diff options
Diffstat (limited to 'vendor/github.com/golang')
| -rw-r--r-- | vendor/github.com/golang/snappy/.gitignore | 16 | ||||
| -rw-r--r-- | vendor/github.com/golang/snappy/README | 107 | ||||
| -rw-r--r-- | vendor/github.com/golang/snappy/decode.go | 24 | ||||
| -rw-r--r-- | vendor/github.com/golang/snappy/decode_amd64.go | 4 | ||||
| -rw-r--r-- | vendor/github.com/golang/snappy/decode_amd64.s | 38 | ||||
| -rw-r--r-- | vendor/github.com/golang/snappy/decode_other.go | 13 | ||||
| -rw-r--r-- | vendor/github.com/golang/snappy/encode.go | 186 | ||||
| -rw-r--r-- | vendor/github.com/golang/snappy/encode_amd64.go | 29 | ||||
| -rw-r--r-- | vendor/github.com/golang/snappy/encode_amd64.s | 730 | ||||
| -rw-r--r-- | vendor/github.com/golang/snappy/encode_other.go | 238 | ||||
| -rw-r--r-- | vendor/github.com/golang/snappy/snappy.go | 22 |
11 files changed, 1225 insertions, 182 deletions
diff --git a/vendor/github.com/golang/snappy/.gitignore b/vendor/github.com/golang/snappy/.gitignore new file mode 100644 index 0000000000..042091d9b3 --- /dev/null +++ b/vendor/github.com/golang/snappy/.gitignore @@ -0,0 +1,16 @@ +cmd/snappytool/snappytool +testdata/bench + +# These explicitly listed benchmark data files are for an obsolete version of +# snappy_test.go. +testdata/alice29.txt +testdata/asyoulik.txt +testdata/fireworks.jpeg +testdata/geo.protodata +testdata/html +testdata/html_x_4 +testdata/kppkn.gtb +testdata/lcet10.txt +testdata/paper-100k.pdf +testdata/plrabn12.txt +testdata/urls.10K diff --git a/vendor/github.com/golang/snappy/README b/vendor/github.com/golang/snappy/README new file mode 100644 index 0000000000..cea12879a0 --- /dev/null +++ b/vendor/github.com/golang/snappy/README @@ -0,0 +1,107 @@ +The Snappy compression format in the Go programming language. + +To download and install from source: +$ go get github.com/golang/snappy + +Unless otherwise noted, the Snappy-Go source files are distributed +under the BSD-style license found in the LICENSE file. + + + +Benchmarks. + +The golang/snappy benchmarks include compressing (Z) and decompressing (U) ten +or so files, the same set used by the C++ Snappy code (github.com/google/snappy +and note the "google", not "golang"). On an "Intel(R) Core(TM) i7-3770 CPU @ +3.40GHz", Go's GOARCH=amd64 numbers as of 2016-05-29: + +"go test -test.bench=." + +_UFlat0-8 2.19GB/s ± 0% html +_UFlat1-8 1.41GB/s ± 0% urls +_UFlat2-8 23.5GB/s ± 2% jpg +_UFlat3-8 1.91GB/s ± 0% jpg_200 +_UFlat4-8 14.0GB/s ± 1% pdf +_UFlat5-8 1.97GB/s ± 0% html4 +_UFlat6-8 814MB/s ± 0% txt1 +_UFlat7-8 785MB/s ± 0% txt2 +_UFlat8-8 857MB/s ± 0% txt3 +_UFlat9-8 719MB/s ± 1% txt4 +_UFlat10-8 2.84GB/s ± 0% pb +_UFlat11-8 1.05GB/s ± 0% gaviota + +_ZFlat0-8 1.04GB/s ± 0% html +_ZFlat1-8 534MB/s ± 0% urls +_ZFlat2-8 15.7GB/s ± 1% jpg +_ZFlat3-8 740MB/s ± 3% jpg_200 +_ZFlat4-8 9.20GB/s ± 1% pdf +_ZFlat5-8 991MB/s ± 0% html4 +_ZFlat6-8 379MB/s ± 0% txt1 +_ZFlat7-8 352MB/s ± 0% txt2 +_ZFlat8-8 396MB/s ± 1% txt3 +_ZFlat9-8 327MB/s ± 1% txt4 +_ZFlat10-8 1.33GB/s ± 1% pb +_ZFlat11-8 605MB/s ± 1% gaviota + + + +"go test -test.bench=. -tags=noasm" + +_UFlat0-8 621MB/s ± 2% html +_UFlat1-8 494MB/s ± 1% urls +_UFlat2-8 23.2GB/s ± 1% jpg +_UFlat3-8 1.12GB/s ± 1% jpg_200 +_UFlat4-8 4.35GB/s ± 1% pdf +_UFlat5-8 609MB/s ± 0% html4 +_UFlat6-8 296MB/s ± 0% txt1 +_UFlat7-8 288MB/s ± 0% txt2 +_UFlat8-8 309MB/s ± 1% txt3 +_UFlat9-8 280MB/s ± 1% txt4 +_UFlat10-8 753MB/s ± 0% pb +_UFlat11-8 400MB/s ± 0% gaviota + +_ZFlat0-8 409MB/s ± 1% html +_ZFlat1-8 250MB/s ± 1% urls +_ZFlat2-8 12.3GB/s ± 1% jpg +_ZFlat3-8 132MB/s ± 0% jpg_200 +_ZFlat4-8 2.92GB/s ± 0% pdf +_ZFlat5-8 405MB/s ± 1% html4 +_ZFlat6-8 179MB/s ± 1% txt1 +_ZFlat7-8 170MB/s ± 1% txt2 +_ZFlat8-8 189MB/s ± 1% txt3 +_ZFlat9-8 164MB/s ± 1% txt4 +_ZFlat10-8 479MB/s ± 1% pb +_ZFlat11-8 270MB/s ± 1% gaviota + + + +For comparison (Go's encoded output is byte-for-byte identical to C++'s), here +are the numbers from C++ Snappy's + +make CXXFLAGS="-O2 -DNDEBUG -g" clean snappy_unittest.log && cat snappy_unittest.log + +BM_UFlat/0 2.4GB/s html +BM_UFlat/1 1.4GB/s urls +BM_UFlat/2 21.8GB/s jpg +BM_UFlat/3 1.5GB/s jpg_200 +BM_UFlat/4 13.3GB/s pdf +BM_UFlat/5 2.1GB/s html4 +BM_UFlat/6 1.0GB/s txt1 +BM_UFlat/7 959.4MB/s txt2 +BM_UFlat/8 1.0GB/s txt3 +BM_UFlat/9 864.5MB/s txt4 +BM_UFlat/10 2.9GB/s pb +BM_UFlat/11 1.2GB/s gaviota + +BM_ZFlat/0 944.3MB/s html (22.31 %) +BM_ZFlat/1 501.6MB/s urls (47.78 %) +BM_ZFlat/2 14.3GB/s jpg (99.95 %) +BM_ZFlat/3 538.3MB/s jpg_200 (73.00 %) +BM_ZFlat/4 8.3GB/s pdf (83.30 %) +BM_ZFlat/5 903.5MB/s html4 (22.52 %) +BM_ZFlat/6 336.0MB/s txt1 (57.88 %) +BM_ZFlat/7 312.3MB/s txt2 (61.91 %) +BM_ZFlat/8 353.1MB/s txt3 (54.99 %) +BM_ZFlat/9 289.9MB/s txt4 (66.26 %) +BM_ZFlat/10 1.2GB/s pb (19.68 %) +BM_ZFlat/11 527.4MB/s gaviota (37.72 %) diff --git a/vendor/github.com/golang/snappy/decode.go b/vendor/github.com/golang/snappy/decode.go index 7be590cee7..72efb0353d 100644 --- a/vendor/github.com/golang/snappy/decode.go +++ b/vendor/github.com/golang/snappy/decode.go @@ -18,7 +18,6 @@ var ( // ErrUnsupported reports that the input isn't supported. ErrUnsupported = errors.New("snappy: unsupported input") - errUnsupportedCopy4Tag = errors.New("snappy: unsupported COPY_4 tag") errUnsupportedLiteralLength = errors.New("snappy: unsupported literal length") ) @@ -46,7 +45,6 @@ func decodedLen(src []byte) (blockLen, headerLen int, err error) { const ( decodeErrCodeCorrupt = 1 decodeErrCodeUnsupportedLiteralLength = 2 - decodeErrCodeUnsupportedCopy4Tag = 3 ) // Decode returns the decoded form of src. The returned slice may be a sub- @@ -69,8 +67,6 @@ func Decode(dst, src []byte) ([]byte, error) { return dst, nil case decodeErrCodeUnsupportedLiteralLength: return nil, errUnsupportedLiteralLength - case decodeErrCodeUnsupportedCopy4Tag: - return nil, errUnsupportedCopy4Tag } return nil, ErrCorrupt } @@ -108,9 +104,9 @@ func (r *Reader) Reset(reader io.Reader) { r.readHeader = false } -func (r *Reader) readFull(p []byte) (ok bool) { +func (r *Reader) readFull(p []byte, allowEOF bool) (ok bool) { if _, r.err = io.ReadFull(r.r, p); r.err != nil { - if r.err == io.ErrUnexpectedEOF { + if r.err == io.ErrUnexpectedEOF || (r.err == io.EOF && !allowEOF) { r.err = ErrCorrupt } return false @@ -129,7 +125,7 @@ func (r *Reader) Read(p []byte) (int, error) { r.i += n return n, nil } - if !r.readFull(r.buf[:4]) { + if !r.readFull(r.buf[:4], true) { return 0, r.err } chunkType := r.buf[0] @@ -156,7 +152,7 @@ func (r *Reader) Read(p []byte) (int, error) { return 0, r.err } buf := r.buf[:chunkLen] - if !r.readFull(buf) { + if !r.readFull(buf, false) { return 0, r.err } checksum := uint32(buf[0]) | uint32(buf[1])<<8 | uint32(buf[2])<<16 | uint32(buf[3])<<24 @@ -189,13 +185,17 @@ func (r *Reader) Read(p []byte) (int, error) { return 0, r.err } buf := r.buf[:checksumSize] - if !r.readFull(buf) { + if !r.readFull(buf, false) { return 0, r.err } checksum := uint32(buf[0]) | uint32(buf[1])<<8 | uint32(buf[2])<<16 | uint32(buf[3])<<24 // Read directly into r.decoded instead of via r.buf. n := chunkLen - checksumSize - if !r.readFull(r.decoded[:n]) { + if n > len(r.decoded) { + r.err = ErrCorrupt + return 0, r.err + } + if !r.readFull(r.decoded[:n], false) { return 0, r.err } if crc(r.decoded[:n]) != checksum { @@ -211,7 +211,7 @@ func (r *Reader) Read(p []byte) (int, error) { r.err = ErrCorrupt return 0, r.err } - if !r.readFull(r.buf[:len(magicBody)]) { + if !r.readFull(r.buf[:len(magicBody)], false) { return 0, r.err } for i := 0; i < len(magicBody); i++ { @@ -230,7 +230,7 @@ func (r *Reader) Read(p []byte) (int, error) { } // Section 4.4 Padding (chunk type 0xfe). // Section 4.6. Reserved skippable chunks (chunk types 0x80-0xfd). - if !r.readFull(r.buf[:chunkLen]) { + if !r.readFull(r.buf[:chunkLen], false) { return 0, r.err } } diff --git a/vendor/github.com/golang/snappy/decode_amd64.go b/vendor/github.com/golang/snappy/decode_amd64.go index 32bce47093..fcd192b849 100644 --- a/vendor/github.com/golang/snappy/decode_amd64.go +++ b/vendor/github.com/golang/snappy/decode_amd64.go @@ -2,6 +2,10 @@ // Use of this source code is governed by a BSD-style // license that can be found in the LICENSE file. +// +build !appengine +// +build gc +// +build !noasm + package snappy // decode has the same semantics as in decode_other.go. diff --git a/vendor/github.com/golang/snappy/decode_amd64.s b/vendor/github.com/golang/snappy/decode_amd64.s index c33f5bf97b..e6179f65e3 100644 --- a/vendor/github.com/golang/snappy/decode_amd64.s +++ b/vendor/github.com/golang/snappy/decode_amd64.s @@ -2,12 +2,16 @@ // Use of this source code is governed by a BSD-style // license that can be found in the LICENSE file. +// +build !appengine +// +build gc +// +build !noasm + #include "textflag.h" -// func decode(dst, src []byte) int -// // The asm code generally follows the pure Go code in decode_other.go, except // where marked with a "!!!". + +// func decode(dst, src []byte) int // // All local variables fit into registers. The non-zero stack size is only to // spill registers and push args when issuing a CALL. The register allocation: @@ -222,6 +226,25 @@ tagLit63: // ---------------------------------------- // The code below handles copy tags. +tagCopy4: + // case tagCopy4: + // s += 5 + ADDQ $5, SI + + // if uint(s) > uint(len(src)) { etc } + MOVQ SI, BX + SUBQ R11, BX + CMPQ BX, R12 + JA errCorrupt + + // length = 1 + int(src[s-5])>>2 + SHRQ $2, CX + INCQ CX + + // offset = int(uint32(src[s-4]) | uint32(src[s-3])<<8 | uint32(src[s-2])<<16 | uint32(src[s-1])<<24) + MOVLQZX -4(SI), DX + JMP doCopy + tagCopy2: // case tagCopy2: // s += 3 @@ -237,7 +260,7 @@ tagCopy2: SHRQ $2, CX INCQ CX - // offset = int(src[s-2]) | int(src[s-1])<<8 + // offset = int(uint32(src[s-2]) | uint32(src[s-1])<<8) MOVWQZX -2(SI), DX JMP doCopy @@ -247,7 +270,7 @@ tagCopy: // - CX == src[s] CMPQ BX, $2 JEQ tagCopy2 - JA errUC4T + JA tagCopy4 // case tagCopy1: // s += 2 @@ -259,7 +282,7 @@ tagCopy: CMPQ BX, R12 JA errCorrupt - // offset = int(src[s-2])&0xe0<<3 | int(src[s-1]) + // offset = int(uint32(src[s-2])&0xe0<<3 | uint32(src[s-1])) MOVQ CX, DX ANDQ $0xe0, DX SHLQ $3, DX @@ -465,8 +488,3 @@ errCorrupt: // return decodeErrCodeCorrupt MOVQ $1, ret+48(FP) RET - -errUC4T: - // return decodeErrCodeUnsupportedCopy4Tag - MOVQ $3, ret+48(FP) - RET diff --git a/vendor/github.com/golang/snappy/decode_other.go b/vendor/github.com/golang/snappy/decode_other.go index 1a8114ab1f..8c9f2049bc 100644 --- a/vendor/github.com/golang/snappy/decode_other.go +++ b/vendor/github.com/golang/snappy/decode_other.go @@ -2,7 +2,7 @@ // Use of this source code is governed by a BSD-style // license that can be found in the LICENSE file. -// +build !amd64 +// +build !amd64 appengine !gc noasm package snappy @@ -63,7 +63,7 @@ func decode(dst, src []byte) int { return decodeErrCodeCorrupt } length = 4 + int(src[s-2])>>2&0x7 - offset = int(src[s-2])&0xe0<<3 | int(src[s-1]) + offset = int(uint32(src[s-2])&0xe0<<3 | uint32(src[s-1])) case tagCopy2: s += 3 @@ -71,10 +71,15 @@ func decode(dst, src []byte) int { return decodeErrCodeCorrupt } length = 1 + int(src[s-3])>>2 - offset = int(src[s-2]) | int(src[s-1])<<8 + offset = int(uint32(src[s-2]) | uint32(src[s-1])<<8) case tagCopy4: - return decodeErrCodeUnsupportedCopy4Tag + s += 5 + if uint(s) > uint(len(src)) { // The uint conversions catch overflow from the previous line. + return decodeErrCodeCorrupt + } + length = 1 + int(src[s-5])>>2 + offset = int(uint32(src[s-4]) | uint32(src[s-3])<<8 | uint32(src[s-2])<<16 | uint32(src[s-1])<<24) } if offset <= 0 || d < offset || length > len(dst)-d { diff --git a/vendor/github.com/golang/snappy/encode.go b/vendor/github.com/golang/snappy/encode.go index 38ebe952e0..8d393e904b 100644 --- a/vendor/github.com/golang/snappy/encode.go +++ b/vendor/github.com/golang/snappy/encode.go @@ -10,78 +10,11 @@ import ( "io" ) -// maxOffset limits how far copy back-references can go, the same as the C++ -// code. -const maxOffset = 1 << 15 - -// emitLiteral writes a literal chunk and returns the number of bytes written. -func emitLiteral(dst, lit []byte) int { - i, n := 0, uint(len(lit)-1) - switch { - case n < 60: - dst[0] = uint8(n)<<2 | tagLiteral - i = 1 - case n < 1<<8: - dst[0] = 60<<2 | tagLiteral - dst[1] = uint8(n) - i = 2 - case n < 1<<16: - dst[0] = 61<<2 | tagLiteral - dst[1] = uint8(n) - dst[2] = uint8(n >> 8) - i = 3 - case n < 1<<24: - dst[0] = 62<<2 | tagLiteral - dst[1] = uint8(n) - dst[2] = uint8(n >> 8) - dst[3] = uint8(n >> 16) - i = 4 - case int64(n) < 1<<32: - dst[0] = 63<<2 | tagLiteral - dst[1] = uint8(n) - dst[2] = uint8(n >> 8) - dst[3] = uint8(n >> 16) - dst[4] = uint8(n >> 24) - i = 5 - default: - panic("snappy: source buffer is too long") - } - if copy(dst[i:], lit) != len(lit) { - panic("snappy: destination buffer is too short") - } - return i + len(lit) -} - -// emitCopy writes a copy chunk and returns the number of bytes written. -func emitCopy(dst []byte, offset, length int32) int { - i := 0 - for length > 0 { - x := length - 4 - if 0 <= x && x < 1<<3 && offset < 1<<11 { - dst[i+0] = uint8(offset>>8)&0x07<<5 | uint8(x)<<2 | tagCopy1 - dst[i+1] = uint8(offset) - i += 2 - break - } - - x = length - if x > 1<<6 { - x = 1 << 6 - } - dst[i+0] = uint8(x-1)<<2 | tagCopy2 - dst[i+1] = uint8(offset) - dst[i+2] = uint8(offset >> 8) - i += 3 - length -= x - } - return i -} - // Encode returns the encoded form of src. The returned slice may be a sub- // slice of dst if dst was large enough to hold the entire encoded block. // Otherwise, a newly allocated slice will be returned. // -// It is valid to pass a nil dst. +// The dst and src must not overlap. It is valid to pass a nil dst. func Encode(dst, src []byte) []byte { if n := MaxEncodedLen(len(src)); n < 0 { panic(ErrTooLarge) @@ -98,94 +31,43 @@ func Encode(dst, src []byte) []byte { if len(p) > maxBlockSize { p, src = p[:maxBlockSize], p[maxBlockSize:] } - d += encodeBlock(dst[d:], p) + if len(p) < minNonLiteralBlockSize { + d += emitLiteral(dst[d:], p) + } else { + d += encodeBlock(dst[d:], p) + } } return dst[:d] } -// encodeBlock encodes a non-empty src to a guaranteed-large-enough dst. It -// assumes that the varint-encoded length of the decompressed bytes has already -// been written. +// inputMargin is the minimum number of extra input bytes to keep, inside +// encodeBlock's inner loop. On some architectures, this margin lets us +// implement a fast path for emitLiteral, where the copy of short (<= 16 byte) +// literals can be implemented as a single load to and store from a 16-byte +// register. That literal's actual length can be as short as 1 byte, so this +// can copy up to 15 bytes too much, but that's OK as subsequent iterations of +// the encoding loop will fix up the copy overrun, and this inputMargin ensures +// that we don't overrun the dst and src buffers. +const inputMargin = 16 - 1 + +// minNonLiteralBlockSize is the minimum size of the input to encodeBlock that +// could be encoded with a copy tag. This is the minimum with respect to the +// algorithm used by encodeBlock, not a minimum enforced by the file format. // -// It also assumes that: -// len(dst) >= MaxEncodedLen(len(src)) && -// 0 < len(src) && len(src) <= maxBlockSize -func encodeBlock(dst, src []byte) (d int) { - // Return early if src is short. - if len(src) <= 4 { - return emitLiteral(dst, src) - } - - // Initialize the hash table. Its size ranges from 1<<8 to 1<<14 inclusive. - const maxTableSize = 1 << 14 - shift, tableSize := uint(32-8), 1<<8 - for tableSize < maxTableSize && tableSize < len(src) { - shift-- - tableSize *= 2 - } - var table [maxTableSize]int32 - - // Iterate over the source bytes. - var ( - s int32 // The iterator position. - t int32 // The last position with the same hash as s. - lit int32 // The start position of any pending literal bytes. - - // Copied from the C++ snappy implementation: - // - // Heuristic match skipping: If 32 bytes are scanned with no matches - // found, start looking only at every other byte. If 32 more bytes are - // scanned, look at every third byte, etc.. When a match is found, - // immediately go back to looking at every byte. This is a small loss - // (~5% performance, ~0.1% density) for compressible data due to more - // bookkeeping, but for non-compressible data (such as JPEG) it's a - // huge win since the compressor quickly "realizes" the data is - // incompressible and doesn't bother looking for matches everywhere. - // - // The "skip" variable keeps track of how many bytes there are since - // the last match; dividing it by 32 (ie. right-shifting by five) gives - // the number of bytes to move ahead for each iteration. - skip uint32 = 32 - ) - for uint32(s+3) < uint32(len(src)) { // The uint32 conversions catch overflow from the +3. - // Update the hash table. - b0, b1, b2, b3 := src[s], src[s+1], src[s+2], src[s+3] - h := uint32(b0) | uint32(b1)<<8 | uint32(b2)<<16 | uint32(b3)<<24 - p := &table[(h*0x1e35a7bd)>>shift] - // We need to to store values in [-1, inf) in table. To save - // some initialization time, (re)use the table's zero value - // and shift the values against this zero: add 1 on writes, - // subtract 1 on reads. - t, *p = *p-1, s+1 - // If t is invalid or src[s:s+4] differs from src[t:t+4], accumulate a literal byte. - if t < 0 || s-t >= maxOffset || b0 != src[t] || b1 != src[t+1] || b2 != src[t+2] || b3 != src[t+3] { - s += int32(skip >> 5) - skip++ - continue - } - skip = 32 - // Otherwise, we have a match. First, emit any pending literal bytes. - if lit != s { - d += emitLiteral(dst[d:], src[lit:s]) - } - // Extend the match to be as long as possible. - s0 := s - s, t = s+4, t+4 - for int(s) < len(src) && src[s] == src[t] { - s++ - t++ - } - // Emit the copied bytes. - d += emitCopy(dst[d:], s-t, s-s0) - lit = s - } - - // Emit any final pending literal bytes and return. - if int(lit) != len(src) { - d += emitLiteral(dst[d:], src[lit:]) - } - return d -} +// The encoded output must start with at least a 1 byte literal, as there are +// no previous bytes to copy. A minimal (1 byte) copy after that, generated +// from an emitCopy call in encodeBlock's main loop, would require at least +// another inputMargin bytes, for the reason above: we want any emitLiteral +// calls inside encodeBlock's main loop to use the fast path if possible, which +// requires being able to overrun by inputMargin bytes. Thus, +// minNonLiteralBlockSize equals 1 + 1 + inputMargin. +// +// The C++ code doesn't use this exact threshold, but it could, as discussed at +// https://groups.google.com/d/topic/snappy-compression/oGbhsdIJSJ8/discussion +// The difference between Go (2+inputMargin) and C++ (inputMargin) is purely an +// optimization. It should not affect the encoded form. This is tested by +// TestSameEncodingAsCppShortCopies. +const minNonLiteralBlockSize = 1 + 1 + inputMargin // MaxEncodedLen returns the maximum length of a snappy block, given its // uncompressed length. @@ -256,7 +138,7 @@ func NewBufferedWriter(w io.Writer) *Writer { } } -// Writer is an io.Writer than can write Snappy-compressed bytes. +// Writer is an io.Writer that can write Snappy-compressed bytes. type Writer struct { w io.Writer err error diff --git a/vendor/github.com/golang/snappy/encode_amd64.go b/vendor/github.com/golang/snappy/encode_amd64.go new file mode 100644 index 0000000000..150d91bc8b --- /dev/null +++ b/vendor/github.com/golang/snappy/encode_amd64.go @@ -0,0 +1,29 @@ +// Copyright 2016 The Snappy-Go Authors. All rights reserved. +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +// +build !appengine +// +build gc +// +build !noasm + +package snappy + +// emitLiteral has the same semantics as in encode_other.go. +// +//go:noescape +func emitLiteral(dst, lit []byte) int + +// emitCopy has the same semantics as in encode_other.go. +// +//go:noescape +func emitCopy(dst []byte, offset, length int) int + +// extendMatch has the same semantics as in encode_other.go. +// +//go:noescape +func extendMatch(src []byte, i, j int) int + +// encodeBlock has the same semantics as in encode_other.go. +// +//go:noescape +func encodeBlock(dst, src []byte) (d int) diff --git a/vendor/github.com/golang/snappy/encode_amd64.s b/vendor/github.com/golang/snappy/encode_amd64.s new file mode 100644 index 0000000000..adfd979fe2 --- /dev/null +++ b/vendor/github.com/golang/snappy/encode_amd64.s @@ -0,0 +1,730 @@ +// Copyright 2016 The Go Authors. All rights reserved. +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +// +build !appengine +// +build gc +// +build !noasm + +#include "textflag.h" + +// The XXX lines assemble on Go 1.4, 1.5 and 1.7, but not 1.6, due to a +// Go toolchain regression. See https://github.com/golang/go/issues/15426 and +// https://github.com/golang/snappy/issues/29 +// +// As a workaround, the package was built with a known good assembler, and +// those instructions were disassembled by "objdump -d" to yield the +// 4e 0f b7 7c 5c 78 movzwq 0x78(%rsp,%r11,2),%r15 +// style comments, in AT&T asm syntax. Note that rsp here is a physical +// register, not Go/asm's SP pseudo-register (see https://golang.org/doc/asm). +// The instructions were then encoded as "BYTE $0x.." sequences, which assemble +// fine on Go 1.6. + +// The asm code generally follows the pure Go code in encode_other.go, except +// where marked with a "!!!". + +// ---------------------------------------------------------------------------- + +// func emitLiteral(dst, lit []byte) int +// +// All local variables fit into registers. The register allocation: +// - AX len(lit) +// - BX n +// - DX return value +// - DI &dst[i] +// - R10 &lit[0] +// +// The 24 bytes of stack space is to call runtime·memmove. +// +// The unusual register allocation of local variables, such as R10 for the +// source pointer, matches the allocation used at the call site in encodeBlock, +// which makes it easier to manually inline this function. +TEXT ·emitLiteral(SB), NOSPLIT, $24-56 + MOVQ dst_base+0(FP), DI + MOVQ lit_base+24(FP), R10 + MOVQ lit_len+32(FP), AX + MOVQ AX, DX + MOVL AX, BX + SUBL $1, BX + + CMPL BX, $60 + JLT oneByte + CMPL BX, $256 + JLT twoBytes + +threeBytes: + MOVB $0xf4, 0(DI) + MOVW BX, 1(DI) + ADDQ $3, DI + ADDQ $3, DX + JMP memmove + +twoBytes: + MOVB $0xf0, 0(DI) + MOVB BX, 1(DI) + ADDQ $2, DI + ADDQ $2, DX + JMP memmove + +oneByte: + SHLB $2, BX + MOVB BX, 0(DI) + ADDQ $1, DI + ADDQ $1, DX + +memmove: + MOVQ DX, ret+48(FP) + + // copy(dst[i:], lit) + // + // This means calling runtime·memmove(&dst[i], &lit[0], len(lit)), so we push + // DI, R10 and AX as arguments. + MOVQ DI, 0(SP) + MOVQ R10, 8(SP) + MOVQ AX, 16(SP) + CALL runtime·memmove(SB) + RET + +// ---------------------------------------------------------------------------- + +// func emitCopy(dst []byte, offset, length int) int +// +// All local variables fit into registers. The register allocation: +// - AX length +// - SI &dst[0] +// - DI &dst[i] +// - R11 offset +// +// The unusual register allocation of local variables, such as R11 for the +// offset, matches the allocation used at the call site in encodeBlock, which +// makes it easier to manually inline this function. +TEXT ·emitCopy(SB), NOSPLIT, $0-48 + MOVQ dst_base+0(FP), DI + MOVQ DI, SI + MOVQ offset+24(FP), R11 + MOVQ length+32(FP), AX + +loop0: + // for length >= 68 { etc } + CMPL AX, $68 + JLT step1 + + // Emit a length 64 copy, encoded as 3 bytes. + MOVB $0xfe, 0(DI) + MOVW R11, 1(DI) + ADDQ $3, DI + SUBL $64, AX + JMP loop0 + +step1: + // if length > 64 { etc } + CMPL AX, $64 + JLE step2 + + // Emit a length 60 copy, encoded as 3 bytes. + MOVB $0xee, 0(DI) + MOVW R11, 1(DI) + ADDQ $3, DI + SUBL $60, AX + +step2: + // if length >= 12 || offset >= 2048 { goto step3 } + CMPL AX, $12 + JGE step3 + CMPL R11, $2048 + JGE step3 + + // Emit the remaining copy, encoded as 2 bytes. + MOVB R11, 1(DI) + SHRL $8, R11 + SHLB $5, R11 + SUBB $4, AX + SHLB $2, AX + ORB AX, R11 + ORB $1, R11 + MOVB R11, 0(DI) + ADDQ $2, DI + + // Return the number of bytes written. + SUBQ SI, DI + MOVQ DI, ret+40(FP) + RET + +step3: + // Emit the remaining copy, encoded as 3 bytes. + SUBL $1, AX + SHLB $2, AX + ORB $2, AX + MOVB AX, 0(DI) + MOVW R11, 1(DI) + ADDQ $3, DI + + // Return the number of bytes written. + SUBQ SI, DI + MOVQ DI, ret+40(FP) + RET + +// ---------------------------------------------------------------------------- + +// func extendMatch(src []byte, i, j int) int +// +// All local variables fit into registers. The register allocation: +// - DX &src[0] +// - SI &src[j] +// - R13 &src[len(src) - 8] +// - R14 &src[len(src)] +// - R15 &src[i] +// +// The unusual register allocation of local variables, such as R15 for a source +// pointer, matches the allocation used at the call site in encodeBlock, which +// makes it easier to manually inline this function. +TEXT ·extendMatch(SB), NOSPLIT, $0-48 + MOVQ src_base+0(FP), DX + MOVQ src_len+8(FP), R14 + MOVQ i+24(FP), R15 + MOVQ j+32(FP), SI + ADDQ DX, R14 + ADDQ DX, R15 + ADDQ DX, SI + MOVQ R14, R13 + SUBQ $8, R13 + +cmp8: + // As long as we are 8 or more bytes before the end of src, we can load and + // compare 8 bytes at a time. If those 8 bytes are equal, repeat. + CMPQ SI, R13 + JA cmp1 + MOVQ (R15), AX + MOVQ (SI), BX + CMPQ AX, BX + JNE bsf + ADDQ $8, R15 + ADDQ $8, SI + JMP cmp8 + +bsf: + // If those 8 bytes were not equal, XOR the two 8 byte values, and return + // the index of the first byte that differs. The BSF instruction finds the + // least significant 1 bit, the amd64 architecture is little-endian, and + // the shift by 3 converts a bit index to a byte index. + XORQ AX, BX + BSFQ BX, BX + SHRQ $3, BX + ADDQ BX, SI + + // Convert from &src[ret] to ret. + SUBQ DX, SI + MOVQ SI, ret+40(FP) + RET + +cmp1: + // In src's tail, compare 1 byte at a time. + CMPQ SI, R14 + JAE extendMatchEnd + MOVB (R15), AX + MOVB (SI), BX + CMPB AX, BX + JNE extendMatchEnd + ADDQ $1, R15 + ADDQ $1, SI + JMP cmp1 + +extendMatchEnd: + // Convert from &src[ret] to ret. + SUBQ DX, SI + MOVQ SI, ret+40(FP) + RET + +// ---------------------------------------------------------------------------- + +// func encodeBlock(dst, src []byte) (d int) +// +// All local variables fit into registers, other than "var table". The register +// allocation: +// - AX . . +// - BX . . +// - CX 56 shift (note that amd64 shifts by non-immediates must use CX). +// - DX 64 &src[0], tableSize +// - SI 72 &src[s] +// - DI 80 &dst[d] +// - R9 88 sLimit +// - R10 . &src[nextEmit] +// - R11 96 prevHash, currHash, nextHash, offset +// - R12 104 &src[base], skip +// - R13 . &src[nextS], &src[len(src) - 8] +// - R14 . len(src), bytesBetweenHashLookups, &src[len(src)], x +// - R15 112 candidate +// +// The second column (56, 64, etc) is the stack offset to spill the registers +// when calling other functions. We could pack this slightly tighter, but it's +// simpler to have a dedicated spill map independent of the function called. +// +// "var table [maxTableSize]uint16" takes up 32768 bytes of stack space. An +// extra 56 bytes, to call other functions, and an extra 64 bytes, to spill +// local variables (registers) during calls gives 32768 + 56 + 64 = 32888. +TEXT ·encodeBlock(SB), 0, $32888-56 + MOVQ dst_base+0(FP), DI + MOVQ src_base+24(FP), SI + MOVQ src_len+32(FP), R14 + + // shift, tableSize := uint32(32-8), 1<<8 + MOVQ $24, CX + MOVQ $256, DX + +calcShift: + // for ; tableSize < maxTableSize && tableSize < len(src); tableSize *= 2 { + // shift-- + // } + CMPQ DX, $16384 + JGE varTable + CMPQ DX, R14 + JGE varTable + SUBQ $1, CX + SHLQ $1, DX + JMP calcShift + +varTable: + // var table [maxTableSize]uint16 + // + // In the asm code, unlike the Go code, we can zero-initialize only the + // first tableSize elements. Each uint16 element is 2 bytes and each MOVOU + // writes 16 bytes, so we can do only tableSize/8 writes instead of the + // 2048 writes that would zero-initialize all of table's 32768 bytes. + SHRQ $3, DX + LEAQ table-32768(SP), BX + PXOR X0, X0 + +memclr: + MOVOU X0, 0(BX) + ADDQ $16, BX + SUBQ $1, DX + JNZ memclr + + // !!! DX = &src[0] + MOVQ SI, DX + + // sLimit := len(src) - inputMargin + MOVQ R14, R9 + SUBQ $15, R9 + + // !!! Pre-emptively spill CX, DX and R9 to the stack. Their values don't + // change for the rest of the function. + MOVQ CX, 56(SP) + MOVQ DX, 64(SP) + MOVQ R9, 88(SP) + + // nextEmit := 0 + MOVQ DX, R10 + + // s := 1 + ADDQ $1, SI + + // nextHash := hash(load32(src, s), shift) + MOVL 0(SI), R11 + IMULL $0x1e35a7bd, R11 + SHRL CX, R11 + +outer: + // for { etc } + + // skip := 32 + MOVQ $32, R12 + + // nextS := s + MOVQ SI, R13 + + // candidate := 0 + MOVQ $0, R15 + +inner0: + // for { etc } + + // s := nextS + MOVQ R13, SI + + // bytesBetweenHashLookups := skip >> 5 + MOVQ R12, R14 + SHRQ $5, R14 + + // nextS = s + bytesBetweenHashLookups + ADDQ R14, R13 + + // skip += bytesBetweenHashLookups + ADDQ R14, R12 + + // if nextS > sLimit { goto emitRemainder } + MOVQ R13, AX + SUBQ DX, AX + CMPQ AX, R9 + JA emitRemainder + + // candidate = int(table[nextHash]) + // XXX: MOVWQZX table-32768(SP)(R11*2), R15 + // XXX: 4e 0f b7 7c 5c 78 movzwq 0x78(%rsp,%r11,2),%r15 + BYTE $0x4e + BYTE $0x0f + BYTE $0xb7 + BYTE $0x7c + BYTE $0x5c + BYTE $0x78 + + // table[nextHash] = uint16(s) + MOVQ SI, AX + SUBQ DX, AX + + // XXX: MOVW AX, table-32768(SP)(R11*2) + // XXX: 66 42 89 44 5c 78 mov %ax,0x78(%rsp,%r11,2) + BYTE $0x66 + BYTE $0x42 + BYTE $0x89 + BYTE $0x44 + BYTE $0x5c + BYTE $0x78 + + // nextHash = hash(load32(src, nextS), shift) + MOVL 0(R13), R11 + IMULL $0x1e35a7bd, R11 + SHRL CX, R11 + + // if load32(src, s) != load32(src, candidate) { continue } break + MOVL 0(SI), AX + MOVL (DX)(R15*1), BX + CMPL AX, BX + JNE inner0 + +fourByteMatch: + // As per the encode_other.go code: + // + // A 4-byte match has been found. We'll later see etc. + + // !!! Jump to a fast path for short (<= 16 byte) literals. See the comment + // on inputMargin in encode.go. + MOVQ SI, AX + SUBQ R10, AX + CMPQ AX, $16 + JLE emitLiteralFastPath + + // ---------------------------------------- + // Begin inline of the emitLiteral call. + // + // d += emitLiteral(dst[d:], src[nextEmit:s]) + + MOVL AX, BX + SUBL $1, BX + + CMPL BX, $60 + JLT inlineEmitLiteralOneByte + CMPL BX, $256 + JLT inlineEmitLiteralTwoBytes + +inlineEmitLiteralThreeBytes: + MOVB $0xf4, 0(DI) + MOVW BX, 1(DI) + ADDQ $3, DI + JMP inlineEmitLiteralMemmove + +inlineEmitLiteralTwoBytes: + MOVB $0xf0, 0(DI) + MOVB BX, 1(DI) + ADDQ $2, DI + JMP inlineEmitLiteralMemmove + +inlineEmitLiteralOneByte: + SHLB $2, BX + MOVB BX, 0(DI) + ADDQ $1, DI + +inlineEmitLiteralMemmove: + // Spill local variables (registers) onto the stack; call; unspill. + // + // copy(dst[i:], lit) + // + // This means calling runtime·memmove(&dst[i], &lit[0], len(lit)), so we push + // DI, R10 and AX as arguments. + MOVQ DI, 0(SP) + MOVQ R10, 8(SP) + MOVQ AX, 16(SP) + ADDQ AX, DI // Finish the "d +=" part of "d += emitLiteral(etc)". + MOVQ SI, 72(SP) + MOVQ DI, 80(SP) + MOVQ R15, 112(SP) + CALL runtime·memmove(SB) + MOVQ 56(SP), CX + MOVQ 64(SP), DX + MOVQ 72(SP), SI + MOVQ 80(SP), DI + MOVQ 88(SP), R9 + MOVQ 112(SP), R15 + JMP inner1 + +inlineEmitLiteralEnd: + // End inline of the emitLiteral call. + // ---------------------------------------- + +emitLiteralFastPath: + // !!! Emit the 1-byte encoding "uint8(len(lit)-1)<<2". + MOVB AX, BX + SUBB $1, BX + SHLB $2, BX + MOVB BX, (DI) + ADDQ $1, DI + + // !!! Implement the copy from lit to dst as a 16-byte load and store. + // (Encode's documentation says that dst and src must not overlap.) + // + // This always copies 16 bytes, instead of only len(lit) bytes, but that's + // OK. Subsequent iterations will fix up the overrun. + // + // Note that on amd64, it is legal and cheap to issue unaligned 8-byte or + // 16-byte loads and stores. This technique probably wouldn't be as + // effective on architectures that are fussier about alignment. + MOVOU 0(R10), X0 + MOVOU X0, 0(DI) + ADDQ AX, DI + +inner1: + // for { etc } + + // base := s + MOVQ SI, R12 + + // !!! offset := base - candidate + MOVQ R12, R11 + SUBQ R15, R11 + SUBQ DX, R11 + + // ---------------------------------------- + // Begin inline of the extendMatch call. + // + // s = extendMatch(src, candidate+4, s+4) + + // !!! R14 = &src[len(src)] + MOVQ src_len+32(FP), R14 + ADDQ DX, R14 + + // !!! R13 = &src[len(src) - 8] + MOVQ R14, R13 + SUBQ $8, R13 + + // !!! R15 = &src[candidate + 4] + ADDQ $4, R15 + ADDQ DX, R15 + + // !!! s += 4 + ADDQ $4, SI + +inlineExtendMatchCmp8: + // As long as we are 8 or more bytes before the end of src, we can load and + // compare 8 bytes at a time. If those 8 bytes are equal, repeat. + CMPQ SI, R13 + JA inlineExtendMatchCmp1 + MOVQ (R15), AX + MOVQ (SI), BX + CMPQ AX, BX + JNE inlineExtendMatchBSF + ADDQ $8, R15 + ADDQ $8, SI + JMP inlineExtendMatchCmp8 + +inlineExtendMatchBSF: + // If those 8 bytes were not equal, XOR the two 8 byte values, and return + // the index of the first byte that differs. The BSF instruction finds the + // least significant 1 bit, the amd64 architecture is little-endian, and + // the shift by 3 converts a bit index to a byte index. + XORQ AX, BX + BSFQ BX, BX + SHRQ $3, BX + ADDQ BX, SI + JMP inlineExtendMatchEnd + +inlineExtendMatchCmp1: + // In src's tail, compare 1 byte at a time. + CMPQ SI, R14 + JAE inlineExtendMatchEnd + MOVB (R15), AX + MOVB (SI), BX + CMPB AX, BX + JNE inlineExtendMatchEnd + ADDQ $1, R15 + ADDQ $1, SI + JMP inlineExtendMatchCmp1 + +inlineExtendMatchEnd: + // End inline of the extendMatch call. + // ---------------------------------------- + + // ---------------------------------------- + // Begin inline of the emitCopy call. + // + // d += emitCopy(dst[d:], base-candidate, s-base) + + // !!! length := s - base + MOVQ SI, AX + SUBQ R12, AX + +inlineEmitCopyLoop0: + // for length >= 68 { etc } + CMPL AX, $68 + JLT inlineEmitCopyStep1 + + // Emit a length 64 copy, encoded as 3 bytes. + MOVB $0xfe, 0(DI) + MOVW R11, 1(DI) + ADDQ $3, DI + SUBL $64, AX + JMP inlineEmitCopyLoop0 + +inlineEmitCopyStep1: + // if length > 64 { etc } + CMPL AX, $64 + JLE inlineEmitCopyStep2 + + // Emit a length 60 copy, encoded as 3 bytes. + MOVB $0xee, 0(DI) + MOVW R11, 1(DI) + ADDQ $3, DI + SUBL $60, AX + +inlineEmitCopyStep2: + // if length >= 12 || offset >= 2048 { goto inlineEmitCopyStep3 } + CMPL AX, $12 + JGE inlineEmitCopyStep3 + CMPL R11, $2048 + JGE inlineEmitCopyStep3 + + // Emit the remaining copy, encoded as 2 bytes. + MOVB R11, 1(DI) + SHRL $8, R11 + SHLB $5, R11 + SUBB $4, AX + SHLB $2, AX + ORB AX, R11 + ORB $1, R11 + MOVB R11, 0(DI) + ADDQ $2, DI + JMP inlineEmitCopyEnd + +inlineEmitCopyStep3: + // Emit the remaining copy, encoded as 3 bytes. + SUBL $1, AX + SHLB $2, AX + ORB $2, AX + MOVB AX, 0(DI) + MOVW R11, 1(DI) + ADDQ $3, DI + +inlineEmitCopyEnd: + // End inline of the emitCopy call. + // ---------------------------------------- + + // nextEmit = s + MOVQ SI, R10 + + // if s >= sLimit { goto emitRemainder } + MOVQ SI, AX + SUBQ DX, AX + CMPQ AX, R9 + JAE emitRemainder + + // As per the encode_other.go code: + // + // We could immediately etc. + + // x := load64(src, s-1) + MOVQ -1(SI), R14 + + // prevHash := hash(uint32(x>>0), shift) + MOVL R14, R11 + IMULL $0x1e35a7bd, R11 + SHRL CX, R11 + + // table[prevHash] = uint16(s-1) + MOVQ SI, AX + SUBQ DX, AX + SUBQ $1, AX + + // XXX: MOVW AX, table-32768(SP)(R11*2) + // XXX: 66 42 89 44 5c 78 mov %ax,0x78(%rsp,%r11,2) + BYTE $0x66 + BYTE $0x42 + BYTE $0x89 + BYTE $0x44 + BYTE $0x5c + BYTE $0x78 + + // currHash := hash(uint32(x>>8), shift) + SHRQ $8, R14 + MOVL R14, R11 + IMULL $0x1e35a7bd, R11 + SHRL CX, R11 + + // candidate = int(table[currHash]) + // XXX: MOVWQZX table-32768(SP)(R11*2), R15 + // XXX: 4e 0f b7 7c 5c 78 movzwq 0x78(%rsp,%r11,2),%r15 + BYTE $0x4e + BYTE $0x0f + BYTE $0xb7 + BYTE $0x7c + BYTE $0x5c + BYTE $0x78 + + // table[currHash] = uint16(s) + ADDQ $1, AX + + // XXX: MOVW AX, table-32768(SP)(R11*2) + // XXX: 66 42 89 44 5c 78 mov %ax,0x78(%rsp,%r11,2) + BYTE $0x66 + BYTE $0x42 + BYTE $0x89 + BYTE $0x44 + BYTE $0x5c + BYTE $0x78 + + // if uint32(x>>8) == load32(src, candidate) { continue } + MOVL (DX)(R15*1), BX + CMPL R14, BX + JEQ inner1 + + // nextHash = hash(uint32(x>>16), shift) + SHRQ $8, R14 + MOVL R14, R11 + IMULL $0x1e35a7bd, R11 + SHRL CX, R11 + + // s++ + ADDQ $1, SI + + // break out of the inner1 for loop, i.e. continue the outer loop. + JMP outer + +emitRemainder: + // if nextEmit < len(src) { etc } + MOVQ src_len+32(FP), AX + ADDQ DX, AX + CMPQ R10, AX + JEQ encodeBlockEnd + + // d += emitLiteral(dst[d:], src[nextEmit:]) + // + // Push args. + MOVQ DI, 0(SP) + MOVQ $0, 8(SP) // Unnecessary, as the callee ignores it, but conservative. + MOVQ $0, 16(SP) // Unnecessary, as the callee ignores it, but conservative. + MOVQ R10, 24(SP) + SUBQ R10, AX + MOVQ AX, 32(SP) + MOVQ AX, 40(SP) // Unnecessary, as the callee ignores it, but conservative. + + // Spill local variables (registers) onto the stack; call; unspill. + MOVQ DI, 80(SP) + CALL ·emitLiteral(SB) + MOVQ 80(SP), DI + + // Finish the "d +=" part of "d += emitLiteral(etc)". + ADDQ 48(SP), DI + +encodeBlockEnd: + MOVQ dst_base+0(FP), AX + SUBQ AX, DI + MOVQ DI, d+48(FP) + RET diff --git a/vendor/github.com/golang/snappy/encode_other.go b/vendor/github.com/golang/snappy/encode_other.go new file mode 100644 index 0000000000..dbcae905e6 --- /dev/null +++ b/vendor/github.com/golang/snappy/encode_other.go @@ -0,0 +1,238 @@ +// Copyright 2016 The Snappy-Go Authors. All rights reserved. +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +// +build !amd64 appengine !gc noasm + +package snappy + +func load32(b []byte, i int) uint32 { + b = b[i : i+4 : len(b)] // Help the compiler eliminate bounds checks on the next line. + return uint32(b[0]) | uint32(b[1])<<8 | uint32(b[2])<<16 | uint32(b[3])<<24 +} + +func load64(b []byte, i int) uint64 { + b = b[i : i+8 : len(b)] // Help the compiler eliminate bounds checks on the next line. + return uint64(b[0]) | uint64(b[1])<<8 | uint64(b[2])<<16 | uint64(b[3])<<24 | + uint64(b[4])<<32 | uint64(b[5])<<40 | uint64(b[6])<<48 | uint64(b[7])<<56 +} + +// emitLiteral writes a literal chunk and returns the number of bytes written. +// +// It assumes that: +// dst is long enough to hold the encoded bytes +// 1 <= len(lit) && len(lit) <= 65536 +func emitLiteral(dst, lit []byte) int { + i, n := 0, uint(len(lit)-1) + switch { + case n < 60: + dst[0] = uint8(n)<<2 | tagLiteral + i = 1 + case n < 1<<8: + dst[0] = 60<<2 | tagLiteral + dst[1] = uint8(n) + i = 2 + default: + dst[0] = 61<<2 | tagLiteral + dst[1] = uint8(n) + dst[2] = uint8(n >> 8) + i = 3 + } + return i + copy(dst[i:], lit) +} + +// emitCopy writes a copy chunk and returns the number of bytes written. +// +// It assumes that: +// dst is long enough to hold the encoded bytes +// 1 <= offset && offset <= 65535 +// 4 <= length && length <= 65535 +func emitCopy(dst []byte, offset, length int) int { + i := 0 + // The maximum length for a single tagCopy1 or tagCopy2 op is 64 bytes. The + // threshold for this loop is a little higher (at 68 = 64 + 4), and the + // length emitted down below is is a little lower (at 60 = 64 - 4), because + // it's shorter to encode a length 67 copy as a length 60 tagCopy2 followed + // by a length 7 tagCopy1 (which encodes as 3+2 bytes) than to encode it as + // a length 64 tagCopy2 followed by a length 3 tagCopy2 (which encodes as + // 3+3 bytes). The magic 4 in the 64±4 is because the minimum length for a + // tagCopy1 op is 4 bytes, which is why a length 3 copy has to be an + // encodes-as-3-bytes tagCopy2 instead of an encodes-as-2-bytes tagCopy1. + for length >= 68 { + // Emit a length 64 copy, encoded as 3 bytes. + dst[i+0] = 63<<2 | tagCopy2 + dst[i+1] = uint8(offset) + dst[i+2] = uint8(offset >> 8) + i += 3 + length -= 64 + } + if length > 64 { + // Emit a length 60 copy, encoded as 3 bytes. + dst[i+0] = 59<<2 | tagCopy2 + dst[i+1] = uint8(offset) + dst[i+2] = uint8(offset >> 8) + i += 3 + length -= 60 + } + if length >= 12 || offset >= 2048 { + // Emit the remaining copy, encoded as 3 bytes. + dst[i+0] = uint8(length-1)<<2 | tagCopy2 + dst[i+1] = uint8(offset) + dst[i+2] = uint8(offset >> 8) + return i + 3 + } + // Emit the remaining copy, encoded as 2 bytes. + dst[i+0] = uint8(offset>>8)<<5 | uint8(length-4)<<2 | tagCopy1 + dst[i+1] = uint8(offset) + return i + 2 +} + +// extendMatch returns the largest k such that k <= len(src) and that +// src[i:i+k-j] and src[j:k] have the same contents. +// +// It assumes that: +// 0 <= i && i < j && j <= len(src) +func extendMatch(src []byte, i, j int) int { + for ; j < len(src) && src[i] == src[j]; i, j = i+1, j+1 { + } + return j +} + +func hash(u, shift uint32) uint32 { + return (u * 0x1e35a7bd) >> shift +} + +// encodeBlock encodes a non-empty src to a guaranteed-large-enough dst. It +// assumes that the varint-encoded length of the decompressed bytes has already +// been written. +// +// It also assumes that: +// len(dst) >= MaxEncodedLen(len(src)) && +// minNonLiteralBlockSize <= len(src) && len(src) <= maxBlockSize +func encodeBlock(dst, src []byte) (d int) { + // Initialize the hash table. Its size ranges from 1<<8 to 1<<14 inclusive. + // The table element type is uint16, as s < sLimit and sLimit < len(src) + // and len(src) <= maxBlockSize and maxBlockSize == 65536. + const ( + maxTableSize = 1 << 14 + // tableMask is redundant, but helps the compiler eliminate bounds + // checks. + tableMask = maxTableSize - 1 + ) + shift := uint32(32 - 8) + for tableSize := 1 << 8; tableSize < maxTableSize && tableSize < len(src); tableSize *= 2 { + shift-- + } + // In Go, all array elements are zero-initialized, so there is no advantage + // to a smaller tableSize per se. However, it matches the C++ algorithm, + // and in the asm versions of this code, we can get away with zeroing only + // the first tableSize elements. + var table [maxTableSize]uint16 + + // sLimit is when to stop looking for offset/length copies. The inputMargin + // lets us use a fast path for emitLiteral in the main loop, while we are + // looking for copies. + sLimit := len(src) - inputMargin + + // nextEmit is where in src the next emitLiteral should start from. + nextEmit := 0 + + // The encoded form must start with a literal, as there are no previous + // bytes to copy, so we start looking for hash matches at s == 1. + s := 1 + nextHash := hash(load32(src, s), shift) + + for { + // Copied from the C++ snappy implementation: + // + // Heuristic match skipping: If 32 bytes are scanned with no matches + // found, start looking only at every other byte. If 32 more bytes are + // scanned (or skipped), look at every third byte, etc.. When a match + // is found, immediately go back to looking at every byte. This is a + // small loss (~5% performance, ~0.1% density) for compressible data + // due to more bookkeeping, but for non-compressible data (such as + // JPEG) it's a huge win since the compressor quickly "realizes" the + // data is incompressible and doesn't bother looking for matches + // everywhere. + // + // The "skip" variable keeps track of how many bytes there are since + // the last match; dividing it by 32 (ie. right-shifting by five) gives + // the number of bytes to move ahead for each iteration. + skip := 32 + + nextS := s + candidate := 0 + for { + s = nextS + bytesBetweenHashLookups := skip >> 5 + nextS = s + bytesBetweenHashLookups + skip += bytesBetweenHashLookups + if nextS > sLimit { + goto emitRemainder + } + candidate = int(table[nextHash&tableMask]) + table[nextHash&tableMask] = uint16(s) + nextHash = hash(load32(src, nextS), shift) + if load32(src, s) == load32(src, candidate) { + break + } + } + + // A 4-byte match has been found. We'll later see if more than 4 bytes + // match. But, prior to the match, src[nextEmit:s] are unmatched. Emit + // them as literal bytes. + d += emitLiteral(dst[d:], src[nextEmit:s]) + + // Call emitCopy, and then see if another emitCopy could be our next + // move. Repeat until we find no match for the input immediately after + // what was consumed by the last emitCopy call. + // + // If we exit this loop normally then we need to call emitLiteral next, + // though we don't yet know how big the literal will be. We handle that + // by proceeding to the next iteration of the main loop. We also can + // exit this loop via goto if we get close to exhausting the input. + for { + // Invariant: we have a 4-byte match at s, and no need to emit any + // literal bytes prior to s. + base := s + + // Extend the 4-byte match as long as possible. + // + // This is an inlined version of: + // s = extendMatch(src, candidate+4, s+4) + s += 4 + for i := candidate + 4; s < len(src) && src[i] == src[s]; i, s = i+1, s+1 { + } + + d += emitCopy(dst[d:], base-candidate, s-base) + nextEmit = s + if s >= sLimit { + goto emitRemainder + } + + // We could immediately start working at s now, but to improve + // compression we first update the hash table at s-1 and at s. If + // another emitCopy is not our next move, also calculate nextHash + // at s+1. At least on GOARCH=amd64, these three hash calculations + // are faster as one load64 call (with some shifts) instead of + // three load32 calls. + x := load64(src, s-1) + prevHash := hash(uint32(x>>0), shift) + table[prevHash&tableMask] = uint16(s - 1) + currHash := hash(uint32(x>>8), shift) + candidate = int(table[currHash&tableMask]) + table[currHash&tableMask] = uint16(s) + if uint32(x>>8) != load32(src, candidate) { + nextHash = hash(uint32(x>>16), shift) + s++ + break + } + } + } + +emitRemainder: + if nextEmit < len(src) { + d += emitLiteral(dst[d:], src[nextEmit:]) + } + return d +} diff --git a/vendor/github.com/golang/snappy/snappy.go b/vendor/github.com/golang/snappy/snappy.go index 010254265f..ece692ea46 100644 --- a/vendor/github.com/golang/snappy/snappy.go +++ b/vendor/github.com/golang/snappy/snappy.go @@ -2,10 +2,21 @@ // Use of this source code is governed by a BSD-style // license that can be found in the LICENSE file. -// Package snappy implements the snappy block-based compression format. -// It aims for very high speeds and reasonable compression. +// Package snappy implements the Snappy compression format. It aims for very +// high speeds and reasonable compression. // -// The C++ snappy implementation is at https://github.com/google/snappy +// There are actually two Snappy formats: block and stream. They are related, +// but different: trying to decompress block-compressed data as a Snappy stream +// will fail, and vice versa. The block format is the Decode and Encode +// functions and the stream format is the Reader and Writer types. +// +// The block format, the more common case, is used when the complete size (the +// number of bytes) of the original data is known upfront, at the time +// compression starts. The stream format, also known as the framing format, is +// for when that isn't always true. +// +// The canonical, C++ implementation is at https://github.com/google/snappy and +// it only implements the block format. package snappy // import "github.com/golang/snappy" import ( @@ -32,7 +43,10 @@ Lempel-Ziv compression algorithms. In particular: - For l == 2, the offset ranges in [0, 1<<16) and the length in [1, 65). The length is 1 + m. The offset is the little-endian unsigned integer denoted by the next 2 bytes. - - For l == 3, this tag is a legacy format that is no longer supported. + - For l == 3, this tag is a legacy format that is no longer issued by most + encoders. Nonetheless, the offset ranges in [0, 1<<32) and the length in + [1, 65). The length is 1 + m. The offset is the little-endian unsigned + integer denoted by the next 4 bytes. */ const ( tagLiteral = 0x00 |