summaryrefslogtreecommitdiffstats
path: root/third_party/zstd/lib/common
diff options
context:
space:
mode:
Diffstat (limited to 'third_party/zstd/lib/common')
-rw-r--r--third_party/zstd/lib/common/allocations.h55
-rw-r--r--third_party/zstd/lib/common/bits.h200
-rw-r--r--third_party/zstd/lib/common/bitstream.h457
-rw-r--r--third_party/zstd/lib/common/compiler.h450
-rw-r--r--third_party/zstd/lib/common/cpu.h249
-rw-r--r--third_party/zstd/lib/common/debug.c30
-rw-r--r--third_party/zstd/lib/common/debug.h116
-rw-r--r--third_party/zstd/lib/common/entropy_common.c340
-rw-r--r--third_party/zstd/lib/common/error_private.c63
-rw-r--r--third_party/zstd/lib/common/error_private.h168
-rw-r--r--third_party/zstd/lib/common/fse.h640
-rw-r--r--third_party/zstd/lib/common/fse_decompress.c313
-rw-r--r--third_party/zstd/lib/common/huf.h286
-rw-r--r--third_party/zstd/lib/common/mem.h426
-rw-r--r--third_party/zstd/lib/common/pool.c371
-rw-r--r--third_party/zstd/lib/common/pool.h90
-rw-r--r--third_party/zstd/lib/common/portability_macros.h158
-rw-r--r--third_party/zstd/lib/common/threading.c182
-rw-r--r--third_party/zstd/lib/common/threading.h150
-rw-r--r--third_party/zstd/lib/common/xxhash.c18
-rw-r--r--third_party/zstd/lib/common/xxhash.h7020
-rw-r--r--third_party/zstd/lib/common/zstd_common.c48
-rw-r--r--third_party/zstd/lib/common/zstd_deps.h111
-rw-r--r--third_party/zstd/lib/common/zstd_internal.h392
-rw-r--r--third_party/zstd/lib/common/zstd_trace.h163
25 files changed, 12496 insertions, 0 deletions
diff --git a/third_party/zstd/lib/common/allocations.h b/third_party/zstd/lib/common/allocations.h
new file mode 100644
index 0000000000..5e89955010
--- /dev/null
+++ b/third_party/zstd/lib/common/allocations.h
@@ -0,0 +1,55 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+/* This file provides custom allocation primitives
+ */
+
+#define ZSTD_DEPS_NEED_MALLOC
+#include "zstd_deps.h" /* ZSTD_malloc, ZSTD_calloc, ZSTD_free, ZSTD_memset */
+
+#include "compiler.h" /* MEM_STATIC */
+#define ZSTD_STATIC_LINKING_ONLY
+#include "../zstd.h" /* ZSTD_customMem */
+
+#ifndef ZSTD_ALLOCATIONS_H
+#define ZSTD_ALLOCATIONS_H
+
+/* custom memory allocation functions */
+
+MEM_STATIC void* ZSTD_customMalloc(size_t size, ZSTD_customMem customMem)
+{
+ if (customMem.customAlloc)
+ return customMem.customAlloc(customMem.opaque, size);
+ return ZSTD_malloc(size);
+}
+
+MEM_STATIC void* ZSTD_customCalloc(size_t size, ZSTD_customMem customMem)
+{
+ if (customMem.customAlloc) {
+ /* calloc implemented as malloc+memset;
+ * not as efficient as calloc, but next best guess for custom malloc */
+ void* const ptr = customMem.customAlloc(customMem.opaque, size);
+ ZSTD_memset(ptr, 0, size);
+ return ptr;
+ }
+ return ZSTD_calloc(1, size);
+}
+
+MEM_STATIC void ZSTD_customFree(void* ptr, ZSTD_customMem customMem)
+{
+ if (ptr!=NULL) {
+ if (customMem.customFree)
+ customMem.customFree(customMem.opaque, ptr);
+ else
+ ZSTD_free(ptr);
+ }
+}
+
+#endif /* ZSTD_ALLOCATIONS_H */
diff --git a/third_party/zstd/lib/common/bits.h b/third_party/zstd/lib/common/bits.h
new file mode 100644
index 0000000000..def56c474c
--- /dev/null
+++ b/third_party/zstd/lib/common/bits.h
@@ -0,0 +1,200 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_BITS_H
+#define ZSTD_BITS_H
+
+#include "mem.h"
+
+MEM_STATIC unsigned ZSTD_countTrailingZeros32_fallback(U32 val)
+{
+ assert(val != 0);
+ {
+ static const U32 DeBruijnBytePos[32] = {0, 1, 28, 2, 29, 14, 24, 3,
+ 30, 22, 20, 15, 25, 17, 4, 8,
+ 31, 27, 13, 23, 21, 19, 16, 7,
+ 26, 12, 18, 6, 11, 5, 10, 9};
+ return DeBruijnBytePos[((U32) ((val & -(S32) val) * 0x077CB531U)) >> 27];
+ }
+}
+
+MEM_STATIC unsigned ZSTD_countTrailingZeros32(U32 val)
+{
+ assert(val != 0);
+# if defined(_MSC_VER)
+# if STATIC_BMI2 == 1
+ return (unsigned)_tzcnt_u32(val);
+# else
+ if (val != 0) {
+ unsigned long r;
+ _BitScanForward(&r, val);
+ return (unsigned)r;
+ } else {
+ /* Should not reach this code path */
+ __assume(0);
+ }
+# endif
+# elif defined(__GNUC__) && (__GNUC__ >= 4)
+ return (unsigned)__builtin_ctz(val);
+# else
+ return ZSTD_countTrailingZeros32_fallback(val);
+# endif
+}
+
+MEM_STATIC unsigned ZSTD_countLeadingZeros32_fallback(U32 val) {
+ assert(val != 0);
+ {
+ static const U32 DeBruijnClz[32] = {0, 9, 1, 10, 13, 21, 2, 29,
+ 11, 14, 16, 18, 22, 25, 3, 30,
+ 8, 12, 20, 28, 15, 17, 24, 7,
+ 19, 27, 23, 6, 26, 5, 4, 31};
+ val |= val >> 1;
+ val |= val >> 2;
+ val |= val >> 4;
+ val |= val >> 8;
+ val |= val >> 16;
+ return 31 - DeBruijnClz[(val * 0x07C4ACDDU) >> 27];
+ }
+}
+
+MEM_STATIC unsigned ZSTD_countLeadingZeros32(U32 val)
+{
+ assert(val != 0);
+# if defined(_MSC_VER)
+# if STATIC_BMI2 == 1
+ return (unsigned)_lzcnt_u32(val);
+# else
+ if (val != 0) {
+ unsigned long r;
+ _BitScanReverse(&r, val);
+ return (unsigned)(31 - r);
+ } else {
+ /* Should not reach this code path */
+ __assume(0);
+ }
+# endif
+# elif defined(__GNUC__) && (__GNUC__ >= 4)
+ return (unsigned)__builtin_clz(val);
+# else
+ return ZSTD_countLeadingZeros32_fallback(val);
+# endif
+}
+
+MEM_STATIC unsigned ZSTD_countTrailingZeros64(U64 val)
+{
+ assert(val != 0);
+# if defined(_MSC_VER) && defined(_WIN64)
+# if STATIC_BMI2 == 1
+ return (unsigned)_tzcnt_u64(val);
+# else
+ if (val != 0) {
+ unsigned long r;
+ _BitScanForward64(&r, val);
+ return (unsigned)r;
+ } else {
+ /* Should not reach this code path */
+ __assume(0);
+ }
+# endif
+# elif defined(__GNUC__) && (__GNUC__ >= 4) && defined(__LP64__)
+ return (unsigned)__builtin_ctzll(val);
+# else
+ {
+ U32 mostSignificantWord = (U32)(val >> 32);
+ U32 leastSignificantWord = (U32)val;
+ if (leastSignificantWord == 0) {
+ return 32 + ZSTD_countTrailingZeros32(mostSignificantWord);
+ } else {
+ return ZSTD_countTrailingZeros32(leastSignificantWord);
+ }
+ }
+# endif
+}
+
+MEM_STATIC unsigned ZSTD_countLeadingZeros64(U64 val)
+{
+ assert(val != 0);
+# if defined(_MSC_VER) && defined(_WIN64)
+# if STATIC_BMI2 == 1
+ return (unsigned)_lzcnt_u64(val);
+# else
+ if (val != 0) {
+ unsigned long r;
+ _BitScanReverse64(&r, val);
+ return (unsigned)(63 - r);
+ } else {
+ /* Should not reach this code path */
+ __assume(0);
+ }
+# endif
+# elif defined(__GNUC__) && (__GNUC__ >= 4)
+ return (unsigned)(__builtin_clzll(val));
+# else
+ {
+ U32 mostSignificantWord = (U32)(val >> 32);
+ U32 leastSignificantWord = (U32)val;
+ if (mostSignificantWord == 0) {
+ return 32 + ZSTD_countLeadingZeros32(leastSignificantWord);
+ } else {
+ return ZSTD_countLeadingZeros32(mostSignificantWord);
+ }
+ }
+# endif
+}
+
+MEM_STATIC unsigned ZSTD_NbCommonBytes(size_t val)
+{
+ if (MEM_isLittleEndian()) {
+ if (MEM_64bits()) {
+ return ZSTD_countTrailingZeros64((U64)val) >> 3;
+ } else {
+ return ZSTD_countTrailingZeros32((U32)val) >> 3;
+ }
+ } else { /* Big Endian CPU */
+ if (MEM_64bits()) {
+ return ZSTD_countLeadingZeros64((U64)val) >> 3;
+ } else {
+ return ZSTD_countLeadingZeros32((U32)val) >> 3;
+ }
+ }
+}
+
+MEM_STATIC unsigned ZSTD_highbit32(U32 val) /* compress, dictBuilder, decodeCorpus */
+{
+ assert(val != 0);
+ return 31 - ZSTD_countLeadingZeros32(val);
+}
+
+/* ZSTD_rotateRight_*():
+ * Rotates a bitfield to the right by "count" bits.
+ * https://en.wikipedia.org/w/index.php?title=Circular_shift&oldid=991635599#Implementing_circular_shifts
+ */
+MEM_STATIC
+U64 ZSTD_rotateRight_U64(U64 const value, U32 count) {
+ assert(count < 64);
+ count &= 0x3F; /* for fickle pattern recognition */
+ return (value >> count) | (U64)(value << ((0U - count) & 0x3F));
+}
+
+MEM_STATIC
+U32 ZSTD_rotateRight_U32(U32 const value, U32 count) {
+ assert(count < 32);
+ count &= 0x1F; /* for fickle pattern recognition */
+ return (value >> count) | (U32)(value << ((0U - count) & 0x1F));
+}
+
+MEM_STATIC
+U16 ZSTD_rotateRight_U16(U16 const value, U32 count) {
+ assert(count < 16);
+ count &= 0x0F; /* for fickle pattern recognition */
+ return (value >> count) | (U16)(value << ((0U - count) & 0x0F));
+}
+
+#endif /* ZSTD_BITS_H */
diff --git a/third_party/zstd/lib/common/bitstream.h b/third_party/zstd/lib/common/bitstream.h
new file mode 100644
index 0000000000..676044989c
--- /dev/null
+++ b/third_party/zstd/lib/common/bitstream.h
@@ -0,0 +1,457 @@
+/* ******************************************************************
+ * bitstream
+ * Part of FSE library
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ *
+ * You can contact the author at :
+ * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+****************************************************************** */
+#ifndef BITSTREAM_H_MODULE
+#define BITSTREAM_H_MODULE
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+/*
+* This API consists of small unitary functions, which must be inlined for best performance.
+* Since link-time-optimization is not available for all compilers,
+* these functions are defined into a .h to be included.
+*/
+
+/*-****************************************
+* Dependencies
+******************************************/
+#include "mem.h" /* unaligned access routines */
+#include "compiler.h" /* UNLIKELY() */
+#include "debug.h" /* assert(), DEBUGLOG(), RAWLOG() */
+#include "error_private.h" /* error codes and messages */
+#include "bits.h" /* ZSTD_highbit32 */
+
+
+/*=========================================
+* Target specific
+=========================================*/
+#ifndef ZSTD_NO_INTRINSICS
+# if (defined(__BMI__) || defined(__BMI2__)) && defined(__GNUC__)
+# include <immintrin.h> /* support for bextr (experimental)/bzhi */
+# elif defined(__ICCARM__)
+# include <intrinsics.h>
+# endif
+#endif
+
+#define STREAM_ACCUMULATOR_MIN_32 25
+#define STREAM_ACCUMULATOR_MIN_64 57
+#define STREAM_ACCUMULATOR_MIN ((U32)(MEM_32bits() ? STREAM_ACCUMULATOR_MIN_32 : STREAM_ACCUMULATOR_MIN_64))
+
+
+/*-******************************************
+* bitStream encoding API (write forward)
+********************************************/
+/* bitStream can mix input from multiple sources.
+ * A critical property of these streams is that they encode and decode in **reverse** direction.
+ * So the first bit sequence you add will be the last to be read, like a LIFO stack.
+ */
+typedef struct {
+ size_t bitContainer;
+ unsigned bitPos;
+ char* startPtr;
+ char* ptr;
+ char* endPtr;
+} BIT_CStream_t;
+
+MEM_STATIC size_t BIT_initCStream(BIT_CStream_t* bitC, void* dstBuffer, size_t dstCapacity);
+MEM_STATIC void BIT_addBits(BIT_CStream_t* bitC, size_t value, unsigned nbBits);
+MEM_STATIC void BIT_flushBits(BIT_CStream_t* bitC);
+MEM_STATIC size_t BIT_closeCStream(BIT_CStream_t* bitC);
+
+/* Start with initCStream, providing the size of buffer to write into.
+* bitStream will never write outside of this buffer.
+* `dstCapacity` must be >= sizeof(bitD->bitContainer), otherwise @return will be an error code.
+*
+* bits are first added to a local register.
+* Local register is size_t, hence 64-bits on 64-bits systems, or 32-bits on 32-bits systems.
+* Writing data into memory is an explicit operation, performed by the flushBits function.
+* Hence keep track how many bits are potentially stored into local register to avoid register overflow.
+* After a flushBits, a maximum of 7 bits might still be stored into local register.
+*
+* Avoid storing elements of more than 24 bits if you want compatibility with 32-bits bitstream readers.
+*
+* Last operation is to close the bitStream.
+* The function returns the final size of CStream in bytes.
+* If data couldn't fit into `dstBuffer`, it will return a 0 ( == not storable)
+*/
+
+
+/*-********************************************
+* bitStream decoding API (read backward)
+**********************************************/
+typedef size_t BitContainerType;
+typedef struct {
+ BitContainerType bitContainer;
+ unsigned bitsConsumed;
+ const char* ptr;
+ const char* start;
+ const char* limitPtr;
+} BIT_DStream_t;
+
+typedef enum { BIT_DStream_unfinished = 0, /* fully refilled */
+ BIT_DStream_endOfBuffer = 1, /* still some bits left in bitstream */
+ BIT_DStream_completed = 2, /* bitstream entirely consumed, bit-exact */
+ BIT_DStream_overflow = 3 /* user requested more bits than present in bitstream */
+ } BIT_DStream_status; /* result of BIT_reloadDStream() */
+
+MEM_STATIC size_t BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, size_t srcSize);
+MEM_STATIC size_t BIT_readBits(BIT_DStream_t* bitD, unsigned nbBits);
+MEM_STATIC BIT_DStream_status BIT_reloadDStream(BIT_DStream_t* bitD);
+MEM_STATIC unsigned BIT_endOfDStream(const BIT_DStream_t* bitD);
+
+
+/* Start by invoking BIT_initDStream().
+* A chunk of the bitStream is then stored into a local register.
+* Local register size is 64-bits on 64-bits systems, 32-bits on 32-bits systems (BitContainerType).
+* You can then retrieve bitFields stored into the local register, **in reverse order**.
+* Local register is explicitly reloaded from memory by the BIT_reloadDStream() method.
+* A reload guarantee a minimum of ((8*sizeof(bitD->bitContainer))-7) bits when its result is BIT_DStream_unfinished.
+* Otherwise, it can be less than that, so proceed accordingly.
+* Checking if DStream has reached its end can be performed with BIT_endOfDStream().
+*/
+
+
+/*-****************************************
+* unsafe API
+******************************************/
+MEM_STATIC void BIT_addBitsFast(BIT_CStream_t* bitC, size_t value, unsigned nbBits);
+/* faster, but works only if value is "clean", meaning all high bits above nbBits are 0 */
+
+MEM_STATIC void BIT_flushBitsFast(BIT_CStream_t* bitC);
+/* unsafe version; does not check buffer overflow */
+
+MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, unsigned nbBits);
+/* faster, but works only if nbBits >= 1 */
+
+/*===== Local Constants =====*/
+static const unsigned BIT_mask[] = {
+ 0, 1, 3, 7, 0xF, 0x1F,
+ 0x3F, 0x7F, 0xFF, 0x1FF, 0x3FF, 0x7FF,
+ 0xFFF, 0x1FFF, 0x3FFF, 0x7FFF, 0xFFFF, 0x1FFFF,
+ 0x3FFFF, 0x7FFFF, 0xFFFFF, 0x1FFFFF, 0x3FFFFF, 0x7FFFFF,
+ 0xFFFFFF, 0x1FFFFFF, 0x3FFFFFF, 0x7FFFFFF, 0xFFFFFFF, 0x1FFFFFFF,
+ 0x3FFFFFFF, 0x7FFFFFFF}; /* up to 31 bits */
+#define BIT_MASK_SIZE (sizeof(BIT_mask) / sizeof(BIT_mask[0]))
+
+/*-**************************************************************
+* bitStream encoding
+****************************************************************/
+/*! BIT_initCStream() :
+ * `dstCapacity` must be > sizeof(size_t)
+ * @return : 0 if success,
+ * otherwise an error code (can be tested using ERR_isError()) */
+MEM_STATIC size_t BIT_initCStream(BIT_CStream_t* bitC,
+ void* startPtr, size_t dstCapacity)
+{
+ bitC->bitContainer = 0;
+ bitC->bitPos = 0;
+ bitC->startPtr = (char*)startPtr;
+ bitC->ptr = bitC->startPtr;
+ bitC->endPtr = bitC->startPtr + dstCapacity - sizeof(bitC->bitContainer);
+ if (dstCapacity <= sizeof(bitC->bitContainer)) return ERROR(dstSize_tooSmall);
+ return 0;
+}
+
+FORCE_INLINE_TEMPLATE size_t BIT_getLowerBits(size_t bitContainer, U32 const nbBits)
+{
+#if defined(STATIC_BMI2) && STATIC_BMI2 == 1 && !defined(ZSTD_NO_INTRINSICS)
+ return _bzhi_u64(bitContainer, nbBits);
+#else
+ assert(nbBits < BIT_MASK_SIZE);
+ return bitContainer & BIT_mask[nbBits];
+#endif
+}
+
+/*! BIT_addBits() :
+ * can add up to 31 bits into `bitC`.
+ * Note : does not check for register overflow ! */
+MEM_STATIC void BIT_addBits(BIT_CStream_t* bitC,
+ size_t value, unsigned nbBits)
+{
+ DEBUG_STATIC_ASSERT(BIT_MASK_SIZE == 32);
+ assert(nbBits < BIT_MASK_SIZE);
+ assert(nbBits + bitC->bitPos < sizeof(bitC->bitContainer) * 8);
+ bitC->bitContainer |= BIT_getLowerBits(value, nbBits) << bitC->bitPos;
+ bitC->bitPos += nbBits;
+}
+
+/*! BIT_addBitsFast() :
+ * works only if `value` is _clean_,
+ * meaning all high bits above nbBits are 0 */
+MEM_STATIC void BIT_addBitsFast(BIT_CStream_t* bitC,
+ size_t value, unsigned nbBits)
+{
+ assert((value>>nbBits) == 0);
+ assert(nbBits + bitC->bitPos < sizeof(bitC->bitContainer) * 8);
+ bitC->bitContainer |= value << bitC->bitPos;
+ bitC->bitPos += nbBits;
+}
+
+/*! BIT_flushBitsFast() :
+ * assumption : bitContainer has not overflowed
+ * unsafe version; does not check buffer overflow */
+MEM_STATIC void BIT_flushBitsFast(BIT_CStream_t* bitC)
+{
+ size_t const nbBytes = bitC->bitPos >> 3;
+ assert(bitC->bitPos < sizeof(bitC->bitContainer) * 8);
+ assert(bitC->ptr <= bitC->endPtr);
+ MEM_writeLEST(bitC->ptr, bitC->bitContainer);
+ bitC->ptr += nbBytes;
+ bitC->bitPos &= 7;
+ bitC->bitContainer >>= nbBytes*8;
+}
+
+/*! BIT_flushBits() :
+ * assumption : bitContainer has not overflowed
+ * safe version; check for buffer overflow, and prevents it.
+ * note : does not signal buffer overflow.
+ * overflow will be revealed later on using BIT_closeCStream() */
+MEM_STATIC void BIT_flushBits(BIT_CStream_t* bitC)
+{
+ size_t const nbBytes = bitC->bitPos >> 3;
+ assert(bitC->bitPos < sizeof(bitC->bitContainer) * 8);
+ assert(bitC->ptr <= bitC->endPtr);
+ MEM_writeLEST(bitC->ptr, bitC->bitContainer);
+ bitC->ptr += nbBytes;
+ if (bitC->ptr > bitC->endPtr) bitC->ptr = bitC->endPtr;
+ bitC->bitPos &= 7;
+ bitC->bitContainer >>= nbBytes*8;
+}
+
+/*! BIT_closeCStream() :
+ * @return : size of CStream, in bytes,
+ * or 0 if it could not fit into dstBuffer */
+MEM_STATIC size_t BIT_closeCStream(BIT_CStream_t* bitC)
+{
+ BIT_addBitsFast(bitC, 1, 1); /* endMark */
+ BIT_flushBits(bitC);
+ if (bitC->ptr >= bitC->endPtr) return 0; /* overflow detected */
+ return (bitC->ptr - bitC->startPtr) + (bitC->bitPos > 0);
+}
+
+
+/*-********************************************************
+* bitStream decoding
+**********************************************************/
+/*! BIT_initDStream() :
+ * Initialize a BIT_DStream_t.
+ * `bitD` : a pointer to an already allocated BIT_DStream_t structure.
+ * `srcSize` must be the *exact* size of the bitStream, in bytes.
+ * @return : size of stream (== srcSize), or an errorCode if a problem is detected
+ */
+MEM_STATIC size_t BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, size_t srcSize)
+{
+ if (srcSize < 1) { ZSTD_memset(bitD, 0, sizeof(*bitD)); return ERROR(srcSize_wrong); }
+
+ bitD->start = (const char*)srcBuffer;
+ bitD->limitPtr = bitD->start + sizeof(bitD->bitContainer);
+
+ if (srcSize >= sizeof(bitD->bitContainer)) { /* normal case */
+ bitD->ptr = (const char*)srcBuffer + srcSize - sizeof(bitD->bitContainer);
+ bitD->bitContainer = MEM_readLEST(bitD->ptr);
+ { BYTE const lastByte = ((const BYTE*)srcBuffer)[srcSize-1];
+ bitD->bitsConsumed = lastByte ? 8 - ZSTD_highbit32(lastByte) : 0; /* ensures bitsConsumed is always set */
+ if (lastByte == 0) return ERROR(GENERIC); /* endMark not present */ }
+ } else {
+ bitD->ptr = bitD->start;
+ bitD->bitContainer = *(const BYTE*)(bitD->start);
+ switch(srcSize)
+ {
+ case 7: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[6]) << (sizeof(bitD->bitContainer)*8 - 16);
+ ZSTD_FALLTHROUGH;
+
+ case 6: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[5]) << (sizeof(bitD->bitContainer)*8 - 24);
+ ZSTD_FALLTHROUGH;
+
+ case 5: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[4]) << (sizeof(bitD->bitContainer)*8 - 32);
+ ZSTD_FALLTHROUGH;
+
+ case 4: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[3]) << 24;
+ ZSTD_FALLTHROUGH;
+
+ case 3: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[2]) << 16;
+ ZSTD_FALLTHROUGH;
+
+ case 2: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[1]) << 8;
+ ZSTD_FALLTHROUGH;
+
+ default: break;
+ }
+ { BYTE const lastByte = ((const BYTE*)srcBuffer)[srcSize-1];
+ bitD->bitsConsumed = lastByte ? 8 - ZSTD_highbit32(lastByte) : 0;
+ if (lastByte == 0) return ERROR(corruption_detected); /* endMark not present */
+ }
+ bitD->bitsConsumed += (U32)(sizeof(bitD->bitContainer) - srcSize)*8;
+ }
+
+ return srcSize;
+}
+
+FORCE_INLINE_TEMPLATE size_t BIT_getUpperBits(BitContainerType bitContainer, U32 const start)
+{
+ return bitContainer >> start;
+}
+
+FORCE_INLINE_TEMPLATE size_t BIT_getMiddleBits(BitContainerType bitContainer, U32 const start, U32 const nbBits)
+{
+ U32 const regMask = sizeof(bitContainer)*8 - 1;
+ /* if start > regMask, bitstream is corrupted, and result is undefined */
+ assert(nbBits < BIT_MASK_SIZE);
+ /* x86 transform & ((1 << nbBits) - 1) to bzhi instruction, it is better
+ * than accessing memory. When bmi2 instruction is not present, we consider
+ * such cpus old (pre-Haswell, 2013) and their performance is not of that
+ * importance.
+ */
+#if defined(__x86_64__) || defined(_M_X86)
+ return (bitContainer >> (start & regMask)) & ((((U64)1) << nbBits) - 1);
+#else
+ return (bitContainer >> (start & regMask)) & BIT_mask[nbBits];
+#endif
+}
+
+/*! BIT_lookBits() :
+ * Provides next n bits from local register.
+ * local register is not modified.
+ * On 32-bits, maxNbBits==24.
+ * On 64-bits, maxNbBits==56.
+ * @return : value extracted */
+FORCE_INLINE_TEMPLATE size_t BIT_lookBits(const BIT_DStream_t* bitD, U32 nbBits)
+{
+ /* arbitrate between double-shift and shift+mask */
+#if 1
+ /* if bitD->bitsConsumed + nbBits > sizeof(bitD->bitContainer)*8,
+ * bitstream is likely corrupted, and result is undefined */
+ return BIT_getMiddleBits(bitD->bitContainer, (sizeof(bitD->bitContainer)*8) - bitD->bitsConsumed - nbBits, nbBits);
+#else
+ /* this code path is slower on my os-x laptop */
+ U32 const regMask = sizeof(bitD->bitContainer)*8 - 1;
+ return ((bitD->bitContainer << (bitD->bitsConsumed & regMask)) >> 1) >> ((regMask-nbBits) & regMask);
+#endif
+}
+
+/*! BIT_lookBitsFast() :
+ * unsafe version; only works if nbBits >= 1 */
+MEM_STATIC size_t BIT_lookBitsFast(const BIT_DStream_t* bitD, U32 nbBits)
+{
+ U32 const regMask = sizeof(bitD->bitContainer)*8 - 1;
+ assert(nbBits >= 1);
+ return (bitD->bitContainer << (bitD->bitsConsumed & regMask)) >> (((regMask+1)-nbBits) & regMask);
+}
+
+FORCE_INLINE_TEMPLATE void BIT_skipBits(BIT_DStream_t* bitD, U32 nbBits)
+{
+ bitD->bitsConsumed += nbBits;
+}
+
+/*! BIT_readBits() :
+ * Read (consume) next n bits from local register and update.
+ * Pay attention to not read more than nbBits contained into local register.
+ * @return : extracted value. */
+FORCE_INLINE_TEMPLATE size_t BIT_readBits(BIT_DStream_t* bitD, unsigned nbBits)
+{
+ size_t const value = BIT_lookBits(bitD, nbBits);
+ BIT_skipBits(bitD, nbBits);
+ return value;
+}
+
+/*! BIT_readBitsFast() :
+ * unsafe version; only works if nbBits >= 1 */
+MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, unsigned nbBits)
+{
+ size_t const value = BIT_lookBitsFast(bitD, nbBits);
+ assert(nbBits >= 1);
+ BIT_skipBits(bitD, nbBits);
+ return value;
+}
+
+/*! BIT_reloadDStream_internal() :
+ * Simple variant of BIT_reloadDStream(), with two conditions:
+ * 1. bitstream is valid : bitsConsumed <= sizeof(bitD->bitContainer)*8
+ * 2. look window is valid after shifted down : bitD->ptr >= bitD->start
+ */
+MEM_STATIC BIT_DStream_status BIT_reloadDStream_internal(BIT_DStream_t* bitD)
+{
+ assert(bitD->bitsConsumed <= sizeof(bitD->bitContainer)*8);
+ bitD->ptr -= bitD->bitsConsumed >> 3;
+ assert(bitD->ptr >= bitD->start);
+ bitD->bitsConsumed &= 7;
+ bitD->bitContainer = MEM_readLEST(bitD->ptr);
+ return BIT_DStream_unfinished;
+}
+
+/*! BIT_reloadDStreamFast() :
+ * Similar to BIT_reloadDStream(), but with two differences:
+ * 1. bitsConsumed <= sizeof(bitD->bitContainer)*8 must hold!
+ * 2. Returns BIT_DStream_overflow when bitD->ptr < bitD->limitPtr, at this
+ * point you must use BIT_reloadDStream() to reload.
+ */
+MEM_STATIC BIT_DStream_status BIT_reloadDStreamFast(BIT_DStream_t* bitD)
+{
+ if (UNLIKELY(bitD->ptr < bitD->limitPtr))
+ return BIT_DStream_overflow;
+ return BIT_reloadDStream_internal(bitD);
+}
+
+/*! BIT_reloadDStream() :
+ * Refill `bitD` from buffer previously set in BIT_initDStream() .
+ * This function is safe, it guarantees it will not never beyond src buffer.
+ * @return : status of `BIT_DStream_t` internal register.
+ * when status == BIT_DStream_unfinished, internal register is filled with at least 25 or 57 bits */
+FORCE_INLINE_TEMPLATE BIT_DStream_status BIT_reloadDStream(BIT_DStream_t* bitD)
+{
+ /* note : once in overflow mode, a bitstream remains in this mode until it's reset */
+ if (UNLIKELY(bitD->bitsConsumed > (sizeof(bitD->bitContainer)*8))) {
+ static const BitContainerType zeroFilled = 0;
+ bitD->ptr = (const char*)&zeroFilled; /* aliasing is allowed for char */
+ /* overflow detected, erroneous scenario or end of stream: no update */
+ return BIT_DStream_overflow;
+ }
+
+ assert(bitD->ptr >= bitD->start);
+
+ if (bitD->ptr >= bitD->limitPtr) {
+ return BIT_reloadDStream_internal(bitD);
+ }
+ if (bitD->ptr == bitD->start) {
+ /* reached end of bitStream => no update */
+ if (bitD->bitsConsumed < sizeof(bitD->bitContainer)*8) return BIT_DStream_endOfBuffer;
+ return BIT_DStream_completed;
+ }
+ /* start < ptr < limitPtr => cautious update */
+ { U32 nbBytes = bitD->bitsConsumed >> 3;
+ BIT_DStream_status result = BIT_DStream_unfinished;
+ if (bitD->ptr - nbBytes < bitD->start) {
+ nbBytes = (U32)(bitD->ptr - bitD->start); /* ptr > start */
+ result = BIT_DStream_endOfBuffer;
+ }
+ bitD->ptr -= nbBytes;
+ bitD->bitsConsumed -= nbBytes*8;
+ bitD->bitContainer = MEM_readLEST(bitD->ptr); /* reminder : srcSize > sizeof(bitD->bitContainer), otherwise bitD->ptr == bitD->start */
+ return result;
+ }
+}
+
+/*! BIT_endOfDStream() :
+ * @return : 1 if DStream has _exactly_ reached its end (all bits consumed).
+ */
+MEM_STATIC unsigned BIT_endOfDStream(const BIT_DStream_t* DStream)
+{
+ return ((DStream->ptr == DStream->start) && (DStream->bitsConsumed == sizeof(DStream->bitContainer)*8));
+}
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* BITSTREAM_H_MODULE */
diff --git a/third_party/zstd/lib/common/compiler.h b/third_party/zstd/lib/common/compiler.h
new file mode 100644
index 0000000000..31880ecbe1
--- /dev/null
+++ b/third_party/zstd/lib/common/compiler.h
@@ -0,0 +1,450 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_COMPILER_H
+#define ZSTD_COMPILER_H
+
+#include <stddef.h>
+
+#include "portability_macros.h"
+
+/*-*******************************************************
+* Compiler specifics
+*********************************************************/
+/* force inlining */
+
+#if !defined(ZSTD_NO_INLINE)
+#if (defined(__GNUC__) && !defined(__STRICT_ANSI__)) || defined(__cplusplus) || defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L /* C99 */
+# define INLINE_KEYWORD inline
+#else
+# define INLINE_KEYWORD
+#endif
+
+#if defined(__GNUC__) || defined(__ICCARM__)
+# define FORCE_INLINE_ATTR __attribute__((always_inline))
+#elif defined(_MSC_VER)
+# define FORCE_INLINE_ATTR __forceinline
+#else
+# define FORCE_INLINE_ATTR
+#endif
+
+#else
+
+#define INLINE_KEYWORD
+#define FORCE_INLINE_ATTR
+
+#endif
+
+/**
+ On MSVC qsort requires that functions passed into it use the __cdecl calling conversion(CC).
+ This explicitly marks such functions as __cdecl so that the code will still compile
+ if a CC other than __cdecl has been made the default.
+*/
+#if defined(_MSC_VER)
+# define WIN_CDECL __cdecl
+#else
+# define WIN_CDECL
+#endif
+
+/* UNUSED_ATTR tells the compiler it is okay if the function is unused. */
+#if defined(__GNUC__)
+# define UNUSED_ATTR __attribute__((unused))
+#else
+# define UNUSED_ATTR
+#endif
+
+/**
+ * FORCE_INLINE_TEMPLATE is used to define C "templates", which take constant
+ * parameters. They must be inlined for the compiler to eliminate the constant
+ * branches.
+ */
+#define FORCE_INLINE_TEMPLATE static INLINE_KEYWORD FORCE_INLINE_ATTR UNUSED_ATTR
+/**
+ * HINT_INLINE is used to help the compiler generate better code. It is *not*
+ * used for "templates", so it can be tweaked based on the compilers
+ * performance.
+ *
+ * gcc-4.8 and gcc-4.9 have been shown to benefit from leaving off the
+ * always_inline attribute.
+ *
+ * clang up to 5.0.0 (trunk) benefit tremendously from the always_inline
+ * attribute.
+ */
+#if !defined(__clang__) && defined(__GNUC__) && __GNUC__ >= 4 && __GNUC_MINOR__ >= 8 && __GNUC__ < 5
+# define HINT_INLINE static INLINE_KEYWORD
+#else
+# define HINT_INLINE FORCE_INLINE_TEMPLATE
+#endif
+
+/* "soft" inline :
+ * The compiler is free to select if it's a good idea to inline or not.
+ * The main objective is to silence compiler warnings
+ * when a defined function in included but not used.
+ *
+ * Note : this macro is prefixed `MEM_` because it used to be provided by `mem.h` unit.
+ * Updating the prefix is probably preferable, but requires a fairly large codemod,
+ * since this name is used everywhere.
+ */
+#ifndef MEM_STATIC /* already defined in Linux Kernel mem.h */
+#if defined(__GNUC__)
+# define MEM_STATIC static __inline UNUSED_ATTR
+#elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+# define MEM_STATIC static inline
+#elif defined(_MSC_VER)
+# define MEM_STATIC static __inline
+#else
+# define MEM_STATIC static /* this version may generate warnings for unused static functions; disable the relevant warning */
+#endif
+#endif
+
+/* force no inlining */
+#ifdef _MSC_VER
+# define FORCE_NOINLINE static __declspec(noinline)
+#else
+# if defined(__GNUC__) || defined(__ICCARM__)
+# define FORCE_NOINLINE static __attribute__((__noinline__))
+# else
+# define FORCE_NOINLINE static
+# endif
+#endif
+
+
+/* target attribute */
+#if defined(__GNUC__) || defined(__ICCARM__)
+# define TARGET_ATTRIBUTE(target) __attribute__((__target__(target)))
+#else
+# define TARGET_ATTRIBUTE(target)
+#endif
+
+/* Target attribute for BMI2 dynamic dispatch.
+ * Enable lzcnt, bmi, and bmi2.
+ * We test for bmi1 & bmi2. lzcnt is included in bmi1.
+ */
+#define BMI2_TARGET_ATTRIBUTE TARGET_ATTRIBUTE("lzcnt,bmi,bmi2")
+
+/* prefetch
+ * can be disabled, by declaring NO_PREFETCH build macro */
+#if defined(NO_PREFETCH)
+# define PREFETCH_L1(ptr) do { (void)(ptr); } while (0) /* disabled */
+# define PREFETCH_L2(ptr) do { (void)(ptr); } while (0) /* disabled */
+#else
+# if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_I86)) && !defined(_M_ARM64EC) /* _mm_prefetch() is not defined outside of x86/x64 */
+# include <mmintrin.h> /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
+# define PREFETCH_L1(ptr) _mm_prefetch((const char*)(ptr), _MM_HINT_T0)
+# define PREFETCH_L2(ptr) _mm_prefetch((const char*)(ptr), _MM_HINT_T1)
+# elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) )
+# define PREFETCH_L1(ptr) __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */)
+# define PREFETCH_L2(ptr) __builtin_prefetch((ptr), 0 /* rw==read */, 2 /* locality */)
+# elif defined(__aarch64__)
+# define PREFETCH_L1(ptr) do { __asm__ __volatile__("prfm pldl1keep, %0" ::"Q"(*(ptr))); } while (0)
+# define PREFETCH_L2(ptr) do { __asm__ __volatile__("prfm pldl2keep, %0" ::"Q"(*(ptr))); } while (0)
+# else
+# define PREFETCH_L1(ptr) do { (void)(ptr); } while (0) /* disabled */
+# define PREFETCH_L2(ptr) do { (void)(ptr); } while (0) /* disabled */
+# endif
+#endif /* NO_PREFETCH */
+
+#define CACHELINE_SIZE 64
+
+#define PREFETCH_AREA(p, s) \
+ do { \
+ const char* const _ptr = (const char*)(p); \
+ size_t const _size = (size_t)(s); \
+ size_t _pos; \
+ for (_pos=0; _pos<_size; _pos+=CACHELINE_SIZE) { \
+ PREFETCH_L2(_ptr + _pos); \
+ } \
+ } while (0)
+
+/* vectorization
+ * older GCC (pre gcc-4.3 picked as the cutoff) uses a different syntax,
+ * and some compilers, like Intel ICC and MCST LCC, do not support it at all. */
+#if !defined(__INTEL_COMPILER) && !defined(__clang__) && defined(__GNUC__) && !defined(__LCC__)
+# if (__GNUC__ == 4 && __GNUC_MINOR__ > 3) || (__GNUC__ >= 5)
+# define DONT_VECTORIZE __attribute__((optimize("no-tree-vectorize")))
+# else
+# define DONT_VECTORIZE _Pragma("GCC optimize(\"no-tree-vectorize\")")
+# endif
+#else
+# define DONT_VECTORIZE
+#endif
+
+/* Tell the compiler that a branch is likely or unlikely.
+ * Only use these macros if it causes the compiler to generate better code.
+ * If you can remove a LIKELY/UNLIKELY annotation without speed changes in gcc
+ * and clang, please do.
+ */
+#if defined(__GNUC__)
+#define LIKELY(x) (__builtin_expect((x), 1))
+#define UNLIKELY(x) (__builtin_expect((x), 0))
+#else
+#define LIKELY(x) (x)
+#define UNLIKELY(x) (x)
+#endif
+
+#if __has_builtin(__builtin_unreachable) || (defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 5)))
+# define ZSTD_UNREACHABLE do { assert(0), __builtin_unreachable(); } while (0)
+#else
+# define ZSTD_UNREACHABLE do { assert(0); } while (0)
+#endif
+
+/* disable warnings */
+#ifdef _MSC_VER /* Visual Studio */
+# include <intrin.h> /* For Visual 2005 */
+# pragma warning(disable : 4100) /* disable: C4100: unreferenced formal parameter */
+# pragma warning(disable : 4127) /* disable: C4127: conditional expression is constant */
+# pragma warning(disable : 4204) /* disable: C4204: non-constant aggregate initializer */
+# pragma warning(disable : 4214) /* disable: C4214: non-int bitfields */
+# pragma warning(disable : 4324) /* disable: C4324: padded structure */
+#endif
+
+/*Like DYNAMIC_BMI2 but for compile time determination of BMI2 support*/
+#ifndef STATIC_BMI2
+# if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_I86))
+# ifdef __AVX2__ //MSVC does not have a BMI2 specific flag, but every CPU that supports AVX2 also supports BMI2
+# define STATIC_BMI2 1
+# endif
+# elif defined(__BMI2__) && defined(__x86_64__) && defined(__GNUC__)
+# define STATIC_BMI2 1
+# endif
+#endif
+
+#ifndef STATIC_BMI2
+ #define STATIC_BMI2 0
+#endif
+
+/* compile time determination of SIMD support */
+#if !defined(ZSTD_NO_INTRINSICS)
+# if defined(__SSE2__) || defined(_M_AMD64) || (defined (_M_IX86) && defined(_M_IX86_FP) && (_M_IX86_FP >= 2))
+# define ZSTD_ARCH_X86_SSE2
+# endif
+# if defined(__ARM_NEON) || defined(_M_ARM64)
+# define ZSTD_ARCH_ARM_NEON
+# endif
+#
+# if defined(ZSTD_ARCH_X86_SSE2)
+# include <emmintrin.h>
+# elif defined(ZSTD_ARCH_ARM_NEON)
+# include <arm_neon.h>
+# endif
+#endif
+
+/* C-language Attributes are added in C23. */
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ > 201710L) && defined(__has_c_attribute)
+# define ZSTD_HAS_C_ATTRIBUTE(x) __has_c_attribute(x)
+#else
+# define ZSTD_HAS_C_ATTRIBUTE(x) 0
+#endif
+
+/* Only use C++ attributes in C++. Some compilers report support for C++
+ * attributes when compiling with C.
+ */
+#if defined(__cplusplus) && defined(__has_cpp_attribute)
+# define ZSTD_HAS_CPP_ATTRIBUTE(x) __has_cpp_attribute(x)
+#else
+# define ZSTD_HAS_CPP_ATTRIBUTE(x) 0
+#endif
+
+/* Define ZSTD_FALLTHROUGH macro for annotating switch case with the 'fallthrough' attribute.
+ * - C23: https://en.cppreference.com/w/c/language/attributes/fallthrough
+ * - CPP17: https://en.cppreference.com/w/cpp/language/attributes/fallthrough
+ * - Else: __attribute__((__fallthrough__))
+ */
+#ifndef ZSTD_FALLTHROUGH
+# if ZSTD_HAS_C_ATTRIBUTE(fallthrough)
+# define ZSTD_FALLTHROUGH [[fallthrough]]
+# elif ZSTD_HAS_CPP_ATTRIBUTE(fallthrough)
+# define ZSTD_FALLTHROUGH [[fallthrough]]
+# elif __has_attribute(__fallthrough__)
+/* Leading semicolon is to satisfy gcc-11 with -pedantic. Without the semicolon
+ * gcc complains about: a label can only be part of a statement and a declaration is not a statement.
+ */
+# define ZSTD_FALLTHROUGH ; __attribute__((__fallthrough__))
+# else
+# define ZSTD_FALLTHROUGH
+# endif
+#endif
+
+/*-**************************************************************
+* Alignment check
+*****************************************************************/
+
+/* this test was initially positioned in mem.h,
+ * but this file is removed (or replaced) for linux kernel
+ * so it's now hosted in compiler.h,
+ * which remains valid for both user & kernel spaces.
+ */
+
+#ifndef ZSTD_ALIGNOF
+# if defined(__GNUC__) || defined(_MSC_VER)
+/* covers gcc, clang & MSVC */
+/* note : this section must come first, before C11,
+ * due to a limitation in the kernel source generator */
+# define ZSTD_ALIGNOF(T) __alignof(T)
+
+# elif defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)
+/* C11 support */
+# include <stdalign.h>
+# define ZSTD_ALIGNOF(T) alignof(T)
+
+# else
+/* No known support for alignof() - imperfect backup */
+# define ZSTD_ALIGNOF(T) (sizeof(void*) < sizeof(T) ? sizeof(void*) : sizeof(T))
+
+# endif
+#endif /* ZSTD_ALIGNOF */
+
+/*-**************************************************************
+* Sanitizer
+*****************************************************************/
+
+/**
+ * Zstd relies on pointer overflow in its decompressor.
+ * We add this attribute to functions that rely on pointer overflow.
+ */
+#ifndef ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+# if __has_attribute(no_sanitize)
+# if !defined(__clang__) && defined(__GNUC__) && __GNUC__ < 8
+ /* gcc < 8 only has signed-integer-overlow which triggers on pointer overflow */
+# define ZSTD_ALLOW_POINTER_OVERFLOW_ATTR __attribute__((no_sanitize("signed-integer-overflow")))
+# else
+ /* older versions of clang [3.7, 5.0) will warn that pointer-overflow is ignored. */
+# define ZSTD_ALLOW_POINTER_OVERFLOW_ATTR __attribute__((no_sanitize("pointer-overflow")))
+# endif
+# else
+# define ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+# endif
+#endif
+
+/**
+ * Helper function to perform a wrapped pointer difference without trigging
+ * UBSAN.
+ *
+ * @returns lhs - rhs with wrapping
+ */
+MEM_STATIC
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+ptrdiff_t ZSTD_wrappedPtrDiff(unsigned char const* lhs, unsigned char const* rhs)
+{
+ return lhs - rhs;
+}
+
+/**
+ * Helper function to perform a wrapped pointer add without triggering UBSAN.
+ *
+ * @return ptr + add with wrapping
+ */
+MEM_STATIC
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+unsigned char const* ZSTD_wrappedPtrAdd(unsigned char const* ptr, ptrdiff_t add)
+{
+ return ptr + add;
+}
+
+/**
+ * Helper function to perform a wrapped pointer subtraction without triggering
+ * UBSAN.
+ *
+ * @return ptr - sub with wrapping
+ */
+MEM_STATIC
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+unsigned char const* ZSTD_wrappedPtrSub(unsigned char const* ptr, ptrdiff_t sub)
+{
+ return ptr - sub;
+}
+
+/**
+ * Helper function to add to a pointer that works around C's undefined behavior
+ * of adding 0 to NULL.
+ *
+ * @returns `ptr + add` except it defines `NULL + 0 == NULL`.
+ */
+MEM_STATIC
+unsigned char* ZSTD_maybeNullPtrAdd(unsigned char* ptr, ptrdiff_t add)
+{
+ return add > 0 ? ptr + add : ptr;
+}
+
+/* Issue #3240 reports an ASAN failure on an llvm-mingw build. Out of an
+ * abundance of caution, disable our custom poisoning on mingw. */
+#ifdef __MINGW32__
+#ifndef ZSTD_ASAN_DONT_POISON_WORKSPACE
+#define ZSTD_ASAN_DONT_POISON_WORKSPACE 1
+#endif
+#ifndef ZSTD_MSAN_DONT_POISON_WORKSPACE
+#define ZSTD_MSAN_DONT_POISON_WORKSPACE 1
+#endif
+#endif
+
+#if ZSTD_MEMORY_SANITIZER && !defined(ZSTD_MSAN_DONT_POISON_WORKSPACE)
+/* Not all platforms that support msan provide sanitizers/msan_interface.h.
+ * We therefore declare the functions we need ourselves, rather than trying to
+ * include the header file... */
+#include <stddef.h> /* size_t */
+#define ZSTD_DEPS_NEED_STDINT
+#include "zstd_deps.h" /* intptr_t */
+
+/* Make memory region fully initialized (without changing its contents). */
+void __msan_unpoison(const volatile void *a, size_t size);
+
+/* Make memory region fully uninitialized (without changing its contents).
+ This is a legacy interface that does not update origin information. Use
+ __msan_allocated_memory() instead. */
+void __msan_poison(const volatile void *a, size_t size);
+
+/* Returns the offset of the first (at least partially) poisoned byte in the
+ memory range, or -1 if the whole range is good. */
+intptr_t __msan_test_shadow(const volatile void *x, size_t size);
+
+/* Print shadow and origin for the memory range to stderr in a human-readable
+ format. */
+void __msan_print_shadow(const volatile void *x, size_t size);
+#endif
+
+#if ZSTD_ADDRESS_SANITIZER && !defined(ZSTD_ASAN_DONT_POISON_WORKSPACE)
+/* Not all platforms that support asan provide sanitizers/asan_interface.h.
+ * We therefore declare the functions we need ourselves, rather than trying to
+ * include the header file... */
+#include <stddef.h> /* size_t */
+
+/**
+ * Marks a memory region (<c>[addr, addr+size)</c>) as unaddressable.
+ *
+ * This memory must be previously allocated by your program. Instrumented
+ * code is forbidden from accessing addresses in this region until it is
+ * unpoisoned. This function is not guaranteed to poison the entire region -
+ * it could poison only a subregion of <c>[addr, addr+size)</c> due to ASan
+ * alignment restrictions.
+ *
+ * \note This function is not thread-safe because no two threads can poison or
+ * unpoison memory in the same memory region simultaneously.
+ *
+ * \param addr Start of memory region.
+ * \param size Size of memory region. */
+void __asan_poison_memory_region(void const volatile *addr, size_t size);
+
+/**
+ * Marks a memory region (<c>[addr, addr+size)</c>) as addressable.
+ *
+ * This memory must be previously allocated by your program. Accessing
+ * addresses in this region is allowed until this region is poisoned again.
+ * This function could unpoison a super-region of <c>[addr, addr+size)</c> due
+ * to ASan alignment restrictions.
+ *
+ * \note This function is not thread-safe because no two threads can
+ * poison or unpoison memory in the same memory region simultaneously.
+ *
+ * \param addr Start of memory region.
+ * \param size Size of memory region. */
+void __asan_unpoison_memory_region(void const volatile *addr, size_t size);
+#endif
+
+#endif /* ZSTD_COMPILER_H */
diff --git a/third_party/zstd/lib/common/cpu.h b/third_party/zstd/lib/common/cpu.h
new file mode 100644
index 0000000000..d9cdf8febe
--- /dev/null
+++ b/third_party/zstd/lib/common/cpu.h
@@ -0,0 +1,249 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_COMMON_CPU_H
+#define ZSTD_COMMON_CPU_H
+
+/**
+ * Implementation taken from folly/CpuId.h
+ * https://github.com/facebook/folly/blob/master/folly/CpuId.h
+ */
+
+#include "mem.h"
+
+#ifdef _MSC_VER
+#include <intrin.h>
+#endif
+
+typedef struct {
+ U32 f1c;
+ U32 f1d;
+ U32 f7b;
+ U32 f7c;
+} ZSTD_cpuid_t;
+
+MEM_STATIC ZSTD_cpuid_t ZSTD_cpuid(void) {
+ U32 f1c = 0;
+ U32 f1d = 0;
+ U32 f7b = 0;
+ U32 f7c = 0;
+#if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86))
+#if !defined(__clang__) || __clang_major__ >= 16
+ int reg[4];
+ __cpuid((int*)reg, 0);
+ {
+ int const n = reg[0];
+ if (n >= 1) {
+ __cpuid((int*)reg, 1);
+ f1c = (U32)reg[2];
+ f1d = (U32)reg[3];
+ }
+ if (n >= 7) {
+ __cpuidex((int*)reg, 7, 0);
+ f7b = (U32)reg[1];
+ f7c = (U32)reg[2];
+ }
+ }
+#else
+ /* Clang compiler has a bug (fixed in https://reviews.llvm.org/D101338) in
+ * which the `__cpuid` intrinsic does not save and restore `rbx` as it needs
+ * to due to being a reserved register. So in that case, do the `cpuid`
+ * ourselves. Clang supports inline assembly anyway.
+ */
+ U32 n;
+ __asm__(
+ "pushq %%rbx\n\t"
+ "cpuid\n\t"
+ "popq %%rbx\n\t"
+ : "=a"(n)
+ : "a"(0)
+ : "rcx", "rdx");
+ if (n >= 1) {
+ U32 f1a;
+ __asm__(
+ "pushq %%rbx\n\t"
+ "cpuid\n\t"
+ "popq %%rbx\n\t"
+ : "=a"(f1a), "=c"(f1c), "=d"(f1d)
+ : "a"(1)
+ :);
+ }
+ if (n >= 7) {
+ __asm__(
+ "pushq %%rbx\n\t"
+ "cpuid\n\t"
+ "movq %%rbx, %%rax\n\t"
+ "popq %%rbx"
+ : "=a"(f7b), "=c"(f7c)
+ : "a"(7), "c"(0)
+ : "rdx");
+ }
+#endif
+#elif defined(__i386__) && defined(__PIC__) && !defined(__clang__) && defined(__GNUC__)
+ /* The following block like the normal cpuid branch below, but gcc
+ * reserves ebx for use of its pic register so we must specially
+ * handle the save and restore to avoid clobbering the register
+ */
+ U32 n;
+ __asm__(
+ "pushl %%ebx\n\t"
+ "cpuid\n\t"
+ "popl %%ebx\n\t"
+ : "=a"(n)
+ : "a"(0)
+ : "ecx", "edx");
+ if (n >= 1) {
+ U32 f1a;
+ __asm__(
+ "pushl %%ebx\n\t"
+ "cpuid\n\t"
+ "popl %%ebx\n\t"
+ : "=a"(f1a), "=c"(f1c), "=d"(f1d)
+ : "a"(1));
+ }
+ if (n >= 7) {
+ __asm__(
+ "pushl %%ebx\n\t"
+ "cpuid\n\t"
+ "movl %%ebx, %%eax\n\t"
+ "popl %%ebx"
+ : "=a"(f7b), "=c"(f7c)
+ : "a"(7), "c"(0)
+ : "edx");
+ }
+#elif defined(__x86_64__) || defined(_M_X64) || defined(__i386__)
+ U32 n;
+ __asm__("cpuid" : "=a"(n) : "a"(0) : "ebx", "ecx", "edx");
+ if (n >= 1) {
+ U32 f1a;
+ __asm__("cpuid" : "=a"(f1a), "=c"(f1c), "=d"(f1d) : "a"(1) : "ebx");
+ }
+ if (n >= 7) {
+ U32 f7a;
+ __asm__("cpuid"
+ : "=a"(f7a), "=b"(f7b), "=c"(f7c)
+ : "a"(7), "c"(0)
+ : "edx");
+ }
+#endif
+ {
+ ZSTD_cpuid_t cpuid;
+ cpuid.f1c = f1c;
+ cpuid.f1d = f1d;
+ cpuid.f7b = f7b;
+ cpuid.f7c = f7c;
+ return cpuid;
+ }
+}
+
+#define X(name, r, bit) \
+ MEM_STATIC int ZSTD_cpuid_##name(ZSTD_cpuid_t const cpuid) { \
+ return ((cpuid.r) & (1U << bit)) != 0; \
+ }
+
+/* cpuid(1): Processor Info and Feature Bits. */
+#define C(name, bit) X(name, f1c, bit)
+ C(sse3, 0)
+ C(pclmuldq, 1)
+ C(dtes64, 2)
+ C(monitor, 3)
+ C(dscpl, 4)
+ C(vmx, 5)
+ C(smx, 6)
+ C(eist, 7)
+ C(tm2, 8)
+ C(ssse3, 9)
+ C(cnxtid, 10)
+ C(fma, 12)
+ C(cx16, 13)
+ C(xtpr, 14)
+ C(pdcm, 15)
+ C(pcid, 17)
+ C(dca, 18)
+ C(sse41, 19)
+ C(sse42, 20)
+ C(x2apic, 21)
+ C(movbe, 22)
+ C(popcnt, 23)
+ C(tscdeadline, 24)
+ C(aes, 25)
+ C(xsave, 26)
+ C(osxsave, 27)
+ C(avx, 28)
+ C(f16c, 29)
+ C(rdrand, 30)
+#undef C
+#define D(name, bit) X(name, f1d, bit)
+ D(fpu, 0)
+ D(vme, 1)
+ D(de, 2)
+ D(pse, 3)
+ D(tsc, 4)
+ D(msr, 5)
+ D(pae, 6)
+ D(mce, 7)
+ D(cx8, 8)
+ D(apic, 9)
+ D(sep, 11)
+ D(mtrr, 12)
+ D(pge, 13)
+ D(mca, 14)
+ D(cmov, 15)
+ D(pat, 16)
+ D(pse36, 17)
+ D(psn, 18)
+ D(clfsh, 19)
+ D(ds, 21)
+ D(acpi, 22)
+ D(mmx, 23)
+ D(fxsr, 24)
+ D(sse, 25)
+ D(sse2, 26)
+ D(ss, 27)
+ D(htt, 28)
+ D(tm, 29)
+ D(pbe, 31)
+#undef D
+
+/* cpuid(7): Extended Features. */
+#define B(name, bit) X(name, f7b, bit)
+ B(bmi1, 3)
+ B(hle, 4)
+ B(avx2, 5)
+ B(smep, 7)
+ B(bmi2, 8)
+ B(erms, 9)
+ B(invpcid, 10)
+ B(rtm, 11)
+ B(mpx, 14)
+ B(avx512f, 16)
+ B(avx512dq, 17)
+ B(rdseed, 18)
+ B(adx, 19)
+ B(smap, 20)
+ B(avx512ifma, 21)
+ B(pcommit, 22)
+ B(clflushopt, 23)
+ B(clwb, 24)
+ B(avx512pf, 26)
+ B(avx512er, 27)
+ B(avx512cd, 28)
+ B(sha, 29)
+ B(avx512bw, 30)
+ B(avx512vl, 31)
+#undef B
+#define C(name, bit) X(name, f7c, bit)
+ C(prefetchwt1, 0)
+ C(avx512vbmi, 1)
+#undef C
+
+#undef X
+
+#endif /* ZSTD_COMMON_CPU_H */
diff --git a/third_party/zstd/lib/common/debug.c b/third_party/zstd/lib/common/debug.c
new file mode 100644
index 0000000000..9d0b7d229c
--- /dev/null
+++ b/third_party/zstd/lib/common/debug.c
@@ -0,0 +1,30 @@
+/* ******************************************************************
+ * debug
+ * Part of FSE library
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ *
+ * You can contact the author at :
+ * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+****************************************************************** */
+
+
+/*
+ * This module only hosts one global variable
+ * which can be used to dynamically influence the verbosity of traces,
+ * such as DEBUGLOG and RAWLOG
+ */
+
+#include "debug.h"
+
+#if !defined(ZSTD_LINUX_KERNEL) || (DEBUGLEVEL>=2)
+/* We only use this when DEBUGLEVEL>=2, but we get -Werror=pedantic errors if a
+ * translation unit is empty. So remove this from Linux kernel builds, but
+ * otherwise just leave it in.
+ */
+int g_debuglevel = DEBUGLEVEL;
+#endif
diff --git a/third_party/zstd/lib/common/debug.h b/third_party/zstd/lib/common/debug.h
new file mode 100644
index 0000000000..a16b69e574
--- /dev/null
+++ b/third_party/zstd/lib/common/debug.h
@@ -0,0 +1,116 @@
+/* ******************************************************************
+ * debug
+ * Part of FSE library
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ *
+ * You can contact the author at :
+ * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+****************************************************************** */
+
+
+/*
+ * The purpose of this header is to enable debug functions.
+ * They regroup assert(), DEBUGLOG() and RAWLOG() for run-time,
+ * and DEBUG_STATIC_ASSERT() for compile-time.
+ *
+ * By default, DEBUGLEVEL==0, which means run-time debug is disabled.
+ *
+ * Level 1 enables assert() only.
+ * Starting level 2, traces can be generated and pushed to stderr.
+ * The higher the level, the more verbose the traces.
+ *
+ * It's possible to dynamically adjust level using variable g_debug_level,
+ * which is only declared if DEBUGLEVEL>=2,
+ * and is a global variable, not multi-thread protected (use with care)
+ */
+
+#ifndef DEBUG_H_12987983217
+#define DEBUG_H_12987983217
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/* static assert is triggered at compile time, leaving no runtime artefact.
+ * static assert only works with compile-time constants.
+ * Also, this variant can only be used inside a function. */
+#define DEBUG_STATIC_ASSERT(c) (void)sizeof(char[(c) ? 1 : -1])
+
+
+/* DEBUGLEVEL is expected to be defined externally,
+ * typically through compiler command line.
+ * Value must be a number. */
+#ifndef DEBUGLEVEL
+# define DEBUGLEVEL 0
+#endif
+
+
+/* recommended values for DEBUGLEVEL :
+ * 0 : release mode, no debug, all run-time checks disabled
+ * 1 : enables assert() only, no display
+ * 2 : reserved, for currently active debug path
+ * 3 : events once per object lifetime (CCtx, CDict, etc.)
+ * 4 : events once per frame
+ * 5 : events once per block
+ * 6 : events once per sequence (verbose)
+ * 7+: events at every position (*very* verbose)
+ *
+ * It's generally inconvenient to output traces > 5.
+ * In which case, it's possible to selectively trigger high verbosity levels
+ * by modifying g_debug_level.
+ */
+
+#if (DEBUGLEVEL>=1)
+# define ZSTD_DEPS_NEED_ASSERT
+# include "zstd_deps.h"
+#else
+# ifndef assert /* assert may be already defined, due to prior #include <assert.h> */
+# define assert(condition) ((void)0) /* disable assert (default) */
+# endif
+#endif
+
+#if (DEBUGLEVEL>=2)
+# define ZSTD_DEPS_NEED_IO
+# include "zstd_deps.h"
+extern int g_debuglevel; /* the variable is only declared,
+ it actually lives in debug.c,
+ and is shared by the whole process.
+ It's not thread-safe.
+ It's useful when enabling very verbose levels
+ on selective conditions (such as position in src) */
+
+# define RAWLOG(l, ...) \
+ do { \
+ if (l<=g_debuglevel) { \
+ ZSTD_DEBUG_PRINT(__VA_ARGS__); \
+ } \
+ } while (0)
+
+#define STRINGIFY(x) #x
+#define TOSTRING(x) STRINGIFY(x)
+#define LINE_AS_STRING TOSTRING(__LINE__)
+
+# define DEBUGLOG(l, ...) \
+ do { \
+ if (l<=g_debuglevel) { \
+ ZSTD_DEBUG_PRINT(__FILE__ ":" LINE_AS_STRING ": " __VA_ARGS__); \
+ ZSTD_DEBUG_PRINT(" \n"); \
+ } \
+ } while (0)
+#else
+# define RAWLOG(l, ...) do { } while (0) /* disabled */
+# define DEBUGLOG(l, ...) do { } while (0) /* disabled */
+#endif
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* DEBUG_H_12987983217 */
diff --git a/third_party/zstd/lib/common/entropy_common.c b/third_party/zstd/lib/common/entropy_common.c
new file mode 100644
index 0000000000..e2173afb0a
--- /dev/null
+++ b/third_party/zstd/lib/common/entropy_common.c
@@ -0,0 +1,340 @@
+/* ******************************************************************
+ * Common functions of New Generation Entropy library
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ *
+ * You can contact the author at :
+ * - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
+ * - Public forum : https://groups.google.com/forum/#!forum/lz4c
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+****************************************************************** */
+
+/* *************************************
+* Dependencies
+***************************************/
+#include "mem.h"
+#include "error_private.h" /* ERR_*, ERROR */
+#define FSE_STATIC_LINKING_ONLY /* FSE_MIN_TABLELOG */
+#include "fse.h"
+#include "huf.h"
+#include "bits.h" /* ZSDT_highbit32, ZSTD_countTrailingZeros32 */
+
+
+/*=== Version ===*/
+unsigned FSE_versionNumber(void) { return FSE_VERSION_NUMBER; }
+
+
+/*=== Error Management ===*/
+unsigned FSE_isError(size_t code) { return ERR_isError(code); }
+const char* FSE_getErrorName(size_t code) { return ERR_getErrorName(code); }
+
+unsigned HUF_isError(size_t code) { return ERR_isError(code); }
+const char* HUF_getErrorName(size_t code) { return ERR_getErrorName(code); }
+
+
+/*-**************************************************************
+* FSE NCount encoding-decoding
+****************************************************************/
+FORCE_INLINE_TEMPLATE
+size_t FSE_readNCount_body(short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+ const void* headerBuffer, size_t hbSize)
+{
+ const BYTE* const istart = (const BYTE*) headerBuffer;
+ const BYTE* const iend = istart + hbSize;
+ const BYTE* ip = istart;
+ int nbBits;
+ int remaining;
+ int threshold;
+ U32 bitStream;
+ int bitCount;
+ unsigned charnum = 0;
+ unsigned const maxSV1 = *maxSVPtr + 1;
+ int previous0 = 0;
+
+ if (hbSize < 8) {
+ /* This function only works when hbSize >= 8 */
+ char buffer[8] = {0};
+ ZSTD_memcpy(buffer, headerBuffer, hbSize);
+ { size_t const countSize = FSE_readNCount(normalizedCounter, maxSVPtr, tableLogPtr,
+ buffer, sizeof(buffer));
+ if (FSE_isError(countSize)) return countSize;
+ if (countSize > hbSize) return ERROR(corruption_detected);
+ return countSize;
+ } }
+ assert(hbSize >= 8);
+
+ /* init */
+ ZSTD_memset(normalizedCounter, 0, (*maxSVPtr+1) * sizeof(normalizedCounter[0])); /* all symbols not present in NCount have a frequency of 0 */
+ bitStream = MEM_readLE32(ip);
+ nbBits = (bitStream & 0xF) + FSE_MIN_TABLELOG; /* extract tableLog */
+ if (nbBits > FSE_TABLELOG_ABSOLUTE_MAX) return ERROR(tableLog_tooLarge);
+ bitStream >>= 4;
+ bitCount = 4;
+ *tableLogPtr = nbBits;
+ remaining = (1<<nbBits)+1;
+ threshold = 1<<nbBits;
+ nbBits++;
+
+ for (;;) {
+ if (previous0) {
+ /* Count the number of repeats. Each time the
+ * 2-bit repeat code is 0b11 there is another
+ * repeat.
+ * Avoid UB by setting the high bit to 1.
+ */
+ int repeats = ZSTD_countTrailingZeros32(~bitStream | 0x80000000) >> 1;
+ while (repeats >= 12) {
+ charnum += 3 * 12;
+ if (LIKELY(ip <= iend-7)) {
+ ip += 3;
+ } else {
+ bitCount -= (int)(8 * (iend - 7 - ip));
+ bitCount &= 31;
+ ip = iend - 4;
+ }
+ bitStream = MEM_readLE32(ip) >> bitCount;
+ repeats = ZSTD_countTrailingZeros32(~bitStream | 0x80000000) >> 1;
+ }
+ charnum += 3 * repeats;
+ bitStream >>= 2 * repeats;
+ bitCount += 2 * repeats;
+
+ /* Add the final repeat which isn't 0b11. */
+ assert((bitStream & 3) < 3);
+ charnum += bitStream & 3;
+ bitCount += 2;
+
+ /* This is an error, but break and return an error
+ * at the end, because returning out of a loop makes
+ * it harder for the compiler to optimize.
+ */
+ if (charnum >= maxSV1) break;
+
+ /* We don't need to set the normalized count to 0
+ * because we already memset the whole buffer to 0.
+ */
+
+ if (LIKELY(ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
+ assert((bitCount >> 3) <= 3); /* For first condition to work */
+ ip += bitCount>>3;
+ bitCount &= 7;
+ } else {
+ bitCount -= (int)(8 * (iend - 4 - ip));
+ bitCount &= 31;
+ ip = iend - 4;
+ }
+ bitStream = MEM_readLE32(ip) >> bitCount;
+ }
+ {
+ int const max = (2*threshold-1) - remaining;
+ int count;
+
+ if ((bitStream & (threshold-1)) < (U32)max) {
+ count = bitStream & (threshold-1);
+ bitCount += nbBits-1;
+ } else {
+ count = bitStream & (2*threshold-1);
+ if (count >= threshold) count -= max;
+ bitCount += nbBits;
+ }
+
+ count--; /* extra accuracy */
+ /* When it matters (small blocks), this is a
+ * predictable branch, because we don't use -1.
+ */
+ if (count >= 0) {
+ remaining -= count;
+ } else {
+ assert(count == -1);
+ remaining += count;
+ }
+ normalizedCounter[charnum++] = (short)count;
+ previous0 = !count;
+
+ assert(threshold > 1);
+ if (remaining < threshold) {
+ /* This branch can be folded into the
+ * threshold update condition because we
+ * know that threshold > 1.
+ */
+ if (remaining <= 1) break;
+ nbBits = ZSTD_highbit32(remaining) + 1;
+ threshold = 1 << (nbBits - 1);
+ }
+ if (charnum >= maxSV1) break;
+
+ if (LIKELY(ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
+ ip += bitCount>>3;
+ bitCount &= 7;
+ } else {
+ bitCount -= (int)(8 * (iend - 4 - ip));
+ bitCount &= 31;
+ ip = iend - 4;
+ }
+ bitStream = MEM_readLE32(ip) >> bitCount;
+ } }
+ if (remaining != 1) return ERROR(corruption_detected);
+ /* Only possible when there are too many zeros. */
+ if (charnum > maxSV1) return ERROR(maxSymbolValue_tooSmall);
+ if (bitCount > 32) return ERROR(corruption_detected);
+ *maxSVPtr = charnum-1;
+
+ ip += (bitCount+7)>>3;
+ return ip-istart;
+}
+
+/* Avoids the FORCE_INLINE of the _body() function. */
+static size_t FSE_readNCount_body_default(
+ short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+ const void* headerBuffer, size_t hbSize)
+{
+ return FSE_readNCount_body(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
+}
+
+#if DYNAMIC_BMI2
+BMI2_TARGET_ATTRIBUTE static size_t FSE_readNCount_body_bmi2(
+ short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+ const void* headerBuffer, size_t hbSize)
+{
+ return FSE_readNCount_body(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
+}
+#endif
+
+size_t FSE_readNCount_bmi2(
+ short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+ const void* headerBuffer, size_t hbSize, int bmi2)
+{
+#if DYNAMIC_BMI2
+ if (bmi2) {
+ return FSE_readNCount_body_bmi2(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
+ }
+#endif
+ (void)bmi2;
+ return FSE_readNCount_body_default(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
+}
+
+size_t FSE_readNCount(
+ short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+ const void* headerBuffer, size_t hbSize)
+{
+ return FSE_readNCount_bmi2(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize, /* bmi2 */ 0);
+}
+
+
+/*! HUF_readStats() :
+ Read compact Huffman tree, saved by HUF_writeCTable().
+ `huffWeight` is destination buffer.
+ `rankStats` is assumed to be a table of at least HUF_TABLELOG_MAX U32.
+ @return : size read from `src` , or an error Code .
+ Note : Needed by HUF_readCTable() and HUF_readDTableX?() .
+*/
+size_t HUF_readStats(BYTE* huffWeight, size_t hwSize, U32* rankStats,
+ U32* nbSymbolsPtr, U32* tableLogPtr,
+ const void* src, size_t srcSize)
+{
+ U32 wksp[HUF_READ_STATS_WORKSPACE_SIZE_U32];
+ return HUF_readStats_wksp(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, wksp, sizeof(wksp), /* flags */ 0);
+}
+
+FORCE_INLINE_TEMPLATE size_t
+HUF_readStats_body(BYTE* huffWeight, size_t hwSize, U32* rankStats,
+ U32* nbSymbolsPtr, U32* tableLogPtr,
+ const void* src, size_t srcSize,
+ void* workSpace, size_t wkspSize,
+ int bmi2)
+{
+ U32 weightTotal;
+ const BYTE* ip = (const BYTE*) src;
+ size_t iSize;
+ size_t oSize;
+
+ if (!srcSize) return ERROR(srcSize_wrong);
+ iSize = ip[0];
+ /* ZSTD_memset(huffWeight, 0, hwSize); *//* is not necessary, even though some analyzer complain ... */
+
+ if (iSize >= 128) { /* special header */
+ oSize = iSize - 127;
+ iSize = ((oSize+1)/2);
+ if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
+ if (oSize >= hwSize) return ERROR(corruption_detected);
+ ip += 1;
+ { U32 n;
+ for (n=0; n<oSize; n+=2) {
+ huffWeight[n] = ip[n/2] >> 4;
+ huffWeight[n+1] = ip[n/2] & 15;
+ } } }
+ else { /* header compressed with FSE (normal case) */
+ if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
+ /* max (hwSize-1) values decoded, as last one is implied */
+ oSize = FSE_decompress_wksp_bmi2(huffWeight, hwSize-1, ip+1, iSize, 6, workSpace, wkspSize, bmi2);
+ if (FSE_isError(oSize)) return oSize;
+ }
+
+ /* collect weight stats */
+ ZSTD_memset(rankStats, 0, (HUF_TABLELOG_MAX + 1) * sizeof(U32));
+ weightTotal = 0;
+ { U32 n; for (n=0; n<oSize; n++) {
+ if (huffWeight[n] > HUF_TABLELOG_MAX) return ERROR(corruption_detected);
+ rankStats[huffWeight[n]]++;
+ weightTotal += (1 << huffWeight[n]) >> 1;
+ } }
+ if (weightTotal == 0) return ERROR(corruption_detected);
+
+ /* get last non-null symbol weight (implied, total must be 2^n) */
+ { U32 const tableLog = ZSTD_highbit32(weightTotal) + 1;
+ if (tableLog > HUF_TABLELOG_MAX) return ERROR(corruption_detected);
+ *tableLogPtr = tableLog;
+ /* determine last weight */
+ { U32 const total = 1 << tableLog;
+ U32 const rest = total - weightTotal;
+ U32 const verif = 1 << ZSTD_highbit32(rest);
+ U32 const lastWeight = ZSTD_highbit32(rest) + 1;
+ if (verif != rest) return ERROR(corruption_detected); /* last value must be a clean power of 2 */
+ huffWeight[oSize] = (BYTE)lastWeight;
+ rankStats[lastWeight]++;
+ } }
+
+ /* check tree construction validity */
+ if ((rankStats[1] < 2) || (rankStats[1] & 1)) return ERROR(corruption_detected); /* by construction : at least 2 elts of rank 1, must be even */
+
+ /* results */
+ *nbSymbolsPtr = (U32)(oSize+1);
+ return iSize+1;
+}
+
+/* Avoids the FORCE_INLINE of the _body() function. */
+static size_t HUF_readStats_body_default(BYTE* huffWeight, size_t hwSize, U32* rankStats,
+ U32* nbSymbolsPtr, U32* tableLogPtr,
+ const void* src, size_t srcSize,
+ void* workSpace, size_t wkspSize)
+{
+ return HUF_readStats_body(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize, 0);
+}
+
+#if DYNAMIC_BMI2
+static BMI2_TARGET_ATTRIBUTE size_t HUF_readStats_body_bmi2(BYTE* huffWeight, size_t hwSize, U32* rankStats,
+ U32* nbSymbolsPtr, U32* tableLogPtr,
+ const void* src, size_t srcSize,
+ void* workSpace, size_t wkspSize)
+{
+ return HUF_readStats_body(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize, 1);
+}
+#endif
+
+size_t HUF_readStats_wksp(BYTE* huffWeight, size_t hwSize, U32* rankStats,
+ U32* nbSymbolsPtr, U32* tableLogPtr,
+ const void* src, size_t srcSize,
+ void* workSpace, size_t wkspSize,
+ int flags)
+{
+#if DYNAMIC_BMI2
+ if (flags & HUF_flags_bmi2) {
+ return HUF_readStats_body_bmi2(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize);
+ }
+#endif
+ (void)flags;
+ return HUF_readStats_body_default(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize);
+}
diff --git a/third_party/zstd/lib/common/error_private.c b/third_party/zstd/lib/common/error_private.c
new file mode 100644
index 0000000000..075fc5ef42
--- /dev/null
+++ b/third_party/zstd/lib/common/error_private.c
@@ -0,0 +1,63 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+/* The purpose of this file is to have a single list of error strings embedded in binary */
+
+#include "error_private.h"
+
+const char* ERR_getErrorString(ERR_enum code)
+{
+#ifdef ZSTD_STRIP_ERROR_STRINGS
+ (void)code;
+ return "Error strings stripped";
+#else
+ static const char* const notErrorCode = "Unspecified error code";
+ switch( code )
+ {
+ case PREFIX(no_error): return "No error detected";
+ case PREFIX(GENERIC): return "Error (generic)";
+ case PREFIX(prefix_unknown): return "Unknown frame descriptor";
+ case PREFIX(version_unsupported): return "Version not supported";
+ case PREFIX(frameParameter_unsupported): return "Unsupported frame parameter";
+ case PREFIX(frameParameter_windowTooLarge): return "Frame requires too much memory for decoding";
+ case PREFIX(corruption_detected): return "Data corruption detected";
+ case PREFIX(checksum_wrong): return "Restored data doesn't match checksum";
+ case PREFIX(literals_headerWrong): return "Header of Literals' block doesn't respect format specification";
+ case PREFIX(parameter_unsupported): return "Unsupported parameter";
+ case PREFIX(parameter_combination_unsupported): return "Unsupported combination of parameters";
+ case PREFIX(parameter_outOfBound): return "Parameter is out of bound";
+ case PREFIX(init_missing): return "Context should be init first";
+ case PREFIX(memory_allocation): return "Allocation error : not enough memory";
+ case PREFIX(workSpace_tooSmall): return "workSpace buffer is not large enough";
+ case PREFIX(stage_wrong): return "Operation not authorized at current processing stage";
+ case PREFIX(tableLog_tooLarge): return "tableLog requires too much memory : unsupported";
+ case PREFIX(maxSymbolValue_tooLarge): return "Unsupported max Symbol Value : too large";
+ case PREFIX(maxSymbolValue_tooSmall): return "Specified maxSymbolValue is too small";
+ case PREFIX(stabilityCondition_notRespected): return "pledged buffer stability condition is not respected";
+ case PREFIX(dictionary_corrupted): return "Dictionary is corrupted";
+ case PREFIX(dictionary_wrong): return "Dictionary mismatch";
+ case PREFIX(dictionaryCreation_failed): return "Cannot create Dictionary from provided samples";
+ case PREFIX(dstSize_tooSmall): return "Destination buffer is too small";
+ case PREFIX(srcSize_wrong): return "Src size is incorrect";
+ case PREFIX(dstBuffer_null): return "Operation on NULL destination buffer";
+ case PREFIX(noForwardProgress_destFull): return "Operation made no progress over multiple calls, due to output buffer being full";
+ case PREFIX(noForwardProgress_inputEmpty): return "Operation made no progress over multiple calls, due to input being empty";
+ /* following error codes are not stable and may be removed or changed in a future version */
+ case PREFIX(frameIndex_tooLarge): return "Frame index is too large";
+ case PREFIX(seekableIO): return "An I/O error occurred when reading/seeking";
+ case PREFIX(dstBuffer_wrong): return "Destination buffer is wrong";
+ case PREFIX(srcBuffer_wrong): return "Source buffer is wrong";
+ case PREFIX(sequenceProducer_failed): return "Block-level external sequence producer returned an error code";
+ case PREFIX(externalSequences_invalid): return "External sequences are not valid";
+ case PREFIX(maxCode):
+ default: return notErrorCode;
+ }
+#endif
+}
diff --git a/third_party/zstd/lib/common/error_private.h b/third_party/zstd/lib/common/error_private.h
new file mode 100644
index 0000000000..0156010c74
--- /dev/null
+++ b/third_party/zstd/lib/common/error_private.h
@@ -0,0 +1,168 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+/* Note : this module is expected to remain private, do not expose it */
+
+#ifndef ERROR_H_MODULE
+#define ERROR_H_MODULE
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+/* ****************************************
+* Dependencies
+******************************************/
+#include "../zstd_errors.h" /* enum list */
+#include "compiler.h"
+#include "debug.h"
+#include "zstd_deps.h" /* size_t */
+
+
+/* ****************************************
+* Compiler-specific
+******************************************/
+#if defined(__GNUC__)
+# define ERR_STATIC static __attribute__((unused))
+#elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+# define ERR_STATIC static inline
+#elif defined(_MSC_VER)
+# define ERR_STATIC static __inline
+#else
+# define ERR_STATIC static /* this version may generate warnings for unused static functions; disable the relevant warning */
+#endif
+
+
+/*-****************************************
+* Customization (error_public.h)
+******************************************/
+typedef ZSTD_ErrorCode ERR_enum;
+#define PREFIX(name) ZSTD_error_##name
+
+
+/*-****************************************
+* Error codes handling
+******************************************/
+#undef ERROR /* already defined on Visual Studio */
+#define ERROR(name) ZSTD_ERROR(name)
+#define ZSTD_ERROR(name) ((size_t)-PREFIX(name))
+
+ERR_STATIC unsigned ERR_isError(size_t code) { return (code > ERROR(maxCode)); }
+
+ERR_STATIC ERR_enum ERR_getErrorCode(size_t code) { if (!ERR_isError(code)) return (ERR_enum)0; return (ERR_enum) (0-code); }
+
+/* check and forward error code */
+#define CHECK_V_F(e, f) \
+ size_t const e = f; \
+ do { \
+ if (ERR_isError(e)) \
+ return e; \
+ } while (0)
+#define CHECK_F(f) do { CHECK_V_F(_var_err__, f); } while (0)
+
+
+/*-****************************************
+* Error Strings
+******************************************/
+
+const char* ERR_getErrorString(ERR_enum code); /* error_private.c */
+
+ERR_STATIC const char* ERR_getErrorName(size_t code)
+{
+ return ERR_getErrorString(ERR_getErrorCode(code));
+}
+
+/**
+ * Ignore: this is an internal helper.
+ *
+ * This is a helper function to help force C99-correctness during compilation.
+ * Under strict compilation modes, variadic macro arguments can't be empty.
+ * However, variadic function arguments can be. Using a function therefore lets
+ * us statically check that at least one (string) argument was passed,
+ * independent of the compilation flags.
+ */
+static INLINE_KEYWORD UNUSED_ATTR
+void _force_has_format_string(const char *format, ...) {
+ (void)format;
+}
+
+/**
+ * Ignore: this is an internal helper.
+ *
+ * We want to force this function invocation to be syntactically correct, but
+ * we don't want to force runtime evaluation of its arguments.
+ */
+#define _FORCE_HAS_FORMAT_STRING(...) \
+ do { \
+ if (0) { \
+ _force_has_format_string(__VA_ARGS__); \
+ } \
+ } while (0)
+
+#define ERR_QUOTE(str) #str
+
+/**
+ * Return the specified error if the condition evaluates to true.
+ *
+ * In debug modes, prints additional information.
+ * In order to do that (particularly, printing the conditional that failed),
+ * this can't just wrap RETURN_ERROR().
+ */
+#define RETURN_ERROR_IF(cond, err, ...) \
+ do { \
+ if (cond) { \
+ RAWLOG(3, "%s:%d: ERROR!: check %s failed, returning %s", \
+ __FILE__, __LINE__, ERR_QUOTE(cond), ERR_QUOTE(ERROR(err))); \
+ _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
+ RAWLOG(3, ": " __VA_ARGS__); \
+ RAWLOG(3, "\n"); \
+ return ERROR(err); \
+ } \
+ } while (0)
+
+/**
+ * Unconditionally return the specified error.
+ *
+ * In debug modes, prints additional information.
+ */
+#define RETURN_ERROR(err, ...) \
+ do { \
+ RAWLOG(3, "%s:%d: ERROR!: unconditional check failed, returning %s", \
+ __FILE__, __LINE__, ERR_QUOTE(ERROR(err))); \
+ _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
+ RAWLOG(3, ": " __VA_ARGS__); \
+ RAWLOG(3, "\n"); \
+ return ERROR(err); \
+ } while(0)
+
+/**
+ * If the provided expression evaluates to an error code, returns that error code.
+ *
+ * In debug modes, prints additional information.
+ */
+#define FORWARD_IF_ERROR(err, ...) \
+ do { \
+ size_t const err_code = (err); \
+ if (ERR_isError(err_code)) { \
+ RAWLOG(3, "%s:%d: ERROR!: forwarding error in %s: %s", \
+ __FILE__, __LINE__, ERR_QUOTE(err), ERR_getErrorName(err_code)); \
+ _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
+ RAWLOG(3, ": " __VA_ARGS__); \
+ RAWLOG(3, "\n"); \
+ return err_code; \
+ } \
+ } while(0)
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* ERROR_H_MODULE */
diff --git a/third_party/zstd/lib/common/fse.h b/third_party/zstd/lib/common/fse.h
new file mode 100644
index 0000000000..2ae128e60d
--- /dev/null
+++ b/third_party/zstd/lib/common/fse.h
@@ -0,0 +1,640 @@
+/* ******************************************************************
+ * FSE : Finite State Entropy codec
+ * Public Prototypes declaration
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ *
+ * You can contact the author at :
+ * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+****************************************************************** */
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+#ifndef FSE_H
+#define FSE_H
+
+
+/*-*****************************************
+* Dependencies
+******************************************/
+#include "zstd_deps.h" /* size_t, ptrdiff_t */
+
+
+/*-*****************************************
+* FSE_PUBLIC_API : control library symbols visibility
+******************************************/
+#if defined(FSE_DLL_EXPORT) && (FSE_DLL_EXPORT==1) && defined(__GNUC__) && (__GNUC__ >= 4)
+# define FSE_PUBLIC_API __attribute__ ((visibility ("default")))
+#elif defined(FSE_DLL_EXPORT) && (FSE_DLL_EXPORT==1) /* Visual expected */
+# define FSE_PUBLIC_API __declspec(dllexport)
+#elif defined(FSE_DLL_IMPORT) && (FSE_DLL_IMPORT==1)
+# define FSE_PUBLIC_API __declspec(dllimport) /* It isn't required but allows to generate better code, saving a function pointer load from the IAT and an indirect jump.*/
+#else
+# define FSE_PUBLIC_API
+#endif
+
+/*------ Version ------*/
+#define FSE_VERSION_MAJOR 0
+#define FSE_VERSION_MINOR 9
+#define FSE_VERSION_RELEASE 0
+
+#define FSE_LIB_VERSION FSE_VERSION_MAJOR.FSE_VERSION_MINOR.FSE_VERSION_RELEASE
+#define FSE_QUOTE(str) #str
+#define FSE_EXPAND_AND_QUOTE(str) FSE_QUOTE(str)
+#define FSE_VERSION_STRING FSE_EXPAND_AND_QUOTE(FSE_LIB_VERSION)
+
+#define FSE_VERSION_NUMBER (FSE_VERSION_MAJOR *100*100 + FSE_VERSION_MINOR *100 + FSE_VERSION_RELEASE)
+FSE_PUBLIC_API unsigned FSE_versionNumber(void); /**< library version number; to be used when checking dll version */
+
+
+/*-*****************************************
+* Tool functions
+******************************************/
+FSE_PUBLIC_API size_t FSE_compressBound(size_t size); /* maximum compressed size */
+
+/* Error Management */
+FSE_PUBLIC_API unsigned FSE_isError(size_t code); /* tells if a return value is an error code */
+FSE_PUBLIC_API const char* FSE_getErrorName(size_t code); /* provides error code string (useful for debugging) */
+
+
+/*-*****************************************
+* FSE detailed API
+******************************************/
+/*!
+FSE_compress() does the following:
+1. count symbol occurrence from source[] into table count[] (see hist.h)
+2. normalize counters so that sum(count[]) == Power_of_2 (2^tableLog)
+3. save normalized counters to memory buffer using writeNCount()
+4. build encoding table 'CTable' from normalized counters
+5. encode the data stream using encoding table 'CTable'
+
+FSE_decompress() does the following:
+1. read normalized counters with readNCount()
+2. build decoding table 'DTable' from normalized counters
+3. decode the data stream using decoding table 'DTable'
+
+The following API allows targeting specific sub-functions for advanced tasks.
+For example, it's possible to compress several blocks using the same 'CTable',
+or to save and provide normalized distribution using external method.
+*/
+
+/* *** COMPRESSION *** */
+
+/*! FSE_optimalTableLog():
+ dynamically downsize 'tableLog' when conditions are met.
+ It saves CPU time, by using smaller tables, while preserving or even improving compression ratio.
+ @return : recommended tableLog (necessarily <= 'maxTableLog') */
+FSE_PUBLIC_API unsigned FSE_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue);
+
+/*! FSE_normalizeCount():
+ normalize counts so that sum(count[]) == Power_of_2 (2^tableLog)
+ 'normalizedCounter' is a table of short, of minimum size (maxSymbolValue+1).
+ useLowProbCount is a boolean parameter which trades off compressed size for
+ faster header decoding. When it is set to 1, the compressed data will be slightly
+ smaller. And when it is set to 0, FSE_readNCount() and FSE_buildDTable() will be
+ faster. If you are compressing a small amount of data (< 2 KB) then useLowProbCount=0
+ is a good default, since header deserialization makes a big speed difference.
+ Otherwise, useLowProbCount=1 is a good default, since the speed difference is small.
+ @return : tableLog,
+ or an errorCode, which can be tested using FSE_isError() */
+FSE_PUBLIC_API size_t FSE_normalizeCount(short* normalizedCounter, unsigned tableLog,
+ const unsigned* count, size_t srcSize, unsigned maxSymbolValue, unsigned useLowProbCount);
+
+/*! FSE_NCountWriteBound():
+ Provides the maximum possible size of an FSE normalized table, given 'maxSymbolValue' and 'tableLog'.
+ Typically useful for allocation purpose. */
+FSE_PUBLIC_API size_t FSE_NCountWriteBound(unsigned maxSymbolValue, unsigned tableLog);
+
+/*! FSE_writeNCount():
+ Compactly save 'normalizedCounter' into 'buffer'.
+ @return : size of the compressed table,
+ or an errorCode, which can be tested using FSE_isError(). */
+FSE_PUBLIC_API size_t FSE_writeNCount (void* buffer, size_t bufferSize,
+ const short* normalizedCounter,
+ unsigned maxSymbolValue, unsigned tableLog);
+
+/*! Constructor and Destructor of FSE_CTable.
+ Note that FSE_CTable size depends on 'tableLog' and 'maxSymbolValue' */
+typedef unsigned FSE_CTable; /* don't allocate that. It's only meant to be more restrictive than void* */
+
+/*! FSE_buildCTable():
+ Builds `ct`, which must be already allocated, using FSE_createCTable().
+ @return : 0, or an errorCode, which can be tested using FSE_isError() */
+FSE_PUBLIC_API size_t FSE_buildCTable(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog);
+
+/*! FSE_compress_usingCTable():
+ Compress `src` using `ct` into `dst` which must be already allocated.
+ @return : size of compressed data (<= `dstCapacity`),
+ or 0 if compressed data could not fit into `dst`,
+ or an errorCode, which can be tested using FSE_isError() */
+FSE_PUBLIC_API size_t FSE_compress_usingCTable (void* dst, size_t dstCapacity, const void* src, size_t srcSize, const FSE_CTable* ct);
+
+/*!
+Tutorial :
+----------
+The first step is to count all symbols. FSE_count() does this job very fast.
+Result will be saved into 'count', a table of unsigned int, which must be already allocated, and have 'maxSymbolValuePtr[0]+1' cells.
+'src' is a table of bytes of size 'srcSize'. All values within 'src' MUST be <= maxSymbolValuePtr[0]
+maxSymbolValuePtr[0] will be updated, with its real value (necessarily <= original value)
+FSE_count() will return the number of occurrence of the most frequent symbol.
+This can be used to know if there is a single symbol within 'src', and to quickly evaluate its compressibility.
+If there is an error, the function will return an ErrorCode (which can be tested using FSE_isError()).
+
+The next step is to normalize the frequencies.
+FSE_normalizeCount() will ensure that sum of frequencies is == 2 ^'tableLog'.
+It also guarantees a minimum of 1 to any Symbol with frequency >= 1.
+You can use 'tableLog'==0 to mean "use default tableLog value".
+If you are unsure of which tableLog value to use, you can ask FSE_optimalTableLog(),
+which will provide the optimal valid tableLog given sourceSize, maxSymbolValue, and a user-defined maximum (0 means "default").
+
+The result of FSE_normalizeCount() will be saved into a table,
+called 'normalizedCounter', which is a table of signed short.
+'normalizedCounter' must be already allocated, and have at least 'maxSymbolValue+1' cells.
+The return value is tableLog if everything proceeded as expected.
+It is 0 if there is a single symbol within distribution.
+If there is an error (ex: invalid tableLog value), the function will return an ErrorCode (which can be tested using FSE_isError()).
+
+'normalizedCounter' can be saved in a compact manner to a memory area using FSE_writeNCount().
+'buffer' must be already allocated.
+For guaranteed success, buffer size must be at least FSE_headerBound().
+The result of the function is the number of bytes written into 'buffer'.
+If there is an error, the function will return an ErrorCode (which can be tested using FSE_isError(); ex : buffer size too small).
+
+'normalizedCounter' can then be used to create the compression table 'CTable'.
+The space required by 'CTable' must be already allocated, using FSE_createCTable().
+You can then use FSE_buildCTable() to fill 'CTable'.
+If there is an error, both functions will return an ErrorCode (which can be tested using FSE_isError()).
+
+'CTable' can then be used to compress 'src', with FSE_compress_usingCTable().
+Similar to FSE_count(), the convention is that 'src' is assumed to be a table of char of size 'srcSize'
+The function returns the size of compressed data (without header), necessarily <= `dstCapacity`.
+If it returns '0', compressed data could not fit into 'dst'.
+If there is an error, the function will return an ErrorCode (which can be tested using FSE_isError()).
+*/
+
+
+/* *** DECOMPRESSION *** */
+
+/*! FSE_readNCount():
+ Read compactly saved 'normalizedCounter' from 'rBuffer'.
+ @return : size read from 'rBuffer',
+ or an errorCode, which can be tested using FSE_isError().
+ maxSymbolValuePtr[0] and tableLogPtr[0] will also be updated with their respective values */
+FSE_PUBLIC_API size_t FSE_readNCount (short* normalizedCounter,
+ unsigned* maxSymbolValuePtr, unsigned* tableLogPtr,
+ const void* rBuffer, size_t rBuffSize);
+
+/*! FSE_readNCount_bmi2():
+ * Same as FSE_readNCount() but pass bmi2=1 when your CPU supports BMI2 and 0 otherwise.
+ */
+FSE_PUBLIC_API size_t FSE_readNCount_bmi2(short* normalizedCounter,
+ unsigned* maxSymbolValuePtr, unsigned* tableLogPtr,
+ const void* rBuffer, size_t rBuffSize, int bmi2);
+
+typedef unsigned FSE_DTable; /* don't allocate that. It's just a way to be more restrictive than void* */
+
+/*!
+Tutorial :
+----------
+(Note : these functions only decompress FSE-compressed blocks.
+ If block is uncompressed, use memcpy() instead
+ If block is a single repeated byte, use memset() instead )
+
+The first step is to obtain the normalized frequencies of symbols.
+This can be performed by FSE_readNCount() if it was saved using FSE_writeNCount().
+'normalizedCounter' must be already allocated, and have at least 'maxSymbolValuePtr[0]+1' cells of signed short.
+In practice, that means it's necessary to know 'maxSymbolValue' beforehand,
+or size the table to handle worst case situations (typically 256).
+FSE_readNCount() will provide 'tableLog' and 'maxSymbolValue'.
+The result of FSE_readNCount() is the number of bytes read from 'rBuffer'.
+Note that 'rBufferSize' must be at least 4 bytes, even if useful information is less than that.
+If there is an error, the function will return an error code, which can be tested using FSE_isError().
+
+The next step is to build the decompression tables 'FSE_DTable' from 'normalizedCounter'.
+This is performed by the function FSE_buildDTable().
+The space required by 'FSE_DTable' must be already allocated using FSE_createDTable().
+If there is an error, the function will return an error code, which can be tested using FSE_isError().
+
+`FSE_DTable` can then be used to decompress `cSrc`, with FSE_decompress_usingDTable().
+`cSrcSize` must be strictly correct, otherwise decompression will fail.
+FSE_decompress_usingDTable() result will tell how many bytes were regenerated (<=`dstCapacity`).
+If there is an error, the function will return an error code, which can be tested using FSE_isError(). (ex: dst buffer too small)
+*/
+
+#endif /* FSE_H */
+
+
+#if defined(FSE_STATIC_LINKING_ONLY) && !defined(FSE_H_FSE_STATIC_LINKING_ONLY)
+#define FSE_H_FSE_STATIC_LINKING_ONLY
+
+/* *** Dependency *** */
+#include "bitstream.h"
+
+
+/* *****************************************
+* Static allocation
+*******************************************/
+/* FSE buffer bounds */
+#define FSE_NCOUNTBOUND 512
+#define FSE_BLOCKBOUND(size) ((size) + ((size)>>7) + 4 /* fse states */ + sizeof(size_t) /* bitContainer */)
+#define FSE_COMPRESSBOUND(size) (FSE_NCOUNTBOUND + FSE_BLOCKBOUND(size)) /* Macro version, useful for static allocation */
+
+/* It is possible to statically allocate FSE CTable/DTable as a table of FSE_CTable/FSE_DTable using below macros */
+#define FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue) (1 + (1<<((maxTableLog)-1)) + (((maxSymbolValue)+1)*2))
+#define FSE_DTABLE_SIZE_U32(maxTableLog) (1 + (1<<(maxTableLog)))
+
+/* or use the size to malloc() space directly. Pay attention to alignment restrictions though */
+#define FSE_CTABLE_SIZE(maxTableLog, maxSymbolValue) (FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue) * sizeof(FSE_CTable))
+#define FSE_DTABLE_SIZE(maxTableLog) (FSE_DTABLE_SIZE_U32(maxTableLog) * sizeof(FSE_DTable))
+
+
+/* *****************************************
+ * FSE advanced API
+ ***************************************** */
+
+unsigned FSE_optimalTableLog_internal(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, unsigned minus);
+/**< same as FSE_optimalTableLog(), which used `minus==2` */
+
+size_t FSE_buildCTable_rle (FSE_CTable* ct, unsigned char symbolValue);
+/**< build a fake FSE_CTable, designed to compress always the same symbolValue */
+
+/* FSE_buildCTable_wksp() :
+ * Same as FSE_buildCTable(), but using an externally allocated scratch buffer (`workSpace`).
+ * `wkspSize` must be >= `FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(maxSymbolValue, tableLog)` of `unsigned`.
+ * See FSE_buildCTable_wksp() for breakdown of workspace usage.
+ */
+#define FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(maxSymbolValue, tableLog) (((maxSymbolValue + 2) + (1ull << (tableLog)))/2 + sizeof(U64)/sizeof(U32) /* additional 8 bytes for potential table overwrite */)
+#define FSE_BUILD_CTABLE_WORKSPACE_SIZE(maxSymbolValue, tableLog) (sizeof(unsigned) * FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(maxSymbolValue, tableLog))
+size_t FSE_buildCTable_wksp(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);
+
+#define FSE_BUILD_DTABLE_WKSP_SIZE(maxTableLog, maxSymbolValue) (sizeof(short) * (maxSymbolValue + 1) + (1ULL << maxTableLog) + 8)
+#define FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) ((FSE_BUILD_DTABLE_WKSP_SIZE(maxTableLog, maxSymbolValue) + sizeof(unsigned) - 1) / sizeof(unsigned))
+FSE_PUBLIC_API size_t FSE_buildDTable_wksp(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);
+/**< Same as FSE_buildDTable(), using an externally allocated `workspace` produced with `FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxSymbolValue)` */
+
+#define FSE_DECOMPRESS_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) (FSE_DTABLE_SIZE_U32(maxTableLog) + 1 + FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) + (FSE_MAX_SYMBOL_VALUE + 1) / 2 + 1)
+#define FSE_DECOMPRESS_WKSP_SIZE(maxTableLog, maxSymbolValue) (FSE_DECOMPRESS_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) * sizeof(unsigned))
+size_t FSE_decompress_wksp_bmi2(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize, int bmi2);
+/**< same as FSE_decompress(), using an externally allocated `workSpace` produced with `FSE_DECOMPRESS_WKSP_SIZE_U32(maxLog, maxSymbolValue)`.
+ * Set bmi2 to 1 if your CPU supports BMI2 or 0 if it doesn't */
+
+typedef enum {
+ FSE_repeat_none, /**< Cannot use the previous table */
+ FSE_repeat_check, /**< Can use the previous table but it must be checked */
+ FSE_repeat_valid /**< Can use the previous table and it is assumed to be valid */
+ } FSE_repeat;
+
+/* *****************************************
+* FSE symbol compression API
+*******************************************/
+/*!
+ This API consists of small unitary functions, which highly benefit from being inlined.
+ Hence their body are included in next section.
+*/
+typedef struct {
+ ptrdiff_t value;
+ const void* stateTable;
+ const void* symbolTT;
+ unsigned stateLog;
+} FSE_CState_t;
+
+static void FSE_initCState(FSE_CState_t* CStatePtr, const FSE_CTable* ct);
+
+static void FSE_encodeSymbol(BIT_CStream_t* bitC, FSE_CState_t* CStatePtr, unsigned symbol);
+
+static void FSE_flushCState(BIT_CStream_t* bitC, const FSE_CState_t* CStatePtr);
+
+/**<
+These functions are inner components of FSE_compress_usingCTable().
+They allow the creation of custom streams, mixing multiple tables and bit sources.
+
+A key property to keep in mind is that encoding and decoding are done **in reverse direction**.
+So the first symbol you will encode is the last you will decode, like a LIFO stack.
+
+You will need a few variables to track your CStream. They are :
+
+FSE_CTable ct; // Provided by FSE_buildCTable()
+BIT_CStream_t bitStream; // bitStream tracking structure
+FSE_CState_t state; // State tracking structure (can have several)
+
+
+The first thing to do is to init bitStream and state.
+ size_t errorCode = BIT_initCStream(&bitStream, dstBuffer, maxDstSize);
+ FSE_initCState(&state, ct);
+
+Note that BIT_initCStream() can produce an error code, so its result should be tested, using FSE_isError();
+You can then encode your input data, byte after byte.
+FSE_encodeSymbol() outputs a maximum of 'tableLog' bits at a time.
+Remember decoding will be done in reverse direction.
+ FSE_encodeByte(&bitStream, &state, symbol);
+
+At any time, you can also add any bit sequence.
+Note : maximum allowed nbBits is 25, for compatibility with 32-bits decoders
+ BIT_addBits(&bitStream, bitField, nbBits);
+
+The above methods don't commit data to memory, they just store it into local register, for speed.
+Local register size is 64-bits on 64-bits systems, 32-bits on 32-bits systems (size_t).
+Writing data to memory is a manual operation, performed by the flushBits function.
+ BIT_flushBits(&bitStream);
+
+Your last FSE encoding operation shall be to flush your last state value(s).
+ FSE_flushState(&bitStream, &state);
+
+Finally, you must close the bitStream.
+The function returns the size of CStream in bytes.
+If data couldn't fit into dstBuffer, it will return a 0 ( == not compressible)
+If there is an error, it returns an errorCode (which can be tested using FSE_isError()).
+ size_t size = BIT_closeCStream(&bitStream);
+*/
+
+
+/* *****************************************
+* FSE symbol decompression API
+*******************************************/
+typedef struct {
+ size_t state;
+ const void* table; /* precise table may vary, depending on U16 */
+} FSE_DState_t;
+
+
+static void FSE_initDState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD, const FSE_DTable* dt);
+
+static unsigned char FSE_decodeSymbol(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD);
+
+static unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr);
+
+/**<
+Let's now decompose FSE_decompress_usingDTable() into its unitary components.
+You will decode FSE-encoded symbols from the bitStream,
+and also any other bitFields you put in, **in reverse order**.
+
+You will need a few variables to track your bitStream. They are :
+
+BIT_DStream_t DStream; // Stream context
+FSE_DState_t DState; // State context. Multiple ones are possible
+FSE_DTable* DTablePtr; // Decoding table, provided by FSE_buildDTable()
+
+The first thing to do is to init the bitStream.
+ errorCode = BIT_initDStream(&DStream, srcBuffer, srcSize);
+
+You should then retrieve your initial state(s)
+(in reverse flushing order if you have several ones) :
+ errorCode = FSE_initDState(&DState, &DStream, DTablePtr);
+
+You can then decode your data, symbol after symbol.
+For information the maximum number of bits read by FSE_decodeSymbol() is 'tableLog'.
+Keep in mind that symbols are decoded in reverse order, like a LIFO stack (last in, first out).
+ unsigned char symbol = FSE_decodeSymbol(&DState, &DStream);
+
+You can retrieve any bitfield you eventually stored into the bitStream (in reverse order)
+Note : maximum allowed nbBits is 25, for 32-bits compatibility
+ size_t bitField = BIT_readBits(&DStream, nbBits);
+
+All above operations only read from local register (which size depends on size_t).
+Refueling the register from memory is manually performed by the reload method.
+ endSignal = FSE_reloadDStream(&DStream);
+
+BIT_reloadDStream() result tells if there is still some more data to read from DStream.
+BIT_DStream_unfinished : there is still some data left into the DStream.
+BIT_DStream_endOfBuffer : Dstream reached end of buffer. Its container may no longer be completely filled.
+BIT_DStream_completed : Dstream reached its exact end, corresponding in general to decompression completed.
+BIT_DStream_tooFar : Dstream went too far. Decompression result is corrupted.
+
+When reaching end of buffer (BIT_DStream_endOfBuffer), progress slowly, notably if you decode multiple symbols per loop,
+to properly detect the exact end of stream.
+After each decoded symbol, check if DStream is fully consumed using this simple test :
+ BIT_reloadDStream(&DStream) >= BIT_DStream_completed
+
+When it's done, verify decompression is fully completed, by checking both DStream and the relevant states.
+Checking if DStream has reached its end is performed by :
+ BIT_endOfDStream(&DStream);
+Check also the states. There might be some symbols left there, if some high probability ones (>50%) are possible.
+ FSE_endOfDState(&DState);
+*/
+
+
+/* *****************************************
+* FSE unsafe API
+*******************************************/
+static unsigned char FSE_decodeSymbolFast(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD);
+/* faster, but works only if nbBits is always >= 1 (otherwise, result will be corrupted) */
+
+
+/* *****************************************
+* Implementation of inlined functions
+*******************************************/
+typedef struct {
+ int deltaFindState;
+ U32 deltaNbBits;
+} FSE_symbolCompressionTransform; /* total 8 bytes */
+
+MEM_STATIC void FSE_initCState(FSE_CState_t* statePtr, const FSE_CTable* ct)
+{
+ const void* ptr = ct;
+ const U16* u16ptr = (const U16*) ptr;
+ const U32 tableLog = MEM_read16(ptr);
+ statePtr->value = (ptrdiff_t)1<<tableLog;
+ statePtr->stateTable = u16ptr+2;
+ statePtr->symbolTT = ct + 1 + (tableLog ? (1<<(tableLog-1)) : 1);
+ statePtr->stateLog = tableLog;
+}
+
+
+/*! FSE_initCState2() :
+* Same as FSE_initCState(), but the first symbol to include (which will be the last to be read)
+* uses the smallest state value possible, saving the cost of this symbol */
+MEM_STATIC void FSE_initCState2(FSE_CState_t* statePtr, const FSE_CTable* ct, U32 symbol)
+{
+ FSE_initCState(statePtr, ct);
+ { const FSE_symbolCompressionTransform symbolTT = ((const FSE_symbolCompressionTransform*)(statePtr->symbolTT))[symbol];
+ const U16* stateTable = (const U16*)(statePtr->stateTable);
+ U32 nbBitsOut = (U32)((symbolTT.deltaNbBits + (1<<15)) >> 16);
+ statePtr->value = (nbBitsOut << 16) - symbolTT.deltaNbBits;
+ statePtr->value = stateTable[(statePtr->value >> nbBitsOut) + symbolTT.deltaFindState];
+ }
+}
+
+MEM_STATIC void FSE_encodeSymbol(BIT_CStream_t* bitC, FSE_CState_t* statePtr, unsigned symbol)
+{
+ FSE_symbolCompressionTransform const symbolTT = ((const FSE_symbolCompressionTransform*)(statePtr->symbolTT))[symbol];
+ const U16* const stateTable = (const U16*)(statePtr->stateTable);
+ U32 const nbBitsOut = (U32)((statePtr->value + symbolTT.deltaNbBits) >> 16);
+ BIT_addBits(bitC, (size_t)statePtr->value, nbBitsOut);
+ statePtr->value = stateTable[ (statePtr->value >> nbBitsOut) + symbolTT.deltaFindState];
+}
+
+MEM_STATIC void FSE_flushCState(BIT_CStream_t* bitC, const FSE_CState_t* statePtr)
+{
+ BIT_addBits(bitC, (size_t)statePtr->value, statePtr->stateLog);
+ BIT_flushBits(bitC);
+}
+
+
+/* FSE_getMaxNbBits() :
+ * Approximate maximum cost of a symbol, in bits.
+ * Fractional get rounded up (i.e. a symbol with a normalized frequency of 3 gives the same result as a frequency of 2)
+ * note 1 : assume symbolValue is valid (<= maxSymbolValue)
+ * note 2 : if freq[symbolValue]==0, @return a fake cost of tableLog+1 bits */
+MEM_STATIC U32 FSE_getMaxNbBits(const void* symbolTTPtr, U32 symbolValue)
+{
+ const FSE_symbolCompressionTransform* symbolTT = (const FSE_symbolCompressionTransform*) symbolTTPtr;
+ return (symbolTT[symbolValue].deltaNbBits + ((1<<16)-1)) >> 16;
+}
+
+/* FSE_bitCost() :
+ * Approximate symbol cost, as fractional value, using fixed-point format (accuracyLog fractional bits)
+ * note 1 : assume symbolValue is valid (<= maxSymbolValue)
+ * note 2 : if freq[symbolValue]==0, @return a fake cost of tableLog+1 bits */
+MEM_STATIC U32 FSE_bitCost(const void* symbolTTPtr, U32 tableLog, U32 symbolValue, U32 accuracyLog)
+{
+ const FSE_symbolCompressionTransform* symbolTT = (const FSE_symbolCompressionTransform*) symbolTTPtr;
+ U32 const minNbBits = symbolTT[symbolValue].deltaNbBits >> 16;
+ U32 const threshold = (minNbBits+1) << 16;
+ assert(tableLog < 16);
+ assert(accuracyLog < 31-tableLog); /* ensure enough room for renormalization double shift */
+ { U32 const tableSize = 1 << tableLog;
+ U32 const deltaFromThreshold = threshold - (symbolTT[symbolValue].deltaNbBits + tableSize);
+ U32 const normalizedDeltaFromThreshold = (deltaFromThreshold << accuracyLog) >> tableLog; /* linear interpolation (very approximate) */
+ U32 const bitMultiplier = 1 << accuracyLog;
+ assert(symbolTT[symbolValue].deltaNbBits + tableSize <= threshold);
+ assert(normalizedDeltaFromThreshold <= bitMultiplier);
+ return (minNbBits+1)*bitMultiplier - normalizedDeltaFromThreshold;
+ }
+}
+
+
+/* ====== Decompression ====== */
+
+typedef struct {
+ U16 tableLog;
+ U16 fastMode;
+} FSE_DTableHeader; /* sizeof U32 */
+
+typedef struct
+{
+ unsigned short newState;
+ unsigned char symbol;
+ unsigned char nbBits;
+} FSE_decode_t; /* size == U32 */
+
+MEM_STATIC void FSE_initDState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD, const FSE_DTable* dt)
+{
+ const void* ptr = dt;
+ const FSE_DTableHeader* const DTableH = (const FSE_DTableHeader*)ptr;
+ DStatePtr->state = BIT_readBits(bitD, DTableH->tableLog);
+ BIT_reloadDStream(bitD);
+ DStatePtr->table = dt + 1;
+}
+
+MEM_STATIC BYTE FSE_peekSymbol(const FSE_DState_t* DStatePtr)
+{
+ FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
+ return DInfo.symbol;
+}
+
+MEM_STATIC void FSE_updateState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD)
+{
+ FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
+ U32 const nbBits = DInfo.nbBits;
+ size_t const lowBits = BIT_readBits(bitD, nbBits);
+ DStatePtr->state = DInfo.newState + lowBits;
+}
+
+MEM_STATIC BYTE FSE_decodeSymbol(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD)
+{
+ FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
+ U32 const nbBits = DInfo.nbBits;
+ BYTE const symbol = DInfo.symbol;
+ size_t const lowBits = BIT_readBits(bitD, nbBits);
+
+ DStatePtr->state = DInfo.newState + lowBits;
+ return symbol;
+}
+
+/*! FSE_decodeSymbolFast() :
+ unsafe, only works if no symbol has a probability > 50% */
+MEM_STATIC BYTE FSE_decodeSymbolFast(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD)
+{
+ FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
+ U32 const nbBits = DInfo.nbBits;
+ BYTE const symbol = DInfo.symbol;
+ size_t const lowBits = BIT_readBitsFast(bitD, nbBits);
+
+ DStatePtr->state = DInfo.newState + lowBits;
+ return symbol;
+}
+
+MEM_STATIC unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr)
+{
+ return DStatePtr->state == 0;
+}
+
+
+
+#ifndef FSE_COMMONDEFS_ONLY
+
+/* **************************************************************
+* Tuning parameters
+****************************************************************/
+/*!MEMORY_USAGE :
+* Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.)
+* Increasing memory usage improves compression ratio
+* Reduced memory usage can improve speed, due to cache effect
+* Recommended max value is 14, for 16KB, which nicely fits into Intel x86 L1 cache */
+#ifndef FSE_MAX_MEMORY_USAGE
+# define FSE_MAX_MEMORY_USAGE 14
+#endif
+#ifndef FSE_DEFAULT_MEMORY_USAGE
+# define FSE_DEFAULT_MEMORY_USAGE 13
+#endif
+#if (FSE_DEFAULT_MEMORY_USAGE > FSE_MAX_MEMORY_USAGE)
+# error "FSE_DEFAULT_MEMORY_USAGE must be <= FSE_MAX_MEMORY_USAGE"
+#endif
+
+/*!FSE_MAX_SYMBOL_VALUE :
+* Maximum symbol value authorized.
+* Required for proper stack allocation */
+#ifndef FSE_MAX_SYMBOL_VALUE
+# define FSE_MAX_SYMBOL_VALUE 255
+#endif
+
+/* **************************************************************
+* template functions type & suffix
+****************************************************************/
+#define FSE_FUNCTION_TYPE BYTE
+#define FSE_FUNCTION_EXTENSION
+#define FSE_DECODE_TYPE FSE_decode_t
+
+
+#endif /* !FSE_COMMONDEFS_ONLY */
+
+
+/* ***************************************************************
+* Constants
+*****************************************************************/
+#define FSE_MAX_TABLELOG (FSE_MAX_MEMORY_USAGE-2)
+#define FSE_MAX_TABLESIZE (1U<<FSE_MAX_TABLELOG)
+#define FSE_MAXTABLESIZE_MASK (FSE_MAX_TABLESIZE-1)
+#define FSE_DEFAULT_TABLELOG (FSE_DEFAULT_MEMORY_USAGE-2)
+#define FSE_MIN_TABLELOG 5
+
+#define FSE_TABLELOG_ABSOLUTE_MAX 15
+#if FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX
+# error "FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX is not supported"
+#endif
+
+#define FSE_TABLESTEP(tableSize) (((tableSize)>>1) + ((tableSize)>>3) + 3)
+
+
+#endif /* FSE_STATIC_LINKING_ONLY */
+
+
+#if defined (__cplusplus)
+}
+#endif
diff --git a/third_party/zstd/lib/common/fse_decompress.c b/third_party/zstd/lib/common/fse_decompress.c
new file mode 100644
index 0000000000..0dcc4640d0
--- /dev/null
+++ b/third_party/zstd/lib/common/fse_decompress.c
@@ -0,0 +1,313 @@
+/* ******************************************************************
+ * FSE : Finite State Entropy decoder
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ *
+ * You can contact the author at :
+ * - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
+ * - Public forum : https://groups.google.com/forum/#!forum/lz4c
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+****************************************************************** */
+
+
+/* **************************************************************
+* Includes
+****************************************************************/
+#include "debug.h" /* assert */
+#include "bitstream.h"
+#include "compiler.h"
+#define FSE_STATIC_LINKING_ONLY
+#include "fse.h"
+#include "error_private.h"
+#include "zstd_deps.h" /* ZSTD_memcpy */
+#include "bits.h" /* ZSTD_highbit32 */
+
+
+/* **************************************************************
+* Error Management
+****************************************************************/
+#define FSE_isError ERR_isError
+#define FSE_STATIC_ASSERT(c) DEBUG_STATIC_ASSERT(c) /* use only *after* variable declarations */
+
+
+/* **************************************************************
+* Templates
+****************************************************************/
+/*
+ designed to be included
+ for type-specific functions (template emulation in C)
+ Objective is to write these functions only once, for improved maintenance
+*/
+
+/* safety checks */
+#ifndef FSE_FUNCTION_EXTENSION
+# error "FSE_FUNCTION_EXTENSION must be defined"
+#endif
+#ifndef FSE_FUNCTION_TYPE
+# error "FSE_FUNCTION_TYPE must be defined"
+#endif
+
+/* Function names */
+#define FSE_CAT(X,Y) X##Y
+#define FSE_FUNCTION_NAME(X,Y) FSE_CAT(X,Y)
+#define FSE_TYPE_NAME(X,Y) FSE_CAT(X,Y)
+
+static size_t FSE_buildDTable_internal(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize)
+{
+ void* const tdPtr = dt+1; /* because *dt is unsigned, 32-bits aligned on 32-bits */
+ FSE_DECODE_TYPE* const tableDecode = (FSE_DECODE_TYPE*) (tdPtr);
+ U16* symbolNext = (U16*)workSpace;
+ BYTE* spread = (BYTE*)(symbolNext + maxSymbolValue + 1);
+
+ U32 const maxSV1 = maxSymbolValue + 1;
+ U32 const tableSize = 1 << tableLog;
+ U32 highThreshold = tableSize-1;
+
+ /* Sanity Checks */
+ if (FSE_BUILD_DTABLE_WKSP_SIZE(tableLog, maxSymbolValue) > wkspSize) return ERROR(maxSymbolValue_tooLarge);
+ if (maxSymbolValue > FSE_MAX_SYMBOL_VALUE) return ERROR(maxSymbolValue_tooLarge);
+ if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);
+
+ /* Init, lay down lowprob symbols */
+ { FSE_DTableHeader DTableH;
+ DTableH.tableLog = (U16)tableLog;
+ DTableH.fastMode = 1;
+ { S16 const largeLimit= (S16)(1 << (tableLog-1));
+ U32 s;
+ for (s=0; s<maxSV1; s++) {
+ if (normalizedCounter[s]==-1) {
+ tableDecode[highThreshold--].symbol = (FSE_FUNCTION_TYPE)s;
+ symbolNext[s] = 1;
+ } else {
+ if (normalizedCounter[s] >= largeLimit) DTableH.fastMode=0;
+ symbolNext[s] = (U16)normalizedCounter[s];
+ } } }
+ ZSTD_memcpy(dt, &DTableH, sizeof(DTableH));
+ }
+
+ /* Spread symbols */
+ if (highThreshold == tableSize - 1) {
+ size_t const tableMask = tableSize-1;
+ size_t const step = FSE_TABLESTEP(tableSize);
+ /* First lay down the symbols in order.
+ * We use a uint64_t to lay down 8 bytes at a time. This reduces branch
+ * misses since small blocks generally have small table logs, so nearly
+ * all symbols have counts <= 8. We ensure we have 8 bytes at the end of
+ * our buffer to handle the over-write.
+ */
+ { U64 const add = 0x0101010101010101ull;
+ size_t pos = 0;
+ U64 sv = 0;
+ U32 s;
+ for (s=0; s<maxSV1; ++s, sv += add) {
+ int i;
+ int const n = normalizedCounter[s];
+ MEM_write64(spread + pos, sv);
+ for (i = 8; i < n; i += 8) {
+ MEM_write64(spread + pos + i, sv);
+ }
+ pos += (size_t)n;
+ } }
+ /* Now we spread those positions across the table.
+ * The benefit of doing it in two stages is that we avoid the
+ * variable size inner loop, which caused lots of branch misses.
+ * Now we can run through all the positions without any branch misses.
+ * We unroll the loop twice, since that is what empirically worked best.
+ */
+ {
+ size_t position = 0;
+ size_t s;
+ size_t const unroll = 2;
+ assert(tableSize % unroll == 0); /* FSE_MIN_TABLELOG is 5 */
+ for (s = 0; s < (size_t)tableSize; s += unroll) {
+ size_t u;
+ for (u = 0; u < unroll; ++u) {
+ size_t const uPosition = (position + (u * step)) & tableMask;
+ tableDecode[uPosition].symbol = spread[s + u];
+ }
+ position = (position + (unroll * step)) & tableMask;
+ }
+ assert(position == 0);
+ }
+ } else {
+ U32 const tableMask = tableSize-1;
+ U32 const step = FSE_TABLESTEP(tableSize);
+ U32 s, position = 0;
+ for (s=0; s<maxSV1; s++) {
+ int i;
+ for (i=0; i<normalizedCounter[s]; i++) {
+ tableDecode[position].symbol = (FSE_FUNCTION_TYPE)s;
+ position = (position + step) & tableMask;
+ while (position > highThreshold) position = (position + step) & tableMask; /* lowprob area */
+ } }
+ if (position!=0) return ERROR(GENERIC); /* position must reach all cells once, otherwise normalizedCounter is incorrect */
+ }
+
+ /* Build Decoding table */
+ { U32 u;
+ for (u=0; u<tableSize; u++) {
+ FSE_FUNCTION_TYPE const symbol = (FSE_FUNCTION_TYPE)(tableDecode[u].symbol);
+ U32 const nextState = symbolNext[symbol]++;
+ tableDecode[u].nbBits = (BYTE) (tableLog - ZSTD_highbit32(nextState) );
+ tableDecode[u].newState = (U16) ( (nextState << tableDecode[u].nbBits) - tableSize);
+ } }
+
+ return 0;
+}
+
+size_t FSE_buildDTable_wksp(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize)
+{
+ return FSE_buildDTable_internal(dt, normalizedCounter, maxSymbolValue, tableLog, workSpace, wkspSize);
+}
+
+
+#ifndef FSE_COMMONDEFS_ONLY
+
+/*-*******************************************************
+* Decompression (Byte symbols)
+*********************************************************/
+
+FORCE_INLINE_TEMPLATE size_t FSE_decompress_usingDTable_generic(
+ void* dst, size_t maxDstSize,
+ const void* cSrc, size_t cSrcSize,
+ const FSE_DTable* dt, const unsigned fast)
+{
+ BYTE* const ostart = (BYTE*) dst;
+ BYTE* op = ostart;
+ BYTE* const omax = op + maxDstSize;
+ BYTE* const olimit = omax-3;
+
+ BIT_DStream_t bitD;
+ FSE_DState_t state1;
+ FSE_DState_t state2;
+
+ /* Init */
+ CHECK_F(BIT_initDStream(&bitD, cSrc, cSrcSize));
+
+ FSE_initDState(&state1, &bitD, dt);
+ FSE_initDState(&state2, &bitD, dt);
+
+#define FSE_GETSYMBOL(statePtr) fast ? FSE_decodeSymbolFast(statePtr, &bitD) : FSE_decodeSymbol(statePtr, &bitD)
+
+ /* 4 symbols per loop */
+ for ( ; (BIT_reloadDStream(&bitD)==BIT_DStream_unfinished) & (op<olimit) ; op+=4) {
+ op[0] = FSE_GETSYMBOL(&state1);
+
+ if (FSE_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8) /* This test must be static */
+ BIT_reloadDStream(&bitD);
+
+ op[1] = FSE_GETSYMBOL(&state2);
+
+ if (FSE_MAX_TABLELOG*4+7 > sizeof(bitD.bitContainer)*8) /* This test must be static */
+ { if (BIT_reloadDStream(&bitD) > BIT_DStream_unfinished) { op+=2; break; } }
+
+ op[2] = FSE_GETSYMBOL(&state1);
+
+ if (FSE_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8) /* This test must be static */
+ BIT_reloadDStream(&bitD);
+
+ op[3] = FSE_GETSYMBOL(&state2);
+ }
+
+ /* tail */
+ /* note : BIT_reloadDStream(&bitD) >= FSE_DStream_partiallyFilled; Ends at exactly BIT_DStream_completed */
+ while (1) {
+ if (op>(omax-2)) return ERROR(dstSize_tooSmall);
+ *op++ = FSE_GETSYMBOL(&state1);
+ if (BIT_reloadDStream(&bitD)==BIT_DStream_overflow) {
+ *op++ = FSE_GETSYMBOL(&state2);
+ break;
+ }
+
+ if (op>(omax-2)) return ERROR(dstSize_tooSmall);
+ *op++ = FSE_GETSYMBOL(&state2);
+ if (BIT_reloadDStream(&bitD)==BIT_DStream_overflow) {
+ *op++ = FSE_GETSYMBOL(&state1);
+ break;
+ } }
+
+ assert(op >= ostart);
+ return (size_t)(op-ostart);
+}
+
+typedef struct {
+ short ncount[FSE_MAX_SYMBOL_VALUE + 1];
+} FSE_DecompressWksp;
+
+
+FORCE_INLINE_TEMPLATE size_t FSE_decompress_wksp_body(
+ void* dst, size_t dstCapacity,
+ const void* cSrc, size_t cSrcSize,
+ unsigned maxLog, void* workSpace, size_t wkspSize,
+ int bmi2)
+{
+ const BYTE* const istart = (const BYTE*)cSrc;
+ const BYTE* ip = istart;
+ unsigned tableLog;
+ unsigned maxSymbolValue = FSE_MAX_SYMBOL_VALUE;
+ FSE_DecompressWksp* const wksp = (FSE_DecompressWksp*)workSpace;
+ size_t const dtablePos = sizeof(FSE_DecompressWksp) / sizeof(FSE_DTable);
+ FSE_DTable* const dtable = (FSE_DTable*)workSpace + dtablePos;
+
+ FSE_STATIC_ASSERT((FSE_MAX_SYMBOL_VALUE + 1) % 2 == 0);
+ if (wkspSize < sizeof(*wksp)) return ERROR(GENERIC);
+
+ /* correct offset to dtable depends on this property */
+ FSE_STATIC_ASSERT(sizeof(FSE_DecompressWksp) % sizeof(FSE_DTable) == 0);
+
+ /* normal FSE decoding mode */
+ { size_t const NCountLength =
+ FSE_readNCount_bmi2(wksp->ncount, &maxSymbolValue, &tableLog, istart, cSrcSize, bmi2);
+ if (FSE_isError(NCountLength)) return NCountLength;
+ if (tableLog > maxLog) return ERROR(tableLog_tooLarge);
+ assert(NCountLength <= cSrcSize);
+ ip += NCountLength;
+ cSrcSize -= NCountLength;
+ }
+
+ if (FSE_DECOMPRESS_WKSP_SIZE(tableLog, maxSymbolValue) > wkspSize) return ERROR(tableLog_tooLarge);
+ assert(sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog) <= wkspSize);
+ workSpace = (BYTE*)workSpace + sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog);
+ wkspSize -= sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog);
+
+ CHECK_F( FSE_buildDTable_internal(dtable, wksp->ncount, maxSymbolValue, tableLog, workSpace, wkspSize) );
+
+ {
+ const void* ptr = dtable;
+ const FSE_DTableHeader* DTableH = (const FSE_DTableHeader*)ptr;
+ const U32 fastMode = DTableH->fastMode;
+
+ /* select fast mode (static) */
+ if (fastMode) return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, dtable, 1);
+ return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, dtable, 0);
+ }
+}
+
+/* Avoids the FORCE_INLINE of the _body() function. */
+static size_t FSE_decompress_wksp_body_default(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize)
+{
+ return FSE_decompress_wksp_body(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize, 0);
+}
+
+#if DYNAMIC_BMI2
+BMI2_TARGET_ATTRIBUTE static size_t FSE_decompress_wksp_body_bmi2(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize)
+{
+ return FSE_decompress_wksp_body(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize, 1);
+}
+#endif
+
+size_t FSE_decompress_wksp_bmi2(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize, int bmi2)
+{
+#if DYNAMIC_BMI2
+ if (bmi2) {
+ return FSE_decompress_wksp_body_bmi2(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize);
+ }
+#endif
+ (void)bmi2;
+ return FSE_decompress_wksp_body_default(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize);
+}
+
+#endif /* FSE_COMMONDEFS_ONLY */
diff --git a/third_party/zstd/lib/common/huf.h b/third_party/zstd/lib/common/huf.h
new file mode 100644
index 0000000000..99bf85d6f4
--- /dev/null
+++ b/third_party/zstd/lib/common/huf.h
@@ -0,0 +1,286 @@
+/* ******************************************************************
+ * huff0 huffman codec,
+ * part of Finite State Entropy library
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ *
+ * You can contact the author at :
+ * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+****************************************************************** */
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+#ifndef HUF_H_298734234
+#define HUF_H_298734234
+
+/* *** Dependencies *** */
+#include "zstd_deps.h" /* size_t */
+#include "mem.h" /* U32 */
+#define FSE_STATIC_LINKING_ONLY
+#include "fse.h"
+
+
+/* *** Tool functions *** */
+#define HUF_BLOCKSIZE_MAX (128 * 1024) /**< maximum input size for a single block compressed with HUF_compress */
+size_t HUF_compressBound(size_t size); /**< maximum compressed size (worst case) */
+
+/* Error Management */
+unsigned HUF_isError(size_t code); /**< tells if a return value is an error code */
+const char* HUF_getErrorName(size_t code); /**< provides error code string (useful for debugging) */
+
+
+#define HUF_WORKSPACE_SIZE ((8 << 10) + 512 /* sorting scratch space */)
+#define HUF_WORKSPACE_SIZE_U64 (HUF_WORKSPACE_SIZE / sizeof(U64))
+
+/* *** Constants *** */
+#define HUF_TABLELOG_MAX 12 /* max runtime value of tableLog (due to static allocation); can be modified up to HUF_TABLELOG_ABSOLUTEMAX */
+#define HUF_TABLELOG_DEFAULT 11 /* default tableLog value when none specified */
+#define HUF_SYMBOLVALUE_MAX 255
+
+#define HUF_TABLELOG_ABSOLUTEMAX 12 /* absolute limit of HUF_MAX_TABLELOG. Beyond that value, code does not work */
+#if (HUF_TABLELOG_MAX > HUF_TABLELOG_ABSOLUTEMAX)
+# error "HUF_TABLELOG_MAX is too large !"
+#endif
+
+
+/* ****************************************
+* Static allocation
+******************************************/
+/* HUF buffer bounds */
+#define HUF_CTABLEBOUND 129
+#define HUF_BLOCKBOUND(size) (size + (size>>8) + 8) /* only true when incompressible is pre-filtered with fast heuristic */
+#define HUF_COMPRESSBOUND(size) (HUF_CTABLEBOUND + HUF_BLOCKBOUND(size)) /* Macro version, useful for static allocation */
+
+/* static allocation of HUF's Compression Table */
+/* this is a private definition, just exposed for allocation and strict aliasing purpose. never EVER access its members directly */
+typedef size_t HUF_CElt; /* consider it an incomplete type */
+#define HUF_CTABLE_SIZE_ST(maxSymbolValue) ((maxSymbolValue)+2) /* Use tables of size_t, for proper alignment */
+#define HUF_CTABLE_SIZE(maxSymbolValue) (HUF_CTABLE_SIZE_ST(maxSymbolValue) * sizeof(size_t))
+#define HUF_CREATE_STATIC_CTABLE(name, maxSymbolValue) \
+ HUF_CElt name[HUF_CTABLE_SIZE_ST(maxSymbolValue)] /* no final ; */
+
+/* static allocation of HUF's DTable */
+typedef U32 HUF_DTable;
+#define HUF_DTABLE_SIZE(maxTableLog) (1 + (1<<(maxTableLog)))
+#define HUF_CREATE_STATIC_DTABLEX1(DTable, maxTableLog) \
+ HUF_DTable DTable[HUF_DTABLE_SIZE((maxTableLog)-1)] = { ((U32)((maxTableLog)-1) * 0x01000001) }
+#define HUF_CREATE_STATIC_DTABLEX2(DTable, maxTableLog) \
+ HUF_DTable DTable[HUF_DTABLE_SIZE(maxTableLog)] = { ((U32)(maxTableLog) * 0x01000001) }
+
+
+/* ****************************************
+* Advanced decompression functions
+******************************************/
+
+/**
+ * Huffman flags bitset.
+ * For all flags, 0 is the default value.
+ */
+typedef enum {
+ /**
+ * If compiled with DYNAMIC_BMI2: Set flag only if the CPU supports BMI2 at runtime.
+ * Otherwise: Ignored.
+ */
+ HUF_flags_bmi2 = (1 << 0),
+ /**
+ * If set: Test possible table depths to find the one that produces the smallest header + encoded size.
+ * If unset: Use heuristic to find the table depth.
+ */
+ HUF_flags_optimalDepth = (1 << 1),
+ /**
+ * If set: If the previous table can encode the input, always reuse the previous table.
+ * If unset: If the previous table can encode the input, reuse the previous table if it results in a smaller output.
+ */
+ HUF_flags_preferRepeat = (1 << 2),
+ /**
+ * If set: Sample the input and check if the sample is uncompressible, if it is then don't attempt to compress.
+ * If unset: Always histogram the entire input.
+ */
+ HUF_flags_suspectUncompressible = (1 << 3),
+ /**
+ * If set: Don't use assembly implementations
+ * If unset: Allow using assembly implementations
+ */
+ HUF_flags_disableAsm = (1 << 4),
+ /**
+ * If set: Don't use the fast decoding loop, always use the fallback decoding loop.
+ * If unset: Use the fast decoding loop when possible.
+ */
+ HUF_flags_disableFast = (1 << 5)
+} HUF_flags_e;
+
+
+/* ****************************************
+ * HUF detailed API
+ * ****************************************/
+#define HUF_OPTIMAL_DEPTH_THRESHOLD ZSTD_btultra
+
+/*! HUF_compress() does the following:
+ * 1. count symbol occurrence from source[] into table count[] using FSE_count() (exposed within "fse.h")
+ * 2. (optional) refine tableLog using HUF_optimalTableLog()
+ * 3. build Huffman table from count using HUF_buildCTable()
+ * 4. save Huffman table to memory buffer using HUF_writeCTable()
+ * 5. encode the data stream using HUF_compress4X_usingCTable()
+ *
+ * The following API allows targeting specific sub-functions for advanced tasks.
+ * For example, it's possible to compress several blocks using the same 'CTable',
+ * or to save and regenerate 'CTable' using external methods.
+ */
+unsigned HUF_minTableLog(unsigned symbolCardinality);
+unsigned HUF_cardinality(const unsigned* count, unsigned maxSymbolValue);
+unsigned HUF_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, void* workSpace,
+ size_t wkspSize, HUF_CElt* table, const unsigned* count, int flags); /* table is used as scratch space for building and testing tables, not a return value */
+size_t HUF_writeCTable_wksp(void* dst, size_t maxDstSize, const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog, void* workspace, size_t workspaceSize);
+size_t HUF_compress4X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int flags);
+size_t HUF_estimateCompressedSize(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue);
+int HUF_validateCTable(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue);
+
+typedef enum {
+ HUF_repeat_none, /**< Cannot use the previous table */
+ HUF_repeat_check, /**< Can use the previous table but it must be checked. Note : The previous table must have been constructed by HUF_compress{1, 4}X_repeat */
+ HUF_repeat_valid /**< Can use the previous table and it is assumed to be valid */
+ } HUF_repeat;
+
+/** HUF_compress4X_repeat() :
+ * Same as HUF_compress4X_wksp(), but considers using hufTable if *repeat != HUF_repeat_none.
+ * If it uses hufTable it does not modify hufTable or repeat.
+ * If it doesn't, it sets *repeat = HUF_repeat_none, and it sets hufTable to the table used.
+ * If preferRepeat then the old table will always be used if valid.
+ * If suspectUncompressible then some sampling checks will be run to potentially skip huffman coding */
+size_t HUF_compress4X_repeat(void* dst, size_t dstSize,
+ const void* src, size_t srcSize,
+ unsigned maxSymbolValue, unsigned tableLog,
+ void* workSpace, size_t wkspSize, /**< `workSpace` must be aligned on 4-bytes boundaries, `wkspSize` must be >= HUF_WORKSPACE_SIZE */
+ HUF_CElt* hufTable, HUF_repeat* repeat, int flags);
+
+/** HUF_buildCTable_wksp() :
+ * Same as HUF_buildCTable(), but using externally allocated scratch buffer.
+ * `workSpace` must be aligned on 4-bytes boundaries, and its size must be >= HUF_CTABLE_WORKSPACE_SIZE.
+ */
+#define HUF_CTABLE_WORKSPACE_SIZE_U32 ((4 * (HUF_SYMBOLVALUE_MAX + 1)) + 192)
+#define HUF_CTABLE_WORKSPACE_SIZE (HUF_CTABLE_WORKSPACE_SIZE_U32 * sizeof(unsigned))
+size_t HUF_buildCTable_wksp (HUF_CElt* tree,
+ const unsigned* count, U32 maxSymbolValue, U32 maxNbBits,
+ void* workSpace, size_t wkspSize);
+
+/*! HUF_readStats() :
+ * Read compact Huffman tree, saved by HUF_writeCTable().
+ * `huffWeight` is destination buffer.
+ * @return : size read from `src` , or an error Code .
+ * Note : Needed by HUF_readCTable() and HUF_readDTableXn() . */
+size_t HUF_readStats(BYTE* huffWeight, size_t hwSize,
+ U32* rankStats, U32* nbSymbolsPtr, U32* tableLogPtr,
+ const void* src, size_t srcSize);
+
+/*! HUF_readStats_wksp() :
+ * Same as HUF_readStats() but takes an external workspace which must be
+ * 4-byte aligned and its size must be >= HUF_READ_STATS_WORKSPACE_SIZE.
+ * If the CPU has BMI2 support, pass bmi2=1, otherwise pass bmi2=0.
+ */
+#define HUF_READ_STATS_WORKSPACE_SIZE_U32 FSE_DECOMPRESS_WKSP_SIZE_U32(6, HUF_TABLELOG_MAX-1)
+#define HUF_READ_STATS_WORKSPACE_SIZE (HUF_READ_STATS_WORKSPACE_SIZE_U32 * sizeof(unsigned))
+size_t HUF_readStats_wksp(BYTE* huffWeight, size_t hwSize,
+ U32* rankStats, U32* nbSymbolsPtr, U32* tableLogPtr,
+ const void* src, size_t srcSize,
+ void* workspace, size_t wkspSize,
+ int flags);
+
+/** HUF_readCTable() :
+ * Loading a CTable saved with HUF_writeCTable() */
+size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void* src, size_t srcSize, unsigned *hasZeroWeights);
+
+/** HUF_getNbBitsFromCTable() :
+ * Read nbBits from CTable symbolTable, for symbol `symbolValue` presumed <= HUF_SYMBOLVALUE_MAX
+ * Note 1 : If symbolValue > HUF_readCTableHeader(symbolTable).maxSymbolValue, returns 0
+ * Note 2 : is not inlined, as HUF_CElt definition is private
+ */
+U32 HUF_getNbBitsFromCTable(const HUF_CElt* symbolTable, U32 symbolValue);
+
+typedef struct {
+ BYTE tableLog;
+ BYTE maxSymbolValue;
+ BYTE unused[sizeof(size_t) - 2];
+} HUF_CTableHeader;
+
+/** HUF_readCTableHeader() :
+ * @returns The header from the CTable specifying the tableLog and the maxSymbolValue.
+ */
+HUF_CTableHeader HUF_readCTableHeader(HUF_CElt const* ctable);
+
+/*
+ * HUF_decompress() does the following:
+ * 1. select the decompression algorithm (X1, X2) based on pre-computed heuristics
+ * 2. build Huffman table from save, using HUF_readDTableX?()
+ * 3. decode 1 or 4 segments in parallel using HUF_decompress?X?_usingDTable()
+ */
+
+/** HUF_selectDecoder() :
+ * Tells which decoder is likely to decode faster,
+ * based on a set of pre-computed metrics.
+ * @return : 0==HUF_decompress4X1, 1==HUF_decompress4X2 .
+ * Assumption : 0 < dstSize <= 128 KB */
+U32 HUF_selectDecoder (size_t dstSize, size_t cSrcSize);
+
+/**
+ * The minimum workspace size for the `workSpace` used in
+ * HUF_readDTableX1_wksp() and HUF_readDTableX2_wksp().
+ *
+ * The space used depends on HUF_TABLELOG_MAX, ranging from ~1500 bytes when
+ * HUF_TABLE_LOG_MAX=12 to ~1850 bytes when HUF_TABLE_LOG_MAX=15.
+ * Buffer overflow errors may potentially occur if code modifications result in
+ * a required workspace size greater than that specified in the following
+ * macro.
+ */
+#define HUF_DECOMPRESS_WORKSPACE_SIZE ((2 << 10) + (1 << 9))
+#define HUF_DECOMPRESS_WORKSPACE_SIZE_U32 (HUF_DECOMPRESS_WORKSPACE_SIZE / sizeof(U32))
+
+
+/* ====================== */
+/* single stream variants */
+/* ====================== */
+
+size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int flags);
+/** HUF_compress1X_repeat() :
+ * Same as HUF_compress1X_wksp(), but considers using hufTable if *repeat != HUF_repeat_none.
+ * If it uses hufTable it does not modify hufTable or repeat.
+ * If it doesn't, it sets *repeat = HUF_repeat_none, and it sets hufTable to the table used.
+ * If preferRepeat then the old table will always be used if valid.
+ * If suspectUncompressible then some sampling checks will be run to potentially skip huffman coding */
+size_t HUF_compress1X_repeat(void* dst, size_t dstSize,
+ const void* src, size_t srcSize,
+ unsigned maxSymbolValue, unsigned tableLog,
+ void* workSpace, size_t wkspSize, /**< `workSpace` must be aligned on 4-bytes boundaries, `wkspSize` must be >= HUF_WORKSPACE_SIZE */
+ HUF_CElt* hufTable, HUF_repeat* repeat, int flags);
+
+size_t HUF_decompress1X_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags);
+#ifndef HUF_FORCE_DECOMPRESS_X1
+size_t HUF_decompress1X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags); /**< double-symbols decoder */
+#endif
+
+/* BMI2 variants.
+ * If the CPU has BMI2 support, pass bmi2=1, otherwise pass bmi2=0.
+ */
+size_t HUF_decompress1X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int flags);
+#ifndef HUF_FORCE_DECOMPRESS_X2
+size_t HUF_decompress1X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags);
+#endif
+size_t HUF_decompress4X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int flags);
+size_t HUF_decompress4X_hufOnly_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags);
+#ifndef HUF_FORCE_DECOMPRESS_X2
+size_t HUF_readDTableX1_wksp(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize, int flags);
+#endif
+#ifndef HUF_FORCE_DECOMPRESS_X1
+size_t HUF_readDTableX2_wksp(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize, int flags);
+#endif
+
+#endif /* HUF_H_298734234 */
+
+#if defined (__cplusplus)
+}
+#endif
diff --git a/third_party/zstd/lib/common/mem.h b/third_party/zstd/lib/common/mem.h
new file mode 100644
index 0000000000..096f4be519
--- /dev/null
+++ b/third_party/zstd/lib/common/mem.h
@@ -0,0 +1,426 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef MEM_H_MODULE
+#define MEM_H_MODULE
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/*-****************************************
+* Dependencies
+******************************************/
+#include <stddef.h> /* size_t, ptrdiff_t */
+#include "compiler.h" /* __has_builtin */
+#include "debug.h" /* DEBUG_STATIC_ASSERT */
+#include "zstd_deps.h" /* ZSTD_memcpy */
+
+
+/*-****************************************
+* Compiler specifics
+******************************************/
+#if defined(_MSC_VER) /* Visual Studio */
+# include <stdlib.h> /* _byteswap_ulong */
+# include <intrin.h> /* _byteswap_* */
+#endif
+
+/*-**************************************************************
+* Basic Types
+*****************************************************************/
+#if !defined (__VMS) && (defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+# if defined(_AIX)
+# include <inttypes.h>
+# else
+# include <stdint.h> /* intptr_t */
+# endif
+ typedef uint8_t BYTE;
+ typedef uint8_t U8;
+ typedef int8_t S8;
+ typedef uint16_t U16;
+ typedef int16_t S16;
+ typedef uint32_t U32;
+ typedef int32_t S32;
+ typedef uint64_t U64;
+ typedef int64_t S64;
+#else
+# include <limits.h>
+#if CHAR_BIT != 8
+# error "this implementation requires char to be exactly 8-bit type"
+#endif
+ typedef unsigned char BYTE;
+ typedef unsigned char U8;
+ typedef signed char S8;
+#if USHRT_MAX != 65535
+# error "this implementation requires short to be exactly 16-bit type"
+#endif
+ typedef unsigned short U16;
+ typedef signed short S16;
+#if UINT_MAX != 4294967295
+# error "this implementation requires int to be exactly 32-bit type"
+#endif
+ typedef unsigned int U32;
+ typedef signed int S32;
+/* note : there are no limits defined for long long type in C90.
+ * limits exist in C99, however, in such case, <stdint.h> is preferred */
+ typedef unsigned long long U64;
+ typedef signed long long S64;
+#endif
+
+
+/*-**************************************************************
+* Memory I/O API
+*****************************************************************/
+/*=== Static platform detection ===*/
+MEM_STATIC unsigned MEM_32bits(void);
+MEM_STATIC unsigned MEM_64bits(void);
+MEM_STATIC unsigned MEM_isLittleEndian(void);
+
+/*=== Native unaligned read/write ===*/
+MEM_STATIC U16 MEM_read16(const void* memPtr);
+MEM_STATIC U32 MEM_read32(const void* memPtr);
+MEM_STATIC U64 MEM_read64(const void* memPtr);
+MEM_STATIC size_t MEM_readST(const void* memPtr);
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value);
+MEM_STATIC void MEM_write32(void* memPtr, U32 value);
+MEM_STATIC void MEM_write64(void* memPtr, U64 value);
+
+/*=== Little endian unaligned read/write ===*/
+MEM_STATIC U16 MEM_readLE16(const void* memPtr);
+MEM_STATIC U32 MEM_readLE24(const void* memPtr);
+MEM_STATIC U32 MEM_readLE32(const void* memPtr);
+MEM_STATIC U64 MEM_readLE64(const void* memPtr);
+MEM_STATIC size_t MEM_readLEST(const void* memPtr);
+
+MEM_STATIC void MEM_writeLE16(void* memPtr, U16 val);
+MEM_STATIC void MEM_writeLE24(void* memPtr, U32 val);
+MEM_STATIC void MEM_writeLE32(void* memPtr, U32 val32);
+MEM_STATIC void MEM_writeLE64(void* memPtr, U64 val64);
+MEM_STATIC void MEM_writeLEST(void* memPtr, size_t val);
+
+/*=== Big endian unaligned read/write ===*/
+MEM_STATIC U32 MEM_readBE32(const void* memPtr);
+MEM_STATIC U64 MEM_readBE64(const void* memPtr);
+MEM_STATIC size_t MEM_readBEST(const void* memPtr);
+
+MEM_STATIC void MEM_writeBE32(void* memPtr, U32 val32);
+MEM_STATIC void MEM_writeBE64(void* memPtr, U64 val64);
+MEM_STATIC void MEM_writeBEST(void* memPtr, size_t val);
+
+/*=== Byteswap ===*/
+MEM_STATIC U32 MEM_swap32(U32 in);
+MEM_STATIC U64 MEM_swap64(U64 in);
+MEM_STATIC size_t MEM_swapST(size_t in);
+
+
+/*-**************************************************************
+* Memory I/O Implementation
+*****************************************************************/
+/* MEM_FORCE_MEMORY_ACCESS : For accessing unaligned memory:
+ * Method 0 : always use `memcpy()`. Safe and portable.
+ * Method 1 : Use compiler extension to set unaligned access.
+ * Method 2 : direct access. This method is portable but violate C standard.
+ * It can generate buggy code on targets depending on alignment.
+ * Default : method 1 if supported, else method 0
+ */
+#ifndef MEM_FORCE_MEMORY_ACCESS /* can be defined externally, on command line for example */
+# ifdef __GNUC__
+# define MEM_FORCE_MEMORY_ACCESS 1
+# endif
+#endif
+
+MEM_STATIC unsigned MEM_32bits(void) { return sizeof(size_t)==4; }
+MEM_STATIC unsigned MEM_64bits(void) { return sizeof(size_t)==8; }
+
+MEM_STATIC unsigned MEM_isLittleEndian(void)
+{
+#if defined(__BYTE_ORDER__) && defined(__ORDER_LITTLE_ENDIAN__) && (__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
+ return 1;
+#elif defined(__BYTE_ORDER__) && defined(__ORDER_BIG_ENDIAN__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+ return 0;
+#elif defined(__clang__) && __LITTLE_ENDIAN__
+ return 1;
+#elif defined(__clang__) && __BIG_ENDIAN__
+ return 0;
+#elif defined(_MSC_VER) && (_M_AMD64 || _M_IX86)
+ return 1;
+#elif defined(__DMC__) && defined(_M_IX86)
+ return 1;
+#else
+ const union { U32 u; BYTE c[4]; } one = { 1 }; /* don't use static : performance detrimental */
+ return one.c[0];
+#endif
+}
+
+#if defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==2)
+
+/* violates C standard, by lying on structure alignment.
+Only use if no other choice to achieve best performance on target platform */
+MEM_STATIC U16 MEM_read16(const void* memPtr) { return *(const U16*) memPtr; }
+MEM_STATIC U32 MEM_read32(const void* memPtr) { return *(const U32*) memPtr; }
+MEM_STATIC U64 MEM_read64(const void* memPtr) { return *(const U64*) memPtr; }
+MEM_STATIC size_t MEM_readST(const void* memPtr) { return *(const size_t*) memPtr; }
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value) { *(U16*)memPtr = value; }
+MEM_STATIC void MEM_write32(void* memPtr, U32 value) { *(U32*)memPtr = value; }
+MEM_STATIC void MEM_write64(void* memPtr, U64 value) { *(U64*)memPtr = value; }
+
+#elif defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==1)
+
+typedef __attribute__((aligned(1))) U16 unalign16;
+typedef __attribute__((aligned(1))) U32 unalign32;
+typedef __attribute__((aligned(1))) U64 unalign64;
+typedef __attribute__((aligned(1))) size_t unalignArch;
+
+MEM_STATIC U16 MEM_read16(const void* ptr) { return *(const unalign16*)ptr; }
+MEM_STATIC U32 MEM_read32(const void* ptr) { return *(const unalign32*)ptr; }
+MEM_STATIC U64 MEM_read64(const void* ptr) { return *(const unalign64*)ptr; }
+MEM_STATIC size_t MEM_readST(const void* ptr) { return *(const unalignArch*)ptr; }
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value) { *(unalign16*)memPtr = value; }
+MEM_STATIC void MEM_write32(void* memPtr, U32 value) { *(unalign32*)memPtr = value; }
+MEM_STATIC void MEM_write64(void* memPtr, U64 value) { *(unalign64*)memPtr = value; }
+
+#else
+
+/* default method, safe and standard.
+ can sometimes prove slower */
+
+MEM_STATIC U16 MEM_read16(const void* memPtr)
+{
+ U16 val; ZSTD_memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC U32 MEM_read32(const void* memPtr)
+{
+ U32 val; ZSTD_memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC U64 MEM_read64(const void* memPtr)
+{
+ U64 val; ZSTD_memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC size_t MEM_readST(const void* memPtr)
+{
+ size_t val; ZSTD_memcpy(&val, memPtr, sizeof(val)); return val;
+}
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value)
+{
+ ZSTD_memcpy(memPtr, &value, sizeof(value));
+}
+
+MEM_STATIC void MEM_write32(void* memPtr, U32 value)
+{
+ ZSTD_memcpy(memPtr, &value, sizeof(value));
+}
+
+MEM_STATIC void MEM_write64(void* memPtr, U64 value)
+{
+ ZSTD_memcpy(memPtr, &value, sizeof(value));
+}
+
+#endif /* MEM_FORCE_MEMORY_ACCESS */
+
+MEM_STATIC U32 MEM_swap32_fallback(U32 in)
+{
+ return ((in << 24) & 0xff000000 ) |
+ ((in << 8) & 0x00ff0000 ) |
+ ((in >> 8) & 0x0000ff00 ) |
+ ((in >> 24) & 0x000000ff );
+}
+
+MEM_STATIC U32 MEM_swap32(U32 in)
+{
+#if defined(_MSC_VER) /* Visual Studio */
+ return _byteswap_ulong(in);
+#elif (defined (__GNUC__) && (__GNUC__ * 100 + __GNUC_MINOR__ >= 403)) \
+ || (defined(__clang__) && __has_builtin(__builtin_bswap32))
+ return __builtin_bswap32(in);
+#else
+ return MEM_swap32_fallback(in);
+#endif
+}
+
+MEM_STATIC U64 MEM_swap64_fallback(U64 in)
+{
+ return ((in << 56) & 0xff00000000000000ULL) |
+ ((in << 40) & 0x00ff000000000000ULL) |
+ ((in << 24) & 0x0000ff0000000000ULL) |
+ ((in << 8) & 0x000000ff00000000ULL) |
+ ((in >> 8) & 0x00000000ff000000ULL) |
+ ((in >> 24) & 0x0000000000ff0000ULL) |
+ ((in >> 40) & 0x000000000000ff00ULL) |
+ ((in >> 56) & 0x00000000000000ffULL);
+}
+
+MEM_STATIC U64 MEM_swap64(U64 in)
+{
+#if defined(_MSC_VER) /* Visual Studio */
+ return _byteswap_uint64(in);
+#elif (defined (__GNUC__) && (__GNUC__ * 100 + __GNUC_MINOR__ >= 403)) \
+ || (defined(__clang__) && __has_builtin(__builtin_bswap64))
+ return __builtin_bswap64(in);
+#else
+ return MEM_swap64_fallback(in);
+#endif
+}
+
+MEM_STATIC size_t MEM_swapST(size_t in)
+{
+ if (MEM_32bits())
+ return (size_t)MEM_swap32((U32)in);
+ else
+ return (size_t)MEM_swap64((U64)in);
+}
+
+/*=== Little endian r/w ===*/
+
+MEM_STATIC U16 MEM_readLE16(const void* memPtr)
+{
+ if (MEM_isLittleEndian())
+ return MEM_read16(memPtr);
+ else {
+ const BYTE* p = (const BYTE*)memPtr;
+ return (U16)(p[0] + (p[1]<<8));
+ }
+}
+
+MEM_STATIC void MEM_writeLE16(void* memPtr, U16 val)
+{
+ if (MEM_isLittleEndian()) {
+ MEM_write16(memPtr, val);
+ } else {
+ BYTE* p = (BYTE*)memPtr;
+ p[0] = (BYTE)val;
+ p[1] = (BYTE)(val>>8);
+ }
+}
+
+MEM_STATIC U32 MEM_readLE24(const void* memPtr)
+{
+ return (U32)MEM_readLE16(memPtr) + ((U32)(((const BYTE*)memPtr)[2]) << 16);
+}
+
+MEM_STATIC void MEM_writeLE24(void* memPtr, U32 val)
+{
+ MEM_writeLE16(memPtr, (U16)val);
+ ((BYTE*)memPtr)[2] = (BYTE)(val>>16);
+}
+
+MEM_STATIC U32 MEM_readLE32(const void* memPtr)
+{
+ if (MEM_isLittleEndian())
+ return MEM_read32(memPtr);
+ else
+ return MEM_swap32(MEM_read32(memPtr));
+}
+
+MEM_STATIC void MEM_writeLE32(void* memPtr, U32 val32)
+{
+ if (MEM_isLittleEndian())
+ MEM_write32(memPtr, val32);
+ else
+ MEM_write32(memPtr, MEM_swap32(val32));
+}
+
+MEM_STATIC U64 MEM_readLE64(const void* memPtr)
+{
+ if (MEM_isLittleEndian())
+ return MEM_read64(memPtr);
+ else
+ return MEM_swap64(MEM_read64(memPtr));
+}
+
+MEM_STATIC void MEM_writeLE64(void* memPtr, U64 val64)
+{
+ if (MEM_isLittleEndian())
+ MEM_write64(memPtr, val64);
+ else
+ MEM_write64(memPtr, MEM_swap64(val64));
+}
+
+MEM_STATIC size_t MEM_readLEST(const void* memPtr)
+{
+ if (MEM_32bits())
+ return (size_t)MEM_readLE32(memPtr);
+ else
+ return (size_t)MEM_readLE64(memPtr);
+}
+
+MEM_STATIC void MEM_writeLEST(void* memPtr, size_t val)
+{
+ if (MEM_32bits())
+ MEM_writeLE32(memPtr, (U32)val);
+ else
+ MEM_writeLE64(memPtr, (U64)val);
+}
+
+/*=== Big endian r/w ===*/
+
+MEM_STATIC U32 MEM_readBE32(const void* memPtr)
+{
+ if (MEM_isLittleEndian())
+ return MEM_swap32(MEM_read32(memPtr));
+ else
+ return MEM_read32(memPtr);
+}
+
+MEM_STATIC void MEM_writeBE32(void* memPtr, U32 val32)
+{
+ if (MEM_isLittleEndian())
+ MEM_write32(memPtr, MEM_swap32(val32));
+ else
+ MEM_write32(memPtr, val32);
+}
+
+MEM_STATIC U64 MEM_readBE64(const void* memPtr)
+{
+ if (MEM_isLittleEndian())
+ return MEM_swap64(MEM_read64(memPtr));
+ else
+ return MEM_read64(memPtr);
+}
+
+MEM_STATIC void MEM_writeBE64(void* memPtr, U64 val64)
+{
+ if (MEM_isLittleEndian())
+ MEM_write64(memPtr, MEM_swap64(val64));
+ else
+ MEM_write64(memPtr, val64);
+}
+
+MEM_STATIC size_t MEM_readBEST(const void* memPtr)
+{
+ if (MEM_32bits())
+ return (size_t)MEM_readBE32(memPtr);
+ else
+ return (size_t)MEM_readBE64(memPtr);
+}
+
+MEM_STATIC void MEM_writeBEST(void* memPtr, size_t val)
+{
+ if (MEM_32bits())
+ MEM_writeBE32(memPtr, (U32)val);
+ else
+ MEM_writeBE64(memPtr, (U64)val);
+}
+
+/* code only tested on 32 and 64 bits systems */
+MEM_STATIC void MEM_check(void) { DEBUG_STATIC_ASSERT((sizeof(size_t)==4) || (sizeof(size_t)==8)); }
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* MEM_H_MODULE */
diff --git a/third_party/zstd/lib/common/pool.c b/third_party/zstd/lib/common/pool.c
new file mode 100644
index 0000000000..3adcefc9a5
--- /dev/null
+++ b/third_party/zstd/lib/common/pool.c
@@ -0,0 +1,371 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+
+/* ====== Dependencies ======= */
+#include "../common/allocations.h" /* ZSTD_customCalloc, ZSTD_customFree */
+#include "zstd_deps.h" /* size_t */
+#include "debug.h" /* assert */
+#include "pool.h"
+
+/* ====== Compiler specifics ====== */
+#if defined(_MSC_VER)
+# pragma warning(disable : 4204) /* disable: C4204: non-constant aggregate initializer */
+#endif
+
+
+#ifdef ZSTD_MULTITHREAD
+
+#include "threading.h" /* pthread adaptation */
+
+/* A job is a function and an opaque argument */
+typedef struct POOL_job_s {
+ POOL_function function;
+ void *opaque;
+} POOL_job;
+
+struct POOL_ctx_s {
+ ZSTD_customMem customMem;
+ /* Keep track of the threads */
+ ZSTD_pthread_t* threads;
+ size_t threadCapacity;
+ size_t threadLimit;
+
+ /* The queue is a circular buffer */
+ POOL_job *queue;
+ size_t queueHead;
+ size_t queueTail;
+ size_t queueSize;
+
+ /* The number of threads working on jobs */
+ size_t numThreadsBusy;
+ /* Indicates if the queue is empty */
+ int queueEmpty;
+
+ /* The mutex protects the queue */
+ ZSTD_pthread_mutex_t queueMutex;
+ /* Condition variable for pushers to wait on when the queue is full */
+ ZSTD_pthread_cond_t queuePushCond;
+ /* Condition variables for poppers to wait on when the queue is empty */
+ ZSTD_pthread_cond_t queuePopCond;
+ /* Indicates if the queue is shutting down */
+ int shutdown;
+};
+
+/* POOL_thread() :
+ * Work thread for the thread pool.
+ * Waits for jobs and executes them.
+ * @returns : NULL on failure else non-null.
+ */
+static void* POOL_thread(void* opaque) {
+ POOL_ctx* const ctx = (POOL_ctx*)opaque;
+ if (!ctx) { return NULL; }
+ for (;;) {
+ /* Lock the mutex and wait for a non-empty queue or until shutdown */
+ ZSTD_pthread_mutex_lock(&ctx->queueMutex);
+
+ while ( ctx->queueEmpty
+ || (ctx->numThreadsBusy >= ctx->threadLimit) ) {
+ if (ctx->shutdown) {
+ /* even if !queueEmpty, (possible if numThreadsBusy >= threadLimit),
+ * a few threads will be shutdown while !queueEmpty,
+ * but enough threads will remain active to finish the queue */
+ ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
+ return opaque;
+ }
+ ZSTD_pthread_cond_wait(&ctx->queuePopCond, &ctx->queueMutex);
+ }
+ /* Pop a job off the queue */
+ { POOL_job const job = ctx->queue[ctx->queueHead];
+ ctx->queueHead = (ctx->queueHead + 1) % ctx->queueSize;
+ ctx->numThreadsBusy++;
+ ctx->queueEmpty = (ctx->queueHead == ctx->queueTail);
+ /* Unlock the mutex, signal a pusher, and run the job */
+ ZSTD_pthread_cond_signal(&ctx->queuePushCond);
+ ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
+
+ job.function(job.opaque);
+
+ /* If the intended queue size was 0, signal after finishing job */
+ ZSTD_pthread_mutex_lock(&ctx->queueMutex);
+ ctx->numThreadsBusy--;
+ ZSTD_pthread_cond_signal(&ctx->queuePushCond);
+ ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
+ }
+ } /* for (;;) */
+ assert(0); /* Unreachable */
+}
+
+/* ZSTD_createThreadPool() : public access point */
+POOL_ctx* ZSTD_createThreadPool(size_t numThreads) {
+ return POOL_create (numThreads, 0);
+}
+
+POOL_ctx* POOL_create(size_t numThreads, size_t queueSize) {
+ return POOL_create_advanced(numThreads, queueSize, ZSTD_defaultCMem);
+}
+
+POOL_ctx* POOL_create_advanced(size_t numThreads, size_t queueSize,
+ ZSTD_customMem customMem)
+{
+ POOL_ctx* ctx;
+ /* Check parameters */
+ if (!numThreads) { return NULL; }
+ /* Allocate the context and zero initialize */
+ ctx = (POOL_ctx*)ZSTD_customCalloc(sizeof(POOL_ctx), customMem);
+ if (!ctx) { return NULL; }
+ /* Initialize the job queue.
+ * It needs one extra space since one space is wasted to differentiate
+ * empty and full queues.
+ */
+ ctx->queueSize = queueSize + 1;
+ ctx->queue = (POOL_job*)ZSTD_customCalloc(ctx->queueSize * sizeof(POOL_job), customMem);
+ ctx->queueHead = 0;
+ ctx->queueTail = 0;
+ ctx->numThreadsBusy = 0;
+ ctx->queueEmpty = 1;
+ {
+ int error = 0;
+ error |= ZSTD_pthread_mutex_init(&ctx->queueMutex, NULL);
+ error |= ZSTD_pthread_cond_init(&ctx->queuePushCond, NULL);
+ error |= ZSTD_pthread_cond_init(&ctx->queuePopCond, NULL);
+ if (error) { POOL_free(ctx); return NULL; }
+ }
+ ctx->shutdown = 0;
+ /* Allocate space for the thread handles */
+ ctx->threads = (ZSTD_pthread_t*)ZSTD_customCalloc(numThreads * sizeof(ZSTD_pthread_t), customMem);
+ ctx->threadCapacity = 0;
+ ctx->customMem = customMem;
+ /* Check for errors */
+ if (!ctx->threads || !ctx->queue) { POOL_free(ctx); return NULL; }
+ /* Initialize the threads */
+ { size_t i;
+ for (i = 0; i < numThreads; ++i) {
+ if (ZSTD_pthread_create(&ctx->threads[i], NULL, &POOL_thread, ctx)) {
+ ctx->threadCapacity = i;
+ POOL_free(ctx);
+ return NULL;
+ } }
+ ctx->threadCapacity = numThreads;
+ ctx->threadLimit = numThreads;
+ }
+ return ctx;
+}
+
+/*! POOL_join() :
+ Shutdown the queue, wake any sleeping threads, and join all of the threads.
+*/
+static void POOL_join(POOL_ctx* ctx) {
+ /* Shut down the queue */
+ ZSTD_pthread_mutex_lock(&ctx->queueMutex);
+ ctx->shutdown = 1;
+ ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
+ /* Wake up sleeping threads */
+ ZSTD_pthread_cond_broadcast(&ctx->queuePushCond);
+ ZSTD_pthread_cond_broadcast(&ctx->queuePopCond);
+ /* Join all of the threads */
+ { size_t i;
+ for (i = 0; i < ctx->threadCapacity; ++i) {
+ ZSTD_pthread_join(ctx->threads[i]); /* note : could fail */
+ } }
+}
+
+void POOL_free(POOL_ctx *ctx) {
+ if (!ctx) { return; }
+ POOL_join(ctx);
+ ZSTD_pthread_mutex_destroy(&ctx->queueMutex);
+ ZSTD_pthread_cond_destroy(&ctx->queuePushCond);
+ ZSTD_pthread_cond_destroy(&ctx->queuePopCond);
+ ZSTD_customFree(ctx->queue, ctx->customMem);
+ ZSTD_customFree(ctx->threads, ctx->customMem);
+ ZSTD_customFree(ctx, ctx->customMem);
+}
+
+/*! POOL_joinJobs() :
+ * Waits for all queued jobs to finish executing.
+ */
+void POOL_joinJobs(POOL_ctx* ctx) {
+ ZSTD_pthread_mutex_lock(&ctx->queueMutex);
+ while(!ctx->queueEmpty || ctx->numThreadsBusy > 0) {
+ ZSTD_pthread_cond_wait(&ctx->queuePushCond, &ctx->queueMutex);
+ }
+ ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
+}
+
+void ZSTD_freeThreadPool (ZSTD_threadPool* pool) {
+ POOL_free (pool);
+}
+
+size_t POOL_sizeof(const POOL_ctx* ctx) {
+ if (ctx==NULL) return 0; /* supports sizeof NULL */
+ return sizeof(*ctx)
+ + ctx->queueSize * sizeof(POOL_job)
+ + ctx->threadCapacity * sizeof(ZSTD_pthread_t);
+}
+
+
+/* @return : 0 on success, 1 on error */
+static int POOL_resize_internal(POOL_ctx* ctx, size_t numThreads)
+{
+ if (numThreads <= ctx->threadCapacity) {
+ if (!numThreads) return 1;
+ ctx->threadLimit = numThreads;
+ return 0;
+ }
+ /* numThreads > threadCapacity */
+ { ZSTD_pthread_t* const threadPool = (ZSTD_pthread_t*)ZSTD_customCalloc(numThreads * sizeof(ZSTD_pthread_t), ctx->customMem);
+ if (!threadPool) return 1;
+ /* replace existing thread pool */
+ ZSTD_memcpy(threadPool, ctx->threads, ctx->threadCapacity * sizeof(ZSTD_pthread_t));
+ ZSTD_customFree(ctx->threads, ctx->customMem);
+ ctx->threads = threadPool;
+ /* Initialize additional threads */
+ { size_t threadId;
+ for (threadId = ctx->threadCapacity; threadId < numThreads; ++threadId) {
+ if (ZSTD_pthread_create(&threadPool[threadId], NULL, &POOL_thread, ctx)) {
+ ctx->threadCapacity = threadId;
+ return 1;
+ } }
+ } }
+ /* successfully expanded */
+ ctx->threadCapacity = numThreads;
+ ctx->threadLimit = numThreads;
+ return 0;
+}
+
+/* @return : 0 on success, 1 on error */
+int POOL_resize(POOL_ctx* ctx, size_t numThreads)
+{
+ int result;
+ if (ctx==NULL) return 1;
+ ZSTD_pthread_mutex_lock(&ctx->queueMutex);
+ result = POOL_resize_internal(ctx, numThreads);
+ ZSTD_pthread_cond_broadcast(&ctx->queuePopCond);
+ ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
+ return result;
+}
+
+/**
+ * Returns 1 if the queue is full and 0 otherwise.
+ *
+ * When queueSize is 1 (pool was created with an intended queueSize of 0),
+ * then a queue is empty if there is a thread free _and_ no job is waiting.
+ */
+static int isQueueFull(POOL_ctx const* ctx) {
+ if (ctx->queueSize > 1) {
+ return ctx->queueHead == ((ctx->queueTail + 1) % ctx->queueSize);
+ } else {
+ return (ctx->numThreadsBusy == ctx->threadLimit) ||
+ !ctx->queueEmpty;
+ }
+}
+
+
+static void
+POOL_add_internal(POOL_ctx* ctx, POOL_function function, void *opaque)
+{
+ POOL_job job;
+ job.function = function;
+ job.opaque = opaque;
+ assert(ctx != NULL);
+ if (ctx->shutdown) return;
+
+ ctx->queueEmpty = 0;
+ ctx->queue[ctx->queueTail] = job;
+ ctx->queueTail = (ctx->queueTail + 1) % ctx->queueSize;
+ ZSTD_pthread_cond_signal(&ctx->queuePopCond);
+}
+
+void POOL_add(POOL_ctx* ctx, POOL_function function, void* opaque)
+{
+ assert(ctx != NULL);
+ ZSTD_pthread_mutex_lock(&ctx->queueMutex);
+ /* Wait until there is space in the queue for the new job */
+ while (isQueueFull(ctx) && (!ctx->shutdown)) {
+ ZSTD_pthread_cond_wait(&ctx->queuePushCond, &ctx->queueMutex);
+ }
+ POOL_add_internal(ctx, function, opaque);
+ ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
+}
+
+
+int POOL_tryAdd(POOL_ctx* ctx, POOL_function function, void* opaque)
+{
+ assert(ctx != NULL);
+ ZSTD_pthread_mutex_lock(&ctx->queueMutex);
+ if (isQueueFull(ctx)) {
+ ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
+ return 0;
+ }
+ POOL_add_internal(ctx, function, opaque);
+ ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
+ return 1;
+}
+
+
+#else /* ZSTD_MULTITHREAD not defined */
+
+/* ========================== */
+/* No multi-threading support */
+/* ========================== */
+
+
+/* We don't need any data, but if it is empty, malloc() might return NULL. */
+struct POOL_ctx_s {
+ int dummy;
+};
+static POOL_ctx g_poolCtx;
+
+POOL_ctx* POOL_create(size_t numThreads, size_t queueSize) {
+ return POOL_create_advanced(numThreads, queueSize, ZSTD_defaultCMem);
+}
+
+POOL_ctx*
+POOL_create_advanced(size_t numThreads, size_t queueSize, ZSTD_customMem customMem)
+{
+ (void)numThreads;
+ (void)queueSize;
+ (void)customMem;
+ return &g_poolCtx;
+}
+
+void POOL_free(POOL_ctx* ctx) {
+ assert(!ctx || ctx == &g_poolCtx);
+ (void)ctx;
+}
+
+void POOL_joinJobs(POOL_ctx* ctx){
+ assert(!ctx || ctx == &g_poolCtx);
+ (void)ctx;
+}
+
+int POOL_resize(POOL_ctx* ctx, size_t numThreads) {
+ (void)ctx; (void)numThreads;
+ return 0;
+}
+
+void POOL_add(POOL_ctx* ctx, POOL_function function, void* opaque) {
+ (void)ctx;
+ function(opaque);
+}
+
+int POOL_tryAdd(POOL_ctx* ctx, POOL_function function, void* opaque) {
+ (void)ctx;
+ function(opaque);
+ return 1;
+}
+
+size_t POOL_sizeof(const POOL_ctx* ctx) {
+ if (ctx==NULL) return 0; /* supports sizeof NULL */
+ assert(ctx == &g_poolCtx);
+ return sizeof(*ctx);
+}
+
+#endif /* ZSTD_MULTITHREAD */
diff --git a/third_party/zstd/lib/common/pool.h b/third_party/zstd/lib/common/pool.h
new file mode 100644
index 0000000000..cca4de73a8
--- /dev/null
+++ b/third_party/zstd/lib/common/pool.h
@@ -0,0 +1,90 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef POOL_H
+#define POOL_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+
+#include "zstd_deps.h"
+#define ZSTD_STATIC_LINKING_ONLY /* ZSTD_customMem */
+#include "../zstd.h"
+
+typedef struct POOL_ctx_s POOL_ctx;
+
+/*! POOL_create() :
+ * Create a thread pool with at most `numThreads` threads.
+ * `numThreads` must be at least 1.
+ * The maximum number of queued jobs before blocking is `queueSize`.
+ * @return : POOL_ctx pointer on success, else NULL.
+*/
+POOL_ctx* POOL_create(size_t numThreads, size_t queueSize);
+
+POOL_ctx* POOL_create_advanced(size_t numThreads, size_t queueSize,
+ ZSTD_customMem customMem);
+
+/*! POOL_free() :
+ * Free a thread pool returned by POOL_create().
+ */
+void POOL_free(POOL_ctx* ctx);
+
+
+/*! POOL_joinJobs() :
+ * Waits for all queued jobs to finish executing.
+ */
+void POOL_joinJobs(POOL_ctx* ctx);
+
+/*! POOL_resize() :
+ * Expands or shrinks pool's number of threads.
+ * This is more efficient than releasing + creating a new context,
+ * since it tries to preserve and reuse existing threads.
+ * `numThreads` must be at least 1.
+ * @return : 0 when resize was successful,
+ * !0 (typically 1) if there is an error.
+ * note : only numThreads can be resized, queueSize remains unchanged.
+ */
+int POOL_resize(POOL_ctx* ctx, size_t numThreads);
+
+/*! POOL_sizeof() :
+ * @return threadpool memory usage
+ * note : compatible with NULL (returns 0 in this case)
+ */
+size_t POOL_sizeof(const POOL_ctx* ctx);
+
+/*! POOL_function :
+ * The function type that can be added to a thread pool.
+ */
+typedef void (*POOL_function)(void*);
+
+/*! POOL_add() :
+ * Add the job `function(opaque)` to the thread pool. `ctx` must be valid.
+ * Possibly blocks until there is room in the queue.
+ * Note : The function may be executed asynchronously,
+ * therefore, `opaque` must live until function has been completed.
+ */
+void POOL_add(POOL_ctx* ctx, POOL_function function, void* opaque);
+
+
+/*! POOL_tryAdd() :
+ * Add the job `function(opaque)` to thread pool _if_ a queue slot is available.
+ * Returns immediately even if not (does not block).
+ * @return : 1 if successful, 0 if not.
+ */
+int POOL_tryAdd(POOL_ctx* ctx, POOL_function function, void* opaque);
+
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif
diff --git a/third_party/zstd/lib/common/portability_macros.h b/third_party/zstd/lib/common/portability_macros.h
new file mode 100644
index 0000000000..e50314a78e
--- /dev/null
+++ b/third_party/zstd/lib/common/portability_macros.h
@@ -0,0 +1,158 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_PORTABILITY_MACROS_H
+#define ZSTD_PORTABILITY_MACROS_H
+
+/**
+ * This header file contains macro definitions to support portability.
+ * This header is shared between C and ASM code, so it MUST only
+ * contain macro definitions. It MUST not contain any C code.
+ *
+ * This header ONLY defines macros to detect platforms/feature support.
+ *
+ */
+
+
+/* compat. with non-clang compilers */
+#ifndef __has_attribute
+ #define __has_attribute(x) 0
+#endif
+
+/* compat. with non-clang compilers */
+#ifndef __has_builtin
+# define __has_builtin(x) 0
+#endif
+
+/* compat. with non-clang compilers */
+#ifndef __has_feature
+# define __has_feature(x) 0
+#endif
+
+/* detects whether we are being compiled under msan */
+#ifndef ZSTD_MEMORY_SANITIZER
+# if __has_feature(memory_sanitizer)
+# define ZSTD_MEMORY_SANITIZER 1
+# else
+# define ZSTD_MEMORY_SANITIZER 0
+# endif
+#endif
+
+/* detects whether we are being compiled under asan */
+#ifndef ZSTD_ADDRESS_SANITIZER
+# if __has_feature(address_sanitizer)
+# define ZSTD_ADDRESS_SANITIZER 1
+# elif defined(__SANITIZE_ADDRESS__)
+# define ZSTD_ADDRESS_SANITIZER 1
+# else
+# define ZSTD_ADDRESS_SANITIZER 0
+# endif
+#endif
+
+/* detects whether we are being compiled under dfsan */
+#ifndef ZSTD_DATAFLOW_SANITIZER
+# if __has_feature(dataflow_sanitizer)
+# define ZSTD_DATAFLOW_SANITIZER 1
+# else
+# define ZSTD_DATAFLOW_SANITIZER 0
+# endif
+#endif
+
+/* Mark the internal assembly functions as hidden */
+#ifdef __ELF__
+# define ZSTD_HIDE_ASM_FUNCTION(func) .hidden func
+#elif defined(__APPLE__)
+# define ZSTD_HIDE_ASM_FUNCTION(func) .private_extern func
+#else
+# define ZSTD_HIDE_ASM_FUNCTION(func)
+#endif
+
+/* Enable runtime BMI2 dispatch based on the CPU.
+ * Enabled for clang & gcc >=4.8 on x86 when BMI2 isn't enabled by default.
+ */
+#ifndef DYNAMIC_BMI2
+ #if ((defined(__clang__) && __has_attribute(__target__)) \
+ || (defined(__GNUC__) \
+ && (__GNUC__ >= 5 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8)))) \
+ && (defined(__x86_64__) || defined(_M_X64)) \
+ && !defined(__BMI2__)
+ # define DYNAMIC_BMI2 1
+ #else
+ # define DYNAMIC_BMI2 0
+ #endif
+#endif
+
+/**
+ * Only enable assembly for GNUC compatible compilers,
+ * because other platforms may not support GAS assembly syntax.
+ *
+ * Only enable assembly for Linux / MacOS, other platforms may
+ * work, but they haven't been tested. This could likely be
+ * extended to BSD systems.
+ *
+ * Disable assembly when MSAN is enabled, because MSAN requires
+ * 100% of code to be instrumented to work.
+ */
+#if defined(__GNUC__)
+# if defined(__linux__) || defined(__linux) || defined(__APPLE__)
+# if ZSTD_MEMORY_SANITIZER
+# define ZSTD_ASM_SUPPORTED 0
+# elif ZSTD_DATAFLOW_SANITIZER
+# define ZSTD_ASM_SUPPORTED 0
+# else
+# define ZSTD_ASM_SUPPORTED 1
+# endif
+# else
+# define ZSTD_ASM_SUPPORTED 0
+# endif
+#else
+# define ZSTD_ASM_SUPPORTED 0
+#endif
+
+/**
+ * Determines whether we should enable assembly for x86-64
+ * with BMI2.
+ *
+ * Enable if all of the following conditions hold:
+ * - ASM hasn't been explicitly disabled by defining ZSTD_DISABLE_ASM
+ * - Assembly is supported
+ * - We are compiling for x86-64 and either:
+ * - DYNAMIC_BMI2 is enabled
+ * - BMI2 is supported at compile time
+ */
+#if !defined(ZSTD_DISABLE_ASM) && \
+ ZSTD_ASM_SUPPORTED && \
+ defined(__x86_64__) && \
+ (DYNAMIC_BMI2 || defined(__BMI2__))
+# define ZSTD_ENABLE_ASM_X86_64_BMI2 1
+#else
+# define ZSTD_ENABLE_ASM_X86_64_BMI2 0
+#endif
+
+/*
+ * For x86 ELF targets, add .note.gnu.property section for Intel CET in
+ * assembly sources when CET is enabled.
+ *
+ * Additionally, any function that may be called indirectly must begin
+ * with ZSTD_CET_ENDBRANCH.
+ */
+#if defined(__ELF__) && (defined(__x86_64__) || defined(__i386__)) \
+ && defined(__has_include)
+# if __has_include(<cet.h>)
+# include <cet.h>
+# define ZSTD_CET_ENDBRANCH _CET_ENDBR
+# endif
+#endif
+
+#ifndef ZSTD_CET_ENDBRANCH
+# define ZSTD_CET_ENDBRANCH
+#endif
+
+#endif /* ZSTD_PORTABILITY_MACROS_H */
diff --git a/third_party/zstd/lib/common/threading.c b/third_party/zstd/lib/common/threading.c
new file mode 100644
index 0000000000..25bb8b9810
--- /dev/null
+++ b/third_party/zstd/lib/common/threading.c
@@ -0,0 +1,182 @@
+/**
+ * Copyright (c) 2016 Tino Reichardt
+ * All rights reserved.
+ *
+ * You can contact the author at:
+ * - zstdmt source repository: https://github.com/mcmilk/zstdmt
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+/**
+ * This file will hold wrapper for systems, which do not support pthreads
+ */
+
+#include "threading.h"
+
+/* create fake symbol to avoid empty translation unit warning */
+int g_ZSTD_threading_useless_symbol;
+
+#if defined(ZSTD_MULTITHREAD) && defined(_WIN32)
+
+/**
+ * Windows minimalist Pthread Wrapper
+ */
+
+
+/* === Dependencies === */
+#include <process.h>
+#include <errno.h>
+
+
+/* === Implementation === */
+
+typedef struct {
+ void* (*start_routine)(void*);
+ void* arg;
+ int initialized;
+ ZSTD_pthread_cond_t initialized_cond;
+ ZSTD_pthread_mutex_t initialized_mutex;
+} ZSTD_thread_params_t;
+
+static unsigned __stdcall worker(void *arg)
+{
+ void* (*start_routine)(void*);
+ void* thread_arg;
+
+ /* Initialized thread_arg and start_routine and signal main thread that we don't need it
+ * to wait any longer.
+ */
+ {
+ ZSTD_thread_params_t* thread_param = (ZSTD_thread_params_t*)arg;
+ thread_arg = thread_param->arg;
+ start_routine = thread_param->start_routine;
+
+ /* Signal main thread that we are running and do not depend on its memory anymore */
+ ZSTD_pthread_mutex_lock(&thread_param->initialized_mutex);
+ thread_param->initialized = 1;
+ ZSTD_pthread_cond_signal(&thread_param->initialized_cond);
+ ZSTD_pthread_mutex_unlock(&thread_param->initialized_mutex);
+ }
+
+ start_routine(thread_arg);
+
+ return 0;
+}
+
+int ZSTD_pthread_create(ZSTD_pthread_t* thread, const void* unused,
+ void* (*start_routine) (void*), void* arg)
+{
+ ZSTD_thread_params_t thread_param;
+ (void)unused;
+
+ if (thread==NULL) return -1;
+ *thread = NULL;
+
+ thread_param.start_routine = start_routine;
+ thread_param.arg = arg;
+ thread_param.initialized = 0;
+
+ /* Setup thread initialization synchronization */
+ if(ZSTD_pthread_cond_init(&thread_param.initialized_cond, NULL)) {
+ /* Should never happen on Windows */
+ return -1;
+ }
+ if(ZSTD_pthread_mutex_init(&thread_param.initialized_mutex, NULL)) {
+ /* Should never happen on Windows */
+ ZSTD_pthread_cond_destroy(&thread_param.initialized_cond);
+ return -1;
+ }
+
+ /* Spawn thread */
+ *thread = (HANDLE)_beginthreadex(NULL, 0, worker, &thread_param, 0, NULL);
+ if (*thread==NULL) {
+ ZSTD_pthread_mutex_destroy(&thread_param.initialized_mutex);
+ ZSTD_pthread_cond_destroy(&thread_param.initialized_cond);
+ return errno;
+ }
+
+ /* Wait for thread to be initialized */
+ ZSTD_pthread_mutex_lock(&thread_param.initialized_mutex);
+ while(!thread_param.initialized) {
+ ZSTD_pthread_cond_wait(&thread_param.initialized_cond, &thread_param.initialized_mutex);
+ }
+ ZSTD_pthread_mutex_unlock(&thread_param.initialized_mutex);
+ ZSTD_pthread_mutex_destroy(&thread_param.initialized_mutex);
+ ZSTD_pthread_cond_destroy(&thread_param.initialized_cond);
+
+ return 0;
+}
+
+int ZSTD_pthread_join(ZSTD_pthread_t thread)
+{
+ DWORD result;
+
+ if (!thread) return 0;
+
+ result = WaitForSingleObject(thread, INFINITE);
+ CloseHandle(thread);
+
+ switch (result) {
+ case WAIT_OBJECT_0:
+ return 0;
+ case WAIT_ABANDONED:
+ return EINVAL;
+ default:
+ return GetLastError();
+ }
+}
+
+#endif /* ZSTD_MULTITHREAD */
+
+#if defined(ZSTD_MULTITHREAD) && DEBUGLEVEL >= 1 && !defined(_WIN32)
+
+#define ZSTD_DEPS_NEED_MALLOC
+#include "zstd_deps.h"
+
+int ZSTD_pthread_mutex_init(ZSTD_pthread_mutex_t* mutex, pthread_mutexattr_t const* attr)
+{
+ assert(mutex != NULL);
+ *mutex = (pthread_mutex_t*)ZSTD_malloc(sizeof(pthread_mutex_t));
+ if (!*mutex)
+ return 1;
+ return pthread_mutex_init(*mutex, attr);
+}
+
+int ZSTD_pthread_mutex_destroy(ZSTD_pthread_mutex_t* mutex)
+{
+ assert(mutex != NULL);
+ if (!*mutex)
+ return 0;
+ {
+ int const ret = pthread_mutex_destroy(*mutex);
+ ZSTD_free(*mutex);
+ return ret;
+ }
+}
+
+int ZSTD_pthread_cond_init(ZSTD_pthread_cond_t* cond, pthread_condattr_t const* attr)
+{
+ assert(cond != NULL);
+ *cond = (pthread_cond_t*)ZSTD_malloc(sizeof(pthread_cond_t));
+ if (!*cond)
+ return 1;
+ return pthread_cond_init(*cond, attr);
+}
+
+int ZSTD_pthread_cond_destroy(ZSTD_pthread_cond_t* cond)
+{
+ assert(cond != NULL);
+ if (!*cond)
+ return 0;
+ {
+ int const ret = pthread_cond_destroy(*cond);
+ ZSTD_free(*cond);
+ return ret;
+ }
+}
+
+#endif
diff --git a/third_party/zstd/lib/common/threading.h b/third_party/zstd/lib/common/threading.h
new file mode 100644
index 0000000000..fb5c1c8787
--- /dev/null
+++ b/third_party/zstd/lib/common/threading.h
@@ -0,0 +1,150 @@
+/**
+ * Copyright (c) 2016 Tino Reichardt
+ * All rights reserved.
+ *
+ * You can contact the author at:
+ * - zstdmt source repository: https://github.com/mcmilk/zstdmt
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef THREADING_H_938743
+#define THREADING_H_938743
+
+#include "debug.h"
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+#if defined(ZSTD_MULTITHREAD) && defined(_WIN32)
+
+/**
+ * Windows minimalist Pthread Wrapper
+ */
+#ifdef WINVER
+# undef WINVER
+#endif
+#define WINVER 0x0600
+
+#ifdef _WIN32_WINNT
+# undef _WIN32_WINNT
+#endif
+#define _WIN32_WINNT 0x0600
+
+#ifndef WIN32_LEAN_AND_MEAN
+# define WIN32_LEAN_AND_MEAN
+#endif
+
+#undef ERROR /* reported already defined on VS 2015 (Rich Geldreich) */
+#include <windows.h>
+#undef ERROR
+#define ERROR(name) ZSTD_ERROR(name)
+
+
+/* mutex */
+#define ZSTD_pthread_mutex_t CRITICAL_SECTION
+#define ZSTD_pthread_mutex_init(a, b) ((void)(b), InitializeCriticalSection((a)), 0)
+#define ZSTD_pthread_mutex_destroy(a) DeleteCriticalSection((a))
+#define ZSTD_pthread_mutex_lock(a) EnterCriticalSection((a))
+#define ZSTD_pthread_mutex_unlock(a) LeaveCriticalSection((a))
+
+/* condition variable */
+#define ZSTD_pthread_cond_t CONDITION_VARIABLE
+#define ZSTD_pthread_cond_init(a, b) ((void)(b), InitializeConditionVariable((a)), 0)
+#define ZSTD_pthread_cond_destroy(a) ((void)(a))
+#define ZSTD_pthread_cond_wait(a, b) SleepConditionVariableCS((a), (b), INFINITE)
+#define ZSTD_pthread_cond_signal(a) WakeConditionVariable((a))
+#define ZSTD_pthread_cond_broadcast(a) WakeAllConditionVariable((a))
+
+/* ZSTD_pthread_create() and ZSTD_pthread_join() */
+typedef HANDLE ZSTD_pthread_t;
+
+int ZSTD_pthread_create(ZSTD_pthread_t* thread, const void* unused,
+ void* (*start_routine) (void*), void* arg);
+
+int ZSTD_pthread_join(ZSTD_pthread_t thread);
+
+/**
+ * add here more wrappers as required
+ */
+
+
+#elif defined(ZSTD_MULTITHREAD) /* posix assumed ; need a better detection method */
+/* === POSIX Systems === */
+# include <pthread.h>
+
+#if DEBUGLEVEL < 1
+
+#define ZSTD_pthread_mutex_t pthread_mutex_t
+#define ZSTD_pthread_mutex_init(a, b) pthread_mutex_init((a), (b))
+#define ZSTD_pthread_mutex_destroy(a) pthread_mutex_destroy((a))
+#define ZSTD_pthread_mutex_lock(a) pthread_mutex_lock((a))
+#define ZSTD_pthread_mutex_unlock(a) pthread_mutex_unlock((a))
+
+#define ZSTD_pthread_cond_t pthread_cond_t
+#define ZSTD_pthread_cond_init(a, b) pthread_cond_init((a), (b))
+#define ZSTD_pthread_cond_destroy(a) pthread_cond_destroy((a))
+#define ZSTD_pthread_cond_wait(a, b) pthread_cond_wait((a), (b))
+#define ZSTD_pthread_cond_signal(a) pthread_cond_signal((a))
+#define ZSTD_pthread_cond_broadcast(a) pthread_cond_broadcast((a))
+
+#define ZSTD_pthread_t pthread_t
+#define ZSTD_pthread_create(a, b, c, d) pthread_create((a), (b), (c), (d))
+#define ZSTD_pthread_join(a) pthread_join((a),NULL)
+
+#else /* DEBUGLEVEL >= 1 */
+
+/* Debug implementation of threading.
+ * In this implementation we use pointers for mutexes and condition variables.
+ * This way, if we forget to init/destroy them the program will crash or ASAN
+ * will report leaks.
+ */
+
+#define ZSTD_pthread_mutex_t pthread_mutex_t*
+int ZSTD_pthread_mutex_init(ZSTD_pthread_mutex_t* mutex, pthread_mutexattr_t const* attr);
+int ZSTD_pthread_mutex_destroy(ZSTD_pthread_mutex_t* mutex);
+#define ZSTD_pthread_mutex_lock(a) pthread_mutex_lock(*(a))
+#define ZSTD_pthread_mutex_unlock(a) pthread_mutex_unlock(*(a))
+
+#define ZSTD_pthread_cond_t pthread_cond_t*
+int ZSTD_pthread_cond_init(ZSTD_pthread_cond_t* cond, pthread_condattr_t const* attr);
+int ZSTD_pthread_cond_destroy(ZSTD_pthread_cond_t* cond);
+#define ZSTD_pthread_cond_wait(a, b) pthread_cond_wait(*(a), *(b))
+#define ZSTD_pthread_cond_signal(a) pthread_cond_signal(*(a))
+#define ZSTD_pthread_cond_broadcast(a) pthread_cond_broadcast(*(a))
+
+#define ZSTD_pthread_t pthread_t
+#define ZSTD_pthread_create(a, b, c, d) pthread_create((a), (b), (c), (d))
+#define ZSTD_pthread_join(a) pthread_join((a),NULL)
+
+#endif
+
+#else /* ZSTD_MULTITHREAD not defined */
+/* No multithreading support */
+
+typedef int ZSTD_pthread_mutex_t;
+#define ZSTD_pthread_mutex_init(a, b) ((void)(a), (void)(b), 0)
+#define ZSTD_pthread_mutex_destroy(a) ((void)(a))
+#define ZSTD_pthread_mutex_lock(a) ((void)(a))
+#define ZSTD_pthread_mutex_unlock(a) ((void)(a))
+
+typedef int ZSTD_pthread_cond_t;
+#define ZSTD_pthread_cond_init(a, b) ((void)(a), (void)(b), 0)
+#define ZSTD_pthread_cond_destroy(a) ((void)(a))
+#define ZSTD_pthread_cond_wait(a, b) ((void)(a), (void)(b))
+#define ZSTD_pthread_cond_signal(a) ((void)(a))
+#define ZSTD_pthread_cond_broadcast(a) ((void)(a))
+
+/* do not use ZSTD_pthread_t */
+
+#endif /* ZSTD_MULTITHREAD */
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* THREADING_H_938743 */
diff --git a/third_party/zstd/lib/common/xxhash.c b/third_party/zstd/lib/common/xxhash.c
new file mode 100644
index 0000000000..052cd52282
--- /dev/null
+++ b/third_party/zstd/lib/common/xxhash.c
@@ -0,0 +1,18 @@
+/*
+ * xxHash - Extremely Fast Hash algorithm
+ * Copyright (c) Yann Collet - Meta Platforms, Inc
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+/*
+ * xxhash.c instantiates functions defined in xxhash.h
+ */
+
+#define XXH_STATIC_LINKING_ONLY /* access advanced declarations */
+#define XXH_IMPLEMENTATION /* access definitions */
+
+#include "xxhash.h"
diff --git a/third_party/zstd/lib/common/xxhash.h b/third_party/zstd/lib/common/xxhash.h
new file mode 100644
index 0000000000..e59e44267c
--- /dev/null
+++ b/third_party/zstd/lib/common/xxhash.h
@@ -0,0 +1,7020 @@
+/*
+ * xxHash - Extremely Fast Hash algorithm
+ * Header File
+ * Copyright (c) Yann Collet - Meta Platforms, Inc
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+/* Local adaptations for Zstandard */
+
+#ifndef XXH_NO_XXH3
+# define XXH_NO_XXH3
+#endif
+
+#ifndef XXH_NAMESPACE
+# define XXH_NAMESPACE ZSTD_
+#endif
+
+/*!
+ * @mainpage xxHash
+ *
+ * xxHash is an extremely fast non-cryptographic hash algorithm, working at RAM speed
+ * limits.
+ *
+ * It is proposed in four flavors, in three families:
+ * 1. @ref XXH32_family
+ * - Classic 32-bit hash function. Simple, compact, and runs on almost all
+ * 32-bit and 64-bit systems.
+ * 2. @ref XXH64_family
+ * - Classic 64-bit adaptation of XXH32. Just as simple, and runs well on most
+ * 64-bit systems (but _not_ 32-bit systems).
+ * 3. @ref XXH3_family
+ * - Modern 64-bit and 128-bit hash function family which features improved
+ * strength and performance across the board, especially on smaller data.
+ * It benefits greatly from SIMD and 64-bit without requiring it.
+ *
+ * Benchmarks
+ * ---
+ * The reference system uses an Intel i7-9700K CPU, and runs Ubuntu x64 20.04.
+ * The open source benchmark program is compiled with clang v10.0 using -O3 flag.
+ *
+ * | Hash Name | ISA ext | Width | Large Data Speed | Small Data Velocity |
+ * | -------------------- | ------- | ----: | ---------------: | ------------------: |
+ * | XXH3_64bits() | @b AVX2 | 64 | 59.4 GB/s | 133.1 |
+ * | MeowHash | AES-NI | 128 | 58.2 GB/s | 52.5 |
+ * | XXH3_128bits() | @b AVX2 | 128 | 57.9 GB/s | 118.1 |
+ * | CLHash | PCLMUL | 64 | 37.1 GB/s | 58.1 |
+ * | XXH3_64bits() | @b SSE2 | 64 | 31.5 GB/s | 133.1 |
+ * | XXH3_128bits() | @b SSE2 | 128 | 29.6 GB/s | 118.1 |
+ * | RAM sequential read | | N/A | 28.0 GB/s | N/A |
+ * | ahash | AES-NI | 64 | 22.5 GB/s | 107.2 |
+ * | City64 | | 64 | 22.0 GB/s | 76.6 |
+ * | T1ha2 | | 64 | 22.0 GB/s | 99.0 |
+ * | City128 | | 128 | 21.7 GB/s | 57.7 |
+ * | FarmHash | AES-NI | 64 | 21.3 GB/s | 71.9 |
+ * | XXH64() | | 64 | 19.4 GB/s | 71.0 |
+ * | SpookyHash | | 64 | 19.3 GB/s | 53.2 |
+ * | Mum | | 64 | 18.0 GB/s | 67.0 |
+ * | CRC32C | SSE4.2 | 32 | 13.0 GB/s | 57.9 |
+ * | XXH32() | | 32 | 9.7 GB/s | 71.9 |
+ * | City32 | | 32 | 9.1 GB/s | 66.0 |
+ * | Blake3* | @b AVX2 | 256 | 4.4 GB/s | 8.1 |
+ * | Murmur3 | | 32 | 3.9 GB/s | 56.1 |
+ * | SipHash* | | 64 | 3.0 GB/s | 43.2 |
+ * | Blake3* | @b SSE2 | 256 | 2.4 GB/s | 8.1 |
+ * | HighwayHash | | 64 | 1.4 GB/s | 6.0 |
+ * | FNV64 | | 64 | 1.2 GB/s | 62.7 |
+ * | Blake2* | | 256 | 1.1 GB/s | 5.1 |
+ * | SHA1* | | 160 | 0.8 GB/s | 5.6 |
+ * | MD5* | | 128 | 0.6 GB/s | 7.8 |
+ * @note
+ * - Hashes which require a specific ISA extension are noted. SSE2 is also noted,
+ * even though it is mandatory on x64.
+ * - Hashes with an asterisk are cryptographic. Note that MD5 is non-cryptographic
+ * by modern standards.
+ * - Small data velocity is a rough average of algorithm's efficiency for small
+ * data. For more accurate information, see the wiki.
+ * - More benchmarks and strength tests are found on the wiki:
+ * https://github.com/Cyan4973/xxHash/wiki
+ *
+ * Usage
+ * ------
+ * All xxHash variants use a similar API. Changing the algorithm is a trivial
+ * substitution.
+ *
+ * @pre
+ * For functions which take an input and length parameter, the following
+ * requirements are assumed:
+ * - The range from [`input`, `input + length`) is valid, readable memory.
+ * - The only exception is if the `length` is `0`, `input` may be `NULL`.
+ * - For C++, the objects must have the *TriviallyCopyable* property, as the
+ * functions access bytes directly as if it was an array of `unsigned char`.
+ *
+ * @anchor single_shot_example
+ * **Single Shot**
+ *
+ * These functions are stateless functions which hash a contiguous block of memory,
+ * immediately returning the result. They are the easiest and usually the fastest
+ * option.
+ *
+ * XXH32(), XXH64(), XXH3_64bits(), XXH3_128bits()
+ *
+ * @code{.c}
+ * #include <string.h>
+ * #include "xxhash.h"
+ *
+ * // Example for a function which hashes a null terminated string with XXH32().
+ * XXH32_hash_t hash_string(const char* string, XXH32_hash_t seed)
+ * {
+ * // NULL pointers are only valid if the length is zero
+ * size_t length = (string == NULL) ? 0 : strlen(string);
+ * return XXH32(string, length, seed);
+ * }
+ * @endcode
+ *
+ *
+ * @anchor streaming_example
+ * **Streaming**
+ *
+ * These groups of functions allow incremental hashing of unknown size, even
+ * more than what would fit in a size_t.
+ *
+ * XXH32_reset(), XXH64_reset(), XXH3_64bits_reset(), XXH3_128bits_reset()
+ *
+ * @code{.c}
+ * #include <stdio.h>
+ * #include <assert.h>
+ * #include "xxhash.h"
+ * // Example for a function which hashes a FILE incrementally with XXH3_64bits().
+ * XXH64_hash_t hashFile(FILE* f)
+ * {
+ * // Allocate a state struct. Do not just use malloc() or new.
+ * XXH3_state_t* state = XXH3_createState();
+ * assert(state != NULL && "Out of memory!");
+ * // Reset the state to start a new hashing session.
+ * XXH3_64bits_reset(state);
+ * char buffer[4096];
+ * size_t count;
+ * // Read the file in chunks
+ * while ((count = fread(buffer, 1, sizeof(buffer), f)) != 0) {
+ * // Run update() as many times as necessary to process the data
+ * XXH3_64bits_update(state, buffer, count);
+ * }
+ * // Retrieve the finalized hash. This will not change the state.
+ * XXH64_hash_t result = XXH3_64bits_digest(state);
+ * // Free the state. Do not use free().
+ * XXH3_freeState(state);
+ * return result;
+ * }
+ * @endcode
+ *
+ * Streaming functions generate the xxHash value from an incremental input.
+ * This method is slower than single-call functions, due to state management.
+ * For small inputs, prefer `XXH32()` and `XXH64()`, which are better optimized.
+ *
+ * An XXH state must first be allocated using `XXH*_createState()`.
+ *
+ * Start a new hash by initializing the state with a seed using `XXH*_reset()`.
+ *
+ * Then, feed the hash state by calling `XXH*_update()` as many times as necessary.
+ *
+ * The function returns an error code, with 0 meaning OK, and any other value
+ * meaning there is an error.
+ *
+ * Finally, a hash value can be produced anytime, by using `XXH*_digest()`.
+ * This function returns the nn-bits hash as an int or long long.
+ *
+ * It's still possible to continue inserting input into the hash state after a
+ * digest, and generate new hash values later on by invoking `XXH*_digest()`.
+ *
+ * When done, release the state using `XXH*_freeState()`.
+ *
+ *
+ * @anchor canonical_representation_example
+ * **Canonical Representation**
+ *
+ * The default return values from XXH functions are unsigned 32, 64 and 128 bit
+ * integers.
+ * This the simplest and fastest format for further post-processing.
+ *
+ * However, this leaves open the question of what is the order on the byte level,
+ * since little and big endian conventions will store the same number differently.
+ *
+ * The canonical representation settles this issue by mandating big-endian
+ * convention, the same convention as human-readable numbers (large digits first).
+ *
+ * When writing hash values to storage, sending them over a network, or printing
+ * them, it's highly recommended to use the canonical representation to ensure
+ * portability across a wider range of systems, present and future.
+ *
+ * The following functions allow transformation of hash values to and from
+ * canonical format.
+ *
+ * XXH32_canonicalFromHash(), XXH32_hashFromCanonical(),
+ * XXH64_canonicalFromHash(), XXH64_hashFromCanonical(),
+ * XXH128_canonicalFromHash(), XXH128_hashFromCanonical(),
+ *
+ * @code{.c}
+ * #include <stdio.h>
+ * #include "xxhash.h"
+ *
+ * // Example for a function which prints XXH32_hash_t in human readable format
+ * void printXxh32(XXH32_hash_t hash)
+ * {
+ * XXH32_canonical_t cano;
+ * XXH32_canonicalFromHash(&cano, hash);
+ * size_t i;
+ * for(i = 0; i < sizeof(cano.digest); ++i) {
+ * printf("%02x", cano.digest[i]);
+ * }
+ * printf("\n");
+ * }
+ *
+ * // Example for a function which converts XXH32_canonical_t to XXH32_hash_t
+ * XXH32_hash_t convertCanonicalToXxh32(XXH32_canonical_t cano)
+ * {
+ * XXH32_hash_t hash = XXH32_hashFromCanonical(&cano);
+ * return hash;
+ * }
+ * @endcode
+ *
+ *
+ * @file xxhash.h
+ * xxHash prototypes and implementation
+ */
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/* ****************************
+ * INLINE mode
+ ******************************/
+/*!
+ * @defgroup public Public API
+ * Contains details on the public xxHash functions.
+ * @{
+ */
+#ifdef XXH_DOXYGEN
+/*!
+ * @brief Gives access to internal state declaration, required for static allocation.
+ *
+ * Incompatible with dynamic linking, due to risks of ABI changes.
+ *
+ * Usage:
+ * @code{.c}
+ * #define XXH_STATIC_LINKING_ONLY
+ * #include "xxhash.h"
+ * @endcode
+ */
+# define XXH_STATIC_LINKING_ONLY
+/* Do not undef XXH_STATIC_LINKING_ONLY for Doxygen */
+
+/*!
+ * @brief Gives access to internal definitions.
+ *
+ * Usage:
+ * @code{.c}
+ * #define XXH_STATIC_LINKING_ONLY
+ * #define XXH_IMPLEMENTATION
+ * #include "xxhash.h"
+ * @endcode
+ */
+# define XXH_IMPLEMENTATION
+/* Do not undef XXH_IMPLEMENTATION for Doxygen */
+
+/*!
+ * @brief Exposes the implementation and marks all functions as `inline`.
+ *
+ * Use these build macros to inline xxhash into the target unit.
+ * Inlining improves performance on small inputs, especially when the length is
+ * expressed as a compile-time constant:
+ *
+ * https://fastcompression.blogspot.com/2018/03/xxhash-for-small-keys-impressive-power.html
+ *
+ * It also keeps xxHash symbols private to the unit, so they are not exported.
+ *
+ * Usage:
+ * @code{.c}
+ * #define XXH_INLINE_ALL
+ * #include "xxhash.h"
+ * @endcode
+ * Do not compile and link xxhash.o as a separate object, as it is not useful.
+ */
+# define XXH_INLINE_ALL
+# undef XXH_INLINE_ALL
+/*!
+ * @brief Exposes the implementation without marking functions as inline.
+ */
+# define XXH_PRIVATE_API
+# undef XXH_PRIVATE_API
+/*!
+ * @brief Emulate a namespace by transparently prefixing all symbols.
+ *
+ * If you want to include _and expose_ xxHash functions from within your own
+ * library, but also want to avoid symbol collisions with other libraries which
+ * may also include xxHash, you can use @ref XXH_NAMESPACE to automatically prefix
+ * any public symbol from xxhash library with the value of @ref XXH_NAMESPACE
+ * (therefore, avoid empty or numeric values).
+ *
+ * Note that no change is required within the calling program as long as it
+ * includes `xxhash.h`: Regular symbol names will be automatically translated
+ * by this header.
+ */
+# define XXH_NAMESPACE /* YOUR NAME HERE */
+# undef XXH_NAMESPACE
+#endif
+
+#if (defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)) \
+ && !defined(XXH_INLINE_ALL_31684351384)
+ /* this section should be traversed only once */
+# define XXH_INLINE_ALL_31684351384
+ /* give access to the advanced API, required to compile implementations */
+# undef XXH_STATIC_LINKING_ONLY /* avoid macro redef */
+# define XXH_STATIC_LINKING_ONLY
+ /* make all functions private */
+# undef XXH_PUBLIC_API
+# if defined(__GNUC__)
+# define XXH_PUBLIC_API static __inline __attribute__((unused))
+# elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+# define XXH_PUBLIC_API static inline
+# elif defined(_MSC_VER)
+# define XXH_PUBLIC_API static __inline
+# else
+ /* note: this version may generate warnings for unused static functions */
+# define XXH_PUBLIC_API static
+# endif
+
+ /*
+ * This part deals with the special case where a unit wants to inline xxHash,
+ * but "xxhash.h" has previously been included without XXH_INLINE_ALL,
+ * such as part of some previously included *.h header file.
+ * Without further action, the new include would just be ignored,
+ * and functions would effectively _not_ be inlined (silent failure).
+ * The following macros solve this situation by prefixing all inlined names,
+ * avoiding naming collision with previous inclusions.
+ */
+ /* Before that, we unconditionally #undef all symbols,
+ * in case they were already defined with XXH_NAMESPACE.
+ * They will then be redefined for XXH_INLINE_ALL
+ */
+# undef XXH_versionNumber
+ /* XXH32 */
+# undef XXH32
+# undef XXH32_createState
+# undef XXH32_freeState
+# undef XXH32_reset
+# undef XXH32_update
+# undef XXH32_digest
+# undef XXH32_copyState
+# undef XXH32_canonicalFromHash
+# undef XXH32_hashFromCanonical
+ /* XXH64 */
+# undef XXH64
+# undef XXH64_createState
+# undef XXH64_freeState
+# undef XXH64_reset
+# undef XXH64_update
+# undef XXH64_digest
+# undef XXH64_copyState
+# undef XXH64_canonicalFromHash
+# undef XXH64_hashFromCanonical
+ /* XXH3_64bits */
+# undef XXH3_64bits
+# undef XXH3_64bits_withSecret
+# undef XXH3_64bits_withSeed
+# undef XXH3_64bits_withSecretandSeed
+# undef XXH3_createState
+# undef XXH3_freeState
+# undef XXH3_copyState
+# undef XXH3_64bits_reset
+# undef XXH3_64bits_reset_withSeed
+# undef XXH3_64bits_reset_withSecret
+# undef XXH3_64bits_update
+# undef XXH3_64bits_digest
+# undef XXH3_generateSecret
+ /* XXH3_128bits */
+# undef XXH128
+# undef XXH3_128bits
+# undef XXH3_128bits_withSeed
+# undef XXH3_128bits_withSecret
+# undef XXH3_128bits_reset
+# undef XXH3_128bits_reset_withSeed
+# undef XXH3_128bits_reset_withSecret
+# undef XXH3_128bits_reset_withSecretandSeed
+# undef XXH3_128bits_update
+# undef XXH3_128bits_digest
+# undef XXH128_isEqual
+# undef XXH128_cmp
+# undef XXH128_canonicalFromHash
+# undef XXH128_hashFromCanonical
+ /* Finally, free the namespace itself */
+# undef XXH_NAMESPACE
+
+ /* employ the namespace for XXH_INLINE_ALL */
+# define XXH_NAMESPACE XXH_INLINE_
+ /*
+ * Some identifiers (enums, type names) are not symbols,
+ * but they must nonetheless be renamed to avoid redeclaration.
+ * Alternative solution: do not redeclare them.
+ * However, this requires some #ifdefs, and has a more dispersed impact.
+ * Meanwhile, renaming can be achieved in a single place.
+ */
+# define XXH_IPREF(Id) XXH_NAMESPACE ## Id
+# define XXH_OK XXH_IPREF(XXH_OK)
+# define XXH_ERROR XXH_IPREF(XXH_ERROR)
+# define XXH_errorcode XXH_IPREF(XXH_errorcode)
+# define XXH32_canonical_t XXH_IPREF(XXH32_canonical_t)
+# define XXH64_canonical_t XXH_IPREF(XXH64_canonical_t)
+# define XXH128_canonical_t XXH_IPREF(XXH128_canonical_t)
+# define XXH32_state_s XXH_IPREF(XXH32_state_s)
+# define XXH32_state_t XXH_IPREF(XXH32_state_t)
+# define XXH64_state_s XXH_IPREF(XXH64_state_s)
+# define XXH64_state_t XXH_IPREF(XXH64_state_t)
+# define XXH3_state_s XXH_IPREF(XXH3_state_s)
+# define XXH3_state_t XXH_IPREF(XXH3_state_t)
+# define XXH128_hash_t XXH_IPREF(XXH128_hash_t)
+ /* Ensure the header is parsed again, even if it was previously included */
+# undef XXHASH_H_5627135585666179
+# undef XXHASH_H_STATIC_13879238742
+#endif /* XXH_INLINE_ALL || XXH_PRIVATE_API */
+
+/* ****************************************************************
+ * Stable API
+ *****************************************************************/
+#ifndef XXHASH_H_5627135585666179
+#define XXHASH_H_5627135585666179 1
+
+/*! @brief Marks a global symbol. */
+#if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API)
+# if defined(WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT))
+# ifdef XXH_EXPORT
+# define XXH_PUBLIC_API __declspec(dllexport)
+# elif XXH_IMPORT
+# define XXH_PUBLIC_API __declspec(dllimport)
+# endif
+# else
+# define XXH_PUBLIC_API /* do nothing */
+# endif
+#endif
+
+#ifdef XXH_NAMESPACE
+# define XXH_CAT(A,B) A##B
+# define XXH_NAME2(A,B) XXH_CAT(A,B)
+# define XXH_versionNumber XXH_NAME2(XXH_NAMESPACE, XXH_versionNumber)
+/* XXH32 */
+# define XXH32 XXH_NAME2(XXH_NAMESPACE, XXH32)
+# define XXH32_createState XXH_NAME2(XXH_NAMESPACE, XXH32_createState)
+# define XXH32_freeState XXH_NAME2(XXH_NAMESPACE, XXH32_freeState)
+# define XXH32_reset XXH_NAME2(XXH_NAMESPACE, XXH32_reset)
+# define XXH32_update XXH_NAME2(XXH_NAMESPACE, XXH32_update)
+# define XXH32_digest XXH_NAME2(XXH_NAMESPACE, XXH32_digest)
+# define XXH32_copyState XXH_NAME2(XXH_NAMESPACE, XXH32_copyState)
+# define XXH32_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH32_canonicalFromHash)
+# define XXH32_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH32_hashFromCanonical)
+/* XXH64 */
+# define XXH64 XXH_NAME2(XXH_NAMESPACE, XXH64)
+# define XXH64_createState XXH_NAME2(XXH_NAMESPACE, XXH64_createState)
+# define XXH64_freeState XXH_NAME2(XXH_NAMESPACE, XXH64_freeState)
+# define XXH64_reset XXH_NAME2(XXH_NAMESPACE, XXH64_reset)
+# define XXH64_update XXH_NAME2(XXH_NAMESPACE, XXH64_update)
+# define XXH64_digest XXH_NAME2(XXH_NAMESPACE, XXH64_digest)
+# define XXH64_copyState XXH_NAME2(XXH_NAMESPACE, XXH64_copyState)
+# define XXH64_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH64_canonicalFromHash)
+# define XXH64_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH64_hashFromCanonical)
+/* XXH3_64bits */
+# define XXH3_64bits XXH_NAME2(XXH_NAMESPACE, XXH3_64bits)
+# define XXH3_64bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecret)
+# define XXH3_64bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSeed)
+# define XXH3_64bits_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecretandSeed)
+# define XXH3_createState XXH_NAME2(XXH_NAMESPACE, XXH3_createState)
+# define XXH3_freeState XXH_NAME2(XXH_NAMESPACE, XXH3_freeState)
+# define XXH3_copyState XXH_NAME2(XXH_NAMESPACE, XXH3_copyState)
+# define XXH3_64bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset)
+# define XXH3_64bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSeed)
+# define XXH3_64bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecret)
+# define XXH3_64bits_reset_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecretandSeed)
+# define XXH3_64bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_update)
+# define XXH3_64bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_digest)
+# define XXH3_generateSecret XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret)
+# define XXH3_generateSecret_fromSeed XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret_fromSeed)
+/* XXH3_128bits */
+# define XXH128 XXH_NAME2(XXH_NAMESPACE, XXH128)
+# define XXH3_128bits XXH_NAME2(XXH_NAMESPACE, XXH3_128bits)
+# define XXH3_128bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSeed)
+# define XXH3_128bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecret)
+# define XXH3_128bits_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecretandSeed)
+# define XXH3_128bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset)
+# define XXH3_128bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSeed)
+# define XXH3_128bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecret)
+# define XXH3_128bits_reset_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecretandSeed)
+# define XXH3_128bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_update)
+# define XXH3_128bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_digest)
+# define XXH128_isEqual XXH_NAME2(XXH_NAMESPACE, XXH128_isEqual)
+# define XXH128_cmp XXH_NAME2(XXH_NAMESPACE, XXH128_cmp)
+# define XXH128_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH128_canonicalFromHash)
+# define XXH128_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH128_hashFromCanonical)
+#endif
+
+
+/* *************************************
+* Compiler specifics
+***************************************/
+
+/* specific declaration modes for Windows */
+#if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API)
+# if defined(WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT))
+# ifdef XXH_EXPORT
+# define XXH_PUBLIC_API __declspec(dllexport)
+# elif XXH_IMPORT
+# define XXH_PUBLIC_API __declspec(dllimport)
+# endif
+# else
+# define XXH_PUBLIC_API /* do nothing */
+# endif
+#endif
+
+#if defined (__GNUC__)
+# define XXH_CONSTF __attribute__((const))
+# define XXH_PUREF __attribute__((pure))
+# define XXH_MALLOCF __attribute__((malloc))
+#else
+# define XXH_CONSTF /* disable */
+# define XXH_PUREF
+# define XXH_MALLOCF
+#endif
+
+/* *************************************
+* Version
+***************************************/
+#define XXH_VERSION_MAJOR 0
+#define XXH_VERSION_MINOR 8
+#define XXH_VERSION_RELEASE 2
+/*! @brief Version number, encoded as two digits each */
+#define XXH_VERSION_NUMBER (XXH_VERSION_MAJOR *100*100 + XXH_VERSION_MINOR *100 + XXH_VERSION_RELEASE)
+
+/*!
+ * @brief Obtains the xxHash version.
+ *
+ * This is mostly useful when xxHash is compiled as a shared library,
+ * since the returned value comes from the library, as opposed to header file.
+ *
+ * @return @ref XXH_VERSION_NUMBER of the invoked library.
+ */
+XXH_PUBLIC_API XXH_CONSTF unsigned XXH_versionNumber (void);
+
+
+/* ****************************
+* Common basic types
+******************************/
+#include <stddef.h> /* size_t */
+/*!
+ * @brief Exit code for the streaming API.
+ */
+typedef enum {
+ XXH_OK = 0, /*!< OK */
+ XXH_ERROR /*!< Error */
+} XXH_errorcode;
+
+
+/*-**********************************************************************
+* 32-bit hash
+************************************************************************/
+#if defined(XXH_DOXYGEN) /* Don't show <stdint.h> include */
+/*!
+ * @brief An unsigned 32-bit integer.
+ *
+ * Not necessarily defined to `uint32_t` but functionally equivalent.
+ */
+typedef uint32_t XXH32_hash_t;
+
+#elif !defined (__VMS) \
+ && (defined (__cplusplus) \
+ || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+# ifdef _AIX
+# include <inttypes.h>
+# else
+# include <stdint.h>
+# endif
+ typedef uint32_t XXH32_hash_t;
+
+#else
+# include <limits.h>
+# if UINT_MAX == 0xFFFFFFFFUL
+ typedef unsigned int XXH32_hash_t;
+# elif ULONG_MAX == 0xFFFFFFFFUL
+ typedef unsigned long XXH32_hash_t;
+# else
+# error "unsupported platform: need a 32-bit type"
+# endif
+#endif
+
+/*!
+ * @}
+ *
+ * @defgroup XXH32_family XXH32 family
+ * @ingroup public
+ * Contains functions used in the classic 32-bit xxHash algorithm.
+ *
+ * @note
+ * XXH32 is useful for older platforms, with no or poor 64-bit performance.
+ * Note that the @ref XXH3_family provides competitive speed for both 32-bit
+ * and 64-bit systems, and offers true 64/128 bit hash results.
+ *
+ * @see @ref XXH64_family, @ref XXH3_family : Other xxHash families
+ * @see @ref XXH32_impl for implementation details
+ * @{
+ */
+
+/*!
+ * @brief Calculates the 32-bit hash of @p input using xxHash32.
+ *
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ * @param seed The 32-bit seed to alter the hash's output predictably.
+ *
+ * @pre
+ * The memory between @p input and @p input + @p length must be valid,
+ * readable, contiguous memory. However, if @p length is `0`, @p input may be
+ * `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 32-bit xxHash32 value.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed);
+
+#ifndef XXH_NO_STREAM
+/*!
+ * @typedef struct XXH32_state_s XXH32_state_t
+ * @brief The opaque state struct for the XXH32 streaming API.
+ *
+ * @see XXH32_state_s for details.
+ */
+typedef struct XXH32_state_s XXH32_state_t;
+
+/*!
+ * @brief Allocates an @ref XXH32_state_t.
+ *
+ * @return An allocated pointer of @ref XXH32_state_t on success.
+ * @return `NULL` on failure.
+ *
+ * @note Must be freed with XXH32_freeState().
+ */
+XXH_PUBLIC_API XXH_MALLOCF XXH32_state_t* XXH32_createState(void);
+/*!
+ * @brief Frees an @ref XXH32_state_t.
+ *
+ * @param statePtr A pointer to an @ref XXH32_state_t allocated with @ref XXH32_createState().
+ *
+ * @return @ref XXH_OK.
+ *
+ * @note @p statePtr must be allocated with XXH32_createState().
+ *
+ */
+XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr);
+/*!
+ * @brief Copies one @ref XXH32_state_t to another.
+ *
+ * @param dst_state The state to copy to.
+ * @param src_state The state to copy from.
+ * @pre
+ * @p dst_state and @p src_state must not be `NULL` and must not overlap.
+ */
+XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dst_state, const XXH32_state_t* src_state);
+
+/*!
+ * @brief Resets an @ref XXH32_state_t to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed The 32-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ * @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note This function resets and seeds a state. Call it before @ref XXH32_update().
+ */
+XXH_PUBLIC_API XXH_errorcode XXH32_reset (XXH32_state_t* statePtr, XXH32_hash_t seed);
+
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH32_state_t.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ * @p statePtr must not be `NULL`.
+ * @pre
+ * The memory between @p input and @p input + @p length must be valid,
+ * readable, contiguous memory. However, if @p length is `0`, @p input may be
+ * `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note Call this to incrementally consume blocks of data.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* statePtr, const void* input, size_t length);
+
+/*!
+ * @brief Returns the calculated hash value from an @ref XXH32_state_t.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ * @p statePtr must not be `NULL`.
+ *
+ * @return The calculated 32-bit xxHash32 value from that state.
+ *
+ * @note
+ * Calling XXH32_digest() will not affect @p statePtr, so you can update,
+ * digest, and update again.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_digest (const XXH32_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
+
+/******* Canonical representation *******/
+
+/*!
+ * @brief Canonical (big endian) representation of @ref XXH32_hash_t.
+ */
+typedef struct {
+ unsigned char digest[4]; /*!< Hash bytes, big endian */
+} XXH32_canonical_t;
+
+/*!
+ * @brief Converts an @ref XXH32_hash_t to a big endian @ref XXH32_canonical_t.
+ *
+ * @param dst The @ref XXH32_canonical_t pointer to be stored to.
+ * @param hash The @ref XXH32_hash_t to be converted.
+ *
+ * @pre
+ * @p dst must not be `NULL`.
+ *
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash);
+
+/*!
+ * @brief Converts an @ref XXH32_canonical_t to a native @ref XXH32_hash_t.
+ *
+ * @param src The @ref XXH32_canonical_t to convert.
+ *
+ * @pre
+ * @p src must not be `NULL`.
+ *
+ * @return The converted hash.
+ *
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src);
+
+
+/*! @cond Doxygen ignores this part */
+#ifdef __has_attribute
+# define XXH_HAS_ATTRIBUTE(x) __has_attribute(x)
+#else
+# define XXH_HAS_ATTRIBUTE(x) 0
+#endif
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+/*
+ * C23 __STDC_VERSION__ number hasn't been specified yet. For now
+ * leave as `201711L` (C17 + 1).
+ * TODO: Update to correct value when its been specified.
+ */
+#define XXH_C23_VN 201711L
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+/* C-language Attributes are added in C23. */
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN) && defined(__has_c_attribute)
+# define XXH_HAS_C_ATTRIBUTE(x) __has_c_attribute(x)
+#else
+# define XXH_HAS_C_ATTRIBUTE(x) 0
+#endif
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+#if defined(__cplusplus) && defined(__has_cpp_attribute)
+# define XXH_HAS_CPP_ATTRIBUTE(x) __has_cpp_attribute(x)
+#else
+# define XXH_HAS_CPP_ATTRIBUTE(x) 0
+#endif
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+/*
+ * Define XXH_FALLTHROUGH macro for annotating switch case with the 'fallthrough' attribute
+ * introduced in CPP17 and C23.
+ * CPP17 : https://en.cppreference.com/w/cpp/language/attributes/fallthrough
+ * C23 : https://en.cppreference.com/w/c/language/attributes/fallthrough
+ */
+#if XXH_HAS_C_ATTRIBUTE(fallthrough) || XXH_HAS_CPP_ATTRIBUTE(fallthrough)
+# define XXH_FALLTHROUGH [[fallthrough]]
+#elif XXH_HAS_ATTRIBUTE(__fallthrough__)
+# define XXH_FALLTHROUGH __attribute__ ((__fallthrough__))
+#else
+# define XXH_FALLTHROUGH /* fallthrough */
+#endif
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+/*
+ * Define XXH_NOESCAPE for annotated pointers in public API.
+ * https://clang.llvm.org/docs/AttributeReference.html#noescape
+ * As of writing this, only supported by clang.
+ */
+#if XXH_HAS_ATTRIBUTE(noescape)
+# define XXH_NOESCAPE __attribute__((noescape))
+#else
+# define XXH_NOESCAPE
+#endif
+/*! @endcond */
+
+
+/*!
+ * @}
+ * @ingroup public
+ * @{
+ */
+
+#ifndef XXH_NO_LONG_LONG
+/*-**********************************************************************
+* 64-bit hash
+************************************************************************/
+#if defined(XXH_DOXYGEN) /* don't include <stdint.h> */
+/*!
+ * @brief An unsigned 64-bit integer.
+ *
+ * Not necessarily defined to `uint64_t` but functionally equivalent.
+ */
+typedef uint64_t XXH64_hash_t;
+#elif !defined (__VMS) \
+ && (defined (__cplusplus) \
+ || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+# ifdef _AIX
+# include <inttypes.h>
+# else
+# include <stdint.h>
+# endif
+ typedef uint64_t XXH64_hash_t;
+#else
+# include <limits.h>
+# if defined(__LP64__) && ULONG_MAX == 0xFFFFFFFFFFFFFFFFULL
+ /* LP64 ABI says uint64_t is unsigned long */
+ typedef unsigned long XXH64_hash_t;
+# else
+ /* the following type must have a width of 64-bit */
+ typedef unsigned long long XXH64_hash_t;
+# endif
+#endif
+
+/*!
+ * @}
+ *
+ * @defgroup XXH64_family XXH64 family
+ * @ingroup public
+ * @{
+ * Contains functions used in the classic 64-bit xxHash algorithm.
+ *
+ * @note
+ * XXH3 provides competitive speed for both 32-bit and 64-bit systems,
+ * and offers true 64/128 bit hash results.
+ * It provides better speed for systems with vector processing capabilities.
+ */
+
+/*!
+ * @brief Calculates the 64-bit hash of @p input using xxHash64.
+ *
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ * @param seed The 64-bit seed to alter the hash's output predictably.
+ *
+ * @pre
+ * The memory between @p input and @p input + @p length must be valid,
+ * readable, contiguous memory. However, if @p length is `0`, @p input may be
+ * `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 64-bit xxHash64 value.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed);
+
+/******* Streaming *******/
+#ifndef XXH_NO_STREAM
+/*!
+ * @brief The opaque state struct for the XXH64 streaming API.
+ *
+ * @see XXH64_state_s for details.
+ */
+typedef struct XXH64_state_s XXH64_state_t; /* incomplete type */
+
+/*!
+ * @brief Allocates an @ref XXH64_state_t.
+ *
+ * @return An allocated pointer of @ref XXH64_state_t on success.
+ * @return `NULL` on failure.
+ *
+ * @note Must be freed with XXH64_freeState().
+ */
+XXH_PUBLIC_API XXH_MALLOCF XXH64_state_t* XXH64_createState(void);
+
+/*!
+ * @brief Frees an @ref XXH64_state_t.
+ *
+ * @param statePtr A pointer to an @ref XXH64_state_t allocated with @ref XXH64_createState().
+ *
+ * @return @ref XXH_OK.
+ *
+ * @note @p statePtr must be allocated with XXH64_createState().
+ */
+XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr);
+
+/*!
+ * @brief Copies one @ref XXH64_state_t to another.
+ *
+ * @param dst_state The state to copy to.
+ * @param src_state The state to copy from.
+ * @pre
+ * @p dst_state and @p src_state must not be `NULL` and must not overlap.
+ */
+XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dst_state, const XXH64_state_t* src_state);
+
+/*!
+ * @brief Resets an @ref XXH64_state_t to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed The 64-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ * @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note This function resets and seeds a state. Call it before @ref XXH64_update().
+ */
+XXH_PUBLIC_API XXH_errorcode XXH64_reset (XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed);
+
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH64_state_t.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ * @p statePtr must not be `NULL`.
+ * @pre
+ * The memory between @p input and @p input + @p length must be valid,
+ * readable, contiguous memory. However, if @p length is `0`, @p input may be
+ * `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note Call this to incrementally consume blocks of data.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH_NOESCAPE XXH64_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
+
+/*!
+ * @brief Returns the calculated hash value from an @ref XXH64_state_t.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ * @p statePtr must not be `NULL`.
+ *
+ * @return The calculated 64-bit xxHash64 value from that state.
+ *
+ * @note
+ * Calling XXH64_digest() will not affect @p statePtr, so you can update,
+ * digest, and update again.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_digest (XXH_NOESCAPE const XXH64_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
+/******* Canonical representation *******/
+
+/*!
+ * @brief Canonical (big endian) representation of @ref XXH64_hash_t.
+ */
+typedef struct { unsigned char digest[sizeof(XXH64_hash_t)]; } XXH64_canonical_t;
+
+/*!
+ * @brief Converts an @ref XXH64_hash_t to a big endian @ref XXH64_canonical_t.
+ *
+ * @param dst The @ref XXH64_canonical_t pointer to be stored to.
+ * @param hash The @ref XXH64_hash_t to be converted.
+ *
+ * @pre
+ * @p dst must not be `NULL`.
+ *
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash);
+
+/*!
+ * @brief Converts an @ref XXH64_canonical_t to a native @ref XXH64_hash_t.
+ *
+ * @param src The @ref XXH64_canonical_t to convert.
+ *
+ * @pre
+ * @p src must not be `NULL`.
+ *
+ * @return The converted hash.
+ *
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src);
+
+#ifndef XXH_NO_XXH3
+
+/*!
+ * @}
+ * ************************************************************************
+ * @defgroup XXH3_family XXH3 family
+ * @ingroup public
+ * @{
+ *
+ * XXH3 is a more recent hash algorithm featuring:
+ * - Improved speed for both small and large inputs
+ * - True 64-bit and 128-bit outputs
+ * - SIMD acceleration
+ * - Improved 32-bit viability
+ *
+ * Speed analysis methodology is explained here:
+ *
+ * https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html
+ *
+ * Compared to XXH64, expect XXH3 to run approximately
+ * ~2x faster on large inputs and >3x faster on small ones,
+ * exact differences vary depending on platform.
+ *
+ * XXH3's speed benefits greatly from SIMD and 64-bit arithmetic,
+ * but does not require it.
+ * Most 32-bit and 64-bit targets that can run XXH32 smoothly can run XXH3
+ * at competitive speeds, even without vector support. Further details are
+ * explained in the implementation.
+ *
+ * XXH3 has a fast scalar implementation, but it also includes accelerated SIMD
+ * implementations for many common platforms:
+ * - AVX512
+ * - AVX2
+ * - SSE2
+ * - ARM NEON
+ * - WebAssembly SIMD128
+ * - POWER8 VSX
+ * - s390x ZVector
+ * This can be controlled via the @ref XXH_VECTOR macro, but it automatically
+ * selects the best version according to predefined macros. For the x86 family, an
+ * automatic runtime dispatcher is included separately in @ref xxh_x86dispatch.c.
+ *
+ * XXH3 implementation is portable:
+ * it has a generic C90 formulation that can be compiled on any platform,
+ * all implementations generate exactly the same hash value on all platforms.
+ * Starting from v0.8.0, it's also labelled "stable", meaning that
+ * any future version will also generate the same hash value.
+ *
+ * XXH3 offers 2 variants, _64bits and _128bits.
+ *
+ * When only 64 bits are needed, prefer invoking the _64bits variant, as it
+ * reduces the amount of mixing, resulting in faster speed on small inputs.
+ * It's also generally simpler to manipulate a scalar return type than a struct.
+ *
+ * The API supports one-shot hashing, streaming mode, and custom secrets.
+ */
+/*-**********************************************************************
+* XXH3 64-bit variant
+************************************************************************/
+
+/*!
+ * @brief Calculates 64-bit unseeded variant of XXH3 hash of @p input.
+ *
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ * The memory between @p input and @p input + @p length must be valid,
+ * readable, contiguous memory. However, if @p length is `0`, @p input may be
+ * `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 64-bit XXH3 hash value.
+ *
+ * @note
+ * This is equivalent to @ref XXH3_64bits_withSeed() with a seed of `0`, however
+ * it may have slightly better performance due to constant propagation of the
+ * defaults.
+ *
+ * @see
+ * XXH3_64bits_withSeed(), XXH3_64bits_withSecret(): other seeding variants
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length);
+
+/*!
+ * @brief Calculates 64-bit seeded variant of XXH3 hash of @p input.
+ *
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ * @param seed The 64-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ * The memory between @p input and @p input + @p length must be valid,
+ * readable, contiguous memory. However, if @p length is `0`, @p input may be
+ * `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 64-bit XXH3 hash value.
+ *
+ * @note
+ * seed == 0 produces the same results as @ref XXH3_64bits().
+ *
+ * This variant generates a custom secret on the fly based on default secret
+ * altered using the @p seed value.
+ *
+ * While this operation is decently fast, note that it's not completely free.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed);
+
+/*!
+ * The bare minimum size for a custom secret.
+ *
+ * @see
+ * XXH3_64bits_withSecret(), XXH3_64bits_reset_withSecret(),
+ * XXH3_128bits_withSecret(), XXH3_128bits_reset_withSecret().
+ */
+#define XXH3_SECRET_SIZE_MIN 136
+
+/*!
+ * @brief Calculates 64-bit variant of XXH3 with a custom "secret".
+ *
+ * @param data The block of data to be hashed, at least @p len bytes in size.
+ * @param len The length of @p data, in bytes.
+ * @param secret The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ *
+ * @return The calculated 64-bit XXH3 hash value.
+ *
+ * @pre
+ * The memory between @p data and @p data + @p len must be valid,
+ * readable, contiguous memory. However, if @p length is `0`, @p data may be
+ * `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * It's possible to provide any blob of bytes as a "secret" to generate the hash.
+ * This makes it more difficult for an external actor to prepare an intentional collision.
+ * The main condition is that @p secretSize *must* be large enough (>= @ref XXH3_SECRET_SIZE_MIN).
+ * However, the quality of the secret impacts the dispersion of the hash algorithm.
+ * Therefore, the secret _must_ look like a bunch of random bytes.
+ * Avoid "trivial" or structured data such as repeated sequences or a text document.
+ * Whenever in doubt about the "randomness" of the blob of bytes,
+ * consider employing @ref XXH3_generateSecret() instead (see below).
+ * It will generate a proper high entropy secret derived from the blob of bytes.
+ * Another advantage of using XXH3_generateSecret() is that
+ * it guarantees that all bits within the initial blob of bytes
+ * will impact every bit of the output.
+ * This is not necessarily the case when using the blob of bytes directly
+ * because, when hashing _small_ inputs, only a portion of the secret is employed.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize);
+
+
+/******* Streaming *******/
+#ifndef XXH_NO_STREAM
+/*
+ * Streaming requires state maintenance.
+ * This operation costs memory and CPU.
+ * As a consequence, streaming is slower than one-shot hashing.
+ * For better performance, prefer one-shot functions whenever applicable.
+ */
+
+/*!
+ * @brief The opaque state struct for the XXH3 streaming API.
+ *
+ * @see XXH3_state_s for details.
+ */
+typedef struct XXH3_state_s XXH3_state_t;
+XXH_PUBLIC_API XXH_MALLOCF XXH3_state_t* XXH3_createState(void);
+XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr);
+
+/*!
+ * @brief Copies one @ref XXH3_state_t to another.
+ *
+ * @param dst_state The state to copy to.
+ * @param src_state The state to copy from.
+ * @pre
+ * @p dst_state and @p src_state must not be `NULL` and must not overlap.
+ */
+XXH_PUBLIC_API void XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ *
+ * @pre
+ * @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ * - This function resets `statePtr` and generate a secret with default parameters.
+ * - Call this function before @ref XXH3_64bits_update().
+ * - Digest will be equivalent to `XXH3_64bits()`.
+ *
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed The 64-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ * @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ * - This function resets `statePtr` and generate a secret from `seed`.
+ * - Call this function before @ref XXH3_64bits_update().
+ * - Digest will be equivalent to `XXH3_64bits_withSeed()`.
+ *
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param secret The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ *
+ * @pre
+ * @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ * `secret` is referenced, it _must outlive_ the hash streaming session.
+ *
+ * Similar to one-shot API, `secretSize` must be >= @ref XXH3_SECRET_SIZE_MIN,
+ * and the quality of produced hash values depends on secret's entropy
+ * (secret's content should look like a bunch of random bytes).
+ * When in doubt about the randomness of a candidate `secret`,
+ * consider employing `XXH3_generateSecret()` instead (see below).
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize);
+
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH3_state_t.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ * @p statePtr must not be `NULL`.
+ * @pre
+ * The memory between @p input and @p input + @p length must be valid,
+ * readable, contiguous memory. However, if @p length is `0`, @p input may be
+ * `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note Call this to incrementally consume blocks of data.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
+
+/*!
+ * @brief Returns the calculated XXH3 64-bit hash value from an @ref XXH3_state_t.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ * @p statePtr must not be `NULL`.
+ *
+ * @return The calculated XXH3 64-bit hash value from that state.
+ *
+ * @note
+ * Calling XXH3_64bits_digest() will not affect @p statePtr, so you can update,
+ * digest, and update again.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
+
+/* note : canonical representation of XXH3 is the same as XXH64
+ * since they both produce XXH64_hash_t values */
+
+
+/*-**********************************************************************
+* XXH3 128-bit variant
+************************************************************************/
+
+/*!
+ * @brief The return value from 128-bit hashes.
+ *
+ * Stored in little endian order, although the fields themselves are in native
+ * endianness.
+ */
+typedef struct {
+ XXH64_hash_t low64; /*!< `value & 0xFFFFFFFFFFFFFFFF` */
+ XXH64_hash_t high64; /*!< `value >> 64` */
+} XXH128_hash_t;
+
+/*!
+ * @brief Calculates 128-bit unseeded variant of XXH3 of @p data.
+ *
+ * @param data The block of data to be hashed, at least @p length bytes in size.
+ * @param len The length of @p data, in bytes.
+ *
+ * @return The calculated 128-bit variant of XXH3 value.
+ *
+ * The 128-bit variant of XXH3 has more strength, but it has a bit of overhead
+ * for shorter inputs.
+ *
+ * This is equivalent to @ref XXH3_128bits_withSeed() with a seed of `0`, however
+ * it may have slightly better performance due to constant propagation of the
+ * defaults.
+ *
+ * @see XXH3_128bits_withSeed(), XXH3_128bits_withSecret(): other seeding variants
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits(XXH_NOESCAPE const void* data, size_t len);
+/*! @brief Calculates 128-bit seeded variant of XXH3 hash of @p data.
+ *
+ * @param data The block of data to be hashed, at least @p length bytes in size.
+ * @param len The length of @p data, in bytes.
+ * @param seed The 64-bit seed to alter the hash result predictably.
+ *
+ * @return The calculated 128-bit variant of XXH3 value.
+ *
+ * @note
+ * seed == 0 produces the same results as @ref XXH3_64bits().
+ *
+ * This variant generates a custom secret on the fly based on default secret
+ * altered using the @p seed value.
+ *
+ * While this operation is decently fast, note that it's not completely free.
+ *
+ * @see XXH3_128bits(), XXH3_128bits_withSecret(): other seeding variants
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSeed(XXH_NOESCAPE const void* data, size_t len, XXH64_hash_t seed);
+/*!
+ * @brief Calculates 128-bit variant of XXH3 with a custom "secret".
+ *
+ * @param data The block of data to be hashed, at least @p len bytes in size.
+ * @param len The length of @p data, in bytes.
+ * @param secret The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ *
+ * @return The calculated 128-bit variant of XXH3 value.
+ *
+ * It's possible to provide any blob of bytes as a "secret" to generate the hash.
+ * This makes it more difficult for an external actor to prepare an intentional collision.
+ * The main condition is that @p secretSize *must* be large enough (>= @ref XXH3_SECRET_SIZE_MIN).
+ * However, the quality of the secret impacts the dispersion of the hash algorithm.
+ * Therefore, the secret _must_ look like a bunch of random bytes.
+ * Avoid "trivial" or structured data such as repeated sequences or a text document.
+ * Whenever in doubt about the "randomness" of the blob of bytes,
+ * consider employing @ref XXH3_generateSecret() instead (see below).
+ * It will generate a proper high entropy secret derived from the blob of bytes.
+ * Another advantage of using XXH3_generateSecret() is that
+ * it guarantees that all bits within the initial blob of bytes
+ * will impact every bit of the output.
+ * This is not necessarily the case when using the blob of bytes directly
+ * because, when hashing _small_ inputs, only a portion of the secret is employed.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize);
+
+/******* Streaming *******/
+#ifndef XXH_NO_STREAM
+/*
+ * Streaming requires state maintenance.
+ * This operation costs memory and CPU.
+ * As a consequence, streaming is slower than one-shot hashing.
+ * For better performance, prefer one-shot functions whenever applicable.
+ *
+ * XXH3_128bits uses the same XXH3_state_t as XXH3_64bits().
+ * Use already declared XXH3_createState() and XXH3_freeState().
+ *
+ * All reset and streaming functions have same meaning as their 64-bit counterpart.
+ */
+
+/*!
+ * @brief Resets an @ref XXH3_state_t to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ *
+ * @pre
+ * @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ * - This function resets `statePtr` and generate a secret with default parameters.
+ * - Call it before @ref XXH3_128bits_update().
+ * - Digest will be equivalent to `XXH3_128bits()`.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed The 64-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ * @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ * - This function resets `statePtr` and generate a secret from `seed`.
+ * - Call it before @ref XXH3_128bits_update().
+ * - Digest will be equivalent to `XXH3_128bits_withSeed()`.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed);
+/*!
+ * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
+ *
+ * @param statePtr The state struct to reset.
+ * @param secret The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ *
+ * @pre
+ * @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * `secret` is referenced, it _must outlive_ the hash streaming session.
+ * Similar to one-shot API, `secretSize` must be >= @ref XXH3_SECRET_SIZE_MIN,
+ * and the quality of produced hash values depends on secret's entropy
+ * (secret's content should look like a bunch of random bytes).
+ * When in doubt about the randomness of a candidate `secret`,
+ * consider employing `XXH3_generateSecret()` instead (see below).
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize);
+
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH3_state_t.
+ *
+ * Call this to incrementally consume blocks of data.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ * @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @note
+ * The memory between @p input and @p input + @p length must be valid,
+ * readable, contiguous memory. However, if @p length is `0`, @p input may be
+ * `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
+
+/*!
+ * @brief Returns the calculated XXH3 128-bit hash value from an @ref XXH3_state_t.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ * @p statePtr must not be `NULL`.
+ *
+ * @return The calculated XXH3 128-bit hash value from that state.
+ *
+ * @note
+ * Calling XXH3_128bits_digest() will not affect @p statePtr, so you can update,
+ * digest, and update again.
+ *
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
+
+/* Following helper functions make it possible to compare XXH128_hast_t values.
+ * Since XXH128_hash_t is a structure, this capability is not offered by the language.
+ * Note: For better performance, these functions can be inlined using XXH_INLINE_ALL */
+
+/*!
+ * @brief Check equality of two XXH128_hash_t values
+ *
+ * @param h1 The 128-bit hash value.
+ * @param h2 Another 128-bit hash value.
+ *
+ * @return `1` if `h1` and `h2` are equal.
+ * @return `0` if they are not.
+ */
+XXH_PUBLIC_API XXH_PUREF int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2);
+
+/*!
+ * @brief Compares two @ref XXH128_hash_t
+ *
+ * This comparator is compatible with stdlib's `qsort()`/`bsearch()`.
+ *
+ * @param h128_1 Left-hand side value
+ * @param h128_2 Right-hand side value
+ *
+ * @return >0 if @p h128_1 > @p h128_2
+ * @return =0 if @p h128_1 == @p h128_2
+ * @return <0 if @p h128_1 < @p h128_2
+ */
+XXH_PUBLIC_API XXH_PUREF int XXH128_cmp(XXH_NOESCAPE const void* h128_1, XXH_NOESCAPE const void* h128_2);
+
+
+/******* Canonical representation *******/
+typedef struct { unsigned char digest[sizeof(XXH128_hash_t)]; } XXH128_canonical_t;
+
+
+/*!
+ * @brief Converts an @ref XXH128_hash_t to a big endian @ref XXH128_canonical_t.
+ *
+ * @param dst The @ref XXH128_canonical_t pointer to be stored to.
+ * @param hash The @ref XXH128_hash_t to be converted.
+ *
+ * @pre
+ * @p dst must not be `NULL`.
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash);
+
+/*!
+ * @brief Converts an @ref XXH128_canonical_t to a native @ref XXH128_hash_t.
+ *
+ * @param src The @ref XXH128_canonical_t to convert.
+ *
+ * @pre
+ * @p src must not be `NULL`.
+ *
+ * @return The converted hash.
+ * @see @ref canonical_representation_example "Canonical Representation Example"
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src);
+
+
+#endif /* !XXH_NO_XXH3 */
+#endif /* XXH_NO_LONG_LONG */
+
+/*!
+ * @}
+ */
+#endif /* XXHASH_H_5627135585666179 */
+
+
+
+#if defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742)
+#define XXHASH_H_STATIC_13879238742
+/* ****************************************************************************
+ * This section contains declarations which are not guaranteed to remain stable.
+ * They may change in future versions, becoming incompatible with a different
+ * version of the library.
+ * These declarations should only be used with static linking.
+ * Never use them in association with dynamic linking!
+ ***************************************************************************** */
+
+/*
+ * These definitions are only present to allow static allocation
+ * of XXH states, on stack or in a struct, for example.
+ * Never **ever** access their members directly.
+ */
+
+/*!
+ * @internal
+ * @brief Structure for XXH32 streaming API.
+ *
+ * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
+ * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. Otherwise it is
+ * an opaque type. This allows fields to safely be changed.
+ *
+ * Typedef'd to @ref XXH32_state_t.
+ * Do not access the members of this struct directly.
+ * @see XXH64_state_s, XXH3_state_s
+ */
+struct XXH32_state_s {
+ XXH32_hash_t total_len_32; /*!< Total length hashed, modulo 2^32 */
+ XXH32_hash_t large_len; /*!< Whether the hash is >= 16 (handles @ref total_len_32 overflow) */
+ XXH32_hash_t v[4]; /*!< Accumulator lanes */
+ XXH32_hash_t mem32[4]; /*!< Internal buffer for partial reads. Treated as unsigned char[16]. */
+ XXH32_hash_t memsize; /*!< Amount of data in @ref mem32 */
+ XXH32_hash_t reserved; /*!< Reserved field. Do not read nor write to it. */
+}; /* typedef'd to XXH32_state_t */
+
+
+#ifndef XXH_NO_LONG_LONG /* defined when there is no 64-bit support */
+
+/*!
+ * @internal
+ * @brief Structure for XXH64 streaming API.
+ *
+ * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
+ * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. Otherwise it is
+ * an opaque type. This allows fields to safely be changed.
+ *
+ * Typedef'd to @ref XXH64_state_t.
+ * Do not access the members of this struct directly.
+ * @see XXH32_state_s, XXH3_state_s
+ */
+struct XXH64_state_s {
+ XXH64_hash_t total_len; /*!< Total length hashed. This is always 64-bit. */
+ XXH64_hash_t v[4]; /*!< Accumulator lanes */
+ XXH64_hash_t mem64[4]; /*!< Internal buffer for partial reads. Treated as unsigned char[32]. */
+ XXH32_hash_t memsize; /*!< Amount of data in @ref mem64 */
+ XXH32_hash_t reserved32; /*!< Reserved field, needed for padding anyways*/
+ XXH64_hash_t reserved64; /*!< Reserved field. Do not read or write to it. */
+}; /* typedef'd to XXH64_state_t */
+
+#ifndef XXH_NO_XXH3
+
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* >= C11 */
+# include <stdalign.h>
+# define XXH_ALIGN(n) alignas(n)
+#elif defined(__cplusplus) && (__cplusplus >= 201103L) /* >= C++11 */
+/* In C++ alignas() is a keyword */
+# define XXH_ALIGN(n) alignas(n)
+#elif defined(__GNUC__)
+# define XXH_ALIGN(n) __attribute__ ((aligned(n)))
+#elif defined(_MSC_VER)
+# define XXH_ALIGN(n) __declspec(align(n))
+#else
+# define XXH_ALIGN(n) /* disabled */
+#endif
+
+/* Old GCC versions only accept the attribute after the type in structures. */
+#if !(defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)) /* C11+ */ \
+ && ! (defined(__cplusplus) && (__cplusplus >= 201103L)) /* >= C++11 */ \
+ && defined(__GNUC__)
+# define XXH_ALIGN_MEMBER(align, type) type XXH_ALIGN(align)
+#else
+# define XXH_ALIGN_MEMBER(align, type) XXH_ALIGN(align) type
+#endif
+
+/*!
+ * @brief The size of the internal XXH3 buffer.
+ *
+ * This is the optimal update size for incremental hashing.
+ *
+ * @see XXH3_64b_update(), XXH3_128b_update().
+ */
+#define XXH3_INTERNALBUFFER_SIZE 256
+
+/*!
+ * @internal
+ * @brief Default size of the secret buffer (and @ref XXH3_kSecret).
+ *
+ * This is the size used in @ref XXH3_kSecret and the seeded functions.
+ *
+ * Not to be confused with @ref XXH3_SECRET_SIZE_MIN.
+ */
+#define XXH3_SECRET_DEFAULT_SIZE 192
+
+/*!
+ * @internal
+ * @brief Structure for XXH3 streaming API.
+ *
+ * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
+ * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined.
+ * Otherwise it is an opaque type.
+ * Never use this definition in combination with dynamic library.
+ * This allows fields to safely be changed in the future.
+ *
+ * @note ** This structure has a strict alignment requirement of 64 bytes!! **
+ * Do not allocate this with `malloc()` or `new`,
+ * it will not be sufficiently aligned.
+ * Use @ref XXH3_createState() and @ref XXH3_freeState(), or stack allocation.
+ *
+ * Typedef'd to @ref XXH3_state_t.
+ * Do never access the members of this struct directly.
+ *
+ * @see XXH3_INITSTATE() for stack initialization.
+ * @see XXH3_createState(), XXH3_freeState().
+ * @see XXH32_state_s, XXH64_state_s
+ */
+struct XXH3_state_s {
+ XXH_ALIGN_MEMBER(64, XXH64_hash_t acc[8]);
+ /*!< The 8 accumulators. See @ref XXH32_state_s::v and @ref XXH64_state_s::v */
+ XXH_ALIGN_MEMBER(64, unsigned char customSecret[XXH3_SECRET_DEFAULT_SIZE]);
+ /*!< Used to store a custom secret generated from a seed. */
+ XXH_ALIGN_MEMBER(64, unsigned char buffer[XXH3_INTERNALBUFFER_SIZE]);
+ /*!< The internal buffer. @see XXH32_state_s::mem32 */
+ XXH32_hash_t bufferedSize;
+ /*!< The amount of memory in @ref buffer, @see XXH32_state_s::memsize */
+ XXH32_hash_t useSeed;
+ /*!< Reserved field. Needed for padding on 64-bit. */
+ size_t nbStripesSoFar;
+ /*!< Number or stripes processed. */
+ XXH64_hash_t totalLen;
+ /*!< Total length hashed. 64-bit even on 32-bit targets. */
+ size_t nbStripesPerBlock;
+ /*!< Number of stripes per block. */
+ size_t secretLimit;
+ /*!< Size of @ref customSecret or @ref extSecret */
+ XXH64_hash_t seed;
+ /*!< Seed for _withSeed variants. Must be zero otherwise, @see XXH3_INITSTATE() */
+ XXH64_hash_t reserved64;
+ /*!< Reserved field. */
+ const unsigned char* extSecret;
+ /*!< Reference to an external secret for the _withSecret variants, NULL
+ * for other variants. */
+ /* note: there may be some padding at the end due to alignment on 64 bytes */
+}; /* typedef'd to XXH3_state_t */
+
+#undef XXH_ALIGN_MEMBER
+
+/*!
+ * @brief Initializes a stack-allocated `XXH3_state_s`.
+ *
+ * When the @ref XXH3_state_t structure is merely emplaced on stack,
+ * it should be initialized with XXH3_INITSTATE() or a memset()
+ * in case its first reset uses XXH3_NNbits_reset_withSeed().
+ * This init can be omitted if the first reset uses default or _withSecret mode.
+ * This operation isn't necessary when the state is created with XXH3_createState().
+ * Note that this doesn't prepare the state for a streaming operation,
+ * it's still necessary to use XXH3_NNbits_reset*() afterwards.
+ */
+#define XXH3_INITSTATE(XXH3_state_ptr) \
+ do { \
+ XXH3_state_t* tmp_xxh3_state_ptr = (XXH3_state_ptr); \
+ tmp_xxh3_state_ptr->seed = 0; \
+ tmp_xxh3_state_ptr->extSecret = NULL; \
+ } while(0)
+
+
+/*!
+ * @brief Calculates the 128-bit hash of @p data using XXH3.
+ *
+ * @param data The block of data to be hashed, at least @p len bytes in size.
+ * @param len The length of @p data, in bytes.
+ * @param seed The 64-bit seed to alter the hash's output predictably.
+ *
+ * @pre
+ * The memory between @p data and @p data + @p len must be valid,
+ * readable, contiguous memory. However, if @p len is `0`, @p data may be
+ * `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 128-bit XXH3 value.
+ *
+ * @see @ref single_shot_example "Single Shot Example" for an example.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128(XXH_NOESCAPE const void* data, size_t len, XXH64_hash_t seed);
+
+
+/* === Experimental API === */
+/* Symbols defined below must be considered tied to a specific library version. */
+
+/*!
+ * @brief Derive a high-entropy secret from any user-defined content, named customSeed.
+ *
+ * @param secretBuffer A writable buffer for derived high-entropy secret data.
+ * @param secretSize Size of secretBuffer, in bytes. Must be >= XXH3_SECRET_DEFAULT_SIZE.
+ * @param customSeed A user-defined content.
+ * @param customSeedSize Size of customSeed, in bytes.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * The generated secret can be used in combination with `*_withSecret()` functions.
+ * The `_withSecret()` variants are useful to provide a higher level of protection
+ * than 64-bit seed, as it becomes much more difficult for an external actor to
+ * guess how to impact the calculation logic.
+ *
+ * The function accepts as input a custom seed of any length and any content,
+ * and derives from it a high-entropy secret of length @p secretSize into an
+ * already allocated buffer @p secretBuffer.
+ *
+ * The generated secret can then be used with any `*_withSecret()` variant.
+ * The functions @ref XXH3_128bits_withSecret(), @ref XXH3_64bits_withSecret(),
+ * @ref XXH3_128bits_reset_withSecret() and @ref XXH3_64bits_reset_withSecret()
+ * are part of this list. They all accept a `secret` parameter
+ * which must be large enough for implementation reasons (>= @ref XXH3_SECRET_SIZE_MIN)
+ * _and_ feature very high entropy (consist of random-looking bytes).
+ * These conditions can be a high bar to meet, so @ref XXH3_generateSecret() can
+ * be employed to ensure proper quality.
+ *
+ * @p customSeed can be anything. It can have any size, even small ones,
+ * and its content can be anything, even "poor entropy" sources such as a bunch
+ * of zeroes. The resulting `secret` will nonetheless provide all required qualities.
+ *
+ * @pre
+ * - @p secretSize must be >= @ref XXH3_SECRET_SIZE_MIN
+ * - When @p customSeedSize > 0, supplying NULL as customSeed is undefined behavior.
+ *
+ * Example code:
+ * @code{.c}
+ * #include <stdio.h>
+ * #include <stdlib.h>
+ * #include <string.h>
+ * #define XXH_STATIC_LINKING_ONLY // expose unstable API
+ * #include "xxhash.h"
+ * // Hashes argv[2] using the entropy from argv[1].
+ * int main(int argc, char* argv[])
+ * {
+ * char secret[XXH3_SECRET_SIZE_MIN];
+ * if (argv != 3) { return 1; }
+ * XXH3_generateSecret(secret, sizeof(secret), argv[1], strlen(argv[1]));
+ * XXH64_hash_t h = XXH3_64bits_withSecret(
+ * argv[2], strlen(argv[2]),
+ * secret, sizeof(secret)
+ * );
+ * printf("%016llx\n", (unsigned long long) h);
+ * }
+ * @endcode
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_generateSecret(XXH_NOESCAPE void* secretBuffer, size_t secretSize, XXH_NOESCAPE const void* customSeed, size_t customSeedSize);
+
+/*!
+ * @brief Generate the same secret as the _withSeed() variants.
+ *
+ * @param secretBuffer A writable buffer of @ref XXH3_SECRET_SIZE_MIN bytes
+ * @param seed The 64-bit seed to alter the hash result predictably.
+ *
+ * The generated secret can be used in combination with
+ *`*_withSecret()` and `_withSecretandSeed()` variants.
+ *
+ * Example C++ `std::string` hash class:
+ * @code{.cpp}
+ * #include <string>
+ * #define XXH_STATIC_LINKING_ONLY // expose unstable API
+ * #include "xxhash.h"
+ * // Slow, seeds each time
+ * class HashSlow {
+ * XXH64_hash_t seed;
+ * public:
+ * HashSlow(XXH64_hash_t s) : seed{s} {}
+ * size_t operator()(const std::string& x) const {
+ * return size_t{XXH3_64bits_withSeed(x.c_str(), x.length(), seed)};
+ * }
+ * };
+ * // Fast, caches the seeded secret for future uses.
+ * class HashFast {
+ * unsigned char secret[XXH3_SECRET_SIZE_MIN];
+ * public:
+ * HashFast(XXH64_hash_t s) {
+ * XXH3_generateSecret_fromSeed(secret, seed);
+ * }
+ * size_t operator()(const std::string& x) const {
+ * return size_t{
+ * XXH3_64bits_withSecret(x.c_str(), x.length(), secret, sizeof(secret))
+ * };
+ * }
+ * };
+ * @endcode
+ */
+XXH_PUBLIC_API void XXH3_generateSecret_fromSeed(XXH_NOESCAPE void* secretBuffer, XXH64_hash_t seed);
+
+/*!
+ * @brief Calculates 64/128-bit seeded variant of XXH3 hash of @p data.
+ *
+ * @param data The block of data to be hashed, at least @p len bytes in size.
+ * @param len The length of @p data, in bytes.
+ * @param secret The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ * @param seed The 64-bit seed to alter the hash result predictably.
+ *
+ * These variants generate hash values using either
+ * @p seed for "short" keys (< @ref XXH3_MIDSIZE_MAX = 240 bytes)
+ * or @p secret for "large" keys (>= @ref XXH3_MIDSIZE_MAX).
+ *
+ * This generally benefits speed, compared to `_withSeed()` or `_withSecret()`.
+ * `_withSeed()` has to generate the secret on the fly for "large" keys.
+ * It's fast, but can be perceptible for "not so large" keys (< 1 KB).
+ * `_withSecret()` has to generate the masks on the fly for "small" keys,
+ * which requires more instructions than _withSeed() variants.
+ * Therefore, _withSecretandSeed variant combines the best of both worlds.
+ *
+ * When @p secret has been generated by XXH3_generateSecret_fromSeed(),
+ * this variant produces *exactly* the same results as `_withSeed()` variant,
+ * hence offering only a pure speed benefit on "large" input,
+ * by skipping the need to regenerate the secret for every large input.
+ *
+ * Another usage scenario is to hash the secret to a 64-bit hash value,
+ * for example with XXH3_64bits(), which then becomes the seed,
+ * and then employ both the seed and the secret in _withSecretandSeed().
+ * On top of speed, an added benefit is that each bit in the secret
+ * has a 50% chance to swap each bit in the output, via its impact to the seed.
+ *
+ * This is not guaranteed when using the secret directly in "small data" scenarios,
+ * because only portions of the secret are employed for small data.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t
+XXH3_64bits_withSecretandSeed(XXH_NOESCAPE const void* data, size_t len,
+ XXH_NOESCAPE const void* secret, size_t secretSize,
+ XXH64_hash_t seed);
+/*!
+ * @brief Calculates 128-bit seeded variant of XXH3 hash of @p data.
+ *
+ * @param input The block of data to be hashed, at least @p len bytes in size.
+ * @param length The length of @p data, in bytes.
+ * @param secret The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ * @param seed64 The 64-bit seed to alter the hash result predictably.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @see XXH3_64bits_withSecretandSeed()
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t
+XXH3_128bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t length,
+ XXH_NOESCAPE const void* secret, size_t secretSize,
+ XXH64_hash_t seed64);
+#ifndef XXH_NO_STREAM
+/*!
+ * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
+ *
+ * @param statePtr A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState().
+ * @param secret The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ * @param seed64 The 64-bit seed to alter the hash result predictably.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @see XXH3_64bits_withSecretandSeed()
+ */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr,
+ XXH_NOESCAPE const void* secret, size_t secretSize,
+ XXH64_hash_t seed64);
+/*!
+ * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
+ *
+ * @param statePtr A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState().
+ * @param secret The secret data.
+ * @param secretSize The length of @p secret, in bytes.
+ * @param seed64 The 64-bit seed to alter the hash result predictably.
+ *
+ * @return @ref XXH_OK on success.
+ * @return @ref XXH_ERROR on failure.
+ *
+ * @see XXH3_64bits_withSecretandSeed()
+ */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr,
+ XXH_NOESCAPE const void* secret, size_t secretSize,
+ XXH64_hash_t seed64);
+#endif /* !XXH_NO_STREAM */
+
+#endif /* !XXH_NO_XXH3 */
+#endif /* XXH_NO_LONG_LONG */
+#if defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)
+# define XXH_IMPLEMENTATION
+#endif
+
+#endif /* defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742) */
+
+
+/* ======================================================================== */
+/* ======================================================================== */
+/* ======================================================================== */
+
+
+/*-**********************************************************************
+ * xxHash implementation
+ *-**********************************************************************
+ * xxHash's implementation used to be hosted inside xxhash.c.
+ *
+ * However, inlining requires implementation to be visible to the compiler,
+ * hence be included alongside the header.
+ * Previously, implementation was hosted inside xxhash.c,
+ * which was then #included when inlining was activated.
+ * This construction created issues with a few build and install systems,
+ * as it required xxhash.c to be stored in /include directory.
+ *
+ * xxHash implementation is now directly integrated within xxhash.h.
+ * As a consequence, xxhash.c is no longer needed in /include.
+ *
+ * xxhash.c is still available and is still useful.
+ * In a "normal" setup, when xxhash is not inlined,
+ * xxhash.h only exposes the prototypes and public symbols,
+ * while xxhash.c can be built into an object file xxhash.o
+ * which can then be linked into the final binary.
+ ************************************************************************/
+
+#if ( defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) \
+ || defined(XXH_IMPLEMENTATION) ) && !defined(XXH_IMPLEM_13a8737387)
+# define XXH_IMPLEM_13a8737387
+
+/* *************************************
+* Tuning parameters
+***************************************/
+
+/*!
+ * @defgroup tuning Tuning parameters
+ * @{
+ *
+ * Various macros to control xxHash's behavior.
+ */
+#ifdef XXH_DOXYGEN
+/*!
+ * @brief Define this to disable 64-bit code.
+ *
+ * Useful if only using the @ref XXH32_family and you have a strict C90 compiler.
+ */
+# define XXH_NO_LONG_LONG
+# undef XXH_NO_LONG_LONG /* don't actually */
+/*!
+ * @brief Controls how unaligned memory is accessed.
+ *
+ * By default, access to unaligned memory is controlled by `memcpy()`, which is
+ * safe and portable.
+ *
+ * Unfortunately, on some target/compiler combinations, the generated assembly
+ * is sub-optimal.
+ *
+ * The below switch allow selection of a different access method
+ * in the search for improved performance.
+ *
+ * @par Possible options:
+ *
+ * - `XXH_FORCE_MEMORY_ACCESS=0` (default): `memcpy`
+ * @par
+ * Use `memcpy()`. Safe and portable. Note that most modern compilers will
+ * eliminate the function call and treat it as an unaligned access.
+ *
+ * - `XXH_FORCE_MEMORY_ACCESS=1`: `__attribute__((aligned(1)))`
+ * @par
+ * Depends on compiler extensions and is therefore not portable.
+ * This method is safe _if_ your compiler supports it,
+ * and *generally* as fast or faster than `memcpy`.
+ *
+ * - `XXH_FORCE_MEMORY_ACCESS=2`: Direct cast
+ * @par
+ * Casts directly and dereferences. This method doesn't depend on the
+ * compiler, but it violates the C standard as it directly dereferences an
+ * unaligned pointer. It can generate buggy code on targets which do not
+ * support unaligned memory accesses, but in some circumstances, it's the
+ * only known way to get the most performance.
+ *
+ * - `XXH_FORCE_MEMORY_ACCESS=3`: Byteshift
+ * @par
+ * Also portable. This can generate the best code on old compilers which don't
+ * inline small `memcpy()` calls, and it might also be faster on big-endian
+ * systems which lack a native byteswap instruction. However, some compilers
+ * will emit literal byteshifts even if the target supports unaligned access.
+ *
+ *
+ * @warning
+ * Methods 1 and 2 rely on implementation-defined behavior. Use these with
+ * care, as what works on one compiler/platform/optimization level may cause
+ * another to read garbage data or even crash.
+ *
+ * See https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html for details.
+ *
+ * Prefer these methods in priority order (0 > 3 > 1 > 2)
+ */
+# define XXH_FORCE_MEMORY_ACCESS 0
+
+/*!
+ * @def XXH_SIZE_OPT
+ * @brief Controls how much xxHash optimizes for size.
+ *
+ * xxHash, when compiled, tends to result in a rather large binary size. This
+ * is mostly due to heavy usage to forced inlining and constant folding of the
+ * @ref XXH3_family to increase performance.
+ *
+ * However, some developers prefer size over speed. This option can
+ * significantly reduce the size of the generated code. When using the `-Os`
+ * or `-Oz` options on GCC or Clang, this is defined to 1 by default,
+ * otherwise it is defined to 0.
+ *
+ * Most of these size optimizations can be controlled manually.
+ *
+ * This is a number from 0-2.
+ * - `XXH_SIZE_OPT` == 0: Default. xxHash makes no size optimizations. Speed
+ * comes first.
+ * - `XXH_SIZE_OPT` == 1: Default for `-Os` and `-Oz`. xxHash is more
+ * conservative and disables hacks that increase code size. It implies the
+ * options @ref XXH_NO_INLINE_HINTS == 1, @ref XXH_FORCE_ALIGN_CHECK == 0,
+ * and @ref XXH3_NEON_LANES == 8 if they are not already defined.
+ * - `XXH_SIZE_OPT` == 2: xxHash tries to make itself as small as possible.
+ * Performance may cry. For example, the single shot functions just use the
+ * streaming API.
+ */
+# define XXH_SIZE_OPT 0
+
+/*!
+ * @def XXH_FORCE_ALIGN_CHECK
+ * @brief If defined to non-zero, adds a special path for aligned inputs (XXH32()
+ * and XXH64() only).
+ *
+ * This is an important performance trick for architectures without decent
+ * unaligned memory access performance.
+ *
+ * It checks for input alignment, and when conditions are met, uses a "fast
+ * path" employing direct 32-bit/64-bit reads, resulting in _dramatically
+ * faster_ read speed.
+ *
+ * The check costs one initial branch per hash, which is generally negligible,
+ * but not zero.
+ *
+ * Moreover, it's not useful to generate an additional code path if memory
+ * access uses the same instruction for both aligned and unaligned
+ * addresses (e.g. x86 and aarch64).
+ *
+ * In these cases, the alignment check can be removed by setting this macro to 0.
+ * Then the code will always use unaligned memory access.
+ * Align check is automatically disabled on x86, x64, ARM64, and some ARM chips
+ * which are platforms known to offer good unaligned memory accesses performance.
+ *
+ * It is also disabled by default when @ref XXH_SIZE_OPT >= 1.
+ *
+ * This option does not affect XXH3 (only XXH32 and XXH64).
+ */
+# define XXH_FORCE_ALIGN_CHECK 0
+
+/*!
+ * @def XXH_NO_INLINE_HINTS
+ * @brief When non-zero, sets all functions to `static`.
+ *
+ * By default, xxHash tries to force the compiler to inline almost all internal
+ * functions.
+ *
+ * This can usually improve performance due to reduced jumping and improved
+ * constant folding, but significantly increases the size of the binary which
+ * might not be favorable.
+ *
+ * Additionally, sometimes the forced inlining can be detrimental to performance,
+ * depending on the architecture.
+ *
+ * XXH_NO_INLINE_HINTS marks all internal functions as static, giving the
+ * compiler full control on whether to inline or not.
+ *
+ * When not optimizing (-O0), using `-fno-inline` with GCC or Clang, or if
+ * @ref XXH_SIZE_OPT >= 1, this will automatically be defined.
+ */
+# define XXH_NO_INLINE_HINTS 0
+
+/*!
+ * @def XXH3_INLINE_SECRET
+ * @brief Determines whether to inline the XXH3 withSecret code.
+ *
+ * When the secret size is known, the compiler can improve the performance
+ * of XXH3_64bits_withSecret() and XXH3_128bits_withSecret().
+ *
+ * However, if the secret size is not known, it doesn't have any benefit. This
+ * happens when xxHash is compiled into a global symbol. Therefore, if
+ * @ref XXH_INLINE_ALL is *not* defined, this will be defined to 0.
+ *
+ * Additionally, this defaults to 0 on GCC 12+, which has an issue with function pointers
+ * that are *sometimes* force inline on -Og, and it is impossible to automatically
+ * detect this optimization level.
+ */
+# define XXH3_INLINE_SECRET 0
+
+/*!
+ * @def XXH32_ENDJMP
+ * @brief Whether to use a jump for `XXH32_finalize`.
+ *
+ * For performance, `XXH32_finalize` uses multiple branches in the finalizer.
+ * This is generally preferable for performance,
+ * but depending on exact architecture, a jmp may be preferable.
+ *
+ * This setting is only possibly making a difference for very small inputs.
+ */
+# define XXH32_ENDJMP 0
+
+/*!
+ * @internal
+ * @brief Redefines old internal names.
+ *
+ * For compatibility with code that uses xxHash's internals before the names
+ * were changed to improve namespacing. There is no other reason to use this.
+ */
+# define XXH_OLD_NAMES
+# undef XXH_OLD_NAMES /* don't actually use, it is ugly. */
+
+/*!
+ * @def XXH_NO_STREAM
+ * @brief Disables the streaming API.
+ *
+ * When xxHash is not inlined and the streaming functions are not used, disabling
+ * the streaming functions can improve code size significantly, especially with
+ * the @ref XXH3_family which tends to make constant folded copies of itself.
+ */
+# define XXH_NO_STREAM
+# undef XXH_NO_STREAM /* don't actually */
+#endif /* XXH_DOXYGEN */
+/*!
+ * @}
+ */
+
+#ifndef XXH_FORCE_MEMORY_ACCESS /* can be defined externally, on command line for example */
+ /* prefer __packed__ structures (method 1) for GCC
+ * < ARMv7 with unaligned access (e.g. Raspbian armhf) still uses byte shifting, so we use memcpy
+ * which for some reason does unaligned loads. */
+# if defined(__GNUC__) && !(defined(__ARM_ARCH) && __ARM_ARCH < 7 && defined(__ARM_FEATURE_UNALIGNED))
+# define XXH_FORCE_MEMORY_ACCESS 1
+# endif
+#endif
+
+#ifndef XXH_SIZE_OPT
+ /* default to 1 for -Os or -Oz */
+# if (defined(__GNUC__) || defined(__clang__)) && defined(__OPTIMIZE_SIZE__)
+# define XXH_SIZE_OPT 1
+# else
+# define XXH_SIZE_OPT 0
+# endif
+#endif
+
+#ifndef XXH_FORCE_ALIGN_CHECK /* can be defined externally */
+ /* don't check on sizeopt, x86, aarch64, or arm when unaligned access is available */
+# if XXH_SIZE_OPT >= 1 || \
+ defined(__i386) || defined(__x86_64__) || defined(__aarch64__) || defined(__ARM_FEATURE_UNALIGNED) \
+ || defined(_M_IX86) || defined(_M_X64) || defined(_M_ARM64) || defined(_M_ARM) /* visual */
+# define XXH_FORCE_ALIGN_CHECK 0
+# else
+# define XXH_FORCE_ALIGN_CHECK 1
+# endif
+#endif
+
+#ifndef XXH_NO_INLINE_HINTS
+# if XXH_SIZE_OPT >= 1 || defined(__NO_INLINE__) /* -O0, -fno-inline */
+# define XXH_NO_INLINE_HINTS 1
+# else
+# define XXH_NO_INLINE_HINTS 0
+# endif
+#endif
+
+#ifndef XXH3_INLINE_SECRET
+# if (defined(__GNUC__) && !defined(__clang__) && __GNUC__ >= 12) \
+ || !defined(XXH_INLINE_ALL)
+# define XXH3_INLINE_SECRET 0
+# else
+# define XXH3_INLINE_SECRET 1
+# endif
+#endif
+
+#ifndef XXH32_ENDJMP
+/* generally preferable for performance */
+# define XXH32_ENDJMP 0
+#endif
+
+/*!
+ * @defgroup impl Implementation
+ * @{
+ */
+
+
+/* *************************************
+* Includes & Memory related functions
+***************************************/
+#if defined(XXH_NO_STREAM)
+/* nothing */
+#elif defined(XXH_NO_STDLIB)
+
+/* When requesting to disable any mention of stdlib,
+ * the library loses the ability to invoked malloc / free.
+ * In practice, it means that functions like `XXH*_createState()`
+ * will always fail, and return NULL.
+ * This flag is useful in situations where
+ * xxhash.h is integrated into some kernel, embedded or limited environment
+ * without access to dynamic allocation.
+ */
+
+static XXH_CONSTF void* XXH_malloc(size_t s) { (void)s; return NULL; }
+static void XXH_free(void* p) { (void)p; }
+
+#else
+
+/*
+ * Modify the local functions below should you wish to use
+ * different memory routines for malloc() and free()
+ */
+#include <stdlib.h>
+
+/*!
+ * @internal
+ * @brief Modify this function to use a different routine than malloc().
+ */
+static XXH_MALLOCF void* XXH_malloc(size_t s) { return malloc(s); }
+
+/*!
+ * @internal
+ * @brief Modify this function to use a different routine than free().
+ */
+static void XXH_free(void* p) { free(p); }
+
+#endif /* XXH_NO_STDLIB */
+
+#include <string.h>
+
+/*!
+ * @internal
+ * @brief Modify this function to use a different routine than memcpy().
+ */
+static void* XXH_memcpy(void* dest, const void* src, size_t size)
+{
+ return memcpy(dest,src,size);
+}
+
+#include <limits.h> /* ULLONG_MAX */
+
+
+/* *************************************
+* Compiler Specific Options
+***************************************/
+#ifdef _MSC_VER /* Visual Studio warning fix */
+# pragma warning(disable : 4127) /* disable: C4127: conditional expression is constant */
+#endif
+
+#if XXH_NO_INLINE_HINTS /* disable inlining hints */
+# if defined(__GNUC__) || defined(__clang__)
+# define XXH_FORCE_INLINE static __attribute__((unused))
+# else
+# define XXH_FORCE_INLINE static
+# endif
+# define XXH_NO_INLINE static
+/* enable inlining hints */
+#elif defined(__GNUC__) || defined(__clang__)
+# define XXH_FORCE_INLINE static __inline__ __attribute__((always_inline, unused))
+# define XXH_NO_INLINE static __attribute__((noinline))
+#elif defined(_MSC_VER) /* Visual Studio */
+# define XXH_FORCE_INLINE static __forceinline
+# define XXH_NO_INLINE static __declspec(noinline)
+#elif defined (__cplusplus) \
+ || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L)) /* C99 */
+# define XXH_FORCE_INLINE static inline
+# define XXH_NO_INLINE static
+#else
+# define XXH_FORCE_INLINE static
+# define XXH_NO_INLINE static
+#endif
+
+#if XXH3_INLINE_SECRET
+# define XXH3_WITH_SECRET_INLINE XXH_FORCE_INLINE
+#else
+# define XXH3_WITH_SECRET_INLINE XXH_NO_INLINE
+#endif
+
+
+/* *************************************
+* Debug
+***************************************/
+/*!
+ * @ingroup tuning
+ * @def XXH_DEBUGLEVEL
+ * @brief Sets the debugging level.
+ *
+ * XXH_DEBUGLEVEL is expected to be defined externally, typically via the
+ * compiler's command line options. The value must be a number.
+ */
+#ifndef XXH_DEBUGLEVEL
+# ifdef DEBUGLEVEL /* backwards compat */
+# define XXH_DEBUGLEVEL DEBUGLEVEL
+# else
+# define XXH_DEBUGLEVEL 0
+# endif
+#endif
+
+#if (XXH_DEBUGLEVEL>=1)
+# include <assert.h> /* note: can still be disabled with NDEBUG */
+# define XXH_ASSERT(c) assert(c)
+#else
+# if defined(__INTEL_COMPILER)
+# define XXH_ASSERT(c) XXH_ASSUME((unsigned char) (c))
+# else
+# define XXH_ASSERT(c) XXH_ASSUME(c)
+# endif
+#endif
+
+/* note: use after variable declarations */
+#ifndef XXH_STATIC_ASSERT
+# if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* C11 */
+# define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { _Static_assert((c),m); } while(0)
+# elif defined(__cplusplus) && (__cplusplus >= 201103L) /* C++11 */
+# define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { static_assert((c),m); } while(0)
+# else
+# define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { struct xxh_sa { char x[(c) ? 1 : -1]; }; } while(0)
+# endif
+# define XXH_STATIC_ASSERT(c) XXH_STATIC_ASSERT_WITH_MESSAGE((c),#c)
+#endif
+
+/*!
+ * @internal
+ * @def XXH_COMPILER_GUARD(var)
+ * @brief Used to prevent unwanted optimizations for @p var.
+ *
+ * It uses an empty GCC inline assembly statement with a register constraint
+ * which forces @p var into a general purpose register (eg eax, ebx, ecx
+ * on x86) and marks it as modified.
+ *
+ * This is used in a few places to avoid unwanted autovectorization (e.g.
+ * XXH32_round()). All vectorization we want is explicit via intrinsics,
+ * and _usually_ isn't wanted elsewhere.
+ *
+ * We also use it to prevent unwanted constant folding for AArch64 in
+ * XXH3_initCustomSecret_scalar().
+ */
+#if defined(__GNUC__) || defined(__clang__)
+# define XXH_COMPILER_GUARD(var) __asm__("" : "+r" (var))
+#else
+# define XXH_COMPILER_GUARD(var) ((void)0)
+#endif
+
+/* Specifically for NEON vectors which use the "w" constraint, on
+ * Clang. */
+#if defined(__clang__) && defined(__ARM_ARCH) && !defined(__wasm__)
+# define XXH_COMPILER_GUARD_CLANG_NEON(var) __asm__("" : "+w" (var))
+#else
+# define XXH_COMPILER_GUARD_CLANG_NEON(var) ((void)0)
+#endif
+
+/* *************************************
+* Basic Types
+***************************************/
+#if !defined (__VMS) \
+ && (defined (__cplusplus) \
+ || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+# ifdef _AIX
+# include <inttypes.h>
+# else
+# include <stdint.h>
+# endif
+ typedef uint8_t xxh_u8;
+#else
+ typedef unsigned char xxh_u8;
+#endif
+typedef XXH32_hash_t xxh_u32;
+
+#ifdef XXH_OLD_NAMES
+# warning "XXH_OLD_NAMES is planned to be removed starting v0.9. If the program depends on it, consider moving away from it by employing newer type names directly"
+# define BYTE xxh_u8
+# define U8 xxh_u8
+# define U32 xxh_u32
+#endif
+
+/* *** Memory access *** */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_read32(const void* ptr)
+ * @brief Reads an unaligned 32-bit integer from @p ptr in native endianness.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ *
+ * @param ptr The pointer to read from.
+ * @return The 32-bit native endian integer from the bytes at @p ptr.
+ */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_readLE32(const void* ptr)
+ * @brief Reads an unaligned 32-bit little endian integer from @p ptr.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ *
+ * @param ptr The pointer to read from.
+ * @return The 32-bit little endian integer from the bytes at @p ptr.
+ */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_readBE32(const void* ptr)
+ * @brief Reads an unaligned 32-bit big endian integer from @p ptr.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ *
+ * @param ptr The pointer to read from.
+ * @return The 32-bit big endian integer from the bytes at @p ptr.
+ */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_readLE32_align(const void* ptr, XXH_alignment align)
+ * @brief Like @ref XXH_readLE32(), but has an option for aligned reads.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ * Note that when @ref XXH_FORCE_ALIGN_CHECK == 0, the @p align parameter is
+ * always @ref XXH_alignment::XXH_unaligned.
+ *
+ * @param ptr The pointer to read from.
+ * @param align Whether @p ptr is aligned.
+ * @pre
+ * If @p align == @ref XXH_alignment::XXH_aligned, @p ptr must be 4 byte
+ * aligned.
+ * @return The 32-bit little endian integer from the bytes at @p ptr.
+ */
+
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+/*
+ * Manual byteshift. Best for old compilers which don't inline memcpy.
+ * We actually directly use XXH_readLE32 and XXH_readBE32.
+ */
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
+
+/*
+ * Force direct memory access. Only works on CPU which support unaligned memory
+ * access in hardware.
+ */
+static xxh_u32 XXH_read32(const void* memPtr) { return *(const xxh_u32*) memPtr; }
+
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
+
+/*
+ * __attribute__((aligned(1))) is supported by gcc and clang. Originally the
+ * documentation claimed that it only increased the alignment, but actually it
+ * can decrease it on gcc, clang, and icc:
+ * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502,
+ * https://gcc.godbolt.org/z/xYez1j67Y.
+ */
+#ifdef XXH_OLD_NAMES
+typedef union { xxh_u32 u32; } __attribute__((packed)) unalign;
+#endif
+static xxh_u32 XXH_read32(const void* ptr)
+{
+ typedef __attribute__((aligned(1))) xxh_u32 xxh_unalign32;
+ return *((const xxh_unalign32*)ptr);
+}
+
+#else
+
+/*
+ * Portable and safe solution. Generally efficient.
+ * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
+ */
+static xxh_u32 XXH_read32(const void* memPtr)
+{
+ xxh_u32 val;
+ XXH_memcpy(&val, memPtr, sizeof(val));
+ return val;
+}
+
+#endif /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
+
+
+/* *** Endianness *** */
+
+/*!
+ * @ingroup tuning
+ * @def XXH_CPU_LITTLE_ENDIAN
+ * @brief Whether the target is little endian.
+ *
+ * Defined to 1 if the target is little endian, or 0 if it is big endian.
+ * It can be defined externally, for example on the compiler command line.
+ *
+ * If it is not defined,
+ * a runtime check (which is usually constant folded) is used instead.
+ *
+ * @note
+ * This is not necessarily defined to an integer constant.
+ *
+ * @see XXH_isLittleEndian() for the runtime check.
+ */
+#ifndef XXH_CPU_LITTLE_ENDIAN
+/*
+ * Try to detect endianness automatically, to avoid the nonstandard behavior
+ * in `XXH_isLittleEndian()`
+ */
+# if defined(_WIN32) /* Windows is always little endian */ \
+ || defined(__LITTLE_ENDIAN__) \
+ || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
+# define XXH_CPU_LITTLE_ENDIAN 1
+# elif defined(__BIG_ENDIAN__) \
+ || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+# define XXH_CPU_LITTLE_ENDIAN 0
+# else
+/*!
+ * @internal
+ * @brief Runtime check for @ref XXH_CPU_LITTLE_ENDIAN.
+ *
+ * Most compilers will constant fold this.
+ */
+static int XXH_isLittleEndian(void)
+{
+ /*
+ * Portable and well-defined behavior.
+ * Don't use static: it is detrimental to performance.
+ */
+ const union { xxh_u32 u; xxh_u8 c[4]; } one = { 1 };
+ return one.c[0];
+}
+# define XXH_CPU_LITTLE_ENDIAN XXH_isLittleEndian()
+# endif
+#endif
+
+
+
+
+/* ****************************************
+* Compiler-specific Functions and Macros
+******************************************/
+#define XXH_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
+
+#ifdef __has_builtin
+# define XXH_HAS_BUILTIN(x) __has_builtin(x)
+#else
+# define XXH_HAS_BUILTIN(x) 0
+#endif
+
+
+
+/*
+ * C23 and future versions have standard "unreachable()".
+ * Once it has been implemented reliably we can add it as an
+ * additional case:
+ *
+ * ```
+ * #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN)
+ * # include <stddef.h>
+ * # ifdef unreachable
+ * # define XXH_UNREACHABLE() unreachable()
+ * # endif
+ * #endif
+ * ```
+ *
+ * Note C++23 also has std::unreachable() which can be detected
+ * as follows:
+ * ```
+ * #if defined(__cpp_lib_unreachable) && (__cpp_lib_unreachable >= 202202L)
+ * # include <utility>
+ * # define XXH_UNREACHABLE() std::unreachable()
+ * #endif
+ * ```
+ * NB: `__cpp_lib_unreachable` is defined in the `<version>` header.
+ * We don't use that as including `<utility>` in `extern "C"` blocks
+ * doesn't work on GCC12
+ */
+
+#if XXH_HAS_BUILTIN(__builtin_unreachable)
+# define XXH_UNREACHABLE() __builtin_unreachable()
+
+#elif defined(_MSC_VER)
+# define XXH_UNREACHABLE() __assume(0)
+
+#else
+# define XXH_UNREACHABLE()
+#endif
+
+#if XXH_HAS_BUILTIN(__builtin_assume)
+# define XXH_ASSUME(c) __builtin_assume(c)
+#else
+# define XXH_ASSUME(c) if (!(c)) { XXH_UNREACHABLE(); }
+#endif
+
+/*!
+ * @internal
+ * @def XXH_rotl32(x,r)
+ * @brief 32-bit rotate left.
+ *
+ * @param x The 32-bit integer to be rotated.
+ * @param r The number of bits to rotate.
+ * @pre
+ * @p r > 0 && @p r < 32
+ * @note
+ * @p x and @p r may be evaluated multiple times.
+ * @return The rotated result.
+ */
+#if !defined(NO_CLANG_BUILTIN) && XXH_HAS_BUILTIN(__builtin_rotateleft32) \
+ && XXH_HAS_BUILTIN(__builtin_rotateleft64)
+# define XXH_rotl32 __builtin_rotateleft32
+# define XXH_rotl64 __builtin_rotateleft64
+/* Note: although _rotl exists for minGW (GCC under windows), performance seems poor */
+#elif defined(_MSC_VER)
+# define XXH_rotl32(x,r) _rotl(x,r)
+# define XXH_rotl64(x,r) _rotl64(x,r)
+#else
+# define XXH_rotl32(x,r) (((x) << (r)) | ((x) >> (32 - (r))))
+# define XXH_rotl64(x,r) (((x) << (r)) | ((x) >> (64 - (r))))
+#endif
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_swap32(xxh_u32 x)
+ * @brief A 32-bit byteswap.
+ *
+ * @param x The 32-bit integer to byteswap.
+ * @return @p x, byteswapped.
+ */
+#if defined(_MSC_VER) /* Visual Studio */
+# define XXH_swap32 _byteswap_ulong
+#elif XXH_GCC_VERSION >= 403
+# define XXH_swap32 __builtin_bswap32
+#else
+static xxh_u32 XXH_swap32 (xxh_u32 x)
+{
+ return ((x << 24) & 0xff000000 ) |
+ ((x << 8) & 0x00ff0000 ) |
+ ((x >> 8) & 0x0000ff00 ) |
+ ((x >> 24) & 0x000000ff );
+}
+#endif
+
+
+/* ***************************
+* Memory reads
+*****************************/
+
+/*!
+ * @internal
+ * @brief Enum to indicate whether a pointer is aligned.
+ */
+typedef enum {
+ XXH_aligned, /*!< Aligned */
+ XXH_unaligned /*!< Possibly unaligned */
+} XXH_alignment;
+
+/*
+ * XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load.
+ *
+ * This is ideal for older compilers which don't inline memcpy.
+ */
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+
+XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* memPtr)
+{
+ const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+ return bytePtr[0]
+ | ((xxh_u32)bytePtr[1] << 8)
+ | ((xxh_u32)bytePtr[2] << 16)
+ | ((xxh_u32)bytePtr[3] << 24);
+}
+
+XXH_FORCE_INLINE xxh_u32 XXH_readBE32(const void* memPtr)
+{
+ const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+ return bytePtr[3]
+ | ((xxh_u32)bytePtr[2] << 8)
+ | ((xxh_u32)bytePtr[1] << 16)
+ | ((xxh_u32)bytePtr[0] << 24);
+}
+
+#else
+XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* ptr)
+{
+ return XXH_CPU_LITTLE_ENDIAN ? XXH_read32(ptr) : XXH_swap32(XXH_read32(ptr));
+}
+
+static xxh_u32 XXH_readBE32(const void* ptr)
+{
+ return XXH_CPU_LITTLE_ENDIAN ? XXH_swap32(XXH_read32(ptr)) : XXH_read32(ptr);
+}
+#endif
+
+XXH_FORCE_INLINE xxh_u32
+XXH_readLE32_align(const void* ptr, XXH_alignment align)
+{
+ if (align==XXH_unaligned) {
+ return XXH_readLE32(ptr);
+ } else {
+ return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u32*)ptr : XXH_swap32(*(const xxh_u32*)ptr);
+ }
+}
+
+
+/* *************************************
+* Misc
+***************************************/
+/*! @ingroup public */
+XXH_PUBLIC_API unsigned XXH_versionNumber (void) { return XXH_VERSION_NUMBER; }
+
+
+/* *******************************************************************
+* 32-bit hash functions
+*********************************************************************/
+/*!
+ * @}
+ * @defgroup XXH32_impl XXH32 implementation
+ * @ingroup impl
+ *
+ * Details on the XXH32 implementation.
+ * @{
+ */
+ /* #define instead of static const, to be used as initializers */
+#define XXH_PRIME32_1 0x9E3779B1U /*!< 0b10011110001101110111100110110001 */
+#define XXH_PRIME32_2 0x85EBCA77U /*!< 0b10000101111010111100101001110111 */
+#define XXH_PRIME32_3 0xC2B2AE3DU /*!< 0b11000010101100101010111000111101 */
+#define XXH_PRIME32_4 0x27D4EB2FU /*!< 0b00100111110101001110101100101111 */
+#define XXH_PRIME32_5 0x165667B1U /*!< 0b00010110010101100110011110110001 */
+
+#ifdef XXH_OLD_NAMES
+# define PRIME32_1 XXH_PRIME32_1
+# define PRIME32_2 XXH_PRIME32_2
+# define PRIME32_3 XXH_PRIME32_3
+# define PRIME32_4 XXH_PRIME32_4
+# define PRIME32_5 XXH_PRIME32_5
+#endif
+
+/*!
+ * @internal
+ * @brief Normal stripe processing routine.
+ *
+ * This shuffles the bits so that any bit from @p input impacts several bits in
+ * @p acc.
+ *
+ * @param acc The accumulator lane.
+ * @param input The stripe of input to mix.
+ * @return The mixed accumulator lane.
+ */
+static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input)
+{
+ acc += input * XXH_PRIME32_2;
+ acc = XXH_rotl32(acc, 13);
+ acc *= XXH_PRIME32_1;
+#if (defined(__SSE4_1__) || defined(__aarch64__) || defined(__wasm_simd128__)) && !defined(XXH_ENABLE_AUTOVECTORIZE)
+ /*
+ * UGLY HACK:
+ * A compiler fence is the only thing that prevents GCC and Clang from
+ * autovectorizing the XXH32 loop (pragmas and attributes don't work for some
+ * reason) without globally disabling SSE4.1.
+ *
+ * The reason we want to avoid vectorization is because despite working on
+ * 4 integers at a time, there are multiple factors slowing XXH32 down on
+ * SSE4:
+ * - There's a ridiculous amount of lag from pmulld (10 cycles of latency on
+ * newer chips!) making it slightly slower to multiply four integers at
+ * once compared to four integers independently. Even when pmulld was
+ * fastest, Sandy/Ivy Bridge, it is still not worth it to go into SSE
+ * just to multiply unless doing a long operation.
+ *
+ * - Four instructions are required to rotate,
+ * movqda tmp, v // not required with VEX encoding
+ * pslld tmp, 13 // tmp <<= 13
+ * psrld v, 19 // x >>= 19
+ * por v, tmp // x |= tmp
+ * compared to one for scalar:
+ * roll v, 13 // reliably fast across the board
+ * shldl v, v, 13 // Sandy Bridge and later prefer this for some reason
+ *
+ * - Instruction level parallelism is actually more beneficial here because
+ * the SIMD actually serializes this operation: While v1 is rotating, v2
+ * can load data, while v3 can multiply. SSE forces them to operate
+ * together.
+ *
+ * This is also enabled on AArch64, as Clang is *very aggressive* in vectorizing
+ * the loop. NEON is only faster on the A53, and with the newer cores, it is less
+ * than half the speed.
+ *
+ * Additionally, this is used on WASM SIMD128 because it JITs to the same
+ * SIMD instructions and has the same issue.
+ */
+ XXH_COMPILER_GUARD(acc);
+#endif
+ return acc;
+}
+
+/*!
+ * @internal
+ * @brief Mixes all bits to finalize the hash.
+ *
+ * The final mix ensures that all input bits have a chance to impact any bit in
+ * the output digest, resulting in an unbiased distribution.
+ *
+ * @param hash The hash to avalanche.
+ * @return The avalanched hash.
+ */
+static xxh_u32 XXH32_avalanche(xxh_u32 hash)
+{
+ hash ^= hash >> 15;
+ hash *= XXH_PRIME32_2;
+ hash ^= hash >> 13;
+ hash *= XXH_PRIME32_3;
+ hash ^= hash >> 16;
+ return hash;
+}
+
+#define XXH_get32bits(p) XXH_readLE32_align(p, align)
+
+/*!
+ * @internal
+ * @brief Processes the last 0-15 bytes of @p ptr.
+ *
+ * There may be up to 15 bytes remaining to consume from the input.
+ * This final stage will digest them to ensure that all input bytes are present
+ * in the final mix.
+ *
+ * @param hash The hash to finalize.
+ * @param ptr The pointer to the remaining input.
+ * @param len The remaining length, modulo 16.
+ * @param align Whether @p ptr is aligned.
+ * @return The finalized hash.
+ * @see XXH64_finalize().
+ */
+static XXH_PUREF xxh_u32
+XXH32_finalize(xxh_u32 hash, const xxh_u8* ptr, size_t len, XXH_alignment align)
+{
+#define XXH_PROCESS1 do { \
+ hash += (*ptr++) * XXH_PRIME32_5; \
+ hash = XXH_rotl32(hash, 11) * XXH_PRIME32_1; \
+} while (0)
+
+#define XXH_PROCESS4 do { \
+ hash += XXH_get32bits(ptr) * XXH_PRIME32_3; \
+ ptr += 4; \
+ hash = XXH_rotl32(hash, 17) * XXH_PRIME32_4; \
+} while (0)
+
+ if (ptr==NULL) XXH_ASSERT(len == 0);
+
+ /* Compact rerolled version; generally faster */
+ if (!XXH32_ENDJMP) {
+ len &= 15;
+ while (len >= 4) {
+ XXH_PROCESS4;
+ len -= 4;
+ }
+ while (len > 0) {
+ XXH_PROCESS1;
+ --len;
+ }
+ return XXH32_avalanche(hash);
+ } else {
+ switch(len&15) /* or switch(bEnd - p) */ {
+ case 12: XXH_PROCESS4;
+ XXH_FALLTHROUGH; /* fallthrough */
+ case 8: XXH_PROCESS4;
+ XXH_FALLTHROUGH; /* fallthrough */
+ case 4: XXH_PROCESS4;
+ return XXH32_avalanche(hash);
+
+ case 13: XXH_PROCESS4;
+ XXH_FALLTHROUGH; /* fallthrough */
+ case 9: XXH_PROCESS4;
+ XXH_FALLTHROUGH; /* fallthrough */
+ case 5: XXH_PROCESS4;
+ XXH_PROCESS1;
+ return XXH32_avalanche(hash);
+
+ case 14: XXH_PROCESS4;
+ XXH_FALLTHROUGH; /* fallthrough */
+ case 10: XXH_PROCESS4;
+ XXH_FALLTHROUGH; /* fallthrough */
+ case 6: XXH_PROCESS4;
+ XXH_PROCESS1;
+ XXH_PROCESS1;
+ return XXH32_avalanche(hash);
+
+ case 15: XXH_PROCESS4;
+ XXH_FALLTHROUGH; /* fallthrough */
+ case 11: XXH_PROCESS4;
+ XXH_FALLTHROUGH; /* fallthrough */
+ case 7: XXH_PROCESS4;
+ XXH_FALLTHROUGH; /* fallthrough */
+ case 3: XXH_PROCESS1;
+ XXH_FALLTHROUGH; /* fallthrough */
+ case 2: XXH_PROCESS1;
+ XXH_FALLTHROUGH; /* fallthrough */
+ case 1: XXH_PROCESS1;
+ XXH_FALLTHROUGH; /* fallthrough */
+ case 0: return XXH32_avalanche(hash);
+ }
+ XXH_ASSERT(0);
+ return hash; /* reaching this point is deemed impossible */
+ }
+}
+
+#ifdef XXH_OLD_NAMES
+# define PROCESS1 XXH_PROCESS1
+# define PROCESS4 XXH_PROCESS4
+#else
+# undef XXH_PROCESS1
+# undef XXH_PROCESS4
+#endif
+
+/*!
+ * @internal
+ * @brief The implementation for @ref XXH32().
+ *
+ * @param input , len , seed Directly passed from @ref XXH32().
+ * @param align Whether @p input is aligned.
+ * @return The calculated hash.
+ */
+XXH_FORCE_INLINE XXH_PUREF xxh_u32
+XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment align)
+{
+ xxh_u32 h32;
+
+ if (input==NULL) XXH_ASSERT(len == 0);
+
+ if (len>=16) {
+ const xxh_u8* const bEnd = input + len;
+ const xxh_u8* const limit = bEnd - 15;
+ xxh_u32 v1 = seed + XXH_PRIME32_1 + XXH_PRIME32_2;
+ xxh_u32 v2 = seed + XXH_PRIME32_2;
+ xxh_u32 v3 = seed + 0;
+ xxh_u32 v4 = seed - XXH_PRIME32_1;
+
+ do {
+ v1 = XXH32_round(v1, XXH_get32bits(input)); input += 4;
+ v2 = XXH32_round(v2, XXH_get32bits(input)); input += 4;
+ v3 = XXH32_round(v3, XXH_get32bits(input)); input += 4;
+ v4 = XXH32_round(v4, XXH_get32bits(input)); input += 4;
+ } while (input < limit);
+
+ h32 = XXH_rotl32(v1, 1) + XXH_rotl32(v2, 7)
+ + XXH_rotl32(v3, 12) + XXH_rotl32(v4, 18);
+ } else {
+ h32 = seed + XXH_PRIME32_5;
+ }
+
+ h32 += (xxh_u32)len;
+
+ return XXH32_finalize(h32, input, len&15, align);
+}
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t len, XXH32_hash_t seed)
+{
+#if !defined(XXH_NO_STREAM) && XXH_SIZE_OPT >= 2
+ /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
+ XXH32_state_t state;
+ XXH32_reset(&state, seed);
+ XXH32_update(&state, (const xxh_u8*)input, len);
+ return XXH32_digest(&state);
+#else
+ if (XXH_FORCE_ALIGN_CHECK) {
+ if ((((size_t)input) & 3) == 0) { /* Input is 4-bytes aligned, leverage the speed benefit */
+ return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_aligned);
+ } }
+
+ return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned);
+#endif
+}
+
+
+
+/******* Hash streaming *******/
+#ifndef XXH_NO_STREAM
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void)
+{
+ return (XXH32_state_t*)XXH_malloc(sizeof(XXH32_state_t));
+}
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr)
+{
+ XXH_free(statePtr);
+ return XXH_OK;
+}
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dstState, const XXH32_state_t* srcState)
+{
+ XXH_memcpy(dstState, srcState, sizeof(*dstState));
+}
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, XXH32_hash_t seed)
+{
+ XXH_ASSERT(statePtr != NULL);
+ memset(statePtr, 0, sizeof(*statePtr));
+ statePtr->v[0] = seed + XXH_PRIME32_1 + XXH_PRIME32_2;
+ statePtr->v[1] = seed + XXH_PRIME32_2;
+ statePtr->v[2] = seed + 0;
+ statePtr->v[3] = seed - XXH_PRIME32_1;
+ return XXH_OK;
+}
+
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH32_update(XXH32_state_t* state, const void* input, size_t len)
+{
+ if (input==NULL) {
+ XXH_ASSERT(len == 0);
+ return XXH_OK;
+ }
+
+ { const xxh_u8* p = (const xxh_u8*)input;
+ const xxh_u8* const bEnd = p + len;
+
+ state->total_len_32 += (XXH32_hash_t)len;
+ state->large_len |= (XXH32_hash_t)((len>=16) | (state->total_len_32>=16));
+
+ if (state->memsize + len < 16) { /* fill in tmp buffer */
+ XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, len);
+ state->memsize += (XXH32_hash_t)len;
+ return XXH_OK;
+ }
+
+ if (state->memsize) { /* some data left from previous update */
+ XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, 16-state->memsize);
+ { const xxh_u32* p32 = state->mem32;
+ state->v[0] = XXH32_round(state->v[0], XXH_readLE32(p32)); p32++;
+ state->v[1] = XXH32_round(state->v[1], XXH_readLE32(p32)); p32++;
+ state->v[2] = XXH32_round(state->v[2], XXH_readLE32(p32)); p32++;
+ state->v[3] = XXH32_round(state->v[3], XXH_readLE32(p32));
+ }
+ p += 16-state->memsize;
+ state->memsize = 0;
+ }
+
+ if (p <= bEnd-16) {
+ const xxh_u8* const limit = bEnd - 16;
+
+ do {
+ state->v[0] = XXH32_round(state->v[0], XXH_readLE32(p)); p+=4;
+ state->v[1] = XXH32_round(state->v[1], XXH_readLE32(p)); p+=4;
+ state->v[2] = XXH32_round(state->v[2], XXH_readLE32(p)); p+=4;
+ state->v[3] = XXH32_round(state->v[3], XXH_readLE32(p)); p+=4;
+ } while (p<=limit);
+
+ }
+
+ if (p < bEnd) {
+ XXH_memcpy(state->mem32, p, (size_t)(bEnd-p));
+ state->memsize = (unsigned)(bEnd-p);
+ }
+ }
+
+ return XXH_OK;
+}
+
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH32_hash_t XXH32_digest(const XXH32_state_t* state)
+{
+ xxh_u32 h32;
+
+ if (state->large_len) {
+ h32 = XXH_rotl32(state->v[0], 1)
+ + XXH_rotl32(state->v[1], 7)
+ + XXH_rotl32(state->v[2], 12)
+ + XXH_rotl32(state->v[3], 18);
+ } else {
+ h32 = state->v[2] /* == seed */ + XXH_PRIME32_5;
+ }
+
+ h32 += state->total_len_32;
+
+ return XXH32_finalize(h32, (const xxh_u8*)state->mem32, state->memsize, XXH_aligned);
+}
+#endif /* !XXH_NO_STREAM */
+
+/******* Canonical representation *******/
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash)
+{
+ XXH_STATIC_ASSERT(sizeof(XXH32_canonical_t) == sizeof(XXH32_hash_t));
+ if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap32(hash);
+ XXH_memcpy(dst, &hash, sizeof(*dst));
+}
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src)
+{
+ return XXH_readBE32(src);
+}
+
+
+#ifndef XXH_NO_LONG_LONG
+
+/* *******************************************************************
+* 64-bit hash functions
+*********************************************************************/
+/*!
+ * @}
+ * @ingroup impl
+ * @{
+ */
+/******* Memory access *******/
+
+typedef XXH64_hash_t xxh_u64;
+
+#ifdef XXH_OLD_NAMES
+# define U64 xxh_u64
+#endif
+
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+/*
+ * Manual byteshift. Best for old compilers which don't inline memcpy.
+ * We actually directly use XXH_readLE64 and XXH_readBE64.
+ */
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
+
+/* Force direct memory access. Only works on CPU which support unaligned memory access in hardware */
+static xxh_u64 XXH_read64(const void* memPtr)
+{
+ return *(const xxh_u64*) memPtr;
+}
+
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
+
+/*
+ * __attribute__((aligned(1))) is supported by gcc and clang. Originally the
+ * documentation claimed that it only increased the alignment, but actually it
+ * can decrease it on gcc, clang, and icc:
+ * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502,
+ * https://gcc.godbolt.org/z/xYez1j67Y.
+ */
+#ifdef XXH_OLD_NAMES
+typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) unalign64;
+#endif
+static xxh_u64 XXH_read64(const void* ptr)
+{
+ typedef __attribute__((aligned(1))) xxh_u64 xxh_unalign64;
+ return *((const xxh_unalign64*)ptr);
+}
+
+#else
+
+/*
+ * Portable and safe solution. Generally efficient.
+ * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
+ */
+static xxh_u64 XXH_read64(const void* memPtr)
+{
+ xxh_u64 val;
+ XXH_memcpy(&val, memPtr, sizeof(val));
+ return val;
+}
+
+#endif /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
+
+#if defined(_MSC_VER) /* Visual Studio */
+# define XXH_swap64 _byteswap_uint64
+#elif XXH_GCC_VERSION >= 403
+# define XXH_swap64 __builtin_bswap64
+#else
+static xxh_u64 XXH_swap64(xxh_u64 x)
+{
+ return ((x << 56) & 0xff00000000000000ULL) |
+ ((x << 40) & 0x00ff000000000000ULL) |
+ ((x << 24) & 0x0000ff0000000000ULL) |
+ ((x << 8) & 0x000000ff00000000ULL) |
+ ((x >> 8) & 0x00000000ff000000ULL) |
+ ((x >> 24) & 0x0000000000ff0000ULL) |
+ ((x >> 40) & 0x000000000000ff00ULL) |
+ ((x >> 56) & 0x00000000000000ffULL);
+}
+#endif
+
+
+/* XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load. */
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+
+XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* memPtr)
+{
+ const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+ return bytePtr[0]
+ | ((xxh_u64)bytePtr[1] << 8)
+ | ((xxh_u64)bytePtr[2] << 16)
+ | ((xxh_u64)bytePtr[3] << 24)
+ | ((xxh_u64)bytePtr[4] << 32)
+ | ((xxh_u64)bytePtr[5] << 40)
+ | ((xxh_u64)bytePtr[6] << 48)
+ | ((xxh_u64)bytePtr[7] << 56);
+}
+
+XXH_FORCE_INLINE xxh_u64 XXH_readBE64(const void* memPtr)
+{
+ const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+ return bytePtr[7]
+ | ((xxh_u64)bytePtr[6] << 8)
+ | ((xxh_u64)bytePtr[5] << 16)
+ | ((xxh_u64)bytePtr[4] << 24)
+ | ((xxh_u64)bytePtr[3] << 32)
+ | ((xxh_u64)bytePtr[2] << 40)
+ | ((xxh_u64)bytePtr[1] << 48)
+ | ((xxh_u64)bytePtr[0] << 56);
+}
+
+#else
+XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* ptr)
+{
+ return XXH_CPU_LITTLE_ENDIAN ? XXH_read64(ptr) : XXH_swap64(XXH_read64(ptr));
+}
+
+static xxh_u64 XXH_readBE64(const void* ptr)
+{
+ return XXH_CPU_LITTLE_ENDIAN ? XXH_swap64(XXH_read64(ptr)) : XXH_read64(ptr);
+}
+#endif
+
+XXH_FORCE_INLINE xxh_u64
+XXH_readLE64_align(const void* ptr, XXH_alignment align)
+{
+ if (align==XXH_unaligned)
+ return XXH_readLE64(ptr);
+ else
+ return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u64*)ptr : XXH_swap64(*(const xxh_u64*)ptr);
+}
+
+
+/******* xxh64 *******/
+/*!
+ * @}
+ * @defgroup XXH64_impl XXH64 implementation
+ * @ingroup impl
+ *
+ * Details on the XXH64 implementation.
+ * @{
+ */
+/* #define rather that static const, to be used as initializers */
+#define XXH_PRIME64_1 0x9E3779B185EBCA87ULL /*!< 0b1001111000110111011110011011000110000101111010111100101010000111 */
+#define XXH_PRIME64_2 0xC2B2AE3D27D4EB4FULL /*!< 0b1100001010110010101011100011110100100111110101001110101101001111 */
+#define XXH_PRIME64_3 0x165667B19E3779F9ULL /*!< 0b0001011001010110011001111011000110011110001101110111100111111001 */
+#define XXH_PRIME64_4 0x85EBCA77C2B2AE63ULL /*!< 0b1000010111101011110010100111011111000010101100101010111001100011 */
+#define XXH_PRIME64_5 0x27D4EB2F165667C5ULL /*!< 0b0010011111010100111010110010111100010110010101100110011111000101 */
+
+#ifdef XXH_OLD_NAMES
+# define PRIME64_1 XXH_PRIME64_1
+# define PRIME64_2 XXH_PRIME64_2
+# define PRIME64_3 XXH_PRIME64_3
+# define PRIME64_4 XXH_PRIME64_4
+# define PRIME64_5 XXH_PRIME64_5
+#endif
+
+/*! @copydoc XXH32_round */
+static xxh_u64 XXH64_round(xxh_u64 acc, xxh_u64 input)
+{
+ acc += input * XXH_PRIME64_2;
+ acc = XXH_rotl64(acc, 31);
+ acc *= XXH_PRIME64_1;
+#if (defined(__AVX512F__)) && !defined(XXH_ENABLE_AUTOVECTORIZE)
+ /*
+ * DISABLE AUTOVECTORIZATION:
+ * A compiler fence is used to prevent GCC and Clang from
+ * autovectorizing the XXH64 loop (pragmas and attributes don't work for some
+ * reason) without globally disabling AVX512.
+ *
+ * Autovectorization of XXH64 tends to be detrimental,
+ * though the exact outcome may change depending on exact cpu and compiler version.
+ * For information, it has been reported as detrimental for Skylake-X,
+ * but possibly beneficial for Zen4.
+ *
+ * The default is to disable auto-vectorization,
+ * but you can select to enable it instead using `XXH_ENABLE_AUTOVECTORIZE` build variable.
+ */
+ XXH_COMPILER_GUARD(acc);
+#endif
+ return acc;
+}
+
+static xxh_u64 XXH64_mergeRound(xxh_u64 acc, xxh_u64 val)
+{
+ val = XXH64_round(0, val);
+ acc ^= val;
+ acc = acc * XXH_PRIME64_1 + XXH_PRIME64_4;
+ return acc;
+}
+
+/*! @copydoc XXH32_avalanche */
+static xxh_u64 XXH64_avalanche(xxh_u64 hash)
+{
+ hash ^= hash >> 33;
+ hash *= XXH_PRIME64_2;
+ hash ^= hash >> 29;
+ hash *= XXH_PRIME64_3;
+ hash ^= hash >> 32;
+ return hash;
+}
+
+
+#define XXH_get64bits(p) XXH_readLE64_align(p, align)
+
+/*!
+ * @internal
+ * @brief Processes the last 0-31 bytes of @p ptr.
+ *
+ * There may be up to 31 bytes remaining to consume from the input.
+ * This final stage will digest them to ensure that all input bytes are present
+ * in the final mix.
+ *
+ * @param hash The hash to finalize.
+ * @param ptr The pointer to the remaining input.
+ * @param len The remaining length, modulo 32.
+ * @param align Whether @p ptr is aligned.
+ * @return The finalized hash
+ * @see XXH32_finalize().
+ */
+static XXH_PUREF xxh_u64
+XXH64_finalize(xxh_u64 hash, const xxh_u8* ptr, size_t len, XXH_alignment align)
+{
+ if (ptr==NULL) XXH_ASSERT(len == 0);
+ len &= 31;
+ while (len >= 8) {
+ xxh_u64 const k1 = XXH64_round(0, XXH_get64bits(ptr));
+ ptr += 8;
+ hash ^= k1;
+ hash = XXH_rotl64(hash,27) * XXH_PRIME64_1 + XXH_PRIME64_4;
+ len -= 8;
+ }
+ if (len >= 4) {
+ hash ^= (xxh_u64)(XXH_get32bits(ptr)) * XXH_PRIME64_1;
+ ptr += 4;
+ hash = XXH_rotl64(hash, 23) * XXH_PRIME64_2 + XXH_PRIME64_3;
+ len -= 4;
+ }
+ while (len > 0) {
+ hash ^= (*ptr++) * XXH_PRIME64_5;
+ hash = XXH_rotl64(hash, 11) * XXH_PRIME64_1;
+ --len;
+ }
+ return XXH64_avalanche(hash);
+}
+
+#ifdef XXH_OLD_NAMES
+# define PROCESS1_64 XXH_PROCESS1_64
+# define PROCESS4_64 XXH_PROCESS4_64
+# define PROCESS8_64 XXH_PROCESS8_64
+#else
+# undef XXH_PROCESS1_64
+# undef XXH_PROCESS4_64
+# undef XXH_PROCESS8_64
+#endif
+
+/*!
+ * @internal
+ * @brief The implementation for @ref XXH64().
+ *
+ * @param input , len , seed Directly passed from @ref XXH64().
+ * @param align Whether @p input is aligned.
+ * @return The calculated hash.
+ */
+XXH_FORCE_INLINE XXH_PUREF xxh_u64
+XXH64_endian_align(const xxh_u8* input, size_t len, xxh_u64 seed, XXH_alignment align)
+{
+ xxh_u64 h64;
+ if (input==NULL) XXH_ASSERT(len == 0);
+
+ if (len>=32) {
+ const xxh_u8* const bEnd = input + len;
+ const xxh_u8* const limit = bEnd - 31;
+ xxh_u64 v1 = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
+ xxh_u64 v2 = seed + XXH_PRIME64_2;
+ xxh_u64 v3 = seed + 0;
+ xxh_u64 v4 = seed - XXH_PRIME64_1;
+
+ do {
+ v1 = XXH64_round(v1, XXH_get64bits(input)); input+=8;
+ v2 = XXH64_round(v2, XXH_get64bits(input)); input+=8;
+ v3 = XXH64_round(v3, XXH_get64bits(input)); input+=8;
+ v4 = XXH64_round(v4, XXH_get64bits(input)); input+=8;
+ } while (input<limit);
+
+ h64 = XXH_rotl64(v1, 1) + XXH_rotl64(v2, 7) + XXH_rotl64(v3, 12) + XXH_rotl64(v4, 18);
+ h64 = XXH64_mergeRound(h64, v1);
+ h64 = XXH64_mergeRound(h64, v2);
+ h64 = XXH64_mergeRound(h64, v3);
+ h64 = XXH64_mergeRound(h64, v4);
+
+ } else {
+ h64 = seed + XXH_PRIME64_5;
+ }
+
+ h64 += (xxh_u64) len;
+
+ return XXH64_finalize(h64, input, len, align);
+}
+
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64 (XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
+{
+#if !defined(XXH_NO_STREAM) && XXH_SIZE_OPT >= 2
+ /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
+ XXH64_state_t state;
+ XXH64_reset(&state, seed);
+ XXH64_update(&state, (const xxh_u8*)input, len);
+ return XXH64_digest(&state);
+#else
+ if (XXH_FORCE_ALIGN_CHECK) {
+ if ((((size_t)input) & 7)==0) { /* Input is aligned, let's leverage the speed advantage */
+ return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_aligned);
+ } }
+
+ return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned);
+
+#endif
+}
+
+/******* Hash Streaming *******/
+#ifndef XXH_NO_STREAM
+/*! @ingroup XXH64_family*/
+XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void)
+{
+ return (XXH64_state_t*)XXH_malloc(sizeof(XXH64_state_t));
+}
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr)
+{
+ XXH_free(statePtr);
+ return XXH_OK;
+}
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dstState, const XXH64_state_t* srcState)
+{
+ XXH_memcpy(dstState, srcState, sizeof(*dstState));
+}
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed)
+{
+ XXH_ASSERT(statePtr != NULL);
+ memset(statePtr, 0, sizeof(*statePtr));
+ statePtr->v[0] = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
+ statePtr->v[1] = seed + XXH_PRIME64_2;
+ statePtr->v[2] = seed + 0;
+ statePtr->v[3] = seed - XXH_PRIME64_1;
+ return XXH_OK;
+}
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH64_update (XXH_NOESCAPE XXH64_state_t* state, XXH_NOESCAPE const void* input, size_t len)
+{
+ if (input==NULL) {
+ XXH_ASSERT(len == 0);
+ return XXH_OK;
+ }
+
+ { const xxh_u8* p = (const xxh_u8*)input;
+ const xxh_u8* const bEnd = p + len;
+
+ state->total_len += len;
+
+ if (state->memsize + len < 32) { /* fill in tmp buffer */
+ XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, len);
+ state->memsize += (xxh_u32)len;
+ return XXH_OK;
+ }
+
+ if (state->memsize) { /* tmp buffer is full */
+ XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, 32-state->memsize);
+ state->v[0] = XXH64_round(state->v[0], XXH_readLE64(state->mem64+0));
+ state->v[1] = XXH64_round(state->v[1], XXH_readLE64(state->mem64+1));
+ state->v[2] = XXH64_round(state->v[2], XXH_readLE64(state->mem64+2));
+ state->v[3] = XXH64_round(state->v[3], XXH_readLE64(state->mem64+3));
+ p += 32 - state->memsize;
+ state->memsize = 0;
+ }
+
+ if (p+32 <= bEnd) {
+ const xxh_u8* const limit = bEnd - 32;
+
+ do {
+ state->v[0] = XXH64_round(state->v[0], XXH_readLE64(p)); p+=8;
+ state->v[1] = XXH64_round(state->v[1], XXH_readLE64(p)); p+=8;
+ state->v[2] = XXH64_round(state->v[2], XXH_readLE64(p)); p+=8;
+ state->v[3] = XXH64_round(state->v[3], XXH_readLE64(p)); p+=8;
+ } while (p<=limit);
+
+ }
+
+ if (p < bEnd) {
+ XXH_memcpy(state->mem64, p, (size_t)(bEnd-p));
+ state->memsize = (unsigned)(bEnd-p);
+ }
+ }
+
+ return XXH_OK;
+}
+
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64_digest(XXH_NOESCAPE const XXH64_state_t* state)
+{
+ xxh_u64 h64;
+
+ if (state->total_len >= 32) {
+ h64 = XXH_rotl64(state->v[0], 1) + XXH_rotl64(state->v[1], 7) + XXH_rotl64(state->v[2], 12) + XXH_rotl64(state->v[3], 18);
+ h64 = XXH64_mergeRound(h64, state->v[0]);
+ h64 = XXH64_mergeRound(h64, state->v[1]);
+ h64 = XXH64_mergeRound(h64, state->v[2]);
+ h64 = XXH64_mergeRound(h64, state->v[3]);
+ } else {
+ h64 = state->v[2] /*seed*/ + XXH_PRIME64_5;
+ }
+
+ h64 += (xxh_u64) state->total_len;
+
+ return XXH64_finalize(h64, (const xxh_u8*)state->mem64, (size_t)state->total_len, XXH_aligned);
+}
+#endif /* !XXH_NO_STREAM */
+
+/******* Canonical representation *******/
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash)
+{
+ XXH_STATIC_ASSERT(sizeof(XXH64_canonical_t) == sizeof(XXH64_hash_t));
+ if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap64(hash);
+ XXH_memcpy(dst, &hash, sizeof(*dst));
+}
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src)
+{
+ return XXH_readBE64(src);
+}
+
+#ifndef XXH_NO_XXH3
+
+/* *********************************************************************
+* XXH3
+* New generation hash designed for speed on small keys and vectorization
+************************************************************************ */
+/*!
+ * @}
+ * @defgroup XXH3_impl XXH3 implementation
+ * @ingroup impl
+ * @{
+ */
+
+/* === Compiler specifics === */
+
+#if ((defined(sun) || defined(__sun)) && __cplusplus) /* Solaris includes __STDC_VERSION__ with C++. Tested with GCC 5.5 */
+# define XXH_RESTRICT /* disable */
+#elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L /* >= C99 */
+# define XXH_RESTRICT restrict
+#elif (defined (__GNUC__) && ((__GNUC__ > 3) || (__GNUC__ == 3 && __GNUC_MINOR__ >= 1))) \
+ || (defined (__clang__)) \
+ || (defined (_MSC_VER) && (_MSC_VER >= 1400)) \
+ || (defined (__INTEL_COMPILER) && (__INTEL_COMPILER >= 1300))
+/*
+ * There are a LOT more compilers that recognize __restrict but this
+ * covers the major ones.
+ */
+# define XXH_RESTRICT __restrict
+#else
+# define XXH_RESTRICT /* disable */
+#endif
+
+#if (defined(__GNUC__) && (__GNUC__ >= 3)) \
+ || (defined(__INTEL_COMPILER) && (__INTEL_COMPILER >= 800)) \
+ || defined(__clang__)
+# define XXH_likely(x) __builtin_expect(x, 1)
+# define XXH_unlikely(x) __builtin_expect(x, 0)
+#else
+# define XXH_likely(x) (x)
+# define XXH_unlikely(x) (x)
+#endif
+
+#ifndef XXH_HAS_INCLUDE
+# ifdef __has_include
+/*
+ * Not defined as XXH_HAS_INCLUDE(x) (function-like) because
+ * this causes segfaults in Apple Clang 4.2 (on Mac OS X 10.7 Lion)
+ */
+# define XXH_HAS_INCLUDE __has_include
+# else
+# define XXH_HAS_INCLUDE(x) 0
+# endif
+#endif
+
+#if defined(__GNUC__) || defined(__clang__)
+# if defined(__ARM_FEATURE_SVE)
+# include <arm_sve.h>
+# endif
+# if defined(__ARM_NEON__) || defined(__ARM_NEON) \
+ || (defined(_M_ARM) && _M_ARM >= 7) \
+ || defined(_M_ARM64) || defined(_M_ARM64EC) \
+ || (defined(__wasm_simd128__) && XXH_HAS_INCLUDE(<arm_neon.h>)) /* WASM SIMD128 via SIMDe */
+# define inline __inline__ /* circumvent a clang bug */
+# include <arm_neon.h>
+# undef inline
+# elif defined(__AVX2__)
+# include <immintrin.h>
+# elif defined(__SSE2__)
+# include <emmintrin.h>
+# endif
+#endif
+
+#if defined(_MSC_VER)
+# include <intrin.h>
+#endif
+
+/*
+ * One goal of XXH3 is to make it fast on both 32-bit and 64-bit, while
+ * remaining a true 64-bit/128-bit hash function.
+ *
+ * This is done by prioritizing a subset of 64-bit operations that can be
+ * emulated without too many steps on the average 32-bit machine.
+ *
+ * For example, these two lines seem similar, and run equally fast on 64-bit:
+ *
+ * xxh_u64 x;
+ * x ^= (x >> 47); // good
+ * x ^= (x >> 13); // bad
+ *
+ * However, to a 32-bit machine, there is a major difference.
+ *
+ * x ^= (x >> 47) looks like this:
+ *
+ * x.lo ^= (x.hi >> (47 - 32));
+ *
+ * while x ^= (x >> 13) looks like this:
+ *
+ * // note: funnel shifts are not usually cheap.
+ * x.lo ^= (x.lo >> 13) | (x.hi << (32 - 13));
+ * x.hi ^= (x.hi >> 13);
+ *
+ * The first one is significantly faster than the second, simply because the
+ * shift is larger than 32. This means:
+ * - All the bits we need are in the upper 32 bits, so we can ignore the lower
+ * 32 bits in the shift.
+ * - The shift result will always fit in the lower 32 bits, and therefore,
+ * we can ignore the upper 32 bits in the xor.
+ *
+ * Thanks to this optimization, XXH3 only requires these features to be efficient:
+ *
+ * - Usable unaligned access
+ * - A 32-bit or 64-bit ALU
+ * - If 32-bit, a decent ADC instruction
+ * - A 32 or 64-bit multiply with a 64-bit result
+ * - For the 128-bit variant, a decent byteswap helps short inputs.
+ *
+ * The first two are already required by XXH32, and almost all 32-bit and 64-bit
+ * platforms which can run XXH32 can run XXH3 efficiently.
+ *
+ * Thumb-1, the classic 16-bit only subset of ARM's instruction set, is one
+ * notable exception.
+ *
+ * First of all, Thumb-1 lacks support for the UMULL instruction which
+ * performs the important long multiply. This means numerous __aeabi_lmul
+ * calls.
+ *
+ * Second of all, the 8 functional registers are just not enough.
+ * Setup for __aeabi_lmul, byteshift loads, pointers, and all arithmetic need
+ * Lo registers, and this shuffling results in thousands more MOVs than A32.
+ *
+ * A32 and T32 don't have this limitation. They can access all 14 registers,
+ * do a 32->64 multiply with UMULL, and the flexible operand allowing free
+ * shifts is helpful, too.
+ *
+ * Therefore, we do a quick sanity check.
+ *
+ * If compiling Thumb-1 for a target which supports ARM instructions, we will
+ * emit a warning, as it is not a "sane" platform to compile for.
+ *
+ * Usually, if this happens, it is because of an accident and you probably need
+ * to specify -march, as you likely meant to compile for a newer architecture.
+ *
+ * Credit: large sections of the vectorial and asm source code paths
+ * have been contributed by @easyaspi314
+ */
+#if defined(__thumb__) && !defined(__thumb2__) && defined(__ARM_ARCH_ISA_ARM)
+# warning "XXH3 is highly inefficient without ARM or Thumb-2."
+#endif
+
+/* ==========================================
+ * Vectorization detection
+ * ========================================== */
+
+#ifdef XXH_DOXYGEN
+/*!
+ * @ingroup tuning
+ * @brief Overrides the vectorization implementation chosen for XXH3.
+ *
+ * Can be defined to 0 to disable SIMD or any of the values mentioned in
+ * @ref XXH_VECTOR_TYPE.
+ *
+ * If this is not defined, it uses predefined macros to determine the best
+ * implementation.
+ */
+# define XXH_VECTOR XXH_SCALAR
+/*!
+ * @ingroup tuning
+ * @brief Possible values for @ref XXH_VECTOR.
+ *
+ * Note that these are actually implemented as macros.
+ *
+ * If this is not defined, it is detected automatically.
+ * internal macro XXH_X86DISPATCH overrides this.
+ */
+enum XXH_VECTOR_TYPE /* fake enum */ {
+ XXH_SCALAR = 0, /*!< Portable scalar version */
+ XXH_SSE2 = 1, /*!<
+ * SSE2 for Pentium 4, Opteron, all x86_64.
+ *
+ * @note SSE2 is also guaranteed on Windows 10, macOS, and
+ * Android x86.
+ */
+ XXH_AVX2 = 2, /*!< AVX2 for Haswell and Bulldozer */
+ XXH_AVX512 = 3, /*!< AVX512 for Skylake and Icelake */
+ XXH_NEON = 4, /*!<
+ * NEON for most ARMv7-A, all AArch64, and WASM SIMD128
+ * via the SIMDeverywhere polyfill provided with the
+ * Emscripten SDK.
+ */
+ XXH_VSX = 5, /*!< VSX and ZVector for POWER8/z13 (64-bit) */
+ XXH_SVE = 6, /*!< SVE for some ARMv8-A and ARMv9-A */
+};
+/*!
+ * @ingroup tuning
+ * @brief Selects the minimum alignment for XXH3's accumulators.
+ *
+ * When using SIMD, this should match the alignment required for said vector
+ * type, so, for example, 32 for AVX2.
+ *
+ * Default: Auto detected.
+ */
+# define XXH_ACC_ALIGN 8
+#endif
+
+/* Actual definition */
+#ifndef XXH_DOXYGEN
+# define XXH_SCALAR 0
+# define XXH_SSE2 1
+# define XXH_AVX2 2
+# define XXH_AVX512 3
+# define XXH_NEON 4
+# define XXH_VSX 5
+# define XXH_SVE 6
+#endif
+
+#ifndef XXH_VECTOR /* can be defined on command line */
+# if defined(__ARM_FEATURE_SVE)
+# define XXH_VECTOR XXH_SVE
+# elif ( \
+ defined(__ARM_NEON__) || defined(__ARM_NEON) /* gcc */ \
+ || defined(_M_ARM) || defined(_M_ARM64) || defined(_M_ARM64EC) /* msvc */ \
+ || (defined(__wasm_simd128__) && XXH_HAS_INCLUDE(<arm_neon.h>)) /* wasm simd128 via SIMDe */ \
+ ) && ( \
+ defined(_WIN32) || defined(__LITTLE_ENDIAN__) /* little endian only */ \
+ || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
+ )
+# define XXH_VECTOR XXH_NEON
+# elif defined(__AVX512F__)
+# define XXH_VECTOR XXH_AVX512
+# elif defined(__AVX2__)
+# define XXH_VECTOR XXH_AVX2
+# elif defined(__SSE2__) || defined(_M_AMD64) || defined(_M_X64) || (defined(_M_IX86_FP) && (_M_IX86_FP == 2))
+# define XXH_VECTOR XXH_SSE2
+# elif (defined(__PPC64__) && defined(__POWER8_VECTOR__)) \
+ || (defined(__s390x__) && defined(__VEC__)) \
+ && defined(__GNUC__) /* TODO: IBM XL */
+# define XXH_VECTOR XXH_VSX
+# else
+# define XXH_VECTOR XXH_SCALAR
+# endif
+#endif
+
+/* __ARM_FEATURE_SVE is only supported by GCC & Clang. */
+#if (XXH_VECTOR == XXH_SVE) && !defined(__ARM_FEATURE_SVE)
+# ifdef _MSC_VER
+# pragma warning(once : 4606)
+# else
+# warning "__ARM_FEATURE_SVE isn't supported. Use SCALAR instead."
+# endif
+# undef XXH_VECTOR
+# define XXH_VECTOR XXH_SCALAR
+#endif
+
+/*
+ * Controls the alignment of the accumulator,
+ * for compatibility with aligned vector loads, which are usually faster.
+ */
+#ifndef XXH_ACC_ALIGN
+# if defined(XXH_X86DISPATCH)
+# define XXH_ACC_ALIGN 64 /* for compatibility with avx512 */
+# elif XXH_VECTOR == XXH_SCALAR /* scalar */
+# define XXH_ACC_ALIGN 8
+# elif XXH_VECTOR == XXH_SSE2 /* sse2 */
+# define XXH_ACC_ALIGN 16
+# elif XXH_VECTOR == XXH_AVX2 /* avx2 */
+# define XXH_ACC_ALIGN 32
+# elif XXH_VECTOR == XXH_NEON /* neon */
+# define XXH_ACC_ALIGN 16
+# elif XXH_VECTOR == XXH_VSX /* vsx */
+# define XXH_ACC_ALIGN 16
+# elif XXH_VECTOR == XXH_AVX512 /* avx512 */
+# define XXH_ACC_ALIGN 64
+# elif XXH_VECTOR == XXH_SVE /* sve */
+# define XXH_ACC_ALIGN 64
+# endif
+#endif
+
+#if defined(XXH_X86DISPATCH) || XXH_VECTOR == XXH_SSE2 \
+ || XXH_VECTOR == XXH_AVX2 || XXH_VECTOR == XXH_AVX512
+# define XXH_SEC_ALIGN XXH_ACC_ALIGN
+#elif XXH_VECTOR == XXH_SVE
+# define XXH_SEC_ALIGN XXH_ACC_ALIGN
+#else
+# define XXH_SEC_ALIGN 8
+#endif
+
+#if defined(__GNUC__) || defined(__clang__)
+# define XXH_ALIASING __attribute__((may_alias))
+#else
+# define XXH_ALIASING /* nothing */
+#endif
+
+/*
+ * UGLY HACK:
+ * GCC usually generates the best code with -O3 for xxHash.
+ *
+ * However, when targeting AVX2, it is overzealous in its unrolling resulting
+ * in code roughly 3/4 the speed of Clang.
+ *
+ * There are other issues, such as GCC splitting _mm256_loadu_si256 into
+ * _mm_loadu_si128 + _mm256_inserti128_si256. This is an optimization which
+ * only applies to Sandy and Ivy Bridge... which don't even support AVX2.
+ *
+ * That is why when compiling the AVX2 version, it is recommended to use either
+ * -O2 -mavx2 -march=haswell
+ * or
+ * -O2 -mavx2 -mno-avx256-split-unaligned-load
+ * for decent performance, or to use Clang instead.
+ *
+ * Fortunately, we can control the first one with a pragma that forces GCC into
+ * -O2, but the other one we can't control without "failed to inline always
+ * inline function due to target mismatch" warnings.
+ */
+#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
+ && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
+ && defined(__OPTIMIZE__) && XXH_SIZE_OPT <= 0 /* respect -O0 and -Os */
+# pragma GCC push_options
+# pragma GCC optimize("-O2")
+#endif
+
+#if XXH_VECTOR == XXH_NEON
+
+/*
+ * UGLY HACK: While AArch64 GCC on Linux does not seem to care, on macOS, GCC -O3
+ * optimizes out the entire hashLong loop because of the aliasing violation.
+ *
+ * However, GCC is also inefficient at load-store optimization with vld1q/vst1q,
+ * so the only option is to mark it as aliasing.
+ */
+typedef uint64x2_t xxh_aliasing_uint64x2_t XXH_ALIASING;
+
+/*!
+ * @internal
+ * @brief `vld1q_u64` but faster and alignment-safe.
+ *
+ * On AArch64, unaligned access is always safe, but on ARMv7-a, it is only
+ * *conditionally* safe (`vld1` has an alignment bit like `movdq[ua]` in x86).
+ *
+ * GCC for AArch64 sees `vld1q_u8` as an intrinsic instead of a load, so it
+ * prohibits load-store optimizations. Therefore, a direct dereference is used.
+ *
+ * Otherwise, `vld1q_u8` is used with `vreinterpretq_u8_u64` to do a safe
+ * unaligned load.
+ */
+#if defined(__aarch64__) && defined(__GNUC__) && !defined(__clang__)
+XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr) /* silence -Wcast-align */
+{
+ return *(xxh_aliasing_uint64x2_t const *)ptr;
+}
+#else
+XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr)
+{
+ return vreinterpretq_u64_u8(vld1q_u8((uint8_t const*)ptr));
+}
+#endif
+
+/*!
+ * @internal
+ * @brief `vmlal_u32` on low and high halves of a vector.
+ *
+ * This is a workaround for AArch64 GCC < 11 which implemented arm_neon.h with
+ * inline assembly and were therefore incapable of merging the `vget_{low, high}_u32`
+ * with `vmlal_u32`.
+ */
+#if defined(__aarch64__) && defined(__GNUC__) && !defined(__clang__) && __GNUC__ < 11
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+ /* Inline assembly is the only way */
+ __asm__("umlal %0.2d, %1.2s, %2.2s" : "+w" (acc) : "w" (lhs), "w" (rhs));
+ return acc;
+}
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+ /* This intrinsic works as expected */
+ return vmlal_high_u32(acc, lhs, rhs);
+}
+#else
+/* Portable intrinsic versions */
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+ return vmlal_u32(acc, vget_low_u32(lhs), vget_low_u32(rhs));
+}
+/*! @copydoc XXH_vmlal_low_u32
+ * Assume the compiler converts this to vmlal_high_u32 on aarch64 */
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+ return vmlal_u32(acc, vget_high_u32(lhs), vget_high_u32(rhs));
+}
+#endif
+
+/*!
+ * @ingroup tuning
+ * @brief Controls the NEON to scalar ratio for XXH3
+ *
+ * This can be set to 2, 4, 6, or 8.
+ *
+ * ARM Cortex CPUs are _very_ sensitive to how their pipelines are used.
+ *
+ * For example, the Cortex-A73 can dispatch 3 micro-ops per cycle, but only 2 of those
+ * can be NEON. If you are only using NEON instructions, you are only using 2/3 of the CPU
+ * bandwidth.
+ *
+ * This is even more noticeable on the more advanced cores like the Cortex-A76 which
+ * can dispatch 8 micro-ops per cycle, but still only 2 NEON micro-ops at once.
+ *
+ * Therefore, to make the most out of the pipeline, it is beneficial to run 6 NEON lanes
+ * and 2 scalar lanes, which is chosen by default.
+ *
+ * This does not apply to Apple processors or 32-bit processors, which run better with
+ * full NEON. These will default to 8. Additionally, size-optimized builds run 8 lanes.
+ *
+ * This change benefits CPUs with large micro-op buffers without negatively affecting
+ * most other CPUs:
+ *
+ * | Chipset | Dispatch type | NEON only | 6:2 hybrid | Diff. |
+ * |:----------------------|:--------------------|----------:|-----------:|------:|
+ * | Snapdragon 730 (A76) | 2 NEON/8 micro-ops | 8.8 GB/s | 10.1 GB/s | ~16% |
+ * | Snapdragon 835 (A73) | 2 NEON/3 micro-ops | 5.1 GB/s | 5.3 GB/s | ~5% |
+ * | Marvell PXA1928 (A53) | In-order dual-issue | 1.9 GB/s | 1.9 GB/s | 0% |
+ * | Apple M1 | 4 NEON/8 micro-ops | 37.3 GB/s | 36.1 GB/s | ~-3% |
+ *
+ * It also seems to fix some bad codegen on GCC, making it almost as fast as clang.
+ *
+ * When using WASM SIMD128, if this is 2 or 6, SIMDe will scalarize 2 of the lanes meaning
+ * it effectively becomes worse 4.
+ *
+ * @see XXH3_accumulate_512_neon()
+ */
+# ifndef XXH3_NEON_LANES
+# if (defined(__aarch64__) || defined(__arm64__) || defined(_M_ARM64) || defined(_M_ARM64EC)) \
+ && !defined(__APPLE__) && XXH_SIZE_OPT <= 0
+# define XXH3_NEON_LANES 6
+# else
+# define XXH3_NEON_LANES XXH_ACC_NB
+# endif
+# endif
+#endif /* XXH_VECTOR == XXH_NEON */
+
+/*
+ * VSX and Z Vector helpers.
+ *
+ * This is very messy, and any pull requests to clean this up are welcome.
+ *
+ * There are a lot of problems with supporting VSX and s390x, due to
+ * inconsistent intrinsics, spotty coverage, and multiple endiannesses.
+ */
+#if XXH_VECTOR == XXH_VSX
+/* Annoyingly, these headers _may_ define three macros: `bool`, `vector`,
+ * and `pixel`. This is a problem for obvious reasons.
+ *
+ * These keywords are unnecessary; the spec literally says they are
+ * equivalent to `__bool`, `__vector`, and `__pixel` and may be undef'd
+ * after including the header.
+ *
+ * We use pragma push_macro/pop_macro to keep the namespace clean. */
+# pragma push_macro("bool")
+# pragma push_macro("vector")
+# pragma push_macro("pixel")
+/* silence potential macro redefined warnings */
+# undef bool
+# undef vector
+# undef pixel
+
+# if defined(__s390x__)
+# include <s390intrin.h>
+# else
+# include <altivec.h>
+# endif
+
+/* Restore the original macro values, if applicable. */
+# pragma pop_macro("pixel")
+# pragma pop_macro("vector")
+# pragma pop_macro("bool")
+
+typedef __vector unsigned long long xxh_u64x2;
+typedef __vector unsigned char xxh_u8x16;
+typedef __vector unsigned xxh_u32x4;
+
+/*
+ * UGLY HACK: Similar to aarch64 macOS GCC, s390x GCC has the same aliasing issue.
+ */
+typedef xxh_u64x2 xxh_aliasing_u64x2 XXH_ALIASING;
+
+# ifndef XXH_VSX_BE
+# if defined(__BIG_ENDIAN__) \
+ || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+# define XXH_VSX_BE 1
+# elif defined(__VEC_ELEMENT_REG_ORDER__) && __VEC_ELEMENT_REG_ORDER__ == __ORDER_BIG_ENDIAN__
+# warning "-maltivec=be is not recommended. Please use native endianness."
+# define XXH_VSX_BE 1
+# else
+# define XXH_VSX_BE 0
+# endif
+# endif /* !defined(XXH_VSX_BE) */
+
+# if XXH_VSX_BE
+# if defined(__POWER9_VECTOR__) || (defined(__clang__) && defined(__s390x__))
+# define XXH_vec_revb vec_revb
+# else
+/*!
+ * A polyfill for POWER9's vec_revb().
+ */
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_revb(xxh_u64x2 val)
+{
+ xxh_u8x16 const vByteSwap = { 0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00,
+ 0x0F, 0x0E, 0x0D, 0x0C, 0x0B, 0x0A, 0x09, 0x08 };
+ return vec_perm(val, val, vByteSwap);
+}
+# endif
+# endif /* XXH_VSX_BE */
+
+/*!
+ * Performs an unaligned vector load and byte swaps it on big endian.
+ */
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(const void *ptr)
+{
+ xxh_u64x2 ret;
+ XXH_memcpy(&ret, ptr, sizeof(xxh_u64x2));
+# if XXH_VSX_BE
+ ret = XXH_vec_revb(ret);
+# endif
+ return ret;
+}
+
+/*
+ * vec_mulo and vec_mule are very problematic intrinsics on PowerPC
+ *
+ * These intrinsics weren't added until GCC 8, despite existing for a while,
+ * and they are endian dependent. Also, their meaning swap depending on version.
+ * */
+# if defined(__s390x__)
+ /* s390x is always big endian, no issue on this platform */
+# define XXH_vec_mulo vec_mulo
+# define XXH_vec_mule vec_mule
+# elif defined(__clang__) && XXH_HAS_BUILTIN(__builtin_altivec_vmuleuw) && !defined(__ibmxl__)
+/* Clang has a better way to control this, we can just use the builtin which doesn't swap. */
+ /* The IBM XL Compiler (which defined __clang__) only implements the vec_* operations */
+# define XXH_vec_mulo __builtin_altivec_vmulouw
+# define XXH_vec_mule __builtin_altivec_vmuleuw
+# else
+/* gcc needs inline assembly */
+/* Adapted from https://github.com/google/highwayhash/blob/master/highwayhash/hh_vsx.h. */
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mulo(xxh_u32x4 a, xxh_u32x4 b)
+{
+ xxh_u64x2 result;
+ __asm__("vmulouw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
+ return result;
+}
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mule(xxh_u32x4 a, xxh_u32x4 b)
+{
+ xxh_u64x2 result;
+ __asm__("vmuleuw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
+ return result;
+}
+# endif /* XXH_vec_mulo, XXH_vec_mule */
+#endif /* XXH_VECTOR == XXH_VSX */
+
+#if XXH_VECTOR == XXH_SVE
+#define ACCRND(acc, offset) \
+do { \
+ svuint64_t input_vec = svld1_u64(mask, xinput + offset); \
+ svuint64_t secret_vec = svld1_u64(mask, xsecret + offset); \
+ svuint64_t mixed = sveor_u64_x(mask, secret_vec, input_vec); \
+ svuint64_t swapped = svtbl_u64(input_vec, kSwap); \
+ svuint64_t mixed_lo = svextw_u64_x(mask, mixed); \
+ svuint64_t mixed_hi = svlsr_n_u64_x(mask, mixed, 32); \
+ svuint64_t mul = svmad_u64_x(mask, mixed_lo, mixed_hi, swapped); \
+ acc = svadd_u64_x(mask, acc, mul); \
+} while (0)
+#endif /* XXH_VECTOR == XXH_SVE */
+
+/* prefetch
+ * can be disabled, by declaring XXH_NO_PREFETCH build macro */
+#if defined(XXH_NO_PREFETCH)
+# define XXH_PREFETCH(ptr) (void)(ptr) /* disabled */
+#else
+# if XXH_SIZE_OPT >= 1
+# define XXH_PREFETCH(ptr) (void)(ptr)
+# elif defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86)) /* _mm_prefetch() not defined outside of x86/x64 */
+# include <mmintrin.h> /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
+# define XXH_PREFETCH(ptr) _mm_prefetch((const char*)(ptr), _MM_HINT_T0)
+# elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) )
+# define XXH_PREFETCH(ptr) __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */)
+# else
+# define XXH_PREFETCH(ptr) (void)(ptr) /* disabled */
+# endif
+#endif /* XXH_NO_PREFETCH */
+
+
+/* ==========================================
+ * XXH3 default settings
+ * ========================================== */
+
+#define XXH_SECRET_DEFAULT_SIZE 192 /* minimum XXH3_SECRET_SIZE_MIN */
+
+#if (XXH_SECRET_DEFAULT_SIZE < XXH3_SECRET_SIZE_MIN)
+# error "default keyset is not large enough"
+#endif
+
+/*! Pseudorandom secret taken directly from FARSH. */
+XXH_ALIGN(64) static const xxh_u8 XXH3_kSecret[XXH_SECRET_DEFAULT_SIZE] = {
+ 0xb8, 0xfe, 0x6c, 0x39, 0x23, 0xa4, 0x4b, 0xbe, 0x7c, 0x01, 0x81, 0x2c, 0xf7, 0x21, 0xad, 0x1c,
+ 0xde, 0xd4, 0x6d, 0xe9, 0x83, 0x90, 0x97, 0xdb, 0x72, 0x40, 0xa4, 0xa4, 0xb7, 0xb3, 0x67, 0x1f,
+ 0xcb, 0x79, 0xe6, 0x4e, 0xcc, 0xc0, 0xe5, 0x78, 0x82, 0x5a, 0xd0, 0x7d, 0xcc, 0xff, 0x72, 0x21,
+ 0xb8, 0x08, 0x46, 0x74, 0xf7, 0x43, 0x24, 0x8e, 0xe0, 0x35, 0x90, 0xe6, 0x81, 0x3a, 0x26, 0x4c,
+ 0x3c, 0x28, 0x52, 0xbb, 0x91, 0xc3, 0x00, 0xcb, 0x88, 0xd0, 0x65, 0x8b, 0x1b, 0x53, 0x2e, 0xa3,
+ 0x71, 0x64, 0x48, 0x97, 0xa2, 0x0d, 0xf9, 0x4e, 0x38, 0x19, 0xef, 0x46, 0xa9, 0xde, 0xac, 0xd8,
+ 0xa8, 0xfa, 0x76, 0x3f, 0xe3, 0x9c, 0x34, 0x3f, 0xf9, 0xdc, 0xbb, 0xc7, 0xc7, 0x0b, 0x4f, 0x1d,
+ 0x8a, 0x51, 0xe0, 0x4b, 0xcd, 0xb4, 0x59, 0x31, 0xc8, 0x9f, 0x7e, 0xc9, 0xd9, 0x78, 0x73, 0x64,
+ 0xea, 0xc5, 0xac, 0x83, 0x34, 0xd3, 0xeb, 0xc3, 0xc5, 0x81, 0xa0, 0xff, 0xfa, 0x13, 0x63, 0xeb,
+ 0x17, 0x0d, 0xdd, 0x51, 0xb7, 0xf0, 0xda, 0x49, 0xd3, 0x16, 0x55, 0x26, 0x29, 0xd4, 0x68, 0x9e,
+ 0x2b, 0x16, 0xbe, 0x58, 0x7d, 0x47, 0xa1, 0xfc, 0x8f, 0xf8, 0xb8, 0xd1, 0x7a, 0xd0, 0x31, 0xce,
+ 0x45, 0xcb, 0x3a, 0x8f, 0x95, 0x16, 0x04, 0x28, 0xaf, 0xd7, 0xfb, 0xca, 0xbb, 0x4b, 0x40, 0x7e,
+};
+
+static const xxh_u64 PRIME_MX1 = 0x165667919E3779F9ULL; /*!< 0b0001011001010110011001111001000110011110001101110111100111111001 */
+static const xxh_u64 PRIME_MX2 = 0x9FB21C651E98DF25ULL; /*!< 0b1001111110110010000111000110010100011110100110001101111100100101 */
+
+#ifdef XXH_OLD_NAMES
+# define kSecret XXH3_kSecret
+#endif
+
+#ifdef XXH_DOXYGEN
+/*!
+ * @brief Calculates a 32-bit to 64-bit long multiply.
+ *
+ * Implemented as a macro.
+ *
+ * Wraps `__emulu` on MSVC x86 because it tends to call `__allmul` when it doesn't
+ * need to (but it shouldn't need to anyways, it is about 7 instructions to do
+ * a 64x64 multiply...). Since we know that this will _always_ emit `MULL`, we
+ * use that instead of the normal method.
+ *
+ * If you are compiling for platforms like Thumb-1 and don't have a better option,
+ * you may also want to write your own long multiply routine here.
+ *
+ * @param x, y Numbers to be multiplied
+ * @return 64-bit product of the low 32 bits of @p x and @p y.
+ */
+XXH_FORCE_INLINE xxh_u64
+XXH_mult32to64(xxh_u64 x, xxh_u64 y)
+{
+ return (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF);
+}
+#elif defined(_MSC_VER) && defined(_M_IX86)
+# define XXH_mult32to64(x, y) __emulu((unsigned)(x), (unsigned)(y))
+#else
+/*
+ * Downcast + upcast is usually better than masking on older compilers like
+ * GCC 4.2 (especially 32-bit ones), all without affecting newer compilers.
+ *
+ * The other method, (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF), will AND both operands
+ * and perform a full 64x64 multiply -- entirely redundant on 32-bit.
+ */
+# define XXH_mult32to64(x, y) ((xxh_u64)(xxh_u32)(x) * (xxh_u64)(xxh_u32)(y))
+#endif
+
+/*!
+ * @brief Calculates a 64->128-bit long multiply.
+ *
+ * Uses `__uint128_t` and `_umul128` if available, otherwise uses a scalar
+ * version.
+ *
+ * @param lhs , rhs The 64-bit integers to be multiplied
+ * @return The 128-bit result represented in an @ref XXH128_hash_t.
+ */
+static XXH128_hash_t
+XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs)
+{
+ /*
+ * GCC/Clang __uint128_t method.
+ *
+ * On most 64-bit targets, GCC and Clang define a __uint128_t type.
+ * This is usually the best way as it usually uses a native long 64-bit
+ * multiply, such as MULQ on x86_64 or MUL + UMULH on aarch64.
+ *
+ * Usually.
+ *
+ * Despite being a 32-bit platform, Clang (and emscripten) define this type
+ * despite not having the arithmetic for it. This results in a laggy
+ * compiler builtin call which calculates a full 128-bit multiply.
+ * In that case it is best to use the portable one.
+ * https://github.com/Cyan4973/xxHash/issues/211#issuecomment-515575677
+ */
+#if (defined(__GNUC__) || defined(__clang__)) && !defined(__wasm__) \
+ && defined(__SIZEOF_INT128__) \
+ || (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128)
+
+ __uint128_t const product = (__uint128_t)lhs * (__uint128_t)rhs;
+ XXH128_hash_t r128;
+ r128.low64 = (xxh_u64)(product);
+ r128.high64 = (xxh_u64)(product >> 64);
+ return r128;
+
+ /*
+ * MSVC for x64's _umul128 method.
+ *
+ * xxh_u64 _umul128(xxh_u64 Multiplier, xxh_u64 Multiplicand, xxh_u64 *HighProduct);
+ *
+ * This compiles to single operand MUL on x64.
+ */
+#elif (defined(_M_X64) || defined(_M_IA64)) && !defined(_M_ARM64EC)
+
+#ifndef _MSC_VER
+# pragma intrinsic(_umul128)
+#endif
+ xxh_u64 product_high;
+ xxh_u64 const product_low = _umul128(lhs, rhs, &product_high);
+ XXH128_hash_t r128;
+ r128.low64 = product_low;
+ r128.high64 = product_high;
+ return r128;
+
+ /*
+ * MSVC for ARM64's __umulh method.
+ *
+ * This compiles to the same MUL + UMULH as GCC/Clang's __uint128_t method.
+ */
+#elif defined(_M_ARM64) || defined(_M_ARM64EC)
+
+#ifndef _MSC_VER
+# pragma intrinsic(__umulh)
+#endif
+ XXH128_hash_t r128;
+ r128.low64 = lhs * rhs;
+ r128.high64 = __umulh(lhs, rhs);
+ return r128;
+
+#else
+ /*
+ * Portable scalar method. Optimized for 32-bit and 64-bit ALUs.
+ *
+ * This is a fast and simple grade school multiply, which is shown below
+ * with base 10 arithmetic instead of base 0x100000000.
+ *
+ * 9 3 // D2 lhs = 93
+ * x 7 5 // D2 rhs = 75
+ * ----------
+ * 1 5 // D2 lo_lo = (93 % 10) * (75 % 10) = 15
+ * 4 5 | // D2 hi_lo = (93 / 10) * (75 % 10) = 45
+ * 2 1 | // D2 lo_hi = (93 % 10) * (75 / 10) = 21
+ * + 6 3 | | // D2 hi_hi = (93 / 10) * (75 / 10) = 63
+ * ---------
+ * 2 7 | // D2 cross = (15 / 10) + (45 % 10) + 21 = 27
+ * + 6 7 | | // D2 upper = (27 / 10) + (45 / 10) + 63 = 67
+ * ---------
+ * 6 9 7 5 // D4 res = (27 * 10) + (15 % 10) + (67 * 100) = 6975
+ *
+ * The reasons for adding the products like this are:
+ * 1. It avoids manual carry tracking. Just like how
+ * (9 * 9) + 9 + 9 = 99, the same applies with this for UINT64_MAX.
+ * This avoids a lot of complexity.
+ *
+ * 2. It hints for, and on Clang, compiles to, the powerful UMAAL
+ * instruction available in ARM's Digital Signal Processing extension
+ * in 32-bit ARMv6 and later, which is shown below:
+ *
+ * void UMAAL(xxh_u32 *RdLo, xxh_u32 *RdHi, xxh_u32 Rn, xxh_u32 Rm)
+ * {
+ * xxh_u64 product = (xxh_u64)*RdLo * (xxh_u64)*RdHi + Rn + Rm;
+ * *RdLo = (xxh_u32)(product & 0xFFFFFFFF);
+ * *RdHi = (xxh_u32)(product >> 32);
+ * }
+ *
+ * This instruction was designed for efficient long multiplication, and
+ * allows this to be calculated in only 4 instructions at speeds
+ * comparable to some 64-bit ALUs.
+ *
+ * 3. It isn't terrible on other platforms. Usually this will be a couple
+ * of 32-bit ADD/ADCs.
+ */
+
+ /* First calculate all of the cross products. */
+ xxh_u64 const lo_lo = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs & 0xFFFFFFFF);
+ xxh_u64 const hi_lo = XXH_mult32to64(lhs >> 32, rhs & 0xFFFFFFFF);
+ xxh_u64 const lo_hi = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs >> 32);
+ xxh_u64 const hi_hi = XXH_mult32to64(lhs >> 32, rhs >> 32);
+
+ /* Now add the products together. These will never overflow. */
+ xxh_u64 const cross = (lo_lo >> 32) + (hi_lo & 0xFFFFFFFF) + lo_hi;
+ xxh_u64 const upper = (hi_lo >> 32) + (cross >> 32) + hi_hi;
+ xxh_u64 const lower = (cross << 32) | (lo_lo & 0xFFFFFFFF);
+
+ XXH128_hash_t r128;
+ r128.low64 = lower;
+ r128.high64 = upper;
+ return r128;
+#endif
+}
+
+/*!
+ * @brief Calculates a 64-bit to 128-bit multiply, then XOR folds it.
+ *
+ * The reason for the separate function is to prevent passing too many structs
+ * around by value. This will hopefully inline the multiply, but we don't force it.
+ *
+ * @param lhs , rhs The 64-bit integers to multiply
+ * @return The low 64 bits of the product XOR'd by the high 64 bits.
+ * @see XXH_mult64to128()
+ */
+static xxh_u64
+XXH3_mul128_fold64(xxh_u64 lhs, xxh_u64 rhs)
+{
+ XXH128_hash_t product = XXH_mult64to128(lhs, rhs);
+ return product.low64 ^ product.high64;
+}
+
+/*! Seems to produce slightly better code on GCC for some reason. */
+XXH_FORCE_INLINE XXH_CONSTF xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift)
+{
+ XXH_ASSERT(0 <= shift && shift < 64);
+ return v64 ^ (v64 >> shift);
+}
+
+/*
+ * This is a fast avalanche stage,
+ * suitable when input bits are already partially mixed
+ */
+static XXH64_hash_t XXH3_avalanche(xxh_u64 h64)
+{
+ h64 = XXH_xorshift64(h64, 37);
+ h64 *= PRIME_MX1;
+ h64 = XXH_xorshift64(h64, 32);
+ return h64;
+}
+
+/*
+ * This is a stronger avalanche,
+ * inspired by Pelle Evensen's rrmxmx
+ * preferable when input has not been previously mixed
+ */
+static XXH64_hash_t XXH3_rrmxmx(xxh_u64 h64, xxh_u64 len)
+{
+ /* this mix is inspired by Pelle Evensen's rrmxmx */
+ h64 ^= XXH_rotl64(h64, 49) ^ XXH_rotl64(h64, 24);
+ h64 *= PRIME_MX2;
+ h64 ^= (h64 >> 35) + len ;
+ h64 *= PRIME_MX2;
+ return XXH_xorshift64(h64, 28);
+}
+
+
+/* ==========================================
+ * Short keys
+ * ==========================================
+ * One of the shortcomings of XXH32 and XXH64 was that their performance was
+ * sub-optimal on short lengths. It used an iterative algorithm which strongly
+ * favored lengths that were a multiple of 4 or 8.
+ *
+ * Instead of iterating over individual inputs, we use a set of single shot
+ * functions which piece together a range of lengths and operate in constant time.
+ *
+ * Additionally, the number of multiplies has been significantly reduced. This
+ * reduces latency, especially when emulating 64-bit multiplies on 32-bit.
+ *
+ * Depending on the platform, this may or may not be faster than XXH32, but it
+ * is almost guaranteed to be faster than XXH64.
+ */
+
+/*
+ * At very short lengths, there isn't enough input to fully hide secrets, or use
+ * the entire secret.
+ *
+ * There is also only a limited amount of mixing we can do before significantly
+ * impacting performance.
+ *
+ * Therefore, we use different sections of the secret and always mix two secret
+ * samples with an XOR. This should have no effect on performance on the
+ * seedless or withSeed variants because everything _should_ be constant folded
+ * by modern compilers.
+ *
+ * The XOR mixing hides individual parts of the secret and increases entropy.
+ *
+ * This adds an extra layer of strength for custom secrets.
+ */
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_1to3_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+ XXH_ASSERT(input != NULL);
+ XXH_ASSERT(1 <= len && len <= 3);
+ XXH_ASSERT(secret != NULL);
+ /*
+ * len = 1: combined = { input[0], 0x01, input[0], input[0] }
+ * len = 2: combined = { input[1], 0x02, input[0], input[1] }
+ * len = 3: combined = { input[2], 0x03, input[0], input[1] }
+ */
+ { xxh_u8 const c1 = input[0];
+ xxh_u8 const c2 = input[len >> 1];
+ xxh_u8 const c3 = input[len - 1];
+ xxh_u32 const combined = ((xxh_u32)c1 << 16) | ((xxh_u32)c2 << 24)
+ | ((xxh_u32)c3 << 0) | ((xxh_u32)len << 8);
+ xxh_u64 const bitflip = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed;
+ xxh_u64 const keyed = (xxh_u64)combined ^ bitflip;
+ return XXH64_avalanche(keyed);
+ }
+}
+
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_4to8_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+ XXH_ASSERT(input != NULL);
+ XXH_ASSERT(secret != NULL);
+ XXH_ASSERT(4 <= len && len <= 8);
+ seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32;
+ { xxh_u32 const input1 = XXH_readLE32(input);
+ xxh_u32 const input2 = XXH_readLE32(input + len - 4);
+ xxh_u64 const bitflip = (XXH_readLE64(secret+8) ^ XXH_readLE64(secret+16)) - seed;
+ xxh_u64 const input64 = input2 + (((xxh_u64)input1) << 32);
+ xxh_u64 const keyed = input64 ^ bitflip;
+ return XXH3_rrmxmx(keyed, len);
+ }
+}
+
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_9to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+ XXH_ASSERT(input != NULL);
+ XXH_ASSERT(secret != NULL);
+ XXH_ASSERT(9 <= len && len <= 16);
+ { xxh_u64 const bitflip1 = (XXH_readLE64(secret+24) ^ XXH_readLE64(secret+32)) + seed;
+ xxh_u64 const bitflip2 = (XXH_readLE64(secret+40) ^ XXH_readLE64(secret+48)) - seed;
+ xxh_u64 const input_lo = XXH_readLE64(input) ^ bitflip1;
+ xxh_u64 const input_hi = XXH_readLE64(input + len - 8) ^ bitflip2;
+ xxh_u64 const acc = len
+ + XXH_swap64(input_lo) + input_hi
+ + XXH3_mul128_fold64(input_lo, input_hi);
+ return XXH3_avalanche(acc);
+ }
+}
+
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_0to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+ XXH_ASSERT(len <= 16);
+ { if (XXH_likely(len > 8)) return XXH3_len_9to16_64b(input, len, secret, seed);
+ if (XXH_likely(len >= 4)) return XXH3_len_4to8_64b(input, len, secret, seed);
+ if (len) return XXH3_len_1to3_64b(input, len, secret, seed);
+ return XXH64_avalanche(seed ^ (XXH_readLE64(secret+56) ^ XXH_readLE64(secret+64)));
+ }
+}
+
+/*
+ * DISCLAIMER: There are known *seed-dependent* multicollisions here due to
+ * multiplication by zero, affecting hashes of lengths 17 to 240.
+ *
+ * However, they are very unlikely.
+ *
+ * Keep this in mind when using the unseeded XXH3_64bits() variant: As with all
+ * unseeded non-cryptographic hashes, it does not attempt to defend itself
+ * against specially crafted inputs, only random inputs.
+ *
+ * Compared to classic UMAC where a 1 in 2^31 chance of 4 consecutive bytes
+ * cancelling out the secret is taken an arbitrary number of times (addressed
+ * in XXH3_accumulate_512), this collision is very unlikely with random inputs
+ * and/or proper seeding:
+ *
+ * This only has a 1 in 2^63 chance of 8 consecutive bytes cancelling out, in a
+ * function that is only called up to 16 times per hash with up to 240 bytes of
+ * input.
+ *
+ * This is not too bad for a non-cryptographic hash function, especially with
+ * only 64 bit outputs.
+ *
+ * The 128-bit variant (which trades some speed for strength) is NOT affected
+ * by this, although it is always a good idea to use a proper seed if you care
+ * about strength.
+ */
+XXH_FORCE_INLINE xxh_u64 XXH3_mix16B(const xxh_u8* XXH_RESTRICT input,
+ const xxh_u8* XXH_RESTRICT secret, xxh_u64 seed64)
+{
+#if defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
+ && defined(__i386__) && defined(__SSE2__) /* x86 + SSE2 */ \
+ && !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable like XXH32 hack */
+ /*
+ * UGLY HACK:
+ * GCC for x86 tends to autovectorize the 128-bit multiply, resulting in
+ * slower code.
+ *
+ * By forcing seed64 into a register, we disrupt the cost model and
+ * cause it to scalarize. See `XXH32_round()`
+ *
+ * FIXME: Clang's output is still _much_ faster -- On an AMD Ryzen 3600,
+ * XXH3_64bits @ len=240 runs at 4.6 GB/s with Clang 9, but 3.3 GB/s on
+ * GCC 9.2, despite both emitting scalar code.
+ *
+ * GCC generates much better scalar code than Clang for the rest of XXH3,
+ * which is why finding a more optimal codepath is an interest.
+ */
+ XXH_COMPILER_GUARD(seed64);
+#endif
+ { xxh_u64 const input_lo = XXH_readLE64(input);
+ xxh_u64 const input_hi = XXH_readLE64(input+8);
+ return XXH3_mul128_fold64(
+ input_lo ^ (XXH_readLE64(secret) + seed64),
+ input_hi ^ (XXH_readLE64(secret+8) - seed64)
+ );
+ }
+}
+
+/* For mid range keys, XXH3 uses a Mum-hash variant. */
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_17to128_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
+ const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+ XXH64_hash_t seed)
+{
+ XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+ XXH_ASSERT(16 < len && len <= 128);
+
+ { xxh_u64 acc = len * XXH_PRIME64_1;
+#if XXH_SIZE_OPT >= 1
+ /* Smaller and cleaner, but slightly slower. */
+ unsigned int i = (unsigned int)(len - 1) / 32;
+ do {
+ acc += XXH3_mix16B(input+16 * i, secret+32*i, seed);
+ acc += XXH3_mix16B(input+len-16*(i+1), secret+32*i+16, seed);
+ } while (i-- != 0);
+#else
+ if (len > 32) {
+ if (len > 64) {
+ if (len > 96) {
+ acc += XXH3_mix16B(input+48, secret+96, seed);
+ acc += XXH3_mix16B(input+len-64, secret+112, seed);
+ }
+ acc += XXH3_mix16B(input+32, secret+64, seed);
+ acc += XXH3_mix16B(input+len-48, secret+80, seed);
+ }
+ acc += XXH3_mix16B(input+16, secret+32, seed);
+ acc += XXH3_mix16B(input+len-32, secret+48, seed);
+ }
+ acc += XXH3_mix16B(input+0, secret+0, seed);
+ acc += XXH3_mix16B(input+len-16, secret+16, seed);
+#endif
+ return XXH3_avalanche(acc);
+ }
+}
+
+/*!
+ * @brief Maximum size of "short" key in bytes.
+ */
+#define XXH3_MIDSIZE_MAX 240
+
+XXH_NO_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
+ const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+ XXH64_hash_t seed)
+{
+ XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+ XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
+
+ #define XXH3_MIDSIZE_STARTOFFSET 3
+ #define XXH3_MIDSIZE_LASTOFFSET 17
+
+ { xxh_u64 acc = len * XXH_PRIME64_1;
+ xxh_u64 acc_end;
+ unsigned int const nbRounds = (unsigned int)len / 16;
+ unsigned int i;
+ XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
+ for (i=0; i<8; i++) {
+ acc += XXH3_mix16B(input+(16*i), secret+(16*i), seed);
+ }
+ /* last bytes */
+ acc_end = XXH3_mix16B(input + len - 16, secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, seed);
+ XXH_ASSERT(nbRounds >= 8);
+ acc = XXH3_avalanche(acc);
+#if defined(__clang__) /* Clang */ \
+ && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \
+ && !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable */
+ /*
+ * UGLY HACK:
+ * Clang for ARMv7-A tries to vectorize this loop, similar to GCC x86.
+ * In everywhere else, it uses scalar code.
+ *
+ * For 64->128-bit multiplies, even if the NEON was 100% optimal, it
+ * would still be slower than UMAAL (see XXH_mult64to128).
+ *
+ * Unfortunately, Clang doesn't handle the long multiplies properly and
+ * converts them to the nonexistent "vmulq_u64" intrinsic, which is then
+ * scalarized into an ugly mess of VMOV.32 instructions.
+ *
+ * This mess is difficult to avoid without turning autovectorization
+ * off completely, but they are usually relatively minor and/or not
+ * worth it to fix.
+ *
+ * This loop is the easiest to fix, as unlike XXH32, this pragma
+ * _actually works_ because it is a loop vectorization instead of an
+ * SLP vectorization.
+ */
+ #pragma clang loop vectorize(disable)
+#endif
+ for (i=8 ; i < nbRounds; i++) {
+ /*
+ * Prevents clang for unrolling the acc loop and interleaving with this one.
+ */
+ XXH_COMPILER_GUARD(acc);
+ acc_end += XXH3_mix16B(input+(16*i), secret+(16*(i-8)) + XXH3_MIDSIZE_STARTOFFSET, seed);
+ }
+ return XXH3_avalanche(acc + acc_end);
+ }
+}
+
+
+/* ======= Long Keys ======= */
+
+#define XXH_STRIPE_LEN 64
+#define XXH_SECRET_CONSUME_RATE 8 /* nb of secret bytes consumed at each accumulation */
+#define XXH_ACC_NB (XXH_STRIPE_LEN / sizeof(xxh_u64))
+
+#ifdef XXH_OLD_NAMES
+# define STRIPE_LEN XXH_STRIPE_LEN
+# define ACC_NB XXH_ACC_NB
+#endif
+
+#ifndef XXH_PREFETCH_DIST
+# ifdef __clang__
+# define XXH_PREFETCH_DIST 320
+# else
+# if (XXH_VECTOR == XXH_AVX512)
+# define XXH_PREFETCH_DIST 512
+# else
+# define XXH_PREFETCH_DIST 384
+# endif
+# endif /* __clang__ */
+#endif /* XXH_PREFETCH_DIST */
+
+/*
+ * These macros are to generate an XXH3_accumulate() function.
+ * The two arguments select the name suffix and target attribute.
+ *
+ * The name of this symbol is XXH3_accumulate_<name>() and it calls
+ * XXH3_accumulate_512_<name>().
+ *
+ * It may be useful to hand implement this function if the compiler fails to
+ * optimize the inline function.
+ */
+#define XXH3_ACCUMULATE_TEMPLATE(name) \
+void \
+XXH3_accumulate_##name(xxh_u64* XXH_RESTRICT acc, \
+ const xxh_u8* XXH_RESTRICT input, \
+ const xxh_u8* XXH_RESTRICT secret, \
+ size_t nbStripes) \
+{ \
+ size_t n; \
+ for (n = 0; n < nbStripes; n++ ) { \
+ const xxh_u8* const in = input + n*XXH_STRIPE_LEN; \
+ XXH_PREFETCH(in + XXH_PREFETCH_DIST); \
+ XXH3_accumulate_512_##name( \
+ acc, \
+ in, \
+ secret + n*XXH_SECRET_CONSUME_RATE); \
+ } \
+}
+
+
+XXH_FORCE_INLINE void XXH_writeLE64(void* dst, xxh_u64 v64)
+{
+ if (!XXH_CPU_LITTLE_ENDIAN) v64 = XXH_swap64(v64);
+ XXH_memcpy(dst, &v64, sizeof(v64));
+}
+
+/* Several intrinsic functions below are supposed to accept __int64 as argument,
+ * as documented in https://software.intel.com/sites/landingpage/IntrinsicsGuide/ .
+ * However, several environments do not define __int64 type,
+ * requiring a workaround.
+ */
+#if !defined (__VMS) \
+ && (defined (__cplusplus) \
+ || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+ typedef int64_t xxh_i64;
+#else
+ /* the following type must have a width of 64-bit */
+ typedef long long xxh_i64;
+#endif
+
+
+/*
+ * XXH3_accumulate_512 is the tightest loop for long inputs, and it is the most optimized.
+ *
+ * It is a hardened version of UMAC, based off of FARSH's implementation.
+ *
+ * This was chosen because it adapts quite well to 32-bit, 64-bit, and SIMD
+ * implementations, and it is ridiculously fast.
+ *
+ * We harden it by mixing the original input to the accumulators as well as the product.
+ *
+ * This means that in the (relatively likely) case of a multiply by zero, the
+ * original input is preserved.
+ *
+ * On 128-bit inputs, we swap 64-bit pairs when we add the input to improve
+ * cross-pollination, as otherwise the upper and lower halves would be
+ * essentially independent.
+ *
+ * This doesn't matter on 64-bit hashes since they all get merged together in
+ * the end, so we skip the extra step.
+ *
+ * Both XXH3_64bits and XXH3_128bits use this subroutine.
+ */
+
+#if (XXH_VECTOR == XXH_AVX512) \
+ || (defined(XXH_DISPATCH_AVX512) && XXH_DISPATCH_AVX512 != 0)
+
+#ifndef XXH_TARGET_AVX512
+# define XXH_TARGET_AVX512 /* disable attribute target */
+#endif
+
+XXH_FORCE_INLINE XXH_TARGET_AVX512 void
+XXH3_accumulate_512_avx512(void* XXH_RESTRICT acc,
+ const void* XXH_RESTRICT input,
+ const void* XXH_RESTRICT secret)
+{
+ __m512i* const xacc = (__m512i *) acc;
+ XXH_ASSERT((((size_t)acc) & 63) == 0);
+ XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i));
+
+ {
+ /* data_vec = input[0]; */
+ __m512i const data_vec = _mm512_loadu_si512 (input);
+ /* key_vec = secret[0]; */
+ __m512i const key_vec = _mm512_loadu_si512 (secret);
+ /* data_key = data_vec ^ key_vec; */
+ __m512i const data_key = _mm512_xor_si512 (data_vec, key_vec);
+ /* data_key_lo = data_key >> 32; */
+ __m512i const data_key_lo = _mm512_srli_epi64 (data_key, 32);
+ /* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
+ __m512i const product = _mm512_mul_epu32 (data_key, data_key_lo);
+ /* xacc[0] += swap(data_vec); */
+ __m512i const data_swap = _mm512_shuffle_epi32(data_vec, (_MM_PERM_ENUM)_MM_SHUFFLE(1, 0, 3, 2));
+ __m512i const sum = _mm512_add_epi64(*xacc, data_swap);
+ /* xacc[0] += product; */
+ *xacc = _mm512_add_epi64(product, sum);
+ }
+}
+XXH_FORCE_INLINE XXH_TARGET_AVX512 XXH3_ACCUMULATE_TEMPLATE(avx512)
+
+/*
+ * XXH3_scrambleAcc: Scrambles the accumulators to improve mixing.
+ *
+ * Multiplication isn't perfect, as explained by Google in HighwayHash:
+ *
+ * // Multiplication mixes/scrambles bytes 0-7 of the 64-bit result to
+ * // varying degrees. In descending order of goodness, bytes
+ * // 3 4 2 5 1 6 0 7 have quality 228 224 164 160 100 96 36 32.
+ * // As expected, the upper and lower bytes are much worse.
+ *
+ * Source: https://github.com/google/highwayhash/blob/0aaf66b/highwayhash/hh_avx2.h#L291
+ *
+ * Since our algorithm uses a pseudorandom secret to add some variance into the
+ * mix, we don't need to (or want to) mix as often or as much as HighwayHash does.
+ *
+ * This isn't as tight as XXH3_accumulate, but still written in SIMD to avoid
+ * extraction.
+ *
+ * Both XXH3_64bits and XXH3_128bits use this subroutine.
+ */
+
+XXH_FORCE_INLINE XXH_TARGET_AVX512 void
+XXH3_scrambleAcc_avx512(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+ XXH_ASSERT((((size_t)acc) & 63) == 0);
+ XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i));
+ { __m512i* const xacc = (__m512i*) acc;
+ const __m512i prime32 = _mm512_set1_epi32((int)XXH_PRIME32_1);
+
+ /* xacc[0] ^= (xacc[0] >> 47) */
+ __m512i const acc_vec = *xacc;
+ __m512i const shifted = _mm512_srli_epi64 (acc_vec, 47);
+ /* xacc[0] ^= secret; */
+ __m512i const key_vec = _mm512_loadu_si512 (secret);
+ __m512i const data_key = _mm512_ternarylogic_epi32(key_vec, acc_vec, shifted, 0x96 /* key_vec ^ acc_vec ^ shifted */);
+
+ /* xacc[0] *= XXH_PRIME32_1; */
+ __m512i const data_key_hi = _mm512_srli_epi64 (data_key, 32);
+ __m512i const prod_lo = _mm512_mul_epu32 (data_key, prime32);
+ __m512i const prod_hi = _mm512_mul_epu32 (data_key_hi, prime32);
+ *xacc = _mm512_add_epi64(prod_lo, _mm512_slli_epi64(prod_hi, 32));
+ }
+}
+
+XXH_FORCE_INLINE XXH_TARGET_AVX512 void
+XXH3_initCustomSecret_avx512(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+ XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 63) == 0);
+ XXH_STATIC_ASSERT(XXH_SEC_ALIGN == 64);
+ XXH_ASSERT(((size_t)customSecret & 63) == 0);
+ (void)(&XXH_writeLE64);
+ { int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m512i);
+ __m512i const seed_pos = _mm512_set1_epi64((xxh_i64)seed64);
+ __m512i const seed = _mm512_mask_sub_epi64(seed_pos, 0xAA, _mm512_set1_epi8(0), seed_pos);
+
+ const __m512i* const src = (const __m512i*) ((const void*) XXH3_kSecret);
+ __m512i* const dest = ( __m512i*) customSecret;
+ int i;
+ XXH_ASSERT(((size_t)src & 63) == 0); /* control alignment */
+ XXH_ASSERT(((size_t)dest & 63) == 0);
+ for (i=0; i < nbRounds; ++i) {
+ dest[i] = _mm512_add_epi64(_mm512_load_si512(src + i), seed);
+ } }
+}
+
+#endif
+
+#if (XXH_VECTOR == XXH_AVX2) \
+ || (defined(XXH_DISPATCH_AVX2) && XXH_DISPATCH_AVX2 != 0)
+
+#ifndef XXH_TARGET_AVX2
+# define XXH_TARGET_AVX2 /* disable attribute target */
+#endif
+
+XXH_FORCE_INLINE XXH_TARGET_AVX2 void
+XXH3_accumulate_512_avx2( void* XXH_RESTRICT acc,
+ const void* XXH_RESTRICT input,
+ const void* XXH_RESTRICT secret)
+{
+ XXH_ASSERT((((size_t)acc) & 31) == 0);
+ { __m256i* const xacc = (__m256i *) acc;
+ /* Unaligned. This is mainly for pointer arithmetic, and because
+ * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
+ const __m256i* const xinput = (const __m256i *) input;
+ /* Unaligned. This is mainly for pointer arithmetic, and because
+ * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
+ const __m256i* const xsecret = (const __m256i *) secret;
+
+ size_t i;
+ for (i=0; i < XXH_STRIPE_LEN/sizeof(__m256i); i++) {
+ /* data_vec = xinput[i]; */
+ __m256i const data_vec = _mm256_loadu_si256 (xinput+i);
+ /* key_vec = xsecret[i]; */
+ __m256i const key_vec = _mm256_loadu_si256 (xsecret+i);
+ /* data_key = data_vec ^ key_vec; */
+ __m256i const data_key = _mm256_xor_si256 (data_vec, key_vec);
+ /* data_key_lo = data_key >> 32; */
+ __m256i const data_key_lo = _mm256_srli_epi64 (data_key, 32);
+ /* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
+ __m256i const product = _mm256_mul_epu32 (data_key, data_key_lo);
+ /* xacc[i] += swap(data_vec); */
+ __m256i const data_swap = _mm256_shuffle_epi32(data_vec, _MM_SHUFFLE(1, 0, 3, 2));
+ __m256i const sum = _mm256_add_epi64(xacc[i], data_swap);
+ /* xacc[i] += product; */
+ xacc[i] = _mm256_add_epi64(product, sum);
+ } }
+}
+XXH_FORCE_INLINE XXH_TARGET_AVX2 XXH3_ACCUMULATE_TEMPLATE(avx2)
+
+XXH_FORCE_INLINE XXH_TARGET_AVX2 void
+XXH3_scrambleAcc_avx2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+ XXH_ASSERT((((size_t)acc) & 31) == 0);
+ { __m256i* const xacc = (__m256i*) acc;
+ /* Unaligned. This is mainly for pointer arithmetic, and because
+ * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
+ const __m256i* const xsecret = (const __m256i *) secret;
+ const __m256i prime32 = _mm256_set1_epi32((int)XXH_PRIME32_1);
+
+ size_t i;
+ for (i=0; i < XXH_STRIPE_LEN/sizeof(__m256i); i++) {
+ /* xacc[i] ^= (xacc[i] >> 47) */
+ __m256i const acc_vec = xacc[i];
+ __m256i const shifted = _mm256_srli_epi64 (acc_vec, 47);
+ __m256i const data_vec = _mm256_xor_si256 (acc_vec, shifted);
+ /* xacc[i] ^= xsecret; */
+ __m256i const key_vec = _mm256_loadu_si256 (xsecret+i);
+ __m256i const data_key = _mm256_xor_si256 (data_vec, key_vec);
+
+ /* xacc[i] *= XXH_PRIME32_1; */
+ __m256i const data_key_hi = _mm256_srli_epi64 (data_key, 32);
+ __m256i const prod_lo = _mm256_mul_epu32 (data_key, prime32);
+ __m256i const prod_hi = _mm256_mul_epu32 (data_key_hi, prime32);
+ xacc[i] = _mm256_add_epi64(prod_lo, _mm256_slli_epi64(prod_hi, 32));
+ }
+ }
+}
+
+XXH_FORCE_INLINE XXH_TARGET_AVX2 void XXH3_initCustomSecret_avx2(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+ XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 31) == 0);
+ XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE / sizeof(__m256i)) == 6);
+ XXH_STATIC_ASSERT(XXH_SEC_ALIGN <= 64);
+ (void)(&XXH_writeLE64);
+ XXH_PREFETCH(customSecret);
+ { __m256i const seed = _mm256_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64, (xxh_i64)(0U - seed64), (xxh_i64)seed64);
+
+ const __m256i* const src = (const __m256i*) ((const void*) XXH3_kSecret);
+ __m256i* dest = ( __m256i*) customSecret;
+
+# if defined(__GNUC__) || defined(__clang__)
+ /*
+ * On GCC & Clang, marking 'dest' as modified will cause the compiler:
+ * - do not extract the secret from sse registers in the internal loop
+ * - use less common registers, and avoid pushing these reg into stack
+ */
+ XXH_COMPILER_GUARD(dest);
+# endif
+ XXH_ASSERT(((size_t)src & 31) == 0); /* control alignment */
+ XXH_ASSERT(((size_t)dest & 31) == 0);
+
+ /* GCC -O2 need unroll loop manually */
+ dest[0] = _mm256_add_epi64(_mm256_load_si256(src+0), seed);
+ dest[1] = _mm256_add_epi64(_mm256_load_si256(src+1), seed);
+ dest[2] = _mm256_add_epi64(_mm256_load_si256(src+2), seed);
+ dest[3] = _mm256_add_epi64(_mm256_load_si256(src+3), seed);
+ dest[4] = _mm256_add_epi64(_mm256_load_si256(src+4), seed);
+ dest[5] = _mm256_add_epi64(_mm256_load_si256(src+5), seed);
+ }
+}
+
+#endif
+
+/* x86dispatch always generates SSE2 */
+#if (XXH_VECTOR == XXH_SSE2) || defined(XXH_X86DISPATCH)
+
+#ifndef XXH_TARGET_SSE2
+# define XXH_TARGET_SSE2 /* disable attribute target */
+#endif
+
+XXH_FORCE_INLINE XXH_TARGET_SSE2 void
+XXH3_accumulate_512_sse2( void* XXH_RESTRICT acc,
+ const void* XXH_RESTRICT input,
+ const void* XXH_RESTRICT secret)
+{
+ /* SSE2 is just a half-scale version of the AVX2 version. */
+ XXH_ASSERT((((size_t)acc) & 15) == 0);
+ { __m128i* const xacc = (__m128i *) acc;
+ /* Unaligned. This is mainly for pointer arithmetic, and because
+ * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
+ const __m128i* const xinput = (const __m128i *) input;
+ /* Unaligned. This is mainly for pointer arithmetic, and because
+ * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
+ const __m128i* const xsecret = (const __m128i *) secret;
+
+ size_t i;
+ for (i=0; i < XXH_STRIPE_LEN/sizeof(__m128i); i++) {
+ /* data_vec = xinput[i]; */
+ __m128i const data_vec = _mm_loadu_si128 (xinput+i);
+ /* key_vec = xsecret[i]; */
+ __m128i const key_vec = _mm_loadu_si128 (xsecret+i);
+ /* data_key = data_vec ^ key_vec; */
+ __m128i const data_key = _mm_xor_si128 (data_vec, key_vec);
+ /* data_key_lo = data_key >> 32; */
+ __m128i const data_key_lo = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
+ /* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
+ __m128i const product = _mm_mul_epu32 (data_key, data_key_lo);
+ /* xacc[i] += swap(data_vec); */
+ __m128i const data_swap = _mm_shuffle_epi32(data_vec, _MM_SHUFFLE(1,0,3,2));
+ __m128i const sum = _mm_add_epi64(xacc[i], data_swap);
+ /* xacc[i] += product; */
+ xacc[i] = _mm_add_epi64(product, sum);
+ } }
+}
+XXH_FORCE_INLINE XXH_TARGET_SSE2 XXH3_ACCUMULATE_TEMPLATE(sse2)
+
+XXH_FORCE_INLINE XXH_TARGET_SSE2 void
+XXH3_scrambleAcc_sse2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+ XXH_ASSERT((((size_t)acc) & 15) == 0);
+ { __m128i* const xacc = (__m128i*) acc;
+ /* Unaligned. This is mainly for pointer arithmetic, and because
+ * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
+ const __m128i* const xsecret = (const __m128i *) secret;
+ const __m128i prime32 = _mm_set1_epi32((int)XXH_PRIME32_1);
+
+ size_t i;
+ for (i=0; i < XXH_STRIPE_LEN/sizeof(__m128i); i++) {
+ /* xacc[i] ^= (xacc[i] >> 47) */
+ __m128i const acc_vec = xacc[i];
+ __m128i const shifted = _mm_srli_epi64 (acc_vec, 47);
+ __m128i const data_vec = _mm_xor_si128 (acc_vec, shifted);
+ /* xacc[i] ^= xsecret[i]; */
+ __m128i const key_vec = _mm_loadu_si128 (xsecret+i);
+ __m128i const data_key = _mm_xor_si128 (data_vec, key_vec);
+
+ /* xacc[i] *= XXH_PRIME32_1; */
+ __m128i const data_key_hi = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
+ __m128i const prod_lo = _mm_mul_epu32 (data_key, prime32);
+ __m128i const prod_hi = _mm_mul_epu32 (data_key_hi, prime32);
+ xacc[i] = _mm_add_epi64(prod_lo, _mm_slli_epi64(prod_hi, 32));
+ }
+ }
+}
+
+XXH_FORCE_INLINE XXH_TARGET_SSE2 void XXH3_initCustomSecret_sse2(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+ XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0);
+ (void)(&XXH_writeLE64);
+ { int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m128i);
+
+# if defined(_MSC_VER) && defined(_M_IX86) && _MSC_VER < 1900
+ /* MSVC 32bit mode does not support _mm_set_epi64x before 2015 */
+ XXH_ALIGN(16) const xxh_i64 seed64x2[2] = { (xxh_i64)seed64, (xxh_i64)(0U - seed64) };
+ __m128i const seed = _mm_load_si128((__m128i const*)seed64x2);
+# else
+ __m128i const seed = _mm_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64);
+# endif
+ int i;
+
+ const void* const src16 = XXH3_kSecret;
+ __m128i* dst16 = (__m128i*) customSecret;
+# if defined(__GNUC__) || defined(__clang__)
+ /*
+ * On GCC & Clang, marking 'dest' as modified will cause the compiler:
+ * - do not extract the secret from sse registers in the internal loop
+ * - use less common registers, and avoid pushing these reg into stack
+ */
+ XXH_COMPILER_GUARD(dst16);
+# endif
+ XXH_ASSERT(((size_t)src16 & 15) == 0); /* control alignment */
+ XXH_ASSERT(((size_t)dst16 & 15) == 0);
+
+ for (i=0; i < nbRounds; ++i) {
+ dst16[i] = _mm_add_epi64(_mm_load_si128((const __m128i *)src16+i), seed);
+ } }
+}
+
+#endif
+
+#if (XXH_VECTOR == XXH_NEON)
+
+/* forward declarations for the scalar routines */
+XXH_FORCE_INLINE void
+XXH3_scalarRound(void* XXH_RESTRICT acc, void const* XXH_RESTRICT input,
+ void const* XXH_RESTRICT secret, size_t lane);
+
+XXH_FORCE_INLINE void
+XXH3_scalarScrambleRound(void* XXH_RESTRICT acc,
+ void const* XXH_RESTRICT secret, size_t lane);
+
+/*!
+ * @internal
+ * @brief The bulk processing loop for NEON and WASM SIMD128.
+ *
+ * The NEON code path is actually partially scalar when running on AArch64. This
+ * is to optimize the pipelining and can have up to 15% speedup depending on the
+ * CPU, and it also mitigates some GCC codegen issues.
+ *
+ * @see XXH3_NEON_LANES for configuring this and details about this optimization.
+ *
+ * NEON's 32-bit to 64-bit long multiply takes a half vector of 32-bit
+ * integers instead of the other platforms which mask full 64-bit vectors,
+ * so the setup is more complicated than just shifting right.
+ *
+ * Additionally, there is an optimization for 4 lanes at once noted below.
+ *
+ * Since, as stated, the most optimal amount of lanes for Cortexes is 6,
+ * there needs to be *three* versions of the accumulate operation used
+ * for the remaining 2 lanes.
+ *
+ * WASM's SIMD128 uses SIMDe's arm_neon.h polyfill because the intrinsics overlap
+ * nearly perfectly.
+ */
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_neon( void* XXH_RESTRICT acc,
+ const void* XXH_RESTRICT input,
+ const void* XXH_RESTRICT secret)
+{
+ XXH_ASSERT((((size_t)acc) & 15) == 0);
+ XXH_STATIC_ASSERT(XXH3_NEON_LANES > 0 && XXH3_NEON_LANES <= XXH_ACC_NB && XXH3_NEON_LANES % 2 == 0);
+ { /* GCC for darwin arm64 does not like aliasing here */
+ xxh_aliasing_uint64x2_t* const xacc = (xxh_aliasing_uint64x2_t*) acc;
+ /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. */
+ uint8_t const* xinput = (const uint8_t *) input;
+ uint8_t const* xsecret = (const uint8_t *) secret;
+
+ size_t i;
+#ifdef __wasm_simd128__
+ /*
+ * On WASM SIMD128, Clang emits direct address loads when XXH3_kSecret
+ * is constant propagated, which results in it converting it to this
+ * inside the loop:
+ *
+ * a = v128.load(XXH3_kSecret + 0 + $secret_offset, offset = 0)
+ * b = v128.load(XXH3_kSecret + 16 + $secret_offset, offset = 0)
+ * ...
+ *
+ * This requires a full 32-bit address immediate (and therefore a 6 byte
+ * instruction) as well as an add for each offset.
+ *
+ * Putting an asm guard prevents it from folding (at the cost of losing
+ * the alignment hint), and uses the free offset in `v128.load` instead
+ * of adding secret_offset each time which overall reduces code size by
+ * about a kilobyte and improves performance.
+ */
+ XXH_COMPILER_GUARD(xsecret);
+#endif
+ /* Scalar lanes use the normal scalarRound routine */
+ for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
+ XXH3_scalarRound(acc, input, secret, i);
+ }
+ i = 0;
+ /* 4 NEON lanes at a time. */
+ for (; i+1 < XXH3_NEON_LANES / 2; i+=2) {
+ /* data_vec = xinput[i]; */
+ uint64x2_t data_vec_1 = XXH_vld1q_u64(xinput + (i * 16));
+ uint64x2_t data_vec_2 = XXH_vld1q_u64(xinput + ((i+1) * 16));
+ /* key_vec = xsecret[i]; */
+ uint64x2_t key_vec_1 = XXH_vld1q_u64(xsecret + (i * 16));
+ uint64x2_t key_vec_2 = XXH_vld1q_u64(xsecret + ((i+1) * 16));
+ /* data_swap = swap(data_vec) */
+ uint64x2_t data_swap_1 = vextq_u64(data_vec_1, data_vec_1, 1);
+ uint64x2_t data_swap_2 = vextq_u64(data_vec_2, data_vec_2, 1);
+ /* data_key = data_vec ^ key_vec; */
+ uint64x2_t data_key_1 = veorq_u64(data_vec_1, key_vec_1);
+ uint64x2_t data_key_2 = veorq_u64(data_vec_2, key_vec_2);
+
+ /*
+ * If we reinterpret the 64x2 vectors as 32x4 vectors, we can use a
+ * de-interleave operation for 4 lanes in 1 step with `vuzpq_u32` to
+ * get one vector with the low 32 bits of each lane, and one vector
+ * with the high 32 bits of each lane.
+ *
+ * The intrinsic returns a double vector because the original ARMv7-a
+ * instruction modified both arguments in place. AArch64 and SIMD128 emit
+ * two instructions from this intrinsic.
+ *
+ * [ dk11L | dk11H | dk12L | dk12H ] -> [ dk11L | dk12L | dk21L | dk22L ]
+ * [ dk21L | dk21H | dk22L | dk22H ] -> [ dk11H | dk12H | dk21H | dk22H ]
+ */
+ uint32x4x2_t unzipped = vuzpq_u32(
+ vreinterpretq_u32_u64(data_key_1),
+ vreinterpretq_u32_u64(data_key_2)
+ );
+ /* data_key_lo = data_key & 0xFFFFFFFF */
+ uint32x4_t data_key_lo = unzipped.val[0];
+ /* data_key_hi = data_key >> 32 */
+ uint32x4_t data_key_hi = unzipped.val[1];
+ /*
+ * Then, we can split the vectors horizontally and multiply which, as for most
+ * widening intrinsics, have a variant that works on both high half vectors
+ * for free on AArch64. A similar instruction is available on SIMD128.
+ *
+ * sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi
+ */
+ uint64x2_t sum_1 = XXH_vmlal_low_u32(data_swap_1, data_key_lo, data_key_hi);
+ uint64x2_t sum_2 = XXH_vmlal_high_u32(data_swap_2, data_key_lo, data_key_hi);
+ /*
+ * Clang reorders
+ * a += b * c; // umlal swap.2d, dkl.2s, dkh.2s
+ * c += a; // add acc.2d, acc.2d, swap.2d
+ * to
+ * c += a; // add acc.2d, acc.2d, swap.2d
+ * c += b * c; // umlal acc.2d, dkl.2s, dkh.2s
+ *
+ * While it would make sense in theory since the addition is faster,
+ * for reasons likely related to umlal being limited to certain NEON
+ * pipelines, this is worse. A compiler guard fixes this.
+ */
+ XXH_COMPILER_GUARD_CLANG_NEON(sum_1);
+ XXH_COMPILER_GUARD_CLANG_NEON(sum_2);
+ /* xacc[i] = acc_vec + sum; */
+ xacc[i] = vaddq_u64(xacc[i], sum_1);
+ xacc[i+1] = vaddq_u64(xacc[i+1], sum_2);
+ }
+ /* Operate on the remaining NEON lanes 2 at a time. */
+ for (; i < XXH3_NEON_LANES / 2; i++) {
+ /* data_vec = xinput[i]; */
+ uint64x2_t data_vec = XXH_vld1q_u64(xinput + (i * 16));
+ /* key_vec = xsecret[i]; */
+ uint64x2_t key_vec = XXH_vld1q_u64(xsecret + (i * 16));
+ /* acc_vec_2 = swap(data_vec) */
+ uint64x2_t data_swap = vextq_u64(data_vec, data_vec, 1);
+ /* data_key = data_vec ^ key_vec; */
+ uint64x2_t data_key = veorq_u64(data_vec, key_vec);
+ /* For two lanes, just use VMOVN and VSHRN. */
+ /* data_key_lo = data_key & 0xFFFFFFFF; */
+ uint32x2_t data_key_lo = vmovn_u64(data_key);
+ /* data_key_hi = data_key >> 32; */
+ uint32x2_t data_key_hi = vshrn_n_u64(data_key, 32);
+ /* sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi; */
+ uint64x2_t sum = vmlal_u32(data_swap, data_key_lo, data_key_hi);
+ /* Same Clang workaround as before */
+ XXH_COMPILER_GUARD_CLANG_NEON(sum);
+ /* xacc[i] = acc_vec + sum; */
+ xacc[i] = vaddq_u64 (xacc[i], sum);
+ }
+ }
+}
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(neon)
+
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_neon(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+ XXH_ASSERT((((size_t)acc) & 15) == 0);
+
+ { xxh_aliasing_uint64x2_t* xacc = (xxh_aliasing_uint64x2_t*) acc;
+ uint8_t const* xsecret = (uint8_t const*) secret;
+
+ size_t i;
+ /* WASM uses operator overloads and doesn't need these. */
+#ifndef __wasm_simd128__
+ /* { prime32_1, prime32_1 } */
+ uint32x2_t const kPrimeLo = vdup_n_u32(XXH_PRIME32_1);
+ /* { 0, prime32_1, 0, prime32_1 } */
+ uint32x4_t const kPrimeHi = vreinterpretq_u32_u64(vdupq_n_u64((xxh_u64)XXH_PRIME32_1 << 32));
+#endif
+
+ /* AArch64 uses both scalar and neon at the same time */
+ for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
+ XXH3_scalarScrambleRound(acc, secret, i);
+ }
+ for (i=0; i < XXH3_NEON_LANES / 2; i++) {
+ /* xacc[i] ^= (xacc[i] >> 47); */
+ uint64x2_t acc_vec = xacc[i];
+ uint64x2_t shifted = vshrq_n_u64(acc_vec, 47);
+ uint64x2_t data_vec = veorq_u64(acc_vec, shifted);
+
+ /* xacc[i] ^= xsecret[i]; */
+ uint64x2_t key_vec = XXH_vld1q_u64(xsecret + (i * 16));
+ uint64x2_t data_key = veorq_u64(data_vec, key_vec);
+ /* xacc[i] *= XXH_PRIME32_1 */
+#ifdef __wasm_simd128__
+ /* SIMD128 has multiply by u64x2, use it instead of expanding and scalarizing */
+ xacc[i] = data_key * XXH_PRIME32_1;
+#else
+ /*
+ * Expanded version with portable NEON intrinsics
+ *
+ * lo(x) * lo(y) + (hi(x) * lo(y) << 32)
+ *
+ * prod_hi = hi(data_key) * lo(prime) << 32
+ *
+ * Since we only need 32 bits of this multiply a trick can be used, reinterpreting the vector
+ * as a uint32x4_t and multiplying by { 0, prime, 0, prime } to cancel out the unwanted bits
+ * and avoid the shift.
+ */
+ uint32x4_t prod_hi = vmulq_u32 (vreinterpretq_u32_u64(data_key), kPrimeHi);
+ /* Extract low bits for vmlal_u32 */
+ uint32x2_t data_key_lo = vmovn_u64(data_key);
+ /* xacc[i] = prod_hi + lo(data_key) * XXH_PRIME32_1; */
+ xacc[i] = vmlal_u32(vreinterpretq_u64_u32(prod_hi), data_key_lo, kPrimeLo);
+#endif
+ }
+ }
+}
+#endif
+
+#if (XXH_VECTOR == XXH_VSX)
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_vsx( void* XXH_RESTRICT acc,
+ const void* XXH_RESTRICT input,
+ const void* XXH_RESTRICT secret)
+{
+ /* presumed aligned */
+ xxh_aliasing_u64x2* const xacc = (xxh_aliasing_u64x2*) acc;
+ xxh_u8 const* const xinput = (xxh_u8 const*) input; /* no alignment restriction */
+ xxh_u8 const* const xsecret = (xxh_u8 const*) secret; /* no alignment restriction */
+ xxh_u64x2 const v32 = { 32, 32 };
+ size_t i;
+ for (i = 0; i < XXH_STRIPE_LEN / sizeof(xxh_u64x2); i++) {
+ /* data_vec = xinput[i]; */
+ xxh_u64x2 const data_vec = XXH_vec_loadu(xinput + 16*i);
+ /* key_vec = xsecret[i]; */
+ xxh_u64x2 const key_vec = XXH_vec_loadu(xsecret + 16*i);
+ xxh_u64x2 const data_key = data_vec ^ key_vec;
+ /* shuffled = (data_key << 32) | (data_key >> 32); */
+ xxh_u32x4 const shuffled = (xxh_u32x4)vec_rl(data_key, v32);
+ /* product = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)shuffled & 0xFFFFFFFF); */
+ xxh_u64x2 const product = XXH_vec_mulo((xxh_u32x4)data_key, shuffled);
+ /* acc_vec = xacc[i]; */
+ xxh_u64x2 acc_vec = xacc[i];
+ acc_vec += product;
+
+ /* swap high and low halves */
+#ifdef __s390x__
+ acc_vec += vec_permi(data_vec, data_vec, 2);
+#else
+ acc_vec += vec_xxpermdi(data_vec, data_vec, 2);
+#endif
+ xacc[i] = acc_vec;
+ }
+}
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(vsx)
+
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_vsx(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+ XXH_ASSERT((((size_t)acc) & 15) == 0);
+
+ { xxh_aliasing_u64x2* const xacc = (xxh_aliasing_u64x2*) acc;
+ const xxh_u8* const xsecret = (const xxh_u8*) secret;
+ /* constants */
+ xxh_u64x2 const v32 = { 32, 32 };
+ xxh_u64x2 const v47 = { 47, 47 };
+ xxh_u32x4 const prime = { XXH_PRIME32_1, XXH_PRIME32_1, XXH_PRIME32_1, XXH_PRIME32_1 };
+ size_t i;
+ for (i = 0; i < XXH_STRIPE_LEN / sizeof(xxh_u64x2); i++) {
+ /* xacc[i] ^= (xacc[i] >> 47); */
+ xxh_u64x2 const acc_vec = xacc[i];
+ xxh_u64x2 const data_vec = acc_vec ^ (acc_vec >> v47);
+
+ /* xacc[i] ^= xsecret[i]; */
+ xxh_u64x2 const key_vec = XXH_vec_loadu(xsecret + 16*i);
+ xxh_u64x2 const data_key = data_vec ^ key_vec;
+
+ /* xacc[i] *= XXH_PRIME32_1 */
+ /* prod_lo = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)prime & 0xFFFFFFFF); */
+ xxh_u64x2 const prod_even = XXH_vec_mule((xxh_u32x4)data_key, prime);
+ /* prod_hi = ((xxh_u64x2)data_key >> 32) * ((xxh_u64x2)prime >> 32); */
+ xxh_u64x2 const prod_odd = XXH_vec_mulo((xxh_u32x4)data_key, prime);
+ xacc[i] = prod_odd + (prod_even << v32);
+ } }
+}
+
+#endif
+
+#if (XXH_VECTOR == XXH_SVE)
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_sve( void* XXH_RESTRICT acc,
+ const void* XXH_RESTRICT input,
+ const void* XXH_RESTRICT secret)
+{
+ uint64_t *xacc = (uint64_t *)acc;
+ const uint64_t *xinput = (const uint64_t *)(const void *)input;
+ const uint64_t *xsecret = (const uint64_t *)(const void *)secret;
+ svuint64_t kSwap = sveor_n_u64_z(svptrue_b64(), svindex_u64(0, 1), 1);
+ uint64_t element_count = svcntd();
+ if (element_count >= 8) {
+ svbool_t mask = svptrue_pat_b64(SV_VL8);
+ svuint64_t vacc = svld1_u64(mask, xacc);
+ ACCRND(vacc, 0);
+ svst1_u64(mask, xacc, vacc);
+ } else if (element_count == 2) { /* sve128 */
+ svbool_t mask = svptrue_pat_b64(SV_VL2);
+ svuint64_t acc0 = svld1_u64(mask, xacc + 0);
+ svuint64_t acc1 = svld1_u64(mask, xacc + 2);
+ svuint64_t acc2 = svld1_u64(mask, xacc + 4);
+ svuint64_t acc3 = svld1_u64(mask, xacc + 6);
+ ACCRND(acc0, 0);
+ ACCRND(acc1, 2);
+ ACCRND(acc2, 4);
+ ACCRND(acc3, 6);
+ svst1_u64(mask, xacc + 0, acc0);
+ svst1_u64(mask, xacc + 2, acc1);
+ svst1_u64(mask, xacc + 4, acc2);
+ svst1_u64(mask, xacc + 6, acc3);
+ } else {
+ svbool_t mask = svptrue_pat_b64(SV_VL4);
+ svuint64_t acc0 = svld1_u64(mask, xacc + 0);
+ svuint64_t acc1 = svld1_u64(mask, xacc + 4);
+ ACCRND(acc0, 0);
+ ACCRND(acc1, 4);
+ svst1_u64(mask, xacc + 0, acc0);
+ svst1_u64(mask, xacc + 4, acc1);
+ }
+}
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_sve(xxh_u64* XXH_RESTRICT acc,
+ const xxh_u8* XXH_RESTRICT input,
+ const xxh_u8* XXH_RESTRICT secret,
+ size_t nbStripes)
+{
+ if (nbStripes != 0) {
+ uint64_t *xacc = (uint64_t *)acc;
+ const uint64_t *xinput = (const uint64_t *)(const void *)input;
+ const uint64_t *xsecret = (const uint64_t *)(const void *)secret;
+ svuint64_t kSwap = sveor_n_u64_z(svptrue_b64(), svindex_u64(0, 1), 1);
+ uint64_t element_count = svcntd();
+ if (element_count >= 8) {
+ svbool_t mask = svptrue_pat_b64(SV_VL8);
+ svuint64_t vacc = svld1_u64(mask, xacc + 0);
+ do {
+ /* svprfd(svbool_t, void *, enum svfprop); */
+ svprfd(mask, xinput + 128, SV_PLDL1STRM);
+ ACCRND(vacc, 0);
+ xinput += 8;
+ xsecret += 1;
+ nbStripes--;
+ } while (nbStripes != 0);
+
+ svst1_u64(mask, xacc + 0, vacc);
+ } else if (element_count == 2) { /* sve128 */
+ svbool_t mask = svptrue_pat_b64(SV_VL2);
+ svuint64_t acc0 = svld1_u64(mask, xacc + 0);
+ svuint64_t acc1 = svld1_u64(mask, xacc + 2);
+ svuint64_t acc2 = svld1_u64(mask, xacc + 4);
+ svuint64_t acc3 = svld1_u64(mask, xacc + 6);
+ do {
+ svprfd(mask, xinput + 128, SV_PLDL1STRM);
+ ACCRND(acc0, 0);
+ ACCRND(acc1, 2);
+ ACCRND(acc2, 4);
+ ACCRND(acc3, 6);
+ xinput += 8;
+ xsecret += 1;
+ nbStripes--;
+ } while (nbStripes != 0);
+
+ svst1_u64(mask, xacc + 0, acc0);
+ svst1_u64(mask, xacc + 2, acc1);
+ svst1_u64(mask, xacc + 4, acc2);
+ svst1_u64(mask, xacc + 6, acc3);
+ } else {
+ svbool_t mask = svptrue_pat_b64(SV_VL4);
+ svuint64_t acc0 = svld1_u64(mask, xacc + 0);
+ svuint64_t acc1 = svld1_u64(mask, xacc + 4);
+ do {
+ svprfd(mask, xinput + 128, SV_PLDL1STRM);
+ ACCRND(acc0, 0);
+ ACCRND(acc1, 4);
+ xinput += 8;
+ xsecret += 1;
+ nbStripes--;
+ } while (nbStripes != 0);
+
+ svst1_u64(mask, xacc + 0, acc0);
+ svst1_u64(mask, xacc + 4, acc1);
+ }
+ }
+}
+
+#endif
+
+/* scalar variants - universal */
+
+#if defined(__aarch64__) && (defined(__GNUC__) || defined(__clang__))
+/*
+ * In XXH3_scalarRound(), GCC and Clang have a similar codegen issue, where they
+ * emit an excess mask and a full 64-bit multiply-add (MADD X-form).
+ *
+ * While this might not seem like much, as AArch64 is a 64-bit architecture, only
+ * big Cortex designs have a full 64-bit multiplier.
+ *
+ * On the little cores, the smaller 32-bit multiplier is used, and full 64-bit
+ * multiplies expand to 2-3 multiplies in microcode. This has a major penalty
+ * of up to 4 latency cycles and 2 stall cycles in the multiply pipeline.
+ *
+ * Thankfully, AArch64 still provides the 32-bit long multiply-add (UMADDL) which does
+ * not have this penalty and does the mask automatically.
+ */
+XXH_FORCE_INLINE xxh_u64
+XXH_mult32to64_add64(xxh_u64 lhs, xxh_u64 rhs, xxh_u64 acc)
+{
+ xxh_u64 ret;
+ /* note: %x = 64-bit register, %w = 32-bit register */
+ __asm__("umaddl %x0, %w1, %w2, %x3" : "=r" (ret) : "r" (lhs), "r" (rhs), "r" (acc));
+ return ret;
+}
+#else
+XXH_FORCE_INLINE xxh_u64
+XXH_mult32to64_add64(xxh_u64 lhs, xxh_u64 rhs, xxh_u64 acc)
+{
+ return XXH_mult32to64((xxh_u32)lhs, (xxh_u32)rhs) + acc;
+}
+#endif
+
+/*!
+ * @internal
+ * @brief Scalar round for @ref XXH3_accumulate_512_scalar().
+ *
+ * This is extracted to its own function because the NEON path uses a combination
+ * of NEON and scalar.
+ */
+XXH_FORCE_INLINE void
+XXH3_scalarRound(void* XXH_RESTRICT acc,
+ void const* XXH_RESTRICT input,
+ void const* XXH_RESTRICT secret,
+ size_t lane)
+{
+ xxh_u64* xacc = (xxh_u64*) acc;
+ xxh_u8 const* xinput = (xxh_u8 const*) input;
+ xxh_u8 const* xsecret = (xxh_u8 const*) secret;
+ XXH_ASSERT(lane < XXH_ACC_NB);
+ XXH_ASSERT(((size_t)acc & (XXH_ACC_ALIGN-1)) == 0);
+ {
+ xxh_u64 const data_val = XXH_readLE64(xinput + lane * 8);
+ xxh_u64 const data_key = data_val ^ XXH_readLE64(xsecret + lane * 8);
+ xacc[lane ^ 1] += data_val; /* swap adjacent lanes */
+ xacc[lane] = XXH_mult32to64_add64(data_key /* & 0xFFFFFFFF */, data_key >> 32, xacc[lane]);
+ }
+}
+
+/*!
+ * @internal
+ * @brief Processes a 64 byte block of data using the scalar path.
+ */
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_scalar(void* XXH_RESTRICT acc,
+ const void* XXH_RESTRICT input,
+ const void* XXH_RESTRICT secret)
+{
+ size_t i;
+ /* ARM GCC refuses to unroll this loop, resulting in a 24% slowdown on ARMv6. */
+#if defined(__GNUC__) && !defined(__clang__) \
+ && (defined(__arm__) || defined(__thumb2__)) \
+ && defined(__ARM_FEATURE_UNALIGNED) /* no unaligned access just wastes bytes */ \
+ && XXH_SIZE_OPT <= 0
+# pragma GCC unroll 8
+#endif
+ for (i=0; i < XXH_ACC_NB; i++) {
+ XXH3_scalarRound(acc, input, secret, i);
+ }
+}
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(scalar)
+
+/*!
+ * @internal
+ * @brief Scalar scramble step for @ref XXH3_scrambleAcc_scalar().
+ *
+ * This is extracted to its own function because the NEON path uses a combination
+ * of NEON and scalar.
+ */
+XXH_FORCE_INLINE void
+XXH3_scalarScrambleRound(void* XXH_RESTRICT acc,
+ void const* XXH_RESTRICT secret,
+ size_t lane)
+{
+ xxh_u64* const xacc = (xxh_u64*) acc; /* presumed aligned */
+ const xxh_u8* const xsecret = (const xxh_u8*) secret; /* no alignment restriction */
+ XXH_ASSERT((((size_t)acc) & (XXH_ACC_ALIGN-1)) == 0);
+ XXH_ASSERT(lane < XXH_ACC_NB);
+ {
+ xxh_u64 const key64 = XXH_readLE64(xsecret + lane * 8);
+ xxh_u64 acc64 = xacc[lane];
+ acc64 = XXH_xorshift64(acc64, 47);
+ acc64 ^= key64;
+ acc64 *= XXH_PRIME32_1;
+ xacc[lane] = acc64;
+ }
+}
+
+/*!
+ * @internal
+ * @brief Scrambles the accumulators after a large chunk has been read
+ */
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_scalar(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+ size_t i;
+ for (i=0; i < XXH_ACC_NB; i++) {
+ XXH3_scalarScrambleRound(acc, secret, i);
+ }
+}
+
+XXH_FORCE_INLINE void
+XXH3_initCustomSecret_scalar(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+ /*
+ * We need a separate pointer for the hack below,
+ * which requires a non-const pointer.
+ * Any decent compiler will optimize this out otherwise.
+ */
+ const xxh_u8* kSecretPtr = XXH3_kSecret;
+ XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0);
+
+#if defined(__GNUC__) && defined(__aarch64__)
+ /*
+ * UGLY HACK:
+ * GCC and Clang generate a bunch of MOV/MOVK pairs for aarch64, and they are
+ * placed sequentially, in order, at the top of the unrolled loop.
+ *
+ * While MOVK is great for generating constants (2 cycles for a 64-bit
+ * constant compared to 4 cycles for LDR), it fights for bandwidth with
+ * the arithmetic instructions.
+ *
+ * I L S
+ * MOVK
+ * MOVK
+ * MOVK
+ * MOVK
+ * ADD
+ * SUB STR
+ * STR
+ * By forcing loads from memory (as the asm line causes the compiler to assume
+ * that XXH3_kSecretPtr has been changed), the pipelines are used more
+ * efficiently:
+ * I L S
+ * LDR
+ * ADD LDR
+ * SUB STR
+ * STR
+ *
+ * See XXH3_NEON_LANES for details on the pipsline.
+ *
+ * XXH3_64bits_withSeed, len == 256, Snapdragon 835
+ * without hack: 2654.4 MB/s
+ * with hack: 3202.9 MB/s
+ */
+ XXH_COMPILER_GUARD(kSecretPtr);
+#endif
+ { int const nbRounds = XXH_SECRET_DEFAULT_SIZE / 16;
+ int i;
+ for (i=0; i < nbRounds; i++) {
+ /*
+ * The asm hack causes the compiler to assume that kSecretPtr aliases with
+ * customSecret, and on aarch64, this prevented LDP from merging two
+ * loads together for free. Putting the loads together before the stores
+ * properly generates LDP.
+ */
+ xxh_u64 lo = XXH_readLE64(kSecretPtr + 16*i) + seed64;
+ xxh_u64 hi = XXH_readLE64(kSecretPtr + 16*i + 8) - seed64;
+ XXH_writeLE64((xxh_u8*)customSecret + 16*i, lo);
+ XXH_writeLE64((xxh_u8*)customSecret + 16*i + 8, hi);
+ } }
+}
+
+
+typedef void (*XXH3_f_accumulate)(xxh_u64* XXH_RESTRICT, const xxh_u8* XXH_RESTRICT, const xxh_u8* XXH_RESTRICT, size_t);
+typedef void (*XXH3_f_scrambleAcc)(void* XXH_RESTRICT, const void*);
+typedef void (*XXH3_f_initCustomSecret)(void* XXH_RESTRICT, xxh_u64);
+
+
+#if (XXH_VECTOR == XXH_AVX512)
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_avx512
+#define XXH3_accumulate XXH3_accumulate_avx512
+#define XXH3_scrambleAcc XXH3_scrambleAcc_avx512
+#define XXH3_initCustomSecret XXH3_initCustomSecret_avx512
+
+#elif (XXH_VECTOR == XXH_AVX2)
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_avx2
+#define XXH3_accumulate XXH3_accumulate_avx2
+#define XXH3_scrambleAcc XXH3_scrambleAcc_avx2
+#define XXH3_initCustomSecret XXH3_initCustomSecret_avx2
+
+#elif (XXH_VECTOR == XXH_SSE2)
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_sse2
+#define XXH3_accumulate XXH3_accumulate_sse2
+#define XXH3_scrambleAcc XXH3_scrambleAcc_sse2
+#define XXH3_initCustomSecret XXH3_initCustomSecret_sse2
+
+#elif (XXH_VECTOR == XXH_NEON)
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_neon
+#define XXH3_accumulate XXH3_accumulate_neon
+#define XXH3_scrambleAcc XXH3_scrambleAcc_neon
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+
+#elif (XXH_VECTOR == XXH_VSX)
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_vsx
+#define XXH3_accumulate XXH3_accumulate_vsx
+#define XXH3_scrambleAcc XXH3_scrambleAcc_vsx
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+
+#elif (XXH_VECTOR == XXH_SVE)
+#define XXH3_accumulate_512 XXH3_accumulate_512_sve
+#define XXH3_accumulate XXH3_accumulate_sve
+#define XXH3_scrambleAcc XXH3_scrambleAcc_scalar
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+
+#else /* scalar */
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_scalar
+#define XXH3_accumulate XXH3_accumulate_scalar
+#define XXH3_scrambleAcc XXH3_scrambleAcc_scalar
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+
+#endif
+
+#if XXH_SIZE_OPT >= 1 /* don't do SIMD for initialization */
+# undef XXH3_initCustomSecret
+# define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+#endif
+
+XXH_FORCE_INLINE void
+XXH3_hashLong_internal_loop(xxh_u64* XXH_RESTRICT acc,
+ const xxh_u8* XXH_RESTRICT input, size_t len,
+ const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+ XXH3_f_accumulate f_acc,
+ XXH3_f_scrambleAcc f_scramble)
+{
+ size_t const nbStripesPerBlock = (secretSize - XXH_STRIPE_LEN) / XXH_SECRET_CONSUME_RATE;
+ size_t const block_len = XXH_STRIPE_LEN * nbStripesPerBlock;
+ size_t const nb_blocks = (len - 1) / block_len;
+
+ size_t n;
+
+ XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
+
+ for (n = 0; n < nb_blocks; n++) {
+ f_acc(acc, input + n*block_len, secret, nbStripesPerBlock);
+ f_scramble(acc, secret + secretSize - XXH_STRIPE_LEN);
+ }
+
+ /* last partial block */
+ XXH_ASSERT(len > XXH_STRIPE_LEN);
+ { size_t const nbStripes = ((len - 1) - (block_len * nb_blocks)) / XXH_STRIPE_LEN;
+ XXH_ASSERT(nbStripes <= (secretSize / XXH_SECRET_CONSUME_RATE));
+ f_acc(acc, input + nb_blocks*block_len, secret, nbStripes);
+
+ /* last stripe */
+ { const xxh_u8* const p = input + len - XXH_STRIPE_LEN;
+#define XXH_SECRET_LASTACC_START 7 /* not aligned on 8, last secret is different from acc & scrambler */
+ XXH3_accumulate_512(acc, p, secret + secretSize - XXH_STRIPE_LEN - XXH_SECRET_LASTACC_START);
+ } }
+}
+
+XXH_FORCE_INLINE xxh_u64
+XXH3_mix2Accs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret)
+{
+ return XXH3_mul128_fold64(
+ acc[0] ^ XXH_readLE64(secret),
+ acc[1] ^ XXH_readLE64(secret+8) );
+}
+
+static XXH64_hash_t
+XXH3_mergeAccs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret, xxh_u64 start)
+{
+ xxh_u64 result64 = start;
+ size_t i = 0;
+
+ for (i = 0; i < 4; i++) {
+ result64 += XXH3_mix2Accs(acc+2*i, secret + 16*i);
+#if defined(__clang__) /* Clang */ \
+ && (defined(__arm__) || defined(__thumb__)) /* ARMv7 */ \
+ && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \
+ && !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable */
+ /*
+ * UGLY HACK:
+ * Prevent autovectorization on Clang ARMv7-a. Exact same problem as
+ * the one in XXH3_len_129to240_64b. Speeds up shorter keys > 240b.
+ * XXH3_64bits, len == 256, Snapdragon 835:
+ * without hack: 2063.7 MB/s
+ * with hack: 2560.7 MB/s
+ */
+ XXH_COMPILER_GUARD(result64);
+#endif
+ }
+
+ return XXH3_avalanche(result64);
+}
+
+#define XXH3_INIT_ACC { XXH_PRIME32_3, XXH_PRIME64_1, XXH_PRIME64_2, XXH_PRIME64_3, \
+ XXH_PRIME64_4, XXH_PRIME32_2, XXH_PRIME64_5, XXH_PRIME32_1 }
+
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_hashLong_64b_internal(const void* XXH_RESTRICT input, size_t len,
+ const void* XXH_RESTRICT secret, size_t secretSize,
+ XXH3_f_accumulate f_acc,
+ XXH3_f_scrambleAcc f_scramble)
+{
+ XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC;
+
+ XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, f_acc, f_scramble);
+
+ /* converge into final hash */
+ XXH_STATIC_ASSERT(sizeof(acc) == 64);
+ /* do not align on 8, so that the secret is different from the accumulator */
+#define XXH_SECRET_MERGEACCS_START 11
+ XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
+ return XXH3_mergeAccs(acc, (const xxh_u8*)secret + XXH_SECRET_MERGEACCS_START, (xxh_u64)len * XXH_PRIME64_1);
+}
+
+/*
+ * It's important for performance to transmit secret's size (when it's static)
+ * so that the compiler can properly optimize the vectorized loop.
+ * This makes a big performance difference for "medium" keys (<1 KB) when using AVX instruction set.
+ * When the secret size is unknown, or on GCC 12 where the mix of NO_INLINE and FORCE_INLINE
+ * breaks -Og, this is XXH_NO_INLINE.
+ */
+XXH3_WITH_SECRET_INLINE XXH64_hash_t
+XXH3_hashLong_64b_withSecret(const void* XXH_RESTRICT input, size_t len,
+ XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
+{
+ (void)seed64;
+ return XXH3_hashLong_64b_internal(input, len, secret, secretLen, XXH3_accumulate, XXH3_scrambleAcc);
+}
+
+/*
+ * It's preferable for performance that XXH3_hashLong is not inlined,
+ * as it results in a smaller function for small data, easier to the instruction cache.
+ * Note that inside this no_inline function, we do inline the internal loop,
+ * and provide a statically defined secret size to allow optimization of vector loop.
+ */
+XXH_NO_INLINE XXH_PUREF XXH64_hash_t
+XXH3_hashLong_64b_default(const void* XXH_RESTRICT input, size_t len,
+ XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
+{
+ (void)seed64; (void)secret; (void)secretLen;
+ return XXH3_hashLong_64b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_accumulate, XXH3_scrambleAcc);
+}
+
+/*
+ * XXH3_hashLong_64b_withSeed():
+ * Generate a custom key based on alteration of default XXH3_kSecret with the seed,
+ * and then use this key for long mode hashing.
+ *
+ * This operation is decently fast but nonetheless costs a little bit of time.
+ * Try to avoid it whenever possible (typically when seed==0).
+ *
+ * It's important for performance that XXH3_hashLong is not inlined. Not sure
+ * why (uop cache maybe?), but the difference is large and easily measurable.
+ */
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_hashLong_64b_withSeed_internal(const void* input, size_t len,
+ XXH64_hash_t seed,
+ XXH3_f_accumulate f_acc,
+ XXH3_f_scrambleAcc f_scramble,
+ XXH3_f_initCustomSecret f_initSec)
+{
+#if XXH_SIZE_OPT <= 0
+ if (seed == 0)
+ return XXH3_hashLong_64b_internal(input, len,
+ XXH3_kSecret, sizeof(XXH3_kSecret),
+ f_acc, f_scramble);
+#endif
+ { XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
+ f_initSec(secret, seed);
+ return XXH3_hashLong_64b_internal(input, len, secret, sizeof(secret),
+ f_acc, f_scramble);
+ }
+}
+
+/*
+ * It's important for performance that XXH3_hashLong is not inlined.
+ */
+XXH_NO_INLINE XXH64_hash_t
+XXH3_hashLong_64b_withSeed(const void* XXH_RESTRICT input, size_t len,
+ XXH64_hash_t seed, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
+{
+ (void)secret; (void)secretLen;
+ return XXH3_hashLong_64b_withSeed_internal(input, len, seed,
+ XXH3_accumulate, XXH3_scrambleAcc, XXH3_initCustomSecret);
+}
+
+
+typedef XXH64_hash_t (*XXH3_hashLong64_f)(const void* XXH_RESTRICT, size_t,
+ XXH64_hash_t, const xxh_u8* XXH_RESTRICT, size_t);
+
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_64bits_internal(const void* XXH_RESTRICT input, size_t len,
+ XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen,
+ XXH3_hashLong64_f f_hashLong)
+{
+ XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN);
+ /*
+ * If an action is to be taken if `secretLen` condition is not respected,
+ * it should be done here.
+ * For now, it's a contract pre-condition.
+ * Adding a check and a branch here would cost performance at every hash.
+ * Also, note that function signature doesn't offer room to return an error.
+ */
+ if (len <= 16)
+ return XXH3_len_0to16_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, seed64);
+ if (len <= 128)
+ return XXH3_len_17to128_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+ if (len <= XXH3_MIDSIZE_MAX)
+ return XXH3_len_129to240_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+ return f_hashLong(input, len, seed64, (const xxh_u8*)secret, secretLen);
+}
+
+
+/* === Public entry point === */
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length)
+{
+ return XXH3_64bits_internal(input, length, 0, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_default);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSecret(XXH_NOESCAPE const void* input, size_t length, XXH_NOESCAPE const void* secret, size_t secretSize)
+{
+ return XXH3_64bits_internal(input, length, 0, secret, secretSize, XXH3_hashLong_64b_withSecret);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed)
+{
+ return XXH3_64bits_internal(input, length, seed, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_withSeed);
+}
+
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t length, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed)
+{
+ if (length <= XXH3_MIDSIZE_MAX)
+ return XXH3_64bits_internal(input, length, seed, XXH3_kSecret, sizeof(XXH3_kSecret), NULL);
+ return XXH3_hashLong_64b_withSecret(input, length, seed, (const xxh_u8*)secret, secretSize);
+}
+
+
+/* === XXH3 streaming === */
+#ifndef XXH_NO_STREAM
+/*
+ * Malloc's a pointer that is always aligned to align.
+ *
+ * This must be freed with `XXH_alignedFree()`.
+ *
+ * malloc typically guarantees 16 byte alignment on 64-bit systems and 8 byte
+ * alignment on 32-bit. This isn't enough for the 32 byte aligned loads in AVX2
+ * or on 32-bit, the 16 byte aligned loads in SSE2 and NEON.
+ *
+ * This underalignment previously caused a rather obvious crash which went
+ * completely unnoticed due to XXH3_createState() not actually being tested.
+ * Credit to RedSpah for noticing this bug.
+ *
+ * The alignment is done manually: Functions like posix_memalign or _mm_malloc
+ * are avoided: To maintain portability, we would have to write a fallback
+ * like this anyways, and besides, testing for the existence of library
+ * functions without relying on external build tools is impossible.
+ *
+ * The method is simple: Overallocate, manually align, and store the offset
+ * to the original behind the returned pointer.
+ *
+ * Align must be a power of 2 and 8 <= align <= 128.
+ */
+static XXH_MALLOCF void* XXH_alignedMalloc(size_t s, size_t align)
+{
+ XXH_ASSERT(align <= 128 && align >= 8); /* range check */
+ XXH_ASSERT((align & (align-1)) == 0); /* power of 2 */
+ XXH_ASSERT(s != 0 && s < (s + align)); /* empty/overflow */
+ { /* Overallocate to make room for manual realignment and an offset byte */
+ xxh_u8* base = (xxh_u8*)XXH_malloc(s + align);
+ if (base != NULL) {
+ /*
+ * Get the offset needed to align this pointer.
+ *
+ * Even if the returned pointer is aligned, there will always be
+ * at least one byte to store the offset to the original pointer.
+ */
+ size_t offset = align - ((size_t)base & (align - 1)); /* base % align */
+ /* Add the offset for the now-aligned pointer */
+ xxh_u8* ptr = base + offset;
+
+ XXH_ASSERT((size_t)ptr % align == 0);
+
+ /* Store the offset immediately before the returned pointer. */
+ ptr[-1] = (xxh_u8)offset;
+ return ptr;
+ }
+ return NULL;
+ }
+}
+/*
+ * Frees an aligned pointer allocated by XXH_alignedMalloc(). Don't pass
+ * normal malloc'd pointers, XXH_alignedMalloc has a specific data layout.
+ */
+static void XXH_alignedFree(void* p)
+{
+ if (p != NULL) {
+ xxh_u8* ptr = (xxh_u8*)p;
+ /* Get the offset byte we added in XXH_malloc. */
+ xxh_u8 offset = ptr[-1];
+ /* Free the original malloc'd pointer */
+ xxh_u8* base = ptr - offset;
+ XXH_free(base);
+ }
+}
+/*! @ingroup XXH3_family */
+/*!
+ * @brief Allocate an @ref XXH3_state_t.
+ *
+ * @return An allocated pointer of @ref XXH3_state_t on success.
+ * @return `NULL` on failure.
+ *
+ * @note Must be freed with XXH3_freeState().
+ */
+XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void)
+{
+ XXH3_state_t* const state = (XXH3_state_t*)XXH_alignedMalloc(sizeof(XXH3_state_t), 64);
+ if (state==NULL) return NULL;
+ XXH3_INITSTATE(state);
+ return state;
+}
+
+/*! @ingroup XXH3_family */
+/*!
+ * @brief Frees an @ref XXH3_state_t.
+ *
+ * @param statePtr A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState().
+ *
+ * @return @ref XXH_OK.
+ *
+ * @note Must be allocated with XXH3_createState().
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr)
+{
+ XXH_alignedFree(statePtr);
+ return XXH_OK;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API void
+XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state)
+{
+ XXH_memcpy(dst_state, src_state, sizeof(*dst_state));
+}
+
+static void
+XXH3_reset_internal(XXH3_state_t* statePtr,
+ XXH64_hash_t seed,
+ const void* secret, size_t secretSize)
+{
+ size_t const initStart = offsetof(XXH3_state_t, bufferedSize);
+ size_t const initLength = offsetof(XXH3_state_t, nbStripesPerBlock) - initStart;
+ XXH_ASSERT(offsetof(XXH3_state_t, nbStripesPerBlock) > initStart);
+ XXH_ASSERT(statePtr != NULL);
+ /* set members from bufferedSize to nbStripesPerBlock (excluded) to 0 */
+ memset((char*)statePtr + initStart, 0, initLength);
+ statePtr->acc[0] = XXH_PRIME32_3;
+ statePtr->acc[1] = XXH_PRIME64_1;
+ statePtr->acc[2] = XXH_PRIME64_2;
+ statePtr->acc[3] = XXH_PRIME64_3;
+ statePtr->acc[4] = XXH_PRIME64_4;
+ statePtr->acc[5] = XXH_PRIME32_2;
+ statePtr->acc[6] = XXH_PRIME64_5;
+ statePtr->acc[7] = XXH_PRIME32_1;
+ statePtr->seed = seed;
+ statePtr->useSeed = (seed != 0);
+ statePtr->extSecret = (const unsigned char*)secret;
+ XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
+ statePtr->secretLimit = secretSize - XXH_STRIPE_LEN;
+ statePtr->nbStripesPerBlock = statePtr->secretLimit / XXH_SECRET_CONSUME_RATE;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr)
+{
+ if (statePtr == NULL) return XXH_ERROR;
+ XXH3_reset_internal(statePtr, 0, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE);
+ return XXH_OK;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize)
+{
+ if (statePtr == NULL) return XXH_ERROR;
+ XXH3_reset_internal(statePtr, 0, secret, secretSize);
+ if (secret == NULL) return XXH_ERROR;
+ if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
+ return XXH_OK;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed)
+{
+ if (statePtr == NULL) return XXH_ERROR;
+ if (seed==0) return XXH3_64bits_reset(statePtr);
+ if ((seed != statePtr->seed) || (statePtr->extSecret != NULL))
+ XXH3_initCustomSecret(statePtr->customSecret, seed);
+ XXH3_reset_internal(statePtr, seed, NULL, XXH_SECRET_DEFAULT_SIZE);
+ return XXH_OK;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed64)
+{
+ if (statePtr == NULL) return XXH_ERROR;
+ if (secret == NULL) return XXH_ERROR;
+ if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
+ XXH3_reset_internal(statePtr, seed64, secret, secretSize);
+ statePtr->useSeed = 1; /* always, even if seed64==0 */
+ return XXH_OK;
+}
+
+/*!
+ * @internal
+ * @brief Processes a large input for XXH3_update() and XXH3_digest_long().
+ *
+ * Unlike XXH3_hashLong_internal_loop(), this can process data that overlaps a block.
+ *
+ * @param acc Pointer to the 8 accumulator lanes
+ * @param nbStripesSoFarPtr In/out pointer to the number of leftover stripes in the block*
+ * @param nbStripesPerBlock Number of stripes in a block
+ * @param input Input pointer
+ * @param nbStripes Number of stripes to process
+ * @param secret Secret pointer
+ * @param secretLimit Offset of the last block in @p secret
+ * @param f_acc Pointer to an XXH3_accumulate implementation
+ * @param f_scramble Pointer to an XXH3_scrambleAcc implementation
+ * @return Pointer past the end of @p input after processing
+ */
+XXH_FORCE_INLINE const xxh_u8 *
+XXH3_consumeStripes(xxh_u64* XXH_RESTRICT acc,
+ size_t* XXH_RESTRICT nbStripesSoFarPtr, size_t nbStripesPerBlock,
+ const xxh_u8* XXH_RESTRICT input, size_t nbStripes,
+ const xxh_u8* XXH_RESTRICT secret, size_t secretLimit,
+ XXH3_f_accumulate f_acc,
+ XXH3_f_scrambleAcc f_scramble)
+{
+ const xxh_u8* initialSecret = secret + *nbStripesSoFarPtr * XXH_SECRET_CONSUME_RATE;
+ /* Process full blocks */
+ if (nbStripes >= (nbStripesPerBlock - *nbStripesSoFarPtr)) {
+ /* Process the initial partial block... */
+ size_t nbStripesThisIter = nbStripesPerBlock - *nbStripesSoFarPtr;
+
+ do {
+ /* Accumulate and scramble */
+ f_acc(acc, input, initialSecret, nbStripesThisIter);
+ f_scramble(acc, secret + secretLimit);
+ input += nbStripesThisIter * XXH_STRIPE_LEN;
+ nbStripes -= nbStripesThisIter;
+ /* Then continue the loop with the full block size */
+ nbStripesThisIter = nbStripesPerBlock;
+ initialSecret = secret;
+ } while (nbStripes >= nbStripesPerBlock);
+ *nbStripesSoFarPtr = 0;
+ }
+ /* Process a partial block */
+ if (nbStripes > 0) {
+ f_acc(acc, input, initialSecret, nbStripes);
+ input += nbStripes * XXH_STRIPE_LEN;
+ *nbStripesSoFarPtr += nbStripes;
+ }
+ /* Return end pointer */
+ return input;
+}
+
+#ifndef XXH3_STREAM_USE_STACK
+# if XXH_SIZE_OPT <= 0 && !defined(__clang__) /* clang doesn't need additional stack space */
+# define XXH3_STREAM_USE_STACK 1
+# endif
+#endif
+/*
+ * Both XXH3_64bits_update and XXH3_128bits_update use this routine.
+ */
+XXH_FORCE_INLINE XXH_errorcode
+XXH3_update(XXH3_state_t* XXH_RESTRICT const state,
+ const xxh_u8* XXH_RESTRICT input, size_t len,
+ XXH3_f_accumulate f_acc,
+ XXH3_f_scrambleAcc f_scramble)
+{
+ if (input==NULL) {
+ XXH_ASSERT(len == 0);
+ return XXH_OK;
+ }
+
+ XXH_ASSERT(state != NULL);
+ { const xxh_u8* const bEnd = input + len;
+ const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
+#if defined(XXH3_STREAM_USE_STACK) && XXH3_STREAM_USE_STACK >= 1
+ /* For some reason, gcc and MSVC seem to suffer greatly
+ * when operating accumulators directly into state.
+ * Operating into stack space seems to enable proper optimization.
+ * clang, on the other hand, doesn't seem to need this trick */
+ XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[8];
+ XXH_memcpy(acc, state->acc, sizeof(acc));
+#else
+ xxh_u64* XXH_RESTRICT const acc = state->acc;
+#endif
+ state->totalLen += len;
+ XXH_ASSERT(state->bufferedSize <= XXH3_INTERNALBUFFER_SIZE);
+
+ /* small input : just fill in tmp buffer */
+ if (len <= XXH3_INTERNALBUFFER_SIZE - state->bufferedSize) {
+ XXH_memcpy(state->buffer + state->bufferedSize, input, len);
+ state->bufferedSize += (XXH32_hash_t)len;
+ return XXH_OK;
+ }
+
+ /* total input is now > XXH3_INTERNALBUFFER_SIZE */
+ #define XXH3_INTERNALBUFFER_STRIPES (XXH3_INTERNALBUFFER_SIZE / XXH_STRIPE_LEN)
+ XXH_STATIC_ASSERT(XXH3_INTERNALBUFFER_SIZE % XXH_STRIPE_LEN == 0); /* clean multiple */
+
+ /*
+ * Internal buffer is partially filled (always, except at beginning)
+ * Complete it, then consume it.
+ */
+ if (state->bufferedSize) {
+ size_t const loadSize = XXH3_INTERNALBUFFER_SIZE - state->bufferedSize;
+ XXH_memcpy(state->buffer + state->bufferedSize, input, loadSize);
+ input += loadSize;
+ XXH3_consumeStripes(acc,
+ &state->nbStripesSoFar, state->nbStripesPerBlock,
+ state->buffer, XXH3_INTERNALBUFFER_STRIPES,
+ secret, state->secretLimit,
+ f_acc, f_scramble);
+ state->bufferedSize = 0;
+ }
+ XXH_ASSERT(input < bEnd);
+ if (bEnd - input > XXH3_INTERNALBUFFER_SIZE) {
+ size_t nbStripes = (size_t)(bEnd - 1 - input) / XXH_STRIPE_LEN;
+ input = XXH3_consumeStripes(acc,
+ &state->nbStripesSoFar, state->nbStripesPerBlock,
+ input, nbStripes,
+ secret, state->secretLimit,
+ f_acc, f_scramble);
+ XXH_memcpy(state->buffer + sizeof(state->buffer) - XXH_STRIPE_LEN, input - XXH_STRIPE_LEN, XXH_STRIPE_LEN);
+
+ }
+ /* Some remaining input (always) : buffer it */
+ XXH_ASSERT(input < bEnd);
+ XXH_ASSERT(bEnd - input <= XXH3_INTERNALBUFFER_SIZE);
+ XXH_ASSERT(state->bufferedSize == 0);
+ XXH_memcpy(state->buffer, input, (size_t)(bEnd-input));
+ state->bufferedSize = (XXH32_hash_t)(bEnd-input);
+#if defined(XXH3_STREAM_USE_STACK) && XXH3_STREAM_USE_STACK >= 1
+ /* save stack accumulators into state */
+ XXH_memcpy(state->acc, acc, sizeof(acc));
+#endif
+ }
+
+ return XXH_OK;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_update(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len)
+{
+ return XXH3_update(state, (const xxh_u8*)input, len,
+ XXH3_accumulate, XXH3_scrambleAcc);
+}
+
+
+XXH_FORCE_INLINE void
+XXH3_digest_long (XXH64_hash_t* acc,
+ const XXH3_state_t* state,
+ const unsigned char* secret)
+{
+ xxh_u8 lastStripe[XXH_STRIPE_LEN];
+ const xxh_u8* lastStripePtr;
+
+ /*
+ * Digest on a local copy. This way, the state remains unaltered, and it can
+ * continue ingesting more input afterwards.
+ */
+ XXH_memcpy(acc, state->acc, sizeof(state->acc));
+ if (state->bufferedSize >= XXH_STRIPE_LEN) {
+ /* Consume remaining stripes then point to remaining data in buffer */
+ size_t const nbStripes = (state->bufferedSize - 1) / XXH_STRIPE_LEN;
+ size_t nbStripesSoFar = state->nbStripesSoFar;
+ XXH3_consumeStripes(acc,
+ &nbStripesSoFar, state->nbStripesPerBlock,
+ state->buffer, nbStripes,
+ secret, state->secretLimit,
+ XXH3_accumulate, XXH3_scrambleAcc);
+ lastStripePtr = state->buffer + state->bufferedSize - XXH_STRIPE_LEN;
+ } else { /* bufferedSize < XXH_STRIPE_LEN */
+ /* Copy to temp buffer */
+ size_t const catchupSize = XXH_STRIPE_LEN - state->bufferedSize;
+ XXH_ASSERT(state->bufferedSize > 0); /* there is always some input buffered */
+ XXH_memcpy(lastStripe, state->buffer + sizeof(state->buffer) - catchupSize, catchupSize);
+ XXH_memcpy(lastStripe + catchupSize, state->buffer, state->bufferedSize);
+ lastStripePtr = lastStripe;
+ }
+ /* Last stripe */
+ XXH3_accumulate_512(acc,
+ lastStripePtr,
+ secret + state->secretLimit - XXH_SECRET_LASTACC_START);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (XXH_NOESCAPE const XXH3_state_t* state)
+{
+ const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
+ if (state->totalLen > XXH3_MIDSIZE_MAX) {
+ XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[XXH_ACC_NB];
+ XXH3_digest_long(acc, state, secret);
+ return XXH3_mergeAccs(acc,
+ secret + XXH_SECRET_MERGEACCS_START,
+ (xxh_u64)state->totalLen * XXH_PRIME64_1);
+ }
+ /* totalLen <= XXH3_MIDSIZE_MAX: digesting a short input */
+ if (state->useSeed)
+ return XXH3_64bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed);
+ return XXH3_64bits_withSecret(state->buffer, (size_t)(state->totalLen),
+ secret, state->secretLimit + XXH_STRIPE_LEN);
+}
+#endif /* !XXH_NO_STREAM */
+
+
+/* ==========================================
+ * XXH3 128 bits (a.k.a XXH128)
+ * ==========================================
+ * XXH3's 128-bit variant has better mixing and strength than the 64-bit variant,
+ * even without counting the significantly larger output size.
+ *
+ * For example, extra steps are taken to avoid the seed-dependent collisions
+ * in 17-240 byte inputs (See XXH3_mix16B and XXH128_mix32B).
+ *
+ * This strength naturally comes at the cost of some speed, especially on short
+ * lengths. Note that longer hashes are about as fast as the 64-bit version
+ * due to it using only a slight modification of the 64-bit loop.
+ *
+ * XXH128 is also more oriented towards 64-bit machines. It is still extremely
+ * fast for a _128-bit_ hash on 32-bit (it usually clears XXH64).
+ */
+
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_1to3_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+ /* A doubled version of 1to3_64b with different constants. */
+ XXH_ASSERT(input != NULL);
+ XXH_ASSERT(1 <= len && len <= 3);
+ XXH_ASSERT(secret != NULL);
+ /*
+ * len = 1: combinedl = { input[0], 0x01, input[0], input[0] }
+ * len = 2: combinedl = { input[1], 0x02, input[0], input[1] }
+ * len = 3: combinedl = { input[2], 0x03, input[0], input[1] }
+ */
+ { xxh_u8 const c1 = input[0];
+ xxh_u8 const c2 = input[len >> 1];
+ xxh_u8 const c3 = input[len - 1];
+ xxh_u32 const combinedl = ((xxh_u32)c1 <<16) | ((xxh_u32)c2 << 24)
+ | ((xxh_u32)c3 << 0) | ((xxh_u32)len << 8);
+ xxh_u32 const combinedh = XXH_rotl32(XXH_swap32(combinedl), 13);
+ xxh_u64 const bitflipl = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed;
+ xxh_u64 const bitfliph = (XXH_readLE32(secret+8) ^ XXH_readLE32(secret+12)) - seed;
+ xxh_u64 const keyed_lo = (xxh_u64)combinedl ^ bitflipl;
+ xxh_u64 const keyed_hi = (xxh_u64)combinedh ^ bitfliph;
+ XXH128_hash_t h128;
+ h128.low64 = XXH64_avalanche(keyed_lo);
+ h128.high64 = XXH64_avalanche(keyed_hi);
+ return h128;
+ }
+}
+
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_4to8_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+ XXH_ASSERT(input != NULL);
+ XXH_ASSERT(secret != NULL);
+ XXH_ASSERT(4 <= len && len <= 8);
+ seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32;
+ { xxh_u32 const input_lo = XXH_readLE32(input);
+ xxh_u32 const input_hi = XXH_readLE32(input + len - 4);
+ xxh_u64 const input_64 = input_lo + ((xxh_u64)input_hi << 32);
+ xxh_u64 const bitflip = (XXH_readLE64(secret+16) ^ XXH_readLE64(secret+24)) + seed;
+ xxh_u64 const keyed = input_64 ^ bitflip;
+
+ /* Shift len to the left to ensure it is even, this avoids even multiplies. */
+ XXH128_hash_t m128 = XXH_mult64to128(keyed, XXH_PRIME64_1 + (len << 2));
+
+ m128.high64 += (m128.low64 << 1);
+ m128.low64 ^= (m128.high64 >> 3);
+
+ m128.low64 = XXH_xorshift64(m128.low64, 35);
+ m128.low64 *= PRIME_MX2;
+ m128.low64 = XXH_xorshift64(m128.low64, 28);
+ m128.high64 = XXH3_avalanche(m128.high64);
+ return m128;
+ }
+}
+
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_9to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+ XXH_ASSERT(input != NULL);
+ XXH_ASSERT(secret != NULL);
+ XXH_ASSERT(9 <= len && len <= 16);
+ { xxh_u64 const bitflipl = (XXH_readLE64(secret+32) ^ XXH_readLE64(secret+40)) - seed;
+ xxh_u64 const bitfliph = (XXH_readLE64(secret+48) ^ XXH_readLE64(secret+56)) + seed;
+ xxh_u64 const input_lo = XXH_readLE64(input);
+ xxh_u64 input_hi = XXH_readLE64(input + len - 8);
+ XXH128_hash_t m128 = XXH_mult64to128(input_lo ^ input_hi ^ bitflipl, XXH_PRIME64_1);
+ /*
+ * Put len in the middle of m128 to ensure that the length gets mixed to
+ * both the low and high bits in the 128x64 multiply below.
+ */
+ m128.low64 += (xxh_u64)(len - 1) << 54;
+ input_hi ^= bitfliph;
+ /*
+ * Add the high 32 bits of input_hi to the high 32 bits of m128, then
+ * add the long product of the low 32 bits of input_hi and XXH_PRIME32_2 to
+ * the high 64 bits of m128.
+ *
+ * The best approach to this operation is different on 32-bit and 64-bit.
+ */
+ if (sizeof(void *) < sizeof(xxh_u64)) { /* 32-bit */
+ /*
+ * 32-bit optimized version, which is more readable.
+ *
+ * On 32-bit, it removes an ADC and delays a dependency between the two
+ * halves of m128.high64, but it generates an extra mask on 64-bit.
+ */
+ m128.high64 += (input_hi & 0xFFFFFFFF00000000ULL) + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2);
+ } else {
+ /*
+ * 64-bit optimized (albeit more confusing) version.
+ *
+ * Uses some properties of addition and multiplication to remove the mask:
+ *
+ * Let:
+ * a = input_hi.lo = (input_hi & 0x00000000FFFFFFFF)
+ * b = input_hi.hi = (input_hi & 0xFFFFFFFF00000000)
+ * c = XXH_PRIME32_2
+ *
+ * a + (b * c)
+ * Inverse Property: x + y - x == y
+ * a + (b * (1 + c - 1))
+ * Distributive Property: x * (y + z) == (x * y) + (x * z)
+ * a + (b * 1) + (b * (c - 1))
+ * Identity Property: x * 1 == x
+ * a + b + (b * (c - 1))
+ *
+ * Substitute a, b, and c:
+ * input_hi.hi + input_hi.lo + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1))
+ *
+ * Since input_hi.hi + input_hi.lo == input_hi, we get this:
+ * input_hi + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1))
+ */
+ m128.high64 += input_hi + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2 - 1);
+ }
+ /* m128 ^= XXH_swap64(m128 >> 64); */
+ m128.low64 ^= XXH_swap64(m128.high64);
+
+ { /* 128x64 multiply: h128 = m128 * XXH_PRIME64_2; */
+ XXH128_hash_t h128 = XXH_mult64to128(m128.low64, XXH_PRIME64_2);
+ h128.high64 += m128.high64 * XXH_PRIME64_2;
+
+ h128.low64 = XXH3_avalanche(h128.low64);
+ h128.high64 = XXH3_avalanche(h128.high64);
+ return h128;
+ } }
+}
+
+/*
+ * Assumption: `secret` size is >= XXH3_SECRET_SIZE_MIN
+ */
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_0to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+ XXH_ASSERT(len <= 16);
+ { if (len > 8) return XXH3_len_9to16_128b(input, len, secret, seed);
+ if (len >= 4) return XXH3_len_4to8_128b(input, len, secret, seed);
+ if (len) return XXH3_len_1to3_128b(input, len, secret, seed);
+ { XXH128_hash_t h128;
+ xxh_u64 const bitflipl = XXH_readLE64(secret+64) ^ XXH_readLE64(secret+72);
+ xxh_u64 const bitfliph = XXH_readLE64(secret+80) ^ XXH_readLE64(secret+88);
+ h128.low64 = XXH64_avalanche(seed ^ bitflipl);
+ h128.high64 = XXH64_avalanche( seed ^ bitfliph);
+ return h128;
+ } }
+}
+
+/*
+ * A bit slower than XXH3_mix16B, but handles multiply by zero better.
+ */
+XXH_FORCE_INLINE XXH128_hash_t
+XXH128_mix32B(XXH128_hash_t acc, const xxh_u8* input_1, const xxh_u8* input_2,
+ const xxh_u8* secret, XXH64_hash_t seed)
+{
+ acc.low64 += XXH3_mix16B (input_1, secret+0, seed);
+ acc.low64 ^= XXH_readLE64(input_2) + XXH_readLE64(input_2 + 8);
+ acc.high64 += XXH3_mix16B (input_2, secret+16, seed);
+ acc.high64 ^= XXH_readLE64(input_1) + XXH_readLE64(input_1 + 8);
+ return acc;
+}
+
+
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_17to128_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
+ const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+ XXH64_hash_t seed)
+{
+ XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+ XXH_ASSERT(16 < len && len <= 128);
+
+ { XXH128_hash_t acc;
+ acc.low64 = len * XXH_PRIME64_1;
+ acc.high64 = 0;
+
+#if XXH_SIZE_OPT >= 1
+ {
+ /* Smaller, but slightly slower. */
+ unsigned int i = (unsigned int)(len - 1) / 32;
+ do {
+ acc = XXH128_mix32B(acc, input+16*i, input+len-16*(i+1), secret+32*i, seed);
+ } while (i-- != 0);
+ }
+#else
+ if (len > 32) {
+ if (len > 64) {
+ if (len > 96) {
+ acc = XXH128_mix32B(acc, input+48, input+len-64, secret+96, seed);
+ }
+ acc = XXH128_mix32B(acc, input+32, input+len-48, secret+64, seed);
+ }
+ acc = XXH128_mix32B(acc, input+16, input+len-32, secret+32, seed);
+ }
+ acc = XXH128_mix32B(acc, input, input+len-16, secret, seed);
+#endif
+ { XXH128_hash_t h128;
+ h128.low64 = acc.low64 + acc.high64;
+ h128.high64 = (acc.low64 * XXH_PRIME64_1)
+ + (acc.high64 * XXH_PRIME64_4)
+ + ((len - seed) * XXH_PRIME64_2);
+ h128.low64 = XXH3_avalanche(h128.low64);
+ h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64);
+ return h128;
+ }
+ }
+}
+
+XXH_NO_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_129to240_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
+ const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+ XXH64_hash_t seed)
+{
+ XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+ XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
+
+ { XXH128_hash_t acc;
+ unsigned i;
+ acc.low64 = len * XXH_PRIME64_1;
+ acc.high64 = 0;
+ /*
+ * We set as `i` as offset + 32. We do this so that unchanged
+ * `len` can be used as upper bound. This reaches a sweet spot
+ * where both x86 and aarch64 get simple agen and good codegen
+ * for the loop.
+ */
+ for (i = 32; i < 160; i += 32) {
+ acc = XXH128_mix32B(acc,
+ input + i - 32,
+ input + i - 16,
+ secret + i - 32,
+ seed);
+ }
+ acc.low64 = XXH3_avalanche(acc.low64);
+ acc.high64 = XXH3_avalanche(acc.high64);
+ /*
+ * NB: `i <= len` will duplicate the last 32-bytes if
+ * len % 32 was zero. This is an unfortunate necessity to keep
+ * the hash result stable.
+ */
+ for (i=160; i <= len; i += 32) {
+ acc = XXH128_mix32B(acc,
+ input + i - 32,
+ input + i - 16,
+ secret + XXH3_MIDSIZE_STARTOFFSET + i - 160,
+ seed);
+ }
+ /* last bytes */
+ acc = XXH128_mix32B(acc,
+ input + len - 16,
+ input + len - 32,
+ secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET - 16,
+ (XXH64_hash_t)0 - seed);
+
+ { XXH128_hash_t h128;
+ h128.low64 = acc.low64 + acc.high64;
+ h128.high64 = (acc.low64 * XXH_PRIME64_1)
+ + (acc.high64 * XXH_PRIME64_4)
+ + ((len - seed) * XXH_PRIME64_2);
+ h128.low64 = XXH3_avalanche(h128.low64);
+ h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64);
+ return h128;
+ }
+ }
+}
+
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_hashLong_128b_internal(const void* XXH_RESTRICT input, size_t len,
+ const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+ XXH3_f_accumulate f_acc,
+ XXH3_f_scrambleAcc f_scramble)
+{
+ XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC;
+
+ XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, secret, secretSize, f_acc, f_scramble);
+
+ /* converge into final hash */
+ XXH_STATIC_ASSERT(sizeof(acc) == 64);
+ XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
+ { XXH128_hash_t h128;
+ h128.low64 = XXH3_mergeAccs(acc,
+ secret + XXH_SECRET_MERGEACCS_START,
+ (xxh_u64)len * XXH_PRIME64_1);
+ h128.high64 = XXH3_mergeAccs(acc,
+ secret + secretSize
+ - sizeof(acc) - XXH_SECRET_MERGEACCS_START,
+ ~((xxh_u64)len * XXH_PRIME64_2));
+ return h128;
+ }
+}
+
+/*
+ * It's important for performance that XXH3_hashLong() is not inlined.
+ */
+XXH_NO_INLINE XXH_PUREF XXH128_hash_t
+XXH3_hashLong_128b_default(const void* XXH_RESTRICT input, size_t len,
+ XXH64_hash_t seed64,
+ const void* XXH_RESTRICT secret, size_t secretLen)
+{
+ (void)seed64; (void)secret; (void)secretLen;
+ return XXH3_hashLong_128b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret),
+ XXH3_accumulate, XXH3_scrambleAcc);
+}
+
+/*
+ * It's important for performance to pass @p secretLen (when it's static)
+ * to the compiler, so that it can properly optimize the vectorized loop.
+ *
+ * When the secret size is unknown, or on GCC 12 where the mix of NO_INLINE and FORCE_INLINE
+ * breaks -Og, this is XXH_NO_INLINE.
+ */
+XXH3_WITH_SECRET_INLINE XXH128_hash_t
+XXH3_hashLong_128b_withSecret(const void* XXH_RESTRICT input, size_t len,
+ XXH64_hash_t seed64,
+ const void* XXH_RESTRICT secret, size_t secretLen)
+{
+ (void)seed64;
+ return XXH3_hashLong_128b_internal(input, len, (const xxh_u8*)secret, secretLen,
+ XXH3_accumulate, XXH3_scrambleAcc);
+}
+
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_hashLong_128b_withSeed_internal(const void* XXH_RESTRICT input, size_t len,
+ XXH64_hash_t seed64,
+ XXH3_f_accumulate f_acc,
+ XXH3_f_scrambleAcc f_scramble,
+ XXH3_f_initCustomSecret f_initSec)
+{
+ if (seed64 == 0)
+ return XXH3_hashLong_128b_internal(input, len,
+ XXH3_kSecret, sizeof(XXH3_kSecret),
+ f_acc, f_scramble);
+ { XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
+ f_initSec(secret, seed64);
+ return XXH3_hashLong_128b_internal(input, len, (const xxh_u8*)secret, sizeof(secret),
+ f_acc, f_scramble);
+ }
+}
+
+/*
+ * It's important for performance that XXH3_hashLong is not inlined.
+ */
+XXH_NO_INLINE XXH128_hash_t
+XXH3_hashLong_128b_withSeed(const void* input, size_t len,
+ XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen)
+{
+ (void)secret; (void)secretLen;
+ return XXH3_hashLong_128b_withSeed_internal(input, len, seed64,
+ XXH3_accumulate, XXH3_scrambleAcc, XXH3_initCustomSecret);
+}
+
+typedef XXH128_hash_t (*XXH3_hashLong128_f)(const void* XXH_RESTRICT, size_t,
+ XXH64_hash_t, const void* XXH_RESTRICT, size_t);
+
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_128bits_internal(const void* input, size_t len,
+ XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen,
+ XXH3_hashLong128_f f_hl128)
+{
+ XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN);
+ /*
+ * If an action is to be taken if `secret` conditions are not respected,
+ * it should be done here.
+ * For now, it's a contract pre-condition.
+ * Adding a check and a branch here would cost performance at every hash.
+ */
+ if (len <= 16)
+ return XXH3_len_0to16_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, seed64);
+ if (len <= 128)
+ return XXH3_len_17to128_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+ if (len <= XXH3_MIDSIZE_MAX)
+ return XXH3_len_129to240_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+ return f_hl128(input, len, seed64, secret, secretLen);
+}
+
+
+/* === Public XXH128 API === */
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(XXH_NOESCAPE const void* input, size_t len)
+{
+ return XXH3_128bits_internal(input, len, 0,
+ XXH3_kSecret, sizeof(XXH3_kSecret),
+ XXH3_hashLong_128b_default);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH3_128bits_withSecret(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize)
+{
+ return XXH3_128bits_internal(input, len, 0,
+ (const xxh_u8*)secret, secretSize,
+ XXH3_hashLong_128b_withSecret);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH3_128bits_withSeed(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
+{
+ return XXH3_128bits_internal(input, len, seed,
+ XXH3_kSecret, sizeof(XXH3_kSecret),
+ XXH3_hashLong_128b_withSeed);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH3_128bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed)
+{
+ if (len <= XXH3_MIDSIZE_MAX)
+ return XXH3_128bits_internal(input, len, seed, XXH3_kSecret, sizeof(XXH3_kSecret), NULL);
+ return XXH3_hashLong_128b_withSecret(input, len, seed, secret, secretSize);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH128(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
+{
+ return XXH3_128bits_withSeed(input, len, seed);
+}
+
+
+/* === XXH3 128-bit streaming === */
+#ifndef XXH_NO_STREAM
+/*
+ * All initialization and update functions are identical to 64-bit streaming variant.
+ * The only difference is the finalization routine.
+ */
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr)
+{
+ return XXH3_64bits_reset(statePtr);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize)
+{
+ return XXH3_64bits_reset_withSecret(statePtr, secret, secretSize);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed)
+{
+ return XXH3_64bits_reset_withSeed(statePtr, seed);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed)
+{
+ return XXH3_64bits_reset_withSecretandSeed(statePtr, secret, secretSize, seed);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_update(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len)
+{
+ return XXH3_64bits_update(state, input, len);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (XXH_NOESCAPE const XXH3_state_t* state)
+{
+ const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
+ if (state->totalLen > XXH3_MIDSIZE_MAX) {
+ XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[XXH_ACC_NB];
+ XXH3_digest_long(acc, state, secret);
+ XXH_ASSERT(state->secretLimit + XXH_STRIPE_LEN >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
+ { XXH128_hash_t h128;
+ h128.low64 = XXH3_mergeAccs(acc,
+ secret + XXH_SECRET_MERGEACCS_START,
+ (xxh_u64)state->totalLen * XXH_PRIME64_1);
+ h128.high64 = XXH3_mergeAccs(acc,
+ secret + state->secretLimit + XXH_STRIPE_LEN
+ - sizeof(acc) - XXH_SECRET_MERGEACCS_START,
+ ~((xxh_u64)state->totalLen * XXH_PRIME64_2));
+ return h128;
+ }
+ }
+ /* len <= XXH3_MIDSIZE_MAX : short code */
+ if (state->seed)
+ return XXH3_128bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed);
+ return XXH3_128bits_withSecret(state->buffer, (size_t)(state->totalLen),
+ secret, state->secretLimit + XXH_STRIPE_LEN);
+}
+#endif /* !XXH_NO_STREAM */
+/* 128-bit utility functions */
+
+#include <string.h> /* memcmp, memcpy */
+
+/* return : 1 is equal, 0 if different */
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2)
+{
+ /* note : XXH128_hash_t is compact, it has no padding byte */
+ return !(memcmp(&h1, &h2, sizeof(h1)));
+}
+
+/* This prototype is compatible with stdlib's qsort().
+ * @return : >0 if *h128_1 > *h128_2
+ * <0 if *h128_1 < *h128_2
+ * =0 if *h128_1 == *h128_2 */
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API int XXH128_cmp(XXH_NOESCAPE const void* h128_1, XXH_NOESCAPE const void* h128_2)
+{
+ XXH128_hash_t const h1 = *(const XXH128_hash_t*)h128_1;
+ XXH128_hash_t const h2 = *(const XXH128_hash_t*)h128_2;
+ int const hcmp = (h1.high64 > h2.high64) - (h2.high64 > h1.high64);
+ /* note : bets that, in most cases, hash values are different */
+ if (hcmp) return hcmp;
+ return (h1.low64 > h2.low64) - (h2.low64 > h1.low64);
+}
+
+
+/*====== Canonical representation ======*/
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API void
+XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash)
+{
+ XXH_STATIC_ASSERT(sizeof(XXH128_canonical_t) == sizeof(XXH128_hash_t));
+ if (XXH_CPU_LITTLE_ENDIAN) {
+ hash.high64 = XXH_swap64(hash.high64);
+ hash.low64 = XXH_swap64(hash.low64);
+ }
+ XXH_memcpy(dst, &hash.high64, sizeof(hash.high64));
+ XXH_memcpy((char*)dst + sizeof(hash.high64), &hash.low64, sizeof(hash.low64));
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src)
+{
+ XXH128_hash_t h;
+ h.high64 = XXH_readBE64(src);
+ h.low64 = XXH_readBE64(src->digest + 8);
+ return h;
+}
+
+
+
+/* ==========================================
+ * Secret generators
+ * ==========================================
+ */
+#define XXH_MIN(x, y) (((x) > (y)) ? (y) : (x))
+
+XXH_FORCE_INLINE void XXH3_combine16(void* dst, XXH128_hash_t h128)
+{
+ XXH_writeLE64( dst, XXH_readLE64(dst) ^ h128.low64 );
+ XXH_writeLE64( (char*)dst+8, XXH_readLE64((char*)dst+8) ^ h128.high64 );
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_generateSecret(XXH_NOESCAPE void* secretBuffer, size_t secretSize, XXH_NOESCAPE const void* customSeed, size_t customSeedSize)
+{
+#if (XXH_DEBUGLEVEL >= 1)
+ XXH_ASSERT(secretBuffer != NULL);
+ XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
+#else
+ /* production mode, assert() are disabled */
+ if (secretBuffer == NULL) return XXH_ERROR;
+ if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
+#endif
+
+ if (customSeedSize == 0) {
+ customSeed = XXH3_kSecret;
+ customSeedSize = XXH_SECRET_DEFAULT_SIZE;
+ }
+#if (XXH_DEBUGLEVEL >= 1)
+ XXH_ASSERT(customSeed != NULL);
+#else
+ if (customSeed == NULL) return XXH_ERROR;
+#endif
+
+ /* Fill secretBuffer with a copy of customSeed - repeat as needed */
+ { size_t pos = 0;
+ while (pos < secretSize) {
+ size_t const toCopy = XXH_MIN((secretSize - pos), customSeedSize);
+ memcpy((char*)secretBuffer + pos, customSeed, toCopy);
+ pos += toCopy;
+ } }
+
+ { size_t const nbSeg16 = secretSize / 16;
+ size_t n;
+ XXH128_canonical_t scrambler;
+ XXH128_canonicalFromHash(&scrambler, XXH128(customSeed, customSeedSize, 0));
+ for (n=0; n<nbSeg16; n++) {
+ XXH128_hash_t const h128 = XXH128(&scrambler, sizeof(scrambler), n);
+ XXH3_combine16((char*)secretBuffer + n*16, h128);
+ }
+ /* last segment */
+ XXH3_combine16((char*)secretBuffer + secretSize - 16, XXH128_hashFromCanonical(&scrambler));
+ }
+ return XXH_OK;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API void
+XXH3_generateSecret_fromSeed(XXH_NOESCAPE void* secretBuffer, XXH64_hash_t seed)
+{
+ XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
+ XXH3_initCustomSecret(secret, seed);
+ XXH_ASSERT(secretBuffer != NULL);
+ memcpy(secretBuffer, secret, XXH_SECRET_DEFAULT_SIZE);
+}
+
+
+
+/* Pop our optimization override from above */
+#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
+ && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
+ && defined(__OPTIMIZE__) && XXH_SIZE_OPT <= 0 /* respect -O0 and -Os */
+# pragma GCC pop_options
+#endif
+
+#endif /* XXH_NO_LONG_LONG */
+
+#endif /* XXH_NO_XXH3 */
+
+/*!
+ * @}
+ */
+#endif /* XXH_IMPLEMENTATION */
+
+
+#if defined (__cplusplus)
+} /* extern "C" */
+#endif
diff --git a/third_party/zstd/lib/common/zstd_common.c b/third_party/zstd/lib/common/zstd_common.c
new file mode 100644
index 0000000000..3f04c22abf
--- /dev/null
+++ b/third_party/zstd/lib/common/zstd_common.c
@@ -0,0 +1,48 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+
+
+/*-*************************************
+* Dependencies
+***************************************/
+#define ZSTD_DEPS_NEED_MALLOC
+#include "error_private.h"
+#include "zstd_internal.h"
+
+
+/*-****************************************
+* Version
+******************************************/
+unsigned ZSTD_versionNumber(void) { return ZSTD_VERSION_NUMBER; }
+
+const char* ZSTD_versionString(void) { return ZSTD_VERSION_STRING; }
+
+
+/*-****************************************
+* ZSTD Error Management
+******************************************/
+#undef ZSTD_isError /* defined within zstd_internal.h */
+/*! ZSTD_isError() :
+ * tells if a return value is an error code
+ * symbol is required for external callers */
+unsigned ZSTD_isError(size_t code) { return ERR_isError(code); }
+
+/*! ZSTD_getErrorName() :
+ * provides error code string from function result (useful for debugging) */
+const char* ZSTD_getErrorName(size_t code) { return ERR_getErrorName(code); }
+
+/*! ZSTD_getError() :
+ * convert a `size_t` function result into a proper ZSTD_errorCode enum */
+ZSTD_ErrorCode ZSTD_getErrorCode(size_t code) { return ERR_getErrorCode(code); }
+
+/*! ZSTD_getErrorString() :
+ * provides error code string from enum */
+const char* ZSTD_getErrorString(ZSTD_ErrorCode code) { return ERR_getErrorString(code); }
diff --git a/third_party/zstd/lib/common/zstd_deps.h b/third_party/zstd/lib/common/zstd_deps.h
new file mode 100644
index 0000000000..4d767ae9b0
--- /dev/null
+++ b/third_party/zstd/lib/common/zstd_deps.h
@@ -0,0 +1,111 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+/* This file provides common libc dependencies that zstd requires.
+ * The purpose is to allow replacing this file with a custom implementation
+ * to compile zstd without libc support.
+ */
+
+/* Need:
+ * NULL
+ * INT_MAX
+ * UINT_MAX
+ * ZSTD_memcpy()
+ * ZSTD_memset()
+ * ZSTD_memmove()
+ */
+#ifndef ZSTD_DEPS_COMMON
+#define ZSTD_DEPS_COMMON
+
+#include <limits.h>
+#include <stddef.h>
+#include <string.h>
+
+#if defined(__GNUC__) && __GNUC__ >= 4
+# define ZSTD_memcpy(d,s,l) __builtin_memcpy((d),(s),(l))
+# define ZSTD_memmove(d,s,l) __builtin_memmove((d),(s),(l))
+# define ZSTD_memset(p,v,l) __builtin_memset((p),(v),(l))
+#else
+# define ZSTD_memcpy(d,s,l) memcpy((d),(s),(l))
+# define ZSTD_memmove(d,s,l) memmove((d),(s),(l))
+# define ZSTD_memset(p,v,l) memset((p),(v),(l))
+#endif
+
+#endif /* ZSTD_DEPS_COMMON */
+
+/* Need:
+ * ZSTD_malloc()
+ * ZSTD_free()
+ * ZSTD_calloc()
+ */
+#ifdef ZSTD_DEPS_NEED_MALLOC
+#ifndef ZSTD_DEPS_MALLOC
+#define ZSTD_DEPS_MALLOC
+
+#include <stdlib.h>
+
+#define ZSTD_malloc(s) malloc(s)
+#define ZSTD_calloc(n,s) calloc((n), (s))
+#define ZSTD_free(p) free((p))
+
+#endif /* ZSTD_DEPS_MALLOC */
+#endif /* ZSTD_DEPS_NEED_MALLOC */
+
+/*
+ * Provides 64-bit math support.
+ * Need:
+ * U64 ZSTD_div64(U64 dividend, U32 divisor)
+ */
+#ifdef ZSTD_DEPS_NEED_MATH64
+#ifndef ZSTD_DEPS_MATH64
+#define ZSTD_DEPS_MATH64
+
+#define ZSTD_div64(dividend, divisor) ((dividend) / (divisor))
+
+#endif /* ZSTD_DEPS_MATH64 */
+#endif /* ZSTD_DEPS_NEED_MATH64 */
+
+/* Need:
+ * assert()
+ */
+#ifdef ZSTD_DEPS_NEED_ASSERT
+#ifndef ZSTD_DEPS_ASSERT
+#define ZSTD_DEPS_ASSERT
+
+#include <assert.h>
+
+#endif /* ZSTD_DEPS_ASSERT */
+#endif /* ZSTD_DEPS_NEED_ASSERT */
+
+/* Need:
+ * ZSTD_DEBUG_PRINT()
+ */
+#ifdef ZSTD_DEPS_NEED_IO
+#ifndef ZSTD_DEPS_IO
+#define ZSTD_DEPS_IO
+
+#include <stdio.h>
+#define ZSTD_DEBUG_PRINT(...) fprintf(stderr, __VA_ARGS__)
+
+#endif /* ZSTD_DEPS_IO */
+#endif /* ZSTD_DEPS_NEED_IO */
+
+/* Only requested when <stdint.h> is known to be present.
+ * Need:
+ * intptr_t
+ */
+#ifdef ZSTD_DEPS_NEED_STDINT
+#ifndef ZSTD_DEPS_STDINT
+#define ZSTD_DEPS_STDINT
+
+#include <stdint.h>
+
+#endif /* ZSTD_DEPS_STDINT */
+#endif /* ZSTD_DEPS_NEED_STDINT */
diff --git a/third_party/zstd/lib/common/zstd_internal.h b/third_party/zstd/lib/common/zstd_internal.h
new file mode 100644
index 0000000000..ecb9cfba87
--- /dev/null
+++ b/third_party/zstd/lib/common/zstd_internal.h
@@ -0,0 +1,392 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_CCOMMON_H_MODULE
+#define ZSTD_CCOMMON_H_MODULE
+
+/* this module contains definitions which must be identical
+ * across compression, decompression and dictBuilder.
+ * It also contains a few functions useful to at least 2 of them
+ * and which benefit from being inlined */
+
+/*-*************************************
+* Dependencies
+***************************************/
+#include "compiler.h"
+#include "cpu.h"
+#include "mem.h"
+#include "debug.h" /* assert, DEBUGLOG, RAWLOG, g_debuglevel */
+#include "error_private.h"
+#define ZSTD_STATIC_LINKING_ONLY
+#include "../zstd.h"
+#define FSE_STATIC_LINKING_ONLY
+#include "fse.h"
+#include "huf.h"
+#ifndef XXH_STATIC_LINKING_ONLY
+# define XXH_STATIC_LINKING_ONLY /* XXH64_state_t */
+#endif
+#include "xxhash.h" /* XXH_reset, update, digest */
+#ifndef ZSTD_NO_TRACE
+# include "zstd_trace.h"
+#else
+# define ZSTD_TRACE 0
+#endif
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/* ---- static assert (debug) --- */
+#define ZSTD_STATIC_ASSERT(c) DEBUG_STATIC_ASSERT(c)
+#define ZSTD_isError ERR_isError /* for inlining */
+#define FSE_isError ERR_isError
+#define HUF_isError ERR_isError
+
+
+/*-*************************************
+* shared macros
+***************************************/
+#undef MIN
+#undef MAX
+#define MIN(a,b) ((a)<(b) ? (a) : (b))
+#define MAX(a,b) ((a)>(b) ? (a) : (b))
+#define BOUNDED(min,val,max) (MAX(min,MIN(val,max)))
+
+
+/*-*************************************
+* Common constants
+***************************************/
+#define ZSTD_OPT_NUM (1<<12)
+
+#define ZSTD_REP_NUM 3 /* number of repcodes */
+static UNUSED_ATTR const U32 repStartValue[ZSTD_REP_NUM] = { 1, 4, 8 };
+
+#define KB *(1 <<10)
+#define MB *(1 <<20)
+#define GB *(1U<<30)
+
+#define BIT7 128
+#define BIT6 64
+#define BIT5 32
+#define BIT4 16
+#define BIT1 2
+#define BIT0 1
+
+#define ZSTD_WINDOWLOG_ABSOLUTEMIN 10
+static UNUSED_ATTR const size_t ZSTD_fcs_fieldSize[4] = { 0, 2, 4, 8 };
+static UNUSED_ATTR const size_t ZSTD_did_fieldSize[4] = { 0, 1, 2, 4 };
+
+#define ZSTD_FRAMEIDSIZE 4 /* magic number size */
+
+#define ZSTD_BLOCKHEADERSIZE 3 /* C standard doesn't allow `static const` variable to be init using another `static const` variable */
+static UNUSED_ATTR const size_t ZSTD_blockHeaderSize = ZSTD_BLOCKHEADERSIZE;
+typedef enum { bt_raw, bt_rle, bt_compressed, bt_reserved } blockType_e;
+
+#define ZSTD_FRAMECHECKSUMSIZE 4
+
+#define MIN_SEQUENCES_SIZE 1 /* nbSeq==0 */
+#define MIN_CBLOCK_SIZE (1 /*litCSize*/ + 1 /* RLE or RAW */) /* for a non-null block */
+#define MIN_LITERALS_FOR_4_STREAMS 6
+
+typedef enum { set_basic, set_rle, set_compressed, set_repeat } symbolEncodingType_e;
+
+#define LONGNBSEQ 0x7F00
+
+#define MINMATCH 3
+
+#define Litbits 8
+#define LitHufLog 11
+#define MaxLit ((1<<Litbits) - 1)
+#define MaxML 52
+#define MaxLL 35
+#define DefaultMaxOff 28
+#define MaxOff 31
+#define MaxSeq MAX(MaxLL, MaxML) /* Assumption : MaxOff < MaxLL,MaxML */
+#define MLFSELog 9
+#define LLFSELog 9
+#define OffFSELog 8
+#define MaxFSELog MAX(MAX(MLFSELog, LLFSELog), OffFSELog)
+#define MaxMLBits 16
+#define MaxLLBits 16
+
+#define ZSTD_MAX_HUF_HEADER_SIZE 128 /* header + <= 127 byte tree description */
+/* Each table cannot take more than #symbols * FSELog bits */
+#define ZSTD_MAX_FSE_HEADERS_SIZE (((MaxML + 1) * MLFSELog + (MaxLL + 1) * LLFSELog + (MaxOff + 1) * OffFSELog + 7) / 8)
+
+static UNUSED_ATTR const U8 LL_bits[MaxLL+1] = {
+ 0, 0, 0, 0, 0, 0, 0, 0,
+ 0, 0, 0, 0, 0, 0, 0, 0,
+ 1, 1, 1, 1, 2, 2, 3, 3,
+ 4, 6, 7, 8, 9,10,11,12,
+ 13,14,15,16
+};
+static UNUSED_ATTR const S16 LL_defaultNorm[MaxLL+1] = {
+ 4, 3, 2, 2, 2, 2, 2, 2,
+ 2, 2, 2, 2, 2, 1, 1, 1,
+ 2, 2, 2, 2, 2, 2, 2, 2,
+ 2, 3, 2, 1, 1, 1, 1, 1,
+ -1,-1,-1,-1
+};
+#define LL_DEFAULTNORMLOG 6 /* for static allocation */
+static UNUSED_ATTR const U32 LL_defaultNormLog = LL_DEFAULTNORMLOG;
+
+static UNUSED_ATTR const U8 ML_bits[MaxML+1] = {
+ 0, 0, 0, 0, 0, 0, 0, 0,
+ 0, 0, 0, 0, 0, 0, 0, 0,
+ 0, 0, 0, 0, 0, 0, 0, 0,
+ 0, 0, 0, 0, 0, 0, 0, 0,
+ 1, 1, 1, 1, 2, 2, 3, 3,
+ 4, 4, 5, 7, 8, 9,10,11,
+ 12,13,14,15,16
+};
+static UNUSED_ATTR const S16 ML_defaultNorm[MaxML+1] = {
+ 1, 4, 3, 2, 2, 2, 2, 2,
+ 2, 1, 1, 1, 1, 1, 1, 1,
+ 1, 1, 1, 1, 1, 1, 1, 1,
+ 1, 1, 1, 1, 1, 1, 1, 1,
+ 1, 1, 1, 1, 1, 1, 1, 1,
+ 1, 1, 1, 1, 1, 1,-1,-1,
+ -1,-1,-1,-1,-1
+};
+#define ML_DEFAULTNORMLOG 6 /* for static allocation */
+static UNUSED_ATTR const U32 ML_defaultNormLog = ML_DEFAULTNORMLOG;
+
+static UNUSED_ATTR const S16 OF_defaultNorm[DefaultMaxOff+1] = {
+ 1, 1, 1, 1, 1, 1, 2, 2,
+ 2, 1, 1, 1, 1, 1, 1, 1,
+ 1, 1, 1, 1, 1, 1, 1, 1,
+ -1,-1,-1,-1,-1
+};
+#define OF_DEFAULTNORMLOG 5 /* for static allocation */
+static UNUSED_ATTR const U32 OF_defaultNormLog = OF_DEFAULTNORMLOG;
+
+
+/*-*******************************************
+* Shared functions to include for inlining
+*********************************************/
+static void ZSTD_copy8(void* dst, const void* src) {
+#if defined(ZSTD_ARCH_ARM_NEON)
+ vst1_u8((uint8_t*)dst, vld1_u8((const uint8_t*)src));
+#else
+ ZSTD_memcpy(dst, src, 8);
+#endif
+}
+#define COPY8(d,s) do { ZSTD_copy8(d,s); d+=8; s+=8; } while (0)
+
+/* Need to use memmove here since the literal buffer can now be located within
+ the dst buffer. In circumstances where the op "catches up" to where the
+ literal buffer is, there can be partial overlaps in this call on the final
+ copy if the literal is being shifted by less than 16 bytes. */
+static void ZSTD_copy16(void* dst, const void* src) {
+#if defined(ZSTD_ARCH_ARM_NEON)
+ vst1q_u8((uint8_t*)dst, vld1q_u8((const uint8_t*)src));
+#elif defined(ZSTD_ARCH_X86_SSE2)
+ _mm_storeu_si128((__m128i*)dst, _mm_loadu_si128((const __m128i*)src));
+#elif defined(__clang__)
+ ZSTD_memmove(dst, src, 16);
+#else
+ /* ZSTD_memmove is not inlined properly by gcc */
+ BYTE copy16_buf[16];
+ ZSTD_memcpy(copy16_buf, src, 16);
+ ZSTD_memcpy(dst, copy16_buf, 16);
+#endif
+}
+#define COPY16(d,s) do { ZSTD_copy16(d,s); d+=16; s+=16; } while (0)
+
+#define WILDCOPY_OVERLENGTH 32
+#define WILDCOPY_VECLEN 16
+
+typedef enum {
+ ZSTD_no_overlap,
+ ZSTD_overlap_src_before_dst
+ /* ZSTD_overlap_dst_before_src, */
+} ZSTD_overlap_e;
+
+/*! ZSTD_wildcopy() :
+ * Custom version of ZSTD_memcpy(), can over read/write up to WILDCOPY_OVERLENGTH bytes (if length==0)
+ * @param ovtype controls the overlap detection
+ * - ZSTD_no_overlap: The source and destination are guaranteed to be at least WILDCOPY_VECLEN bytes apart.
+ * - ZSTD_overlap_src_before_dst: The src and dst may overlap, but they MUST be at least 8 bytes apart.
+ * The src buffer must be before the dst buffer.
+ */
+MEM_STATIC FORCE_INLINE_ATTR
+void ZSTD_wildcopy(void* dst, const void* src, ptrdiff_t length, ZSTD_overlap_e const ovtype)
+{
+ ptrdiff_t diff = (BYTE*)dst - (const BYTE*)src;
+ const BYTE* ip = (const BYTE*)src;
+ BYTE* op = (BYTE*)dst;
+ BYTE* const oend = op + length;
+
+ if (ovtype == ZSTD_overlap_src_before_dst && diff < WILDCOPY_VECLEN) {
+ /* Handle short offset copies. */
+ do {
+ COPY8(op, ip);
+ } while (op < oend);
+ } else {
+ assert(diff >= WILDCOPY_VECLEN || diff <= -WILDCOPY_VECLEN);
+ /* Separate out the first COPY16() call because the copy length is
+ * almost certain to be short, so the branches have different
+ * probabilities. Since it is almost certain to be short, only do
+ * one COPY16() in the first call. Then, do two calls per loop since
+ * at that point it is more likely to have a high trip count.
+ */
+ ZSTD_copy16(op, ip);
+ if (16 >= length) return;
+ op += 16;
+ ip += 16;
+ do {
+ COPY16(op, ip);
+ COPY16(op, ip);
+ }
+ while (op < oend);
+ }
+}
+
+MEM_STATIC size_t ZSTD_limitCopy(void* dst, size_t dstCapacity, const void* src, size_t srcSize)
+{
+ size_t const length = MIN(dstCapacity, srcSize);
+ if (length > 0) {
+ ZSTD_memcpy(dst, src, length);
+ }
+ return length;
+}
+
+/* define "workspace is too large" as this number of times larger than needed */
+#define ZSTD_WORKSPACETOOLARGE_FACTOR 3
+
+/* when workspace is continuously too large
+ * during at least this number of times,
+ * context's memory usage is considered wasteful,
+ * because it's sized to handle a worst case scenario which rarely happens.
+ * In which case, resize it down to free some memory */
+#define ZSTD_WORKSPACETOOLARGE_MAXDURATION 128
+
+/* Controls whether the input/output buffer is buffered or stable. */
+typedef enum {
+ ZSTD_bm_buffered = 0, /* Buffer the input/output */
+ ZSTD_bm_stable = 1 /* ZSTD_inBuffer/ZSTD_outBuffer is stable */
+} ZSTD_bufferMode_e;
+
+
+/*-*******************************************
+* Private declarations
+*********************************************/
+typedef struct seqDef_s {
+ U32 offBase; /* offBase == Offset + ZSTD_REP_NUM, or repcode 1,2,3 */
+ U16 litLength;
+ U16 mlBase; /* mlBase == matchLength - MINMATCH */
+} seqDef;
+
+/* Controls whether seqStore has a single "long" litLength or matchLength. See seqStore_t. */
+typedef enum {
+ ZSTD_llt_none = 0, /* no longLengthType */
+ ZSTD_llt_literalLength = 1, /* represents a long literal */
+ ZSTD_llt_matchLength = 2 /* represents a long match */
+} ZSTD_longLengthType_e;
+
+typedef struct {
+ seqDef* sequencesStart;
+ seqDef* sequences; /* ptr to end of sequences */
+ BYTE* litStart;
+ BYTE* lit; /* ptr to end of literals */
+ BYTE* llCode;
+ BYTE* mlCode;
+ BYTE* ofCode;
+ size_t maxNbSeq;
+ size_t maxNbLit;
+
+ /* longLengthPos and longLengthType to allow us to represent either a single litLength or matchLength
+ * in the seqStore that has a value larger than U16 (if it exists). To do so, we increment
+ * the existing value of the litLength or matchLength by 0x10000.
+ */
+ ZSTD_longLengthType_e longLengthType;
+ U32 longLengthPos; /* Index of the sequence to apply long length modification to */
+} seqStore_t;
+
+typedef struct {
+ U32 litLength;
+ U32 matchLength;
+} ZSTD_sequenceLength;
+
+/**
+ * Returns the ZSTD_sequenceLength for the given sequences. It handles the decoding of long sequences
+ * indicated by longLengthPos and longLengthType, and adds MINMATCH back to matchLength.
+ */
+MEM_STATIC ZSTD_sequenceLength ZSTD_getSequenceLength(seqStore_t const* seqStore, seqDef const* seq)
+{
+ ZSTD_sequenceLength seqLen;
+ seqLen.litLength = seq->litLength;
+ seqLen.matchLength = seq->mlBase + MINMATCH;
+ if (seqStore->longLengthPos == (U32)(seq - seqStore->sequencesStart)) {
+ if (seqStore->longLengthType == ZSTD_llt_literalLength) {
+ seqLen.litLength += 0x10000;
+ }
+ if (seqStore->longLengthType == ZSTD_llt_matchLength) {
+ seqLen.matchLength += 0x10000;
+ }
+ }
+ return seqLen;
+}
+
+/**
+ * Contains the compressed frame size and an upper-bound for the decompressed frame size.
+ * Note: before using `compressedSize`, check for errors using ZSTD_isError().
+ * similarly, before using `decompressedBound`, check for errors using:
+ * `decompressedBound != ZSTD_CONTENTSIZE_ERROR`
+ */
+typedef struct {
+ size_t nbBlocks;
+ size_t compressedSize;
+ unsigned long long decompressedBound;
+} ZSTD_frameSizeInfo; /* decompress & legacy */
+
+const seqStore_t* ZSTD_getSeqStore(const ZSTD_CCtx* ctx); /* compress & dictBuilder */
+int ZSTD_seqToCodes(const seqStore_t* seqStorePtr); /* compress, dictBuilder, decodeCorpus (shouldn't get its definition from here) */
+
+
+/* ZSTD_invalidateRepCodes() :
+ * ensures next compression will not use repcodes from previous block.
+ * Note : only works with regular variant;
+ * do not use with extDict variant ! */
+void ZSTD_invalidateRepCodes(ZSTD_CCtx* cctx); /* zstdmt, adaptive_compression (shouldn't get this definition from here) */
+
+
+typedef struct {
+ blockType_e blockType;
+ U32 lastBlock;
+ U32 origSize;
+} blockProperties_t; /* declared here for decompress and fullbench */
+
+/*! ZSTD_getcBlockSize() :
+ * Provides the size of compressed block from block header `src` */
+/* Used by: decompress, fullbench */
+size_t ZSTD_getcBlockSize(const void* src, size_t srcSize,
+ blockProperties_t* bpPtr);
+
+/*! ZSTD_decodeSeqHeaders() :
+ * decode sequence header from src */
+/* Used by: zstd_decompress_block, fullbench */
+size_t ZSTD_decodeSeqHeaders(ZSTD_DCtx* dctx, int* nbSeqPtr,
+ const void* src, size_t srcSize);
+
+/**
+ * @returns true iff the CPU supports dynamic BMI2 dispatch.
+ */
+MEM_STATIC int ZSTD_cpuSupportsBmi2(void)
+{
+ ZSTD_cpuid_t cpuid = ZSTD_cpuid();
+ return ZSTD_cpuid_bmi1(cpuid) && ZSTD_cpuid_bmi2(cpuid);
+}
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* ZSTD_CCOMMON_H_MODULE */
diff --git a/third_party/zstd/lib/common/zstd_trace.h b/third_party/zstd/lib/common/zstd_trace.h
new file mode 100644
index 0000000000..da20534ebd
--- /dev/null
+++ b/third_party/zstd/lib/common/zstd_trace.h
@@ -0,0 +1,163 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_TRACE_H
+#define ZSTD_TRACE_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+#include <stddef.h>
+
+/* weak symbol support
+ * For now, enable conservatively:
+ * - Only GNUC
+ * - Only ELF
+ * - Only x86-64, i386 and aarch64
+ * Also, explicitly disable on platforms known not to work so they aren't
+ * forgotten in the future.
+ */
+#if !defined(ZSTD_HAVE_WEAK_SYMBOLS) && \
+ defined(__GNUC__) && defined(__ELF__) && \
+ (defined(__x86_64__) || defined(_M_X64) || defined(__i386__) || defined(_M_IX86) || defined(__aarch64__)) && \
+ !defined(__APPLE__) && !defined(_WIN32) && !defined(__MINGW32__) && \
+ !defined(__CYGWIN__) && !defined(_AIX)
+# define ZSTD_HAVE_WEAK_SYMBOLS 1
+#else
+# define ZSTD_HAVE_WEAK_SYMBOLS 0
+#endif
+#if ZSTD_HAVE_WEAK_SYMBOLS
+# define ZSTD_WEAK_ATTR __attribute__((__weak__))
+#else
+# define ZSTD_WEAK_ATTR
+#endif
+
+/* Only enable tracing when weak symbols are available. */
+#ifndef ZSTD_TRACE
+# define ZSTD_TRACE ZSTD_HAVE_WEAK_SYMBOLS
+#endif
+
+#if ZSTD_TRACE
+
+struct ZSTD_CCtx_s;
+struct ZSTD_DCtx_s;
+struct ZSTD_CCtx_params_s;
+
+typedef struct {
+ /**
+ * ZSTD_VERSION_NUMBER
+ *
+ * This is guaranteed to be the first member of ZSTD_trace.
+ * Otherwise, this struct is not stable between versions. If
+ * the version number does not match your expectation, you
+ * should not interpret the rest of the struct.
+ */
+ unsigned version;
+ /**
+ * Non-zero if streaming (de)compression is used.
+ */
+ unsigned streaming;
+ /**
+ * The dictionary ID.
+ */
+ unsigned dictionaryID;
+ /**
+ * Is the dictionary cold?
+ * Only set on decompression.
+ */
+ unsigned dictionaryIsCold;
+ /**
+ * The dictionary size or zero if no dictionary.
+ */
+ size_t dictionarySize;
+ /**
+ * The uncompressed size of the data.
+ */
+ size_t uncompressedSize;
+ /**
+ * The compressed size of the data.
+ */
+ size_t compressedSize;
+ /**
+ * The fully resolved CCtx parameters (NULL on decompression).
+ */
+ struct ZSTD_CCtx_params_s const* params;
+ /**
+ * The ZSTD_CCtx pointer (NULL on decompression).
+ */
+ struct ZSTD_CCtx_s const* cctx;
+ /**
+ * The ZSTD_DCtx pointer (NULL on compression).
+ */
+ struct ZSTD_DCtx_s const* dctx;
+} ZSTD_Trace;
+
+/**
+ * A tracing context. It must be 0 when tracing is disabled.
+ * Otherwise, any non-zero value returned by a tracing begin()
+ * function is presented to any subsequent calls to end().
+ *
+ * Any non-zero value is treated as tracing is enabled and not
+ * interpreted by the library.
+ *
+ * Two possible uses are:
+ * * A timestamp for when the begin() function was called.
+ * * A unique key identifying the (de)compression, like the
+ * address of the [dc]ctx pointer if you need to track
+ * more information than just a timestamp.
+ */
+typedef unsigned long long ZSTD_TraceCtx;
+
+/**
+ * Trace the beginning of a compression call.
+ * @param cctx The dctx pointer for the compression.
+ * It can be used as a key to map begin() to end().
+ * @returns Non-zero if tracing is enabled. The return value is
+ * passed to ZSTD_trace_compress_end().
+ */
+ZSTD_WEAK_ATTR ZSTD_TraceCtx ZSTD_trace_compress_begin(
+ struct ZSTD_CCtx_s const* cctx);
+
+/**
+ * Trace the end of a compression call.
+ * @param ctx The return value of ZSTD_trace_compress_begin().
+ * @param trace The zstd tracing info.
+ */
+ZSTD_WEAK_ATTR void ZSTD_trace_compress_end(
+ ZSTD_TraceCtx ctx,
+ ZSTD_Trace const* trace);
+
+/**
+ * Trace the beginning of a decompression call.
+ * @param dctx The dctx pointer for the decompression.
+ * It can be used as a key to map begin() to end().
+ * @returns Non-zero if tracing is enabled. The return value is
+ * passed to ZSTD_trace_compress_end().
+ */
+ZSTD_WEAK_ATTR ZSTD_TraceCtx ZSTD_trace_decompress_begin(
+ struct ZSTD_DCtx_s const* dctx);
+
+/**
+ * Trace the end of a decompression call.
+ * @param ctx The return value of ZSTD_trace_decompress_begin().
+ * @param trace The zstd tracing info.
+ */
+ZSTD_WEAK_ATTR void ZSTD_trace_decompress_end(
+ ZSTD_TraceCtx ctx,
+ ZSTD_Trace const* trace);
+
+#endif /* ZSTD_TRACE */
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* ZSTD_TRACE_H */