.nextCBlock->entropy,
           &zc->appliedParams,
           &entropyMetadata,
-          zc->entropyWorkspace, HUF_WORKSPACE_SIZE /* statically allocated in resetCCtx */), "");
+          zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */), "");
 
     return ZSTD_compressSubBlock_multi(&zc->seqStore,
             zc->blockState.prevCBlock,
@@ -841,5 +573,5 @@ size_t ZSTD_compressSuperBlock(ZSTD_CCtx* zc,
             dst, dstCapacity,
             src, srcSize,
             zc->bmi2, lastBlock,
-            zc->entropyWorkspace, HUF_WORKSPACE_SIZE /* statically allocated in resetCCtx */);
+            zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */);
 }
diff --git a/contrib/zstd/zstd_compress_superblock.h b/contrib/zstd/zstd_compress_superblock.h
index 35d207299..1ac733b75 100644
--- a/contrib/zstd/zstd_compress_superblock.h
+++ b/contrib/zstd/zstd_compress_superblock.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/contrib/zstd/zstd_cwksp.h b/contrib/zstd/zstd_cwksp.h
index 91f812fa3..4d007c099 100644
--- a/contrib/zstd/zstd_cwksp.h
+++ b/contrib/zstd/zstd_cwksp.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -35,6 +35,10 @@ extern "C" {
 #define ZSTD_CWKSP_ASAN_REDZONE_SIZE 128
 #endif
 
+
+/* Set our tables and aligneds to align by 64 bytes */
+#define ZSTD_CWKSP_ALIGNMENT_BYTES 64
+
 /*-*************************************
 *  Structures
 ***************************************/
@@ -45,6 +49,16 @@ typedef enum {
 } ZSTD_cwksp_alloc_phase_e;
 
 /**
+ * Used to describe whether the workspace is statically allocated (and will not
+ * necessarily ever be freed), or if it's dynamically allocated and we can
+ * expect a well-formed caller to free this.
+ */
+typedef enum {
+    ZSTD_cwksp_dynamic_alloc,
+    ZSTD_cwksp_static_alloc
+} ZSTD_cwksp_static_alloc_e;
+
+/**
  * Zstd fits all its internal datastructures into a single continuous buffer,
  * so that it only needs to perform a single OS allocation (or so that a buffer
  * can be provided to it and it can perform no allocations at all). This buffer
@@ -92,7 +106,7 @@ typedef enum {
  *
  * - Static objects: this is optionally the enclosing ZSTD_CCtx or ZSTD_CDict,
  *   so that literally everything fits in a single buffer. Note: if present,
- *   this must be the first object in the workspace, since ZSTD_free{CCtx,
+ *   this must be the first object in the workspace, since ZSTD_customFree{CCtx,
  *   CDict}() rely on a pointer comparison to see whether one or two frees are
  *   required.
  *
@@ -107,10 +121,11 @@ typedef enum {
  * - Tables: these are any of several different datastructures (hash tables,
  *   chain tables, binary trees) that all respect a common format: they are
  *   uint32_t arrays, all of whose values are between 0 and (nextSrc - base).
- *   Their sizes depend on the cparams.
+ *   Their sizes depend on the cparams. These tables are 64-byte aligned.
  *
  * - Aligned: these buffers are used for various purposes that require 4 byte
- *   alignment, but don't require any initialization before they're used.
+ *   alignment, but don't require any initialization before they're used. These
+ *   buffers are each aligned to 64 bytes.
  *
  * - Buffers: these buffers are used for various purposes that don't require
  *   any alignment or initialization before they're used. This means they can
@@ -123,8 +138,7 @@ typedef enum {
  *
  * 1. Objects
  * 2. Buffers
- * 3. Aligned
- * 4. Tables
+ * 3. Aligned/Tables
  *
  * Attempts to reserve objects of different types out of order will fail.
  */
@@ -137,9 +151,10 @@ typedef struct {
     void* tableValidEnd;
     void* allocStart;
 
-    int allocFailed;
+    BYTE allocFailed;
     int workspaceOversizedDuration;
     ZSTD_cwksp_alloc_phase_e phase;
+    ZSTD_cwksp_static_alloc_e isStatic;
 } ZSTD_cwksp;
 
 /*-*************************************
@@ -176,82 +191,168 @@ MEM_STATIC size_t ZSTD_cwksp_align(size_t size, size_t const align) {
  * Since tables aren't currently redzoned, you don't need to call through this
  * to figure out how much space you need for the matchState tables. Everything
  * else is though.
+ *
+ * Do not use for sizing aligned buffers. Instead, use ZSTD_cwksp_aligned_alloc_size().
  */
 MEM_STATIC size_t ZSTD_cwksp_alloc_size(size_t size) {
-#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
+    if (size == 0)
+        return 0;
+#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
     return size + 2 * ZSTD_CWKSP_ASAN_REDZONE_SIZE;
 #else
     return size;
 #endif
 }
 
-MEM_STATIC void ZSTD_cwksp_internal_advance_phase(
-        ZSTD_cwksp* ws, ZSTD_cwksp_alloc_phase_e phase) {
+/**
+ * Returns an adjusted alloc size that is the nearest larger multiple of 64 bytes.
+ * Used to determine the number of bytes required for a given "aligned".
+ */
+MEM_STATIC size_t ZSTD_cwksp_aligned_alloc_size(size_t size) {
+    return ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(size, ZSTD_CWKSP_ALIGNMENT_BYTES));
+}
+
+/**
+ * Returns the amount of additional space the cwksp must allocate
+ * for internal purposes (currently only alignment).
+ */
+MEM_STATIC size_t ZSTD_cwksp_slack_space_required(void) {
+    /* For alignment, the wksp will always allocate an additional n_1=[1, 64] bytes
+     * to align the beginning of tables section, as well as another n_2=[0, 63] bytes
+     * to align the beginning of the aligned section.
+     *
+     * n_1 + n_2 == 64 bytes if the cwksp is freshly allocated, due to tables and
+     * aligneds being sized in multiples of 64 bytes.
+     */
+    size_t const slackSpace = ZSTD_CWKSP_ALIGNMENT_BYTES;
+    return slackSpace;
+}
+
+
+/**
+ * Return the number of additional bytes required to align a pointer to the given number of bytes.
+ * alignBytes must be a power of two.
+ */
+MEM_STATIC size_t ZSTD_cwksp_bytes_to_align_ptr(void* ptr, const size_t alignBytes) {
+    size_t const alignBytesMask = alignBytes - 1;
+    size_t const bytes = (alignBytes - ((size_t)ptr & (alignBytesMask))) & alignBytesMask;
+    assert((alignBytes & alignBytesMask) == 0);
+    assert(bytes != ZSTD_CWKSP_ALIGNMENT_BYTES);
+    return bytes;
+}
+
+/**
+ * Internal function. Do not use directly.
+ * Reserves the given number of bytes within the aligned/buffer segment of the wksp,
+ * which counts from the end of the wksp (as opposed to the object/table segment).
+ *
+ * Returns a pointer to the beginning of that space.
+ */
+MEM_STATIC void*
+ZSTD_cwksp_reserve_internal_buffer_space(ZSTD_cwksp* ws, size_t const bytes)
+{
+    void* const alloc = (BYTE*)ws->allocStart - bytes;
+    void* const bottom = ws->tableEnd;
+    DEBUGLOG(5, "cwksp: reserving %p %zd bytes, %zd bytes remaining",
+        alloc, bytes, ZSTD_cwksp_available_space(ws) - bytes);
+    ZSTD_cwksp_assert_internal_consistency(ws);
+    assert(alloc >= bottom);
+    if (alloc < bottom) {
+        DEBUGLOG(4, "cwksp: alloc failed!");
+        ws->allocFailed = 1;
+        return NULL;
+    }
+    /* the area is reserved from the end of wksp.
+     * If it overlaps with tableValidEnd, it voids guarantees on values' range */
+    if (alloc < ws->tableValidEnd) {
+        ws->tableValidEnd = alloc;
+    }
+    ws->allocStart = alloc;
+    return alloc;
+}
+
+/**
+ * Moves the cwksp to the next phase, and does any necessary allocations.
+ * cwksp initialization must necessarily go through each phase in order.
+ * Returns a 0 on success, or zstd error
+ */
+MEM_STATIC size_t
+ZSTD_cwksp_internal_advance_phase(ZSTD_cwksp* ws, ZSTD_cwksp_alloc_phase_e phase)
+{
     assert(phase >= ws->phase);
     if (phase > ws->phase) {
+        /* Going from allocating objects to allocating buffers */
         if (ws->phase < ZSTD_cwksp_alloc_buffers &&
                 phase >= ZSTD_cwksp_alloc_buffers) {
             ws->tableValidEnd = ws->objectEnd;
         }
+
+        /* Going from allocating buffers to allocating aligneds/tables */
         if (ws->phase < ZSTD_cwksp_alloc_aligned &&
                 phase >= ZSTD_cwksp_alloc_aligned) {
-            /* If unaligned allocations down from a too-large top have left us
-             * unaligned, we need to realign our alloc ptr. Technically, this
-             * can consume space that is unaccounted for in the neededSpace
-             * calculation. However, I believe this can only happen when the
-             * workspace is too large, and specifically when it is too large
-             * by a larger margin than the space that will be consumed. */
-            /* TODO: cleaner, compiler warning friendly way to do this??? */
-            ws->allocStart = (BYTE*)ws->allocStart - ((size_t)ws->allocStart & (sizeof(U32)-1));
-            if (ws->allocStart < ws->tableValidEnd) {
-                ws->tableValidEnd = ws->allocStart;
+            {   /* Align the start of the "aligned" to 64 bytes. Use [1, 64] bytes. */
+                size_t const bytesToAlign =
+                    ZSTD_CWKSP_ALIGNMENT_BYTES - ZSTD_cwksp_bytes_to_align_ptr(ws->allocStart, ZSTD_CWKSP_ALIGNMENT_BYTES);
+                DEBUGLOG(5, "reserving aligned alignment addtl space: %zu", bytesToAlign);
+                ZSTD_STATIC_ASSERT((ZSTD_CWKSP_ALIGNMENT_BYTES & (ZSTD_CWKSP_ALIGNMENT_BYTES - 1)) == 0); /* power of 2 */
+                RETURN_ERROR_IF(!ZSTD_cwksp_reserve_internal_buffer_space(ws, bytesToAlign),
+                                memory_allocation, "aligned phase - alignment initial allocation failed!");
             }
-        }
+            {   /* Align the start of the tables to 64 bytes. Use [0, 63] bytes */
+                void* const alloc = ws->objectEnd;
+                size_t const bytesToAlign = ZSTD_cwksp_bytes_to_align_ptr(alloc, ZSTD_CWKSP_ALIGNMENT_BYTES);
+                void* const objectEnd = (BYTE*)alloc + bytesToAlign;
+                DEBUGLOG(5, "reserving table alignment addtl space: %zu", bytesToAlign);
+                RETURN_ERROR_IF(objectEnd > ws->workspaceEnd, memory_allocation,
+                                "table phase - alignment initial allocation failed!");
+                ws->objectEnd = objectEnd;
+                ws->tableEnd = objectEnd;  /* table area starts being empty */
+                if (ws->tableValidEnd < ws->tableEnd) {
+                    ws->tableValidEnd = ws->tableEnd;
+        }   }   }
         ws->phase = phase;
+        ZSTD_cwksp_assert_internal_consistency(ws);
     }
+    return 0;
 }
 
 /**
  * Returns whether this object/buffer/etc was allocated in this workspace.
  */
-MEM_STATIC int ZSTD_cwksp_owns_buffer(const ZSTD_cwksp* ws, const void* ptr) {
+MEM_STATIC int ZSTD_cwksp_owns_buffer(const ZSTD_cwksp* ws, const void* ptr)
+{
     return (ptr != NULL) && (ws->workspace <= ptr) && (ptr <= ws->workspaceEnd);
 }
 
 /**
  * Internal function. Do not use directly.
  */
-MEM_STATIC void* ZSTD_cwksp_reserve_internal(
-        ZSTD_cwksp* ws, size_t bytes, ZSTD_cwksp_alloc_phase_e phase) {
+MEM_STATIC void*
+ZSTD_cwksp_reserve_internal(ZSTD_cwksp* ws, size_t bytes, ZSTD_cwksp_alloc_phase_e phase)
+{
     void* alloc;
-    void* bottom = ws->tableEnd;
-    ZSTD_cwksp_internal_advance_phase(ws, phase);
-    alloc = (BYTE *)ws->allocStart - bytes;
+    if (ZSTD_isError(ZSTD_cwksp_internal_advance_phase(ws, phase)) || bytes == 0) {
+        return NULL;
+    }
 
-#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
+#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
     /* over-reserve space */
-    alloc = (BYTE *)alloc - 2 * ZSTD_CWKSP_ASAN_REDZONE_SIZE;
+    bytes += 2 * ZSTD_CWKSP_ASAN_REDZONE_SIZE;
 #endif
 
-    DEBUGLOG(5, "cwksp: reserving %p %zd bytes, %zd bytes remaining",
-        alloc, bytes, ZSTD_cwksp_available_space(ws) - bytes);
-    ZSTD_cwksp_assert_internal_consistency(ws);
-    assert(alloc >= bottom);
-    if (alloc < bottom) {
-        DEBUGLOG(4, "cwksp: alloc failed!");
-        ws->allocFailed = 1;
-        return NULL;
-    }
-    if (alloc < ws->tableValidEnd) {
-        ws->tableValidEnd = alloc;
-    }
-    ws->allocStart = alloc;
+    alloc = ZSTD_cwksp_reserve_internal_buffer_space(ws, bytes);
 
-#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
+#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
     /* Move alloc so there's ZSTD_CWKSP_ASAN_REDZONE_SIZE unused space on
      * either size. */
-    alloc = (BYTE *)alloc + ZSTD_CWKSP_ASAN_REDZONE_SIZE;
-    __asan_unpoison_memory_region(alloc, bytes);
+    if (alloc) {
+        alloc = (BYTE *)alloc + ZSTD_CWKSP_ASAN_REDZONE_SIZE;
+        if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) {
+            /* We need to keep the redzone poisoned while unpoisoning the bytes that
+             * are actually allocated. */
+            __asan_unpoison_memory_region(alloc, bytes - 2 * ZSTD_CWKSP_ASAN_REDZONE_SIZE);
+        }
+    }
 #endif
 
     return alloc;
@@ -260,33 +361,44 @@ MEM_STATIC void* ZSTD_cwksp_reserve_internal(
 /**
  * Reserves and returns unaligned memory.
  */
-MEM_STATIC BYTE* ZSTD_cwksp_reserve_buffer(ZSTD_cwksp* ws, size_t bytes) {
+MEM_STATIC BYTE* ZSTD_cwksp_reserve_buffer(ZSTD_cwksp* ws, size_t bytes)
+{
     return (BYTE*)ZSTD_cwksp_reserve_internal(ws, bytes, ZSTD_cwksp_alloc_buffers);
 }
 
 /**
- * Reserves and returns memory sized on and aligned on sizeof(unsigned).
+ * Reserves and returns memory sized on and aligned on ZSTD_CWKSP_ALIGNMENT_BYTES (64 bytes).
  */
-MEM_STATIC void* ZSTD_cwksp_reserve_aligned(ZSTD_cwksp* ws, size_t bytes) {
-    assert((bytes & (sizeof(U32)-1)) == 0);
-    return ZSTD_cwksp_reserve_internal(ws, ZSTD_cwksp_align(bytes, sizeof(U32)), ZSTD_cwksp_alloc_aligned);
+MEM_STATIC void* ZSTD_cwksp_reserve_aligned(ZSTD_cwksp* ws, size_t bytes)
+{
+    void* ptr = ZSTD_cwksp_reserve_internal(ws, ZSTD_cwksp_align(bytes, ZSTD_CWKSP_ALIGNMENT_BYTES),
+                                            ZSTD_cwksp_alloc_aligned);
+    assert(((size_t)ptr & (ZSTD_CWKSP_ALIGNMENT_BYTES-1))== 0);
+    return ptr;
 }
 
 /**
- * Aligned on sizeof(unsigned). These buffers have the special property that
+ * Aligned on 64 bytes. These buffers have the special property that
  * their values remain constrained, allowing us to re-use them without
  * memset()-ing them.
  */
-MEM_STATIC void* ZSTD_cwksp_reserve_table(ZSTD_cwksp* ws, size_t bytes) {
+MEM_STATIC void* ZSTD_cwksp_reserve_table(ZSTD_cwksp* ws, size_t bytes)
+{
     const ZSTD_cwksp_alloc_phase_e phase = ZSTD_cwksp_alloc_aligned;
-    void* alloc = ws->tableEnd;
-    void* end = (BYTE *)alloc + bytes;
-    void* top = ws->allocStart;
+    void* alloc;
+    void* end;
+    void* top;
+
+    if (ZSTD_isError(ZSTD_cwksp_internal_advance_phase(ws, phase))) {
+        return NULL;
+    }
+    alloc = ws->tableEnd;
+    end = (BYTE *)alloc + bytes;
+    top = ws->allocStart;
 
     DEBUGLOG(5, "cwksp: reserving %p table %zd bytes, %zd bytes remaining",
         alloc, bytes, ZSTD_cwksp_available_space(ws) - bytes);
     assert((bytes & (sizeof(U32)-1)) == 0);
-    ZSTD_cwksp_internal_advance_phase(ws, phase);
     ZSTD_cwksp_assert_internal_consistency(ws);
     assert(end <= top);
     if (end > top) {
@@ -296,35 +408,41 @@ MEM_STATIC void* ZSTD_cwksp_reserve_table(ZSTD_cwksp* ws, size_t bytes) {
     }
     ws->tableEnd = end;
 
-#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
-    __asan_unpoison_memory_region(alloc, bytes);
+#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
+    if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) {
+        __asan_unpoison_memory_region(alloc, bytes);
+    }
 #endif
 
+    assert((bytes & (ZSTD_CWKSP_ALIGNMENT_BYTES-1)) == 0);
+    assert(((size_t)alloc & (ZSTD_CWKSP_ALIGNMENT_BYTES-1))== 0);
     return alloc;
 }
 
 /**
  * Aligned on sizeof(void*).
+ * Note : should happen only once, at workspace first initialization
  */
-MEM_STATIC void* ZSTD_cwksp_reserve_object(ZSTD_cwksp* ws, size_t bytes) {
-    size_t roundedBytes = ZSTD_cwksp_align(bytes, sizeof(void*));
+MEM_STATIC void* ZSTD_cwksp_reserve_object(ZSTD_cwksp* ws, size_t bytes)
+{
+    size_t const roundedBytes = ZSTD_cwksp_align(bytes, sizeof(void*));
     void* alloc = ws->objectEnd;
     void* end = (BYTE*)alloc + roundedBytes;
 
-#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
+#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
     /* over-reserve space */
     end = (BYTE *)end + 2 * ZSTD_CWKSP_ASAN_REDZONE_SIZE;
 #endif
 
-    DEBUGLOG(5,
+    DEBUGLOG(4,
         "cwksp: reserving %p object %zd bytes (rounded to %zd), %zd bytes remaining",
         alloc, bytes, roundedBytes, ZSTD_cwksp_available_space(ws) - roundedBytes);
-    assert(((size_t)alloc & (sizeof(void*)-1)) == 0);
-    assert((bytes & (sizeof(void*)-1)) == 0);
+    assert((size_t)alloc % ZSTD_ALIGNOF(void*) == 0);
+    assert(bytes % ZSTD_ALIGNOF(void*) == 0);
     ZSTD_cwksp_assert_internal_consistency(ws);
     /* we must be in the first phase, no advance is possible */
     if (ws->phase != ZSTD_cwksp_alloc_objects || end > ws->workspaceEnd) {
-        DEBUGLOG(4, "cwksp: object alloc failed!");
+        DEBUGLOG(3, "cwksp: object alloc failed!");
         ws->allocFailed = 1;
         return NULL;
     }
@@ -332,20 +450,23 @@ MEM_STATIC void* ZSTD_cwksp_reserve_object(ZSTD_cwksp* ws, size_t bytes) {
     ws->tableEnd = end;
     ws->tableValidEnd = end;
 
-#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
+#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
     /* Move alloc so there's ZSTD_CWKSP_ASAN_REDZONE_SIZE unused space on
      * either size. */
-    alloc = (BYTE *)alloc + ZSTD_CWKSP_ASAN_REDZONE_SIZE;
-    __asan_unpoison_memory_region(alloc, bytes);
+    alloc = (BYTE*)alloc + ZSTD_CWKSP_ASAN_REDZONE_SIZE;
+    if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) {
+        __asan_unpoison_memory_region(alloc, bytes);
+    }
 #endif
 
     return alloc;
 }
 
-MEM_STATIC void ZSTD_cwksp_mark_tables_dirty(ZSTD_cwksp* ws) {
+MEM_STATIC void ZSTD_cwksp_mark_tables_dirty(ZSTD_cwksp* ws)
+{
     DEBUGLOG(4, "cwksp: ZSTD_cwksp_mark_tables_dirty");
 
-#if defined (MEMORY_SANITIZER) && !defined (ZSTD_MSAN_DONT_POISON_WORKSPACE)
+#if ZSTD_MEMORY_SANITIZER && !defined (ZSTD_MSAN_DONT_POISON_WORKSPACE)
     /* To validate that the table re-use logic is sound, and that we don't
      * access table space that we haven't cleaned, we re-"poison" the table
      * space every time we mark it dirty. */
@@ -380,7 +501,7 @@ MEM_STATIC void ZSTD_cwksp_clean_tables(ZSTD_cwksp* ws) {
     assert(ws->tableValidEnd >= ws->objectEnd);
     assert(ws->tableValidEnd <= ws->allocStart);
     if (ws->tableValidEnd < ws->tableEnd) {
-        memset(ws->tableValidEnd, 0, (BYTE*)ws->tableEnd - (BYTE*)ws->tableValidEnd);
+        ZSTD_memset(ws->tableValidEnd, 0, (size_t)((BYTE*)ws->tableEnd - (BYTE*)ws->tableValidEnd));
     }
     ZSTD_cwksp_mark_tables_clean(ws);
 }
@@ -392,8 +513,12 @@ MEM_STATIC void ZSTD_cwksp_clean_tables(ZSTD_cwksp* ws) {
 MEM_STATIC void ZSTD_cwksp_clear_tables(ZSTD_cwksp* ws) {
     DEBUGLOG(4, "cwksp: clearing tables!");
 
-#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
-    {
+#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
+    /* We don't do this when the workspace is statically allocated, because
+     * when that is the case, we have no capability to hook into the end of the
+     * workspace's lifecycle to unpoison the memory.
+     */
+    if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) {
         size_t size = (BYTE*)ws->tableValidEnd - (BYTE*)ws->objectEnd;
         __asan_poison_memory_region(ws->objectEnd, size);
     }
@@ -410,7 +535,7 @@ MEM_STATIC void ZSTD_cwksp_clear_tables(ZSTD_cwksp* ws) {
 MEM_STATIC void ZSTD_cwksp_clear(ZSTD_cwksp* ws) {
     DEBUGLOG(4, "cwksp: clearing!");
 
-#if defined (MEMORY_SANITIZER) && !defined (ZSTD_MSAN_DONT_POISON_WORKSPACE)
+#if ZSTD_MEMORY_SANITIZER && !defined (ZSTD_MSAN_DONT_POISON_WORKSPACE)
     /* To validate that the context re-use logic is sound, and that we don't
      * access stuff that this compression hasn't initialized, we re-"poison"
      * the workspace (or at least the non-static, non-table parts of it)
@@ -421,8 +546,12 @@ MEM_STATIC void ZSTD_cwksp_clear(ZSTD_cwksp* ws) {
     }
 #endif
 
-#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
-    {
+#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
+    /* We don't do this when the workspace is statically allocated, because
+     * when that is the case, we have no capability to hook into the end of the
+     * workspace's lifecycle to unpoison the memory.
+     */
+    if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) {
         size_t size = (BYTE*)ws->workspaceEnd - (BYTE*)ws->objectEnd;
         __asan_poison_memory_region(ws->objectEnd, size);
     }
@@ -442,7 +571,7 @@ MEM_STATIC void ZSTD_cwksp_clear(ZSTD_cwksp* ws) {
  * Any existing values in the workspace are ignored (the previously managed
  * buffer, if present, must be separately freed).
  */
-MEM_STATIC void ZSTD_cwksp_init(ZSTD_cwksp* ws, void* start, size_t size) {
+MEM_STATIC void ZSTD_cwksp_init(ZSTD_cwksp* ws, void* start, size_t size, ZSTD_cwksp_static_alloc_e isStatic) {
     DEBUGLOG(4, "cwksp: init'ing workspace with %zd bytes", size);
     assert(((size_t)start & (sizeof(void*)-1)) == 0); /* ensure correct alignment */
     ws->workspace = start;
@@ -450,39 +579,45 @@ MEM_STATIC void ZSTD_cwksp_init(ZSTD_cwksp* ws, void* start, size_t size) {
     ws->objectEnd = ws->workspace;
     ws->tableValidEnd = ws->objectEnd;
     ws->phase = ZSTD_cwksp_alloc_objects;
+    ws->isStatic = isStatic;
     ZSTD_cwksp_clear(ws);
     ws->workspaceOversizedDuration = 0;
     ZSTD_cwksp_assert_internal_consistency(ws);
 }
 
 MEM_STATIC size_t ZSTD_cwksp_create(ZSTD_cwksp* ws, size_t size, ZSTD_customMem customMem) {
-    void* workspace = ZSTD_malloc(size, customMem);
+    void* workspace = ZSTD_customMalloc(size, customMem);
     DEBUGLOG(4, "cwksp: creating new workspace with %zd bytes", size);
     RETURN_ERROR_IF(workspace == NULL, memory_allocation, "NULL pointer!");
-    ZSTD_cwksp_init(ws, workspace, size);
+    ZSTD_cwksp_init(ws, workspace, size, ZSTD_cwksp_dynamic_alloc);
     return 0;
 }
 
 MEM_STATIC void ZSTD_cwksp_free(ZSTD_cwksp* ws, ZSTD_customMem customMem) {
     void *ptr = ws->workspace;
     DEBUGLOG(4, "cwksp: freeing workspace");
-    memset(ws, 0, sizeof(ZSTD_cwksp));
-    ZSTD_free(ptr, customMem);
+    ZSTD_memset(ws, 0, sizeof(ZSTD_cwksp));
+    ZSTD_customFree(ptr, customMem);
 }
 
 /**
  * Moves the management of a workspace from one cwksp to another. The src cwksp
- * is left in an invalid state (src must be re-init()'ed before its used again).
+ * is left in an invalid state (src must be re-init()'ed before it's used again).
  */
 MEM_STATIC void ZSTD_cwksp_move(ZSTD_cwksp* dst, ZSTD_cwksp* src) {
     *dst = *src;
-    memset(src, 0, sizeof(ZSTD_cwksp));
+    ZSTD_memset(src, 0, sizeof(ZSTD_cwksp));
 }
 
 MEM_STATIC size_t ZSTD_cwksp_sizeof(const ZSTD_cwksp* ws) {
     return (size_t)((BYTE*)ws->workspaceEnd - (BYTE*)ws->workspace);
 }
 
+MEM_STATIC size_t ZSTD_cwksp_used(const ZSTD_cwksp* ws) {
+    return (size_t)((BYTE*)ws->tableEnd - (BYTE*)ws->workspace)
+         + (size_t)((BYTE*)ws->workspaceEnd - (BYTE*)ws->allocStart);
+}
+
 MEM_STATIC int ZSTD_cwksp_reserve_failed(const ZSTD_cwksp* ws) {
     return ws->allocFailed;
 }
@@ -491,6 +626,24 @@ MEM_STATIC int ZSTD_cwksp_reserve_failed(const ZSTD_cwksp* ws) {
 *  Functions Checking Free Space
 ***************************************/
 
+/* ZSTD_alignmentSpaceWithinBounds() :
+ * Returns if the estimated space needed for a wksp is within an acceptable limit of the
+ * actual amount of space used.
+ */
+MEM_STATIC int ZSTD_cwksp_estimated_space_within_bounds(const ZSTD_cwksp* const ws,
+                                                        size_t const estimatedSpace, int resizedWorkspace) {
+    if (resizedWorkspace) {
+        /* Resized/newly allocated wksp should have exact bounds */
+        return ZSTD_cwksp_used(ws) == estimatedSpace;
+    } else {
+        /* Due to alignment, when reusing a workspace, we can actually consume 63 fewer or more bytes
+         * than estimatedSpace. See the comments in zstd_cwksp.h for details.
+         */
+        return (ZSTD_cwksp_used(ws) >= estimatedSpace - 63) && (ZSTD_cwksp_used(ws) <= estimatedSpace + 63);
+    }
+}
+
+
 MEM_STATIC size_t ZSTD_cwksp_available_space(ZSTD_cwksp* ws) {
     return (size_t)((BYTE*)ws->allocStart - (BYTE*)ws->tableEnd);
 }
diff --git a/contrib/zstd/zstd_ddict.c b/contrib/zstd/zstd_ddict.c
index 47e985fca..06bd996c0 100644
--- a/contrib/zstd/zstd_ddict.c
+++ b/contrib/zstd/zstd_ddict.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -14,12 +14,11 @@
 /*-*******************************************************
 *  Dependencies
 *********************************************************/
-#include <string.h>      /* memcpy, memmove, memset */
+#include "zstd_deps.h"   /* ZSTD_memcpy, ZSTD_memmove, ZSTD_memset */
 #include "cpu.h"         /* bmi2 */
 #include "mem.h"         /* low level memory routines */
 #define FSE_STATIC_LINKING_ONLY
 #include "fse.h"
-#define HUF_STATIC_LINKING_ONLY
 #include "huf.h"
 #include "zstd_decompress_internal.h"
 #include "zstd_ddict.h"
@@ -127,14 +126,14 @@ static size_t ZSTD_initDDict_internal(ZSTD_DDict* ddict,
         ddict->dictContent = dict;
         if (!dict) dictSize = 0;
     } else {
-        void* const internalBuffer = ZSTD_malloc(dictSize, ddict->cMem);
+        void* const internalBuffer = ZSTD_customMalloc(dictSize, ddict->cMem);
         ddict->dictBuffer = internalBuffer;
         ddict->dictContent = internalBuffer;
         if (!internalBuffer) return ERROR(memory_allocation);
-        memcpy(internalBuffer, dict, dictSize);
+        ZSTD_memcpy(internalBuffer, dict, dictSize);
     }
     ddict->dictSize = dictSize;
-    ddict->entropy.hufTable[0] = (HUF_DTable)((HufLog)*0x1000001);  /* cover both little and big endian */
+    ddict->entropy.hufTable[0] = (HUF_DTable)((ZSTD_HUFFDTABLE_CAPACITY_LOG)*0x1000001);  /* cover both little and big endian */
 
     /* parse dictionary content */
     FORWARD_IF_ERROR( ZSTD_loadEntropy_intoDDict(ddict, dictContentType) , "");
@@ -147,9 +146,9 @@ ZSTD_DDict* ZSTD_createDDict_advanced(const void* dict, size_t dictSize,
                                       ZSTD_dictContentType_e dictContentType,
                                       ZSTD_customMem customMem)
 {
-    if (!customMem.customAlloc ^ !customMem.customFree) return NULL;
+    if ((!customMem.customAlloc) ^ (!customMem.customFree)) return NULL;
 
-    {   ZSTD_DDict* const ddict = (ZSTD_DDict*) ZSTD_malloc(sizeof(ZSTD_DDict), customMem);
+    {   ZSTD_DDict* const ddict = (ZSTD_DDict*) ZSTD_customMalloc(sizeof(ZSTD_DDict), customMem);
         if (ddict == NULL) return NULL;
         ddict->cMem = customMem;
         {   size_t const initResult = ZSTD_initDDict_internal(ddict,
@@ -198,7 +197,7 @@ const ZSTD_DDict* ZSTD_initStaticDDict(
     if ((size_t)sBuffer & 7) return NULL;   /* 8-aligned */
     if (sBufferSize < neededSpace) return NULL;
     if (dictLoadMethod == ZSTD_dlm_byCopy) {
-        memcpy(ddict+1, dict, dictSize);  /* local copy */
+        ZSTD_memcpy(ddict+1, dict, dictSize);  /* local copy */
         dict = ddict+1;
     }
     if (ZSTD_isError( ZSTD_initDDict_internal(ddict,
@@ -213,8 +212,8 @@ size_t ZSTD_freeDDict(ZSTD_DDict* ddict)
 {
     if (ddict==NULL) return 0;   /* support free on NULL */
     {   ZSTD_customMem const cMem = ddict->cMem;
-        ZSTD_free(ddict->dictBuffer, cMem);
-        ZSTD_free(ddict, cMem);
+        ZSTD_customFree(ddict->dictBuffer, cMem);
+        ZSTD_customFree(ddict, cMem);
         return 0;
     }
 }
@@ -240,5 +239,5 @@ size_t ZSTD_sizeof_DDict(const ZSTD_DDict* ddict)
 unsigned ZSTD_getDictID_fromDDict(const ZSTD_DDict* ddict)
 {
     if (ddict==NULL) return 0;
-    return ZSTD_getDictID_fromDict(ddict->dictContent, ddict->dictSize);
+    return ddict->dictID;
 }
diff --git a/contrib/zstd/zstd_ddict.h b/contrib/zstd/zstd_ddict.h
index 3ab7532ee..f49c57037 100644
--- a/contrib/zstd/zstd_ddict.h
+++ b/contrib/zstd/zstd_ddict.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -15,7 +15,7 @@
 /*-*******************************************************
  *  Dependencies
  *********************************************************/
-#include <stddef.h>   /* size_t */
+#include "zstd_deps.h"   /* size_t */
 #include "zstd.h"     /* ZSTD_DDict, and several public functions */
 
 
diff --git a/contrib/zstd/zstd_decompress.c b/contrib/zstd/zstd_decompress.c
index e466647f0..05704ce4d 100644
--- a/contrib/zstd/zstd_decompress.c
+++ b/contrib/zstd/zstd_decompress.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -55,23 +55,164 @@
 /*-*******************************************************
 *  Dependencies
 *********************************************************/
-#include <string.h>      /* memcpy, memmove, memset */
-#include "cpu.h"         /* bmi2 */
+#include "zstd_deps.h"   /* ZSTD_memcpy, ZSTD_memmove, ZSTD_memset */
 #include "mem.h"         /* low level memory routines */
 #define FSE_STATIC_LINKING_ONLY
 #include "fse.h"
-#define HUF_STATIC_LINKING_ONLY
 #include "huf.h"
+#include "xxhash.h" /* XXH64_reset, XXH64_update, XXH64_digest, XXH64 */
 #include "zstd_internal.h"  /* blockProperties_t */
 #include "zstd_decompress_internal.h"   /* ZSTD_DCtx */
 #include "zstd_ddict.h"  /* ZSTD_DDictDictContent */
 #include "zstd_decompress_block.h"   /* ZSTD_decompressBlock_internal */
+#include "bits.h"  /* ZSTD_highbit32 */
 
 #if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
 #  include "../legacy/zstd_legacy.h"
 #endif
 
 
+
+/*************************************
+ * Multiple DDicts Hashset internals *
+ *************************************/
+
+#define DDICT_HASHSET_MAX_LOAD_FACTOR_COUNT_MULT 4
+#define DDICT_HASHSET_MAX_LOAD_FACTOR_SIZE_MULT 3  /* These two constants represent SIZE_MULT/COUNT_MULT load factor without using a float.
+                                                    * Currently, that means a 0.75 load factor.
+                                                    * So, if count * COUNT_MULT / size * SIZE_MULT != 0, then we've exceeded
+                                                    * the load factor of the ddict hash set.
+                                                    */
+
+#define DDICT_HASHSET_TABLE_BASE_SIZE 64
+#define DDICT_HASHSET_RESIZE_FACTOR 2
+
+/* Hash function to determine starting position of dict insertion within the table
+ * Returns an index between [0, hashSet->ddictPtrTableSize]
+ */
+static size_t ZSTD_DDictHashSet_getIndex(const ZSTD_DDictHashSet* hashSet, U32 dictID) {
+    const U64 hash = XXH64(&dictID, sizeof(U32), 0);
+    /* DDict ptr table size is a multiple of 2, use size - 1 as mask to get index within [0, hashSet->ddictPtrTableSize) */
+    return hash & (hashSet->ddictPtrTableSize - 1);
+}
+
+/* Adds DDict to a hashset without resizing it.
+ * If inserting a DDict with a dictID that already exists in the set, replaces the one in the set.
+ * Returns 0 if successful, or a zstd error code if something went wrong.
+ */
+static size_t ZSTD_DDictHashSet_emplaceDDict(ZSTD_DDictHashSet* hashSet, const ZSTD_DDict* ddict) {
+    const U32 dictID = ZSTD_getDictID_fromDDict(ddict);
+    size_t idx = ZSTD_DDictHashSet_getIndex(hashSet, dictID);
+    const size_t idxRangeMask = hashSet->ddictPtrTableSize - 1;
+    RETURN_ERROR_IF(hashSet->ddictPtrCount == hashSet->ddictPtrTableSize, GENERIC, "Hash set is full!");
+    DEBUGLOG(4, "Hashed index: for dictID: %u is %zu", dictID, idx);
+    while (hashSet->ddictPtrTable[idx] != NULL) {
+        /* Replace existing ddict if inserting ddict with same dictID */
+        if (ZSTD_getDictID_fromDDict(hashSet->ddictPtrTable[idx]) == dictID) {
+            DEBUGLOG(4, "DictID already exists, replacing rather than adding");
+            hashSet->ddictPtrTable[idx] = ddict;
+            return 0;
+        }
+        idx &= idxRangeMask;
+        idx++;
+    }
+    DEBUGLOG(4, "Final idx after probing for dictID %u is: %zu", dictID, idx);
+    hashSet->ddictPtrTable[idx] = ddict;
+    hashSet->ddictPtrCount++;
+    return 0;
+}
+
+/* Expands hash table by factor of DDICT_HASHSET_RESIZE_FACTOR and
+ * rehashes all values, allocates new table, frees old table.
+ * Returns 0 on success, otherwise a zstd error code.
+ */
+static size_t ZSTD_DDictHashSet_expand(ZSTD_DDictHashSet* hashSet, ZSTD_customMem customMem) {
+    size_t newTableSize = hashSet->ddictPtrTableSize * DDICT_HASHSET_RESIZE_FACTOR;
+    const ZSTD_DDict** newTable = (const ZSTD_DDict**)ZSTD_customCalloc(sizeof(ZSTD_DDict*) * newTableSize, customMem);
+    const ZSTD_DDict** oldTable = hashSet->ddictPtrTable;
+    size_t oldTableSize = hashSet->ddictPtrTableSize;
+    size_t i;
+
+    DEBUGLOG(4, "Expanding DDict hash table! Old size: %zu new size: %zu", oldTableSize, newTableSize);
+    RETURN_ERROR_IF(!newTable, memory_allocation, "Expanded hashset allocation failed!");
+    hashSet->ddictPtrTable = newTable;
+    hashSet->ddictPtrTableSize = newTableSize;
+    hashSet->ddictPtrCount = 0;
+    for (i = 0; i < oldTableSize; ++i) {
+        if (oldTable[i] != NULL) {
+            FORWARD_IF_ERROR(ZSTD_DDictHashSet_emplaceDDict(hashSet, oldTable[i]), "");
+        }
+    }
+    ZSTD_customFree((void*)oldTable, customMem);
+    DEBUGLOG(4, "Finished re-hash");
+    return 0;
+}
+
+/* Fetches a DDict with the given dictID
+ * Returns the ZSTD_DDict* with the requested dictID. If it doesn't exist, then returns NULL.
+ */
+static const ZSTD_DDict* ZSTD_DDictHashSet_getDDict(ZSTD_DDictHashSet* hashSet, U32 dictID) {
+    size_t idx = ZSTD_DDictHashSet_getIndex(hashSet, dictID);
+    const size_t idxRangeMask = hashSet->ddictPtrTableSize - 1;
+    DEBUGLOG(4, "Hashed index: for dictID: %u is %zu", dictID, idx);
+    for (;;) {
+        size_t currDictID = ZSTD_getDictID_fromDDict(hashSet->ddictPtrTable[idx]);
+        if (currDictID == dictID || currDictID == 0) {
+            /* currDictID == 0 implies a NULL ddict entry */
+            break;
+        } else {
+            idx &= idxRangeMask;    /* Goes to start of table when we reach the end */
+            idx++;
+        }
+    }
+    DEBUGLOG(4, "Final idx after probing for dictID %u is: %zu", dictID, idx);
+    return hashSet->ddictPtrTable[idx];
+}
+
+/* Allocates space for and returns a ddict hash set
+ * The hash set's ZSTD_DDict* table has all values automatically set to NULL to begin with.
+ * Returns NULL if allocation failed.
+ */
+static ZSTD_DDictHashSet* ZSTD_createDDictHashSet(ZSTD_customMem customMem) {
+    ZSTD_DDictHashSet* ret = (ZSTD_DDictHashSet*)ZSTD_customMalloc(sizeof(ZSTD_DDictHashSet), customMem);
+    DEBUGLOG(4, "Allocating new hash set");
+    if (!ret)
+        return NULL;
+    ret->ddictPtrTable = (const ZSTD_DDict**)ZSTD_customCalloc(DDICT_HASHSET_TABLE_BASE_SIZE * sizeof(ZSTD_DDict*), customMem);
+    if (!ret->ddictPtrTable) {
+        ZSTD_customFree(ret, customMem);
+        return NULL;
+    }
+    ret->ddictPtrTableSize = DDICT_HASHSET_TABLE_BASE_SIZE;
+    ret->ddictPtrCount = 0;
+    return ret;
+}
+
+/* Frees the table of ZSTD_DDict* within a hashset, then frees the hashset itself.
+ * Note: The ZSTD_DDict* within the table are NOT freed.
+ */
+static void ZSTD_freeDDictHashSet(ZSTD_DDictHashSet* hashSet, ZSTD_customMem customMem) {
+    DEBUGLOG(4, "Freeing ddict hash set");
+    if (hashSet && hashSet->ddictPtrTable) {
+        ZSTD_customFree((void*)hashSet->ddictPtrTable, customMem);
+    }
+    if (hashSet) {
+        ZSTD_customFree(hashSet, customMem);
+    }
+}
+
+/* Public function: Adds a DDict into the ZSTD_DDictHashSet, possibly triggering a resize of the hash set.
+ * Returns 0 on success, or a ZSTD error.
+ */
+static size_t ZSTD_DDictHashSet_addDDict(ZSTD_DDictHashSet* hashSet, const ZSTD_DDict* ddict, ZSTD_customMem customMem) {
+    DEBUGLOG(4, "Adding dict ID: %u to hashset with - Count: %zu Tablesize: %zu", ZSTD_getDictID_fromDDict(ddict), hashSet->ddictPtrCount, hashSet->ddictPtrTableSize);
+    if (hashSet->ddictPtrCount * DDICT_HASHSET_MAX_LOAD_FACTOR_COUNT_MULT / hashSet->ddictPtrTableSize * DDICT_HASHSET_MAX_LOAD_FACTOR_SIZE_MULT != 0) {
+        FORWARD_IF_ERROR(ZSTD_DDictHashSet_expand(hashSet, customMem), "");
+    }
+    FORWARD_IF_ERROR(ZSTD_DDictHashSet_emplaceDDict(hashSet, ddict), "");
+    return 0;
+}
+
 /*-*************************************************************
 *   Context management
 ***************************************************************/
@@ -94,11 +235,20 @@ static size_t ZSTD_startingInputLength(ZSTD_format_e format)
     return startingInputLength;
 }
 
+static void ZSTD_DCtx_resetParameters(ZSTD_DCtx* dctx)
+{
+    assert(dctx->streamStage == zdss_init);
+    dctx->format = ZSTD_f_zstd1;
+    dctx->maxWindowSize = ZSTD_MAXWINDOWSIZE_DEFAULT;
+    dctx->outBufferMode = ZSTD_bm_buffered;
+    dctx->forceIgnoreChecksum = ZSTD_d_validateChecksum;
+    dctx->refMultipleDDicts = ZSTD_rmd_refSingleDDict;
+    dctx->disableHufAsm = 0;
+}
+
 static void ZSTD_initDCtx_internal(ZSTD_DCtx* dctx)
 {
-    dctx->format = ZSTD_f_zstd1;  /* ZSTD_decompressBegin() invokes ZSTD_startingInputLength() with argument dctx->format */
     dctx->staticSize  = 0;
-    dctx->maxWindowSize = ZSTD_MAXWINDOWSIZE_DEFAULT;
     dctx->ddict       = NULL;
     dctx->ddictLocal  = NULL;
     dctx->dictEnd     = NULL;
@@ -108,12 +258,17 @@ static void ZSTD_initDCtx_internal(ZSTD_DCtx* dctx)
     dctx->inBuffSize  = 0;
     dctx->outBuffSize = 0;
     dctx->streamStage = zdss_init;
+#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
     dctx->legacyContext = NULL;
     dctx->previousLegacyVersion = 0;
+#endif
     dctx->noForwardProgress = 0;
     dctx->oversizedDuration = 0;
-    dctx->bmi2 = ZSTD_cpuid_bmi2(ZSTD_cpuid());
-    dctx->outBufferMode = ZSTD_obm_buffered;
+#if DYNAMIC_BMI2
+    dctx->bmi2 = ZSTD_cpuSupportsBmi2();
+#endif
+    dctx->ddictSet = NULL;
+    ZSTD_DCtx_resetParameters(dctx);
 #ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
     dctx->dictContentEndForFuzzing = NULL;
 #endif
@@ -132,11 +287,10 @@ ZSTD_DCtx* ZSTD_initStaticDCtx(void *workspace, size_t workspaceSize)
     return dctx;
 }
 
-ZSTD_DCtx* ZSTD_createDCtx_advanced(ZSTD_customMem customMem)
-{
-    if (!customMem.customAlloc ^ !customMem.customFree) return NULL;
+static ZSTD_DCtx* ZSTD_createDCtx_internal(ZSTD_customMem customMem) {
+    if ((!customMem.customAlloc) ^ (!customMem.customFree)) return NULL;
 
-    {   ZSTD_DCtx* const dctx = (ZSTD_DCtx*)ZSTD_malloc(sizeof(*dctx), customMem);
+    {   ZSTD_DCtx* const dctx = (ZSTD_DCtx*)ZSTD_customMalloc(sizeof(*dctx), customMem);
         if (!dctx) return NULL;
         dctx->customMem = customMem;
         ZSTD_initDCtx_internal(dctx);
@@ -144,10 +298,15 @@ ZSTD_DCtx* ZSTD_createDCtx_advanced(ZSTD_customMem customMem)
     }
 }
 
+ZSTD_DCtx* ZSTD_createDCtx_advanced(ZSTD_customMem customMem)
+{
+    return ZSTD_createDCtx_internal(customMem);
+}
+
 ZSTD_DCtx* ZSTD_createDCtx(void)
 {
     DEBUGLOG(3, "ZSTD_createDCtx");
-    return ZSTD_createDCtx_advanced(ZSTD_defaultCMem);
+    return ZSTD_createDCtx_internal(ZSTD_defaultCMem);
 }
 
 static void ZSTD_clearDict(ZSTD_DCtx* dctx)
@@ -164,13 +323,17 @@ size_t ZSTD_freeDCtx(ZSTD_DCtx* dctx)
     RETURN_ERROR_IF(dctx->staticSize, memory_allocation, "not compatible with static DCtx");
     {   ZSTD_customMem const cMem = dctx->customMem;
         ZSTD_clearDict(dctx);
-        ZSTD_free(dctx->inBuff, cMem);
+        ZSTD_customFree(dctx->inBuff, cMem);
         dctx->inBuff = NULL;
 #if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT >= 1)
         if (dctx->legacyContext)
             ZSTD_freeLegacyStreamContext(dctx->legacyContext, dctx->previousLegacyVersion);
 #endif
-        ZSTD_free(dctx, cMem);
+        if (dctx->ddictSet) {
+            ZSTD_freeDDictHashSet(dctx->ddictSet, cMem);
+            dctx->ddictSet = NULL;
+        }
+        ZSTD_customFree(dctx, cMem);
         return 0;
     }
 }
@@ -179,7 +342,30 @@ size_t ZSTD_freeDCtx(ZSTD_DCtx* dctx)
 void ZSTD_copyDCtx(ZSTD_DCtx* dstDCtx, const ZSTD_DCtx* srcDCtx)
 {
     size_t const toCopy = (size_t)((char*)(&dstDCtx->inBuff) - (char*)dstDCtx);
-    memcpy(dstDCtx, srcDCtx, toCopy);  /* no need to copy workspace */
+    ZSTD_memcpy(dstDCtx, srcDCtx, toCopy);  /* no need to copy workspace */
+}
+
+/* Given a dctx with a digested frame params, re-selects the correct ZSTD_DDict based on
+ * the requested dict ID from the frame. If there exists a reference to the correct ZSTD_DDict, then
+ * accordingly sets the ddict to be used to decompress the frame.
+ *
+ * If no DDict is found, then no action is taken, and the ZSTD_DCtx::ddict remains as-is.
+ *
+ * ZSTD_d_refMultipleDDicts must be enabled for this function to be called.
+ */
+static void ZSTD_DCtx_selectFrameDDict(ZSTD_DCtx* dctx) {
+    assert(dctx->refMultipleDDicts && dctx->ddictSet);
+    DEBUGLOG(4, "Adjusting DDict based on requested dict ID from frame");
+    if (dctx->ddict) {
+        const ZSTD_DDict* frameDDict = ZSTD_DDictHashSet_getDDict(dctx->ddictSet, dctx->fParams.dictID);
+        if (frameDDict) {
+            DEBUGLOG(4, "DDict found!");
+            ZSTD_clearDict(dctx);
+            dctx->dictID = dctx->fParams.dictID;
+            dctx->ddict = frameDDict;
+            dctx->dictUses = ZSTD_use_indefinitely;
+        }
+    }
 }
 
 
@@ -205,6 +391,19 @@ unsigned ZSTD_isFrame(const void* buffer, size_t size)
     return 0;
 }
 
+/*! ZSTD_isSkippableFrame() :
+ *  Tells if the content of `buffer` starts with a valid Frame Identifier for a skippable frame.
+ *  Note : Frame Identifier is 4 bytes. If `size < 4`, @return will always be 0.
+ */
+unsigned ZSTD_isSkippableFrame(const void* buffer, size_t size)
+{
+    if (size < ZSTD_FRAMEIDSIZE) return 0;
+    {   U32 const magic = MEM_readLE32(buffer);
+        if ((magic & ZSTD_MAGIC_SKIPPABLE_MASK) == ZSTD_MAGIC_SKIPPABLE_START) return 1;
+    }
+    return 0;
+}
+
 /** ZSTD_frameHeaderSize_internal() :
  *  srcSize must be large enough to reach header size fields.
  *  note : only works for formats ZSTD_f_zstd1 and ZSTD_f_zstd1_magicless.
@@ -240,23 +439,47 @@ size_t ZSTD_frameHeaderSize(const void* src, size_t srcSize)
  *  note : only works for formats ZSTD_f_zstd1 and ZSTD_f_zstd1_magicless
  * @return : 0, `zfhPtr` is correctly filled,
  *          >0, `srcSize` is too small, value is wanted `srcSize` amount,
- *           or an error code, which can be tested using ZSTD_isError() */
+**           or an error code, which can be tested using ZSTD_isError() */
 size_t ZSTD_getFrameHeader_advanced(ZSTD_frameHeader* zfhPtr, const void* src, size_t srcSize, ZSTD_format_e format)
 {
     const BYTE* ip = (const BYTE*)src;
     size_t const minInputSize = ZSTD_startingInputLength(format);
 
-    memset(zfhPtr, 0, sizeof(*zfhPtr));   /* not strictly necessary, but static analyzer do not understand that zfhPtr is only going to be read only if return value is zero, since they are 2 different signals */
-    if (srcSize < minInputSize) return minInputSize;
-    RETURN_ERROR_IF(src==NULL, GENERIC, "invalid parameter");
+    DEBUGLOG(5, "ZSTD_getFrameHeader_advanced: minInputSize = %zu, srcSize = %zu", minInputSize, srcSize);
 
+    if (srcSize > 0) {
+        /* note : technically could be considered an assert(), since it's an invalid entry */
+        RETURN_ERROR_IF(src==NULL, GENERIC, "invalid parameter : src==NULL, but srcSize>0");
+    }
+    if (srcSize < minInputSize) {
+        if (srcSize > 0 && format != ZSTD_f_zstd1_magicless) {
+            /* when receiving less than @minInputSize bytes,
+             * control these bytes at least correspond to a supported magic number
+             * in order to error out early if they don't.
+            **/
+            size_t const toCopy = MIN(4, srcSize);
+            unsigned char hbuf[4]; MEM_writeLE32(hbuf, ZSTD_MAGICNUMBER);
+            assert(src != NULL);
+            ZSTD_memcpy(hbuf, src, toCopy);
+            if ( MEM_readLE32(hbuf) != ZSTD_MAGICNUMBER ) {
+                /* not a zstd frame : let's check if it's a skippable frame */
+                MEM_writeLE32(hbuf, ZSTD_MAGIC_SKIPPABLE_START);
+                ZSTD_memcpy(hbuf, src, toCopy);
+                if ((MEM_readLE32(hbuf) & ZSTD_MAGIC_SKIPPABLE_MASK) != ZSTD_MAGIC_SKIPPABLE_START) {
+                    RETURN_ERROR(prefix_unknown,
+                                "first bytes don't correspond to any supported magic number");
+        }   }   }
+        return minInputSize;
+    }
+
+    ZSTD_memset(zfhPtr, 0, sizeof(*zfhPtr));   /* not strictly necessary, but static analyzers may not understand that zfhPtr will be read only if return value is zero, since they are 2 different signals */
     if ( (format != ZSTD_f_zstd1_magicless)
       && (MEM_readLE32(src) != ZSTD_MAGICNUMBER) ) {
         if ((MEM_readLE32(src) & ZSTD_MAGIC_SKIPPABLE_MASK) == ZSTD_MAGIC_SKIPPABLE_START) {
             /* skippable frame */
             if (srcSize < ZSTD_SKIPPABLEHEADERSIZE)
                 return ZSTD_SKIPPABLEHEADERSIZE; /* magic number + frame length */
-            memset(zfhPtr, 0, sizeof(*zfhPtr));
+            ZSTD_memset(zfhPtr, 0, sizeof(*zfhPtr));
             zfhPtr->frameContentSize = MEM_readLE32((const char *)src + ZSTD_FRAMEIDSIZE);
             zfhPtr->frameType = ZSTD_skippableFrame;
             return 0;
@@ -291,7 +514,9 @@ size_t ZSTD_getFrameHeader_advanced(ZSTD_frameHeader* zfhPtr, const void* src, s
         }
         switch(dictIDSizeCode)
         {
-            default: assert(0);  /* impossible */
+            default:
+                assert(0);  /* impossible */
+                ZSTD_FALLTHROUGH;
             case 0 : break;
             case 1 : dictID = ip[pos]; pos++; break;
             case 2 : dictID = MEM_readLE16(ip+pos); pos+=2; break;
@@ -299,7 +524,9 @@ size_t ZSTD_getFrameHeader_advanced(ZSTD_frameHeader* zfhPtr, const void* src, s
         }
         switch(fcsID)
         {
-            default: assert(0);  /* impossible */
+            default:
+                assert(0);  /* impossible */
+                ZSTD_FALLTHROUGH;
             case 0 : if (singleSegment) frameContentSize = ip[pos]; break;
             case 1 : frameContentSize = MEM_readLE16(ip+pos)+256; break;
             case 2 : frameContentSize = MEM_readLE32(ip+pos); break;
@@ -328,7 +555,6 @@ size_t ZSTD_getFrameHeader(ZSTD_frameHeader* zfhPtr, const void* src, size_t src
     return ZSTD_getFrameHeader_advanced(zfhPtr, src, srcSize, ZSTD_f_zstd1);
 }
 
-
 /** ZSTD_getFrameContentSize() :
  *  compatible with legacy mode
  * @return : decompressed size of the single frame pointed to be `src` if known, otherwise
@@ -369,6 +595,37 @@ static size_t readSkippableFrameSize(void const* src, size_t srcSize)
     }
 }
 
+/*! ZSTD_readSkippableFrame() :
+ * Retrieves a zstd skippable frame containing data given by src, and writes it to dst buffer.
+ *
+ * The parameter magicVariant will receive the magicVariant that was supplied when the frame was written,
+ * i.e. magicNumber - ZSTD_MAGIC_SKIPPABLE_START.  This can be NULL if the caller is not interested
+ * in the magicVariant.
+ *
+ * Returns an error if destination buffer is not large enough, or if the frame is not skippable.
+ *
+ * @return : number of bytes written or a ZSTD error.
+ */
+ZSTDLIB_API size_t ZSTD_readSkippableFrame(void* dst, size_t dstCapacity, unsigned* magicVariant,
+                                            const void* src, size_t srcSize)
+{
+    U32 const magicNumber = MEM_readLE32(src);
+    size_t skippableFrameSize = readSkippableFrameSize(src, srcSize);
+    size_t skippableContentSize = skippableFrameSize - ZSTD_SKIPPABLEHEADERSIZE;
+
+    /* check input validity */
+    RETURN_ERROR_IF(!ZSTD_isSkippableFrame(src, srcSize), frameParameter_unsupported, "");
+    RETURN_ERROR_IF(skippableFrameSize < ZSTD_SKIPPABLEHEADERSIZE || skippableFrameSize > srcSize, srcSize_wrong, "");
+    RETURN_ERROR_IF(skippableContentSize > dstCapacity, dstSize_tooSmall, "");
+
+    /* deliver payload */
+    if (skippableContentSize > 0  && dst != NULL)
+        ZSTD_memcpy(dst, (const BYTE *)src + ZSTD_SKIPPABLEHEADERSIZE, skippableContentSize);
+    if (magicVariant != NULL)
+        *magicVariant = magicNumber - ZSTD_MAGIC_SKIPPABLE_START;
+    return skippableContentSize;
+}
+
 /** ZSTD_findDecompressedSize() :
  *  compatible with legacy mode
  *  `srcSize` must be the exact length of some number of ZSTD compressed and/or
@@ -433,12 +690,19 @@ unsigned long long ZSTD_getDecompressedSize(const void* src, size_t srcSize)
 
 /** ZSTD_decodeFrameHeader() :
  * `headerSize` must be the size provided by ZSTD_frameHeaderSize().
+ * If multiple DDict references are enabled, also will choose the correct DDict to use.
  * @return : 0 if success, or an error code, which can be tested using ZSTD_isError() */
 static size_t ZSTD_decodeFrameHeader(ZSTD_DCtx* dctx, const void* src, size_t headerSize)
 {
     size_t const result = ZSTD_getFrameHeader_advanced(&(dctx->fParams), src, headerSize, dctx->format);
     if (ZSTD_isError(result)) return result;    /* invalid header */
     RETURN_ERROR_IF(result>0, srcSize_wrong, "headerSize too small");
+
+    /* Reference DDict requested by frame if dctx references multiple ddicts */
+    if (dctx->refMultipleDDicts == ZSTD_rmd_refMultipleDDicts && dctx->ddictSet) {
+        ZSTD_DCtx_selectFrameDDict(dctx);
+    }
+
 #ifndef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
     /* Skip the dictID check in fuzzing mode, because it makes the search
      * harder.
@@ -446,7 +710,9 @@ static size_t ZSTD_decodeFrameHeader(ZSTD_DCtx* dctx, const void* src, size_t he
     RETURN_ERROR_IF(dctx->fParams.dictID && (dctx->dictID != dctx->fParams.dictID),
                     dictionary_wrong, "");
 #endif
-    if (dctx->fParams.checksumFlag) XXH64_reset(&dctx->xxhState, 0);
+    dctx->validateChecksum = (dctx->fParams.checksumFlag && !dctx->forceIgnoreChecksum) ? 1 : 0;
+    if (dctx->validateChecksum) XXH64_reset(&dctx->xxhState, 0);
+    dctx->processedCSize += headerSize;
     return 0;
 }
 
@@ -461,7 +727,7 @@ static ZSTD_frameSizeInfo ZSTD_errorFrameSizeInfo(size_t ret)
 static ZSTD_frameSizeInfo ZSTD_findFrameSizeInfo(const void* src, size_t srcSize)
 {
     ZSTD_frameSizeInfo frameSizeInfo;
-    memset(&frameSizeInfo, 0, sizeof(ZSTD_frameSizeInfo));
+    ZSTD_memset(&frameSizeInfo, 0, sizeof(ZSTD_frameSizeInfo));
 
 #if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT >= 1)
     if (ZSTD_isLegacy(src, srcSize))
@@ -516,10 +782,11 @@ static ZSTD_frameSizeInfo ZSTD_findFrameSizeInfo(const void* src, size_t srcSize
             ip += 4;
         }
 
-        frameSizeInfo.compressedSize = ip - ipstart;
+        frameSizeInfo.nbBlocks = nbBlocks;
+        frameSizeInfo.compressedSize = (size_t)(ip - ipstart);
         frameSizeInfo.decompressedBound = (zfh.frameContentSize != ZSTD_CONTENTSIZE_UNKNOWN)
                                         ? zfh.frameContentSize
-                                        : nbBlocks * zfh.blockSizeMax;
+                                        : (unsigned long long)nbBlocks * zfh.blockSizeMax;
         return frameSizeInfo;
     }
 }
@@ -559,6 +826,48 @@ unsigned long long ZSTD_decompressBound(const void* src, size_t srcSize)
     return bound;
 }
 
+size_t ZSTD_decompressionMargin(void const* src, size_t srcSize)
+{
+    size_t margin = 0;
+    unsigned maxBlockSize = 0;
+
+    /* Iterate over each frame */
+    while (srcSize > 0) {
+        ZSTD_frameSizeInfo const frameSizeInfo = ZSTD_findFrameSizeInfo(src, srcSize);
+        size_t const compressedSize = frameSizeInfo.compressedSize;
+        unsigned long long const decompressedBound = frameSizeInfo.decompressedBound;
+        ZSTD_frameHeader zfh;
+
+        FORWARD_IF_ERROR(ZSTD_getFrameHeader(&zfh, src, srcSize), "");
+        if (ZSTD_isError(compressedSize) || decompressedBound == ZSTD_CONTENTSIZE_ERROR)
+            return ERROR(corruption_detected);
+
+        if (zfh.frameType == ZSTD_frame) {
+            /* Add the frame header to our margin */
+            margin += zfh.headerSize;
+            /* Add the checksum to our margin */
+            margin += zfh.checksumFlag ? 4 : 0;
+            /* Add 3 bytes per block */
+            margin += 3 * frameSizeInfo.nbBlocks;
+
+            /* Compute the max block size */
+            maxBlockSize = MAX(maxBlockSize, zfh.blockSizeMax);
+        } else {
+            assert(zfh.frameType == ZSTD_skippableFrame);
+            /* Add the entire skippable frame size to our margin. */
+            margin += compressedSize;
+        }
+
+        assert(srcSize >= compressedSize);
+        src = (const BYTE*)src + compressedSize;
+        srcSize -= compressedSize;
+    }
+
+    /* Add the max block size back to the margin. */
+    margin += maxBlockSize;
+
+    return margin;
+}
 
 /*-*************************************************************
  *   Frame decoding
@@ -569,7 +878,7 @@ unsigned long long ZSTD_decompressBound(const void* src, size_t srcSize)
 size_t ZSTD_insertBlock(ZSTD_DCtx* dctx, const void* blockStart, size_t blockSize)
 {
     DEBUGLOG(5, "ZSTD_insertBlock: %u bytes", (unsigned)blockSize);
-    ZSTD_checkContinuity(dctx, blockStart);
+    ZSTD_checkContinuity(dctx, blockStart, blockSize);
     dctx->previousDstEnd = (const char*)blockStart + blockSize;
     return blockSize;
 }
@@ -579,12 +888,12 @@ static size_t ZSTD_copyRawBlock(void* dst, size_t dstCapacity,
                           const void* src, size_t srcSize)
 {
     DEBUGLOG(5, "ZSTD_copyRawBlock");
+    RETURN_ERROR_IF(srcSize > dstCapacity, dstSize_tooSmall, "");
     if (dst == NULL) {
         if (srcSize == 0) return 0;
         RETURN_ERROR(dstBuffer_null, "");
     }
-    RETURN_ERROR_IF(srcSize > dstCapacity, dstSize_tooSmall, "");
-    memcpy(dst, src, srcSize);
+    ZSTD_memmove(dst, src, srcSize);
     return srcSize;
 }
 
@@ -592,15 +901,41 @@ static size_t ZSTD_setRleBlock(void* dst, size_t dstCapacity,
                                BYTE b,
                                size_t regenSize)
 {
+    RETURN_ERROR_IF(regenSize > dstCapacity, dstSize_tooSmall, "");
     if (dst == NULL) {
         if (regenSize == 0) return 0;
         RETURN_ERROR(dstBuffer_null, "");
     }
-    RETURN_ERROR_IF(regenSize > dstCapacity, dstSize_tooSmall, "");
-    memset(dst, b, regenSize);
+    ZSTD_memset(dst, b, regenSize);
     return regenSize;
 }
 
+static void ZSTD_DCtx_trace_end(ZSTD_DCtx const* dctx, U64 uncompressedSize, U64 compressedSize, unsigned streaming)
+{
+#if ZSTD_TRACE
+    if (dctx->traceCtx && ZSTD_trace_decompress_end != NULL) {
+        ZSTD_Trace trace;
+        ZSTD_memset(&trace, 0, sizeof(trace));
+        trace.version = ZSTD_VERSION_NUMBER;
+        trace.streaming = streaming;
+        if (dctx->ddict) {
+            trace.dictionaryID = ZSTD_getDictID_fromDDict(dctx->ddict);
+            trace.dictionarySize = ZSTD_DDict_dictSize(dctx->ddict);
+            trace.dictionaryIsCold = dctx->ddictIsCold;
+        }
+        trace.uncompressedSize = (size_t)uncompressedSize;
+        trace.compressedSize = (size_t)compressedSize;
+        trace.dctx = dctx;
+        ZSTD_trace_decompress_end(dctx->traceCtx, &trace);
+    }
+#else
+    (void)dctx;
+    (void)uncompressedSize;
+    (void)compressedSize;
+    (void)streaming;
+#endif
+}
+
 
 /*! ZSTD_decompressFrame() :
  * @dctx must be properly initialized
@@ -610,8 +945,9 @@ static size_t ZSTD_decompressFrame(ZSTD_DCtx* dctx,
                                    void* dst, size_t dstCapacity,
                              const void** srcPtr, size_t *srcSizePtr)
 {
-    const BYTE* ip = (const BYTE*)(*srcPtr);
-    BYTE* const ostart = (BYTE* const)dst;
+    const BYTE* const istart = (const BYTE*)(*srcPtr);
+    const BYTE* ip = istart;
+    BYTE* const ostart = (BYTE*)dst;
     BYTE* const oend = dstCapacity != 0 ? ostart + dstCapacity : ostart;
     BYTE* op = ostart;
     size_t remainingSrcSize = *srcSizePtr;
@@ -635,6 +971,7 @@ static size_t ZSTD_decompressFrame(ZSTD_DCtx* dctx,
 
     /* Loop on each block */
     while (1) {
+        BYTE* oBlockEnd = oend;
         size_t decodedSize;
         blockProperties_t blockProperties;
         size_t const cBlockSize = ZSTD_getcBlockSize(ip, remainingSrcSize, &blockProperties);
@@ -644,16 +981,34 @@ static size_t ZSTD_decompressFrame(ZSTD_DCtx* dctx,
         remainingSrcSize -= ZSTD_blockHeaderSize;
         RETURN_ERROR_IF(cBlockSize > remainingSrcSize, srcSize_wrong, "");
 
+        if (ip >= op && ip < oBlockEnd) {
+            /* We are decompressing in-place. Limit the output pointer so that we
+             * don't overwrite the block that we are currently reading. This will
+             * fail decompression if the input & output pointers aren't spaced
+             * far enough apart.
+             *
+             * This is important to set, even when the pointers are far enough
+             * apart, because ZSTD_decompressBlock_internal() can decide to store
+             * literals in the output buffer, after the block it is decompressing.
+             * Since we don't want anything to overwrite our input, we have to tell
+             * ZSTD_decompressBlock_internal to never write past ip.
+             *
+             * See ZSTD_allocateLiteralsBuffer() for reference.
+             */
+            oBlockEnd = op + (ip - op);
+        }
+
         switch(blockProperties.blockType)
         {
         case bt_compressed:
-            decodedSize = ZSTD_decompressBlock_internal(dctx, op, oend-op, ip, cBlockSize, /* frame */ 1);
+            decodedSize = ZSTD_decompressBlock_internal(dctx, op, (size_t)(oBlockEnd-op), ip, cBlockSize, /* frame */ 1, not_streaming);
             break;
         case bt_raw :
-            decodedSize = ZSTD_copyRawBlock(op, oend-op, ip, cBlockSize);
+            /* Use oend instead of oBlockEnd because this function is safe to overlap. It uses memmove. */
+            decodedSize = ZSTD_copyRawBlock(op, (size_t)(oend-op), ip, cBlockSize);
             break;
         case bt_rle :
-            decodedSize = ZSTD_setRleBlock(op, oend-op, *ip, blockProperties.origSize);
+            decodedSize = ZSTD_setRleBlock(op, (size_t)(oBlockEnd-op), *ip, blockProperties.origSize);
             break;
         case bt_reserved :
         default:
@@ -661,7 +1016,7 @@ static size_t ZSTD_decompressFrame(ZSTD_DCtx* dctx,
         }
 
         if (ZSTD_isError(decodedSize)) return decodedSize;
-        if (dctx->fParams.checksumFlag)
+        if (dctx->validateChecksum)
             XXH64_update(&dctx->xxhState, op, decodedSize);
         if (decodedSize != 0)
             op += decodedSize;
@@ -676,19 +1031,22 @@ static size_t ZSTD_decompressFrame(ZSTD_DCtx* dctx,
                         corruption_detected, "");
     }
     if (dctx->fParams.checksumFlag) { /* Frame content checksum verification */
-        U32 const checkCalc = (U32)XXH64_digest(&dctx->xxhState);
-        U32 checkRead;
         RETURN_ERROR_IF(remainingSrcSize<4, checksum_wrong, "");
-        checkRead = MEM_readLE32(ip);
-        RETURN_ERROR_IF(checkRead != checkCalc, checksum_wrong, "");
+        if (!dctx->forceIgnoreChecksum) {
+            U32 const checkCalc = (U32)XXH64_digest(&dctx->xxhState);
+            U32 checkRead;
+            checkRead = MEM_readLE32(ip);
+            RETURN_ERROR_IF(checkRead != checkCalc, checksum_wrong, "");
+        }
         ip += 4;
         remainingSrcSize -= 4;
     }
-
+    ZSTD_DCtx_trace_end(dctx, (U64)(op-ostart), (U64)(ip-istart), /* streaming */ 0);
     /* Allow caller to get size read */
+    DEBUGLOG(4, "ZSTD_decompressFrame: decompressed frame of size %zi, consuming %zi bytes of input", op-ostart, ip - (const BYTE*)*srcPtr);
     *srcPtr = ip;
     *srcSizePtr = remainingSrcSize;
-    return op-ostart;
+    return (size_t)(op-ostart);
 }
 
 static size_t ZSTD_decompressMultiFrame(ZSTD_DCtx* dctx,
@@ -721,7 +1079,7 @@ static size_t ZSTD_decompressMultiFrame(ZSTD_DCtx* dctx,
             decodedSize = ZSTD_decompressLegacy(dst, dstCapacity, src, frameSize, dict, dictSize);
             if (ZSTD_isError(decodedSize)) return decodedSize;
 
-            assert(decodedSize <=- dstCapacity);
+            assert(decodedSize <= dstCapacity);
             dst = (BYTE*)dst + decodedSize;
             dstCapacity -= decodedSize;
 
@@ -753,7 +1111,7 @@ static size_t ZSTD_decompressMultiFrame(ZSTD_DCtx* dctx,
              * use this in all cases but ddict */
             FORWARD_IF_ERROR(ZSTD_decompressBegin_usingDict(dctx, dict, dictSize), "");
         }
-        ZSTD_checkContinuity(dctx, dst);
+        ZSTD_checkContinuity(dctx, dst, dstCapacity);
 
         {   const size_t res = ZSTD_decompressFrame(dctx, dst, dstCapacity,
                                                     &src, &srcSize);
@@ -761,15 +1119,13 @@ static size_t ZSTD_decompressMultiFrame(ZSTD_DCtx* dctx,
                 (ZSTD_getErrorCode(res) == ZSTD_error_prefix_unknown)
              && (moreThan1Frame==1),
                 srcSize_wrong,
-                "at least one frame successfully completed, but following "
-                "bytes are garbage: it's more likely to be a srcSize error, "
-                "specifying more bytes than compressed size of frame(s). This "
-                "error message replaces ERROR(prefix_unknown), which would be "
-                "confusing, as the first header is actually correct. Note that "
-                "one could be unlucky, it might be a corruption error instead, "
-                "happening right at the place where we expect zstd magic "
-                "bytes. But this is _much_ less likely than a srcSize field "
-                "error.");
+                "At least one frame successfully completed, "
+                "but following bytes are garbage: "
+                "it's more likely to be a srcSize error, "
+                "specifying more input bytes than size of frame(s). "
+                "Note: one could be unlucky, it might be a corruption error instead, "
+                "happening right at the place where we expect zstd magic bytes. "
+                "But this is _much_ less likely than a srcSize field error.");
             if (ZSTD_isError(res)) return res;
             assert(res <= dstCapacity);
             if (res != 0)
@@ -781,7 +1137,7 @@ static size_t ZSTD_decompressMultiFrame(ZSTD_DCtx* dctx,
 
     RETURN_ERROR_IF(srcSize, srcSize_wrong, "input not entirely consumed");
 
-    return (BYTE*)dst - (BYTE*)dststart;
+    return (size_t)((BYTE*)dst - (BYTE*)dststart);
 }
 
 size_t ZSTD_decompress_usingDict(ZSTD_DCtx* dctx,
@@ -798,7 +1154,7 @@ static ZSTD_DDict const* ZSTD_getDDict(ZSTD_DCtx* dctx)
     switch (dctx->dictUses) {
     default:
         assert(0 /* Impossible */);
-        /* fall-through */
+        ZSTD_FALLTHROUGH;
     case ZSTD_dont_use:
         ZSTD_clearDict(dctx);
         return NULL;
@@ -820,7 +1176,7 @@ size_t ZSTD_decompress(void* dst, size_t dstCapacity, const void* src, size_t sr
 {
 #if defined(ZSTD_HEAPMODE) && (ZSTD_HEAPMODE>=1)
     size_t regenSize;
-    ZSTD_DCtx* const dctx = ZSTD_createDCtx();
+    ZSTD_DCtx* const dctx =  ZSTD_createDCtx_internal(ZSTD_defaultCMem);
     RETURN_ERROR_IF(dctx==NULL, memory_allocation, "NULL pointer!");
     regenSize = ZSTD_decompressDCtx(dctx, dst, dstCapacity, src, srcSize);
     ZSTD_freeDCtx(dctx);
@@ -840,8 +1196,8 @@ size_t ZSTD_decompress(void* dst, size_t dstCapacity, const void* src, size_t sr
 size_t ZSTD_nextSrcSizeToDecompress(ZSTD_DCtx* dctx) { return dctx->expected; }
 
 /**
- * Similar to ZSTD_nextSrcSizeToDecompress(), but when when a block input can be streamed,
- * we allow taking a partial block as the input. Currently only raw uncompressed blocks can
+ * Similar to ZSTD_nextSrcSizeToDecompress(), but when a block input can be streamed, we
+ * allow taking a partial block as the input. Currently only raw uncompressed blocks can
  * be streamed.
  *
  * For blocks that can be streamed, this allows us to reduce the latency until we produce
@@ -854,7 +1210,7 @@ static size_t ZSTD_nextSrcSizeToDecompressWithInputSize(ZSTD_DCtx* dctx, size_t
         return dctx->expected;
     if (dctx->bType != bt_raw)
         return dctx->expected;
-    return MIN(MAX(inputSize, 1), dctx->expected);
+    return BOUNDED(1, inputSize, dctx->expected);
 }
 
 ZSTD_nextInputType_e ZSTD_nextInputType(ZSTD_DCtx* dctx) {
@@ -862,7 +1218,9 @@ ZSTD_nextInputType_e ZSTD_nextInputType(ZSTD_DCtx* dctx) {
     {
     default:   /* should not happen */
         assert(0);
+        ZSTD_FALLTHROUGH;
     case ZSTDds_getFrameHeaderSize:
+        ZSTD_FALLTHROUGH;
     case ZSTDds_decodeFrameHeader:
         return ZSTDnit_frameHeader;
     case ZSTDds_decodeBlockHeader:
@@ -874,6 +1232,7 @@ ZSTD_nextInputType_e ZSTD_nextInputType(ZSTD_DCtx* dctx) {
     case ZSTDds_checkChecksum:
         return ZSTDnit_checksum;
     case ZSTDds_decodeSkippableHeader:
+        ZSTD_FALLTHROUGH;
     case ZSTDds_skipFrame:
         return ZSTDnit_skippableFrame;
     }
@@ -890,7 +1249,9 @@ size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, c
     DEBUGLOG(5, "ZSTD_decompressContinue (srcSize:%u)", (unsigned)srcSize);
     /* Sanity check */
     RETURN_ERROR_IF(srcSize != ZSTD_nextSrcSizeToDecompressWithInputSize(dctx, srcSize), srcSize_wrong, "not allowed");
-    if (dstCapacity) ZSTD_checkContinuity(dctx, dst);
+    ZSTD_checkContinuity(dctx, dst, dstCapacity);
+
+    dctx->processedCSize += srcSize;
 
     switch (dctx->stage)
     {
@@ -899,21 +1260,21 @@ size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, c
         if (dctx->format == ZSTD_f_zstd1) {  /* allows header */
             assert(srcSize >= ZSTD_FRAMEIDSIZE);  /* to read skippable magic number */
             if ((MEM_readLE32(src) & ZSTD_MAGIC_SKIPPABLE_MASK) == ZSTD_MAGIC_SKIPPABLE_START) {        /* skippable frame */
-                memcpy(dctx->headerBuffer, src, srcSize);
+                ZSTD_memcpy(dctx->headerBuffer, src, srcSize);
                 dctx->expected = ZSTD_SKIPPABLEHEADERSIZE - srcSize;  /* remaining to load to get full skippable frame header */
                 dctx->stage = ZSTDds_decodeSkippableHeader;
                 return 0;
         }   }
         dctx->headerSize = ZSTD_frameHeaderSize_internal(src, srcSize, dctx->format);
         if (ZSTD_isError(dctx->headerSize)) return dctx->headerSize;
-        memcpy(dctx->headerBuffer, src, srcSize);
+        ZSTD_memcpy(dctx->headerBuffer, src, srcSize);
         dctx->expected = dctx->headerSize - srcSize;
         dctx->stage = ZSTDds_decodeFrameHeader;
         return 0;
 
     case ZSTDds_decodeFrameHeader:
         assert(src != NULL);
-        memcpy(dctx->headerBuffer + (dctx->headerSize - srcSize), src, srcSize);
+        ZSTD_memcpy(dctx->headerBuffer + (dctx->headerSize - srcSize), src, srcSize);
         FORWARD_IF_ERROR(ZSTD_decodeFrameHeader(dctx, dctx->headerBuffer, dctx->headerSize), "");
         dctx->expected = ZSTD_blockHeaderSize;
         dctx->stage = ZSTDds_decodeBlockHeader;
@@ -955,7 +1316,7 @@ size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, c
             {
             case bt_compressed:
                 DEBUGLOG(5, "ZSTD_decompressContinue: case bt_compressed");
-                rSize = ZSTD_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize, /* frame */ 1);
+                rSize = ZSTD_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize, /* frame */ 1, is_streaming);
                 dctx->expected = 0;  /* Streaming not supported */
                 break;
             case bt_raw :
@@ -977,7 +1338,7 @@ size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, c
             RETURN_ERROR_IF(rSize > dctx->fParams.blockSizeMax, corruption_detected, "Decompressed Block Size Exceeds Maximum");
             DEBUGLOG(5, "ZSTD_decompressContinue: decoded size from block : %u", (unsigned)rSize);
             dctx->decodedSize += rSize;
-            if (dctx->fParams.checksumFlag) XXH64_update(&dctx->xxhState, dst, rSize);
+            if (dctx->validateChecksum) XXH64_update(&dctx->xxhState, dst, rSize);
             dctx->previousDstEnd = (char*)dst + rSize;
 
             /* Stay on the same stage until we are finished streaming the block. */
@@ -995,6 +1356,7 @@ size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, c
                     dctx->expected = 4;
                     dctx->stage = ZSTDds_checkChecksum;
                 } else {
+                    ZSTD_DCtx_trace_end(dctx, dctx->decodedSize, dctx->processedCSize, /* streaming */ 1);
                     dctx->expected = 0;   /* ends here */
                     dctx->stage = ZSTDds_getFrameHeaderSize;
                 }
@@ -1007,10 +1369,14 @@ size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, c
 
     case ZSTDds_checkChecksum:
         assert(srcSize == 4);  /* guaranteed by dctx->expected */
-        {   U32 const h32 = (U32)XXH64_digest(&dctx->xxhState);
-            U32 const check32 = MEM_readLE32(src);
-            DEBUGLOG(4, "ZSTD_decompressContinue: checksum : calculated %08X :: %08X read", (unsigned)h32, (unsigned)check32);
-            RETURN_ERROR_IF(check32 != h32, checksum_wrong, "");
+        {
+            if (dctx->validateChecksum) {
+                U32 const h32 = (U32)XXH64_digest(&dctx->xxhState);
+                U32 const check32 = MEM_readLE32(src);
+                DEBUGLOG(4, "ZSTD_decompressContinue: checksum : calculated %08X :: %08X read", (unsigned)h32, (unsigned)check32);
+                RETURN_ERROR_IF(check32 != h32, checksum_wrong, "");
+            }
+            ZSTD_DCtx_trace_end(dctx, dctx->decodedSize, dctx->processedCSize, /* streaming */ 1);
             dctx->expected = 0;
             dctx->stage = ZSTDds_getFrameHeaderSize;
             return 0;
@@ -1019,7 +1385,7 @@ size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, c
     case ZSTDds_decodeSkippableHeader:
         assert(src != NULL);
         assert(srcSize <= ZSTD_SKIPPABLEHEADERSIZE);
-        memcpy(dctx->headerBuffer + (ZSTD_SKIPPABLEHEADERSIZE - srcSize), src, srcSize);   /* complete skippable header */
+        ZSTD_memcpy(dctx->headerBuffer + (ZSTD_SKIPPABLEHEADERSIZE - srcSize), src, srcSize);   /* complete skippable header */
         dctx->expected = MEM_readLE32(dctx->headerBuffer + ZSTD_FRAMEIDSIZE);   /* note : dctx->expected can grow seriously large, beyond local buffer size */
         dctx->stage = ZSTDds_skipFrame;
         return 0;
@@ -1031,7 +1397,7 @@ size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, c
 
     default:
         assert(0);   /* impossible */
-        RETURN_ERROR(GENERIC, "impossible to reach");   /* some compiler require default to do something */
+        RETURN_ERROR(GENERIC, "impossible to reach");   /* some compilers require default to do something */
     }
 }
 
@@ -1072,11 +1438,11 @@ ZSTD_loadDEntropy(ZSTD_entropyDTables_t* entropy,
         /* in minimal huffman, we always use X1 variants */
         size_t const hSize = HUF_readDTableX1_wksp(entropy->hufTable,
                                                 dictPtr, dictEnd - dictPtr,
-                                                workspace, workspaceSize);
+                                                workspace, workspaceSize, /* flags */ 0);
 #else
         size_t const hSize = HUF_readDTableX2_wksp(entropy->hufTable,
-                                                dictPtr, dictEnd - dictPtr,
-                                                workspace, workspaceSize);
+                                                dictPtr, (size_t)(dictEnd - dictPtr),
+                                                workspace, workspaceSize, /* flags */ 0);
 #endif
         RETURN_ERROR_IF(HUF_isError(hSize), dictionary_corrupted, "");
         dictPtr += hSize;
@@ -1084,40 +1450,46 @@ ZSTD_loadDEntropy(ZSTD_entropyDTables_t* entropy,
 
     {   short offcodeNCount[MaxOff+1];
         unsigned offcodeMaxValue = MaxOff, offcodeLog;
-        size_t const offcodeHeaderSize = FSE_readNCount(offcodeNCount, &offcodeMaxValue, &offcodeLog, dictPtr, dictEnd-dictPtr);
+        size_t const offcodeHeaderSize = FSE_readNCount(offcodeNCount, &offcodeMaxValue, &offcodeLog, dictPtr, (size_t)(dictEnd-dictPtr));
         RETURN_ERROR_IF(FSE_isError(offcodeHeaderSize), dictionary_corrupted, "");
         RETURN_ERROR_IF(offcodeMaxValue > MaxOff, dictionary_corrupted, "");
         RETURN_ERROR_IF(offcodeLog > OffFSELog, dictionary_corrupted, "");
         ZSTD_buildFSETable( entropy->OFTable,
                             offcodeNCount, offcodeMaxValue,
                             OF_base, OF_bits,
-                            offcodeLog);
+                            offcodeLog,
+                            entropy->workspace, sizeof(entropy->workspace),
+                            /* bmi2 */0);
         dictPtr += offcodeHeaderSize;
     }
 
     {   short matchlengthNCount[MaxML+1];
         unsigned matchlengthMaxValue = MaxML, matchlengthLog;
-        size_t const matchlengthHeaderSize = FSE_readNCount(matchlengthNCount, &matchlengthMaxValue, &matchlengthLog, dictPtr, dictEnd-dictPtr);
+        size_t const matchlengthHeaderSize = FSE_readNCount(matchlengthNCount, &matchlengthMaxValue, &matchlengthLog, dictPtr, (size_t)(dictEnd-dictPtr));
         RETURN_ERROR_IF(FSE_isError(matchlengthHeaderSize), dictionary_corrupted, "");
         RETURN_ERROR_IF(matchlengthMaxValue > MaxML, dictionary_corrupted, "");
         RETURN_ERROR_IF(matchlengthLog > MLFSELog, dictionary_corrupted, "");
         ZSTD_buildFSETable( entropy->MLTable,
                             matchlengthNCount, matchlengthMaxValue,
                             ML_base, ML_bits,
-                            matchlengthLog);
+                            matchlengthLog,
+                            entropy->workspace, sizeof(entropy->workspace),
+                            /* bmi2 */ 0);
         dictPtr += matchlengthHeaderSize;
     }
 
     {   short litlengthNCount[MaxLL+1];
         unsigned litlengthMaxValue = MaxLL, litlengthLog;
-        size_t const litlengthHeaderSize = FSE_readNCount(litlengthNCount, &litlengthMaxValue, &litlengthLog, dictPtr, dictEnd-dictPtr);
+        size_t const litlengthHeaderSize = FSE_readNCount(litlengthNCount, &litlengthMaxValue, &litlengthLog, dictPtr, (size_t)(dictEnd-dictPtr));
         RETURN_ERROR_IF(FSE_isError(litlengthHeaderSize), dictionary_corrupted, "");
         RETURN_ERROR_IF(litlengthMaxValue > MaxLL, dictionary_corrupted, "");
         RETURN_ERROR_IF(litlengthLog > LLFSELog, dictionary_corrupted, "");
         ZSTD_buildFSETable( entropy->LLTable,
                             litlengthNCount, litlengthMaxValue,
                             LL_base, LL_bits,
-                            litlengthLog);
+                            litlengthLog,
+                            entropy->workspace, sizeof(entropy->workspace),
+                            /* bmi2 */ 0);
         dictPtr += litlengthHeaderSize;
     }
 
@@ -1131,7 +1503,7 @@ ZSTD_loadDEntropy(ZSTD_entropyDTables_t* entropy,
             entropy->rep[i] = rep;
     }   }
 
-    return dictPtr - (const BYTE*)dict;
+    return (size_t)(dictPtr - (const BYTE*)dict);
 }
 
 static size_t ZSTD_decompress_insertDictionary(ZSTD_DCtx* dctx, const void* dict, size_t dictSize)
@@ -1158,19 +1530,23 @@ static size_t ZSTD_decompress_insertDictionary(ZSTD_DCtx* dctx, const void* dict
 size_t ZSTD_decompressBegin(ZSTD_DCtx* dctx)
 {
     assert(dctx != NULL);
+#if ZSTD_TRACE
+    dctx->traceCtx = (ZSTD_trace_decompress_begin != NULL) ? ZSTD_trace_decompress_begin(dctx) : 0;
+#endif
     dctx->expected = ZSTD_startingInputLength(dctx->format);  /* dctx->format must be properly set */
     dctx->stage = ZSTDds_getFrameHeaderSize;
+    dctx->processedCSize = 0;
     dctx->decodedSize = 0;
     dctx->previousDstEnd = NULL;
     dctx->prefixStart = NULL;
     dctx->virtualStart = NULL;
     dctx->dictEnd = NULL;
-    dctx->entropy.hufTable[0] = (HUF_DTable)((HufLog)*0x1000001);  /* cover both little and big endian */
+    dctx->entropy.hufTable[0] = (HUF_DTable)((ZSTD_HUFFDTABLE_CAPACITY_LOG)*0x1000001);  /* cover both little and big endian */
     dctx->litEntropy = dctx->fseEntropy = 0;
     dctx->dictID = 0;
     dctx->bType = bt_reserved;
     ZSTD_STATIC_ASSERT(sizeof(dctx->entropy.rep) == sizeof(repStartValue));
-    memcpy(dctx->entropy.rep, repStartValue, sizeof(repStartValue));  /* initial repcodes */
+    ZSTD_memcpy(dctx->entropy.rep, repStartValue, sizeof(repStartValue));  /* initial repcodes */
     dctx->LLTptr = dctx->entropy.LLTable;
     dctx->MLTptr = dctx->entropy.MLTable;
     dctx->OFTptr = dctx->entropy.OFTable;
@@ -1227,7 +1603,7 @@ unsigned ZSTD_getDictID_fromDict(const void* dict, size_t dictSize)
  *  This could for one of the following reasons :
  *  - The frame does not require a dictionary (most common case).
  *  - The frame was built with dictID intentionally removed.
- *    Needed dictionary is a hidden information.
+ *    Needed dictionary is a hidden piece of information.
  *    Note : this use case also happens when using a non-conformant dictionary.
  *  - `srcSize` is too small, and as a result, frame header could not be decoded.
  *    Note : possible if `srcSize < ZSTD_FRAMEHEADERSIZE_MAX`.
@@ -1236,7 +1612,7 @@ unsigned ZSTD_getDictID_fromDict(const void* dict, size_t dictSize)
  *  ZSTD_getFrameHeader(), which will provide a more precise error code. */
 unsigned ZSTD_getDictID_fromFrame(const void* src, size_t srcSize)
 {
-    ZSTD_frameHeader zfp = { 0, 0, 0, ZSTD_frame, 0, 0, 0 };
+    ZSTD_frameHeader zfp = { 0, 0, 0, ZSTD_frame, 0, 0, 0, 0, 0 };
     size_t const hError = ZSTD_getFrameHeader(&zfp, src, srcSize);
     if (ZSTD_isError(hError)) return 0;
     return zfp.dictID;
@@ -1265,7 +1641,7 @@ size_t ZSTD_decompress_usingDDict(ZSTD_DCtx* dctx,
 ZSTD_DStream* ZSTD_createDStream(void)
 {
     DEBUGLOG(3, "ZSTD_createDStream");
-    return ZSTD_createDStream_advanced(ZSTD_defaultCMem);
+    return ZSTD_createDCtx_internal(ZSTD_defaultCMem);
 }
 
 ZSTD_DStream* ZSTD_initStaticDStream(void *workspace, size_t workspaceSize)
@@ -1275,7 +1651,7 @@ ZSTD_DStream* ZSTD_initStaticDStream(void *workspace, size_t workspaceSize)
 
 ZSTD_DStream* ZSTD_createDStream_advanced(ZSTD_customMem customMem)
 {
-    return ZSTD_createDCtx_advanced(customMem);
+    return ZSTD_createDCtx_internal(customMem);
 }
 
 size_t ZSTD_freeDStream(ZSTD_DStream* zds)
@@ -1343,7 +1719,9 @@ size_t ZSTD_initDStream_usingDict(ZSTD_DStream* zds, const void* dict, size_t di
 size_t ZSTD_initDStream(ZSTD_DStream* zds)
 {
     DEBUGLOG(4, "ZSTD_initDStream");
-    return ZSTD_initDStream_usingDDict(zds, NULL);
+    FORWARD_IF_ERROR(ZSTD_DCtx_reset(zds, ZSTD_reset_session_only), "");
+    FORWARD_IF_ERROR(ZSTD_DCtx_refDDict(zds, NULL), "");
+    return ZSTD_startingInputLength(zds->format);
 }
 
 /* ZSTD_initDStream_usingDDict() :
@@ -1351,6 +1729,7 @@ size_t ZSTD_initDStream(ZSTD_DStream* zds)
  * this function cannot fail */
 size_t ZSTD_initDStream_usingDDict(ZSTD_DStream* dctx, const ZSTD_DDict* ddict)
 {
+    DEBUGLOG(4, "ZSTD_initDStream_usingDDict");
     FORWARD_IF_ERROR( ZSTD_DCtx_reset(dctx, ZSTD_reset_session_only) , "");
     FORWARD_IF_ERROR( ZSTD_DCtx_refDDict(dctx, ddict) , "");
     return ZSTD_startingInputLength(dctx->format);
@@ -1361,6 +1740,7 @@ size_t ZSTD_initDStream_usingDDict(ZSTD_DStream* dctx, const ZSTD_DDict* ddict)
  * this function cannot fail */
 size_t ZSTD_resetDStream(ZSTD_DStream* dctx)
 {
+    DEBUGLOG(4, "ZSTD_resetDStream");
     FORWARD_IF_ERROR(ZSTD_DCtx_reset(dctx, ZSTD_reset_session_only), "");
     return ZSTD_startingInputLength(dctx->format);
 }
@@ -1373,6 +1753,16 @@ size_t ZSTD_DCtx_refDDict(ZSTD_DCtx* dctx, const ZSTD_DDict* ddict)
     if (ddict) {
         dctx->ddict = ddict;
         dctx->dictUses = ZSTD_use_indefinitely;
+        if (dctx->refMultipleDDicts == ZSTD_rmd_refMultipleDDicts) {
+            if (dctx->ddictSet == NULL) {
+                dctx->ddictSet = ZSTD_createDDictHashSet(dctx->customMem);
+                if (!dctx->ddictSet) {
+                    RETURN_ERROR(memory_allocation, "Failed to allocate memory for hash set!");
+                }
+            }
+            assert(!dctx->staticSize);  /* Impossible: ddictSet cannot have been allocated if static dctx */
+            FORWARD_IF_ERROR(ZSTD_DDictHashSet_addDDict(dctx->ddictSet, ddict, dctx->customMem), "");
+        }
     }
     return 0;
 }
@@ -1394,7 +1784,7 @@ size_t ZSTD_DCtx_setMaxWindowSize(ZSTD_DCtx* dctx, size_t maxWindowSize)
 
 size_t ZSTD_DCtx_setFormat(ZSTD_DCtx* dctx, ZSTD_format_e format)
 {
-    return ZSTD_DCtx_setParameter(dctx, ZSTD_d_format, format);
+    return ZSTD_DCtx_setParameter(dctx, ZSTD_d_format, (int)format);
 }
 
 ZSTD_bounds ZSTD_dParam_getBounds(ZSTD_dParameter dParam)
@@ -1411,9 +1801,22 @@ ZSTD_bounds ZSTD_dParam_getBounds(ZSTD_dParameter dParam)
             ZSTD_STATIC_ASSERT(ZSTD_f_zstd1 < ZSTD_f_zstd1_magicless);
             return bounds;
         case ZSTD_d_stableOutBuffer:
-            bounds.lowerBound = (int)ZSTD_obm_buffered;
-            bounds.upperBound = (int)ZSTD_obm_stable;
+            bounds.lowerBound = (int)ZSTD_bm_buffered;
+            bounds.upperBound = (int)ZSTD_bm_stable;
+            return bounds;
+        case ZSTD_d_forceIgnoreChecksum:
+            bounds.lowerBound = (int)ZSTD_d_validateChecksum;
+            bounds.upperBound = (int)ZSTD_d_ignoreChecksum;
             return bounds;
+        case ZSTD_d_refMultipleDDicts:
+            bounds.lowerBound = (int)ZSTD_rmd_refSingleDDict;
+            bounds.upperBound = (int)ZSTD_rmd_refMultipleDDicts;
+            return bounds;
+        case ZSTD_d_disableHuffmanAssembly:
+            bounds.lowerBound = 0;
+            bounds.upperBound = 1;
+            return bounds;
+
         default:;
     }
     bounds.error = ERROR(parameter_unsupported);
@@ -1436,6 +1839,32 @@ static int ZSTD_dParam_withinBounds(ZSTD_dParameter dParam, int value)
     RETURN_ERROR_IF(!ZSTD_dParam_withinBounds(p, v), parameter_outOfBound, ""); \
 }
 
+size_t ZSTD_DCtx_getParameter(ZSTD_DCtx* dctx, ZSTD_dParameter param, int* value)
+{
+    switch (param) {
+        case ZSTD_d_windowLogMax:
+            *value = (int)ZSTD_highbit32((U32)dctx->maxWindowSize);
+            return 0;
+        case ZSTD_d_format:
+            *value = (int)dctx->format;
+            return 0;
+        case ZSTD_d_stableOutBuffer:
+            *value = (int)dctx->outBufferMode;
+            return 0;
+        case ZSTD_d_forceIgnoreChecksum:
+            *value = (int)dctx->forceIgnoreChecksum;
+            return 0;
+        case ZSTD_d_refMultipleDDicts:
+            *value = (int)dctx->refMultipleDDicts;
+            return 0;
+        case ZSTD_d_disableHuffmanAssembly:
+            *value = (int)dctx->disableHufAsm;
+            return 0;
+        default:;
+    }
+    RETURN_ERROR(parameter_unsupported, "");
+}
+
 size_t ZSTD_DCtx_setParameter(ZSTD_DCtx* dctx, ZSTD_dParameter dParam, int value)
 {
     RETURN_ERROR_IF(dctx->streamStage != zdss_init, stage_wrong, "");
@@ -1451,7 +1880,22 @@ size_t ZSTD_DCtx_setParameter(ZSTD_DCtx* dctx, ZSTD_dParameter dParam, int value
             return 0;
         case ZSTD_d_stableOutBuffer:
             CHECK_DBOUNDS(ZSTD_d_stableOutBuffer, value);
-            dctx->outBufferMode = (ZSTD_outBufferMode_e)value;
+            dctx->outBufferMode = (ZSTD_bufferMode_e)value;
+            return 0;
+        case ZSTD_d_forceIgnoreChecksum:
+            CHECK_DBOUNDS(ZSTD_d_forceIgnoreChecksum, value);
+            dctx->forceIgnoreChecksum = (ZSTD_forceIgnoreChecksum_e)value;
+            return 0;
+        case ZSTD_d_refMultipleDDicts:
+            CHECK_DBOUNDS(ZSTD_d_refMultipleDDicts, value);
+            if (dctx->staticSize != 0) {
+                RETURN_ERROR(parameter_unsupported, "Static dctx does not support multiple DDicts!");
+            }
+            dctx->refMultipleDDicts = (ZSTD_refMultipleDDicts_e)value;
+            return 0;
+        case ZSTD_d_disableHuffmanAssembly:
+            CHECK_DBOUNDS(ZSTD_d_disableHuffmanAssembly, value);
+            dctx->disableHufAsm = value != 0;
             return 0;
         default:;
     }
@@ -1469,8 +1913,7 @@ size_t ZSTD_DCtx_reset(ZSTD_DCtx* dctx, ZSTD_ResetDirective reset)
       || (reset == ZSTD_reset_session_and_parameters) ) {
         RETURN_ERROR_IF(dctx->streamStage != zdss_init, stage_wrong, "");
         ZSTD_clearDict(dctx);
-        dctx->format = ZSTD_f_zstd1;
-        dctx->maxWindowSize = ZSTD_MAXWINDOWSIZE_DEFAULT;
+        ZSTD_DCtx_resetParameters(dctx);
     }
     return 0;
 }
@@ -1484,7 +1927,8 @@ size_t ZSTD_sizeof_DStream(const ZSTD_DStream* dctx)
 size_t ZSTD_decodingBufferSize_min(unsigned long long windowSize, unsigned long long frameContentSize)
 {
     size_t const blockSize = (size_t) MIN(windowSize, ZSTD_BLOCKSIZE_MAX);
-    unsigned long long const neededRBSize = windowSize + blockSize + (WILDCOPY_OVERLENGTH * 2);
+    /* space is needed to store the litbuffer after the output of a given block without stomping the extDict of a previous run, as well as to cover both windows against wildcopy*/
+    unsigned long long const neededRBSize = windowSize + blockSize + ZSTD_BLOCKSIZE_MAX + (WILDCOPY_OVERLENGTH * 2);
     unsigned long long const neededSize = MIN(frameContentSize, neededRBSize);
     size_t const minRBSize = (size_t) neededSize;
     RETURN_ERROR_IF((unsigned long long)minRBSize != neededSize,
@@ -1538,7 +1982,7 @@ static size_t ZSTD_checkOutBuffer(ZSTD_DStream const* zds, ZSTD_outBuffer const*
 {
     ZSTD_outBuffer const expect = zds->expectedOutBuffer;
     /* No requirement when ZSTD_obm_stable is not enabled. */
-    if (zds->outBufferMode != ZSTD_obm_stable)
+    if (zds->outBufferMode != ZSTD_bm_stable)
         return 0;
     /* Any buffer is allowed in zdss_init, this must be the same for every other call until
      * the context is reset.
@@ -1548,7 +1992,7 @@ static size_t ZSTD_checkOutBuffer(ZSTD_DStream const* zds, ZSTD_outBuffer const*
     /* The buffer must match our expectation exactly. */
     if (expect.dst == output->dst && expect.pos == output->pos && expect.size == output->size)
         return 0;
-    RETURN_ERROR(dstBuffer_wrong, "ZSTD_obm_stable enabled but output differs!");
+    RETURN_ERROR(dstBuffer_wrong, "ZSTD_d_stableOutBuffer enabled but output differs!");
 }
 
 /* Calls ZSTD_decompressContinue() with the right parameters for ZSTD_decompressStream()
@@ -1560,7 +2004,7 @@ static size_t ZSTD_decompressContinueStream(
             ZSTD_DStream* zds, char** op, char* oend,
             void const* src, size_t srcSize) {
     int const isSkipFrame = ZSTD_isSkipFrame(zds);
-    if (zds->outBufferMode == ZSTD_obm_buffered) {
+    if (zds->outBufferMode == ZSTD_bm_buffered) {
         size_t const dstSize = isSkipFrame ? 0 : zds->outBuffSize - zds->outStart;
         size_t const decodedSize = ZSTD_decompressContinue(zds,
                 zds->outBuff + zds->outStart, dstSize, src, srcSize);
@@ -1573,14 +2017,14 @@ static size_t ZSTD_decompressContinueStream(
         }
     } else {
         /* Write directly into the output buffer */
-        size_t const dstSize = isSkipFrame ? 0 : oend - *op;
+        size_t const dstSize = isSkipFrame ? 0 : (size_t)(oend - *op);
         size_t const decodedSize = ZSTD_decompressContinue(zds, *op, dstSize, src, srcSize);
         FORWARD_IF_ERROR(decodedSize, "");
         *op += decodedSize;
         /* Flushing is not needed. */
         zds->streamStage = zdss_read;
         assert(*op <= oend);
-        assert(zds->outBufferMode == ZSTD_obm_stable);
+        assert(zds->outBufferMode == ZSTD_bm_stable);
     }
     return 0;
 }
@@ -1618,10 +2062,12 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
             DEBUGLOG(5, "stage zdss_init => transparent reset ");
             zds->streamStage = zdss_loadHeader;
             zds->lhSize = zds->inPos = zds->outStart = zds->outEnd = 0;
+#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
             zds->legacyVersion = 0;
+#endif
             zds->hostageByte = 0;
             zds->expectedOutBuffer = *output;
-            /* fall-through */
+            ZSTD_FALLTHROUGH;
 
         case zdss_loadHeader :
             DEBUGLOG(5, "stage zdss_loadHeader (srcSize : %u)", (U32)(iend - ip));
@@ -1635,7 +2081,9 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
             }   }
 #endif
             {   size_t const hSize = ZSTD_getFrameHeader_advanced(&zds->fParams, zds->headerBuffer, zds->lhSize, zds->format);
-                DEBUGLOG(5, "header size : %u", (U32)hSize);
+                if (zds->refMultipleDDicts && zds->ddictSet) {
+                    ZSTD_DCtx_selectFrameDDict(zds);
+                }
                 if (ZSTD_isError(hSize)) {
 #if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
                     U32 const legacyVersion = ZSTD_isLegacy(istart, iend-istart);
@@ -1663,14 +2111,19 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
                     assert(iend >= ip);
                     if (toLoad > remainingInput) {   /* not enough input to load full header */
                         if (remainingInput > 0) {
-                            memcpy(zds->headerBuffer + zds->lhSize, ip, remainingInput);
+                            ZSTD_memcpy(zds->headerBuffer + zds->lhSize, ip, remainingInput);
                             zds->lhSize += remainingInput;
                         }
                         input->pos = input->size;
+                        /* check first few bytes */
+                        FORWARD_IF_ERROR(
+                            ZSTD_getFrameHeader_advanced(&zds->fParams, zds->headerBuffer, zds->lhSize, zds->format),
+                            "First few bytes detected incorrect" );
+                        /* return hint input size */
                         return (MAX((size_t)ZSTD_FRAMEHEADERSIZE_MIN(zds->format), hSize) - zds->lhSize) + ZSTD_blockHeaderSize;   /* remaining header bytes + next block header */
                     }
                     assert(ip != NULL);
-                    memcpy(zds->headerBuffer + zds->lhSize, ip, toLoad); zds->lhSize = hSize; ip += toLoad;
+                    ZSTD_memcpy(zds->headerBuffer + zds->lhSize, ip, toLoad); zds->lhSize = hSize; ip += toLoad;
                     break;
             }   }
 
@@ -1678,14 +2131,15 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
             if (zds->fParams.frameContentSize != ZSTD_CONTENTSIZE_UNKNOWN
                 && zds->fParams.frameType != ZSTD_skippableFrame
                 && (U64)(size_t)(oend-op) >= zds->fParams.frameContentSize) {
-                size_t const cSize = ZSTD_findFrameCompressedSize(istart, iend-istart);
+                size_t const cSize = ZSTD_findFrameCompressedSize(istart, (size_t)(iend-istart));
                 if (cSize <= (size_t)(iend-istart)) {
                     /* shortcut : using single-pass mode */
-                    size_t const decompressedSize = ZSTD_decompress_usingDDict(zds, op, oend-op, istart, cSize, ZSTD_getDDict(zds));
+                    size_t const decompressedSize = ZSTD_decompress_usingDDict(zds, op, (size_t)(oend-op), istart, cSize, ZSTD_getDDict(zds));
                     if (ZSTD_isError(decompressedSize)) return decompressedSize;
                     DEBUGLOG(4, "shortcut to single-pass ZSTD_decompress_usingDDict()")
+                    assert(istart != NULL);
                     ip = istart + cSize;
-                    op += decompressedSize;
+                    op = op ? op + decompressedSize : op; /* can occur if frameContentSize = 0 (empty frame) */
                     zds->expected = 0;
                     zds->streamStage = zdss_init;
                     someMoreWork = 0;
@@ -1693,7 +2147,7 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
             }   }
 
             /* Check output buffer is large enough for ZSTD_odm_stable. */
-            if (zds->outBufferMode == ZSTD_obm_stable
+            if (zds->outBufferMode == ZSTD_bm_stable
                 && zds->fParams.frameType != ZSTD_skippableFrame
                 && zds->fParams.frameContentSize != ZSTD_CONTENTSIZE_UNKNOWN
                 && (U64)(size_t)(oend-op) < zds->fParams.frameContentSize) {
@@ -1723,7 +2177,7 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
 
             /* Adapt buffer sizes to frame header instructions */
             {   size_t const neededInBuffSize = MAX(zds->fParams.blockSizeMax, 4 /* frame checksum */);
-                size_t const neededOutBuffSize = zds->outBufferMode == ZSTD_obm_buffered
+                size_t const neededOutBuffSize = zds->outBufferMode == ZSTD_bm_buffered
                         ? ZSTD_decodingBufferSize_min(zds->fParams.windowSize, zds->fParams.frameContentSize)
                         : 0;
 
@@ -1745,10 +2199,10 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
                                 bufferSize > zds->staticSize - sizeof(ZSTD_DCtx),
                                 memory_allocation, "");
                         } else {
-                            ZSTD_free(zds->inBuff, zds->customMem);
+                            ZSTD_customFree(zds->inBuff, zds->customMem);
                             zds->inBuffSize = 0;
                             zds->outBuffSize = 0;
-                            zds->inBuff = (char*)ZSTD_malloc(bufferSize, zds->customMem);
+                            zds->inBuff = (char*)ZSTD_customMalloc(bufferSize, zds->customMem);
                             RETURN_ERROR_IF(zds->inBuff == NULL, memory_allocation, "");
                         }
                         zds->inBuffSize = neededInBuffSize;
@@ -1756,11 +2210,11 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
                         zds->outBuffSize = neededOutBuffSize;
             }   }   }
             zds->streamStage = zdss_read;
-            /* fall-through */
+            ZSTD_FALLTHROUGH;
 
         case zdss_read:
             DEBUGLOG(5, "stage zdss_read");
-            {   size_t const neededInSize = ZSTD_nextSrcSizeToDecompressWithInputSize(zds, iend - ip);
+            {   size_t const neededInSize = ZSTD_nextSrcSizeToDecompressWithInputSize(zds, (size_t)(iend - ip));
                 DEBUGLOG(5, "neededInSize = %u", (U32)neededInSize);
                 if (neededInSize==0) {  /* end of frame */
                     zds->streamStage = zdss_init;
@@ -1769,13 +2223,14 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
                 }
                 if ((size_t)(iend-ip) >= neededInSize) {  /* decode directly from src */
                     FORWARD_IF_ERROR(ZSTD_decompressContinueStream(zds, &op, oend, ip, neededInSize), "");
+                    assert(ip != NULL);
                     ip += neededInSize;
                     /* Function modifies the stage so we must break */
                     break;
             }   }
             if (ip==iend) { someMoreWork = 0; break; }   /* no more input */
             zds->streamStage = zdss_load;
-            /* fall-through */
+            ZSTD_FALLTHROUGH;
 
         case zdss_load:
             {   size_t const neededInSize = ZSTD_nextSrcSizeToDecompress(zds);
@@ -1783,17 +2238,20 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
                 int const isSkipFrame = ZSTD_isSkipFrame(zds);
                 size_t loadedSize;
                 /* At this point we shouldn't be decompressing a block that we can stream. */
-                assert(neededInSize == ZSTD_nextSrcSizeToDecompressWithInputSize(zds, iend - ip));
+                assert(neededInSize == ZSTD_nextSrcSizeToDecompressWithInputSize(zds, (size_t)(iend - ip)));
                 if (isSkipFrame) {
                     loadedSize = MIN(toLoad, (size_t)(iend-ip));
                 } else {
                     RETURN_ERROR_IF(toLoad > zds->inBuffSize - zds->inPos,
                                     corruption_detected,
                                     "should never happen");
-                    loadedSize = ZSTD_limitCopy(zds->inBuff + zds->inPos, toLoad, ip, iend-ip);
+                    loadedSize = ZSTD_limitCopy(zds->inBuff + zds->inPos, toLoad, ip, (size_t)(iend-ip));
+                }
+                if (loadedSize != 0) {
+                    /* ip may be NULL */
+                    ip += loadedSize;
+                    zds->inPos += loadedSize;
                 }
-                ip += loadedSize;
-                zds->inPos += loadedSize;
                 if (loadedSize < toLoad) { someMoreWork = 0; break; }   /* not enough input, wait for more */
 
                 /* decode loaded input */
@@ -1803,14 +2261,17 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
                 break;
             }
         case zdss_flush:
-            {   size_t const toFlushSize = zds->outEnd - zds->outStart;
-                size_t const flushedSize = ZSTD_limitCopy(op, oend-op, zds->outBuff + zds->outStart, toFlushSize);
-                op += flushedSize;
+            {
+                size_t const toFlushSize = zds->outEnd - zds->outStart;
+                size_t const flushedSize = ZSTD_limitCopy(op, (size_t)(oend-op), zds->outBuff + zds->outStart, toFlushSize);
+
+                op = op ? op + flushedSize : op;
+
                 zds->outStart += flushedSize;
                 if (flushedSize == toFlushSize) {  /* flush completed */
                     zds->streamStage = zdss_read;
                     if ( (zds->outBuffSize < zds->fParams.frameContentSize)
-                      && (zds->outStart + zds->fParams.blockSizeMax > zds->outBuffSize) ) {
+                        && (zds->outStart + zds->fParams.blockSizeMax > zds->outBuffSize) ) {
                         DEBUGLOG(5, "restart filling outBuff from beginning (left:%i, needed:%u)",
                                 (int)(zds->outBuffSize - zds->outStart),
                                 (U32)zds->fParams.blockSizeMax);
@@ -1824,7 +2285,7 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
 
         default:
             assert(0);    /* impossible */
-            RETURN_ERROR(GENERIC, "impossible to reach");   /* some compiler require default to do something */
+            RETURN_ERROR(GENERIC, "impossible to reach");   /* some compilers require default to do something */
     }   }
 
     /* result */
@@ -1837,8 +2298,8 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
     if ((ip==istart) && (op==ostart)) {  /* no forward progress */
         zds->noForwardProgress ++;
         if (zds->noForwardProgress >= ZSTD_NO_FORWARD_PROGRESS_MAX) {
-            RETURN_ERROR_IF(op==oend, dstSize_tooSmall, "");
-            RETURN_ERROR_IF(ip==iend, srcSize_wrong, "");
+            RETURN_ERROR_IF(op==oend, noForwardProgress_destFull, "");
+            RETURN_ERROR_IF(ip==iend, noForwardProgress_inputEmpty, "");
             assert(0);
         }
     } else {
@@ -1875,11 +2336,17 @@ size_t ZSTD_decompressStream_simpleArgs (
                             void* dst, size_t dstCapacity, size_t* dstPos,
                       const void* src, size_t srcSize, size_t* srcPos)
 {
-    ZSTD_outBuffer output = { dst, dstCapacity, *dstPos };
-    ZSTD_inBuffer  input  = { src, srcSize, *srcPos };
-    /* ZSTD_compress_generic() will check validity of dstPos and srcPos */
-    size_t const cErr = ZSTD_decompressStream(dctx, &output, &input);
-    *dstPos = output.pos;
-    *srcPos = input.pos;
-    return cErr;
+    ZSTD_outBuffer output;
+    ZSTD_inBuffer  input;
+    output.dst = dst;
+    output.size = dstCapacity;
+    output.pos = *dstPos;
+    input.src = src;
+    input.size = srcSize;
+    input.pos = *srcPos;
+    {   size_t const cErr = ZSTD_decompressStream(dctx, &output, &input);
+        *dstPos = output.pos;
+        *srcPos = input.pos;
+        return cErr;
+    }
 }
diff --git a/contrib/zstd/zstd_decompress_block.c b/contrib/zstd/zstd_decompress_block.c
index fcb381b99..add277e4f 100644
--- a/contrib/zstd/zstd_decompress_block.c
+++ b/contrib/zstd/zstd_decompress_block.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -14,18 +14,18 @@
 /*-*******************************************************
 *  Dependencies
 *********************************************************/
-#include <string.h>      /* memcpy, memmove, memset */
+#include "zstd_deps.h"   /* ZSTD_memcpy, ZSTD_memmove, ZSTD_memset */
 #include "compiler.h"    /* prefetch */
 #include "cpu.h"         /* bmi2 */
 #include "mem.h"         /* low level memory routines */
 #define FSE_STATIC_LINKING_ONLY
 #include "fse.h"
-#define HUF_STATIC_LINKING_ONLY
 #include "huf.h"
 #include "zstd_internal.h"
 #include "zstd_decompress_internal.h"   /* ZSTD_DCtx */
 #include "zstd_ddict.h"  /* ZSTD_DDictDictContent */
 #include "zstd_decompress_block.h"
+#include "bits.h"  /* ZSTD_highbit32 */
 
 /*_*******************************************************
 *  Macros
@@ -44,7 +44,7 @@
 /*_*******************************************************
 *  Memory operations
 **********************************************************/
-static void ZSTD_copy4(void* dst, const void* src) { memcpy(dst, src, 4); }
+static void ZSTD_copy4(void* dst, const void* src) { ZSTD_memcpy(dst, src, 4); }
 
 
 /*-*************************************************************
@@ -69,15 +69,56 @@ size_t ZSTD_getcBlockSize(const void* src, size_t srcSize,
     }
 }
 
+/* Allocate buffer for literals, either overlapping current dst, or split between dst and litExtraBuffer, or stored entirely within litExtraBuffer */
+static void ZSTD_allocateLiteralsBuffer(ZSTD_DCtx* dctx, void* const dst, const size_t dstCapacity, const size_t litSize,
+    const streaming_operation streaming, const size_t expectedWriteSize, const unsigned splitImmediately)
+{
+    if (streaming == not_streaming && dstCapacity > ZSTD_BLOCKSIZE_MAX + WILDCOPY_OVERLENGTH + litSize + WILDCOPY_OVERLENGTH)
+    {
+        /* room for litbuffer to fit without read faulting */
+        dctx->litBuffer = (BYTE*)dst + ZSTD_BLOCKSIZE_MAX + WILDCOPY_OVERLENGTH;
+        dctx->litBufferEnd = dctx->litBuffer + litSize;
+        dctx->litBufferLocation = ZSTD_in_dst;
+    }
+    else if (litSize > ZSTD_LITBUFFEREXTRASIZE)
+    {
+        /* won't fit in litExtraBuffer, so it will be split between end of dst and extra buffer */
+        if (splitImmediately) {
+            /* won't fit in litExtraBuffer, so it will be split between end of dst and extra buffer */
+            dctx->litBuffer = (BYTE*)dst + expectedWriteSize - litSize + ZSTD_LITBUFFEREXTRASIZE - WILDCOPY_OVERLENGTH;
+            dctx->litBufferEnd = dctx->litBuffer + litSize - ZSTD_LITBUFFEREXTRASIZE;
+        }
+        else {
+            /* initially this will be stored entirely in dst during huffman decoding, it will partially be shifted to litExtraBuffer after */
+            dctx->litBuffer = (BYTE*)dst + expectedWriteSize - litSize;
+            dctx->litBufferEnd = (BYTE*)dst + expectedWriteSize;
+        }
+        dctx->litBufferLocation = ZSTD_split;
+    }
+    else
+    {
+        /* fits entirely within litExtraBuffer, so no split is necessary */
+        dctx->litBuffer = dctx->litExtraBuffer;
+        dctx->litBufferEnd = dctx->litBuffer + litSize;
+        dctx->litBufferLocation = ZSTD_not_in_dst;
+    }
+}
 
 /* Hidden declaration for fullbench */
 size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
-                          const void* src, size_t srcSize);
+                          const void* src, size_t srcSize,
+                          void* dst, size_t dstCapacity, const streaming_operation streaming);
 /*! ZSTD_decodeLiteralsBlock() :
+ * Where it is possible to do so without being stomped by the output during decompression, the literals block will be stored
+ * in the dstBuffer.  If there is room to do so, it will be stored in full in the excess dst space after where the current
+ * block will be output.  Otherwise it will be stored at the end of the current dst blockspace, with a small portion being
+ * stored in dctx->litExtraBuffer to help keep it "ahead" of the current output write.
+ *
  * @return : nb of bytes read from src (< srcSize )
  *  note : symbol not declared but exposed for fullbench */
 size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
-                          const void* src, size_t srcSize)   /* note : srcSize < BLOCKSIZE */
+                          const void* src, size_t srcSize,   /* note : srcSize < BLOCKSIZE */
+                          void* dst, size_t dstCapacity, const streaming_operation streaming)
 {
     DEBUGLOG(5, "ZSTD_decodeLiteralsBlock");
     RETURN_ERROR_IF(srcSize < MIN_CBLOCK_SIZE, corruption_detected, "");
@@ -90,15 +131,19 @@ size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
         case set_repeat:
             DEBUGLOG(5, "set_repeat flag : re-using stats from previous compressed literals block");
             RETURN_ERROR_IF(dctx->litEntropy==0, dictionary_corrupted, "");
-            /* fall-through */
+            ZSTD_FALLTHROUGH;
 
         case set_compressed:
-            RETURN_ERROR_IF(srcSize < 5, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 3; here we need up to 5 for case 3");
+            RETURN_ERROR_IF(srcSize < 5, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need up to 5 for case 3");
             {   size_t lhSize, litSize, litCSize;
                 U32 singleStream=0;
                 U32 const lhlCode = (istart[0] >> 2) & 3;
                 U32 const lhc = MEM_readLE32(istart);
                 size_t hufSuccess;
+                size_t expectedWriteSize = MIN(ZSTD_BLOCKSIZE_MAX, dstCapacity);
+                int const flags = 0
+                    | (ZSTD_DCtx_get_bmi2(dctx) ? HUF_flags_bmi2 : 0)
+                    | (dctx->disableHufAsm ? HUF_flags_disableAsm : 0);
                 switch(lhlCode)
                 {
                 case 0: case 1: default:   /* note : default is impossible, since lhlCode into [0..3] */
@@ -121,8 +166,15 @@ size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
                     litCSize = (lhc >> 22) + ((size_t)istart[4] << 10);
                     break;
                 }
+                RETURN_ERROR_IF(litSize > 0 && dst == NULL, dstSize_tooSmall, "NULL not handled");
                 RETURN_ERROR_IF(litSize > ZSTD_BLOCKSIZE_MAX, corruption_detected, "");
+                if (!singleStream)
+                    RETURN_ERROR_IF(litSize < MIN_LITERALS_FOR_4_STREAMS, literals_headerWrong,
+                        "Not enough literals (%zu) for the 4-streams mode (min %u)",
+                        litSize, MIN_LITERALS_FOR_4_STREAMS);
                 RETURN_ERROR_IF(litCSize + lhSize > srcSize, corruption_detected, "");
+                RETURN_ERROR_IF(expectedWriteSize < litSize , dstSize_tooSmall, "");
+                ZSTD_allocateLiteralsBuffer(dctx, dst, dstCapacity, litSize, streaming, expectedWriteSize, 0);
 
                 /* prefetch huffman table if cold */
                 if (dctx->ddictIsCold && (litSize > 768 /* heuristic */)) {
@@ -131,13 +183,14 @@ size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
 
                 if (litEncType==set_repeat) {
                     if (singleStream) {
-                        hufSuccess = HUF_decompress1X_usingDTable_bmi2(
+                        hufSuccess = HUF_decompress1X_usingDTable(
                             dctx->litBuffer, litSize, istart+lhSize, litCSize,
-                            dctx->HUFptr, dctx->bmi2);
+                            dctx->HUFptr, flags);
                     } else {
-                        hufSuccess = HUF_decompress4X_usingDTable_bmi2(
+                        assert(litSize >= MIN_LITERALS_FOR_4_STREAMS);
+                        hufSuccess = HUF_decompress4X_usingDTable(
                             dctx->litBuffer, litSize, istart+lhSize, litCSize,
-                            dctx->HUFptr, dctx->bmi2);
+                            dctx->HUFptr, flags);
                     }
                 } else {
                     if (singleStream) {
@@ -145,20 +198,27 @@ size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
                         hufSuccess = HUF_decompress1X_DCtx_wksp(
                             dctx->entropy.hufTable, dctx->litBuffer, litSize,
                             istart+lhSize, litCSize, dctx->workspace,
-                            sizeof(dctx->workspace));
+                            sizeof(dctx->workspace), flags);
 #else
-                        hufSuccess = HUF_decompress1X1_DCtx_wksp_bmi2(
+                        hufSuccess = HUF_decompress1X1_DCtx_wksp(
                             dctx->entropy.hufTable, dctx->litBuffer, litSize,
                             istart+lhSize, litCSize, dctx->workspace,
-                            sizeof(dctx->workspace), dctx->bmi2);
+                            sizeof(dctx->workspace), flags);
 #endif
                     } else {
-                        hufSuccess = HUF_decompress4X_hufOnly_wksp_bmi2(
+                        hufSuccess = HUF_decompress4X_hufOnly_wksp(
                             dctx->entropy.hufTable, dctx->litBuffer, litSize,
                             istart+lhSize, litCSize, dctx->workspace,
-                            sizeof(dctx->workspace), dctx->bmi2);
+                            sizeof(dctx->workspace), flags);
                     }
                 }
+                if (dctx->litBufferLocation == ZSTD_split)
+                {
+                    ZSTD_memcpy(dctx->litExtraBuffer, dctx->litBufferEnd - ZSTD_LITBUFFEREXTRASIZE, ZSTD_LITBUFFEREXTRASIZE);
+                    ZSTD_memmove(dctx->litBuffer + ZSTD_LITBUFFEREXTRASIZE - WILDCOPY_OVERLENGTH, dctx->litBuffer, litSize - ZSTD_LITBUFFEREXTRASIZE);
+                    dctx->litBuffer += ZSTD_LITBUFFEREXTRASIZE - WILDCOPY_OVERLENGTH;
+                    dctx->litBufferEnd -= WILDCOPY_OVERLENGTH;
+                }
 
                 RETURN_ERROR_IF(HUF_isError(hufSuccess), corruption_detected, "");
 
@@ -166,13 +226,13 @@ size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
                 dctx->litSize = litSize;
                 dctx->litEntropy = 1;
                 if (litEncType==set_compressed) dctx->HUFptr = dctx->entropy.hufTable;
-                memset(dctx->litBuffer + dctx->litSize, 0, WILDCOPY_OVERLENGTH);
                 return litCSize + lhSize;
             }
 
         case set_basic:
             {   size_t litSize, lhSize;
                 U32 const lhlCode = ((istart[0]) >> 2) & 3;
+                size_t expectedWriteSize = MIN(ZSTD_BLOCKSIZE_MAX, dstCapacity);
                 switch(lhlCode)
                 {
                 case 0: case 2: default:   /* note : default is impossible, since lhlCode into [0..3] */
@@ -185,27 +245,41 @@ size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
                     break;
                 case 3:
                     lhSize = 3;
+                    RETURN_ERROR_IF(srcSize<3, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need lhSize = 3");
                     litSize = MEM_readLE24(istart) >> 4;
                     break;
                 }
 
+                RETURN_ERROR_IF(litSize > 0 && dst == NULL, dstSize_tooSmall, "NULL not handled");
+                RETURN_ERROR_IF(expectedWriteSize < litSize, dstSize_tooSmall, "");
+                ZSTD_allocateLiteralsBuffer(dctx, dst, dstCapacity, litSize, streaming, expectedWriteSize, 1);
                 if (lhSize+litSize+WILDCOPY_OVERLENGTH > srcSize) {  /* risk reading beyond src buffer with wildcopy */
                     RETURN_ERROR_IF(litSize+lhSize > srcSize, corruption_detected, "");
-                    memcpy(dctx->litBuffer, istart+lhSize, litSize);
+                    if (dctx->litBufferLocation == ZSTD_split)
+                    {
+                        ZSTD_memcpy(dctx->litBuffer, istart + lhSize, litSize - ZSTD_LITBUFFEREXTRASIZE);
+                        ZSTD_memcpy(dctx->litExtraBuffer, istart + lhSize + litSize - ZSTD_LITBUFFEREXTRASIZE, ZSTD_LITBUFFEREXTRASIZE);
+                    }
+                    else
+                    {
+                        ZSTD_memcpy(dctx->litBuffer, istart + lhSize, litSize);
+                    }
                     dctx->litPtr = dctx->litBuffer;
                     dctx->litSize = litSize;
-                    memset(dctx->litBuffer + dctx->litSize, 0, WILDCOPY_OVERLENGTH);
                     return lhSize+litSize;
                 }
                 /* direct reference into compressed stream */
                 dctx->litPtr = istart+lhSize;
                 dctx->litSize = litSize;
+                dctx->litBufferEnd = dctx->litPtr + litSize;
+                dctx->litBufferLocation = ZSTD_not_in_dst;
                 return lhSize+litSize;
             }
 
         case set_rle:
             {   U32 const lhlCode = ((istart[0]) >> 2) & 3;
                 size_t litSize, lhSize;
+                size_t expectedWriteSize = MIN(ZSTD_BLOCKSIZE_MAX, dstCapacity);
                 switch(lhlCode)
                 {
                 case 0: case 2: default:   /* note : default is impossible, since lhlCode into [0..3] */
@@ -214,16 +288,28 @@ size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
                     break;
                 case 1:
                     lhSize = 2;
+                    RETURN_ERROR_IF(srcSize<3, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need lhSize+1 = 3");
                     litSize = MEM_readLE16(istart) >> 4;
                     break;
                 case 3:
                     lhSize = 3;
+                    RETURN_ERROR_IF(srcSize<4, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need lhSize+1 = 4");
                     litSize = MEM_readLE24(istart) >> 4;
-                    RETURN_ERROR_IF(srcSize<4, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 3; here we need lhSize+1 = 4");
                     break;
                 }
+                RETURN_ERROR_IF(litSize > 0 && dst == NULL, dstSize_tooSmall, "NULL not handled");
                 RETURN_ERROR_IF(litSize > ZSTD_BLOCKSIZE_MAX, corruption_detected, "");
-                memset(dctx->litBuffer, istart[lhSize], litSize + WILDCOPY_OVERLENGTH);
+                RETURN_ERROR_IF(expectedWriteSize < litSize, dstSize_tooSmall, "");
+                ZSTD_allocateLiteralsBuffer(dctx, dst, dstCapacity, litSize, streaming, expectedWriteSize, 1);
+                if (dctx->litBufferLocation == ZSTD_split)
+                {
+                    ZSTD_memset(dctx->litBuffer, istart[lhSize], litSize - ZSTD_LITBUFFEREXTRASIZE);
+                    ZSTD_memset(dctx->litExtraBuffer, istart[lhSize], ZSTD_LITBUFFEREXTRASIZE);
+                }
+                else
+                {
+                    ZSTD_memset(dctx->litBuffer, istart[lhSize], litSize);
+                }
                 dctx->litPtr = dctx->litBuffer;
                 dctx->litSize = litSize;
                 return lhSize+1;
@@ -236,7 +322,7 @@ size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
 
 /* Default FSE distribution tables.
  * These are pre-calculated FSE decoding tables using default distributions as defined in specification :
- * https://github.com/facebook/zstd/blob/master/doc/zstd_compression_format.md#default-distributions
+ * https://github.com/facebook/zstd/blob/release/doc/zstd_compression_format.md#default-distributions
  * They were generated programmatically with following method :
  * - start from default distributions, present in /lib/common/zstd_internal.h
  * - generate tables normally, using ZSTD_buildFSETable()
@@ -343,7 +429,7 @@ static const ZSTD_seqSymbol ML_defaultDTable[(1<<ML_DEFAULTNORMLOG)+1] = {
 };   /* ML_defaultDTable */
 
 
-static void ZSTD_buildSeqTable_rle(ZSTD_seqSymbol* dt, U32 baseValue, U32 nbAddBits)
+static void ZSTD_buildSeqTable_rle(ZSTD_seqSymbol* dt, U32 baseValue, U8 nbAddBits)
 {
     void* ptr = dt;
     ZSTD_seqSymbol_header* const DTableH = (ZSTD_seqSymbol_header*)ptr;
@@ -355,7 +441,7 @@ static void ZSTD_buildSeqTable_rle(ZSTD_seqSymbol* dt, U32 baseValue, U32 nbAddB
     cell->nbBits = 0;
     cell->nextState = 0;
     assert(nbAddBits < 255);
-    cell->nbAdditionalBits = (BYTE)nbAddBits;
+    cell->nbAdditionalBits = nbAddBits;
     cell->baseValue = baseValue;
 }
 
@@ -364,23 +450,26 @@ static void ZSTD_buildSeqTable_rle(ZSTD_seqSymbol* dt, U32 baseValue, U32 nbAddB
  * generate FSE decoding table for one symbol (ll, ml or off)
  * cannot fail if input is valid =>
  * all inputs are presumed validated at this stage */
-void
-ZSTD_buildFSETable(ZSTD_seqSymbol* dt,
+FORCE_INLINE_TEMPLATE
+void ZSTD_buildFSETable_body(ZSTD_seqSymbol* dt,
             const short* normalizedCounter, unsigned maxSymbolValue,
-            const U32* baseValue, const U32* nbAdditionalBits,
-            unsigned tableLog)
+            const U32* baseValue, const U8* nbAdditionalBits,
+            unsigned tableLog, void* wksp, size_t wkspSize)
 {
     ZSTD_seqSymbol* const tableDecode = dt+1;
-    U16 symbolNext[MaxSeq+1];
-
     U32 const maxSV1 = maxSymbolValue + 1;
     U32 const tableSize = 1 << tableLog;
-    U32 highThreshold = tableSize-1;
+
+    U16* symbolNext = (U16*)wksp;
+    BYTE* spread = (BYTE*)(symbolNext + MaxSeq + 1);
+    U32 highThreshold = tableSize - 1;
+
 
     /* Sanity Checks */
     assert(maxSymbolValue <= MaxSeq);
     assert(tableLog <= MaxFSELog);
-
+    assert(wkspSize >= ZSTD_BUILD_FSE_TABLE_WKSP_SIZE);
+    (void)wkspSize;
     /* Init, lay down lowprob symbols */
     {   ZSTD_seqSymbol_header DTableH;
         DTableH.tableLog = tableLog;
@@ -396,34 +485,128 @@ ZSTD_buildFSETable(ZSTD_seqSymbol* dt,
                     assert(normalizedCounter[s]>=0);
                     symbolNext[s] = (U16)normalizedCounter[s];
         }   }   }
-        memcpy(dt, &DTableH, sizeof(DTableH));
+        ZSTD_memcpy(dt, &DTableH, sizeof(DTableH));
     }
 
     /* Spread symbols */
-    {   U32 const tableMask = tableSize-1;
+    assert(tableSize <= 512);
+    /* Specialized symbol spreading for the case when there are
+     * no low probability (-1 count) symbols. When compressing
+     * small blocks we avoid low probability symbols to hit this
+     * case, since header decoding speed matters more.
+     */
+    if (highThreshold == tableSize - 1) {
+        size_t const tableMask = tableSize-1;
+        size_t const step = FSE_TABLESTEP(tableSize);
+        /* First lay down the symbols in order.
+         * We use a uint64_t to lay down 8 bytes at a time. This reduces branch
+         * misses since small blocks generally have small table logs, so nearly
+         * all symbols have counts <= 8. We ensure we have 8 bytes at the end of
+         * our buffer to handle the over-write.
+         */
+        {
+            U64 const add = 0x0101010101010101ull;
+            size_t pos = 0;
+            U64 sv = 0;
+            U32 s;
+            for (s=0; s<maxSV1; ++s, sv += add) {
+                int i;
+                int const n = normalizedCounter[s];
+                MEM_write64(spread + pos, sv);
+                for (i = 8; i < n; i += 8) {
+                    MEM_write64(spread + pos + i, sv);
+                }
+                assert(n>=0);
+                pos += (size_t)n;
+            }
+        }
+        /* Now we spread those positions across the table.
+         * The benefit of doing it in two stages is that we avoid the
+         * variable size inner loop, which caused lots of branch misses.
+         * Now we can run through all the positions without any branch misses.
+         * We unroll the loop twice, since that is what empirically worked best.
+         */
+        {
+            size_t position = 0;
+            size_t s;
+            size_t const unroll = 2;
+            assert(tableSize % unroll == 0); /* FSE_MIN_TABLELOG is 5 */
+            for (s = 0; s < (size_t)tableSize; s += unroll) {
+                size_t u;
+                for (u = 0; u < unroll; ++u) {
+                    size_t const uPosition = (position + (u * step)) & tableMask;
+                    tableDecode[uPosition].baseValue = spread[s + u];
+                }
+                position = (position + (unroll * step)) & tableMask;
+            }
+            assert(position == 0);
+        }
+    } else {
+        U32 const tableMask = tableSize-1;
         U32 const step = FSE_TABLESTEP(tableSize);
         U32 s, position = 0;
         for (s=0; s<maxSV1; s++) {
             int i;
-            for (i=0; i<normalizedCounter[s]; i++) {
+            int const n = normalizedCounter[s];
+            for (i=0; i<n; i++) {
                 tableDecode[position].baseValue = s;
                 position = (position + step) & tableMask;
-                while (position > highThreshold) position = (position + step) & tableMask;   /* lowprob area */
+                while (UNLIKELY(position > highThreshold)) position = (position + step) & tableMask;   /* lowprob area */
         }   }
         assert(position == 0); /* position must reach all cells once, otherwise normalizedCounter is incorrect */
     }
 
     /* Build Decoding table */
-    {   U32 u;
+    {
+        U32 u;
         for (u=0; u<tableSize; u++) {
             U32 const symbol = tableDecode[u].baseValue;
             U32 const nextState = symbolNext[symbol]++;
-            tableDecode[u].nbBits = (BYTE) (tableLog - BIT_highbit32(nextState) );
+            tableDecode[u].nbBits = (BYTE) (tableLog - ZSTD_highbit32(nextState) );
             tableDecode[u].nextState = (U16) ( (nextState << tableDecode[u].nbBits) - tableSize);
             assert(nbAdditionalBits[symbol] < 255);
-            tableDecode[u].nbAdditionalBits = (BYTE)nbAdditionalBits[symbol];
+            tableDecode[u].nbAdditionalBits = nbAdditionalBits[symbol];
             tableDecode[u].baseValue = baseValue[symbol];
-    }   }
+        }
+    }
+}
+
+/* Avoids the FORCE_INLINE of the _body() function. */
+static void ZSTD_buildFSETable_body_default(ZSTD_seqSymbol* dt,
+            const short* normalizedCounter, unsigned maxSymbolValue,
+            const U32* baseValue, const U8* nbAdditionalBits,
+            unsigned tableLog, void* wksp, size_t wkspSize)
+{
+    ZSTD_buildFSETable_body(dt, normalizedCounter, maxSymbolValue,
+            baseValue, nbAdditionalBits, tableLog, wksp, wkspSize);
+}
+
+#if DYNAMIC_BMI2
+BMI2_TARGET_ATTRIBUTE static void ZSTD_buildFSETable_body_bmi2(ZSTD_seqSymbol* dt,
+            const short* normalizedCounter, unsigned maxSymbolValue,
+            const U32* baseValue, const U8* nbAdditionalBits,
+            unsigned tableLog, void* wksp, size_t wkspSize)
+{
+    ZSTD_buildFSETable_body(dt, normalizedCounter, maxSymbolValue,
+            baseValue, nbAdditionalBits, tableLog, wksp, wkspSize);
+}
+#endif
+
+void ZSTD_buildFSETable(ZSTD_seqSymbol* dt,
+            const short* normalizedCounter, unsigned maxSymbolValue,
+            const U32* baseValue, const U8* nbAdditionalBits,
+            unsigned tableLog, void* wksp, size_t wkspSize, int bmi2)
+{
+#if DYNAMIC_BMI2
+    if (bmi2) {
+        ZSTD_buildFSETable_body_bmi2(dt, normalizedCounter, maxSymbolValue,
+                baseValue, nbAdditionalBits, tableLog, wksp, wkspSize);
+        return;
+    }
+#endif
+    (void)bmi2;
+    ZSTD_buildFSETable_body_default(dt, normalizedCounter, maxSymbolValue,
+            baseValue, nbAdditionalBits, tableLog, wksp, wkspSize);
 }
 
 
@@ -433,9 +616,10 @@ ZSTD_buildFSETable(ZSTD_seqSymbol* dt,
 static size_t ZSTD_buildSeqTable(ZSTD_seqSymbol* DTableSpace, const ZSTD_seqSymbol** DTablePtr,
                                  symbolEncodingType_e type, unsigned max, U32 maxLog,
                                  const void* src, size_t srcSize,
-                                 const U32* baseValue, const U32* nbAdditionalBits,
+                                 const U32* baseValue, const U8* nbAdditionalBits,
                                  const ZSTD_seqSymbol* defaultTable, U32 flagRepeatTable,
-                                 int ddictIsCold, int nbSeq)
+                                 int ddictIsCold, int nbSeq, U32* wksp, size_t wkspSize,
+                                 int bmi2)
 {
     switch(type)
     {
@@ -444,7 +628,7 @@ static size_t ZSTD_buildSeqTable(ZSTD_seqSymbol* DTableSpace, const ZSTD_seqSymb
         RETURN_ERROR_IF((*(const BYTE*)src) > max, corruption_detected, "");
         {   U32 const symbol = *(const BYTE*)src;
             U32 const baseline = baseValue[symbol];
-            U32 const nbBits = nbAdditionalBits[symbol];
+            U8 const nbBits = nbAdditionalBits[symbol];
             ZSTD_buildSeqTable_rle(DTableSpace, baseline, nbBits);
         }
         *DTablePtr = DTableSpace;
@@ -467,7 +651,7 @@ static size_t ZSTD_buildSeqTable(ZSTD_seqSymbol* DTableSpace, const ZSTD_seqSymb
             size_t const headerSize = FSE_readNCount(norm, &max, &tableLog, src, srcSize);
             RETURN_ERROR_IF(FSE_isError(headerSize), corruption_detected, "");
             RETURN_ERROR_IF(tableLog > maxLog, corruption_detected, "");
-            ZSTD_buildFSETable(DTableSpace, norm, max, baseValue, nbAdditionalBits, tableLog);
+            ZSTD_buildFSETable(DTableSpace, norm, max, baseValue, nbAdditionalBits, tableLog, wksp, wkspSize, bmi2);
             *DTablePtr = DTableSpace;
             return headerSize;
         }
@@ -480,7 +664,7 @@ static size_t ZSTD_buildSeqTable(ZSTD_seqSymbol* DTableSpace, const ZSTD_seqSymb
 size_t ZSTD_decodeSeqHeaders(ZSTD_DCtx* dctx, int* nbSeqPtr,
                              const void* src, size_t srcSize)
 {
-    const BYTE* const istart = (const BYTE* const)src;
+    const BYTE* const istart = (const BYTE*)src;
     const BYTE* const iend = istart + srcSize;
     const BYTE* ip = istart;
     int nbSeq;
@@ -499,7 +683,8 @@ size_t ZSTD_decodeSeqHeaders(ZSTD_DCtx* dctx, int* nbSeqPtr,
     if (nbSeq > 0x7F) {
         if (nbSeq == 0xFF) {
             RETURN_ERROR_IF(ip+2 > iend, srcSize_wrong, "");
-            nbSeq = MEM_readLE16(ip) + LONGNBSEQ, ip+=2;
+            nbSeq = MEM_readLE16(ip) + LONGNBSEQ;
+            ip+=2;
         } else {
             RETURN_ERROR_IF(ip >= iend, srcSize_wrong, "");
             nbSeq = ((nbSeq-0x80)<<8) + *ip++;
@@ -520,7 +705,9 @@ size_t ZSTD_decodeSeqHeaders(ZSTD_DCtx* dctx, int* nbSeqPtr,
                                                       ip, iend-ip,
                                                       LL_base, LL_bits,
                                                       LL_defaultDTable, dctx->fseEntropy,
-                                                      dctx->ddictIsCold, nbSeq);
+                                                      dctx->ddictIsCold, nbSeq,
+                                                      dctx->workspace, sizeof(dctx->workspace),
+                                                      ZSTD_DCtx_get_bmi2(dctx));
             RETURN_ERROR_IF(ZSTD_isError(llhSize), corruption_detected, "ZSTD_buildSeqTable failed");
             ip += llhSize;
         }
@@ -530,7 +717,9 @@ size_t ZSTD_decodeSeqHeaders(ZSTD_DCtx* dctx, int* nbSeqPtr,
                                                       ip, iend-ip,
                                                       OF_base, OF_bits,
                                                       OF_defaultDTable, dctx->fseEntropy,
-                                                      dctx->ddictIsCold, nbSeq);
+                                                      dctx->ddictIsCold, nbSeq,
+                                                      dctx->workspace, sizeof(dctx->workspace),
+                                                      ZSTD_DCtx_get_bmi2(dctx));
             RETURN_ERROR_IF(ZSTD_isError(ofhSize), corruption_detected, "ZSTD_buildSeqTable failed");
             ip += ofhSize;
         }
@@ -540,7 +729,9 @@ size_t ZSTD_decodeSeqHeaders(ZSTD_DCtx* dctx, int* nbSeqPtr,
                                                       ip, iend-ip,
                                                       ML_base, ML_bits,
                                                       ML_defaultDTable, dctx->fseEntropy,
-                                                      dctx->ddictIsCold, nbSeq);
+                                                      dctx->ddictIsCold, nbSeq,
+                                                      dctx->workspace, sizeof(dctx->workspace),
+                                                      ZSTD_DCtx_get_bmi2(dctx));
             RETURN_ERROR_IF(ZSTD_isError(mlhSize), corruption_detected, "ZSTD_buildSeqTable failed");
             ip += mlhSize;
         }
@@ -554,7 +745,6 @@ typedef struct {
     size_t litLength;
     size_t matchLength;
     size_t offset;
-    const BYTE* match;
 } seq_t;
 
 typedef struct {
@@ -568,9 +758,6 @@ typedef struct {
     ZSTD_fseState stateOffb;
     ZSTD_fseState stateML;
     size_t prevOffset[ZSTD_REP_NUM];
-    const BYTE* prefixStart;
-    const BYTE* dictEnd;
-    size_t pos;
 } seqState_t;
 
 /*! ZSTD_overlapCopy8() :
@@ -613,7 +800,7 @@ HINT_INLINE void ZSTD_overlapCopy8(BYTE** op, BYTE const** ip, size_t offset) {
  *         - ZSTD_overlap_src_before_dst: The src and dst may overlap and may be any distance apart.
  *           The src buffer must be before the dst buffer.
  */
-static void ZSTD_safecopy(BYTE* op, BYTE* const oend_w, BYTE const* ip, ptrdiff_t length, ZSTD_overlap_e ovtype) {
+static void ZSTD_safecopy(BYTE* op, const BYTE* const oend_w, BYTE const* ip, ptrdiff_t length, ZSTD_overlap_e ovtype) {
     ptrdiff_t const diff = op - ip;
     BYTE* const oend = op + length;
 
@@ -629,6 +816,7 @@ static void ZSTD_safecopy(BYTE* op, BYTE* const oend_w, BYTE const* ip, ptrdiff_
         /* Copy 8 bytes and ensure the offset >= 8 when there can be overlap. */
         assert(length >= 8);
         ZSTD_overlapCopy8(&op, &ip, diff);
+        length -= 8;
         assert(op - ip >= 8);
         assert(op <= oend);
     }
@@ -643,12 +831,35 @@ static void ZSTD_safecopy(BYTE* op, BYTE* const oend_w, BYTE const* ip, ptrdiff_
         assert(oend > oend_w);
         ZSTD_wildcopy(op, ip, oend_w - op, ovtype);
         ip += oend_w - op;
-        op = oend_w;
+        op += oend_w - op;
     }
     /* Handle the leftovers. */
     while (op < oend) *op++ = *ip++;
 }
 
+/* ZSTD_safecopyDstBeforeSrc():
+ * This version allows overlap with dst before src, or handles the non-overlap case with dst after src
+ * Kept separate from more common ZSTD_safecopy case to avoid performance impact to the safecopy common case */
+static void ZSTD_safecopyDstBeforeSrc(BYTE* op, BYTE const* ip, ptrdiff_t length) {
+    ptrdiff_t const diff = op - ip;
+    BYTE* const oend = op + length;
+
+    if (length < 8 || diff > -8) {
+        /* Handle short lengths, close overlaps, and dst not before src. */
+        while (op < oend) *op++ = *ip++;
+        return;
+    }
+
+    if (op <= oend - WILDCOPY_OVERLENGTH && diff < -WILDCOPY_VECLEN) {
+        ZSTD_wildcopy(op, ip, oend - WILDCOPY_OVERLENGTH - op, ZSTD_no_overlap);
+        ip += oend - WILDCOPY_OVERLENGTH - op;
+        op += oend - WILDCOPY_OVERLENGTH - op;
+    }
+
+    /* Handle the leftovers. */
+    while (op < oend) *op++ = *ip++;
+}
+
 /* ZSTD_execSequenceEnd():
  * This version handles cases that are near the end of the output buffer. It requires
  * more careful checks to make sure there is no overflow. By separating out these hard
@@ -659,9 +870,9 @@ static void ZSTD_safecopy(BYTE* op, BYTE* const oend_w, BYTE const* ip, ptrdiff_
  */
 FORCE_NOINLINE
 size_t ZSTD_execSequenceEnd(BYTE* op,
-                            BYTE* const oend, seq_t sequence,
-                            const BYTE** litPtr, const BYTE* const litLimit,
-                            const BYTE* const prefixStart, const BYTE* const virtualStart, const BYTE* const dictEnd)
+    BYTE* const oend, seq_t sequence,
+    const BYTE** litPtr, const BYTE* const litLimit,
+    const BYTE* const prefixStart, const BYTE* const virtualStart, const BYTE* const dictEnd)
 {
     BYTE* const oLitEnd = op + sequence.litLength;
     size_t const sequenceLength = sequence.litLength + sequence.matchLength;
@@ -684,27 +895,76 @@ size_t ZSTD_execSequenceEnd(BYTE* op,
     if (sequence.offset > (size_t)(oLitEnd - prefixStart)) {
         /* offset beyond prefix */
         RETURN_ERROR_IF(sequence.offset > (size_t)(oLitEnd - virtualStart), corruption_detected, "");
-        match = dictEnd - (prefixStart-match);
+        match = dictEnd - (prefixStart - match);
         if (match + sequence.matchLength <= dictEnd) {
-            memmove(oLitEnd, match, sequence.matchLength);
+            ZSTD_memmove(oLitEnd, match, sequence.matchLength);
             return sequenceLength;
         }
         /* span extDict & currentPrefixSegment */
         {   size_t const length1 = dictEnd - match;
-            memmove(oLitEnd, match, length1);
-            op = oLitEnd + length1;
-            sequence.matchLength -= length1;
-            match = prefixStart;
-    }   }
+        ZSTD_memmove(oLitEnd, match, length1);
+        op = oLitEnd + length1;
+        sequence.matchLength -= length1;
+        match = prefixStart;
+        }
+    }
+    ZSTD_safecopy(op, oend_w, match, sequence.matchLength, ZSTD_overlap_src_before_dst);
+    return sequenceLength;
+}
+
+/* ZSTD_execSequenceEndSplitLitBuffer():
+ * This version is intended to be used during instances where the litBuffer is still split.  It is kept separate to avoid performance impact for the good case.
+ */
+FORCE_NOINLINE
+size_t ZSTD_execSequenceEndSplitLitBuffer(BYTE* op,
+    BYTE* const oend, const BYTE* const oend_w, seq_t sequence,
+    const BYTE** litPtr, const BYTE* const litLimit,
+    const BYTE* const prefixStart, const BYTE* const virtualStart, const BYTE* const dictEnd)
+{
+    BYTE* const oLitEnd = op + sequence.litLength;
+    size_t const sequenceLength = sequence.litLength + sequence.matchLength;
+    const BYTE* const iLitEnd = *litPtr + sequence.litLength;
+    const BYTE* match = oLitEnd - sequence.offset;
+
+
+    /* bounds checks : careful of address space overflow in 32-bit mode */
+    RETURN_ERROR_IF(sequenceLength > (size_t)(oend - op), dstSize_tooSmall, "last match must fit within dstBuffer");
+    RETURN_ERROR_IF(sequence.litLength > (size_t)(litLimit - *litPtr), corruption_detected, "try to read beyond literal buffer");
+    assert(op < op + sequenceLength);
+    assert(oLitEnd < op + sequenceLength);
+
+    /* copy literals */
+    RETURN_ERROR_IF(op > *litPtr && op < *litPtr + sequence.litLength, dstSize_tooSmall, "output should not catch up to and overwrite literal buffer");
+    ZSTD_safecopyDstBeforeSrc(op, *litPtr, sequence.litLength);
+    op = oLitEnd;
+    *litPtr = iLitEnd;
+
+    /* copy Match */
+    if (sequence.offset > (size_t)(oLitEnd - prefixStart)) {
+        /* offset beyond prefix */
+        RETURN_ERROR_IF(sequence.offset > (size_t)(oLitEnd - virtualStart), corruption_detected, "");
+        match = dictEnd - (prefixStart - match);
+        if (match + sequence.matchLength <= dictEnd) {
+            ZSTD_memmove(oLitEnd, match, sequence.matchLength);
+            return sequenceLength;
+        }
+        /* span extDict & currentPrefixSegment */
+        {   size_t const length1 = dictEnd - match;
+        ZSTD_memmove(oLitEnd, match, length1);
+        op = oLitEnd + length1;
+        sequence.matchLength -= length1;
+        match = prefixStart;
+        }
+    }
     ZSTD_safecopy(op, oend_w, match, sequence.matchLength, ZSTD_overlap_src_before_dst);
     return sequenceLength;
 }
 
 HINT_INLINE
 size_t ZSTD_execSequence(BYTE* op,
-                         BYTE* const oend, seq_t sequence,
-                         const BYTE** litPtr, const BYTE* const litLimit,
-                         const BYTE* const prefixStart, const BYTE* const virtualStart, const BYTE* const dictEnd)
+    BYTE* const oend, seq_t sequence,
+    const BYTE** litPtr, const BYTE* const litLimit,
+    const BYTE* const prefixStart, const BYTE* const virtualStart, const BYTE* const dictEnd)
 {
     BYTE* const oLitEnd = op + sequence.litLength;
     size_t const sequenceLength = sequence.litLength + sequence.matchLength;
@@ -715,6 +975,103 @@ size_t ZSTD_execSequence(BYTE* op,
 
     assert(op != NULL /* Precondition */);
     assert(oend_w < oend /* No underflow */);
+
+#if defined(__aarch64__)
+    /* prefetch sequence starting from match that will be used for copy later */
+    PREFETCH_L1(match);
+#endif
+    /* Handle edge cases in a slow path:
+     *   - Read beyond end of literals
+     *   - Match end is within WILDCOPY_OVERLIMIT of oend
+     *   - 32-bit mode and the match length overflows
+     */
+    if (UNLIKELY(
+        iLitEnd > litLimit ||
+        oMatchEnd > oend_w ||
+        (MEM_32bits() && (size_t)(oend - op) < sequenceLength + WILDCOPY_OVERLENGTH)))
+        return ZSTD_execSequenceEnd(op, oend, sequence, litPtr, litLimit, prefixStart, virtualStart, dictEnd);
+
+    /* Assumptions (everything else goes into ZSTD_execSequenceEnd()) */
+    assert(op <= oLitEnd /* No overflow */);
+    assert(oLitEnd < oMatchEnd /* Non-zero match & no overflow */);
+    assert(oMatchEnd <= oend /* No underflow */);
+    assert(iLitEnd <= litLimit /* Literal length is in bounds */);
+    assert(oLitEnd <= oend_w /* Can wildcopy literals */);
+    assert(oMatchEnd <= oend_w /* Can wildcopy matches */);
+
+    /* Copy Literals:
+     * Split out litLength <= 16 since it is nearly always true. +1.6% on gcc-9.
+     * We likely don't need the full 32-byte wildcopy.
+     */
+    assert(WILDCOPY_OVERLENGTH >= 16);
+    ZSTD_copy16(op, (*litPtr));
+    if (UNLIKELY(sequence.litLength > 16)) {
+        ZSTD_wildcopy(op + 16, (*litPtr) + 16, sequence.litLength - 16, ZSTD_no_overlap);
+    }
+    op = oLitEnd;
+    *litPtr = iLitEnd;   /* update for next sequence */
+
+    /* Copy Match */
+    if (sequence.offset > (size_t)(oLitEnd - prefixStart)) {
+        /* offset beyond prefix -> go into extDict */
+        RETURN_ERROR_IF(UNLIKELY(sequence.offset > (size_t)(oLitEnd - virtualStart)), corruption_detected, "");
+        match = dictEnd + (match - prefixStart);
+        if (match + sequence.matchLength <= dictEnd) {
+            ZSTD_memmove(oLitEnd, match, sequence.matchLength);
+            return sequenceLength;
+        }
+        /* span extDict & currentPrefixSegment */
+        {   size_t const length1 = dictEnd - match;
+        ZSTD_memmove(oLitEnd, match, length1);
+        op = oLitEnd + length1;
+        sequence.matchLength -= length1;
+        match = prefixStart;
+        }
+    }
+    /* Match within prefix of 1 or more bytes */
+    assert(op <= oMatchEnd);
+    assert(oMatchEnd <= oend_w);
+    assert(match >= prefixStart);
+    assert(sequence.matchLength >= 1);
+
+    /* Nearly all offsets are >= WILDCOPY_VECLEN bytes, which means we can use wildcopy
+     * without overlap checking.
+     */
+    if (LIKELY(sequence.offset >= WILDCOPY_VECLEN)) {
+        /* We bet on a full wildcopy for matches, since we expect matches to be
+         * longer than literals (in general). In silesia, ~10% of matches are longer
+         * than 16 bytes.
+         */
+        ZSTD_wildcopy(op, match, (ptrdiff_t)sequence.matchLength, ZSTD_no_overlap);
+        return sequenceLength;
+    }
+    assert(sequence.offset < WILDCOPY_VECLEN);
+
+    /* Copy 8 bytes and spread the offset to be >= 8. */
+    ZSTD_overlapCopy8(&op, &match, sequence.offset);
+
+    /* If the match length is > 8 bytes, then continue with the wildcopy. */
+    if (sequence.matchLength > 8) {
+        assert(op < oMatchEnd);
+        ZSTD_wildcopy(op, match, (ptrdiff_t)sequence.matchLength - 8, ZSTD_overlap_src_before_dst);
+    }
+    return sequenceLength;
+}
+
+HINT_INLINE
+size_t ZSTD_execSequenceSplitLitBuffer(BYTE* op,
+    BYTE* const oend, const BYTE* const oend_w, seq_t sequence,
+    const BYTE** litPtr, const BYTE* const litLimit,
+    const BYTE* const prefixStart, const BYTE* const virtualStart, const BYTE* const dictEnd)
+{
+    BYTE* const oLitEnd = op + sequence.litLength;
+    size_t const sequenceLength = sequence.litLength + sequence.matchLength;
+    BYTE* const oMatchEnd = op + sequenceLength;   /* risk : address space overflow (32-bits) */
+    const BYTE* const iLitEnd = *litPtr + sequence.litLength;
+    const BYTE* match = oLitEnd - sequence.offset;
+
+    assert(op != NULL /* Precondition */);
+    assert(oend_w < oend /* No underflow */);
     /* Handle edge cases in a slow path:
      *   - Read beyond end of literals
      *   - Match end is within WILDCOPY_OVERLIMIT of oend
@@ -724,7 +1081,7 @@ size_t ZSTD_execSequence(BYTE* op,
             iLitEnd > litLimit ||
             oMatchEnd > oend_w ||
             (MEM_32bits() && (size_t)(oend - op) < sequenceLength + WILDCOPY_OVERLENGTH)))
-        return ZSTD_execSequenceEnd(op, oend, sequence, litPtr, litLimit, prefixStart, virtualStart, dictEnd);
+        return ZSTD_execSequenceEndSplitLitBuffer(op, oend, oend_w, sequence, litPtr, litLimit, prefixStart, virtualStart, dictEnd);
 
     /* Assumptions (everything else goes into ZSTD_execSequenceEnd()) */
     assert(op <= oLitEnd /* No overflow */);
@@ -752,12 +1109,12 @@ size_t ZSTD_execSequence(BYTE* op,
         RETURN_ERROR_IF(UNLIKELY(sequence.offset > (size_t)(oLitEnd - virtualStart)), corruption_detected, "");
         match = dictEnd + (match - prefixStart);
         if (match + sequence.matchLength <= dictEnd) {
-            memmove(oLitEnd, match, sequence.matchLength);
+            ZSTD_memmove(oLitEnd, match, sequence.matchLength);
             return sequenceLength;
         }
         /* span extDict & currentPrefixSegment */
         {   size_t const length1 = dictEnd - match;
-            memmove(oLitEnd, match, length1);
+            ZSTD_memmove(oLitEnd, match, length1);
             op = oLitEnd + length1;
             sequence.matchLength -= length1;
             match = prefixStart;
@@ -792,6 +1149,7 @@ size_t ZSTD_execSequence(BYTE* op,
     return sequenceLength;
 }
 
+
 static void
 ZSTD_initFseState(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD, const ZSTD_seqSymbol* dt)
 {
@@ -805,24 +1163,14 @@ ZSTD_initFseState(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD, const ZSTD_seqS
 }
 
 FORCE_INLINE_TEMPLATE void
-ZSTD_updateFseState(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD)
-{
-    ZSTD_seqSymbol const DInfo = DStatePtr->table[DStatePtr->state];
-    U32 const nbBits = DInfo.nbBits;
-    size_t const lowBits = BIT_readBits(bitD, nbBits);
-    DStatePtr->state = DInfo.nextState + lowBits;
-}
-
-FORCE_INLINE_TEMPLATE void
-ZSTD_updateFseStateWithDInfo(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD, ZSTD_seqSymbol const DInfo)
+ZSTD_updateFseStateWithDInfo(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD, U16 nextState, U32 nbBits)
 {
-    U32 const nbBits = DInfo.nbBits;
     size_t const lowBits = BIT_readBits(bitD, nbBits);
-    DStatePtr->state = DInfo.nextState + lowBits;
+    DStatePtr->state = nextState + lowBits;
 }
 
 /* We need to add at most (ZSTD_WINDOWLOG_MAX_32 - 1) bits to read the maximum
- * offset bits. But we can only read at most (STREAM_ACCUMULATOR_MIN_32 - 1)
+ * offset bits. But we can only read at most STREAM_ACCUMULATOR_MIN_32
  * bits before reloading. This value is the maximum number of bytes we read
  * after reloading when we are decoding long offsets.
  */
@@ -832,123 +1180,137 @@ ZSTD_updateFseStateWithDInfo(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD, ZSTD
         : 0)
 
 typedef enum { ZSTD_lo_isRegularOffset, ZSTD_lo_isLongOffset=1 } ZSTD_longOffset_e;
-typedef enum { ZSTD_p_noPrefetch=0, ZSTD_p_prefetch=1 } ZSTD_prefetch_e;
 
 FORCE_INLINE_TEMPLATE seq_t
-ZSTD_decodeSequence(seqState_t* seqState, const ZSTD_longOffset_e longOffsets, const ZSTD_prefetch_e prefetch)
+ZSTD_decodeSequence(seqState_t* seqState, const ZSTD_longOffset_e longOffsets)
 {
     seq_t seq;
-    ZSTD_seqSymbol const llDInfo = seqState->stateLL.table[seqState->stateLL.state];
-    ZSTD_seqSymbol const mlDInfo = seqState->stateML.table[seqState->stateML.state];
-    ZSTD_seqSymbol const ofDInfo = seqState->stateOffb.table[seqState->stateOffb.state];
-    U32 const llBase = llDInfo.baseValue;
-    U32 const mlBase = mlDInfo.baseValue;
-    U32 const ofBase = ofDInfo.baseValue;
-    BYTE const llBits = llDInfo.nbAdditionalBits;
-    BYTE const mlBits = mlDInfo.nbAdditionalBits;
-    BYTE const ofBits = ofDInfo.nbAdditionalBits;
-    BYTE const totalBits = llBits+mlBits+ofBits;
-
-    /* sequence */
-    {   size_t offset;
-        if (ofBits > 1) {
-            ZSTD_STATIC_ASSERT(ZSTD_lo_isLongOffset == 1);
-            ZSTD_STATIC_ASSERT(LONG_OFFSETS_MAX_EXTRA_BITS_32 == 5);
-            assert(ofBits <= MaxOff);
-            if (MEM_32bits() && longOffsets && (ofBits >= STREAM_ACCUMULATOR_MIN_32)) {
-                U32 const extraBits = ofBits - MIN(ofBits, 32 - seqState->DStream.bitsConsumed);
-                offset = ofBase + (BIT_readBitsFast(&seqState->DStream, ofBits - extraBits) << extraBits);
-                BIT_reloadDStream(&seqState->DStream);
-                if (extraBits) offset += BIT_readBitsFast(&seqState->DStream, extraBits);
-                assert(extraBits <= LONG_OFFSETS_MAX_EXTRA_BITS_32);   /* to avoid another reload */
-            } else {
-                offset = ofBase + BIT_readBitsFast(&seqState->DStream, ofBits/*>0*/);   /* <=  (ZSTD_WINDOWLOG_MAX-1) bits */
-                if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream);
-            }
-            seqState->prevOffset[2] = seqState->prevOffset[1];
-            seqState->prevOffset[1] = seqState->prevOffset[0];
-            seqState->prevOffset[0] = offset;
-        } else {
-            U32 const ll0 = (llBase == 0);
-            if (LIKELY((ofBits == 0))) {
-                if (LIKELY(!ll0))
-                    offset = seqState->prevOffset[0];
-                else {
-                    offset = seqState->prevOffset[1];
-                    seqState->prevOffset[1] = seqState->prevOffset[0];
-                    seqState->prevOffset[0] = offset;
-                }
-            } else {
-                offset = ofBase + ll0 + BIT_readBitsFast(&seqState->DStream, 1);
-                {   size_t temp = (offset==3) ? seqState->prevOffset[0] - 1 : seqState->prevOffset[offset];
-                    temp += !temp;   /* 0 is not valid; input is corrupted; force offset to 1 */
-                    if (offset != 1) seqState->prevOffset[2] = seqState->prevOffset[1];
-                    seqState->prevOffset[1] = seqState->prevOffset[0];
-                    seqState->prevOffset[0] = offset = temp;
-        }   }   }
-        seq.offset = offset;
-    }
-
-    seq.matchLength = mlBase;
-    if (mlBits > 0)
-        seq.matchLength += BIT_readBitsFast(&seqState->DStream, mlBits/*>0*/);
-
-    if (MEM_32bits() && (mlBits+llBits >= STREAM_ACCUMULATOR_MIN_32-LONG_OFFSETS_MAX_EXTRA_BITS_32))
-        BIT_reloadDStream(&seqState->DStream);
-    if (MEM_64bits() && UNLIKELY(totalBits >= STREAM_ACCUMULATOR_MIN_64-(LLFSELog+MLFSELog+OffFSELog)))
-        BIT_reloadDStream(&seqState->DStream);
-    /* Ensure there are enough bits to read the rest of data in 64-bit mode. */
-    ZSTD_STATIC_ASSERT(16+LLFSELog+MLFSELog+OffFSELog < STREAM_ACCUMULATOR_MIN_64);
-
-    seq.litLength = llBase;
-    if (llBits > 0)
-        seq.litLength += BIT_readBitsFast(&seqState->DStream, llBits/*>0*/);
-
-    if (MEM_32bits())
-        BIT_reloadDStream(&seqState->DStream);
-
-    DEBUGLOG(6, "seq: litL=%u, matchL=%u, offset=%u",
-                (U32)seq.litLength, (U32)seq.matchLength, (U32)seq.offset);
-
-    if (prefetch == ZSTD_p_prefetch) {
-        size_t const pos = seqState->pos + seq.litLength;
-        const BYTE* const matchBase = (seq.offset > pos) ? seqState->dictEnd : seqState->prefixStart;
-        seq.match = matchBase + pos - seq.offset;  /* note : this operation can overflow when seq.offset is really too large, which can only happen when input is corrupted.
-                                                    * No consequence though : no memory access will occur, offset is only used for prefetching */
-        seqState->pos = pos + seq.matchLength;
-    }
-
-    /* ANS state update
-     * gcc-9.0.0 does 2.5% worse with ZSTD_updateFseStateWithDInfo().
-     * clang-9.2.0 does 7% worse with ZSTD_updateFseState().
-     * Naturally it seems like ZSTD_updateFseStateWithDInfo() should be the
-     * better option, so it is the default for other compilers. But, if you
-     * measure that it is worse, please put up a pull request.
+    /*
+     * ZSTD_seqSymbol is a structure with a total of 64 bits wide. So it can be
+     * loaded in one operation and extracted its fields by simply shifting or
+     * bit-extracting on aarch64.
+     * GCC doesn't recognize this and generates more unnecessary ldr/ldrb/ldrh
+     * operations that cause performance drop. This can be avoided by using this
+     * ZSTD_memcpy hack.
      */
-    {
-#if defined(__GNUC__) && !defined(__clang__)
-        const int kUseUpdateFseState = 1;
+#if defined(__aarch64__) && (defined(__GNUC__) && !defined(__clang__))
+    ZSTD_seqSymbol llDInfoS, mlDInfoS, ofDInfoS;
+    ZSTD_seqSymbol* const llDInfo = &llDInfoS;
+    ZSTD_seqSymbol* const mlDInfo = &mlDInfoS;
+    ZSTD_seqSymbol* const ofDInfo = &ofDInfoS;
+    ZSTD_memcpy(llDInfo, seqState->stateLL.table + seqState->stateLL.state, sizeof(ZSTD_seqSymbol));
+    ZSTD_memcpy(mlDInfo, seqState->stateML.table + seqState->stateML.state, sizeof(ZSTD_seqSymbol));
+    ZSTD_memcpy(ofDInfo, seqState->stateOffb.table + seqState->stateOffb.state, sizeof(ZSTD_seqSymbol));
 #else
-        const int kUseUpdateFseState = 0;
+    const ZSTD_seqSymbol* const llDInfo = seqState->stateLL.table + seqState->stateLL.state;
+    const ZSTD_seqSymbol* const mlDInfo = seqState->stateML.table + seqState->stateML.state;
+    const ZSTD_seqSymbol* const ofDInfo = seqState->stateOffb.table + seqState->stateOffb.state;
 #endif
-        if (kUseUpdateFseState) {
-            ZSTD_updateFseState(&seqState->stateLL, &seqState->DStream);    /* <=  9 bits */
-            ZSTD_updateFseState(&seqState->stateML, &seqState->DStream);    /* <=  9 bits */
-            if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream);    /* <= 18 bits */
-            ZSTD_updateFseState(&seqState->stateOffb, &seqState->DStream);  /* <=  8 bits */
-        } else {
-            ZSTD_updateFseStateWithDInfo(&seqState->stateLL, &seqState->DStream, llDInfo);    /* <=  9 bits */
-            ZSTD_updateFseStateWithDInfo(&seqState->stateML, &seqState->DStream, mlDInfo);    /* <=  9 bits */
-            if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream);    /* <= 18 bits */
-            ZSTD_updateFseStateWithDInfo(&seqState->stateOffb, &seqState->DStream, ofDInfo);  /* <=  8 bits */
+    seq.matchLength = mlDInfo->baseValue;
+    seq.litLength = llDInfo->baseValue;
+    {   U32 const ofBase = ofDInfo->baseValue;
+        BYTE const llBits = llDInfo->nbAdditionalBits;
+        BYTE const mlBits = mlDInfo->nbAdditionalBits;
+        BYTE const ofBits = ofDInfo->nbAdditionalBits;
+        BYTE const totalBits = llBits+mlBits+ofBits;
+
+        U16 const llNext = llDInfo->nextState;
+        U16 const mlNext = mlDInfo->nextState;
+        U16 const ofNext = ofDInfo->nextState;
+        U32 const llnbBits = llDInfo->nbBits;
+        U32 const mlnbBits = mlDInfo->nbBits;
+        U32 const ofnbBits = ofDInfo->nbBits;
+
+        assert(llBits <= MaxLLBits);
+        assert(mlBits <= MaxMLBits);
+        assert(ofBits <= MaxOff);
+        /*
+         * As gcc has better branch and block analyzers, sometimes it is only
+         * valuable to mark likeliness for clang, it gives around 3-4% of
+         * performance.
+         */
+
+        /* sequence */
+        {   size_t offset;
+    #if defined(__clang__)
+            if (LIKELY(ofBits > 1)) {
+    #else
+            if (ofBits > 1) {
+    #endif
+                ZSTD_STATIC_ASSERT(ZSTD_lo_isLongOffset == 1);
+                ZSTD_STATIC_ASSERT(LONG_OFFSETS_MAX_EXTRA_BITS_32 == 5);
+                ZSTD_STATIC_ASSERT(STREAM_ACCUMULATOR_MIN_32 > LONG_OFFSETS_MAX_EXTRA_BITS_32);
+                ZSTD_STATIC_ASSERT(STREAM_ACCUMULATOR_MIN_32 - LONG_OFFSETS_MAX_EXTRA_BITS_32 >= MaxMLBits);
+                if (MEM_32bits() && longOffsets && (ofBits >= STREAM_ACCUMULATOR_MIN_32)) {
+                    /* Always read extra bits, this keeps the logic simple,
+                     * avoids branches, and avoids accidentally reading 0 bits.
+                     */
+                    U32 const extraBits = LONG_OFFSETS_MAX_EXTRA_BITS_32;
+                    offset = ofBase + (BIT_readBitsFast(&seqState->DStream, ofBits - extraBits) << extraBits);
+                    BIT_reloadDStream(&seqState->DStream);
+                    offset += BIT_readBitsFast(&seqState->DStream, extraBits);
+                } else {
+                    offset = ofBase + BIT_readBitsFast(&seqState->DStream, ofBits/*>0*/);   /* <=  (ZSTD_WINDOWLOG_MAX-1) bits */
+                    if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream);
+                }
+                seqState->prevOffset[2] = seqState->prevOffset[1];
+                seqState->prevOffset[1] = seqState->prevOffset[0];
+                seqState->prevOffset[0] = offset;
+            } else {
+                U32 const ll0 = (llDInfo->baseValue == 0);
+                if (LIKELY((ofBits == 0))) {
+                    offset = seqState->prevOffset[ll0];
+                    seqState->prevOffset[1] = seqState->prevOffset[!ll0];
+                    seqState->prevOffset[0] = offset;
+                } else {
+                    offset = ofBase + ll0 + BIT_readBitsFast(&seqState->DStream, 1);
+                    {   size_t temp = (offset==3) ? seqState->prevOffset[0] - 1 : seqState->prevOffset[offset];
+                        temp += !temp;   /* 0 is not valid; input is corrupted; force offset to 1 */
+                        if (offset != 1) seqState->prevOffset[2] = seqState->prevOffset[1];
+                        seqState->prevOffset[1] = seqState->prevOffset[0];
+                        seqState->prevOffset[0] = offset = temp;
+            }   }   }
+            seq.offset = offset;
         }
+
+    #if defined(__clang__)
+        if (UNLIKELY(mlBits > 0))
+    #else
+        if (mlBits > 0)
+    #endif
+            seq.matchLength += BIT_readBitsFast(&seqState->DStream, mlBits/*>0*/);
+
+        if (MEM_32bits() && (mlBits+llBits >= STREAM_ACCUMULATOR_MIN_32-LONG_OFFSETS_MAX_EXTRA_BITS_32))
+            BIT_reloadDStream(&seqState->DStream);
+        if (MEM_64bits() && UNLIKELY(totalBits >= STREAM_ACCUMULATOR_MIN_64-(LLFSELog+MLFSELog+OffFSELog)))
+            BIT_reloadDStream(&seqState->DStream);
+        /* Ensure there are enough bits to read the rest of data in 64-bit mode. */
+        ZSTD_STATIC_ASSERT(16+LLFSELog+MLFSELog+OffFSELog < STREAM_ACCUMULATOR_MIN_64);
+
+    #if defined(__clang__)
+        if (UNLIKELY(llBits > 0))
+    #else
+        if (llBits > 0)
+    #endif
+            seq.litLength += BIT_readBitsFast(&seqState->DStream, llBits/*>0*/);
+
+        if (MEM_32bits())
+            BIT_reloadDStream(&seqState->DStream);
+
+        DEBUGLOG(6, "seq: litL=%u, matchL=%u, offset=%u",
+                    (U32)seq.litLength, (U32)seq.matchLength, (U32)seq.offset);
+
+        ZSTD_updateFseStateWithDInfo(&seqState->stateLL, &seqState->DStream, llNext, llnbBits);    /* <=  9 bits */
+        ZSTD_updateFseStateWithDInfo(&seqState->stateML, &seqState->DStream, mlNext, mlnbBits);    /* <=  9 bits */
+        if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream);    /* <= 18 bits */
+        ZSTD_updateFseStateWithDInfo(&seqState->stateOffb, &seqState->DStream, ofNext, ofnbBits);  /* <=  8 bits */
     }
 
     return seq;
 }
 
 #ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
-static int ZSTD_dictionaryIsActive(ZSTD_DCtx const* dctx, BYTE const* prefixStart, BYTE const* oLitEnd)
+MEM_STATIC int ZSTD_dictionaryIsActive(ZSTD_DCtx const* dctx, BYTE const* prefixStart, BYTE const* oLitEnd)
 {
     size_t const windowSize = dctx->fParams.windowSize;
     /* No dictionary used. */
@@ -969,6 +1331,7 @@ MEM_STATIC void ZSTD_assertValidSequence(
         seq_t const seq,
         BYTE const* prefixStart, BYTE const* virtualStart)
 {
+#if DEBUGLEVEL >= 1
     size_t const windowSize = dctx->fParams.windowSize;
     size_t const sequenceSize = seq.litLength + seq.matchLength;
     BYTE const* const oLitEnd = op + seq.litLength;
@@ -986,13 +1349,18 @@ MEM_STATIC void ZSTD_assertValidSequence(
         /* Offset must be within our window. */
         assert(seq.offset <= windowSize);
     }
+#else
+    (void)dctx, (void)op, (void)oend, (void)seq, (void)prefixStart, (void)virtualStart;
+#endif
 }
 #endif
 
 #ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG
+
+
 FORCE_INLINE_TEMPLATE size_t
 DONT_VECTORIZE
-ZSTD_decompressSequences_body( ZSTD_DCtx* dctx,
+ZSTD_decompressSequences_bodySplitLitBuffer( ZSTD_DCtx* dctx,
                                void* dst, size_t maxDstSize,
                          const void* seqStart, size_t seqSize, int nbSeq,
                          const ZSTD_longOffset_e isLongOffset,
@@ -1000,21 +1368,20 @@ ZSTD_decompressSequences_body( ZSTD_DCtx* dctx,
 {
     const BYTE* ip = (const BYTE*)seqStart;
     const BYTE* const iend = ip + seqSize;
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     BYTE* const oend = ostart + maxDstSize;
     BYTE* op = ostart;
     const BYTE* litPtr = dctx->litPtr;
-    const BYTE* const litEnd = litPtr + dctx->litSize;
+    const BYTE* litBufferEnd = dctx->litBufferEnd;
     const BYTE* const prefixStart = (const BYTE*) (dctx->prefixStart);
     const BYTE* const vBase = (const BYTE*) (dctx->virtualStart);
     const BYTE* const dictEnd = (const BYTE*) (dctx->dictEnd);
-    DEBUGLOG(5, "ZSTD_decompressSequences_body");
+    DEBUGLOG(5, "ZSTD_decompressSequences_bodySplitLitBuffer");
     (void)frame;
 
     /* Regen sequences */
     if (nbSeq) {
         seqState_t seqState;
-        size_t error = 0;
         dctx->fseEntropy = 1;
         { U32 i; for (i=0; i<ZSTD_REP_NUM; i++) seqState.prevOffset[i] = dctx->entropy.rep[i]; }
         RETURN_ERROR_IF(
@@ -1030,70 +1397,255 @@ ZSTD_decompressSequences_body( ZSTD_DCtx* dctx,
                 BIT_DStream_endOfBuffer < BIT_DStream_completed &&
                 BIT_DStream_completed < BIT_DStream_overflow);
 
+        /* decompress without overrunning litPtr begins */
+        {
+            seq_t sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
+            /* Align the decompression loop to 32 + 16 bytes.
+                *
+                * zstd compiled with gcc-9 on an Intel i9-9900k shows 10% decompression
+                * speed swings based on the alignment of the decompression loop. This
+                * performance swing is caused by parts of the decompression loop falling
+                * out of the DSB. The entire decompression loop should fit in the DSB,
+                * when it can't we get much worse performance. You can measure if you've
+                * hit the good case or the bad case with this perf command for some
+                * compressed file test.zst:
+                *
+                *   perf stat -e cycles -e instructions -e idq.all_dsb_cycles_any_uops \
+                *             -e idq.all_mite_cycles_any_uops -- ./zstd -tq test.zst
+                *
+                * If you see most cycles served out of the MITE you've hit the bad case.
+                * If you see most cycles served out of the DSB you've hit the good case.
+                * If it is pretty even then you may be in an okay case.
+                *
+                * This issue has been reproduced on the following CPUs:
+                *   - Kabylake: Macbook Pro (15-inch, 2019) 2.4 GHz Intel Core i9
+                *               Use Instruments->Counters to get DSB/MITE cycles.
+                *               I never got performance swings, but I was able to
+                *               go from the good case of mostly DSB to half of the
+                *               cycles served from MITE.
+                *   - Coffeelake: Intel i9-9900k
+                *   - Coffeelake: Intel i7-9700k
+                *
+                * I haven't been able to reproduce the instability or DSB misses on any
+                * of the following CPUS:
+                *   - Haswell
+                *   - Broadwell: Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GH
+                *   - Skylake
+                *
+                * Alignment is done for each of the three major decompression loops:
+                *   - ZSTD_decompressSequences_bodySplitLitBuffer - presplit section of the literal buffer
+                *   - ZSTD_decompressSequences_bodySplitLitBuffer - postsplit section of the literal buffer
+                *   - ZSTD_decompressSequences_body
+                * Alignment choices are made to minimize large swings on bad cases and influence on performance
+                * from changes external to this code, rather than to overoptimize on the current commit.
+                *
+                * If you are seeing performance stability this script can help test.
+                * It tests on 4 commits in zstd where I saw performance change.
+                *
+                *   https://gist.github.com/terrelln/9889fc06a423fd5ca6e99351564473f4
+                */
 #if defined(__GNUC__) && defined(__x86_64__)
-        /* Align the decompression loop to 32 + 16 bytes.
-         *
-         * zstd compiled with gcc-9 on an Intel i9-9900k shows 10% decompression
-         * speed swings based on the alignment of the decompression loop. This
-         * performance swing is caused by parts of the decompression loop falling
-         * out of the DSB. The entire decompression loop should fit in the DSB,
-         * when it can't we get much worse performance. You can measure if you've
-         * hit the good case or the bad case with this perf command for some
-         * compressed file test.zst:
-         *
-         *   perf stat -e cycles -e instructions -e idq.all_dsb_cycles_any_uops \
-         *             -e idq.all_mite_cycles_any_uops -- ./zstd -tq test.zst
-         *
-         * If you see most cycles served out of the MITE you've hit the bad case.
-         * If you see most cycles served out of the DSB you've hit the good case.
-         * If it is pretty even then you may be in an okay case.
-         *
-         * I've been able to reproduce this issue on the following CPUs:
-         *   - Kabylake: Macbook Pro (15-inch, 2019) 2.4 GHz Intel Core i9
-         *               Use Instruments->Counters to get DSB/MITE cycles.
-         *               I never got performance swings, but I was able to
-         *               go from the good case of mostly DSB to half of the
-         *               cycles served from MITE.
-         *   - Coffeelake: Intel i9-9900k
-         *
-         * I haven't been able to reproduce the instability or DSB misses on any
-         * of the following CPUS:
-         *   - Haswell
-         *   - Broadwell: Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GH
-         *   - Skylake
-         *
-         * If you are seeing performance stability this script can help test.
-         * It tests on 4 commits in zstd where I saw performance change.
-         *
-         *   https://gist.github.com/terrelln/9889fc06a423fd5ca6e99351564473f4
-         */
-        __asm__(".p2align 5");
-        __asm__("nop");
-        __asm__(".p2align 4");
+            __asm__(".p2align 6");
+#  if __GNUC__ >= 7
+	    /* good for gcc-7, gcc-9, and gcc-11 */
+            __asm__("nop");
+            __asm__(".p2align 5");
+            __asm__("nop");
+            __asm__(".p2align 4");
+#    if __GNUC__ == 8 || __GNUC__ == 10
+	    /* good for gcc-8 and gcc-10 */
+            __asm__("nop");
+            __asm__(".p2align 3");
+#    endif
+#  endif
+#endif
+
+            /* Handle the initial state where litBuffer is currently split between dst and litExtraBuffer */
+            for (; litPtr + sequence.litLength <= dctx->litBufferEnd; ) {
+                size_t const oneSeqSize = ZSTD_execSequenceSplitLitBuffer(op, oend, litPtr + sequence.litLength - WILDCOPY_OVERLENGTH, sequence, &litPtr, litBufferEnd, prefixStart, vBase, dictEnd);
+#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
+                assert(!ZSTD_isError(oneSeqSize));
+                if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
+#endif
+                if (UNLIKELY(ZSTD_isError(oneSeqSize)))
+                    return oneSeqSize;
+                DEBUGLOG(6, "regenerated sequence size : %u", (U32)oneSeqSize);
+                op += oneSeqSize;
+                if (UNLIKELY(!--nbSeq))
+                    break;
+                BIT_reloadDStream(&(seqState.DStream));
+                sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
+            }
+
+            /* If there are more sequences, they will need to read literals from litExtraBuffer; copy over the remainder from dst and update litPtr and litEnd */
+            if (nbSeq > 0) {
+                const size_t leftoverLit = dctx->litBufferEnd - litPtr;
+                if (leftoverLit)
+                {
+                    RETURN_ERROR_IF(leftoverLit > (size_t)(oend - op), dstSize_tooSmall, "remaining lit must fit within dstBuffer");
+                    ZSTD_safecopyDstBeforeSrc(op, litPtr, leftoverLit);
+                    sequence.litLength -= leftoverLit;
+                    op += leftoverLit;
+                }
+                litPtr = dctx->litExtraBuffer;
+                litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
+                dctx->litBufferLocation = ZSTD_not_in_dst;
+                {
+                    size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequence, &litPtr, litBufferEnd, prefixStart, vBase, dictEnd);
+#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
+                    assert(!ZSTD_isError(oneSeqSize));
+                    if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
+#endif
+                    if (UNLIKELY(ZSTD_isError(oneSeqSize)))
+                        return oneSeqSize;
+                    DEBUGLOG(6, "regenerated sequence size : %u", (U32)oneSeqSize);
+                    op += oneSeqSize;
+                    if (--nbSeq)
+                        BIT_reloadDStream(&(seqState.DStream));
+                }
+            }
+        }
+
+        if (nbSeq > 0) /* there is remaining lit from extra buffer */
+        {
+
+#if defined(__GNUC__) && defined(__x86_64__)
+            __asm__(".p2align 6");
+            __asm__("nop");
+#  if __GNUC__ != 7
+            /* worse for gcc-7 better for gcc-8, gcc-9, and gcc-10 and clang */
+            __asm__(".p2align 4");
+            __asm__("nop");
+            __asm__(".p2align 3");
+#  elif __GNUC__ >= 11
+            __asm__(".p2align 3");
+#  else
+            __asm__(".p2align 5");
+            __asm__("nop");
+            __asm__(".p2align 3");
+#  endif
+#endif
+
+            for (; ; ) {
+                seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
+                size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequence, &litPtr, litBufferEnd, prefixStart, vBase, dictEnd);
+#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
+                assert(!ZSTD_isError(oneSeqSize));
+                if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
+#endif
+                if (UNLIKELY(ZSTD_isError(oneSeqSize)))
+                    return oneSeqSize;
+                DEBUGLOG(6, "regenerated sequence size : %u", (U32)oneSeqSize);
+                op += oneSeqSize;
+                if (UNLIKELY(!--nbSeq))
+                    break;
+                BIT_reloadDStream(&(seqState.DStream));
+            }
+        }
+
+        /* check if reached exact end */
+        DEBUGLOG(5, "ZSTD_decompressSequences_bodySplitLitBuffer: after decode loop, remaining nbSeq : %i", nbSeq);
+        RETURN_ERROR_IF(nbSeq, corruption_detected, "");
+        RETURN_ERROR_IF(BIT_reloadDStream(&seqState.DStream) < BIT_DStream_completed, corruption_detected, "");
+        /* save reps for next block */
+        { U32 i; for (i=0; i<ZSTD_REP_NUM; i++) dctx->entropy.rep[i] = (U32)(seqState.prevOffset[i]); }
+    }
+
+    /* last literal segment */
+    if (dctx->litBufferLocation == ZSTD_split)  /* split hasn't been reached yet, first get dst then copy litExtraBuffer */
+    {
+        size_t const lastLLSize = litBufferEnd - litPtr;
+        RETURN_ERROR_IF(lastLLSize > (size_t)(oend - op), dstSize_tooSmall, "");
+        if (op != NULL) {
+            ZSTD_memmove(op, litPtr, lastLLSize);
+            op += lastLLSize;
+        }
+        litPtr = dctx->litExtraBuffer;
+        litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
+        dctx->litBufferLocation = ZSTD_not_in_dst;
+    }
+    {   size_t const lastLLSize = litBufferEnd - litPtr;
+        RETURN_ERROR_IF(lastLLSize > (size_t)(oend-op), dstSize_tooSmall, "");
+        if (op != NULL) {
+            ZSTD_memcpy(op, litPtr, lastLLSize);
+            op += lastLLSize;
+        }
+    }
+
+    return op-ostart;
+}
+
+FORCE_INLINE_TEMPLATE size_t
+DONT_VECTORIZE
+ZSTD_decompressSequences_body(ZSTD_DCtx* dctx,
+    void* dst, size_t maxDstSize,
+    const void* seqStart, size_t seqSize, int nbSeq,
+    const ZSTD_longOffset_e isLongOffset,
+    const int frame)
+{
+    const BYTE* ip = (const BYTE*)seqStart;
+    const BYTE* const iend = ip + seqSize;
+    BYTE* const ostart = (BYTE*)dst;
+    BYTE* const oend = dctx->litBufferLocation == ZSTD_not_in_dst ? ostart + maxDstSize : dctx->litBuffer;
+    BYTE* op = ostart;
+    const BYTE* litPtr = dctx->litPtr;
+    const BYTE* const litEnd = litPtr + dctx->litSize;
+    const BYTE* const prefixStart = (const BYTE*)(dctx->prefixStart);
+    const BYTE* const vBase = (const BYTE*)(dctx->virtualStart);
+    const BYTE* const dictEnd = (const BYTE*)(dctx->dictEnd);
+    DEBUGLOG(5, "ZSTD_decompressSequences_body: nbSeq = %d", nbSeq);
+    (void)frame;
+
+    /* Regen sequences */
+    if (nbSeq) {
+        seqState_t seqState;
+        dctx->fseEntropy = 1;
+        { U32 i; for (i = 0; i < ZSTD_REP_NUM; i++) seqState.prevOffset[i] = dctx->entropy.rep[i]; }
+        RETURN_ERROR_IF(
+            ERR_isError(BIT_initDStream(&seqState.DStream, ip, iend - ip)),
+            corruption_detected, "");
+        ZSTD_initFseState(&seqState.stateLL, &seqState.DStream, dctx->LLTptr);
+        ZSTD_initFseState(&seqState.stateOffb, &seqState.DStream, dctx->OFTptr);
+        ZSTD_initFseState(&seqState.stateML, &seqState.DStream, dctx->MLTptr);
+        assert(dst != NULL);
+
+        ZSTD_STATIC_ASSERT(
+            BIT_DStream_unfinished < BIT_DStream_completed &&
+            BIT_DStream_endOfBuffer < BIT_DStream_completed &&
+            BIT_DStream_completed < BIT_DStream_overflow);
+
+#if defined(__GNUC__) && defined(__x86_64__)
+            __asm__(".p2align 6");
+            __asm__("nop");
+#  if __GNUC__ >= 7
+            __asm__(".p2align 5");
+            __asm__("nop");
+            __asm__(".p2align 3");
+#  else
+            __asm__(".p2align 4");
+            __asm__("nop");
+            __asm__(".p2align 3");
+#  endif
 #endif
+
         for ( ; ; ) {
-            seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset, ZSTD_p_noPrefetch);
+            seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
             size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequence, &litPtr, litEnd, prefixStart, vBase, dictEnd);
 #if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
             assert(!ZSTD_isError(oneSeqSize));
             if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
 #endif
+            if (UNLIKELY(ZSTD_isError(oneSeqSize)))
+                return oneSeqSize;
             DEBUGLOG(6, "regenerated sequence size : %u", (U32)oneSeqSize);
+            op += oneSeqSize;
+            if (UNLIKELY(!--nbSeq))
+                break;
             BIT_reloadDStream(&(seqState.DStream));
-            /* gcc and clang both don't like early returns in this loop.
-             * gcc doesn't like early breaks either.
-             * Instead save an error and report it at the end.
-             * When there is an error, don't increment op, so we don't
-             * overwrite.
-             */
-            if (UNLIKELY(ZSTD_isError(oneSeqSize))) error = oneSeqSize;
-            else op += oneSeqSize;
-            if (UNLIKELY(!--nbSeq)) break;
         }
 
         /* check if reached exact end */
         DEBUGLOG(5, "ZSTD_decompressSequences_body: after decode loop, remaining nbSeq : %i", nbSeq);
-        if (ZSTD_isError(error)) return error;
         RETURN_ERROR_IF(nbSeq, corruption_detected, "");
         RETURN_ERROR_IF(BIT_reloadDStream(&seqState.DStream) < BIT_DStream_completed, corruption_detected, "");
         /* save reps for next block */
@@ -1104,7 +1656,7 @@ ZSTD_decompressSequences_body( ZSTD_DCtx* dctx,
     {   size_t const lastLLSize = litEnd - litPtr;
         RETURN_ERROR_IF(lastLLSize > (size_t)(oend-op), dstSize_tooSmall, "");
         if (op != NULL) {
-            memcpy(op, litPtr, lastLLSize);
+            ZSTD_memcpy(op, litPtr, lastLLSize);
             op += lastLLSize;
         }
     }
@@ -1121,9 +1673,37 @@ ZSTD_decompressSequences_default(ZSTD_DCtx* dctx,
 {
     return ZSTD_decompressSequences_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
 }
+
+static size_t
+ZSTD_decompressSequencesSplitLitBuffer_default(ZSTD_DCtx* dctx,
+                                               void* dst, size_t maxDstSize,
+                                         const void* seqStart, size_t seqSize, int nbSeq,
+                                         const ZSTD_longOffset_e isLongOffset,
+                                         const int frame)
+{
+    return ZSTD_decompressSequences_bodySplitLitBuffer(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+}
 #endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG */
 
 #ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT
+
+FORCE_INLINE_TEMPLATE size_t
+ZSTD_prefetchMatch(size_t prefetchPos, seq_t const sequence,
+                   const BYTE* const prefixStart, const BYTE* const dictEnd)
+{
+    prefetchPos += sequence.litLength;
+    {   const BYTE* const matchBase = (sequence.offset > prefetchPos) ? dictEnd : prefixStart;
+        const BYTE* const match = matchBase + prefetchPos - sequence.offset; /* note : this operation can overflow when seq.offset is really too large, which can only happen when input is corrupted.
+                                                                              * No consequence though : memory address is only used for prefetching, not for dereferencing */
+        PREFETCH_L1(match); PREFETCH_L1(match+CACHELINE_SIZE);   /* note : it's safe to invoke PREFETCH() on any memory address, including invalid ones */
+    }
+    return prefetchPos + sequence.matchLength;
+}
+
+/* This decoding function employs prefetching
+ * to reduce latency impact of cache misses.
+ * It's generally employed when block contains a significant portion of long-distance matches
+ * or when coupled with a "cold" dictionary */
 FORCE_INLINE_TEMPLATE size_t
 ZSTD_decompressSequencesLong_body(
                                ZSTD_DCtx* dctx,
@@ -1134,11 +1714,11 @@ ZSTD_decompressSequencesLong_body(
 {
     const BYTE* ip = (const BYTE*)seqStart;
     const BYTE* const iend = ip + seqSize;
-    BYTE* const ostart = (BYTE* const)dst;
-    BYTE* const oend = ostart + maxDstSize;
+    BYTE* const ostart = (BYTE*)dst;
+    BYTE* const oend = dctx->litBufferLocation == ZSTD_in_dst ? dctx->litBuffer : ostart + maxDstSize;
     BYTE* op = ostart;
     const BYTE* litPtr = dctx->litPtr;
-    const BYTE* const litEnd = litPtr + dctx->litSize;
+    const BYTE* litBufferEnd = dctx->litBufferEnd;
     const BYTE* const prefixStart = (const BYTE*) (dctx->prefixStart);
     const BYTE* const dictStart = (const BYTE*) (dctx->virtualStart);
     const BYTE* const dictEnd = (const BYTE*) (dctx->dictEnd);
@@ -1146,18 +1726,17 @@ ZSTD_decompressSequencesLong_body(
 
     /* Regen sequences */
     if (nbSeq) {
-#define STORED_SEQS 4
+#define STORED_SEQS 8
 #define STORED_SEQS_MASK (STORED_SEQS-1)
-#define ADVANCED_SEQS 4
+#define ADVANCED_SEQS STORED_SEQS
         seq_t sequences[STORED_SEQS];
         int const seqAdvance = MIN(nbSeq, ADVANCED_SEQS);
         seqState_t seqState;
         int seqNb;
+        size_t prefetchPos = (size_t)(op-prefixStart); /* track position relative to prefixStart */
+
         dctx->fseEntropy = 1;
         { int i; for (i=0; i<ZSTD_REP_NUM; i++) seqState.prevOffset[i] = dctx->entropy.rep[i]; }
-        seqState.prefixStart = prefixStart;
-        seqState.pos = (size_t)(op-prefixStart);
-        seqState.dictEnd = dictEnd;
         assert(dst != NULL);
         assert(iend >= ip);
         RETURN_ERROR_IF(
@@ -1169,36 +1748,100 @@ ZSTD_decompressSequencesLong_body(
 
         /* prepare in advance */
         for (seqNb=0; (BIT_reloadDStream(&seqState.DStream) <= BIT_DStream_completed) && (seqNb<seqAdvance); seqNb++) {
-            sequences[seqNb] = ZSTD_decodeSequence(&seqState, isLongOffset, ZSTD_p_prefetch);
-            PREFETCH_L1(sequences[seqNb].match); PREFETCH_L1(sequences[seqNb].match + sequences[seqNb].matchLength - 1); /* note : it's safe to invoke PREFETCH() on any memory address, including invalid ones */
+            seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
+            prefetchPos = ZSTD_prefetchMatch(prefetchPos, sequence, prefixStart, dictEnd);
+            sequences[seqNb] = sequence;
         }
         RETURN_ERROR_IF(seqNb<seqAdvance, corruption_detected, "");
 
-        /* decode and decompress */
-        for ( ; (BIT_reloadDStream(&(seqState.DStream)) <= BIT_DStream_completed) && (seqNb<nbSeq) ; seqNb++) {
-            seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset, ZSTD_p_prefetch);
-            size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequences[(seqNb-ADVANCED_SEQS) & STORED_SEQS_MASK], &litPtr, litEnd, prefixStart, dictStart, dictEnd);
+        /* decompress without stomping litBuffer */
+        for (; (BIT_reloadDStream(&(seqState.DStream)) <= BIT_DStream_completed) && (seqNb < nbSeq); seqNb++) {
+            seq_t sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
+            size_t oneSeqSize;
+
+            if (dctx->litBufferLocation == ZSTD_split && litPtr + sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK].litLength > dctx->litBufferEnd)
+            {
+                /* lit buffer is reaching split point, empty out the first buffer and transition to litExtraBuffer */
+                const size_t leftoverLit = dctx->litBufferEnd - litPtr;
+                if (leftoverLit)
+                {
+                    RETURN_ERROR_IF(leftoverLit > (size_t)(oend - op), dstSize_tooSmall, "remaining lit must fit within dstBuffer");
+                    ZSTD_safecopyDstBeforeSrc(op, litPtr, leftoverLit);
+                    sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK].litLength -= leftoverLit;
+                    op += leftoverLit;
+                }
+                litPtr = dctx->litExtraBuffer;
+                litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
+                dctx->litBufferLocation = ZSTD_not_in_dst;
+                oneSeqSize = ZSTD_execSequence(op, oend, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd);
 #if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
-            assert(!ZSTD_isError(oneSeqSize));
-            if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequences[(seqNb-ADVANCED_SEQS) & STORED_SEQS_MASK], prefixStart, dictStart);
+                assert(!ZSTD_isError(oneSeqSize));
+                if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], prefixStart, dictStart);
 #endif
-            if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
-            PREFETCH_L1(sequence.match); PREFETCH_L1(sequence.match + sequence.matchLength - 1); /* note : it's safe to invoke PREFETCH() on any memory address, including invalid ones */
-            sequences[seqNb & STORED_SEQS_MASK] = sequence;
-            op += oneSeqSize;
+                if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
+
+                prefetchPos = ZSTD_prefetchMatch(prefetchPos, sequence, prefixStart, dictEnd);
+                sequences[seqNb & STORED_SEQS_MASK] = sequence;
+                op += oneSeqSize;
+            }
+            else
+            {
+                /* lit buffer is either wholly contained in first or second split, or not split at all*/
+                oneSeqSize = dctx->litBufferLocation == ZSTD_split ?
+                    ZSTD_execSequenceSplitLitBuffer(op, oend, litPtr + sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK].litLength - WILDCOPY_OVERLENGTH, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd) :
+                    ZSTD_execSequence(op, oend, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd);
+#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
+                assert(!ZSTD_isError(oneSeqSize));
+                if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], prefixStart, dictStart);
+#endif
+                if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
+
+                prefetchPos = ZSTD_prefetchMatch(prefetchPos, sequence, prefixStart, dictEnd);
+                sequences[seqNb & STORED_SEQS_MASK] = sequence;
+                op += oneSeqSize;
+            }
         }
         RETURN_ERROR_IF(seqNb<nbSeq, corruption_detected, "");
 
         /* finish queue */
         seqNb -= seqAdvance;
         for ( ; seqNb<nbSeq ; seqNb++) {
-            size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequences[seqNb&STORED_SEQS_MASK], &litPtr, litEnd, prefixStart, dictStart, dictEnd);
+            seq_t *sequence = &(sequences[seqNb&STORED_SEQS_MASK]);
+            if (dctx->litBufferLocation == ZSTD_split && litPtr + sequence->litLength > dctx->litBufferEnd)
+            {
+                const size_t leftoverLit = dctx->litBufferEnd - litPtr;
+                if (leftoverLit)
+                {
+                    RETURN_ERROR_IF(leftoverLit > (size_t)(oend - op), dstSize_tooSmall, "remaining lit must fit within dstBuffer");
+                    ZSTD_safecopyDstBeforeSrc(op, litPtr, leftoverLit);
+                    sequence->litLength -= leftoverLit;
+                    op += leftoverLit;
+                }
+                litPtr = dctx->litExtraBuffer;
+                litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
+                dctx->litBufferLocation = ZSTD_not_in_dst;
+                {
+                    size_t const oneSeqSize = ZSTD_execSequence(op, oend, *sequence, &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd);
 #if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
-            assert(!ZSTD_isError(oneSeqSize));
-            if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequences[seqNb&STORED_SEQS_MASK], prefixStart, dictStart);
+                    assert(!ZSTD_isError(oneSeqSize));
+                    if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequences[seqNb&STORED_SEQS_MASK], prefixStart, dictStart);
 #endif
-            if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
-            op += oneSeqSize;
+                    if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
+                    op += oneSeqSize;
+                }
+            }
+            else
+            {
+                size_t const oneSeqSize = dctx->litBufferLocation == ZSTD_split ?
+                    ZSTD_execSequenceSplitLitBuffer(op, oend, litPtr + sequence->litLength - WILDCOPY_OVERLENGTH, *sequence, &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd) :
+                    ZSTD_execSequence(op, oend, *sequence, &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd);
+#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
+                assert(!ZSTD_isError(oneSeqSize));
+                if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequences[seqNb&STORED_SEQS_MASK], prefixStart, dictStart);
+#endif
+                if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
+                op += oneSeqSize;
+            }
         }
 
         /* save reps for next block */
@@ -1206,10 +1849,21 @@ ZSTD_decompressSequencesLong_body(
     }
 
     /* last literal segment */
-    {   size_t const lastLLSize = litEnd - litPtr;
+    if (dctx->litBufferLocation == ZSTD_split)  /* first deplete literal buffer in dst, then copy litExtraBuffer */
+    {
+        size_t const lastLLSize = litBufferEnd - litPtr;
+        RETURN_ERROR_IF(lastLLSize > (size_t)(oend - op), dstSize_tooSmall, "");
+        if (op != NULL) {
+            ZSTD_memmove(op, litPtr, lastLLSize);
+            op += lastLLSize;
+        }
+        litPtr = dctx->litExtraBuffer;
+        litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
+    }
+    {   size_t const lastLLSize = litBufferEnd - litPtr;
         RETURN_ERROR_IF(lastLLSize > (size_t)(oend-op), dstSize_tooSmall, "");
         if (op != NULL) {
-            memcpy(op, litPtr, lastLLSize);
+            ZSTD_memmove(op, litPtr, lastLLSize);
             op += lastLLSize;
         }
     }
@@ -1233,7 +1887,7 @@ ZSTD_decompressSequencesLong_default(ZSTD_DCtx* dctx,
 #if DYNAMIC_BMI2
 
 #ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG
-static TARGET_ATTRIBUTE("bmi2") size_t
+static BMI2_TARGET_ATTRIBUTE size_t
 DONT_VECTORIZE
 ZSTD_decompressSequences_bmi2(ZSTD_DCtx* dctx,
                                  void* dst, size_t maxDstSize,
@@ -1243,10 +1897,20 @@ ZSTD_decompressSequences_bmi2(ZSTD_DCtx* dctx,
 {
     return ZSTD_decompressSequences_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
 }
+static BMI2_TARGET_ATTRIBUTE size_t
+DONT_VECTORIZE
+ZSTD_decompressSequencesSplitLitBuffer_bmi2(ZSTD_DCtx* dctx,
+                                 void* dst, size_t maxDstSize,
+                           const void* seqStart, size_t seqSize, int nbSeq,
+                           const ZSTD_longOffset_e isLongOffset,
+                           const int frame)
+{
+    return ZSTD_decompressSequences_bodySplitLitBuffer(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+}
 #endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG */
 
 #ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT
-static TARGET_ATTRIBUTE("bmi2") size_t
+static BMI2_TARGET_ATTRIBUTE size_t
 ZSTD_decompressSequencesLong_bmi2(ZSTD_DCtx* dctx,
                                  void* dst, size_t maxDstSize,
                            const void* seqStart, size_t seqSize, int nbSeq,
@@ -1275,11 +1939,25 @@ ZSTD_decompressSequences(ZSTD_DCtx* dctx, void* dst, size_t maxDstSize,
 {
     DEBUGLOG(5, "ZSTD_decompressSequences");
 #if DYNAMIC_BMI2
-    if (dctx->bmi2) {
+    if (ZSTD_DCtx_get_bmi2(dctx)) {
         return ZSTD_decompressSequences_bmi2(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
     }
 #endif
-  return ZSTD_decompressSequences_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+    return ZSTD_decompressSequences_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+}
+static size_t
+ZSTD_decompressSequencesSplitLitBuffer(ZSTD_DCtx* dctx, void* dst, size_t maxDstSize,
+                                 const void* seqStart, size_t seqSize, int nbSeq,
+                                 const ZSTD_longOffset_e isLongOffset,
+                                 const int frame)
+{
+    DEBUGLOG(5, "ZSTD_decompressSequencesSplitLitBuffer");
+#if DYNAMIC_BMI2
+    if (ZSTD_DCtx_get_bmi2(dctx)) {
+        return ZSTD_decompressSequencesSplitLitBuffer_bmi2(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
+    }
+#endif
+    return ZSTD_decompressSequencesSplitLitBuffer_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
 }
 #endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG */
 
@@ -1299,7 +1977,7 @@ ZSTD_decompressSequencesLong(ZSTD_DCtx* dctx,
 {
     DEBUGLOG(5, "ZSTD_decompressSequencesLong");
 #if DYNAMIC_BMI2
-    if (dctx->bmi2) {
+    if (ZSTD_DCtx_get_bmi2(dctx)) {
         return ZSTD_decompressSequencesLong_bmi2(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
     }
 #endif
@@ -1308,55 +1986,101 @@ ZSTD_decompressSequencesLong(ZSTD_DCtx* dctx,
 #endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT */
 
 
+/**
+ * @returns The total size of the history referencable by zstd, including
+ * both the prefix and the extDict. At @p op any offset larger than this
+ * is invalid.
+ */
+static size_t ZSTD_totalHistorySize(BYTE* op, BYTE const* virtualStart)
+{
+    return (size_t)(op - virtualStart);
+}
 
-#if !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \
-    !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG)
-/* ZSTD_getLongOffsetsShare() :
+typedef struct {
+    unsigned longOffsetShare;
+    unsigned maxNbAdditionalBits;
+} ZSTD_OffsetInfo;
+
+/* ZSTD_getOffsetInfo() :
  * condition : offTable must be valid
  * @return : "share" of long offsets (arbitrarily defined as > (1<<23))
- *           compared to maximum possible of (1<<OffFSELog) */
-static unsigned
-ZSTD_getLongOffsetsShare(const ZSTD_seqSymbol* offTable)
+ *           compared to maximum possible of (1<<OffFSELog),
+ *           as well as the maximum number additional bits required.
+ */
+static ZSTD_OffsetInfo
+ZSTD_getOffsetInfo(const ZSTD_seqSymbol* offTable, int nbSeq)
 {
-    const void* ptr = offTable;
-    U32 const tableLog = ((const ZSTD_seqSymbol_header*)ptr)[0].tableLog;
-    const ZSTD_seqSymbol* table = offTable + 1;
-    U32 const max = 1 << tableLog;
-    U32 u, total = 0;
-    DEBUGLOG(5, "ZSTD_getLongOffsetsShare: (tableLog=%u)", tableLog);
-
-    assert(max <= (1 << OffFSELog));  /* max not too large */
-    for (u=0; u<max; u++) {
-        if (table[u].nbAdditionalBits > 22) total += 1;
+    ZSTD_OffsetInfo info = {0, 0};
+    /* If nbSeq == 0, then the offTable is uninitialized, but we have
+     * no sequences, so both values should be 0.
+     */
+    if (nbSeq != 0) {
+        const void* ptr = offTable;
+        U32 const tableLog = ((const ZSTD_seqSymbol_header*)ptr)[0].tableLog;
+        const ZSTD_seqSymbol* table = offTable + 1;
+        U32 const max = 1 << tableLog;
+        U32 u;
+        DEBUGLOG(5, "ZSTD_getLongOffsetsShare: (tableLog=%u)", tableLog);
+
+        assert(max <= (1 << OffFSELog));  /* max not too large */
+        for (u=0; u<max; u++) {
+            info.maxNbAdditionalBits = MAX(info.maxNbAdditionalBits, table[u].nbAdditionalBits);
+            if (table[u].nbAdditionalBits > 22) info.longOffsetShare += 1;
+        }
+
+        assert(tableLog <= OffFSELog);
+        info.longOffsetShare <<= (OffFSELog - tableLog);  /* scale to OffFSELog */
     }
 
-    assert(tableLog <= OffFSELog);
-    total <<= (OffFSELog - tableLog);  /* scale to OffFSELog */
+    return info;
+}
 
-    return total;
+/**
+ * @returns The maximum offset we can decode in one read of our bitstream, without
+ * reloading more bits in the middle of the offset bits read. Any offsets larger
+ * than this must use the long offset decoder.
+ */
+static size_t ZSTD_maxShortOffset(void)
+{
+    if (MEM_64bits()) {
+        /* We can decode any offset without reloading bits.
+         * This might change if the max window size grows.
+         */
+        ZSTD_STATIC_ASSERT(ZSTD_WINDOWLOG_MAX <= 31);
+        return (size_t)-1;
+    } else {
+        /* The maximum offBase is (1 << (STREAM_ACCUMULATOR_MIN + 1)) - 1.
+         * This offBase would require STREAM_ACCUMULATOR_MIN extra bits.
+         * Then we have to subtract ZSTD_REP_NUM to get the maximum possible offset.
+         */
+        size_t const maxOffbase = ((size_t)1 << (STREAM_ACCUMULATOR_MIN + 1)) - 1;
+        size_t const maxOffset = maxOffbase - ZSTD_REP_NUM;
+        assert(ZSTD_highbit32((U32)maxOffbase) == STREAM_ACCUMULATOR_MIN);
+        return maxOffset;
+    }
 }
-#endif
 
 size_t
 ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
                               void* dst, size_t dstCapacity,
-                        const void* src, size_t srcSize, const int frame)
+                        const void* src, size_t srcSize, const int frame, const streaming_operation streaming)
 {   /* blockType == blockCompressed */
     const BYTE* ip = (const BYTE*)src;
-    /* isLongOffset must be true if there are long offsets.
-     * Offsets are long if they are larger than 2^STREAM_ACCUMULATOR_MIN.
-     * We don't expect that to be the case in 64-bit mode.
-     * In block mode, window size is not known, so we have to be conservative.
-     * (note: but it could be evaluated from current-lowLimit)
-     */
-    ZSTD_longOffset_e const isLongOffset = (ZSTD_longOffset_e)(MEM_32bits() && (!frame || (dctx->fParams.windowSize > (1ULL << STREAM_ACCUMULATOR_MIN))));
     DEBUGLOG(5, "ZSTD_decompressBlock_internal (size : %u)", (U32)srcSize);
 
-    RETURN_ERROR_IF(srcSize >= ZSTD_BLOCKSIZE_MAX, srcSize_wrong, "");
+    /* Note : the wording of the specification
+     * allows compressed block to be sized exactly ZSTD_BLOCKSIZE_MAX.
+     * This generally does not happen, as it makes little sense,
+     * since an uncompressed block would feature same size and have no decompression cost.
+     * Also, note that decoder from reference libzstd before < v1.5.4
+     * would consider this edge case as an error.
+     * As a consequence, avoid generating compressed blocks of size ZSTD_BLOCKSIZE_MAX
+     * for broader compatibility with the deployed ecosystem of zstd decoders */
+    RETURN_ERROR_IF(srcSize > ZSTD_BLOCKSIZE_MAX, srcSize_wrong, "");
 
     /* Decode literals section */
-    {   size_t const litCSize = ZSTD_decodeLiteralsBlock(dctx, src, srcSize);
-        DEBUGLOG(5, "ZSTD_decodeLiteralsBlock : %u", (U32)litCSize);
+    {   size_t const litCSize = ZSTD_decodeLiteralsBlock(dctx, src, srcSize, dst, dstCapacity, streaming);
+        DEBUGLOG(5, "ZSTD_decodeLiteralsBlock : cSize=%u, nbLiterals=%zu", (U32)litCSize, dctx->litSize);
         if (ZSTD_isError(litCSize)) return litCSize;
         ip += litCSize;
         srcSize -= litCSize;
@@ -1364,6 +2088,23 @@ ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
 
     /* Build Decoding Tables */
     {
+        /* Compute the maximum block size, which must also work when !frame and fParams are unset.
+         * Additionally, take the min with dstCapacity to ensure that the totalHistorySize fits in a size_t.
+         */
+        size_t const blockSizeMax = MIN(dstCapacity, (frame ? dctx->fParams.blockSizeMax : ZSTD_BLOCKSIZE_MAX));
+        size_t const totalHistorySize = ZSTD_totalHistorySize((BYTE*)dst + blockSizeMax, (BYTE const*)dctx->virtualStart);
+        /* isLongOffset must be true if there are long offsets.
+         * Offsets are long if they are larger than ZSTD_maxShortOffset().
+         * We don't expect that to be the case in 64-bit mode.
+         *
+         * We check here to see if our history is large enough to allow long offsets.
+         * If it isn't, then we can't possible have (valid) long offsets. If the offset
+         * is invalid, then it is okay to read it incorrectly.
+         *
+         * If isLongOffsets is true, then we will later check our decoding table to see
+         * if it is even possible to generate long offsets.
+         */
+        ZSTD_longOffset_e isLongOffset = (ZSTD_longOffset_e)(MEM_32bits() && (totalHistorySize > ZSTD_maxShortOffset()));
         /* These macros control at build-time which decompressor implementation
          * we use. If neither is defined, we do some inspection and dispatch at
          * runtime.
@@ -1371,6 +2112,11 @@ ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
 #if !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \
     !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG)
         int usePrefetchDecoder = dctx->ddictIsCold;
+#else
+        /* Set to 1 to avoid computing offset info if we don't need to.
+         * Otherwise this value is ignored.
+         */
+        int usePrefetchDecoder = 1;
 #endif
         int nbSeq;
         size_t const seqHSize = ZSTD_decodeSeqHeaders(dctx, &nbSeq, ip, srcSize);
@@ -1380,38 +2126,53 @@ ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
 
         RETURN_ERROR_IF(dst == NULL && nbSeq > 0, dstSize_tooSmall, "NULL not handled");
 
-#if !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \
-    !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG)
-        if ( !usePrefetchDecoder
-          && (!frame || (dctx->fParams.windowSize > (1<<24)))
-          && (nbSeq>ADVANCED_SEQS) ) {  /* could probably use a larger nbSeq limit */
-            U32 const shareLongOffsets = ZSTD_getLongOffsetsShare(dctx->OFTptr);
-            U32 const minShare = MEM_64bits() ? 7 : 20; /* heuristic values, correspond to 2.73% and 7.81% */
-            usePrefetchDecoder = (shareLongOffsets >= minShare);
+        /* If we could potentially have long offsets, or we might want to use the prefetch decoder,
+         * compute information about the share of long offsets, and the maximum nbAdditionalBits.
+         * NOTE: could probably use a larger nbSeq limit
+         */
+        if (isLongOffset || (!usePrefetchDecoder && (totalHistorySize > (1u << 24)) && (nbSeq > 8))) {
+            ZSTD_OffsetInfo const info = ZSTD_getOffsetInfo(dctx->OFTptr, nbSeq);
+            if (isLongOffset && info.maxNbAdditionalBits <= STREAM_ACCUMULATOR_MIN) {
+                /* If isLongOffset, but the maximum number of additional bits that we see in our table is small
+                 * enough, then we know it is impossible to have too long an offset in this block, so we can
+                 * use the regular offset decoder.
+                 */
+                isLongOffset = ZSTD_lo_isRegularOffset;
+            }
+            if (!usePrefetchDecoder) {
+                U32 const minShare = MEM_64bits() ? 7 : 20; /* heuristic values, correspond to 2.73% and 7.81% */
+                usePrefetchDecoder = (info.longOffsetShare >= minShare);
+            }
         }
-#endif
 
         dctx->ddictIsCold = 0;
 
 #if !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \
     !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG)
-        if (usePrefetchDecoder)
+        if (usePrefetchDecoder) {
+#else
+        (void)usePrefetchDecoder;
+        {
 #endif
 #ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT
             return ZSTD_decompressSequencesLong(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset, frame);
 #endif
+        }
 
 #ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG
         /* else */
-        return ZSTD_decompressSequences(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset, frame);
+        if (dctx->litBufferLocation == ZSTD_split)
+            return ZSTD_decompressSequencesSplitLitBuffer(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset, frame);
+        else
+            return ZSTD_decompressSequences(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset, frame);
 #endif
     }
 }
 
 
-void ZSTD_checkContinuity(ZSTD_DCtx* dctx, const void* dst)
+void ZSTD_checkContinuity(ZSTD_DCtx* dctx, const void* dst, size_t dstSize)
 {
-    if (dst != dctx->previousDstEnd) {   /* not contiguous */
+    if (dst != dctx->previousDstEnd && dstSize > 0) {   /* not contiguous */
         dctx->dictEnd = dctx->previousDstEnd;
         dctx->virtualStart = (const char*)dst - ((const char*)(dctx->previousDstEnd) - (const char*)(dctx->prefixStart));
         dctx->prefixStart = dst;
@@ -1425,8 +2186,8 @@ size_t ZSTD_decompressBlock(ZSTD_DCtx* dctx,
                       const void* src, size_t srcSize)
 {
     size_t dSize;
-    ZSTD_checkContinuity(dctx, dst);
-    dSize = ZSTD_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize, /* frame */ 0);
+    ZSTD_checkContinuity(dctx, dst, dstCapacity);
+    dSize = ZSTD_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize, /* frame */ 0, not_streaming);
     dctx->previousDstEnd = (char*)dst + dSize;
     return dSize;
 }
diff --git a/contrib/zstd/zstd_decompress_block.h b/contrib/zstd/zstd_decompress_block.h
index a065f8d93..f640ff951 100644
--- a/contrib/zstd/zstd_decompress_block.h
+++ b/contrib/zstd/zstd_decompress_block.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -15,8 +15,8 @@
 /*-*******************************************************
  *  Dependencies
  *********************************************************/
-#include <stddef.h>   /* size_t */
-#include "zstd.h"    /* DCtx, and some public functions */
+#include "zstd_deps.h"   /* size_t */
+#include "../zstd.h"    /* DCtx, and some public functions */
 #include "zstd_internal.h"  /* blockProperties_t, and some public functions */
 #include "zstd_decompress_internal.h"  /* ZSTD_seqSymbol */
 
@@ -33,6 +33,12 @@
  */
 
 
+ /* Streaming state is used to inform allocation of the literal buffer */
+typedef enum {
+    not_streaming = 0,
+    is_streaming = 1
+} streaming_operation;
+
 /* ZSTD_decompressBlock_internal() :
  * decompress block, starting at `src`,
  * into destination buffer `dst`.
@@ -41,19 +47,22 @@
  */
 size_t ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
                                void* dst, size_t dstCapacity,
-                         const void* src, size_t srcSize, const int frame);
+                         const void* src, size_t srcSize, const int frame, const streaming_operation streaming);
 
 /* ZSTD_buildFSETable() :
  * generate FSE decoding table for one symbol (ll, ml or off)
  * this function must be called with valid parameters only
  * (dt is large enough, normalizedCounter distribution total is a power of 2, max is within range, etc.)
  * in which case it cannot fail.
+ * The workspace must be 4-byte aligned and at least ZSTD_BUILD_FSE_TABLE_WKSP_SIZE bytes, which is
+ * defined in zstd_decompress_internal.h.
  * Internal use only.
  */
 void ZSTD_buildFSETable(ZSTD_seqSymbol* dt,
              const short* normalizedCounter, unsigned maxSymbolValue,
-             const U32* baseValue, const U32* nbAdditionalBits,
-                   unsigned tableLog);
+             const U32* baseValue, const U8* nbAdditionalBits,
+                   unsigned tableLog, void* wksp, size_t wkspSize,
+                   int bmi2);
 
 
 #endif /* ZSTD_DEC_BLOCK_H */
diff --git a/contrib/zstd/zstd_decompress_internal.h b/contrib/zstd/zstd_decompress_internal.h
index a5cafc04e..a164de3ef 100644
--- a/contrib/zstd/zstd_decompress_internal.h
+++ b/contrib/zstd/zstd_decompress_internal.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -20,33 +20,33 @@
  *  Dependencies
  *********************************************************/
 #include "mem.h"             /* BYTE, U16, U32 */
-#include "zstd_internal.h"   /* ZSTD_seqSymbol */
+#include "zstd_internal.h"   /* constants : MaxLL, MaxML, MaxOff, LLFSELog, etc. */
 
 
 
 /*-*******************************************************
  *  Constants
  *********************************************************/
-static const U32 LL_base[MaxLL+1] = {
+static UNUSED_ATTR const U32 LL_base[MaxLL+1] = {
                  0,    1,    2,     3,     4,     5,     6,      7,
                  8,    9,   10,    11,    12,    13,    14,     15,
                 16,   18,   20,    22,    24,    28,    32,     40,
                 48,   64, 0x80, 0x100, 0x200, 0x400, 0x800, 0x1000,
                 0x2000, 0x4000, 0x8000, 0x10000 };
 
-static const U32 OF_base[MaxOff+1] = {
+static UNUSED_ATTR const U32 OF_base[MaxOff+1] = {
                  0,        1,       1,       5,     0xD,     0x1D,     0x3D,     0x7D,
                  0xFD,   0x1FD,   0x3FD,   0x7FD,   0xFFD,   0x1FFD,   0x3FFD,   0x7FFD,
                  0xFFFD, 0x1FFFD, 0x3FFFD, 0x7FFFD, 0xFFFFD, 0x1FFFFD, 0x3FFFFD, 0x7FFFFD,
                  0xFFFFFD, 0x1FFFFFD, 0x3FFFFFD, 0x7FFFFFD, 0xFFFFFFD, 0x1FFFFFFD, 0x3FFFFFFD, 0x7FFFFFFD };
 
-static const U32 OF_bits[MaxOff+1] = {
+static UNUSED_ATTR const U8 OF_bits[MaxOff+1] = {
                      0,  1,  2,  3,  4,  5,  6,  7,
                      8,  9, 10, 11, 12, 13, 14, 15,
                     16, 17, 18, 19, 20, 21, 22, 23,
                     24, 25, 26, 27, 28, 29, 30, 31 };
 
-static const U32 ML_base[MaxML+1] = {
+static UNUSED_ATTR const U32 ML_base[MaxML+1] = {
                      3,  4,  5,    6,     7,     8,     9,    10,
                     11, 12, 13,   14,    15,    16,    17,    18,
                     19, 20, 21,   22,    23,    24,    25,    26,
@@ -73,12 +73,17 @@ static const U32 ML_base[MaxML+1] = {
 
  #define SEQSYMBOL_TABLE_SIZE(log)   (1 + (1 << (log)))
 
+#define ZSTD_BUILD_FSE_TABLE_WKSP_SIZE (sizeof(S16) * (MaxSeq + 1) + (1u << MaxFSELog) + sizeof(U64))
+#define ZSTD_BUILD_FSE_TABLE_WKSP_SIZE_U32 ((ZSTD_BUILD_FSE_TABLE_WKSP_SIZE + sizeof(U32) - 1) / sizeof(U32))
+#define ZSTD_HUFFDTABLE_CAPACITY_LOG 12
+
 typedef struct {
     ZSTD_seqSymbol LLTable[SEQSYMBOL_TABLE_SIZE(LLFSELog)];    /* Note : Space reserved for FSE Tables */
     ZSTD_seqSymbol OFTable[SEQSYMBOL_TABLE_SIZE(OffFSELog)];   /* is also used as temporary workspace while building hufTable during DDict creation */
     ZSTD_seqSymbol MLTable[SEQSYMBOL_TABLE_SIZE(MLFSELog)];    /* and therefore must be at least HUF_DECOMPRESS_WORKSPACE_SIZE large */
-    HUF_DTable hufTable[HUF_DTABLE_SIZE(HufLog)];  /* can accommodate HUF_decompress4X */
+    HUF_DTable hufTable[HUF_DTABLE_SIZE(ZSTD_HUFFDTABLE_CAPACITY_LOG)];  /* can accommodate HUF_decompress4X */
     U32 rep[ZSTD_REP_NUM];
+    U32 workspace[ZSTD_BUILD_FSE_TABLE_WKSP_SIZE_U32];
 } ZSTD_entropyDTables_t;
 
 typedef enum { ZSTDds_getFrameHeaderSize, ZSTDds_decodeFrameHeader,
@@ -95,10 +100,28 @@ typedef enum {
     ZSTD_use_once = 1            /* Use the dictionary once and set to ZSTD_dont_use */
 } ZSTD_dictUses_e;
 
+/* Hashset for storing references to multiple ZSTD_DDict within ZSTD_DCtx */
+typedef struct {
+    const ZSTD_DDict** ddictPtrTable;
+    size_t ddictPtrTableSize;
+    size_t ddictPtrCount;
+} ZSTD_DDictHashSet;
+
+#ifndef ZSTD_DECODER_INTERNAL_BUFFER
+#  define ZSTD_DECODER_INTERNAL_BUFFER  (1 << 16)
+#endif
+
+#define ZSTD_LBMIN 64
+#define ZSTD_LBMAX (128 << 10)
+
+/* extra buffer, compensates when dst is not large enough to store litBuffer */
+#define ZSTD_LITBUFFEREXTRASIZE  BOUNDED(ZSTD_LBMIN, ZSTD_DECODER_INTERNAL_BUFFER, ZSTD_LBMAX)
+
 typedef enum {
-    ZSTD_obm_buffered = 0,  /* Buffer the output */
-    ZSTD_obm_stable = 1     /* ZSTD_outBuffer is stable */
-} ZSTD_outBufferMode_e;
+    ZSTD_not_in_dst = 0,  /* Stored entirely within litExtraBuffer */
+    ZSTD_in_dst = 1,           /* Stored entirely within dst (in memory after current output write) */
+    ZSTD_split = 2            /* Split between litExtraBuffer and dst */
+} ZSTD_litLocation_e;
 
 struct ZSTD_DCtx_s
 {
@@ -114,6 +137,7 @@ struct ZSTD_DCtx_s
     const void* dictEnd;          /* end of previous segment */
     size_t expected;
     ZSTD_frameHeader fParams;
+    U64 processedCSize;
     U64 decodedSize;
     blockType_e bType;            /* used in ZSTD_decompressContinue(), store blockType between block header decoding and block decompression stages */
     ZSTD_dStage stage;
@@ -122,12 +146,16 @@ struct ZSTD_DCtx_s
     XXH64_state_t xxhState;
     size_t headerSize;
     ZSTD_format_e format;
+    ZSTD_forceIgnoreChecksum_e forceIgnoreChecksum;   /* User specified: if == 1, will ignore checksums in compressed frame. Default == 0 */
+    U32 validateChecksum;         /* if == 1, will validate checksum. Is == 1 if (fParams.checksumFlag == 1) and (forceIgnoreChecksum == 0). */
     const BYTE* litPtr;
     ZSTD_customMem customMem;
     size_t litSize;
     size_t rleSize;
     size_t staticSize;
+#if DYNAMIC_BMI2 != 0
     int bmi2;                     /* == 1 if the CPU supports BMI2 and 0 otherwise. CPU support is determined dynamically once per context lifetime. */
+#endif
 
     /* dictionary */
     ZSTD_DDict* ddictLocal;
@@ -135,6 +163,9 @@ struct ZSTD_DCtx_s
     U32 dictID;
     int ddictIsCold;             /* if == 1 : dictionary is "new" for working context, and presumed "cold" (not in cpu cache) */
     ZSTD_dictUses_e dictUses;
+    ZSTD_DDictHashSet* ddictSet;                    /* Hash set for multiple ddicts */
+    ZSTD_refMultipleDDicts_e refMultipleDDicts;     /* User specified: if == 1, will allow references to multiple DDicts. Default == 0 (disabled) */
+    int disableHufAsm;
 
     /* streaming */
     ZSTD_dStreamStage streamStage;
@@ -147,16 +178,21 @@ struct ZSTD_DCtx_s
     size_t outStart;
     size_t outEnd;
     size_t lhSize;
+#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
     void* legacyContext;
     U32 previousLegacyVersion;
     U32 legacyVersion;
+#endif
     U32 hostageByte;
     int noForwardProgress;
-    ZSTD_outBufferMode_e outBufferMode;
+    ZSTD_bufferMode_e outBufferMode;
     ZSTD_outBuffer expectedOutBuffer;
 
     /* workspace */
-    BYTE litBuffer[ZSTD_BLOCKSIZE_MAX + WILDCOPY_OVERLENGTH];
+    BYTE* litBuffer;
+    const BYTE* litBufferEnd;
+    ZSTD_litLocation_e litBufferLocation;
+    BYTE litExtraBuffer[ZSTD_LITBUFFEREXTRASIZE + WILDCOPY_OVERLENGTH]; /* literal buffer can be split between storage within dst and within this scratch buffer */
     BYTE headerBuffer[ZSTD_FRAMEHEADERSIZE_MAX];
 
     size_t oversizedDuration;
@@ -165,8 +201,21 @@ struct ZSTD_DCtx_s
     void const* dictContentBeginForFuzzing;
     void const* dictContentEndForFuzzing;
 #endif
+
+    /* Tracing */
+#if ZSTD_TRACE
+    ZSTD_TraceCtx traceCtx;
+#endif
 };  /* typedef'd to ZSTD_DCtx within "zstd.h" */
 
+MEM_STATIC int ZSTD_DCtx_get_bmi2(const struct ZSTD_DCtx_s *dctx) {
+#if DYNAMIC_BMI2 != 0
+	return dctx->bmi2;
+#else
+    (void)dctx;
+	return 0;
+#endif
+}
 
 /*-*******************************************************
  *  Shared internal functions
@@ -183,7 +232,7 @@ size_t ZSTD_loadDEntropy(ZSTD_entropyDTables_t* entropy,
  *  If yes, do nothing (continue on current segment).
  *  If not, classify previous segment as "external dictionary", and start a new segment.
  *  This function cannot fail. */
-void ZSTD_checkContinuity(ZSTD_DCtx* dctx, const void* dst);
+void ZSTD_checkContinuity(ZSTD_DCtx* dctx, const void* dst, size_t dstSize);
 
 
 #endif /* ZSTD_DECOMPRESS_INTERNAL_H */
diff --git a/contrib/zstd/zstd_deps.h b/contrib/zstd/zstd_deps.h
new file mode 100644
index 000000000..4d767ae9b
--- /dev/null
+++ b/contrib/zstd/zstd_deps.h
@@ -0,0 +1,111 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+/* This file provides common libc dependencies that zstd requires.
+ * The purpose is to allow replacing this file with a custom implementation
+ * to compile zstd without libc support.
+ */
+
+/* Need:
+ * NULL
+ * INT_MAX
+ * UINT_MAX
+ * ZSTD_memcpy()
+ * ZSTD_memset()
+ * ZSTD_memmove()
+ */
+#ifndef ZSTD_DEPS_COMMON
+#define ZSTD_DEPS_COMMON
+
+#include <limits.h>
+#include <stddef.h>
+#include <string.h>
+
+#if defined(__GNUC__) && __GNUC__ >= 4
+# define ZSTD_memcpy(d,s,l) __builtin_memcpy((d),(s),(l))
+# define ZSTD_memmove(d,s,l) __builtin_memmove((d),(s),(l))
+# define ZSTD_memset(p,v,l) __builtin_memset((p),(v),(l))
+#else
+# define ZSTD_memcpy(d,s,l) memcpy((d),(s),(l))
+# define ZSTD_memmove(d,s,l) memmove((d),(s),(l))
+# define ZSTD_memset(p,v,l) memset((p),(v),(l))
+#endif
+
+#endif /* ZSTD_DEPS_COMMON */
+
+/* Need:
+ * ZSTD_malloc()
+ * ZSTD_free()
+ * ZSTD_calloc()
+ */
+#ifdef ZSTD_DEPS_NEED_MALLOC
+#ifndef ZSTD_DEPS_MALLOC
+#define ZSTD_DEPS_MALLOC
+
+#include <stdlib.h>
+
+#define ZSTD_malloc(s) malloc(s)
+#define ZSTD_calloc(n,s) calloc((n), (s))
+#define ZSTD_free(p) free((p))
+
+#endif /* ZSTD_DEPS_MALLOC */
+#endif /* ZSTD_DEPS_NEED_MALLOC */
+
+/*
+ * Provides 64-bit math support.
+ * Need:
+ * U64 ZSTD_div64(U64 dividend, U32 divisor)
+ */
+#ifdef ZSTD_DEPS_NEED_MATH64
+#ifndef ZSTD_DEPS_MATH64
+#define ZSTD_DEPS_MATH64
+
+#define ZSTD_div64(dividend, divisor) ((dividend) / (divisor))
+
+#endif /* ZSTD_DEPS_MATH64 */
+#endif /* ZSTD_DEPS_NEED_MATH64 */
+
+/* Need:
+ * assert()
+ */
+#ifdef ZSTD_DEPS_NEED_ASSERT
+#ifndef ZSTD_DEPS_ASSERT
+#define ZSTD_DEPS_ASSERT
+
+#include <assert.h>
+
+#endif /* ZSTD_DEPS_ASSERT */
+#endif /* ZSTD_DEPS_NEED_ASSERT */
+
+/* Need:
+ * ZSTD_DEBUG_PRINT()
+ */
+#ifdef ZSTD_DEPS_NEED_IO
+#ifndef ZSTD_DEPS_IO
+#define ZSTD_DEPS_IO
+
+#include <stdio.h>
+#define ZSTD_DEBUG_PRINT(...) fprintf(stderr, __VA_ARGS__)
+
+#endif /* ZSTD_DEPS_IO */
+#endif /* ZSTD_DEPS_NEED_IO */
+
+/* Only requested when <stdint.h> is known to be present.
+ * Need:
+ * intptr_t
+ */
+#ifdef ZSTD_DEPS_NEED_STDINT
+#ifndef ZSTD_DEPS_STDINT
+#define ZSTD_DEPS_STDINT
+
+#include <stdint.h>
+
+#endif /* ZSTD_DEPS_STDINT */
+#endif /* ZSTD_DEPS_NEED_STDINT */
diff --git a/contrib/zstd/zstd_double_fast.c b/contrib/zstd/zstd_double_fast.c
index 27eed66cf..0ad88ffc7 100644
--- a/contrib/zstd/zstd_double_fast.c
+++ b/contrib/zstd/zstd_double_fast.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -11,8 +11,43 @@
 #include "zstd_compress_internal.h"
 #include "zstd_double_fast.h"
 
+static void ZSTD_fillDoubleHashTableForCDict(ZSTD_matchState_t* ms,
+                              void const* end, ZSTD_dictTableLoadMethod_e dtlm)
+{
+    const ZSTD_compressionParameters* const cParams = &ms->cParams;
+    U32* const hashLarge = ms->hashTable;
+    U32  const hBitsL = cParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS;
+    U32  const mls = cParams->minMatch;
+    U32* const hashSmall = ms->chainTable;
+    U32  const hBitsS = cParams->chainLog + ZSTD_SHORT_CACHE_TAG_BITS;
+    const BYTE* const base = ms->window.base;
+    const BYTE* ip = base + ms->nextToUpdate;
+    const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE;
+    const U32 fastHashFillStep = 3;
 
-void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms,
+    /* Always insert every fastHashFillStep position into the hash tables.
+     * Insert the other positions into the large hash table if their entry
+     * is empty.
+     */
+    for (; ip + fastHashFillStep - 1 <= iend; ip += fastHashFillStep) {
+        U32 const curr = (U32)(ip - base);
+        U32 i;
+        for (i = 0; i < fastHashFillStep; ++i) {
+            size_t const smHashAndTag = ZSTD_hashPtr(ip + i, hBitsS, mls);
+            size_t const lgHashAndTag = ZSTD_hashPtr(ip + i, hBitsL, 8);
+            if (i == 0) {
+                ZSTD_writeTaggedIndex(hashSmall, smHashAndTag, curr + i);
+            }
+            if (i == 0 || hashLarge[lgHashAndTag >> ZSTD_SHORT_CACHE_TAG_BITS] == 0) {
+                ZSTD_writeTaggedIndex(hashLarge, lgHashAndTag, curr + i);
+            }
+            /* Only load extra positions for ZSTD_dtlm_full */
+            if (dtlm == ZSTD_dtlm_fast)
+                break;
+    }   }
+}
+
+static void ZSTD_fillDoubleHashTableForCCtx(ZSTD_matchState_t* ms,
                               void const* end, ZSTD_dictTableLoadMethod_e dtlm)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
@@ -31,27 +66,249 @@ void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms,
      * is empty.
      */
     for (; ip + fastHashFillStep - 1 <= iend; ip += fastHashFillStep) {
-        U32 const current = (U32)(ip - base);
+        U32 const curr = (U32)(ip - base);
         U32 i;
         for (i = 0; i < fastHashFillStep; ++i) {
             size_t const smHash = ZSTD_hashPtr(ip + i, hBitsS, mls);
             size_t const lgHash = ZSTD_hashPtr(ip + i, hBitsL, 8);
             if (i == 0)
-                hashSmall[smHash] = current + i;
+                hashSmall[smHash] = curr + i;
             if (i == 0 || hashLarge[lgHash] == 0)
-                hashLarge[lgHash] = current + i;
+                hashLarge[lgHash] = curr + i;
             /* Only load extra positions for ZSTD_dtlm_full */
             if (dtlm == ZSTD_dtlm_fast)
                 break;
-    }   }
+        }   }
+}
+
+void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms,
+                        const void* const end,
+                        ZSTD_dictTableLoadMethod_e dtlm,
+                        ZSTD_tableFillPurpose_e tfp)
+{
+    if (tfp == ZSTD_tfp_forCDict) {
+        ZSTD_fillDoubleHashTableForCDict(ms, end, dtlm);
+    } else {
+        ZSTD_fillDoubleHashTableForCCtx(ms, end, dtlm);
+    }
+}
+
+
+FORCE_INLINE_TEMPLATE
+size_t ZSTD_compressBlock_doubleFast_noDict_generic(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize, U32 const mls /* template */)
+{
+    ZSTD_compressionParameters const* cParams = &ms->cParams;
+    U32* const hashLong = ms->hashTable;
+    const U32 hBitsL = cParams->hashLog;
+    U32* const hashSmall = ms->chainTable;
+    const U32 hBitsS = cParams->chainLog;
+    const BYTE* const base = ms->window.base;
+    const BYTE* const istart = (const BYTE*)src;
+    const BYTE* anchor = istart;
+    const U32 endIndex = (U32)((size_t)(istart - base) + srcSize);
+    /* presumes that, if there is a dictionary, it must be using Attach mode */
+    const U32 prefixLowestIndex = ZSTD_getLowestPrefixIndex(ms, endIndex, cParams->windowLog);
+    const BYTE* const prefixLowest = base + prefixLowestIndex;
+    const BYTE* const iend = istart + srcSize;
+    const BYTE* const ilimit = iend - HASH_READ_SIZE;
+    U32 offset_1=rep[0], offset_2=rep[1];
+    U32 offsetSaved1 = 0, offsetSaved2 = 0;
+
+    size_t mLength;
+    U32 offset;
+    U32 curr;
+
+    /* how many positions to search before increasing step size */
+    const size_t kStepIncr = 1 << kSearchStrength;
+    /* the position at which to increment the step size if no match is found */
+    const BYTE* nextStep;
+    size_t step; /* the current step size */
+
+    size_t hl0; /* the long hash at ip */
+    size_t hl1; /* the long hash at ip1 */
+
+    U32 idxl0; /* the long match index for ip */
+    U32 idxl1; /* the long match index for ip1 */
+
+    const BYTE* matchl0; /* the long match for ip */
+    const BYTE* matchs0; /* the short match for ip */
+    const BYTE* matchl1; /* the long match for ip1 */
+
+    const BYTE* ip = istart; /* the current position */
+    const BYTE* ip1; /* the next position */
+
+    DEBUGLOG(5, "ZSTD_compressBlock_doubleFast_noDict_generic");
+
+    /* init */
+    ip += ((ip - prefixLowest) == 0);
+    {
+        U32 const current = (U32)(ip - base);
+        U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, current, cParams->windowLog);
+        U32 const maxRep = current - windowLow;
+        if (offset_2 > maxRep) offsetSaved2 = offset_2, offset_2 = 0;
+        if (offset_1 > maxRep) offsetSaved1 = offset_1, offset_1 = 0;
+    }
+
+    /* Outer Loop: one iteration per match found and stored */
+    while (1) {
+        step = 1;
+        nextStep = ip + kStepIncr;
+        ip1 = ip + step;
+
+        if (ip1 > ilimit) {
+            goto _cleanup;
+        }
+
+        hl0 = ZSTD_hashPtr(ip, hBitsL, 8);
+        idxl0 = hashLong[hl0];
+        matchl0 = base + idxl0;
+
+        /* Inner Loop: one iteration per search / position */
+        do {
+            const size_t hs0 = ZSTD_hashPtr(ip, hBitsS, mls);
+            const U32 idxs0 = hashSmall[hs0];
+            curr = (U32)(ip-base);
+            matchs0 = base + idxs0;
+
+            hashLong[hl0] = hashSmall[hs0] = curr;   /* update hash tables */
+
+            /* check noDict repcode */
+            if ((offset_1 > 0) & (MEM_read32(ip+1-offset_1) == MEM_read32(ip+1))) {
+                mLength = ZSTD_count(ip+1+4, ip+1+4-offset_1, iend) + 4;
+                ip++;
+                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength);
+                goto _match_stored;
+            }
+
+            hl1 = ZSTD_hashPtr(ip1, hBitsL, 8);
+
+            if (idxl0 > prefixLowestIndex) {
+                /* check prefix long match */
+                if (MEM_read64(matchl0) == MEM_read64(ip)) {
+                    mLength = ZSTD_count(ip+8, matchl0+8, iend) + 8;
+                    offset = (U32)(ip-matchl0);
+                    while (((ip>anchor) & (matchl0>prefixLowest)) && (ip[-1] == matchl0[-1])) { ip--; matchl0--; mLength++; } /* catch up */
+                    goto _match_found;
+                }
+            }
+
+            idxl1 = hashLong[hl1];
+            matchl1 = base + idxl1;
+
+            if (idxs0 > prefixLowestIndex) {
+                /* check prefix short match */
+                if (MEM_read32(matchs0) == MEM_read32(ip)) {
+                    goto _search_next_long;
+                }
+            }
+
+            if (ip1 >= nextStep) {
+                PREFETCH_L1(ip1 + 64);
+                PREFETCH_L1(ip1 + 128);
+                step++;
+                nextStep += kStepIncr;
+            }
+            ip = ip1;
+            ip1 += step;
+
+            hl0 = hl1;
+            idxl0 = idxl1;
+            matchl0 = matchl1;
+    #if defined(__aarch64__)
+            PREFETCH_L1(ip+256);
+    #endif
+        } while (ip1 <= ilimit);
+
+_cleanup:
+        /* If offset_1 started invalid (offsetSaved1 != 0) and became valid (offset_1 != 0),
+         * rotate saved offsets. See comment in ZSTD_compressBlock_fast_noDict for more context. */
+        offsetSaved2 = ((offsetSaved1 != 0) && (offset_1 != 0)) ? offsetSaved1 : offsetSaved2;
+
+        /* save reps for next block */
+        rep[0] = offset_1 ? offset_1 : offsetSaved1;
+        rep[1] = offset_2 ? offset_2 : offsetSaved2;
+
+        /* Return the last literals size */
+        return (size_t)(iend - anchor);
+
+_search_next_long:
+
+        /* check prefix long +1 match */
+        if (idxl1 > prefixLowestIndex) {
+            if (MEM_read64(matchl1) == MEM_read64(ip1)) {
+                ip = ip1;
+                mLength = ZSTD_count(ip+8, matchl1+8, iend) + 8;
+                offset = (U32)(ip-matchl1);
+                while (((ip>anchor) & (matchl1>prefixLowest)) && (ip[-1] == matchl1[-1])) { ip--; matchl1--; mLength++; } /* catch up */
+                goto _match_found;
+            }
+        }
+
+        /* if no long +1 match, explore the short match we found */
+        mLength = ZSTD_count(ip+4, matchs0+4, iend) + 4;
+        offset = (U32)(ip - matchs0);
+        while (((ip>anchor) & (matchs0>prefixLowest)) && (ip[-1] == matchs0[-1])) { ip--; matchs0--; mLength++; } /* catch up */
+
+        /* fall-through */
+
+_match_found: /* requires ip, offset, mLength */
+        offset_2 = offset_1;
+        offset_1 = offset;
+
+        if (step < 4) {
+            /* It is unsafe to write this value back to the hashtable when ip1 is
+             * greater than or equal to the new ip we will have after we're done
+             * processing this match. Rather than perform that test directly
+             * (ip1 >= ip + mLength), which costs speed in practice, we do a simpler
+             * more predictable test. The minmatch even if we take a short match is
+             * 4 bytes, so as long as step, the distance between ip and ip1
+             * (initially) is less than 4, we know ip1 < new ip. */
+            hashLong[hl1] = (U32)(ip1 - base);
+        }
+
+        ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
+
+_match_stored:
+        /* match found */
+        ip += mLength;
+        anchor = ip;
+
+        if (ip <= ilimit) {
+            /* Complementary insertion */
+            /* done after iLimit test, as candidates could be > iend-8 */
+            {   U32 const indexToInsert = curr+2;
+                hashLong[ZSTD_hashPtr(base+indexToInsert, hBitsL, 8)] = indexToInsert;
+                hashLong[ZSTD_hashPtr(ip-2, hBitsL, 8)] = (U32)(ip-2-base);
+                hashSmall[ZSTD_hashPtr(base+indexToInsert, hBitsS, mls)] = indexToInsert;
+                hashSmall[ZSTD_hashPtr(ip-1, hBitsS, mls)] = (U32)(ip-1-base);
+            }
+
+            /* check immediate repcode */
+            while ( (ip <= ilimit)
+                 && ( (offset_2>0)
+                    & (MEM_read32(ip) == MEM_read32(ip - offset_2)) )) {
+                /* store sequence */
+                size_t const rLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4;
+                U32 const tmpOff = offset_2; offset_2 = offset_1; offset_1 = tmpOff;  /* swap offset_2 <=> offset_1 */
+                hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = (U32)(ip-base);
+                hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = (U32)(ip-base);
+                ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, rLength);
+                ip += rLength;
+                anchor = ip;
+                continue;   /* faster when present ... (?) */
+            }
+        }
+    }
 }
 
 
 FORCE_INLINE_TEMPLATE
-size_t ZSTD_compressBlock_doubleFast_generic(
+size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize,
-        U32 const mls /* template */, ZSTD_dictMode_e const dictMode)
+        U32 const mls /* template */)
 {
     ZSTD_compressionParameters const* cParams = &ms->cParams;
     U32* const hashLong = ms->hashTable;
@@ -69,57 +326,39 @@ size_t ZSTD_compressBlock_doubleFast_generic(
     const BYTE* const iend = istart + srcSize;
     const BYTE* const ilimit = iend - HASH_READ_SIZE;
     U32 offset_1=rep[0], offset_2=rep[1];
-    U32 offsetSaved = 0;
 
     const ZSTD_matchState_t* const dms = ms->dictMatchState;
-    const ZSTD_compressionParameters* const dictCParams =
-                                     dictMode == ZSTD_dictMatchState ?
-                                     &dms->cParams : NULL;
-    const U32* const dictHashLong  = dictMode == ZSTD_dictMatchState ?
-                                     dms->hashTable : NULL;
-    const U32* const dictHashSmall = dictMode == ZSTD_dictMatchState ?
-                                     dms->chainTable : NULL;
-    const U32 dictStartIndex       = dictMode == ZSTD_dictMatchState ?
-                                     dms->window.dictLimit : 0;
-    const BYTE* const dictBase     = dictMode == ZSTD_dictMatchState ?
-                                     dms->window.base : NULL;
-    const BYTE* const dictStart    = dictMode == ZSTD_dictMatchState ?
-                                     dictBase + dictStartIndex : NULL;
-    const BYTE* const dictEnd      = dictMode == ZSTD_dictMatchState ?
-                                     dms->window.nextSrc : NULL;
-    const U32 dictIndexDelta       = dictMode == ZSTD_dictMatchState ?
-                                     prefixLowestIndex - (U32)(dictEnd - dictBase) :
-                                     0;
-    const U32 dictHBitsL           = dictMode == ZSTD_dictMatchState ?
-                                     dictCParams->hashLog : hBitsL;
-    const U32 dictHBitsS           = dictMode == ZSTD_dictMatchState ?
-                                     dictCParams->chainLog : hBitsS;
+    const ZSTD_compressionParameters* const dictCParams = &dms->cParams;
+    const U32* const dictHashLong  = dms->hashTable;
+    const U32* const dictHashSmall = dms->chainTable;
+    const U32 dictStartIndex       = dms->window.dictLimit;
+    const BYTE* const dictBase     = dms->window.base;
+    const BYTE* const dictStart    = dictBase + dictStartIndex;
+    const BYTE* const dictEnd      = dms->window.nextSrc;
+    const U32 dictIndexDelta       = prefixLowestIndex - (U32)(dictEnd - dictBase);
+    const U32 dictHBitsL           = dictCParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS;
+    const U32 dictHBitsS           = dictCParams->chainLog + ZSTD_SHORT_CACHE_TAG_BITS;
     const U32 dictAndPrefixLength  = (U32)((ip - prefixLowest) + (dictEnd - dictStart));
 
-    DEBUGLOG(5, "ZSTD_compressBlock_doubleFast_generic");
-
-    assert(dictMode == ZSTD_noDict || dictMode == ZSTD_dictMatchState);
+    DEBUGLOG(5, "ZSTD_compressBlock_doubleFast_dictMatchState_generic");
 
     /* if a dictionary is attached, it must be within window range */
-    if (dictMode == ZSTD_dictMatchState) {
-        assert(ms->window.dictLimit + (1U << cParams->windowLog) >= endIndex);
+    assert(ms->window.dictLimit + (1U << cParams->windowLog) >= endIndex);
+
+    if (ms->prefetchCDictTables) {
+        size_t const hashTableBytes = (((size_t)1) << dictCParams->hashLog) * sizeof(U32);
+        size_t const chainTableBytes = (((size_t)1) << dictCParams->chainLog) * sizeof(U32);
+        PREFETCH_AREA(dictHashLong, hashTableBytes)
+        PREFETCH_AREA(dictHashSmall, chainTableBytes)
     }
 
     /* init */
     ip += (dictAndPrefixLength == 0);
-    if (dictMode == ZSTD_noDict) {
-        U32 const current = (U32)(ip - base);
-        U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, current, cParams->windowLog);
-        U32 const maxRep = current - windowLow;
-        if (offset_2 > maxRep) offsetSaved = offset_2, offset_2 = 0;
-        if (offset_1 > maxRep) offsetSaved = offset_1, offset_1 = 0;
-    }
-    if (dictMode == ZSTD_dictMatchState) {
-        /* dictMatchState repCode checks don't currently handle repCode == 0
-         * disabling. */
-        assert(offset_1 <= dictAndPrefixLength);
-        assert(offset_2 <= dictAndPrefixLength);
-    }
+
+    /* dictMatchState repCode checks don't currently handle repCode == 0
+     * disabling. */
+    assert(offset_1 <= dictAndPrefixLength);
+    assert(offset_2 <= dictAndPrefixLength);
 
     /* Main Search Loop */
     while (ip < ilimit) {   /* < instead of <=, because repcode check at (ip+1) */
@@ -127,37 +366,30 @@ size_t ZSTD_compressBlock_doubleFast_generic(
         U32 offset;
         size_t const h2 = ZSTD_hashPtr(ip, hBitsL, 8);
         size_t const h = ZSTD_hashPtr(ip, hBitsS, mls);
-        size_t const dictHL = ZSTD_hashPtr(ip, dictHBitsL, 8);
-        size_t const dictHS = ZSTD_hashPtr(ip, dictHBitsS, mls);
-        U32 const current = (U32)(ip-base);
+        size_t const dictHashAndTagL = ZSTD_hashPtr(ip, dictHBitsL, 8);
+        size_t const dictHashAndTagS = ZSTD_hashPtr(ip, dictHBitsS, mls);
+        U32 const dictMatchIndexAndTagL = dictHashLong[dictHashAndTagL >> ZSTD_SHORT_CACHE_TAG_BITS];
+        U32 const dictMatchIndexAndTagS = dictHashSmall[dictHashAndTagS >> ZSTD_SHORT_CACHE_TAG_BITS];
+        int const dictTagsMatchL = ZSTD_comparePackedTags(dictMatchIndexAndTagL, dictHashAndTagL);
+        int const dictTagsMatchS = ZSTD_comparePackedTags(dictMatchIndexAndTagS, dictHashAndTagS);
+        U32 const curr = (U32)(ip-base);
         U32 const matchIndexL = hashLong[h2];
         U32 matchIndexS = hashSmall[h];
         const BYTE* matchLong = base + matchIndexL;
         const BYTE* match = base + matchIndexS;
-        const U32 repIndex = current + 1 - offset_1;
-        const BYTE* repMatch = (dictMode == ZSTD_dictMatchState
-                            && repIndex < prefixLowestIndex) ?
+        const U32 repIndex = curr + 1 - offset_1;
+        const BYTE* repMatch = (repIndex < prefixLowestIndex) ?
                                dictBase + (repIndex - dictIndexDelta) :
                                base + repIndex;
-        hashLong[h2] = hashSmall[h] = current;   /* update hash tables */
+        hashLong[h2] = hashSmall[h] = curr;   /* update hash tables */
 
-        /* check dictMatchState repcode */
-        if (dictMode == ZSTD_dictMatchState
-            && ((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
+        /* check repcode */
+        if (((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
             && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
             const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
             mLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
             ip++;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, 0, mLength-MINMATCH);
-            goto _match_stored;
-        }
-
-        /* check noDict repcode */
-        if ( dictMode == ZSTD_noDict
-          && ((offset_1 > 0) & (MEM_read32(ip+1-offset_1) == MEM_read32(ip+1)))) {
-            mLength = ZSTD_count(ip+1+4, ip+1+4-offset_1, iend) + 4;
-            ip++;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, 0, mLength-MINMATCH);
+            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength);
             goto _match_stored;
         }
 
@@ -169,15 +401,15 @@ size_t ZSTD_compressBlock_doubleFast_generic(
                 while (((ip>anchor) & (matchLong>prefixLowest)) && (ip[-1] == matchLong[-1])) { ip--; matchLong--; mLength++; } /* catch up */
                 goto _match_found;
             }
-        } else if (dictMode == ZSTD_dictMatchState) {
+        } else if (dictTagsMatchL) {
             /* check dictMatchState long match */
-            U32 const dictMatchIndexL = dictHashLong[dictHL];
+            U32 const dictMatchIndexL = dictMatchIndexAndTagL >> ZSTD_SHORT_CACHE_TAG_BITS;
             const BYTE* dictMatchL = dictBase + dictMatchIndexL;
             assert(dictMatchL < dictEnd);
 
             if (dictMatchL > dictStart && MEM_read64(dictMatchL) == MEM_read64(ip)) {
                 mLength = ZSTD_count_2segments(ip+8, dictMatchL+8, iend, dictEnd, prefixLowest) + 8;
-                offset = (U32)(current - dictMatchIndexL - dictIndexDelta);
+                offset = (U32)(curr - dictMatchIndexL - dictIndexDelta);
                 while (((ip>anchor) & (dictMatchL>dictStart)) && (ip[-1] == dictMatchL[-1])) { ip--; dictMatchL--; mLength++; } /* catch up */
                 goto _match_found;
         }   }
@@ -187,9 +419,9 @@ size_t ZSTD_compressBlock_doubleFast_generic(
             if (MEM_read32(match) == MEM_read32(ip)) {
                 goto _search_next_long;
             }
-        } else if (dictMode == ZSTD_dictMatchState) {
+        } else if (dictTagsMatchS) {
             /* check dictMatchState short match */
-            U32 const dictMatchIndexS = dictHashSmall[dictHS];
+            U32 const dictMatchIndexS = dictMatchIndexAndTagS >> ZSTD_SHORT_CACHE_TAG_BITS;
             match = dictBase + dictMatchIndexS;
             matchIndexS = dictMatchIndexS + dictIndexDelta;
 
@@ -204,12 +436,13 @@ size_t ZSTD_compressBlock_doubleFast_generic(
         continue;
 
 _search_next_long:
-
         {   size_t const hl3 = ZSTD_hashPtr(ip+1, hBitsL, 8);
-            size_t const dictHLNext = ZSTD_hashPtr(ip+1, dictHBitsL, 8);
+            size_t const dictHashAndTagL3 = ZSTD_hashPtr(ip+1, dictHBitsL, 8);
             U32 const matchIndexL3 = hashLong[hl3];
+            U32 const dictMatchIndexAndTagL3 = dictHashLong[dictHashAndTagL3 >> ZSTD_SHORT_CACHE_TAG_BITS];
+            int const dictTagsMatchL3 = ZSTD_comparePackedTags(dictMatchIndexAndTagL3, dictHashAndTagL3);
             const BYTE* matchL3 = base + matchIndexL3;
-            hashLong[hl3] = current + 1;
+            hashLong[hl3] = curr + 1;
 
             /* check prefix long +1 match */
             if (matchIndexL3 > prefixLowestIndex) {
@@ -220,23 +453,23 @@ _search_next_long:
                     while (((ip>anchor) & (matchL3>prefixLowest)) && (ip[-1] == matchL3[-1])) { ip--; matchL3--; mLength++; } /* catch up */
                     goto _match_found;
                 }
-            } else if (dictMode == ZSTD_dictMatchState) {
+            } else if (dictTagsMatchL3) {
                 /* check dict long +1 match */
-                U32 const dictMatchIndexL3 = dictHashLong[dictHLNext];
+                U32 const dictMatchIndexL3 = dictMatchIndexAndTagL3 >> ZSTD_SHORT_CACHE_TAG_BITS;
                 const BYTE* dictMatchL3 = dictBase + dictMatchIndexL3;
                 assert(dictMatchL3 < dictEnd);
                 if (dictMatchL3 > dictStart && MEM_read64(dictMatchL3) == MEM_read64(ip+1)) {
                     mLength = ZSTD_count_2segments(ip+1+8, dictMatchL3+8, iend, dictEnd, prefixLowest) + 8;
                     ip++;
-                    offset = (U32)(current + 1 - dictMatchIndexL3 - dictIndexDelta);
+                    offset = (U32)(curr + 1 - dictMatchIndexL3 - dictIndexDelta);
                     while (((ip>anchor) & (dictMatchL3>dictStart)) && (ip[-1] == dictMatchL3[-1])) { ip--; dictMatchL3--; mLength++; } /* catch up */
                     goto _match_found;
         }   }   }
 
         /* if no long +1 match, explore the short match we found */
-        if (dictMode == ZSTD_dictMatchState && matchIndexS < prefixLowestIndex) {
+        if (matchIndexS < prefixLowestIndex) {
             mLength = ZSTD_count_2segments(ip+4, match+4, iend, dictEnd, prefixLowest) + 4;
-            offset = (U32)(current - matchIndexS);
+            offset = (U32)(curr - matchIndexS);
             while (((ip>anchor) & (match>dictStart)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; } /* catch up */
         } else {
             mLength = ZSTD_count(ip+4, match+4, iend) + 4;
@@ -244,13 +477,11 @@ _search_next_long:
             while (((ip>anchor) & (match>prefixLowest)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; } /* catch up */
         }
 
-        /* fall-through */
-
 _match_found:
         offset_2 = offset_1;
         offset_1 = offset;
 
-        ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH);
+        ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
 
 _match_stored:
         /* match found */
@@ -260,7 +491,7 @@ _match_stored:
         if (ip <= ilimit) {
             /* Complementary insertion */
             /* done after iLimit test, as candidates could be > iend-8 */
-            {   U32 const indexToInsert = current+2;
+            {   U32 const indexToInsert = curr+2;
                 hashLong[ZSTD_hashPtr(base+indexToInsert, hBitsL, 8)] = indexToInsert;
                 hashLong[ZSTD_hashPtr(ip-2, hBitsL, 8)] = (U32)(ip-2-base);
                 hashSmall[ZSTD_hashPtr(base+indexToInsert, hBitsS, mls)] = indexToInsert;
@@ -268,53 +499,55 @@ _match_stored:
             }
 
             /* check immediate repcode */
-            if (dictMode == ZSTD_dictMatchState) {
-                while (ip <= ilimit) {
-                    U32 const current2 = (U32)(ip-base);
-                    U32 const repIndex2 = current2 - offset_2;
-                    const BYTE* repMatch2 = dictMode == ZSTD_dictMatchState
-                        && repIndex2 < prefixLowestIndex ?
-                            dictBase + repIndex2 - dictIndexDelta :
-                            base + repIndex2;
-                    if ( ((U32)((prefixLowestIndex-1) - (U32)repIndex2) >= 3 /* intentional overflow */)
-                       && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
-                        const BYTE* const repEnd2 = repIndex2 < prefixLowestIndex ? dictEnd : iend;
-                        size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixLowest) + 4;
-                        U32 tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset;   /* swap offset_2 <=> offset_1 */
-                        ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, repLength2-MINMATCH);
-                        hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = current2;
-                        hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = current2;
-                        ip += repLength2;
-                        anchor = ip;
-                        continue;
-                    }
-                    break;
-            }   }
-
-            if (dictMode == ZSTD_noDict) {
-                while ( (ip <= ilimit)
-                     && ( (offset_2>0)
-                        & (MEM_read32(ip) == MEM_read32(ip - offset_2)) )) {
-                    /* store sequence */
-                    size_t const rLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4;
-                    U32 const tmpOff = offset_2; offset_2 = offset_1; offset_1 = tmpOff;  /* swap offset_2 <=> offset_1 */
-                    hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = (U32)(ip-base);
-                    hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = (U32)(ip-base);
-                    ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, rLength-MINMATCH);
-                    ip += rLength;
+            while (ip <= ilimit) {
+                U32 const current2 = (U32)(ip-base);
+                U32 const repIndex2 = current2 - offset_2;
+                const BYTE* repMatch2 = repIndex2 < prefixLowestIndex ?
+                        dictBase + repIndex2 - dictIndexDelta :
+                        base + repIndex2;
+                if ( ((U32)((prefixLowestIndex-1) - (U32)repIndex2) >= 3 /* intentional overflow */)
+                   && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
+                    const BYTE* const repEnd2 = repIndex2 < prefixLowestIndex ? dictEnd : iend;
+                    size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixLowest) + 4;
+                    U32 tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset;   /* swap offset_2 <=> offset_1 */
+                    ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
+                    hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = current2;
+                    hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = current2;
+                    ip += repLength2;
                     anchor = ip;
-                    continue;   /* faster when present ... (?) */
-        }   }   }
+                    continue;
+                }
+                break;
+            }
+        }
     }   /* while (ip < ilimit) */
 
     /* save reps for next block */
-    rep[0] = offset_1 ? offset_1 : offsetSaved;
-    rep[1] = offset_2 ? offset_2 : offsetSaved;
+    rep[0] = offset_1;
+    rep[1] = offset_2;
 
     /* Return the last literals size */
     return (size_t)(iend - anchor);
 }
 
+#define ZSTD_GEN_DFAST_FN(dictMode, mls)                                                                 \
+    static size_t ZSTD_compressBlock_doubleFast_##dictMode##_##mls(                                      \
+            ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],                          \
+            void const* src, size_t srcSize)                                                             \
+    {                                                                                                    \
+        return ZSTD_compressBlock_doubleFast_##dictMode##_generic(ms, seqStore, rep, src, srcSize, mls); \
+    }
+
+ZSTD_GEN_DFAST_FN(noDict, 4)
+ZSTD_GEN_DFAST_FN(noDict, 5)
+ZSTD_GEN_DFAST_FN(noDict, 6)
+ZSTD_GEN_DFAST_FN(noDict, 7)
+
+ZSTD_GEN_DFAST_FN(dictMatchState, 4)
+ZSTD_GEN_DFAST_FN(dictMatchState, 5)
+ZSTD_GEN_DFAST_FN(dictMatchState, 6)
+ZSTD_GEN_DFAST_FN(dictMatchState, 7)
+
 
 size_t ZSTD_compressBlock_doubleFast(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
@@ -325,13 +558,13 @@ size_t ZSTD_compressBlock_doubleFast(
     {
     default: /* includes case 3 */
     case 4 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 4, ZSTD_noDict);
+        return ZSTD_compressBlock_doubleFast_noDict_4(ms, seqStore, rep, src, srcSize);
     case 5 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 5, ZSTD_noDict);
+        return ZSTD_compressBlock_doubleFast_noDict_5(ms, seqStore, rep, src, srcSize);
     case 6 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 6, ZSTD_noDict);
+        return ZSTD_compressBlock_doubleFast_noDict_6(ms, seqStore, rep, src, srcSize);
     case 7 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 7, ZSTD_noDict);
+        return ZSTD_compressBlock_doubleFast_noDict_7(ms, seqStore, rep, src, srcSize);
     }
 }
 
@@ -345,13 +578,13 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState(
     {
     default: /* includes case 3 */
     case 4 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 4, ZSTD_dictMatchState);
+        return ZSTD_compressBlock_doubleFast_dictMatchState_4(ms, seqStore, rep, src, srcSize);
     case 5 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 5, ZSTD_dictMatchState);
+        return ZSTD_compressBlock_doubleFast_dictMatchState_5(ms, seqStore, rep, src, srcSize);
     case 6 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 6, ZSTD_dictMatchState);
+        return ZSTD_compressBlock_doubleFast_dictMatchState_6(ms, seqStore, rep, src, srcSize);
     case 7 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 7, ZSTD_dictMatchState);
+        return ZSTD_compressBlock_doubleFast_dictMatchState_7(ms, seqStore, rep, src, srcSize);
     }
 }
 
@@ -387,7 +620,7 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
 
     /* if extDict is invalidated due to maxDistance, switch to "regular" variant */
     if (prefixStartIndex == dictStartIndex)
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, mls, ZSTD_noDict);
+        return ZSTD_compressBlock_doubleFast(ms, seqStore, rep, src, srcSize);
 
     /* Search Loop */
     while (ip < ilimit) {  /* < instead of <=, because (ip+1) */
@@ -401,31 +634,31 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
         const BYTE* const matchLongBase = matchLongIndex < prefixStartIndex ? dictBase : base;
         const BYTE* matchLong = matchLongBase + matchLongIndex;
 
-        const U32 current = (U32)(ip-base);
-        const U32 repIndex = current + 1 - offset_1;   /* offset_1 expected <= current +1 */
+        const U32 curr = (U32)(ip-base);
+        const U32 repIndex = curr + 1 - offset_1;   /* offset_1 expected <= curr +1 */
         const BYTE* const repBase = repIndex < prefixStartIndex ? dictBase : base;
         const BYTE* const repMatch = repBase + repIndex;
         size_t mLength;
-        hashSmall[hSmall] = hashLong[hLong] = current;   /* update hash table */
+        hashSmall[hSmall] = hashLong[hLong] = curr;   /* update hash table */
 
         if ((((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow : ensure repIndex doesn't overlap dict + prefix */
-            & (repIndex > dictStartIndex))
+            & (offset_1 <= curr+1 - dictStartIndex)) /* note: we are searching at curr+1 */
           && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
             const BYTE* repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
             mLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixStart) + 4;
             ip++;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, 0, mLength-MINMATCH);
+            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength);
         } else {
             if ((matchLongIndex > dictStartIndex) && (MEM_read64(matchLong) == MEM_read64(ip))) {
                 const BYTE* const matchEnd = matchLongIndex < prefixStartIndex ? dictEnd : iend;
                 const BYTE* const lowMatchPtr = matchLongIndex < prefixStartIndex ? dictStart : prefixStart;
                 U32 offset;
                 mLength = ZSTD_count_2segments(ip+8, matchLong+8, iend, matchEnd, prefixStart) + 8;
-                offset = current - matchLongIndex;
+                offset = curr - matchLongIndex;
                 while (((ip>anchor) & (matchLong>lowMatchPtr)) && (ip[-1] == matchLong[-1])) { ip--; matchLong--; mLength++; }   /* catch up */
                 offset_2 = offset_1;
                 offset_1 = offset;
-                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH);
+                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
 
             } else if ((matchIndex > dictStartIndex) && (MEM_read32(match) == MEM_read32(ip))) {
                 size_t const h3 = ZSTD_hashPtr(ip+1, hBitsL, 8);
@@ -433,24 +666,24 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
                 const BYTE* const match3Base = matchIndex3 < prefixStartIndex ? dictBase : base;
                 const BYTE* match3 = match3Base + matchIndex3;
                 U32 offset;
-                hashLong[h3] = current + 1;
+                hashLong[h3] = curr + 1;
                 if ( (matchIndex3 > dictStartIndex) && (MEM_read64(match3) == MEM_read64(ip+1)) ) {
                     const BYTE* const matchEnd = matchIndex3 < prefixStartIndex ? dictEnd : iend;
                     const BYTE* const lowMatchPtr = matchIndex3 < prefixStartIndex ? dictStart : prefixStart;
                     mLength = ZSTD_count_2segments(ip+9, match3+8, iend, matchEnd, prefixStart) + 8;
                     ip++;
-                    offset = current+1 - matchIndex3;
+                    offset = curr+1 - matchIndex3;
                     while (((ip>anchor) & (match3>lowMatchPtr)) && (ip[-1] == match3[-1])) { ip--; match3--; mLength++; } /* catch up */
                 } else {
                     const BYTE* const matchEnd = matchIndex < prefixStartIndex ? dictEnd : iend;
                     const BYTE* const lowMatchPtr = matchIndex < prefixStartIndex ? dictStart : prefixStart;
                     mLength = ZSTD_count_2segments(ip+4, match+4, iend, matchEnd, prefixStart) + 4;
-                    offset = current - matchIndex;
+                    offset = curr - matchIndex;
                     while (((ip>anchor) & (match>lowMatchPtr)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; }   /* catch up */
                 }
                 offset_2 = offset_1;
                 offset_1 = offset;
-                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH);
+                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
 
             } else {
                 ip += ((ip-anchor) >> kSearchStrength) + 1;
@@ -464,7 +697,7 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
         if (ip <= ilimit) {
             /* Complementary insertion */
             /* done after iLimit test, as candidates could be > iend-8 */
-            {   U32 const indexToInsert = current+2;
+            {   U32 const indexToInsert = curr+2;
                 hashLong[ZSTD_hashPtr(base+indexToInsert, hBitsL, 8)] = indexToInsert;
                 hashLong[ZSTD_hashPtr(ip-2, hBitsL, 8)] = (U32)(ip-2-base);
                 hashSmall[ZSTD_hashPtr(base+indexToInsert, hBitsS, mls)] = indexToInsert;
@@ -477,12 +710,12 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
                 U32 const repIndex2 = current2 - offset_2;
                 const BYTE* repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2;
                 if ( (((U32)((prefixStartIndex-1) - repIndex2) >= 3)   /* intentional overflow : ensure repIndex2 doesn't overlap dict + prefix */
-                    & (repIndex2 > dictStartIndex))
+                    & (offset_2 <= current2 - dictStartIndex))
                   && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
                     const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
                     size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
                     U32 const tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset;   /* swap offset_2 <=> offset_1 */
-                    ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, repLength2-MINMATCH);
+                    ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
                     hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = current2;
                     hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = current2;
                     ip += repLength2;
@@ -500,6 +733,10 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
     return (size_t)(iend - anchor);
 }
 
+ZSTD_GEN_DFAST_FN(extDict, 4)
+ZSTD_GEN_DFAST_FN(extDict, 5)
+ZSTD_GEN_DFAST_FN(extDict, 6)
+ZSTD_GEN_DFAST_FN(extDict, 7)
 
 size_t ZSTD_compressBlock_doubleFast_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
@@ -510,12 +747,12 @@ size_t ZSTD_compressBlock_doubleFast_extDict(
     {
     default: /* includes case 3 */
     case 4 :
-        return ZSTD_compressBlock_doubleFast_extDict_generic(ms, seqStore, rep, src, srcSize, 4);
+        return ZSTD_compressBlock_doubleFast_extDict_4(ms, seqStore, rep, src, srcSize);
     case 5 :
-        return ZSTD_compressBlock_doubleFast_extDict_generic(ms, seqStore, rep, src, srcSize, 5);
+        return ZSTD_compressBlock_doubleFast_extDict_5(ms, seqStore, rep, src, srcSize);
     case 6 :
-        return ZSTD_compressBlock_doubleFast_extDict_generic(ms, seqStore, rep, src, srcSize, 6);
+        return ZSTD_compressBlock_doubleFast_extDict_6(ms, seqStore, rep, src, srcSize);
     case 7 :
-        return ZSTD_compressBlock_doubleFast_extDict_generic(ms, seqStore, rep, src, srcSize, 7);
+        return ZSTD_compressBlock_doubleFast_extDict_7(ms, seqStore, rep, src, srcSize);
     }
 }
diff --git a/contrib/zstd/zstd_double_fast.h b/contrib/zstd/zstd_double_fast.h
index 73a2002f3..e8ce20c70 100644
--- a/contrib/zstd/zstd_double_fast.h
+++ b/contrib/zstd/zstd_double_fast.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -19,7 +19,8 @@ extern "C" {
 #include "zstd_compress_internal.h"     /* ZSTD_CCtx, size_t */
 
 void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms,
-                              void const* end, ZSTD_dictTableLoadMethod_e dtlm);
+                              void const* end, ZSTD_dictTableLoadMethod_e dtlm,
+                              ZSTD_tableFillPurpose_e tfp);
 size_t ZSTD_compressBlock_doubleFast(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
diff --git a/contrib/zstd/zstd_errors.h b/contrib/zstd/zstd_errors.h
index 998398e7e..dc75eeeba 100644
--- a/contrib/zstd/zstd_errors.h
+++ b/contrib/zstd/zstd_errors.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -20,19 +20,31 @@ extern "C" {
 
 
 /* =====   ZSTDERRORLIB_API : control library symbols visibility   ===== */
-#ifndef ZSTDERRORLIB_VISIBILITY
-#  if defined(__GNUC__) && (__GNUC__ >= 4)
-#    define ZSTDERRORLIB_VISIBILITY __attribute__ ((visibility ("default")))
+#ifndef ZSTDERRORLIB_VISIBLE
+   /* Backwards compatibility with old macro name */
+#  ifdef ZSTDERRORLIB_VISIBILITY
+#    define ZSTDERRORLIB_VISIBLE ZSTDERRORLIB_VISIBILITY
+#  elif defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
+#    define ZSTDERRORLIB_VISIBLE __attribute__ ((visibility ("default")))
 #  else
-#    define ZSTDERRORLIB_VISIBILITY
+#    define ZSTDERRORLIB_VISIBLE
 #  endif
 #endif
+
+#ifndef ZSTDERRORLIB_HIDDEN
+#  if defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
+#    define ZSTDERRORLIB_HIDDEN __attribute__ ((visibility ("hidden")))
+#  else
+#    define ZSTDERRORLIB_HIDDEN
+#  endif
+#endif
+
 #if defined(ZSTD_DLL_EXPORT) && (ZSTD_DLL_EXPORT==1)
-#  define ZSTDERRORLIB_API __declspec(dllexport) ZSTDERRORLIB_VISIBILITY
+#  define ZSTDERRORLIB_API __declspec(dllexport) ZSTDERRORLIB_VISIBLE
 #elif defined(ZSTD_DLL_IMPORT) && (ZSTD_DLL_IMPORT==1)
-#  define ZSTDERRORLIB_API __declspec(dllimport) ZSTDERRORLIB_VISIBILITY /* It isn't required but allows to generate better code, saving a function pointer load from the IAT and an indirect jump.*/
+#  define ZSTDERRORLIB_API __declspec(dllimport) ZSTDERRORLIB_VISIBLE /* It isn't required but allows to generate better code, saving a function pointer load from the IAT and an indirect jump.*/
 #else
-#  define ZSTDERRORLIB_API ZSTDERRORLIB_VISIBILITY
+#  define ZSTDERRORLIB_API ZSTDERRORLIB_VISIBLE
 #endif
 
 /*-*********************************************
@@ -58,14 +70,17 @@ typedef enum {
   ZSTD_error_frameParameter_windowTooLarge = 16,
   ZSTD_error_corruption_detected = 20,
   ZSTD_error_checksum_wrong      = 22,
+  ZSTD_error_literals_headerWrong = 24,
   ZSTD_error_dictionary_corrupted      = 30,
   ZSTD_error_dictionary_wrong          = 32,
   ZSTD_error_dictionaryCreation_failed = 34,
   ZSTD_error_parameter_unsupported   = 40,
+  ZSTD_error_parameter_combination_unsupported = 41,
   ZSTD_error_parameter_outOfBound    = 42,
   ZSTD_error_tableLog_tooLarge       = 44,
   ZSTD_error_maxSymbolValue_tooLarge = 46,
   ZSTD_error_maxSymbolValue_tooSmall = 48,
+  ZSTD_error_stabilityCondition_notRespected = 50,
   ZSTD_error_stage_wrong       = 60,
   ZSTD_error_init_missing      = 62,
   ZSTD_error_memory_allocation = 64,
@@ -73,10 +88,15 @@ typedef enum {
   ZSTD_error_dstSize_tooSmall = 70,
   ZSTD_error_srcSize_wrong    = 72,
   ZSTD_error_dstBuffer_null   = 74,
+  ZSTD_error_noForwardProgress_destFull = 80,
+  ZSTD_error_noForwardProgress_inputEmpty = 82,
   /* following error codes are __NOT STABLE__, they can be removed or changed in future versions */
   ZSTD_error_frameIndex_tooLarge = 100,
   ZSTD_error_seekableIO          = 102,
   ZSTD_error_dstBuffer_wrong     = 104,
+  ZSTD_error_srcBuffer_wrong     = 105,
+  ZSTD_error_sequenceProducer_failed = 106,
+  ZSTD_error_externalSequences_invalid = 107,
   ZSTD_error_maxCode = 120  /* never EVER use this value directly, it can change in future versions! Use ZSTD_isError() instead */
 } ZSTD_ErrorCode;
 
diff --git a/contrib/zstd/zstd_fast.c b/contrib/zstd/zstd_fast.c
index 85a3a7a91..5f2c6a2ed 100644
--- a/contrib/zstd/zstd_fast.c
+++ b/contrib/zstd/zstd_fast.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -11,8 +11,42 @@
 #include "zstd_compress_internal.h"  /* ZSTD_hashPtr, ZSTD_count, ZSTD_storeSeq */
 #include "zstd_fast.h"
 
+static void ZSTD_fillHashTableForCDict(ZSTD_matchState_t* ms,
+                        const void* const end,
+                        ZSTD_dictTableLoadMethod_e dtlm)
+{
+    const ZSTD_compressionParameters* const cParams = &ms->cParams;
+    U32* const hashTable = ms->hashTable;
+    U32  const hBits = cParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS;
+    U32  const mls = cParams->minMatch;
+    const BYTE* const base = ms->window.base;
+    const BYTE* ip = base + ms->nextToUpdate;
+    const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE;
+    const U32 fastHashFillStep = 3;
 
-void ZSTD_fillHashTable(ZSTD_matchState_t* ms,
+    /* Currently, we always use ZSTD_dtlm_full for filling CDict tables.
+     * Feel free to remove this assert if there's a good reason! */
+    assert(dtlm == ZSTD_dtlm_full);
+
+    /* Always insert every fastHashFillStep position into the hash table.
+     * Insert the other positions if their hash entry is empty.
+     */
+    for ( ; ip + fastHashFillStep < iend + 2; ip += fastHashFillStep) {
+        U32 const curr = (U32)(ip - base);
+        {   size_t const hashAndTag = ZSTD_hashPtr(ip, hBits, mls);
+            ZSTD_writeTaggedIndex(hashTable, hashAndTag, curr);   }
+
+        if (dtlm == ZSTD_dtlm_fast) continue;
+        /* Only load extra positions for ZSTD_dtlm_full */
+        {   U32 p;
+            for (p = 1; p < fastHashFillStep; ++p) {
+                size_t const hashAndTag = ZSTD_hashPtr(ip + p, hBits, mls);
+                if (hashTable[hashAndTag >> ZSTD_SHORT_CACHE_TAG_BITS] == 0) {  /* not yet filled */
+                    ZSTD_writeTaggedIndex(hashTable, hashAndTag, curr + p);
+                }   }   }   }
+}
+
+static void ZSTD_fillHashTableForCCtx(ZSTD_matchState_t* ms,
                         const void* const end,
                         ZSTD_dictTableLoadMethod_e dtlm)
 {
@@ -25,163 +59,365 @@ void ZSTD_fillHashTable(ZSTD_matchState_t* ms,
     const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE;
     const U32 fastHashFillStep = 3;
 
+    /* Currently, we always use ZSTD_dtlm_fast for filling CCtx tables.
+     * Feel free to remove this assert if there's a good reason! */
+    assert(dtlm == ZSTD_dtlm_fast);
+
     /* Always insert every fastHashFillStep position into the hash table.
      * Insert the other positions if their hash entry is empty.
      */
     for ( ; ip + fastHashFillStep < iend + 2; ip += fastHashFillStep) {
-        U32 const current = (U32)(ip - base);
+        U32 const curr = (U32)(ip - base);
         size_t const hash0 = ZSTD_hashPtr(ip, hBits, mls);
-        hashTable[hash0] = current;
+        hashTable[hash0] = curr;
         if (dtlm == ZSTD_dtlm_fast) continue;
         /* Only load extra positions for ZSTD_dtlm_full */
         {   U32 p;
             for (p = 1; p < fastHashFillStep; ++p) {
                 size_t const hash = ZSTD_hashPtr(ip + p, hBits, mls);
                 if (hashTable[hash] == 0) {  /* not yet filled */
-                    hashTable[hash] = current + p;
+                    hashTable[hash] = curr + p;
     }   }   }   }
 }
 
+void ZSTD_fillHashTable(ZSTD_matchState_t* ms,
+                        const void* const end,
+                        ZSTD_dictTableLoadMethod_e dtlm,
+                        ZSTD_tableFillPurpose_e tfp)
+{
+    if (tfp == ZSTD_tfp_forCDict) {
+        ZSTD_fillHashTableForCDict(ms, end, dtlm);
+    } else {
+        ZSTD_fillHashTableForCCtx(ms, end, dtlm);
+    }
+}
+
 
+/**
+ * If you squint hard enough (and ignore repcodes), the search operation at any
+ * given position is broken into 4 stages:
+ *
+ * 1. Hash   (map position to hash value via input read)
+ * 2. Lookup (map hash val to index via hashtable read)
+ * 3. Load   (map index to value at that position via input read)
+ * 4. Compare
+ *
+ * Each of these steps involves a memory read at an address which is computed
+ * from the previous step. This means these steps must be sequenced and their
+ * latencies are cumulative.
+ *
+ * Rather than do 1->2->3->4 sequentially for a single position before moving
+ * onto the next, this implementation interleaves these operations across the
+ * next few positions:
+ *
+ * R = Repcode Read & Compare
+ * H = Hash
+ * T = Table Lookup
+ * M = Match Read & Compare
+ *
+ * Pos | Time -->
+ * ----+-------------------
+ * N   | ... M
+ * N+1 | ...   TM
+ * N+2 |    R H   T M
+ * N+3 |         H    TM
+ * N+4 |           R H   T M
+ * N+5 |                H   ...
+ * N+6 |                  R ...
+ *
+ * This is very much analogous to the pipelining of execution in a CPU. And just
+ * like a CPU, we have to dump the pipeline when we find a match (i.e., take a
+ * branch).
+ *
+ * When this happens, we throw away our current state, and do the following prep
+ * to re-enter the loop:
+ *
+ * Pos | Time -->
+ * ----+-------------------
+ * N   | H T
+ * N+1 |  H
+ *
+ * This is also the work we do at the beginning to enter the loop initially.
+ */
 FORCE_INLINE_TEMPLATE size_t
-ZSTD_compressBlock_fast_generic(
+ZSTD_compressBlock_fast_noDict_generic(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize,
-        U32 const mls)
+        U32 const mls, U32 const hasStep)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
     U32* const hashTable = ms->hashTable;
     U32 const hlog = cParams->hashLog;
     /* support stepSize of 0 */
-    size_t const stepSize = cParams->targetLength + !(cParams->targetLength) + 1;
+    size_t const stepSize = hasStep ? (cParams->targetLength + !(cParams->targetLength) + 1) : 2;
     const BYTE* const base = ms->window.base;
     const BYTE* const istart = (const BYTE*)src;
-    /* We check ip0 (ip + 0) and ip1 (ip + 1) each loop */
-    const BYTE* ip0 = istart;
-    const BYTE* ip1;
-    const BYTE* anchor = istart;
     const U32   endIndex = (U32)((size_t)(istart - base) + srcSize);
     const U32   prefixStartIndex = ZSTD_getLowestPrefixIndex(ms, endIndex, cParams->windowLog);
     const BYTE* const prefixStart = base + prefixStartIndex;
     const BYTE* const iend = istart + srcSize;
     const BYTE* const ilimit = iend - HASH_READ_SIZE;
-    U32 offset_1=rep[0], offset_2=rep[1];
-    U32 offsetSaved = 0;
 
-    /* init */
+    const BYTE* anchor = istart;
+    const BYTE* ip0 = istart;
+    const BYTE* ip1;
+    const BYTE* ip2;
+    const BYTE* ip3;
+    U32 current0;
+
+    U32 rep_offset1 = rep[0];
+    U32 rep_offset2 = rep[1];
+    U32 offsetSaved1 = 0, offsetSaved2 = 0;
+
+    size_t hash0; /* hash for ip0 */
+    size_t hash1; /* hash for ip1 */
+    U32 idx; /* match idx for ip0 */
+    U32 mval; /* src value at match idx */
+
+    U32 offcode;
+    const BYTE* match0;
+    size_t mLength;
+
+    /* ip0 and ip1 are always adjacent. The targetLength skipping and
+     * uncompressibility acceleration is applied to every other position,
+     * matching the behavior of #1562. step therefore represents the gap
+     * between pairs of positions, from ip0 to ip2 or ip1 to ip3. */
+    size_t step;
+    const BYTE* nextStep;
+    const size_t kStepIncr = (1 << (kSearchStrength - 1));
+
     DEBUGLOG(5, "ZSTD_compressBlock_fast_generic");
     ip0 += (ip0 == prefixStart);
+    {   U32 const curr = (U32)(ip0 - base);
+        U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, curr, cParams->windowLog);
+        U32 const maxRep = curr - windowLow;
+        if (rep_offset2 > maxRep) offsetSaved2 = rep_offset2, rep_offset2 = 0;
+        if (rep_offset1 > maxRep) offsetSaved1 = rep_offset1, rep_offset1 = 0;
+    }
+
+    /* start each op */
+_start: /* Requires: ip0 */
+
+    step = stepSize;
+    nextStep = ip0 + kStepIncr;
+
+    /* calculate positions, ip0 - anchor == 0, so we skip step calc */
     ip1 = ip0 + 1;
-    {   U32 const current = (U32)(ip0 - base);
-        U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, current, cParams->windowLog);
-        U32 const maxRep = current - windowLow;
-        if (offset_2 > maxRep) offsetSaved = offset_2, offset_2 = 0;
-        if (offset_1 > maxRep) offsetSaved = offset_1, offset_1 = 0;
+    ip2 = ip0 + step;
+    ip3 = ip2 + 1;
+
+    if (ip3 >= ilimit) {
+        goto _cleanup;
     }
 
-    /* Main Search Loop */
-#ifdef __INTEL_COMPILER
-    /* From intel 'The vector pragma indicates that the loop should be
-     * vectorized if it is legal to do so'. Can be used together with
-     * #pragma ivdep (but have opted to exclude that because intel
-     * warns against using it).*/
-    #pragma vector always
-#endif
-    while (ip1 < ilimit) {   /* < instead of <=, because check at ip0+2 */
-        size_t mLength;
-        BYTE const* ip2 = ip0 + 2;
-        size_t const h0 = ZSTD_hashPtr(ip0, hlog, mls);
-        U32 const val0 = MEM_read32(ip0);
-        size_t const h1 = ZSTD_hashPtr(ip1, hlog, mls);
-        U32 const val1 = MEM_read32(ip1);
-        U32 const current0 = (U32)(ip0-base);
-        U32 const current1 = (U32)(ip1-base);
-        U32 const matchIndex0 = hashTable[h0];
-        U32 const matchIndex1 = hashTable[h1];
-        BYTE const* repMatch = ip2 - offset_1;
-        const BYTE* match0 = base + matchIndex0;
-        const BYTE* match1 = base + matchIndex1;
-        U32 offcode;
-
-#if defined(__aarch64__)
-        PREFETCH_L1(ip0+256);
-#endif
-
-        hashTable[h0] = current0;   /* update hash table */
-        hashTable[h1] = current1;   /* update hash table */
-
-        assert(ip0 + 1 == ip1);
-
-        if ((offset_1 > 0) & (MEM_read32(repMatch) == MEM_read32(ip2))) {
-            mLength = (ip2[-1] == repMatch[-1]) ? 1 : 0;
-            ip0 = ip2 - mLength;
-            match0 = repMatch - mLength;
+    hash0 = ZSTD_hashPtr(ip0, hlog, mls);
+    hash1 = ZSTD_hashPtr(ip1, hlog, mls);
+
+    idx = hashTable[hash0];
+
+    do {
+        /* load repcode match for ip[2]*/
+        const U32 rval = MEM_read32(ip2 - rep_offset1);
+
+        /* write back hash table entry */
+        current0 = (U32)(ip0 - base);
+        hashTable[hash0] = current0;
+
+        /* check repcode at ip[2] */
+        if ((MEM_read32(ip2) == rval) & (rep_offset1 > 0)) {
+            ip0 = ip2;
+            match0 = ip0 - rep_offset1;
+            mLength = ip0[-1] == match0[-1];
+            ip0 -= mLength;
+            match0 -= mLength;
+            offcode = REPCODE1_TO_OFFBASE;
             mLength += 4;
-            offcode = 0;
+
+            /* First write next hash table entry; we've already calculated it.
+             * This write is known to be safe because the ip1 is before the
+             * repcode (ip2). */
+            hashTable[hash1] = (U32)(ip1 - base);
+
             goto _match;
         }
-        if ((matchIndex0 > prefixStartIndex) && MEM_read32(match0) == val0) {
-            /* found a regular match */
-            goto _offset;
+
+        /* load match for ip[0] */
+        if (idx >= prefixStartIndex) {
+            mval = MEM_read32(base + idx);
+        } else {
+            mval = MEM_read32(ip0) ^ 1; /* guaranteed to not match. */
         }
-        if ((matchIndex1 > prefixStartIndex) && MEM_read32(match1) == val1) {
-            /* found a regular match after one literal */
-            ip0 = ip1;
-            match0 = match1;
+
+        /* check match at ip[0] */
+        if (MEM_read32(ip0) == mval) {
+            /* found a match! */
+
+            /* First write next hash table entry; we've already calculated it.
+             * This write is known to be safe because the ip1 == ip0 + 1, so
+             * we know we will resume searching after ip1 */
+            hashTable[hash1] = (U32)(ip1 - base);
+
             goto _offset;
         }
-        {   size_t const step = ((size_t)(ip0-anchor) >> (kSearchStrength - 1)) + stepSize;
-            assert(step >= 2);
-            ip0 += step;
-            ip1 += step;
-            continue;
+
+        /* lookup ip[1] */
+        idx = hashTable[hash1];
+
+        /* hash ip[2] */
+        hash0 = hash1;
+        hash1 = ZSTD_hashPtr(ip2, hlog, mls);
+
+        /* advance to next positions */
+        ip0 = ip1;
+        ip1 = ip2;
+        ip2 = ip3;
+
+        /* write back hash table entry */
+        current0 = (U32)(ip0 - base);
+        hashTable[hash0] = current0;
+
+        /* load match for ip[0] */
+        if (idx >= prefixStartIndex) {
+            mval = MEM_read32(base + idx);
+        } else {
+            mval = MEM_read32(ip0) ^ 1; /* guaranteed to not match. */
         }
-_offset: /* Requires: ip0, match0 */
-        /* Compute the offset code */
-        offset_2 = offset_1;
-        offset_1 = (U32)(ip0-match0);
-        offcode = offset_1 + ZSTD_REP_MOVE;
-        mLength = 4;
-        /* Count the backwards match length */
-        while (((ip0>anchor) & (match0>prefixStart))
-             && (ip0[-1] == match0[-1])) { ip0--; match0--; mLength++; } /* catch up */
 
-_match: /* Requires: ip0, match0, offcode */
-        /* Count the forward length */
-        mLength += ZSTD_count(ip0+mLength, match0+mLength, iend);
-        ZSTD_storeSeq(seqStore, (size_t)(ip0-anchor), anchor, iend, offcode, mLength-MINMATCH);
-        /* match found */
-        ip0 += mLength;
-        anchor = ip0;
+        /* check match at ip[0] */
+        if (MEM_read32(ip0) == mval) {
+            /* found a match! */
+
+            /* first write next hash table entry; we've already calculated it */
+            if (step <= 4) {
+                /* We need to avoid writing an index into the hash table >= the
+                 * position at which we will pick up our searching after we've
+                 * taken this match.
+                 *
+                 * The minimum possible match has length 4, so the earliest ip0
+                 * can be after we take this match will be the current ip0 + 4.
+                 * ip1 is ip0 + step - 1. If ip1 is >= ip0 + 4, we can't safely
+                 * write this position.
+                 */
+                hashTable[hash1] = (U32)(ip1 - base);
+            }
 
-        if (ip0 <= ilimit) {
-            /* Fill Table */
-            assert(base+current0+2 > istart);  /* check base overflow */
-            hashTable[ZSTD_hashPtr(base+current0+2, hlog, mls)] = current0+2;  /* here because current+2 could be > iend-8 */
-            hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base);
+            goto _offset;
+        }
 
-            if (offset_2 > 0) { /* offset_2==0 means offset_2 is invalidated */
-                while ( (ip0 <= ilimit) && (MEM_read32(ip0) == MEM_read32(ip0 - offset_2)) ) {
-                    /* store sequence */
-                    size_t const rLength = ZSTD_count(ip0+4, ip0+4-offset_2, iend) + 4;
-                    { U32 const tmpOff = offset_2; offset_2 = offset_1; offset_1 = tmpOff; } /* swap offset_2 <=> offset_1 */
-                    hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = (U32)(ip0-base);
-                    ip0 += rLength;
-                    ZSTD_storeSeq(seqStore, 0 /*litLen*/, anchor, iend, 0 /*offCode*/, rLength-MINMATCH);
-                    anchor = ip0;
-                    continue;   /* faster when present (confirmed on gcc-8) ... (?) */
-        }   }   }
-        ip1 = ip0 + 1;
-    }
+        /* lookup ip[1] */
+        idx = hashTable[hash1];
+
+        /* hash ip[2] */
+        hash0 = hash1;
+        hash1 = ZSTD_hashPtr(ip2, hlog, mls);
+
+        /* advance to next positions */
+        ip0 = ip1;
+        ip1 = ip2;
+        ip2 = ip0 + step;
+        ip3 = ip1 + step;
+
+        /* calculate step */
+        if (ip2 >= nextStep) {
+            step++;
+            PREFETCH_L1(ip1 + 64);
+            PREFETCH_L1(ip1 + 128);
+            nextStep += kStepIncr;
+        }
+    } while (ip3 < ilimit);
+
+_cleanup:
+    /* Note that there are probably still a couple positions we could search.
+     * However, it seems to be a meaningful performance hit to try to search
+     * them. So let's not. */
+
+    /* When the repcodes are outside of the prefix, we set them to zero before the loop.
+     * When the offsets are still zero, we need to restore them after the block to have a correct
+     * repcode history. If only one offset was invalid, it is easy. The tricky case is when both
+     * offsets were invalid. We need to figure out which offset to refill with.
+     *     - If both offsets are zero they are in the same order.
+     *     - If both offsets are non-zero, we won't restore the offsets from `offsetSaved[12]`.
+     *     - If only one is zero, we need to decide which offset to restore.
+     *         - If rep_offset1 is non-zero, then rep_offset2 must be offsetSaved1.
+     *         - It is impossible for rep_offset2 to be non-zero.
+     *
+     * So if rep_offset1 started invalid (offsetSaved1 != 0) and became valid (rep_offset1 != 0), then
+     * set rep[0] = rep_offset1 and rep[1] = offsetSaved1.
+     */
+    offsetSaved2 = ((offsetSaved1 != 0) && (rep_offset1 != 0)) ? offsetSaved1 : offsetSaved2;
 
     /* save reps for next block */
-    rep[0] = offset_1 ? offset_1 : offsetSaved;
-    rep[1] = offset_2 ? offset_2 : offsetSaved;
+    rep[0] = rep_offset1 ? rep_offset1 : offsetSaved1;
+    rep[1] = rep_offset2 ? rep_offset2 : offsetSaved2;
 
     /* Return the last literals size */
     return (size_t)(iend - anchor);
+
+_offset: /* Requires: ip0, idx */
+
+    /* Compute the offset code. */
+    match0 = base + idx;
+    rep_offset2 = rep_offset1;
+    rep_offset1 = (U32)(ip0-match0);
+    offcode = OFFSET_TO_OFFBASE(rep_offset1);
+    mLength = 4;
+
+    /* Count the backwards match length. */
+    while (((ip0>anchor) & (match0>prefixStart)) && (ip0[-1] == match0[-1])) {
+        ip0--;
+        match0--;
+        mLength++;
+    }
+
+_match: /* Requires: ip0, match0, offcode */
+
+    /* Count the forward length. */
+    mLength += ZSTD_count(ip0 + mLength, match0 + mLength, iend);
+
+    ZSTD_storeSeq(seqStore, (size_t)(ip0 - anchor), anchor, iend, offcode, mLength);
+
+    ip0 += mLength;
+    anchor = ip0;
+
+    /* Fill table and check for immediate repcode. */
+    if (ip0 <= ilimit) {
+        /* Fill Table */
+        assert(base+current0+2 > istart);  /* check base overflow */
+        hashTable[ZSTD_hashPtr(base+current0+2, hlog, mls)] = current0+2;  /* here because current+2 could be > iend-8 */
+        hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base);
+
+        if (rep_offset2 > 0) { /* rep_offset2==0 means rep_offset2 is invalidated */
+            while ( (ip0 <= ilimit) && (MEM_read32(ip0) == MEM_read32(ip0 - rep_offset2)) ) {
+                /* store sequence */
+                size_t const rLength = ZSTD_count(ip0+4, ip0+4-rep_offset2, iend) + 4;
+                { U32 const tmpOff = rep_offset2; rep_offset2 = rep_offset1; rep_offset1 = tmpOff; } /* swap rep_offset2 <=> rep_offset1 */
+                hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = (U32)(ip0-base);
+                ip0 += rLength;
+                ZSTD_storeSeq(seqStore, 0 /*litLen*/, anchor, iend, REPCODE1_TO_OFFBASE, rLength);
+                anchor = ip0;
+                continue;   /* faster when present (confirmed on gcc-8) ... (?) */
+    }   }   }
+
+    goto _start;
 }
 
+#define ZSTD_GEN_FAST_FN(dictMode, mls, step)                                                            \
+    static size_t ZSTD_compressBlock_fast_##dictMode##_##mls##_##step(                                      \
+            ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],                    \
+            void const* src, size_t srcSize)                                                       \
+    {                                                                                              \
+        return ZSTD_compressBlock_fast_##dictMode##_generic(ms, seqStore, rep, src, srcSize, mls, step); \
+    }
+
+ZSTD_GEN_FAST_FN(noDict, 4, 1)
+ZSTD_GEN_FAST_FN(noDict, 5, 1)
+ZSTD_GEN_FAST_FN(noDict, 6, 1)
+ZSTD_GEN_FAST_FN(noDict, 7, 1)
+
+ZSTD_GEN_FAST_FN(noDict, 4, 0)
+ZSTD_GEN_FAST_FN(noDict, 5, 0)
+ZSTD_GEN_FAST_FN(noDict, 6, 0)
+ZSTD_GEN_FAST_FN(noDict, 7, 0)
 
 size_t ZSTD_compressBlock_fast(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
@@ -189,24 +425,40 @@ size_t ZSTD_compressBlock_fast(
 {
     U32 const mls = ms->cParams.minMatch;
     assert(ms->dictMatchState == NULL);
-    switch(mls)
-    {
-    default: /* includes case 3 */
-    case 4 :
-        return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, 4);
-    case 5 :
-        return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, 5);
-    case 6 :
-        return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, 6);
-    case 7 :
-        return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, 7);
+    if (ms->cParams.targetLength > 1) {
+        switch(mls)
+        {
+        default: /* includes case 3 */
+        case 4 :
+            return ZSTD_compressBlock_fast_noDict_4_1(ms, seqStore, rep, src, srcSize);
+        case 5 :
+            return ZSTD_compressBlock_fast_noDict_5_1(ms, seqStore, rep, src, srcSize);
+        case 6 :
+            return ZSTD_compressBlock_fast_noDict_6_1(ms, seqStore, rep, src, srcSize);
+        case 7 :
+            return ZSTD_compressBlock_fast_noDict_7_1(ms, seqStore, rep, src, srcSize);
+        }
+    } else {
+        switch(mls)
+        {
+        default: /* includes case 3 */
+        case 4 :
+            return ZSTD_compressBlock_fast_noDict_4_0(ms, seqStore, rep, src, srcSize);
+        case 5 :
+            return ZSTD_compressBlock_fast_noDict_5_0(ms, seqStore, rep, src, srcSize);
+        case 6 :
+            return ZSTD_compressBlock_fast_noDict_6_0(ms, seqStore, rep, src, srcSize);
+        case 7 :
+            return ZSTD_compressBlock_fast_noDict_7_0(ms, seqStore, rep, src, srcSize);
+        }
+
     }
 }
 
 FORCE_INLINE_TEMPLATE
 size_t ZSTD_compressBlock_fast_dictMatchState_generic(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize, U32 const mls)
+        void const* src, size_t srcSize, U32 const mls, U32 const hasStep)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
     U32* const hashTable = ms->hashTable;
@@ -215,14 +467,14 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic(
     U32 const stepSize = cParams->targetLength + !(cParams->targetLength);
     const BYTE* const base = ms->window.base;
     const BYTE* const istart = (const BYTE*)src;
-    const BYTE* ip = istart;
+    const BYTE* ip0 = istart;
+    const BYTE* ip1 = ip0 + stepSize; /* we assert below that stepSize >= 1 */
     const BYTE* anchor = istart;
     const U32   prefixStartIndex = ms->window.dictLimit;
     const BYTE* const prefixStart = base + prefixStartIndex;
     const BYTE* const iend = istart + srcSize;
     const BYTE* const ilimit = iend - HASH_READ_SIZE;
     U32 offset_1=rep[0], offset_2=rep[1];
-    U32 offsetSaved = 0;
 
     const ZSTD_matchState_t* const dms = ms->dictMatchState;
     const ZSTD_compressionParameters* const dictCParams = &dms->cParams ;
@@ -232,125 +484,182 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic(
     const BYTE* const dictStart    = dictBase + dictStartIndex;
     const BYTE* const dictEnd      = dms->window.nextSrc;
     const U32 dictIndexDelta       = prefixStartIndex - (U32)(dictEnd - dictBase);
-    const U32 dictAndPrefixLength  = (U32)(ip - prefixStart + dictEnd - dictStart);
-    const U32 dictHLog             = dictCParams->hashLog;
+    const U32 dictAndPrefixLength  = (U32)(istart - prefixStart + dictEnd - dictStart);
+    const U32 dictHBits            = dictCParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS;
 
     /* if a dictionary is still attached, it necessarily means that
      * it is within window size. So we just check it. */
     const U32 maxDistance = 1U << cParams->windowLog;
-    const U32 endIndex = (U32)((size_t)(ip - base) + srcSize);
+    const U32 endIndex = (U32)((size_t)(istart - base) + srcSize);
     assert(endIndex - prefixStartIndex <= maxDistance);
     (void)maxDistance; (void)endIndex;   /* these variables are not used when assert() is disabled */
 
-    /* ensure there will be no no underflow
+    (void)hasStep; /* not currently specialized on whether it's accelerated */
+
+    /* ensure there will be no underflow
      * when translating a dict index into a local index */
     assert(prefixStartIndex >= (U32)(dictEnd - dictBase));
 
+    if (ms->prefetchCDictTables) {
+        size_t const hashTableBytes = (((size_t)1) << dictCParams->hashLog) * sizeof(U32);
+        PREFETCH_AREA(dictHashTable, hashTableBytes)
+    }
+
     /* init */
     DEBUGLOG(5, "ZSTD_compressBlock_fast_dictMatchState_generic");
-    ip += (dictAndPrefixLength == 0);
+    ip0 += (dictAndPrefixLength == 0);
     /* dictMatchState repCode checks don't currently handle repCode == 0
      * disabling. */
     assert(offset_1 <= dictAndPrefixLength);
     assert(offset_2 <= dictAndPrefixLength);
 
-    /* Main Search Loop */
-    while (ip < ilimit) {   /* < instead of <=, because repcode check at (ip+1) */
+    /* Outer search loop */
+    assert(stepSize >= 1);
+    while (ip1 <= ilimit) {   /* repcode check at (ip0 + 1) is safe because ip0 < ip1 */
         size_t mLength;
-        size_t const h = ZSTD_hashPtr(ip, hlog, mls);
-        U32 const current = (U32)(ip-base);
-        U32 const matchIndex = hashTable[h];
-        const BYTE* match = base + matchIndex;
-        const U32 repIndex = current + 1 - offset_1;
-        const BYTE* repMatch = (repIndex < prefixStartIndex) ?
-                               dictBase + (repIndex - dictIndexDelta) :
-                               base + repIndex;
-        hashTable[h] = current;   /* update hash table */
-
-        if ( ((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow : ensure repIndex isn't overlapping dict + prefix */
-          && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
-            const BYTE* const repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
-            mLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixStart) + 4;
-            ip++;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, 0, mLength-MINMATCH);
-        } else if ( (matchIndex <= prefixStartIndex) ) {
-            size_t const dictHash = ZSTD_hashPtr(ip, dictHLog, mls);
-            U32 const dictMatchIndex = dictHashTable[dictHash];
-            const BYTE* dictMatch = dictBase + dictMatchIndex;
-            if (dictMatchIndex <= dictStartIndex ||
-                MEM_read32(dictMatch) != MEM_read32(ip)) {
-                assert(stepSize >= 1);
-                ip += ((ip-anchor) >> kSearchStrength) + stepSize;
-                continue;
-            } else {
-                /* found a dict match */
-                U32 const offset = (U32)(current-dictMatchIndex-dictIndexDelta);
-                mLength = ZSTD_count_2segments(ip+4, dictMatch+4, iend, dictEnd, prefixStart) + 4;
-                while (((ip>anchor) & (dictMatch>dictStart))
-                     && (ip[-1] == dictMatch[-1])) {
-                    ip--; dictMatch--; mLength++;
+        size_t hash0 = ZSTD_hashPtr(ip0, hlog, mls);
+
+        size_t const dictHashAndTag0 = ZSTD_hashPtr(ip0, dictHBits, mls);
+        U32 dictMatchIndexAndTag = dictHashTable[dictHashAndTag0 >> ZSTD_SHORT_CACHE_TAG_BITS];
+        int dictTagsMatch = ZSTD_comparePackedTags(dictMatchIndexAndTag, dictHashAndTag0);
+
+        U32 matchIndex = hashTable[hash0];
+        U32 curr = (U32)(ip0 - base);
+        size_t step = stepSize;
+        const size_t kStepIncr = 1 << kSearchStrength;
+        const BYTE* nextStep = ip0 + kStepIncr;
+
+        /* Inner search loop */
+        while (1) {
+            const BYTE* match = base + matchIndex;
+            const U32 repIndex = curr + 1 - offset_1;
+            const BYTE* repMatch = (repIndex < prefixStartIndex) ?
+                                   dictBase + (repIndex - dictIndexDelta) :
+                                   base + repIndex;
+            const size_t hash1 = ZSTD_hashPtr(ip1, hlog, mls);
+            size_t const dictHashAndTag1 = ZSTD_hashPtr(ip1, dictHBits, mls);
+            hashTable[hash0] = curr;   /* update hash table */
+
+            if (((U32) ((prefixStartIndex - 1) - repIndex) >=
+                 3) /* intentional underflow : ensure repIndex isn't overlapping dict + prefix */
+                && (MEM_read32(repMatch) == MEM_read32(ip0 + 1))) {
+                const BYTE* const repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
+                mLength = ZSTD_count_2segments(ip0 + 1 + 4, repMatch + 4, iend, repMatchEnd, prefixStart) + 4;
+                ip0++;
+                ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength);
+                break;
+            }
+
+            if (dictTagsMatch) {
+                /* Found a possible dict match */
+                const U32 dictMatchIndex = dictMatchIndexAndTag >> ZSTD_SHORT_CACHE_TAG_BITS;
+                const BYTE* dictMatch = dictBase + dictMatchIndex;
+                if (dictMatchIndex > dictStartIndex &&
+                    MEM_read32(dictMatch) == MEM_read32(ip0)) {
+                    /* To replicate extDict parse behavior, we only use dict matches when the normal matchIndex is invalid */
+                    if (matchIndex <= prefixStartIndex) {
+                        U32 const offset = (U32) (curr - dictMatchIndex - dictIndexDelta);
+                        mLength = ZSTD_count_2segments(ip0 + 4, dictMatch + 4, iend, dictEnd, prefixStart) + 4;
+                        while (((ip0 > anchor) & (dictMatch > dictStart))
+                            && (ip0[-1] == dictMatch[-1])) {
+                            ip0--;
+                            dictMatch--;
+                            mLength++;
+                        } /* catch up */
+                        offset_2 = offset_1;
+                        offset_1 = offset;
+                        ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
+                        break;
+                    }
+                }
+            }
+
+            if (matchIndex > prefixStartIndex && MEM_read32(match) == MEM_read32(ip0)) {
+                /* found a regular match */
+                U32 const offset = (U32) (ip0 - match);
+                mLength = ZSTD_count(ip0 + 4, match + 4, iend) + 4;
+                while (((ip0 > anchor) & (match > prefixStart))
+                       && (ip0[-1] == match[-1])) {
+                    ip0--;
+                    match--;
+                    mLength++;
                 } /* catch up */
                 offset_2 = offset_1;
                 offset_1 = offset;
-                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH);
+                ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
+                break;
             }
-        } else if (MEM_read32(match) != MEM_read32(ip)) {
-            /* it's not a match, and we're not going to check the dictionary */
-            assert(stepSize >= 1);
-            ip += ((ip-anchor) >> kSearchStrength) + stepSize;
-            continue;
-        } else {
-            /* found a regular match */
-            U32 const offset = (U32)(ip-match);
-            mLength = ZSTD_count(ip+4, match+4, iend) + 4;
-            while (((ip>anchor) & (match>prefixStart))
-                 && (ip[-1] == match[-1])) { ip--; match--; mLength++; } /* catch up */
-            offset_2 = offset_1;
-            offset_1 = offset;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH);
-        }
+
+            /* Prepare for next iteration */
+            dictMatchIndexAndTag = dictHashTable[dictHashAndTag1 >> ZSTD_SHORT_CACHE_TAG_BITS];
+            dictTagsMatch = ZSTD_comparePackedTags(dictMatchIndexAndTag, dictHashAndTag1);
+            matchIndex = hashTable[hash1];
+
+            if (ip1 >= nextStep) {
+                step++;
+                nextStep += kStepIncr;
+            }
+            ip0 = ip1;
+            ip1 = ip1 + step;
+            if (ip1 > ilimit) goto _cleanup;
+
+            curr = (U32)(ip0 - base);
+            hash0 = hash1;
+        }   /* end inner search loop */
 
         /* match found */
-        ip += mLength;
-        anchor = ip;
+        assert(mLength);
+        ip0 += mLength;
+        anchor = ip0;
 
-        if (ip <= ilimit) {
+        if (ip0 <= ilimit) {
             /* Fill Table */
-            assert(base+current+2 > istart);  /* check base overflow */
-            hashTable[ZSTD_hashPtr(base+current+2, hlog, mls)] = current+2;  /* here because current+2 could be > iend-8 */
-            hashTable[ZSTD_hashPtr(ip-2, hlog, mls)] = (U32)(ip-2-base);
+            assert(base+curr+2 > istart);  /* check base overflow */
+            hashTable[ZSTD_hashPtr(base+curr+2, hlog, mls)] = curr+2;  /* here because curr+2 could be > iend-8 */
+            hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base);
 
             /* check immediate repcode */
-            while (ip <= ilimit) {
-                U32 const current2 = (U32)(ip-base);
+            while (ip0 <= ilimit) {
+                U32 const current2 = (U32)(ip0-base);
                 U32 const repIndex2 = current2 - offset_2;
                 const BYTE* repMatch2 = repIndex2 < prefixStartIndex ?
                         dictBase - dictIndexDelta + repIndex2 :
                         base + repIndex2;
                 if ( ((U32)((prefixStartIndex-1) - (U32)repIndex2) >= 3 /* intentional overflow */)
-                   && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
+                   && (MEM_read32(repMatch2) == MEM_read32(ip0))) {
                     const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
-                    size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
+                    size_t const repLength2 = ZSTD_count_2segments(ip0+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
                     U32 tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset;   /* swap offset_2 <=> offset_1 */
-                    ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, repLength2-MINMATCH);
-                    hashTable[ZSTD_hashPtr(ip, hlog, mls)] = current2;
-                    ip += repLength2;
-                    anchor = ip;
+                    ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
+                    hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = current2;
+                    ip0 += repLength2;
+                    anchor = ip0;
                     continue;
                 }
                 break;
             }
         }
+
+        /* Prepare for next iteration */
+        assert(ip0 == anchor);
+        ip1 = ip0 + stepSize;
     }
 
+_cleanup:
     /* save reps for next block */
-    rep[0] = offset_1 ? offset_1 : offsetSaved;
-    rep[1] = offset_2 ? offset_2 : offsetSaved;
+    rep[0] = offset_1;
+    rep[1] = offset_2;
 
     /* Return the last literals size */
     return (size_t)(iend - anchor);
 }
 
+
+ZSTD_GEN_FAST_FN(dictMatchState, 4, 0)
+ZSTD_GEN_FAST_FN(dictMatchState, 5, 0)
+ZSTD_GEN_FAST_FN(dictMatchState, 6, 0)
+ZSTD_GEN_FAST_FN(dictMatchState, 7, 0)
+
 size_t ZSTD_compressBlock_fast_dictMatchState(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
@@ -361,30 +670,29 @@ size_t ZSTD_compressBlock_fast_dictMatchState(
     {
     default: /* includes case 3 */
     case 4 :
-        return ZSTD_compressBlock_fast_dictMatchState_generic(ms, seqStore, rep, src, srcSize, 4);
+        return ZSTD_compressBlock_fast_dictMatchState_4_0(ms, seqStore, rep, src, srcSize);
     case 5 :
-        return ZSTD_compressBlock_fast_dictMatchState_generic(ms, seqStore, rep, src, srcSize, 5);
+        return ZSTD_compressBlock_fast_dictMatchState_5_0(ms, seqStore, rep, src, srcSize);
     case 6 :
-        return ZSTD_compressBlock_fast_dictMatchState_generic(ms, seqStore, rep, src, srcSize, 6);
+        return ZSTD_compressBlock_fast_dictMatchState_6_0(ms, seqStore, rep, src, srcSize);
     case 7 :
-        return ZSTD_compressBlock_fast_dictMatchState_generic(ms, seqStore, rep, src, srcSize, 7);
+        return ZSTD_compressBlock_fast_dictMatchState_7_0(ms, seqStore, rep, src, srcSize);
     }
 }
 
 
 static size_t ZSTD_compressBlock_fast_extDict_generic(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize, U32 const mls)
+        void const* src, size_t srcSize, U32 const mls, U32 const hasStep)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
     U32* const hashTable = ms->hashTable;
     U32 const hlog = cParams->hashLog;
     /* support stepSize of 0 */
-    U32 const stepSize = cParams->targetLength + !(cParams->targetLength);
+    size_t const stepSize = cParams->targetLength + !(cParams->targetLength) + 1;
     const BYTE* const base = ms->window.base;
     const BYTE* const dictBase = ms->window.dictBase;
     const BYTE* const istart = (const BYTE*)src;
-    const BYTE* ip = istart;
     const BYTE* anchor = istart;
     const U32   endIndex = (U32)((size_t)(istart - base) + srcSize);
     const U32   lowLimit = ZSTD_getLowestMatchIndex(ms, endIndex, cParams->windowLog);
@@ -397,100 +705,256 @@ static size_t ZSTD_compressBlock_fast_extDict_generic(
     const BYTE* const iend = istart + srcSize;
     const BYTE* const ilimit = iend - 8;
     U32 offset_1=rep[0], offset_2=rep[1];
+    U32 offsetSaved1 = 0, offsetSaved2 = 0;
+
+    const BYTE* ip0 = istart;
+    const BYTE* ip1;
+    const BYTE* ip2;
+    const BYTE* ip3;
+    U32 current0;
+
+
+    size_t hash0; /* hash for ip0 */
+    size_t hash1; /* hash for ip1 */
+    U32 idx; /* match idx for ip0 */
+    const BYTE* idxBase; /* base pointer for idx */
+
+    U32 offcode;
+    const BYTE* match0;
+    size_t mLength;
+    const BYTE* matchEnd = 0; /* initialize to avoid warning, assert != 0 later */
+
+    size_t step;
+    const BYTE* nextStep;
+    const size_t kStepIncr = (1 << (kSearchStrength - 1));
+
+    (void)hasStep; /* not currently specialized on whether it's accelerated */
 
     DEBUGLOG(5, "ZSTD_compressBlock_fast_extDict_generic (offset_1=%u)", offset_1);
 
     /* switch to "regular" variant if extDict is invalidated due to maxDistance */
     if (prefixStartIndex == dictStartIndex)
-        return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, mls);
-
-    /* Search Loop */
-    while (ip < ilimit) {  /* < instead of <=, because (ip+1) */
-        const size_t h = ZSTD_hashPtr(ip, hlog, mls);
-        const U32    matchIndex = hashTable[h];
-        const BYTE* const matchBase = matchIndex < prefixStartIndex ? dictBase : base;
-        const BYTE*  match = matchBase + matchIndex;
-        const U32    current = (U32)(ip-base);
-        const U32    repIndex = current + 1 - offset_1;
-        const BYTE* const repBase = repIndex < prefixStartIndex ? dictBase : base;
-        const BYTE* const repMatch = repBase + repIndex;
-        hashTable[h] = current;   /* update hash table */
-        DEBUGLOG(7, "offset_1 = %u , current = %u", offset_1, current);
-        assert(offset_1 <= current +1);   /* check repIndex */
-
-        if ( (((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow */ & (repIndex > dictStartIndex))
-           && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
-            const BYTE* const repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
-            size_t const rLength = ZSTD_count_2segments(ip+1 +4, repMatch +4, iend, repMatchEnd, prefixStart) + 4;
-            ip++;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, 0, rLength-MINMATCH);
-            ip += rLength;
-            anchor = ip;
-        } else {
-            if ( (matchIndex < dictStartIndex) ||
-                 (MEM_read32(match) != MEM_read32(ip)) ) {
-                assert(stepSize >= 1);
-                ip += ((ip-anchor) >> kSearchStrength) + stepSize;
-                continue;
+        return ZSTD_compressBlock_fast(ms, seqStore, rep, src, srcSize);
+
+    {   U32 const curr = (U32)(ip0 - base);
+        U32 const maxRep = curr - dictStartIndex;
+        if (offset_2 >= maxRep) offsetSaved2 = offset_2, offset_2 = 0;
+        if (offset_1 >= maxRep) offsetSaved1 = offset_1, offset_1 = 0;
+    }
+
+    /* start each op */
+_start: /* Requires: ip0 */
+
+    step = stepSize;
+    nextStep = ip0 + kStepIncr;
+
+    /* calculate positions, ip0 - anchor == 0, so we skip step calc */
+    ip1 = ip0 + 1;
+    ip2 = ip0 + step;
+    ip3 = ip2 + 1;
+
+    if (ip3 >= ilimit) {
+        goto _cleanup;
+    }
+
+    hash0 = ZSTD_hashPtr(ip0, hlog, mls);
+    hash1 = ZSTD_hashPtr(ip1, hlog, mls);
+
+    idx = hashTable[hash0];
+    idxBase = idx < prefixStartIndex ? dictBase : base;
+
+    do {
+        {   /* load repcode match for ip[2] */
+            U32 const current2 = (U32)(ip2 - base);
+            U32 const repIndex = current2 - offset_1;
+            const BYTE* const repBase = repIndex < prefixStartIndex ? dictBase : base;
+            U32 rval;
+            if ( ((U32)(prefixStartIndex - repIndex) >= 4) /* intentional underflow */
+                 & (offset_1 > 0) ) {
+                rval = MEM_read32(repBase + repIndex);
+            } else {
+                rval = MEM_read32(ip2) ^ 1; /* guaranteed to not match. */
             }
-            {   const BYTE* const matchEnd = matchIndex < prefixStartIndex ? dictEnd : iend;
-                const BYTE* const lowMatchPtr = matchIndex < prefixStartIndex ? dictStart : prefixStart;
-                U32 const offset = current - matchIndex;
-                size_t mLength = ZSTD_count_2segments(ip+4, match+4, iend, matchEnd, prefixStart) + 4;
-                while (((ip>anchor) & (match>lowMatchPtr)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; }   /* catch up */
-                offset_2 = offset_1; offset_1 = offset;  /* update offset history */
-                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH);
-                ip += mLength;
-                anchor = ip;
+
+            /* write back hash table entry */
+            current0 = (U32)(ip0 - base);
+            hashTable[hash0] = current0;
+
+            /* check repcode at ip[2] */
+            if (MEM_read32(ip2) == rval) {
+                ip0 = ip2;
+                match0 = repBase + repIndex;
+                matchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
+                assert((match0 != prefixStart) & (match0 != dictStart));
+                mLength = ip0[-1] == match0[-1];
+                ip0 -= mLength;
+                match0 -= mLength;
+                offcode = REPCODE1_TO_OFFBASE;
+                mLength += 4;
+                goto _match;
         }   }
 
-        if (ip <= ilimit) {
-            /* Fill Table */
-            hashTable[ZSTD_hashPtr(base+current+2, hlog, mls)] = current+2;
-            hashTable[ZSTD_hashPtr(ip-2, hlog, mls)] = (U32)(ip-2-base);
-            /* check immediate repcode */
-            while (ip <= ilimit) {
-                U32 const current2 = (U32)(ip-base);
-                U32 const repIndex2 = current2 - offset_2;
-                const BYTE* const repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2;
-                if ( (((U32)((prefixStartIndex-1) - repIndex2) >= 3) & (repIndex2 > dictStartIndex))  /* intentional overflow */
-                   && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
-                    const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
-                    size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
-                    { U32 const tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; }  /* swap offset_2 <=> offset_1 */
-                    ZSTD_storeSeq(seqStore, 0 /*litlen*/, anchor, iend, 0 /*offcode*/, repLength2-MINMATCH);
-                    hashTable[ZSTD_hashPtr(ip, hlog, mls)] = current2;
-                    ip += repLength2;
-                    anchor = ip;
-                    continue;
-                }
-                break;
-    }   }   }
+        {   /* load match for ip[0] */
+            U32 const mval = idx >= dictStartIndex ?
+                    MEM_read32(idxBase + idx) :
+                    MEM_read32(ip0) ^ 1; /* guaranteed not to match */
+
+            /* check match at ip[0] */
+            if (MEM_read32(ip0) == mval) {
+                /* found a match! */
+                goto _offset;
+        }   }
+
+        /* lookup ip[1] */
+        idx = hashTable[hash1];
+        idxBase = idx < prefixStartIndex ? dictBase : base;
+
+        /* hash ip[2] */
+        hash0 = hash1;
+        hash1 = ZSTD_hashPtr(ip2, hlog, mls);
+
+        /* advance to next positions */
+        ip0 = ip1;
+        ip1 = ip2;
+        ip2 = ip3;
+
+        /* write back hash table entry */
+        current0 = (U32)(ip0 - base);
+        hashTable[hash0] = current0;
+
+        {   /* load match for ip[0] */
+            U32 const mval = idx >= dictStartIndex ?
+                    MEM_read32(idxBase + idx) :
+                    MEM_read32(ip0) ^ 1; /* guaranteed not to match */
+
+            /* check match at ip[0] */
+            if (MEM_read32(ip0) == mval) {
+                /* found a match! */
+                goto _offset;
+        }   }
+
+        /* lookup ip[1] */
+        idx = hashTable[hash1];
+        idxBase = idx < prefixStartIndex ? dictBase : base;
+
+        /* hash ip[2] */
+        hash0 = hash1;
+        hash1 = ZSTD_hashPtr(ip2, hlog, mls);
+
+        /* advance to next positions */
+        ip0 = ip1;
+        ip1 = ip2;
+        ip2 = ip0 + step;
+        ip3 = ip1 + step;
+
+        /* calculate step */
+        if (ip2 >= nextStep) {
+            step++;
+            PREFETCH_L1(ip1 + 64);
+            PREFETCH_L1(ip1 + 128);
+            nextStep += kStepIncr;
+        }
+    } while (ip3 < ilimit);
+
+_cleanup:
+    /* Note that there are probably still a couple positions we could search.
+     * However, it seems to be a meaningful performance hit to try to search
+     * them. So let's not. */
+
+    /* If offset_1 started invalid (offsetSaved1 != 0) and became valid (offset_1 != 0),
+     * rotate saved offsets. See comment in ZSTD_compressBlock_fast_noDict for more context. */
+    offsetSaved2 = ((offsetSaved1 != 0) && (offset_1 != 0)) ? offsetSaved1 : offsetSaved2;
 
     /* save reps for next block */
-    rep[0] = offset_1;
-    rep[1] = offset_2;
+    rep[0] = offset_1 ? offset_1 : offsetSaved1;
+    rep[1] = offset_2 ? offset_2 : offsetSaved2;
 
     /* Return the last literals size */
     return (size_t)(iend - anchor);
+
+_offset: /* Requires: ip0, idx, idxBase */
+
+    /* Compute the offset code. */
+    {   U32 const offset = current0 - idx;
+        const BYTE* const lowMatchPtr = idx < prefixStartIndex ? dictStart : prefixStart;
+        matchEnd = idx < prefixStartIndex ? dictEnd : iend;
+        match0 = idxBase + idx;
+        offset_2 = offset_1;
+        offset_1 = offset;
+        offcode = OFFSET_TO_OFFBASE(offset);
+        mLength = 4;
+
+        /* Count the backwards match length. */
+        while (((ip0>anchor) & (match0>lowMatchPtr)) && (ip0[-1] == match0[-1])) {
+            ip0--;
+            match0--;
+            mLength++;
+    }   }
+
+_match: /* Requires: ip0, match0, offcode, matchEnd */
+
+    /* Count the forward length. */
+    assert(matchEnd != 0);
+    mLength += ZSTD_count_2segments(ip0 + mLength, match0 + mLength, iend, matchEnd, prefixStart);
+
+    ZSTD_storeSeq(seqStore, (size_t)(ip0 - anchor), anchor, iend, offcode, mLength);
+
+    ip0 += mLength;
+    anchor = ip0;
+
+    /* write next hash table entry */
+    if (ip1 < ip0) {
+        hashTable[hash1] = (U32)(ip1 - base);
+    }
+
+    /* Fill table and check for immediate repcode. */
+    if (ip0 <= ilimit) {
+        /* Fill Table */
+        assert(base+current0+2 > istart);  /* check base overflow */
+        hashTable[ZSTD_hashPtr(base+current0+2, hlog, mls)] = current0+2;  /* here because current+2 could be > iend-8 */
+        hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base);
+
+        while (ip0 <= ilimit) {
+            U32 const repIndex2 = (U32)(ip0-base) - offset_2;
+            const BYTE* const repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2;
+            if ( (((U32)((prefixStartIndex-1) - repIndex2) >= 3) & (offset_2 > 0))  /* intentional underflow */
+                 && (MEM_read32(repMatch2) == MEM_read32(ip0)) ) {
+                const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
+                size_t const repLength2 = ZSTD_count_2segments(ip0+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
+                { U32 const tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; }  /* swap offset_2 <=> offset_1 */
+                ZSTD_storeSeq(seqStore, 0 /*litlen*/, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
+                hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = (U32)(ip0-base);
+                ip0 += repLength2;
+                anchor = ip0;
+                continue;
+            }
+            break;
+    }   }
+
+    goto _start;
 }
 
+ZSTD_GEN_FAST_FN(extDict, 4, 0)
+ZSTD_GEN_FAST_FN(extDict, 5, 0)
+ZSTD_GEN_FAST_FN(extDict, 6, 0)
+ZSTD_GEN_FAST_FN(extDict, 7, 0)
 
 size_t ZSTD_compressBlock_fast_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
 {
     U32 const mls = ms->cParams.minMatch;
+    assert(ms->dictMatchState == NULL);
     switch(mls)
     {
     default: /* includes case 3 */
     case 4 :
-        return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, 4);
+        return ZSTD_compressBlock_fast_extDict_4_0(ms, seqStore, rep, src, srcSize);
     case 5 :
-        return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, 5);
+        return ZSTD_compressBlock_fast_extDict_5_0(ms, seqStore, rep, src, srcSize);
     case 6 :
-        return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, 6);
+        return ZSTD_compressBlock_fast_extDict_6_0(ms, seqStore, rep, src, srcSize);
     case 7 :
-        return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, 7);
+        return ZSTD_compressBlock_fast_extDict_7_0(ms, seqStore, rep, src, srcSize);
     }
 }
diff --git a/contrib/zstd/zstd_fast.h b/contrib/zstd/zstd_fast.h
index 3e15cc85c..3e81de496 100644
--- a/contrib/zstd/zstd_fast.h
+++ b/contrib/zstd/zstd_fast.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -19,7 +19,8 @@ extern "C" {
 #include "zstd_compress_internal.h"
 
 void ZSTD_fillHashTable(ZSTD_matchState_t* ms,
-                        void const* end, ZSTD_dictTableLoadMethod_e dtlm);
+                        void const* end, ZSTD_dictTableLoadMethod_e dtlm,
+                        ZSTD_tableFillPurpose_e tfp);
 size_t ZSTD_compressBlock_fast(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
diff --git a/contrib/zstd/zstd_internal.h b/contrib/zstd/zstd_internal.h
index d45e88257..8f4f64b0f 100644
--- a/contrib/zstd/zstd_internal.h
+++ b/contrib/zstd/zstd_internal.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -19,25 +19,25 @@
 /*-*************************************
 *  Dependencies
 ***************************************/
-#ifdef __aarch64__
-#include <arm_neon.h>
-#endif
 #include "compiler.h"
+#include "cpu.h"
 #include "mem.h"
 #include "debug.h"                 /* assert, DEBUGLOG, RAWLOG, g_debuglevel */
 #include "error_private.h"
-#ifndef ZSTD_STATIC_LINKING_ONLY
 #define ZSTD_STATIC_LINKING_ONLY
-#endif
 #include "zstd.h"
 #define FSE_STATIC_LINKING_ONLY
 #include "fse.h"
-#define HUF_STATIC_LINKING_ONLY
 #include "huf.h"
 #ifndef XXH_STATIC_LINKING_ONLY
 #  define XXH_STATIC_LINKING_ONLY  /* XXH64_state_t */
 #endif
 #include "xxhash.h"                /* XXH_reset, update, digest */
+#ifndef ZSTD_NO_TRACE
+#  include "zstd_trace.h"
+#else
+#  define ZSTD_TRACE 0
+#endif
 
 #if defined (__cplusplus)
 extern "C" {
@@ -57,81 +57,7 @@ extern "C" {
 #undef MAX
 #define MIN(a,b) ((a)<(b) ? (a) : (b))
 #define MAX(a,b) ((a)>(b) ? (a) : (b))
-
-/**
- * Ignore: this is an internal helper.
- *
- * This is a helper function to help force C99-correctness during compilation.
- * Under strict compilation modes, variadic macro arguments can't be empty.
- * However, variadic function arguments can be. Using a function therefore lets
- * us statically check that at least one (string) argument was passed,
- * independent of the compilation flags.
- */
-static INLINE_KEYWORD UNUSED_ATTR
-void _force_has_format_string(const char *format, ...) {
-  (void)format;
-}
-
-/**
- * Ignore: this is an internal helper.
- *
- * We want to force this function invocation to be syntactically correct, but
- * we don't want to force runtime evaluation of its arguments.
- */
-#define _FORCE_HAS_FORMAT_STRING(...) \
-  if (0) { \
-    _force_has_format_string(__VA_ARGS__); \
-  }
-
-/**
- * Return the specified error if the condition evaluates to true.
- *
- * In debug modes, prints additional information.
- * In order to do that (particularly, printing the conditional that failed),
- * this can't just wrap RETURN_ERROR().
- */
-#define RETURN_ERROR_IF(cond, err, ...) \
-  if (cond) { \
-    RAWLOG(3, "%s:%d: ERROR!: check %s failed, returning %s", \
-           __FILE__, __LINE__, ZSTD_QUOTE(cond), ZSTD_QUOTE(ERROR(err))); \
-    _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
-    RAWLOG(3, ": " __VA_ARGS__); \
-    RAWLOG(3, "\n"); \
-    return ERROR(err); \
-  }
-
-/**
- * Unconditionally return the specified error.
- *
- * In debug modes, prints additional information.
- */
-#define RETURN_ERROR(err, ...) \
-  do { \
-    RAWLOG(3, "%s:%d: ERROR!: unconditional check failed, returning %s", \
-           __FILE__, __LINE__, ZSTD_QUOTE(ERROR(err))); \
-    _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
-    RAWLOG(3, ": " __VA_ARGS__); \
-    RAWLOG(3, "\n"); \
-    return ERROR(err); \
-  } while(0);
-
-/**
- * If the provided expression evaluates to an error code, returns that error code.
- *
- * In debug modes, prints additional information.
- */
-#define FORWARD_IF_ERROR(err, ...) \
-  do { \
-    size_t const err_code = (err); \
-    if (ERR_isError(err_code)) { \
-      RAWLOG(3, "%s:%d: ERROR!: forwarding error in %s: %s", \
-             __FILE__, __LINE__, ZSTD_QUOTE(err), ERR_getErrorName(err_code)); \
-      _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
-      RAWLOG(3, ": " __VA_ARGS__); \
-      RAWLOG(3, "\n"); \
-      return err_code; \
-    } \
-  } while(0);
+#define BOUNDED(min,val,max) (MAX(min,MIN(val,max)))
 
 
 /*-*************************************
@@ -140,8 +66,7 @@ void _force_has_format_string(const char *format, ...) {
 #define ZSTD_OPT_NUM    (1<<12)
 
 #define ZSTD_REP_NUM      3                 /* number of repcodes */
-#define ZSTD_REP_MOVE     (ZSTD_REP_NUM-1)
-static const U32 repStartValue[ZSTD_REP_NUM] = { 1, 4, 8 };
+static UNUSED_ATTR const U32 repStartValue[ZSTD_REP_NUM] = { 1, 4, 8 };
 
 #define KB *(1 <<10)
 #define MB *(1 <<20)
@@ -155,21 +80,21 @@ static const U32 repStartValue[ZSTD_REP_NUM] = { 1, 4, 8 };
 #define BIT0   1
 
 #define ZSTD_WINDOWLOG_ABSOLUTEMIN 10
-static const size_t ZSTD_fcs_fieldSize[4] = { 0, 2, 4, 8 };
-static const size_t ZSTD_did_fieldSize[4] = { 0, 1, 2, 4 };
+static UNUSED_ATTR const size_t ZSTD_fcs_fieldSize[4] = { 0, 2, 4, 8 };
+static UNUSED_ATTR const size_t ZSTD_did_fieldSize[4] = { 0, 1, 2, 4 };
 
 #define ZSTD_FRAMEIDSIZE 4   /* magic number size */
 
 #define ZSTD_BLOCKHEADERSIZE 3   /* C standard doesn't allow `static const` variable to be init using another `static const` variable */
-static const size_t ZSTD_blockHeaderSize = ZSTD_BLOCKHEADERSIZE;
+static UNUSED_ATTR const size_t ZSTD_blockHeaderSize = ZSTD_BLOCKHEADERSIZE;
 typedef enum { bt_raw, bt_rle, bt_compressed, bt_reserved } blockType_e;
 
 #define ZSTD_FRAMECHECKSUMSIZE 4
 
 #define MIN_SEQUENCES_SIZE 1 /* nbSeq==0 */
-#define MIN_CBLOCK_SIZE (1 /*litCSize*/ + 1 /* RLE or RAW */ + MIN_SEQUENCES_SIZE /* nbSeq==0 */)   /* for a non-null block */
+#define MIN_CBLOCK_SIZE (1 /*litCSize*/ + 1 /* RLE or RAW */)   /* for a non-null block */
+#define MIN_LITERALS_FOR_4_STREAMS 6
 
-#define HufLog 12
 typedef enum { set_basic, set_rle, set_compressed, set_repeat } symbolEncodingType_e;
 
 #define LONGNBSEQ 0x7F00
@@ -177,6 +102,7 @@ typedef enum { set_basic, set_rle, set_compressed, set_repeat } symbolEncodingTy
 #define MINMATCH 3
 
 #define Litbits  8
+#define LitHufLog 11
 #define MaxLit ((1<<Litbits) - 1)
 #define MaxML   52
 #define MaxLL   35
@@ -187,62 +113,89 @@ typedef enum { set_basic, set_rle, set_compressed, set_repeat } symbolEncodingTy
 #define LLFSELog    9
 #define OffFSELog   8
 #define MaxFSELog  MAX(MAX(MLFSELog, LLFSELog), OffFSELog)
-
-static const U32 LL_bits[MaxLL+1] = { 0, 0, 0, 0, 0, 0, 0, 0,
-                                      0, 0, 0, 0, 0, 0, 0, 0,
-                                      1, 1, 1, 1, 2, 2, 3, 3,
-                                      4, 6, 7, 8, 9,10,11,12,
-                                     13,14,15,16 };
-static const S16 LL_defaultNorm[MaxLL+1] = { 4, 3, 2, 2, 2, 2, 2, 2,
-                                             2, 2, 2, 2, 2, 1, 1, 1,
-                                             2, 2, 2, 2, 2, 2, 2, 2,
-                                             2, 3, 2, 1, 1, 1, 1, 1,
-                                            -1,-1,-1,-1 };
+#define MaxMLBits 16
+#define MaxLLBits 16
+
+#define ZSTD_MAX_HUF_HEADER_SIZE 128 /* header + <= 127 byte tree description */
+/* Each table cannot take more than #symbols * FSELog bits */
+#define ZSTD_MAX_FSE_HEADERS_SIZE (((MaxML + 1) * MLFSELog + (MaxLL + 1) * LLFSELog + (MaxOff + 1) * OffFSELog + 7) / 8)
+
+static UNUSED_ATTR const U8 LL_bits[MaxLL+1] = {
+     0, 0, 0, 0, 0, 0, 0, 0,
+     0, 0, 0, 0, 0, 0, 0, 0,
+     1, 1, 1, 1, 2, 2, 3, 3,
+     4, 6, 7, 8, 9,10,11,12,
+    13,14,15,16
+};
+static UNUSED_ATTR const S16 LL_defaultNorm[MaxLL+1] = {
+     4, 3, 2, 2, 2, 2, 2, 2,
+     2, 2, 2, 2, 2, 1, 1, 1,
+     2, 2, 2, 2, 2, 2, 2, 2,
+     2, 3, 2, 1, 1, 1, 1, 1,
+    -1,-1,-1,-1
+};
 #define LL_DEFAULTNORMLOG 6  /* for static allocation */
-static const U32 LL_defaultNormLog = LL_DEFAULTNORMLOG;
-
-static const U32 ML_bits[MaxML+1] = { 0, 0, 0, 0, 0, 0, 0, 0,
-                                      0, 0, 0, 0, 0, 0, 0, 0,
-                                      0, 0, 0, 0, 0, 0, 0, 0,
-                                      0, 0, 0, 0, 0, 0, 0, 0,
-                                      1, 1, 1, 1, 2, 2, 3, 3,
-                                      4, 4, 5, 7, 8, 9,10,11,
-                                     12,13,14,15,16 };
-static const S16 ML_defaultNorm[MaxML+1] = { 1, 4, 3, 2, 2, 2, 2, 2,
-                                             2, 1, 1, 1, 1, 1, 1, 1,
-                                             1, 1, 1, 1, 1, 1, 1, 1,
-                                             1, 1, 1, 1, 1, 1, 1, 1,
-                                             1, 1, 1, 1, 1, 1, 1, 1,
-                                             1, 1, 1, 1, 1, 1,-1,-1,
-                                            -1,-1,-1,-1,-1 };
+static UNUSED_ATTR const U32 LL_defaultNormLog = LL_DEFAULTNORMLOG;
+
+static UNUSED_ATTR const U8 ML_bits[MaxML+1] = {
+     0, 0, 0, 0, 0, 0, 0, 0,
+     0, 0, 0, 0, 0, 0, 0, 0,
+     0, 0, 0, 0, 0, 0, 0, 0,
+     0, 0, 0, 0, 0, 0, 0, 0,
+     1, 1, 1, 1, 2, 2, 3, 3,
+     4, 4, 5, 7, 8, 9,10,11,
+    12,13,14,15,16
+};
+static UNUSED_ATTR const S16 ML_defaultNorm[MaxML+1] = {
+     1, 4, 3, 2, 2, 2, 2, 2,
+     2, 1, 1, 1, 1, 1, 1, 1,
+     1, 1, 1, 1, 1, 1, 1, 1,
+     1, 1, 1, 1, 1, 1, 1, 1,
+     1, 1, 1, 1, 1, 1, 1, 1,
+     1, 1, 1, 1, 1, 1,-1,-1,
+    -1,-1,-1,-1,-1
+};
 #define ML_DEFAULTNORMLOG 6  /* for static allocation */
-static const U32 ML_defaultNormLog = ML_DEFAULTNORMLOG;
-
-static const S16 OF_defaultNorm[DefaultMaxOff+1] = { 1, 1, 1, 1, 1, 1, 2, 2,
-                                                     2, 1, 1, 1, 1, 1, 1, 1,
-                                                     1, 1, 1, 1, 1, 1, 1, 1,
-                                                    -1,-1,-1,-1,-1 };
+static UNUSED_ATTR const U32 ML_defaultNormLog = ML_DEFAULTNORMLOG;
+
+static UNUSED_ATTR const S16 OF_defaultNorm[DefaultMaxOff+1] = {
+     1, 1, 1, 1, 1, 1, 2, 2,
+     2, 1, 1, 1, 1, 1, 1, 1,
+     1, 1, 1, 1, 1, 1, 1, 1,
+    -1,-1,-1,-1,-1
+};
 #define OF_DEFAULTNORMLOG 5  /* for static allocation */
-static const U32 OF_defaultNormLog = OF_DEFAULTNORMLOG;
+static UNUSED_ATTR const U32 OF_defaultNormLog = OF_DEFAULTNORMLOG;
 
 
 /*-*******************************************
 *  Shared functions to include for inlining
 *********************************************/
 static void ZSTD_copy8(void* dst, const void* src) {
-#ifdef __aarch64__
+#if defined(ZSTD_ARCH_ARM_NEON)
     vst1_u8((uint8_t*)dst, vld1_u8((const uint8_t*)src));
 #else
-    memcpy(dst, src, 8);
+    ZSTD_memcpy(dst, src, 8);
 #endif
 }
-
 #define COPY8(d,s) { ZSTD_copy8(d,s); d+=8; s+=8; }
+
+/* Need to use memmove here since the literal buffer can now be located within
+   the dst buffer. In circumstances where the op "catches up" to where the
+   literal buffer is, there can be partial overlaps in this call on the final
+   copy if the literal is being shifted by less than 16 bytes. */
 static void ZSTD_copy16(void* dst, const void* src) {
-#ifdef __aarch64__
+#if defined(ZSTD_ARCH_ARM_NEON)
     vst1q_u8((uint8_t*)dst, vld1q_u8((const uint8_t*)src));
+#elif defined(ZSTD_ARCH_X86_SSE2)
+    _mm_storeu_si128((__m128i*)dst, _mm_loadu_si128((const __m128i*)src));
+#elif defined(__clang__)
+    ZSTD_memmove(dst, src, 16);
 #else
-    memcpy(dst, src, 16);
+    /* ZSTD_memmove is not inlined properly by gcc */
+    BYTE copy16_buf[16];
+    ZSTD_memcpy(copy16_buf, src, 16);
+    ZSTD_memcpy(dst, copy16_buf, 16);
 #endif
 }
 #define COPY16(d,s) { ZSTD_copy16(d,s); d+=16; s+=16; }
@@ -257,7 +210,7 @@ typedef enum {
 } ZSTD_overlap_e;
 
 /*! ZSTD_wildcopy() :
- *  Custom version of memcpy(), can over read/write up to WILDCOPY_OVERLENGTH bytes (if length==0)
+ *  Custom version of ZSTD_memcpy(), can over read/write up to WILDCOPY_OVERLENGTH bytes (if length==0)
  *  @param ovtype controls the overlap detection
  *         - ZSTD_no_overlap: The source and destination are guaranteed to be at least WILDCOPY_VECLEN bytes apart.
  *         - ZSTD_overlap_src_before_dst: The src and dst may overlap, but they MUST be at least 8 bytes apart.
@@ -271,8 +224,6 @@ void ZSTD_wildcopy(void* dst, const void* src, ptrdiff_t length, ZSTD_overlap_e
     BYTE* op = (BYTE*)dst;
     BYTE* const oend = op + length;
 
-    assert(diff >= 8 || (ovtype == ZSTD_no_overlap && diff <= -WILDCOPY_VECLEN));
-
     if (ovtype == ZSTD_overlap_src_before_dst && diff < WILDCOPY_VECLEN) {
         /* Handle short offset copies. */
         do {
@@ -286,20 +237,15 @@ void ZSTD_wildcopy(void* dst, const void* src, ptrdiff_t length, ZSTD_overlap_e
          * one COPY16() in the first call. Then, do two calls per loop since
          * at that point it is more likely to have a high trip count.
          */
-#ifndef __aarch64__
-        do {
-            COPY16(op, ip);
-        }
-        while (op < oend);
-#else
-        COPY16(op, ip);
-        if (op >= oend) return;
+        ZSTD_copy16(op, ip);
+        if (16 >= length) return;
+        op += 16;
+        ip += 16;
         do {
             COPY16(op, ip);
             COPY16(op, ip);
         }
         while (op < oend);
-#endif
     }
 }
 
@@ -307,7 +253,7 @@ MEM_STATIC size_t ZSTD_limitCopy(void* dst, size_t dstCapacity, const void* src,
 {
     size_t const length = MIN(dstCapacity, srcSize);
     if (length > 0) {
-        memcpy(dst, src, length);
+        ZSTD_memcpy(dst, src, length);
     }
     return length;
 }
@@ -322,28 +268,46 @@ MEM_STATIC size_t ZSTD_limitCopy(void* dst, size_t dstCapacity, const void* src,
  * In which case, resize it down to free some memory */
 #define ZSTD_WORKSPACETOOLARGE_MAXDURATION 128
 
+/* Controls whether the input/output buffer is buffered or stable. */
+typedef enum {
+    ZSTD_bm_buffered = 0,  /* Buffer the input/output */
+    ZSTD_bm_stable = 1     /* ZSTD_inBuffer/ZSTD_outBuffer is stable */
+} ZSTD_bufferMode_e;
+
 
 /*-*******************************************
 *  Private declarations
 *********************************************/
 typedef struct seqDef_s {
-    U32 offset;
+    U32 offBase;   /* offBase == Offset + ZSTD_REP_NUM, or repcode 1,2,3 */
     U16 litLength;
-    U16 matchLength;
+    U16 mlBase;    /* mlBase == matchLength - MINMATCH */
 } seqDef;
 
+/* Controls whether seqStore has a single "long" litLength or matchLength. See seqStore_t. */
+typedef enum {
+    ZSTD_llt_none = 0,             /* no longLengthType */
+    ZSTD_llt_literalLength = 1,    /* represents a long literal */
+    ZSTD_llt_matchLength = 2       /* represents a long match */
+} ZSTD_longLengthType_e;
+
 typedef struct {
     seqDef* sequencesStart;
-    seqDef* sequences;
-    BYTE* litStart;
-    BYTE* lit;
-    BYTE* llCode;
-    BYTE* mlCode;
-    BYTE* ofCode;
+    seqDef* sequences;      /* ptr to end of sequences */
+    BYTE*  litStart;
+    BYTE*  lit;             /* ptr to end of literals */
+    BYTE*  llCode;
+    BYTE*  mlCode;
+    BYTE*  ofCode;
     size_t maxNbSeq;
     size_t maxNbLit;
-    U32   longLengthID;   /* 0 == no longLength; 1 == Lit.longLength; 2 == Match.longLength; */
-    U32   longLengthPos;
+
+    /* longLengthPos and longLengthType to allow us to represent either a single litLength or matchLength
+     * in the seqStore that has a value larger than U16 (if it exists). To do so, we increment
+     * the existing value of the litLength or matchLength by 0x10000.
+     */
+    ZSTD_longLengthType_e longLengthType;
+    U32                   longLengthPos;  /* Index of the sequence to apply long length modification to */
 } seqStore_t;
 
 typedef struct {
@@ -353,19 +317,19 @@ typedef struct {
 
 /**
  * Returns the ZSTD_sequenceLength for the given sequences. It handles the decoding of long sequences
- * indicated by longLengthPos and longLengthID, and adds MINMATCH back to matchLength.
+ * indicated by longLengthPos and longLengthType, and adds MINMATCH back to matchLength.
  */
 MEM_STATIC ZSTD_sequenceLength ZSTD_getSequenceLength(seqStore_t const* seqStore, seqDef const* seq)
 {
     ZSTD_sequenceLength seqLen;
     seqLen.litLength = seq->litLength;
-    seqLen.matchLength = seq->matchLength + MINMATCH;
+    seqLen.matchLength = seq->mlBase + MINMATCH;
     if (seqStore->longLengthPos == (U32)(seq - seqStore->sequencesStart)) {
-        if (seqStore->longLengthID == 1) {
-            seqLen.litLength += 0xFFFF;
+        if (seqStore->longLengthType == ZSTD_llt_literalLength) {
+            seqLen.litLength += 0x10000;
         }
-        if (seqStore->longLengthID == 2) {
-            seqLen.matchLength += 0xFFFF;
+        if (seqStore->longLengthType == ZSTD_llt_matchLength) {
+            seqLen.matchLength += 0x10000;
         }
     }
     return seqLen;
@@ -378,42 +342,18 @@ MEM_STATIC ZSTD_sequenceLength ZSTD_getSequenceLength(seqStore_t const* seqStore
  *          `decompressedBound != ZSTD_CONTENTSIZE_ERROR`
  */
 typedef struct {
+    size_t nbBlocks;
     size_t compressedSize;
     unsigned long long decompressedBound;
 } ZSTD_frameSizeInfo;   /* decompress & legacy */
 
 const seqStore_t* ZSTD_getSeqStore(const ZSTD_CCtx* ctx);   /* compress & dictBuilder */
-void ZSTD_seqToCodes(const seqStore_t* seqStorePtr);   /* compress, dictBuilder, decodeCorpus (shouldn't get its definition from here) */
+int ZSTD_seqToCodes(const seqStore_t* seqStorePtr);   /* compress, dictBuilder, decodeCorpus (shouldn't get its definition from here) */
 
 /* custom memory allocation functions */
-void* ZSTD_malloc(size_t size, ZSTD_customMem customMem);
-void* ZSTD_calloc(size_t size, ZSTD_customMem customMem);
-void ZSTD_free(void* ptr, ZSTD_customMem customMem);
-
-
-MEM_STATIC U32 ZSTD_highbit32(U32 val)   /* compress, dictBuilder, decodeCorpus */
-{
-    assert(val != 0);
-    {
-#   if defined(_MSC_VER)   /* Visual */
-        unsigned long r=0;
-        return _BitScanReverse(&r, val) ? (unsigned)r : 0;
-#   elif defined(__GNUC__) && (__GNUC__ >= 3)   /* GCC Intrinsic */
-        return __builtin_clz (val) ^ 31;
-#   elif defined(__ICCARM__)    /* IAR Intrinsic */
-        return 31 - __CLZ(val);
-#   else   /* Software version */
-        static const U32 DeBruijnClz[32] = { 0, 9, 1, 10, 13, 21, 2, 29, 11, 14, 16, 18, 22, 25, 3, 30, 8, 12, 20, 28, 15, 17, 24, 7, 19, 27, 23, 6, 26, 5, 4, 31 };
-        U32 v = val;
-        v |= v >> 1;
-        v |= v >> 2;
-        v |= v >> 4;
-        v |= v >> 8;
-        v |= v >> 16;
-        return DeBruijnClz[(v * 0x07C4ACDDU) >> 27];
-#   endif
-    }
-}
+void* ZSTD_customMalloc(size_t size, ZSTD_customMem customMem);
+void* ZSTD_customCalloc(size_t size, ZSTD_customMem customMem);
+void ZSTD_customFree(void* ptr, ZSTD_customMem customMem);
 
 
 /* ZSTD_invalidateRepCodes() :
@@ -441,6 +381,14 @@ size_t ZSTD_getcBlockSize(const void* src, size_t srcSize,
 size_t ZSTD_decodeSeqHeaders(ZSTD_DCtx* dctx, int* nbSeqPtr,
                        const void* src, size_t srcSize);
 
+/**
+ * @returns true iff the CPU supports dynamic BMI2 dispatch.
+ */
+MEM_STATIC int ZSTD_cpuSupportsBmi2(void)
+{
+    ZSTD_cpuid_t cpuid = ZSTD_cpuid();
+    return ZSTD_cpuid_bmi1(cpuid) && ZSTD_cpuid_bmi2(cpuid);
+}
 
 #if defined (__cplusplus)
 }
diff --git a/contrib/zstd/zstd_lazy.c b/contrib/zstd/zstd_lazy.c
index 4cf5c88b5..297f52007 100644
--- a/contrib/zstd/zstd_lazy.c
+++ b/contrib/zstd/zstd_lazy.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -10,6 +10,7 @@
 
 #include "zstd_compress_internal.h"
 #include "zstd_lazy.h"
+#include "bits.h" /* ZSTD_countTrailingZeros64 */
 
 
 /*-*************************************
@@ -58,11 +59,11 @@ ZSTD_updateDUBT(ZSTD_matchState_t* ms,
 
 /** ZSTD_insertDUBT1() :
  *  sort one already inserted but unsorted position
- *  assumption : current >= btlow == (current - btmask)
+ *  assumption : curr >= btlow == (curr - btmask)
  *  doesn't fail */
 static void
-ZSTD_insertDUBT1(ZSTD_matchState_t* ms,
-                 U32 current, const BYTE* inputEnd,
+ZSTD_insertDUBT1(const ZSTD_matchState_t* ms,
+                 U32 curr, const BYTE* inputEnd,
                  U32 nbCompares, U32 btLow,
                  const ZSTD_dictMode_e dictMode)
 {
@@ -74,41 +75,41 @@ ZSTD_insertDUBT1(ZSTD_matchState_t* ms,
     const BYTE* const base = ms->window.base;
     const BYTE* const dictBase = ms->window.dictBase;
     const U32 dictLimit = ms->window.dictLimit;
-    const BYTE* const ip = (current>=dictLimit) ? base + current : dictBase + current;
-    const BYTE* const iend = (current>=dictLimit) ? inputEnd : dictBase + dictLimit;
+    const BYTE* const ip = (curr>=dictLimit) ? base + curr : dictBase + curr;
+    const BYTE* const iend = (curr>=dictLimit) ? inputEnd : dictBase + dictLimit;
     const BYTE* const dictEnd = dictBase + dictLimit;
     const BYTE* const prefixStart = base + dictLimit;
     const BYTE* match;
-    U32* smallerPtr = bt + 2*(current&btMask);
+    U32* smallerPtr = bt + 2*(curr&btMask);
     U32* largerPtr  = smallerPtr + 1;
     U32 matchIndex = *smallerPtr;   /* this candidate is unsorted : next sorted candidate is reached through *smallerPtr, while *largerPtr contains previous unsorted candidate (which is already saved and can be overwritten) */
     U32 dummy32;   /* to be nullified at the end */
     U32 const windowValid = ms->window.lowLimit;
     U32 const maxDistance = 1U << cParams->windowLog;
-    U32 const windowLow = (current - windowValid > maxDistance) ? current - maxDistance : windowValid;
+    U32 const windowLow = (curr - windowValid > maxDistance) ? curr - maxDistance : windowValid;
 
 
     DEBUGLOG(8, "ZSTD_insertDUBT1(%u) (dictLimit=%u, lowLimit=%u)",
-                current, dictLimit, windowLow);
-    assert(current >= btLow);
+                curr, dictLimit, windowLow);
+    assert(curr >= btLow);
     assert(ip < iend);   /* condition for ZSTD_count */
 
-    while (nbCompares-- && (matchIndex > windowLow)) {
+    for (; nbCompares && (matchIndex > windowLow); --nbCompares) {
         U32* const nextPtr = bt + 2*(matchIndex & btMask);
         size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
-        assert(matchIndex < current);
+        assert(matchIndex < curr);
         /* note : all candidates are now supposed sorted,
          * but it's still possible to have nextPtr[1] == ZSTD_DUBT_UNSORTED_MARK
          * when a real index has the same value as ZSTD_DUBT_UNSORTED_MARK */
 
         if ( (dictMode != ZSTD_extDict)
           || (matchIndex+matchLength >= dictLimit)  /* both in current segment*/
-          || (current < dictLimit) /* both in extDict */) {
+          || (curr < dictLimit) /* both in extDict */) {
             const BYTE* const mBase = ( (dictMode != ZSTD_extDict)
                                      || (matchIndex+matchLength >= dictLimit)) ?
                                         base : dictBase;
             assert( (matchIndex+matchLength >= dictLimit)   /* might be wrong if extDict is incorrectly set to 0 */
-                 || (current < dictLimit) );
+                 || (curr < dictLimit) );
             match = mBase + matchIndex;
             matchLength += ZSTD_count(ip+matchLength, match+matchLength, iend);
         } else {
@@ -119,7 +120,7 @@ ZSTD_insertDUBT1(ZSTD_matchState_t* ms,
         }
 
         DEBUGLOG(8, "ZSTD_insertDUBT1: comparing %u with %u : found %u common bytes ",
-                    current, matchIndex, (U32)matchLength);
+                    curr, matchIndex, (U32)matchLength);
 
         if (ip+matchLength == iend) {   /* equal : no way to know if inf or sup */
             break;   /* drop , to guarantee consistency ; miss a bit of compression, but other solutions can corrupt tree */
@@ -151,7 +152,7 @@ ZSTD_insertDUBT1(ZSTD_matchState_t* ms,
 
 static size_t
 ZSTD_DUBT_findBetterDictMatch (
-        ZSTD_matchState_t* ms,
+        const ZSTD_matchState_t* ms,
         const BYTE* const ip, const BYTE* const iend,
         size_t* offsetPtr,
         size_t bestLength,
@@ -168,7 +169,7 @@ ZSTD_DUBT_findBetterDictMatch (
 
     const BYTE* const base = ms->window.base;
     const BYTE* const prefixStart = base + ms->window.dictLimit;
-    U32         const current = (U32)(ip-base);
+    U32         const curr = (U32)(ip-base);
     const BYTE* const dictBase = dms->window.base;
     const BYTE* const dictEnd = dms->window.nextSrc;
     U32         const dictHighLimit = (U32)(dms->window.nextSrc - dms->window.base);
@@ -185,7 +186,7 @@ ZSTD_DUBT_findBetterDictMatch (
     (void)dictMode;
     assert(dictMode == ZSTD_dictMatchState);
 
-    while (nbCompares-- && (dictMatchIndex > dictLowLimit)) {
+    for (; nbCompares && (dictMatchIndex > dictLowLimit); --nbCompares) {
         U32* const nextPtr = dictBt + 2*(dictMatchIndex & btMask);
         size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
         const BYTE* match = dictBase + dictMatchIndex;
@@ -195,10 +196,10 @@ ZSTD_DUBT_findBetterDictMatch (
 
         if (matchLength > bestLength) {
             U32 matchIndex = dictMatchIndex + dictIndexDelta;
-            if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(current-matchIndex+1) - ZSTD_highbit32((U32)offsetPtr[0]+1)) ) {
+            if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(curr-matchIndex+1) - ZSTD_highbit32((U32)offsetPtr[0]+1)) ) {
                 DEBUGLOG(9, "ZSTD_DUBT_findBetterDictMatch(%u) : found better match length %u -> %u and offsetCode %u -> %u (dictMatchIndex %u, matchIndex %u)",
-                    current, (U32)bestLength, (U32)matchLength, (U32)*offsetPtr, ZSTD_REP_MOVE + current - matchIndex, dictMatchIndex, matchIndex);
-                bestLength = matchLength, *offsetPtr = ZSTD_REP_MOVE + current - matchIndex;
+                    curr, (U32)bestLength, (U32)matchLength, (U32)*offsetPtr, OFFSET_TO_OFFBASE(curr - matchIndex), dictMatchIndex, matchIndex);
+                bestLength = matchLength, *offsetPtr = OFFSET_TO_OFFBASE(curr - matchIndex);
             }
             if (ip+matchLength == iend) {   /* reached end of input : ip[matchLength] is not valid, no way to know if it's larger or smaller than match */
                 break;   /* drop, to guarantee consistency (miss a little bit of compression) */
@@ -218,9 +219,9 @@ ZSTD_DUBT_findBetterDictMatch (
     }
 
     if (bestLength >= MINMATCH) {
-        U32 const mIndex = current - ((U32)*offsetPtr - ZSTD_REP_MOVE); (void)mIndex;
+        U32 const mIndex = curr - (U32)OFFBASE_TO_OFFSET(*offsetPtr); (void)mIndex;
         DEBUGLOG(8, "ZSTD_DUBT_findBetterDictMatch(%u) : found match of length %u and offsetCode %u (pos %u)",
-                    current, (U32)bestLength, (U32)*offsetPtr, mIndex);
+                    curr, (U32)bestLength, (U32)*offsetPtr, mIndex);
     }
     return bestLength;
 
@@ -230,7 +231,7 @@ ZSTD_DUBT_findBetterDictMatch (
 static size_t
 ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms,
                         const BYTE* const ip, const BYTE* const iend,
-                        size_t* offsetPtr,
+                        size_t* offBasePtr,
                         U32 const mls,
                         const ZSTD_dictMode_e dictMode)
 {
@@ -241,13 +242,13 @@ ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms,
     U32          matchIndex  = hashTable[h];
 
     const BYTE* const base = ms->window.base;
-    U32    const current = (U32)(ip-base);
-    U32    const windowLow = ZSTD_getLowestMatchIndex(ms, current, cParams->windowLog);
+    U32    const curr = (U32)(ip-base);
+    U32    const windowLow = ZSTD_getLowestMatchIndex(ms, curr, cParams->windowLog);
 
     U32*   const bt = ms->chainTable;
     U32    const btLog  = cParams->chainLog - 1;
     U32    const btMask = (1 << btLog) - 1;
-    U32    const btLow = (btMask >= current) ? 0 : current - btMask;
+    U32    const btLow = (btMask >= curr) ? 0 : curr - btMask;
     U32    const unsortLimit = MAX(btLow, windowLow);
 
     U32*         nextCandidate = bt + 2*(matchIndex&btMask);
@@ -256,8 +257,9 @@ ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms,
     U32          nbCandidates = nbCompares;
     U32          previousCandidate = 0;
 
-    DEBUGLOG(7, "ZSTD_DUBT_findBestMatch (%u) ", current);
+    DEBUGLOG(7, "ZSTD_DUBT_findBestMatch (%u) ", curr);
     assert(ip <= iend-8);   /* required for h calculation */
+    assert(dictMode != ZSTD_dedicatedDictSearch);
 
     /* reach end of unsorted candidates list */
     while ( (matchIndex > unsortLimit)
@@ -299,16 +301,16 @@ ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms,
         const U32 dictLimit = ms->window.dictLimit;
         const BYTE* const dictEnd = dictBase + dictLimit;
         const BYTE* const prefixStart = base + dictLimit;
-        U32* smallerPtr = bt + 2*(current&btMask);
-        U32* largerPtr  = bt + 2*(current&btMask) + 1;
-        U32 matchEndIdx = current + 8 + 1;
+        U32* smallerPtr = bt + 2*(curr&btMask);
+        U32* largerPtr  = bt + 2*(curr&btMask) + 1;
+        U32 matchEndIdx = curr + 8 + 1;
         U32 dummy32;   /* to be nullified at the end */
         size_t bestLength = 0;
 
         matchIndex  = hashTable[h];
-        hashTable[h] = current;   /* Update Hash Table */
+        hashTable[h] = curr;   /* Update Hash Table */
 
-        while (nbCompares-- && (matchIndex > windowLow)) {
+        for (; nbCompares && (matchIndex > windowLow); --nbCompares) {
             U32* const nextPtr = bt + 2*(matchIndex & btMask);
             size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
             const BYTE* match;
@@ -326,8 +328,8 @@ ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms,
             if (matchLength > bestLength) {
                 if (matchLength > matchEndIdx - matchIndex)
                     matchEndIdx = matchIndex + (U32)matchLength;
-                if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(current-matchIndex+1) - ZSTD_highbit32((U32)offsetPtr[0]+1)) )
-                    bestLength = matchLength, *offsetPtr = ZSTD_REP_MOVE + current - matchIndex;
+                if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(curr - matchIndex + 1) - ZSTD_highbit32((U32)*offBasePtr)) )
+                    bestLength = matchLength, *offBasePtr = OFFSET_TO_OFFBASE(curr - matchIndex);
                 if (ip+matchLength == iend) {   /* equal : no way to know if inf or sup */
                     if (dictMode == ZSTD_dictMatchState) {
                         nbCompares = 0; /* in addition to avoiding checking any
@@ -356,19 +358,20 @@ ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms,
 
         *smallerPtr = *largerPtr = 0;
 
+        assert(nbCompares <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */
         if (dictMode == ZSTD_dictMatchState && nbCompares) {
             bestLength = ZSTD_DUBT_findBetterDictMatch(
                     ms, ip, iend,
-                    offsetPtr, bestLength, nbCompares,
+                    offBasePtr, bestLength, nbCompares,
                     mls, dictMode);
         }
 
-        assert(matchEndIdx > current+8); /* ensure nextToUpdate is increased */
+        assert(matchEndIdx > curr+8); /* ensure nextToUpdate is increased */
         ms->nextToUpdate = matchEndIdx - 8;   /* skip repetitive patterns */
         if (bestLength >= MINMATCH) {
-            U32 const mIndex = current - ((U32)*offsetPtr - ZSTD_REP_MOVE); (void)mIndex;
+            U32 const mIndex = curr - (U32)OFFBASE_TO_OFFSET(*offBasePtr); (void)mIndex;
             DEBUGLOG(8, "ZSTD_DUBT_findBestMatch(%u) : found match of length %u and offsetCode %u (pos %u)",
-                        current, (U32)bestLength, (U32)*offsetPtr, mIndex);
+                        curr, (U32)bestLength, (U32)*offBasePtr, mIndex);
         }
         return bestLength;
     }
@@ -379,66 +382,232 @@ ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms,
 FORCE_INLINE_TEMPLATE size_t
 ZSTD_BtFindBestMatch( ZSTD_matchState_t* ms,
                 const BYTE* const ip, const BYTE* const iLimit,
-                      size_t* offsetPtr,
+                      size_t* offBasePtr,
                 const U32 mls /* template */,
                 const ZSTD_dictMode_e dictMode)
 {
     DEBUGLOG(7, "ZSTD_BtFindBestMatch");
     if (ip < ms->window.base + ms->nextToUpdate) return 0;   /* skipped area */
     ZSTD_updateDUBT(ms, ip, iLimit, mls);
-    return ZSTD_DUBT_findBestMatch(ms, ip, iLimit, offsetPtr, mls, dictMode);
+    return ZSTD_DUBT_findBestMatch(ms, ip, iLimit, offBasePtr, mls, dictMode);
 }
 
+/***********************************
+* Dedicated dict search
+***********************************/
 
-static size_t
-ZSTD_BtFindBestMatch_selectMLS (  ZSTD_matchState_t* ms,
-                            const BYTE* ip, const BYTE* const iLimit,
-                                  size_t* offsetPtr)
+void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t* ms, const BYTE* const ip)
 {
-    switch(ms->cParams.minMatch)
+    const BYTE* const base = ms->window.base;
+    U32 const target = (U32)(ip - base);
+    U32* const hashTable = ms->hashTable;
+    U32* const chainTable = ms->chainTable;
+    U32 const chainSize = 1 << ms->cParams.chainLog;
+    U32 idx = ms->nextToUpdate;
+    U32 const minChain = chainSize < target - idx ? target - chainSize : idx;
+    U32 const bucketSize = 1 << ZSTD_LAZY_DDSS_BUCKET_LOG;
+    U32 const cacheSize = bucketSize - 1;
+    U32 const chainAttempts = (1 << ms->cParams.searchLog) - cacheSize;
+    U32 const chainLimit = chainAttempts > 255 ? 255 : chainAttempts;
+
+    /* We know the hashtable is oversized by a factor of `bucketSize`.
+     * We are going to temporarily pretend `bucketSize == 1`, keeping only a
+     * single entry. We will use the rest of the space to construct a temporary
+     * chaintable.
+     */
+    U32 const hashLog = ms->cParams.hashLog - ZSTD_LAZY_DDSS_BUCKET_LOG;
+    U32* const tmpHashTable = hashTable;
+    U32* const tmpChainTable = hashTable + ((size_t)1 << hashLog);
+    U32 const tmpChainSize = (U32)((1 << ZSTD_LAZY_DDSS_BUCKET_LOG) - 1) << hashLog;
+    U32 const tmpMinChain = tmpChainSize < target ? target - tmpChainSize : idx;
+    U32 hashIdx;
+
+    assert(ms->cParams.chainLog <= 24);
+    assert(ms->cParams.hashLog > ms->cParams.chainLog);
+    assert(idx != 0);
+    assert(tmpMinChain <= minChain);
+
+    /* fill conventional hash table and conventional chain table */
+    for ( ; idx < target; idx++) {
+        U32 const h = (U32)ZSTD_hashPtr(base + idx, hashLog, ms->cParams.minMatch);
+        if (idx >= tmpMinChain) {
+            tmpChainTable[idx - tmpMinChain] = hashTable[h];
+        }
+        tmpHashTable[h] = idx;
+    }
+
+    /* sort chains into ddss chain table */
     {
-    default : /* includes case 3 */
-    case 4 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 4, ZSTD_noDict);
-    case 5 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 5, ZSTD_noDict);
-    case 7 :
-    case 6 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 6, ZSTD_noDict);
+        U32 chainPos = 0;
+        for (hashIdx = 0; hashIdx < (1U << hashLog); hashIdx++) {
+            U32 count;
+            U32 countBeyondMinChain = 0;
+            U32 i = tmpHashTable[hashIdx];
+            for (count = 0; i >= tmpMinChain && count < cacheSize; count++) {
+                /* skip through the chain to the first position that won't be
+                 * in the hash cache bucket */
+                if (i < minChain) {
+                    countBeyondMinChain++;
+                }
+                i = tmpChainTable[i - tmpMinChain];
+            }
+            if (count == cacheSize) {
+                for (count = 0; count < chainLimit;) {
+                    if (i < minChain) {
+                        if (!i || ++countBeyondMinChain > cacheSize) {
+                            /* only allow pulling `cacheSize` number of entries
+                             * into the cache or chainTable beyond `minChain`,
+                             * to replace the entries pulled out of the
+                             * chainTable into the cache. This lets us reach
+                             * back further without increasing the total number
+                             * of entries in the chainTable, guaranteeing the
+                             * DDSS chain table will fit into the space
+                             * allocated for the regular one. */
+                            break;
+                        }
+                    }
+                    chainTable[chainPos++] = i;
+                    count++;
+                    if (i < tmpMinChain) {
+                        break;
+                    }
+                    i = tmpChainTable[i - tmpMinChain];
+                }
+            } else {
+                count = 0;
+            }
+            if (count) {
+                tmpHashTable[hashIdx] = ((chainPos - count) << 8) + count;
+            } else {
+                tmpHashTable[hashIdx] = 0;
+            }
+        }
+        assert(chainPos <= chainSize); /* I believe this is guaranteed... */
+    }
+
+    /* move chain pointers into the last entry of each hash bucket */
+    for (hashIdx = (1 << hashLog); hashIdx; ) {
+        U32 const bucketIdx = --hashIdx << ZSTD_LAZY_DDSS_BUCKET_LOG;
+        U32 const chainPackedPointer = tmpHashTable[hashIdx];
+        U32 i;
+        for (i = 0; i < cacheSize; i++) {
+            hashTable[bucketIdx + i] = 0;
+        }
+        hashTable[bucketIdx + bucketSize - 1] = chainPackedPointer;
+    }
+
+    /* fill the buckets of the hash table */
+    for (idx = ms->nextToUpdate; idx < target; idx++) {
+        U32 const h = (U32)ZSTD_hashPtr(base + idx, hashLog, ms->cParams.minMatch)
+                   << ZSTD_LAZY_DDSS_BUCKET_LOG;
+        U32 i;
+        /* Shift hash cache down 1. */
+        for (i = cacheSize - 1; i; i--)
+            hashTable[h + i] = hashTable[h + i - 1];
+        hashTable[h] = idx;
     }
+
+    ms->nextToUpdate = target;
 }
 
+/* Returns the longest match length found in the dedicated dict search structure.
+ * If none are longer than the argument ml, then ml will be returned.
+ */
+FORCE_INLINE_TEMPLATE
+size_t ZSTD_dedicatedDictSearch_lazy_search(size_t* offsetPtr, size_t ml, U32 nbAttempts,
+                                            const ZSTD_matchState_t* const dms,
+                                            const BYTE* const ip, const BYTE* const iLimit,
+                                            const BYTE* const prefixStart, const U32 curr,
+                                            const U32 dictLimit, const size_t ddsIdx) {
+    const U32 ddsLowestIndex  = dms->window.dictLimit;
+    const BYTE* const ddsBase = dms->window.base;
+    const BYTE* const ddsEnd  = dms->window.nextSrc;
+    const U32 ddsSize         = (U32)(ddsEnd - ddsBase);
+    const U32 ddsIndexDelta   = dictLimit - ddsSize;
+    const U32 bucketSize      = (1 << ZSTD_LAZY_DDSS_BUCKET_LOG);
+    const U32 bucketLimit     = nbAttempts < bucketSize - 1 ? nbAttempts : bucketSize - 1;
+    U32 ddsAttempt;
+    U32 matchIndex;
+
+    for (ddsAttempt = 0; ddsAttempt < bucketSize - 1; ddsAttempt++) {
+        PREFETCH_L1(ddsBase + dms->hashTable[ddsIdx + ddsAttempt]);
+    }
 
-static size_t ZSTD_BtFindBestMatch_dictMatchState_selectMLS (
-                        ZSTD_matchState_t* ms,
-                        const BYTE* ip, const BYTE* const iLimit,
-                        size_t* offsetPtr)
-{
-    switch(ms->cParams.minMatch)
     {
-    default : /* includes case 3 */
-    case 4 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 4, ZSTD_dictMatchState);
-    case 5 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 5, ZSTD_dictMatchState);
-    case 7 :
-    case 6 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 6, ZSTD_dictMatchState);
+        U32 const chainPackedPointer = dms->hashTable[ddsIdx + bucketSize - 1];
+        U32 const chainIndex = chainPackedPointer >> 8;
+
+        PREFETCH_L1(&dms->chainTable[chainIndex]);
     }
-}
 
+    for (ddsAttempt = 0; ddsAttempt < bucketLimit; ddsAttempt++) {
+        size_t currentMl=0;
+        const BYTE* match;
+        matchIndex = dms->hashTable[ddsIdx + ddsAttempt];
+        match = ddsBase + matchIndex;
+
+        if (!matchIndex) {
+            return ml;
+        }
+
+        /* guaranteed by table construction */
+        (void)ddsLowestIndex;
+        assert(matchIndex >= ddsLowestIndex);
+        assert(match+4 <= ddsEnd);
+        if (MEM_read32(match) == MEM_read32(ip)) {
+            /* assumption : matchIndex <= dictLimit-4 (by table construction) */
+            currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, ddsEnd, prefixStart) + 4;
+        }
+
+        /* save best solution */
+        if (currentMl > ml) {
+            ml = currentMl;
+            *offsetPtr = OFFSET_TO_OFFBASE(curr - (matchIndex + ddsIndexDelta));
+            if (ip+currentMl == iLimit) {
+                /* best possible, avoids read overflow on next attempt */
+                return ml;
+            }
+        }
+    }
 
-static size_t ZSTD_BtFindBestMatch_extDict_selectMLS (
-                        ZSTD_matchState_t* ms,
-                        const BYTE* ip, const BYTE* const iLimit,
-                        size_t* offsetPtr)
-{
-    switch(ms->cParams.minMatch)
     {
-    default : /* includes case 3 */
-    case 4 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 4, ZSTD_extDict);
-    case 5 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 5, ZSTD_extDict);
-    case 7 :
-    case 6 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 6, ZSTD_extDict);
+        U32 const chainPackedPointer = dms->hashTable[ddsIdx + bucketSize - 1];
+        U32 chainIndex = chainPackedPointer >> 8;
+        U32 const chainLength = chainPackedPointer & 0xFF;
+        U32 const chainAttempts = nbAttempts - ddsAttempt;
+        U32 const chainLimit = chainAttempts > chainLength ? chainLength : chainAttempts;
+        U32 chainAttempt;
+
+        for (chainAttempt = 0 ; chainAttempt < chainLimit; chainAttempt++) {
+            PREFETCH_L1(ddsBase + dms->chainTable[chainIndex + chainAttempt]);
+        }
+
+        for (chainAttempt = 0 ; chainAttempt < chainLimit; chainAttempt++, chainIndex++) {
+            size_t currentMl=0;
+            const BYTE* match;
+            matchIndex = dms->chainTable[chainIndex];
+            match = ddsBase + matchIndex;
+
+            /* guaranteed by table construction */
+            assert(matchIndex >= ddsLowestIndex);
+            assert(match+4 <= ddsEnd);
+            if (MEM_read32(match) == MEM_read32(ip)) {
+                /* assumption : matchIndex <= dictLimit-4 (by table construction) */
+                currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, ddsEnd, prefixStart) + 4;
+            }
+
+            /* save best solution */
+            if (currentMl > ml) {
+                ml = currentMl;
+                *offsetPtr = OFFSET_TO_OFFBASE(curr - (matchIndex + ddsIndexDelta));
+                if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
+            }
+        }
     }
+    return ml;
 }
 
 
-
 /* *********************************
 *  Hash Chain
 ***********************************/
@@ -446,7 +615,7 @@ static size_t ZSTD_BtFindBestMatch_extDict_selectMLS (
 
 /* Update chains up to ip (excluded)
    Assumption : always within prefix (i.e. not within extDict) */
-static U32 ZSTD_insertAndFindFirstIndex_internal(
+FORCE_INLINE_TEMPLATE U32 ZSTD_insertAndFindFirstIndex_internal(
                         ZSTD_matchState_t* ms,
                         const ZSTD_compressionParameters* const cParams,
                         const BYTE* ip, U32 const mls)
@@ -475,10 +644,9 @@ U32 ZSTD_insertAndFindFirstIndex(ZSTD_matchState_t* ms, const BYTE* ip) {
     return ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, ms->cParams.minMatch);
 }
 
-
 /* inlining is important to hardwire a hot branch (template emulation) */
 FORCE_INLINE_TEMPLATE
-size_t ZSTD_HcFindBestMatch_generic (
+size_t ZSTD_HcFindBestMatch(
                         ZSTD_matchState_t* ms,
                         const BYTE* const ip, const BYTE* const iLimit,
                         size_t* offsetPtr,
@@ -493,25 +661,39 @@ size_t ZSTD_HcFindBestMatch_generic (
     const U32 dictLimit = ms->window.dictLimit;
     const BYTE* const prefixStart = base + dictLimit;
     const BYTE* const dictEnd = dictBase + dictLimit;
-    const U32 current = (U32)(ip-base);
+    const U32 curr = (U32)(ip-base);
     const U32 maxDistance = 1U << cParams->windowLog;
     const U32 lowestValid = ms->window.lowLimit;
-    const U32 withinMaxDistance = (current - lowestValid > maxDistance) ? current - maxDistance : lowestValid;
+    const U32 withinMaxDistance = (curr - lowestValid > maxDistance) ? curr - maxDistance : lowestValid;
     const U32 isDictionary = (ms->loadedDictEnd != 0);
     const U32 lowLimit = isDictionary ? lowestValid : withinMaxDistance;
-    const U32 minChain = current > chainSize ? current - chainSize : 0;
+    const U32 minChain = curr > chainSize ? curr - chainSize : 0;
     U32 nbAttempts = 1U << cParams->searchLog;
     size_t ml=4-1;
 
+    const ZSTD_matchState_t* const dms = ms->dictMatchState;
+    const U32 ddsHashLog = dictMode == ZSTD_dedicatedDictSearch
+                         ? dms->cParams.hashLog - ZSTD_LAZY_DDSS_BUCKET_LOG : 0;
+    const size_t ddsIdx = dictMode == ZSTD_dedicatedDictSearch
+                        ? ZSTD_hashPtr(ip, ddsHashLog, mls) << ZSTD_LAZY_DDSS_BUCKET_LOG : 0;
+
+    U32 matchIndex;
+
+    if (dictMode == ZSTD_dedicatedDictSearch) {
+        const U32* entry = &dms->hashTable[ddsIdx];
+        PREFETCH_L1(entry);
+    }
+
     /* HC4 match finder */
-    U32 matchIndex = ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, mls);
+    matchIndex = ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, mls);
 
-    for ( ; (matchIndex>lowLimit) & (nbAttempts>0) ; nbAttempts--) {
+    for ( ; (matchIndex>=lowLimit) & (nbAttempts>0) ; nbAttempts--) {
         size_t currentMl=0;
         if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) {
             const BYTE* const match = base + matchIndex;
             assert(matchIndex >= dictLimit);   /* ensures this is true if dictMode != ZSTD_extDict */
-            if (match[ml] == ip[ml])   /* potentially better */
+            /* read 4B starting from (match + ml + 1 - sizeof(U32)) */
+            if (MEM_read32(match + ml - 3) == MEM_read32(ip + ml - 3))   /* potentially better */
                 currentMl = ZSTD_count(ip, match, iLimit);
         } else {
             const BYTE* const match = dictBase + matchIndex;
@@ -523,7 +705,7 @@ size_t ZSTD_HcFindBestMatch_generic (
         /* save best solution */
         if (currentMl > ml) {
             ml = currentMl;
-            *offsetPtr = current - matchIndex + ZSTD_REP_MOVE;
+            *offsetPtr = OFFSET_TO_OFFBASE(curr - matchIndex);
             if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
         }
 
@@ -531,8 +713,11 @@ size_t ZSTD_HcFindBestMatch_generic (
         matchIndex = NEXT_IN_CHAIN(matchIndex, chainMask);
     }
 
-    if (dictMode == ZSTD_dictMatchState) {
-        const ZSTD_matchState_t* const dms = ms->dictMatchState;
+    assert(nbAttempts <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */
+    if (dictMode == ZSTD_dedicatedDictSearch) {
+        ml = ZSTD_dedicatedDictSearch_lazy_search(offsetPtr, ml, nbAttempts, dms,
+                                                  ip, iLimit, prefixStart, curr, dictLimit, ddsIdx);
+    } else if (dictMode == ZSTD_dictMatchState) {
         const U32* const dmsChainTable = dms->chainTable;
         const U32 dmsChainSize         = (1 << dms->cParams.chainLog);
         const U32 dmsChainMask         = dmsChainSize - 1;
@@ -545,7 +730,7 @@ size_t ZSTD_HcFindBestMatch_generic (
 
         matchIndex = dms->hashTable[ZSTD_hashPtr(ip, dms->cParams.hashLog, mls)];
 
-        for ( ; (matchIndex>dmsLowestIndex) & (nbAttempts>0) ; nbAttempts--) {
+        for ( ; (matchIndex>=dmsLowestIndex) & (nbAttempts>0) ; nbAttempts--) {
             size_t currentMl=0;
             const BYTE* const match = dmsBase + matchIndex;
             assert(match+4 <= dmsEnd);
@@ -555,11 +740,13 @@ size_t ZSTD_HcFindBestMatch_generic (
             /* save best solution */
             if (currentMl > ml) {
                 ml = currentMl;
-                *offsetPtr = current - (matchIndex + dmsIndexDelta) + ZSTD_REP_MOVE;
+                assert(curr > matchIndex + dmsIndexDelta);
+                *offsetPtr = OFFSET_TO_OFFBASE(curr - (matchIndex + dmsIndexDelta));
                 if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
             }
 
             if (matchIndex <= dmsMinChain) break;
+
             matchIndex = dmsChainTable[matchIndex & dmsChainMask];
         }
     }
@@ -567,59 +754,744 @@ size_t ZSTD_HcFindBestMatch_generic (
     return ml;
 }
 
+/* *********************************
+* (SIMD) Row-based matchfinder
+***********************************/
+/* Constants for row-based hash */
+#define ZSTD_ROW_HASH_TAG_OFFSET 16     /* byte offset of hashes in the match state's tagTable from the beginning of a row */
+#define ZSTD_ROW_HASH_TAG_MASK ((1u << ZSTD_ROW_HASH_TAG_BITS) - 1)
+#define ZSTD_ROW_HASH_MAX_ENTRIES 64    /* absolute maximum number of entries per row, for all configurations */
 
-FORCE_INLINE_TEMPLATE size_t ZSTD_HcFindBestMatch_selectMLS (
-                        ZSTD_matchState_t* ms,
-                        const BYTE* ip, const BYTE* const iLimit,
-                        size_t* offsetPtr)
+#define ZSTD_ROW_HASH_CACHE_MASK (ZSTD_ROW_HASH_CACHE_SIZE - 1)
+
+typedef U64 ZSTD_VecMask;   /* Clarifies when we are interacting with a U64 representing a mask of matches */
+
+/* ZSTD_VecMask_next():
+ * Starting from the LSB, returns the idx of the next non-zero bit.
+ * Basically counting the nb of trailing zeroes.
+ */
+MEM_STATIC U32 ZSTD_VecMask_next(ZSTD_VecMask val) {
+    return ZSTD_countTrailingZeros64(val);
+}
+
+/* ZSTD_rotateRight_*():
+ * Rotates a bitfield to the right by "count" bits.
+ * https://en.wikipedia.org/w/index.php?title=Circular_shift&oldid=991635599#Implementing_circular_shifts
+ */
+FORCE_INLINE_TEMPLATE
+U64 ZSTD_rotateRight_U64(U64 const value, U32 count) {
+    assert(count < 64);
+    count &= 0x3F; /* for fickle pattern recognition */
+    return (value >> count) | (U64)(value << ((0U - count) & 0x3F));
+}
+
+FORCE_INLINE_TEMPLATE
+U32 ZSTD_rotateRight_U32(U32 const value, U32 count) {
+    assert(count < 32);
+    count &= 0x1F; /* for fickle pattern recognition */
+    return (value >> count) | (U32)(value << ((0U - count) & 0x1F));
+}
+
+FORCE_INLINE_TEMPLATE
+U16 ZSTD_rotateRight_U16(U16 const value, U32 count) {
+    assert(count < 16);
+    count &= 0x0F; /* for fickle pattern recognition */
+    return (value >> count) | (U16)(value << ((0U - count) & 0x0F));
+}
+
+/* ZSTD_row_nextIndex():
+ * Returns the next index to insert at within a tagTable row, and updates the "head"
+ * value to reflect the update. Essentially cycles backwards from [0, {entries per row})
+ */
+FORCE_INLINE_TEMPLATE U32 ZSTD_row_nextIndex(BYTE* const tagRow, U32 const rowMask) {
+  U32 const next = (*tagRow - 1) & rowMask;
+  *tagRow = (BYTE)next;
+  return next;
+}
+
+/* ZSTD_isAligned():
+ * Checks that a pointer is aligned to "align" bytes which must be a power of 2.
+ */
+MEM_STATIC int ZSTD_isAligned(void const* ptr, size_t align) {
+    assert((align & (align - 1)) == 0);
+    return (((size_t)ptr) & (align - 1)) == 0;
+}
+
+/* ZSTD_row_prefetch():
+ * Performs prefetching for the hashTable and tagTable at a given row.
+ */
+FORCE_INLINE_TEMPLATE void ZSTD_row_prefetch(U32 const* hashTable, U16 const* tagTable, U32 const relRow, U32 const rowLog) {
+    PREFETCH_L1(hashTable + relRow);
+    if (rowLog >= 5) {
+        PREFETCH_L1(hashTable + relRow + 16);
+        /* Note: prefetching more of the hash table does not appear to be beneficial for 128-entry rows */
+    }
+    PREFETCH_L1(tagTable + relRow);
+    if (rowLog == 6) {
+        PREFETCH_L1(tagTable + relRow + 32);
+    }
+    assert(rowLog == 4 || rowLog == 5 || rowLog == 6);
+    assert(ZSTD_isAligned(hashTable + relRow, 64));                 /* prefetched hash row always 64-byte aligned */
+    assert(ZSTD_isAligned(tagTable + relRow, (size_t)1 << rowLog)); /* prefetched tagRow sits on correct multiple of bytes (32,64,128) */
+}
+
+/* ZSTD_row_fillHashCache():
+ * Fill up the hash cache starting at idx, prefetching up to ZSTD_ROW_HASH_CACHE_SIZE entries,
+ * but not beyond iLimit.
+ */
+FORCE_INLINE_TEMPLATE void ZSTD_row_fillHashCache(ZSTD_matchState_t* ms, const BYTE* base,
+                                   U32 const rowLog, U32 const mls,
+                                   U32 idx, const BYTE* const iLimit)
 {
-    switch(ms->cParams.minMatch)
-    {
-    default : /* includes case 3 */
-    case 4 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 4, ZSTD_noDict);
-    case 5 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 5, ZSTD_noDict);
-    case 7 :
-    case 6 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 6, ZSTD_noDict);
+    U32 const* const hashTable = ms->hashTable;
+    U16 const* const tagTable = ms->tagTable;
+    U32 const hashLog = ms->rowHashLog;
+    U32 const maxElemsToPrefetch = (base + idx) > iLimit ? 0 : (U32)(iLimit - (base + idx) + 1);
+    U32 const lim = idx + MIN(ZSTD_ROW_HASH_CACHE_SIZE, maxElemsToPrefetch);
+
+    for (; idx < lim; ++idx) {
+        U32 const hash = (U32)ZSTD_hashPtr(base + idx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls);
+        U32 const row = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
+        ZSTD_row_prefetch(hashTable, tagTable, row, rowLog);
+        ms->hashCache[idx & ZSTD_ROW_HASH_CACHE_MASK] = hash;
     }
+
+    DEBUGLOG(6, "ZSTD_row_fillHashCache(): [%u %u %u %u %u %u %u %u]", ms->hashCache[0], ms->hashCache[1],
+                                                     ms->hashCache[2], ms->hashCache[3], ms->hashCache[4],
+                                                     ms->hashCache[5], ms->hashCache[6], ms->hashCache[7]);
 }
 
+/* ZSTD_row_nextCachedHash():
+ * Returns the hash of base + idx, and replaces the hash in the hash cache with the byte at
+ * base + idx + ZSTD_ROW_HASH_CACHE_SIZE. Also prefetches the appropriate rows from hashTable and tagTable.
+ */
+FORCE_INLINE_TEMPLATE U32 ZSTD_row_nextCachedHash(U32* cache, U32 const* hashTable,
+                                                  U16 const* tagTable, BYTE const* base,
+                                                  U32 idx, U32 const hashLog,
+                                                  U32 const rowLog, U32 const mls)
+{
+    U32 const newHash = (U32)ZSTD_hashPtr(base+idx+ZSTD_ROW_HASH_CACHE_SIZE, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls);
+    U32 const row = (newHash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
+    ZSTD_row_prefetch(hashTable, tagTable, row, rowLog);
+    {   U32 const hash = cache[idx & ZSTD_ROW_HASH_CACHE_MASK];
+        cache[idx & ZSTD_ROW_HASH_CACHE_MASK] = newHash;
+        return hash;
+    }
+}
 
-static size_t ZSTD_HcFindBestMatch_dictMatchState_selectMLS (
-                        ZSTD_matchState_t* ms,
-                        const BYTE* ip, const BYTE* const iLimit,
-                        size_t* offsetPtr)
+/* ZSTD_row_update_internalImpl():
+ * Updates the hash table with positions starting from updateStartIdx until updateEndIdx.
+ */
+FORCE_INLINE_TEMPLATE void ZSTD_row_update_internalImpl(ZSTD_matchState_t* ms,
+                                                        U32 updateStartIdx, U32 const updateEndIdx,
+                                                        U32 const mls, U32 const rowLog,
+                                                        U32 const rowMask, U32 const useCache)
 {
-    switch(ms->cParams.minMatch)
-    {
-    default : /* includes case 3 */
-    case 4 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 4, ZSTD_dictMatchState);
-    case 5 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 5, ZSTD_dictMatchState);
-    case 7 :
-    case 6 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 6, ZSTD_dictMatchState);
+    U32* const hashTable = ms->hashTable;
+    U16* const tagTable = ms->tagTable;
+    U32 const hashLog = ms->rowHashLog;
+    const BYTE* const base = ms->window.base;
+
+    DEBUGLOG(6, "ZSTD_row_update_internalImpl(): updateStartIdx=%u, updateEndIdx=%u", updateStartIdx, updateEndIdx);
+    for (; updateStartIdx < updateEndIdx; ++updateStartIdx) {
+        U32 const hash = useCache ? ZSTD_row_nextCachedHash(ms->hashCache, hashTable, tagTable, base, updateStartIdx, hashLog, rowLog, mls)
+                                  : (U32)ZSTD_hashPtr(base + updateStartIdx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls);
+        U32 const relRow = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
+        U32* const row = hashTable + relRow;
+        BYTE* tagRow = (BYTE*)(tagTable + relRow);  /* Though tagTable is laid out as a table of U16, each tag is only 1 byte.
+                                                       Explicit cast allows us to get exact desired position within each row */
+        U32 const pos = ZSTD_row_nextIndex(tagRow, rowMask);
+
+        assert(hash == ZSTD_hashPtr(base + updateStartIdx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls));
+        ((BYTE*)tagRow)[pos + ZSTD_ROW_HASH_TAG_OFFSET] = hash & ZSTD_ROW_HASH_TAG_MASK;
+        row[pos] = updateStartIdx;
+    }
+}
+
+/* ZSTD_row_update_internal():
+ * Inserts the byte at ip into the appropriate position in the hash table, and updates ms->nextToUpdate.
+ * Skips sections of long matches as is necessary.
+ */
+FORCE_INLINE_TEMPLATE void ZSTD_row_update_internal(ZSTD_matchState_t* ms, const BYTE* ip,
+                                                    U32 const mls, U32 const rowLog,
+                                                    U32 const rowMask, U32 const useCache)
+{
+    U32 idx = ms->nextToUpdate;
+    const BYTE* const base = ms->window.base;
+    const U32 target = (U32)(ip - base);
+    const U32 kSkipThreshold = 384;
+    const U32 kMaxMatchStartPositionsToUpdate = 96;
+    const U32 kMaxMatchEndPositionsToUpdate = 32;
+
+    if (useCache) {
+        /* Only skip positions when using hash cache, i.e.
+         * if we are loading a dict, don't skip anything.
+         * If we decide to skip, then we only update a set number
+         * of positions at the beginning and end of the match.
+         */
+        if (UNLIKELY(target - idx > kSkipThreshold)) {
+            U32 const bound = idx + kMaxMatchStartPositionsToUpdate;
+            ZSTD_row_update_internalImpl(ms, idx, bound, mls, rowLog, rowMask, useCache);
+            idx = target - kMaxMatchEndPositionsToUpdate;
+            ZSTD_row_fillHashCache(ms, base, rowLog, mls, idx, ip+1);
+        }
     }
+    assert(target >= idx);
+    ZSTD_row_update_internalImpl(ms, idx, target, mls, rowLog, rowMask, useCache);
+    ms->nextToUpdate = target;
 }
 
+/* ZSTD_row_update():
+ * External wrapper for ZSTD_row_update_internal(). Used for filling the hashtable during dictionary
+ * processing.
+ */
+void ZSTD_row_update(ZSTD_matchState_t* const ms, const BYTE* ip) {
+    const U32 rowLog = BOUNDED(4, ms->cParams.searchLog, 6);
+    const U32 rowMask = (1u << rowLog) - 1;
+    const U32 mls = MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */);
+
+    DEBUGLOG(5, "ZSTD_row_update(), rowLog=%u", rowLog);
+    ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 0 /* don't use cache */);
+}
 
-FORCE_INLINE_TEMPLATE size_t ZSTD_HcFindBestMatch_extDict_selectMLS (
+/* Returns the mask width of bits group of which will be set to 1. Given not all
+ * architectures have easy movemask instruction, this helps to iterate over
+ * groups of bits easier and faster.
+ */
+FORCE_INLINE_TEMPLATE U32
+ZSTD_row_matchMaskGroupWidth(const U32 rowEntries)
+{
+    assert((rowEntries == 16) || (rowEntries == 32) || rowEntries == 64);
+    assert(rowEntries <= ZSTD_ROW_HASH_MAX_ENTRIES);
+    (void)rowEntries;
+#if defined(ZSTD_ARCH_ARM_NEON)
+    /* NEON path only works for little endian */
+    if (!MEM_isLittleEndian()) {
+        return 1;
+    }
+    if (rowEntries == 16) {
+        return 4;
+    }
+    if (rowEntries == 32) {
+        return 2;
+    }
+    if (rowEntries == 64) {
+        return 1;
+    }
+#endif
+    return 1;
+}
+
+#if defined(ZSTD_ARCH_X86_SSE2)
+FORCE_INLINE_TEMPLATE ZSTD_VecMask
+ZSTD_row_getSSEMask(int nbChunks, const BYTE* const src, const BYTE tag, const U32 head)
+{
+    const __m128i comparisonMask = _mm_set1_epi8((char)tag);
+    int matches[4] = {0};
+    int i;
+    assert(nbChunks == 1 || nbChunks == 2 || nbChunks == 4);
+    for (i=0; i<nbChunks; i++) {
+        const __m128i chunk = _mm_loadu_si128((const __m128i*)(const void*)(src + 16*i));
+        const __m128i equalMask = _mm_cmpeq_epi8(chunk, comparisonMask);
+        matches[i] = _mm_movemask_epi8(equalMask);
+    }
+    if (nbChunks == 1) return ZSTD_rotateRight_U16((U16)matches[0], head);
+    if (nbChunks == 2) return ZSTD_rotateRight_U32((U32)matches[1] << 16 | (U32)matches[0], head);
+    assert(nbChunks == 4);
+    return ZSTD_rotateRight_U64((U64)matches[3] << 48 | (U64)matches[2] << 32 | (U64)matches[1] << 16 | (U64)matches[0], head);
+}
+#endif
+
+#if defined(ZSTD_ARCH_ARM_NEON)
+FORCE_INLINE_TEMPLATE ZSTD_VecMask
+ZSTD_row_getNEONMask(const U32 rowEntries, const BYTE* const src, const BYTE tag, const U32 headGrouped)
+{
+    assert((rowEntries == 16) || (rowEntries == 32) || rowEntries == 64);
+    if (rowEntries == 16) {
+        /* vshrn_n_u16 shifts by 4 every u16 and narrows to 8 lower bits.
+         * After that groups of 4 bits represent the equalMask. We lower
+         * all bits except the highest in these groups by doing AND with
+         * 0x88 = 0b10001000.
+         */
+        const uint8x16_t chunk = vld1q_u8(src);
+        const uint16x8_t equalMask = vreinterpretq_u16_u8(vceqq_u8(chunk, vdupq_n_u8(tag)));
+        const uint8x8_t res = vshrn_n_u16(equalMask, 4);
+        const U64 matches = vget_lane_u64(vreinterpret_u64_u8(res), 0);
+        return ZSTD_rotateRight_U64(matches, headGrouped) & 0x8888888888888888ull;
+    } else if (rowEntries == 32) {
+        /* Same idea as with rowEntries == 16 but doing AND with
+         * 0x55 = 0b01010101.
+         */
+        const uint16x8x2_t chunk = vld2q_u16((const uint16_t*)(const void*)src);
+        const uint8x16_t chunk0 = vreinterpretq_u8_u16(chunk.val[0]);
+        const uint8x16_t chunk1 = vreinterpretq_u8_u16(chunk.val[1]);
+        const uint8x16_t dup = vdupq_n_u8(tag);
+        const uint8x8_t t0 = vshrn_n_u16(vreinterpretq_u16_u8(vceqq_u8(chunk0, dup)), 6);
+        const uint8x8_t t1 = vshrn_n_u16(vreinterpretq_u16_u8(vceqq_u8(chunk1, dup)), 6);
+        const uint8x8_t res = vsli_n_u8(t0, t1, 4);
+        const U64 matches = vget_lane_u64(vreinterpret_u64_u8(res), 0) ;
+        return ZSTD_rotateRight_U64(matches, headGrouped) & 0x5555555555555555ull;
+    } else { /* rowEntries == 64 */
+        const uint8x16x4_t chunk = vld4q_u8(src);
+        const uint8x16_t dup = vdupq_n_u8(tag);
+        const uint8x16_t cmp0 = vceqq_u8(chunk.val[0], dup);
+        const uint8x16_t cmp1 = vceqq_u8(chunk.val[1], dup);
+        const uint8x16_t cmp2 = vceqq_u8(chunk.val[2], dup);
+        const uint8x16_t cmp3 = vceqq_u8(chunk.val[3], dup);
+
+        const uint8x16_t t0 = vsriq_n_u8(cmp1, cmp0, 1);
+        const uint8x16_t t1 = vsriq_n_u8(cmp3, cmp2, 1);
+        const uint8x16_t t2 = vsriq_n_u8(t1, t0, 2);
+        const uint8x16_t t3 = vsriq_n_u8(t2, t2, 4);
+        const uint8x8_t t4 = vshrn_n_u16(vreinterpretq_u16_u8(t3), 4);
+        const U64 matches = vget_lane_u64(vreinterpret_u64_u8(t4), 0);
+        return ZSTD_rotateRight_U64(matches, headGrouped);
+    }
+}
+#endif
+
+/* Returns a ZSTD_VecMask (U64) that has the nth group (determined by
+ * ZSTD_row_matchMaskGroupWidth) of bits set to 1 if the newly-computed "tag"
+ * matches the hash at the nth position in a row of the tagTable.
+ * Each row is a circular buffer beginning at the value of "headGrouped". So we
+ * must rotate the "matches" bitfield to match up with the actual layout of the
+ * entries within the hashTable */
+FORCE_INLINE_TEMPLATE ZSTD_VecMask
+ZSTD_row_getMatchMask(const BYTE* const tagRow, const BYTE tag, const U32 headGrouped, const U32 rowEntries)
+{
+    const BYTE* const src = tagRow + ZSTD_ROW_HASH_TAG_OFFSET;
+    assert((rowEntries == 16) || (rowEntries == 32) || rowEntries == 64);
+    assert(rowEntries <= ZSTD_ROW_HASH_MAX_ENTRIES);
+    assert(ZSTD_row_matchMaskGroupWidth(rowEntries) * rowEntries <= sizeof(ZSTD_VecMask) * 8);
+
+#if defined(ZSTD_ARCH_X86_SSE2)
+
+    return ZSTD_row_getSSEMask(rowEntries / 16, src, tag, headGrouped);
+
+#else /* SW or NEON-LE */
+
+# if defined(ZSTD_ARCH_ARM_NEON)
+  /* This NEON path only works for little endian - otherwise use SWAR below */
+    if (MEM_isLittleEndian()) {
+        return ZSTD_row_getNEONMask(rowEntries, src, tag, headGrouped);
+    }
+# endif /* ZSTD_ARCH_ARM_NEON */
+    /* SWAR */
+    {   const int chunkSize = sizeof(size_t);
+        const size_t shiftAmount = ((chunkSize * 8) - chunkSize);
+        const size_t xFF = ~((size_t)0);
+        const size_t x01 = xFF / 0xFF;
+        const size_t x80 = x01 << 7;
+        const size_t splatChar = tag * x01;
+        ZSTD_VecMask matches = 0;
+        int i = rowEntries - chunkSize;
+        assert((sizeof(size_t) == 4) || (sizeof(size_t) == 8));
+        if (MEM_isLittleEndian()) { /* runtime check so have two loops */
+            const size_t extractMagic = (xFF / 0x7F) >> chunkSize;
+            do {
+                size_t chunk = MEM_readST(&src[i]);
+                chunk ^= splatChar;
+                chunk = (((chunk | x80) - x01) | chunk) & x80;
+                matches <<= chunkSize;
+                matches |= (chunk * extractMagic) >> shiftAmount;
+                i -= chunkSize;
+            } while (i >= 0);
+        } else { /* big endian: reverse bits during extraction */
+            const size_t msb = xFF ^ (xFF >> 1);
+            const size_t extractMagic = (msb / 0x1FF) | msb;
+            do {
+                size_t chunk = MEM_readST(&src[i]);
+                chunk ^= splatChar;
+                chunk = (((chunk | x80) - x01) | chunk) & x80;
+                matches <<= chunkSize;
+                matches |= ((chunk >> 7) * extractMagic) >> shiftAmount;
+                i -= chunkSize;
+            } while (i >= 0);
+        }
+        matches = ~matches;
+        if (rowEntries == 16) {
+            return ZSTD_rotateRight_U16((U16)matches, headGrouped);
+        } else if (rowEntries == 32) {
+            return ZSTD_rotateRight_U32((U32)matches, headGrouped);
+        } else {
+            return ZSTD_rotateRight_U64((U64)matches, headGrouped);
+        }
+    }
+#endif
+}
+
+/* The high-level approach of the SIMD row based match finder is as follows:
+ * - Figure out where to insert the new entry:
+ *      - Generate a hash from a byte along with an additional 1-byte "short hash". The additional byte is our "tag"
+ *      - The hashTable is effectively split into groups or "rows" of 16 or 32 entries of U32, and the hash determines
+ *        which row to insert into.
+ *      - Determine the correct position within the row to insert the entry into. Each row of 16 or 32 can
+ *        be considered as a circular buffer with a "head" index that resides in the tagTable.
+ *      - Also insert the "tag" into the equivalent row and position in the tagTable.
+ *          - Note: The tagTable has 17 or 33 1-byte entries per row, due to 16 or 32 tags, and 1 "head" entry.
+ *                  The 17 or 33 entry rows are spaced out to occur every 32 or 64 bytes, respectively,
+ *                  for alignment/performance reasons, leaving some bytes unused.
+ * - Use SIMD to efficiently compare the tags in the tagTable to the 1-byte "short hash" and
+ *   generate a bitfield that we can cycle through to check the collisions in the hash table.
+ * - Pick the longest match.
+ */
+FORCE_INLINE_TEMPLATE
+size_t ZSTD_RowFindBestMatch(
                         ZSTD_matchState_t* ms,
-                        const BYTE* ip, const BYTE* const iLimit,
-                        size_t* offsetPtr)
+                        const BYTE* const ip, const BYTE* const iLimit,
+                        size_t* offsetPtr,
+                        const U32 mls, const ZSTD_dictMode_e dictMode,
+                        const U32 rowLog)
 {
-    switch(ms->cParams.minMatch)
-    {
-    default : /* includes case 3 */
-    case 4 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 4, ZSTD_extDict);
-    case 5 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 5, ZSTD_extDict);
-    case 7 :
-    case 6 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 6, ZSTD_extDict);
+    U32* const hashTable = ms->hashTable;
+    U16* const tagTable = ms->tagTable;
+    U32* const hashCache = ms->hashCache;
+    const U32 hashLog = ms->rowHashLog;
+    const ZSTD_compressionParameters* const cParams = &ms->cParams;
+    const BYTE* const base = ms->window.base;
+    const BYTE* const dictBase = ms->window.dictBase;
+    const U32 dictLimit = ms->window.dictLimit;
+    const BYTE* const prefixStart = base + dictLimit;
+    const BYTE* const dictEnd = dictBase + dictLimit;
+    const U32 curr = (U32)(ip-base);
+    const U32 maxDistance = 1U << cParams->windowLog;
+    const U32 lowestValid = ms->window.lowLimit;
+    const U32 withinMaxDistance = (curr - lowestValid > maxDistance) ? curr - maxDistance : lowestValid;
+    const U32 isDictionary = (ms->loadedDictEnd != 0);
+    const U32 lowLimit = isDictionary ? lowestValid : withinMaxDistance;
+    const U32 rowEntries = (1U << rowLog);
+    const U32 rowMask = rowEntries - 1;
+    const U32 cappedSearchLog = MIN(cParams->searchLog, rowLog); /* nb of searches is capped at nb entries per row */
+    const U32 groupWidth = ZSTD_row_matchMaskGroupWidth(rowEntries);
+    U32 nbAttempts = 1U << cappedSearchLog;
+    size_t ml=4-1;
+
+    /* DMS/DDS variables that may be referenced laster */
+    const ZSTD_matchState_t* const dms = ms->dictMatchState;
+
+    /* Initialize the following variables to satisfy static analyzer */
+    size_t ddsIdx = 0;
+    U32 ddsExtraAttempts = 0; /* cctx hash tables are limited in searches, but allow extra searches into DDS */
+    U32 dmsTag = 0;
+    U32* dmsRow = NULL;
+    BYTE* dmsTagRow = NULL;
+
+    if (dictMode == ZSTD_dedicatedDictSearch) {
+        const U32 ddsHashLog = dms->cParams.hashLog - ZSTD_LAZY_DDSS_BUCKET_LOG;
+        {   /* Prefetch DDS hashtable entry */
+            ddsIdx = ZSTD_hashPtr(ip, ddsHashLog, mls) << ZSTD_LAZY_DDSS_BUCKET_LOG;
+            PREFETCH_L1(&dms->hashTable[ddsIdx]);
+        }
+        ddsExtraAttempts = cParams->searchLog > rowLog ? 1U << (cParams->searchLog - rowLog) : 0;
+    }
+
+    if (dictMode == ZSTD_dictMatchState) {
+        /* Prefetch DMS rows */
+        U32* const dmsHashTable = dms->hashTable;
+        U16* const dmsTagTable = dms->tagTable;
+        U32 const dmsHash = (U32)ZSTD_hashPtr(ip, dms->rowHashLog + ZSTD_ROW_HASH_TAG_BITS, mls);
+        U32 const dmsRelRow = (dmsHash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
+        dmsTag = dmsHash & ZSTD_ROW_HASH_TAG_MASK;
+        dmsTagRow = (BYTE*)(dmsTagTable + dmsRelRow);
+        dmsRow = dmsHashTable + dmsRelRow;
+        ZSTD_row_prefetch(dmsHashTable, dmsTagTable, dmsRelRow, rowLog);
+    }
+
+    /* Update the hashTable and tagTable up to (but not including) ip */
+    ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 1 /* useCache */);
+    {   /* Get the hash for ip, compute the appropriate row */
+        U32 const hash = ZSTD_row_nextCachedHash(hashCache, hashTable, tagTable, base, curr, hashLog, rowLog, mls);
+        U32 const relRow = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
+        U32 const tag = hash & ZSTD_ROW_HASH_TAG_MASK;
+        U32* const row = hashTable + relRow;
+        BYTE* tagRow = (BYTE*)(tagTable + relRow);
+        U32 const headGrouped = (*tagRow & rowMask) * groupWidth;
+        U32 matchBuffer[ZSTD_ROW_HASH_MAX_ENTRIES];
+        size_t numMatches = 0;
+        size_t currMatch = 0;
+        ZSTD_VecMask matches = ZSTD_row_getMatchMask(tagRow, (BYTE)tag, headGrouped, rowEntries);
+
+        /* Cycle through the matches and prefetch */
+        for (; (matches > 0) && (nbAttempts > 0); --nbAttempts, matches &= (matches - 1)) {
+            U32 const matchPos = ((headGrouped + ZSTD_VecMask_next(matches)) / groupWidth) & rowMask;
+            U32 const matchIndex = row[matchPos];
+            assert(numMatches < rowEntries);
+            if (matchIndex < lowLimit)
+                break;
+            if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) {
+                PREFETCH_L1(base + matchIndex);
+            } else {
+                PREFETCH_L1(dictBase + matchIndex);
+            }
+            matchBuffer[numMatches++] = matchIndex;
+        }
+
+        /* Speed opt: insert current byte into hashtable too. This allows us to avoid one iteration of the loop
+           in ZSTD_row_update_internal() at the next search. */
+        {
+            U32 const pos = ZSTD_row_nextIndex(tagRow, rowMask);
+            tagRow[pos + ZSTD_ROW_HASH_TAG_OFFSET] = (BYTE)tag;
+            row[pos] = ms->nextToUpdate++;
+        }
+
+        /* Return the longest match */
+        for (; currMatch < numMatches; ++currMatch) {
+            U32 const matchIndex = matchBuffer[currMatch];
+            size_t currentMl=0;
+            assert(matchIndex < curr);
+            assert(matchIndex >= lowLimit);
+
+            if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) {
+                const BYTE* const match = base + matchIndex;
+                assert(matchIndex >= dictLimit);   /* ensures this is true if dictMode != ZSTD_extDict */
+                /* read 4B starting from (match + ml + 1 - sizeof(U32)) */
+                if (MEM_read32(match + ml - 3) == MEM_read32(ip + ml - 3))   /* potentially better */
+                    currentMl = ZSTD_count(ip, match, iLimit);
+            } else {
+                const BYTE* const match = dictBase + matchIndex;
+                assert(match+4 <= dictEnd);
+                if (MEM_read32(match) == MEM_read32(ip))   /* assumption : matchIndex <= dictLimit-4 (by table construction) */
+                    currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dictEnd, prefixStart) + 4;
+            }
+
+            /* Save best solution */
+            if (currentMl > ml) {
+                ml = currentMl;
+                *offsetPtr = OFFSET_TO_OFFBASE(curr - matchIndex);
+                if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
+            }
+        }
     }
+
+    assert(nbAttempts <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */
+    if (dictMode == ZSTD_dedicatedDictSearch) {
+        ml = ZSTD_dedicatedDictSearch_lazy_search(offsetPtr, ml, nbAttempts + ddsExtraAttempts, dms,
+                                                  ip, iLimit, prefixStart, curr, dictLimit, ddsIdx);
+    } else if (dictMode == ZSTD_dictMatchState) {
+        /* TODO: Measure and potentially add prefetching to DMS */
+        const U32 dmsLowestIndex       = dms->window.dictLimit;
+        const BYTE* const dmsBase      = dms->window.base;
+        const BYTE* const dmsEnd       = dms->window.nextSrc;
+        const U32 dmsSize              = (U32)(dmsEnd - dmsBase);
+        const U32 dmsIndexDelta        = dictLimit - dmsSize;
+
+        {   U32 const headGrouped = (*dmsTagRow & rowMask) * groupWidth;
+            U32 matchBuffer[ZSTD_ROW_HASH_MAX_ENTRIES];
+            size_t numMatches = 0;
+            size_t currMatch = 0;
+            ZSTD_VecMask matches = ZSTD_row_getMatchMask(dmsTagRow, (BYTE)dmsTag, headGrouped, rowEntries);
+
+            for (; (matches > 0) && (nbAttempts > 0); --nbAttempts, matches &= (matches - 1)) {
+                U32 const matchPos = ((headGrouped + ZSTD_VecMask_next(matches)) / groupWidth) & rowMask;
+                U32 const matchIndex = dmsRow[matchPos];
+                if (matchIndex < dmsLowestIndex)
+                    break;
+                PREFETCH_L1(dmsBase + matchIndex);
+                matchBuffer[numMatches++] = matchIndex;
+            }
+
+            /* Return the longest match */
+            for (; currMatch < numMatches; ++currMatch) {
+                U32 const matchIndex = matchBuffer[currMatch];
+                size_t currentMl=0;
+                assert(matchIndex >= dmsLowestIndex);
+                assert(matchIndex < curr);
+
+                {   const BYTE* const match = dmsBase + matchIndex;
+                    assert(match+4 <= dmsEnd);
+                    if (MEM_read32(match) == MEM_read32(ip))
+                        currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dmsEnd, prefixStart) + 4;
+                }
+
+                if (currentMl > ml) {
+                    ml = currentMl;
+                    assert(curr > matchIndex + dmsIndexDelta);
+                    *offsetPtr = OFFSET_TO_OFFBASE(curr - (matchIndex + dmsIndexDelta));
+                    if (ip+currentMl == iLimit) break;
+                }
+            }
+        }
+    }
+    return ml;
 }
 
 
+/**
+ * Generate search functions templated on (dictMode, mls, rowLog).
+ * These functions are outlined for code size & compilation time.
+ * ZSTD_searchMax() dispatches to the correct implementation function.
+ *
+ * TODO: The start of the search function involves loading and calculating a
+ * bunch of constants from the ZSTD_matchState_t. These computations could be
+ * done in an initialization function, and saved somewhere in the match state.
+ * Then we could pass a pointer to the saved state instead of the match state,
+ * and avoid duplicate computations.
+ *
+ * TODO: Move the match re-winding into searchMax. This improves compression
+ * ratio, and unlocks further simplifications with the next TODO.
+ *
+ * TODO: Try moving the repcode search into searchMax. After the re-winding
+ * and repcode search are in searchMax, there is no more logic in the match
+ * finder loop that requires knowledge about the dictMode. So we should be
+ * able to avoid force inlining it, and we can join the extDict loop with
+ * the single segment loop. It should go in searchMax instead of its own
+ * function to avoid having multiple virtual function calls per search.
+ */
+
+#define ZSTD_BT_SEARCH_FN(dictMode, mls) ZSTD_BtFindBestMatch_##dictMode##_##mls
+#define ZSTD_HC_SEARCH_FN(dictMode, mls) ZSTD_HcFindBestMatch_##dictMode##_##mls
+#define ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog) ZSTD_RowFindBestMatch_##dictMode##_##mls##_##rowLog
+
+#define ZSTD_SEARCH_FN_ATTRS FORCE_NOINLINE
+
+#define GEN_ZSTD_BT_SEARCH_FN(dictMode, mls)                                           \
+    ZSTD_SEARCH_FN_ATTRS size_t ZSTD_BT_SEARCH_FN(dictMode, mls)(                      \
+            ZSTD_matchState_t* ms,                                                     \
+            const BYTE* ip, const BYTE* const iLimit,                                  \
+            size_t* offBasePtr)                                                        \
+    {                                                                                  \
+        assert(MAX(4, MIN(6, ms->cParams.minMatch)) == mls);                           \
+        return ZSTD_BtFindBestMatch(ms, ip, iLimit, offBasePtr, mls, ZSTD_##dictMode); \
+    }                                                                                  \
+
+#define GEN_ZSTD_HC_SEARCH_FN(dictMode, mls)                                          \
+    ZSTD_SEARCH_FN_ATTRS size_t ZSTD_HC_SEARCH_FN(dictMode, mls)(                     \
+            ZSTD_matchState_t* ms,                                                    \
+            const BYTE* ip, const BYTE* const iLimit,                                 \
+            size_t* offsetPtr)                                                        \
+    {                                                                                 \
+        assert(MAX(4, MIN(6, ms->cParams.minMatch)) == mls);                          \
+        return ZSTD_HcFindBestMatch(ms, ip, iLimit, offsetPtr, mls, ZSTD_##dictMode); \
+    }                                                                                 \
+
+#define GEN_ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog)                                          \
+    ZSTD_SEARCH_FN_ATTRS size_t ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog)(                     \
+            ZSTD_matchState_t* ms,                                                             \
+            const BYTE* ip, const BYTE* const iLimit,                                          \
+            size_t* offsetPtr)                                                                 \
+    {                                                                                          \
+        assert(MAX(4, MIN(6, ms->cParams.minMatch)) == mls);                                   \
+        assert(MAX(4, MIN(6, ms->cParams.searchLog)) == rowLog);                               \
+        return ZSTD_RowFindBestMatch(ms, ip, iLimit, offsetPtr, mls, ZSTD_##dictMode, rowLog); \
+    }                                                                                          \
+
+#define ZSTD_FOR_EACH_ROWLOG(X, dictMode, mls) \
+    X(dictMode, mls, 4)                        \
+    X(dictMode, mls, 5)                        \
+    X(dictMode, mls, 6)
+
+#define ZSTD_FOR_EACH_MLS_ROWLOG(X, dictMode) \
+    ZSTD_FOR_EACH_ROWLOG(X, dictMode, 4)      \
+    ZSTD_FOR_EACH_ROWLOG(X, dictMode, 5)      \
+    ZSTD_FOR_EACH_ROWLOG(X, dictMode, 6)
+
+#define ZSTD_FOR_EACH_MLS(X, dictMode) \
+    X(dictMode, 4)                     \
+    X(dictMode, 5)                     \
+    X(dictMode, 6)
+
+#define ZSTD_FOR_EACH_DICT_MODE(X, ...) \
+    X(__VA_ARGS__, noDict)              \
+    X(__VA_ARGS__, extDict)             \
+    X(__VA_ARGS__, dictMatchState)      \
+    X(__VA_ARGS__, dedicatedDictSearch)
+
+/* Generate row search fns for each combination of (dictMode, mls, rowLog) */
+ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS_ROWLOG, GEN_ZSTD_ROW_SEARCH_FN)
+/* Generate binary Tree search fns for each combination of (dictMode, mls) */
+ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS, GEN_ZSTD_BT_SEARCH_FN)
+/* Generate hash chain search fns for each combination of (dictMode, mls) */
+ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS, GEN_ZSTD_HC_SEARCH_FN)
+
+typedef enum { search_hashChain=0, search_binaryTree=1, search_rowHash=2 } searchMethod_e;
+
+#define GEN_ZSTD_CALL_BT_SEARCH_FN(dictMode, mls)                         \
+    case mls:                                                             \
+        return ZSTD_BT_SEARCH_FN(dictMode, mls)(ms, ip, iend, offsetPtr);
+#define GEN_ZSTD_CALL_HC_SEARCH_FN(dictMode, mls)                         \
+    case mls:                                                             \
+        return ZSTD_HC_SEARCH_FN(dictMode, mls)(ms, ip, iend, offsetPtr);
+#define GEN_ZSTD_CALL_ROW_SEARCH_FN(dictMode, mls, rowLog)                         \
+    case rowLog:                                                                   \
+        return ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog)(ms, ip, iend, offsetPtr);
+
+#define ZSTD_SWITCH_MLS(X, dictMode)   \
+    switch (mls) {                     \
+        ZSTD_FOR_EACH_MLS(X, dictMode) \
+    }
+
+#define ZSTD_SWITCH_ROWLOG(dictMode, mls)                                    \
+    case mls:                                                                \
+        switch (rowLog) {                                                    \
+            ZSTD_FOR_EACH_ROWLOG(GEN_ZSTD_CALL_ROW_SEARCH_FN, dictMode, mls) \
+        }                                                                    \
+        ZSTD_UNREACHABLE;                                                    \
+        break;
+
+#define ZSTD_SWITCH_SEARCH_METHOD(dictMode)                       \
+    switch (searchMethod) {                                       \
+        case search_hashChain:                                    \
+            ZSTD_SWITCH_MLS(GEN_ZSTD_CALL_HC_SEARCH_FN, dictMode) \
+            break;                                                \
+        case search_binaryTree:                                   \
+            ZSTD_SWITCH_MLS(GEN_ZSTD_CALL_BT_SEARCH_FN, dictMode) \
+            break;                                                \
+        case search_rowHash:                                      \
+            ZSTD_SWITCH_MLS(ZSTD_SWITCH_ROWLOG, dictMode)         \
+            break;                                                \
+    }                                                             \
+    ZSTD_UNREACHABLE;
+
+/**
+ * Searches for the longest match at @p ip.
+ * Dispatches to the correct implementation function based on the
+ * (searchMethod, dictMode, mls, rowLog). We use switch statements
+ * here instead of using an indirect function call through a function
+ * pointer because after Spectre and Meltdown mitigations, indirect
+ * function calls can be very costly, especially in the kernel.
+ *
+ * NOTE: dictMode and searchMethod should be templated, so those switch
+ * statements should be optimized out. Only the mls & rowLog switches
+ * should be left.
+ *
+ * @param ms The match state.
+ * @param ip The position to search at.
+ * @param iend The end of the input data.
+ * @param[out] offsetPtr Stores the match offset into this pointer.
+ * @param mls The minimum search length, in the range [4, 6].
+ * @param rowLog The row log (if applicable), in the range [4, 6].
+ * @param searchMethod The search method to use (templated).
+ * @param dictMode The dictMode (templated).
+ *
+ * @returns The length of the longest match found, or < mls if no match is found.
+ * If a match is found its offset is stored in @p offsetPtr.
+ */
+FORCE_INLINE_TEMPLATE size_t ZSTD_searchMax(
+    ZSTD_matchState_t* ms,
+    const BYTE* ip,
+    const BYTE* iend,
+    size_t* offsetPtr,
+    U32 const mls,
+    U32 const rowLog,
+    searchMethod_e const searchMethod,
+    ZSTD_dictMode_e const dictMode)
+{
+    if (dictMode == ZSTD_noDict) {
+        ZSTD_SWITCH_SEARCH_METHOD(noDict)
+    } else if (dictMode == ZSTD_extDict) {
+        ZSTD_SWITCH_SEARCH_METHOD(extDict)
+    } else if (dictMode == ZSTD_dictMatchState) {
+        ZSTD_SWITCH_SEARCH_METHOD(dictMatchState)
+    } else if (dictMode == ZSTD_dedicatedDictSearch) {
+        ZSTD_SWITCH_SEARCH_METHOD(dedicatedDictSearch)
+    }
+    ZSTD_UNREACHABLE;
+    return 0;
+}
+
 /* *******************************
 *  Common parser - lazy strategy
 *********************************/
-typedef enum { search_hashChain, search_binaryTree } searchMethod_e;
 
 FORCE_INLINE_TEMPLATE size_t
 ZSTD_compressBlock_lazy_generic(
@@ -633,53 +1505,51 @@ ZSTD_compressBlock_lazy_generic(
     const BYTE* ip = istart;
     const BYTE* anchor = istart;
     const BYTE* const iend = istart + srcSize;
-    const BYTE* const ilimit = iend - 8;
+    const BYTE* const ilimit = (searchMethod == search_rowHash) ? iend - 8 - ZSTD_ROW_HASH_CACHE_SIZE : iend - 8;
     const BYTE* const base = ms->window.base;
     const U32 prefixLowestIndex = ms->window.dictLimit;
     const BYTE* const prefixLowest = base + prefixLowestIndex;
+    const U32 mls = BOUNDED(4, ms->cParams.minMatch, 6);
+    const U32 rowLog = BOUNDED(4, ms->cParams.searchLog, 6);
 
-    typedef size_t (*searchMax_f)(
-                        ZSTD_matchState_t* ms,
-                        const BYTE* ip, const BYTE* iLimit, size_t* offsetPtr);
-    searchMax_f const searchMax = dictMode == ZSTD_dictMatchState ?
-        (searchMethod==search_binaryTree ? ZSTD_BtFindBestMatch_dictMatchState_selectMLS
-                                         : ZSTD_HcFindBestMatch_dictMatchState_selectMLS) :
-        (searchMethod==search_binaryTree ? ZSTD_BtFindBestMatch_selectMLS
-                                         : ZSTD_HcFindBestMatch_selectMLS);
-    U32 offset_1 = rep[0], offset_2 = rep[1], savedOffset=0;
+    U32 offset_1 = rep[0], offset_2 = rep[1];
+    U32 offsetSaved1 = 0, offsetSaved2 = 0;
 
+    const int isDMS = dictMode == ZSTD_dictMatchState;
+    const int isDDS = dictMode == ZSTD_dedicatedDictSearch;
+    const int isDxS = isDMS || isDDS;
     const ZSTD_matchState_t* const dms = ms->dictMatchState;
-    const U32 dictLowestIndex      = dictMode == ZSTD_dictMatchState ?
-                                     dms->window.dictLimit : 0;
-    const BYTE* const dictBase     = dictMode == ZSTD_dictMatchState ?
-                                     dms->window.base : NULL;
-    const BYTE* const dictLowest   = dictMode == ZSTD_dictMatchState ?
-                                     dictBase + dictLowestIndex : NULL;
-    const BYTE* const dictEnd      = dictMode == ZSTD_dictMatchState ?
-                                     dms->window.nextSrc : NULL;
-    const U32 dictIndexDelta       = dictMode == ZSTD_dictMatchState ?
+    const U32 dictLowestIndex      = isDxS ? dms->window.dictLimit : 0;
+    const BYTE* const dictBase     = isDxS ? dms->window.base : NULL;
+    const BYTE* const dictLowest   = isDxS ? dictBase + dictLowestIndex : NULL;
+    const BYTE* const dictEnd      = isDxS ? dms->window.nextSrc : NULL;
+    const U32 dictIndexDelta       = isDxS ?
                                      prefixLowestIndex - (U32)(dictEnd - dictBase) :
                                      0;
     const U32 dictAndPrefixLength = (U32)((ip - prefixLowest) + (dictEnd - dictLowest));
 
-    DEBUGLOG(5, "ZSTD_compressBlock_lazy_generic (dictMode=%u)", (U32)dictMode);
-
-    /* init */
+    DEBUGLOG(5, "ZSTD_compressBlock_lazy_generic (dictMode=%u) (searchFunc=%u)", (U32)dictMode, (U32)searchMethod);
     ip += (dictAndPrefixLength == 0);
     if (dictMode == ZSTD_noDict) {
-        U32 const current = (U32)(ip - base);
-        U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, current, ms->cParams.windowLog);
-        U32 const maxRep = current - windowLow;
-        if (offset_2 > maxRep) savedOffset = offset_2, offset_2 = 0;
-        if (offset_1 > maxRep) savedOffset = offset_1, offset_1 = 0;
+        U32 const curr = (U32)(ip - base);
+        U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, curr, ms->cParams.windowLog);
+        U32 const maxRep = curr - windowLow;
+        if (offset_2 > maxRep) offsetSaved2 = offset_2, offset_2 = 0;
+        if (offset_1 > maxRep) offsetSaved1 = offset_1, offset_1 = 0;
     }
-    if (dictMode == ZSTD_dictMatchState) {
+    if (isDxS) {
         /* dictMatchState repCode checks don't currently handle repCode == 0
          * disabling. */
         assert(offset_1 <= dictAndPrefixLength);
         assert(offset_2 <= dictAndPrefixLength);
     }
 
+    if (searchMethod == search_rowHash) {
+        ZSTD_row_fillHashCache(ms, base, rowLog,
+                            MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */),
+                            ms->nextToUpdate, ilimit);
+    }
+
     /* Match Loop */
 #if defined(__GNUC__) && defined(__x86_64__)
     /* I've measured random a 5% speed loss on levels 5 & 6 (greedy) when the
@@ -689,13 +1559,14 @@ ZSTD_compressBlock_lazy_generic(
 #endif
     while (ip < ilimit) {
         size_t matchLength=0;
-        size_t offset=0;
+        size_t offBase = REPCODE1_TO_OFFBASE;
         const BYTE* start=ip+1;
+        DEBUGLOG(7, "search baseline (depth 0)");
 
         /* check repCode */
-        if (dictMode == ZSTD_dictMatchState) {
+        if (isDxS) {
             const U32 repIndex = (U32)(ip - base) + 1 - offset_1;
-            const BYTE* repMatch = (dictMode == ZSTD_dictMatchState
+            const BYTE* repMatch = ((dictMode == ZSTD_dictMatchState || dictMode == ZSTD_dedicatedDictSearch)
                                 && repIndex < prefixLowestIndex) ?
                                    dictBase + (repIndex - dictIndexDelta) :
                                    base + repIndex;
@@ -713,10 +1584,10 @@ ZSTD_compressBlock_lazy_generic(
         }
 
         /* first search (depth 0) */
-        {   size_t offsetFound = 999999999;
-            size_t const ml2 = searchMax(ms, ip, iend, &offsetFound);
+        {   size_t offbaseFound = 999999999;
+            size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &offbaseFound, mls, rowLog, searchMethod, dictMode);
             if (ml2 > matchLength)
-                matchLength = ml2, start = ip, offset=offsetFound;
+                matchLength = ml2, start = ip, offBase = offbaseFound;
         }
 
         if (matchLength < 4) {
@@ -727,16 +1598,17 @@ ZSTD_compressBlock_lazy_generic(
         /* let's try to find a better solution */
         if (depth>=1)
         while (ip<ilimit) {
+            DEBUGLOG(7, "search depth 1");
             ip ++;
             if ( (dictMode == ZSTD_noDict)
-              && (offset) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
+              && (offBase) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
                 size_t const mlRep = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4;
                 int const gain2 = (int)(mlRep * 3);
-                int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offset+1) + 1);
+                int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offBase) + 1);
                 if ((mlRep >= 4) && (gain2 > gain1))
-                    matchLength = mlRep, offset = 0, start = ip;
+                    matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
             }
-            if (dictMode == ZSTD_dictMatchState) {
+            if (isDxS) {
                 const U32 repIndex = (U32)(ip - base) - offset_1;
                 const BYTE* repMatch = repIndex < prefixLowestIndex ?
                                dictBase + (repIndex - dictIndexDelta) :
@@ -746,32 +1618,33 @@ ZSTD_compressBlock_lazy_generic(
                     const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
                     size_t const mlRep = ZSTD_count_2segments(ip+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
                     int const gain2 = (int)(mlRep * 3);
-                    int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offset+1) + 1);
+                    int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offBase) + 1);
                     if ((mlRep >= 4) && (gain2 > gain1))
-                        matchLength = mlRep, offset = 0, start = ip;
+                        matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
                 }
             }
-            {   size_t offset2=999999999;
-                size_t const ml2 = searchMax(ms, ip, iend, &offset2);
-                int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1));   /* raw approx */
-                int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 4);
+            {   size_t ofbCandidate=999999999;
+                size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, dictMode);
+                int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */
+                int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 4);
                 if ((ml2 >= 4) && (gain2 > gain1)) {
-                    matchLength = ml2, offset = offset2, start = ip;
+                    matchLength = ml2, offBase = ofbCandidate, start = ip;
                     continue;   /* search a better one */
             }   }
 
             /* let's find an even better one */
             if ((depth==2) && (ip<ilimit)) {
+                DEBUGLOG(7, "search depth 2");
                 ip ++;
                 if ( (dictMode == ZSTD_noDict)
-                  && (offset) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
+                  && (offBase) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
                     size_t const mlRep = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4;
                     int const gain2 = (int)(mlRep * 4);
-                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 1);
+                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 1);
                     if ((mlRep >= 4) && (gain2 > gain1))
-                        matchLength = mlRep, offset = 0, start = ip;
+                        matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
                 }
-                if (dictMode == ZSTD_dictMatchState) {
+                if (isDxS) {
                     const U32 repIndex = (U32)(ip - base) - offset_1;
                     const BYTE* repMatch = repIndex < prefixLowestIndex ?
                                    dictBase + (repIndex - dictIndexDelta) :
@@ -781,64 +1654,62 @@ ZSTD_compressBlock_lazy_generic(
                         const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
                         size_t const mlRep = ZSTD_count_2segments(ip+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
                         int const gain2 = (int)(mlRep * 4);
-                        int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 1);
+                        int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 1);
                         if ((mlRep >= 4) && (gain2 > gain1))
-                            matchLength = mlRep, offset = 0, start = ip;
+                            matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
                     }
                 }
-                {   size_t offset2=999999999;
-                    size_t const ml2 = searchMax(ms, ip, iend, &offset2);
-                    int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1));   /* raw approx */
-                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 7);
+                {   size_t ofbCandidate=999999999;
+                    size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, dictMode);
+                    int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */
+                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 7);
                     if ((ml2 >= 4) && (gain2 > gain1)) {
-                        matchLength = ml2, offset = offset2, start = ip;
+                        matchLength = ml2, offBase = ofbCandidate, start = ip;
                         continue;
             }   }   }
             break;  /* nothing found : store previous solution */
         }
 
         /* NOTE:
-         * start[-offset+ZSTD_REP_MOVE-1] is undefined behavior.
-         * (-offset+ZSTD_REP_MOVE-1) is unsigned, and is added to start, which
-         * overflows the pointer, which is undefined behavior.
+         * Pay attention that `start[-value]` can lead to strange undefined behavior
+         * notably if `value` is unsigned, resulting in a large positive `-value`.
          */
         /* catch up */
-        if (offset) {
+        if (OFFBASE_IS_OFFSET(offBase)) {
             if (dictMode == ZSTD_noDict) {
-                while ( ((start > anchor) & (start - (offset-ZSTD_REP_MOVE) > prefixLowest))
-                     && (start[-1] == (start-(offset-ZSTD_REP_MOVE))[-1]) )  /* only search for offset within prefix */
+                while ( ((start > anchor) & (start - OFFBASE_TO_OFFSET(offBase) > prefixLowest))
+                     && (start[-1] == (start-OFFBASE_TO_OFFSET(offBase))[-1]) )  /* only search for offset within prefix */
                     { start--; matchLength++; }
             }
-            if (dictMode == ZSTD_dictMatchState) {
-                U32 const matchIndex = (U32)((start-base) - (offset - ZSTD_REP_MOVE));
+            if (isDxS) {
+                U32 const matchIndex = (U32)((size_t)(start-base) - OFFBASE_TO_OFFSET(offBase));
                 const BYTE* match = (matchIndex < prefixLowestIndex) ? dictBase + matchIndex - dictIndexDelta : base + matchIndex;
                 const BYTE* const mStart = (matchIndex < prefixLowestIndex) ? dictLowest : prefixLowest;
                 while ((start>anchor) && (match>mStart) && (start[-1] == match[-1])) { start--; match--; matchLength++; }  /* catch up */
             }
-            offset_2 = offset_1; offset_1 = (U32)(offset - ZSTD_REP_MOVE);
+            offset_2 = offset_1; offset_1 = (U32)OFFBASE_TO_OFFSET(offBase);
         }
         /* store sequence */
 _storeSequence:
-        {   size_t const litLength = start - anchor;
-            ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offset, matchLength-MINMATCH);
+        {   size_t const litLength = (size_t)(start - anchor);
+            ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offBase, matchLength);
             anchor = ip = start + matchLength;
         }
 
         /* check immediate repcode */
-        if (dictMode == ZSTD_dictMatchState) {
+        if (isDxS) {
             while (ip <= ilimit) {
                 U32 const current2 = (U32)(ip-base);
                 U32 const repIndex = current2 - offset_2;
-                const BYTE* repMatch = dictMode == ZSTD_dictMatchState
-                    && repIndex < prefixLowestIndex ?
+                const BYTE* repMatch = repIndex < prefixLowestIndex ?
                         dictBase - dictIndexDelta + repIndex :
                         base + repIndex;
                 if ( ((U32)((prefixLowestIndex-1) - (U32)repIndex) >= 3 /* intentional overflow */)
                    && (MEM_read32(repMatch) == MEM_read32(ip)) ) {
                     const BYTE* const repEnd2 = repIndex < prefixLowestIndex ? dictEnd : iend;
                     matchLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd2, prefixLowest) + 4;
-                    offset = offset_2; offset_2 = offset_1; offset_1 = (U32)offset;   /* swap offset_2 <=> offset_1 */
-                    ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, matchLength-MINMATCH);
+                    offBase = offset_2; offset_2 = offset_1; offset_1 = (U32)offBase;   /* swap offset_2 <=> offset_1 */
+                    ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, matchLength);
                     ip += matchLength;
                     anchor = ip;
                     continue;
@@ -852,16 +1723,20 @@ _storeSequence:
                  && (MEM_read32(ip) == MEM_read32(ip - offset_2)) ) {
                 /* store sequence */
                 matchLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4;
-                offset = offset_2; offset_2 = offset_1; offset_1 = (U32)offset; /* swap repcodes */
-                ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, matchLength-MINMATCH);
+                offBase = offset_2; offset_2 = offset_1; offset_1 = (U32)offBase; /* swap repcodes */
+                ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, matchLength);
                 ip += matchLength;
                 anchor = ip;
                 continue;   /* faster when present ... (?) */
     }   }   }
 
-    /* Save reps for next block */
-    rep[0] = offset_1 ? offset_1 : savedOffset;
-    rep[1] = offset_2 ? offset_2 : savedOffset;
+    /* If offset_1 started invalid (offsetSaved1 != 0) and became valid (offset_1 != 0),
+     * rotate saved offsets. See comment in ZSTD_compressBlock_fast_noDict for more context. */
+    offsetSaved2 = ((offsetSaved1 != 0) && (offset_1 != 0)) ? offsetSaved1 : offsetSaved2;
+
+    /* save reps for next block */
+    rep[0] = offset_1 ? offset_1 : offsetSaved1;
+    rep[1] = offset_2 ? offset_2 : offsetSaved2;
 
     /* Return the last literals size */
     return (size_t)(iend - anchor);
@@ -925,6 +1800,92 @@ size_t ZSTD_compressBlock_greedy_dictMatchState(
 }
 
 
+size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_dedicatedDictSearch);
+}
+
+size_t ZSTD_compressBlock_lazy_dedicatedDictSearch(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_dedicatedDictSearch);
+}
+
+size_t ZSTD_compressBlock_greedy_dedicatedDictSearch(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_dedicatedDictSearch);
+}
+
+/* Row-based matchfinder */
+size_t ZSTD_compressBlock_lazy2_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_noDict);
+}
+
+size_t ZSTD_compressBlock_lazy_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_noDict);
+}
+
+size_t ZSTD_compressBlock_greedy_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_noDict);
+}
+
+size_t ZSTD_compressBlock_lazy2_dictMatchState_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_dictMatchState);
+}
+
+size_t ZSTD_compressBlock_lazy_dictMatchState_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_dictMatchState);
+}
+
+size_t ZSTD_compressBlock_greedy_dictMatchState_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_dictMatchState);
+}
+
+
+size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_dedicatedDictSearch);
+}
+
+size_t ZSTD_compressBlock_lazy_dedicatedDictSearch_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_dedicatedDictSearch);
+}
+
+size_t ZSTD_compressBlock_greedy_dedicatedDictSearch_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_dedicatedDictSearch);
+}
+
 FORCE_INLINE_TEMPLATE
 size_t ZSTD_compressBlock_lazy_extDict_generic(
                         ZSTD_matchState_t* ms, seqStore_t* seqStore,
@@ -936,7 +1897,7 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
     const BYTE* ip = istart;
     const BYTE* anchor = istart;
     const BYTE* const iend = istart + srcSize;
-    const BYTE* const ilimit = iend - 8;
+    const BYTE* const ilimit = searchMethod == search_rowHash ? iend - 8 - ZSTD_ROW_HASH_CACHE_SIZE : iend - 8;
     const BYTE* const base = ms->window.base;
     const U32 dictLimit = ms->window.dictLimit;
     const BYTE* const prefixStart = base + dictLimit;
@@ -944,18 +1905,20 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
     const BYTE* const dictEnd  = dictBase + dictLimit;
     const BYTE* const dictStart  = dictBase + ms->window.lowLimit;
     const U32 windowLog = ms->cParams.windowLog;
-
-    typedef size_t (*searchMax_f)(
-                        ZSTD_matchState_t* ms,
-                        const BYTE* ip, const BYTE* iLimit, size_t* offsetPtr);
-    searchMax_f searchMax = searchMethod==search_binaryTree ? ZSTD_BtFindBestMatch_extDict_selectMLS : ZSTD_HcFindBestMatch_extDict_selectMLS;
+    const U32 mls = BOUNDED(4, ms->cParams.minMatch, 6);
+    const U32 rowLog = BOUNDED(4, ms->cParams.searchLog, 6);
 
     U32 offset_1 = rep[0], offset_2 = rep[1];
 
-    DEBUGLOG(5, "ZSTD_compressBlock_lazy_extDict_generic");
+    DEBUGLOG(5, "ZSTD_compressBlock_lazy_extDict_generic (searchFunc=%u)", (U32)searchMethod);
 
     /* init */
     ip += (ip == prefixStart);
+    if (searchMethod == search_rowHash) {
+        ZSTD_row_fillHashCache(ms, base, rowLog,
+                               MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */),
+                               ms->nextToUpdate, ilimit);
+    }
 
     /* Match Loop */
 #if defined(__GNUC__) && defined(__x86_64__)
@@ -966,16 +1929,17 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
 #endif
     while (ip < ilimit) {
         size_t matchLength=0;
-        size_t offset=0;
+        size_t offBase = REPCODE1_TO_OFFBASE;
         const BYTE* start=ip+1;
-        U32 current = (U32)(ip-base);
+        U32 curr = (U32)(ip-base);
 
         /* check repCode */
-        {   const U32 windowLow = ZSTD_getLowestMatchIndex(ms, current+1, windowLog);
-            const U32 repIndex = (U32)(current+1 - offset_1);
+        {   const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr+1, windowLog);
+            const U32 repIndex = (U32)(curr+1 - offset_1);
             const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
             const BYTE* const repMatch = repBase + repIndex;
-            if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > windowLow))   /* intentional overflow */
+            if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow */
+               & (offset_1 <= curr+1 - windowLow) ) /* note: we are searching at curr+1 */
             if (MEM_read32(ip+1) == MEM_read32(repMatch)) {
                 /* repcode detected we should take it */
                 const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
@@ -984,13 +1948,13 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
         }   }
 
         /* first search (depth 0) */
-        {   size_t offsetFound = 999999999;
-            size_t const ml2 = searchMax(ms, ip, iend, &offsetFound);
+        {   size_t ofbCandidate = 999999999;
+            size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, ZSTD_extDict);
             if (ml2 > matchLength)
-                matchLength = ml2, start = ip, offset=offsetFound;
+                matchLength = ml2, start = ip, offBase = ofbCandidate;
         }
 
-         if (matchLength < 4) {
+        if (matchLength < 4) {
             ip += ((ip-anchor) >> kSearchStrength) + 1;   /* jump faster over incompressible sections */
             continue;
         }
@@ -999,80 +1963,82 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
         if (depth>=1)
         while (ip<ilimit) {
             ip ++;
-            current++;
+            curr++;
             /* check repCode */
-            if (offset) {
-                const U32 windowLow = ZSTD_getLowestMatchIndex(ms, current, windowLog);
-                const U32 repIndex = (U32)(current - offset_1);
+            if (offBase) {
+                const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr, windowLog);
+                const U32 repIndex = (U32)(curr - offset_1);
                 const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
                 const BYTE* const repMatch = repBase + repIndex;
-                if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > windowLow))  /* intentional overflow */
+                if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow : do not test positions overlapping 2 memory segments  */
+                   & (offset_1 <= curr - windowLow) ) /* equivalent to `curr > repIndex >= windowLow` */
                 if (MEM_read32(ip) == MEM_read32(repMatch)) {
                     /* repcode detected */
                     const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
                     size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
                     int const gain2 = (int)(repLength * 3);
-                    int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offset+1) + 1);
+                    int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offBase) + 1);
                     if ((repLength >= 4) && (gain2 > gain1))
-                        matchLength = repLength, offset = 0, start = ip;
+                        matchLength = repLength, offBase = REPCODE1_TO_OFFBASE, start = ip;
             }   }
 
             /* search match, depth 1 */
-            {   size_t offset2=999999999;
-                size_t const ml2 = searchMax(ms, ip, iend, &offset2);
-                int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1));   /* raw approx */
-                int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 4);
+            {   size_t ofbCandidate = 999999999;
+                size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, ZSTD_extDict);
+                int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */
+                int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 4);
                 if ((ml2 >= 4) && (gain2 > gain1)) {
-                    matchLength = ml2, offset = offset2, start = ip;
+                    matchLength = ml2, offBase = ofbCandidate, start = ip;
                     continue;   /* search a better one */
             }   }
 
             /* let's find an even better one */
             if ((depth==2) && (ip<ilimit)) {
                 ip ++;
-                current++;
+                curr++;
                 /* check repCode */
-                if (offset) {
-                    const U32 windowLow = ZSTD_getLowestMatchIndex(ms, current, windowLog);
-                    const U32 repIndex = (U32)(current - offset_1);
+                if (offBase) {
+                    const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr, windowLog);
+                    const U32 repIndex = (U32)(curr - offset_1);
                     const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
                     const BYTE* const repMatch = repBase + repIndex;
-                    if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > windowLow))  /* intentional overflow */
+                    if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow : do not test positions overlapping 2 memory segments  */
+                       & (offset_1 <= curr - windowLow) ) /* equivalent to `curr > repIndex >= windowLow` */
                     if (MEM_read32(ip) == MEM_read32(repMatch)) {
                         /* repcode detected */
                         const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
                         size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
                         int const gain2 = (int)(repLength * 4);
-                        int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 1);
+                        int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 1);
                         if ((repLength >= 4) && (gain2 > gain1))
-                            matchLength = repLength, offset = 0, start = ip;
+                            matchLength = repLength, offBase = REPCODE1_TO_OFFBASE, start = ip;
                 }   }
 
                 /* search match, depth 2 */
-                {   size_t offset2=999999999;
-                    size_t const ml2 = searchMax(ms, ip, iend, &offset2);
-                    int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1));   /* raw approx */
-                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 7);
+                {   size_t ofbCandidate = 999999999;
+                    size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, ZSTD_extDict);
+                    int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */
+                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 7);
                     if ((ml2 >= 4) && (gain2 > gain1)) {
-                        matchLength = ml2, offset = offset2, start = ip;
+                        matchLength = ml2, offBase = ofbCandidate, start = ip;
                         continue;
             }   }   }
             break;  /* nothing found : store previous solution */
         }
 
         /* catch up */
-        if (offset) {
-            U32 const matchIndex = (U32)((start-base) - (offset - ZSTD_REP_MOVE));
+        if (OFFBASE_IS_OFFSET(offBase)) {
+            U32 const matchIndex = (U32)((size_t)(start-base) - OFFBASE_TO_OFFSET(offBase));
             const BYTE* match = (matchIndex < dictLimit) ? dictBase + matchIndex : base + matchIndex;
             const BYTE* const mStart = (matchIndex < dictLimit) ? dictStart : prefixStart;
             while ((start>anchor) && (match>mStart) && (start[-1] == match[-1])) { start--; match--; matchLength++; }  /* catch up */
-            offset_2 = offset_1; offset_1 = (U32)(offset - ZSTD_REP_MOVE);
+            offset_2 = offset_1; offset_1 = (U32)OFFBASE_TO_OFFSET(offBase);
         }
 
         /* store sequence */
 _storeSequence:
-        {   size_t const litLength = start - anchor;
-            ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offset, matchLength-MINMATCH);
+        {   size_t const litLength = (size_t)(start - anchor);
+            ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offBase, matchLength);
             anchor = ip = start + matchLength;
         }
 
@@ -1083,13 +2049,14 @@ _storeSequence:
             const U32 repIndex = repCurrent - offset_2;
             const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
             const BYTE* const repMatch = repBase + repIndex;
-            if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > windowLow))  /* intentional overflow */
+            if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow : do not test positions overlapping 2 memory segments  */
+               & (offset_2 <= repCurrent - windowLow) ) /* equivalent to `curr > repIndex >= windowLow` */
             if (MEM_read32(ip) == MEM_read32(repMatch)) {
                 /* repcode detected we should take it */
                 const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
                 matchLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
-                offset = offset_2; offset_2 = offset_1; offset_1 = (U32)offset;   /* swap offset history */
-                ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, matchLength-MINMATCH);
+                offBase = offset_2; offset_2 = offset_1; offset_1 = (U32)offBase;   /* swap offset history */
+                ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, matchLength);
                 ip += matchLength;
                 anchor = ip;
                 continue;   /* faster when present ... (?) */
@@ -1136,3 +2103,25 @@ size_t ZSTD_compressBlock_btlazy2_extDict(
 {
     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2);
 }
+
+size_t ZSTD_compressBlock_greedy_extDict_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0);
+}
+
+size_t ZSTD_compressBlock_lazy_extDict_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+
+{
+    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1);
+}
+
+size_t ZSTD_compressBlock_lazy2_extDict_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2);
+}
diff --git a/contrib/zstd/zstd_lazy.h b/contrib/zstd/zstd_lazy.h
index 581936f03..3bde67331 100644
--- a/contrib/zstd/zstd_lazy.h
+++ b/contrib/zstd/zstd_lazy.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -17,7 +17,20 @@ extern "C" {
 
 #include "zstd_compress_internal.h"
 
+/**
+ * Dedicated Dictionary Search Structure bucket log. In the
+ * ZSTD_dedicatedDictSearch mode, the hashTable has
+ * 2 ** ZSTD_LAZY_DDSS_BUCKET_LOG entries in each bucket, rather than just
+ * one.
+ */
+#define ZSTD_LAZY_DDSS_BUCKET_LOG 2
+
+#define ZSTD_ROW_HASH_TAG_BITS 8        /* nb bits to use for the tag */
+
 U32 ZSTD_insertAndFindFirstIndex(ZSTD_matchState_t* ms, const BYTE* ip);
+void ZSTD_row_update(ZSTD_matchState_t* const ms, const BYTE* ip);
+
+void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t* ms, const BYTE* const ip);
 
 void ZSTD_preserveUnsortedMark (U32* const table, U32 const size, U32 const reducerValue);  /*! used in ZSTD_reduceIndex(). preemptively increase value of ZSTD_DUBT_UNSORTED_MARK */
 
@@ -33,6 +46,15 @@ size_t ZSTD_compressBlock_lazy(
 size_t ZSTD_compressBlock_greedy(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy2_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_greedy_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
 
 size_t ZSTD_compressBlock_btlazy2_dictMatchState(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
@@ -46,6 +68,34 @@ size_t ZSTD_compressBlock_lazy_dictMatchState(
 size_t ZSTD_compressBlock_greedy_dictMatchState(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy2_dictMatchState_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy_dictMatchState_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_greedy_dictMatchState_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+
+size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy_dedicatedDictSearch(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_greedy_dedicatedDictSearch(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy_dedicatedDictSearch_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_greedy_dedicatedDictSearch_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
 
 size_t ZSTD_compressBlock_greedy_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
@@ -56,10 +106,20 @@ size_t ZSTD_compressBlock_lazy_extDict(
 size_t ZSTD_compressBlock_lazy2_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_greedy_extDict_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy_extDict_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy2_extDict_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
 size_t ZSTD_compressBlock_btlazy2_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
 
+
 #if defined (__cplusplus)
 }
 #endif
diff --git a/contrib/zstd/zstd_ldm.c b/contrib/zstd/zstd_ldm.c
index 3916da494..b8dcbcada 100644
--- a/contrib/zstd/zstd_ldm.c
+++ b/contrib/zstd/zstd_ldm.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -11,13 +11,126 @@
 #include "zstd_ldm.h"
 
 #include "debug.h"
+#include "xxhash.h"
 #include "zstd_fast.h"          /* ZSTD_fillHashTable() */
 #include "zstd_double_fast.h"   /* ZSTD_fillDoubleHashTable() */
+#include "zstd_ldm_geartab.h"
 
 #define LDM_BUCKET_SIZE_LOG 3
 #define LDM_MIN_MATCH_LENGTH 64
 #define LDM_HASH_RLOG 7
-#define LDM_HASH_CHAR_OFFSET 10
+
+typedef struct {
+    U64 rolling;
+    U64 stopMask;
+} ldmRollingHashState_t;
+
+/** ZSTD_ldm_gear_init():
+ *
+ * Initializes the rolling hash state such that it will honor the
+ * settings in params. */
+static void ZSTD_ldm_gear_init(ldmRollingHashState_t* state, ldmParams_t const* params)
+{
+    unsigned maxBitsInMask = MIN(params->minMatchLength, 64);
+    unsigned hashRateLog = params->hashRateLog;
+
+    state->rolling = ~(U32)0;
+
+    /* The choice of the splitting criterion is subject to two conditions:
+     *   1. it has to trigger on average every 2^(hashRateLog) bytes;
+     *   2. ideally, it has to depend on a window of minMatchLength bytes.
+     *
+     * In the gear hash algorithm, bit n depends on the last n bytes;
+     * so in order to obtain a good quality splitting criterion it is
+     * preferable to use bits with high weight.
+     *
+     * To match condition 1 we use a mask with hashRateLog bits set
+     * and, because of the previous remark, we make sure these bits
+     * have the highest possible weight while still respecting
+     * condition 2.
+     */
+    if (hashRateLog > 0 && hashRateLog <= maxBitsInMask) {
+        state->stopMask = (((U64)1 << hashRateLog) - 1) << (maxBitsInMask - hashRateLog);
+    } else {
+        /* In this degenerate case we simply honor the hash rate. */
+        state->stopMask = ((U64)1 << hashRateLog) - 1;
+    }
+}
+
+/** ZSTD_ldm_gear_reset()
+ * Feeds [data, data + minMatchLength) into the hash without registering any
+ * splits. This effectively resets the hash state. This is used when skipping
+ * over data, either at the beginning of a block, or skipping sections.
+ */
+static void ZSTD_ldm_gear_reset(ldmRollingHashState_t* state,
+                                BYTE const* data, size_t minMatchLength)
+{
+    U64 hash = state->rolling;
+    size_t n = 0;
+
+#define GEAR_ITER_ONCE() do {                                  \
+        hash = (hash << 1) + ZSTD_ldm_gearTab[data[n] & 0xff]; \
+        n += 1;                                                \
+    } while (0)
+    while (n + 3 < minMatchLength) {
+        GEAR_ITER_ONCE();
+        GEAR_ITER_ONCE();
+        GEAR_ITER_ONCE();
+        GEAR_ITER_ONCE();
+    }
+    while (n < minMatchLength) {
+        GEAR_ITER_ONCE();
+    }
+#undef GEAR_ITER_ONCE
+}
+
+/** ZSTD_ldm_gear_feed():
+ *
+ * Registers in the splits array all the split points found in the first
+ * size bytes following the data pointer. This function terminates when
+ * either all the data has been processed or LDM_BATCH_SIZE splits are
+ * present in the splits array.
+ *
+ * Precondition: The splits array must not be full.
+ * Returns: The number of bytes processed. */
+static size_t ZSTD_ldm_gear_feed(ldmRollingHashState_t* state,
+                                 BYTE const* data, size_t size,
+                                 size_t* splits, unsigned* numSplits)
+{
+    size_t n;
+    U64 hash, mask;
+
+    hash = state->rolling;
+    mask = state->stopMask;
+    n = 0;
+
+#define GEAR_ITER_ONCE() do { \
+        hash = (hash << 1) + ZSTD_ldm_gearTab[data[n] & 0xff]; \
+        n += 1; \
+        if (UNLIKELY((hash & mask) == 0)) { \
+            splits[*numSplits] = n; \
+            *numSplits += 1; \
+            if (*numSplits == LDM_BATCH_SIZE) \
+                goto done; \
+        } \
+    } while (0)
+
+    while (n + 3 < size) {
+        GEAR_ITER_ONCE();
+        GEAR_ITER_ONCE();
+        GEAR_ITER_ONCE();
+        GEAR_ITER_ONCE();
+    }
+    while (n < size) {
+        GEAR_ITER_ONCE();
+    }
+
+#undef GEAR_ITER_ONCE
+
+done:
+    state->rolling = hash;
+    return n;
+}
 
 void ZSTD_ldm_adjustParameters(ldmParams_t* params,
                                ZSTD_compressionParameters const* cParams)
@@ -27,13 +140,6 @@ void ZSTD_ldm_adjustParameters(ldmParams_t* params,
     DEBUGLOG(4, "ZSTD_ldm_adjustParameters");
     if (!params->bucketSizeLog) params->bucketSizeLog = LDM_BUCKET_SIZE_LOG;
     if (!params->minMatchLength) params->minMatchLength = LDM_MIN_MATCH_LENGTH;
-    if (cParams->strategy >= ZSTD_btopt) {
-      /* Get out of the way of the optimal parser */
-      U32 const minMatch = MAX(cParams->targetLength, params->minMatchLength);
-      assert(minMatch >= ZSTD_LDM_MINMATCH_MIN);
-      assert(minMatch <= ZSTD_LDM_MINMATCH_MAX);
-      params->minMatchLength = minMatch;
-    }
     if (params->hashLog == 0) {
         params->hashLog = MAX(ZSTD_HASHLOG_MIN, params->windowLog - LDM_HASH_RLOG);
         assert(params->hashLog <= ZSTD_HASHLOG_MAX);
@@ -53,47 +159,12 @@ size_t ZSTD_ldm_getTableSize(ldmParams_t params)
     size_t const ldmBucketSize = ((size_t)1) << (params.hashLog - ldmBucketSizeLog);
     size_t const totalSize = ZSTD_cwksp_alloc_size(ldmBucketSize)
                            + ZSTD_cwksp_alloc_size(ldmHSize * sizeof(ldmEntry_t));
-    return params.enableLdm ? totalSize : 0;
+    return params.enableLdm == ZSTD_ps_enable ? totalSize : 0;
 }
 
 size_t ZSTD_ldm_getMaxNbSeq(ldmParams_t params, size_t maxChunkSize)
 {
-    return params.enableLdm ? (maxChunkSize / params.minMatchLength) : 0;
-}
-
-/** ZSTD_ldm_getSmallHash() :
- *  numBits should be <= 32
- *  If numBits==0, returns 0.
- *  @return : the most significant numBits of value. */
-static U32 ZSTD_ldm_getSmallHash(U64 value, U32 numBits)
-{
-    assert(numBits <= 32);
-    return numBits == 0 ? 0 : (U32)(value >> (64 - numBits));
-}
-
-/** ZSTD_ldm_getChecksum() :
- *  numBitsToDiscard should be <= 32
- *  @return : the next most significant 32 bits after numBitsToDiscard */
-static U32 ZSTD_ldm_getChecksum(U64 hash, U32 numBitsToDiscard)
-{
-    assert(numBitsToDiscard <= 32);
-    return (hash >> (64 - 32 - numBitsToDiscard)) & 0xFFFFFFFF;
-}
-
-/** ZSTD_ldm_getTag() ;
- *  Given the hash, returns the most significant numTagBits bits
- *  after (32 + hbits) bits.
- *
- *  If there are not enough bits remaining, return the last
- *  numTagBits bits. */
-static U32 ZSTD_ldm_getTag(U64 hash, U32 hbits, U32 numTagBits)
-{
-    assert(numTagBits < 32 && hbits <= 32);
-    if (32 - hbits < numTagBits) {
-        return hash & (((U32)1 << numTagBits) - 1);
-    } else {
-        return (hash >> (32 - hbits - numTagBits)) & (((U32)1 << numTagBits) - 1);
-    }
+    return params.enableLdm == ZSTD_ps_enable ? (maxChunkSize / params.minMatchLength) : 0;
 }
 
 /** ZSTD_ldm_getBucket() :
@@ -110,38 +181,12 @@ static void ZSTD_ldm_insertEntry(ldmState_t* ldmState,
                                  size_t const hash, const ldmEntry_t entry,
                                  ldmParams_t const ldmParams)
 {
-    BYTE* const bucketOffsets = ldmState->bucketOffsets;
-    *(ZSTD_ldm_getBucket(ldmState, hash, ldmParams) + bucketOffsets[hash]) = entry;
-    bucketOffsets[hash]++;
-    bucketOffsets[hash] &= ((U32)1 << ldmParams.bucketSizeLog) - 1;
-}
+    BYTE* const pOffset = ldmState->bucketOffsets + hash;
+    unsigned const offset = *pOffset;
+
+    *(ZSTD_ldm_getBucket(ldmState, hash, ldmParams) + offset) = entry;
+    *pOffset = (BYTE)((offset + 1) & ((1u << ldmParams.bucketSizeLog) - 1));
 
-/** ZSTD_ldm_makeEntryAndInsertByTag() :
- *
- *  Gets the small hash, checksum, and tag from the rollingHash.
- *
- *  If the tag matches (1 << ldmParams.hashRateLog)-1, then
- *  creates an ldmEntry from the offset, and inserts it into the hash table.
- *
- *  hBits is the length of the small hash, which is the most significant hBits
- *  of rollingHash. The checksum is the next 32 most significant bits, followed
- *  by ldmParams.hashRateLog bits that make up the tag. */
-static void ZSTD_ldm_makeEntryAndInsertByTag(ldmState_t* ldmState,
-                                             U64 const rollingHash,
-                                             U32 const hBits,
-                                             U32 const offset,
-                                             ldmParams_t const ldmParams)
-{
-    U32 const tag = ZSTD_ldm_getTag(rollingHash, hBits, ldmParams.hashRateLog);
-    U32 const tagMask = ((U32)1 << ldmParams.hashRateLog) - 1;
-    if (tag == tagMask) {
-        U32 const hash = ZSTD_ldm_getSmallHash(rollingHash, hBits);
-        U32 const checksum = ZSTD_ldm_getChecksum(rollingHash, hBits);
-        ldmEntry_t entry;
-        entry.offset = offset;
-        entry.checksum = checksum;
-        ZSTD_ldm_insertEntry(ldmState, hash, entry, ldmParams);
-    }
 }
 
 /** ZSTD_ldm_countBackwardsMatch() :
@@ -150,10 +195,10 @@ static void ZSTD_ldm_makeEntryAndInsertByTag(ldmState_t* ldmState,
  *  We count only bytes where pMatch >= pBase and pIn >= pAnchor. */
 static size_t ZSTD_ldm_countBackwardsMatch(
             const BYTE* pIn, const BYTE* pAnchor,
-            const BYTE* pMatch, const BYTE* pBase)
+            const BYTE* pMatch, const BYTE* pMatchBase)
 {
     size_t matchLength = 0;
-    while (pIn > pAnchor && pMatch > pBase && pIn[-1] == pMatch[-1]) {
+    while (pIn > pAnchor && pMatch > pMatchBase && pIn[-1] == pMatch[-1]) {
         pIn--;
         pMatch--;
         matchLength++;
@@ -161,6 +206,27 @@ static size_t ZSTD_ldm_countBackwardsMatch(
     return matchLength;
 }
 
+/** ZSTD_ldm_countBackwardsMatch_2segments() :
+ *  Returns the number of bytes that match backwards from pMatch,
+ *  even with the backwards match spanning 2 different segments.
+ *
+ *  On reaching `pMatchBase`, start counting from mEnd */
+static size_t ZSTD_ldm_countBackwardsMatch_2segments(
+                    const BYTE* pIn, const BYTE* pAnchor,
+                    const BYTE* pMatch, const BYTE* pMatchBase,
+                    const BYTE* pExtDictStart, const BYTE* pExtDictEnd)
+{
+    size_t matchLength = ZSTD_ldm_countBackwardsMatch(pIn, pAnchor, pMatch, pMatchBase);
+    if (pMatch - matchLength != pMatchBase || pMatchBase == pExtDictStart) {
+        /* If backwards match is entirely in the extDict or prefix, immediately return */
+        return matchLength;
+    }
+    DEBUGLOG(7, "ZSTD_ldm_countBackwardsMatch_2segments: found 2-parts backwards match (length in prefix==%zu)", matchLength);
+    matchLength += ZSTD_ldm_countBackwardsMatch(pIn - matchLength, pAnchor, pExtDictEnd, pExtDictStart);
+    DEBUGLOG(7, "final backwards match length = %zu", matchLength);
+    return matchLength;
+}
+
 /** ZSTD_ldm_fillFastTables() :
  *
  *  Fills the relevant tables for the ZSTD_fast and ZSTD_dfast strategies.
@@ -176,11 +242,11 @@ static size_t ZSTD_ldm_fillFastTables(ZSTD_matchState_t* ms,
     switch(ms->cParams.strategy)
     {
     case ZSTD_fast:
-        ZSTD_fillHashTable(ms, iend, ZSTD_dtlm_fast);
+        ZSTD_fillHashTable(ms, iend, ZSTD_dtlm_fast, ZSTD_tfp_forCCtx);
         break;
 
     case ZSTD_dfast:
-        ZSTD_fillDoubleHashTable(ms, iend, ZSTD_dtlm_fast);
+        ZSTD_fillDoubleHashTable(ms, iend, ZSTD_dtlm_fast, ZSTD_tfp_forCCtx);
         break;
 
     case ZSTD_greedy:
@@ -198,43 +264,42 @@ static size_t ZSTD_ldm_fillFastTables(ZSTD_matchState_t* ms,
     return 0;
 }
 
-/** ZSTD_ldm_fillLdmHashTable() :
- *
- *  Fills hashTable from (lastHashed + 1) to iend (non-inclusive).
- *  lastHash is the rolling hash that corresponds to lastHashed.
- *
- *  Returns the rolling hash corresponding to position iend-1. */
-static U64 ZSTD_ldm_fillLdmHashTable(ldmState_t* state,
-                                     U64 lastHash, const BYTE* lastHashed,
-                                     const BYTE* iend, const BYTE* base,
-                                     U32 hBits, ldmParams_t const ldmParams)
-{
-    U64 rollingHash = lastHash;
-    const BYTE* cur = lastHashed + 1;
-
-    while (cur < iend) {
-        rollingHash = ZSTD_rollingHash_rotate(rollingHash, cur[-1],
-                                              cur[ldmParams.minMatchLength-1],
-                                              state->hashPower);
-        ZSTD_ldm_makeEntryAndInsertByTag(state,
-                                         rollingHash, hBits,
-                                         (U32)(cur - base), ldmParams);
-        ++cur;
-    }
-    return rollingHash;
-}
-
 void ZSTD_ldm_fillHashTable(
-            ldmState_t* state, const BYTE* ip,
+            ldmState_t* ldmState, const BYTE* ip,
             const BYTE* iend, ldmParams_t const* params)
 {
+    U32 const minMatchLength = params->minMatchLength;
+    U32 const hBits = params->hashLog - params->bucketSizeLog;
+    BYTE const* const base = ldmState->window.base;
+    BYTE const* const istart = ip;
+    ldmRollingHashState_t hashState;
+    size_t* const splits = ldmState->splitIndices;
+    unsigned numSplits;
+
     DEBUGLOG(5, "ZSTD_ldm_fillHashTable");
-    if ((size_t)(iend - ip) >= params->minMatchLength) {
-        U64 startingHash = ZSTD_rollingHash_compute(ip, params->minMatchLength);
-        ZSTD_ldm_fillLdmHashTable(
-            state, startingHash, ip, iend - params->minMatchLength, state->window.base,
-            params->hashLog - params->bucketSizeLog,
-            *params);
+
+    ZSTD_ldm_gear_init(&hashState, params);
+    while (ip < iend) {
+        size_t hashed;
+        unsigned n;
+
+        numSplits = 0;
+        hashed = ZSTD_ldm_gear_feed(&hashState, ip, iend - ip, splits, &numSplits);
+
+        for (n = 0; n < numSplits; n++) {
+            if (ip + splits[n] >= istart + minMatchLength) {
+                BYTE const* const split = ip + splits[n] - minMatchLength;
+                U64 const xxhash = XXH64(split, minMatchLength, 0);
+                U32 const hash = (U32)(xxhash & (((U32)1 << hBits) - 1));
+                ldmEntry_t entry;
+
+                entry.offset = (U32)(split - base);
+                entry.checksum = (U32)(xxhash >> 32);
+                ZSTD_ldm_insertEntry(ldmState, hash, entry, *params);
+            }
+        }
+
+        ip += hashed;
     }
 }
 
@@ -246,10 +311,10 @@ void ZSTD_ldm_fillHashTable(
  *  (after a long match, only update tables a limited amount). */
 static void ZSTD_ldm_limitTableUpdate(ZSTD_matchState_t* ms, const BYTE* anchor)
 {
-    U32 const current = (U32)(anchor - ms->window.base);
-    if (current > ms->nextToUpdate + 1024) {
+    U32 const curr = (U32)(anchor - ms->window.base);
+    if (curr > ms->nextToUpdate + 1024) {
         ms->nextToUpdate =
-            current - MIN(512, current - ms->nextToUpdate - 1024);
+            curr - MIN(512, curr - ms->nextToUpdate - 1024);
     }
 }
 
@@ -260,11 +325,8 @@ static size_t ZSTD_ldm_generateSequences_internal(
     /* LDM parameters */
     int const extDict = ZSTD_window_hasExtDict(ldmState->window);
     U32 const minMatchLength = params->minMatchLength;
-    U64 const hashPower = ldmState->hashPower;
+    U32 const entsPerBucket = 1U << params->bucketSizeLog;
     U32 const hBits = params->hashLog - params->bucketSizeLog;
-    U32 const ldmBucketSize = 1U << params->bucketSizeLog;
-    U32 const hashRateLog = params->hashRateLog;
-    U32 const ldmTagMask = (1U << params->hashRateLog) - 1;
     /* Prefix and extDict parameters */
     U32 const dictLimit = ldmState->window.dictLimit;
     U32 const lowestIndex = extDict ? ldmState->window.lowLimit : dictLimit;
@@ -276,45 +338,69 @@ static size_t ZSTD_ldm_generateSequences_internal(
     /* Input bounds */
     BYTE const* const istart = (BYTE const*)src;
     BYTE const* const iend = istart + srcSize;
-    BYTE const* const ilimit = iend - MAX(minMatchLength, HASH_READ_SIZE);
+    BYTE const* const ilimit = iend - HASH_READ_SIZE;
     /* Input positions */
     BYTE const* anchor = istart;
     BYTE const* ip = istart;
-    /* Rolling hash */
-    BYTE const* lastHashed = NULL;
-    U64 rollingHash = 0;
-
-    while (ip <= ilimit) {
-        size_t mLength;
-        U32 const current = (U32)(ip - base);
-        size_t forwardMatchLength = 0, backwardMatchLength = 0;
-        ldmEntry_t* bestEntry = NULL;
-        if (ip != istart) {
-            rollingHash = ZSTD_rollingHash_rotate(rollingHash, lastHashed[0],
-                                                  lastHashed[minMatchLength],
-                                                  hashPower);
-        } else {
-            rollingHash = ZSTD_rollingHash_compute(ip, minMatchLength);
+    /* Rolling hash state */
+    ldmRollingHashState_t hashState;
+    /* Arrays for staged-processing */
+    size_t* const splits = ldmState->splitIndices;
+    ldmMatchCandidate_t* const candidates = ldmState->matchCandidates;
+    unsigned numSplits;
+
+    if (srcSize < minMatchLength)
+        return iend - anchor;
+
+    /* Initialize the rolling hash state with the first minMatchLength bytes */
+    ZSTD_ldm_gear_init(&hashState, params);
+    ZSTD_ldm_gear_reset(&hashState, ip, minMatchLength);
+    ip += minMatchLength;
+
+    while (ip < ilimit) {
+        size_t hashed;
+        unsigned n;
+
+        numSplits = 0;
+        hashed = ZSTD_ldm_gear_feed(&hashState, ip, ilimit - ip,
+                                    splits, &numSplits);
+
+        for (n = 0; n < numSplits; n++) {
+            BYTE const* const split = ip + splits[n] - minMatchLength;
+            U64 const xxhash = XXH64(split, minMatchLength, 0);
+            U32 const hash = (U32)(xxhash & (((U32)1 << hBits) - 1));
+
+            candidates[n].split = split;
+            candidates[n].hash = hash;
+            candidates[n].checksum = (U32)(xxhash >> 32);
+            candidates[n].bucket = ZSTD_ldm_getBucket(ldmState, hash, *params);
+            PREFETCH_L1(candidates[n].bucket);
         }
-        lastHashed = ip;
 
-        /* Do not insert and do not look for a match */
-        if (ZSTD_ldm_getTag(rollingHash, hBits, hashRateLog) != ldmTagMask) {
-           ip++;
-           continue;
-        }
+        for (n = 0; n < numSplits; n++) {
+            size_t forwardMatchLength = 0, backwardMatchLength = 0,
+                   bestMatchLength = 0, mLength;
+            U32 offset;
+            BYTE const* const split = candidates[n].split;
+            U32 const checksum = candidates[n].checksum;
+            U32 const hash = candidates[n].hash;
+            ldmEntry_t* const bucket = candidates[n].bucket;
+            ldmEntry_t const* cur;
+            ldmEntry_t const* bestEntry = NULL;
+            ldmEntry_t newEntry;
+
+            newEntry.offset = (U32)(split - base);
+            newEntry.checksum = checksum;
+
+            /* If a split point would generate a sequence overlapping with
+             * the previous one, we merely register it in the hash table and
+             * move on */
+            if (split < anchor) {
+                ZSTD_ldm_insertEntry(ldmState, hash, newEntry, *params);
+                continue;
+            }
 
-        /* Get the best entry and compute the match lengths */
-        {
-            ldmEntry_t* const bucket =
-                ZSTD_ldm_getBucket(ldmState,
-                                   ZSTD_ldm_getSmallHash(rollingHash, hBits),
-                                   *params);
-            ldmEntry_t* cur;
-            size_t bestMatchLength = 0;
-            U32 const checksum = ZSTD_ldm_getChecksum(rollingHash, hBits);
-
-            for (cur = bucket; cur < bucket + ldmBucketSize; ++cur) {
+            for (cur = bucket; cur < bucket + entsPerBucket; cur++) {
                 size_t curForwardMatchLength, curBackwardMatchLength,
                        curTotalMatchLength;
                 if (cur->checksum != checksum || cur->offset <= lowestIndex) {
@@ -328,30 +414,23 @@ static size_t ZSTD_ldm_generateSequences_internal(
                         cur->offset < dictLimit ? dictEnd : iend;
                     BYTE const* const lowMatchPtr =
                         cur->offset < dictLimit ? dictStart : lowPrefixPtr;
-
-                    curForwardMatchLength = ZSTD_count_2segments(
-                                                ip, pMatch, iend,
-                                                matchEnd, lowPrefixPtr);
+                    curForwardMatchLength =
+                        ZSTD_count_2segments(split, pMatch, iend, matchEnd, lowPrefixPtr);
                     if (curForwardMatchLength < minMatchLength) {
                         continue;
                     }
-                    curBackwardMatchLength =
-                        ZSTD_ldm_countBackwardsMatch(ip, anchor, pMatch,
-                                                     lowMatchPtr);
-                    curTotalMatchLength = curForwardMatchLength +
-                                          curBackwardMatchLength;
+                    curBackwardMatchLength = ZSTD_ldm_countBackwardsMatch_2segments(
+                            split, anchor, pMatch, lowMatchPtr, dictStart, dictEnd);
                 } else { /* !extDict */
                     BYTE const* const pMatch = base + cur->offset;
-                    curForwardMatchLength = ZSTD_count(ip, pMatch, iend);
+                    curForwardMatchLength = ZSTD_count(split, pMatch, iend);
                     if (curForwardMatchLength < minMatchLength) {
                         continue;
                     }
                     curBackwardMatchLength =
-                        ZSTD_ldm_countBackwardsMatch(ip, anchor, pMatch,
-                                                     lowPrefixPtr);
-                    curTotalMatchLength = curForwardMatchLength +
-                                          curBackwardMatchLength;
+                        ZSTD_ldm_countBackwardsMatch(split, anchor, pMatch, lowPrefixPtr);
                 }
+                curTotalMatchLength = curForwardMatchLength + curBackwardMatchLength;
 
                 if (curTotalMatchLength > bestMatchLength) {
                     bestMatchLength = curTotalMatchLength;
@@ -360,57 +439,54 @@ static size_t ZSTD_ldm_generateSequences_internal(
                     bestEntry = cur;
                 }
             }
-        }
-
-        /* No match found -- continue searching */
-        if (bestEntry == NULL) {
-            ZSTD_ldm_makeEntryAndInsertByTag(ldmState, rollingHash,
-                                             hBits, current,
-                                             *params);
-            ip++;
-            continue;
-        }
 
-        /* Match found */
-        mLength = forwardMatchLength + backwardMatchLength;
-        ip -= backwardMatchLength;
+            /* No match found -- insert an entry into the hash table
+             * and process the next candidate match */
+            if (bestEntry == NULL) {
+                ZSTD_ldm_insertEntry(ldmState, hash, newEntry, *params);
+                continue;
+            }
 
-        {
-            /* Store the sequence:
-             * ip = current - backwardMatchLength
-             * The match is at (bestEntry->offset - backwardMatchLength)
-             */
-            U32 const matchIndex = bestEntry->offset;
-            U32 const offset = current - matchIndex;
-            rawSeq* const seq = rawSeqStore->seq + rawSeqStore->size;
-
-            /* Out of sequence storage */
-            if (rawSeqStore->size == rawSeqStore->capacity)
-                return ERROR(dstSize_tooSmall);
-            seq->litLength = (U32)(ip - anchor);
-            seq->matchLength = (U32)mLength;
-            seq->offset = offset;
-            rawSeqStore->size++;
-        }
+            /* Match found */
+            offset = (U32)(split - base) - bestEntry->offset;
+            mLength = forwardMatchLength + backwardMatchLength;
+            {
+                rawSeq* const seq = rawSeqStore->seq + rawSeqStore->size;
+
+                /* Out of sequence storage */
+                if (rawSeqStore->size == rawSeqStore->capacity)
+                    return ERROR(dstSize_tooSmall);
+                seq->litLength = (U32)(split - backwardMatchLength - anchor);
+                seq->matchLength = (U32)mLength;
+                seq->offset = offset;
+                rawSeqStore->size++;
+            }
 
-        /* Insert the current entry into the hash table */
-        ZSTD_ldm_makeEntryAndInsertByTag(ldmState, rollingHash, hBits,
-                                         (U32)(lastHashed - base),
-                                         *params);
+            /* Insert the current entry into the hash table --- it must be
+             * done after the previous block to avoid clobbering bestEntry */
+            ZSTD_ldm_insertEntry(ldmState, hash, newEntry, *params);
 
-        assert(ip + backwardMatchLength == lastHashed);
+            anchor = split + forwardMatchLength;
 
-        /* Fill the hash table from lastHashed+1 to ip+mLength*/
-        /* Heuristic: don't need to fill the entire table at end of block */
-        if (ip + mLength <= ilimit) {
-            rollingHash = ZSTD_ldm_fillLdmHashTable(
-                              ldmState, rollingHash, lastHashed,
-                              ip + mLength, base, hBits, *params);
-            lastHashed = ip + mLength - 1;
+            /* If we find a match that ends after the data that we've hashed
+             * then we have a repeating, overlapping, pattern. E.g. all zeros.
+             * If one repetition of the pattern matches our `stopMask` then all
+             * repetitions will. We don't need to insert them all into out table,
+             * only the first one. So skip over overlapping matches.
+             * This is a major speed boost (20x) for compressing a single byte
+             * repeated, when that byte ends up in the table.
+             */
+            if (anchor > ip + hashed) {
+                ZSTD_ldm_gear_reset(&hashState, anchor - minMatchLength, minMatchLength);
+                /* Continue the outer loop at anchor (ip + hashed == anchor). */
+                ip = anchor - hashed;
+                break;
+            }
         }
-        ip += mLength;
-        anchor = ip;
+
+        ip += hashed;
     }
+
     return iend - anchor;
 }
 
@@ -459,7 +535,7 @@ size_t ZSTD_ldm_generateSequences(
 
         assert(chunkStart < iend);
         /* 1. Perform overflow correction if necessary. */
-        if (ZSTD_window_needOverflowCorrection(ldmState->window, chunkEnd)) {
+        if (ZSTD_window_needOverflowCorrection(ldmState->window, 0, maxDist, ldmState->loadedDictEnd, chunkStart, chunkEnd)) {
             U32 const ldmHSize = 1U << params->hashLog;
             U32 const correction = ZSTD_window_correctOverflow(
                 &ldmState->window, /* cycleLog */ 0, maxDist, chunkStart);
@@ -473,7 +549,7 @@ size_t ZSTD_ldm_generateSequences(
          * the window through early invalidation.
          * TODO: * Test the chunk size.
          *       * Try invalidation after the sequence generation and test the
-         *         the offset against maxDist directly.
+         *         offset against maxDist directly.
          *
          * NOTE: Because of dictionaries + sequence splitting we MUST make sure
          * that any offset used is valid at the END of the sequence, since it may
@@ -503,7 +579,9 @@ size_t ZSTD_ldm_generateSequences(
     return 0;
 }
 
-void ZSTD_ldm_skipSequences(rawSeqStore_t* rawSeqStore, size_t srcSize, U32 const minMatch) {
+void
+ZSTD_ldm_skipSequences(rawSeqStore_t* rawSeqStore, size_t srcSize, U32 const minMatch)
+{
     while (srcSize > 0 && rawSeqStore->pos < rawSeqStore->size) {
         rawSeq* seq = rawSeqStore->seq + rawSeqStore->pos;
         if (srcSize <= seq->litLength) {
@@ -562,14 +640,32 @@ static rawSeq maybeSplitSequence(rawSeqStore_t* rawSeqStore,
     return sequence;
 }
 
+void ZSTD_ldm_skipRawSeqStoreBytes(rawSeqStore_t* rawSeqStore, size_t nbBytes) {
+    U32 currPos = (U32)(rawSeqStore->posInSequence + nbBytes);
+    while (currPos && rawSeqStore->pos < rawSeqStore->size) {
+        rawSeq currSeq = rawSeqStore->seq[rawSeqStore->pos];
+        if (currPos >= currSeq.litLength + currSeq.matchLength) {
+            currPos -= currSeq.litLength + currSeq.matchLength;
+            rawSeqStore->pos++;
+        } else {
+            rawSeqStore->posInSequence = currPos;
+            break;
+        }
+    }
+    if (currPos == 0 || rawSeqStore->pos == rawSeqStore->size) {
+        rawSeqStore->posInSequence = 0;
+    }
+}
+
 size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore,
     ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+    ZSTD_paramSwitch_e useRowMatchFinder,
     void const* src, size_t srcSize)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
     unsigned const minMatch = cParams->minMatch;
     ZSTD_blockCompressor const blockCompressor =
-        ZSTD_selectBlockCompressor(cParams->strategy, ZSTD_matchState_dictMode(ms));
+        ZSTD_selectBlockCompressor(cParams->strategy, useRowMatchFinder, ZSTD_matchState_dictMode(ms));
     /* Input bounds */
     BYTE const* const istart = (BYTE const*)src;
     BYTE const* const iend = istart + srcSize;
@@ -577,9 +673,18 @@ size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore,
     BYTE const* ip = istart;
 
     DEBUGLOG(5, "ZSTD_ldm_blockCompress: srcSize=%zu", srcSize);
+    /* If using opt parser, use LDMs only as candidates rather than always accepting them */
+    if (cParams->strategy >= ZSTD_btopt) {
+        size_t lastLLSize;
+        ms->ldmSeqStore = rawSeqStore;
+        lastLLSize = blockCompressor(ms, seqStore, rep, src, srcSize);
+        ZSTD_ldm_skipRawSeqStoreBytes(rawSeqStore, srcSize);
+        return lastLLSize;
+    }
+
     assert(rawSeqStore->pos <= rawSeqStore->size);
     assert(rawSeqStore->size <= rawSeqStore->capacity);
-    /* Loop through each sequence and apply the block compressor to the lits */
+    /* Loop through each sequence and apply the block compressor to the literals */
     while (rawSeqStore->pos < rawSeqStore->size && ip < iend) {
         /* maybeSplitSequence updates rawSeqStore->pos */
         rawSeq const sequence = maybeSplitSequence(rawSeqStore,
@@ -606,8 +711,8 @@ size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore,
             rep[0] = sequence.offset;
             /* Store the sequence */
             ZSTD_storeSeq(seqStore, newLitLength, ip - newLitLength, iend,
-                          sequence.offset + ZSTD_REP_MOVE,
-                          sequence.matchLength - MINMATCH);
+                          OFFSET_TO_OFFBASE(sequence.offset),
+                          sequence.matchLength);
             ip += sequence.matchLength;
         }
     }
diff --git a/contrib/zstd/zstd_ldm.h b/contrib/zstd/zstd_ldm.h
index aa2b6c52e..38ef9fb36 100644
--- a/contrib/zstd/zstd_ldm.h
+++ b/contrib/zstd/zstd_ldm.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -66,6 +66,7 @@ size_t ZSTD_ldm_generateSequences(
  */
 size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore,
             ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+            ZSTD_paramSwitch_e useRowMatchFinder,
             void const* src, size_t srcSize);
 
 /**
@@ -73,11 +74,17 @@ size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore,
  *
  * Skip past `srcSize` bytes worth of sequences in `rawSeqStore`.
  * Avoids emitting matches less than `minMatch` bytes.
- * Must be called for data with is not passed to ZSTD_ldm_blockCompress().
+ * Must be called for data that is not passed to ZSTD_ldm_blockCompress().
  */
 void ZSTD_ldm_skipSequences(rawSeqStore_t* rawSeqStore, size_t srcSize,
     U32 const minMatch);
 
+/* ZSTD_ldm_skipRawSeqStoreBytes():
+ * Moves forward in rawSeqStore by nbBytes, updating fields 'pos' and 'posInSequence'.
+ * Not to be used in conjunction with ZSTD_ldm_skipSequences().
+ * Must be called for data with is not passed to ZSTD_ldm_blockCompress().
+ */
+void ZSTD_ldm_skipRawSeqStoreBytes(rawSeqStore_t* rawSeqStore, size_t nbBytes);
 
 /** ZSTD_ldm_getTableSize() :
  *  Estimate the space needed for long distance matching tables or 0 if LDM is
diff --git a/contrib/zstd/zstd_ldm_geartab.h b/contrib/zstd/zstd_ldm_geartab.h
new file mode 100644
index 000000000..92455a0b8
--- /dev/null
+++ b/contrib/zstd/zstd_ldm_geartab.h
@@ -0,0 +1,106 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_LDM_GEARTAB_H
+#define ZSTD_LDM_GEARTAB_H
+
+#include "compiler.h" /* UNUSED_ATTR */
+#include "mem.h"      /* U64 */
+
+static UNUSED_ATTR const U64 ZSTD_ldm_gearTab[256] = {
+    0xf5b8f72c5f77775c, 0x84935f266b7ac412, 0xb647ada9ca730ccc,
+    0xb065bb4b114fb1de, 0x34584e7e8c3a9fd0, 0x4e97e17c6ae26b05,
+    0x3a03d743bc99a604, 0xcecd042422c4044f, 0x76de76c58524259e,
+    0x9c8528f65badeaca, 0x86563706e2097529, 0x2902475fa375d889,
+    0xafb32a9739a5ebe6, 0xce2714da3883e639, 0x21eaf821722e69e,
+    0x37b628620b628,    0x49a8d455d88caf5,  0x8556d711e6958140,
+    0x4f7ae74fc605c1f,  0x829f0c3468bd3a20, 0x4ffdc885c625179e,
+    0x8473de048a3daf1b, 0x51008822b05646b2, 0x69d75d12b2d1cc5f,
+    0x8c9d4a19159154bc, 0xc3cc10f4abbd4003, 0xd06ddc1cecb97391,
+    0xbe48e6e7ed80302e, 0x3481db31cee03547, 0xacc3f67cdaa1d210,
+    0x65cb771d8c7f96cc, 0x8eb27177055723dd, 0xc789950d44cd94be,
+    0x934feadc3700b12b, 0x5e485f11edbdf182, 0x1e2e2a46fd64767a,
+    0x2969ca71d82efa7c, 0x9d46e9935ebbba2e, 0xe056b67e05e6822b,
+    0x94d73f55739d03a0, 0xcd7010bdb69b5a03, 0x455ef9fcd79b82f4,
+    0x869cb54a8749c161, 0x38d1a4fa6185d225, 0xb475166f94bbe9bb,
+    0xa4143548720959f1, 0x7aed4780ba6b26ba, 0xd0ce264439e02312,
+    0x84366d746078d508, 0xa8ce973c72ed17be, 0x21c323a29a430b01,
+    0x9962d617e3af80ee, 0xab0ce91d9c8cf75b, 0x530e8ee6d19a4dbc,
+    0x2ef68c0cf53f5d72, 0xc03a681640a85506, 0x496e4e9f9c310967,
+    0x78580472b59b14a0, 0x273824c23b388577, 0x66bf923ad45cb553,
+    0x47ae1a5a2492ba86, 0x35e304569e229659, 0x4765182a46870b6f,
+    0x6cbab625e9099412, 0xddac9a2e598522c1, 0x7172086e666624f2,
+    0xdf5003ca503b7837, 0x88c0c1db78563d09, 0x58d51865acfc289d,
+    0x177671aec65224f1, 0xfb79d8a241e967d7, 0x2be1e101cad9a49a,
+    0x6625682f6e29186b, 0x399553457ac06e50, 0x35dffb4c23abb74,
+    0x429db2591f54aade, 0xc52802a8037d1009, 0x6acb27381f0b25f3,
+    0xf45e2551ee4f823b, 0x8b0ea2d99580c2f7, 0x3bed519cbcb4e1e1,
+    0xff452823dbb010a,  0x9d42ed614f3dd267, 0x5b9313c06257c57b,
+    0xa114b8008b5e1442, 0xc1fe311c11c13d4b, 0x66e8763ea34c5568,
+    0x8b982af1c262f05d, 0xee8876faaa75fbb7, 0x8a62a4d0d172bb2a,
+    0xc13d94a3b7449a97, 0x6dbbba9dc15d037c, 0xc786101f1d92e0f1,
+    0xd78681a907a0b79b, 0xf61aaf2962c9abb9, 0x2cfd16fcd3cb7ad9,
+    0x868c5b6744624d21, 0x25e650899c74ddd7, 0xba042af4a7c37463,
+    0x4eb1a539465a3eca, 0xbe09dbf03b05d5ca, 0x774e5a362b5472ba,
+    0x47a1221229d183cd, 0x504b0ca18ef5a2df, 0xdffbdfbde2456eb9,
+    0x46cd2b2fbee34634, 0xf2aef8fe819d98c3, 0x357f5276d4599d61,
+    0x24a5483879c453e3, 0x88026889192b4b9,  0x28da96671782dbec,
+    0x4ef37c40588e9aaa, 0x8837b90651bc9fb3, 0xc164f741d3f0e5d6,
+    0xbc135a0a704b70ba, 0x69cd868f7622ada,  0xbc37ba89e0b9c0ab,
+    0x47c14a01323552f6, 0x4f00794bacee98bb, 0x7107de7d637a69d5,
+    0x88af793bb6f2255e, 0xf3c6466b8799b598, 0xc288c616aa7f3b59,
+    0x81ca63cf42fca3fd, 0x88d85ace36a2674b, 0xd056bd3792389e7,
+    0xe55c396c4e9dd32d, 0xbefb504571e6c0a6, 0x96ab32115e91e8cc,
+    0xbf8acb18de8f38d1, 0x66dae58801672606, 0x833b6017872317fb,
+    0xb87c16f2d1c92864, 0xdb766a74e58b669c, 0x89659f85c61417be,
+    0xc8daad856011ea0c, 0x76a4b565b6fe7eae, 0xa469d085f6237312,
+    0xaaf0365683a3e96c, 0x4dbb746f8424f7b8, 0x638755af4e4acc1,
+    0x3d7807f5bde64486, 0x17be6d8f5bbb7639, 0x903f0cd44dc35dc,
+    0x67b672eafdf1196c, 0xa676ff93ed4c82f1, 0x521d1004c5053d9d,
+    0x37ba9ad09ccc9202, 0x84e54d297aacfb51, 0xa0b4b776a143445,
+    0x820d471e20b348e,  0x1874383cb83d46dc, 0x97edeec7a1efe11c,
+    0xb330e50b1bdc42aa, 0x1dd91955ce70e032, 0xa514cdb88f2939d5,
+    0x2791233fd90db9d3, 0x7b670a4cc50f7a9b, 0x77c07d2a05c6dfa5,
+    0xe3778b6646d0a6fa, 0xb39c8eda47b56749, 0x933ed448addbef28,
+    0xaf846af6ab7d0bf4, 0xe5af208eb666e49,  0x5e6622f73534cd6a,
+    0x297daeca42ef5b6e, 0x862daef3d35539a6, 0xe68722498f8e1ea9,
+    0x981c53093dc0d572, 0xfa09b0bfbf86fbf5, 0x30b1e96166219f15,
+    0x70e7d466bdc4fb83, 0x5a66736e35f2a8e9, 0xcddb59d2b7c1baef,
+    0xd6c7d247d26d8996, 0xea4e39eac8de1ba3, 0x539c8bb19fa3aff2,
+    0x9f90e4c5fd508d8,  0xa34e5956fbaf3385, 0x2e2f8e151d3ef375,
+    0x173691e9b83faec1, 0xb85a8d56bf016379, 0x8382381267408ae3,
+    0xb90f901bbdc0096d, 0x7c6ad32933bcec65, 0x76bb5e2f2c8ad595,
+    0x390f851a6cf46d28, 0xc3e6064da1c2da72, 0xc52a0c101cfa5389,
+    0xd78eaf84a3fbc530, 0x3781b9e2288b997e, 0x73c2f6dea83d05c4,
+    0x4228e364c5b5ed7,  0x9d7a3edf0da43911, 0x8edcfeda24686756,
+    0x5e7667a7b7a9b3a1, 0x4c4f389fa143791d, 0xb08bc1023da7cddc,
+    0x7ab4be3ae529b1cc, 0x754e6132dbe74ff9, 0x71635442a839df45,
+    0x2f6fb1643fbe52de, 0x961e0a42cf7a8177, 0xf3b45d83d89ef2ea,
+    0xee3de4cf4a6e3e9b, 0xcd6848542c3295e7, 0xe4cee1664c78662f,
+    0x9947548b474c68c4, 0x25d73777a5ed8b0b, 0xc915b1d636b7fc,
+    0x21c2ba75d9b0d2da, 0x5f6b5dcf608a64a1, 0xdcf333255ff9570c,
+    0x633b922418ced4ee, 0xc136dde0b004b34a, 0x58cc83b05d4b2f5a,
+    0x5eb424dda28e42d2, 0x62df47369739cd98, 0xb4e0b42485e4ce17,
+    0x16e1f0c1f9a8d1e7, 0x8ec3916707560ebf, 0x62ba6e2df2cc9db3,
+    0xcbf9f4ff77d83a16, 0x78d9d7d07d2bbcc4, 0xef554ce1e02c41f4,
+    0x8d7581127eccf94d, 0xa9b53336cb3c8a05, 0x38c42c0bf45c4f91,
+    0x640893cdf4488863, 0x80ec34bc575ea568, 0x39f324f5b48eaa40,
+    0xe9d9ed1f8eff527f, 0x9224fc058cc5a214, 0xbaba00b04cfe7741,
+    0x309a9f120fcf52af, 0xa558f3ec65626212, 0x424bec8b7adabe2f,
+    0x41622513a6aea433, 0xb88da2d5324ca798, 0xd287733b245528a4,
+    0x9a44697e6d68aec3, 0x7b1093be2f49bb28, 0x50bbec632e3d8aad,
+    0x6cd90723e1ea8283, 0x897b9e7431b02bf3, 0x219efdcb338a7047,
+    0x3b0311f0a27c0656, 0xdb17bf91c0db96e7, 0x8cd4fd6b4e85a5b2,
+    0xfab071054ba6409d, 0x40d6fe831fa9dfd9, 0xaf358debad7d791e,
+    0xeb8d0e25a65e3e58, 0xbbcbd3df14e08580, 0xcf751f27ecdab2b,
+    0x2b4da14f2613d8f4
+};
+
+#endif /* ZSTD_LDM_GEARTAB_H */
diff --git a/contrib/zstd/zstd_opt.c b/contrib/zstd/zstd_opt.c
index 36fff050c..fdd7f9d8b 100644
--- a/contrib/zstd/zstd_opt.c
+++ b/contrib/zstd/zstd_opt.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Przemyslaw Skibinski, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -14,40 +14,47 @@
 
 
 #define ZSTD_LITFREQ_ADD    2   /* scaling factor for litFreq, so that frequencies adapt faster to new stats */
-#define ZSTD_FREQ_DIV       4   /* log factor when using previous stats to init next stats */
 #define ZSTD_MAX_PRICE     (1<<30)
 
-#define ZSTD_PREDEF_THRESHOLD 1024   /* if srcSize < ZSTD_PREDEF_THRESHOLD, symbols' cost is assumed static, directly determined by pre-defined distributions */
+#define ZSTD_PREDEF_THRESHOLD 8   /* if srcSize < ZSTD_PREDEF_THRESHOLD, symbols' cost is assumed static, directly determined by pre-defined distributions */
 
 
 /*-*************************************
 *  Price functions for optimal parser
 ***************************************/
 
-#if 0    /* approximation at bit level */
+#if 0    /* approximation at bit level (for tests) */
 #  define BITCOST_ACCURACY 0
 #  define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY)
-#  define WEIGHT(stat)  ((void)opt, ZSTD_bitWeight(stat))
-#elif 0  /* fractional bit accuracy */
+#  define WEIGHT(stat, opt) ((void)(opt), ZSTD_bitWeight(stat))
+#elif 0  /* fractional bit accuracy (for tests) */
 #  define BITCOST_ACCURACY 8
 #  define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY)
-#  define WEIGHT(stat,opt) ((void)opt, ZSTD_fracWeight(stat))
+#  define WEIGHT(stat,opt) ((void)(opt), ZSTD_fracWeight(stat))
 #else    /* opt==approx, ultra==accurate */
 #  define BITCOST_ACCURACY 8
 #  define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY)
-#  define WEIGHT(stat,opt) (opt ? ZSTD_fracWeight(stat) : ZSTD_bitWeight(stat))
+#  define WEIGHT(stat,opt) ((opt) ? ZSTD_fracWeight(stat) : ZSTD_bitWeight(stat))
 #endif
 
+/* ZSTD_bitWeight() :
+ * provide estimated "cost" of a stat in full bits only */
 MEM_STATIC U32 ZSTD_bitWeight(U32 stat)
 {
     return (ZSTD_highbit32(stat+1) * BITCOST_MULTIPLIER);
 }
 
+/* ZSTD_fracWeight() :
+ * provide fractional-bit "cost" of a stat,
+ * using linear interpolation approximation */
 MEM_STATIC U32 ZSTD_fracWeight(U32 rawStat)
 {
     U32 const stat = rawStat + 1;
     U32 const hb = ZSTD_highbit32(stat);
     U32 const BWeight = hb * BITCOST_MULTIPLIER;
+    /* Fweight was meant for "Fractional weight"
+     * but it's effectively a value between 1 and 2
+     * using fixed point arithmetic */
     U32 const FWeight = (stat << BITCOST_ACCURACY) >> hb;
     U32 const weight = BWeight + FWeight;
     assert(hb + BITCOST_ACCURACY < 31);
@@ -58,7 +65,7 @@ MEM_STATIC U32 ZSTD_fracWeight(U32 rawStat)
 /* debugging function,
  * @return price in bytes as fractional value
  * for debug messages only */
-MEM_STATIC double ZSTD_fCost(U32 price)
+MEM_STATIC double ZSTD_fCost(int price)
 {
     return (double)price / (BITCOST_MULTIPLIER*8);
 }
@@ -66,7 +73,7 @@ MEM_STATIC double ZSTD_fCost(U32 price)
 
 static int ZSTD_compressedLiterals(optState_t const* const optPtr)
 {
-    return optPtr->literalCompressionMode != ZSTD_lcm_uncompressed;
+    return optPtr->literalCompressionMode != ZSTD_ps_disable;
 }
 
 static void ZSTD_setBasePrices(optState_t* optPtr, int optLevel)
@@ -79,25 +86,52 @@ static void ZSTD_setBasePrices(optState_t* optPtr, int optLevel)
 }
 
 
-/* ZSTD_downscaleStat() :
- * reduce all elements in table by a factor 2^(ZSTD_FREQ_DIV+malus)
- * return the resulting sum of elements */
-static U32 ZSTD_downscaleStat(unsigned* table, U32 lastEltIndex, int malus)
+static U32 sum_u32(const unsigned table[], size_t nbElts)
+{
+    size_t n;
+    U32 total = 0;
+    for (n=0; n<nbElts; n++) {
+        total += table[n];
+    }
+    return total;
+}
+
+typedef enum { base_0possible=0, base_1guaranteed=1 } base_directive_e;
+
+static U32
+ZSTD_downscaleStats(unsigned* table, U32 lastEltIndex, U32 shift, base_directive_e base1)
 {
     U32 s, sum=0;
-    DEBUGLOG(5, "ZSTD_downscaleStat (nbElts=%u)", (unsigned)lastEltIndex+1);
-    assert(ZSTD_FREQ_DIV+malus > 0 && ZSTD_FREQ_DIV+malus < 31);
+    DEBUGLOG(5, "ZSTD_downscaleStats (nbElts=%u, shift=%u)",
+            (unsigned)lastEltIndex+1, (unsigned)shift );
+    assert(shift < 30);
     for (s=0; s<lastEltIndex+1; s++) {
-        table[s] = 1 + (table[s] >> (ZSTD_FREQ_DIV+malus));
-        sum += table[s];
+        unsigned const base = base1 ? 1 : (table[s]>0);
+        unsigned const newStat = base + (table[s] >> shift);
+        sum += newStat;
+        table[s] = newStat;
     }
     return sum;
 }
 
+/* ZSTD_scaleStats() :
+ * reduce all elt frequencies in table if sum too large
+ * return the resulting sum of elements */
+static U32 ZSTD_scaleStats(unsigned* table, U32 lastEltIndex, U32 logTarget)
+{
+    U32 const prevsum = sum_u32(table, lastEltIndex+1);
+    U32 const factor = prevsum >> logTarget;
+    DEBUGLOG(5, "ZSTD_scaleStats (nbElts=%u, target=%u)", (unsigned)lastEltIndex+1, (unsigned)logTarget);
+    assert(logTarget < 30);
+    if (factor <= 1) return prevsum;
+    return ZSTD_downscaleStats(table, lastEltIndex, ZSTD_highbit32(factor), base_1guaranteed);
+}
+
 /* ZSTD_rescaleFreqs() :
  * if first block (detected by optPtr->litLengthSum == 0) : init statistics
  *    take hints from dictionary if there is one
- *    or init from zero, using src for literals stats, or flat 1 for match symbols
+ *    and init from zero if there is none,
+ *    using src for literals stats, and baseline stats for sequence symbols
  * otherwise downscale existing stats, to be used as seed for next block.
  */
 static void
@@ -109,24 +143,28 @@ ZSTD_rescaleFreqs(optState_t* const optPtr,
     DEBUGLOG(5, "ZSTD_rescaleFreqs (srcSize=%u)", (unsigned)srcSize);
     optPtr->priceType = zop_dynamic;
 
-    if (optPtr->litLengthSum == 0) {  /* first block : init */
-        if (srcSize <= ZSTD_PREDEF_THRESHOLD) {  /* heuristic */
-            DEBUGLOG(5, "(srcSize <= ZSTD_PREDEF_THRESHOLD) => zop_predef");
+    if (optPtr->litLengthSum == 0) {  /* no literals stats collected -> first block assumed -> init */
+
+        /* heuristic: use pre-defined stats for too small inputs */
+        if (srcSize <= ZSTD_PREDEF_THRESHOLD) {
+            DEBUGLOG(5, "srcSize <= %i : use predefined stats", ZSTD_PREDEF_THRESHOLD);
             optPtr->priceType = zop_predef;
         }
 
         assert(optPtr->symbolCosts != NULL);
         if (optPtr->symbolCosts->huf.repeatMode == HUF_repeat_valid) {
-            /* huffman table presumed generated by dictionary */
+
+            /* huffman stats covering the full value set : table presumed generated by dictionary */
             optPtr->priceType = zop_dynamic;
 
             if (compressedLiterals) {
+                /* generate literals statistics from huffman table */
                 unsigned lit;
                 assert(optPtr->litFreq != NULL);
                 optPtr->litSum = 0;
                 for (lit=0; lit<=MaxLit; lit++) {
                     U32 const scaleLog = 11;   /* scale to 2K */
-                    U32 const bitCost = HUF_getNbBits(optPtr->symbolCosts->huf.CTable, lit);
+                    U32 const bitCost = HUF_getNbBitsFromCTable(optPtr->symbolCosts->huf.CTable, lit);
                     assert(bitCost <= scaleLog);
                     optPtr->litFreq[lit] = bitCost ? 1 << (scaleLog-bitCost) : 1 /*minimum to calculate cost*/;
                     optPtr->litSum += optPtr->litFreq[lit];
@@ -168,20 +206,26 @@ ZSTD_rescaleFreqs(optState_t* const optPtr,
                     optPtr->offCodeSum += optPtr->offCodeFreq[of];
             }   }
 
-        } else {  /* not a dictionary */
+        } else {  /* first block, no dictionary */
 
             assert(optPtr->litFreq != NULL);
             if (compressedLiterals) {
+                /* base initial cost of literals on direct frequency within src */
                 unsigned lit = MaxLit;
                 HIST_count_simple(optPtr->litFreq, &lit, src, srcSize);   /* use raw first block to init statistics */
-                optPtr->litSum = ZSTD_downscaleStat(optPtr->litFreq, MaxLit, 1);
+                optPtr->litSum = ZSTD_downscaleStats(optPtr->litFreq, MaxLit, 8, base_0possible);
             }
 
-            {   unsigned ll;
-                for (ll=0; ll<=MaxLL; ll++)
-                    optPtr->litLengthFreq[ll] = 1;
+            {   unsigned const baseLLfreqs[MaxLL+1] = {
+                    4, 2, 1, 1, 1, 1, 1, 1,
+                    1, 1, 1, 1, 1, 1, 1, 1,
+                    1, 1, 1, 1, 1, 1, 1, 1,
+                    1, 1, 1, 1, 1, 1, 1, 1,
+                    1, 1, 1, 1
+                };
+                ZSTD_memcpy(optPtr->litLengthFreq, baseLLfreqs, sizeof(baseLLfreqs));
+                optPtr->litLengthSum = sum_u32(baseLLfreqs, MaxLL+1);
             }
-            optPtr->litLengthSum = MaxLL+1;
 
             {   unsigned ml;
                 for (ml=0; ml<=MaxML; ml++)
@@ -189,21 +233,25 @@ ZSTD_rescaleFreqs(optState_t* const optPtr,
             }
             optPtr->matchLengthSum = MaxML+1;
 
-            {   unsigned of;
-                for (of=0; of<=MaxOff; of++)
-                    optPtr->offCodeFreq[of] = 1;
+            {   unsigned const baseOFCfreqs[MaxOff+1] = {
+                    6, 2, 1, 1, 2, 3, 4, 4,
+                    4, 3, 2, 1, 1, 1, 1, 1,
+                    1, 1, 1, 1, 1, 1, 1, 1,
+                    1, 1, 1, 1, 1, 1, 1, 1
+                };
+                ZSTD_memcpy(optPtr->offCodeFreq, baseOFCfreqs, sizeof(baseOFCfreqs));
+                optPtr->offCodeSum = sum_u32(baseOFCfreqs, MaxOff+1);
             }
-            optPtr->offCodeSum = MaxOff+1;
 
         }
 
-    } else {   /* new block : re-use previous statistics, scaled down */
+    } else {   /* new block : scale down accumulated statistics */
 
         if (compressedLiterals)
-            optPtr->litSum = ZSTD_downscaleStat(optPtr->litFreq, MaxLit, 1);
-        optPtr->litLengthSum = ZSTD_downscaleStat(optPtr->litLengthFreq, MaxLL, 0);
-        optPtr->matchLengthSum = ZSTD_downscaleStat(optPtr->matchLengthFreq, MaxML, 0);
-        optPtr->offCodeSum = ZSTD_downscaleStat(optPtr->offCodeFreq, MaxOff, 0);
+            optPtr->litSum = ZSTD_scaleStats(optPtr->litFreq, MaxLit, 12);
+        optPtr->litLengthSum = ZSTD_scaleStats(optPtr->litLengthFreq, MaxLL, 11);
+        optPtr->matchLengthSum = ZSTD_scaleStats(optPtr->matchLengthFreq, MaxML, 11);
+        optPtr->offCodeSum = ZSTD_scaleStats(optPtr->offCodeFreq, MaxOff, 11);
     }
 
     ZSTD_setBasePrices(optPtr, optLevel);
@@ -225,11 +273,14 @@ static U32 ZSTD_rawLiteralsCost(const BYTE* const literals, U32 const litLength,
         return (litLength*6) * BITCOST_MULTIPLIER;  /* 6 bit per literal - no statistic used */
 
     /* dynamic statistics */
-    {   U32 price = litLength * optPtr->litSumBasePrice;
+    {   U32 price = optPtr->litSumBasePrice * litLength;
+        U32 const litPriceMax = optPtr->litSumBasePrice - BITCOST_MULTIPLIER;
         U32 u;
+        assert(optPtr->litSumBasePrice >= BITCOST_MULTIPLIER);
         for (u=0; u < litLength; u++) {
-            assert(WEIGHT(optPtr->litFreq[literals[u]], optLevel) <= optPtr->litSumBasePrice);   /* literal cost should never be negative */
-            price -= WEIGHT(optPtr->litFreq[literals[u]], optLevel);
+            U32 litPrice = WEIGHT(optPtr->litFreq[literals[u]], optLevel);
+            if (UNLIKELY(litPrice > litPriceMax)) litPrice = litPriceMax;
+            price -= litPrice;
         }
         return price;
     }
@@ -239,7 +290,17 @@ static U32 ZSTD_rawLiteralsCost(const BYTE* const literals, U32 const litLength,
  * cost of literalLength symbol */
 static U32 ZSTD_litLengthPrice(U32 const litLength, const optState_t* const optPtr, int optLevel)
 {
-    if (optPtr->priceType == zop_predef) return WEIGHT(litLength, optLevel);
+    assert(litLength <= ZSTD_BLOCKSIZE_MAX);
+    if (optPtr->priceType == zop_predef)
+        return WEIGHT(litLength, optLevel);
+
+    /* ZSTD_LLcode() can't compute litLength price for sizes >= ZSTD_BLOCKSIZE_MAX
+     * because it isn't representable in the zstd format.
+     * So instead just pretend it would cost 1 bit more than ZSTD_BLOCKSIZE_MAX - 1.
+     * In such a case, the block would be all literals.
+     */
+    if (litLength == ZSTD_BLOCKSIZE_MAX)
+        return BITCOST_MULTIPLIER + ZSTD_litLengthPrice(ZSTD_BLOCKSIZE_MAX - 1, optPtr, optLevel);
 
     /* dynamic statistics */
     {   U32 const llCode = ZSTD_LLcode(litLength);
@@ -250,22 +311,25 @@ static U32 ZSTD_litLengthPrice(U32 const litLength, const optState_t* const optP
 }
 
 /* ZSTD_getMatchPrice() :
- * Provides the cost of the match part (offset + matchLength) of a sequence
+ * Provides the cost of the match part (offset + matchLength) of a sequence.
  * Must be combined with ZSTD_fullLiteralsCost() to get the full cost of a sequence.
- * optLevel: when <2, favors small offset for decompression speed (improved cache efficiency) */
+ * @offBase : sumtype, representing an offset or a repcode, and using numeric representation of ZSTD_storeSeq()
+ * @optLevel: when <2, favors small offset for decompression speed (improved cache efficiency)
+ */
 FORCE_INLINE_TEMPLATE U32
-ZSTD_getMatchPrice(U32 const offset,
+ZSTD_getMatchPrice(U32 const offBase,
                    U32 const matchLength,
              const optState_t* const optPtr,
                    int const optLevel)
 {
     U32 price;
-    U32 const offCode = ZSTD_highbit32(offset+1);
+    U32 const offCode = ZSTD_highbit32(offBase);
     U32 const mlBase = matchLength - MINMATCH;
     assert(matchLength >= MINMATCH);
 
-    if (optPtr->priceType == zop_predef)  /* fixed scheme, do not use statistics */
-        return WEIGHT(mlBase, optLevel) + ((16 + offCode) * BITCOST_MULTIPLIER);
+    if (optPtr->priceType == zop_predef)  /* fixed scheme, does not use statistics */
+        return WEIGHT(mlBase, optLevel)
+             + ((16 + offCode) * BITCOST_MULTIPLIER); /* emulated offset cost */
 
     /* dynamic statistics */
     price = (offCode * BITCOST_MULTIPLIER) + (optPtr->offCodeSumBasePrice - WEIGHT(optPtr->offCodeFreq[offCode], optLevel));
@@ -284,10 +348,10 @@ ZSTD_getMatchPrice(U32 const offset,
 }
 
 /* ZSTD_updateStats() :
- * assumption : literals + litLengtn <= iend */
+ * assumption : literals + litLength <= iend */
 static void ZSTD_updateStats(optState_t* const optPtr,
                              U32 litLength, const BYTE* literals,
-                             U32 offsetCode, U32 matchLength)
+                             U32 offBase, U32 matchLength)
 {
     /* literals */
     if (ZSTD_compressedLiterals(optPtr)) {
@@ -303,8 +367,8 @@ static void ZSTD_updateStats(optState_t* const optPtr,
         optPtr->litLengthSum++;
     }
 
-    /* match offset code (0-2=>repCode; 3+=>offset+2) */
-    {   U32 const offCode = ZSTD_highbit32(offsetCode+1);
+    /* offset code : follows storeSeq() numeric representation */
+    {   U32 const offCode = ZSTD_highbit32(offBase);
         assert(offCode <= MaxOff);
         optPtr->offCodeFreq[offCode]++;
         optPtr->offCodeSum++;
@@ -338,7 +402,7 @@ MEM_STATIC U32 ZSTD_readMINMATCH(const void* memPtr, U32 length)
 
 /* Update hashTable3 up to ip (excluded)
    Assumption : always within prefix (i.e. not within extDict) */
-static U32 ZSTD_insertAndFindFirstIndexHash3 (ZSTD_matchState_t* ms,
+static U32 ZSTD_insertAndFindFirstIndexHash3 (const ZSTD_matchState_t* ms,
                                               U32* nextToUpdate3,
                                               const BYTE* const ip)
 {
@@ -364,11 +428,13 @@ static U32 ZSTD_insertAndFindFirstIndexHash3 (ZSTD_matchState_t* ms,
 *  Binary Tree search
 ***************************************/
 /** ZSTD_insertBt1() : add one or multiple positions to tree.
- *  ip : assumed <= iend-8 .
+ * @param ip assumed <= iend-8 .
+ * @param target The target of ZSTD_updateTree_internal() - we are filling to this position
  * @return : nb of positions added */
 static U32 ZSTD_insertBt1(
-                ZSTD_matchState_t* ms,
+                const ZSTD_matchState_t* ms,
                 const BYTE* const ip, const BYTE* const iend,
+                U32 const target,
                 U32 const mls, const int extDict)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
@@ -386,32 +452,36 @@ static U32 ZSTD_insertBt1(
     const BYTE* const dictEnd = dictBase + dictLimit;
     const BYTE* const prefixStart = base + dictLimit;
     const BYTE* match;
-    const U32 current = (U32)(ip-base);
-    const U32 btLow = btMask >= current ? 0 : current - btMask;
-    U32* smallerPtr = bt + 2*(current&btMask);
+    const U32 curr = (U32)(ip-base);
+    const U32 btLow = btMask >= curr ? 0 : curr - btMask;
+    U32* smallerPtr = bt + 2*(curr&btMask);
     U32* largerPtr  = smallerPtr + 1;
     U32 dummy32;   /* to be nullified at the end */
-    U32 const windowLow = ms->window.lowLimit;
-    U32 matchEndIdx = current+8+1;
+    /* windowLow is based on target because
+     * we only need positions that will be in the window at the end of the tree update.
+     */
+    U32 const windowLow = ZSTD_getLowestMatchIndex(ms, target, cParams->windowLog);
+    U32 matchEndIdx = curr+8+1;
     size_t bestLength = 8;
     U32 nbCompares = 1U << cParams->searchLog;
 #ifdef ZSTD_C_PREDICT
-    U32 predictedSmall = *(bt + 2*((current-1)&btMask) + 0);
-    U32 predictedLarge = *(bt + 2*((current-1)&btMask) + 1);
+    U32 predictedSmall = *(bt + 2*((curr-1)&btMask) + 0);
+    U32 predictedLarge = *(bt + 2*((curr-1)&btMask) + 1);
     predictedSmall += (predictedSmall>0);
     predictedLarge += (predictedLarge>0);
 #endif /* ZSTD_C_PREDICT */
 
-    DEBUGLOG(8, "ZSTD_insertBt1 (%u)", current);
+    DEBUGLOG(8, "ZSTD_insertBt1 (%u)", curr);
 
+    assert(curr <= target);
     assert(ip <= iend-8);   /* required for h calculation */
-    hashTable[h] = current;   /* Update Hash Table */
+    hashTable[h] = curr;   /* Update Hash Table */
 
     assert(windowLow > 0);
-    while (nbCompares-- && (matchIndex >= windowLow)) {
+    for (; nbCompares && (matchIndex >= windowLow); --nbCompares) {
         U32* const nextPtr = bt + 2*(matchIndex & btMask);
         size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
-        assert(matchIndex < current);
+        assert(matchIndex < curr);
 
 #ifdef ZSTD_C_PREDICT   /* note : can create issues when hlog small <= 11 */
         const U32* predictPtr = bt + 2*((matchIndex-1) & btMask);   /* written this way, as bt is a roll buffer */
@@ -474,8 +544,8 @@ static U32 ZSTD_insertBt1(
     *smallerPtr = *largerPtr = 0;
     {   U32 positions = 0;
         if (bestLength > 384) positions = MIN(192, (U32)(bestLength - 384));   /* speed optimization */
-        assert(matchEndIdx > current + 8);
-        return MAX(positions, matchEndIdx - (current + 8));
+        assert(matchEndIdx > curr + 8);
+        return MAX(positions, matchEndIdx - (curr + 8));
     }
 }
 
@@ -492,7 +562,7 @@ void ZSTD_updateTree_internal(
                 idx, target, dictMode);
 
     while(idx < target) {
-        U32 const forward = ZSTD_insertBt1(ms, base+idx, iend, mls, dictMode == ZSTD_extDict);
+        U32 const forward = ZSTD_insertBt1(ms, base+idx, iend, target, mls, dictMode == ZSTD_extDict);
         assert(idx < (U32)(idx + forward));
         idx += forward;
     }
@@ -505,21 +575,22 @@ void ZSTD_updateTree(ZSTD_matchState_t* ms, const BYTE* ip, const BYTE* iend) {
     ZSTD_updateTree_internal(ms, ip, iend, ms->cParams.minMatch, ZSTD_noDict);
 }
 
-FORCE_INLINE_TEMPLATE
-U32 ZSTD_insertBtAndGetAllMatches (
-                    ZSTD_match_t* matches,   /* store result (found matches) in this table (presumed large enough) */
-                    ZSTD_matchState_t* ms,
-                    U32* nextToUpdate3,
-                    const BYTE* const ip, const BYTE* const iLimit, const ZSTD_dictMode_e dictMode,
-                    const U32 rep[ZSTD_REP_NUM],
-                    U32 const ll0,   /* tells if associated literal length is 0 or not. This value must be 0 or 1 */
-                    const U32 lengthToBeat,
-                    U32 const mls /* template */)
+FORCE_INLINE_TEMPLATE U32
+ZSTD_insertBtAndGetAllMatches (
+                ZSTD_match_t* matches,  /* store result (found matches) in this table (presumed large enough) */
+                ZSTD_matchState_t* ms,
+                U32* nextToUpdate3,
+                const BYTE* const ip, const BYTE* const iLimit,
+                const ZSTD_dictMode_e dictMode,
+                const U32 rep[ZSTD_REP_NUM],
+                const U32 ll0,  /* tells if associated literal length is 0 or not. This value must be 0 or 1 */
+                const U32 lengthToBeat,
+                const U32 mls /* template */)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
     U32 const sufficient_len = MIN(cParams->targetLength, ZSTD_OPT_NUM -1);
     const BYTE* const base = ms->window.base;
-    U32 const current = (U32)(ip-base);
+    U32 const curr = (U32)(ip-base);
     U32 const hashLog = cParams->hashLog;
     U32 const minMatch = (mls==3) ? 3 : 4;
     U32* const hashTable = ms->hashTable;
@@ -533,12 +604,12 @@ U32 ZSTD_insertBtAndGetAllMatches (
     U32 const dictLimit = ms->window.dictLimit;
     const BYTE* const dictEnd = dictBase + dictLimit;
     const BYTE* const prefixStart = base + dictLimit;
-    U32 const btLow = (btMask >= current) ? 0 : current - btMask;
-    U32 const windowLow = ZSTD_getLowestMatchIndex(ms, current, cParams->windowLog);
+    U32 const btLow = (btMask >= curr) ? 0 : curr - btMask;
+    U32 const windowLow = ZSTD_getLowestMatchIndex(ms, curr, cParams->windowLog);
     U32 const matchLow = windowLow ? windowLow : 1;
-    U32* smallerPtr = bt + 2*(current&btMask);
-    U32* largerPtr  = bt + 2*(current&btMask) + 1;
-    U32 matchEndIdx = current+8+1;   /* farthest referenced position of any match => detects repetitive patterns */
+    U32* smallerPtr = bt + 2*(curr&btMask);
+    U32* largerPtr  = bt + 2*(curr&btMask) + 1;
+    U32 matchEndIdx = curr+8+1;   /* farthest referenced position of any match => detects repetitive patterns */
     U32 dummy32;   /* to be nullified at the end */
     U32 mnum = 0;
     U32 nbCompares = 1U << cParams->searchLog;
@@ -557,7 +628,7 @@ U32 ZSTD_insertBtAndGetAllMatches (
     U32         const dmsBtLow      = dictMode == ZSTD_dictMatchState && dmsBtMask < dmsHighLimit - dmsLowLimit ? dmsHighLimit - dmsBtMask : dmsLowLimit;
 
     size_t bestLength = lengthToBeat-1;
-    DEBUGLOG(8, "ZSTD_insertBtAndGetAllMatches: current=%u", current);
+    DEBUGLOG(8, "ZSTD_insertBtAndGetAllMatches: current=%u", curr);
 
     /* check repCode */
     assert(ll0 <= 1);   /* necessarily 1 or 0 */
@@ -565,29 +636,29 @@ U32 ZSTD_insertBtAndGetAllMatches (
         U32 repCode;
         for (repCode = ll0; repCode < lastR; repCode++) {
             U32 const repOffset = (repCode==ZSTD_REP_NUM) ? (rep[0] - 1) : rep[repCode];
-            U32 const repIndex = current - repOffset;
+            U32 const repIndex = curr - repOffset;
             U32 repLen = 0;
-            assert(current >= dictLimit);
-            if (repOffset-1 /* intentional overflow, discards 0 and -1 */ < current-dictLimit) {  /* equivalent to `current > repIndex >= dictLimit` */
+            assert(curr >= dictLimit);
+            if (repOffset-1 /* intentional overflow, discards 0 and -1 */ < curr-dictLimit) {  /* equivalent to `curr > repIndex >= dictLimit` */
                 /* We must validate the repcode offset because when we're using a dictionary the
                  * valid offset range shrinks when the dictionary goes out of bounds.
                  */
                 if ((repIndex >= windowLow) & (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(ip - repOffset, minMatch))) {
                     repLen = (U32)ZSTD_count(ip+minMatch, ip+minMatch-repOffset, iLimit) + minMatch;
                 }
-            } else {  /* repIndex < dictLimit || repIndex >= current */
+            } else {  /* repIndex < dictLimit || repIndex >= curr */
                 const BYTE* const repMatch = dictMode == ZSTD_dictMatchState ?
                                              dmsBase + repIndex - dmsIndexDelta :
                                              dictBase + repIndex;
-                assert(current >= windowLow);
+                assert(curr >= windowLow);
                 if ( dictMode == ZSTD_extDict
-                  && ( ((repOffset-1) /*intentional overflow*/ < current - windowLow)  /* equivalent to `current > repIndex >= windowLow` */
+                  && ( ((repOffset-1) /*intentional overflow*/ < curr - windowLow)  /* equivalent to `curr > repIndex >= windowLow` */
                      & (((U32)((dictLimit-1) - repIndex) >= 3) ) /* intentional overflow : do not test positions overlapping 2 memory segments */)
                   && (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(repMatch, minMatch)) ) {
                     repLen = (U32)ZSTD_count_2segments(ip+minMatch, repMatch+minMatch, iLimit, dictEnd, prefixStart) + minMatch;
                 }
                 if (dictMode == ZSTD_dictMatchState
-                  && ( ((repOffset-1) /*intentional overflow*/ < current - (dmsLowLimit + dmsIndexDelta))  /* equivalent to `current > repIndex >= dmsLowLimit` */
+                  && ( ((repOffset-1) /*intentional overflow*/ < curr - (dmsLowLimit + dmsIndexDelta))  /* equivalent to `curr > repIndex >= dmsLowLimit` */
                      & ((U32)((dictLimit-1) - repIndex) >= 3) ) /* intentional overflow : do not test positions overlapping 2 memory segments */
                   && (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(repMatch, minMatch)) ) {
                     repLen = (U32)ZSTD_count_2segments(ip+minMatch, repMatch+minMatch, iLimit, dmsEnd, prefixStart) + minMatch;
@@ -597,7 +668,7 @@ U32 ZSTD_insertBtAndGetAllMatches (
                 DEBUGLOG(8, "found repCode %u (ll0:%u, offset:%u) of length %u",
                             repCode, ll0, repOffset, repLen);
                 bestLength = repLen;
-                matches[mnum].off = repCode - ll0;
+                matches[mnum].off = REPCODE_TO_OFFBASE(repCode - ll0 + 1);  /* expect value between 1 and 3 */
                 matches[mnum].len = (U32)repLen;
                 mnum++;
                 if ( (repLen > sufficient_len)
@@ -609,7 +680,7 @@ U32 ZSTD_insertBtAndGetAllMatches (
     if ((mls == 3) /*static*/ && (bestLength < mls)) {
         U32 const matchIndex3 = ZSTD_insertAndFindFirstIndexHash3(ms, nextToUpdate3, ip);
         if ((matchIndex3 >= matchLow)
-          & (current - matchIndex3 < (1<<18)) /*heuristic : longer distance likely too expensive*/ ) {
+          & (curr - matchIndex3 < (1<<18)) /*heuristic : longer distance likely too expensive*/ ) {
             size_t mlen;
             if ((dictMode == ZSTD_noDict) /*static*/ || (dictMode == ZSTD_dictMatchState) /*static*/ || (matchIndex3 >= dictLimit)) {
                 const BYTE* const match = base + matchIndex3;
@@ -624,26 +695,26 @@ U32 ZSTD_insertBtAndGetAllMatches (
                 DEBUGLOG(8, "found small match with hlog3, of length %u",
                             (U32)mlen);
                 bestLength = mlen;
-                assert(current > matchIndex3);
+                assert(curr > matchIndex3);
                 assert(mnum==0);  /* no prior solution */
-                matches[0].off = (current - matchIndex3) + ZSTD_REP_MOVE;
+                matches[0].off = OFFSET_TO_OFFBASE(curr - matchIndex3);
                 matches[0].len = (U32)mlen;
                 mnum = 1;
                 if ( (mlen > sufficient_len) |
                      (ip+mlen == iLimit) ) {  /* best possible length */
-                    ms->nextToUpdate = current+1;  /* skip insertion */
+                    ms->nextToUpdate = curr+1;  /* skip insertion */
                     return 1;
         }   }   }
         /* no dictMatchState lookup: dicts don't have a populated HC3 table */
-    }
+    }  /* if (mls == 3) */
 
-    hashTable[h] = current;   /* Update Hash Table */
+    hashTable[h] = curr;   /* Update Hash Table */
 
-    while (nbCompares-- && (matchIndex >= matchLow)) {
+    for (; nbCompares && (matchIndex >= matchLow); --nbCompares) {
         U32* const nextPtr = bt + 2*(matchIndex & btMask);
         const BYTE* match;
         size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
-        assert(current > matchIndex);
+        assert(curr > matchIndex);
 
         if ((dictMode == ZSTD_noDict) || (dictMode == ZSTD_dictMatchState) || (matchIndex+matchLength >= dictLimit)) {
             assert(matchIndex+matchLength >= dictLimit);  /* ensure the condition is correct when !extDict */
@@ -659,21 +730,20 @@ U32 ZSTD_insertBtAndGetAllMatches (
         }
 
         if (matchLength > bestLength) {
-            DEBUGLOG(8, "found match of length %u at distance %u (offCode=%u)",
-                    (U32)matchLength, current - matchIndex, current - matchIndex + ZSTD_REP_MOVE);
+            DEBUGLOG(8, "found match of length %u at distance %u (offBase=%u)",
+                    (U32)matchLength, curr - matchIndex, OFFSET_TO_OFFBASE(curr - matchIndex));
             assert(matchEndIdx > matchIndex);
             if (matchLength > matchEndIdx - matchIndex)
                 matchEndIdx = matchIndex + (U32)matchLength;
             bestLength = matchLength;
-            matches[mnum].off = (current - matchIndex) + ZSTD_REP_MOVE;
+            matches[mnum].off = OFFSET_TO_OFFBASE(curr - matchIndex);
             matches[mnum].len = (U32)matchLength;
             mnum++;
             if ( (matchLength > ZSTD_OPT_NUM)
                | (ip+matchLength == iLimit) /* equal : no way to know if inf or sup */) {
                 if (dictMode == ZSTD_dictMatchState) nbCompares = 0; /* break should also skip searching dms */
                 break; /* drop, to preserve bt consistency (miss a little bit of compression) */
-            }
-        }
+        }   }
 
         if (match[matchLength] < ip[matchLength]) {
             /* match smaller than current */
@@ -692,12 +762,13 @@ U32 ZSTD_insertBtAndGetAllMatches (
 
     *smallerPtr = *largerPtr = 0;
 
+    assert(nbCompares <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */
     if (dictMode == ZSTD_dictMatchState && nbCompares) {
         size_t const dmsH = ZSTD_hashPtr(ip, dmsHashLog, mls);
         U32 dictMatchIndex = dms->hashTable[dmsH];
         const U32* const dmsBt = dms->chainTable;
         commonLengthSmaller = commonLengthLarger = 0;
-        while (nbCompares-- && (dictMatchIndex > dmsLowLimit)) {
+        for (; nbCompares && (dictMatchIndex > dmsLowLimit); --nbCompares) {
             const U32* const nextPtr = dmsBt + 2*(dictMatchIndex & dmsBtMask);
             size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
             const BYTE* match = dmsBase + dictMatchIndex;
@@ -707,19 +778,18 @@ U32 ZSTD_insertBtAndGetAllMatches (
 
             if (matchLength > bestLength) {
                 matchIndex = dictMatchIndex + dmsIndexDelta;
-                DEBUGLOG(8, "found dms match of length %u at distance %u (offCode=%u)",
-                        (U32)matchLength, current - matchIndex, current - matchIndex + ZSTD_REP_MOVE);
+                DEBUGLOG(8, "found dms match of length %u at distance %u (offBase=%u)",
+                        (U32)matchLength, curr - matchIndex, OFFSET_TO_OFFBASE(curr - matchIndex));
                 if (matchLength > matchEndIdx - matchIndex)
                     matchEndIdx = matchIndex + (U32)matchLength;
                 bestLength = matchLength;
-                matches[mnum].off = (current - matchIndex) + ZSTD_REP_MOVE;
+                matches[mnum].off = OFFSET_TO_OFFBASE(curr - matchIndex);
                 matches[mnum].len = (U32)matchLength;
                 mnum++;
                 if ( (matchLength > ZSTD_OPT_NUM)
                    | (ip+matchLength == iLimit) /* equal : no way to know if inf or sup */) {
                     break;   /* drop, to guarantee consistency (miss a little bit of compression) */
-                }
-            }
+            }   }
 
             if (dictMatchIndex <= dmsBtLow) { break; }   /* beyond tree size, stop the search */
             if (match[matchLength] < ip[matchLength]) {
@@ -729,47 +799,242 @@ U32 ZSTD_insertBtAndGetAllMatches (
                 /* match is larger than current */
                 commonLengthLarger = matchLength;
                 dictMatchIndex = nextPtr[0];
-            }
-        }
-    }
+    }   }   }  /* if (dictMode == ZSTD_dictMatchState) */
 
-    assert(matchEndIdx > current+8);
+    assert(matchEndIdx > curr+8);
     ms->nextToUpdate = matchEndIdx - 8;  /* skip repetitive patterns */
     return mnum;
 }
 
+typedef U32 (*ZSTD_getAllMatchesFn)(
+    ZSTD_match_t*,
+    ZSTD_matchState_t*,
+    U32*,
+    const BYTE*,
+    const BYTE*,
+    const U32 rep[ZSTD_REP_NUM],
+    U32 const ll0,
+    U32 const lengthToBeat);
+
+FORCE_INLINE_TEMPLATE U32 ZSTD_btGetAllMatches_internal(
+        ZSTD_match_t* matches,
+        ZSTD_matchState_t* ms,
+        U32* nextToUpdate3,
+        const BYTE* ip,
+        const BYTE* const iHighLimit,
+        const U32 rep[ZSTD_REP_NUM],
+        U32 const ll0,
+        U32 const lengthToBeat,
+        const ZSTD_dictMode_e dictMode,
+        const U32 mls)
+{
+    assert(BOUNDED(3, ms->cParams.minMatch, 6) == mls);
+    DEBUGLOG(8, "ZSTD_BtGetAllMatches(dictMode=%d, mls=%u)", (int)dictMode, mls);
+    if (ip < ms->window.base + ms->nextToUpdate)
+        return 0;   /* skipped area */
+    ZSTD_updateTree_internal(ms, ip, iHighLimit, mls, dictMode);
+    return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, mls);
+}
+
+#define ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, mls) ZSTD_btGetAllMatches_##dictMode##_##mls
+
+#define GEN_ZSTD_BT_GET_ALL_MATCHES_(dictMode, mls)            \
+    static U32 ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, mls)(      \
+            ZSTD_match_t* matches,                             \
+            ZSTD_matchState_t* ms,                             \
+            U32* nextToUpdate3,                                \
+            const BYTE* ip,                                    \
+            const BYTE* const iHighLimit,                      \
+            const U32 rep[ZSTD_REP_NUM],                       \
+            U32 const ll0,                                     \
+            U32 const lengthToBeat)                            \
+    {                                                          \
+        return ZSTD_btGetAllMatches_internal(                  \
+                matches, ms, nextToUpdate3, ip, iHighLimit,    \
+                rep, ll0, lengthToBeat, ZSTD_##dictMode, mls); \
+    }
+
+#define GEN_ZSTD_BT_GET_ALL_MATCHES(dictMode)  \
+    GEN_ZSTD_BT_GET_ALL_MATCHES_(dictMode, 3)  \
+    GEN_ZSTD_BT_GET_ALL_MATCHES_(dictMode, 4)  \
+    GEN_ZSTD_BT_GET_ALL_MATCHES_(dictMode, 5)  \
+    GEN_ZSTD_BT_GET_ALL_MATCHES_(dictMode, 6)
+
+GEN_ZSTD_BT_GET_ALL_MATCHES(noDict)
+GEN_ZSTD_BT_GET_ALL_MATCHES(extDict)
+GEN_ZSTD_BT_GET_ALL_MATCHES(dictMatchState)
+
+#define ZSTD_BT_GET_ALL_MATCHES_ARRAY(dictMode)  \
+    {                                            \
+        ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, 3), \
+        ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, 4), \
+        ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, 5), \
+        ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, 6)  \
+    }
 
-FORCE_INLINE_TEMPLATE U32 ZSTD_BtGetAllMatches (
-                        ZSTD_match_t* matches,   /* store result (match found, increasing size) in this table */
-                        ZSTD_matchState_t* ms,
-                        U32* nextToUpdate3,
-                        const BYTE* ip, const BYTE* const iHighLimit, const ZSTD_dictMode_e dictMode,
-                        const U32 rep[ZSTD_REP_NUM],
-                        U32 const ll0,
-                        U32 const lengthToBeat)
+static ZSTD_getAllMatchesFn
+ZSTD_selectBtGetAllMatches(ZSTD_matchState_t const* ms, ZSTD_dictMode_e const dictMode)
 {
-    const ZSTD_compressionParameters* const cParams = &ms->cParams;
-    U32 const matchLengthSearch = cParams->minMatch;
-    DEBUGLOG(8, "ZSTD_BtGetAllMatches");
-    if (ip < ms->window.base + ms->nextToUpdate) return 0;   /* skipped area */
-    ZSTD_updateTree_internal(ms, ip, iHighLimit, matchLengthSearch, dictMode);
-    switch(matchLengthSearch)
-    {
-    case 3 : return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, 3);
-    default :
-    case 4 : return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, 4);
-    case 5 : return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, 5);
-    case 7 :
-    case 6 : return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, 6);
+    ZSTD_getAllMatchesFn const getAllMatchesFns[3][4] = {
+        ZSTD_BT_GET_ALL_MATCHES_ARRAY(noDict),
+        ZSTD_BT_GET_ALL_MATCHES_ARRAY(extDict),
+        ZSTD_BT_GET_ALL_MATCHES_ARRAY(dictMatchState)
+    };
+    U32 const mls = BOUNDED(3, ms->cParams.minMatch, 6);
+    assert((U32)dictMode < 3);
+    assert(mls - 3 < 4);
+    return getAllMatchesFns[(int)dictMode][mls - 3];
+}
+
+/*************************
+*  LDM helper functions  *
+*************************/
+
+/* Struct containing info needed to make decision about ldm inclusion */
+typedef struct {
+    rawSeqStore_t seqStore;   /* External match candidates store for this block */
+    U32 startPosInBlock;      /* Start position of the current match candidate */
+    U32 endPosInBlock;        /* End position of the current match candidate */
+    U32 offset;               /* Offset of the match candidate */
+} ZSTD_optLdm_t;
+
+/* ZSTD_optLdm_skipRawSeqStoreBytes():
+ * Moves forward in @rawSeqStore by @nbBytes,
+ * which will update the fields 'pos' and 'posInSequence'.
+ */
+static void ZSTD_optLdm_skipRawSeqStoreBytes(rawSeqStore_t* rawSeqStore, size_t nbBytes)
+{
+    U32 currPos = (U32)(rawSeqStore->posInSequence + nbBytes);
+    while (currPos && rawSeqStore->pos < rawSeqStore->size) {
+        rawSeq currSeq = rawSeqStore->seq[rawSeqStore->pos];
+        if (currPos >= currSeq.litLength + currSeq.matchLength) {
+            currPos -= currSeq.litLength + currSeq.matchLength;
+            rawSeqStore->pos++;
+        } else {
+            rawSeqStore->posInSequence = currPos;
+            break;
+        }
+    }
+    if (currPos == 0 || rawSeqStore->pos == rawSeqStore->size) {
+        rawSeqStore->posInSequence = 0;
+    }
+}
+
+/* ZSTD_opt_getNextMatchAndUpdateSeqStore():
+ * Calculates the beginning and end of the next match in the current block.
+ * Updates 'pos' and 'posInSequence' of the ldmSeqStore.
+ */
+static void
+ZSTD_opt_getNextMatchAndUpdateSeqStore(ZSTD_optLdm_t* optLdm, U32 currPosInBlock,
+                                       U32 blockBytesRemaining)
+{
+    rawSeq currSeq;
+    U32 currBlockEndPos;
+    U32 literalsBytesRemaining;
+    U32 matchBytesRemaining;
+
+    /* Setting match end position to MAX to ensure we never use an LDM during this block */
+    if (optLdm->seqStore.size == 0 || optLdm->seqStore.pos >= optLdm->seqStore.size) {
+        optLdm->startPosInBlock = UINT_MAX;
+        optLdm->endPosInBlock = UINT_MAX;
+        return;
+    }
+    /* Calculate appropriate bytes left in matchLength and litLength
+     * after adjusting based on ldmSeqStore->posInSequence */
+    currSeq = optLdm->seqStore.seq[optLdm->seqStore.pos];
+    assert(optLdm->seqStore.posInSequence <= currSeq.litLength + currSeq.matchLength);
+    currBlockEndPos = currPosInBlock + blockBytesRemaining;
+    literalsBytesRemaining = (optLdm->seqStore.posInSequence < currSeq.litLength) ?
+            currSeq.litLength - (U32)optLdm->seqStore.posInSequence :
+            0;
+    matchBytesRemaining = (literalsBytesRemaining == 0) ?
+            currSeq.matchLength - ((U32)optLdm->seqStore.posInSequence - currSeq.litLength) :
+            currSeq.matchLength;
+
+    /* If there are more literal bytes than bytes remaining in block, no ldm is possible */
+    if (literalsBytesRemaining >= blockBytesRemaining) {
+        optLdm->startPosInBlock = UINT_MAX;
+        optLdm->endPosInBlock = UINT_MAX;
+        ZSTD_optLdm_skipRawSeqStoreBytes(&optLdm->seqStore, blockBytesRemaining);
+        return;
+    }
+
+    /* Matches may be < MINMATCH by this process. In that case, we will reject them
+       when we are deciding whether or not to add the ldm */
+    optLdm->startPosInBlock = currPosInBlock + literalsBytesRemaining;
+    optLdm->endPosInBlock = optLdm->startPosInBlock + matchBytesRemaining;
+    optLdm->offset = currSeq.offset;
+
+    if (optLdm->endPosInBlock > currBlockEndPos) {
+        /* Match ends after the block ends, we can't use the whole match */
+        optLdm->endPosInBlock = currBlockEndPos;
+        ZSTD_optLdm_skipRawSeqStoreBytes(&optLdm->seqStore, currBlockEndPos - currPosInBlock);
+    } else {
+        /* Consume nb of bytes equal to size of sequence left */
+        ZSTD_optLdm_skipRawSeqStoreBytes(&optLdm->seqStore, literalsBytesRemaining + matchBytesRemaining);
+    }
+}
+
+/* ZSTD_optLdm_maybeAddMatch():
+ * Adds a match if it's long enough,
+ * based on it's 'matchStartPosInBlock' and 'matchEndPosInBlock',
+ * into 'matches'. Maintains the correct ordering of 'matches'.
+ */
+static void ZSTD_optLdm_maybeAddMatch(ZSTD_match_t* matches, U32* nbMatches,
+                                      const ZSTD_optLdm_t* optLdm, U32 currPosInBlock)
+{
+    U32 const posDiff = currPosInBlock - optLdm->startPosInBlock;
+    /* Note: ZSTD_match_t actually contains offBase and matchLength (before subtracting MINMATCH) */
+    U32 const candidateMatchLength = optLdm->endPosInBlock - optLdm->startPosInBlock - posDiff;
+
+    /* Ensure that current block position is not outside of the match */
+    if (currPosInBlock < optLdm->startPosInBlock
+      || currPosInBlock >= optLdm->endPosInBlock
+      || candidateMatchLength < MINMATCH) {
+        return;
+    }
+
+    if (*nbMatches == 0 || ((candidateMatchLength > matches[*nbMatches-1].len) && *nbMatches < ZSTD_OPT_NUM)) {
+        U32 const candidateOffBase = OFFSET_TO_OFFBASE(optLdm->offset);
+        DEBUGLOG(6, "ZSTD_optLdm_maybeAddMatch(): Adding ldm candidate match (offBase: %u matchLength %u) at block position=%u",
+                 candidateOffBase, candidateMatchLength, currPosInBlock);
+        matches[*nbMatches].len = candidateMatchLength;
+        matches[*nbMatches].off = candidateOffBase;
+        (*nbMatches)++;
     }
 }
 
+/* ZSTD_optLdm_processMatchCandidate():
+ * Wrapper function to update ldm seq store and call ldm functions as necessary.
+ */
+static void
+ZSTD_optLdm_processMatchCandidate(ZSTD_optLdm_t* optLdm,
+                                  ZSTD_match_t* matches, U32* nbMatches,
+                                  U32 currPosInBlock, U32 remainingBytes)
+{
+    if (optLdm->seqStore.size == 0 || optLdm->seqStore.pos >= optLdm->seqStore.size) {
+        return;
+    }
+
+    if (currPosInBlock >= optLdm->endPosInBlock) {
+        if (currPosInBlock > optLdm->endPosInBlock) {
+            /* The position at which ZSTD_optLdm_processMatchCandidate() is called is not necessarily
+             * at the end of a match from the ldm seq store, and will often be some bytes
+             * over beyond matchEndPosInBlock. As such, we need to correct for these "overshoots"
+             */
+            U32 const posOvershoot = currPosInBlock - optLdm->endPosInBlock;
+            ZSTD_optLdm_skipRawSeqStoreBytes(&optLdm->seqStore, posOvershoot);
+        }
+        ZSTD_opt_getNextMatchAndUpdateSeqStore(optLdm, currPosInBlock, remainingBytes);
+    }
+    ZSTD_optLdm_maybeAddMatch(matches, nbMatches, optLdm, currPosInBlock);
+}
+
 
 /*-*******************************
 *  Optimal parser
 *********************************/
 
-
 static U32 ZSTD_totalLen(ZSTD_optimal_t sol)
 {
     return sol.litlen + sol.mlen;
@@ -810,6 +1075,8 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
     const BYTE* const prefixStart = base + ms->window.dictLimit;
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
 
+    ZSTD_getAllMatchesFn getAllMatches = ZSTD_selectBtGetAllMatches(ms, dictMode);
+
     U32 const sufficient_len = MIN(cParams->targetLength, ZSTD_OPT_NUM -1);
     U32 const minMatch = (cParams->minMatch == 3) ? 3 : 4;
     U32 nextToUpdate3 = ms->nextToUpdate;
@@ -817,6 +1084,11 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
     ZSTD_optimal_t* const opt = optStatePtr->priceTable;
     ZSTD_match_t* const matches = optStatePtr->matchTable;
     ZSTD_optimal_t lastSequence;
+    ZSTD_optLdm_t optLdm;
+
+    optLdm.seqStore = ms->ldmSeqStore ? *ms->ldmSeqStore : kNullRawSeqStore;
+    optLdm.endPosInBlock = optLdm.startPosInBlock = optLdm.offset = 0;
+    ZSTD_opt_getNextMatchAndUpdateSeqStore(&optLdm, (U32)(ip-istart), (U32)(iend-ip));
 
     /* init */
     DEBUGLOG(5, "ZSTD_compressBlock_opt_generic: current=%u, prefix=%u, nextToUpdate=%u",
@@ -832,7 +1104,9 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
         /* find first match */
         {   U32 const litlen = (U32)(ip - anchor);
             U32 const ll0 = !litlen;
-            U32 const nbMatches = ZSTD_BtGetAllMatches(matches, ms, &nextToUpdate3, ip, iend, dictMode, rep, ll0, minMatch);
+            U32 nbMatches = getAllMatches(matches, ms, &nextToUpdate3, ip, iend, rep, ll0, minMatch);
+            ZSTD_optLdm_processMatchCandidate(&optLdm, matches, &nbMatches,
+                                              (U32)(ip-istart), (U32)(iend - ip));
             if (!nbMatches) { ip++; continue; }
 
             /* initialize opt[0] */
@@ -844,18 +1118,18 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
              * in every price. We include the literal length to avoid negative
              * prices when we subtract the previous literal length.
              */
-            opt[0].price = ZSTD_litLengthPrice(litlen, optStatePtr, optLevel);
+            opt[0].price = (int)ZSTD_litLengthPrice(litlen, optStatePtr, optLevel);
 
             /* large match -> immediate encoding */
             {   U32 const maxML = matches[nbMatches-1].len;
-                U32 const maxOffset = matches[nbMatches-1].off;
-                DEBUGLOG(6, "found %u matches of maxLength=%u and maxOffCode=%u at cPos=%u => start new series",
-                            nbMatches, maxML, maxOffset, (U32)(ip-prefixStart));
+                U32 const maxOffBase = matches[nbMatches-1].off;
+                DEBUGLOG(6, "found %u matches of maxLength=%u and maxOffBase=%u at cPos=%u => start new series",
+                            nbMatches, maxML, maxOffBase, (U32)(ip-prefixStart));
 
                 if (maxML > sufficient_len) {
                     lastSequence.litlen = litlen;
                     lastSequence.mlen = maxML;
-                    lastSequence.off = maxOffset;
+                    lastSequence.off = maxOffBase;
                     DEBUGLOG(6, "large match (%u>%u), immediate encoding",
                                 maxML, sufficient_len);
                     cur = 0;
@@ -864,24 +1138,25 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
             }   }
 
             /* set prices for first matches starting position == 0 */
-            {   U32 const literalsPrice = opt[0].price + ZSTD_litLengthPrice(0, optStatePtr, optLevel);
+            assert(opt[0].price >= 0);
+            {   U32 const literalsPrice = (U32)opt[0].price + ZSTD_litLengthPrice(0, optStatePtr, optLevel);
                 U32 pos;
                 U32 matchNb;
                 for (pos = 1; pos < minMatch; pos++) {
                     opt[pos].price = ZSTD_MAX_PRICE;   /* mlen, litlen and price will be fixed during forward scanning */
                 }
                 for (matchNb = 0; matchNb < nbMatches; matchNb++) {
-                    U32 const offset = matches[matchNb].off;
+                    U32 const offBase = matches[matchNb].off;
                     U32 const end = matches[matchNb].len;
                     for ( ; pos <= end ; pos++ ) {
-                        U32 const matchPrice = ZSTD_getMatchPrice(offset, pos, optStatePtr, optLevel);
+                        U32 const matchPrice = ZSTD_getMatchPrice(offBase, pos, optStatePtr, optLevel);
                         U32 const sequencePrice = literalsPrice + matchPrice;
                         DEBUGLOG(7, "rPos:%u => set initial price : %.2f",
-                                    pos, ZSTD_fCost(sequencePrice));
+                                    pos, ZSTD_fCost((int)sequencePrice));
                         opt[pos].mlen = pos;
-                        opt[pos].off = offset;
+                        opt[pos].off = offBase;
                         opt[pos].litlen = litlen;
-                        opt[pos].price = sequencePrice;
+                        opt[pos].price = (int)sequencePrice;
                 }   }
                 last_pos = pos-1;
             }
@@ -896,9 +1171,9 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
             /* Fix current position with one literal if cheaper */
             {   U32 const litlen = (opt[cur-1].mlen == 0) ? opt[cur-1].litlen + 1 : 1;
                 int const price = opt[cur-1].price
-                                + ZSTD_rawLiteralsCost(ip+cur-1, 1, optStatePtr, optLevel)
-                                + ZSTD_litLengthPrice(litlen, optStatePtr, optLevel)
-                                - ZSTD_litLengthPrice(litlen-1, optStatePtr, optLevel);
+                                + (int)ZSTD_rawLiteralsCost(ip+cur-1, 1, optStatePtr, optLevel)
+                                + (int)ZSTD_litLengthPrice(litlen, optStatePtr, optLevel)
+                                - (int)ZSTD_litLengthPrice(litlen-1, optStatePtr, optLevel);
                 assert(price < 1000000000); /* overflow check */
                 if (price <= opt[cur].price) {
                     DEBUGLOG(7, "cPos:%zi==rPos:%u : better price (%.2f<=%.2f) using literal (ll==%u) (hist:%u,%u,%u)",
@@ -924,10 +1199,10 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
             assert(cur >= opt[cur].mlen);
             if (opt[cur].mlen != 0) {
                 U32 const prev = cur - opt[cur].mlen;
-                repcodes_t newReps = ZSTD_updateRep(opt[prev].rep, opt[cur].off, opt[cur].litlen==0);
-                memcpy(opt[cur].rep, &newReps, sizeof(repcodes_t));
+                repcodes_t const newReps = ZSTD_newRep(opt[prev].rep, opt[cur].off, opt[cur].litlen==0);
+                ZSTD_memcpy(opt[cur].rep, &newReps, sizeof(repcodes_t));
             } else {
-                memcpy(opt[cur].rep, opt[cur - 1].rep, sizeof(repcodes_t));
+                ZSTD_memcpy(opt[cur].rep, opt[cur - 1].rep, sizeof(repcodes_t));
             }
 
             /* last match must start at a minimum distance of 8 from oend */
@@ -941,12 +1216,17 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
                 continue;  /* skip unpromising positions; about ~+6% speed, -0.01 ratio */
             }
 
+            assert(opt[cur].price >= 0);
             {   U32 const ll0 = (opt[cur].mlen != 0);
                 U32 const litlen = (opt[cur].mlen == 0) ? opt[cur].litlen : 0;
-                U32 const previousPrice = opt[cur].price;
+                U32 const previousPrice = (U32)opt[cur].price;
                 U32 const basePrice = previousPrice + ZSTD_litLengthPrice(0, optStatePtr, optLevel);
-                U32 const nbMatches = ZSTD_BtGetAllMatches(matches, ms, &nextToUpdate3, inr, iend, dictMode, opt[cur].rep, ll0, minMatch);
+                U32 nbMatches = getAllMatches(matches, ms, &nextToUpdate3, inr, iend, opt[cur].rep, ll0, minMatch);
                 U32 matchNb;
+
+                ZSTD_optLdm_processMatchCandidate(&optLdm, matches, &nbMatches,
+                                                  (U32)(inr-istart), (U32)(iend-inr));
+
                 if (!nbMatches) {
                     DEBUGLOG(7, "rPos:%u : no match found", cur);
                     continue;
@@ -974,12 +1254,12 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
                     U32 const startML = (matchNb>0) ? matches[matchNb-1].len+1 : minMatch;
                     U32 mlen;
 
-                    DEBUGLOG(7, "testing match %u => offCode=%4u, mlen=%2u, llen=%2u",
+                    DEBUGLOG(7, "testing match %u => offBase=%4u, mlen=%2u, llen=%2u",
                                 matchNb, matches[matchNb].off, lastML, litlen);
 
                     for (mlen = lastML; mlen >= startML; mlen--) {  /* scan downward */
                         U32 const pos = cur + mlen;
-                        int const price = basePrice + ZSTD_getMatchPrice(offset, mlen, optStatePtr, optLevel);
+                        int const price = (int)basePrice + (int)ZSTD_getMatchPrice(offset, mlen, optStatePtr, optLevel);
 
                         if ((pos > last_pos) || (price < opt[pos].price)) {
                             DEBUGLOG(7, "rPos:%u (ml=%2u) => new better price (%.2f<%.2f)",
@@ -1009,10 +1289,10 @@ _shortestPath:   /* cur, last_pos, best_mlen, best_off have to be set */
          * update them while traversing the sequences.
          */
         if (lastSequence.mlen != 0) {
-            repcodes_t reps = ZSTD_updateRep(opt[cur].rep, lastSequence.off, lastSequence.litlen==0);
-            memcpy(rep, &reps, sizeof(reps));
+            repcodes_t const reps = ZSTD_newRep(opt[cur].rep, lastSequence.off, lastSequence.litlen==0);
+            ZSTD_memcpy(rep, &reps, sizeof(reps));
         } else {
-            memcpy(rep, opt[cur].rep, sizeof(repcodes_t));
+            ZSTD_memcpy(rep, opt[cur].rep, sizeof(repcodes_t));
         }
 
         {   U32 const storeEnd = cur + 1;
@@ -1040,7 +1320,7 @@ _shortestPath:   /* cur, last_pos, best_mlen, best_off have to be set */
                 for (storePos=storeStart; storePos <= storeEnd; storePos++) {
                     U32 const llen = opt[storePos].litlen;
                     U32 const mlen = opt[storePos].mlen;
-                    U32 const offCode = opt[storePos].off;
+                    U32 const offBase = opt[storePos].off;
                     U32 const advance = llen + mlen;
                     DEBUGLOG(6, "considering seq starting at %zi, llen=%u, mlen=%u",
                                 anchor - istart, (unsigned)llen, (unsigned)mlen);
@@ -1052,8 +1332,8 @@ _shortestPath:   /* cur, last_pos, best_mlen, best_off have to be set */
                     }
 
                     assert(anchor + llen <= iend);
-                    ZSTD_updateStats(optStatePtr, llen, anchor, offCode, mlen);
-                    ZSTD_storeSeq(seqStore, llen, anchor, iend, offCode, mlen-MINMATCH);
+                    ZSTD_updateStats(optStatePtr, llen, anchor, offBase, mlen);
+                    ZSTD_storeSeq(seqStore, llen, anchor, iend, offBase, mlen);
                     anchor += advance;
                     ip = anchor;
             }   }
@@ -1065,43 +1345,35 @@ _shortestPath:   /* cur, last_pos, best_mlen, best_off have to be set */
     return (size_t)(iend - anchor);
 }
 
+static size_t ZSTD_compressBlock_opt0(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        const void* src, size_t srcSize, const ZSTD_dictMode_e dictMode)
+{
+    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /* optLevel */, dictMode);
+}
+
+static size_t ZSTD_compressBlock_opt2(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        const void* src, size_t srcSize, const ZSTD_dictMode_e dictMode)
+{
+    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /* optLevel */, dictMode);
+}
 
 size_t ZSTD_compressBlock_btopt(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         const void* src, size_t srcSize)
 {
     DEBUGLOG(5, "ZSTD_compressBlock_btopt");
-    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /*optLevel*/, ZSTD_noDict);
+    return ZSTD_compressBlock_opt0(ms, seqStore, rep, src, srcSize, ZSTD_noDict);
 }
 
 
-/* used in 2-pass strategy */
-static U32 ZSTD_upscaleStat(unsigned* table, U32 lastEltIndex, int bonus)
-{
-    U32 s, sum=0;
-    assert(ZSTD_FREQ_DIV+bonus >= 0);
-    for (s=0; s<lastEltIndex+1; s++) {
-        table[s] <<= ZSTD_FREQ_DIV+bonus;
-        table[s]--;
-        sum += table[s];
-    }
-    return sum;
-}
 
-/* used in 2-pass strategy */
-MEM_STATIC void ZSTD_upscaleStats(optState_t* optPtr)
-{
-    if (ZSTD_compressedLiterals(optPtr))
-        optPtr->litSum = ZSTD_upscaleStat(optPtr->litFreq, MaxLit, 0);
-    optPtr->litLengthSum = ZSTD_upscaleStat(optPtr->litLengthFreq, MaxLL, 0);
-    optPtr->matchLengthSum = ZSTD_upscaleStat(optPtr->matchLengthFreq, MaxML, 0);
-    optPtr->offCodeSum = ZSTD_upscaleStat(optPtr->offCodeFreq, MaxOff, 0);
-}
 
 /* ZSTD_initStats_ultra():
  * make a first compression pass, just to seed stats with more accurate starting values.
  * only works on first block, with no dictionary and no ldm.
- * this function cannot error, hence its contract must be respected.
+ * this function cannot error out, its narrow contract must be respected.
  */
 static void
 ZSTD_initStats_ultra(ZSTD_matchState_t* ms,
@@ -1110,7 +1382,7 @@ ZSTD_initStats_ultra(ZSTD_matchState_t* ms,
                const void* src, size_t srcSize)
 {
     U32 tmpRep[ZSTD_REP_NUM];  /* updated rep codes will sink here */
-    memcpy(tmpRep, rep, sizeof(tmpRep));
+    ZSTD_memcpy(tmpRep, rep, sizeof(tmpRep));
 
     DEBUGLOG(4, "ZSTD_initStats_ultra (srcSize=%zu)", srcSize);
     assert(ms->opt.litLengthSum == 0);    /* first block */
@@ -1118,17 +1390,15 @@ ZSTD_initStats_ultra(ZSTD_matchState_t* ms,
     assert(ms->window.dictLimit == ms->window.lowLimit);   /* no dictionary */
     assert(ms->window.dictLimit - ms->nextToUpdate <= 1);  /* no prefix (note: intentional overflow, defined as 2-complement) */
 
-    ZSTD_compressBlock_opt_generic(ms, seqStore, tmpRep, src, srcSize, 2 /*optLevel*/, ZSTD_noDict);   /* generate stats into ms->opt*/
+    ZSTD_compressBlock_opt2(ms, seqStore, tmpRep, src, srcSize, ZSTD_noDict);   /* generate stats into ms->opt*/
 
-    /* invalidate first scan from history */
+    /* invalidate first scan from history, only keep entropy stats */
     ZSTD_resetSeqStore(seqStore);
     ms->window.base -= srcSize;
     ms->window.dictLimit += (U32)srcSize;
     ms->window.lowLimit = ms->window.dictLimit;
     ms->nextToUpdate = ms->window.dictLimit;
 
-    /* re-inforce weight of collected statistics */
-    ZSTD_upscaleStats(&ms->opt);
 }
 
 size_t ZSTD_compressBlock_btultra(
@@ -1136,63 +1406,63 @@ size_t ZSTD_compressBlock_btultra(
         const void* src, size_t srcSize)
 {
     DEBUGLOG(5, "ZSTD_compressBlock_btultra (srcSize=%zu)", srcSize);
-    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_noDict);
+    return ZSTD_compressBlock_opt2(ms, seqStore, rep, src, srcSize, ZSTD_noDict);
 }
 
 size_t ZSTD_compressBlock_btultra2(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         const void* src, size_t srcSize)
 {
-    U32 const current = (U32)((const BYTE*)src - ms->window.base);
+    U32 const curr = (U32)((const BYTE*)src - ms->window.base);
     DEBUGLOG(5, "ZSTD_compressBlock_btultra2 (srcSize=%zu)", srcSize);
 
-    /* 2-pass strategy:
+    /* 2-passes strategy:
      * this strategy makes a first pass over first block to collect statistics
-     * and seed next round's statistics with it.
-     * After 1st pass, function forgets everything, and starts a new block.
+     * in order to seed next round's statistics with it.
+     * After 1st pass, function forgets history, and starts a new block.
      * Consequently, this can only work if no data has been previously loaded in tables,
      * aka, no dictionary, no prefix, no ldm preprocessing.
      * The compression ratio gain is generally small (~0.5% on first block),
-     * the cost is 2x cpu time on first block. */
+    ** the cost is 2x cpu time on first block. */
     assert(srcSize <= ZSTD_BLOCKSIZE_MAX);
     if ( (ms->opt.litLengthSum==0)   /* first block */
       && (seqStore->sequences == seqStore->sequencesStart)  /* no ldm */
       && (ms->window.dictLimit == ms->window.lowLimit)   /* no dictionary */
-      && (current == ms->window.dictLimit)   /* start of frame, nothing already loaded nor skipped */
-      && (srcSize > ZSTD_PREDEF_THRESHOLD)
+      && (curr == ms->window.dictLimit)    /* start of frame, nothing already loaded nor skipped */
+      && (srcSize > ZSTD_PREDEF_THRESHOLD) /* input large enough to not employ default stats */
       ) {
         ZSTD_initStats_ultra(ms, seqStore, rep, src, srcSize);
     }
 
-    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_noDict);
+    return ZSTD_compressBlock_opt2(ms, seqStore, rep, src, srcSize, ZSTD_noDict);
 }
 
 size_t ZSTD_compressBlock_btopt_dictMatchState(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         const void* src, size_t srcSize)
 {
-    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /*optLevel*/, ZSTD_dictMatchState);
+    return ZSTD_compressBlock_opt0(ms, seqStore, rep, src, srcSize, ZSTD_dictMatchState);
 }
 
 size_t ZSTD_compressBlock_btultra_dictMatchState(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         const void* src, size_t srcSize)
 {
-    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_dictMatchState);
+    return ZSTD_compressBlock_opt2(ms, seqStore, rep, src, srcSize, ZSTD_dictMatchState);
 }
 
 size_t ZSTD_compressBlock_btopt_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         const void* src, size_t srcSize)
 {
-    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /*optLevel*/, ZSTD_extDict);
+    return ZSTD_compressBlock_opt0(ms, seqStore, rep, src, srcSize, ZSTD_extDict);
 }
 
 size_t ZSTD_compressBlock_btultra_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         const void* src, size_t srcSize)
 {
-    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_extDict);
+    return ZSTD_compressBlock_opt2(ms, seqStore, rep, src, srcSize, ZSTD_extDict);
 }
 
 /* note : no btultra2 variant for extDict nor dictMatchState,
diff --git a/contrib/zstd/zstd_opt.h b/contrib/zstd/zstd_opt.h
index 9aba8a901..342e5a311 100644
--- a/contrib/zstd/zstd_opt.h
+++ b/contrib/zstd/zstd_opt.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/contrib/zstd/zstd_trace.h b/contrib/zstd/zstd_trace.h
new file mode 100644
index 000000000..da20534eb
--- /dev/null
+++ b/contrib/zstd/zstd_trace.h
@@ -0,0 +1,163 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_TRACE_H
+#define ZSTD_TRACE_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+#include <stddef.h>
+
+/* weak symbol support
+ * For now, enable conservatively:
+ * - Only GNUC
+ * - Only ELF
+ * - Only x86-64, i386 and aarch64
+ * Also, explicitly disable on platforms known not to work so they aren't
+ * forgotten in the future.
+ */
+#if !defined(ZSTD_HAVE_WEAK_SYMBOLS) && \
+    defined(__GNUC__) && defined(__ELF__) && \
+    (defined(__x86_64__) || defined(_M_X64) || defined(__i386__) || defined(_M_IX86) || defined(__aarch64__)) && \
+    !defined(__APPLE__) && !defined(_WIN32) && !defined(__MINGW32__) && \
+    !defined(__CYGWIN__) && !defined(_AIX)
+#  define ZSTD_HAVE_WEAK_SYMBOLS 1
+#else
+#  define ZSTD_HAVE_WEAK_SYMBOLS 0
+#endif
+#if ZSTD_HAVE_WEAK_SYMBOLS
+#  define ZSTD_WEAK_ATTR __attribute__((__weak__))
+#else
+#  define ZSTD_WEAK_ATTR
+#endif
+
+/* Only enable tracing when weak symbols are available. */
+#ifndef ZSTD_TRACE
+#  define ZSTD_TRACE ZSTD_HAVE_WEAK_SYMBOLS
+#endif
+
+#if ZSTD_TRACE
+
+struct ZSTD_CCtx_s;
+struct ZSTD_DCtx_s;
+struct ZSTD_CCtx_params_s;
+
+typedef struct {
+    /**
+     * ZSTD_VERSION_NUMBER
+     *
+     * This is guaranteed to be the first member of ZSTD_trace.
+     * Otherwise, this struct is not stable between versions. If
+     * the version number does not match your expectation, you
+     * should not interpret the rest of the struct.
+     */
+    unsigned version;
+    /**
+     * Non-zero if streaming (de)compression is used.
+     */
+    unsigned streaming;
+    /**
+     * The dictionary ID.
+     */
+    unsigned dictionaryID;
+    /**
+     * Is the dictionary cold?
+     * Only set on decompression.
+     */
+    unsigned dictionaryIsCold;
+    /**
+     * The dictionary size or zero if no dictionary.
+     */
+    size_t dictionarySize;
+    /**
+     * The uncompressed size of the data.
+     */
+    size_t uncompressedSize;
+    /**
+     * The compressed size of the data.
+     */
+    size_t compressedSize;
+    /**
+     * The fully resolved CCtx parameters (NULL on decompression).
+     */
+    struct ZSTD_CCtx_params_s const* params;
+    /**
+     * The ZSTD_CCtx pointer (NULL on decompression).
+     */
+    struct ZSTD_CCtx_s const* cctx;
+    /**
+     * The ZSTD_DCtx pointer (NULL on compression).
+     */
+    struct ZSTD_DCtx_s const* dctx;
+} ZSTD_Trace;
+
+/**
+ * A tracing context. It must be 0 when tracing is disabled.
+ * Otherwise, any non-zero value returned by a tracing begin()
+ * function is presented to any subsequent calls to end().
+ *
+ * Any non-zero value is treated as tracing is enabled and not
+ * interpreted by the library.
+ *
+ * Two possible uses are:
+ * * A timestamp for when the begin() function was called.
+ * * A unique key identifying the (de)compression, like the
+ *   address of the [dc]ctx pointer if you need to track
+ *   more information than just a timestamp.
+ */
+typedef unsigned long long ZSTD_TraceCtx;
+
+/**
+ * Trace the beginning of a compression call.
+ * @param cctx The dctx pointer for the compression.
+ *             It can be used as a key to map begin() to end().
+ * @returns Non-zero if tracing is enabled. The return value is
+ *          passed to ZSTD_trace_compress_end().
+ */
+ZSTD_WEAK_ATTR ZSTD_TraceCtx ZSTD_trace_compress_begin(
+    struct ZSTD_CCtx_s const* cctx);
+
+/**
+ * Trace the end of a compression call.
+ * @param ctx The return value of ZSTD_trace_compress_begin().
+ * @param trace The zstd tracing info.
+ */
+ZSTD_WEAK_ATTR void ZSTD_trace_compress_end(
+    ZSTD_TraceCtx ctx,
+    ZSTD_Trace const* trace);
+
+/**
+ * Trace the beginning of a decompression call.
+ * @param dctx The dctx pointer for the decompression.
+ *             It can be used as a key to map begin() to end().
+ * @returns Non-zero if tracing is enabled. The return value is
+ *          passed to ZSTD_trace_compress_end().
+ */
+ZSTD_WEAK_ATTR ZSTD_TraceCtx ZSTD_trace_decompress_begin(
+    struct ZSTD_DCtx_s const* dctx);
+
+/**
+ * Trace the end of a decompression call.
+ * @param ctx The return value of ZSTD_trace_decompress_begin().
+ * @param trace The zstd tracing info.
+ */
+ZSTD_WEAK_ATTR void ZSTD_trace_decompress_end(
+    ZSTD_TraceCtx ctx,
+    ZSTD_Trace const* trace);
+
+#endif /* ZSTD_TRACE */
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* ZSTD_TRACE_H */