rspamd/contrib/stringzilla/include/stringzilla/stringzilla.h

/**
 *  @brief  StringZilla is a collection of simple string algorithms, designed to be used in Big Data applications.
 *          It may be slower than LibC, but has a broader & cleaner interface, and a very short implementation
 *          targeting modern x86 CPUs with AVX-512 and Arm NEON and older CPUs with SWAR and auto-vectorization.
 *
 *  Consider overriding the following macros to customize the library:
 *
 *  - `SZ_DEBUG=0` - whether to enable debug assertions and logging.
 *  - `SZ_DYNAMIC_DISPATCH=0` - whether to use runtime dispatching of the most advanced SIMD backend.
 *  - `SZ_USE_MISALIGNED_LOADS=0` - whether to use misaligned loads on platforms that support them.
 *  - `SZ_SWAR_THRESHOLD=24` - threshold for switching to SWAR backend over serial byte-level for-loops.
 *  - `SZ_USE_X86_AVX512=?` - whether to use AVX-512 instructions on x86_64.
 *  - `SZ_USE_X86_AVX2=?` - whether to use AVX2 instructions on x86_64.
 *  - `SZ_USE_ARM_NEON=?` - whether to use NEON instructions on ARM.
 *  - `SZ_USE_ARM_SVE=?` - whether to use SVE instructions on ARM.
 *
 *  @see    StringZilla: https://github.com/ashvardanian/StringZilla/blob/main/README.md
 *  @see    LibC String: https://pubs.opengroup.org/onlinepubs/009695399/basedefs/string.h.html
 *
 *  @file   stringzilla.h
 *  @author Ash Vardanian
 */
#ifndef STRINGZILLA_H_
#define STRINGZILLA_H_

#define STRINGZILLA_VERSION_MAJOR 3
#define STRINGZILLA_VERSION_MINOR 5
#define STRINGZILLA_VERSION_PATCH 0

/**
 *  @brief  When set to 1, the library will include the following LibC headers: <stddef.h> and <stdint.h>.
 *          In debug builds (SZ_DEBUG=1), the library will also include <stdio.h> and <stdlib.h>.
 *
 *  You may want to disable this compiling for use in the kernel, or in embedded systems.
 *  You may also avoid them, if you are very sensitive to compilation time and avoid pre-compiled headers.
 *  https://artificial-mind.net/projects/compile-health/
 */
#ifndef SZ_AVOID_LIBC
#define SZ_AVOID_LIBC (0) // true or false
#endif

/**
 *  @brief  A misaligned load can be - trying to fetch eight consecutive bytes from an address
 *          that is not divisible by eight. On x86 enabled by default. On ARM it's not.
 *
 *  Most platforms support it, but there is no industry standard way to check for those.
 *  This value will mostly affect the performance of the serial (SWAR) backend.
 */
#ifndef SZ_USE_MISALIGNED_LOADS
#if defined(__x86_64__) || defined(_M_X64) || defined(__i386__) || defined(_M_IX86)
#define SZ_USE_MISALIGNED_LOADS (1) // true or false
#else
#define SZ_USE_MISALIGNED_LOADS (0) // true or false
#endif
#endif

/**
 *  @brief  Removes compile-time dispatching, and replaces it with runtime dispatching.
 *          So the `sz_find` function will invoke the most advanced backend supported by the CPU,
 *          that runs the program, rather than the most advanced backend supported by the CPU
 *          used to compile the library or the downstream application.
 */
#ifndef SZ_DYNAMIC_DISPATCH
#define SZ_DYNAMIC_DISPATCH (0) // true or false
#endif

/**
 *  @brief  Analogous to `size_t` and `std::size_t`, unsigned integer, identical to pointer size.
 *          64-bit on most platforms where pointers are 64-bit.
 *          32-bit on platforms where pointers are 32-bit.
 */
#if defined(__LP64__) || defined(_LP64) || defined(__x86_64__) || defined(_WIN64)
#define SZ_DETECT_64_BIT (1)
#define SZ_SIZE_MAX (0xFFFFFFFFFFFFFFFFull)  // Largest unsigned integer that fits into 64 bits.
#define SZ_SSIZE_MAX (0x7FFFFFFFFFFFFFFFull) // Largest signed integer that fits into 64 bits.
#else
#define SZ_DETECT_64_BIT (0)
#define SZ_SIZE_MAX (0xFFFFFFFFu)  // Largest unsigned integer that fits into 32 bits.
#define SZ_SSIZE_MAX (0x7FFFFFFFu) // Largest signed integer that fits into 32 bits.
#endif

/**
 *  @brief  On Big-Endian machines StringZilla will work in compatibility mode.
 *          This disables SWAR hacks to minimize code duplication, assuming practically
 *          all modern popular platforms are Little-Endian.
 *
 *  This variable is hard to infer from macros reliably. It's best to set it manually.
 *  For that CMake provides the `TestBigEndian` and `CMAKE_<LANG>_BYTE_ORDER` (from 3.20 onwards).
 *  In Python one can check `sys.byteorder == 'big'` in the `setup.py` script and pass the appropriate macro.
 *  https://stackoverflow.com/a/27054190
 */
#ifndef SZ_DETECT_BIG_ENDIAN
#if defined(__BYTE_ORDER) && __BYTE_ORDER == __BIG_ENDIAN || defined(__BIG_ENDIAN__) || defined(__ARMEB__) || \
    defined(__THUMBEB__) || defined(__AARCH64EB__) || defined(_MIBSEB) || defined(__MIBSEB) || defined(__MIBSEB__)
#define SZ_DETECT_BIG_ENDIAN (1) //< It's a big-endian target architecture
#else
#define SZ_DETECT_BIG_ENDIAN (0) //< It's a little-endian target architecture
#endif
#endif

/*
 *  Debugging and testing.
 */
#ifndef SZ_DEBUG
#if defined(DEBUG) || defined(_DEBUG) // This means "Not using DEBUG information".
#define SZ_DEBUG (1)
#else
#define SZ_DEBUG (0)
#endif
#endif

/**
 *  @brief  Threshold for switching to SWAR (8-bytes at a time) backend over serial byte-level for-loops.
 *          On very short strings, under 16 bytes long, at most a single word will be processed with SWAR.
 *          Assuming potentially misaligned loads, SWAR makes sense only after ~24 bytes.
 */
#ifndef SZ_SWAR_THRESHOLD
#if SZ_DEBUG
#define SZ_SWAR_THRESHOLD (8u) // 8 bytes in debug builds
#else
#define SZ_SWAR_THRESHOLD (24u) // 24 bytes in release builds
#endif
#endif

/*  Annotation for the public API symbols:
 *
 *  - `SZ_PUBLIC` is used for functions that are part of the public API.
 *  - `SZ_INTERNAL` is used for internal helper functions with unstable APIs.
 *  - `SZ_DYNAMIC` is used for functions that are part of the public API, but are dispatched at runtime.
 */
#ifndef SZ_DYNAMIC
#if SZ_DYNAMIC_DISPATCH
#if defined(_WIN32) || defined(__CYGWIN__)
#define SZ_DYNAMIC __declspec(dllexport)
#define SZ_PUBLIC inline static
#define SZ_INTERNAL inline static
#else
#define SZ_DYNAMIC __attribute__((visibility("default")))
#define SZ_PUBLIC __attribute__((unused)) inline static
#define SZ_INTERNAL __attribute__((always_inline)) inline static
#endif // _WIN32 || __CYGWIN__
#else
#define SZ_DYNAMIC inline static
#define SZ_PUBLIC inline static
#define SZ_INTERNAL inline static
#endif // SZ_DYNAMIC_DISPATCH
#endif // SZ_DYNAMIC

#ifdef __cplusplus
extern "C" {
#endif

/*
 *  Let's infer the integer types or pull them from LibC,
 *  if that is allowed by the user.
 */
#if !SZ_AVOID_LIBC
#include <stddef.h>           // `size_t`
#include <stdint.h>           // `uint8_t`
typedef int8_t sz_i8_t;       // Always 8 bits
typedef uint8_t sz_u8_t;      // Always 8 bits
typedef uint16_t sz_u16_t;    // Always 16 bits
typedef int32_t sz_i32_t;     // Always 32 bits
typedef uint32_t sz_u32_t;    // Always 32 bits
typedef uint64_t sz_u64_t;    // Always 64 bits
typedef int64_t sz_i64_t;     // Always 64 bits
typedef size_t sz_size_t;     // Pointer-sized unsigned integer, 32 or 64 bits
typedef ptrdiff_t sz_ssize_t; // Signed version of `sz_size_t`, 32 or 64 bits

#else // if SZ_AVOID_LIBC:

typedef signed char sz_i8_t;         // Always 8 bits
typedef unsigned char sz_u8_t;       // Always 8 bits
typedef unsigned short sz_u16_t;     // Always 16 bits
typedef int sz_i32_t;                // Always 32 bits
typedef unsigned int sz_u32_t;       // Always 32 bits
typedef long long sz_i64_t;          // Always 64 bits
typedef unsigned long long sz_u64_t; // Always 64 bits

#if SZ_DETECT_64_BIT
typedef unsigned long long sz_size_t; // 64-bit.
typedef long long sz_ssize_t;         // 64-bit.
#else
typedef unsigned sz_size_t;  // 32-bit.
typedef unsigned sz_ssize_t; // 32-bit.
#endif // SZ_DETECT_64_BIT

#endif // SZ_AVOID_LIBC

/**
 *  @brief  Compile-time assert macro similar to `static_assert` in C++.
 */
#define sz_static_assert(condition, name)                \
    typedef struct {                                     \
        int static_assert_##name : (condition) ? 1 : -1; \
    } sz_static_assert_##name##_t

sz_static_assert(sizeof(sz_size_t) == sizeof(void *), sz_size_t_must_be_pointer_size);
sz_static_assert(sizeof(sz_ssize_t) == sizeof(void *), sz_ssize_t_must_be_pointer_size);

#pragma region Public API

typedef char *sz_ptr_t;          // A type alias for `char *`
typedef char const *sz_cptr_t;   // A type alias for `char const *`
typedef sz_i8_t sz_error_cost_t; // Character mismatch cost for fuzzy matching functions

typedef sz_u64_t sz_sorted_idx_t; // Index of a sorted string in a list of strings

typedef enum { sz_false_k = 0, sz_true_k = 1 } sz_bool_t;                        // Only one relevant bit
typedef enum { sz_less_k = -1, sz_equal_k = 0, sz_greater_k = 1 } sz_ordering_t; // Only three possible states: <=>

/**
 *  @brief  Tiny string-view structure. It's POD type, unlike the `std::string_view`.
 */
typedef struct sz_string_view_t {
    sz_cptr_t start;
    sz_size_t length;
} sz_string_view_t;

/**
 *  @brief  Enumeration of SIMD capabilities of the target architecture.
 *          Used to introspect the supported functionality of the dynamic library.
 */
typedef enum sz_capability_t {
    sz_cap_serial_k = 1,       /// Serial (non-SIMD) capability
    sz_cap_any_k = 0x7FFFFFFF, /// Mask representing any capability

    sz_cap_arm_neon_k = 1 << 10, /// ARM NEON capability
    sz_cap_arm_sve_k = 1 << 11,  /// ARM SVE capability TODO: Not yet supported or used

    sz_cap_x86_avx2_k = 1 << 20,       /// x86 AVX2 capability
    sz_cap_x86_avx512f_k = 1 << 21,    /// x86 AVX512 F capability
    sz_cap_x86_avx512bw_k = 1 << 22,   /// x86 AVX512 BW instruction capability
    sz_cap_x86_avx512vl_k = 1 << 23,   /// x86 AVX512 VL instruction capability
    sz_cap_x86_avx512vbmi_k = 1 << 24, /// x86 AVX512 VBMI instruction capability
    sz_cap_x86_gfni_k = 1 << 25,       /// x86 AVX512 GFNI instruction capability

} sz_capability_t;

/**
 *  @brief  Function to determine the SIMD capabilities of the current machine @b only at @b runtime.
 *  @return A bitmask of the SIMD capabilities represented as a `sz_capability_t` enum value.
 */
SZ_DYNAMIC sz_capability_t sz_capabilities(void);

/**
 *  @brief  Bit-set structure for 256 possible byte values. Useful for filtering and search.
 *  @see    sz_charset_init, sz_charset_add, sz_charset_contains, sz_charset_invert
 */
typedef union sz_charset_t {
    sz_u64_t _u64s[4];
    sz_u32_t _u32s[8];
    sz_u16_t _u16s[16];
    sz_u8_t _u8s[32];
} sz_charset_t;

/** @brief  Initializes a bit-set to an empty collection, meaning - all characters are banned. */
SZ_PUBLIC void sz_charset_init(sz_charset_t *s) { s->_u64s[0] = s->_u64s[1] = s->_u64s[2] = s->_u64s[3] = 0; }

/** @brief  Adds a character to the set and accepts @b unsigned integers. */
SZ_PUBLIC void sz_charset_add_u8(sz_charset_t *s, sz_u8_t c) { s->_u64s[c >> 6] |= (1ull << (c & 63u)); }

/** @brief  Adds a character to the set. Consider @b sz_charset_add_u8. */
SZ_PUBLIC void sz_charset_add(sz_charset_t *s, char c) { sz_charset_add_u8(s, *(sz_u8_t *)(&c)); } // bitcast

/** @brief  Checks if the set contains a given character and accepts @b unsigned integers. */
SZ_PUBLIC sz_bool_t sz_charset_contains_u8(sz_charset_t const *s, sz_u8_t c) {
    // Checking the bit can be done in different ways:
    // - (s->_u64s[c >> 6] & (1ull << (c & 63u))) != 0
    // - (s->_u32s[c >> 5] & (1u << (c & 31u))) != 0
    // - (s->_u16s[c >> 4] & (1u << (c & 15u))) != 0
    // - (s->_u8s[c >> 3] & (1u << (c & 7u))) != 0
    return (sz_bool_t)((s->_u64s[c >> 6] & (1ull << (c & 63u))) != 0);
}

/** @brief  Checks if the set contains a given character. Consider @b sz_charset_contains_u8. */
SZ_PUBLIC sz_bool_t sz_charset_contains(sz_charset_t const *s, char c) {
    return sz_charset_contains_u8(s, *(sz_u8_t *)(&c)); // bitcast
}

/** @brief  Inverts the contents of the set, so allowed character get disallowed, and vice versa. */
SZ_PUBLIC void sz_charset_invert(sz_charset_t *s) {
    s->_u64s[0] ^= 0xFFFFFFFFFFFFFFFFull, s->_u64s[1] ^= 0xFFFFFFFFFFFFFFFFull, //
        s->_u64s[2] ^= 0xFFFFFFFFFFFFFFFFull, s->_u64s[3] ^= 0xFFFFFFFFFFFFFFFFull;
}

typedef void *(*sz_memory_allocate_t)(sz_size_t, void *);
typedef void (*sz_memory_free_t)(void *, sz_size_t, void *);
typedef sz_u64_t (*sz_random_generator_t)(void *);

/**
 *  @brief  Some complex pattern matching algorithms may require memory allocations.
 *          This structure is used to pass the memory allocator to those functions.
 *  @see    sz_memory_allocator_init_fixed
 */
typedef struct sz_memory_allocator_t {
    sz_memory_allocate_t allocate;
    sz_memory_free_t free;
    void *handle;
} sz_memory_allocator_t;

/**
 *  @brief  Initializes a memory allocator to use the system default `malloc` and `free`.
 *          ! The function is not available if the library was compiled with `SZ_AVOID_LIBC`.
 *
 *  @param alloc    Memory allocator to initialize.
 */
SZ_PUBLIC void sz_memory_allocator_init_default(sz_memory_allocator_t *alloc);

/**
 *  @brief  Initializes a memory allocator to use a static-capacity buffer.
 *          No dynamic allocations will be performed.
 *
 *  @param alloc    Memory allocator to initialize.
 *  @param buffer   Buffer to use for allocations.
 *  @param length   Length of the buffer. @b Must be greater than 8 bytes. Different values would be optimal for
 *                  different algorithms and input lengths, but 4096 bytes (one RAM page) is a good default.
 */
SZ_PUBLIC void sz_memory_allocator_init_fixed(sz_memory_allocator_t *alloc, void *buffer, sz_size_t length);

/**
 *  @brief  The number of bytes a stack-allocated string can hold, including the SZ_NULL termination character.
 *          ! This can't be changed from outside. Don't use the `#error` as it may already be included and set.
 */
#ifdef SZ_STRING_INTERNAL_SPACE
#undef SZ_STRING_INTERNAL_SPACE
#endif
#define SZ_STRING_INTERNAL_SPACE (sizeof(sz_size_t) * 3 - 1) // 3 pointers minus one byte for an 8-bit length

/**
 *  @brief  Tiny memory-owning string structure with a Small String Optimization (SSO).
 *          Differs in layout from Folly, Clang, GCC, and probably most other implementations.
 *          It's designed to avoid any branches on read-only operations, and can store up
 *          to 22 characters on stack on 64-bit machines, followed by the SZ_NULL-termination character.
 *
 *  @section Changing Length
 *
 *  One nice thing about this design, is that you can, in many cases, change the length of the string
 *  without any branches, invoking a `+=` or `-=` on the 64-bit `length` field. If the string is on heap,
 *  the solution is obvious. If it's on stack, inplace decrement wouldn't affect the top bytes of the string,
 *  only changing the last byte containing the length.
 */
typedef union sz_string_t {

#if !SZ_DETECT_BIG_ENDIAN

    struct external {
        sz_ptr_t start;
        sz_size_t length;
        sz_size_t space;
        sz_size_t padding;
    } external;

    struct internal {
        sz_ptr_t start;
        sz_u8_t length;
        char chars[SZ_STRING_INTERNAL_SPACE];
    } internal;

#else

    struct external {
        sz_ptr_t start;
        sz_size_t space;
        sz_size_t padding;
        sz_size_t length;
    } external;

    struct internal {
        sz_ptr_t start;
        char chars[SZ_STRING_INTERNAL_SPACE];
        sz_u8_t length;
    } internal;

#endif

    sz_size_t words[4];

} sz_string_t;

typedef sz_u64_t (*sz_hash_t)(sz_cptr_t, sz_size_t);
typedef sz_bool_t (*sz_equal_t)(sz_cptr_t, sz_cptr_t, sz_size_t);
typedef sz_ordering_t (*sz_order_t)(sz_cptr_t, sz_size_t, sz_cptr_t, sz_size_t);
typedef void (*sz_to_converter_t)(sz_cptr_t, sz_size_t, sz_ptr_t);

/**
 *  @brief  Computes the 64-bit unsigned hash of a string. Fairly fast for short strings,
 *          simple implementation, and supports rolling computation, reused in other APIs.
 *          Similar to `std::hash` in C++.
 *
 *  @param text     String to hash.
 *  @param length   Number of bytes in the text.
 *  @return         64-bit hash value.
 *
 *  @see    sz_hashes, sz_hashes_fingerprint, sz_hashes_intersection
 */
SZ_PUBLIC sz_u64_t sz_hash(sz_cptr_t text, sz_size_t length);

/** @copydoc sz_hash */
SZ_PUBLIC sz_u64_t sz_hash_serial(sz_cptr_t text, sz_size_t length);

/**
 *  @brief  Checks if two string are equal.
 *          Similar to `memcmp(a, b, length) == 0` in LibC and `a == b` in STL.
 *
 *  The implementation of this function is very similar to `sz_order`, but the usage patterns are different.
 *  This function is more often used in parsing, while `sz_order` is often used in sorting.
 *  It works best on platforms with cheap
 *
 *  @param a        First string to compare.
 *  @param b        Second string to compare.
 *  @param length   Number of bytes in both strings.
 *  @return         1 if strings match, 0 otherwise.
 */
SZ_DYNAMIC sz_bool_t sz_equal(sz_cptr_t a, sz_cptr_t b, sz_size_t length);

/** @copydoc sz_equal */
SZ_PUBLIC sz_bool_t sz_equal_serial(sz_cptr_t a, sz_cptr_t b, sz_size_t length);

/**
 *  @brief  Estimates the relative order of two strings. Equivalent to `memcmp(a, b, length)` in LibC.
 *          Can be used on different length strings.
 *
 *  @param a        First string to compare.
 *  @param a_length Number of bytes in the first string.
 *  @param b        Second string to compare.
 *  @param b_length Number of bytes in the second string.
 *  @return         Negative if (a < b), positive if (a > b), zero if they are equal.
 */
SZ_DYNAMIC sz_ordering_t sz_order(sz_cptr_t a, sz_size_t a_length, sz_cptr_t b, sz_size_t b_length);

/** @copydoc sz_order */
SZ_PUBLIC sz_ordering_t sz_order_serial(sz_cptr_t a, sz_size_t a_length, sz_cptr_t b, sz_size_t b_length);

/**
 *  @brief  Equivalent to `for (char & c : text) c = tolower(c)`.
 *
 *  ASCII characters [A, Z] map to decimals [65, 90], and [a, z] map to [97, 122].
 *  So there are 26 english letters, shifted by 32 values, meaning that a conversion
 *  can be done by flipping the 5th bit each inappropriate character byte. This, however,
 *  breaks for extended ASCII, so a different solution is needed.
 *  http://0x80.pl/notesen/2016-01-06-swar-swap-case.html
 *
 *  @param text     String to be normalized.
 *  @param length   Number of bytes in the string.
 *  @param result   Output string, can point to the same address as ::text.
 */
SZ_PUBLIC void sz_tolower(sz_cptr_t text, sz_size_t length, sz_ptr_t result);

/**
 *  @brief  Equivalent to `for (char & c : text) c = toupper(c)`.
 *
 *  ASCII characters [A, Z] map to decimals [65, 90], and [a, z] map to [97, 122].
 *  So there are 26 english letters, shifted by 32 values, meaning that a conversion
 *  can be done by flipping the 5th bit each inappropriate character byte. This, however,
 *  breaks for extended ASCII, so a different solution is needed.
 *  http://0x80.pl/notesen/2016-01-06-swar-swap-case.html
 *
 *  @param text     String to be normalized.
 *  @param length   Number of bytes in the string.
 *  @param result   Output string, can point to the same address as ::text.
 */
SZ_PUBLIC void sz_toupper(sz_cptr_t text, sz_size_t length, sz_ptr_t result);

/**
 *  @brief  Equivalent to `for (char & c : text) c = toascii(c)`.
 *
 *  @param text     String to be normalized.
 *  @param length   Number of bytes in the string.
 *  @param result   Output string, can point to the same address as ::text.
 */
SZ_PUBLIC void sz_toascii(sz_cptr_t text, sz_size_t length, sz_ptr_t result);

/**
 *  @brief  Checks if all characters in the range are valid ASCII characters.
 *
 *  @param text     String to be analyzed.
 *  @param length   Number of bytes in the string.
 *  @return         Whether all characters are valid ASCII characters.
 */
SZ_PUBLIC sz_bool_t sz_isascii(sz_cptr_t text, sz_size_t length);

/**
 *  @brief  Generates a random string for a given alphabet, avoiding integer division and modulo operations.
 *          Similar to `text[i] = alphabet[rand() % cardinality]`.
 *
 *  The modulo operation is expensive, and should be avoided in performance-critical code.
 *  We avoid it using small lookup tables and replacing it with a multiplication and shifts, similar to `libdivide`.
 *  Alternative algorithms would include:
 *      - Montgomery form: https://en.algorithmica.org/hpc/number-theory/montgomery/
 *      - Barret reduction: https://www.nayuki.io/page/barrett-reduction-algorithm
 *      - Lemire's trick: https://lemire.me/blog/2016/06/27/a-fast-alternative-to-the-modulo-reduction/
 *
 *  @param alphabet     Set of characters to sample from.
 *  @param cardinality  Number of characters to sample from.
 *  @param text         Output string, can point to the same address as ::text.
 *  @param generate     Callback producing random numbers given the generator state.
 *  @param generator    Generator state, can be a pointer to a seed, or a pointer to a random number generator.
 */
SZ_DYNAMIC void sz_generate(sz_cptr_t alphabet, sz_size_t cardinality, sz_ptr_t text, sz_size_t length,
                            sz_random_generator_t generate, void *generator);

/** @copydoc sz_generate */
SZ_PUBLIC void sz_generate_serial(sz_cptr_t alphabet, sz_size_t cardinality, sz_ptr_t text, sz_size_t length,
                                  sz_random_generator_t generate, void *generator);

/**
 *  @brief  Similar to `memcpy`, copies contents of one string into another.
 *          The behavior is undefined if the strings overlap.
 *
 *  @param target   String to copy into.
 *  @param length   Number of bytes to copy.
 *  @param source   String to copy from.
 */
SZ_DYNAMIC void sz_copy(sz_ptr_t target, sz_cptr_t source, sz_size_t length);

/** @copydoc sz_copy */
SZ_PUBLIC void sz_copy_serial(sz_ptr_t target, sz_cptr_t source, sz_size_t length);

/**
 *  @brief  Similar to `memmove`, copies (moves) contents of one string into another.
 *          Unlike `sz_copy`, allows overlapping strings as arguments.
 *
 *  @param target   String to copy into.
 *  @param length   Number of bytes to copy.
 *  @param source   String to copy from.
 */
SZ_DYNAMIC void sz_move(sz_ptr_t target, sz_cptr_t source, sz_size_t length);

/** @copydoc sz_move */
SZ_PUBLIC void sz_move_serial(sz_ptr_t target, sz_cptr_t source, sz_size_t length);

typedef void (*sz_move_t)(sz_ptr_t, sz_cptr_t, sz_size_t);

/**
 *  @brief  Similar to `memset`, fills a string with a given value.
 *
 *  @param target   String to fill.
 *  @param length   Number of bytes to fill.
 *  @param value    Value to fill with.
 */
SZ_DYNAMIC void sz_fill(sz_ptr_t target, sz_size_t length, sz_u8_t value);

/** @copydoc sz_fill */
SZ_PUBLIC void sz_fill_serial(sz_ptr_t target, sz_size_t length, sz_u8_t value);

typedef void (*sz_fill_t)(sz_ptr_t, sz_size_t, sz_u8_t);

/**
 *  @brief  Initializes a string class instance to an empty value.
 */
SZ_PUBLIC void sz_string_init(sz_string_t *string);

/**
 *  @brief  Convenience function checking if the provided string is stored inside of the ::string instance itself,
 *          alternative being - allocated in a remote region of the heap.
 */
SZ_PUBLIC sz_bool_t sz_string_is_on_stack(sz_string_t const *string);

/**
 *  @brief  Unpacks the opaque instance of a string class into its components.
 *          Recommended to use only in read-only operations.
 *
 *  @param string       String to unpack.
 *  @param start        Pointer to the start of the string.
 *  @param length       Number of bytes in the string, before the SZ_NULL character.
 *  @param space        Number of bytes allocated for the string (heap or stack), including the SZ_NULL character.
 *  @param is_external  Whether the string is allocated on the heap externally, or fits withing ::string instance.
 */
SZ_PUBLIC void sz_string_unpack(sz_string_t const *string, sz_ptr_t *start, sz_size_t *length, sz_size_t *space,
                                sz_bool_t *is_external);

/**
 *  @brief  Unpacks only the start and length of the string.
 *          Recommended to use only in read-only operations.
 *
 * @param string       String to unpack.
 * @param start        Pointer to the start of the string.
 * @param length       Number of bytes in the string, before the SZ_NULL character.
 */
SZ_PUBLIC void sz_string_range(sz_string_t const *string, sz_ptr_t *start, sz_size_t *length);

/**
 *  @brief  Constructs a string of a given ::length with noisy contents.
 *          Use the returned character pointer to populate the string.
 *
 *  @param string       String to initialize.
 *  @param length       Number of bytes in the string, before the SZ_NULL character.
 *  @param allocator    Memory allocator to use for the allocation.
 *  @return             SZ_NULL if the operation failed, pointer to the start of the string otherwise.
 */
SZ_PUBLIC sz_ptr_t sz_string_init_length(sz_string_t *string, sz_size_t length, sz_memory_allocator_t *allocator);

/**
 *  @brief  Doesn't change the contents or the length of the string, but grows the available memory capacity.
 *          This is beneficial, if several insertions are expected, and we want to minimize allocations.
 *
 *  @param string       String to grow.
 *  @param new_capacity The number of characters to reserve space for, including existing ones.
 *  @param allocator    Memory allocator to use for the allocation.
 *  @return             SZ_NULL if the operation failed, pointer to the new start of the string otherwise.
 */
SZ_PUBLIC sz_ptr_t sz_string_reserve(sz_string_t *string, sz_size_t new_capacity, sz_memory_allocator_t *allocator);

/**
 *  @brief  Grows the string by adding an uninitialized region of ::added_length at the given ::offset.
 *          Would often be used in conjunction with one or more `sz_copy` calls to populate the allocated region.
 *          Similar to `sz_string_reserve`, but changes the length of the ::string.
 *
 *  @param string       String to grow.
 *  @param offset       Offset of the first byte to reserve space for.
 *                      If provided offset is larger than the length, it will be capped.
 *  @param added_length The number of new characters to reserve space for.
 *  @param allocator    Memory allocator to use for the allocation.
 *  @return             SZ_NULL if the operation failed, pointer to the new start of the string otherwise.
 */
SZ_PUBLIC sz_ptr_t sz_string_expand(sz_string_t *string, sz_size_t offset, sz_size_t added_length,
                                    sz_memory_allocator_t *allocator);

/**
 *  @brief  Removes a range from a string. Changes the length, but not the capacity.
 *          Performs no allocations or deallocations and can't fail.
 *
 *  @param string       String to clean.
 *  @param offset       Offset of the first byte to remove.
 *  @param length       Number of bytes to remove. Out-of-bound ranges will be capped.
 *  @return             Number of bytes removed.
 */
SZ_PUBLIC sz_size_t sz_string_erase(sz_string_t *string, sz_size_t offset, sz_size_t length);

/**
 *  @brief  Shrinks the string to fit the current length, if it's allocated on the heap.
 *          Teh reverse operation of ::sz_string_reserve.
 *
 *  @param string       String to shrink.
 *  @param allocator    Memory allocator to use for the allocation.
 *  @return             Whether the operation was successful. The only failures can come from the allocator.
 */
SZ_PUBLIC sz_ptr_t sz_string_shrink_to_fit(sz_string_t *string, sz_memory_allocator_t *allocator);

/**
 *  @brief  Frees the string, if it's allocated on the heap.
 *          If the string is on the stack, the function clears/resets the state.
 */
SZ_PUBLIC void sz_string_free(sz_string_t *string, sz_memory_allocator_t *allocator);

#pragma endregion

#pragma region Fast Substring Search API

typedef sz_cptr_t (*sz_find_byte_t)(sz_cptr_t, sz_size_t, sz_cptr_t);
typedef sz_cptr_t (*sz_find_t)(sz_cptr_t, sz_size_t, sz_cptr_t, sz_size_t);
typedef sz_cptr_t (*sz_find_set_t)(sz_cptr_t, sz_size_t, sz_charset_t const *);

/**
 *  @brief  Locates first matching byte in a string. Equivalent to `memchr(haystack, *needle, h_length)` in LibC.
 *
 *  X86_64 implementation: https://github.com/lattera/glibc/blob/master/sysdeps/x86_64/memchr.S
 *  Aarch64 implementation: https://github.com/lattera/glibc/blob/master/sysdeps/aarch64/memchr.S
 *
 *  @param haystack Haystack - the string to search in.
 *  @param h_length Number of bytes in the haystack.
 *  @param needle   Needle - single-byte substring to find.
 *  @return         Address of the first match.
 */
SZ_DYNAMIC sz_cptr_t sz_find_byte(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle);

/** @copydoc sz_find_byte */
SZ_PUBLIC sz_cptr_t sz_find_byte_serial(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle);

/**
 *  @brief  Locates last matching byte in a string. Equivalent to `memrchr(haystack, *needle, h_length)` in LibC.
 *
 *  X86_64 implementation: https://github.com/lattera/glibc/blob/master/sysdeps/x86_64/memrchr.S
 *  Aarch64 implementation: missing
 *
 *  @param haystack Haystack - the string to search in.
 *  @param h_length Number of bytes in the haystack.
 *  @param needle   Needle - single-byte substring to find.
 *  @return         Address of the last match.
 */
SZ_DYNAMIC sz_cptr_t sz_rfind_byte(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle);

/** @copydoc sz_rfind_byte */
SZ_PUBLIC sz_cptr_t sz_rfind_byte_serial(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle);

/**
 *  @brief  Locates first matching substring.
 *          Equivalent to `memmem(haystack, h_length, needle, n_length)` in LibC.
 *          Similar to `strstr(haystack, needle)` in LibC, but requires known length.
 *
 *  @param haystack Haystack - the string to search in.
 *  @param h_length Number of bytes in the haystack.
 *  @param needle   Needle - substring to find.
 *  @param n_length Number of bytes in the needle.
 *  @return         Address of the first match.
 */
SZ_DYNAMIC sz_cptr_t sz_find(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle, sz_size_t n_length);

/** @copydoc sz_find */
SZ_PUBLIC sz_cptr_t sz_find_serial(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle, sz_size_t n_length);

/**
 *  @brief  Locates the last matching substring.
 *
 *  @param haystack Haystack - the string to search in.
 *  @param h_length Number of bytes in the haystack.
 *  @param needle   Needle - substring to find.
 *  @param n_length Number of bytes in the needle.
 *  @return         Address of the last match.
 */
SZ_DYNAMIC sz_cptr_t sz_rfind(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle, sz_size_t n_length);

/** @copydoc sz_rfind */
SZ_PUBLIC sz_cptr_t sz_rfind_serial(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle, sz_size_t n_length);

/**
 *  @brief  Finds the first character present from the ::set, present in ::text.
 *          Equivalent to `strspn(text, accepted)` and `strcspn(text, rejected)` in LibC.
 *          May have identical implementation and performance to ::sz_rfind_charset.
 *
 *  @param text     String to be trimmed.
 *  @param accepted Set of accepted characters.
 *  @return         Number of bytes forming the prefix.
 */
SZ_DYNAMIC sz_cptr_t sz_find_charset(sz_cptr_t text, sz_size_t length, sz_charset_t const *set);

/** @copydoc sz_find_charset */
SZ_PUBLIC sz_cptr_t sz_find_charset_serial(sz_cptr_t text, sz_size_t length, sz_charset_t const *set);

/**
 *  @brief  Finds the last character present from the ::set, present in ::text.
 *          Equivalent to `strspn(text, accepted)` and `strcspn(text, rejected)` in LibC.
 *          May have identical implementation and performance to ::sz_find_charset.
 *
 *  Useful for parsing, when we want to skip a set of characters. Examples:
 *  * 6 whitespaces: " \t\n\r\v\f".
 *  * 16 digits forming a float number: "0123456789,.eE+-".
 *  * 5 HTML reserved characters: "\"'&<>", of which "<>" can be useful for parsing.
 *  * 2 JSON string special characters useful to locate the end of the string: "\"\\".
 *
 *  @param text     String to be trimmed.
 *  @param rejected Set of rejected characters.
 *  @return         Number of bytes forming the prefix.
 */
SZ_DYNAMIC sz_cptr_t sz_rfind_charset(sz_cptr_t text, sz_size_t length, sz_charset_t const *set);

/** @copydoc sz_rfind_charset */
SZ_PUBLIC sz_cptr_t sz_rfind_charset_serial(sz_cptr_t text, sz_size_t length, sz_charset_t const *set);

#pragma endregion

#pragma region String Similarity Measures API

/**
 *  @brief  Computes the Hamming distance between two strings - number of not matching characters.
 *          Difference in length is is counted as a mismatch.
 *
 *  @param a        First string to compare.
 *  @param a_length Number of bytes in the first string.
 *  @param b        Second string to compare.
 *  @param b_length Number of bytes in the second string.
 *
 *  @param bound    Upper bound on the distance, that allows us to exit early.
 *                  If zero is passed, the maximum possible distance will be equal to the length of the longer input.
 *  @return         Unsigned integer for the distance, the `bound` if was exceeded.
 *
 *  @see    sz_hamming_distance_utf8
 *  @see    https://en.wikipedia.org/wiki/Hamming_distance
 */
SZ_DYNAMIC sz_size_t sz_hamming_distance(sz_cptr_t a, sz_size_t a_length, sz_cptr_t b, sz_size_t b_length,
                                         sz_size_t bound);

/** @copydoc sz_hamming_distance */
SZ_PUBLIC sz_size_t sz_hamming_distance_serial(sz_cptr_t a, sz_size_t a_length, sz_cptr_t b, sz_size_t b_length,
                                               sz_size_t bound);

/**
 *  @brief  Computes the Hamming distance between two @b UTF8 strings - number of not matching characters.
 *          Difference in length is is counted as a mismatch.
 *
 *  @param a        First string to compare.
 *  @param a_length Number of bytes in the first string.
 *  @param b        Second string to compare.
 *  @param b_length Number of bytes in the second string.
 *
 *  @param bound    Upper bound on the distance, that allows us to exit early.
 *                  If zero is passed, the maximum possible distance will be equal to the length of the longer input.
 *  @return         Unsigned integer for the distance, the `bound` if was exceeded.
 *
 *  @see    sz_hamming_distance
 *  @see    https://en.wikipedia.org/wiki/Hamming_distance
 */
SZ_DYNAMIC sz_size_t sz_hamming_distance_utf8(sz_cptr_t a, sz_size_t a_length, sz_cptr_t b, sz_size_t b_length,
                                              sz_size_t bound);

/** @copydoc sz_hamming_distance_utf8 */
SZ_PUBLIC sz_size_t sz_hamming_distance_utf8_serial(sz_cptr_t a, sz_size_t a_length, sz_cptr_t b, sz_size_t b_length,
                                                    sz_size_t bound);

typedef sz_size_t (*sz_hamming_distance_t)(sz_cptr_t, sz_size_t, sz_cptr_t, sz_size_t, sz_size_t);

/**
 *  @brief  Computes the Levenshtein edit-distance between two strings using the Wagner-Fisher algorithm.
 *          Similar to the Needleman-Wunsch alignment algorithm. Often used in fuzzy string matching.
 *
 *  @param a        First string to compare.
 *  @param a_length Number of bytes in the first string.
 *  @param b        Second string to compare.
 *  @param b_length Number of bytes in the second string.
 *
 *  @param alloc    Temporary memory allocator. Only some of the rows of the matrix will be allocated,
 *                  so the memory usage is linear in relation to ::a_length and ::b_length.
 *                  If SZ_NULL is passed, will initialize to the systems default `malloc`.
 *  @param bound    Upper bound on the distance, that allows us to exit early.
 *                  If zero is passed, the maximum possible distance will be equal to the length of the longer input.
 *  @return         Unsigned integer for edit distance, the `bound` if was exceeded or `SZ_SIZE_MAX`
 *                  if the memory allocation failed.
 *
 *  @see    sz_memory_allocator_init_fixed, sz_memory_allocator_init_default
 *  @see    https://en.wikipedia.org/wiki/Levenshtein_distance
 */
SZ_DYNAMIC sz_size_t sz_edit_distance(sz_cptr_t a, sz_size_t a_length, sz_cptr_t b, sz_size_t b_length, //
                                      sz_size_t bound, sz_memory_allocator_t *alloc);

/** @copydoc sz_edit_distance */
SZ_PUBLIC sz_size_t sz_edit_distance_serial(sz_cptr_t a, sz_size_t a_length, sz_cptr_t b, sz_size_t b_length, //
                                            sz_size_t bound, sz_memory_allocator_t *alloc);

/**
 *  @brief  Computes the Levenshtein edit-distance between two @b UTF8 strings.
 *          Unlike `sz_edit_distance`, reports the distance in Unicode codepoints, and not in bytes.
 *
 *  @param a        First string to compare.
 *  @param a_length Number of bytes in the first string.
 *  @param b        Second string to compare.
 *  @param b_length Number of bytes in the second string.
 *
 *  @param alloc    Temporary memory allocator. Only some of the rows of the matrix will be allocated,
 *                  so the memory usage is linear in relation to ::a_length and ::b_length.
 *                  If SZ_NULL is passed, will initialize to the systems default `malloc`.
 *  @param bound    Upper bound on the distance, that allows us to exit early.
 *                  If zero is passed, the maximum possible distance will be equal to the length of the longer input.
 *  @return         Unsigned integer for edit distance, the `bound` if was exceeded or `SZ_SIZE_MAX`
 *                  if the memory allocation failed.
 *
 *  @see    sz_memory_allocator_init_fixed, sz_memory_allocator_init_default, sz_edit_distance
 *  @see    https://en.wikipedia.org/wiki/Levenshtein_distance
 */
SZ_DYNAMIC sz_size_t sz_edit_distance_utf8(sz_cptr_t a, sz_size_t a_length, sz_cptr_t b, sz_size_t b_length, //
                                           sz_size_t bound, sz_memory_allocator_t *alloc);

typedef sz_size_t (*sz_edit_distance_t)(sz_cptr_t, sz_size_t, sz_cptr_t, sz_size_t, sz_size_t, sz_memory_allocator_t *);

/** @copydoc sz_edit_distance_utf8 */
SZ_PUBLIC sz_size_t sz_edit_distance_utf8_serial(sz_cptr_t a, sz_size_t a_length, sz_cptr_t b, sz_size_t b_length, //
                                                 sz_size_t bound, sz_memory_allocator_t *alloc);

/**
 *  @brief  Computes Needleman–Wunsch alignment score for two string. Often used in bioinformatics and cheminformatics.
 *          Similar to the Levenshtein edit-distance, parameterized for gap and substitution penalties.
 *
 *  Not commutative in the general case, as the order of the strings matters, as `sz_alignment_score(a, b)` may
 *  not be equal to `sz_alignment_score(b, a)`. Becomes @b commutative, if the substitution costs are symmetric.
 *  Equivalent to the negative Levenshtein distance, if: `gap == -1` and `subs[i][j] == (i == j ? 0: -1)`.
 *
 *  @param a        First string to compare.
 *  @param a_length Number of bytes in the first string.
 *  @param b        Second string to compare.
 *  @param b_length Number of bytes in the second string.
 *  @param gap      Penalty cost for gaps - insertions and removals.
 *  @param subs     Substitution costs matrix with 256 x 256 values for all pairs of characters.
 *
 *  @param alloc    Temporary memory allocator. Only some of the rows of the matrix will be allocated,
 *                  so the memory usage is linear in relation to ::a_length and ::b_length.
 *                  If SZ_NULL is passed, will initialize to the systems default `malloc`.
 *  @return         Signed similarity score. Can be negative, depending on the substitution costs.
 *                  If the memory allocation fails, the function returns `SZ_SSIZE_MAX`.
 *
 *  @see    sz_memory_allocator_init_fixed, sz_memory_allocator_init_default
 *  @see    https://en.wikipedia.org/wiki/Needleman%E2%80%93Wunsch_algorithm
 */
SZ_DYNAMIC sz_ssize_t sz_alignment_score(sz_cptr_t a, sz_size_t a_length, sz_cptr_t b, sz_size_t b_length, //
                                         sz_error_cost_t const *subs, sz_error_cost_t gap,                 //
                                         sz_memory_allocator_t *alloc);

/** @copydoc sz_alignment_score */
SZ_PUBLIC sz_ssize_t sz_alignment_score_serial(sz_cptr_t a, sz_size_t a_length, sz_cptr_t b, sz_size_t b_length, //
                                               sz_error_cost_t const *subs, sz_error_cost_t gap,                 //
                                               sz_memory_allocator_t *alloc);

typedef sz_ssize_t (*sz_alignment_score_t)(sz_cptr_t, sz_size_t, sz_cptr_t, sz_size_t, sz_error_cost_t const *,
                                           sz_error_cost_t, sz_memory_allocator_t *);

typedef void (*sz_hash_callback_t)(sz_cptr_t, sz_size_t, sz_u64_t, void *user);

/**
 *  @brief  Computes the Karp-Rabin rolling hashes of a string supplying them to the provided `callback`.
 *          Can be used for similarity scores, search, ranking, etc.
 *
 *  Rabin-Karp-like rolling hashes can have very high-level of collisions and depend
 *  on the choice of bases and the prime number. That's why, often two hashes from the same
 *  family are used with different bases.
 *
 *       1. Kernighan and Ritchie's function uses 31, a prime close to the size of English alphabet.
 *       2. To be friendlier to byte-arrays and UTF8, we use 257 for the second function.
 *
 *  Choosing the right ::window_length is task- and domain-dependant. For example, most English words are
 *  between 3 and 7 characters long, so a window of 4 bytes would be a good choice. For DNA sequences,
 *  the ::window_length might be a multiple of 3, as the codons are 3 (nucleotides) bytes long.
 *  With such minimalistic alphabets of just four characters (AGCT) longer windows might be needed.
 *  For protein sequences the alphabet is 20 characters long, so the window can be shorter, than for DNAs.
 *
 *  @param text             String to hash.
 *  @param length           Number of bytes in the string.
 *  @param window_length    Length of the rolling window in bytes.
 *  @param window_step      Step of reported hashes. @b Must be power of two. Should be smaller than `window_length`.
 *  @param callback         Function receiving the start & length of a substring, the hash, and the `callback_handle`.
 *  @param callback_handle  Optional user-provided pointer to be passed to the `callback`.
 *  @see                    sz_hashes_fingerprint, sz_hashes_intersection
 */
SZ_DYNAMIC void sz_hashes(sz_cptr_t text, sz_size_t length, sz_size_t window_length, sz_size_t window_step, //
                          sz_hash_callback_t callback, void *callback_handle);

/** @copydoc sz_hashes */
SZ_PUBLIC void sz_hashes_serial(sz_cptr_t text, sz_size_t length, sz_size_t window_length, sz_size_t window_step, //
                                sz_hash_callback_t callback, void *callback_handle);

typedef void (*sz_hashes_t)(sz_cptr_t, sz_size_t, sz_size_t, sz_size_t, sz_hash_callback_t, void *);

/**
 *  @brief  Computes the Karp-Rabin rolling hashes of a string outputting a binary fingerprint.
 *          Such fingerprints can be compared with Hamming or Jaccard (Tanimoto) distance for similarity.
 *
 *  The algorithm doesn't clear the fingerprint buffer on start, so it can be invoked multiple times
 *  to produce a fingerprint of a longer string, by passing the previous fingerprint as the ::fingerprint.
 *  It can also be reused to produce multi-resolution fingerprints by changing the ::window_length
 *  and calling the same function multiple times for the same input ::text.
 *
 *  Processes large strings in parts to maximize the cache utilization, using a small on-stack buffer,
 *  avoiding cache-coherency penalties of remote on-heap buffers.
 *
 *  @param text                 String to hash.
 *  @param length               Number of bytes in the string.
 *  @param fingerprint          Output fingerprint buffer.
 *  @param fingerprint_bytes    Number of bytes in the fingerprint buffer.
 *  @param window_length        Length of the rolling window in bytes.
 *  @see                        sz_hashes, sz_hashes_intersection
 */
SZ_PUBLIC void sz_hashes_fingerprint(sz_cptr_t text, sz_size_t length, sz_size_t window_length, //
                                     sz_ptr_t fingerprint, sz_size_t fingerprint_bytes);

typedef void (*sz_hashes_fingerprint_t)(sz_cptr_t, sz_size_t, sz_size_t, sz_ptr_t, sz_size_t);

/**
 *  @brief  Given a hash-fingerprint of a textual document, computes the number of intersecting hashes
 *          of the incoming document. Can be used for document scoring and search.
 *
 *  Processes large strings in parts to maximize the cache utilization, using a small on-stack buffer,
 *  avoiding cache-coherency penalties of remote on-heap buffers.
 *
 *  @param text                 Input document.
 *  @param length               Number of bytes in the input document.
 *  @param fingerprint          Reference document fingerprint.
 *  @param fingerprint_bytes    Number of bytes in the reference documents fingerprint.
 *  @param window_length        Length of the rolling window in bytes.
 *  @see                        sz_hashes, sz_hashes_fingerprint
 */
SZ_PUBLIC sz_size_t sz_hashes_intersection(sz_cptr_t text, sz_size_t length, sz_size_t window_length, //
                                           sz_cptr_t fingerprint, sz_size_t fingerprint_bytes);

typedef sz_size_t (*sz_hashes_intersection_t)(sz_cptr_t, sz_size_t, sz_size_t, sz_cptr_t, sz_size_t);

#pragma endregion

#pragma region Convenience API

/**
 *  @brief  Finds the first character in the haystack, that is present in the needle.
 *          Convenience function, reused across different language bindings.
 *  @see    sz_find_charset
 */
SZ_DYNAMIC sz_cptr_t sz_find_char_from(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n, sz_size_t n_length);

/**
 *  @brief  Finds the first character in the haystack, that is @b not present in the needle.
 *          Convenience function, reused across different language bindings.
 *  @see    sz_find_charset
 */
SZ_DYNAMIC sz_cptr_t sz_find_char_not_from(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n, sz_size_t n_length);

/**
 *  @brief  Finds the last character in the haystack, that is present in the needle.
 *          Convenience function, reused across different language bindings.
 *  @see    sz_find_charset
 */
SZ_DYNAMIC sz_cptr_t sz_rfind_char_from(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n, sz_size_t n_length);

/**
 *  @brief  Finds the last character in the haystack, that is @b not present in the needle.
 *          Convenience function, reused across different language bindings.
 *  @see    sz_find_charset
 */
SZ_DYNAMIC sz_cptr_t sz_rfind_char_not_from(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n, sz_size_t n_length);

#pragma endregion

#pragma region String Sequences API

struct sz_sequence_t;

typedef sz_cptr_t (*sz_sequence_member_start_t)(struct sz_sequence_t const *, sz_size_t);
typedef sz_size_t (*sz_sequence_member_length_t)(struct sz_sequence_t const *, sz_size_t);
typedef sz_bool_t (*sz_sequence_predicate_t)(struct sz_sequence_t const *, sz_size_t);
typedef sz_bool_t (*sz_sequence_comparator_t)(struct sz_sequence_t const *, sz_size_t, sz_size_t);
typedef sz_bool_t (*sz_string_is_less_t)(sz_cptr_t, sz_size_t, sz_cptr_t, sz_size_t);

typedef struct sz_sequence_t {
    sz_sorted_idx_t *order;
    sz_size_t count;
    sz_sequence_member_start_t get_start;
    sz_sequence_member_length_t get_length;
    void const *handle;
} sz_sequence_t;

/**
 *  @brief  Initiates the sequence structure from a tape layout, used by Apache Arrow.
 *          Expects ::offsets to contains `count + 1` entries, the last pointing at the end
 *          of the last string, indicating the total length of the ::tape.
 */
SZ_PUBLIC void sz_sequence_from_u32tape(sz_cptr_t *start, sz_u32_t const *offsets, sz_size_t count,
                                        sz_sequence_t *sequence);

/**
 *  @brief  Initiates the sequence structure from a tape layout, used by Apache Arrow.
 *          Expects ::offsets to contains `count + 1` entries, the last pointing at the end
 *          of the last string, indicating the total length of the ::tape.
 */
SZ_PUBLIC void sz_sequence_from_u64tape(sz_cptr_t *start, sz_u64_t const *offsets, sz_size_t count,
                                        sz_sequence_t *sequence);

/**
 *  @brief  Similar to `std::partition`, given a predicate splits the sequence into two parts.
 *          The algorithm is unstable, meaning that elements may change relative order, as long
 *          as they are in the right partition. This is the simpler algorithm for partitioning.
 */
SZ_PUBLIC sz_size_t sz_partition(sz_sequence_t *sequence, sz_sequence_predicate_t predicate);

/**
 *  @brief  Inplace `std::set_union` for two consecutive chunks forming the same continuous `sequence`.
 *
 *  @param partition The number of elements in the first sub-sequence in `sequence`.
 *  @param less Comparison function, to determine the lexicographic ordering.
 */
SZ_PUBLIC void sz_merge(sz_sequence_t *sequence, sz_size_t partition, sz_sequence_comparator_t less);

/**
 *  @brief  Sorting algorithm, combining Radix Sort for the first 32 bits of every word
 *          and a follow-up by a more conventional sorting procedure on equally prefixed parts.
 */
SZ_PUBLIC void sz_sort(sz_sequence_t *sequence);

/**
 *  @brief  Partial sorting algorithm, combining Radix Sort for the first 32 bits of every word
 *          and a follow-up by a more conventional sorting procedure on equally prefixed parts.
 */
SZ_PUBLIC void sz_sort_partial(sz_sequence_t *sequence, sz_size_t n);

/**
 *  @brief  Intro-Sort algorithm that supports custom comparators.
 */
SZ_PUBLIC void sz_sort_intro(sz_sequence_t *sequence, sz_sequence_comparator_t less);

#pragma endregion

/*
 *  Hardware feature detection.
 *  All of those can be controlled by the user.
 */
#ifndef SZ_USE_X86_AVX512
#ifdef __AVX512BW__
#define SZ_USE_X86_AVX512 1
#else
#define SZ_USE_X86_AVX512 0
#endif
#endif

#ifndef SZ_USE_X86_AVX2
#ifdef __AVX2__
#define SZ_USE_X86_AVX2 1
#else
#define SZ_USE_X86_AVX2 0
#endif
#endif

#ifndef SZ_USE_ARM_NEON
#ifdef __ARM_NEON
#define SZ_USE_ARM_NEON 1
#else
#define SZ_USE_ARM_NEON 0
#endif
#endif

#ifndef SZ_USE_ARM_SVE
#ifdef __ARM_FEATURE_SVE
#define SZ_USE_ARM_SVE 1
#else
#define SZ_USE_ARM_SVE 0
#endif
#endif

/*
 *  Include hardware-specific headers.
 */
#if SZ_USE_X86_AVX512 || SZ_USE_X86_AVX2
#include <immintrin.h>
#endif // SZ_USE_X86...
#if SZ_USE_ARM_NEON
#include <arm_acle.h>
#include <arm_neon.h>
#endif // SZ_USE_ARM_NEON
#if SZ_USE_ARM_SVE
#include <arm_sve.h>
#endif // SZ_USE_ARM_SVE

#pragma region Hardware-Specific API

#if SZ_USE_X86_AVX512

/** @copydoc sz_equal_serial */
SZ_PUBLIC sz_bool_t sz_equal_avx512(sz_cptr_t a, sz_cptr_t b, sz_size_t length);
/** @copydoc sz_order_serial */
SZ_PUBLIC sz_ordering_t sz_order_avx512(sz_cptr_t a, sz_size_t a_length, sz_cptr_t b, sz_size_t b_length);
/** @copydoc sz_copy_serial */
SZ_PUBLIC void sz_copy_avx512(sz_ptr_t target, sz_cptr_t source, sz_size_t length);
/** @copydoc sz_move_serial */
SZ_PUBLIC void sz_move_avx512(sz_ptr_t target, sz_cptr_t source, sz_size_t length);
/** @copydoc sz_fill_serial */
SZ_PUBLIC void sz_fill_avx512(sz_ptr_t target, sz_size_t length, sz_u8_t value);
/** @copydoc sz_find_byte */
SZ_PUBLIC sz_cptr_t sz_find_byte_avx512(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle);
/** @copydoc sz_rfind_byte */
SZ_PUBLIC sz_cptr_t sz_rfind_byte_avx512(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle);
/** @copydoc sz_find */
SZ_PUBLIC sz_cptr_t sz_find_avx512(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle, sz_size_t n_length);
/** @copydoc sz_rfind */
SZ_PUBLIC sz_cptr_t sz_rfind_avx512(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle, sz_size_t n_length);
/** @copydoc sz_find_charset */
SZ_PUBLIC sz_cptr_t sz_find_charset_avx512(sz_cptr_t text, sz_size_t length, sz_charset_t const *set);
/** @copydoc sz_rfind_charset */
SZ_PUBLIC sz_cptr_t sz_rfind_charset_avx512(sz_cptr_t text, sz_size_t length, sz_charset_t const *set);
/** @copydoc sz_edit_distance */
SZ_PUBLIC sz_size_t sz_edit_distance_avx512(sz_cptr_t a, sz_size_t a_length, sz_cptr_t b, sz_size_t b_length, //
                                            sz_size_t bound, sz_memory_allocator_t *alloc);
/** @copydoc sz_alignment_score */
SZ_PUBLIC sz_ssize_t sz_alignment_score_avx512(sz_cptr_t a, sz_size_t a_length, sz_cptr_t b, sz_size_t b_length, //
                                               sz_error_cost_t const *subs, sz_error_cost_t gap,                 //
                                               sz_memory_allocator_t *alloc);
/** @copydoc sz_hashes */
SZ_PUBLIC void sz_hashes_avx512(sz_cptr_t text, sz_size_t length, sz_size_t window_length, sz_size_t step, //
                                sz_hash_callback_t callback, void *callback_handle);
#endif

#if SZ_USE_X86_AVX2
/** @copydoc sz_equal */
SZ_PUBLIC void sz_copy_avx2(sz_ptr_t target, sz_cptr_t source, sz_size_t length);
/** @copydoc sz_move */
SZ_PUBLIC void sz_move_avx2(sz_ptr_t target, sz_cptr_t source, sz_size_t length);
/** @copydoc sz_fill */
SZ_PUBLIC void sz_fill_avx2(sz_ptr_t target, sz_size_t length, sz_u8_t value);
/** @copydoc sz_find_byte */
SZ_PUBLIC sz_cptr_t sz_find_byte_avx2(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle);
/** @copydoc sz_rfind_byte */
SZ_PUBLIC sz_cptr_t sz_rfind_byte_avx2(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle);
/** @copydoc sz_find */
SZ_PUBLIC sz_cptr_t sz_find_avx2(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle, sz_size_t n_length);
/** @copydoc sz_rfind */
SZ_PUBLIC sz_cptr_t sz_rfind_avx2(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle, sz_size_t n_length);
/** @copydoc sz_hashes */
SZ_PUBLIC void sz_hashes_avx2(sz_cptr_t text, sz_size_t length, sz_size_t window_length, sz_size_t step, //
                              sz_hash_callback_t callback, void *callback_handle);
#endif

#if SZ_USE_ARM_NEON
/** @copydoc sz_equal */
SZ_PUBLIC sz_bool_t sz_equal_neon(sz_cptr_t a, sz_cptr_t b, sz_size_t length);
/** @copydoc sz_find_byte */
SZ_PUBLIC sz_cptr_t sz_find_byte_neon(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle);
/** @copydoc sz_rfind_byte */
SZ_PUBLIC sz_cptr_t sz_rfind_byte_neon(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle);
/** @copydoc sz_find */
SZ_PUBLIC sz_cptr_t sz_find_neon(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle, sz_size_t n_length);
/** @copydoc sz_rfind */
SZ_PUBLIC sz_cptr_t sz_rfind_neon(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle, sz_size_t n_length);
/** @copydoc sz_find_charset */
SZ_PUBLIC sz_cptr_t sz_find_charset_neon(sz_cptr_t text, sz_size_t length, sz_charset_t const *set);
/** @copydoc sz_rfind_charset */
SZ_PUBLIC sz_cptr_t sz_rfind_charset_neon(sz_cptr_t text, sz_size_t length, sz_charset_t const *set);
#endif

#pragma endregion

#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wconversion"

/*
 **********************************************************************************************************************
 **********************************************************************************************************************
 **********************************************************************************************************************
 *
 *  This is where we the actual implementation begins.
 *  The rest of the file is hidden from the public API.
 *
 **********************************************************************************************************************
 **********************************************************************************************************************
 **********************************************************************************************************************
 */

#pragma region Compiler Extensions and Helper Functions

#pragma GCC visibility push(hidden)

/**
 *  @brief  Helper-macro to mark potentially unused variables.
 */
#define sz_unused(x) ((void)(x))

/**
 *  @brief  Helper-macro casting a variable to another type of the same size.
 */
#define sz_bitcast(type, value) (*((type *)&(value)))

/**
 *  @brief  Defines `SZ_NULL`, analogous to `NULL`.
 *          The default often comes from locale.h, stddef.h,
 *          stdio.h, stdlib.h, string.h, time.h, or wchar.h.
 */
#ifdef __GNUG__
#define SZ_NULL __null
#define SZ_NULL_CHAR __null
#else
#define SZ_NULL ((void *)0)
#define SZ_NULL_CHAR ((char *)0)
#endif

/**
 *  @brief  Cache-line width, that will affect the execution of some algorithms,
 *          like equality checks and relative order computing.
 */
#define SZ_CACHE_LINE_WIDTH (64) // bytes

/**
 *  @brief  Similar to `assert`, the `sz_assert` is used in the SZ_DEBUG mode
 *          to check the invariants of the library. It's a no-op in the SZ_RELEASE mode.
 *  @note   If you want to catch it, put a breakpoint at @b `__GI_exit`
 */
#if SZ_DEBUG
#include <stdio.h>  // `fprintf`
#include <stdlib.h> // `EXIT_FAILURE`
#define sz_assert(condition)                                                                                \
    do {                                                                                                    \
        if (!(condition)) {                                                                                 \
            fprintf(stderr, "Assertion failed: %s, in file %s, line %d\n", #condition, __FILE__, __LINE__); \
            exit(EXIT_FAILURE);                                                                             \
        }                                                                                                   \
    } while (0)
#else
#define sz_assert(condition) ((void)0)
#endif

/*  Intrinsics aliases for MSVC, GCC, Clang, and Clang-Cl.
 *  The following section of compiler intrinsics comes in 2 flavors.
 */
#if defined(_MSC_VER) && !defined(__clang__) // On Clang-CL
#include <intrin.h>

// Sadly, when building Win32 images, we can't use the `_tzcnt_u64`, `_lzcnt_u64`,
// `_BitScanForward64`, or `_BitScanReverse64` intrinsics. For now it's a simple `for`-loop.
// In the future we can switch to a more efficient De Bruijn's algorithm.
// https://www.chessprogramming.org/BitScan
// https://www.chessprogramming.org/De_Bruijn_Sequence
// https://gist.github.com/resilar/e722d4600dbec9752771ab4c9d47044f
//
// Use the serial version on 32-bit x86 and on Arm.
#if (defined(_WIN32) && !defined(_WIN64)) || defined(_M_ARM) || defined(_M_ARM64)
SZ_INTERNAL int sz_u64_ctz(sz_u64_t x) {
    sz_assert(x != 0);
    int n = 0;
    while ((x & 1) == 0) { n++, x >>= 1; }
    return n;
}
SZ_INTERNAL int sz_u64_clz(sz_u64_t x) {
    sz_assert(x != 0);
    int n = 0;
    while ((x & 0x8000000000000000ULL) == 0) { n++, x <<= 1; }
    return n;
}
SZ_INTERNAL int sz_u64_popcount(sz_u64_t x) {
    x = x - ((x >> 1) & 0x5555555555555555);
    x = (x & 0x3333333333333333) + ((x >> 2) & 0x3333333333333333);
    return (((x + (x >> 4)) & 0x0F0F0F0F0F0F0F0F) * 0x0101010101010101) >> 56;
}
SZ_INTERNAL int sz_u32_popcount(sz_u32_t x) {
    x = x - ((x >> 1) & 0x55555555);
    x = (x & 0x33333333) + ((x >> 2) & 0x33333333);
    return (((x + (x >> 4)) & 0x0F0F0F0F) * 0x01010101) >> 24;
}
#else
SZ_INTERNAL int sz_u64_ctz(sz_u64_t x) { return _tzcnt_u64(x); }
SZ_INTERNAL int sz_u64_clz(sz_u64_t x) { return _lzcnt_u64(x); }
SZ_INTERNAL int sz_u64_popcount(sz_u64_t x) { return __popcnt64(x); }
SZ_INTERNAL int sz_u32_popcount(sz_u32_t x) { return __popcnt(x); }
#endif
SZ_INTERNAL int sz_u32_ctz(sz_u32_t x) { return _tzcnt_u32(x); }
SZ_INTERNAL int sz_u32_clz(sz_u32_t x) { return _lzcnt_u32(x); }
SZ_INTERNAL sz_u64_t sz_u64_bytes_reverse(sz_u64_t val) { return _byteswap_uint64(val); }
SZ_INTERNAL sz_u32_t sz_u32_bytes_reverse(sz_u32_t val) { return _byteswap_ulong(val); }
#else
SZ_INTERNAL int sz_u64_popcount(sz_u64_t x) { return __builtin_popcountll(x); }
SZ_INTERNAL int sz_u32_popcount(sz_u32_t x) { return __builtin_popcount(x); }
SZ_INTERNAL int sz_u64_ctz(sz_u64_t x) { return __builtin_ctzll(x); }
SZ_INTERNAL int sz_u64_clz(sz_u64_t x) { return __builtin_clzll(x); }
SZ_INTERNAL int sz_u32_ctz(sz_u32_t x) { return __builtin_ctz(x); } // ! Undefined if `x == 0`
SZ_INTERNAL int sz_u32_clz(sz_u32_t x) { return __builtin_clz(x); } // ! Undefined if `x == 0`
SZ_INTERNAL sz_u64_t sz_u64_bytes_reverse(sz_u64_t val) { return __builtin_bswap64(val); }
SZ_INTERNAL sz_u32_t sz_u32_bytes_reverse(sz_u32_t val) { return __builtin_bswap32(val); }
#endif

SZ_INTERNAL sz_u64_t sz_u64_rotl(sz_u64_t x, sz_u64_t r) { return (x << r) | (x >> (64 - r)); }

/**
 *  @brief  Select bits from either ::a or ::b depending on the value of ::mask bits.
 *
 *  Similar to `_mm_blend_epi16` intrinsic on x86.
 *  Described in the "Bit Twiddling Hacks" by Sean Eron Anderson.
 *  https://graphics.stanford.edu/~seander/bithacks.html#ConditionalSetOrClearBitsWithoutBranching
 */
SZ_INTERNAL sz_u64_t sz_u64_blend(sz_u64_t a, sz_u64_t b, sz_u64_t mask) { return a ^ ((a ^ b) & mask); }

/*
 *  Efficiently computing the minimum and maximum of two or three values can be tricky.
 *  The simple branching baseline would be:
 *
 *      x < y ? x : y                               // can replace with 1 conditional move
 *
 *  Branchless approach is well known for signed integers, but it doesn't apply to unsigned ones.
 *  https://stackoverflow.com/questions/514435/templatized-branchless-int-max-min-function
 *  https://graphics.stanford.edu/~seander/bithacks.html#IntegerMinOrMax
 *  Using only bit-shifts for singed integers it would be:
 *
 *      y + ((x - y) & (x - y) >> 31)               // 4 unique operations
 *
 *  Alternatively, for any integers using multiplication:
 *
 *      (x > y) * y + (x <= y) * x                  // 5 operations
 *
 *  Alternatively, to avoid multiplication:
 *
 *      x & ~((x < y) - 1) + y & ((x < y) - 1)      // 6 unique operations
 */
#define sz_min_of_two(x, y) (x < y ? x : y)
#define sz_max_of_two(x, y) (x < y ? y : x)
#define sz_min_of_three(x, y, z) sz_min_of_two(x, sz_min_of_two(y, z))
#define sz_max_of_three(x, y, z) sz_max_of_two(x, sz_max_of_two(y, z))

/** @brief  Branchless minimum function for two signed 32-bit integers. */
SZ_INTERNAL sz_i32_t sz_i32_min_of_two(sz_i32_t x, sz_i32_t y) { return y + ((x - y) & (x - y) >> 31); }

/** @brief  Branchless minimum function for two signed 32-bit integers. */
SZ_INTERNAL sz_i32_t sz_i32_max_of_two(sz_i32_t x, sz_i32_t y) { return x - ((x - y) & (x - y) >> 31); }

/**
 *  @brief  Clamps signed offsets in a string to a valid range. Used for Pythonic-style slicing.
 */
SZ_INTERNAL void sz_ssize_clamp_interval(sz_size_t length, sz_ssize_t start, sz_ssize_t end,
                                         sz_size_t *normalized_offset, sz_size_t *normalized_length) {
    // TODO: Remove branches.
    // Normalize negative indices
    if (start < 0) start += length;
    if (end < 0) end += length;

    // Clamp indices to a valid range
    if (start < 0) start = 0;
    if (end < 0) end = 0;
    if (start > (sz_ssize_t)length) start = length;
    if (end > (sz_ssize_t)length) end = length;

    // Ensure start <= end
    if (start > end) start = end;

    *normalized_offset = start;
    *normalized_length = end - start;
}

/**
 *  @brief  Compute the logarithm base 2 of a positive integer, rounding down.
 */
SZ_INTERNAL sz_size_t sz_size_log2i_nonzero(sz_size_t x) {
    sz_assert(x > 0 && "Non-positive numbers have no defined logarithm");
    sz_size_t leading_zeros = sz_u64_clz(x);
    return 63 - leading_zeros;
}

/**
 *  @brief  Compute the smallest power of two greater than or equal to ::x.
 */
SZ_INTERNAL sz_size_t sz_size_bit_ceil(sz_size_t x) {
    // Unlike the commonly used trick with `clz` intrinsics, is valid across the whole range of `x`.
    // https://stackoverflow.com/a/10143264
    x--;
    x |= x >> 1;
    x |= x >> 2;
    x |= x >> 4;
    x |= x >> 8;
    x |= x >> 16;
#if SZ_DETECT_64_BIT
    x |= x >> 32;
#endif
    x++;
    return x;
}

/**
 *  @brief  Transposes an 8x8 bit matrix packed in a `sz_u64_t`.
 *
 *  There is a well known SWAR sequence for that known to chess programmers,
 *  willing to flip a bit-matrix of pieces along the main A1-H8 diagonal.
 *  https://www.chessprogramming.org/Flipping_Mirroring_and_Rotating
 *  https://lukas-prokop.at/articles/2021-07-23-transpose
 */
SZ_INTERNAL sz_u64_t sz_u64_transpose(sz_u64_t x) {
    sz_u64_t t;
    t = x ^ (x << 36);
    x ^= 0xf0f0f0f00f0f0f0full & (t ^ (x >> 36));
    t = 0xcccc0000cccc0000ull & (x ^ (x << 18));
    x ^= t ^ (t >> 18);
    t = 0xaa00aa00aa00aa00ull & (x ^ (x << 9));
    x ^= t ^ (t >> 9);
    return x;
}

/**
 *  @brief  Helper, that swaps two 64-bit integers representing the order of elements in the sequence.
 */
SZ_INTERNAL void sz_u64_swap(sz_u64_t *a, sz_u64_t *b) {
    sz_u64_t t = *a;
    *a = *b;
    *b = t;
}

/**
 *  @brief  Helper, that swaps two 64-bit integers representing the order of elements in the sequence.
 */
SZ_INTERNAL void sz_pointer_swap(void **a, void **b) {
    void *t = *a;
    *a = *b;
    *b = t;
}

/**
 *  @brief  Helper structure to simplify work with 16-bit words.
 *  @see    sz_u16_load
 */
typedef union sz_u16_vec_t {
    sz_u16_t u16;
    sz_u8_t u8s[2];
} sz_u16_vec_t;

/**
 *  @brief Load a 16-bit unsigned integer from a potentially unaligned pointer, can be expensive on some platforms.
 */
SZ_INTERNAL sz_u16_vec_t sz_u16_load(sz_cptr_t ptr) {
#if !SZ_USE_MISALIGNED_LOADS
    sz_u16_vec_t result;
    result.u8s[0] = ptr[0];
    result.u8s[1] = ptr[1];
    return result;
#elif defined(_MSC_VER) && !defined(__clang__)
#if defined(_M_IX86) //< The __unaligned modifier isn't valid for the x86 platform.
    return *((sz_u16_vec_t *)ptr);
#else
    return *((__unaligned sz_u16_vec_t *)ptr);
#endif
#else
    __attribute__((aligned(1))) sz_u16_vec_t const *result = (sz_u16_vec_t const *)ptr;
    return *result;
#endif
}

/**
 *  @brief  Helper structure to simplify work with 32-bit words.
 *  @see    sz_u32_load
 */
typedef union sz_u32_vec_t {
    sz_u32_t u32;
    sz_u16_t u16s[2];
    sz_u8_t u8s[4];
} sz_u32_vec_t;

/**
 *  @brief Load a 32-bit unsigned integer from a potentially unaligned pointer, can be expensive on some platforms.
 */
SZ_INTERNAL sz_u32_vec_t sz_u32_load(sz_cptr_t ptr) {
#if !SZ_USE_MISALIGNED_LOADS
    sz_u32_vec_t result;
    result.u8s[0] = ptr[0];
    result.u8s[1] = ptr[1];
    result.u8s[2] = ptr[2];
    result.u8s[3] = ptr[3];
    return result;
#elif defined(_MSC_VER) && !defined(__clang__)
#if defined(_M_IX86) //< The __unaligned modifier isn't valid for the x86 platform.
    return *((sz_u32_vec_t *)ptr);
#else
    return *((__unaligned sz_u32_vec_t *)ptr);
#endif
#else
    __attribute__((aligned(1))) sz_u32_vec_t const *result = (sz_u32_vec_t const *)ptr;
    return *result;
#endif
}

/**
 *  @brief  Helper structure to simplify work with 64-bit words.
 *  @see    sz_u64_load
 */
typedef union sz_u64_vec_t {
    sz_u64_t u64;
    sz_u32_t u32s[2];
    sz_u16_t u16s[4];
    sz_u8_t u8s[8];
} sz_u64_vec_t;

/**
 *  @brief Load a 64-bit unsigned integer from a potentially unaligned pointer, can be expensive on some platforms.
 */
SZ_INTERNAL sz_u64_vec_t sz_u64_load(sz_cptr_t ptr) {
#if !SZ_USE_MISALIGNED_LOADS
    sz_u64_vec_t result;
    result.u8s[0] = ptr[0];
    result.u8s[1] = ptr[1];
    result.u8s[2] = ptr[2];
    result.u8s[3] = ptr[3];
    result.u8s[4] = ptr[4];
    result.u8s[5] = ptr[5];
    result.u8s[6] = ptr[6];
    result.u8s[7] = ptr[7];
    return result;
#elif defined(_MSC_VER) && !defined(__clang__)
#if defined(_M_IX86) //< The __unaligned modifier isn't valid for the x86 platform.
    return *((sz_u64_vec_t *)ptr);
#else
    return *((__unaligned sz_u64_vec_t *)ptr);
#endif
#else
    __attribute__((aligned(1))) sz_u64_vec_t const *result = (sz_u64_vec_t const *)ptr;
    return *result;
#endif
}

/** @brief  Helper function, using the supplied fixed-capacity buffer to allocate memory. */
SZ_INTERNAL sz_ptr_t _sz_memory_allocate_fixed(sz_size_t length, void *handle) {
    sz_size_t capacity;
    sz_copy((sz_ptr_t)&capacity, (sz_cptr_t)handle, sizeof(sz_size_t));
    sz_size_t consumed_capacity = sizeof(sz_size_t);
    if (consumed_capacity + length > capacity) return SZ_NULL_CHAR;
    return (sz_ptr_t)handle + consumed_capacity;
}

/** @brief  Helper "no-op" function, simulating memory deallocation when we use a "static" memory buffer. */
SZ_INTERNAL void _sz_memory_free_fixed(sz_ptr_t start, sz_size_t length, void *handle) {
    sz_unused(start && length && handle);
}

/** @brief  An internal callback used to set a bit in a power-of-two length binary fingerprint of a string. */
SZ_INTERNAL void _sz_hashes_fingerprint_pow2_callback(sz_cptr_t start, sz_size_t length, sz_u64_t hash, void *handle) {
    sz_string_view_t *fingerprint_buffer = (sz_string_view_t *)handle;
    sz_u8_t *fingerprint_u8s = (sz_u8_t *)fingerprint_buffer->start;
    sz_size_t fingerprint_bytes = fingerprint_buffer->length;
    fingerprint_u8s[(hash / 8) & (fingerprint_bytes - 1)] |= (1 << (hash & 7));
    sz_unused(start && length);
}

/** @brief  An internal callback used to set a bit in a @b non power-of-two length binary fingerprint of a string. */
SZ_INTERNAL void _sz_hashes_fingerprint_non_pow2_callback(sz_cptr_t start, sz_size_t length, sz_u64_t hash,
                                                          void *handle) {
    sz_string_view_t *fingerprint_buffer = (sz_string_view_t *)handle;
    sz_u8_t *fingerprint_u8s = (sz_u8_t *)fingerprint_buffer->start;
    sz_size_t fingerprint_bytes = fingerprint_buffer->length;
    fingerprint_u8s[(hash / 8) % fingerprint_bytes] |= (1 << (hash & 7));
    sz_unused(start && length);
}

/** @brief  An internal callback, used to mix all the running hashes into one pointer-size value. */
SZ_INTERNAL void _sz_hashes_fingerprint_scalar_callback(sz_cptr_t start, sz_size_t length, sz_u64_t hash,
                                                        void *scalar_handle) {
    sz_unused(start && length && hash && scalar_handle);
    sz_size_t *scalar_ptr = (sz_size_t *)scalar_handle;
    *scalar_ptr ^= hash;
}

/**
 *  @brief  Chooses the offsets of the most interesting characters in a search needle.
 *
 *  Search throughput can significantly deteriorate if we are matching the wrong characters.
 *  Say the needle is "aXaYa", and we are comparing the first, second, and last character.
 *  If we use SIMD and compare many offsets at a time, comparing against "a" in every register is a waste.
 *
 *  Similarly, dealing with UTF8 inputs, we know that the lower bits of each character code carry more information.
 *  Cyrillic alphabet, for example, falls into [0x0410, 0x042F] code range for uppercase [А, Я], and
 *  into [0x0430, 0x044F] for lowercase [а, я]. Scanning through a text written in Russian, half of the
 *  bytes will carry absolutely no value and will be equal to 0x04.
 */
SZ_INTERNAL void _sz_locate_needle_anomalies(sz_cptr_t start, sz_size_t length, //
                                             sz_size_t *first, sz_size_t *second, sz_size_t *third) {
    *first = 0;
    *second = length / 2;
    *third = length - 1;

    //
    int has_duplicates =                   //
        start[*first] == start[*second] || //
        start[*first] == start[*third] ||  //
        start[*second] == start[*third];

    // Loop through letters to find non-colliding variants.
    if (length > 3 && has_duplicates) {
        // Pivot the middle point right, until we find a character different from the first one.
        for (; start[*second] == start[*first] && *second + 1 < *third; ++(*second)) {}
        // Pivot the third (last) point left, until we find a different character.
        for (; (start[*third] == start[*second] || start[*third] == start[*first]) && *third > (*second + 1);
             --(*third)) {}
    }

    // TODO: Investigate alternative strategies for long needles.
    // On very long needles we have the luxury to choose!
    // Often dealing with UTF8, we will likely benfit from shifting the first and second characters
    // further to the right, to achieve not only uniqness within the needle, but also avoid common
    // rune prefixes of 2-, 3-, and 4-byte codes.
    if (length > 8) {
        // Pivot the first and second points right, until we find a character, that:
        // > is different from others.
        // > doesn't start with 0b'110x'xxxx - only 5 bits of relevant info.
        // > doesn't start with 0b'1110'xxxx - only 4 bits of relevant info.
        // > doesn't start with 0b'1111'0xxx - only 3 bits of relevant info.
        //
        // So we are practically searching for byte values that start with 0b0xxx'xxxx or 0b'10xx'xxxx.
        // Meaning they fall in the range [0, 127] and [128, 191], in other words any unsigned int up to 191.
        sz_u8_t const *start_u8 = (sz_u8_t const *)start;
        sz_size_t vibrant_first = *first, vibrant_second = *second, vibrant_third = *third;

        // Let's begin with the seccond character, as the termination criterea there is more obvious
        // and we may end up with more variants to check for the first candidate.
        for (; (start_u8[vibrant_second] > 191 || start_u8[vibrant_second] == start_u8[vibrant_third]) &&
               (vibrant_second + 1 < vibrant_third);
             ++vibrant_second) {}

        // Now check if we've indeed found a good candidate or should revert the `vibrant_second` to `second`.
        if (start_u8[vibrant_second] < 191) { *second = vibrant_second; }
        else { vibrant_second = *second; }

        // Now check the first character.
        for (; (start_u8[vibrant_first] > 191 || start_u8[vibrant_first] == start_u8[vibrant_second] ||
                start_u8[vibrant_first] == start_u8[vibrant_third]) &&
               (vibrant_first + 1 < vibrant_second);
             ++vibrant_first) {}

        // Now check if we've indeed found a good candidate or should revert the `vibrant_first` to `first`.
        // We don't need to shift the third one when dealing with texts as the last byte of the text is
        // also the last byte of a rune and contains the most information.
        if (start_u8[vibrant_first] < 191) { *first = vibrant_first; }
    }
}

#pragma GCC visibility pop
#pragma endregion

#pragma region Serial Implementation

#if !SZ_AVOID_LIBC
#include <stdio.h>  // `fprintf`
#include <stdlib.h> // `malloc`, `EXIT_FAILURE`
#endif

SZ_PUBLIC void sz_memory_allocator_init_default(sz_memory_allocator_t *alloc) {
#if !SZ_AVOID_LIBC
    alloc->allocate = (sz_memory_allocate_t)malloc;
    alloc->free = (sz_memory_free_t)free;
#else
    alloc->allocate = (sz_memory_allocate_t)SZ_NULL;
    alloc->free = (sz_memory_free_t)SZ_NULL;
#endif
    alloc->handle = SZ_NULL;
}

SZ_PUBLIC void sz_memory_allocator_init_fixed(sz_memory_allocator_t *alloc, void *buffer, sz_size_t length) {
    // The logic here is simple - put the buffer length in the first slots of the buffer.
    // Later use it for bounds checking.
    alloc->allocate = (sz_memory_allocate_t)_sz_memory_allocate_fixed;
    alloc->free = (sz_memory_free_t)_sz_memory_free_fixed;
    alloc->handle = &buffer;
    sz_copy((sz_ptr_t)buffer, (sz_cptr_t)&length, sizeof(sz_size_t));
}

/**
 *  @brief  Byte-level equality comparison between two strings.
 *          If unaligned loads are allowed, uses a switch-table to avoid loops on short strings.
 */
SZ_PUBLIC sz_bool_t sz_equal_serial(sz_cptr_t a, sz_cptr_t b, sz_size_t length) {
    sz_cptr_t const a_end = a + length;
#if SZ_USE_MISALIGNED_LOADS
    if (length >= SZ_SWAR_THRESHOLD) {
        sz_u64_vec_t a_vec, b_vec;
        for (; a + 8 <= a_end; a += 8, b += 8) {
            a_vec = sz_u64_load(a);
            b_vec = sz_u64_load(b);
            if (a_vec.u64 != b_vec.u64) return sz_false_k;
        }
    }
#endif
    while (a != a_end && *a == *b) a++, b++;
    return (sz_bool_t)(a_end == a);
}

SZ_PUBLIC sz_cptr_t sz_find_charset_serial(sz_cptr_t text, sz_size_t length, sz_charset_t const *set) {
    for (sz_cptr_t const end = text + length; text != end; ++text)
        if (sz_charset_contains(set, *text)) return text;
    return SZ_NULL_CHAR;
}

SZ_PUBLIC sz_cptr_t sz_rfind_charset_serial(sz_cptr_t text, sz_size_t length, sz_charset_t const *set) {
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Warray-bounds"
    sz_cptr_t const end = text;
    for (text += length; text != end;)
        if (sz_charset_contains(set, *(text -= 1))) return text;
    return SZ_NULL_CHAR;
#pragma GCC diagnostic pop
}

/**
 *  One option to avoid branching is to use conditional moves and lookup the comparison result in a table:
 *       sz_ordering_t ordering_lookup[2] = {sz_greater_k, sz_less_k};
 *       for (; a != min_end; ++a, ++b)
 *           if (*a != *b) return ordering_lookup[*a < *b];
 *  That, however, introduces a data-dependency.
 *  A cleaner option is to perform two comparisons and a subtraction.
 *  One instruction more, but no data-dependency.
 */
#define _sz_order_scalars(a, b) ((sz_ordering_t)((a > b) - (a < b)))

SZ_PUBLIC sz_ordering_t sz_order_serial(sz_cptr_t a, sz_size_t a_length, sz_cptr_t b, sz_size_t b_length) {
    sz_bool_t a_shorter = (sz_bool_t)(a_length < b_length);
    sz_size_t min_length = a_shorter ? a_length : b_length;
    sz_cptr_t min_end = a + min_length;
#if SZ_USE_MISALIGNED_LOADS && !SZ_DETECT_BIG_ENDIAN
    for (sz_u64_vec_t a_vec, b_vec; a + 8 <= min_end; a += 8, b += 8) {
        a_vec = sz_u64_load(a);
        b_vec = sz_u64_load(b);
        if (a_vec.u64 != b_vec.u64)
            return _sz_order_scalars(sz_u64_bytes_reverse(a_vec.u64), sz_u64_bytes_reverse(b_vec.u64));
    }
#endif
    for (; a != min_end; ++a, ++b)
        if (*a != *b) return _sz_order_scalars(*a, *b);

    // If the strings are equal up to `min_end`, then the shorter string is smaller
    return _sz_order_scalars(a_length, b_length);
}

/**
 *  @brief  Byte-level equality comparison between two 64-bit integers.
 *  @return 64-bit integer, where every top bit in each byte signifies a match.
 */
SZ_INTERNAL sz_u64_vec_t _sz_u64_each_byte_equal(sz_u64_vec_t a, sz_u64_vec_t b) {
    sz_u64_vec_t vec;
    vec.u64 = ~(a.u64 ^ b.u64);
    // The match is valid, if every bit within each byte is set.
    // For that take the bottom 7 bits of each byte, add one to them,
    // and if this sets the top bit to one, then all the 7 bits are ones as well.
    vec.u64 = ((vec.u64 & 0x7F7F7F7F7F7F7F7Full) + 0x0101010101010101ull) & ((vec.u64 & 0x8080808080808080ull));
    return vec;
}

/**
 *  @brief  Find the first occurrence of a @b single-character needle in an arbitrary length haystack.
 *          This implementation uses hardware-agnostic SWAR technique, to process 8 characters at a time.
 *          Identical to `memchr(haystack, needle[0], haystack_length)`.
 */
SZ_PUBLIC sz_cptr_t sz_find_byte_serial(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n) {

    if (!h_length) return SZ_NULL_CHAR;
    sz_cptr_t const h_end = h + h_length;

#if !SZ_DETECT_BIG_ENDIAN    // Use SWAR only on little-endian platforms for brevety.
#if !SZ_USE_MISALIGNED_LOADS // Process the misaligned head, to void UB on unaligned 64-bit loads.
    for (; ((sz_size_t)h & 7ull) && h < h_end; ++h)
        if (*h == *n) return h;
#endif

    // Broadcast the n into every byte of a 64-bit integer to use SWAR
    // techniques and process eight characters at a time.
    sz_u64_vec_t h_vec, n_vec, match_vec;
    match_vec.u64 = 0;
    n_vec.u64 = (sz_u64_t)n[0] * 0x0101010101010101ull;
    for (; h + 8 <= h_end; h += 8) {
        h_vec.u64 = *(sz_u64_t const *)h;
        match_vec = _sz_u64_each_byte_equal(h_vec, n_vec);
        if (match_vec.u64) return h + sz_u64_ctz(match_vec.u64) / 8;
    }
#endif

    // Handle the misaligned tail.
    for (; h < h_end; ++h)
        if (*h == *n) return h;
    return SZ_NULL_CHAR;
}

/**
 *  @brief  Find the last occurrence of a @b single-character needle in an arbitrary length haystack.
 *          This implementation uses hardware-agnostic SWAR technique, to process 8 characters at a time.
 *          Identical to `memrchr(haystack, needle[0], haystack_length)`.
 */
sz_cptr_t sz_rfind_byte_serial(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n) {

    if (!h_length) return SZ_NULL_CHAR;
    sz_cptr_t const h_start = h;

    // Reposition the `h` pointer to the end, as we will be walking backwards.
    h = h + h_length - 1;

#if !SZ_DETECT_BIG_ENDIAN    // Use SWAR only on little-endian platforms for brevety.
#if !SZ_USE_MISALIGNED_LOADS // Process the misaligned head, to void UB on unaligned 64-bit loads.
    for (; ((sz_size_t)(h + 1) & 7ull) && h >= h_start; --h)
        if (*h == *n) return h;
#endif

    // Broadcast the n into every byte of a 64-bit integer to use SWAR
    // techniques and process eight characters at a time.
    sz_u64_vec_t h_vec, n_vec, match_vec;
    n_vec.u64 = (sz_u64_t)n[0] * 0x0101010101010101ull;
    for (; h >= h_start + 7; h -= 8) {
        h_vec.u64 = *(sz_u64_t const *)(h - 7);
        match_vec = _sz_u64_each_byte_equal(h_vec, n_vec);
        if (match_vec.u64) return h - sz_u64_clz(match_vec.u64) / 8;
    }
#endif

    for (; h >= h_start; --h)
        if (*h == *n) return h;
    return SZ_NULL_CHAR;
}

/**
 *  @brief  2Byte-level equality comparison between two 64-bit integers.
 *  @return 64-bit integer, where every top bit in each 2byte signifies a match.
 */
SZ_INTERNAL sz_u64_vec_t _sz_u64_each_2byte_equal(sz_u64_vec_t a, sz_u64_vec_t b) {
    sz_u64_vec_t vec;
    vec.u64 = ~(a.u64 ^ b.u64);
    // The match is valid, if every bit within each 2byte is set.
    // For that take the bottom 15 bits of each 2byte, add one to them,
    // and if this sets the top bit to one, then all the 15 bits are ones as well.
    vec.u64 = ((vec.u64 & 0x7FFF7FFF7FFF7FFFull) + 0x0001000100010001ull) & ((vec.u64 & 0x8000800080008000ull));
    return vec;
}

/**
 *  @brief  Find the first occurrence of a @b two-character needle in an arbitrary length haystack.
 *          This implementation uses hardware-agnostic SWAR technique, to process 8 possible offsets at a time.
 */
SZ_INTERNAL sz_cptr_t _sz_find_2byte_serial(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n) {

    // This is an internal method, and the haystack is guaranteed to be at least 2 bytes long.
    sz_assert(h_length >= 2 && "The haystack is too short.");
    sz_cptr_t const h_end = h + h_length;

#if !SZ_USE_MISALIGNED_LOADS
    // Process the misaligned head, to void UB on unaligned 64-bit loads.
    for (; ((sz_size_t)h & 7ull) && h + 2 <= h_end; ++h)
        if ((h[0] == n[0]) + (h[1] == n[1]) == 2) return h;
#endif

    sz_u64_vec_t h_even_vec, h_odd_vec, n_vec, matches_even_vec, matches_odd_vec;
    n_vec.u64 = 0;
    n_vec.u8s[0] = n[0], n_vec.u8s[1] = n[1];
    n_vec.u64 *= 0x0001000100010001ull; // broadcast

    // This code simulates hyper-scalar execution, analyzing 8 offsets at a time.
    for (; h + 9 <= h_end; h += 8) {
        h_even_vec.u64 = *(sz_u64_t *)h;
        h_odd_vec.u64 = (h_even_vec.u64 >> 8) | ((sz_u64_t)h[8] << 56);
        matches_even_vec = _sz_u64_each_2byte_equal(h_even_vec, n_vec);
        matches_odd_vec = _sz_u64_each_2byte_equal(h_odd_vec, n_vec);

        matches_even_vec.u64 >>= 8;
        if (matches_even_vec.u64 + matches_odd_vec.u64) {
            sz_u64_t match_indicators = matches_even_vec.u64 | matches_odd_vec.u64;
            return h + sz_u64_ctz(match_indicators) / 8;
        }
    }

    for (; h + 2 <= h_end; ++h)
        if ((h[0] == n[0]) + (h[1] == n[1]) == 2) return h;
    return SZ_NULL_CHAR;
}

/**
 *  @brief  4Byte-level equality comparison between two 64-bit integers.
 *  @return 64-bit integer, where every top bit in each 4byte signifies a match.
 */
SZ_INTERNAL sz_u64_vec_t _sz_u64_each_4byte_equal(sz_u64_vec_t a, sz_u64_vec_t b) {
    sz_u64_vec_t vec;
    vec.u64 = ~(a.u64 ^ b.u64);
    // The match is valid, if every bit within each 4byte is set.
    // For that take the bottom 31 bits of each 4byte, add one to them,
    // and if this sets the top bit to one, then all the 31 bits are ones as well.
    vec.u64 = ((vec.u64 & 0x7FFFFFFF7FFFFFFFull) + 0x0000000100000001ull) & ((vec.u64 & 0x8000000080000000ull));
    return vec;
}

/**
 *  @brief  Find the first occurrence of a @b four-character needle in an arbitrary length haystack.
 *          This implementation uses hardware-agnostic SWAR technique, to process 8 possible offsets at a time.
 */
SZ_INTERNAL sz_cptr_t _sz_find_4byte_serial(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n) {

    // This is an internal method, and the haystack is guaranteed to be at least 4 bytes long.
    sz_assert(h_length >= 4 && "The haystack is too short.");
    sz_cptr_t const h_end = h + h_length;

#if !SZ_USE_MISALIGNED_LOADS
    // Process the misaligned head, to void UB on unaligned 64-bit loads.
    for (; ((sz_size_t)h & 7ull) && h + 4 <= h_end; ++h)
        if ((h[0] == n[0]) + (h[1] == n[1]) + (h[2] == n[2]) + (h[3] == n[3]) == 4) return h;
#endif

    sz_u64_vec_t h0_vec, h1_vec, h2_vec, h3_vec, n_vec, matches0_vec, matches1_vec, matches2_vec, matches3_vec;
    n_vec.u64 = 0;
    n_vec.u8s[0] = n[0], n_vec.u8s[1] = n[1], n_vec.u8s[2] = n[2], n_vec.u8s[3] = n[3];
    n_vec.u64 *= 0x0000000100000001ull; // broadcast

    // This code simulates hyper-scalar execution, analyzing 8 offsets at a time using four 64-bit words.
    // We load the subsequent four-byte word as well, taking its first bytes. Think of it as a glorified prefetch :)
    sz_u64_t h_page_current, h_page_next;
    for (; h + sizeof(sz_u64_t) + sizeof(sz_u32_t) <= h_end; h += sizeof(sz_u64_t)) {
        h_page_current = *(sz_u64_t *)h;
        h_page_next = *(sz_u32_t *)(h + 8);
        h0_vec.u64 = (h_page_current);
        h1_vec.u64 = (h_page_current >> 8) | (h_page_next << 56);
        h2_vec.u64 = (h_page_current >> 16) | (h_page_next << 48);
        h3_vec.u64 = (h_page_current >> 24) | (h_page_next << 40);
        matches0_vec = _sz_u64_each_4byte_equal(h0_vec, n_vec);
        matches1_vec = _sz_u64_each_4byte_equal(h1_vec, n_vec);
        matches2_vec = _sz_u64_each_4byte_equal(h2_vec, n_vec);
        matches3_vec = _sz_u64_each_4byte_equal(h3_vec, n_vec);

        if (matches0_vec.u64 | matches1_vec.u64 | matches2_vec.u64 | matches3_vec.u64) {
            matches0_vec.u64 >>= 24;
            matches1_vec.u64 >>= 16;
            matches2_vec.u64 >>= 8;
            sz_u64_t match_indicators = matches0_vec.u64 | matches1_vec.u64 | matches2_vec.u64 | matches3_vec.u64;
            return h + sz_u64_ctz(match_indicators) / 8;
        }
    }

    for (; h + 4 <= h_end; ++h)
        if ((h[0] == n[0]) + (h[1] == n[1]) + (h[2] == n[2]) + (h[3] == n[3]) == 4) return h;
    return SZ_NULL_CHAR;
}

/**
 *  @brief  3Byte-level equality comparison between two 64-bit integers.
 *  @return 64-bit integer, where every top bit in each 3byte signifies a match.
 */
SZ_INTERNAL sz_u64_vec_t _sz_u64_each_3byte_equal(sz_u64_vec_t a, sz_u64_vec_t b) {
    sz_u64_vec_t vec;
    vec.u64 = ~(a.u64 ^ b.u64);
    // The match is valid, if every bit within each 4byte is set.
    // For that take the bottom 31 bits of each 4byte, add one to them,
    // and if this sets the top bit to one, then all the 31 bits are ones as well.
    vec.u64 = ((vec.u64 & 0xFFFF7FFFFF7FFFFFull) + 0x0000000001000001ull) & ((vec.u64 & 0x0000800000800000ull));
    return vec;
}

/**
 *  @brief  Find the first occurrence of a @b three-character needle in an arbitrary length haystack.
 *          This implementation uses hardware-agnostic SWAR technique, to process 8 possible offsets at a time.
 */
SZ_INTERNAL sz_cptr_t _sz_find_3byte_serial(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n) {

    // This is an internal method, and the haystack is guaranteed to be at least 4 bytes long.
    sz_assert(h_length >= 3 && "The haystack is too short.");
    sz_cptr_t const h_end = h + h_length;

#if !SZ_USE_MISALIGNED_LOADS
    // Process the misaligned head, to void UB on unaligned 64-bit loads.
    for (; ((sz_size_t)h & 7ull) && h + 3 <= h_end; ++h)
        if ((h[0] == n[0]) + (h[1] == n[1]) + (h[2] == n[2]) == 3) return h;
#endif

    // We fetch 12
    sz_u64_vec_t h0_vec, h1_vec, h2_vec, h3_vec, h4_vec;
    sz_u64_vec_t matches0_vec, matches1_vec, matches2_vec, matches3_vec, matches4_vec;
    sz_u64_vec_t n_vec;
    n_vec.u64 = 0;
    n_vec.u8s[0] = n[0], n_vec.u8s[1] = n[1], n_vec.u8s[2] = n[2], n_vec.u8s[3] = n[3];
    n_vec.u64 *= 0x0000000001000001ull; // broadcast

    // This code simulates hyper-scalar execution, analyzing 8 offsets at a time using three 64-bit words.
    // We load the subsequent two-byte word as well.
    sz_u64_t h_page_current, h_page_next;
    for (; h + sizeof(sz_u64_t) + sizeof(sz_u16_t) <= h_end; h += sizeof(sz_u64_t)) {
        h_page_current = *(sz_u64_t *)h;
        h_page_next = *(sz_u16_t *)(h + 8);
        h0_vec.u64 = (h_page_current);
        h1_vec.u64 = (h_page_current >> 8) | (h_page_next << 56);
        h2_vec.u64 = (h_page_current >> 16) | (h_page_next << 48);
        h3_vec.u64 = (h_page_current >> 24) | (h_page_next << 40);
        h4_vec.u64 = (h_page_current >> 32) | (h_page_next << 32);
        matches0_vec = _sz_u64_each_3byte_equal(h0_vec, n_vec);
        matches1_vec = _sz_u64_each_3byte_equal(h1_vec, n_vec);
        matches2_vec = _sz_u64_each_3byte_equal(h2_vec, n_vec);
        matches3_vec = _sz_u64_each_3byte_equal(h3_vec, n_vec);
        matches4_vec = _sz_u64_each_3byte_equal(h4_vec, n_vec);

        if (matches0_vec.u64 | matches1_vec.u64 | matches2_vec.u64 | matches3_vec.u64 | matches4_vec.u64) {
            matches0_vec.u64 >>= 16;
            matches1_vec.u64 >>= 8;
            matches3_vec.u64 <<= 8;
            matches4_vec.u64 <<= 16;
            sz_u64_t match_indicators =
                matches0_vec.u64 | matches1_vec.u64 | matches2_vec.u64 | matches3_vec.u64 | matches4_vec.u64;
            return h + sz_u64_ctz(match_indicators) / 8;
        }
    }

    for (; h + 3 <= h_end; ++h)
        if ((h[0] == n[0]) + (h[1] == n[1]) + (h[2] == n[2]) == 3) return h;
    return SZ_NULL_CHAR;
}

/**
 *  @brief  Boyer-Moore-Horspool algorithm for exact matching of patterns up to @b 256-bytes long.
 *          Uses the Raita heuristic to match the first two, the last, and the middle character of the pattern.
 */
SZ_INTERNAL sz_cptr_t _sz_find_horspool_upto_256bytes_serial(sz_cptr_t h_chars, sz_size_t h_length, //
                                                             sz_cptr_t n_chars, sz_size_t n_length) {
    sz_assert(n_length <= 256 && "The pattern is too long.");
    // Several popular string matching algorithms are using a bad-character shift table.
    // Boyer Moore: https://www-igm.univ-mlv.fr/~lecroq/string/node14.html
    // Quick Search: https://www-igm.univ-mlv.fr/~lecroq/string/node19.html
    // Smith: https://www-igm.univ-mlv.fr/~lecroq/string/node21.html
    union {
        sz_u8_t jumps[256];
        sz_u64_vec_t vecs[64];
    } bad_shift_table;

    // Let's initialize the table using SWAR to the total length of the string.
    sz_u8_t const *h = (sz_u8_t const *)h_chars;
    sz_u8_t const *n = (sz_u8_t const *)n_chars;
    {
        sz_u64_vec_t n_length_vec;
        n_length_vec.u64 = n_length;
        n_length_vec.u64 *= 0x0101010101010101ull; // broadcast
        for (sz_size_t i = 0; i != 64; ++i) bad_shift_table.vecs[i].u64 = n_length_vec.u64;
        for (sz_size_t i = 0; i + 1 < n_length; ++i) bad_shift_table.jumps[n[i]] = (sz_u8_t)(n_length - i - 1);
    }

    // Another common heuristic is to match a few characters from different parts of a string.
    // Raita suggests to use the first two, the last, and the middle character of the pattern.
    sz_u32_vec_t h_vec, n_vec;

    // Pick the parts of the needle that are worth comparing.
    sz_size_t offset_first, offset_mid, offset_last;
    _sz_locate_needle_anomalies(n_chars, n_length, &offset_first, &offset_mid, &offset_last);

    // Broadcast those characters into an unsigned integer.
    n_vec.u8s[0] = n[offset_first];
    n_vec.u8s[1] = n[offset_first + 1];
    n_vec.u8s[2] = n[offset_mid];
    n_vec.u8s[3] = n[offset_last];

    // Scan through the whole haystack, skipping the last `n_length - 1` bytes.
    for (sz_size_t i = 0; i <= h_length - n_length;) {
        h_vec.u8s[0] = h[i + offset_first];
        h_vec.u8s[1] = h[i + offset_first + 1];
        h_vec.u8s[2] = h[i + offset_mid];
        h_vec.u8s[3] = h[i + offset_last];
        if (h_vec.u32 == n_vec.u32 && sz_equal((sz_cptr_t)h + i, n_chars, n_length)) return (sz_cptr_t)h + i;
        i += bad_shift_table.jumps[h[i + n_length - 1]];
    }
    return SZ_NULL_CHAR;
}

/**
 *  @brief  Boyer-Moore-Horspool algorithm for @b reverse-order exact matching of patterns up to @b 256-bytes long.
 *          Uses the Raita heuristic to match the first two, the last, and the middle character of the pattern.
 */
SZ_INTERNAL sz_cptr_t _sz_rfind_horspool_upto_256bytes_serial(sz_cptr_t h_chars, sz_size_t h_length, //
                                                              sz_cptr_t n_chars, sz_size_t n_length) {
    sz_assert(n_length <= 256 && "The pattern is too long.");
    union {
        sz_u8_t jumps[256];
        sz_u64_vec_t vecs[64];
    } bad_shift_table;

    // Let's initialize the table using SWAR to the total length of the string.
    sz_u8_t const *h = (sz_u8_t const *)h_chars;
    sz_u8_t const *n = (sz_u8_t const *)n_chars;
    {
        sz_u64_vec_t n_length_vec;
        n_length_vec.u64 = n_length;
        n_length_vec.u64 *= 0x0101010101010101ull; // broadcast
        for (sz_size_t i = 0; i != 64; ++i) bad_shift_table.vecs[i].u64 = n_length_vec.u64;
        for (sz_size_t i = 0; i + 1 < n_length; ++i)
            bad_shift_table.jumps[n[n_length - i - 1]] = (sz_u8_t)(n_length - i - 1);
    }

    // Another common heuristic is to match a few characters from different parts of a string.
    // Raita suggests to use the first two, the last, and the middle character of the pattern.
    sz_u32_vec_t h_vec, n_vec;

    // Pick the parts of the needle that are worth comparing.
    sz_size_t offset_first, offset_mid, offset_last;
    _sz_locate_needle_anomalies(n_chars, n_length, &offset_first, &offset_mid, &offset_last);

    // Broadcast those characters into an unsigned integer.
    n_vec.u8s[0] = n[offset_first];
    n_vec.u8s[1] = n[offset_first + 1];
    n_vec.u8s[2] = n[offset_mid];
    n_vec.u8s[3] = n[offset_last];

    // Scan through the whole haystack, skipping the first `n_length - 1` bytes.
    for (sz_size_t j = 0; j <= h_length - n_length;) {
        sz_size_t i = h_length - n_length - j;
        h_vec.u8s[0] = h[i + offset_first];
        h_vec.u8s[1] = h[i + offset_first + 1];
        h_vec.u8s[2] = h[i + offset_mid];
        h_vec.u8s[3] = h[i + offset_last];
        if (h_vec.u32 == n_vec.u32 && sz_equal((sz_cptr_t)h + i, n_chars, n_length)) return (sz_cptr_t)h + i;
        j += bad_shift_table.jumps[h[i]];
    }
    return SZ_NULL_CHAR;
}

/**
 *  @brief  Exact substring search helper function, that finds the first occurrence of a prefix of the needle
 *          using a given search function, and then verifies the remaining part of the needle.
 */
SZ_INTERNAL sz_cptr_t _sz_find_with_prefix(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n, sz_size_t n_length,
                                           sz_find_t find_prefix, sz_size_t prefix_length) {

    sz_size_t suffix_length = n_length - prefix_length;
    while (1) {
        sz_cptr_t found = find_prefix(h, h_length, n, prefix_length);
        if (!found) return SZ_NULL_CHAR;

        // Verify the remaining part of the needle
        sz_size_t remaining = h_length - (found - h);
        if (remaining < suffix_length) return SZ_NULL_CHAR;
        if (sz_equal(found + prefix_length, n + prefix_length, suffix_length)) return found;

        // Adjust the position.
        h = found + 1;
        h_length = remaining - 1;
    }

    // Unreachable, but helps silence compiler warnings:
    return SZ_NULL_CHAR;
}

/**
 *  @brief  Exact reverse-order substring search helper function, that finds the last occurrence of a suffix of the
 *          needle using a given search function, and then verifies the remaining part of the needle.
 */
SZ_INTERNAL sz_cptr_t _sz_rfind_with_suffix(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n, sz_size_t n_length,
                                            sz_find_t find_suffix, sz_size_t suffix_length) {

    sz_size_t prefix_length = n_length - suffix_length;
    while (1) {
        sz_cptr_t found = find_suffix(h, h_length, n + prefix_length, suffix_length);
        if (!found) return SZ_NULL_CHAR;

        // Verify the remaining part of the needle
        sz_size_t remaining = found - h;
        if (remaining < prefix_length) return SZ_NULL_CHAR;
        if (sz_equal(found - prefix_length, n, prefix_length)) return found - prefix_length;

        // Adjust the position.
        h_length = remaining - 1;
    }

    // Unreachable, but helps silence compiler warnings:
    return SZ_NULL_CHAR;
}

SZ_INTERNAL sz_cptr_t _sz_find_over_4bytes_serial(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n, sz_size_t n_length) {
    return _sz_find_with_prefix(h, h_length, n, n_length, (sz_find_t)_sz_find_4byte_serial, 4);
}

SZ_INTERNAL sz_cptr_t _sz_find_horspool_over_256bytes_serial(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n,
                                                             sz_size_t n_length) {
    return _sz_find_with_prefix(h, h_length, n, n_length, _sz_find_horspool_upto_256bytes_serial, 256);
}

SZ_INTERNAL sz_cptr_t _sz_rfind_horspool_over_256bytes_serial(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n,
                                                              sz_size_t n_length) {
    return _sz_rfind_with_suffix(h, h_length, n, n_length, _sz_rfind_horspool_upto_256bytes_serial, 256);
}

SZ_PUBLIC sz_cptr_t sz_find_serial(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n, sz_size_t n_length) {

    // This almost never fires, but it's better to be safe than sorry.
    if (h_length < n_length || !n_length) return SZ_NULL_CHAR;

#if SZ_DETECT_BIG_ENDIAN
    sz_find_t backends[] = {
        (sz_find_t)sz_find_byte_serial,
        (sz_find_t)_sz_find_horspool_upto_256bytes_serial,
        (sz_find_t)_sz_find_horspool_over_256bytes_serial,
    };

    return backends[(n_length > 1) + (n_length > 256)](h, h_length, n, n_length);
#else
    sz_find_t backends[] = {
        // For very short strings brute-force SWAR makes sense.
        (sz_find_t)sz_find_byte_serial,
        (sz_find_t)_sz_find_2byte_serial,
        (sz_find_t)_sz_find_3byte_serial,
        (sz_find_t)_sz_find_4byte_serial,
        // To avoid constructing the skip-table, let's use the prefixed approach.
        (sz_find_t)_sz_find_over_4bytes_serial,
        // For longer needles - use skip tables.
        (sz_find_t)_sz_find_horspool_upto_256bytes_serial,
        (sz_find_t)_sz_find_horspool_over_256bytes_serial,
    };

    return backends[
        // For very short strings brute-force SWAR makes sense.
        (n_length > 1) + (n_length > 2) + (n_length > 3) +
        // To avoid constructing the skip-table, let's use the prefixed approach.
        (n_length > 4) +
        // For longer needles - use skip tables.
        (n_length > 8) + (n_length > 256)](h, h_length, n, n_length);
#endif
}

SZ_PUBLIC sz_cptr_t sz_rfind_serial(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n, sz_size_t n_length) {

    // This almost never fires, but it's better to be safe than sorry.
    if (h_length < n_length || !n_length) return SZ_NULL_CHAR;

    sz_find_t backends[] = {
        // For very short strings brute-force SWAR makes sense.
        (sz_find_t)sz_rfind_byte_serial,
        //  TODO: implement reverse-order SWAR for 2/3/4 byte variants.
        //  TODO: (sz_find_t)_sz_rfind_2byte_serial,
        //  TODO: (sz_find_t)_sz_rfind_3byte_serial,
        //  TODO: (sz_find_t)_sz_rfind_4byte_serial,
        // To avoid constructing the skip-table, let's use the prefixed approach.
        // (sz_find_t)_sz_rfind_over_4bytes_serial,
        // For longer needles - use skip tables.
        (sz_find_t)_sz_rfind_horspool_upto_256bytes_serial,
        (sz_find_t)_sz_rfind_horspool_over_256bytes_serial,
    };

    return backends[
        // For very short strings brute-force SWAR makes sense.
        0 +
        // To avoid constructing the skip-table, let's use the prefixed approach.
        (n_length > 1) +
        // For longer needles - use skip tables.
        (n_length > 256)](h, h_length, n, n_length);
}

SZ_INTERNAL sz_size_t _sz_edit_distance_skewed_diagonals_serial( //
    sz_cptr_t shorter, sz_size_t shorter_length,                 //
    sz_cptr_t longer, sz_size_t longer_length,                   //
    sz_size_t bound, sz_memory_allocator_t *alloc) {

    // Simplify usage in higher-level libraries, where wrapping custom allocators may be troublesome.
    sz_memory_allocator_t global_alloc;
    if (!alloc) {
        sz_memory_allocator_init_default(&global_alloc);
        alloc = &global_alloc;
    }

    // TODO: Generalize to remove the following asserts!
    sz_assert(!bound && "For bounded search the method should only evaluate one band of the matrix.");
    sz_assert(shorter_length == longer_length && "The method hasn't been generalized to different length inputs yet.");
    sz_unused(longer_length && bound);

    // We are going to store 3 diagonals of the matrix.
    // The length of the longest (main) diagonal would be `n = (shorter_length + 1)`.
    sz_size_t n = shorter_length + 1;
    sz_size_t buffer_length = sizeof(sz_size_t) * n * 3;
    sz_size_t *distances = (sz_size_t *)alloc->allocate(buffer_length, alloc->handle);
    if (!distances) return SZ_SIZE_MAX;

    sz_size_t *previous_distances = distances;
    sz_size_t *current_distances = previous_distances + n;
    sz_size_t *next_distances = previous_distances + n * 2;

    // Initialize the first two diagonals:
    previous_distances[0] = 0;
    current_distances[0] = current_distances[1] = 1;

    // Progress through the upper triangle of the Levenshtein matrix.
    sz_size_t next_skew_diagonal_index = 2;
    for (; next_skew_diagonal_index != n; ++next_skew_diagonal_index) {
        sz_size_t const next_skew_diagonal_length = next_skew_diagonal_index + 1;
        for (sz_size_t i = 0; i + 2 < next_skew_diagonal_length; ++i) {
            sz_size_t cost_of_substitution = shorter[next_skew_diagonal_index - i - 2] != longer[i];
            sz_size_t cost_if_substitution = previous_distances[i] + cost_of_substitution;
            sz_size_t cost_if_deletion_or_insertion = sz_min_of_two(current_distances[i], current_distances[i + 1]) + 1;
            next_distances[i + 1] = sz_min_of_two(cost_if_deletion_or_insertion, cost_if_substitution);
        }
        // Don't forget to populate the first row and the fiest column of the Levenshtein matrix.
        next_distances[0] = next_distances[next_skew_diagonal_length - 1] = next_skew_diagonal_index;
        // Perform a circular rotarion of those buffers, to reuse the memory.
        sz_size_t *temporary = previous_distances;
        previous_distances = current_distances;
        current_distances = next_distances;
        next_distances = temporary;
    }

    // By now we've scanned through the upper triangle of the matrix, where each subsequent iteration results in a
    // larger diagonal. From now onwards, we will be shrinking. Instead of adding value equal to the skewed diagonal
    // index on either side, we will be cropping those values out.
    sz_size_t total_diagonals = n + n - 1;
    for (; next_skew_diagonal_index != total_diagonals; ++next_skew_diagonal_index) {
        sz_size_t const next_skew_diagonal_length = total_diagonals - next_skew_diagonal_index;
        for (sz_size_t i = 0; i != next_skew_diagonal_length; ++i) {
            sz_size_t cost_of_substitution =
                shorter[shorter_length - 1 - i] != longer[next_skew_diagonal_index - n + i];
            sz_size_t cost_if_substitution = previous_distances[i] + cost_of_substitution;
            sz_size_t cost_if_deletion_or_insertion = sz_min_of_two(current_distances[i], current_distances[i + 1]) + 1;
            next_distances[i] = sz_min_of_two(cost_if_deletion_or_insertion, cost_if_substitution);
        }
        // Perform a circular rotarion of those buffers, to reuse the memory, this time, with a shift,
        // dropping the first element in the current array.
        sz_size_t *temporary = previous_distances;
        previous_distances = current_distances + 1;
        current_distances = next_distances;
        next_distances = temporary;
    }

    // Cache scalar before `free` call.
    sz_size_t result = current_distances[0];
    alloc->free(distances, buffer_length, alloc->handle);
    return result;
}

/**
 *  @brief  Describes the length of a UTF8 character / codepoint / rune in bytes.
 */
typedef enum {
    sz_utf8_invalid_k = 0,     //!< Invalid UTF8 character.
    sz_utf8_rune_1byte_k = 1,  //!< 1-byte UTF8 character.
    sz_utf8_rune_2bytes_k = 2, //!< 2-byte UTF8 character.
    sz_utf8_rune_3bytes_k = 3, //!< 3-byte UTF8 character.
    sz_utf8_rune_4bytes_k = 4, //!< 4-byte UTF8 character.
} sz_rune_length_t;

typedef sz_u32_t sz_rune_t;

/**
 *  @brief  Extracts just one UTF8 codepoint from a UTF8 string into a 32-bit unsigned integer.
 */
SZ_INTERNAL void _sz_extract_utf8_rune(sz_cptr_t utf8, sz_rune_t *code, sz_rune_length_t *code_length) {
    sz_u8_t const *current = (sz_u8_t const *)utf8;
    sz_u8_t leading_byte = *current++;
    sz_rune_t ch;
    sz_rune_length_t ch_length;

    // TODO: This can be made entirely branchless using 32-bit SWAR.
    if (leading_byte < 0x80) {
        // Single-byte rune (0xxxxxxx)
        ch = leading_byte;
        ch_length = sz_utf8_rune_1byte_k;
    }
    else if ((leading_byte & 0xE0) == 0xC0) {
        // Two-byte rune (110xxxxx 10xxxxxx)
        ch = (leading_byte & 0x1F) << 6;
        ch |= (*current++ & 0x3F);
        ch_length = sz_utf8_rune_2bytes_k;
    }
    else if ((leading_byte & 0xF0) == 0xE0) {
        // Three-byte rune (1110xxxx 10xxxxxx 10xxxxxx)
        ch = (leading_byte & 0x0F) << 12;
        ch |= (*current++ & 0x3F) << 6;
        ch |= (*current++ & 0x3F);
        ch_length = sz_utf8_rune_3bytes_k;
    }
    else if ((leading_byte & 0xF8) == 0xF0) {
        // Four-byte rune (11110xxx 10xxxxxx 10xxxxxx 10xxxxxx)
        ch = (leading_byte & 0x07) << 18;
        ch |= (*current++ & 0x3F) << 12;
        ch |= (*current++ & 0x3F) << 6;
        ch |= (*current++ & 0x3F);
        ch_length = sz_utf8_rune_4bytes_k;
    }
    else {
        // Invalid UTF8 rune.
        ch = 0;
        ch_length = sz_utf8_invalid_k;
    }
    *code = ch;
    *code_length = ch_length;
}

/**
 *  @brief  Exports a UTF8 string into a UTF32 buffer.
 *          ! The result is undefined id the UTF8 string is corrupted.
 *  @return The length in the number of codepoints.
 */
SZ_INTERNAL sz_size_t _sz_export_utf8_to_utf32(sz_cptr_t utf8, sz_size_t utf8_length, sz_rune_t *utf32) {
    sz_cptr_t const end = utf8 + utf8_length;
    sz_size_t count = 0;
    sz_rune_length_t rune_length;
    for (; utf8 != end; utf8 += rune_length, utf32++, count++) _sz_extract_utf8_rune(utf8, utf32, &rune_length);
    return count;
}

/**
 *  @brief  Compute the Levenshtein distance between two strings using the Wagner-Fisher algorithm.
 *          Stores only 2 rows of the Levenshtein matrix, but uses 64-bit integers for the distance values,
 *          and upcasts UTF8 variable-length codepoints to 64-bit integers for faster addressing.
 *
 *  ! In the worst case for 2 strings of length 100, that contain just one 16-bit codepoint this will result in extra:
 *      + 2 rows * 100 slots * 8 bytes/slot = 1600 bytes of memory for the two rows of the Levenshtein matrix rows.
 *      + 100 codepoints * 2 strings * 4 bytes/codepoint = 800 bytes of memory for the UTF8 buffer.
 *      = 2400 bytes of memory or @b 12x memory amplification!
 */
SZ_INTERNAL sz_size_t _sz_edit_distance_wagner_fisher_serial( //
    sz_cptr_t longer, sz_size_t longer_length,                //
    sz_cptr_t shorter, sz_size_t shorter_length,              //
    sz_size_t bound, sz_bool_t can_be_unicode, sz_memory_allocator_t *alloc) {

    // Simplify usage in higher-level libraries, where wrapping custom allocators may be troublesome.
    sz_memory_allocator_t global_alloc;
    if (!alloc) {
        sz_memory_allocator_init_default(&global_alloc);
        alloc = &global_alloc;
    }

    // A good idea may be to dispatch different kernels for different string lengths.
    // Like using `uint8_t` counters for strings under 255 characters long.
    // Good in theory, this results in frequent upcasts and downcasts in serial code.
    // On strings over 20 bytes, using `uint8` over `uint64` on 64-bit x86 CPU doubles the execution time.
    // So one must be very cautious with such optimizations.
    typedef sz_size_t _distance_t;

    // Compute the number of columns in our Levenshtein matrix.
    sz_size_t const n = shorter_length + 1;

    // If a buffering memory-allocator is provided, this operation is practically free,
    // and cheaper than allocating even 512 bytes (for small distance matrices) on stack.
    sz_size_t buffer_length = sizeof(_distance_t) * (n * 2);

    // If the strings contain Unicode characters, let's estimate the max character width,
    // and use it to allocate a larger buffer to decode UTF8.
    if ((can_be_unicode == sz_true_k) &&
        (sz_isascii(longer, longer_length) == sz_false_k || sz_isascii(shorter, shorter_length) == sz_false_k)) {
        buffer_length += (shorter_length + longer_length) * sizeof(sz_rune_t);
    }
    else { can_be_unicode = sz_false_k; }

    // If the allocation fails, return the maximum distance.
    sz_ptr_t const buffer = (sz_ptr_t)alloc->allocate(buffer_length, alloc->handle);
    if (!buffer) return SZ_SIZE_MAX;

    // Let's export the UTF8 sequence into the newly allocated buffer at the end.
    if (can_be_unicode == sz_true_k) {
        sz_rune_t *const longer_utf32 = (sz_rune_t *)(buffer + sizeof(_distance_t) * (n * 2));
        sz_rune_t *const shorter_utf32 = longer_utf32 + longer_length;
        // Export the UTF8 sequences into the newly allocated buffer.
        longer_length = _sz_export_utf8_to_utf32(longer, longer_length, longer_utf32);
        shorter_length = _sz_export_utf8_to_utf32(shorter, shorter_length, shorter_utf32);
        longer = (sz_cptr_t)longer_utf32;
        shorter = (sz_cptr_t)shorter_utf32;
    }

    // Let's parameterize the core logic for different character types and distance types.
#define _wagner_fisher_unbounded(_distance_t, _char_t)                                                                \
    /* Now let's cast our pointer to avoid it in subsequent sections. */                                              \
    _char_t const *const longer_chars = (_char_t const *)longer;                                                      \
    _char_t const *const shorter_chars = (_char_t const *)shorter;                                                    \
    _distance_t *previous_distances = (_distance_t *)buffer;                                                          \
    _distance_t *current_distances = previous_distances + n;                                                          \
    /*  Initialize the first row of the Levenshtein matrix with `iota`-style arithmetic progression. */               \
    for (_distance_t idx_shorter = 0; idx_shorter != n; ++idx_shorter) previous_distances[idx_shorter] = idx_shorter; \
    /* The main loop of the algorithm with quadratic complexity. */                                                   \
    for (_distance_t idx_longer = 0; idx_longer != longer_length; ++idx_longer) {                                     \
        _char_t const longer_char = longer_chars[idx_longer];                                                         \
        /* Using pure pointer arithmetic is faster than iterating with an index. */                                   \
        _char_t const *shorter_ptr = shorter_chars;                                                                   \
        _distance_t const *previous_ptr = previous_distances;                                                         \
        _distance_t *current_ptr = current_distances;                                                                 \
        _distance_t *const current_end = current_ptr + shorter_length;                                                \
        current_ptr[0] = idx_longer + 1;                                                                              \
        for (; current_ptr != current_end; ++previous_ptr, ++current_ptr, ++shorter_ptr) {                            \
            _distance_t cost_substitution = previous_ptr[0] + (_distance_t)(longer_char != shorter_ptr[0]);           \
            /* We can avoid `+1` for costs here, shifting it to post-minimum computation, */                          \
            /* saving one increment operation. */                                                                     \
            _distance_t cost_deletion = previous_ptr[1];                                                              \
            _distance_t cost_insertion = current_ptr[0];                                                              \
            /* ? It might be a good idea to enforce branchless execution here. */                                     \
            /* ? The caveat being that the benchmarks on longer sequences backfire and more research is needed. */    \
            current_ptr[1] = sz_min_of_two(cost_substitution, sz_min_of_two(cost_deletion, cost_insertion) + 1);      \
        }                                                                                                             \
        /* Swap `previous_distances` and `current_distances` pointers. */                                             \
        _distance_t *temporary = previous_distances;                                                                  \
        previous_distances = current_distances;                                                                       \
        current_distances = temporary;                                                                                \
    }                                                                                                                 \
    /* Cache scalar before `free` call. */                                                                            \
    sz_size_t result = previous_distances[shorter_length];                                                            \
    alloc->free(buffer, buffer_length, alloc->handle);                                                                \
    return result;

    // Let's define a separate variant for bounded distance computation.
    // Practically the same as unbounded, but also collecting the running minimum within each row for early exit.
#define _wagner_fisher_bounded(_distance_t, _char_t)                                                                  \
    _char_t const *const longer_chars = (_char_t const *)longer;                                                      \
    _char_t const *const shorter_chars = (_char_t const *)shorter;                                                    \
    _distance_t *previous_distances = (_distance_t *)buffer;                                                          \
    _distance_t *current_distances = previous_distances + n;                                                          \
    for (_distance_t idx_shorter = 0; idx_shorter != n; ++idx_shorter) previous_distances[idx_shorter] = idx_shorter; \
    for (_distance_t idx_longer = 0; idx_longer != longer_length; ++idx_longer) {                                     \
        _char_t const longer_char = longer_chars[idx_longer];                                                         \
        _char_t const *shorter_ptr = shorter_chars;                                                                   \
        _distance_t const *previous_ptr = previous_distances;                                                         \
        _distance_t *current_ptr = current_distances;                                                                 \
        _distance_t *const current_end = current_ptr + shorter_length;                                                \
        current_ptr[0] = idx_longer + 1;                                                                              \
        /* Initialize min_distance with a value greater than bound */                                                 \
        _distance_t min_distance = bound - 1;                                                                         \
        for (; current_ptr != current_end; ++previous_ptr, ++current_ptr, ++shorter_ptr) {                            \
            _distance_t cost_substitution = previous_ptr[0] + (_distance_t)(longer_char != shorter_ptr[0]);           \
            _distance_t cost_deletion = previous_ptr[1];                                                              \
            _distance_t cost_insertion = current_ptr[0];                                                              \
            current_ptr[1] = sz_min_of_two(cost_substitution, sz_min_of_two(cost_deletion, cost_insertion) + 1);      \
            /* Keep track of the minimum distance seen so far in this row */                                          \
            min_distance = sz_min_of_two(current_ptr[1], min_distance);                                               \
        }                                                                                                             \
        /* If the minimum distance in this row exceeded the bound, return early */                                    \
        if (min_distance >= bound) {                                                                                  \
            alloc->free(buffer, buffer_length, alloc->handle);                                                        \
            return bound;                                                                                             \
        }                                                                                                             \
        _distance_t *temporary = previous_distances;                                                                  \
        previous_distances = current_distances;                                                                       \
        current_distances = temporary;                                                                                \
    }                                                                                                                 \
    sz_size_t result = previous_distances[shorter_length];                                                            \
    alloc->free(buffer, buffer_length, alloc->handle);                                                                \
    return sz_min_of_two(result, bound);

    // Dispatch the actual computation.
    if (!bound) {
        if (can_be_unicode == sz_true_k) { _wagner_fisher_unbounded(sz_size_t, sz_rune_t); }
        else { _wagner_fisher_unbounded(sz_size_t, sz_u8_t); }
    }
    else {
        if (can_be_unicode == sz_true_k) { _wagner_fisher_bounded(sz_size_t, sz_rune_t); }
        else { _wagner_fisher_bounded(sz_size_t, sz_u8_t); }
    }
}

SZ_PUBLIC sz_size_t sz_edit_distance_serial(     //
    sz_cptr_t longer, sz_size_t longer_length,   //
    sz_cptr_t shorter, sz_size_t shorter_length, //
    sz_size_t bound, sz_memory_allocator_t *alloc) {

    // Let's make sure that we use the amount proportional to the
    // number of elements in the shorter string, not the larger.
    if (shorter_length > longer_length) {
        sz_pointer_swap((void **)&longer_length, (void **)&shorter_length);
        sz_pointer_swap((void **)&longer, (void **)&shorter);
    }

    // Skip the matching prefixes and suffixes, they won't affect the distance.
    for (sz_cptr_t a_end = longer + longer_length, b_end = shorter + shorter_length;
         longer != a_end && shorter != b_end && *longer == *shorter;
         ++longer, ++shorter, --longer_length, --shorter_length)
        ;
    for (; longer_length && shorter_length && longer[longer_length - 1] == shorter[shorter_length - 1];
         --longer_length, --shorter_length)
        ;

    // Bounded computations may exit early.
    if (bound) {
        // If one of the strings is empty - the edit distance is equal to the length of the other one.
        if (longer_length == 0) return sz_min_of_two(shorter_length, bound);
        if (shorter_length == 0) return sz_min_of_two(longer_length, bound);
        // If the difference in length is beyond the `bound`, there is no need to check at all.
        if (longer_length - shorter_length > bound) return bound;
    }

    if (shorter_length == 0) return longer_length; // If no mismatches were found - the distance is zero.
    if (shorter_length == longer_length && !bound)
        return _sz_edit_distance_skewed_diagonals_serial(longer, longer_length, shorter, shorter_length, bound, alloc);
    return _sz_edit_distance_wagner_fisher_serial(longer, longer_length, shorter, shorter_length, bound, sz_false_k,
                                                  alloc);
}

SZ_PUBLIC sz_ssize_t sz_alignment_score_serial(       //
    sz_cptr_t longer, sz_size_t longer_length,        //
    sz_cptr_t shorter, sz_size_t shorter_length,      //
    sz_error_cost_t const *subs, sz_error_cost_t gap, //
    sz_memory_allocator_t *alloc) {

    // If one of the strings is empty - the edit distance is equal to the length of the other one
    if (longer_length == 0) return (sz_ssize_t)shorter_length * gap;
    if (shorter_length == 0) return (sz_ssize_t)longer_length * gap;

    // Let's make sure that we use the amount proportional to the
    // number of elements in the shorter string, not the larger.
    if (shorter_length > longer_length) {
        sz_pointer_swap((void **)&longer_length, (void **)&shorter_length);
        sz_pointer_swap((void **)&longer, (void **)&shorter);
    }

    // Simplify usage in higher-level libraries, where wrapping custom allocators may be troublesome.
    sz_memory_allocator_t global_alloc;
    if (!alloc) {
        sz_memory_allocator_init_default(&global_alloc);
        alloc = &global_alloc;
    }

    sz_size_t n = shorter_length + 1;
    sz_size_t buffer_length = sizeof(sz_ssize_t) * n * 2;
    sz_ssize_t *distances = (sz_ssize_t *)alloc->allocate(buffer_length, alloc->handle);
    sz_ssize_t *previous_distances = distances;
    sz_ssize_t *current_distances = previous_distances + n;

    for (sz_size_t idx_shorter = 0; idx_shorter != n; ++idx_shorter)
        previous_distances[idx_shorter] = (sz_ssize_t)idx_shorter * gap;

    sz_u8_t const *shorter_unsigned = (sz_u8_t const *)shorter;
    sz_u8_t const *longer_unsigned = (sz_u8_t const *)longer;
    for (sz_size_t idx_longer = 0; idx_longer != longer_length; ++idx_longer) {
        current_distances[0] = ((sz_ssize_t)idx_longer + 1) * gap;

        // Initialize min_distance with a value greater than bound
        sz_error_cost_t const *a_subs = subs + longer_unsigned[idx_longer] * 256ul;
        for (sz_size_t idx_shorter = 0; idx_shorter != shorter_length; ++idx_shorter) {
            sz_ssize_t cost_deletion = previous_distances[idx_shorter + 1] + gap;
            sz_ssize_t cost_insertion = current_distances[idx_shorter] + gap;
            sz_ssize_t cost_substitution = previous_distances[idx_shorter] + a_subs[shorter_unsigned[idx_shorter]];
            current_distances[idx_shorter + 1] = sz_max_of_three(cost_deletion, cost_insertion, cost_substitution);
        }

        // Swap previous_distances and current_distances pointers
        sz_pointer_swap((void **)&previous_distances, (void **)&current_distances);
    }

    // Cache scalar before `free` call.
    sz_ssize_t result = previous_distances[shorter_length];
    alloc->free(distances, buffer_length, alloc->handle);
    return result;
}

SZ_PUBLIC sz_size_t sz_hamming_distance_serial( //
    sz_cptr_t a, sz_size_t a_length,            //
    sz_cptr_t b, sz_size_t b_length,            //
    sz_size_t bound) {

    sz_size_t const min_length = sz_min_of_two(a_length, b_length);
    sz_size_t const max_length = sz_max_of_two(a_length, b_length);
    sz_cptr_t const a_end = a + min_length;
    bound = bound == 0 ? max_length : bound;

    // Walk through both strings using SWAR and counting the number of differing characters.
    sz_size_t distance = max_length - min_length;
#if SZ_USE_MISALIGNED_LOADS && !SZ_DETECT_BIG_ENDIAN
    if (min_length >= SZ_SWAR_THRESHOLD) {
        sz_u64_vec_t a_vec, b_vec, match_vec;
        for (; a + 8 <= a_end && distance < bound; a += 8, b += 8) {
            a_vec.u64 = sz_u64_load(a).u64;
            b_vec.u64 = sz_u64_load(b).u64;
            match_vec = _sz_u64_each_byte_equal(a_vec, b_vec);
            distance += sz_u64_popcount((~match_vec.u64) & 0x8080808080808080ull);
        }
    }
#endif

    for (; a != a_end && distance < bound; ++a, ++b) { distance += (*a != *b); }
    return sz_min_of_two(distance, bound);
}

SZ_PUBLIC sz_size_t sz_hamming_distance_utf8_serial( //
    sz_cptr_t a, sz_size_t a_length,                 //
    sz_cptr_t b, sz_size_t b_length,                 //
    sz_size_t bound) {

    sz_cptr_t const a_end = a + a_length;
    sz_cptr_t const b_end = b + b_length;
    sz_size_t distance = 0;

    sz_rune_t a_rune, b_rune;
    sz_rune_length_t a_rune_length, b_rune_length;

    if (bound) {
        for (; a < a_end && b < b_end && distance < bound; a += a_rune_length, b += b_rune_length) {
            _sz_extract_utf8_rune(a, &a_rune, &a_rune_length);
            _sz_extract_utf8_rune(b, &b_rune, &b_rune_length);
            distance += (a_rune != b_rune);
        }
        // If one string has more runes, we need to go through the tail.
        if (distance < bound) {
            for (; a < a_end && distance < bound; a += a_rune_length, ++distance)
                _sz_extract_utf8_rune(a, &a_rune, &a_rune_length);

            for (; b < b_end && distance < bound; b += b_rune_length, ++distance)
                _sz_extract_utf8_rune(b, &b_rune, &b_rune_length);
        }
    }
    else {
        for (; a < a_end && b < b_end; a += a_rune_length, b += b_rune_length) {
            _sz_extract_utf8_rune(a, &a_rune, &a_rune_length);
            _sz_extract_utf8_rune(b, &b_rune, &b_rune_length);
            distance += (a_rune != b_rune);
        }
        // If one string has more runes, we need to go through the tail.
        for (; a < a_end; a += a_rune_length, ++distance) _sz_extract_utf8_rune(a, &a_rune, &a_rune_length);
        for (; b < b_end; b += b_rune_length, ++distance) _sz_extract_utf8_rune(b, &b_rune, &b_rune_length);
    }
    return distance;
}

/**
 *  @brief  Largest prime number that fits into 31 bits.
 *  @see    https://mersenneforum.org/showthread.php?t=3471
 */
#define SZ_U32_MAX_PRIME (2147483647u)

/**
 *  @brief  Largest prime number that fits into 64 bits.
 *  @see    https://mersenneforum.org/showthread.php?t=3471
 *
 *  2^64 = 18,446,744,073,709,551,616
 *  this = 18,446,744,073,709,551,557
 *  diff = 59
 */
#define SZ_U64_MAX_PRIME (18446744073709551557ull)

/*
 *  One hardware-accelerated way of mixing hashes can be CRC, but it's only implemented for 32-bit values.
 *  Using a Boost-like mixer works very poorly in such case:
 *
 *       hash_first ^ (hash_second + 0x517cc1b727220a95 + (hash_first << 6) + (hash_first >> 2));
 *
 *  Let's stick to the Fibonacci hash trick using the golden ratio.
 *  https://probablydance.com/2018/06/16/fibonacci-hashing-the-optimization-that-the-world-forgot-or-a-better-alternative-to-integer-modulo/
 */
#define _sz_hash_mix(first, second) ((first * 11400714819323198485ull) ^ (second * 11400714819323198485ull))
#define _sz_shift_low(x) (x)
#define _sz_shift_high(x) ((x + 77ull) & 0xFFull)
#define _sz_prime_mod(x) (x % SZ_U64_MAX_PRIME)

SZ_PUBLIC sz_u64_t sz_hash_serial(sz_cptr_t start, sz_size_t length) {

    sz_u64_t hash_low = 0;
    sz_u64_t hash_high = 0;
    sz_u8_t const *text = (sz_u8_t const *)start;
    sz_u8_t const *text_end = text + length;

    switch (length) {
    case 0: return 0;

    // Texts under 7 bytes long are definitely below the largest prime.
    case 1:
        hash_low = _sz_shift_low(text[0]);
        hash_high = _sz_shift_high(text[0]);
        break;
    case 2:
        hash_low = _sz_shift_low(text[0]) * 31ull + _sz_shift_low(text[1]);
        hash_high = _sz_shift_high(text[0]) * 257ull + _sz_shift_high(text[1]);
        break;
    case 3:
        hash_low = _sz_shift_low(text[0]) * 31ull * 31ull + //
                   _sz_shift_low(text[1]) * 31ull +         //
                   _sz_shift_low(text[2]);
        hash_high = _sz_shift_high(text[0]) * 257ull * 257ull + //
                    _sz_shift_high(text[1]) * 257ull +          //
                    _sz_shift_high(text[2]);
        break;
    case 4:
        hash_low = _sz_shift_low(text[0]) * 31ull * 31ull * 31ull + //
                   _sz_shift_low(text[1]) * 31ull * 31ull +         //
                   _sz_shift_low(text[2]) * 31ull +                 //
                   _sz_shift_low(text[3]);
        hash_high = _sz_shift_high(text[0]) * 257ull * 257ull * 257ull + //
                    _sz_shift_high(text[1]) * 257ull * 257ull +          //
                    _sz_shift_high(text[2]) * 257ull +                   //
                    _sz_shift_high(text[3]);
        break;
    case 5:
        hash_low = _sz_shift_low(text[0]) * 31ull * 31ull * 31ull * 31ull + //
                   _sz_shift_low(text[1]) * 31ull * 31ull * 31ull +         //
                   _sz_shift_low(text[2]) * 31ull * 31ull +                 //
                   _sz_shift_low(text[3]) * 31ull +                         //
                   _sz_shift_low(text[4]);
        hash_high = _sz_shift_high(text[0]) * 257ull * 257ull * 257ull * 257ull + //
                    _sz_shift_high(text[1]) * 257ull * 257ull * 257ull +          //
                    _sz_shift_high(text[2]) * 257ull * 257ull +                   //
                    _sz_shift_high(text[3]) * 257ull +                            //
                    _sz_shift_high(text[4]);
        break;
    case 6:
        hash_low = _sz_shift_low(text[0]) * 31ull * 31ull * 31ull * 31ull * 31ull + //
                   _sz_shift_low(text[1]) * 31ull * 31ull * 31ull * 31ull +         //
                   _sz_shift_low(text[2]) * 31ull * 31ull * 31ull +                 //
                   _sz_shift_low(text[3]) * 31ull * 31ull +                         //
                   _sz_shift_low(text[4]) * 31ull +                                 //
                   _sz_shift_low(text[5]);
        hash_high = _sz_shift_high(text[0]) * 257ull * 257ull * 257ull * 257ull * 257ull + //
                    _sz_shift_high(text[1]) * 257ull * 257ull * 257ull * 257ull +          //
                    _sz_shift_high(text[2]) * 257ull * 257ull * 257ull +                   //
                    _sz_shift_high(text[3]) * 257ull * 257ull +                            //
                    _sz_shift_high(text[4]) * 257ull +                                     //
                    _sz_shift_high(text[5]);
        break;
    case 7:
        hash_low = _sz_shift_low(text[0]) * 31ull * 31ull * 31ull * 31ull * 31ull * 31ull + //
                   _sz_shift_low(text[1]) * 31ull * 31ull * 31ull * 31ull * 31ull +         //
                   _sz_shift_low(text[2]) * 31ull * 31ull * 31ull * 31ull +                 //
                   _sz_shift_low(text[3]) * 31ull * 31ull * 31ull +                         //
                   _sz_shift_low(text[4]) * 31ull * 31ull +                                 //
                   _sz_shift_low(text[5]) * 31ull +                                         //
                   _sz_shift_low(text[6]);
        hash_high = _sz_shift_high(text[0]) * 257ull * 257ull * 257ull * 257ull * 257ull * 257ull + //
                    _sz_shift_high(text[1]) * 257ull * 257ull * 257ull * 257ull * 257ull +          //
                    _sz_shift_high(text[2]) * 257ull * 257ull * 257ull * 257ull +                   //
                    _sz_shift_high(text[3]) * 257ull * 257ull * 257ull +                            //
                    _sz_shift_high(text[4]) * 257ull * 257ull +                                     //
                    _sz_shift_high(text[5]) * 257ull +                                              //
                    _sz_shift_high(text[6]);
        break;
    default:
        // Unroll the first seven cycles:
        hash_low = hash_low * 31ull + _sz_shift_low(text[0]);
        hash_high = hash_high * 257ull + _sz_shift_high(text[0]);
        hash_low = hash_low * 31ull + _sz_shift_low(text[1]);
        hash_high = hash_high * 257ull + _sz_shift_high(text[1]);
        hash_low = hash_low * 31ull + _sz_shift_low(text[2]);
        hash_high = hash_high * 257ull + _sz_shift_high(text[2]);
        hash_low = hash_low * 31ull + _sz_shift_low(text[3]);
        hash_high = hash_high * 257ull + _sz_shift_high(text[3]);
        hash_low = hash_low * 31ull + _sz_shift_low(text[4]);
        hash_high = hash_high * 257ull + _sz_shift_high(text[4]);
        hash_low = hash_low * 31ull + _sz_shift_low(text[5]);
        hash_high = hash_high * 257ull + _sz_shift_high(text[5]);
        hash_low = hash_low * 31ull + _sz_shift_low(text[6]);
        hash_high = hash_high * 257ull + _sz_shift_high(text[6]);
        text += 7;

        // Iterate throw the rest with the modulus:
        for (; text != text_end; ++text) {
            hash_low = hash_low * 31ull + _sz_shift_low(text[0]);
            hash_high = hash_high * 257ull + _sz_shift_high(text[0]);
            // Wrap the hashes around:
            hash_low = _sz_prime_mod(hash_low);
            hash_high = _sz_prime_mod(hash_high);
        }
        break;
    }

    return _sz_hash_mix(hash_low, hash_high);
}

SZ_PUBLIC void sz_hashes_serial(sz_cptr_t start, sz_size_t length, sz_size_t window_length, sz_size_t step, //
                                sz_hash_callback_t callback, void *callback_handle) {

    if (length < window_length || !window_length) return;
    sz_u8_t const *text = (sz_u8_t const *)start;
    sz_u8_t const *text_end = text + length;

    // Prepare the `prime ^ window_length` values, that we are going to use for modulo arithmetic.
    sz_u64_t prime_power_low = 1, prime_power_high = 1;
    for (sz_size_t i = 0; i + 1 < window_length; ++i)
        prime_power_low = (prime_power_low * 31ull) % SZ_U64_MAX_PRIME,
        prime_power_high = (prime_power_high * 257ull) % SZ_U64_MAX_PRIME;

    // Compute the initial hash value for the first window.
    sz_u64_t hash_low = 0, hash_high = 0, hash_mix;
    for (sz_u8_t const *first_end = text + window_length; text < first_end; ++text)
        hash_low = (hash_low * 31ull + _sz_shift_low(*text)) % SZ_U64_MAX_PRIME,
        hash_high = (hash_high * 257ull + _sz_shift_high(*text)) % SZ_U64_MAX_PRIME;

    // In most cases the fingerprint length will be a power of two.
    hash_mix = _sz_hash_mix(hash_low, hash_high);
    callback((sz_cptr_t)text, window_length, hash_mix, callback_handle);

    // Compute the hash value for every window, exporting into the fingerprint,
    // using the expensive modulo operation.
    sz_size_t cycles = 1;
    sz_size_t const step_mask = step - 1;
    for (; text < text_end; ++text, ++cycles) {
        // Discard one character:
        hash_low -= _sz_shift_low(*(text - window_length)) * prime_power_low;
        hash_high -= _sz_shift_high(*(text - window_length)) * prime_power_high;
        // And add a new one:
        hash_low = 31ull * hash_low + _sz_shift_low(*text);
        hash_high = 257ull * hash_high + _sz_shift_high(*text);
        // Wrap the hashes around:
        hash_low = _sz_prime_mod(hash_low);
        hash_high = _sz_prime_mod(hash_high);
        // Mix only if we've skipped enough hashes.
        if ((cycles & step_mask) == 0) {
            hash_mix = _sz_hash_mix(hash_low, hash_high);
            callback((sz_cptr_t)text, window_length, hash_mix, callback_handle);
        }
    }
}

#undef _sz_shift_low
#undef _sz_shift_high
#undef _sz_hash_mix
#undef _sz_prime_mod

/**
 *  @brief  Uses a small lookup-table to convert a lowercase character to uppercase.
 */
SZ_INTERNAL sz_u8_t sz_u8_tolower(sz_u8_t c) {
    static sz_u8_t const lowered[256] = {
        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,  //
        32,  33,  34,  35,  36,  37,  38,  39,  40,  41,  42,  43,  44,  45,  46,  47,  //
        48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  58,  59,  60,  61,  62,  63,  //
        64,  97,  98,  99,  100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, //
        112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 91,  92,  93,  94,  95,  //
        96,  97,  98,  99,  100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, //
        112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, //
        128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, //
        144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, //
        160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, //
        176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, //
        224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, //
        240, 241, 242, 243, 244, 245, 246, 215, 248, 249, 250, 251, 252, 253, 254, 223, //
        224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, //
        240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255, //
    };
    return lowered[c];
}

/**
 *  @brief  Uses a small lookup-table to convert an uppercase character to lowercase.
 */
SZ_INTERNAL sz_u8_t sz_u8_toupper(sz_u8_t c) {
    static sz_u8_t const upped[256] = {
        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,  //
        32,  33,  34,  35,  36,  37,  38,  39,  40,  41,  42,  43,  44,  45,  46,  47,  //
        48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  58,  59,  60,  61,  62,  63,  //
        64,  97,  98,  99,  100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, //
        112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 91,  92,  93,  94,  95,  //
        96,  65,  66,  67,  68,  69,  70,  71,  72,  73,  74,  75,  76,  77,  78,  79,  //
        80,  81,  82,  83,  84,  85,  86,  87,  88,  89,  90,  123, 124, 125, 126, 127, //
        128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, //
        144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, //
        160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, //
        176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, //
        224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, //
        240, 241, 242, 243, 244, 245, 246, 215, 248, 249, 250, 251, 252, 253, 254, 223, //
        224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, //
        240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255, //
    };
    return upped[c];
}

/**
 *  @brief  Uses two small lookup tables (768 bytes total) to accelerate division by a small
 *          unsigned integer. Performs two lookups, one multiplication, two shifts, and two accumulations.
 *
 *  @param  divisor Integral value @b larger than one.
 *  @param  number  Integral value to divide.
 */
SZ_INTERNAL sz_u8_t sz_u8_divide(sz_u8_t number, sz_u8_t divisor) {
    sz_assert(divisor > 1);
    static sz_u16_t const multipliers[256] = {
        0,     0,     0,     21846, 0,     39322, 21846, 9363,  0,     50973, 39322, 29790, 21846, 15124, 9363,  4370,
        0,     57826, 50973, 44841, 39322, 34329, 29790, 25645, 21846, 18351, 15124, 12137, 9363,  6780,  4370,  2115,
        0,     61565, 57826, 54302, 50973, 47824, 44841, 42011, 39322, 36765, 34329, 32006, 29790, 27671, 25645, 23705,
        21846, 20063, 18351, 16706, 15124, 13602, 12137, 10725, 9363,  8049,  6780,  5554,  4370,  3224,  2115,  1041,
        0,     63520, 61565, 59668, 57826, 56039, 54302, 52614, 50973, 49377, 47824, 46313, 44841, 43407, 42011, 40649,
        39322, 38028, 36765, 35532, 34329, 33154, 32006, 30885, 29790, 28719, 27671, 26647, 25645, 24665, 23705, 22766,
        21846, 20945, 20063, 19198, 18351, 17520, 16706, 15907, 15124, 14356, 13602, 12863, 12137, 11424, 10725, 10038,
        9363,  8700,  8049,  7409,  6780,  6162,  5554,  4957,  4370,  3792,  3224,  2665,  2115,  1573,  1041,  517,
        0,     64520, 63520, 62535, 61565, 60609, 59668, 58740, 57826, 56926, 56039, 55164, 54302, 53452, 52614, 51788,
        50973, 50169, 49377, 48595, 47824, 47063, 46313, 45572, 44841, 44120, 43407, 42705, 42011, 41326, 40649, 39982,
        39322, 38671, 38028, 37392, 36765, 36145, 35532, 34927, 34329, 33738, 33154, 32577, 32006, 31443, 30885, 30334,
        29790, 29251, 28719, 28192, 27671, 27156, 26647, 26143, 25645, 25152, 24665, 24182, 23705, 23233, 22766, 22303,
        21846, 21393, 20945, 20502, 20063, 19628, 19198, 18772, 18351, 17933, 17520, 17111, 16706, 16305, 15907, 15514,
        15124, 14738, 14356, 13977, 13602, 13231, 12863, 12498, 12137, 11779, 11424, 11073, 10725, 10380, 10038, 9699,
        9363,  9030,  8700,  8373,  8049,  7727,  7409,  7093,  6780,  6470,  6162,  5857,  5554,  5254,  4957,  4662,
        4370,  4080,  3792,  3507,  3224,  2943,  2665,  2388,  2115,  1843,  1573,  1306,  1041,  778,   517,   258,
    };
    // This table can be avoided using a single addition and counting trailing zeros.
    static sz_u8_t const shifts[256] = {
        0, 0, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, //
        4, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, //
        5, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, //
        6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, //
        6, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, //
        7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, //
        7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, //
        7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, //
    };
    sz_u32_t multiplier = multipliers[divisor];
    sz_u8_t shift = shifts[divisor];

    sz_u16_t q = (sz_u16_t)((multiplier * number) >> 16);
    sz_u16_t t = ((number - q) >> 1) + q;
    return (sz_u8_t)(t >> shift);
}

SZ_PUBLIC void sz_tolower_serial(sz_cptr_t text, sz_size_t length, sz_ptr_t result) {
    sz_u8_t *unsigned_result = (sz_u8_t *)result;
    sz_u8_t const *unsigned_text = (sz_u8_t const *)text;
    sz_u8_t const *end = unsigned_text + length;
    for (; unsigned_text != end; ++unsigned_text, ++unsigned_result) *unsigned_result = sz_u8_tolower(*unsigned_text);
}

SZ_PUBLIC void sz_toupper_serial(sz_cptr_t text, sz_size_t length, sz_ptr_t result) {
    sz_u8_t *unsigned_result = (sz_u8_t *)result;
    sz_u8_t const *unsigned_text = (sz_u8_t const *)text;
    sz_u8_t const *end = unsigned_text + length;
    for (; unsigned_text != end; ++unsigned_text, ++unsigned_result) *unsigned_result = sz_u8_toupper(*unsigned_text);
}

SZ_PUBLIC void sz_toascii_serial(sz_cptr_t text, sz_size_t length, sz_ptr_t result) {
    sz_u8_t *unsigned_result = (sz_u8_t *)result;
    sz_u8_t const *unsigned_text = (sz_u8_t const *)text;
    sz_u8_t const *end = unsigned_text + length;
    for (; unsigned_text != end; ++unsigned_text, ++unsigned_result) *unsigned_result = *unsigned_text & 0x7F;
}

/**
 *  @brief  Check if there is a byte in this buffer, that exceeds 127 and can't be an ASCII character.
 *          This implementation uses hardware-agnostic SWAR technique, to process 8 characters at a time.
 */
SZ_PUBLIC sz_bool_t sz_isascii_serial(sz_cptr_t text, sz_size_t length) {

    if (!length) return sz_true_k;
    sz_u8_t const *h = (sz_u8_t const *)text;
    sz_u8_t const *const h_end = h + length;

#if !SZ_USE_MISALIGNED_LOADS
    // Process the misaligned head, to void UB on unaligned 64-bit loads.
    for (; ((sz_size_t)h & 7ull) && h < h_end; ++h)
        if (*h & 0x80ull) return sz_false_k;
#endif

    // Validate eight bytes at once using SWAR.
    sz_u64_vec_t text_vec;
    for (; h + 8 <= h_end; h += 8) {
        text_vec.u64 = *(sz_u64_t const *)h;
        if (text_vec.u64 & 0x8080808080808080ull) return sz_false_k;
    }

    // Handle the misaligned tail.
    for (; h < h_end; ++h)
        if (*h & 0x80ull) return sz_false_k;
    return sz_true_k;
}

SZ_PUBLIC void sz_generate_serial(sz_cptr_t alphabet, sz_size_t alphabet_size, sz_ptr_t result, sz_size_t result_length,
                                  sz_random_generator_t generator, void *generator_user_data) {

    sz_assert(alphabet_size > 0 && alphabet_size <= 256 && "Inadequate alphabet size");

    if (alphabet_size == 1) sz_fill(result, result_length, *alphabet);

    else {
        sz_assert(generator && "Expects a valid random generator");
        sz_u8_t divisor = (sz_u8_t)alphabet_size;
        for (sz_cptr_t end = result + result_length; result != end; ++result) {
            sz_u8_t random = generator(generator_user_data) & 0xFF;
            sz_u8_t quotient = sz_u8_divide(random, divisor);
            *result = alphabet[random - quotient * divisor];
        }
    }
}

#pragma endregion

/*
 *  Serial implementation of string class operations.
 */
#pragma region Serial Implementation for the String Class

SZ_PUBLIC sz_bool_t sz_string_is_on_stack(sz_string_t const *string) {
    // It doesn't matter if it's on stack or heap, the pointer location is the same.
    return (sz_bool_t)((sz_cptr_t)string->internal.start == (sz_cptr_t)&string->internal.chars[0]);
}

SZ_PUBLIC void sz_string_range(sz_string_t const *string, sz_ptr_t *start, sz_size_t *length) {
    sz_size_t is_small = (sz_cptr_t)string->internal.start == (sz_cptr_t)&string->internal.chars[0];
    sz_size_t is_big_mask = is_small - 1ull;
    *start = string->external.start; // It doesn't matter if it's on stack or heap, the pointer location is the same.
    // If the string is small, use branch-less approach to mask-out the top 7 bytes of the length.
    *length = string->external.length & (0x00000000000000FFull | is_big_mask);
}

SZ_PUBLIC void sz_string_unpack(sz_string_t const *string, sz_ptr_t *start, sz_size_t *length, sz_size_t *space,
                                sz_bool_t *is_external) {
    sz_size_t is_small = (sz_cptr_t)string->internal.start == (sz_cptr_t)&string->internal.chars[0];
    sz_size_t is_big_mask = is_small - 1ull;
    *start = string->external.start; // It doesn't matter if it's on stack or heap, the pointer location is the same.
    // If the string is small, use branch-less approach to mask-out the top 7 bytes of the length.
    *length = string->external.length & (0x00000000000000FFull | is_big_mask);
    // In case the string is small, the `is_small - 1ull` will become 0xFFFFFFFFFFFFFFFFull.
    *space = sz_u64_blend(SZ_STRING_INTERNAL_SPACE, string->external.space, is_big_mask);
    *is_external = (sz_bool_t)!is_small;
}

SZ_PUBLIC sz_bool_t sz_string_equal(sz_string_t const *a, sz_string_t const *b) {
    // Tempting to say that the external.length is bitwise the same even if it includes
    // some bytes of the on-stack payload, but we don't at this writing maintain that invariant.
    // (An on-stack string includes noise bytes in the high-order bits of external.length. So do this
    // the hard/correct way.

#if SZ_USE_MISALIGNED_LOADS
    // Dealing with StringZilla strings, we know that the `start` pointer always points
    // to a word at least 8 bytes long. Therefore, we can compare the first 8 bytes at once.

#endif
    // Alternatively, fall back to byte-by-byte comparison.
    sz_ptr_t a_start, b_start;
    sz_size_t a_length, b_length;
    sz_string_range(a, &a_start, &a_length);
    sz_string_range(b, &b_start, &b_length);
    return (sz_bool_t)(a_length == b_length && sz_equal(a_start, b_start, b_length));
}

SZ_PUBLIC sz_ordering_t sz_string_order(sz_string_t const *a, sz_string_t const *b) {
#if SZ_USE_MISALIGNED_LOADS
    // Dealing with StringZilla strings, we know that the `start` pointer always points
    // to a word at least 8 bytes long. Therefore, we can compare the first 8 bytes at once.

#endif
    // Alternatively, fall back to byte-by-byte comparison.
    sz_ptr_t a_start, b_start;
    sz_size_t a_length, b_length;
    sz_string_range(a, &a_start, &a_length);
    sz_string_range(b, &b_start, &b_length);
    return sz_order(a_start, a_length, b_start, b_length);
}

SZ_PUBLIC void sz_string_init(sz_string_t *string) {
    sz_assert(string && "String can't be SZ_NULL.");

    // Only 8 + 1 + 1 need to be initialized.
    string->internal.start = &string->internal.chars[0];
    // But for safety let's initialize the entire structure to zeros.
    // string->internal.chars[0] = 0;
    // string->internal.length = 0;
    string->words[1] = 0;
    string->words[2] = 0;
    string->words[3] = 0;
}

SZ_PUBLIC sz_ptr_t sz_string_init_length(sz_string_t *string, sz_size_t length, sz_memory_allocator_t *allocator) {
    sz_size_t space_needed = length + 1; // space for trailing \0
    sz_assert(string && allocator && "String and allocator can't be SZ_NULL.");
    // Initialize the string to zeros for safety.
    string->words[1] = 0;
    string->words[2] = 0;
    string->words[3] = 0;
    // If we are lucky, no memory allocations will be needed.
    if (space_needed <= SZ_STRING_INTERNAL_SPACE) {
        string->internal.start = &string->internal.chars[0];
        string->internal.length = (sz_u8_t)length;
    }
    else {
        // If we are not lucky, we need to allocate memory.
        string->external.start = (sz_ptr_t)allocator->allocate(space_needed, allocator->handle);
        if (!string->external.start) return SZ_NULL_CHAR;
        string->external.length = length;
        string->external.space = space_needed;
    }
    sz_assert(&string->internal.start == &string->external.start && "Alignment confusion");
    string->external.start[length] = 0;
    return string->external.start;
}

SZ_PUBLIC sz_ptr_t sz_string_reserve(sz_string_t *string, sz_size_t new_capacity, sz_memory_allocator_t *allocator) {

    sz_assert(string && "String can't be SZ_NULL.");

    sz_size_t new_space = new_capacity + 1;
    if (new_space <= SZ_STRING_INTERNAL_SPACE) return string->external.start;

    sz_ptr_t string_start;
    sz_size_t string_length;
    sz_size_t string_space;
    sz_bool_t string_is_external;
    sz_string_unpack(string, &string_start, &string_length, &string_space, &string_is_external);
    sz_assert(new_space > string_space && "New space must be larger than current.");

    sz_ptr_t new_start = (sz_ptr_t)allocator->allocate(new_space, allocator->handle);
    if (!new_start) return SZ_NULL_CHAR;

    sz_copy(new_start, string_start, string_length);
    string->external.start = new_start;
    string->external.space = new_space;
    string->external.padding = 0;
    string->external.length = string_length;

    // Deallocate the old string.
    if (string_is_external) allocator->free(string_start, string_space, allocator->handle);
    return string->external.start;
}

SZ_PUBLIC sz_ptr_t sz_string_expand(sz_string_t *string, sz_size_t offset, sz_size_t added_length,
                                    sz_memory_allocator_t *allocator) {

    sz_assert(string && allocator && "String and allocator can't be SZ_NULL.");

    sz_ptr_t string_start;
    sz_size_t string_length;
    sz_size_t string_space;
    sz_bool_t string_is_external;
    sz_string_unpack(string, &string_start, &string_length, &string_space, &string_is_external);

    // The user intended to extend the string.
    offset = sz_min_of_two(offset, string_length);

    // If we are lucky, no memory allocations will be needed.
    if (offset + string_length + added_length < string_space) {
        sz_move(string_start + offset + added_length, string_start + offset, string_length - offset);
        string_start[string_length + added_length] = 0;
        // Even if the string is on the stack, the `+=` won't affect the tail of the string.
        string->external.length += added_length;
    }
    // If we are not lucky, we need to allocate more memory.
    else {
        sz_size_t next_planned_size = sz_max_of_two(SZ_CACHE_LINE_WIDTH, string_space * 2ull);
        sz_size_t min_needed_space = sz_size_bit_ceil(offset + string_length + added_length + 1);
        sz_size_t new_space = sz_max_of_two(min_needed_space, next_planned_size);
        string_start = sz_string_reserve(string, new_space - 1, allocator);
        if (!string_start) return SZ_NULL_CHAR;

        // Copy into the new buffer.
        sz_move(string_start + offset + added_length, string_start + offset, string_length - offset);
        string_start[string_length + added_length] = 0;
        string->external.length = string_length + added_length;
    }

    return string_start;
}

SZ_PUBLIC sz_size_t sz_string_erase(sz_string_t *string, sz_size_t offset, sz_size_t length) {

    sz_assert(string && "String can't be SZ_NULL.");

    sz_ptr_t string_start;
    sz_size_t string_length;
    sz_size_t string_space;
    sz_bool_t string_is_external;
    sz_string_unpack(string, &string_start, &string_length, &string_space, &string_is_external);

    // Normalize the offset, it can't be larger than the length.
    offset = sz_min_of_two(offset, string_length);

    // We shouldn't normalize the length, to avoid overflowing on `offset + length >= string_length`,
    // if receiving `length == SZ_SIZE_MAX`. After following expression the `length` will contain
    // exactly the delta between original and final length of this `string`.
    length = sz_min_of_two(length, string_length - offset);

    // There are 2 common cases, that wouldn't even require a `memmove`:
    //      1.  Erasing the entire contents of the string.
    //          In that case `length` argument will be equal or greater than `length` member.
    //      2.  Removing the tail of the string with something like `string.pop_back()` in C++.
    //
    // In both of those, regardless of the location of the string - stack or heap,
    // the erasing is as easy as setting the length to the offset.
    // In every other case, we must `memmove` the tail of the string to the left.
    if (offset + length < string_length)
        sz_move(string_start + offset, string_start + offset + length, string_length - offset - length);

    // The `string->external.length = offset` assignment would discard last characters
    // of the on-the-stack string, but inplace subtraction would work.
    string->external.length -= length;
    string_start[string_length - length] = 0;
    return length;
}

SZ_PUBLIC void sz_string_free(sz_string_t *string, sz_memory_allocator_t *allocator) {
    if (!sz_string_is_on_stack(string))
        allocator->free(string->external.start, string->external.space, allocator->handle);
    sz_string_init(string);
}

SZ_PUBLIC void sz_fill_serial(sz_ptr_t target, sz_size_t length, sz_u8_t value) {
    sz_ptr_t end = target + length;
    // Dealing with short strings, a single sequential pass would be faster.
    // If the size is larger than 2 words, then at least 1 of them will be aligned.
    // But just one aligned word may not be worth SWAR.
    if (length < SZ_SWAR_THRESHOLD)
        while (target != end) *(target++) = value;

    // In case of long strings, skip unaligned bytes, and then fill the rest in 64-bit chunks.
    else {
        sz_u64_t value64 = (sz_u64_t)(value) * 0x0101010101010101ull;
        while ((sz_size_t)target & 7ull) *(target++) = value;
        while (target + 8 <= end) *(sz_u64_t *)target = value64, target += 8;
        while (target != end) *(target++) = value;
    }
}

SZ_PUBLIC void sz_copy_serial(sz_ptr_t target, sz_cptr_t source, sz_size_t length) {
#if SZ_USE_MISALIGNED_LOADS
    while (length >= 8) *(sz_u64_t *)target = *(sz_u64_t const *)source, target += 8, source += 8, length -= 8;
#endif
    while (length--) *(target++) = *(source++);
}

SZ_PUBLIC void sz_move_serial(sz_ptr_t target, sz_cptr_t source, sz_size_t length) {
    // Implementing `memmove` is trickier, than `memcpy`, as the ranges may overlap.
    // Existing implementations often have two passes, in normal and reversed order,
    // depending on the relation of `target` and `source` addresses.
    // https://student.cs.uwaterloo.ca/~cs350/common/os161-src-html/doxygen/html/memmove_8c_source.html
    // https://marmota.medium.com/c-language-making-memmove-def8792bb8d5
    //
    // We can use the `memcpy` like left-to-right pass if we know that the `target` is before `source`.
    // Or if we know that they don't intersect! In that case the traversal order is irrelevant,
    // but older CPUs may predict and fetch forward-passes better.
    if (target < source || target >= source + length) {
#if SZ_USE_MISALIGNED_LOADS
        while (length >= 8) *(sz_u64_t *)target = *(sz_u64_t const *)(source), target += 8, source += 8, length -= 8;
#endif
        while (length--) *(target++) = *(source++);
    }
    else {
        // Jump to the end and walk backwards.
        target += length, source += length;
#if SZ_USE_MISALIGNED_LOADS
        while (length >= 8) *(sz_u64_t *)(target -= 8) = *(sz_u64_t const *)(source -= 8), length -= 8;
#endif
        while (length--) *(--target) = *(--source);
    }
}

#pragma endregion

/*
 *  @brief  Serial implementation for strings sequence processing.
 */
#pragma region Serial Implementation for Sequences

SZ_PUBLIC sz_size_t sz_partition(sz_sequence_t *sequence, sz_sequence_predicate_t predicate) {

    sz_size_t matches = 0;
    while (matches != sequence->count && predicate(sequence, sequence->order[matches])) ++matches;

    for (sz_size_t i = matches + 1; i < sequence->count; ++i)
        if (predicate(sequence, sequence->order[i]))
            sz_u64_swap(sequence->order + i, sequence->order + matches), ++matches;

    return matches;
}

SZ_PUBLIC void sz_merge(sz_sequence_t *sequence, sz_size_t partition, sz_sequence_comparator_t less) {

    sz_size_t start_b = partition + 1;

    // If the direct merge is already sorted
    if (!less(sequence, sequence->order[start_b], sequence->order[partition])) return;

    sz_size_t start_a = 0;
    while (start_a <= partition && start_b <= sequence->count) {

        // If element 1 is in right place
        if (!less(sequence, sequence->order[start_b], sequence->order[start_a])) { start_a++; }
        else {
            sz_size_t value = sequence->order[start_b];
            sz_size_t index = start_b;

            // Shift all the elements between element 1
            // element 2, right by 1.
            while (index != start_a) { sequence->order[index] = sequence->order[index - 1], index--; }
            sequence->order[start_a] = value;

            // Update all the pointers
            start_a++;
            partition++;
            start_b++;
        }
    }
}

SZ_PUBLIC void sz_sort_insertion(sz_sequence_t *sequence, sz_sequence_comparator_t less) {
    sz_u64_t *keys = sequence->order;
    sz_size_t keys_count = sequence->count;
    for (sz_size_t i = 1; i < keys_count; i++) {
        sz_u64_t i_key = keys[i];
        sz_size_t j = i;
        for (; j > 0 && less(sequence, i_key, keys[j - 1]); --j) keys[j] = keys[j - 1];
        keys[j] = i_key;
    }
}

SZ_INTERNAL void _sz_sift_down(sz_sequence_t *sequence, sz_sequence_comparator_t less, sz_u64_t *order, sz_size_t start,
                               sz_size_t end) {
    sz_size_t root = start;
    while (2 * root + 1 <= end) {
        sz_size_t child = 2 * root + 1;
        if (child + 1 <= end && less(sequence, order[child], order[child + 1])) { child++; }
        if (!less(sequence, order[root], order[child])) { return; }
        sz_u64_swap(order + root, order + child);
        root = child;
    }
}

SZ_INTERNAL void _sz_heapify(sz_sequence_t *sequence, sz_sequence_comparator_t less, sz_u64_t *order, sz_size_t count) {
    sz_size_t start = (count - 2) / 2;
    while (1) {
        _sz_sift_down(sequence, less, order, start, count - 1);
        if (start == 0) return;
        start--;
    }
}

SZ_INTERNAL void _sz_heapsort(sz_sequence_t *sequence, sz_sequence_comparator_t less, sz_size_t first, sz_size_t last) {
    sz_u64_t *order = sequence->order;
    sz_size_t count = last - first;
    _sz_heapify(sequence, less, order + first, count);
    sz_size_t end = count - 1;
    while (end > 0) {
        sz_u64_swap(order + first, order + first + end);
        end--;
        _sz_sift_down(sequence, less, order + first, 0, end);
    }
}

SZ_PUBLIC void sz_sort_introsort_recursion(sz_sequence_t *sequence, sz_sequence_comparator_t less, sz_size_t first,
                                           sz_size_t last, sz_size_t depth) {

    sz_size_t length = last - first;
    switch (length) {
    case 0:
    case 1: return;
    case 2:
        if (less(sequence, sequence->order[first + 1], sequence->order[first]))
            sz_u64_swap(&sequence->order[first], &sequence->order[first + 1]);
        return;
    case 3: {
        sz_u64_t a = sequence->order[first];
        sz_u64_t b = sequence->order[first + 1];
        sz_u64_t c = sequence->order[first + 2];
        if (less(sequence, b, a)) sz_u64_swap(&a, &b);
        if (less(sequence, c, b)) sz_u64_swap(&c, &b);
        if (less(sequence, b, a)) sz_u64_swap(&a, &b);
        sequence->order[first] = a;
        sequence->order[first + 1] = b;
        sequence->order[first + 2] = c;
        return;
    }
    }
    // Until a certain length, the quadratic-complexity insertion-sort is fine
    if (length <= 16) {
        sz_sequence_t sub_seq = *sequence;
        sub_seq.order += first;
        sub_seq.count = length;
        sz_sort_insertion(&sub_seq, less);
        return;
    }

    // Fallback to N-logN-complexity heap-sort
    if (depth == 0) {
        _sz_heapsort(sequence, less, first, last);
        return;
    }

    --depth;

    // Median-of-three logic to choose pivot
    sz_size_t median = first + length / 2;
    if (less(sequence, sequence->order[median], sequence->order[first]))
        sz_u64_swap(&sequence->order[first], &sequence->order[median]);
    if (less(sequence, sequence->order[last - 1], sequence->order[first]))
        sz_u64_swap(&sequence->order[first], &sequence->order[last - 1]);
    if (less(sequence, sequence->order[median], sequence->order[last - 1]))
        sz_u64_swap(&sequence->order[median], &sequence->order[last - 1]);

    // Partition using the median-of-three as the pivot
    sz_u64_t pivot = sequence->order[median];
    sz_size_t left = first;
    sz_size_t right = last - 1;
    while (1) {
        while (less(sequence, sequence->order[left], pivot)) left++;
        while (less(sequence, pivot, sequence->order[right])) right--;
        if (left >= right) break;
        sz_u64_swap(&sequence->order[left], &sequence->order[right]);
        left++;
        right--;
    }

    // Recursively sort the partitions
    sz_sort_introsort_recursion(sequence, less, first, left, depth);
    sz_sort_introsort_recursion(sequence, less, right + 1, last, depth);
}

SZ_PUBLIC void sz_sort_introsort(sz_sequence_t *sequence, sz_sequence_comparator_t less) {
    if (sequence->count == 0) return;
    sz_size_t size_is_not_power_of_two = (sequence->count & (sequence->count - 1)) != 0;
    sz_size_t depth_limit = sz_size_log2i_nonzero(sequence->count) + size_is_not_power_of_two;
    sz_sort_introsort_recursion(sequence, less, 0, sequence->count, depth_limit);
}

SZ_PUBLIC void sz_sort_recursion( //
    sz_sequence_t *sequence, sz_size_t bit_idx, sz_size_t bit_max, sz_sequence_comparator_t comparator,
    sz_size_t partial_order_length) {

    if (!sequence->count) return;

    // Array of size one doesn't need sorting - only needs the prefix to be discarded.
    if (sequence->count == 1) {
        sz_u32_t *order_half_words = (sz_u32_t *)sequence->order;
        order_half_words[1] = 0;
        return;
    }

    // Partition a range of integers according to a specific bit value
    sz_size_t split = 0;
    sz_u64_t mask = (1ull << 63) >> bit_idx;

    // The clean approach would be to perform a single pass over the sequence.
    //
    //    while (split != sequence->count && !(sequence->order[split] & mask)) ++split;
    //    for (sz_size_t i = split + 1; i < sequence->count; ++i)
    //        if (!(sequence->order[i] & mask)) sz_u64_swap(sequence->order + i, sequence->order + split), ++split;
    //
    // This, however, doesn't take into account the high relative cost of writes and swaps.
    // To circumvent that, we can first count the total number entries to be mapped into either part.
    // And then walk through both parts, swapping the entries that are in the wrong part.
    // This would often lead to ~15% performance gain.
    sz_size_t count_with_bit_set = 0;
    for (sz_size_t i = 0; i != sequence->count; ++i) count_with_bit_set += (sequence->order[i] & mask) != 0;
    split = sequence->count - count_with_bit_set;

    // It's possible that the sequence is already partitioned.
    if (split != 0 && split != sequence->count) {
        // Use two pointers to efficiently reposition elements.
        // On pointer walks left-to-right from the start, and the other walks right-to-left from the end.
        sz_size_t left = 0;
        sz_size_t right = sequence->count - 1;
        while (1) {
            // Find the next element with the bit set on the left side.
            while (left < split && !(sequence->order[left] & mask)) ++left;
            // Find the next element without the bit set on the right side.
            while (right >= split && (sequence->order[right] & mask)) --right;
            // Swap the mispositioned elements.
            if (left < split && right >= split) {
                sz_u64_swap(sequence->order + left, sequence->order + right);
                ++left;
                --right;
            }
            else { break; }
        }
    }

    // Go down recursively.
    if (bit_idx < bit_max) {
        sz_sequence_t a = *sequence;
        a.count = split;
        sz_sort_recursion(&a, bit_idx + 1, bit_max, comparator, partial_order_length);

        sz_sequence_t b = *sequence;
        b.order += split;
        b.count -= split;
        sz_sort_recursion(&b, bit_idx + 1, bit_max, comparator, partial_order_length);
    }
    // Reached the end of recursion.
    else {
        // Discard the prefixes.
        sz_u32_t *order_half_words = (sz_u32_t *)sequence->order;
        for (sz_size_t i = 0; i != sequence->count; ++i) { order_half_words[i * 2 + 1] = 0; }

        sz_sequence_t a = *sequence;
        a.count = split;
        sz_sort_introsort(&a, comparator);

        sz_sequence_t b = *sequence;
        b.order += split;
        b.count -= split;
        sz_sort_introsort(&b, comparator);
    }
}

SZ_INTERNAL sz_bool_t _sz_sort_is_less(sz_sequence_t *sequence, sz_size_t i_key, sz_size_t j_key) {
    sz_cptr_t i_str = sequence->get_start(sequence, i_key);
    sz_cptr_t j_str = sequence->get_start(sequence, j_key);
    sz_size_t i_len = sequence->get_length(sequence, i_key);
    sz_size_t j_len = sequence->get_length(sequence, j_key);
    return (sz_bool_t)(sz_order_serial(i_str, i_len, j_str, j_len) == sz_less_k);
}

SZ_PUBLIC void sz_sort_partial(sz_sequence_t *sequence, sz_size_t partial_order_length) {

#if SZ_DETECT_BIG_ENDIAN
    // TODO: Implement partial sort for big-endian systems. For now this sorts the whole thing.
    sz_unused(partial_order_length);
    sz_sort_introsort(sequence, (sz_sequence_comparator_t)_sz_sort_is_less);
#else

    // Export up to 4 bytes into the `sequence` bits themselves
    for (sz_size_t i = 0; i != sequence->count; ++i) {
        sz_cptr_t begin = sequence->get_start(sequence, sequence->order[i]);
        sz_size_t length = sequence->get_length(sequence, sequence->order[i]);
        length = length > 4u ? 4u : length;
        sz_ptr_t prefix = (sz_ptr_t)&sequence->order[i];
        for (sz_size_t j = 0; j != length; ++j) prefix[7 - j] = begin[j];
    }

    // Perform optionally-parallel radix sort on them
    sz_sort_recursion(sequence, 0, 32, (sz_sequence_comparator_t)_sz_sort_is_less, partial_order_length);
#endif
}

SZ_PUBLIC void sz_sort(sz_sequence_t *sequence) {
#if SZ_DETECT_BIG_ENDIAN
    sz_sort_introsort(sequence, (sz_sequence_comparator_t)_sz_sort_is_less);
#else
    sz_sort_partial(sequence, sequence->count);
#endif
}

#pragma endregion

/*
 *  @brief  AVX2 implementation of the string search algorithms.
 *          Very minimalistic, but still faster than the serial implementation.
 */
#pragma region AVX2 Implementation

#if SZ_USE_X86_AVX2
#pragma GCC push_options
#pragma GCC target("avx2")
#pragma clang attribute push(__attribute__((target("avx2"))), apply_to = function)
#include <immintrin.h>

/**
 *  @brief  Helper structure to simplify work with 256-bit registers.
 */
typedef union sz_u256_vec_t {
    __m256i ymm;
    __m128i xmms[2];
    sz_u64_t u64s[4];
    sz_u32_t u32s[8];
    sz_u16_t u16s[16];
    sz_u8_t u8s[32];
} sz_u256_vec_t;

SZ_PUBLIC void sz_fill_avx2(sz_ptr_t target, sz_size_t length, sz_u8_t value) {
    for (; length >= 32; target += 32, length -= 32) _mm256_storeu_si256((__m256i *)target, _mm256_set1_epi8(value));
    sz_fill_serial(target, length, value);
}

SZ_PUBLIC void sz_copy_avx2(sz_ptr_t target, sz_cptr_t source, sz_size_t length) {
    for (; length >= 32; target += 32, source += 32, length -= 32)
        _mm256_storeu_si256((__m256i *)target, _mm256_lddqu_si256((__m256i const *)source));
    sz_copy_serial(target, source, length);
}

SZ_PUBLIC void sz_move_avx2(sz_ptr_t target, sz_cptr_t source, sz_size_t length) {
    if (target < source || target >= source + length) {
        for (; length >= 32; target += 32, source += 32, length -= 32)
            _mm256_storeu_si256((__m256i *)target, _mm256_lddqu_si256((__m256i const *)source));
        while (length--) *(target++) = *(source++);
    }
    else {
        // Jump to the end and walk backwards.
        for (target += length, source += length; length >= 32; length -= 32)
            _mm256_storeu_si256((__m256i *)(target -= 32), _mm256_lddqu_si256((__m256i const *)(source -= 32)));
        while (length--) *(--target) = *(--source);
    }
}

SZ_PUBLIC sz_cptr_t sz_find_byte_avx2(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n) {
    int mask;
    sz_u256_vec_t h_vec, n_vec;
    n_vec.ymm = _mm256_set1_epi8(n[0]);

    while (h_length >= 32) {
        h_vec.ymm = _mm256_lddqu_si256((__m256i const *)h);
        mask = _mm256_movemask_epi8(_mm256_cmpeq_epi8(h_vec.ymm, n_vec.ymm));
        if (mask) return h + sz_u32_ctz(mask);
        h += 32, h_length -= 32;
    }

    return sz_find_byte_serial(h, h_length, n);
}

SZ_PUBLIC sz_cptr_t sz_rfind_byte_avx2(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n) {
    int mask;
    sz_u256_vec_t h_vec, n_vec;
    n_vec.ymm = _mm256_set1_epi8(n[0]);

    while (h_length >= 32) {
        h_vec.ymm = _mm256_lddqu_si256((__m256i const *)(h + h_length - 32));
        mask = _mm256_movemask_epi8(_mm256_cmpeq_epi8(h_vec.ymm, n_vec.ymm));
        if (mask) return h + h_length - 1 - sz_u32_clz(mask);
        h_length -= 32;
    }

    return sz_rfind_byte_serial(h, h_length, n);
}

SZ_PUBLIC sz_cptr_t sz_find_avx2(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n, sz_size_t n_length) {

    // This almost never fires, but it's better to be safe than sorry.
    if (h_length < n_length || !n_length) return SZ_NULL_CHAR;
    if (n_length == 1) return sz_find_byte_avx2(h, h_length, n);

    // Pick the parts of the needle that are worth comparing.
    sz_size_t offset_first, offset_mid, offset_last;
    _sz_locate_needle_anomalies(n, n_length, &offset_first, &offset_mid, &offset_last);

    // Broadcast those characters into YMM registers.
    int matches;
    sz_u256_vec_t h_first_vec, h_mid_vec, h_last_vec, n_first_vec, n_mid_vec, n_last_vec;
    n_first_vec.ymm = _mm256_set1_epi8(n[offset_first]);
    n_mid_vec.ymm = _mm256_set1_epi8(n[offset_mid]);
    n_last_vec.ymm = _mm256_set1_epi8(n[offset_last]);

    // Scan through the string.
    for (; h_length >= n_length + 32; h += 32, h_length -= 32) {
        h_first_vec.ymm = _mm256_lddqu_si256((__m256i const *)(h + offset_first));
        h_mid_vec.ymm = _mm256_lddqu_si256((__m256i const *)(h + offset_mid));
        h_last_vec.ymm = _mm256_lddqu_si256((__m256i const *)(h + offset_last));
        matches = _mm256_movemask_epi8(_mm256_cmpeq_epi8(h_first_vec.ymm, n_first_vec.ymm)) &
                  _mm256_movemask_epi8(_mm256_cmpeq_epi8(h_mid_vec.ymm, n_mid_vec.ymm)) &
                  _mm256_movemask_epi8(_mm256_cmpeq_epi8(h_last_vec.ymm, n_last_vec.ymm));
        while (matches) {
            int potential_offset = sz_u32_ctz(matches);
            if (sz_equal(h + potential_offset, n, n_length)) return h + potential_offset;
            matches &= matches - 1;
        }
    }

    return sz_find_serial(h, h_length, n, n_length);
}

SZ_PUBLIC sz_cptr_t sz_rfind_avx2(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n, sz_size_t n_length) {

    // This almost never fires, but it's better to be safe than sorry.
    if (h_length < n_length || !n_length) return SZ_NULL_CHAR;
    if (n_length == 1) return sz_rfind_byte_avx2(h, h_length, n);

    // Pick the parts of the needle that are worth comparing.
    sz_size_t offset_first, offset_mid, offset_last;
    _sz_locate_needle_anomalies(n, n_length, &offset_first, &offset_mid, &offset_last);

    // Broadcast those characters into YMM registers.
    int matches;
    sz_u256_vec_t h_first_vec, h_mid_vec, h_last_vec, n_first_vec, n_mid_vec, n_last_vec;
    n_first_vec.ymm = _mm256_set1_epi8(n[offset_first]);
    n_mid_vec.ymm = _mm256_set1_epi8(n[offset_mid]);
    n_last_vec.ymm = _mm256_set1_epi8(n[offset_last]);

    // Scan through the string.
    sz_cptr_t h_reversed;
    for (; h_length >= n_length + 32; h_length -= 32) {
        h_reversed = h + h_length - n_length - 32 + 1;
        h_first_vec.ymm = _mm256_lddqu_si256((__m256i const *)(h_reversed + offset_first));
        h_mid_vec.ymm = _mm256_lddqu_si256((__m256i const *)(h_reversed + offset_mid));
        h_last_vec.ymm = _mm256_lddqu_si256((__m256i const *)(h_reversed + offset_last));
        matches = _mm256_movemask_epi8(_mm256_cmpeq_epi8(h_first_vec.ymm, n_first_vec.ymm)) &
                  _mm256_movemask_epi8(_mm256_cmpeq_epi8(h_mid_vec.ymm, n_mid_vec.ymm)) &
                  _mm256_movemask_epi8(_mm256_cmpeq_epi8(h_last_vec.ymm, n_last_vec.ymm));
        while (matches) {
            int potential_offset = sz_u32_clz(matches);
            if (sz_equal(h + h_length - n_length - potential_offset, n, n_length))
                return h + h_length - n_length - potential_offset;
            matches &= ~(1 << (31 - potential_offset));
        }
    }

    return sz_rfind_serial(h, h_length, n, n_length);
}

/**
 *  @brief  There is no AVX2 instruction for fast multiplication of 64-bit integers.
 *          This implementation is coming from Agner Fog's Vector Class Library.
 */
SZ_INTERNAL __m256i _mm256_mul_epu64(__m256i a, __m256i b) {
    __m256i bswap = _mm256_shuffle_epi32(b, 0xB1);
    __m256i prodlh = _mm256_mullo_epi32(a, bswap);
    __m256i zero = _mm256_setzero_si256();
    __m256i prodlh2 = _mm256_hadd_epi32(prodlh, zero);
    __m256i prodlh3 = _mm256_shuffle_epi32(prodlh2, 0x73);
    __m256i prodll = _mm256_mul_epu32(a, b);
    __m256i prod = _mm256_add_epi64(prodll, prodlh3);
    return prod;
}

SZ_PUBLIC void sz_hashes_avx2(sz_cptr_t start, sz_size_t length, sz_size_t window_length, sz_size_t step, //
                              sz_hash_callback_t callback, void *callback_handle) {

    if (length < window_length || !window_length) return;
    if (length < 4 * window_length) {
        sz_hashes_serial(start, length, window_length, step, callback, callback_handle);
        return;
    }

    // Using AVX2, we can perform 4 long integer multiplications and additions within one register.
    // So let's slice the entire string into 4 overlapping windows, to slide over them in parallel.
    sz_size_t const max_hashes = length - window_length + 1;
    sz_size_t const min_hashes_per_thread = max_hashes / 4; // At most one sequence can overlap between 2 threads.
    sz_u8_t const *text_first = (sz_u8_t const *)start;
    sz_u8_t const *text_second = text_first + min_hashes_per_thread;
    sz_u8_t const *text_third = text_first + min_hashes_per_thread * 2;
    sz_u8_t const *text_fourth = text_first + min_hashes_per_thread * 3;
    sz_u8_t const *text_end = text_first + length;

    // Prepare the `prime ^ window_length` values, that we are going to use for modulo arithmetic.
    sz_u64_t prime_power_low = 1, prime_power_high = 1;
    for (sz_size_t i = 0; i + 1 < window_length; ++i)
        prime_power_low = (prime_power_low * 31ull) % SZ_U64_MAX_PRIME,
        prime_power_high = (prime_power_high * 257ull) % SZ_U64_MAX_PRIME;

    // Broadcast the constants into the registers.
    sz_u256_vec_t prime_vec, golden_ratio_vec;
    sz_u256_vec_t base_low_vec, base_high_vec, prime_power_low_vec, prime_power_high_vec, shift_high_vec;
    base_low_vec.ymm = _mm256_set1_epi64x(31ull);
    base_high_vec.ymm = _mm256_set1_epi64x(257ull);
    shift_high_vec.ymm = _mm256_set1_epi64x(77ull);
    prime_vec.ymm = _mm256_set1_epi64x(SZ_U64_MAX_PRIME);
    golden_ratio_vec.ymm = _mm256_set1_epi64x(11400714819323198485ull);
    prime_power_low_vec.ymm = _mm256_set1_epi64x(prime_power_low);
    prime_power_high_vec.ymm = _mm256_set1_epi64x(prime_power_high);

    // Compute the initial hash values for every one of the four windows.
    sz_u256_vec_t hash_low_vec, hash_high_vec, hash_mix_vec, chars_low_vec, chars_high_vec;
    hash_low_vec.ymm = _mm256_setzero_si256();
    hash_high_vec.ymm = _mm256_setzero_si256();
    for (sz_u8_t const *prefix_end = text_first + window_length; text_first < prefix_end;
         ++text_first, ++text_second, ++text_third, ++text_fourth) {

        // 1. Multiply the hashes by the base.
        hash_low_vec.ymm = _mm256_mul_epu64(hash_low_vec.ymm, base_low_vec.ymm);
        hash_high_vec.ymm = _mm256_mul_epu64(hash_high_vec.ymm, base_high_vec.ymm);

        // 2. Load the four characters from `text_first`, `text_first + max_hashes_per_thread`,
        //   `text_first + max_hashes_per_thread * 2`, `text_first + max_hashes_per_thread * 3`.
        chars_low_vec.ymm = _mm256_set_epi64x(text_fourth[0], text_third[0], text_second[0], text_first[0]);
        chars_high_vec.ymm = _mm256_add_epi8(chars_low_vec.ymm, shift_high_vec.ymm);

        // 3. Add the incoming characters.
        hash_low_vec.ymm = _mm256_add_epi64(hash_low_vec.ymm, chars_low_vec.ymm);
        hash_high_vec.ymm = _mm256_add_epi64(hash_high_vec.ymm, chars_high_vec.ymm);

        // 4. Compute the modulo. Assuming there are only 59 values between our prime
        //    and the 2^64 value, we can simply compute the modulo by conditionally subtracting the prime.
        hash_low_vec.ymm = _mm256_blendv_epi8(hash_low_vec.ymm, _mm256_sub_epi64(hash_low_vec.ymm, prime_vec.ymm),
                                              _mm256_cmpgt_epi64(hash_low_vec.ymm, prime_vec.ymm));
        hash_high_vec.ymm = _mm256_blendv_epi8(hash_high_vec.ymm, _mm256_sub_epi64(hash_high_vec.ymm, prime_vec.ymm),
                                               _mm256_cmpgt_epi64(hash_high_vec.ymm, prime_vec.ymm));
    }

    // 5. Compute the hash mix, that will be used to index into the fingerprint.
    //    This includes a serial step at the end.
    hash_low_vec.ymm = _mm256_mul_epu64(hash_low_vec.ymm, golden_ratio_vec.ymm);
    hash_high_vec.ymm = _mm256_mul_epu64(hash_high_vec.ymm, golden_ratio_vec.ymm);
    hash_mix_vec.ymm = _mm256_xor_si256(hash_low_vec.ymm, hash_high_vec.ymm);
    callback((sz_cptr_t)text_first, window_length, hash_mix_vec.u64s[0], callback_handle);
    callback((sz_cptr_t)text_second, window_length, hash_mix_vec.u64s[1], callback_handle);
    callback((sz_cptr_t)text_third, window_length, hash_mix_vec.u64s[2], callback_handle);
    callback((sz_cptr_t)text_fourth, window_length, hash_mix_vec.u64s[3], callback_handle);

    // Now repeat that operation for the remaining characters, discarding older characters.
    sz_size_t cycle = 1;
    sz_size_t const step_mask = step - 1;
    for (; text_fourth != text_end; ++text_first, ++text_second, ++text_third, ++text_fourth, ++cycle) {
        // 0. Load again the four characters we are dropping, shift them, and subtract.
        chars_low_vec.ymm = _mm256_set_epi64x(text_fourth[-window_length], text_third[-window_length],
                                              text_second[-window_length], text_first[-window_length]);
        chars_high_vec.ymm = _mm256_add_epi8(chars_low_vec.ymm, shift_high_vec.ymm);
        hash_low_vec.ymm =
            _mm256_sub_epi64(hash_low_vec.ymm, _mm256_mul_epu64(chars_low_vec.ymm, prime_power_low_vec.ymm));
        hash_high_vec.ymm =
            _mm256_sub_epi64(hash_high_vec.ymm, _mm256_mul_epu64(chars_high_vec.ymm, prime_power_high_vec.ymm));

        // 1. Multiply the hashes by the base.
        hash_low_vec.ymm = _mm256_mul_epu64(hash_low_vec.ymm, base_low_vec.ymm);
        hash_high_vec.ymm = _mm256_mul_epu64(hash_high_vec.ymm, base_high_vec.ymm);

        // 2. Load the four characters from `text_first`, `text_first + max_hashes_per_thread`,
        //   `text_first + max_hashes_per_thread * 2`, `text_first + max_hashes_per_thread * 3`.
        chars_low_vec.ymm = _mm256_set_epi64x(text_fourth[0], text_third[0], text_second[0], text_first[0]);
        chars_high_vec.ymm = _mm256_add_epi8(chars_low_vec.ymm, shift_high_vec.ymm);

        // 3. Add the incoming characters.
        hash_low_vec.ymm = _mm256_add_epi64(hash_low_vec.ymm, chars_low_vec.ymm);
        hash_high_vec.ymm = _mm256_add_epi64(hash_high_vec.ymm, chars_high_vec.ymm);

        // 4. Compute the modulo. Assuming there are only 59 values between our prime
        //    and the 2^64 value, we can simply compute the modulo by conditionally subtracting the prime.
        hash_low_vec.ymm = _mm256_blendv_epi8(hash_low_vec.ymm, _mm256_sub_epi64(hash_low_vec.ymm, prime_vec.ymm),
                                              _mm256_cmpgt_epi64(hash_low_vec.ymm, prime_vec.ymm));
        hash_high_vec.ymm = _mm256_blendv_epi8(hash_high_vec.ymm, _mm256_sub_epi64(hash_high_vec.ymm, prime_vec.ymm),
                                               _mm256_cmpgt_epi64(hash_high_vec.ymm, prime_vec.ymm));

        // 5. Compute the hash mix, that will be used to index into the fingerprint.
        //    This includes a serial step at the end.
        hash_low_vec.ymm = _mm256_mul_epu64(hash_low_vec.ymm, golden_ratio_vec.ymm);
        hash_high_vec.ymm = _mm256_mul_epu64(hash_high_vec.ymm, golden_ratio_vec.ymm);
        hash_mix_vec.ymm = _mm256_xor_si256(hash_low_vec.ymm, hash_high_vec.ymm);
        if ((cycle & step_mask) == 0) {
            callback((sz_cptr_t)text_first, window_length, hash_mix_vec.u64s[0], callback_handle);
            callback((sz_cptr_t)text_second, window_length, hash_mix_vec.u64s[1], callback_handle);
            callback((sz_cptr_t)text_third, window_length, hash_mix_vec.u64s[2], callback_handle);
            callback((sz_cptr_t)text_fourth, window_length, hash_mix_vec.u64s[3], callback_handle);
        }
    }
}

#pragma clang attribute pop
#pragma GCC pop_options
#endif
#pragma endregion

/*
 *  @brief  AVX-512 implementation of the string search algorithms.
 *
 *  Different subsets of AVX-512 were introduced in different years:
 *  * 2017 SkyLake: F, CD, ER, PF, VL, DQ, BW
 *  * 2018 CannonLake: IFMA, VBMI
 *  * 2019 IceLake: VPOPCNTDQ, VNNI, VBMI2, BITALG, GFNI, VPCLMULQDQ, VAES
 *  * 2020 TigerLake: VP2INTERSECT
 */
#pragma region AVX-512 Implementation

#if SZ_USE_X86_AVX512
#pragma GCC push_options
#pragma GCC target("avx", "avx512f", "avx512vl", "avx512bw", "bmi", "bmi2")
#pragma clang attribute push(__attribute__((target("avx,avx512f,avx512vl,avx512bw,bmi,bmi2"))), apply_to = function)
#include <immintrin.h>

/**
 *  @brief  Helper structure to simplify work with 512-bit registers.
 */
typedef union sz_u512_vec_t {
    __m512i zmm;
    __m256i ymms[2];
    __m128i xmms[4];
    sz_u64_t u64s[8];
    sz_u32_t u32s[16];
    sz_u16_t u16s[32];
    sz_u8_t u8s[64];
    sz_i64_t i64s[8];
    sz_i32_t i32s[16];
} sz_u512_vec_t;

SZ_INTERNAL __mmask64 _sz_u64_clamp_mask_until(sz_size_t n) {
    // The simplest approach to compute this if we know that `n` is blow or equal 64:
    //      return (1ull << n) - 1;
    // A slightly more complex approach, if we don't know that `n` is under 64:
    return _bzhi_u64(0xFFFFFFFFFFFFFFFF, n < 64 ? n : 64);
}

SZ_INTERNAL __mmask32 _sz_u32_clamp_mask_until(sz_size_t n) {
    // The simplest approach to compute this if we know that `n` is blow or equal 32:
    //      return (1ull << n) - 1;
    // A slightly more complex approach, if we don't know that `n` is under 32:
    return _bzhi_u32(0xFFFFFFFF, n < 32 ? n : 32);
}

SZ_INTERNAL __mmask16 _sz_u16_clamp_mask_until(sz_size_t n) {
    // The simplest approach to compute this if we know that `n` is blow or equal 16:
    //      return (1ull << n) - 1;
    // A slightly more complex approach, if we don't know that `n` is under 16:
    return _bzhi_u32(0xFFFFFFFF, n < 16 ? n : 16);
}

SZ_INTERNAL __mmask64 _sz_u64_mask_until(sz_size_t n) {
    // The simplest approach to compute this if we know that `n` is blow or equal 64:
    //      return (1ull << n) - 1;
    // A slightly more complex approach, if we don't know that `n` is under 64:
    return _bzhi_u64(0xFFFFFFFFFFFFFFFF, n);
}

SZ_PUBLIC sz_ordering_t sz_order_avx512(sz_cptr_t a, sz_size_t a_length, sz_cptr_t b, sz_size_t b_length) {
    sz_u512_vec_t a_vec, b_vec;
    __mmask64 a_mask, b_mask, mask_not_equal;

    // The rare case, when both string are very long.
    while ((a_length >= 64) & (b_length >= 64)) {
        a_vec.zmm = _mm512_loadu_epi8(a);
        b_vec.zmm = _mm512_loadu_epi8(b);
        mask_not_equal = _mm512_cmpneq_epi8_mask(a_vec.zmm, b_vec.zmm);
        if (mask_not_equal != 0) {
            int first_diff = _tzcnt_u64(mask_not_equal);
            char a_char = a[first_diff];
            char b_char = b[first_diff];
            return _sz_order_scalars(a_char, b_char);
        }
        a += 64, b += 64, a_length -= 64, b_length -= 64;
    }

    // In most common scenarios at least one of the strings is under 64 bytes.
    if (a_length | b_length) {
        a_mask = _sz_u64_clamp_mask_until(a_length);
        b_mask = _sz_u64_clamp_mask_until(b_length);
        a_vec.zmm = _mm512_maskz_loadu_epi8(a_mask, a);
        b_vec.zmm = _mm512_maskz_loadu_epi8(b_mask, b);
        // The AVX-512 `_mm512_mask_cmpneq_epi8_mask` intrinsics are generally handy in such environments.
        // They, however, have latency 3 on most modern CPUs. Using AVX2: `_mm256_cmpeq_epi8` would have
        // been cheaper, if we didn't have to apply `_mm256_movemask_epi8` afterwards.
        mask_not_equal = _mm512_cmpneq_epi8_mask(a_vec.zmm, b_vec.zmm);
        if (mask_not_equal != 0) {
            int first_diff = _tzcnt_u64(mask_not_equal);
            char a_char = a[first_diff];
            char b_char = b[first_diff];
            return _sz_order_scalars(a_char, b_char);
        }
        else
            // From logic perspective, the hardest cases are "abc\0" and "abc".
            // The result must be `sz_greater_k`, as the latter is shorter.
            return _sz_order_scalars(a_length, b_length);
    }
    else
        return sz_equal_k;
}

SZ_PUBLIC sz_bool_t sz_equal_avx512(sz_cptr_t a, sz_cptr_t b, sz_size_t length) {
    __mmask64 mask;
    sz_u512_vec_t a_vec, b_vec;

    while (length >= 64) {
        a_vec.zmm = _mm512_loadu_epi8(a);
        b_vec.zmm = _mm512_loadu_epi8(b);
        mask = _mm512_cmpneq_epi8_mask(a_vec.zmm, b_vec.zmm);
        if (mask != 0) return sz_false_k;
        a += 64, b += 64, length -= 64;
    }

    if (length) {
        mask = _sz_u64_mask_until(length);
        a_vec.zmm = _mm512_maskz_loadu_epi8(mask, a);
        b_vec.zmm = _mm512_maskz_loadu_epi8(mask, b);
        // Reuse the same `mask` variable to find the bit that doesn't match
        mask = _mm512_mask_cmpneq_epi8_mask(mask, a_vec.zmm, b_vec.zmm);
        return (sz_bool_t)(mask == 0);
    }
    else
        return sz_true_k;
}

SZ_PUBLIC void sz_fill_avx512(sz_ptr_t target, sz_size_t length, sz_u8_t value) {
    for (; length >= 64; target += 64, length -= 64) _mm512_storeu_epi8(target, _mm512_set1_epi8(value));
    // At this point the length is guaranteed to be under 64.
    _mm512_mask_storeu_epi8(target, _sz_u64_mask_until(length), _mm512_set1_epi8(value));
}

SZ_PUBLIC void sz_copy_avx512(sz_ptr_t target, sz_cptr_t source, sz_size_t length) {
    for (; length >= 64; target += 64, source += 64, length -= 64)
        _mm512_storeu_epi8(target, _mm512_loadu_epi8(source));
    // At this point the length is guaranteed to be under 64.
    __mmask64 mask = _sz_u64_mask_until(length);
    _mm512_mask_storeu_epi8(target, mask, _mm512_maskz_loadu_epi8(mask, source));
}

SZ_PUBLIC void sz_move_avx512(sz_ptr_t target, sz_cptr_t source, sz_size_t length) {
    if (target < source || target >= source + length) {
        for (; length >= 64; target += 64, source += 64, length -= 64)
            _mm512_storeu_epi8(target, _mm512_loadu_epi8(source));
        // At this point the length is guaranteed to be under 64.
        __mmask64 mask = _sz_u64_mask_until(length);
        _mm512_mask_storeu_epi8(target, mask, _mm512_maskz_loadu_epi8(mask, source));
    }
    else {
        // Jump to the end and walk backwards.
        for (target += length, source += length; length >= 64; length -= 64)
            _mm512_storeu_epi8(target -= 64, _mm512_loadu_epi8(source -= 64));
        // At this point the length is guaranteed to be under 64.
        __mmask64 mask = _sz_u64_mask_until(length);
        _mm512_mask_storeu_epi8(target - length, mask, _mm512_maskz_loadu_epi8(mask, source - length));
    }
}

SZ_PUBLIC sz_cptr_t sz_find_byte_avx512(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n) {
    __mmask64 mask;
    sz_u512_vec_t h_vec, n_vec;
    n_vec.zmm = _mm512_set1_epi8(n[0]);

    while (h_length >= 64) {
        h_vec.zmm = _mm512_loadu_epi8(h);
        mask = _mm512_cmpeq_epi8_mask(h_vec.zmm, n_vec.zmm);
        if (mask) return h + sz_u64_ctz(mask);
        h += 64, h_length -= 64;
    }

    if (h_length) {
        mask = _sz_u64_mask_until(h_length);
        h_vec.zmm = _mm512_maskz_loadu_epi8(mask, h);
        // Reuse the same `mask` variable to find the bit that doesn't match
        mask = _mm512_mask_cmpeq_epu8_mask(mask, h_vec.zmm, n_vec.zmm);
        if (mask) return h + sz_u64_ctz(mask);
    }

    return SZ_NULL_CHAR;
}

SZ_PUBLIC sz_cptr_t sz_find_avx512(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n, sz_size_t n_length) {

    // This almost never fires, but it's better to be safe than sorry.
    if (h_length < n_length || !n_length) return SZ_NULL_CHAR;
    if (n_length == 1) return sz_find_byte_avx512(h, h_length, n);

    // Pick the parts of the needle that are worth comparing.
    sz_size_t offset_first, offset_mid, offset_last;
    _sz_locate_needle_anomalies(n, n_length, &offset_first, &offset_mid, &offset_last);

    // Broadcast those characters into ZMM registers.
    __mmask64 matches;
    __mmask64 mask;
    sz_u512_vec_t h_first_vec, h_mid_vec, h_last_vec, n_first_vec, n_mid_vec, n_last_vec;
    n_first_vec.zmm = _mm512_set1_epi8(n[offset_first]);
    n_mid_vec.zmm = _mm512_set1_epi8(n[offset_mid]);
    n_last_vec.zmm = _mm512_set1_epi8(n[offset_last]);

    // Scan through the string.
    for (; h_length >= n_length + 64; h += 64, h_length -= 64) {
        h_first_vec.zmm = _mm512_loadu_epi8(h + offset_first);
        h_mid_vec.zmm = _mm512_loadu_epi8(h + offset_mid);
        h_last_vec.zmm = _mm512_loadu_epi8(h + offset_last);
        matches = _kand_mask64(_kand_mask64( // Intersect the masks
                                   _mm512_cmpeq_epi8_mask(h_first_vec.zmm, n_first_vec.zmm),
                                   _mm512_cmpeq_epi8_mask(h_mid_vec.zmm, n_mid_vec.zmm)),
                               _mm512_cmpeq_epi8_mask(h_last_vec.zmm, n_last_vec.zmm));
        while (matches) {
            int potential_offset = sz_u64_ctz(matches);
            if (n_length <= 3 || sz_equal_avx512(h + potential_offset, n, n_length)) return h + potential_offset;
            matches &= matches - 1;
        }

        // TODO: If the last character contains a bad byte, we can reposition the start of the next iteration.
        // This will be very helpful for very long needles.
    }

    // The "tail" of the function uses masked loads to process the remaining bytes.
    {
        mask = _sz_u64_mask_until(h_length - n_length + 1);
        h_first_vec.zmm = _mm512_maskz_loadu_epi8(mask, h + offset_first);
        h_mid_vec.zmm = _mm512_maskz_loadu_epi8(mask, h + offset_mid);
        h_last_vec.zmm = _mm512_maskz_loadu_epi8(mask, h + offset_last);
        matches = _kand_mask64(_kand_mask64( // Intersect the masks
                                   _mm512_cmpeq_epi8_mask(h_first_vec.zmm, n_first_vec.zmm),
                                   _mm512_cmpeq_epi8_mask(h_mid_vec.zmm, n_mid_vec.zmm)),
                               _mm512_cmpeq_epi8_mask(h_last_vec.zmm, n_last_vec.zmm));
        while (matches) {
            int potential_offset = sz_u64_ctz(matches);
            if (n_length <= 3 || sz_equal_avx512(h + potential_offset, n, n_length)) return h + potential_offset;
            matches &= matches - 1;
        }
    }
    return SZ_NULL_CHAR;
}

SZ_PUBLIC sz_cptr_t sz_rfind_byte_avx512(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n) {
    __mmask64 mask;
    sz_u512_vec_t h_vec, n_vec;
    n_vec.zmm = _mm512_set1_epi8(n[0]);

    while (h_length >= 64) {
        h_vec.zmm = _mm512_loadu_epi8(h + h_length - 64);
        mask = _mm512_cmpeq_epi8_mask(h_vec.zmm, n_vec.zmm);
        if (mask) return h + h_length - 1 - sz_u64_clz(mask);
        h_length -= 64;
    }

    if (h_length) {
        mask = _sz_u64_mask_until(h_length);
        h_vec.zmm = _mm512_maskz_loadu_epi8(mask, h);
        // Reuse the same `mask` variable to find the bit that doesn't match
        mask = _mm512_mask_cmpeq_epu8_mask(mask, h_vec.zmm, n_vec.zmm);
        if (mask) return h + 64 - sz_u64_clz(mask) - 1;
    }

    return SZ_NULL_CHAR;
}

SZ_PUBLIC sz_cptr_t sz_rfind_avx512(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n, sz_size_t n_length) {

    // This almost never fires, but it's better to be safe than sorry.
    if (h_length < n_length || !n_length) return SZ_NULL_CHAR;
    if (n_length == 1) return sz_rfind_byte_avx512(h, h_length, n);

    // Pick the parts of the needle that are worth comparing.
    sz_size_t offset_first, offset_mid, offset_last;
    _sz_locate_needle_anomalies(n, n_length, &offset_first, &offset_mid, &offset_last);

    // Broadcast those characters into ZMM registers.
    __mmask64 mask;
    __mmask64 matches;
    sz_u512_vec_t h_first_vec, h_mid_vec, h_last_vec, n_first_vec, n_mid_vec, n_last_vec;
    n_first_vec.zmm = _mm512_set1_epi8(n[offset_first]);
    n_mid_vec.zmm = _mm512_set1_epi8(n[offset_mid]);
    n_last_vec.zmm = _mm512_set1_epi8(n[offset_last]);

    // Scan through the string.
    sz_cptr_t h_reversed;
    for (; h_length >= n_length + 64; h_length -= 64) {
        h_reversed = h + h_length - n_length - 64 + 1;
        h_first_vec.zmm = _mm512_loadu_epi8(h_reversed + offset_first);
        h_mid_vec.zmm = _mm512_loadu_epi8(h_reversed + offset_mid);
        h_last_vec.zmm = _mm512_loadu_epi8(h_reversed + offset_last);
        matches = _kand_mask64(_kand_mask64( // Intersect the masks
                                   _mm512_cmpeq_epi8_mask(h_first_vec.zmm, n_first_vec.zmm),
                                   _mm512_cmpeq_epi8_mask(h_mid_vec.zmm, n_mid_vec.zmm)),
                               _mm512_cmpeq_epi8_mask(h_last_vec.zmm, n_last_vec.zmm));
        while (matches) {
            int potential_offset = sz_u64_clz(matches);
            if (n_length <= 3 || sz_equal_avx512(h + h_length - n_length - potential_offset, n, n_length))
                return h + h_length - n_length - potential_offset;
            sz_assert((matches & ((sz_u64_t)1 << (63 - potential_offset))) != 0 &&
                      "The bit must be set before we squash it");
            matches &= ~((sz_u64_t)1 << (63 - potential_offset));
        }
    }

    // The "tail" of the function uses masked loads to process the remaining bytes.
    {
        mask = _sz_u64_mask_until(h_length - n_length + 1);
        h_first_vec.zmm = _mm512_maskz_loadu_epi8(mask, h + offset_first);
        h_mid_vec.zmm = _mm512_maskz_loadu_epi8(mask, h + offset_mid);
        h_last_vec.zmm = _mm512_maskz_loadu_epi8(mask, h + offset_last);
        matches = _kand_mask64(_kand_mask64( // Intersect the masks
                                   _mm512_cmpeq_epi8_mask(h_first_vec.zmm, n_first_vec.zmm),
                                   _mm512_cmpeq_epi8_mask(h_mid_vec.zmm, n_mid_vec.zmm)),
                               _mm512_cmpeq_epi8_mask(h_last_vec.zmm, n_last_vec.zmm));
        while (matches) {
            int potential_offset = sz_u64_clz(matches);
            if (n_length <= 3 || sz_equal_avx512(h + 64 - potential_offset - 1, n, n_length))
                return h + 64 - potential_offset - 1;
            sz_assert((matches & ((sz_u64_t)1 << (63 - potential_offset))) != 0 &&
                      "The bit must be set before we squash it");
            matches &= ~((sz_u64_t)1 << (63 - potential_offset));
        }
    }

    return SZ_NULL_CHAR;
}

SZ_INTERNAL sz_size_t _sz_edit_distance_skewed_diagonals_upto65k_avx512( //
    sz_cptr_t shorter, sz_size_t shorter_length,                         //
    sz_cptr_t longer, sz_size_t longer_length,                           //
    sz_size_t bound, sz_memory_allocator_t *alloc) {

    // Simplify usage in higher-level libraries, where wrapping custom allocators may be troublesome.
    sz_memory_allocator_t global_alloc;
    if (!alloc) {
        sz_memory_allocator_init_default(&global_alloc);
        alloc = &global_alloc;
    }

    // TODO: Generalize!
    sz_size_t max_length = 256u * 256u;
    sz_assert(!bound && "For bounded search the method should only evaluate one band of the matrix.");
    sz_assert(shorter_length == longer_length && "The method hasn't been generalized to different length inputs yet.");
    sz_assert(shorter_length < max_length && "The length must fit into 16-bit integer. Otherwise use serial variant.");
    sz_unused(longer_length && bound && max_length);

    // We are going to store 3 diagonals of the matrix.
    // The length of the longest (main) diagonal would be `n = (shorter_length + 1)`.
    sz_size_t n = shorter_length + 1;
    // Unlike the serial version, we also want to avoid reverse-order iteration over teh shorter string.
    // So let's allocate a bit more memory and reverse-export our shorter string into that buffer.
    sz_size_t buffer_length = sizeof(sz_u16_t) * n * 3 + shorter_length;
    sz_u16_t *distances = (sz_u16_t *)alloc->allocate(buffer_length, alloc->handle);
    if (!distances) return SZ_SIZE_MAX;

    sz_u16_t *previous_distances = distances;
    sz_u16_t *current_distances = previous_distances + n;
    sz_u16_t *next_distances = current_distances + n;
    sz_ptr_t shorter_reversed = (sz_ptr_t)(next_distances + n);

    // Export the reversed string into the buffer.
    for (sz_size_t i = 0; i != shorter_length; ++i) shorter_reversed[i] = shorter[shorter_length - 1 - i];

    // Initialize the first two diagonals:
    previous_distances[0] = 0;
    current_distances[0] = current_distances[1] = 1;

    // Using ZMM registers, we can process 32x 16-bit values at once,
    // storing 16 bytes of each string in YMM registers.
    sz_u512_vec_t insertions_vec, deletions_vec, substitutions_vec, next_vec;
    sz_u512_vec_t ones_u16_vec;
    ones_u16_vec.zmm = _mm512_set1_epi16(1);
    // This is a mixed-precision implementation, using 8-bit representations for part of the operations.
    // Even there, in case `SZ_USE_X86_AVX2=0`, let's use the `sz_u512_vec_t` type, addressing the first YMM halfs.
    sz_u512_vec_t shorter_vec, longer_vec;
    sz_u512_vec_t ones_u8_vec;
    ones_u8_vec.ymms[0] = _mm256_set1_epi8(1);

    // Progress through the upper triangle of the Levenshtein matrix.
    sz_size_t next_skew_diagonal_index = 2;
    for (; next_skew_diagonal_index != n; ++next_skew_diagonal_index) {
        sz_size_t const next_skew_diagonal_length = next_skew_diagonal_index + 1;
        for (sz_size_t i = 0; i + 2 < next_skew_diagonal_length;) {
            sz_size_t remaining_length = next_skew_diagonal_length - i - 2;
            sz_size_t register_length = remaining_length < 32 ? remaining_length : 32;
            sz_u32_t remaining_length_mask = _bzhi_u32(0xFFFFFFFFu, register_length);
            longer_vec.ymms[0] = _mm256_maskz_loadu_epi8(remaining_length_mask, longer + i);
            // Our original code addressed the shorter string `[next_skew_diagonal_index - i - 2]` for growing `i`.
            // If the `shorter` string was reversed, the `[next_skew_diagonal_index - i - 2]` would
            // be equal to `[shorter_length - 1 - next_skew_diagonal_index + i + 2]`.
            // Which simplified would be equal to `[shorter_length - next_skew_diagonal_index + i + 1]`.
            shorter_vec.ymms[0] = _mm256_maskz_loadu_epi8(
                remaining_length_mask, shorter_reversed + shorter_length - next_skew_diagonal_index + i + 1);
            // For substitutions, perform the equality comparison using AVX2 instead of AVX-512
            // to get the result as a vector, instead of a bitmask. Adding 1 to every scalar we can overflow
            // transforming from {0xFF, 0} values to {0, 1} values - exactly what we need. Then - upcast to 16-bit.
            substitutions_vec.zmm = _mm512_cvtepi8_epi16( //
                _mm256_add_epi8(_mm256_cmpeq_epi8(longer_vec.ymms[0], shorter_vec.ymms[0]), ones_u8_vec.ymms[0]));
            substitutions_vec.zmm = _mm512_add_epi16( //
                substitutions_vec.zmm, _mm512_maskz_loadu_epi16(remaining_length_mask, previous_distances + i));
            // For insertions and deletions, on modern hardware, it's faster to issue two separate loads,
            // than rotate the bytes in the ZMM register.
            insertions_vec.zmm = _mm512_maskz_loadu_epi16(remaining_length_mask, current_distances + i);
            deletions_vec.zmm = _mm512_maskz_loadu_epi16(remaining_length_mask, current_distances + i + 1);
            // First get the minimum of insertions and deletions.
            next_vec.zmm = _mm512_add_epi16(_mm512_min_epu16(insertions_vec.zmm, deletions_vec.zmm), ones_u16_vec.zmm);
            next_vec.zmm = _mm512_min_epu16(next_vec.zmm, substitutions_vec.zmm);
            _mm512_mask_storeu_epi16(next_distances + i + 1, remaining_length_mask, next_vec.zmm);
            i += register_length;
        }
        // Don't forget to populate the first row and the fiest column of the Levenshtein matrix.
        next_distances[0] = next_distances[next_skew_diagonal_length - 1] = next_skew_diagonal_index;
        // Perform a circular rotarion of those buffers, to reuse the memory.
        sz_u16_t *temporary = previous_distances;
        previous_distances = current_distances;
        current_distances = next_distances;
        next_distances = temporary;
    }

    // By now we've scanned through the upper triangle of the matrix, where each subsequent iteration results in a
    // larger diagonal. From now onwards, we will be shrinking. Instead of adding value equal to the skewed diagonal
    // index on either side, we will be cropping those values out.
    sz_size_t total_diagonals = n + n - 1;
    for (; next_skew_diagonal_index != total_diagonals; ++next_skew_diagonal_index) {
        sz_size_t const next_skew_diagonal_length = total_diagonals - next_skew_diagonal_index;
        for (sz_size_t i = 0; i != next_skew_diagonal_length;) {
            sz_size_t remaining_length = next_skew_diagonal_length - i;
            sz_size_t register_length = remaining_length < 32 ? remaining_length : 32;
            sz_u32_t remaining_length_mask = _bzhi_u32(0xFFFFFFFFu, register_length);
            longer_vec.ymms[0] =
                _mm256_maskz_loadu_epi8(remaining_length_mask, longer + next_skew_diagonal_index - n + i);
            // Our original code addressed the shorter string `[shorter_length - 1 - i]` for growing `i`.
            // If the `shorter` string was reversed, the `[shorter_length - 1 - i]` would
            // be equal to `[shorter_length - 1 - shorter_length + 1 + i]`.
            // Which simplified would be equal to just `[i]`. Beautiful!
            shorter_vec.ymms[0] = _mm256_maskz_loadu_epi8(remaining_length_mask, shorter_reversed + i);
            // For substitutions, perform the equality comparison using AVX2 instead of AVX-512
            // to get the result as a vector, instead of a bitmask. The compare it against the accumulated
            // substitution costs.
            substitutions_vec.zmm = _mm512_cvtepi8_epi16( //
                _mm256_add_epi8(_mm256_cmpeq_epi8(longer_vec.ymms[0], shorter_vec.ymms[0]), ones_u8_vec.ymms[0]));
            substitutions_vec.zmm = _mm512_add_epi16( //
                substitutions_vec.zmm, _mm512_maskz_loadu_epi16(remaining_length_mask, previous_distances + i));
            // For insertions and deletions, on modern hardware, it's faster to issue two separate loads,
            // than rotate the bytes in the ZMM register.
            insertions_vec.zmm = _mm512_maskz_loadu_epi16(remaining_length_mask, current_distances + i);
            deletions_vec.zmm = _mm512_maskz_loadu_epi16(remaining_length_mask, current_distances + i + 1);
            // First get the minimum of insertions and deletions.
            next_vec.zmm = _mm512_add_epi16(_mm512_min_epu16(insertions_vec.zmm, deletions_vec.zmm), ones_u16_vec.zmm);
            next_vec.zmm = _mm512_min_epu16(next_vec.zmm, substitutions_vec.zmm);
            _mm512_mask_storeu_epi16(next_distances + i, remaining_length_mask, next_vec.zmm);
            i += register_length;
        }

        // Perform a circular rotarion of those buffers, to reuse the memory, this time, with a shift,
        // dropping the first element in the current array.
        sz_u16_t *temporary = previous_distances;
        previous_distances = current_distances + 1;
        current_distances = next_distances;
        next_distances = temporary;
    }

    // Cache scalar before `free` call.
    sz_size_t result = current_distances[0];
    alloc->free(distances, buffer_length, alloc->handle);
    return result;
}

SZ_INTERNAL sz_size_t sz_edit_distance_avx512(   //
    sz_cptr_t shorter, sz_size_t shorter_length, //
    sz_cptr_t longer, sz_size_t longer_length,   //
    sz_size_t bound, sz_memory_allocator_t *alloc) {

    if (shorter_length == longer_length && !bound && shorter_length && shorter_length < 256u * 256u)
        return _sz_edit_distance_skewed_diagonals_upto65k_avx512(shorter, shorter_length, longer, longer_length, bound,
                                                                 alloc);
    else
        return sz_edit_distance_serial(shorter, shorter_length, longer, longer_length, bound, alloc);
}

#pragma clang attribute pop
#pragma GCC pop_options

#pragma GCC push_options
#pragma GCC target("avx", "avx512f", "avx512vl", "avx512bw", "avx512dq", "bmi", "bmi2")
#pragma clang attribute push(__attribute__((target("avx,avx512f,avx512vl,avx512bw,avx512dq,bmi,bmi2"))), \
                             apply_to = function)

SZ_PUBLIC void sz_hashes_avx512(sz_cptr_t start, sz_size_t length, sz_size_t window_length, sz_size_t step, //
                                sz_hash_callback_t callback, void *callback_handle) {

    if (length < window_length || !window_length) return;
    if (length < 4 * window_length) {
        sz_hashes_serial(start, length, window_length, step, callback, callback_handle);
        return;
    }

    // Using AVX2, we can perform 4 long integer multiplications and additions within one register.
    // So let's slice the entire string into 4 overlapping windows, to slide over them in parallel.
    sz_size_t const max_hashes = length - window_length + 1;
    sz_size_t const min_hashes_per_thread = max_hashes / 4; // At most one sequence can overlap between 2 threads.
    sz_u8_t const *text_first = (sz_u8_t const *)start;
    sz_u8_t const *text_second = text_first + min_hashes_per_thread;
    sz_u8_t const *text_third = text_first + min_hashes_per_thread * 2;
    sz_u8_t const *text_fourth = text_first + min_hashes_per_thread * 3;
    sz_u8_t const *text_end = text_first + length;

    // Broadcast the global constants into the registers.
    // Both high and low hashes will work with the same prime and golden ratio.
    sz_u512_vec_t prime_vec, golden_ratio_vec;
    prime_vec.zmm = _mm512_set1_epi64(SZ_U64_MAX_PRIME);
    golden_ratio_vec.zmm = _mm512_set1_epi64(11400714819323198485ull);

    // Prepare the `prime ^ window_length` values, that we are going to use for modulo arithmetic.
    sz_u64_t prime_power_low = 1, prime_power_high = 1;
    for (sz_size_t i = 0; i + 1 < window_length; ++i)
        prime_power_low = (prime_power_low * 31ull) % SZ_U64_MAX_PRIME,
        prime_power_high = (prime_power_high * 257ull) % SZ_U64_MAX_PRIME;

    // We will be evaluating 4 offsets at a time with 2 different hash functions.
    // We can fit all those 8 state variables in each of the following ZMM registers.
    sz_u512_vec_t base_vec, prime_power_vec, shift_vec;
    base_vec.zmm = _mm512_set_epi64(31ull, 31ull, 31ull, 31ull, 257ull, 257ull, 257ull, 257ull);
    shift_vec.zmm = _mm512_set_epi64(0ull, 0ull, 0ull, 0ull, 77ull, 77ull, 77ull, 77ull);
    prime_power_vec.zmm = _mm512_set_epi64(prime_power_low, prime_power_low, prime_power_low, prime_power_low,
                                           prime_power_high, prime_power_high, prime_power_high, prime_power_high);

    // Compute the initial hash values for every one of the four windows.
    sz_u512_vec_t hash_vec, chars_vec;
    hash_vec.zmm = _mm512_setzero_si512();
    for (sz_u8_t const *prefix_end = text_first + window_length; text_first < prefix_end;
         ++text_first, ++text_second, ++text_third, ++text_fourth) {

        // 1. Multiply the hashes by the base.
        hash_vec.zmm = _mm512_mullo_epi64(hash_vec.zmm, base_vec.zmm);

        // 2. Load the four characters from `text_first`, `text_first + max_hashes_per_thread`,
        //   `text_first + max_hashes_per_thread * 2`, `text_first + max_hashes_per_thread * 3`...
        chars_vec.zmm = _mm512_set_epi64(text_fourth[0], text_third[0], text_second[0], text_first[0], //
                                         text_fourth[0], text_third[0], text_second[0], text_first[0]);
        chars_vec.zmm = _mm512_add_epi8(chars_vec.zmm, shift_vec.zmm);

        // 3. Add the incoming characters.
        hash_vec.zmm = _mm512_add_epi64(hash_vec.zmm, chars_vec.zmm);

        // 4. Compute the modulo. Assuming there are only 59 values between our prime
        //    and the 2^64 value, we can simply compute the modulo by conditionally subtracting the prime.
        hash_vec.zmm = _mm512_mask_blend_epi8(_mm512_cmpgt_epi64_mask(hash_vec.zmm, prime_vec.zmm), hash_vec.zmm,
                                              _mm512_sub_epi64(hash_vec.zmm, prime_vec.zmm));
    }

    // 5. Compute the hash mix, that will be used to index into the fingerprint.
    //    This includes a serial step at the end.
    sz_u512_vec_t hash_mix_vec;
    hash_mix_vec.zmm = _mm512_mullo_epi64(hash_vec.zmm, golden_ratio_vec.zmm);
    hash_mix_vec.ymms[0] = _mm256_xor_si256(_mm512_extracti64x4_epi64(hash_mix_vec.zmm, 1), //
                                            _mm512_extracti64x4_epi64(hash_mix_vec.zmm, 0));

    callback((sz_cptr_t)text_first, window_length, hash_mix_vec.u64s[0], callback_handle);
    callback((sz_cptr_t)text_second, window_length, hash_mix_vec.u64s[1], callback_handle);
    callback((sz_cptr_t)text_third, window_length, hash_mix_vec.u64s[2], callback_handle);
    callback((sz_cptr_t)text_fourth, window_length, hash_mix_vec.u64s[3], callback_handle);

    // Now repeat that operation for the remaining characters, discarding older characters.
    sz_size_t cycle = 1;
    sz_size_t step_mask = step - 1;
    for (; text_fourth != text_end; ++text_first, ++text_second, ++text_third, ++text_fourth, ++cycle) {
        // 0. Load again the four characters we are dropping, shift them, and subtract.
        chars_vec.zmm = _mm512_set_epi64(text_fourth[-window_length], text_third[-window_length],
                                         text_second[-window_length], text_first[-window_length], //
                                         text_fourth[-window_length], text_third[-window_length],
                                         text_second[-window_length], text_first[-window_length]);
        chars_vec.zmm = _mm512_add_epi8(chars_vec.zmm, shift_vec.zmm);
        hash_vec.zmm = _mm512_sub_epi64(hash_vec.zmm, _mm512_mullo_epi64(chars_vec.zmm, prime_power_vec.zmm));

        // 1. Multiply the hashes by the base.
        hash_vec.zmm = _mm512_mullo_epi64(hash_vec.zmm, base_vec.zmm);

        // 2. Load the four characters from `text_first`, `text_first + max_hashes_per_thread`,
        //   `text_first + max_hashes_per_thread * 2`, `text_first + max_hashes_per_thread * 3`.
        chars_vec.zmm = _mm512_set_epi64(text_fourth[0], text_third[0], text_second[0], text_first[0], //
                                         text_fourth[0], text_third[0], text_second[0], text_first[0]);
        chars_vec.zmm = _mm512_add_epi8(chars_vec.zmm, shift_vec.zmm);

        // ... and prefetch the next four characters into Level 2 or higher.
        _mm_prefetch((sz_cptr_t)text_fourth + 1, _MM_HINT_T1);
        _mm_prefetch((sz_cptr_t)text_third + 1, _MM_HINT_T1);
        _mm_prefetch((sz_cptr_t)text_second + 1, _MM_HINT_T1);
        _mm_prefetch((sz_cptr_t)text_first + 1, _MM_HINT_T1);

        // 3. Add the incoming characters.
        hash_vec.zmm = _mm512_add_epi64(hash_vec.zmm, chars_vec.zmm);

        // 4. Compute the modulo. Assuming there are only 59 values between our prime
        //    and the 2^64 value, we can simply compute the modulo by conditionally subtracting the prime.
        hash_vec.zmm = _mm512_mask_blend_epi8(_mm512_cmpgt_epi64_mask(hash_vec.zmm, prime_vec.zmm), hash_vec.zmm,
                                              _mm512_sub_epi64(hash_vec.zmm, prime_vec.zmm));

        // 5. Compute the hash mix, that will be used to index into the fingerprint.
        //    This includes a serial step at the end.
        hash_mix_vec.zmm = _mm512_mullo_epi64(hash_vec.zmm, golden_ratio_vec.zmm);
        hash_mix_vec.ymms[0] = _mm256_xor_si256(_mm512_extracti64x4_epi64(hash_mix_vec.zmm, 1), //
                                                _mm512_castsi512_si256(hash_mix_vec.zmm));

        if ((cycle & step_mask) == 0) {
            callback((sz_cptr_t)text_first, window_length, hash_mix_vec.u64s[0], callback_handle);
            callback((sz_cptr_t)text_second, window_length, hash_mix_vec.u64s[1], callback_handle);
            callback((sz_cptr_t)text_third, window_length, hash_mix_vec.u64s[2], callback_handle);
            callback((sz_cptr_t)text_fourth, window_length, hash_mix_vec.u64s[3], callback_handle);
        }
    }
}

#pragma clang attribute pop
#pragma GCC pop_options

#pragma GCC push_options
#pragma GCC target("avx", "avx512f", "avx512vl", "avx512bw", "avx512vbmi", "bmi", "bmi2", "gfni")
#pragma clang attribute push(__attribute__((target("avx,avx512f,avx512vl,avx512bw,avx512vbmi,bmi,bmi2,gfni"))), \
                             apply_to = function)

SZ_PUBLIC sz_cptr_t sz_find_charset_avx512(sz_cptr_t text, sz_size_t length, sz_charset_t const *filter) {

    sz_size_t load_length;
    __mmask32 load_mask, matches_mask;
    // To store the set in the register we need just 256 bits, but the `VPERMB` instruction
    // we are going to invoke is surprisingly cheaper on ZMM registers.
    sz_u512_vec_t text_vec, filter_vec;
    filter_vec.ymms[0] = _mm256_loadu_epi64(&filter->_u64s[0]);

    // We are going to view the `filter` at 8-bit word granularity.
    sz_u512_vec_t filter_slice_offsets_vec;
    sz_u512_vec_t filter_slice_vec;
    sz_u512_vec_t offset_within_slice_vec;
    sz_u512_vec_t mask_in_filter_slice_vec;
    sz_u512_vec_t matches_vec;

    while (length) {
        // For every byte:
        // 1. Find corresponding word in a set.
        // 2. Produce a bitmask to check against that word.
        load_length = sz_min_of_two(length, 32);
        load_mask = _sz_u64_mask_until(load_length);
        text_vec.ymms[0] = _mm256_maskz_loadu_epi8(load_mask, text);

        // To shift right every byte by 3 bits we can use the GF2 affine transformations.
        // https://wunkolo.github.io/post/2020/11/gf2p8affineqb-int8-shifting/
        // After next line, all 8-bit offsets in the `filter_slice_offsets_vec` should be under 32.
        filter_slice_offsets_vec.ymms[0] =
            _mm256_gf2p8affine_epi64_epi8(text_vec.ymms[0], _mm256_set1_epi64x(0x0102040810204080ull << (3 * 8)), 0);

        // After next line, `filter_slice_vec` will contain the right word from the set,
        // needed to filter the presence of the byte in the set.
        filter_slice_vec.ymms[0] = _mm256_permutexvar_epi8(filter_slice_offsets_vec.ymms[0], filter_vec.ymms[0]);

        // After next line, all 8-bit offsets in the `filter_slice_offsets_vec` should be under 8.
        offset_within_slice_vec.ymms[0] = _mm256_and_si256(text_vec.ymms[0], _mm256_set1_epi64x(0x0707070707070707ull));

        // Instead of performing one more Galois Field operation, we can upcast to 16-bit integers,
        // and perform the fift and intersection there.
        filter_slice_vec.zmm = _mm512_cvtepi8_epi16(filter_slice_vec.ymms[0]);
        offset_within_slice_vec.zmm = _mm512_cvtepi8_epi16(offset_within_slice_vec.ymms[0]);
        mask_in_filter_slice_vec.zmm = _mm512_sllv_epi16(_mm512_set1_epi16(1), offset_within_slice_vec.zmm);
        matches_vec.zmm = _mm512_and_si512(filter_slice_vec.zmm, mask_in_filter_slice_vec.zmm);

        matches_mask = _mm512_mask_cmpneq_epi16_mask(load_mask, matches_vec.zmm, _mm512_setzero_si512());
        if (matches_mask) {
            int offset = sz_u32_ctz(matches_mask);
            return text + offset;
        }
        else { text += load_length, length -= load_length; }
    }

    return SZ_NULL_CHAR;
}

SZ_PUBLIC sz_cptr_t sz_rfind_charset_avx512(sz_cptr_t text, sz_size_t length, sz_charset_t const *filter) {

    sz_size_t load_length;
    __mmask32 load_mask, matches_mask;
    // To store the set in the register we need just 256 bits, but the `VPERMB` instruction
    // we are going to invoke is surprisingly cheaper on ZMM registers.
    sz_u512_vec_t text_vec, filter_vec;
    filter_vec.ymms[0] = _mm256_loadu_epi64(&filter->_u64s[0]);

    // We are going to view the `filter` at 8-bit word granularity.
    sz_u512_vec_t filter_slice_offsets_vec;
    sz_u512_vec_t filter_slice_vec;
    sz_u512_vec_t offset_within_slice_vec;
    sz_u512_vec_t mask_in_filter_slice_vec;
    sz_u512_vec_t matches_vec;

    while (length) {
        // For every byte:
        // 1. Find corresponding word in a set.
        // 2. Produce a bitmask to check against that word.
        load_length = sz_min_of_two(length, 32);
        load_mask = _sz_u64_mask_until(load_length);
        text_vec.ymms[0] = _mm256_maskz_loadu_epi8(load_mask, text + length - load_length);

        // To shift right every byte by 3 bits we can use the GF2 affine transformations.
        // https://wunkolo.github.io/post/2020/11/gf2p8affineqb-int8-shifting/
        // After next line, all 8-bit offsets in the `filter_slice_offsets_vec` should be under 32.
        filter_slice_offsets_vec.ymms[0] =
            _mm256_gf2p8affine_epi64_epi8(text_vec.ymms[0], _mm256_set1_epi64x(0x0102040810204080ull << (3 * 8)), 0);

        // After next line, `filter_slice_vec` will contain the right word from the set,
        // needed to filter the presence of the byte in the set.
        filter_slice_vec.ymms[0] = _mm256_permutexvar_epi8(filter_slice_offsets_vec.ymms[0], filter_vec.ymms[0]);

        // After next line, all 8-bit offsets in the `filter_slice_offsets_vec` should be under 8.
        offset_within_slice_vec.ymms[0] = _mm256_and_si256(text_vec.ymms[0], _mm256_set1_epi64x(0x0707070707070707ull));

        // Instead of performing one more Galois Field operation, we can upcast to 16-bit integers,
        // and perform the fift and intersection there.
        filter_slice_vec.zmm = _mm512_cvtepi8_epi16(filter_slice_vec.ymms[0]);
        offset_within_slice_vec.zmm = _mm512_cvtepi8_epi16(offset_within_slice_vec.ymms[0]);
        mask_in_filter_slice_vec.zmm = _mm512_sllv_epi16(_mm512_set1_epi16(1), offset_within_slice_vec.zmm);
        matches_vec.zmm = _mm512_and_si512(filter_slice_vec.zmm, mask_in_filter_slice_vec.zmm);

        matches_mask = _mm512_mask_cmpneq_epi16_mask(load_mask, matches_vec.zmm, _mm512_setzero_si512());
        if (matches_mask) {
            int offset = sz_u32_clz(matches_mask);
            return text + length - load_length + 32 - offset - 1;
        }
        else { length -= load_length; }
    }

    return SZ_NULL_CHAR;
}

/**
 *  Computes the Needleman Wunsch alignment score between two strings.
 *  The method uses 32-bit integers to accumulate the running score for every cell in the matrix.
 *  Assuming the costs of substitutions can be arbitrary signed 8-bit integers, the method is expected to be used
 *  on strings not exceeding 2^24 length or 16.7 million characters.
 *
 *  Unlike the `_sz_edit_distance_skewed_diagonals_upto65k_avx512` method, this one uses signed integers to store
 *  the accumulated score. Moreover, it's primary bottleneck is the latency of gathering the substitution costs
 *  from the substitution matrix. If we use the diagonal order, we will be comparing a slice of the first string with
 *  a slice of the second. If we stick to the conventional horizontal order, we will be comparing one character against
 *  a slice, which is much easier to optimize. In that case we are sampling costs not from arbitrary parts of
 *  a 256 x 256 matrix, but from a single row!
 */
SZ_INTERNAL sz_ssize_t _sz_alignment_score_wagner_fisher_upto17m_avx512( //
    sz_cptr_t shorter, sz_size_t shorter_length,                         //
    sz_cptr_t longer, sz_size_t longer_length,                           //
    sz_error_cost_t const *subs, sz_error_cost_t gap, sz_memory_allocator_t *alloc) {

    // If one of the strings is empty - the edit distance is equal to the length of the other one
    if (longer_length == 0) return (sz_ssize_t)shorter_length * gap;
    if (shorter_length == 0) return (sz_ssize_t)longer_length * gap;

    // Let's make sure that we use the amount proportional to the
    // number of elements in the shorter string, not the larger.
    if (shorter_length > longer_length) {
        sz_pointer_swap((void **)&longer_length, (void **)&shorter_length);
        sz_pointer_swap((void **)&longer, (void **)&shorter);
    }

    // Simplify usage in higher-level libraries, where wrapping custom allocators may be troublesome.
    sz_memory_allocator_t global_alloc;
    if (!alloc) {
        sz_memory_allocator_init_default(&global_alloc);
        alloc = &global_alloc;
    }

    sz_size_t const max_length = 256ull * 256ull * 256ull;
    sz_size_t const n = longer_length + 1;
    sz_assert(n < max_length && "The length must fit into 24-bit integer. Otherwise use serial variant.");
    sz_unused(longer_length && max_length);

    sz_size_t buffer_length = sizeof(sz_i32_t) * n * 2;
    sz_i32_t *distances = (sz_i32_t *)alloc->allocate(buffer_length, alloc->handle);
    sz_i32_t *previous_distances = distances;
    sz_i32_t *current_distances = previous_distances + n;

    // Intialize the first row of the Levenshtein matrix with `iota`.
    for (sz_size_t idx_longer = 0; idx_longer != n; ++idx_longer)
        previous_distances[idx_longer] = (sz_ssize_t)idx_longer * gap;

    /// Contains up to 16 consecutive characters from the longer string.
    sz_u512_vec_t longer_vec;
    sz_u512_vec_t cost_deletion_vec, cost_substitution_vec, lookup_substitution_vec, current_vec;
    sz_u512_vec_t row_first_subs_vec, row_second_subs_vec, row_third_subs_vec, row_fourth_subs_vec;
    sz_u512_vec_t shuffled_first_subs_vec, shuffled_second_subs_vec, shuffled_third_subs_vec, shuffled_fourth_subs_vec;

    // Prepare constants and masks.
    sz_u512_vec_t is_third_or_fourth_vec, is_second_or_fourth_vec, gap_vec;
    {
        char is_third_or_fourth_check, is_second_or_fourth_check;
        *(sz_u8_t *)&is_third_or_fourth_check = 0x80, *(sz_u8_t *)&is_second_or_fourth_check = 0x40;
        is_third_or_fourth_vec.zmm = _mm512_set1_epi8(is_third_or_fourth_check);
        is_second_or_fourth_vec.zmm = _mm512_set1_epi8(is_second_or_fourth_check);
        gap_vec.zmm = _mm512_set1_epi32(gap);
    }

    sz_u8_t const *shorter_unsigned = (sz_u8_t const *)shorter;
    for (sz_size_t idx_shorter = 0; idx_shorter != shorter_length; ++idx_shorter) {
        sz_i32_t last_in_row = current_distances[0] = (sz_ssize_t)(idx_shorter + 1) * gap;

        // Load one row of the substitution matrix into four ZMM registers.
        sz_error_cost_t const *row_subs = subs + shorter_unsigned[idx_shorter] * 256u;
        row_first_subs_vec.zmm = _mm512_loadu_epi8(row_subs + 64 * 0);
        row_second_subs_vec.zmm = _mm512_loadu_epi8(row_subs + 64 * 1);
        row_third_subs_vec.zmm = _mm512_loadu_epi8(row_subs + 64 * 2);
        row_fourth_subs_vec.zmm = _mm512_loadu_epi8(row_subs + 64 * 3);

        // In the serial version we have one forward pass, that computes the deletion,
        // insertion, and substitution costs at once.
        //    for (sz_size_t idx_longer = 0; idx_longer < longer_length; ++idx_longer) {
        //        sz_ssize_t cost_deletion = previous_distances[idx_longer + 1] + gap;
        //        sz_ssize_t cost_insertion = current_distances[idx_longer] + gap;
        //        sz_ssize_t cost_substitution = previous_distances[idx_longer] + row_subs[longer_unsigned[idx_longer]];
        //        current_distances[idx_longer + 1] = sz_min_of_three(cost_deletion, cost_insertion, cost_substitution);
        //    }
        //
        // Given the complexity of handling the data-dependency between consecutive insertion cost computations
        // within a Levenshtein matrix, the simplest design would be to vectorize every kind of cost computation
        // separately.
        //      1. Compute substitution costs for up to 64 characters at once, upcasting from 8-bit integers to 32.
        //      2. Compute the pairwise minimum with deletion costs.
        //      3. Inclusive prefix minimum computation to combine with addition costs.
        // Proceeding with substitutions:
        for (sz_size_t idx_longer = 0; idx_longer < longer_length; idx_longer += 64) {
            sz_size_t register_length = sz_min_of_two(longer_length - idx_longer, 64);
            __mmask64 mask = _sz_u64_mask_until(register_length);
            longer_vec.zmm = _mm512_maskz_loadu_epi8(mask, longer + idx_longer);

            // Blend the `row_(first|second|third|fourth)_subs_vec` into `current_vec`, picking the right source
            // for every character in `longer_vec`. Before that, we need to permute the subsititution vectors.
            // Only the bottom 6 bits of a byte are used in VPERB, so we don't even need to mask.
            shuffled_first_subs_vec.zmm = _mm512_maskz_permutexvar_epi8(mask, longer_vec.zmm, row_first_subs_vec.zmm);
            shuffled_second_subs_vec.zmm = _mm512_maskz_permutexvar_epi8(mask, longer_vec.zmm, row_second_subs_vec.zmm);
            shuffled_third_subs_vec.zmm = _mm512_maskz_permutexvar_epi8(mask, longer_vec.zmm, row_third_subs_vec.zmm);
            shuffled_fourth_subs_vec.zmm = _mm512_maskz_permutexvar_epi8(mask, longer_vec.zmm, row_fourth_subs_vec.zmm);

            // To blend we can invoke three `_mm512_cmplt_epu8_mask`, but we can also achieve the same using
            // the AND logical operation, checking the top two bits of every byte.
            // Continuing this thought, we can use the VPTESTMB instruction to output the mask after the AND.
            __mmask64 is_third_or_fourth = _mm512_mask_test_epi8_mask(mask, longer_vec.zmm, is_third_or_fourth_vec.zmm);
            __mmask64 is_second_or_fourth =
                _mm512_mask_test_epi8_mask(mask, longer_vec.zmm, is_second_or_fourth_vec.zmm);
            lookup_substitution_vec.zmm = _mm512_mask_blend_epi8(
                is_third_or_fourth,
                // Choose between the first and the second.
                _mm512_mask_blend_epi8(is_second_or_fourth, shuffled_first_subs_vec.zmm, shuffled_second_subs_vec.zmm),
                // Choose between the third and the fourth.
                _mm512_mask_blend_epi8(is_second_or_fourth, shuffled_third_subs_vec.zmm, shuffled_fourth_subs_vec.zmm));

            // First, sign-extend lower and upper 16 bytes to 16-bit integers.
            __m512i current_0_31_vec = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(lookup_substitution_vec.zmm, 0));
            __m512i current_32_63_vec = _mm512_cvtepi8_epi16(_mm512_extracti64x4_epi64(lookup_substitution_vec.zmm, 1));

            // Now extend those 16-bit integers to 32-bit.
            // This isn't free, same as the subsequent store, so we only want to do that for the populated lanes.
            // To minimize the number of loads and stores, we can combine our substitution costs with the previous
            // distances, containing the deletion costs.
            {
                cost_substitution_vec.zmm = _mm512_maskz_loadu_epi32(mask, previous_distances + idx_longer);
                cost_substitution_vec.zmm = _mm512_add_epi32(
                    cost_substitution_vec.zmm, _mm512_cvtepi16_epi32(_mm512_extracti64x4_epi64(current_0_31_vec, 0)));
                cost_deletion_vec.zmm = _mm512_maskz_loadu_epi32(mask, previous_distances + 1 + idx_longer);
                cost_deletion_vec.zmm = _mm512_add_epi32(cost_deletion_vec.zmm, gap_vec.zmm);
                current_vec.zmm = _mm512_max_epi32(cost_substitution_vec.zmm, cost_deletion_vec.zmm);

                // Inclusive prefix minimum computation to combine with insertion costs.
                // Simply disabling this operation results in 5x performance improvement, meaning
                // that this operation is responsible for 80% of the total runtime.
                //    for (sz_size_t idx_longer = 0; idx_longer < longer_length; ++idx_longer) {
                //        current_distances[idx_longer + 1] =
                //            sz_max_of_two(current_distances[idx_longer] + gap, current_distances[idx_longer + 1]);
                //    }
                //
                // To perform the same operation in vectorized form, we need to perform a tree-like reduction,
                // that will involve multiple steps. It's quite expensive and should be first tested in the
                // "experimental" section.
                //
                // Another approach might be loop unrolling:
                //      current_vec.i32s[0] = last_in_row = sz_i32_max_of_two(current_vec.i32s[0], last_in_row + gap);
                //      current_vec.i32s[1] = last_in_row = sz_i32_max_of_two(current_vec.i32s[1], last_in_row + gap);
                //      current_vec.i32s[2] = last_in_row = sz_i32_max_of_two(current_vec.i32s[2], last_in_row + gap);
                //      ... yet this approach is also quite expensive.
                for (int i = 0; i != 16; ++i)
                    current_vec.i32s[i] = last_in_row = sz_max_of_two(current_vec.i32s[i], last_in_row + gap);
                _mm512_mask_storeu_epi32(current_distances + idx_longer + 1, mask, current_vec.zmm);
            }

            // Export the values from 16 to 31.
            if (register_length > 16) {
                mask = _kshiftri_mask64(mask, 16);
                cost_substitution_vec.zmm = _mm512_maskz_loadu_epi32(mask, previous_distances + idx_longer + 16);
                cost_substitution_vec.zmm = _mm512_add_epi32(
                    cost_substitution_vec.zmm, _mm512_cvtepi16_epi32(_mm512_extracti64x4_epi64(current_0_31_vec, 1)));
                cost_deletion_vec.zmm = _mm512_maskz_loadu_epi32(mask, previous_distances + 1 + idx_longer + 16);
                cost_deletion_vec.zmm = _mm512_add_epi32(cost_deletion_vec.zmm, gap_vec.zmm);
                current_vec.zmm = _mm512_max_epi32(cost_substitution_vec.zmm, cost_deletion_vec.zmm);

                // Aggregate running insertion costs within the register.
                for (int i = 0; i != 16; ++i)
                    current_vec.i32s[i] = last_in_row = sz_max_of_two(current_vec.i32s[i], last_in_row + gap);
                _mm512_mask_storeu_epi32(current_distances + idx_longer + 1 + 16, mask, current_vec.zmm);
            }

            // Export the values from 32 to 47.
            if (register_length > 32) {
                mask = _kshiftri_mask64(mask, 16);
                cost_substitution_vec.zmm = _mm512_maskz_loadu_epi32(mask, previous_distances + idx_longer + 32);
                cost_substitution_vec.zmm = _mm512_add_epi32(
                    cost_substitution_vec.zmm, _mm512_cvtepi16_epi32(_mm512_extracti64x4_epi64(current_32_63_vec, 0)));
                cost_deletion_vec.zmm = _mm512_maskz_loadu_epi32(mask, previous_distances + 1 + idx_longer + 32);
                cost_deletion_vec.zmm = _mm512_add_epi32(cost_deletion_vec.zmm, gap_vec.zmm);
                current_vec.zmm = _mm512_max_epi32(cost_substitution_vec.zmm, cost_deletion_vec.zmm);

                // Aggregate running insertion costs within the register.
                for (int i = 0; i != 16; ++i)
                    current_vec.i32s[i] = last_in_row = sz_max_of_two(current_vec.i32s[i], last_in_row + gap);
                _mm512_mask_storeu_epi32(current_distances + idx_longer + 1 + 32, mask, current_vec.zmm);
            }

            // Export the values from 32 to 47.
            if (register_length > 48) {
                mask = _kshiftri_mask64(mask, 16);
                cost_substitution_vec.zmm = _mm512_maskz_loadu_epi32(mask, previous_distances + idx_longer + 48);
                cost_substitution_vec.zmm = _mm512_add_epi32(
                    cost_substitution_vec.zmm, _mm512_cvtepi16_epi32(_mm512_extracti64x4_epi64(current_32_63_vec, 1)));
                cost_deletion_vec.zmm = _mm512_maskz_loadu_epi32(mask, previous_distances + 1 + idx_longer + 48);
                cost_deletion_vec.zmm = _mm512_add_epi32(cost_deletion_vec.zmm, gap_vec.zmm);
                current_vec.zmm = _mm512_max_epi32(cost_substitution_vec.zmm, cost_deletion_vec.zmm);

                // Aggregate running insertion costs within the register.
                for (int i = 0; i != 16; ++i)
                    current_vec.i32s[i] = last_in_row = sz_max_of_two(current_vec.i32s[i], last_in_row + gap);
                _mm512_mask_storeu_epi32(current_distances + idx_longer + 1 + 48, mask, current_vec.zmm);
            }
        }

        // Swap previous_distances and current_distances pointers
        sz_pointer_swap((void **)&previous_distances, (void **)&current_distances);
    }

    // Cache scalar before `free` call.
    sz_ssize_t result = previous_distances[longer_length];
    alloc->free(distances, buffer_length, alloc->handle);
    return result;
}

SZ_INTERNAL sz_ssize_t sz_alignment_score_avx512( //
    sz_cptr_t shorter, sz_size_t shorter_length,  //
    sz_cptr_t longer, sz_size_t longer_length,    //
    sz_error_cost_t const *subs, sz_error_cost_t gap, sz_memory_allocator_t *alloc) {

    if (sz_max_of_two(shorter_length, longer_length) < (256ull * 256ull * 256ull))
        return _sz_alignment_score_wagner_fisher_upto17m_avx512(shorter, shorter_length, longer, longer_length, subs,
                                                                gap, alloc);
    else
        return sz_alignment_score_serial(shorter, shorter_length, longer, longer_length, subs, gap, alloc);
}

#pragma clang attribute pop
#pragma GCC pop_options
#endif

#pragma endregion

/*  @brief  Implementation of the string search algorithms using the Arm NEON instruction set, available on 64-bit
 *          Arm processors. Implements: {substring search, character search, character set search} x {forward, reverse}.
 */
#pragma region ARM NEON

#if SZ_USE_ARM_NEON

/**
 *  @brief  Helper structure to simplify work with 64-bit words.
 */
typedef union sz_u128_vec_t {
    uint8x16_t u8x16;
    uint16x8_t u16x8;
    uint32x4_t u32x4;
    uint64x2_t u64x2;
    sz_u64_t u64s[2];
    sz_u32_t u32s[4];
    sz_u16_t u16s[8];
    sz_u8_t u8s[16];
} sz_u128_vec_t;

SZ_INTERNAL sz_u64_t vreinterpretq_u8_u4(uint8x16_t vec) {
    // Use `vshrn` to produce a bitmask, similar to `movemask` in SSE.
    // https://community.arm.com/arm-community-blogs/b/infrastructure-solutions-blog/posts/porting-x86-vector-bitmask-optimizations-to-arm-neon
    return vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(vec), 4)), 0) & 0x8888888888888888ull;
}

SZ_PUBLIC sz_cptr_t sz_find_byte_neon(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n) {
    sz_u64_t matches;
    sz_u128_vec_t h_vec, n_vec, matches_vec;
    n_vec.u8x16 = vld1q_dup_u8((sz_u8_t const *)n);

    while (h_length >= 16) {
        h_vec.u8x16 = vld1q_u8((sz_u8_t const *)h);
        matches_vec.u8x16 = vceqq_u8(h_vec.u8x16, n_vec.u8x16);
        // In Arm NEON we don't have a `movemask` to combine it with `ctz` and get the offset of the match.
        // But assuming the `vmaxvq` is cheap, we can use it to find the first match, by blending (bitwise selecting)
        // the vector with a relative offsets array.
        matches = vreinterpretq_u8_u4(matches_vec.u8x16);
        if (matches) return h + sz_u64_ctz(matches) / 4;

        h += 16, h_length -= 16;
    }

    return sz_find_byte_serial(h, h_length, n);
}

SZ_PUBLIC sz_cptr_t sz_rfind_byte_neon(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n) {
    sz_u64_t matches;
    sz_u128_vec_t h_vec, n_vec, matches_vec;
    n_vec.u8x16 = vld1q_dup_u8((sz_u8_t const *)n);

    while (h_length >= 16) {
        h_vec.u8x16 = vld1q_u8((sz_u8_t const *)h + h_length - 16);
        matches_vec.u8x16 = vceqq_u8(h_vec.u8x16, n_vec.u8x16);
        matches = vreinterpretq_u8_u4(matches_vec.u8x16);
        if (matches) return h + h_length - 1 - sz_u64_clz(matches) / 4;
        h_length -= 16;
    }

    return sz_rfind_byte_serial(h, h_length, n);
}

SZ_PUBLIC sz_u64_t _sz_find_charset_neon_register(sz_u128_vec_t h_vec, uint8x16_t set_top_vec_u8x16,
                                                  uint8x16_t set_bottom_vec_u8x16) {

    // Once we've read the characters in the haystack, we want to
    // compare them against our bitset. The serial version of that code
    // would look like: `(set_->_u8s[c >> 3] & (1u << (c & 7u))) != 0`.
    uint8x16_t byte_index_vec = vshrq_n_u8(h_vec.u8x16, 3);
    uint8x16_t byte_mask_vec = vshlq_u8(vdupq_n_u8(1), vreinterpretq_s8_u8(vandq_u8(h_vec.u8x16, vdupq_n_u8(7))));
    uint8x16_t matches_top_vec = vqtbl1q_u8(set_top_vec_u8x16, byte_index_vec);
    // The table lookup instruction in NEON replies to out-of-bound requests with zeros.
    // The values in `byte_index_vec` all fall in [0; 32). So for values under 16, substracting 16 will underflow
    // and map into interval [240, 256). Meaning that those will be populated with zeros and we can safely
    // merge `matches_top_vec` and `matches_bottom_vec` with a bitwise OR.
    uint8x16_t matches_bottom_vec = vqtbl1q_u8(set_bottom_vec_u8x16, vsubq_u8(byte_index_vec, vdupq_n_u8(16)));
    uint8x16_t matches_vec = vorrq_u8(matches_top_vec, matches_bottom_vec);
    // Istead of pure `vandq_u8`, we can immediately broadcast a match presence across each 8-bit word.
    matches_vec = vtstq_u8(matches_vec, byte_mask_vec);
    return vreinterpretq_u8_u4(matches_vec);
}

SZ_PUBLIC sz_cptr_t sz_find_neon(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n, sz_size_t n_length) {

    // This almost never fires, but it's better to be safe than sorry.
    if (h_length < n_length || !n_length) return SZ_NULL_CHAR;
    if (n_length == 1) return sz_find_byte_neon(h, h_length, n);

    // Scan through the string.
    // Assuming how tiny the Arm NEON registers are, we should avoid internal branches at all costs.
    // That's why, for smaller needles, we use different loops.
    if (n_length == 2) {
        // Broadcast needle characters into SIMD registers.
        sz_u64_t matches;
        sz_u128_vec_t h_first_vec, h_last_vec, n_first_vec, n_last_vec, matches_vec;
        // Dealing with 16-bit values, we can load 2 registers at a time and compare 31 possible offsets
        // in a single loop iteration.
        n_first_vec.u8x16 = vld1q_dup_u8((sz_u8_t const *)&n[0]);
        n_last_vec.u8x16 = vld1q_dup_u8((sz_u8_t const *)&n[1]);
        for (; h_length >= 17; h += 16, h_length -= 16) {
            h_first_vec.u8x16 = vld1q_u8((sz_u8_t const *)(h + 0));
            h_last_vec.u8x16 = vld1q_u8((sz_u8_t const *)(h + 1));
            matches_vec.u8x16 =
                vandq_u8(vceqq_u8(h_first_vec.u8x16, n_first_vec.u8x16), vceqq_u8(h_last_vec.u8x16, n_last_vec.u8x16));
            matches = vreinterpretq_u8_u4(matches_vec.u8x16);
            if (matches) return h + sz_u64_ctz(matches) / 4;
        }
    }
    else if (n_length == 3) {
        // Broadcast needle characters into SIMD registers.
        sz_u64_t matches;
        sz_u128_vec_t h_first_vec, h_mid_vec, h_last_vec, n_first_vec, n_mid_vec, n_last_vec, matches_vec;
        // Comparing 24-bit values is a bumer. Being lazy, I went with the same approach
        // as when searching for string over 4 characters long. I only avoid the last comparison.
        n_first_vec.u8x16 = vld1q_dup_u8((sz_u8_t const *)&n[0]);
        n_mid_vec.u8x16 = vld1q_dup_u8((sz_u8_t const *)&n[1]);
        n_last_vec.u8x16 = vld1q_dup_u8((sz_u8_t const *)&n[2]);
        for (; h_length >= 18; h += 16, h_length -= 16) {
            h_first_vec.u8x16 = vld1q_u8((sz_u8_t const *)(h + 0));
            h_mid_vec.u8x16 = vld1q_u8((sz_u8_t const *)(h + 1));
            h_last_vec.u8x16 = vld1q_u8((sz_u8_t const *)(h + 2));
            matches_vec.u8x16 = vandq_u8(                           //
                vandq_u8(                                           //
                    vceqq_u8(h_first_vec.u8x16, n_first_vec.u8x16), //
                    vceqq_u8(h_mid_vec.u8x16, n_mid_vec.u8x16)),
                vceqq_u8(h_last_vec.u8x16, n_last_vec.u8x16));
            matches = vreinterpretq_u8_u4(matches_vec.u8x16);
            if (matches) return h + sz_u64_ctz(matches) / 4;
        }
    }
    else {
        // Pick the parts of the needle that are worth comparing.
        sz_size_t offset_first, offset_mid, offset_last;
        _sz_locate_needle_anomalies(n, n_length, &offset_first, &offset_mid, &offset_last);
        // Broadcast those characters into SIMD registers.
        sz_u64_t matches;
        sz_u128_vec_t h_first_vec, h_mid_vec, h_last_vec, n_first_vec, n_mid_vec, n_last_vec, matches_vec;
        n_first_vec.u8x16 = vld1q_dup_u8((sz_u8_t const *)&n[offset_first]);
        n_mid_vec.u8x16 = vld1q_dup_u8((sz_u8_t const *)&n[offset_mid]);
        n_last_vec.u8x16 = vld1q_dup_u8((sz_u8_t const *)&n[offset_last]);
        // Walk through the string.
        for (; h_length >= n_length + 16; h += 16, h_length -= 16) {
            h_first_vec.u8x16 = vld1q_u8((sz_u8_t const *)(h + offset_first));
            h_mid_vec.u8x16 = vld1q_u8((sz_u8_t const *)(h + offset_mid));
            h_last_vec.u8x16 = vld1q_u8((sz_u8_t const *)(h + offset_last));
            matches_vec.u8x16 = vandq_u8(                           //
                vandq_u8(                                           //
                    vceqq_u8(h_first_vec.u8x16, n_first_vec.u8x16), //
                    vceqq_u8(h_mid_vec.u8x16, n_mid_vec.u8x16)),
                vceqq_u8(h_last_vec.u8x16, n_last_vec.u8x16));
            matches = vreinterpretq_u8_u4(matches_vec.u8x16);
            while (matches) {
                int potential_offset = sz_u64_ctz(matches) / 4;
                if (sz_equal(h + potential_offset, n, n_length)) return h + potential_offset;
                matches &= matches - 1;
            }
        }
    }

    return sz_find_serial(h, h_length, n, n_length);
}

SZ_PUBLIC sz_cptr_t sz_rfind_neon(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n, sz_size_t n_length) {

    // This almost never fires, but it's better to be safe than sorry.
    if (h_length < n_length || !n_length) return SZ_NULL_CHAR;
    if (n_length == 1) return sz_rfind_byte_neon(h, h_length, n);

    // Pick the parts of the needle that are worth comparing.
    sz_size_t offset_first, offset_mid, offset_last;
    _sz_locate_needle_anomalies(n, n_length, &offset_first, &offset_mid, &offset_last);

    // Will contain 4 bits per character.
    sz_u64_t matches;
    sz_u128_vec_t h_first_vec, h_mid_vec, h_last_vec, n_first_vec, n_mid_vec, n_last_vec, matches_vec;
    n_first_vec.u8x16 = vld1q_dup_u8((sz_u8_t const *)&n[offset_first]);
    n_mid_vec.u8x16 = vld1q_dup_u8((sz_u8_t const *)&n[offset_mid]);
    n_last_vec.u8x16 = vld1q_dup_u8((sz_u8_t const *)&n[offset_last]);

    sz_cptr_t h_reversed;
    for (; h_length >= n_length + 16; h_length -= 16) {
        h_reversed = h + h_length - n_length - 16 + 1;
        h_first_vec.u8x16 = vld1q_u8((sz_u8_t const *)(h_reversed + offset_first));
        h_mid_vec.u8x16 = vld1q_u8((sz_u8_t const *)(h_reversed + offset_mid));
        h_last_vec.u8x16 = vld1q_u8((sz_u8_t const *)(h_reversed + offset_last));
        matches_vec.u8x16 = vandq_u8(                           //
            vandq_u8(                                           //
                vceqq_u8(h_first_vec.u8x16, n_first_vec.u8x16), //
                vceqq_u8(h_mid_vec.u8x16, n_mid_vec.u8x16)),
            vceqq_u8(h_last_vec.u8x16, n_last_vec.u8x16));
        matches = vreinterpretq_u8_u4(matches_vec.u8x16);
        while (matches) {
            int potential_offset = sz_u64_clz(matches) / 4;
            if (sz_equal(h + h_length - n_length - potential_offset, n, n_length))
                return h + h_length - n_length - potential_offset;
            sz_assert((matches & (1ull << (63 - potential_offset * 4))) != 0 &&
                      "The bit must be set before we squash it");
            matches &= ~(1ull << (63 - potential_offset * 4));
        }
    }

    return sz_rfind_serial(h, h_length, n, n_length);
}

SZ_PUBLIC sz_cptr_t sz_find_charset_neon(sz_cptr_t h, sz_size_t h_length, sz_charset_t const *set) {
    sz_u64_t matches;
    sz_u128_vec_t h_vec;
    uint8x16_t set_top_vec_u8x16 = vld1q_u8(&set->_u8s[0]);
    uint8x16_t set_bottom_vec_u8x16 = vld1q_u8(&set->_u8s[16]);

    for (; h_length >= 16; h += 16, h_length -= 16) {
        h_vec.u8x16 = vld1q_u8((sz_u8_t const *)(h));
        matches = _sz_find_charset_neon_register(h_vec, set_top_vec_u8x16, set_bottom_vec_u8x16);
        if (matches) return h + sz_u64_ctz(matches) / 4;
    }

    return sz_find_charset_serial(h, h_length, set);
}

SZ_PUBLIC sz_cptr_t sz_rfind_charset_neon(sz_cptr_t h, sz_size_t h_length, sz_charset_t const *set) {
    sz_u64_t matches;
    sz_u128_vec_t h_vec;
    uint8x16_t set_top_vec_u8x16 = vld1q_u8(&set->_u8s[0]);
    uint8x16_t set_bottom_vec_u8x16 = vld1q_u8(&set->_u8s[16]);

    // Check `sz_find_charset_neon` for explanations.
    for (; h_length >= 16; h_length -= 16) {
        h_vec.u8x16 = vld1q_u8((sz_u8_t const *)(h) + h_length - 16);
        matches = _sz_find_charset_neon_register(h_vec, set_top_vec_u8x16, set_bottom_vec_u8x16);
        if (matches) return h + h_length - 1 - sz_u64_clz(matches) / 4;
    }

    return sz_rfind_charset_serial(h, h_length, set);
}

#endif // Arm Neon

#pragma endregion

/*
 *  @brief  Pick the right implementation for the string search algorithms.
 */
#pragma region Compile-Time Dispatching

SZ_PUBLIC sz_u64_t sz_hash(sz_cptr_t ins, sz_size_t length) { return sz_hash_serial(ins, length); }
SZ_PUBLIC void sz_tolower(sz_cptr_t ins, sz_size_t length, sz_ptr_t outs) { sz_tolower_serial(ins, length, outs); }
SZ_PUBLIC void sz_toupper(sz_cptr_t ins, sz_size_t length, sz_ptr_t outs) { sz_toupper_serial(ins, length, outs); }
SZ_PUBLIC void sz_toascii(sz_cptr_t ins, sz_size_t length, sz_ptr_t outs) { sz_toascii_serial(ins, length, outs); }
SZ_PUBLIC sz_bool_t sz_isascii(sz_cptr_t ins, sz_size_t length) { return sz_isascii_serial(ins, length); }

SZ_PUBLIC void sz_hashes_fingerprint(sz_cptr_t start, sz_size_t length, sz_size_t window_length, sz_ptr_t fingerprint,
                                     sz_size_t fingerprint_bytes) {

    sz_bool_t fingerprint_length_is_power_of_two = (sz_bool_t)((fingerprint_bytes & (fingerprint_bytes - 1)) == 0);
    sz_string_view_t fingerprint_buffer = {fingerprint, fingerprint_bytes};

    // There are several issues related to the fingerprinting algorithm.
    // First, the memory traversal order is important.
    // https://blog.stuffedcow.net/2015/08/pagewalk-coherence/

    // In most cases the fingerprint length will be a power of two.
    if (fingerprint_length_is_power_of_two == sz_false_k)
        sz_hashes(start, length, window_length, 1, _sz_hashes_fingerprint_non_pow2_callback, &fingerprint_buffer);
    else
        sz_hashes(start, length, window_length, 1, _sz_hashes_fingerprint_pow2_callback, &fingerprint_buffer);
}

#if !SZ_DYNAMIC_DISPATCH

SZ_DYNAMIC sz_bool_t sz_equal(sz_cptr_t a, sz_cptr_t b, sz_size_t length) {
#if SZ_USE_X86_AVX512
    return sz_equal_avx512(a, b, length);
#else
    return sz_equal_serial(a, b, length);
#endif
}

SZ_DYNAMIC sz_ordering_t sz_order(sz_cptr_t a, sz_size_t a_length, sz_cptr_t b, sz_size_t b_length) {
#if SZ_USE_X86_AVX512
    return sz_order_avx512(a, a_length, b, b_length);
#else
    return sz_order_serial(a, a_length, b, b_length);
#endif
}

SZ_DYNAMIC void sz_copy(sz_ptr_t target, sz_cptr_t source, sz_size_t length) {
#if SZ_USE_X86_AVX512
    sz_copy_avx512(target, source, length);
#elif SZ_USE_X86_AVX2
    sz_copy_avx2(target, source, length);
#else
    sz_copy_serial(target, source, length);
#endif
}

SZ_DYNAMIC void sz_move(sz_ptr_t target, sz_cptr_t source, sz_size_t length) {
#if SZ_USE_X86_AVX512
    sz_move_avx512(target, source, length);
#elif SZ_USE_X86_AVX2
    sz_move_avx2(target, source, length);
#else
    sz_move_serial(target, source, length);
#endif
}

SZ_DYNAMIC void sz_fill(sz_ptr_t target, sz_size_t length, sz_u8_t value) {
#if SZ_USE_X86_AVX512
    sz_fill_avx512(target, length, value);
#elif SZ_USE_X86_AVX2
    sz_fill_avx2(target, length, value);
#else
    sz_fill_serial(target, length, value);
#endif
}

SZ_DYNAMIC sz_cptr_t sz_find_byte(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle) {
#if SZ_USE_X86_AVX512
    return sz_find_byte_avx512(haystack, h_length, needle);
#elif SZ_USE_X86_AVX2
    return sz_find_byte_avx2(haystack, h_length, needle);
#elif SZ_USE_ARM_NEON
    return sz_find_byte_neon(haystack, h_length, needle);
#else
    return sz_find_byte_serial(haystack, h_length, needle);
#endif
}

SZ_DYNAMIC sz_cptr_t sz_rfind_byte(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle) {
#if SZ_USE_X86_AVX512
    return sz_rfind_byte_avx512(haystack, h_length, needle);
#elif SZ_USE_X86_AVX2
    return sz_rfind_byte_avx2(haystack, h_length, needle);
#elif SZ_USE_ARM_NEON
    return sz_rfind_byte_neon(haystack, h_length, needle);
#else
    return sz_rfind_byte_serial(haystack, h_length, needle);
#endif
}

SZ_DYNAMIC sz_cptr_t sz_find(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle, sz_size_t n_length) {
#if SZ_USE_X86_AVX512
    return sz_find_avx512(haystack, h_length, needle, n_length);
#elif SZ_USE_X86_AVX2
    return sz_find_avx2(haystack, h_length, needle, n_length);
#elif SZ_USE_ARM_NEON
    return sz_find_neon(haystack, h_length, needle, n_length);
#else
    return sz_find_serial(haystack, h_length, needle, n_length);
#endif
}

SZ_DYNAMIC sz_cptr_t sz_rfind(sz_cptr_t haystack, sz_size_t h_length, sz_cptr_t needle, sz_size_t n_length) {
#if SZ_USE_X86_AVX512
    return sz_rfind_avx512(haystack, h_length, needle, n_length);
#elif SZ_USE_X86_AVX2
    return sz_rfind_avx2(haystack, h_length, needle, n_length);
#elif SZ_USE_ARM_NEON
    return sz_rfind_neon(haystack, h_length, needle, n_length);
#else
    return sz_rfind_serial(haystack, h_length, needle, n_length);
#endif
}

SZ_DYNAMIC sz_cptr_t sz_find_charset(sz_cptr_t text, sz_size_t length, sz_charset_t const *set) {
#if SZ_USE_X86_AVX512
    return sz_find_charset_avx512(text, length, set);
#elif SZ_USE_ARM_NEON
    return sz_find_charset_neon(text, length, set);
#else
    return sz_find_charset_serial(text, length, set);
#endif
}

SZ_DYNAMIC sz_cptr_t sz_rfind_charset(sz_cptr_t text, sz_size_t length, sz_charset_t const *set) {
#if SZ_USE_X86_AVX512
    return sz_rfind_charset_avx512(text, length, set);
#elif SZ_USE_ARM_NEON
    return sz_rfind_charset_neon(text, length, set);
#else
    return sz_rfind_charset_serial(text, length, set);
#endif
}

SZ_DYNAMIC sz_size_t sz_hamming_distance( //
    sz_cptr_t a, sz_size_t a_length,      //
    sz_cptr_t b, sz_size_t b_length,      //
    sz_size_t bound) {
    return sz_hamming_distance_serial(a, a_length, b, b_length, bound);
}

SZ_DYNAMIC sz_size_t sz_hamming_distance_utf8( //
    sz_cptr_t a, sz_size_t a_length,           //
    sz_cptr_t b, sz_size_t b_length,           //
    sz_size_t bound) {
    return sz_hamming_distance_utf8_serial(a, a_length, b, b_length, bound);
}

SZ_DYNAMIC sz_size_t sz_edit_distance( //
    sz_cptr_t a, sz_size_t a_length,   //
    sz_cptr_t b, sz_size_t b_length,   //
    sz_size_t bound, sz_memory_allocator_t *alloc) {
#if SZ_USE_X86_AVX512
    return sz_edit_distance_avx512(a, a_length, b, b_length, bound, alloc);
#else
    return sz_edit_distance_serial(a, a_length, b, b_length, bound, alloc);
#endif
}

SZ_DYNAMIC sz_size_t sz_edit_distance_utf8( //
    sz_cptr_t a, sz_size_t a_length,        //
    sz_cptr_t b, sz_size_t b_length,        //
    sz_size_t bound, sz_memory_allocator_t *alloc) {
    return _sz_edit_distance_wagner_fisher_serial(a, a_length, b, b_length, bound, sz_true_k, alloc);
}

SZ_DYNAMIC sz_ssize_t sz_alignment_score(sz_cptr_t a, sz_size_t a_length, sz_cptr_t b, sz_size_t b_length,
                                         sz_error_cost_t const *subs, sz_error_cost_t gap,
                                         sz_memory_allocator_t *alloc) {
#if SZ_USE_X86_AVX512
    return sz_alignment_score_avx512(a, a_length, b, b_length, subs, gap, alloc);
#else
    return sz_alignment_score_serial(a, a_length, b, b_length, subs, gap, alloc);
#endif
}

SZ_DYNAMIC void sz_hashes(sz_cptr_t text, sz_size_t length, sz_size_t window_length, sz_size_t window_step, //
                          sz_hash_callback_t callback, void *callback_handle) {
#if SZ_USE_X86_AVX512
    sz_hashes_avx512(text, length, window_length, window_step, callback, callback_handle);
#elif SZ_USE_X86_AVX2
    sz_hashes_avx2(text, length, window_length, window_step, callback, callback_handle);
#else
    sz_hashes_serial(text, length, window_length, window_step, callback, callback_handle);
#endif
}

SZ_DYNAMIC sz_cptr_t sz_find_char_from(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n, sz_size_t n_length) {
    sz_charset_t set;
    sz_charset_init(&set);
    for (; n_length; ++n, --n_length) sz_charset_add(&set, *n);
    return sz_find_charset(h, h_length, &set);
}

SZ_DYNAMIC sz_cptr_t sz_find_char_not_from(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n, sz_size_t n_length) {
    sz_charset_t set;
    sz_charset_init(&set);
    for (; n_length; ++n, --n_length) sz_charset_add(&set, *n);
    sz_charset_invert(&set);
    return sz_find_charset(h, h_length, &set);
}

SZ_DYNAMIC sz_cptr_t sz_rfind_char_from(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n, sz_size_t n_length) {
    sz_charset_t set;
    sz_charset_init(&set);
    for (; n_length; ++n, --n_length) sz_charset_add(&set, *n);
    return sz_rfind_charset(h, h_length, &set);
}

SZ_DYNAMIC sz_cptr_t sz_rfind_char_not_from(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n, sz_size_t n_length) {
    sz_charset_t set;
    sz_charset_init(&set);
    for (; n_length; ++n, --n_length) sz_charset_add(&set, *n);
    sz_charset_invert(&set);
    return sz_rfind_charset(h, h_length, &set);
}

SZ_DYNAMIC void sz_generate(sz_cptr_t alphabet, sz_size_t alphabet_size, sz_ptr_t result, sz_size_t result_length,
                            sz_random_generator_t generator, void *generator_user_data) {
    sz_generate_serial(alphabet, alphabet_size, result, result_length, generator, generator_user_data);
}

#endif
#pragma endregion

#ifdef __cplusplus
#pragma GCC diagnostic pop
}
#endif // __cplusplus

#endif // STRINGZILLA_H_