Merge pull request #8066 from ThomasWaldmann/upgrade-bundled-code-1.4

Upgrade bundled code (1.4-maint)

src/borg/algorithms/lz4/lib/lz4.h

@@ -1,7 +1,7 @@
 /*
  *  LZ4 - Fast LZ compression algorithm
  *  Header File
- *  Copyright (C) 2011-present, Yann Collet.
+ *  Copyright (C) 2011-2020, Yann Collet.
 
    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
 
@@ -97,36 +97,77 @@ extern "C" {
 #  define LZ4LIB_API LZ4LIB_VISIBILITY
 #endif
 
+/*! LZ4_FREESTANDING :
+ *  When this macro is set to 1, it enables "freestanding mode" that is
+ *  suitable for typical freestanding environment which doesn't support
+ *  standard C library.
+ *
+ *  - LZ4_FREESTANDING is a compile-time switch.
+ *  - It requires the following macros to be defined:
+ *    LZ4_memcpy, LZ4_memmove, LZ4_memset.
+ *  - It only enables LZ4/HC functions which don't use heap.
+ *    All LZ4F_* functions are not supported.
+ *  - See tests/freestanding.c to check its basic setup.
+ */
+#if defined(LZ4_FREESTANDING) && (LZ4_FREESTANDING == 1)
+#  define LZ4_HEAPMODE 0
+#  define LZ4HC_HEAPMODE 0
+#  define LZ4_STATIC_LINKING_ONLY_DISABLE_MEMORY_ALLOCATION 1
+#  if !defined(LZ4_memcpy)
+#    error "LZ4_FREESTANDING requires macro 'LZ4_memcpy'."
+#  endif
+#  if !defined(LZ4_memset)
+#    error "LZ4_FREESTANDING requires macro 'LZ4_memset'."
+#  endif
+#  if !defined(LZ4_memmove)
+#    error "LZ4_FREESTANDING requires macro 'LZ4_memmove'."
+#  endif
+#elif ! defined(LZ4_FREESTANDING)
+#  define LZ4_FREESTANDING 0
+#endif
+
+
 /*------   Version   ------*/
 #define LZ4_VERSION_MAJOR    1    /* for breaking interface changes  */
 #define LZ4_VERSION_MINOR    9    /* for new (non-breaking) interface capabilities */
-#define LZ4_VERSION_RELEASE  2    /* for tweaks, bug-fixes, or development */
+#define LZ4_VERSION_RELEASE  4    /* for tweaks, bug-fixes, or development */
 
 #define LZ4_VERSION_NUMBER (LZ4_VERSION_MAJOR *100*100 + LZ4_VERSION_MINOR *100 + LZ4_VERSION_RELEASE)
 
 #define LZ4_LIB_VERSION LZ4_VERSION_MAJOR.LZ4_VERSION_MINOR.LZ4_VERSION_RELEASE
 #define LZ4_QUOTE(str) #str
 #define LZ4_EXPAND_AND_QUOTE(str) LZ4_QUOTE(str)
-#define LZ4_VERSION_STRING LZ4_EXPAND_AND_QUOTE(LZ4_LIB_VERSION)
+#define LZ4_VERSION_STRING LZ4_EXPAND_AND_QUOTE(LZ4_LIB_VERSION)  /* requires v1.7.3+ */
 
-LZ4LIB_API int LZ4_versionNumber (void);  /**< library version number; useful to check dll version */
-LZ4LIB_API const char* LZ4_versionString (void);   /**< library version string; useful to check dll version */
+LZ4LIB_API int LZ4_versionNumber (void);  /**< library version number; useful to check dll version; requires v1.3.0+ */
+LZ4LIB_API const char* LZ4_versionString (void);   /**< library version string; useful to check dll version; requires v1.7.5+ */
 
 
 /*-************************************
 *  Tuning parameter
 **************************************/
+#define LZ4_MEMORY_USAGE_MIN 10
+#define LZ4_MEMORY_USAGE_DEFAULT 14
+#define LZ4_MEMORY_USAGE_MAX 20
+
 /*!
  * LZ4_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 may improve speed, thanks to better cache locality.
+ * Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; )
+ * Increasing memory usage improves compression ratio, at the cost of speed.
+ * Reduced memory usage may improve speed at the cost of ratio, thanks to better cache locality.
  * Default value is 14, for 16KB, which nicely fits into Intel x86 L1 cache
  */
 #ifndef LZ4_MEMORY_USAGE
-# define LZ4_MEMORY_USAGE 14
+# define LZ4_MEMORY_USAGE LZ4_MEMORY_USAGE_DEFAULT
 #endif
 
+#if (LZ4_MEMORY_USAGE < LZ4_MEMORY_USAGE_MIN)
+#  error "LZ4_MEMORY_USAGE is too small !"
+#endif
+
+#if (LZ4_MEMORY_USAGE > LZ4_MEMORY_USAGE_MAX)
+#  error "LZ4_MEMORY_USAGE is too large !"
+#endif
 
 /*-************************************
 *  Simple Functions
@@ -186,7 +227,8 @@ LZ4LIB_API int LZ4_compressBound(int inputSize);
     The larger the acceleration value, the faster the algorithm, but also the lesser the compression.
     It's a trade-off. It can be fine tuned, with each successive value providing roughly +~3% to speed.
     An acceleration value of "1" is the same as regular LZ4_compress_default()
-    Values <= 0 will be replaced by ACCELERATION_DEFAULT (currently == 1, see lz4.c).
+    Values <= 0 will be replaced by LZ4_ACCELERATION_DEFAULT (currently == 1, see lz4.c).
+    Values > LZ4_ACCELERATION_MAX will be replaced by LZ4_ACCELERATION_MAX (currently == 65537, see lz4.c).
 */
 LZ4LIB_API int LZ4_compress_fast (const char* src, char* dst, int srcSize, int dstCapacity, int acceleration);
 
@@ -212,7 +254,18 @@ LZ4LIB_API int LZ4_compress_fast_extState (void* state, const char* src, char* d
  *               New value is necessarily <= input value.
  * @return : Nb bytes written into 'dst' (necessarily <= targetDestSize)
  *           or 0 if compression fails.
-*/
+ *
+ * Note : from v1.8.2 to v1.9.1, this function had a bug (fixed un v1.9.2+):
+ *        the produced compressed content could, in specific circumstances,
+ *        require to be decompressed into a destination buffer larger
+ *        by at least 1 byte than the content to decompress.
+ *        If an application uses `LZ4_compress_destSize()`,
+ *        it's highly recommended to update liblz4 to v1.9.2 or better.
+ *        If this can't be done or ensured,
+ *        the receiving decompression function should provide
+ *        a dstCapacity which is > decompressedSize, by at least 1 byte.
+ *        See https://github.com/lz4/lz4/issues/859 for details
+ */
 LZ4LIB_API int LZ4_compress_destSize (const char* src, char* dst, int* srcSizePtr, int targetDstSize);
 
 
@@ -220,25 +273,35 @@ LZ4LIB_API int LZ4_compress_destSize (const char* src, char* dst, int* srcSizePt
  *  Decompress an LZ4 compressed block, of size 'srcSize' at position 'src',
  *  into destination buffer 'dst' of size 'dstCapacity'.
  *  Up to 'targetOutputSize' bytes will be decoded.
- *  The function stops decoding on reaching this objective,
- *  which can boost performance when only the beginning of a block is required.
+ *  The function stops decoding on reaching this objective.
+ *  This can be useful to boost performance
+ *  whenever only the beginning of a block is required.
  *
- * @return : the number of bytes decoded in `dst` (necessarily <= dstCapacity)
+ * @return : the number of bytes decoded in `dst` (necessarily <= targetOutputSize)
  *           If source stream is detected malformed, function returns a negative result.
  *
- *  Note : @return can be < targetOutputSize, if compressed block contains less data.
+ *  Note 1 : @return can be < targetOutputSize, if compressed block contains less data.
  *
- *  Note 2 : this function features 2 parameters, targetOutputSize and dstCapacity,
- *           and expects targetOutputSize <= dstCapacity.
- *           It effectively stops decoding on reaching targetOutputSize,
+ *  Note 2 : targetOutputSize must be <= dstCapacity
+ *
+ *  Note 3 : this function effectively stops decoding on reaching targetOutputSize,
  *           so dstCapacity is kind of redundant.
- *           This is because in a previous version of this function,
- *           decoding operation would not "break" a sequence in the middle.
- *           As a consequence, there was no guarantee that decoding would stop at exactly targetOutputSize,
+ *           This is because in older versions of this function,
+ *           decoding operation would still write complete sequences.
+ *           Therefore, there was no guarantee that it would stop writing at exactly targetOutputSize,
  *           it could write more bytes, though only up to dstCapacity.
  *           Some "margin" used to be required for this operation to work properly.
- *           This is no longer necessary.
- *           The function nonetheless keeps its signature, in an effort to not break API.
+ *           Thankfully, this is no longer necessary.
+ *           The function nonetheless keeps the same signature, in an effort to preserve API compatibility.
+ *
+ *  Note 4 : If srcSize is the exact size of the block,
+ *           then targetOutputSize can be any value,
+ *           including larger than the block's decompressed size.
+ *           The function will, at most, generate block's decompressed size.
+ *
+ *  Note 5 : If srcSize is _larger_ than block's compressed size,
+ *           then targetOutputSize **MUST** be <= block's decompressed size.
+ *           Otherwise, *silent corruption will occur*.
  */
 LZ4LIB_API int LZ4_decompress_safe_partial (const char* src, char* dst, int srcSize, int targetOutputSize, int dstCapacity);
 
@@ -248,8 +311,25 @@ LZ4LIB_API int LZ4_decompress_safe_partial (const char* src, char* dst, int srcS
 ***********************************************/
 typedef union LZ4_stream_u LZ4_stream_t;  /* incomplete type (defined later) */
 
+/**
+ Note about RC_INVOKED
+
+ - RC_INVOKED is predefined symbol of rc.exe (the resource compiler which is part of MSVC/Visual Studio).
+   https://docs.microsoft.com/en-us/windows/win32/menurc/predefined-macros
+
+ - Since rc.exe is a legacy compiler, it truncates long symbol (> 30 chars)
+   and reports warning "RC4011: identifier truncated".
+
+ - To eliminate the warning, we surround long preprocessor symbol with
+   "#if !defined(RC_INVOKED) ... #endif" block that means
+   "skip this block when rc.exe is trying to read it".
+*/
+#if !defined(RC_INVOKED) /* https://docs.microsoft.com/en-us/windows/win32/menurc/predefined-macros */
+#if !defined(LZ4_STATIC_LINKING_ONLY_DISABLE_MEMORY_ALLOCATION)
 LZ4LIB_API LZ4_stream_t* LZ4_createStream(void);
 LZ4LIB_API int           LZ4_freeStream (LZ4_stream_t* streamPtr);
+#endif /* !defined(LZ4_STATIC_LINKING_ONLY_DISABLE_MEMORY_ALLOCATION) */
+#endif
 
 /*! LZ4_resetStream_fast() : v1.9.0+
  *  Use this to prepare an LZ4_stream_t for a new chain of dependent blocks
@@ -333,8 +413,12 @@ typedef union LZ4_streamDecode_u LZ4_streamDecode_t;   /* tracking context */
  *  creation / destruction of streaming decompression tracking context.
  *  A tracking context can be re-used multiple times.
  */
+#if !defined(RC_INVOKED) /* https://docs.microsoft.com/en-us/windows/win32/menurc/predefined-macros */
+#if !defined(LZ4_STATIC_LINKING_ONLY_DISABLE_MEMORY_ALLOCATION)
 LZ4LIB_API LZ4_streamDecode_t* LZ4_createStreamDecode(void);
 LZ4LIB_API int                 LZ4_freeStreamDecode (LZ4_streamDecode_t* LZ4_stream);
+#endif /* !defined(LZ4_STATIC_LINKING_ONLY_DISABLE_MEMORY_ALLOCATION) */
+#endif
 
 /*! LZ4_setStreamDecode() :
  *  An LZ4_streamDecode_t context can be allocated once and re-used multiple times.
@@ -384,7 +468,10 @@ LZ4LIB_API int LZ4_decoderRingBufferSize(int maxBlockSize);
  *  save the last 64KB of decoded data into a safe buffer where it can't be modified during decompression,
  *  then indicate where this data is saved using LZ4_setStreamDecode(), before decompressing next block.
 */
-LZ4LIB_API int LZ4_decompress_safe_continue (LZ4_streamDecode_t* LZ4_streamDecode, const char* src, char* dst, int srcSize, int dstCapacity);
+LZ4LIB_API int
+LZ4_decompress_safe_continue (LZ4_streamDecode_t* LZ4_streamDecode,
+                        const char* src, char* dst,
+                        int srcSize, int dstCapacity);
 
 
 /*! LZ4_decompress_*_usingDict() :
@@ -395,7 +482,16 @@ LZ4LIB_API int LZ4_decompress_safe_continue (LZ4_streamDecode_t* LZ4_streamDecod
  *  Performance tip : Decompression speed can be substantially increased
  *                    when dst == dictStart + dictSize.
  */
-LZ4LIB_API int LZ4_decompress_safe_usingDict (const char* src, char* dst, int srcSize, int dstCapcity, const char* dictStart, int dictSize);
+LZ4LIB_API int
+LZ4_decompress_safe_usingDict(const char* src, char* dst,
+                              int srcSize, int dstCapacity,
+                              const char* dictStart, int dictSize);
+
+LZ4LIB_API int
+LZ4_decompress_safe_partial_usingDict(const char* src, char* dst,
+                                      int compressedSize,
+                                      int targetOutputSize, int maxOutputSize,
+                                      const char* dictStart, int dictSize);
 
 #endif /* LZ4_H_2983827168210 */
 
@@ -474,13 +570,15 @@ LZ4LIB_STATIC_API int LZ4_compress_fast_extState_fastReset (void* state, const c
  *  stream (and source buffer) must remain in-place / accessible / unchanged
  *  through the completion of the first compression call on the stream.
  */
-LZ4LIB_STATIC_API void LZ4_attach_dictionary(LZ4_stream_t* workingStream, const LZ4_stream_t* dictionaryStream);
+LZ4LIB_STATIC_API void
+LZ4_attach_dictionary(LZ4_stream_t* workingStream,
+                const LZ4_stream_t* dictionaryStream);
 
 
 /*! In-place compression and decompression
  *
  * It's possible to have input and output sharing the same buffer,
- * for highly contrained memory environments.
+ * for highly constrained memory environments.
  * In both cases, it requires input to lay at the end of the buffer,
  * and decompression to start at beginning of the buffer.
  * Buffer size must feature some margin, hence be larger than final size.
@@ -547,74 +645,52 @@ LZ4LIB_STATIC_API void LZ4_attach_dictionary(LZ4_stream_t* workingStream, const
 #define LZ4_H_98237428734687
 
 /*-************************************************************
- *  PRIVATE DEFINITIONS
+ *  Private Definitions
  **************************************************************
  * Do not use these definitions directly.
  * They are only exposed to allow static allocation of `LZ4_stream_t` and `LZ4_streamDecode_t`.
- * Accessing members will expose code to API and/or ABI break in future versions of the library.
+ * Accessing members will expose user code to API and/or ABI break in future versions of the library.
  **************************************************************/
 #define LZ4_HASHLOG   (LZ4_MEMORY_USAGE-2)
 #define LZ4_HASHTABLESIZE (1 << LZ4_MEMORY_USAGE)
 #define LZ4_HASH_SIZE_U32 (1 << LZ4_HASHLOG)       /* required as macro for static allocation */
 
 #if defined(__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
-#include <stdint.h>
-
-typedef struct LZ4_stream_t_internal LZ4_stream_t_internal;
-struct LZ4_stream_t_internal {
-    uint32_t hashTable[LZ4_HASH_SIZE_U32];
-    uint32_t currentOffset;
-    uint16_t dirty;
-    uint16_t tableType;
-    const uint8_t* dictionary;
-    const LZ4_stream_t_internal* dictCtx;
-    uint32_t dictSize;
-};
-
-typedef struct {
-    const uint8_t* externalDict;
-    size_t extDictSize;
-    const uint8_t* prefixEnd;
-    size_t prefixSize;
-} LZ4_streamDecode_t_internal;
-
+# include <stdint.h>
+  typedef  int8_t  LZ4_i8;
+  typedef uint8_t  LZ4_byte;
+  typedef uint16_t LZ4_u16;
+  typedef uint32_t LZ4_u32;
 #else
+  typedef   signed char  LZ4_i8;
+  typedef unsigned char  LZ4_byte;
+  typedef unsigned short LZ4_u16;
+  typedef unsigned int   LZ4_u32;
+#endif
+
+/*! LZ4_stream_t :
+ *  Never ever use below internal definitions directly !
+ *  These definitions are not API/ABI safe, and may change in future versions.
+ *  If you need static allocation, declare or allocate an LZ4_stream_t object.
+**/
 
 typedef struct LZ4_stream_t_internal LZ4_stream_t_internal;
 struct LZ4_stream_t_internal {
-    unsigned int hashTable[LZ4_HASH_SIZE_U32];
-    unsigned int currentOffset;
-    unsigned short dirty;
-    unsigned short tableType;
-    const unsigned char* dictionary;
+    LZ4_u32 hashTable[LZ4_HASH_SIZE_U32];
+    const LZ4_byte* dictionary;
     const LZ4_stream_t_internal* dictCtx;
-    unsigned int dictSize;
+    LZ4_u32 currentOffset;
+    LZ4_u32 tableType;
+    LZ4_u32 dictSize;
+    /* Implicit padding to ensure structure is aligned */
 };
 
-typedef struct {
-    const unsigned char* externalDict;
-    const unsigned char* prefixEnd;
-    size_t extDictSize;
-    size_t prefixSize;
-} LZ4_streamDecode_t_internal;
-
-#endif
-
-/*! LZ4_stream_t :
- *  information structure to track an LZ4 stream.
- *  LZ4_stream_t can also be created using LZ4_createStream(), which is recommended.
- *  The structure definition can be convenient for static allocation
- *  (on stack, or as part of larger structure).
- *  Init this structure with LZ4_initStream() before first use.
- *  note : only use this definition in association with static linking !
- *    this definition is not API/ABI safe, and may change in a future version.
- */
-#define LZ4_STREAMSIZE_U64 ((1 << (LZ4_MEMORY_USAGE-3)) + 4 + ((sizeof(void*)==16) ? 4 : 0) /*AS-400*/ )
-#define LZ4_STREAMSIZE     (LZ4_STREAMSIZE_U64 * sizeof(unsigned long long))
+#define LZ4_STREAM_MINSIZE  ((1UL << LZ4_MEMORY_USAGE) + 32)  /* static size, for inter-version compatibility */
 union LZ4_stream_u {
-    unsigned long long table[LZ4_STREAMSIZE_U64];
+    char minStateSize[LZ4_STREAM_MINSIZE];
     LZ4_stream_t_internal internal_donotuse;
-} ;  /* previously typedef'd to LZ4_stream_t */
+}; /* previously typedef'd to LZ4_stream_t */
+
 
 /*! LZ4_initStream() : v1.9.0+
  *  An LZ4_stream_t structure must be initialized at least once.
@@ -629,21 +705,25 @@ union LZ4_stream_u {
  *         In which case, the function will @return NULL.
  *  Note2: An LZ4_stream_t structure guarantees correct alignment and size.
  *  Note3: Before v1.9.0, use LZ4_resetStream() instead
- */
+**/
 LZ4LIB_API LZ4_stream_t* LZ4_initStream (void* buffer, size_t size);
 
 
 /*! LZ4_streamDecode_t :
- *  information structure to track an LZ4 stream during decompression.
- *  init this structure  using LZ4_setStreamDecode() before first use.
- *  note : only use in association with static linking !
- *         this definition is not API/ABI safe,
- *         and may change in a future version !
- */
-#define LZ4_STREAMDECODESIZE_U64 (4 + ((sizeof(void*)==16) ? 2 : 0) /*AS-400*/ )
-#define LZ4_STREAMDECODESIZE     (LZ4_STREAMDECODESIZE_U64 * sizeof(unsigned long long))
+ *  Never ever use below internal definitions directly !
+ *  These definitions are not API/ABI safe, and may change in future versions.
+ *  If you need static allocation, declare or allocate an LZ4_streamDecode_t object.
+**/
+typedef struct {
+    const LZ4_byte* externalDict;
+    const LZ4_byte* prefixEnd;
+    size_t extDictSize;
+    size_t prefixSize;
+} LZ4_streamDecode_t_internal;
+
+#define LZ4_STREAMDECODE_MINSIZE 32
 union LZ4_streamDecode_u {
-    unsigned long long table[LZ4_STREAMDECODESIZE_U64];
+    char minStateSize[LZ4_STREAMDECODE_MINSIZE];
     LZ4_streamDecode_t_internal internal_donotuse;
 } ;   /* previously typedef'd to LZ4_streamDecode_t */
 
@@ -667,22 +747,21 @@ union LZ4_streamDecode_u {
 #ifdef LZ4_DISABLE_DEPRECATE_WARNINGS
 #  define LZ4_DEPRECATED(message)   /* disable deprecation warnings */
 #else
-#  define LZ4_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
 #  if defined (__cplusplus) && (__cplusplus >= 201402) /* C++14 or greater */
 #    define LZ4_DEPRECATED(message) [[deprecated(message)]]
-#  elif (LZ4_GCC_VERSION >= 405) || defined(__clang__)
-#    define LZ4_DEPRECATED(message) __attribute__((deprecated(message)))
-#  elif (LZ4_GCC_VERSION >= 301)
-#    define LZ4_DEPRECATED(message) __attribute__((deprecated))
 #  elif defined(_MSC_VER)
 #    define LZ4_DEPRECATED(message) __declspec(deprecated(message))
+#  elif defined(__clang__) || (defined(__GNUC__) && (__GNUC__ * 10 + __GNUC_MINOR__ >= 45))
+#    define LZ4_DEPRECATED(message) __attribute__((deprecated(message)))
+#  elif defined(__GNUC__) && (__GNUC__ * 10 + __GNUC_MINOR__ >= 31)
+#    define LZ4_DEPRECATED(message) __attribute__((deprecated))
 #  else
-#    pragma message("WARNING: You need to implement LZ4_DEPRECATED for this compiler")
-#    define LZ4_DEPRECATED(message)
+#    pragma message("WARNING: LZ4_DEPRECATED needs custom implementation for this compiler")
+#    define LZ4_DEPRECATED(message)   /* disabled */
 #  endif
 #endif /* LZ4_DISABLE_DEPRECATE_WARNINGS */
 
-/* Obsolete compression functions */
+/*! Obsolete compression functions (since v1.7.3) */
 LZ4_DEPRECATED("use LZ4_compress_default() instead")       LZ4LIB_API int LZ4_compress               (const char* src, char* dest, int srcSize);
 LZ4_DEPRECATED("use LZ4_compress_default() instead")       LZ4LIB_API int LZ4_compress_limitedOutput (const char* src, char* dest, int srcSize, int maxOutputSize);
 LZ4_DEPRECATED("use LZ4_compress_fast_extState() instead") LZ4LIB_API int LZ4_compress_withState               (void* state, const char* source, char* dest, int inputSize);
@@ -690,11 +769,12 @@ LZ4_DEPRECATED("use LZ4_compress_fast_extState() instead") LZ4LIB_API int LZ4_co
 LZ4_DEPRECATED("use LZ4_compress_fast_continue() instead") LZ4LIB_API int LZ4_compress_continue                (LZ4_stream_t* LZ4_streamPtr, const char* source, char* dest, int inputSize);
 LZ4_DEPRECATED("use LZ4_compress_fast_continue() instead") LZ4LIB_API int LZ4_compress_limitedOutput_continue  (LZ4_stream_t* LZ4_streamPtr, const char* source, char* dest, int inputSize, int maxOutputSize);
 
-/* Obsolete decompression functions */
+/*! Obsolete decompression functions (since v1.8.0) */
 LZ4_DEPRECATED("use LZ4_decompress_fast() instead") LZ4LIB_API int LZ4_uncompress (const char* source, char* dest, int outputSize);
 LZ4_DEPRECATED("use LZ4_decompress_safe() instead") LZ4LIB_API int LZ4_uncompress_unknownOutputSize (const char* source, char* dest, int isize, int maxOutputSize);
 
-/* Obsolete streaming functions; degraded functionality; do not use!
+/* Obsolete streaming functions (since v1.7.0)
+ * degraded functionality; do not use!
  *
  * In order to perform streaming compression, these functions depended on data
  * that is no longer tracked in the state. They have been preserved as well as
@@ -708,23 +788,22 @@ LZ4_DEPRECATED("Use LZ4_createStream() instead") LZ4LIB_API int   LZ4_sizeofStre
 LZ4_DEPRECATED("Use LZ4_resetStream() instead")  LZ4LIB_API int   LZ4_resetStreamState(void* state, char* inputBuffer);
 LZ4_DEPRECATED("Use LZ4_saveDict() instead")     LZ4LIB_API char* LZ4_slideInputBuffer (void* state);
 
-/* Obsolete streaming decoding functions */
+/*! Obsolete streaming decoding functions (since v1.7.0) */
 LZ4_DEPRECATED("use LZ4_decompress_safe_usingDict() instead") LZ4LIB_API int LZ4_decompress_safe_withPrefix64k (const char* src, char* dst, int compressedSize, int maxDstSize);
 LZ4_DEPRECATED("use LZ4_decompress_fast_usingDict() instead") LZ4LIB_API int LZ4_decompress_fast_withPrefix64k (const char* src, char* dst, int originalSize);
 
-/*! LZ4_decompress_fast() : **unsafe!**
+/*! Obsolete LZ4_decompress_fast variants (since v1.9.0) :
  *  These functions used to be faster than LZ4_decompress_safe(),
- *  but it has changed, and they are now slower than LZ4_decompress_safe().
+ *  but this is no longer the case. They are now slower.
  *  This is because LZ4_decompress_fast() doesn't know the input size,
- *  and therefore must progress more cautiously in the input buffer to not read beyond the end of block.
+ *  and therefore must progress more cautiously into the input buffer to not read beyond the end of block.
  *  On top of that `LZ4_decompress_fast()` is not protected vs malformed or malicious inputs, making it a security liability.
  *  As a consequence, LZ4_decompress_fast() is strongly discouraged, and deprecated.
  *
  *  The last remaining LZ4_decompress_fast() specificity is that
  *  it can decompress a block without knowing its compressed size.
- *  Such functionality could be achieved in a more secure manner,
- *  by also providing the maximum size of input buffer,
- *  but it would require new prototypes, and adaptation of the implementation to this new use case.
+ *  Such functionality can be achieved in a more secure manner
+ *  by employing LZ4_decompress_safe_partial().
  *
  *  Parameters:
  *  originalSize : is the uncompressed size to regenerate.
@@ -739,7 +818,6 @@ LZ4_DEPRECATED("use LZ4_decompress_fast_usingDict() instead") LZ4LIB_API int LZ4
  *         But they may happen if input data is invalid (error or intentional tampering).
  *         As a consequence, use these functions in trusted environments with trusted data **only**.
  */
-
 LZ4_DEPRECATED("This function is deprecated and unsafe. Consider using LZ4_decompress_safe() instead")
 LZ4LIB_API int LZ4_decompress_fast (const char* src, char* dst, int originalSize);
 LZ4_DEPRECATED("This function is deprecated and unsafe. Consider using LZ4_decompress_safe_continue() instead")

src/borg/algorithms/xxh64/xxhash.c

@@ -1,6 +1,6 @@
 /*
  * xxHash - Extremely Fast Hash algorithm
- * Copyright (C) 2012-2020 Yann Collet
+ * Copyright (C) 2012-2021 Yann Collet
  *
  * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php)
  *

src/borg/algorithms/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 "mem.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 */

src/borg/algorithms/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 */

src/borg/algorithms/zstd/lib/common/bitstream.h

@@ -1,7 +1,7 @@
 /* ******************************************************************
  * bitstream
  * Part of FSE library
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -17,7 +17,6 @@
 #if defined (__cplusplus)
 extern "C" {
 #endif
-
 /*
 *  This API consists of small unitary functions, which must be inlined for best performance.
 *  Since link-time-optimization is not available for all compilers,
@@ -31,15 +30,18 @@ extern "C" {
 #include "compiler.h"       /* UNLIKELY() */
 #include "debug.h"          /* assert(), DEBUGLOG(), RAWLOG() */
 #include "error_private.h"  /* error codes and messages */
+#include "bits.h"           /* ZSTD_highbit32 */
 
 
 /*=========================================
 *  Target specific
 =========================================*/
-#if defined(__BMI__) && defined(__GNUC__)
-#  include <immintrin.h>   /* support for bextr (experimental) */
-#elif defined(__ICCARM__)
-#  include <intrinsics.h>
+#ifndef ZSTD_NO_INTRINSICS
+#  if (defined(__BMI__) || defined(__BMI2__)) && defined(__GNUC__)
+#    include <immintrin.h>   /* support for bextr (experimental)/bzhi */
+#  elif defined(__ICCARM__)
+#    include <intrinsics.h>
+#  endif
 #endif
 
 #define STREAM_ACCUMULATOR_MIN_32  25
@@ -131,38 +133,6 @@ MEM_STATIC void BIT_flushBitsFast(BIT_CStream_t* bitC);
 MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, unsigned nbBits);
 /* faster, but works only if nbBits >= 1 */
 
-
-
-/*-**************************************************************
-*  Internal functions
-****************************************************************/
-MEM_STATIC unsigned BIT_highbit32 (U32 val)
-{
-    assert(val != 0);
-    {
-#   if defined(_MSC_VER)   /* Visual */
-        unsigned long r=0;
-        return _BitScanReverse ( &r, val ) ? (unsigned)r : 0;
-#   elif defined(__GNUC__) && (__GNUC__ >= 3)   /* Use GCC Intrinsic */
-        return __builtin_clz (val) ^ 31;
-#   elif defined(__ICCARM__)    /* IAR Intrinsic */
-        return 31 - __CLZ(val);
-#   else   /* Software version */
-        static const unsigned DeBruijnClz[32] = { 0,  9,  1, 10, 13, 21,  2, 29,
-                                                 11, 14, 16, 18, 22, 25,  3, 30,
-                                                  8, 12, 20, 28, 15, 17, 24,  7,
-                                                 19, 27, 23,  6, 26,  5,  4, 31 };
-        U32 v = val;
-        v |= v >> 1;
-        v |= v >> 2;
-        v |= v >> 4;
-        v |= v >> 8;
-        v |= v >> 16;
-        return DeBruijnClz[ (U32) (v * 0x07C4ACDDU) >> 27];
-#   endif
-    }
-}
-
 /*=====    Local Constants   =====*/
 static const unsigned BIT_mask[] = {
     0,          1,         3,         7,         0xF,       0x1F,
@@ -192,16 +162,26 @@ MEM_STATIC size_t BIT_initCStream(BIT_CStream_t* bitC,
     return 0;
 }
 
+MEM_STATIC FORCE_INLINE_ATTR size_t BIT_getLowerBits(size_t bitContainer, U32 const nbBits)
+{
+#if defined(STATIC_BMI2) && STATIC_BMI2 == 1 && !defined(ZSTD_NO_INTRINSICS)
+    return  _bzhi_u64(bitContainer, nbBits);
+#else
+    assert(nbBits < BIT_MASK_SIZE);
+    return bitContainer & BIT_mask[nbBits];
+#endif
+}
+
 /*! BIT_addBits() :
  *  can add up to 31 bits into `bitC`.
  *  Note : does not check for register overflow ! */
 MEM_STATIC void BIT_addBits(BIT_CStream_t* bitC,
                             size_t value, unsigned nbBits)
 {
-    MEM_STATIC_ASSERT(BIT_MASK_SIZE == 32);
+    DEBUG_STATIC_ASSERT(BIT_MASK_SIZE == 32);
     assert(nbBits < BIT_MASK_SIZE);
     assert(nbBits + bitC->bitPos < sizeof(bitC->bitContainer) * 8);
-    bitC->bitContainer |= (value & BIT_mask[nbBits]) << bitC->bitPos;
+    bitC->bitContainer |= BIT_getLowerBits(value, nbBits) << bitC->bitPos;
     bitC->bitPos += nbBits;
 }
 
@@ -271,7 +251,7 @@ MEM_STATIC size_t BIT_closeCStream(BIT_CStream_t* bitC)
  */
 MEM_STATIC size_t BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, size_t srcSize)
 {
-    if (srcSize < 1) { memset(bitD, 0, sizeof(*bitD)); return ERROR(srcSize_wrong); }
+    if (srcSize < 1) { ZSTD_memset(bitD, 0, sizeof(*bitD)); return ERROR(srcSize_wrong); }
 
     bitD->start = (const char*)srcBuffer;
     bitD->limitPtr = bitD->start + sizeof(bitD->bitContainer);
@@ -280,7 +260,7 @@ MEM_STATIC size_t BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, si
         bitD->ptr   = (const char*)srcBuffer + srcSize - sizeof(bitD->bitContainer);
         bitD->bitContainer = MEM_readLEST(bitD->ptr);
         { BYTE const lastByte = ((const BYTE*)srcBuffer)[srcSize-1];
-          bitD->bitsConsumed = lastByte ? 8 - BIT_highbit32(lastByte) : 0;  /* ensures bitsConsumed is always set */
+          bitD->bitsConsumed = lastByte ? 8 - ZSTD_highbit32(lastByte) : 0;  /* ensures bitsConsumed is always set */
           if (lastByte == 0) return ERROR(GENERIC); /* endMark not present */ }
     } else {
         bitD->ptr   = bitD->start;
@@ -288,27 +268,27 @@ MEM_STATIC size_t BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, si
         switch(srcSize)
         {
         case 7: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[6]) << (sizeof(bitD->bitContainer)*8 - 16);
-                /* fall-through */
+                ZSTD_FALLTHROUGH;
 
         case 6: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[5]) << (sizeof(bitD->bitContainer)*8 - 24);
-                /* fall-through */
+                ZSTD_FALLTHROUGH;
 
         case 5: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[4]) << (sizeof(bitD->bitContainer)*8 - 32);
-                /* fall-through */
+                ZSTD_FALLTHROUGH;
 
         case 4: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[3]) << 24;
-                /* fall-through */
+                ZSTD_FALLTHROUGH;
 
         case 3: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[2]) << 16;
-                /* fall-through */
+                ZSTD_FALLTHROUGH;
 
         case 2: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[1]) <<  8;
-                /* fall-through */
+                ZSTD_FALLTHROUGH;
 
         default: break;
         }
         {   BYTE const lastByte = ((const BYTE*)srcBuffer)[srcSize-1];
-            bitD->bitsConsumed = lastByte ? 8 - BIT_highbit32(lastByte) : 0;
+            bitD->bitsConsumed = lastByte ? 8 - ZSTD_highbit32(lastByte) : 0;
             if (lastByte == 0) return ERROR(corruption_detected);  /* endMark not present */
         }
         bitD->bitsConsumed += (U32)(sizeof(bitD->bitContainer) - srcSize)*8;
@@ -317,23 +297,26 @@ MEM_STATIC size_t BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, si
     return srcSize;
 }
 
-MEM_STATIC size_t BIT_getUpperBits(size_t bitContainer, U32 const start)
+MEM_STATIC FORCE_INLINE_ATTR size_t BIT_getUpperBits(size_t bitContainer, U32 const start)
 {
     return bitContainer >> start;
 }
 
-MEM_STATIC size_t BIT_getMiddleBits(size_t bitContainer, U32 const start, U32 const nbBits)
+MEM_STATIC FORCE_INLINE_ATTR size_t BIT_getMiddleBits(size_t bitContainer, U32 const start, U32 const nbBits)
 {
     U32 const regMask = sizeof(bitContainer)*8 - 1;
     /* if start > regMask, bitstream is corrupted, and result is undefined */
     assert(nbBits < BIT_MASK_SIZE);
+    /* x86 transform & ((1 << nbBits) - 1) to bzhi instruction, it is better
+     * than accessing memory. When bmi2 instruction is not present, we consider
+     * such cpus old (pre-Haswell, 2013) and their performance is not of that
+     * importance.
+     */
+#if defined(__x86_64__) || defined(_M_X86)
+    return (bitContainer >> (start & regMask)) & ((((U64)1) << nbBits) - 1);
+#else
     return (bitContainer >> (start & regMask)) & BIT_mask[nbBits];
-}
-
-MEM_STATIC size_t BIT_getLowerBits(size_t bitContainer, U32 const nbBits)
-{
-    assert(nbBits < BIT_MASK_SIZE);
-    return bitContainer & BIT_mask[nbBits];
+#endif
 }
 
 /*! BIT_lookBits() :
@@ -342,7 +325,7 @@ MEM_STATIC size_t BIT_getLowerBits(size_t bitContainer, U32 const nbBits)
  *  On 32-bits, maxNbBits==24.
  *  On 64-bits, maxNbBits==56.
  * @return : value extracted */
-MEM_STATIC size_t BIT_lookBits(const BIT_DStream_t* bitD, U32 nbBits)
+MEM_STATIC  FORCE_INLINE_ATTR size_t BIT_lookBits(const BIT_DStream_t*  bitD, U32 nbBits)
 {
     /* arbitrate between double-shift and shift+mask */
 #if 1
@@ -365,7 +348,7 @@ MEM_STATIC size_t BIT_lookBitsFast(const BIT_DStream_t* bitD, U32 nbBits)
     return (bitD->bitContainer << (bitD->bitsConsumed & regMask)) >> (((regMask+1)-nbBits) & regMask);
 }
 
-MEM_STATIC void BIT_skipBits(BIT_DStream_t* bitD, U32 nbBits)
+MEM_STATIC FORCE_INLINE_ATTR void BIT_skipBits(BIT_DStream_t* bitD, U32 nbBits)
 {
     bitD->bitsConsumed += nbBits;
 }
@@ -374,7 +357,7 @@ MEM_STATIC void BIT_skipBits(BIT_DStream_t* bitD, U32 nbBits)
  *  Read (consume) next n bits from local register and update.
  *  Pay attention to not read more than nbBits contained into local register.
  * @return : extracted value. */
-MEM_STATIC size_t BIT_readBits(BIT_DStream_t* bitD, unsigned nbBits)
+MEM_STATIC FORCE_INLINE_ATTR size_t BIT_readBits(BIT_DStream_t* bitD, unsigned nbBits)
 {
     size_t const value = BIT_lookBits(bitD, nbBits);
     BIT_skipBits(bitD, nbBits);
@@ -382,7 +365,7 @@ MEM_STATIC size_t BIT_readBits(BIT_DStream_t* bitD, unsigned nbBits)
 }
 
 /*! BIT_readBitsFast() :
- *  unsafe version; only works only if nbBits >= 1 */
+ *  unsafe version; only works if nbBits >= 1 */
 MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, unsigned nbBits)
 {
     size_t const value = BIT_lookBitsFast(bitD, nbBits);
@@ -413,7 +396,7 @@ MEM_STATIC BIT_DStream_status BIT_reloadDStreamFast(BIT_DStream_t* bitD)
  *  This function is safe, it guarantees it will not read 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 */
-MEM_STATIC BIT_DStream_status BIT_reloadDStream(BIT_DStream_t* bitD)
+MEM_STATIC FORCE_INLINE_ATTR BIT_DStream_status BIT_reloadDStream(BIT_DStream_t* bitD)
 {
     if (bitD->bitsConsumed > (sizeof(bitD->bitContainer)*8))  /* overflow detected, like end of stream */
         return BIT_DStream_overflow;

src/borg/algorithms/zstd/lib/common/compiler.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -11,6 +11,8 @@
 #ifndef ZSTD_COMPILER_H
 #define ZSTD_COMPILER_H
 
+#include "portability_macros.h"
+
 /*-*******************************************************
 *  Compiler specifics
 *********************************************************/
@@ -38,6 +40,17 @@
 
 #endif
 
+/**
+  On MSVC qsort requires that functions passed into it use the __cdecl calling conversion(CC).
+  This explicitly marks such functions as __cdecl so that the code will still compile
+  if a CC other than __cdecl has been made the default.
+*/
+#if  defined(_MSC_VER)
+#  define WIN_CDECL __cdecl
+#else
+#  define WIN_CDECL
+#endif
+
 /**
  * FORCE_INLINE_TEMPLATE is used to define C "templates", which take constant
  * parameters. They must be inlined for the compiler to eliminate the constant
@@ -79,30 +92,19 @@
 #  endif
 #endif
 
+
 /* target attribute */
-#ifndef __has_attribute
-  #define __has_attribute(x) 0  /* Compatibility with non-clang compilers. */
-#endif
 #if defined(__GNUC__) || defined(__ICCARM__)
 #  define TARGET_ATTRIBUTE(target) __attribute__((__target__(target)))
 #else
 #  define TARGET_ATTRIBUTE(target)
 #endif
 
-/* Enable runtime BMI2 dispatch based on the CPU.
- * Enabled for clang & gcc >=4.8 on x86 when BMI2 isn't enabled by default.
+/* Target attribute for BMI2 dynamic dispatch.
+ * Enable lzcnt, bmi, and bmi2.
+ * We test for bmi1 & bmi2. lzcnt is included in bmi1.
  */
-#ifndef DYNAMIC_BMI2
-  #if ((defined(__clang__) && __has_attribute(__target__)) \
-      || (defined(__GNUC__) \
-          && (__GNUC__ >= 5 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8)))) \
-      && (defined(__x86_64__) || defined(_M_X86)) \
-      && !defined(__BMI2__)
-  #  define DYNAMIC_BMI2 1
-  #else
-  #  define DYNAMIC_BMI2 0
-  #endif
-#endif
+#define BMI2_TARGET_ATTRIBUTE TARGET_ATTRIBUTE("lzcnt,bmi,bmi2")
 
 /* prefetch
  * can be disabled, by declaring NO_PREFETCH build macro */
@@ -114,12 +116,12 @@
 #    include <mmintrin.h>   /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
 #    define PREFETCH_L1(ptr)  _mm_prefetch((const char*)(ptr), _MM_HINT_T0)
 #    define PREFETCH_L2(ptr)  _mm_prefetch((const char*)(ptr), _MM_HINT_T1)
-#    elif defined(__aarch64__)
-#     define PREFETCH_L1(ptr)  __asm__ __volatile__("prfm pldl1keep, %0" ::"Q"(*(ptr)))
-#     define PREFETCH_L2(ptr)  __asm__ __volatile__("prfm pldl2keep, %0" ::"Q"(*(ptr)))
 #  elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) )
 #    define PREFETCH_L1(ptr)  __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */)
 #    define PREFETCH_L2(ptr)  __builtin_prefetch((ptr), 0 /* rw==read */, 2 /* locality */)
+#  elif defined(__aarch64__)
+#    define PREFETCH_L1(ptr)  __asm__ __volatile__("prfm pldl1keep, %0" ::"Q"(*(ptr)))
+#    define PREFETCH_L2(ptr)  __asm__ __volatile__("prfm pldl2keep, %0" ::"Q"(*(ptr)))
 #  else
 #    define PREFETCH_L1(ptr) (void)(ptr)  /* disabled */
 #    define PREFETCH_L2(ptr) (void)(ptr)  /* disabled */
@@ -138,8 +140,9 @@
 }
 
 /* vectorization
- * older GCC (pre gcc-4.3 picked as the cutoff) uses a different syntax */
-#if !defined(__INTEL_COMPILER) && !defined(__clang__) && defined(__GNUC__)
+ * older GCC (pre gcc-4.3 picked as the cutoff) uses a different syntax,
+ * and some compilers, like Intel ICC and MCST LCC, do not support it at all. */
+#if !defined(__INTEL_COMPILER) && !defined(__clang__) && defined(__GNUC__) && !defined(__LCC__)
 #  if (__GNUC__ == 4 && __GNUC_MINOR__ > 3) || (__GNUC__ >= 5)
 #    define DONT_VECTORIZE __attribute__((optimize("no-tree-vectorize")))
 #  else
@@ -162,6 +165,12 @@
 #define UNLIKELY(x) (x)
 #endif
 
+#if __has_builtin(__builtin_unreachable) || (defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 5)))
+#  define ZSTD_UNREACHABLE { assert(0), __builtin_unreachable(); }
+#else
+#  define ZSTD_UNREACHABLE { assert(0); }
+#endif
+
 /* disable warnings */
 #ifdef _MSC_VER    /* Visual Studio */
 #  include <intrin.h>                    /* For Visual 2005 */
@@ -172,4 +181,178 @@
 #  pragma warning(disable : 4324)        /* disable: C4324: padded structure */
 #endif
 
+/*Like DYNAMIC_BMI2 but for compile time determination of BMI2 support*/
+#ifndef STATIC_BMI2
+#  if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_I86))
+#    ifdef __AVX2__  //MSVC does not have a BMI2 specific flag, but every CPU that supports AVX2 also supports BMI2
+#       define STATIC_BMI2 1
+#    endif
+#  elif defined(__BMI2__) && defined(__x86_64__) && defined(__GNUC__)
+#    define STATIC_BMI2 1
+#  endif
+#endif
+
+#ifndef STATIC_BMI2
+    #define STATIC_BMI2 0
+#endif
+
+/* compile time determination of SIMD support */
+#if !defined(ZSTD_NO_INTRINSICS)
+#  if defined(__SSE2__) || defined(_M_AMD64) || (defined (_M_IX86) && defined(_M_IX86_FP) && (_M_IX86_FP >= 2))
+#    define ZSTD_ARCH_X86_SSE2
+#  endif
+#  if defined(__ARM_NEON) || defined(_M_ARM64)
+#    define ZSTD_ARCH_ARM_NEON
+#  endif
+#
+#  if defined(ZSTD_ARCH_X86_SSE2)
+#    include <emmintrin.h>
+#  elif defined(ZSTD_ARCH_ARM_NEON)
+#    include <arm_neon.h>
+#  endif
+#endif
+
+/* C-language Attributes are added in C23. */
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ > 201710L) && defined(__has_c_attribute)
+# define ZSTD_HAS_C_ATTRIBUTE(x) __has_c_attribute(x)
+#else
+# define ZSTD_HAS_C_ATTRIBUTE(x) 0
+#endif
+
+/* Only use C++ attributes in C++. Some compilers report support for C++
+ * attributes when compiling with C.
+ */
+#if defined(__cplusplus) && defined(__has_cpp_attribute)
+# define ZSTD_HAS_CPP_ATTRIBUTE(x) __has_cpp_attribute(x)
+#else
+# define ZSTD_HAS_CPP_ATTRIBUTE(x) 0
+#endif
+
+/* Define ZSTD_FALLTHROUGH macro for annotating switch case with the 'fallthrough' attribute.
+ * - C23: https://en.cppreference.com/w/c/language/attributes/fallthrough
+ * - CPP17: https://en.cppreference.com/w/cpp/language/attributes/fallthrough
+ * - Else: __attribute__((__fallthrough__))
+ */
+#ifndef ZSTD_FALLTHROUGH
+# if ZSTD_HAS_C_ATTRIBUTE(fallthrough)
+#  define ZSTD_FALLTHROUGH [[fallthrough]]
+# elif ZSTD_HAS_CPP_ATTRIBUTE(fallthrough)
+#  define ZSTD_FALLTHROUGH [[fallthrough]]
+# elif __has_attribute(__fallthrough__)
+/* Leading semicolon is to satisfy gcc-11 with -pedantic. Without the semicolon
+ * gcc complains about: a label can only be part of a statement and a declaration is not a statement.
+ */
+#  define ZSTD_FALLTHROUGH ; __attribute__((__fallthrough__))
+# else
+#  define ZSTD_FALLTHROUGH
+# endif
+#endif
+
+/*-**************************************************************
+*  Alignment check
+*****************************************************************/
+
+/* this test was initially positioned in mem.h,
+ * but this file is removed (or replaced) for linux kernel
+ * so it's now hosted in compiler.h,
+ * which remains valid for both user & kernel spaces.
+ */
+
+#ifndef ZSTD_ALIGNOF
+# if defined(__GNUC__) || defined(_MSC_VER)
+/* covers gcc, clang & MSVC */
+/* note : this section must come first, before C11,
+ * due to a limitation in the kernel source generator */
+#  define ZSTD_ALIGNOF(T) __alignof(T)
+
+# elif defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)
+/* C11 support */
+#  include <stdalign.h>
+#  define ZSTD_ALIGNOF(T) alignof(T)
+
+# else
+/* No known support for alignof() - imperfect backup */
+#  define ZSTD_ALIGNOF(T) (sizeof(void*) < sizeof(T) ? sizeof(void*) : sizeof(T))
+
+# endif
+#endif /* ZSTD_ALIGNOF */
+
+/*-**************************************************************
+*  Sanitizer
+*****************************************************************/
+
+/* Issue #3240 reports an ASAN failure on an llvm-mingw build. Out of an
+ * abundance of caution, disable our custom poisoning on mingw. */
+#ifdef __MINGW32__
+#ifndef ZSTD_ASAN_DONT_POISON_WORKSPACE
+#define ZSTD_ASAN_DONT_POISON_WORKSPACE 1
+#endif
+#ifndef ZSTD_MSAN_DONT_POISON_WORKSPACE
+#define ZSTD_MSAN_DONT_POISON_WORKSPACE 1
+#endif
+#endif
+
+#if ZSTD_MEMORY_SANITIZER && !defined(ZSTD_MSAN_DONT_POISON_WORKSPACE)
+/* Not all platforms that support msan provide sanitizers/msan_interface.h.
+ * We therefore declare the functions we need ourselves, rather than trying to
+ * include the header file... */
+#include <stddef.h>  /* size_t */
+#define ZSTD_DEPS_NEED_STDINT
+#include "zstd_deps.h"  /* intptr_t */
+
+/* Make memory region fully initialized (without changing its contents). */
+void __msan_unpoison(const volatile void *a, size_t size);
+
+/* Make memory region fully uninitialized (without changing its contents).
+   This is a legacy interface that does not update origin information. Use
+   __msan_allocated_memory() instead. */
+void __msan_poison(const volatile void *a, size_t size);
+
+/* Returns the offset of the first (at least partially) poisoned byte in the
+   memory range, or -1 if the whole range is good. */
+intptr_t __msan_test_shadow(const volatile void *x, size_t size);
+
+/* 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 */

src/borg/algorithms/zstd/lib/common/cpu.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2020, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -16,8 +16,6 @@
  * https://github.com/facebook/folly/blob/master/folly/CpuId.h
  */
 
-#include <string.h>
-
 #include "mem.h"
 
 #ifdef _MSC_VER

src/borg/algorithms/zstd/lib/common/debug.c

@@ -1,7 +1,7 @@
 /* ******************************************************************
  * debug
  * Part of FSE library
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy

src/borg/algorithms/zstd/lib/common/debug.h

@@ -1,7 +1,7 @@
 /* ******************************************************************
  * debug
  * Part of FSE library
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -51,15 +51,6 @@ extern "C" {
 #endif
 
 
-/* DEBUGFILE can be defined externally,
- * typically through compiler command line.
- * note : currently useless.
- * Value must be stderr or stdout */
-#ifndef DEBUGFILE
-#  define DEBUGFILE stderr
-#endif
-
-
 /* recommended values for DEBUGLEVEL :
  * 0 : release mode, no debug, all run-time checks disabled
  * 1 : enables assert() only, no display
@@ -76,7 +67,8 @@ extern "C" {
  */
 
 #if (DEBUGLEVEL>=1)
-#  include <assert.h>
+#  define ZSTD_DEPS_NEED_ASSERT
+#  include "zstd_deps.h"
 #else
 #  ifndef assert   /* assert may be already defined, due to prior #include <assert.h> */
 #    define assert(condition) ((void)0)   /* disable assert (default) */
@@ -84,7 +76,8 @@ extern "C" {
 #endif
 
 #if (DEBUGLEVEL>=2)
-#  include <stdio.h>
+#  define ZSTD_DEPS_NEED_IO
+#  include "zstd_deps.h"
 extern int g_debuglevel; /* the variable is only declared,
                             it actually lives in debug.c,
                             and is shared by the whole process.
@@ -92,14 +85,14 @@ extern int g_debuglevel; /* the variable is only declared,
                             It's useful when enabling very verbose levels
                             on selective conditions (such as position in src) */
 
-#  define RAWLOG(l, ...) {                                      \
-                if (l<=g_debuglevel) {                          \
-                    fprintf(stderr, __VA_ARGS__);               \
+#  define RAWLOG(l, ...) {                                       \
+                if (l<=g_debuglevel) {                           \
+                    ZSTD_DEBUG_PRINT(__VA_ARGS__);               \
             }   }
-#  define DEBUGLOG(l, ...) {                                    \
-                if (l<=g_debuglevel) {                          \
-                    fprintf(stderr, __FILE__ ": " __VA_ARGS__); \
-                    fprintf(stderr, " \n");                     \
+#  define DEBUGLOG(l, ...) {                                     \
+                if (l<=g_debuglevel) {                           \
+                    ZSTD_DEBUG_PRINT(__FILE__ ": " __VA_ARGS__); \
+                    ZSTD_DEBUG_PRINT(" \n");                     \
             }   }
 #else
 #  define RAWLOG(l, ...)      {}    /* disabled */

src/borg/algorithms/zstd/lib/common/entropy_common.c

@@ -1,6 +1,6 @@
 /* ******************************************************************
  * Common functions of New Generation Entropy library
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  *  You can contact the author at :
  *  - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -19,8 +19,8 @@
 #include "error_private.h"       /* ERR_*, ERROR */
 #define FSE_STATIC_LINKING_ONLY  /* FSE_MIN_TABLELOG */
 #include "fse.h"
-#define HUF_STATIC_LINKING_ONLY  /* HUF_TABLELOG_ABSOLUTEMAX */
 #include "huf.h"
+#include "bits.h"                /* ZSDT_highbit32, ZSTD_countTrailingZeros32 */
 
 
 /*===   Version   ===*/
@@ -38,8 +38,9 @@ const char* HUF_getErrorName(size_t code) { return ERR_getErrorName(code); }
 /*-**************************************************************
 *  FSE NCount encoding-decoding
 ****************************************************************/
-size_t FSE_readNCount (short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
-                 const void* headerBuffer, size_t hbSize)
+FORCE_INLINE_TEMPLATE
+size_t FSE_readNCount_body(short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+                           const void* headerBuffer, size_t hbSize)
 {
     const BYTE* const istart = (const BYTE*) headerBuffer;
     const BYTE* const iend = istart + hbSize;
@@ -50,23 +51,23 @@ size_t FSE_readNCount (short* normalizedCounter, unsigned* maxSVPtr, unsigned* t
     U32 bitStream;
     int bitCount;
     unsigned charnum = 0;
+    unsigned const maxSV1 = *maxSVPtr + 1;
     int previous0 = 0;
 
-    if (hbSize < 4) {
-        /* This function only works when hbSize >= 4 */
-        char buffer[4];
-        memset(buffer, 0, sizeof(buffer));
-        memcpy(buffer, headerBuffer, hbSize);
+    if (hbSize < 8) {
+        /* This function only works when hbSize >= 8 */
+        char buffer[8] = {0};
+        ZSTD_memcpy(buffer, headerBuffer, hbSize);
         {   size_t const countSize = FSE_readNCount(normalizedCounter, maxSVPtr, tableLogPtr,
                                                     buffer, sizeof(buffer));
             if (FSE_isError(countSize)) return countSize;
             if (countSize > hbSize) return ERROR(corruption_detected);
             return countSize;
     }   }
-    assert(hbSize >= 4);
+    assert(hbSize >= 8);
 
     /* init */
-    memset(normalizedCounter, 0, (*maxSVPtr+1) * sizeof(normalizedCounter[0]));   /* all symbols not present in NCount have a frequency of 0 */
+    ZSTD_memset(normalizedCounter, 0, (*maxSVPtr+1) * sizeof(normalizedCounter[0]));   /* all symbols not present in NCount have a frequency of 0 */
     bitStream = MEM_readLE32(ip);
     nbBits = (bitStream & 0xF) + FSE_MIN_TABLELOG;   /* extract tableLog */
     if (nbBits > FSE_TABLELOG_ABSOLUTE_MAX) return ERROR(tableLog_tooLarge);
@@ -77,36 +78,58 @@ size_t FSE_readNCount (short* normalizedCounter, unsigned* maxSVPtr, unsigned* t
     threshold = 1<<nbBits;
     nbBits++;
 
-    while ((remaining>1) & (charnum<=*maxSVPtr)) {
+    for (;;) {
         if (previous0) {
-            unsigned n0 = charnum;
-            while ((bitStream & 0xFFFF) == 0xFFFF) {
-                n0 += 24;
-                if (ip < iend-5) {
-                    ip += 2;
-                    bitStream = MEM_readLE32(ip) >> bitCount;
+            /* Count the number of repeats. Each time the
+             * 2-bit repeat code is 0b11 there is another
+             * repeat.
+             * Avoid UB by setting the high bit to 1.
+             */
+            int repeats = ZSTD_countTrailingZeros32(~bitStream | 0x80000000) >> 1;
+            while (repeats >= 12) {
+                charnum += 3 * 12;
+                if (LIKELY(ip <= iend-7)) {
+                    ip += 3;
                 } else {
-                    bitStream >>= 16;
-                    bitCount   += 16;
-            }   }
-            while ((bitStream & 3) == 3) {
-                n0 += 3;
-                bitStream >>= 2;
-                bitCount += 2;
+                    bitCount -= (int)(8 * (iend - 7 - ip));
+                    bitCount &= 31;
+                    ip = iend - 4;
+                }
+                bitStream = MEM_readLE32(ip) >> bitCount;
+                repeats = ZSTD_countTrailingZeros32(~bitStream | 0x80000000) >> 1;
             }
-            n0 += bitStream & 3;
+            charnum += 3 * repeats;
+            bitStream >>= 2 * repeats;
+            bitCount += 2 * repeats;
+
+            /* Add the final repeat which isn't 0b11. */
+            assert((bitStream & 3) < 3);
+            charnum += bitStream & 3;
             bitCount += 2;
-            if (n0 > *maxSVPtr) return ERROR(maxSymbolValue_tooSmall);
-            while (charnum < n0) normalizedCounter[charnum++] = 0;
-            if ((ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
+
+            /* This is an error, but break and return an error
+             * at the end, because returning out of a loop makes
+             * it harder for the compiler to optimize.
+             */
+            if (charnum >= maxSV1) break;
+
+            /* We don't need to set the normalized count to 0
+             * because we already memset the whole buffer to 0.
+             */
+
+            if (LIKELY(ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
                 assert((bitCount >> 3) <= 3); /* For first condition to work */
                 ip += bitCount>>3;
                 bitCount &= 7;
-                bitStream = MEM_readLE32(ip) >> bitCount;
             } else {
-                bitStream >>= 2;
-        }   }
-        {   int const max = (2*threshold-1) - remaining;
+                bitCount -= (int)(8 * (iend - 4 - ip));
+                bitCount &= 31;
+                ip = iend - 4;
+            }
+            bitStream = MEM_readLE32(ip) >> bitCount;
+        }
+        {
+            int const max = (2*threshold-1) - remaining;
             int count;
 
             if ((bitStream & (threshold-1)) < (U32)max) {
@@ -119,24 +142,43 @@ size_t FSE_readNCount (short* normalizedCounter, unsigned* maxSVPtr, unsigned* t
             }
 
             count--;   /* extra accuracy */
-            remaining -= count < 0 ? -count : count;   /* -1 means +1 */
+            /* When it matters (small blocks), this is a
+             * predictable branch, because we don't use -1.
+             */
+            if (count >= 0) {
+                remaining -= count;
+            } else {
+                assert(count == -1);
+                remaining += count;
+            }
             normalizedCounter[charnum++] = (short)count;
             previous0 = !count;
-            while (remaining < threshold) {
-                nbBits--;
-                threshold >>= 1;
+
+            assert(threshold > 1);
+            if (remaining < threshold) {
+                /* This branch can be folded into the
+                 * threshold update condition because we
+                 * know that threshold > 1.
+                 */
+                if (remaining <= 1) break;
+                nbBits = ZSTD_highbit32(remaining) + 1;
+                threshold = 1 << (nbBits - 1);
             }
+            if (charnum >= maxSV1) break;
 
-            if ((ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
+            if (LIKELY(ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
                 ip += bitCount>>3;
                 bitCount &= 7;
             } else {
                 bitCount -= (int)(8 * (iend - 4 - ip));
+                bitCount &= 31;
                 ip = iend - 4;
             }
-            bitStream = MEM_readLE32(ip) >> (bitCount & 31);
-    }   }   /* while ((remaining>1) & (charnum<=*maxSVPtr)) */
+            bitStream = MEM_readLE32(ip) >> bitCount;
+    }   }
     if (remaining != 1) return ERROR(corruption_detected);
+    /* Only possible when there are too many zeros. */
+    if (charnum > maxSV1) return ERROR(maxSymbolValue_tooSmall);
     if (bitCount > 32) return ERROR(corruption_detected);
     *maxSVPtr = charnum-1;
 
@@ -144,6 +186,43 @@ size_t FSE_readNCount (short* normalizedCounter, unsigned* maxSVPtr, unsigned* t
     return ip-istart;
 }
 
+/* Avoids the FORCE_INLINE of the _body() function. */
+static size_t FSE_readNCount_body_default(
+        short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+        const void* headerBuffer, size_t hbSize)
+{
+    return FSE_readNCount_body(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
+}
+
+#if DYNAMIC_BMI2
+BMI2_TARGET_ATTRIBUTE static size_t FSE_readNCount_body_bmi2(
+        short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+        const void* headerBuffer, size_t hbSize)
+{
+    return FSE_readNCount_body(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
+}
+#endif
+
+size_t FSE_readNCount_bmi2(
+        short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+        const void* headerBuffer, size_t hbSize, int bmi2)
+{
+#if DYNAMIC_BMI2
+    if (bmi2) {
+        return FSE_readNCount_body_bmi2(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
+    }
+#endif
+    (void)bmi2;
+    return FSE_readNCount_body_default(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
+}
+
+size_t FSE_readNCount(
+        short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
+        const void* headerBuffer, size_t hbSize)
+{
+    return FSE_readNCount_bmi2(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize, /* bmi2 */ 0);
+}
+
 
 /*! HUF_readStats() :
     Read compact Huffman tree, saved by HUF_writeCTable().
@@ -155,6 +234,17 @@ size_t FSE_readNCount (short* normalizedCounter, unsigned* maxSVPtr, unsigned* t
 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;
@@ -163,7 +253,7 @@ size_t HUF_readStats(BYTE* huffWeight, size_t hwSize, U32* rankStats,
 
     if (!srcSize) return ERROR(srcSize_wrong);
     iSize = ip[0];
-    /* memset(huffWeight, 0, hwSize);   *//* is not necessary, even though some analyzer complain ... */
+    /* ZSTD_memset(huffWeight, 0, hwSize);   *//* is not necessary, even though some analyzer complain ... */
 
     if (iSize >= 128) {  /* special header */
         oSize = iSize - 127;
@@ -177,31 +267,31 @@ size_t HUF_readStats(BYTE* huffWeight, size_t hwSize, U32* rankStats,
                 huffWeight[n+1] = ip[n/2] & 15;
     }   }   }
     else  {   /* header compressed with FSE (normal case) */
-        FSE_DTable fseWorkspace[FSE_DTABLE_SIZE_U32(6)];  /* 6 is max possible tableLog for HUF header (maybe even 5, to be tested) */
         if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
-        oSize = FSE_decompress_wksp(huffWeight, hwSize-1, ip+1, iSize, fseWorkspace, 6);   /* max (hwSize-1) values decoded, as last one is implied */
+        /* max (hwSize-1) values decoded, as last one is implied */
+        oSize = FSE_decompress_wksp_bmi2(huffWeight, hwSize-1, ip+1, iSize, 6, workSpace, wkspSize, bmi2);
         if (FSE_isError(oSize)) return oSize;
     }
 
     /* collect weight stats */
-    memset(rankStats, 0, (HUF_TABLELOG_MAX + 1) * sizeof(U32));
+    ZSTD_memset(rankStats, 0, (HUF_TABLELOG_MAX + 1) * sizeof(U32));
     weightTotal = 0;
     {   U32 n; for (n=0; n<oSize; n++) {
-            if (huffWeight[n] >= HUF_TABLELOG_MAX) return ERROR(corruption_detected);
+            if (huffWeight[n] > HUF_TABLELOG_MAX) return ERROR(corruption_detected);
             rankStats[huffWeight[n]]++;
             weightTotal += (1 << huffWeight[n]) >> 1;
     }   }
     if (weightTotal == 0) return ERROR(corruption_detected);
 
     /* get last non-null symbol weight (implied, total must be 2^n) */
-    {   U32 const tableLog = BIT_highbit32(weightTotal) + 1;
+    {   U32 const tableLog = ZSTD_highbit32(weightTotal) + 1;
         if (tableLog > HUF_TABLELOG_MAX) return ERROR(corruption_detected);
         *tableLogPtr = tableLog;
         /* determine last weight */
         {   U32 const total = 1 << tableLog;
             U32 const rest = total - weightTotal;
-            U32 const verif = 1 << BIT_highbit32(rest);
-            U32 const lastWeight = BIT_highbit32(rest) + 1;
+            U32 const verif = 1 << ZSTD_highbit32(rest);
+            U32 const lastWeight = ZSTD_highbit32(rest) + 1;
             if (verif != rest) return ERROR(corruption_detected);    /* last value must be a clean power of 2 */
             huffWeight[oSize] = (BYTE)lastWeight;
             rankStats[lastWeight]++;
@@ -214,3 +304,37 @@ size_t HUF_readStats(BYTE* huffWeight, size_t hwSize, U32* rankStats,
     *nbSymbolsPtr = (U32)(oSize+1);
     return iSize+1;
 }
+
+/* Avoids the FORCE_INLINE of the _body() function. */
+static size_t HUF_readStats_body_default(BYTE* huffWeight, size_t hwSize, U32* rankStats,
+                     U32* nbSymbolsPtr, U32* tableLogPtr,
+                     const void* src, size_t srcSize,
+                     void* workSpace, size_t wkspSize)
+{
+    return HUF_readStats_body(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize, 0);
+}
+
+#if DYNAMIC_BMI2
+static BMI2_TARGET_ATTRIBUTE size_t HUF_readStats_body_bmi2(BYTE* huffWeight, size_t hwSize, U32* rankStats,
+                     U32* nbSymbolsPtr, U32* tableLogPtr,
+                     const void* src, size_t srcSize,
+                     void* workSpace, size_t wkspSize)
+{
+    return HUF_readStats_body(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize, 1);
+}
+#endif
+
+size_t HUF_readStats_wksp(BYTE* huffWeight, size_t hwSize, U32* rankStats,
+                     U32* nbSymbolsPtr, U32* tableLogPtr,
+                     const void* src, size_t srcSize,
+                     void* workSpace, size_t wkspSize,
+                     int flags)
+{
+#if DYNAMIC_BMI2
+    if (flags & HUF_flags_bmi2) {
+        return HUF_readStats_body_bmi2(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize);
+    }
+#endif
+    (void)flags;
+    return HUF_readStats_body_default(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize);
+}

src/borg/algorithms/zstd/lib/common/error_private.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -27,9 +27,11 @@ const char* ERR_getErrorString(ERR_enum code)
     case PREFIX(version_unsupported): return "Version not supported";
     case PREFIX(frameParameter_unsupported): return "Unsupported frame parameter";
     case PREFIX(frameParameter_windowTooLarge): return "Frame requires too much memory for decoding";
-    case PREFIX(corruption_detected): return "Corrupted block detected";
+    case PREFIX(corruption_detected): return "Data corruption detected";
     case PREFIX(checksum_wrong): return "Restored data doesn't match checksum";
+    case PREFIX(literals_headerWrong): return "Header of Literals' block doesn't respect format specification";
     case PREFIX(parameter_unsupported): return "Unsupported parameter";
+    case PREFIX(parameter_combination_unsupported): return "Unsupported combination of parameters";
     case PREFIX(parameter_outOfBound): return "Parameter is out of bound";
     case PREFIX(init_missing): return "Context should be init first";
     case PREFIX(memory_allocation): return "Allocation error : not enough memory";
@@ -38,16 +40,22 @@ const char* ERR_getErrorString(ERR_enum code)
     case PREFIX(tableLog_tooLarge): return "tableLog requires too much memory : unsupported";
     case PREFIX(maxSymbolValue_tooLarge): return "Unsupported max Symbol Value : too large";
     case PREFIX(maxSymbolValue_tooSmall): return "Specified maxSymbolValue is too small";
+    case PREFIX(stabilityCondition_notRespected): return "pledged buffer stability condition is not respected";
     case PREFIX(dictionary_corrupted): return "Dictionary is corrupted";
     case PREFIX(dictionary_wrong): return "Dictionary mismatch";
     case PREFIX(dictionaryCreation_failed): return "Cannot create Dictionary from provided samples";
     case PREFIX(dstSize_tooSmall): return "Destination buffer is too small";
     case PREFIX(srcSize_wrong): return "Src size is incorrect";
     case PREFIX(dstBuffer_null): return "Operation on NULL destination buffer";
+    case PREFIX(noForwardProgress_destFull): return "Operation made no progress over multiple calls, due to output buffer being full";
+    case PREFIX(noForwardProgress_inputEmpty): return "Operation made no progress over multiple calls, due to input being empty";
         /* following error codes are not stable and may be removed or changed in a future version */
     case PREFIX(frameIndex_tooLarge): return "Frame index is too large";
     case PREFIX(seekableIO): return "An I/O error occurred when reading/seeking";
     case PREFIX(dstBuffer_wrong): return "Destination buffer is wrong";
+    case PREFIX(srcBuffer_wrong): return "Source buffer is wrong";
+    case PREFIX(sequenceProducer_failed): return "Block-level external sequence producer returned an error code";
+    case PREFIX(externalSequences_invalid): return "External sequences are not valid";
     case PREFIX(maxCode):
     default: return notErrorCode;
     }

src/borg/algorithms/zstd/lib/common/error_private.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -21,8 +21,10 @@ extern "C" {
 /* ****************************************
 *  Dependencies
 ******************************************/
-#include <stddef.h>        /* size_t */
-#include "zstd_errors.h"  /* enum list */
+#include "../zstd_errors.h"  /* enum list */
+#include "compiler.h"
+#include "debug.h"
+#include "zstd_deps.h"       /* size_t */
 
 
 /* ****************************************
@@ -73,6 +75,83 @@ ERR_STATIC const char* ERR_getErrorName(size_t code)
     return ERR_getErrorString(ERR_getErrorCode(code));
 }
 
+/**
+ * Ignore: this is an internal helper.
+ *
+ * This is a helper function to help force C99-correctness during compilation.
+ * Under strict compilation modes, variadic macro arguments can't be empty.
+ * However, variadic function arguments can be. Using a function therefore lets
+ * us statically check that at least one (string) argument was passed,
+ * independent of the compilation flags.
+ */
+static INLINE_KEYWORD UNUSED_ATTR
+void _force_has_format_string(const char *format, ...) {
+  (void)format;
+}
+
+/**
+ * Ignore: this is an internal helper.
+ *
+ * We want to force this function invocation to be syntactically correct, but
+ * we don't want to force runtime evaluation of its arguments.
+ */
+#define _FORCE_HAS_FORMAT_STRING(...) \
+  if (0) { \
+    _force_has_format_string(__VA_ARGS__); \
+  }
+
+#define ERR_QUOTE(str) #str
+
+/**
+ * Return the specified error if the condition evaluates to true.
+ *
+ * In debug modes, prints additional information.
+ * In order to do that (particularly, printing the conditional that failed),
+ * this can't just wrap RETURN_ERROR().
+ */
+#define RETURN_ERROR_IF(cond, err, ...) \
+  if (cond) { \
+    RAWLOG(3, "%s:%d: ERROR!: check %s failed, returning %s", \
+           __FILE__, __LINE__, ERR_QUOTE(cond), ERR_QUOTE(ERROR(err))); \
+    _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
+    RAWLOG(3, ": " __VA_ARGS__); \
+    RAWLOG(3, "\n"); \
+    return ERROR(err); \
+  }
+
+/**
+ * Unconditionally return the specified error.
+ *
+ * In debug modes, prints additional information.
+ */
+#define RETURN_ERROR(err, ...) \
+  do { \
+    RAWLOG(3, "%s:%d: ERROR!: unconditional check failed, returning %s", \
+           __FILE__, __LINE__, ERR_QUOTE(ERROR(err))); \
+    _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
+    RAWLOG(3, ": " __VA_ARGS__); \
+    RAWLOG(3, "\n"); \
+    return ERROR(err); \
+  } while(0);
+
+/**
+ * If the provided expression evaluates to an error code, returns that error code.
+ *
+ * In debug modes, prints additional information.
+ */
+#define FORWARD_IF_ERROR(err, ...) \
+  do { \
+    size_t const err_code = (err); \
+    if (ERR_isError(err_code)) { \
+      RAWLOG(3, "%s:%d: ERROR!: forwarding error in %s: %s", \
+             __FILE__, __LINE__, ERR_QUOTE(err), ERR_getErrorName(err_code)); \
+      _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
+      RAWLOG(3, ": " __VA_ARGS__); \
+      RAWLOG(3, "\n"); \
+      return err_code; \
+    } \
+  } while(0);
+
 #if defined (__cplusplus)
 }
 #endif

src/borg/algorithms/zstd/lib/common/fse.h

@@ -1,7 +1,7 @@
 /* ******************************************************************
  * FSE : Finite State Entropy codec
  * Public Prototypes declaration
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -23,7 +23,7 @@ extern "C" {
 /*-*****************************************
 *  Dependencies
 ******************************************/
-#include <stddef.h>    /* size_t, ptrdiff_t */
+#include "zstd_deps.h"    /* size_t, ptrdiff_t */
 
 
 /*-*****************************************
@@ -53,34 +53,6 @@ extern "C" {
 FSE_PUBLIC_API unsigned FSE_versionNumber(void);   /**< library version number; to be used when checking dll version */
 
 
-/*-****************************************
-*  FSE simple functions
-******************************************/
-/*! FSE_compress() :
-    Compress content of buffer 'src', of size 'srcSize', into destination buffer 'dst'.
-    'dst' buffer must be already allocated. Compression runs faster is dstCapacity >= FSE_compressBound(srcSize).
-    @return : size of compressed data (<= dstCapacity).
-    Special values : if return == 0, srcData is not compressible => Nothing is stored within dst !!!
-                     if return == 1, srcData is a single byte symbol * srcSize times. Use RLE compression instead.
-                     if FSE_isError(return), compression failed (more details using FSE_getErrorName())
-*/
-FSE_PUBLIC_API size_t FSE_compress(void* dst, size_t dstCapacity,
-                             const void* src, size_t srcSize);
-
-/*! FSE_decompress():
-    Decompress FSE data from buffer 'cSrc', of size 'cSrcSize',
-    into already allocated destination buffer 'dst', of size 'dstCapacity'.
-    @return : size of regenerated data (<= maxDstSize),
-              or an error code, which can be tested using FSE_isError() .
-
-    ** Important ** : FSE_decompress() does not decompress non-compressible nor RLE data !!!
-    Why ? : making this distinction requires a header.
-    Header management is intentionally delegated to the user layer, which can better manage special cases.
-*/
-FSE_PUBLIC_API size_t FSE_decompress(void* dst,  size_t dstCapacity,
-                               const void* cSrc, size_t cSrcSize);
-
-
 /*-*****************************************
 *  Tool functions
 ******************************************/
@@ -91,20 +63,6 @@ FSE_PUBLIC_API unsigned    FSE_isError(size_t code);        /* tells if a return
 FSE_PUBLIC_API const char* FSE_getErrorName(size_t code);   /* provides error code string (useful for debugging) */
 
 
-/*-*****************************************
-*  FSE advanced functions
-******************************************/
-/*! FSE_compress2() :
-    Same as FSE_compress(), but allows the selection of 'maxSymbolValue' and 'tableLog'
-    Both parameters can be defined as '0' to mean : use default value
-    @return : size of compressed data
-    Special values : if return == 0, srcData is not compressible => Nothing is stored within cSrc !!!
-                     if return == 1, srcData is a single byte symbol * srcSize times. Use RLE compression.
-                     if FSE_isError(return), it's an error code.
-*/
-FSE_PUBLIC_API size_t FSE_compress2 (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog);
-
-
 /*-*****************************************
 *  FSE detailed API
 ******************************************/
@@ -137,10 +95,16 @@ FSE_PUBLIC_API unsigned FSE_optimalTableLog(unsigned maxTableLog, size_t srcSize
 /*! FSE_normalizeCount():
     normalize counts so that sum(count[]) == Power_of_2 (2^tableLog)
     'normalizedCounter' is a table of short, of minimum size (maxSymbolValue+1).
+    useLowProbCount is a boolean parameter which trades off compressed size for
+    faster header decoding. When it is set to 1, the compressed data will be slightly
+    smaller. And when it is set to 0, FSE_readNCount() and FSE_buildDTable() will be
+    faster. If you are compressing a small amount of data (< 2 KB) then useLowProbCount=0
+    is a good default, since header deserialization makes a big speed difference.
+    Otherwise, useLowProbCount=1 is a good default, since the speed difference is small.
     @return : tableLog,
               or an errorCode, which can be tested using FSE_isError() */
 FSE_PUBLIC_API size_t FSE_normalizeCount(short* normalizedCounter, unsigned tableLog,
-                    const unsigned* count, size_t srcSize, unsigned maxSymbolValue);
+                    const unsigned* count, size_t srcSize, unsigned maxSymbolValue, unsigned useLowProbCount);
 
 /*! FSE_NCountWriteBound():
     Provides the maximum possible size of an FSE normalized table, given 'maxSymbolValue' and 'tableLog'.
@@ -158,8 +122,6 @@ FSE_PUBLIC_API size_t FSE_writeNCount (void* buffer, size_t bufferSize,
 /*! Constructor and Destructor of FSE_CTable.
     Note that FSE_CTable size depends on 'tableLog' and 'maxSymbolValue' */
 typedef unsigned FSE_CTable;   /* don't allocate that. It's only meant to be more restrictive than void* */
-FSE_PUBLIC_API FSE_CTable* FSE_createCTable (unsigned maxSymbolValue, unsigned tableLog);
-FSE_PUBLIC_API void        FSE_freeCTable (FSE_CTable* ct);
 
 /*! FSE_buildCTable():
     Builds `ct`, which must be already allocated, using FSE_createCTable().
@@ -228,23 +190,14 @@ FSE_PUBLIC_API size_t FSE_readNCount (short* normalizedCounter,
                            unsigned* maxSymbolValuePtr, unsigned* tableLogPtr,
                            const void* rBuffer, size_t rBuffSize);
 
-/*! Constructor and Destructor of FSE_DTable.
-    Note that its size depends on 'tableLog' */
+/*! FSE_readNCount_bmi2():
+ * Same as FSE_readNCount() but pass bmi2=1 when your CPU supports BMI2 and 0 otherwise.
+ */
+FSE_PUBLIC_API size_t FSE_readNCount_bmi2(short* normalizedCounter,
+                           unsigned* maxSymbolValuePtr, unsigned* tableLogPtr,
+                           const void* rBuffer, size_t rBuffSize, int bmi2);
+
 typedef unsigned FSE_DTable;   /* don't allocate that. It's just a way to be more restrictive than void* */
-FSE_PUBLIC_API FSE_DTable* FSE_createDTable(unsigned tableLog);
-FSE_PUBLIC_API void        FSE_freeDTable(FSE_DTable* dt);
-
-/*! FSE_buildDTable():
-    Builds 'dt', which must be already allocated, using FSE_createDTable().
-    return : 0, or an errorCode, which can be tested using FSE_isError() */
-FSE_PUBLIC_API size_t FSE_buildDTable (FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog);
-
-/*! FSE_decompress_usingDTable():
-    Decompress compressed source `cSrc` of size `cSrcSize` using `dt`
-    into `dst` which must be already allocated.
-    @return : size of regenerated data (necessarily <= `dstCapacity`),
-              or an errorCode, which can be tested using FSE_isError() */
-FSE_PUBLIC_API size_t FSE_decompress_usingDTable(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, const FSE_DTable* dt);
 
 /*!
 Tutorial :
@@ -288,12 +241,12 @@ If there is an error, the function will return an error code, which can be teste
 *******************************************/
 /* FSE buffer bounds */
 #define FSE_NCOUNTBOUND 512
-#define FSE_BLOCKBOUND(size) (size + (size>>7) + 4 /* fse states */ + sizeof(size_t) /* bitContainer */)
+#define FSE_BLOCKBOUND(size) ((size) + ((size)>>7) + 4 /* fse states */ + sizeof(size_t) /* bitContainer */)
 #define FSE_COMPRESSBOUND(size) (FSE_NCOUNTBOUND + FSE_BLOCKBOUND(size))   /* Macro version, useful for static allocation */
 
 /* It is possible to statically allocate FSE CTable/DTable as a table of FSE_CTable/FSE_DTable using below macros */
-#define FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue)   (1 + (1<<(maxTableLog-1)) + ((maxSymbolValue+1)*2))
-#define FSE_DTABLE_SIZE_U32(maxTableLog)                   (1 + (1<<maxTableLog))
+#define FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue)   (1 + (1<<((maxTableLog)-1)) + (((maxSymbolValue)+1)*2))
+#define FSE_DTABLE_SIZE_U32(maxTableLog)                   (1 + (1<<(maxTableLog)))
 
 /* or use the size to malloc() space directly. Pay attention to alignment restrictions though */
 #define FSE_CTABLE_SIZE(maxTableLog, maxSymbolValue)   (FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue) * sizeof(FSE_CTable))
@@ -307,33 +260,28 @@ If there is an error, the function will return an error code, which can be teste
 unsigned FSE_optimalTableLog_internal(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, unsigned minus);
 /**< same as FSE_optimalTableLog(), which used `minus==2` */
 
-/* FSE_compress_wksp() :
- * Same as FSE_compress2(), but using an externally allocated scratch buffer (`workSpace`).
- * FSE_WKSP_SIZE_U32() provides the minimum size required for `workSpace` as a table of FSE_CTable.
- */
-#define FSE_WKSP_SIZE_U32(maxTableLog, maxSymbolValue)   ( FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue) + ((maxTableLog > 12) ? (1 << (maxTableLog - 2)) : 1024) )
-size_t FSE_compress_wksp (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);
-
-size_t FSE_buildCTable_raw (FSE_CTable* ct, unsigned nbBits);
-/**< build a fake FSE_CTable, designed for a flat distribution, where each symbol uses nbBits */
-
 size_t FSE_buildCTable_rle (FSE_CTable* ct, unsigned char symbolValue);
 /**< build a fake FSE_CTable, designed to compress always the same symbolValue */
 
 /* FSE_buildCTable_wksp() :
  * Same as FSE_buildCTable(), but using an externally allocated scratch buffer (`workSpace`).
- * `wkspSize` must be >= `(1<<tableLog)`.
+ * `wkspSize` must be >= `FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(maxSymbolValue, tableLog)` of `unsigned`.
+ * See FSE_buildCTable_wksp() for breakdown of workspace usage.
  */
+#define FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(maxSymbolValue, tableLog) (((maxSymbolValue + 2) + (1ull << (tableLog)))/2 + sizeof(U64)/sizeof(U32) /* additional 8 bytes for potential table overwrite */)
+#define FSE_BUILD_CTABLE_WORKSPACE_SIZE(maxSymbolValue, tableLog) (sizeof(unsigned) * FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(maxSymbolValue, tableLog))
 size_t FSE_buildCTable_wksp(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);
 
-size_t FSE_buildDTable_raw (FSE_DTable* dt, unsigned nbBits);
-/**< build a fake FSE_DTable, designed to read a flat distribution where each symbol uses nbBits */
+#define FSE_BUILD_DTABLE_WKSP_SIZE(maxTableLog, maxSymbolValue) (sizeof(short) * (maxSymbolValue + 1) + (1ULL << maxTableLog) + 8)
+#define FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) ((FSE_BUILD_DTABLE_WKSP_SIZE(maxTableLog, maxSymbolValue) + sizeof(unsigned) - 1) / sizeof(unsigned))
+FSE_PUBLIC_API size_t FSE_buildDTable_wksp(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);
+/**< Same as FSE_buildDTable(), using an externally allocated `workspace` produced with `FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxSymbolValue)` */
 
-size_t FSE_buildDTable_rle (FSE_DTable* dt, unsigned char symbolValue);
-/**< build a fake FSE_DTable, designed to always generate the same symbolValue */
-
-size_t FSE_decompress_wksp(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, FSE_DTable* workSpace, unsigned maxLog);
-/**< same as FSE_decompress(), using an externally allocated `workSpace` produced with `FSE_DTABLE_SIZE_U32(maxLog)` */
+#define FSE_DECOMPRESS_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) (FSE_DTABLE_SIZE_U32(maxTableLog) + 1 + FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) + (FSE_MAX_SYMBOL_VALUE + 1) / 2 + 1)
+#define FSE_DECOMPRESS_WKSP_SIZE(maxTableLog, maxSymbolValue) (FSE_DECOMPRESS_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) * sizeof(unsigned))
+size_t FSE_decompress_wksp_bmi2(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize, int bmi2);
+/**< same as FSE_decompress(), using an externally allocated `workSpace` produced with `FSE_DECOMPRESS_WKSP_SIZE_U32(maxLog, maxSymbolValue)`.
+ * Set bmi2 to 1 if your CPU supports BMI2 or 0 if it doesn't */
 
 typedef enum {
    FSE_repeat_none,  /**< Cannot use the previous table */
@@ -529,7 +477,7 @@ MEM_STATIC void FSE_flushCState(BIT_CStream_t* bitC, const FSE_CState_t* statePt
 
 /* FSE_getMaxNbBits() :
  * Approximate maximum cost of a symbol, in bits.
- * Fractional get rounded up (i.e : a symbol with a normalized frequency of 3 gives the same result as a frequency of 2)
+ * Fractional get rounded up (i.e. a symbol with a normalized frequency of 3 gives the same result as a frequency of 2)
  * note 1 : assume symbolValue is valid (<= maxSymbolValue)
  * note 2 : if freq[symbolValue]==0, @return a fake cost of tableLog+1 bits */
 MEM_STATIC U32 FSE_getMaxNbBits(const void* symbolTTPtr, U32 symbolValue)
@@ -644,6 +592,9 @@ MEM_STATIC unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr)
 #ifndef FSE_DEFAULT_MEMORY_USAGE
 #  define FSE_DEFAULT_MEMORY_USAGE 13
 #endif
+#if (FSE_DEFAULT_MEMORY_USAGE > FSE_MAX_MEMORY_USAGE)
+#  error "FSE_DEFAULT_MEMORY_USAGE must be <= FSE_MAX_MEMORY_USAGE"
+#endif
 
 /*!FSE_MAX_SYMBOL_VALUE :
 *  Maximum symbol value authorized.
@@ -677,7 +628,7 @@ MEM_STATIC unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr)
 #  error "FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX is not supported"
 #endif
 
-#define FSE_TABLESTEP(tableSize) ((tableSize>>1) + (tableSize>>3) + 3)
+#define FSE_TABLESTEP(tableSize) (((tableSize)>>1) + ((tableSize)>>3) + 3)
 
 
 #endif /* FSE_STATIC_LINKING_ONLY */

src/borg/algorithms/zstd/lib/common/fse_decompress.c

@@ -1,6 +1,6 @@
 /* ******************************************************************
  * FSE : Finite State Entropy decoder
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  *  You can contact the author at :
  *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -16,13 +16,15 @@
 /* **************************************************************
 *  Includes
 ****************************************************************/
-#include <stdlib.h>     /* malloc, free, qsort */
-#include <string.h>     /* memcpy, memset */
+#include "debug.h"      /* assert */
 #include "bitstream.h"
 #include "compiler.h"
 #define FSE_STATIC_LINKING_ONLY
 #include "fse.h"
 #include "error_private.h"
+#define ZSTD_DEPS_NEED_MALLOC
+#include "zstd_deps.h"
+#include "bits.h"       /* ZSTD_highbit32 */
 
 
 /* **************************************************************
@@ -54,30 +56,19 @@
 #define FSE_FUNCTION_NAME(X,Y) FSE_CAT(X,Y)
 #define FSE_TYPE_NAME(X,Y) FSE_CAT(X,Y)
 
-
-/* Function templates */
-FSE_DTable* FSE_createDTable (unsigned tableLog)
-{
-    if (tableLog > FSE_TABLELOG_ABSOLUTE_MAX) tableLog = FSE_TABLELOG_ABSOLUTE_MAX;
-    return (FSE_DTable*)malloc( FSE_DTABLE_SIZE_U32(tableLog) * sizeof (U32) );
-}
-
-void FSE_freeDTable (FSE_DTable* dt)
-{
-    free(dt);
-}
-
-size_t FSE_buildDTable(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog)
+static size_t FSE_buildDTable_internal(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize)
 {
     void* const tdPtr = dt+1;   /* because *dt is unsigned, 32-bits aligned on 32-bits */
     FSE_DECODE_TYPE* const tableDecode = (FSE_DECODE_TYPE*) (tdPtr);
-    U16 symbolNext[FSE_MAX_SYMBOL_VALUE+1];
+    U16* symbolNext = (U16*)workSpace;
+    BYTE* spread = (BYTE*)(symbolNext + maxSymbolValue + 1);
 
     U32 const maxSV1 = maxSymbolValue + 1;
     U32 const tableSize = 1 << tableLog;
     U32 highThreshold = tableSize-1;
 
     /* Sanity Checks */
+    if (FSE_BUILD_DTABLE_WKSP_SIZE(tableLog, maxSymbolValue) > wkspSize) return ERROR(maxSymbolValue_tooLarge);
     if (maxSymbolValue > FSE_MAX_SYMBOL_VALUE) return ERROR(maxSymbolValue_tooLarge);
     if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);
 
@@ -95,11 +86,57 @@ size_t FSE_buildDTable(FSE_DTable* dt, const short* normalizedCounter, unsigned
                     if (normalizedCounter[s] >= largeLimit) DTableH.fastMode=0;
                     symbolNext[s] = normalizedCounter[s];
         }   }   }
-        memcpy(dt, &DTableH, sizeof(DTableH));
+        ZSTD_memcpy(dt, &DTableH, sizeof(DTableH));
     }
 
     /* Spread symbols */
-    {   U32 const tableMask = tableSize-1;
+    if (highThreshold == tableSize - 1) {
+        size_t const tableMask = tableSize-1;
+        size_t const step = FSE_TABLESTEP(tableSize);
+        /* First lay down the symbols in order.
+         * We use a uint64_t to lay down 8 bytes at a time. This reduces branch
+         * misses since small blocks generally have small table logs, so nearly
+         * all symbols have counts <= 8. We ensure we have 8 bytes at the end of
+         * our buffer to handle the over-write.
+         */
+        {
+            U64 const add = 0x0101010101010101ull;
+            size_t pos = 0;
+            U64 sv = 0;
+            U32 s;
+            for (s=0; s<maxSV1; ++s, sv += add) {
+                int i;
+                int const n = normalizedCounter[s];
+                MEM_write64(spread + pos, sv);
+                for (i = 8; i < n; i += 8) {
+                    MEM_write64(spread + pos + i, sv);
+                }
+                pos += n;
+            }
+        }
+        /* Now we spread those positions across the table.
+         * The benefit of doing it in two stages is that we avoid the
+         * variable size inner loop, which caused lots of branch misses.
+         * Now we can run through all the positions without any branch misses.
+         * We unroll the loop twice, since that is what empirically worked best.
+         */
+        {
+            size_t position = 0;
+            size_t s;
+            size_t const unroll = 2;
+            assert(tableSize % unroll == 0); /* FSE_MIN_TABLELOG is 5 */
+            for (s = 0; s < (size_t)tableSize; s += unroll) {
+                size_t u;
+                for (u = 0; u < unroll; ++u) {
+                    size_t const uPosition = (position + (u * step)) & tableMask;
+                    tableDecode[uPosition].symbol = spread[s + u];
+                }
+                position = (position + (unroll * step)) & tableMask;
+            }
+            assert(position == 0);
+        }
+    } else {
+        U32 const tableMask = tableSize-1;
         U32 const step = FSE_TABLESTEP(tableSize);
         U32 s, position = 0;
         for (s=0; s<maxSV1; s++) {
@@ -117,62 +154,24 @@ size_t FSE_buildDTable(FSE_DTable* dt, const short* normalizedCounter, unsigned
         for (u=0; u<tableSize; u++) {
             FSE_FUNCTION_TYPE const symbol = (FSE_FUNCTION_TYPE)(tableDecode[u].symbol);
             U32 const nextState = symbolNext[symbol]++;
-            tableDecode[u].nbBits = (BYTE) (tableLog - BIT_highbit32(nextState) );
+            tableDecode[u].nbBits = (BYTE) (tableLog - ZSTD_highbit32(nextState) );
             tableDecode[u].newState = (U16) ( (nextState << tableDecode[u].nbBits) - tableSize);
     }   }
 
     return 0;
 }
 
+size_t FSE_buildDTable_wksp(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize)
+{
+    return FSE_buildDTable_internal(dt, normalizedCounter, maxSymbolValue, tableLog, workSpace, wkspSize);
+}
+
 
 #ifndef FSE_COMMONDEFS_ONLY
 
 /*-*******************************************************
 *  Decompression (Byte symbols)
 *********************************************************/
-size_t FSE_buildDTable_rle (FSE_DTable* dt, BYTE symbolValue)
-{
-    void* ptr = dt;
-    FSE_DTableHeader* const DTableH = (FSE_DTableHeader*)ptr;
-    void* dPtr = dt + 1;
-    FSE_decode_t* const cell = (FSE_decode_t*)dPtr;
-
-    DTableH->tableLog = 0;
-    DTableH->fastMode = 0;
-
-    cell->newState = 0;
-    cell->symbol = symbolValue;
-    cell->nbBits = 0;
-
-    return 0;
-}
-
-
-size_t FSE_buildDTable_raw (FSE_DTable* dt, unsigned nbBits)
-{
-    void* ptr = dt;
-    FSE_DTableHeader* const DTableH = (FSE_DTableHeader*)ptr;
-    void* dPtr = dt + 1;
-    FSE_decode_t* const dinfo = (FSE_decode_t*)dPtr;
-    const unsigned tableSize = 1 << nbBits;
-    const unsigned tableMask = tableSize - 1;
-    const unsigned maxSV1 = tableMask+1;
-    unsigned s;
-
-    /* Sanity checks */
-    if (nbBits < 1) return ERROR(GENERIC);         /* min size */
-
-    /* Build Decoding Table */
-    DTableH->tableLog = (U16)nbBits;
-    DTableH->fastMode = 1;
-    for (s=0; s<maxSV1; s++) {
-        dinfo[s].newState = 0;
-        dinfo[s].symbol = (BYTE)s;
-        dinfo[s].nbBits = (BYTE)nbBits;
-    }
-
-    return 0;
-}
 
 FORCE_INLINE_TEMPLATE size_t FSE_decompress_usingDTable_generic(
           void* dst, size_t maxDstSize,
@@ -236,51 +235,77 @@ FORCE_INLINE_TEMPLATE size_t FSE_decompress_usingDTable_generic(
     return op-ostart;
 }
 
+typedef struct {
+    short ncount[FSE_MAX_SYMBOL_VALUE + 1];
+    FSE_DTable dtable[1]; /* Dynamically sized */
+} FSE_DecompressWksp;
 
-size_t FSE_decompress_usingDTable(void* dst, size_t originalSize,
-                            const void* cSrc, size_t cSrcSize,
-                            const FSE_DTable* dt)
-{
-    const void* ptr = dt;
-    const FSE_DTableHeader* DTableH = (const FSE_DTableHeader*)ptr;
-    const U32 fastMode = DTableH->fastMode;
 
-    /* select fast mode (static) */
-    if (fastMode) return FSE_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 1);
-    return FSE_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 0);
-}
-
-
-size_t FSE_decompress_wksp(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, FSE_DTable* workSpace, unsigned maxLog)
+FORCE_INLINE_TEMPLATE size_t FSE_decompress_wksp_body(
+        void* dst, size_t dstCapacity,
+        const void* cSrc, size_t cSrcSize,
+        unsigned maxLog, void* workSpace, size_t wkspSize,
+        int bmi2)
 {
     const BYTE* const istart = (const BYTE*)cSrc;
     const BYTE* ip = istart;
-    short counting[FSE_MAX_SYMBOL_VALUE+1];
     unsigned tableLog;
     unsigned maxSymbolValue = FSE_MAX_SYMBOL_VALUE;
+    FSE_DecompressWksp* const wksp = (FSE_DecompressWksp*)workSpace;
+
+    DEBUG_STATIC_ASSERT((FSE_MAX_SYMBOL_VALUE + 1) % 2 == 0);
+    if (wkspSize < sizeof(*wksp)) return ERROR(GENERIC);
 
     /* normal FSE decoding mode */
-    size_t const NCountLength = FSE_readNCount (counting, &maxSymbolValue, &tableLog, istart, cSrcSize);
-    if (FSE_isError(NCountLength)) return NCountLength;
-    /* if (NCountLength >= cSrcSize) return ERROR(srcSize_wrong); */  /* too small input size; supposed to be already checked in NCountLength, only remaining case : NCountLength==cSrcSize */
-    if (tableLog > maxLog) return ERROR(tableLog_tooLarge);
-    ip += NCountLength;
-    cSrcSize -= NCountLength;
+    {
+        size_t const NCountLength = FSE_readNCount_bmi2(wksp->ncount, &maxSymbolValue, &tableLog, istart, cSrcSize, bmi2);
+        if (FSE_isError(NCountLength)) return NCountLength;
+        if (tableLog > maxLog) return ERROR(tableLog_tooLarge);
+        assert(NCountLength <= cSrcSize);
+        ip += NCountLength;
+        cSrcSize -= NCountLength;
+    }
 
-    CHECK_F( FSE_buildDTable (workSpace, counting, maxSymbolValue, tableLog) );
+    if (FSE_DECOMPRESS_WKSP_SIZE(tableLog, maxSymbolValue) > wkspSize) return ERROR(tableLog_tooLarge);
+    assert(sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog) <= wkspSize);
+    workSpace = (BYTE*)workSpace + sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog);
+    wkspSize -= sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog);
 
-    return FSE_decompress_usingDTable (dst, dstCapacity, ip, cSrcSize, workSpace);   /* always return, even if it is an error code */
-}
+    CHECK_F( FSE_buildDTable_internal(wksp->dtable, wksp->ncount, maxSymbolValue, tableLog, workSpace, wkspSize) );
 
+    {
+        const void* ptr = wksp->dtable;
+        const FSE_DTableHeader* DTableH = (const FSE_DTableHeader*)ptr;
+        const U32 fastMode = DTableH->fastMode;
 
-typedef FSE_DTable DTable_max_t[FSE_DTABLE_SIZE_U32(FSE_MAX_TABLELOG)];
+        /* select fast mode (static) */
+        if (fastMode) return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, wksp->dtable, 1);
+        return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, wksp->dtable, 0);
+    }
+}
 
-size_t FSE_decompress(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize)
+/* Avoids the FORCE_INLINE of the _body() function. */
+static size_t FSE_decompress_wksp_body_default(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize)
 {
-    DTable_max_t dt;   /* Static analyzer seems unable to understand this table will be properly initialized later */
-    return FSE_decompress_wksp(dst, dstCapacity, cSrc, cSrcSize, dt, FSE_MAX_TABLELOG);
+    return FSE_decompress_wksp_body(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize, 0);
 }
 
+#if DYNAMIC_BMI2
+BMI2_TARGET_ATTRIBUTE static size_t FSE_decompress_wksp_body_bmi2(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize)
+{
+    return FSE_decompress_wksp_body(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize, 1);
+}
+#endif
 
+size_t FSE_decompress_wksp_bmi2(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize, int bmi2)
+{
+#if DYNAMIC_BMI2
+    if (bmi2) {
+        return FSE_decompress_wksp_body_bmi2(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize);
+    }
+#endif
+    (void)bmi2;
+    return FSE_decompress_wksp_body_default(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize);
+}
 
 #endif   /* FSE_COMMONDEFS_ONLY */

src/borg/algorithms/zstd/lib/common/huf.h

@@ -1,7 +1,7 @@
 /* ******************************************************************
  * huff0 huffman codec,
  * part of Finite State Entropy library
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -20,105 +20,30 @@ extern "C" {
 #define HUF_H_298734234
 
 /* *** Dependencies *** */
-#include <stddef.h>    /* size_t */
-
-
-/* *** library symbols visibility *** */
-/* Note : when linking with -fvisibility=hidden on gcc, or by default on Visual,
- *        HUF symbols remain "private" (internal symbols for library only).
- *        Set macro FSE_DLL_EXPORT to 1 if you want HUF symbols visible on DLL interface */
-#if defined(FSE_DLL_EXPORT) && (FSE_DLL_EXPORT==1) && defined(__GNUC__) && (__GNUC__ >= 4)
-#  define HUF_PUBLIC_API __attribute__ ((visibility ("default")))
-#elif defined(FSE_DLL_EXPORT) && (FSE_DLL_EXPORT==1)   /* Visual expected */
-#  define HUF_PUBLIC_API __declspec(dllexport)
-#elif defined(FSE_DLL_IMPORT) && (FSE_DLL_IMPORT==1)
-#  define HUF_PUBLIC_API __declspec(dllimport)  /* not required, just to generate faster code (saves a function pointer load from IAT and an indirect jump) */
-#else
-#  define HUF_PUBLIC_API
-#endif
-
-
-/* ========================== */
-/* ***  simple functions  *** */
-/* ========================== */
-
-/** HUF_compress() :
- *  Compress content from buffer 'src', of size 'srcSize', into buffer 'dst'.
- * 'dst' buffer must be already allocated.
- *  Compression runs faster if `dstCapacity` >= HUF_compressBound(srcSize).
- * `srcSize` must be <= `HUF_BLOCKSIZE_MAX` == 128 KB.
- * @return : size of compressed data (<= `dstCapacity`).
- *  Special values : if return == 0, srcData is not compressible => Nothing is stored within dst !!!
- *                   if HUF_isError(return), compression failed (more details using HUF_getErrorName())
- */
-HUF_PUBLIC_API size_t HUF_compress(void* dst, size_t dstCapacity,
-                             const void* src, size_t srcSize);
-
-/** HUF_decompress() :
- *  Decompress HUF data from buffer 'cSrc', of size 'cSrcSize',
- *  into already allocated buffer 'dst', of minimum size 'dstSize'.
- * `originalSize` : **must** be the ***exact*** size of original (uncompressed) data.
- *  Note : in contrast with FSE, HUF_decompress can regenerate
- *         RLE (cSrcSize==1) and uncompressed (cSrcSize==dstSize) data,
- *         because it knows size to regenerate (originalSize).
- * @return : size of regenerated data (== originalSize),
- *           or an error code, which can be tested using HUF_isError()
- */
-HUF_PUBLIC_API size_t HUF_decompress(void* dst,  size_t originalSize,
-                               const void* cSrc, size_t cSrcSize);
+#include "zstd_deps.h"    /* size_t */
+#include "mem.h"          /* U32 */
+#define FSE_STATIC_LINKING_ONLY
+#include "fse.h"
 
 
 /* ***   Tool functions *** */
-#define HUF_BLOCKSIZE_MAX (128 * 1024)                  /**< maximum input size for a single block compressed with HUF_compress */
-HUF_PUBLIC_API size_t HUF_compressBound(size_t size);   /**< maximum compressed size (worst case) */
+#define HUF_BLOCKSIZE_MAX (128 * 1024)   /**< maximum input size for a single block compressed with HUF_compress */
+size_t HUF_compressBound(size_t size);   /**< maximum compressed size (worst case) */
 
 /* Error Management */
-HUF_PUBLIC_API unsigned    HUF_isError(size_t code);       /**< tells if a return value is an error code */
-HUF_PUBLIC_API const char* HUF_getErrorName(size_t code);  /**< provides error code string (useful for debugging) */
-
-
-/* ***   Advanced function   *** */
-
-/** HUF_compress2() :
- *  Same as HUF_compress(), but offers control over `maxSymbolValue` and `tableLog`.
- * `maxSymbolValue` must be <= HUF_SYMBOLVALUE_MAX .
- * `tableLog` must be `<= HUF_TABLELOG_MAX` . */
-HUF_PUBLIC_API size_t HUF_compress2 (void* dst, size_t dstCapacity,
-                               const void* src, size_t srcSize,
-                               unsigned maxSymbolValue, unsigned tableLog);
-
-/** HUF_compress4X_wksp() :
- *  Same as HUF_compress2(), but uses externally allocated `workSpace`.
- * `workspace` must have minimum alignment of 4, and be at least as large as HUF_WORKSPACE_SIZE */
-#define HUF_WORKSPACE_SIZE ((6 << 10) + 256)
-#define HUF_WORKSPACE_SIZE_U32 (HUF_WORKSPACE_SIZE / sizeof(U32))
-HUF_PUBLIC_API size_t HUF_compress4X_wksp (void* dst, size_t dstCapacity,
-                                     const void* src, size_t srcSize,
-                                     unsigned maxSymbolValue, unsigned tableLog,
-                                     void* workSpace, size_t wkspSize);
+unsigned    HUF_isError(size_t code);       /**< tells if a return value is an error code */
+const char* HUF_getErrorName(size_t code);  /**< provides error code string (useful for debugging) */
 
-#endif   /* HUF_H_298734234 */
-
-/* ******************************************************************
- *  WARNING !!
- *  The following section contains advanced and experimental definitions
- *  which shall never be used in the context of a dynamic library,
- *  because they are not guaranteed to remain stable in the future.
- *  Only consider them in association with static linking.
- * *****************************************************************/
-#if defined(HUF_STATIC_LINKING_ONLY) && !defined(HUF_H_HUF_STATIC_LINKING_ONLY)
-#define HUF_H_HUF_STATIC_LINKING_ONLY
-
-/* *** Dependencies *** */
-#include "mem.h"   /* U32 */
 
+#define HUF_WORKSPACE_SIZE ((8 << 10) + 512 /* sorting scratch space */)
+#define HUF_WORKSPACE_SIZE_U64 (HUF_WORKSPACE_SIZE / sizeof(U64))
 
 /* *** Constants *** */
-#define HUF_TABLELOG_MAX      12      /* max runtime value of tableLog (due to static allocation); can be modified up to HUF_ABSOLUTEMAX_TABLELOG */
+#define HUF_TABLELOG_MAX      12      /* max runtime value of tableLog (due to static allocation); can be modified up to HUF_TABLELOG_ABSOLUTEMAX */
 #define HUF_TABLELOG_DEFAULT  11      /* default tableLog value when none specified */
 #define HUF_SYMBOLVALUE_MAX  255
 
-#define HUF_TABLELOG_ABSOLUTEMAX  15  /* absolute limit of HUF_MAX_TABLELOG. Beyond that value, code does not work */
+#define HUF_TABLELOG_ABSOLUTEMAX  12  /* absolute limit of HUF_MAX_TABLELOG. Beyond that value, code does not work */
 #if (HUF_TABLELOG_MAX > HUF_TABLELOG_ABSOLUTEMAX)
 #  error "HUF_TABLELOG_MAX is too large !"
 #endif
@@ -133,12 +58,12 @@ HUF_PUBLIC_API size_t HUF_compress4X_wksp (void* dst, size_t dstCapacity,
 #define HUF_COMPRESSBOUND(size) (HUF_CTABLEBOUND + HUF_BLOCKBOUND(size))   /* Macro version, useful for static allocation */
 
 /* static allocation of HUF's Compression Table */
-#define HUF_CTABLE_SIZE_U32(maxSymbolValue)   ((maxSymbolValue)+1)   /* Use tables of U32, for proper alignment */
-#define HUF_CTABLE_SIZE(maxSymbolValue)       (HUF_CTABLE_SIZE_U32(maxSymbolValue) * sizeof(U32))
+/* this is a private definition, just exposed for allocation and strict aliasing purpose. never EVER access its members directly */
+typedef size_t HUF_CElt;   /* consider it an incomplete type */
+#define HUF_CTABLE_SIZE_ST(maxSymbolValue)   ((maxSymbolValue)+2)   /* Use tables of size_t, for proper alignment */
+#define HUF_CTABLE_SIZE(maxSymbolValue)       (HUF_CTABLE_SIZE_ST(maxSymbolValue) * sizeof(size_t))
 #define HUF_CREATE_STATIC_CTABLE(name, maxSymbolValue) \
-    U32 name##hb[HUF_CTABLE_SIZE_U32(maxSymbolValue)]; \
-    void* name##hv = &(name##hb); \
-    HUF_CElt* name = (HUF_CElt*)(name##hv)   /* no final ; */
+    HUF_CElt name[HUF_CTABLE_SIZE_ST(maxSymbolValue)] /* no final ; */
 
 /* static allocation of HUF's DTable */
 typedef U32 HUF_DTable;
@@ -152,25 +77,49 @@ typedef U32 HUF_DTable;
 /* ****************************************
 *  Advanced decompression functions
 ******************************************/
-size_t HUF_decompress4X1 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< single-symbol decoder */
-#ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_decompress4X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< double-symbols decoder */
-#endif
 
-size_t HUF_decompress4X_DCtx (HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< decodes RLE and uncompressed */
-size_t HUF_decompress4X_hufOnly(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize); /**< considers RLE and uncompressed as errors */
-size_t HUF_decompress4X_hufOnly_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize); /**< considers RLE and uncompressed as errors */
-size_t HUF_decompress4X1_DCtx(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< single-symbol decoder */
-size_t HUF_decompress4X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize);   /**< single-symbol decoder */
-#ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_decompress4X2_DCtx(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< double-symbols decoder */
-size_t HUF_decompress4X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize);   /**< double-symbols decoder */
-#endif
+/**
+ * Huffman flags bitset.
+ * For all flags, 0 is the default value.
+ */
+typedef enum {
+    /**
+     * If compiled with DYNAMIC_BMI2: Set flag only if the CPU supports BMI2 at runtime.
+     * Otherwise: Ignored.
+     */
+    HUF_flags_bmi2 = (1 << 0),
+    /**
+     * If set: Test possible table depths to find the one that produces the smallest header + encoded size.
+     * If unset: Use heuristic to find the table depth.
+     */
+    HUF_flags_optimalDepth = (1 << 1),
+    /**
+     * If set: If the previous table can encode the input, always reuse the previous table.
+     * If unset: If the previous table can encode the input, reuse the previous table if it results in a smaller output.
+     */
+    HUF_flags_preferRepeat = (1 << 2),
+    /**
+     * If set: Sample the input and check if the sample is uncompressible, if it is then don't attempt to compress.
+     * If unset: Always histogram the entire input.
+     */
+    HUF_flags_suspectUncompressible = (1 << 3),
+    /**
+     * If set: Don't use assembly implementations
+     * If unset: Allow using assembly implementations
+     */
+    HUF_flags_disableAsm = (1 << 4),
+    /**
+     * If set: Don't use the fast decoding loop, always use the fallback decoding loop.
+     * If unset: Use the fast decoding loop when possible.
+     */
+    HUF_flags_disableFast = (1 << 5)
+} HUF_flags_e;
 
 
 /* ****************************************
  *  HUF detailed API
  * ****************************************/
+#define HUF_OPTIMAL_DEPTH_THRESHOLD ZSTD_btultra
 
 /*! HUF_compress() does the following:
  *  1. count symbol occurrence from source[] into table count[] using FSE_count() (exposed within "fse.h")
@@ -183,11 +132,12 @@ size_t HUF_decompress4X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
  *  For example, it's possible to compress several blocks using the same 'CTable',
  *  or to save and regenerate 'CTable' using external methods.
  */
-unsigned HUF_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue);
-typedef struct HUF_CElt_s HUF_CElt;   /* incomplete type */
-size_t HUF_buildCTable (HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue, unsigned maxNbBits);   /* @return : maxNbBits; CTable and count can overlap. In which case, CTable will overwrite count content */
-size_t HUF_writeCTable (void* dst, size_t maxDstSize, const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog);
-size_t HUF_compress4X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable);
+unsigned HUF_minTableLog(unsigned symbolCardinality);
+unsigned HUF_cardinality(const unsigned* count, unsigned maxSymbolValue);
+unsigned HUF_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, void* workSpace,
+ size_t wkspSize, HUF_CElt* table, const unsigned* count, int flags); /* table is used as scratch space for building and testing tables, not a return value */
+size_t HUF_writeCTable_wksp(void* dst, size_t maxDstSize, const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog, void* workspace, size_t workspaceSize);
+size_t HUF_compress4X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int flags);
 size_t HUF_estimateCompressedSize(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue);
 int HUF_validateCTable(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue);
 
@@ -196,22 +146,24 @@ typedef enum {
    HUF_repeat_check, /**< Can use the previous table but it must be checked. Note : The previous table must have been constructed by HUF_compress{1, 4}X_repeat */
    HUF_repeat_valid  /**< Can use the previous table and it is assumed to be valid */
  } HUF_repeat;
+
 /** HUF_compress4X_repeat() :
  *  Same as HUF_compress4X_wksp(), but considers using hufTable if *repeat != HUF_repeat_none.
  *  If it uses hufTable it does not modify hufTable or repeat.
  *  If it doesn't, it sets *repeat = HUF_repeat_none, and it sets hufTable to the table used.
- *  If preferRepeat then the old table will always be used if valid. */
+ *  If preferRepeat then the old table will always be used if valid.
+ *  If suspectUncompressible then some sampling checks will be run to potentially skip huffman coding */
 size_t HUF_compress4X_repeat(void* dst, size_t dstSize,
                        const void* src, size_t srcSize,
                        unsigned maxSymbolValue, unsigned tableLog,
                        void* workSpace, size_t wkspSize,    /**< `workSpace` must be aligned on 4-bytes boundaries, `wkspSize` must be >= HUF_WORKSPACE_SIZE */
-                       HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2);
+                       HUF_CElt* hufTable, HUF_repeat* repeat, int flags);
 
 /** HUF_buildCTable_wksp() :
  *  Same as HUF_buildCTable(), but using externally allocated scratch buffer.
  * `workSpace` must be aligned on 4-bytes boundaries, and its size must be >= HUF_CTABLE_WORKSPACE_SIZE.
  */
-#define HUF_CTABLE_WORKSPACE_SIZE_U32 (2*HUF_SYMBOLVALUE_MAX +1 +1)
+#define HUF_CTABLE_WORKSPACE_SIZE_U32 ((4 * (HUF_SYMBOLVALUE_MAX + 1)) + 192)
 #define HUF_CTABLE_WORKSPACE_SIZE (HUF_CTABLE_WORKSPACE_SIZE_U32 * sizeof(unsigned))
 size_t HUF_buildCTable_wksp (HUF_CElt* tree,
                        const unsigned* count, U32 maxSymbolValue, U32 maxNbBits,
@@ -226,15 +178,27 @@ size_t HUF_readStats(BYTE* huffWeight, size_t hwSize,
                      U32* rankStats, U32* nbSymbolsPtr, U32* tableLogPtr,
                      const void* src, size_t srcSize);
 
+/*! HUF_readStats_wksp() :
+ * Same as HUF_readStats() but takes an external workspace which must be
+ * 4-byte aligned and its size must be >= HUF_READ_STATS_WORKSPACE_SIZE.
+ * If the CPU has BMI2 support, pass bmi2=1, otherwise pass bmi2=0.
+ */
+#define HUF_READ_STATS_WORKSPACE_SIZE_U32 FSE_DECOMPRESS_WKSP_SIZE_U32(6, HUF_TABLELOG_MAX-1)
+#define HUF_READ_STATS_WORKSPACE_SIZE (HUF_READ_STATS_WORKSPACE_SIZE_U32 * sizeof(unsigned))
+size_t HUF_readStats_wksp(BYTE* huffWeight, size_t hwSize,
+                          U32* rankStats, U32* nbSymbolsPtr, U32* tableLogPtr,
+                          const void* src, size_t srcSize,
+                          void* workspace, size_t wkspSize,
+                          int flags);
+
 /** HUF_readCTable() :
  *  Loading a CTable saved with HUF_writeCTable() */
 size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void* src, size_t srcSize, unsigned *hasZeroWeights);
 
-/** HUF_getNbBits() :
+/** HUF_getNbBitsFromCTable() :
  *  Read nbBits from CTable symbolTable, for symbol `symbolValue` presumed <= HUF_SYMBOLVALUE_MAX
- *  Note 1 : is not inlined, as HUF_CElt definition is private
- *  Note 2 : const void* used, so that it can provide a statically allocated table as argument (which uses type U32) */
-U32 HUF_getNbBits(const void* symbolTable, U32 symbolValue);
+ *  Note 1 : is not inlined, as HUF_CElt definition is private */
+U32 HUF_getNbBitsFromCTable(const HUF_CElt* symbolTable, U32 symbolValue);
 
 /*
  * HUF_decompress() does the following:
@@ -260,80 +224,49 @@ U32 HUF_selectDecoder (size_t dstSize, size_t cSrcSize);
  *  a required workspace size greater than that specified in the following
  *  macro.
  */
-#define HUF_DECOMPRESS_WORKSPACE_SIZE (2 << 10)
+#define HUF_DECOMPRESS_WORKSPACE_SIZE ((2 << 10) + (1 << 9))
 #define HUF_DECOMPRESS_WORKSPACE_SIZE_U32 (HUF_DECOMPRESS_WORKSPACE_SIZE / sizeof(U32))
 
-#ifndef HUF_FORCE_DECOMPRESS_X2
-size_t HUF_readDTableX1 (HUF_DTable* DTable, const void* src, size_t srcSize);
-size_t HUF_readDTableX1_wksp (HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize);
-#endif
-#ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_readDTableX2 (HUF_DTable* DTable, const void* src, size_t srcSize);
-size_t HUF_readDTableX2_wksp (HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize);
-#endif
-
-size_t HUF_decompress4X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);
-#ifndef HUF_FORCE_DECOMPRESS_X2
-size_t HUF_decompress4X1_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);
-#endif
-#ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_decompress4X2_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);
-#endif
-
 
 /* ====================== */
 /* single stream variants */
 /* ====================== */
 
-size_t HUF_compress1X (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog);
-size_t HUF_compress1X_wksp (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);  /**< `workSpace` must be a table of at least HUF_WORKSPACE_SIZE_U32 unsigned */
-size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable);
+size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int flags);
 /** HUF_compress1X_repeat() :
  *  Same as HUF_compress1X_wksp(), but considers using hufTable if *repeat != HUF_repeat_none.
  *  If it uses hufTable it does not modify hufTable or repeat.
  *  If it doesn't, it sets *repeat = HUF_repeat_none, and it sets hufTable to the table used.
- *  If preferRepeat then the old table will always be used if valid. */
+ *  If preferRepeat then the old table will always be used if valid.
+ *  If suspectUncompressible then some sampling checks will be run to potentially skip huffman coding */
 size_t HUF_compress1X_repeat(void* dst, size_t dstSize,
                        const void* src, size_t srcSize,
                        unsigned maxSymbolValue, unsigned tableLog,
                        void* workSpace, size_t wkspSize,   /**< `workSpace` must be aligned on 4-bytes boundaries, `wkspSize` must be >= HUF_WORKSPACE_SIZE */
-                       HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2);
+                       HUF_CElt* hufTable, HUF_repeat* repeat, int flags);
 
-size_t HUF_decompress1X1 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* single-symbol decoder */
+size_t HUF_decompress1X_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags);
 #ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_decompress1X2 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* double-symbol decoder */
+size_t HUF_decompress1X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags);   /**< double-symbols decoder */
 #endif
 
-size_t HUF_decompress1X_DCtx (HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);
-size_t HUF_decompress1X_DCtx_wksp (HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize);
+/* BMI2 variants.
+ * If the CPU has BMI2 support, pass bmi2=1, otherwise pass bmi2=0.
+ */
+size_t HUF_decompress1X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int flags);
 #ifndef HUF_FORCE_DECOMPRESS_X2
-size_t HUF_decompress1X1_DCtx(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< single-symbol decoder */
-size_t HUF_decompress1X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize);   /**< single-symbol decoder */
-#endif
-#ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_decompress1X2_DCtx(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /**< double-symbols decoder */
-size_t HUF_decompress1X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize);   /**< double-symbols decoder */
+size_t HUF_decompress1X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags);
 #endif
-
-size_t HUF_decompress1X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);   /**< automatic selection of sing or double symbol decoder, based on DTable */
+size_t HUF_decompress4X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int flags);
+size_t HUF_decompress4X_hufOnly_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags);
 #ifndef HUF_FORCE_DECOMPRESS_X2
-size_t HUF_decompress1X1_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);
+size_t HUF_readDTableX1_wksp(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize, int flags);
 #endif
 #ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_decompress1X2_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable);
+size_t HUF_readDTableX2_wksp(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize, int flags);
 #endif
 
-/* BMI2 variants.
- * If the CPU has BMI2 support, pass bmi2=1, otherwise pass bmi2=0.
- */
-size_t HUF_decompress1X_usingDTable_bmi2(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int bmi2);
-#ifndef HUF_FORCE_DECOMPRESS_X2
-size_t HUF_decompress1X1_DCtx_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int bmi2);
-#endif
-size_t HUF_decompress4X_usingDTable_bmi2(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int bmi2);
-size_t HUF_decompress4X_hufOnly_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int bmi2);
-
-#endif /* HUF_STATIC_LINKING_ONLY */
+#endif   /* HUF_H_298734234 */
 
 #if defined (__cplusplus)
 }

src/borg/algorithms/zstd/lib/common/mem.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -18,8 +18,10 @@ extern "C" {
 /*-****************************************
 *  Dependencies
 ******************************************/
-#include <stddef.h>     /* size_t, ptrdiff_t */
-#include <string.h>     /* memcpy */
+#include <stddef.h>  /* size_t, ptrdiff_t */
+#include "compiler.h"  /* __has_builtin */
+#include "debug.h"  /* DEBUG_STATIC_ASSERT */
+#include "zstd_deps.h"  /* ZSTD_memcpy */
 
 
 /*-****************************************
@@ -39,94 +41,18 @@ extern "C" {
 #  define MEM_STATIC static  /* this version may generate warnings for unused static functions; disable the relevant warning */
 #endif
 
-#ifndef __has_builtin
-#  define __has_builtin(x) 0  /* compat. with non-clang compilers */
-#endif
-
-/* code only tested on 32 and 64 bits systems */
-#define MEM_STATIC_ASSERT(c)   { enum { MEM_static_assert = 1/(int)(!!(c)) }; }
-MEM_STATIC void MEM_check(void) { MEM_STATIC_ASSERT((sizeof(size_t)==4) || (sizeof(size_t)==8)); }
-
-/* detects whether we are being compiled under msan */
-#if defined (__has_feature)
-#  if __has_feature(memory_sanitizer)
-#    define MEMORY_SANITIZER 1
-#  endif
-#endif
-
-#if defined (MEMORY_SANITIZER)
-/* Not all platforms that support msan provide sanitizers/msan_interface.h.
- * We therefore declare the functions we need ourselves, rather than trying to
- * include the header file... */
-
-#include <stdint.h> /* intptr_t */
-
-/* Make memory region fully initialized (without changing its contents). */
-void __msan_unpoison(const volatile void *a, size_t size);
-
-/* Make memory region fully uninitialized (without changing its contents).
-   This is a legacy interface that does not update origin information. Use
-   __msan_allocated_memory() instead. */
-void __msan_poison(const volatile void *a, size_t size);
-
-/* Returns the offset of the first (at least partially) poisoned byte in the
-   memory range, or -1 if the whole range is good. */
-intptr_t __msan_test_shadow(const volatile void *x, size_t size);
-#endif
-
-/* detects whether we are being compiled under asan */
-#if defined (__has_feature)
-#  if __has_feature(address_sanitizer)
-#    define ADDRESS_SANITIZER 1
-#  endif
-#elif defined(__SANITIZE_ADDRESS__)
-#  define ADDRESS_SANITIZER 1
-#endif
-
-#if defined (ADDRESS_SANITIZER)
-/* Not all platforms that support asan provide sanitizers/asan_interface.h.
- * We therefore declare the functions we need ourselves, rather than trying to
- * include the header file... */
-
-/**
- * Marks a memory region (<c>[addr, addr+size)</c>) as unaddressable.
- *
- * This memory must be previously allocated by your program. Instrumented
- * code is forbidden from accessing addresses in this region until it is
- * unpoisoned. This function is not guaranteed to poison the entire region -
- * it could poison only a subregion of <c>[addr, addr+size)</c> due to ASan
- * alignment restrictions.
- *
- * \note This function is not thread-safe because no two threads can poison or
- * unpoison memory in the same memory region simultaneously.
- *
- * \param addr Start of memory region.
- * \param size Size of memory region. */
-void __asan_poison_memory_region(void const volatile *addr, size_t size);
-
-/**
- * Marks a memory region (<c>[addr, addr+size)</c>) as addressable.
- *
- * This memory must be previously allocated by your program. Accessing
- * addresses in this region is allowed until this region is poisoned again.
- * This function could unpoison a super-region of <c>[addr, addr+size)</c> due
- * to ASan alignment restrictions.
- *
- * \note This function is not thread-safe because no two threads can
- * poison or unpoison memory in the same memory region simultaneously.
- *
- * \param addr Start of memory region.
- * \param size Size of memory region. */
-void __asan_unpoison_memory_region(void const volatile *addr, size_t size);
-#endif
-
-
 /*-**************************************************************
 *  Basic Types
 *****************************************************************/
 #if  !defined (__VMS) && (defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
-# include <stdint.h>
+#  if defined(_AIX)
+#    include <inttypes.h>
+#  else
+#    include <stdint.h> /* intptr_t */
+#  endif
   typedef   uint8_t BYTE;
+  typedef   uint8_t U8;
+  typedef    int8_t S8;
   typedef  uint16_t U16;
   typedef   int16_t S16;
   typedef  uint32_t U32;
@@ -139,6 +65,8 @@ void __asan_unpoison_memory_region(void const volatile *addr, size_t size);
 #  error "this implementation requires char to be exactly 8-bit type"
 #endif
   typedef unsigned char      BYTE;
+  typedef unsigned char      U8;
+  typedef   signed char      S8;
 #if USHRT_MAX != 65535
 #  error "this implementation requires short to be exactly 16-bit type"
 #endif
@@ -157,25 +85,63 @@ void __asan_unpoison_memory_region(void const volatile *addr, size_t size);
 
 
 /*-**************************************************************
-*  Memory I/O
+*  Memory I/O API
+*****************************************************************/
+/*=== Static platform detection ===*/
+MEM_STATIC unsigned MEM_32bits(void);
+MEM_STATIC unsigned MEM_64bits(void);
+MEM_STATIC unsigned MEM_isLittleEndian(void);
+
+/*=== Native unaligned read/write ===*/
+MEM_STATIC U16 MEM_read16(const void* memPtr);
+MEM_STATIC U32 MEM_read32(const void* memPtr);
+MEM_STATIC U64 MEM_read64(const void* memPtr);
+MEM_STATIC size_t MEM_readST(const void* memPtr);
+
+MEM_STATIC void MEM_write16(void* memPtr, U16 value);
+MEM_STATIC void MEM_write32(void* memPtr, U32 value);
+MEM_STATIC void MEM_write64(void* memPtr, U64 value);
+
+/*=== Little endian unaligned read/write ===*/
+MEM_STATIC U16 MEM_readLE16(const void* memPtr);
+MEM_STATIC U32 MEM_readLE24(const void* memPtr);
+MEM_STATIC U32 MEM_readLE32(const void* memPtr);
+MEM_STATIC U64 MEM_readLE64(const void* memPtr);
+MEM_STATIC size_t MEM_readLEST(const void* memPtr);
+
+MEM_STATIC void MEM_writeLE16(void* memPtr, U16 val);
+MEM_STATIC void MEM_writeLE24(void* memPtr, U32 val);
+MEM_STATIC void MEM_writeLE32(void* memPtr, U32 val32);
+MEM_STATIC void MEM_writeLE64(void* memPtr, U64 val64);
+MEM_STATIC void MEM_writeLEST(void* memPtr, size_t val);
+
+/*=== Big endian unaligned read/write ===*/
+MEM_STATIC U32 MEM_readBE32(const void* memPtr);
+MEM_STATIC U64 MEM_readBE64(const void* memPtr);
+MEM_STATIC size_t MEM_readBEST(const void* memPtr);
+
+MEM_STATIC void MEM_writeBE32(void* memPtr, U32 val32);
+MEM_STATIC void MEM_writeBE64(void* memPtr, U64 val64);
+MEM_STATIC void MEM_writeBEST(void* memPtr, size_t val);
+
+/*=== Byteswap ===*/
+MEM_STATIC U32 MEM_swap32(U32 in);
+MEM_STATIC U64 MEM_swap64(U64 in);
+MEM_STATIC size_t MEM_swapST(size_t in);
+
+
+/*-**************************************************************
+*  Memory I/O Implementation
 *****************************************************************/
-/* MEM_FORCE_MEMORY_ACCESS :
- * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
- * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
- * The below switch allow to select different access method for improved performance.
- * Method 0 (default) : use `memcpy()`. Safe and portable.
- * Method 1 : `__packed` statement. It depends on compiler extension (i.e., not portable).
- *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
+/* MEM_FORCE_MEMORY_ACCESS : For accessing unaligned memory:
+ * Method 0 : always use `memcpy()`. Safe and portable.
+ * Method 1 : Use compiler extension to set unaligned access.
  * Method 2 : direct access. This method is portable but violate C standard.
  *            It can generate buggy code on targets depending on alignment.
- *            In some circumstances, it's the only known way to get the most performance (i.e. GCC + ARMv6)
- * See http://fastcompression.blogspot.fr/2015/08/accessing-unaligned-memory.html for details.
- * Prefer these methods in priority order (0 > 1 > 2)
+ * Default  : method 1 if supported, else method 0
  */
 #ifndef MEM_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
-#  if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
-#    define MEM_FORCE_MEMORY_ACCESS 2
-#  elif defined(__INTEL_COMPILER) || defined(__GNUC__) || defined(__ICCARM__)
+#  ifdef __GNUC__
 #    define MEM_FORCE_MEMORY_ACCESS 1
 #  endif
 #endif
@@ -185,8 +151,22 @@ MEM_STATIC unsigned MEM_64bits(void) { return sizeof(size_t)==8; }
 
 MEM_STATIC unsigned MEM_isLittleEndian(void)
 {
+#if defined(__BYTE_ORDER__) && defined(__ORDER_LITTLE_ENDIAN__) && (__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
+    return 1;
+#elif defined(__BYTE_ORDER__) && defined(__ORDER_BIG_ENDIAN__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+    return 0;
+#elif defined(__clang__) && __LITTLE_ENDIAN__
+    return 1;
+#elif defined(__clang__) && __BIG_ENDIAN__
+    return 0;
+#elif defined(_MSC_VER) && (_M_AMD64 || _M_IX86)
+    return 1;
+#elif defined(__DMC__) && defined(_M_IX86)
+    return 1;
+#else
     const union { U32 u; BYTE c[4]; } one = { 1 };   /* don't use static : performance detrimental  */
     return one.c[0];
+#endif
 }
 
 #if defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==2)
@@ -204,30 +184,19 @@ MEM_STATIC void MEM_write64(void* memPtr, U64 value) { *(U64*)memPtr = value; }
 
 #elif defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==1)
 
-/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
-/* currently only defined for gcc and icc */
-#if defined(_MSC_VER) || (defined(__INTEL_COMPILER) && defined(WIN32))
-    __pragma( pack(push, 1) )
-    typedef struct { U16 v; } unalign16;
-    typedef struct { U32 v; } unalign32;
-    typedef struct { U64 v; } unalign64;
-    typedef struct { size_t v; } unalignArch;
-    __pragma( pack(pop) )
-#else
-    typedef struct { U16 v; } __attribute__((packed)) unalign16;
-    typedef struct { U32 v; } __attribute__((packed)) unalign32;
-    typedef struct { U64 v; } __attribute__((packed)) unalign64;
-    typedef struct { size_t v; } __attribute__((packed)) unalignArch;
-#endif
+typedef __attribute__((aligned(1))) U16 unalign16;
+typedef __attribute__((aligned(1))) U32 unalign32;
+typedef __attribute__((aligned(1))) U64 unalign64;
+typedef __attribute__((aligned(1))) size_t unalignArch;
 
-MEM_STATIC U16 MEM_read16(const void* ptr) { return ((const unalign16*)ptr)->v; }
-MEM_STATIC U32 MEM_read32(const void* ptr) { return ((const unalign32*)ptr)->v; }
-MEM_STATIC U64 MEM_read64(const void* ptr) { return ((const unalign64*)ptr)->v; }
-MEM_STATIC size_t MEM_readST(const void* ptr) { return ((const unalignArch*)ptr)->v; }
+MEM_STATIC U16 MEM_read16(const void* ptr) { return *(const unalign16*)ptr; }
+MEM_STATIC U32 MEM_read32(const void* ptr) { return *(const unalign32*)ptr; }
+MEM_STATIC U64 MEM_read64(const void* ptr) { return *(const unalign64*)ptr; }
+MEM_STATIC size_t MEM_readST(const void* ptr) { return *(const unalignArch*)ptr; }
 
-MEM_STATIC void MEM_write16(void* memPtr, U16 value) { ((unalign16*)memPtr)->v = value; }
-MEM_STATIC void MEM_write32(void* memPtr, U32 value) { ((unalign32*)memPtr)->v = value; }
-MEM_STATIC void MEM_write64(void* memPtr, U64 value) { ((unalign64*)memPtr)->v = value; }
+MEM_STATIC void MEM_write16(void* memPtr, U16 value) { *(unalign16*)memPtr = value; }
+MEM_STATIC void MEM_write32(void* memPtr, U32 value) { *(unalign32*)memPtr = value; }
+MEM_STATIC void MEM_write64(void* memPtr, U64 value) { *(unalign64*)memPtr = value; }
 
 #else
 
@@ -236,41 +205,49 @@ MEM_STATIC void MEM_write64(void* memPtr, U64 value) { ((unalign64*)memPtr)->v =
 
 MEM_STATIC U16 MEM_read16(const void* memPtr)
 {
-    U16 val; memcpy(&val, memPtr, sizeof(val)); return val;
+    U16 val; ZSTD_memcpy(&val, memPtr, sizeof(val)); return val;
 }
 
 MEM_STATIC U32 MEM_read32(const void* memPtr)
 {
-    U32 val; memcpy(&val, memPtr, sizeof(val)); return val;
+    U32 val; ZSTD_memcpy(&val, memPtr, sizeof(val)); return val;
 }
 
 MEM_STATIC U64 MEM_read64(const void* memPtr)
 {
-    U64 val; memcpy(&val, memPtr, sizeof(val)); return val;
+    U64 val; ZSTD_memcpy(&val, memPtr, sizeof(val)); return val;
 }
 
 MEM_STATIC size_t MEM_readST(const void* memPtr)
 {
-    size_t val; memcpy(&val, memPtr, sizeof(val)); return val;
+    size_t val; ZSTD_memcpy(&val, memPtr, sizeof(val)); return val;
 }
 
 MEM_STATIC void MEM_write16(void* memPtr, U16 value)
 {
-    memcpy(memPtr, &value, sizeof(value));
+    ZSTD_memcpy(memPtr, &value, sizeof(value));
 }
 
 MEM_STATIC void MEM_write32(void* memPtr, U32 value)
 {
-    memcpy(memPtr, &value, sizeof(value));
+    ZSTD_memcpy(memPtr, &value, sizeof(value));
 }
 
 MEM_STATIC void MEM_write64(void* memPtr, U64 value)
 {
-    memcpy(memPtr, &value, sizeof(value));
+    ZSTD_memcpy(memPtr, &value, sizeof(value));
 }
 
 #endif /* MEM_FORCE_MEMORY_ACCESS */
 
+MEM_STATIC U32 MEM_swap32_fallback(U32 in)
+{
+    return  ((in << 24) & 0xff000000 ) |
+            ((in <<  8) & 0x00ff0000 ) |
+            ((in >>  8) & 0x0000ff00 ) |
+            ((in >> 24) & 0x000000ff );
+}
+
 MEM_STATIC U32 MEM_swap32(U32 in)
 {
 #if defined(_MSC_VER)     /* Visual Studio */
@@ -279,22 +256,13 @@ MEM_STATIC U32 MEM_swap32(U32 in)
   || (defined(__clang__) && __has_builtin(__builtin_bswap32))
     return __builtin_bswap32(in);
 #else
-    return  ((in << 24) & 0xff000000 ) |
-            ((in <<  8) & 0x00ff0000 ) |
-            ((in >>  8) & 0x0000ff00 ) |
-            ((in >> 24) & 0x000000ff );
+    return MEM_swap32_fallback(in);
 #endif
 }
 
-MEM_STATIC U64 MEM_swap64(U64 in)
+MEM_STATIC U64 MEM_swap64_fallback(U64 in)
 {
-#if defined(_MSC_VER)     /* Visual Studio */
-    return _byteswap_uint64(in);
-#elif (defined (__GNUC__) && (__GNUC__ * 100 + __GNUC_MINOR__ >= 403)) \
-  || (defined(__clang__) && __has_builtin(__builtin_bswap64))
-    return __builtin_bswap64(in);
-#else
-    return  ((in << 56) & 0xff00000000000000ULL) |
+     return  ((in << 56) & 0xff00000000000000ULL) |
             ((in << 40) & 0x00ff000000000000ULL) |
             ((in << 24) & 0x0000ff0000000000ULL) |
             ((in << 8)  & 0x000000ff00000000ULL) |
@@ -302,6 +270,17 @@ MEM_STATIC U64 MEM_swap64(U64 in)
             ((in >> 24) & 0x0000000000ff0000ULL) |
             ((in >> 40) & 0x000000000000ff00ULL) |
             ((in >> 56) & 0x00000000000000ffULL);
+}
+
+MEM_STATIC U64 MEM_swap64(U64 in)
+{
+#if defined(_MSC_VER)     /* Visual Studio */
+    return _byteswap_uint64(in);
+#elif (defined (__GNUC__) && (__GNUC__ * 100 + __GNUC_MINOR__ >= 403)) \
+  || (defined(__clang__) && __has_builtin(__builtin_bswap64))
+    return __builtin_bswap64(in);
+#else
+    return MEM_swap64_fallback(in);
 #endif
 }
 
@@ -338,7 +317,7 @@ MEM_STATIC void MEM_writeLE16(void* memPtr, U16 val)
 
 MEM_STATIC U32 MEM_readLE24(const void* memPtr)
 {
-    return MEM_readLE16(memPtr) + (((const BYTE*)memPtr)[2] << 16);
+    return (U32)MEM_readLE16(memPtr) + ((U32)(((const BYTE*)memPtr)[2]) << 16);
 }
 
 MEM_STATIC void MEM_writeLE24(void* memPtr, U32 val)
@@ -445,6 +424,9 @@ MEM_STATIC void MEM_writeBEST(void* memPtr, size_t val)
         MEM_writeBE64(memPtr, (U64)val);
 }
 
+/* code only tested on 32 and 64 bits systems */
+MEM_STATIC void MEM_check(void) { DEBUG_STATIC_ASSERT((sizeof(size_t)==4) || (sizeof(size_t)==8)); }
+
 
 #if defined (__cplusplus)
 }

src/borg/algorithms/zstd/lib/common/pool.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -10,9 +10,9 @@
 
 
 /* ======   Dependencies   ======= */
-#include <stddef.h>    /* size_t */
+#include "../common/allocations.h"  /* ZSTD_customCalloc, ZSTD_customFree */
+#include "zstd_deps.h" /* size_t */
 #include "debug.h"     /* assert */
-#include "zstd_internal.h"  /* ZSTD_malloc, ZSTD_free */
 #include "pool.h"
 
 /* ======   Compiler specifics   ====== */
@@ -86,7 +86,7 @@ static void* POOL_thread(void* opaque) {
         {   POOL_job const job = ctx->queue[ctx->queueHead];
             ctx->queueHead = (ctx->queueHead + 1) % ctx->queueSize;
             ctx->numThreadsBusy++;
-            ctx->queueEmpty = ctx->queueHead == ctx->queueTail;
+            ctx->queueEmpty = (ctx->queueHead == ctx->queueTail);
             /* Unlock the mutex, signal a pusher, and run the job */
             ZSTD_pthread_cond_signal(&ctx->queuePushCond);
             ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
@@ -96,33 +96,37 @@ static void* POOL_thread(void* opaque) {
             /* If the intended queue size was 0, signal after finishing job */
             ZSTD_pthread_mutex_lock(&ctx->queueMutex);
             ctx->numThreadsBusy--;
-            if (ctx->queueSize == 1) {
-                ZSTD_pthread_cond_signal(&ctx->queuePushCond);
-            }
+            ZSTD_pthread_cond_signal(&ctx->queuePushCond);
             ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
         }
     }  /* for (;;) */
     assert(0);  /* Unreachable */
 }
 
+/* ZSTD_createThreadPool() : public access point */
+POOL_ctx* ZSTD_createThreadPool(size_t numThreads) {
+    return POOL_create (numThreads, 0);
+}
+
 POOL_ctx* POOL_create(size_t numThreads, size_t queueSize) {
     return POOL_create_advanced(numThreads, queueSize, ZSTD_defaultCMem);
 }
 
 POOL_ctx* POOL_create_advanced(size_t numThreads, size_t queueSize,
-                               ZSTD_customMem customMem) {
+                               ZSTD_customMem customMem)
+{
     POOL_ctx* ctx;
     /* Check parameters */
     if (!numThreads) { return NULL; }
     /* Allocate the context and zero initialize */
-    ctx = (POOL_ctx*)ZSTD_calloc(sizeof(POOL_ctx), customMem);
+    ctx = (POOL_ctx*)ZSTD_customCalloc(sizeof(POOL_ctx), customMem);
     if (!ctx) { return NULL; }
     /* Initialize the job queue.
      * It needs one extra space since one space is wasted to differentiate
      * empty and full queues.
      */
     ctx->queueSize = queueSize + 1;
-    ctx->queue = (POOL_job*)ZSTD_malloc(ctx->queueSize * sizeof(POOL_job), customMem);
+    ctx->queue = (POOL_job*)ZSTD_customCalloc(ctx->queueSize * sizeof(POOL_job), customMem);
     ctx->queueHead = 0;
     ctx->queueTail = 0;
     ctx->numThreadsBusy = 0;
@@ -136,7 +140,7 @@ POOL_ctx* POOL_create_advanced(size_t numThreads, size_t queueSize,
     }
     ctx->shutdown = 0;
     /* Allocate space for the thread handles */
-    ctx->threads = (ZSTD_pthread_t*)ZSTD_malloc(numThreads * sizeof(ZSTD_pthread_t), customMem);
+    ctx->threads = (ZSTD_pthread_t*)ZSTD_customCalloc(numThreads * sizeof(ZSTD_pthread_t), customMem);
     ctx->threadCapacity = 0;
     ctx->customMem = customMem;
     /* Check for errors */
@@ -169,7 +173,7 @@ static void POOL_join(POOL_ctx* ctx) {
     /* Join all of the threads */
     {   size_t i;
         for (i = 0; i < ctx->threadCapacity; ++i) {
-            ZSTD_pthread_join(ctx->threads[i], NULL);  /* note : could fail */
+            ZSTD_pthread_join(ctx->threads[i]);  /* note : could fail */
     }   }
 }
 
@@ -179,14 +183,27 @@ void POOL_free(POOL_ctx *ctx) {
     ZSTD_pthread_mutex_destroy(&ctx->queueMutex);
     ZSTD_pthread_cond_destroy(&ctx->queuePushCond);
     ZSTD_pthread_cond_destroy(&ctx->queuePopCond);
-    ZSTD_free(ctx->queue, ctx->customMem);
-    ZSTD_free(ctx->threads, ctx->customMem);
-    ZSTD_free(ctx, ctx->customMem);
+    ZSTD_customFree(ctx->queue, ctx->customMem);
+    ZSTD_customFree(ctx->threads, ctx->customMem);
+    ZSTD_customFree(ctx, ctx->customMem);
 }
 
+/*! POOL_joinJobs() :
+ *  Waits for all queued jobs to finish executing.
+ */
+void POOL_joinJobs(POOL_ctx* ctx) {
+    ZSTD_pthread_mutex_lock(&ctx->queueMutex);
+    while(!ctx->queueEmpty || ctx->numThreadsBusy > 0) {
+        ZSTD_pthread_cond_wait(&ctx->queuePushCond, &ctx->queueMutex);
+    }
+    ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
+}
 
+void ZSTD_freeThreadPool (ZSTD_threadPool* pool) {
+  POOL_free (pool);
+}
 
-size_t POOL_sizeof(POOL_ctx *ctx) {
+size_t POOL_sizeof(const POOL_ctx* ctx) {
     if (ctx==NULL) return 0;  /* supports sizeof NULL */
     return sizeof(*ctx)
         + ctx->queueSize * sizeof(POOL_job)
@@ -203,11 +220,11 @@ static int POOL_resize_internal(POOL_ctx* ctx, size_t numThreads)
         return 0;
     }
     /* numThreads > threadCapacity */
-    {   ZSTD_pthread_t* const threadPool = (ZSTD_pthread_t*)ZSTD_malloc(numThreads * sizeof(ZSTD_pthread_t), ctx->customMem);
+    {   ZSTD_pthread_t* const threadPool = (ZSTD_pthread_t*)ZSTD_customCalloc(numThreads * sizeof(ZSTD_pthread_t), ctx->customMem);
         if (!threadPool) return 1;
         /* replace existing thread pool */
-        memcpy(threadPool, ctx->threads, ctx->threadCapacity * sizeof(*threadPool));
-        ZSTD_free(ctx->threads, ctx->customMem);
+        ZSTD_memcpy(threadPool, ctx->threads, ctx->threadCapacity * sizeof(*threadPool));
+        ZSTD_customFree(ctx->threads, ctx->customMem);
         ctx->threads = threadPool;
         /* Initialize additional threads */
         {   size_t threadId;
@@ -251,9 +268,12 @@ static int isQueueFull(POOL_ctx const* ctx) {
 }
 
 
-static void POOL_add_internal(POOL_ctx* ctx, POOL_function function, void *opaque)
+static void
+POOL_add_internal(POOL_ctx* ctx, POOL_function function, void *opaque)
 {
-    POOL_job const job = {function, opaque};
+    POOL_job job;
+    job.function = function;
+    job.opaque = opaque;
     assert(ctx != NULL);
     if (ctx->shutdown) return;
 
@@ -301,21 +321,28 @@ int POOL_tryAdd(POOL_ctx* ctx, POOL_function function, void* opaque)
 struct POOL_ctx_s {
     int dummy;
 };
-static POOL_ctx g_ctx;
+static POOL_ctx g_poolCtx;
 
 POOL_ctx* POOL_create(size_t numThreads, size_t queueSize) {
     return POOL_create_advanced(numThreads, queueSize, ZSTD_defaultCMem);
 }
 
-POOL_ctx* POOL_create_advanced(size_t numThreads, size_t queueSize, ZSTD_customMem customMem) {
+POOL_ctx*
+POOL_create_advanced(size_t numThreads, size_t queueSize, ZSTD_customMem customMem)
+{
     (void)numThreads;
     (void)queueSize;
     (void)customMem;
-    return &g_ctx;
+    return &g_poolCtx;
 }
 
 void POOL_free(POOL_ctx* ctx) {
-    assert(!ctx || ctx == &g_ctx);
+    assert(!ctx || ctx == &g_poolCtx);
+    (void)ctx;
+}
+
+void POOL_joinJobs(POOL_ctx* ctx){
+    assert(!ctx || ctx == &g_poolCtx);
     (void)ctx;
 }
 
@@ -335,9 +362,9 @@ int POOL_tryAdd(POOL_ctx* ctx, POOL_function function, void* opaque) {
     return 1;
 }
 
-size_t POOL_sizeof(POOL_ctx* ctx) {
+size_t POOL_sizeof(const POOL_ctx* ctx) {
     if (ctx==NULL) return 0;  /* supports sizeof NULL */
-    assert(ctx == &g_ctx);
+    assert(ctx == &g_poolCtx);
     return sizeof(*ctx);
 }
 

src/borg/algorithms/zstd/lib/common/pool.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -16,7 +16,7 @@ extern "C" {
 #endif
 
 
-#include <stddef.h>   /* size_t */
+#include "zstd_deps.h"
 #define ZSTD_STATIC_LINKING_ONLY   /* ZSTD_customMem */
 #include "../zstd.h"
 
@@ -38,6 +38,12 @@ POOL_ctx* POOL_create_advanced(size_t numThreads, size_t queueSize,
  */
 void POOL_free(POOL_ctx* ctx);
 
+
+/*! POOL_joinJobs() :
+ *  Waits for all queued jobs to finish executing.
+ */
+void POOL_joinJobs(POOL_ctx* ctx);
+
 /*! POOL_resize() :
  *  Expands or shrinks pool's number of threads.
  *  This is more efficient than releasing + creating a new context,
@@ -53,7 +59,7 @@ int POOL_resize(POOL_ctx* ctx, size_t numThreads);
  * @return threadpool memory usage
  *  note : compatible with NULL (returns 0 in this case)
  */
-size_t POOL_sizeof(POOL_ctx* ctx);
+size_t POOL_sizeof(const POOL_ctx* ctx);
 
 /*! POOL_function :
  *  The function type that can be added to a thread pool.
@@ -70,7 +76,7 @@ void POOL_add(POOL_ctx* ctx, POOL_function function, void* opaque);
 
 
 /*! POOL_tryAdd() :
- *  Add the job `function(opaque)` to thread pool _if_ a worker is available.
+ *  Add the job `function(opaque)` to thread pool _if_ a queue slot is available.
  *  Returns immediately even if not (does not block).
  * @return : 1 if successful, 0 if not.
  */

src/borg/algorithms/zstd/lib/common/portability_macros.h

@@ -0,0 +1,156 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_PORTABILITY_MACROS_H
+#define ZSTD_PORTABILITY_MACROS_H
+
+/**
+ * This header file contains macro definitions to support portability.
+ * This header is shared between C and ASM code, so it MUST only
+ * contain macro definitions. It MUST not contain any C code.
+ *
+ * This header ONLY defines macros to detect platforms/feature support.
+ *
+ */
+
+
+/* compat. with non-clang compilers */
+#ifndef __has_attribute
+  #define __has_attribute(x) 0
+#endif
+
+/* compat. with non-clang compilers */
+#ifndef __has_builtin
+#  define __has_builtin(x) 0
+#endif
+
+/* compat. with non-clang compilers */
+#ifndef __has_feature
+#  define __has_feature(x) 0
+#endif
+
+/* detects whether we are being compiled under msan */
+#ifndef ZSTD_MEMORY_SANITIZER
+#  if __has_feature(memory_sanitizer)
+#    define ZSTD_MEMORY_SANITIZER 1
+#  else
+#    define ZSTD_MEMORY_SANITIZER 0
+#  endif
+#endif
+
+/* detects whether we are being compiled under asan */
+#ifndef ZSTD_ADDRESS_SANITIZER
+#  if __has_feature(address_sanitizer)
+#    define ZSTD_ADDRESS_SANITIZER 1
+#  elif defined(__SANITIZE_ADDRESS__)
+#    define ZSTD_ADDRESS_SANITIZER 1
+#  else
+#    define ZSTD_ADDRESS_SANITIZER 0
+#  endif
+#endif
+
+/* detects whether we are being compiled under dfsan */
+#ifndef ZSTD_DATAFLOW_SANITIZER
+# if __has_feature(dataflow_sanitizer)
+#  define ZSTD_DATAFLOW_SANITIZER 1
+# else
+#  define ZSTD_DATAFLOW_SANITIZER 0
+# endif
+#endif
+
+/* Mark the internal assembly functions as hidden  */
+#ifdef __ELF__
+# define ZSTD_HIDE_ASM_FUNCTION(func) .hidden func
+#else
+# define ZSTD_HIDE_ASM_FUNCTION(func)
+#endif
+
+/* Enable runtime BMI2 dispatch based on the CPU.
+ * Enabled for clang & gcc >=4.8 on x86 when BMI2 isn't enabled by default.
+ */
+#ifndef DYNAMIC_BMI2
+  #if ((defined(__clang__) && __has_attribute(__target__)) \
+      || (defined(__GNUC__) \
+          && (__GNUC__ >= 5 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8)))) \
+      && (defined(__x86_64__) || defined(_M_X64)) \
+      && !defined(__BMI2__)
+  #  define DYNAMIC_BMI2 1
+  #else
+  #  define DYNAMIC_BMI2 0
+  #endif
+#endif
+
+/**
+ * Only enable assembly for GNUC compatible compilers,
+ * because other platforms may not support GAS assembly syntax.
+ *
+ * Only enable assembly for Linux / MacOS, other platforms may
+ * work, but they haven't been tested. This could likely be
+ * extended to BSD systems.
+ *
+ * Disable assembly when MSAN is enabled, because MSAN requires
+ * 100% of code to be instrumented to work.
+ */
+#if defined(__GNUC__)
+#  if defined(__linux__) || defined(__linux) || defined(__APPLE__)
+#    if ZSTD_MEMORY_SANITIZER
+#      define ZSTD_ASM_SUPPORTED 0
+#    elif ZSTD_DATAFLOW_SANITIZER
+#      define ZSTD_ASM_SUPPORTED 0
+#    else
+#      define ZSTD_ASM_SUPPORTED 1
+#    endif
+#  else
+#    define ZSTD_ASM_SUPPORTED 0
+#  endif
+#else
+#  define ZSTD_ASM_SUPPORTED 0
+#endif
+
+/**
+ * Determines whether we should enable assembly for x86-64
+ * with BMI2.
+ *
+ * Enable if all of the following conditions hold:
+ * - ASM hasn't been explicitly disabled by defining ZSTD_DISABLE_ASM
+ * - Assembly is supported
+ * - We are compiling for x86-64 and either:
+ *   - DYNAMIC_BMI2 is enabled
+ *   - BMI2 is supported at compile time
+ */
+#if !defined(ZSTD_DISABLE_ASM) &&                                 \
+    ZSTD_ASM_SUPPORTED &&                                         \
+    defined(__x86_64__) &&                                        \
+    (DYNAMIC_BMI2 || defined(__BMI2__))
+# define ZSTD_ENABLE_ASM_X86_64_BMI2 1
+#else
+# define ZSTD_ENABLE_ASM_X86_64_BMI2 0
+#endif
+
+/*
+ * For x86 ELF targets, add .note.gnu.property section for Intel CET in
+ * assembly sources when CET is enabled.
+ *
+ * Additionally, any function that may be called indirectly must begin
+ * with ZSTD_CET_ENDBRANCH.
+ */
+#if defined(__ELF__) && (defined(__x86_64__) || defined(__i386__)) \
+    && defined(__has_include)
+# if __has_include(<cet.h>)
+#  include <cet.h>
+#  define ZSTD_CET_ENDBRANCH _CET_ENDBR
+# endif
+#endif
+
+#ifndef ZSTD_CET_ENDBRANCH
+# define ZSTD_CET_ENDBRANCH
+#endif
+
+#endif /* ZSTD_PORTABILITY_MACROS_H */

src/borg/algorithms/zstd/lib/common/threading.c

@@ -23,8 +23,7 @@ int g_ZSTD_threading_useless_symbol;
 #if defined(ZSTD_MULTITHREAD) && defined(_WIN32)
 
 /**
- * Windows minimalist Pthread Wrapper, based on :
- * http://www.cse.wustl.edu/~schmidt/win32-cv-1.html
+ * Windows minimalist Pthread Wrapper
  */
 
 
@@ -35,37 +34,92 @@ int g_ZSTD_threading_useless_symbol;
 
 /* ===  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)
 {
-    ZSTD_pthread_t* const thread = (ZSTD_pthread_t*) arg;
-    thread->arg = thread->start_routine(thread->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;
-    thread->arg = arg;
-    thread->start_routine = start_routine;
-    thread->handle = (HANDLE) _beginthreadex(NULL, 0, worker, thread, 0, NULL);
 
-    if (!thread->handle)
+    thread_param.start_routine = start_routine;
+    thread_param.arg = arg;
+    thread_param.initialized = 0;
+    *thread = NULL;
+
+    /* 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) {
+        ZSTD_pthread_mutex_destroy(&thread_param.initialized_mutex);
+        ZSTD_pthread_cond_destroy(&thread_param.initialized_cond);
         return errno;
-    else
-        return 0;
+    }
+
+    /* 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, void **value_ptr)
+int ZSTD_pthread_join(ZSTD_pthread_t thread)
 {
     DWORD result;
 
-    if (!thread.handle) return 0;
+    if (!thread) return 0;
+
+    result = WaitForSingleObject(thread, INFINITE);
+    CloseHandle(thread);
 
-    result = WaitForSingleObject(thread.handle, INFINITE);
     switch (result) {
     case WAIT_OBJECT_0:
-        if (value_ptr) *value_ptr = thread.arg;
         return 0;
     case WAIT_ABANDONED:
         return EINVAL;
@@ -78,11 +132,12 @@ int ZSTD_pthread_join(ZSTD_pthread_t thread, void **value_ptr)
 
 #if defined(ZSTD_MULTITHREAD) && DEBUGLEVEL >= 1 && !defined(_WIN32)
 
-#include <stdlib.h>
+#define ZSTD_DEPS_NEED_MALLOC
+#include "zstd_deps.h"
 
 int ZSTD_pthread_mutex_init(ZSTD_pthread_mutex_t* mutex, pthread_mutexattr_t const* attr)
 {
-    *mutex = (pthread_mutex_t*)malloc(sizeof(pthread_mutex_t));
+    *mutex = (pthread_mutex_t*)ZSTD_malloc(sizeof(pthread_mutex_t));
     if (!*mutex)
         return 1;
     return pthread_mutex_init(*mutex, attr);
@@ -94,14 +149,14 @@ int ZSTD_pthread_mutex_destroy(ZSTD_pthread_mutex_t* mutex)
         return 0;
     {
         int const ret = pthread_mutex_destroy(*mutex);
-        free(*mutex);
+        ZSTD_free(*mutex);
         return ret;
     }
 }
 
 int ZSTD_pthread_cond_init(ZSTD_pthread_cond_t* cond, pthread_condattr_t const* attr)
 {
-    *cond = (pthread_cond_t*)malloc(sizeof(pthread_cond_t));
+    *cond = (pthread_cond_t*)ZSTD_malloc(sizeof(pthread_cond_t));
     if (!*cond)
         return 1;
     return pthread_cond_init(*cond, attr);
@@ -113,7 +168,7 @@ int ZSTD_pthread_cond_destroy(ZSTD_pthread_cond_t* cond)
         return 0;
     {
         int const ret = pthread_cond_destroy(*cond);
-        free(*cond);
+        ZSTD_free(*cond);
         return ret;
     }
 }

src/borg/algorithms/zstd/lib/common/threading.h

@@ -23,8 +23,7 @@ extern "C" {
 #if defined(ZSTD_MULTITHREAD) && defined(_WIN32)
 
 /**
- * Windows minimalist Pthread Wrapper, based on :
- * http://www.cse.wustl.edu/~schmidt/win32-cv-1.html
+ * Windows minimalist Pthread Wrapper
  */
 #ifdef WINVER
 #  undef WINVER
@@ -62,16 +61,12 @@ extern "C" {
 #define ZSTD_pthread_cond_broadcast(a)  WakeAllConditionVariable((a))
 
 /* ZSTD_pthread_create() and ZSTD_pthread_join() */
-typedef struct {
-    HANDLE handle;
-    void* (*start_routine)(void*);
-    void* arg;
-} ZSTD_pthread_t;
+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, void** value_ptr);
+int ZSTD_pthread_join(ZSTD_pthread_t thread);
 
 /**
  * add here more wrappers as required
@@ -99,7 +94,7 @@ int ZSTD_pthread_join(ZSTD_pthread_t thread, void** value_ptr);
 
 #define ZSTD_pthread_t                  pthread_t
 #define ZSTD_pthread_create(a, b, c, d) pthread_create((a), (b), (c), (d))
-#define ZSTD_pthread_join(a, b)         pthread_join((a),(b))
+#define ZSTD_pthread_join(a)         pthread_join((a),NULL)
 
 #else /* DEBUGLEVEL >= 1 */
 
@@ -124,7 +119,7 @@ int ZSTD_pthread_cond_destroy(ZSTD_pthread_cond_t* cond);
 
 #define ZSTD_pthread_t                  pthread_t
 #define ZSTD_pthread_create(a, b, c, d) pthread_create((a), (b), (c), (d))
-#define ZSTD_pthread_join(a, b)         pthread_join((a),(b))
+#define ZSTD_pthread_join(a)         pthread_join((a),NULL)
 
 #endif
 

src/borg/algorithms/zstd/lib/common/xxhash.c

@@ -1,11 +1,11 @@
 /*
  *  xxHash - Fast Hash algorithm
- *  Copyright (c) 2012-2020, Yann Collet, Facebook, Inc.
+ *  Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  *  You can contact the author at :
- *  - xxHash homepage: http://www.xxhash.com
+ *  - xxHash homepage: https://cyan4973.github.io/xxHash/
  *  - xxHash source repository : https://github.com/Cyan4973/xxHash
- * 
+ *
  * 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).
@@ -13,852 +13,12 @@
 */
 
 
-/* *************************************
-*  Tuning parameters
-***************************************/
-/*!XXH_FORCE_MEMORY_ACCESS :
- * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
- * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
- * The below switch allow to select different access method for improved performance.
- * Method 0 (default) : use `memcpy()`. Safe and portable.
- * Method 1 : `__packed` statement. It depends on compiler extension (ie, not portable).
- *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
- * Method 2 : direct access. This method doesn't depend on compiler but violate C standard.
- *            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 (ie GCC + ARMv6)
- * See http://stackoverflow.com/a/32095106/646947 for details.
- * Prefer these methods in priority order (0 > 1 > 2)
- */
-#ifndef XXH_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
-#  if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
-#    define XXH_FORCE_MEMORY_ACCESS 2
-#  elif (defined(__INTEL_COMPILER) && !defined(WIN32)) || \
-  (defined(__GNUC__) && ( defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) || defined(__ARM_ARCH_7S__) )) || \
-  defined(__ICCARM__)
-#    define XXH_FORCE_MEMORY_ACCESS 1
-#  endif
-#endif
-
-/*!XXH_ACCEPT_NULL_INPUT_POINTER :
- * If the input pointer is a null pointer, xxHash default behavior is to trigger a memory access error, since it is a bad pointer.
- * When this option is enabled, xxHash output for null input pointers will be the same as a null-length input.
- * By default, this option is disabled. To enable it, uncomment below define :
- */
-/* #define XXH_ACCEPT_NULL_INPUT_POINTER 1 */
-
-/*!XXH_FORCE_NATIVE_FORMAT :
- * By default, xxHash library provides endian-independent Hash values, based on little-endian convention.
- * Results are therefore identical for little-endian and big-endian CPU.
- * This comes at a performance cost for big-endian CPU, since some swapping is required to emulate little-endian format.
- * Should endian-independence be of no importance for your application, you may set the #define below to 1,
- * to improve speed for Big-endian CPU.
- * This option has no impact on Little_Endian CPU.
- */
-#ifndef XXH_FORCE_NATIVE_FORMAT   /* can be defined externally */
-#  define XXH_FORCE_NATIVE_FORMAT 0
-#endif
 
-/*!XXH_FORCE_ALIGN_CHECK :
- * This is a minor performance trick, only useful with lots of very small keys.
- * It means : check for aligned/unaligned input.
- * The check costs one initial branch per hash; set to 0 when the input data
- * is guaranteed to be aligned.
+/*
+ * xxhash.c instantiates functions defined in xxhash.h
  */
-#ifndef XXH_FORCE_ALIGN_CHECK /* can be defined externally */
-#  if defined(__i386) || defined(_M_IX86) || defined(__x86_64__) || defined(_M_X64)
-#    define XXH_FORCE_ALIGN_CHECK 0
-#  else
-#    define XXH_FORCE_ALIGN_CHECK 1
-#  endif
-#endif
-
 
-/* *************************************
-*  Includes & Memory related functions
-***************************************/
-/* Modify the local functions below should you wish to use some other memory routines */
-/* for malloc(), free() */
-#include <stdlib.h>
-#include <stddef.h>     /* size_t */
-static void* XXH_malloc(size_t s) { return malloc(s); }
-static void  XXH_free  (void* p)  { free(p); }
-/* for memcpy() */
-#include <string.h>
-static void* XXH_memcpy(void* dest, const void* src, size_t size) { return memcpy(dest,src,size); }
+#define XXH_STATIC_LINKING_ONLY   /* access advanced declarations */
+#define XXH_IMPLEMENTATION   /* access definitions */
 
-#ifndef XXH_STATIC_LINKING_ONLY
-#  define XXH_STATIC_LINKING_ONLY
-#endif
 #include "xxhash.h"
-
-
-/* *************************************
-*  Compiler Specific Options
-***************************************/
-#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
-
-#define FORCE_INLINE_TEMPLATE static INLINE_KEYWORD FORCE_INLINE_ATTR
-
-
-#ifdef _MSC_VER
-#  pragma warning(disable : 4127)      /* disable: C4127: conditional expression is constant */
-#endif
-
-
-/* *************************************
-*  Basic Types
-***************************************/
-#ifndef MEM_MODULE
-# define MEM_MODULE
-# if !defined (__VMS) && (defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
-#   include <stdint.h>
-    typedef uint8_t  BYTE;
-    typedef uint16_t U16;
-    typedef uint32_t U32;
-    typedef  int32_t S32;
-    typedef uint64_t U64;
-#  else
-    typedef unsigned char      BYTE;
-    typedef unsigned short     U16;
-    typedef unsigned int       U32;
-    typedef   signed int       S32;
-    typedef unsigned long long U64;   /* if your compiler doesn't support unsigned long long, replace by another 64-bit type here. Note that xxhash.h will also need to be updated. */
-#  endif
-#endif
-
-
-#if (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 U32 XXH_read32(const void* memPtr) { return *(const U32*) memPtr; }
-static U64 XXH_read64(const void* memPtr) { return *(const U64*) memPtr; }
-
-#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
-
-/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
-/* currently only defined for gcc and icc */
-typedef union { U32 u32; U64 u64; } __attribute__((packed)) unalign;
-
-static U32 XXH_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
-static U64 XXH_read64(const void* ptr) { return ((const unalign*)ptr)->u64; }
-
-#else
-
-/* portable and safe solution. Generally efficient.
- * see : http://stackoverflow.com/a/32095106/646947
- */
-
-static U32 XXH_read32(const void* memPtr)
-{
-    U32 val;
-    memcpy(&val, memPtr, sizeof(val));
-    return val;
-}
-
-static U64 XXH_read64(const void* memPtr)
-{
-    U64 val;
-    memcpy(&val, memPtr, sizeof(val));
-    return val;
-}
-
-#endif   /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
-
-
-/* ****************************************
-*  Compiler-specific Functions and Macros
-******************************************/
-#define GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
-
-/* Note : although _rotl exists for minGW (GCC under windows), performance seems poor */
-#if defined(_MSC_VER)
-#  define XXH_rotl32(x,r) _rotl(x,r)
-#  define XXH_rotl64(x,r) _rotl64(x,r)
-#else
-#if defined(__ICCARM__)
-#  include <intrinsics.h>
-#  define XXH_rotl32(x,r) __ROR(x,(32 - r))
-#else
-#  define XXH_rotl32(x,r) ((x << r) | (x >> (32 - r)))
-#endif
-#  define XXH_rotl64(x,r) ((x << r) | (x >> (64 - r)))
-#endif
-
-#if defined(_MSC_VER)     /* Visual Studio */
-#  define XXH_swap32 _byteswap_ulong
-#  define XXH_swap64 _byteswap_uint64
-#elif GCC_VERSION >= 403
-#  define XXH_swap32 __builtin_bswap32
-#  define XXH_swap64 __builtin_bswap64
-#else
-static U32 XXH_swap32 (U32 x)
-{
-    return  ((x << 24) & 0xff000000 ) |
-            ((x <<  8) & 0x00ff0000 ) |
-            ((x >>  8) & 0x0000ff00 ) |
-            ((x >> 24) & 0x000000ff );
-}
-static U64 XXH_swap64 (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
-
-
-/* *************************************
-*  Architecture Macros
-***************************************/
-typedef enum { XXH_bigEndian=0, XXH_littleEndian=1 } XXH_endianess;
-
-/* XXH_CPU_LITTLE_ENDIAN can be defined externally, for example on the compiler command line */
-#ifndef XXH_CPU_LITTLE_ENDIAN
-    static const int g_one = 1;
-#   define XXH_CPU_LITTLE_ENDIAN   (*(const char*)(&g_one))
-#endif
-
-
-/* ***************************
-*  Memory reads
-*****************************/
-typedef enum { XXH_aligned, XXH_unaligned } XXH_alignment;
-
-FORCE_INLINE_TEMPLATE U32 XXH_readLE32_align(const void* ptr, XXH_endianess endian, XXH_alignment align)
-{
-    if (align==XXH_unaligned)
-        return endian==XXH_littleEndian ? XXH_read32(ptr) : XXH_swap32(XXH_read32(ptr));
-    else
-        return endian==XXH_littleEndian ? *(const U32*)ptr : XXH_swap32(*(const U32*)ptr);
-}
-
-FORCE_INLINE_TEMPLATE U32 XXH_readLE32(const void* ptr, XXH_endianess endian)
-{
-    return XXH_readLE32_align(ptr, endian, XXH_unaligned);
-}
-
-static U32 XXH_readBE32(const void* ptr)
-{
-    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap32(XXH_read32(ptr)) : XXH_read32(ptr);
-}
-
-FORCE_INLINE_TEMPLATE U64 XXH_readLE64_align(const void* ptr, XXH_endianess endian, XXH_alignment align)
-{
-    if (align==XXH_unaligned)
-        return endian==XXH_littleEndian ? XXH_read64(ptr) : XXH_swap64(XXH_read64(ptr));
-    else
-        return endian==XXH_littleEndian ? *(const U64*)ptr : XXH_swap64(*(const U64*)ptr);
-}
-
-FORCE_INLINE_TEMPLATE U64 XXH_readLE64(const void* ptr, XXH_endianess endian)
-{
-    return XXH_readLE64_align(ptr, endian, XXH_unaligned);
-}
-
-static U64 XXH_readBE64(const void* ptr)
-{
-    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap64(XXH_read64(ptr)) : XXH_read64(ptr);
-}
-
-
-/* *************************************
-*  Macros
-***************************************/
-#define XXH_STATIC_ASSERT(c)   { enum { XXH_static_assert = 1/(int)(!!(c)) }; }    /* use only *after* variable declarations */
-
-
-/* *************************************
-*  Constants
-***************************************/
-static const U32 PRIME32_1 = 2654435761U;
-static const U32 PRIME32_2 = 2246822519U;
-static const U32 PRIME32_3 = 3266489917U;
-static const U32 PRIME32_4 =  668265263U;
-static const U32 PRIME32_5 =  374761393U;
-
-static const U64 PRIME64_1 = 11400714785074694791ULL;
-static const U64 PRIME64_2 = 14029467366897019727ULL;
-static const U64 PRIME64_3 =  1609587929392839161ULL;
-static const U64 PRIME64_4 =  9650029242287828579ULL;
-static const U64 PRIME64_5 =  2870177450012600261ULL;
-
-XXH_PUBLIC_API unsigned XXH_versionNumber (void) { return XXH_VERSION_NUMBER; }
-
-
-/* **************************
-*  Utils
-****************************/
-XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* restrict dstState, const XXH32_state_t* restrict srcState)
-{
-    memcpy(dstState, srcState, sizeof(*dstState));
-}
-
-XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t* restrict dstState, const XXH64_state_t* restrict srcState)
-{
-    memcpy(dstState, srcState, sizeof(*dstState));
-}
-
-
-/* ***************************
-*  Simple Hash Functions
-*****************************/
-
-static U32 XXH32_round(U32 seed, U32 input)
-{
-    seed += input * PRIME32_2;
-    seed  = XXH_rotl32(seed, 13);
-    seed *= PRIME32_1;
-    return seed;
-}
-
-FORCE_INLINE_TEMPLATE U32 XXH32_endian_align(const void* input, size_t len, U32 seed, XXH_endianess endian, XXH_alignment align)
-{
-    const BYTE* p = (const BYTE*)input;
-    const BYTE* bEnd = p + len;
-    U32 h32;
-#define XXH_get32bits(p) XXH_readLE32_align(p, endian, align)
-
-#ifdef XXH_ACCEPT_NULL_INPUT_POINTER
-    if (p==NULL) {
-        len=0;
-        bEnd=p=(const BYTE*)(size_t)16;
-    }
-#endif
-
-    if (len>=16) {
-        const BYTE* const limit = bEnd - 16;
-        U32 v1 = seed + PRIME32_1 + PRIME32_2;
-        U32 v2 = seed + PRIME32_2;
-        U32 v3 = seed + 0;
-        U32 v4 = seed - PRIME32_1;
-
-        do {
-            v1 = XXH32_round(v1, XXH_get32bits(p)); p+=4;
-            v2 = XXH32_round(v2, XXH_get32bits(p)); p+=4;
-            v3 = XXH32_round(v3, XXH_get32bits(p)); p+=4;
-            v4 = XXH32_round(v4, XXH_get32bits(p)); p+=4;
-        } while (p<=limit);
-
-        h32 = XXH_rotl32(v1, 1) + XXH_rotl32(v2, 7) + XXH_rotl32(v3, 12) + XXH_rotl32(v4, 18);
-    } else {
-        h32  = seed + PRIME32_5;
-    }
-
-    h32 += (U32) len;
-
-    while (p+4<=bEnd) {
-        h32 += XXH_get32bits(p) * PRIME32_3;
-        h32  = XXH_rotl32(h32, 17) * PRIME32_4 ;
-        p+=4;
-    }
-
-    while (p<bEnd) {
-        h32 += (*p) * PRIME32_5;
-        h32 = XXH_rotl32(h32, 11) * PRIME32_1 ;
-        p++;
-    }
-
-    h32 ^= h32 >> 15;
-    h32 *= PRIME32_2;
-    h32 ^= h32 >> 13;
-    h32 *= PRIME32_3;
-    h32 ^= h32 >> 16;
-
-    return h32;
-}
-
-
-XXH_PUBLIC_API unsigned int XXH32 (const void* input, size_t len, unsigned int seed)
-{
-#if 0
-    /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
-    XXH32_CREATESTATE_STATIC(state);
-    XXH32_reset(state, seed);
-    XXH32_update(state, input, len);
-    return XXH32_digest(state);
-#else
-    XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN;
-
-    if (XXH_FORCE_ALIGN_CHECK) {
-        if ((((size_t)input) & 3) == 0) {   /* Input is 4-bytes aligned, leverage the speed benefit */
-            if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
-                return XXH32_endian_align(input, len, seed, XXH_littleEndian, XXH_aligned);
-            else
-                return XXH32_endian_align(input, len, seed, XXH_bigEndian, XXH_aligned);
-    }   }
-
-    if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
-        return XXH32_endian_align(input, len, seed, XXH_littleEndian, XXH_unaligned);
-    else
-        return XXH32_endian_align(input, len, seed, XXH_bigEndian, XXH_unaligned);
-#endif
-}
-
-
-static U64 XXH64_round(U64 acc, U64 input)
-{
-    acc += input * PRIME64_2;
-    acc  = XXH_rotl64(acc, 31);
-    acc *= PRIME64_1;
-    return acc;
-}
-
-static U64 XXH64_mergeRound(U64 acc, U64 val)
-{
-    val  = XXH64_round(0, val);
-    acc ^= val;
-    acc  = acc * PRIME64_1 + PRIME64_4;
-    return acc;
-}
-
-FORCE_INLINE_TEMPLATE U64 XXH64_endian_align(const void* input, size_t len, U64 seed, XXH_endianess endian, XXH_alignment align)
-{
-    const BYTE* p = (const BYTE*)input;
-    const BYTE* const bEnd = p + len;
-    U64 h64;
-#define XXH_get64bits(p) XXH_readLE64_align(p, endian, align)
-
-#ifdef XXH_ACCEPT_NULL_INPUT_POINTER
-    if (p==NULL) {
-        len=0;
-        bEnd=p=(const BYTE*)(size_t)32;
-    }
-#endif
-
-    if (len>=32) {
-        const BYTE* const limit = bEnd - 32;
-        U64 v1 = seed + PRIME64_1 + PRIME64_2;
-        U64 v2 = seed + PRIME64_2;
-        U64 v3 = seed + 0;
-        U64 v4 = seed - PRIME64_1;
-
-        do {
-            v1 = XXH64_round(v1, XXH_get64bits(p)); p+=8;
-            v2 = XXH64_round(v2, XXH_get64bits(p)); p+=8;
-            v3 = XXH64_round(v3, XXH_get64bits(p)); p+=8;
-            v4 = XXH64_round(v4, XXH_get64bits(p)); p+=8;
-        } while (p<=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 + PRIME64_5;
-    }
-
-    h64 += (U64) len;
-
-    while (p+8<=bEnd) {
-        U64 const k1 = XXH64_round(0, XXH_get64bits(p));
-        h64 ^= k1;
-        h64  = XXH_rotl64(h64,27) * PRIME64_1 + PRIME64_4;
-        p+=8;
-    }
-
-    if (p+4<=bEnd) {
-        h64 ^= (U64)(XXH_get32bits(p)) * PRIME64_1;
-        h64 = XXH_rotl64(h64, 23) * PRIME64_2 + PRIME64_3;
-        p+=4;
-    }
-
-    while (p<bEnd) {
-        h64 ^= (*p) * PRIME64_5;
-        h64 = XXH_rotl64(h64, 11) * PRIME64_1;
-        p++;
-    }
-
-    h64 ^= h64 >> 33;
-    h64 *= PRIME64_2;
-    h64 ^= h64 >> 29;
-    h64 *= PRIME64_3;
-    h64 ^= h64 >> 32;
-
-    return h64;
-}
-
-
-XXH_PUBLIC_API unsigned long long XXH64 (const void* input, size_t len, unsigned long long seed)
-{
-#if 0
-    /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
-    XXH64_CREATESTATE_STATIC(state);
-    XXH64_reset(state, seed);
-    XXH64_update(state, input, len);
-    return XXH64_digest(state);
-#else
-    XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN;
-
-    if (XXH_FORCE_ALIGN_CHECK) {
-        if ((((size_t)input) & 7)==0) {  /* Input is aligned, let's leverage the speed advantage */
-            if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
-                return XXH64_endian_align(input, len, seed, XXH_littleEndian, XXH_aligned);
-            else
-                return XXH64_endian_align(input, len, seed, XXH_bigEndian, XXH_aligned);
-    }   }
-
-    if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
-        return XXH64_endian_align(input, len, seed, XXH_littleEndian, XXH_unaligned);
-    else
-        return XXH64_endian_align(input, len, seed, XXH_bigEndian, XXH_unaligned);
-#endif
-}
-
-
-/* **************************************************
-*  Advanced Hash Functions
-****************************************************/
-
-XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void)
-{
-    return (XXH32_state_t*)XXH_malloc(sizeof(XXH32_state_t));
-}
-XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr)
-{
-    XXH_free(statePtr);
-    return XXH_OK;
-}
-
-XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void)
-{
-    return (XXH64_state_t*)XXH_malloc(sizeof(XXH64_state_t));
-}
-XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr)
-{
-    XXH_free(statePtr);
-    return XXH_OK;
-}
-
-
-/*** Hash feed ***/
-
-XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, unsigned int seed)
-{
-    XXH32_state_t state;   /* using a local state to memcpy() in order to avoid strict-aliasing warnings */
-    memset(&state, 0, sizeof(state)-4);   /* do not write into reserved, for future removal */
-    state.v1 = seed + PRIME32_1 + PRIME32_2;
-    state.v2 = seed + PRIME32_2;
-    state.v3 = seed + 0;
-    state.v4 = seed - PRIME32_1;
-    memcpy(statePtr, &state, sizeof(state));
-    return XXH_OK;
-}
-
-
-XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH64_state_t* statePtr, unsigned long long seed)
-{
-    XXH64_state_t state;   /* using a local state to memcpy() in order to avoid strict-aliasing warnings */
-    memset(&state, 0, sizeof(state)-8);   /* do not write into reserved, for future removal */
-    state.v1 = seed + PRIME64_1 + PRIME64_2;
-    state.v2 = seed + PRIME64_2;
-    state.v3 = seed + 0;
-    state.v4 = seed - PRIME64_1;
-    memcpy(statePtr, &state, sizeof(state));
-    return XXH_OK;
-}
-
-
-FORCE_INLINE_TEMPLATE XXH_errorcode XXH32_update_endian (XXH32_state_t* state, const void* input, size_t len, XXH_endianess endian)
-{
-    const BYTE* p = (const BYTE*)input;
-    const BYTE* const bEnd = p + len;
-
-#ifdef XXH_ACCEPT_NULL_INPUT_POINTER
-    if (input==NULL) return XXH_ERROR;
-#endif
-
-    state->total_len_32 += (unsigned)len;
-    state->large_len |= (len>=16) | (state->total_len_32>=16);
-
-    if (state->memsize + len < 16)  {   /* fill in tmp buffer */
-        XXH_memcpy((BYTE*)(state->mem32) + state->memsize, input, len);
-        state->memsize += (unsigned)len;
-        return XXH_OK;
-    }
-
-    if (state->memsize) {   /* some data left from previous update */
-        XXH_memcpy((BYTE*)(state->mem32) + state->memsize, input, 16-state->memsize);
-        {   const U32* p32 = state->mem32;
-            state->v1 = XXH32_round(state->v1, XXH_readLE32(p32, endian)); p32++;
-            state->v2 = XXH32_round(state->v2, XXH_readLE32(p32, endian)); p32++;
-            state->v3 = XXH32_round(state->v3, XXH_readLE32(p32, endian)); p32++;
-            state->v4 = XXH32_round(state->v4, XXH_readLE32(p32, endian)); p32++;
-        }
-        p += 16-state->memsize;
-        state->memsize = 0;
-    }
-
-    if (p <= bEnd-16) {
-        const BYTE* const limit = bEnd - 16;
-        U32 v1 = state->v1;
-        U32 v2 = state->v2;
-        U32 v3 = state->v3;
-        U32 v4 = state->v4;
-
-        do {
-            v1 = XXH32_round(v1, XXH_readLE32(p, endian)); p+=4;
-            v2 = XXH32_round(v2, XXH_readLE32(p, endian)); p+=4;
-            v3 = XXH32_round(v3, XXH_readLE32(p, endian)); p+=4;
-            v4 = XXH32_round(v4, XXH_readLE32(p, endian)); p+=4;
-        } while (p<=limit);
-
-        state->v1 = v1;
-        state->v2 = v2;
-        state->v3 = v3;
-        state->v4 = v4;
-    }
-
-    if (p < bEnd) {
-        XXH_memcpy(state->mem32, p, (size_t)(bEnd-p));
-        state->memsize = (unsigned)(bEnd-p);
-    }
-
-    return XXH_OK;
-}
-
-XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* state_in, const void* input, size_t len)
-{
-    XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN;
-
-    if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
-        return XXH32_update_endian(state_in, input, len, XXH_littleEndian);
-    else
-        return XXH32_update_endian(state_in, input, len, XXH_bigEndian);
-}
-
-
-
-FORCE_INLINE_TEMPLATE U32 XXH32_digest_endian (const XXH32_state_t* state, XXH_endianess endian)
-{
-    const BYTE * p = (const BYTE*)state->mem32;
-    const BYTE* const bEnd = (const BYTE*)(state->mem32) + state->memsize;
-    U32 h32;
-
-    if (state->large_len) {
-        h32 = XXH_rotl32(state->v1, 1) + XXH_rotl32(state->v2, 7) + XXH_rotl32(state->v3, 12) + XXH_rotl32(state->v4, 18);
-    } else {
-        h32 = state->v3 /* == seed */ + PRIME32_5;
-    }
-
-    h32 += state->total_len_32;
-
-    while (p+4<=bEnd) {
-        h32 += XXH_readLE32(p, endian) * PRIME32_3;
-        h32  = XXH_rotl32(h32, 17) * PRIME32_4;
-        p+=4;
-    }
-
-    while (p<bEnd) {
-        h32 += (*p) * PRIME32_5;
-        h32  = XXH_rotl32(h32, 11) * PRIME32_1;
-        p++;
-    }
-
-    h32 ^= h32 >> 15;
-    h32 *= PRIME32_2;
-    h32 ^= h32 >> 13;
-    h32 *= PRIME32_3;
-    h32 ^= h32 >> 16;
-
-    return h32;
-}
-
-
-XXH_PUBLIC_API unsigned int XXH32_digest (const XXH32_state_t* state_in)
-{
-    XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN;
-
-    if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
-        return XXH32_digest_endian(state_in, XXH_littleEndian);
-    else
-        return XXH32_digest_endian(state_in, XXH_bigEndian);
-}
-
-
-
-/* **** XXH64 **** */
-
-FORCE_INLINE_TEMPLATE XXH_errorcode XXH64_update_endian (XXH64_state_t* state, const void* input, size_t len, XXH_endianess endian)
-{
-    const BYTE* p = (const BYTE*)input;
-    const BYTE* const bEnd = p + len;
-
-#ifdef XXH_ACCEPT_NULL_INPUT_POINTER
-    if (input==NULL) return XXH_ERROR;
-#endif
-
-    state->total_len += len;
-
-    if (state->memsize + len < 32) {  /* fill in tmp buffer */
-        if (input != NULL) {
-            XXH_memcpy(((BYTE*)state->mem64) + state->memsize, input, len);
-        }
-        state->memsize += (U32)len;
-        return XXH_OK;
-    }
-
-    if (state->memsize) {   /* tmp buffer is full */
-        XXH_memcpy(((BYTE*)state->mem64) + state->memsize, input, 32-state->memsize);
-        state->v1 = XXH64_round(state->v1, XXH_readLE64(state->mem64+0, endian));
-        state->v2 = XXH64_round(state->v2, XXH_readLE64(state->mem64+1, endian));
-        state->v3 = XXH64_round(state->v3, XXH_readLE64(state->mem64+2, endian));
-        state->v4 = XXH64_round(state->v4, XXH_readLE64(state->mem64+3, endian));
-        p += 32-state->memsize;
-        state->memsize = 0;
-    }
-
-    if (p+32 <= bEnd) {
-        const BYTE* const limit = bEnd - 32;
-        U64 v1 = state->v1;
-        U64 v2 = state->v2;
-        U64 v3 = state->v3;
-        U64 v4 = state->v4;
-
-        do {
-            v1 = XXH64_round(v1, XXH_readLE64(p, endian)); p+=8;
-            v2 = XXH64_round(v2, XXH_readLE64(p, endian)); p+=8;
-            v3 = XXH64_round(v3, XXH_readLE64(p, endian)); p+=8;
-            v4 = XXH64_round(v4, XXH_readLE64(p, endian)); p+=8;
-        } while (p<=limit);
-
-        state->v1 = v1;
-        state->v2 = v2;
-        state->v3 = v3;
-        state->v4 = v4;
-    }
-
-    if (p < bEnd) {
-        XXH_memcpy(state->mem64, p, (size_t)(bEnd-p));
-        state->memsize = (unsigned)(bEnd-p);
-    }
-
-    return XXH_OK;
-}
-
-XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH64_state_t* state_in, const void* input, size_t len)
-{
-    XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN;
-
-    if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
-        return XXH64_update_endian(state_in, input, len, XXH_littleEndian);
-    else
-        return XXH64_update_endian(state_in, input, len, XXH_bigEndian);
-}
-
-
-
-FORCE_INLINE_TEMPLATE U64 XXH64_digest_endian (const XXH64_state_t* state, XXH_endianess endian)
-{
-    const BYTE * p = (const BYTE*)state->mem64;
-    const BYTE* const bEnd = (const BYTE*)state->mem64 + state->memsize;
-    U64 h64;
-
-    if (state->total_len >= 32) {
-        U64 const v1 = state->v1;
-        U64 const v2 = state->v2;
-        U64 const v3 = state->v3;
-        U64 const v4 = state->v4;
-
-        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  = state->v3 + PRIME64_5;
-    }
-
-    h64 += (U64) state->total_len;
-
-    while (p+8<=bEnd) {
-        U64 const k1 = XXH64_round(0, XXH_readLE64(p, endian));
-        h64 ^= k1;
-        h64  = XXH_rotl64(h64,27) * PRIME64_1 + PRIME64_4;
-        p+=8;
-    }
-
-    if (p+4<=bEnd) {
-        h64 ^= (U64)(XXH_readLE32(p, endian)) * PRIME64_1;
-        h64  = XXH_rotl64(h64, 23) * PRIME64_2 + PRIME64_3;
-        p+=4;
-    }
-
-    while (p<bEnd) {
-        h64 ^= (*p) * PRIME64_5;
-        h64  = XXH_rotl64(h64, 11) * PRIME64_1;
-        p++;
-    }
-
-    h64 ^= h64 >> 33;
-    h64 *= PRIME64_2;
-    h64 ^= h64 >> 29;
-    h64 *= PRIME64_3;
-    h64 ^= h64 >> 32;
-
-    return h64;
-}
-
-
-XXH_PUBLIC_API unsigned long long XXH64_digest (const XXH64_state_t* state_in)
-{
-    XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN;
-
-    if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
-        return XXH64_digest_endian(state_in, XXH_littleEndian);
-    else
-        return XXH64_digest_endian(state_in, XXH_bigEndian);
-}
-
-
-/* **************************
-*  Canonical representation
-****************************/
-
-/*! Default XXH result types are basic unsigned 32 and 64 bits.
-*   The canonical representation follows human-readable write convention, aka big-endian (large digits first).
-*   These functions allow transformation of hash result into and from its canonical format.
-*   This way, hash values can be written into a file or buffer, and remain comparable across different systems and programs.
-*/
-
-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);
-    memcpy(dst, &hash, sizeof(*dst));
-}
-
-XXH_PUBLIC_API void XXH64_canonicalFromHash(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);
-    memcpy(dst, &hash, sizeof(*dst));
-}
-
-XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src)
-{
-    return XXH_readBE32(src);
-}
-
-XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src)
-{
-    return XXH_readBE64(src);
-}

src/borg/algorithms/zstd/lib/common/xxhash.h

@@ -1,20 +1,36 @@
 /*
- * xxHash - Extremely Fast Hash algorithm
- * Header File
- * Copyright (c) 2012-2020, Yann Collet, Facebook, Inc.
+ *  xxHash - Fast Hash algorithm
+ *  Copyright (c) Meta Platforms, Inc. and affiliates.
+ *
+ *  You can contact the author at :
+ *  - xxHash homepage: https://cyan4973.github.io/xxHash/
+ *  - xxHash source repository : https://github.com/Cyan4973/xxHash
  *
- * You can contact the author at :
- * - xxHash source repository : https://github.com/Cyan4973/xxHash
- * 
  * 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.
 */
 
-/* Notice extracted from xxHash homepage :
 
-xxHash is an extremely fast Hash algorithm, running at RAM speed limits.
+#ifndef XXH_NO_XXH3
+# define XXH_NO_XXH3
+#endif
+
+#ifndef XXH_NAMESPACE
+# define XXH_NAMESPACE ZSTD_
+#endif
+
+/*!
+ * @mainpage xxHash
+ *
+ * @file xxhash.h
+ * xxHash prototypes and implementation
+ */
+/* TODO: update */
+/* Notice extracted from xxHash homepage:
+
+xxHash is an extremely fast hash algorithm, running at RAM speed limits.
 It also successfully passes all tests from the SMHasher suite.
 
 Comparison (single thread, Windows Seven 32 bits, using SMHasher on a Core 2 Duo @3GHz)
@@ -22,7 +38,7 @@ Comparison (single thread, Windows Seven 32 bits, using SMHasher on a Core 2 Duo
 Name            Speed       Q.Score   Author
 xxHash          5.4 GB/s     10
 CrapWow         3.2 GB/s      2       Andrew
-MumurHash 3a    2.7 GB/s     10       Austin Appleby
+MurmurHash 3a   2.7 GB/s     10       Austin Appleby
 SpookyHash      2.0 GB/s     10       Bob Jenkins
 SBox            1.4 GB/s      9       Bret Mulvey
 Lookup3         1.2 GB/s      9       Bob Jenkins
@@ -37,8 +53,13 @@ Q.Score is a measure of quality of the hash function.
 It depends on successfully passing SMHasher test set.
 10 is a perfect score.
 
-A 64-bits version, named XXH64, is available since r35.
-It offers much better speed, but for 64-bits applications only.
+Note: SMHasher's CRC32 implementation is not the fastest one.
+Other speed-oriented implementations can be faster,
+especially in combination with PCLMUL instruction:
+https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html?showComment=1552696407071#c3490092340461170735
+
+A 64-bit version, named XXH64, is available since r35.
+It offers much better speed, but for 64-bit applications only.
 Name     Speed on 64 bits    Speed on 32 bits
 XXH64       13.8 GB/s            1.9 GB/s
 XXH32        6.8 GB/s            6.0 GB/s
@@ -48,33 +69,34 @@ XXH32        6.8 GB/s            6.0 GB/s
 extern "C" {
 #endif
 
-#ifndef XXHASH_H_5627135585666179
-#define XXHASH_H_5627135585666179 1
-
-
-/* ****************************
-*  Definitions
-******************************/
-#include <stddef.h>   /* size_t */
-typedef enum { XXH_OK=0, XXH_ERROR } XXH_errorcode;
-
-
 /* ****************************
-*  API modifier
-******************************/
-/** XXH_PRIVATE_API
-*   This is useful if you want to include xxhash functions in `static` mode
-*   in order to inline them, and remove their symbol from the public list.
-*   Methodology :
-*     #define XXH_PRIVATE_API
-*     #include "xxhash.h"
-*   `xxhash.c` is automatically included.
-*   It's not useful to compile and link it as a separate module anymore.
-*/
-#ifdef XXH_PRIVATE_API
-#  ifndef XXH_STATIC_LINKING_ONLY
-#    define XXH_STATIC_LINKING_ONLY
-#  endif
+ *  INLINE mode
+ ******************************/
+/*!
+ * XXH_INLINE_ALL (and XXH_PRIVATE_API)
+ * 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:
+ *     #define XXH_INLINE_ALL
+ *     #include "xxhash.h"
+ *
+ * Do not compile and link xxhash.o as a separate object, as it is not useful.
+ */
+#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 */)
@@ -82,45 +104,205 @@ typedef enum { XXH_OK=0, XXH_ERROR } XXH_errorcode;
 #  elif defined(_MSC_VER)
 #    define XXH_PUBLIC_API static __inline
 #  else
-#    define XXH_PUBLIC_API static   /* this version may generate warnings for unused static functions; disable the relevant warning */
+     /* note: this version may generate warnings for unused static functions */
+#    define XXH_PUBLIC_API static
 #  endif
-#else
-#  define XXH_PUBLIC_API   /* do nothing */
-#endif /* XXH_PRIVATE_API */
 
-/*!XXH_NAMESPACE, aka Namespace Emulation :
+   /*
+    * 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
 
-If you want to include _and expose_ xxHash functions from within your own library,
-but also want to avoid symbol collisions with another library which also includes xxHash,
+    /* 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 */
 
-you can use XXH_NAMESPACE, to automatically prefix any public symbol from xxhash library
-with the value of XXH_NAMESPACE (so avoid to keep it NULL and avoid numeric values).
 
-Note that no change is required within the calling program as long as it includes `xxhash.h` :
-regular symbol name will be automatically translated by this header.
-*/
+
+/* ****************************************************************
+ *  Stable API
+ *****************************************************************/
+#ifndef XXHASH_H_5627135585666179
+#define XXHASH_H_5627135585666179 1
+
+
+/*!
+ * @defgroup public Public API
+ * Contains details on the public xxHash functions.
+ * @{
+ */
+/* 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
+
+#ifdef XXH_DOXYGEN
+/*!
+ * @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 XXH_NAMESPACE to automatically prefix
+ * any public symbol from xxhash library with the value of 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
+
 #ifdef XXH_NAMESPACE
 #  define XXH_CAT(A,B) A##B
 #  define XXH_NAME2(A,B) XXH_CAT(A,B)
-#  define XXH32 XXH_NAME2(XXH_NAMESPACE, XXH32)
-#  define XXH64 XXH_NAME2(XXH_NAMESPACE, XXH64)
 #  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 XXH64_createState XXH_NAME2(XXH_NAMESPACE, XXH64_createState)
 #  define XXH32_freeState XXH_NAME2(XXH_NAMESPACE, XXH32_freeState)
-#  define XXH64_freeState XXH_NAME2(XXH_NAMESPACE, XXH64_freeState)
 #  define XXH32_reset XXH_NAME2(XXH_NAMESPACE, XXH32_reset)
-#  define XXH64_reset XXH_NAME2(XXH_NAMESPACE, XXH64_reset)
 #  define XXH32_update XXH_NAME2(XXH_NAMESPACE, XXH32_update)
-#  define XXH64_update XXH_NAME2(XXH_NAMESPACE, XXH64_update)
 #  define XXH32_digest XXH_NAME2(XXH_NAMESPACE, XXH32_digest)
-#  define XXH64_digest XXH_NAME2(XXH_NAMESPACE, XXH64_digest)
 #  define XXH32_copyState XXH_NAME2(XXH_NAMESPACE, XXH32_copyState)
-#  define XXH64_copyState XXH_NAME2(XXH_NAMESPACE, XXH64_copyState)
 #  define XXH32_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH32_canonicalFromHash)
-#  define XXH64_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH64_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
 
 
@@ -128,156 +310,5375 @@ regular symbol name will be automatically translated by this header.
 *  Version
 ***************************************/
 #define XXH_VERSION_MAJOR    0
-#define XXH_VERSION_MINOR    6
-#define XXH_VERSION_RELEASE  2
+#define XXH_VERSION_MINOR    8
+#define XXH_VERSION_RELEASE  1
 #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 `XXH_VERSION_NUMBER` of the invoked library.
+ */
 XXH_PUBLIC_API unsigned XXH_versionNumber (void);
 
 
 /* ****************************
-*  Simple Hash Functions
+*  Common basic types
 ******************************/
-typedef unsigned int       XXH32_hash_t;
-typedef unsigned long long XXH64_hash_t;
-
-XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t length, unsigned int seed);
-XXH_PUBLIC_API XXH64_hash_t XXH64 (const void* input, size_t length, unsigned long long seed);
-
-/*!
-XXH32() :
-    Calculate the 32-bits hash of sequence "length" bytes stored at memory address "input".
-    The memory between input & input+length must be valid (allocated and read-accessible).
-    "seed" can be used to alter the result predictably.
-    Speed on Core 2 Duo @ 3 GHz (single thread, SMHasher benchmark) : 5.4 GB/s
-XXH64() :
-    Calculate the 64-bits hash of sequence of length "len" stored at memory address "input".
-    "seed" can be used to alter the result predictably.
-    This function runs 2x faster on 64-bits systems, but slower on 32-bits systems (see benchmark).
-*/
+#include <stddef.h>   /* size_t */
+typedef enum { XXH_OK=0, XXH_ERROR } XXH_errorcode;
 
 
-/* ****************************
-*  Streaming Hash Functions
-******************************/
-typedef struct XXH32_state_s XXH32_state_t;   /* incomplete type */
-typedef struct XXH64_state_s XXH64_state_t;   /* incomplete type */
+/*-**********************************************************************
+*  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;
 
-/*! State allocation, compatible with dynamic libraries */
+#elif !defined (__VMS) \
+  && (defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+#   include <stdint.h>
+    typedef uint32_t XXH32_hash_t;
 
-XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void);
-XXH_PUBLIC_API XXH_errorcode  XXH32_freeState(XXH32_state_t* statePtr);
+#else
+#   include <limits.h>
+#   if UINT_MAX == 0xFFFFFFFFUL
+      typedef unsigned int XXH32_hash_t;
+#   else
+#     if ULONG_MAX == 0xFFFFFFFFUL
+        typedef unsigned long XXH32_hash_t;
+#     else
+#       error "unsupported platform: need a 32-bit type"
+#     endif
+#   endif
+#endif
 
-XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void);
-XXH_PUBLIC_API XXH_errorcode  XXH64_freeState(XXH64_state_t* statePtr);
+/*!
+ * @}
+ *
+ * @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 @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.
+ *
+ * Speed on Core 2 Duo @ 3 GHz (single thread, SMHasher benchmark): 5.4 GB/s
+ *
+ * @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 hash value.
+ *
+ * @see
+ *    XXH64(), XXH3_64bits_withSeed(), XXH3_128bits_withSeed(), XXH128():
+ *    Direct equivalents for the other variants of xxHash.
+ * @see
+ *    XXH32_createState(), XXH32_update(), XXH32_digest(): Streaming version.
+ */
+XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed);
+
+/*!
+ * 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()`.
+ *
+ * Example code for incrementally hashing a file:
+ * @code{.c}
+ *    #include <stdio.h>
+ *    #include <xxhash.h>
+ *    #define BUFFER_SIZE 256
+ *
+ *    // Note: XXH64 and XXH3 use the same interface.
+ *    XXH32_hash_t
+ *    hashFile(FILE* stream)
+ *    {
+ *        XXH32_state_t* state;
+ *        unsigned char buf[BUFFER_SIZE];
+ *        size_t amt;
+ *        XXH32_hash_t hash;
+ *
+ *        state = XXH32_createState();       // Create a state
+ *        assert(state != NULL);             // Error check here
+ *        XXH32_reset(state, 0xbaad5eed);    // Reset state with our seed
+ *        while ((amt = fread(buf, 1, sizeof(buf), stream)) != 0) {
+ *            XXH32_update(state, buf, amt); // Hash the file in chunks
+ *        }
+ *        hash = XXH32_digest(state);        // Finalize the hash
+ *        XXH32_freeState(state);            // Clean up
+ *        return hash;
+ *    }
+ * @endcode
+ */
+
+/*!
+ * @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.
+ *
+ * Must be freed with XXH32_freeState().
+ * @return An allocated XXH32_state_t on success, `NULL` on failure.
+ */
+XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void);
+/*!
+ * @brief Frees an @ref XXH32_state_t.
+ *
+ * Must be allocated with XXH32_createState().
+ * @param statePtr A pointer to an @ref XXH32_state_t allocated with @ref XXH32_createState().
+ * @return XXH_OK.
+ */
+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);
 
-/* hash streaming */
+/*!
+ * @brief Resets an @ref XXH32_state_t to begin a new hash.
+ *
+ * This function resets and seeds a state. Call it before @ref XXH32_update().
+ *
+ * @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, @ref XXH_ERROR on failure.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH32_reset  (XXH32_state_t* statePtr, XXH32_hash_t seed);
 
-XXH_PUBLIC_API XXH_errorcode XXH32_reset  (XXH32_state_t* statePtr, unsigned int seed);
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH32_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`.
+ * @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, @ref XXH_ERROR on failure.
+ */
 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.
+ *
+ * @note
+ *   Calling XXH32_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ *  @p statePtr must not be `NULL`.
+ *
+ * @return The calculated xxHash32 value from that state.
+ */
 XXH_PUBLIC_API XXH32_hash_t  XXH32_digest (const XXH32_state_t* statePtr);
 
-XXH_PUBLIC_API XXH_errorcode XXH64_reset  (XXH64_state_t* statePtr, unsigned long long seed);
-XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH64_state_t* statePtr, const void* input, size_t length);
-XXH_PUBLIC_API XXH64_hash_t  XXH64_digest (const XXH64_state_t* statePtr);
+/*******   Canonical representation   *******/
 
 /*
-These functions generate the xxHash of an input provided in multiple segments.
-Note that, for small input, they are slower than single-call functions, due to state management.
-For small input, prefer `XXH32()` and `XXH64()` .
+ * The default return values from XXH functions are unsigned 32 and 64 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.
+ */
 
-XXH state must first be allocated, using XXH*_createState() .
+/*!
+ * @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`.
+ */
+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.
+ */
+XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src);
 
-Start a new hash by initializing state with a seed, using XXH*_reset().
 
-Then, feed the hash state by calling XXH*_update() as many times as necessary.
-Obviously, input must be allocated and read accessible.
-The function returns an error code, with 0 meaning OK, and any other value meaning there is an error.
+#ifdef __has_attribute
+# define XXH_HAS_ATTRIBUTE(x) __has_attribute(x)
+#else
+# define XXH_HAS_ATTRIBUTE(x) 0
+#endif
 
-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.
+/* C-language Attributes are added in C23. */
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ > 201710L) && defined(__has_c_attribute)
+# define XXH_HAS_C_ATTRIBUTE(x) __has_c_attribute(x)
+#else
+# define XXH_HAS_C_ATTRIBUTE(x) 0
+#endif
 
-It's still possible to continue inserting input into the hash state after a digest,
-and generate some new hashes later on, by calling again XXH*_digest().
+#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
 
-When done, free XXH state space if it was allocated dynamically.
+/*
+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(x)
+# define XXH_FALLTHROUGH [[fallthrough]]
+#elif XXH_HAS_CPP_ATTRIBUTE(x)
+# define XXH_FALLTHROUGH [[fallthrough]]
+#elif XXH_HAS_ATTRIBUTE(__fallthrough__)
+# define XXH_FALLTHROUGH __attribute__ ((fallthrough))
+#else
+# define XXH_FALLTHROUGH
+#endif
 
+/*!
+ * @}
+ * @ingroup public
+ * @{
+ */
 
-/* **************************
-*  Utils
-****************************/
-#if !(defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L))   /* ! C99 */
-#  define restrict   /* disable restrict */
+#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 */) )
+#  include <stdint.h>
+   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
 
-XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* restrict dst_state, const XXH32_state_t* restrict src_state);
-XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t* restrict dst_state, const XXH64_state_t* restrict src_state);
+/*!
+ * @}
+ *
+ * @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.
+ */
 
 
-/* **************************
-*  Canonical representation
-****************************/
-/* Default result type for XXH functions are primitive unsigned 32 and 64 bits.
-*  The canonical representation uses human-readable write convention, aka big-endian (large digits first).
-*  These functions allow transformation of hash result into and from its canonical format.
-*  This way, hash values can be written into a file / memory, and remain comparable on different systems and programs.
-*/
-typedef struct { unsigned char digest[4]; } XXH32_canonical_t;
-typedef struct { unsigned char digest[8]; } XXH64_canonical_t;
+/*!
+ * @brief Calculates the 64-bit hash of @p input using xxHash64.
+ *
+ * This function usually runs faster on 64-bit systems, but slower on 32-bit
+ * systems (see benchmark).
+ *
+ * @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 hash.
+ *
+ * @see
+ *    XXH32(), XXH3_64bits_withSeed(), XXH3_128bits_withSeed(), XXH128():
+ *    Direct equivalents for the other variants of xxHash.
+ * @see
+ *    XXH64_createState(), XXH64_update(), XXH64_digest(): Streaming version.
+ */
+XXH_PUBLIC_API XXH64_hash_t XXH64(const void* input, size_t length, XXH64_hash_t seed);
 
-XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash);
-XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t* dst, XXH64_hash_t hash);
+/*******   Streaming   *******/
+/*!
+ * @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 */
+XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void);
+XXH_PUBLIC_API XXH_errorcode  XXH64_freeState(XXH64_state_t* statePtr);
+XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t* dst_state, const XXH64_state_t* src_state);
 
-XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src);
+XXH_PUBLIC_API XXH_errorcode XXH64_reset  (XXH64_state_t* statePtr, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH64_state_t* statePtr, const void* input, size_t length);
+XXH_PUBLIC_API XXH64_hash_t  XXH64_digest (const XXH64_state_t* statePtr);
+
+/*******   Canonical representation   *******/
+typedef struct { unsigned char digest[sizeof(XXH64_hash_t)]; } XXH64_canonical_t;
+XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t* dst, XXH64_hash_t hash);
 XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(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.
+ * Any 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.
+ *
+ * Optimized implementations are provided for AVX512, AVX2, SSE2, NEON, POWER8,
+ * ZVector and scalar targets. This can be controlled via the XXH_VECTOR macro.
+ *
+ * XXH3 implementation is portable:
+ * it has a generic C90 formulation that can be compiled on any platform,
+ * all implementations generage 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
+************************************************************************/
+
+/* XXH3_64bits():
+ * default 64-bit variant, using default secret and default seed of 0.
+ * It's the fastest variant. */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(const void* data, size_t len);
+
+/*
+ * XXH3_64bits_withSeed():
+ * This variant generates a custom secret on the fly
+ * based on default secret altered using the `seed` value.
+ * While this operation is decently fast, note that it's not completely free.
+ * Note: seed==0 produces the same results as XXH3_64bits().
+ */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSeed(const void* data, size_t len, 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
+
+/*
+ * XXH3_64bits_withSecret():
+ * 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 secretSize *must* be large enough (>= 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 "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.
+ */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSecret(const void* data, size_t len, const void* secret, size_t secretSize);
+
+
+/*******   Streaming   *******/
+/*
+ * 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 state struct for the XXH3 streaming API.
+ *
+ * @see XXH3_state_s for details.
+ */
+typedef struct XXH3_state_s XXH3_state_t;
+XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void);
+XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr);
+XXH_PUBLIC_API void XXH3_copyState(XXH3_state_t* dst_state, const XXH3_state_t* src_state);
+
+/*
+ * XXH3_64bits_reset():
+ * Initialize with default parameters.
+ * digest will be equivalent to `XXH3_64bits()`.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH3_state_t* statePtr);
+/*
+ * XXH3_64bits_reset_withSeed():
+ * Generate a custom secret from `seed`, and store it into `statePtr`.
+ * digest will be equivalent to `XXH3_64bits_withSeed()`.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed);
+/*
+ * XXH3_64bits_reset_withSecret():
+ * `secret` is referenced, it _must outlive_ the hash streaming session.
+ * Similar to one-shot API, `secretSize` must be >= `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(XXH3_state_t* statePtr, const void* secret, size_t secretSize);
+
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update (XXH3_state_t* statePtr, const void* input, size_t length);
+XXH_PUBLIC_API XXH64_hash_t  XXH3_64bits_digest (const XXH3_state_t* statePtr);
+
+/* 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;
+
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void* data, size_t len);
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSeed(const void* data, size_t len, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSecret(const void* data, size_t len, const void* secret, size_t secretSize);
+
+/*******   Streaming   *******/
+/*
+ * 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.
+ */
+
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH3_state_t* statePtr);
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed);
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize);
+
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update (XXH3_state_t* statePtr, const void* input, size_t length);
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (const XXH3_state_t* statePtr);
+
+/* 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 */
+
+/*!
+ * XXH128_isEqual():
+ * Return: 1 if `h1` and `h2` are equal, 0 if they are not.
+ */
+XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2);
+
+/*!
+ * XXH128_cmp():
+ *
+ * This comparator is compatible with stdlib's `qsort()`/`bsearch()`.
+ *
+ * return: >0 if *h128_1  > *h128_2
+ *         =0 if *h128_1 == *h128_2
+ *         <0 if *h128_1  < *h128_2
+ */
+XXH_PUBLIC_API int XXH128_cmp(const void* h128_1, const void* h128_2);
+
+
+/*******   Canonical representation   *******/
+typedef struct { unsigned char digest[sizeof(XXH128_hash_t)]; } XXH128_canonical_t;
+XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH128_canonical_t* dst, XXH128_hash_t hash);
+XXH_PUBLIC_API XXH128_hash_t XXH128_hashFromCanonical(const XXH128_canonical_t* src);
+
+
+#endif  /* !XXH_NO_XXH3 */
+#endif  /* XXH_NO_LONG_LONG */
+
+/*!
+ * @}
+ */
 #endif /* XXHASH_H_5627135585666179 */
 
 
 
-/* ================================================================================================
-   This section contains definitions which are not guaranteed to remain stable.
-   They may change in future versions, becoming incompatible with a different version of the library.
-   They shall only be used with static linking.
-   Never use these definitions in association with dynamic linking !
-=================================================================================================== */
-#if defined(XXH_STATIC_LINKING_ONLY) && !defined(XXH_STATIC_H_3543687687345)
-#define XXH_STATIC_H_3543687687345
-
-/* These definitions are only meant to allow allocation of XXH state
-   statically, on stack, or in a struct for example.
-   Do not use members directly. */
-
-   struct XXH32_state_s {
-       unsigned total_len_32;
-       unsigned large_len;
-       unsigned v1;
-       unsigned v2;
-       unsigned v3;
-       unsigned v4;
-       unsigned mem32[4];   /* buffer defined as U32 for alignment */
-       unsigned memsize;
-       unsigned reserved;   /* never read nor write, will be removed in a future version */
-   };   /* typedef'd to XXH32_state_t */
-
-   struct XXH64_state_s {
-       unsigned long long total_len;
-       unsigned long long v1;
-       unsigned long long v2;
-       unsigned long long v3;
-       unsigned long long v4;
-       unsigned long long mem64[4];   /* buffer defined as U64 for alignment */
-       unsigned memsize;
-       unsigned reserved[2];          /* never read nor write, will be removed in a future version */
-   };   /* typedef'd to XXH64_state_t */
-
-
-#  ifdef XXH_PRIVATE_API
-#    include "xxhash.c"   /* include xxhash functions as `static`, for inlining */
-#  endif
+#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 */
+
 
-#endif /* XXH_STATIC_LINKING_ONLY && XXH_STATIC_H_3543687687345 */
+#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
+
+/*!
+ * @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. Similar to `vN` in @ref XXH32_state_s::v1 and @ref XXH64_state_s */
+   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)   { (XXH3_state_ptr)->seed = 0; }
+
+
+/* XXH128() :
+ * simple alias to pre-selected XXH3_128bits variant
+ */
+XXH_PUBLIC_API XXH128_hash_t XXH128(const void* data, size_t len, XXH64_hash_t seed);
+
+
+/* ===   Experimental API   === */
+/* Symbols defined below must be considered tied to a specific library version. */
+
+/*
+ * XXH3_generateSecret():
+ *
+ * Derive a high-entropy secret from any user-defined content, named customSeed.
+ * 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 @secretSize
+ * into an already allocated buffer @secretBuffer.
+ * @secretSize must be >= XXH3_SECRET_SIZE_MIN
+ *
+ * The generated secret can then be used with any `*_withSecret()` variant.
+ * Functions `XXH3_128bits_withSecret()`, `XXH3_64bits_withSecret()`,
+ * `XXH3_128bits_reset_withSecret()` and `XXH3_64bits_reset_withSecret()`
+ * are part of this list. They all accept a `secret` parameter
+ * which must be large enough for implementation reasons (>= XXH3_SECRET_SIZE_MIN)
+ * _and_ feature very high entropy (consist of random-looking bytes).
+ * These conditions can be a high bar to meet, so
+ * XXH3_generateSecret() can be employed to ensure proper quality.
+ *
+ * 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.
+ *
+ * When customSeedSize > 0, supplying NULL as customSeed is undefined behavior.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_generateSecret(void* secretBuffer, size_t secretSize, const void* customSeed, size_t customSeedSize);
+
+
+/*
+ * XXH3_generateSecret_fromSeed():
+ *
+ * Generate the same secret as the _withSeed() variants.
+ *
+ * The resulting secret has a length of XXH3_SECRET_DEFAULT_SIZE (necessarily).
+ * @secretBuffer must be already allocated, of size at least XXH3_SECRET_DEFAULT_SIZE bytes.
+ *
+ * The generated secret can be used in combination with
+ *`*_withSecret()` and `_withSecretandSeed()` variants.
+ * This generator is notably useful in combination with `_withSecretandSeed()`,
+ * as a way to emulate a faster `_withSeed()` variant.
+ */
+XXH_PUBLIC_API void XXH3_generateSecret_fromSeed(void* secretBuffer, XXH64_hash_t seed);
+
+/*
+ * *_withSecretandSeed() :
+ * These variants generate hash values using either
+ * @seed for "short" keys (< XXH3_MIDSIZE_MAX = 240 bytes)
+ * or @secret for "large" keys (>= 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 @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 XXH64_hash_t
+XXH3_64bits_withSecretandSeed(const void* data, size_t len,
+                              const void* secret, size_t secretSize,
+                              XXH64_hash_t seed);
+
+XXH_PUBLIC_API XXH128_hash_t
+XXH3_128bits_withSecretandSeed(const void* data, size_t len,
+                               const void* secret, size_t secretSize,
+                               XXH64_hash_t seed64);
+
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSecretandSeed(XXH3_state_t* statePtr,
+                                    const void* secret, size_t secretSize,
+                                    XXH64_hash_t seed64);
+
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSecretandSeed(XXH3_state_t* statePtr,
+                                     const void* secret, size_t secretSize,
+                                     XXH64_hash_t seed64);
+
+
+#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__((packed))`
+ *   @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_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,
+ * which are platforms known to offer good unaligned memory accesses performance.
+ *
+ * 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), optimizing for size (-Os, -Oz), or using
+ * -fno-inline with GCC or Clang, this will automatically be defined.
+ */
+#  define XXH_NO_INLINE_HINTS 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. */
+#endif /* XXH_DOXYGEN */
+/*!
+ * @}
+ */
+
+#ifndef XXH_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
+   /* prefer __packed__ structures (method 1) for gcc on armv7+ and mips */
+#  if !defined(__clang__) && \
+( \
+    (defined(__INTEL_COMPILER) && !defined(_WIN32)) || \
+    ( \
+        defined(__GNUC__) && ( \
+            (defined(__ARM_ARCH) && __ARM_ARCH >= 7) || \
+            ( \
+                defined(__mips__) && \
+                (__mips <= 5 || __mips_isa_rev < 6) && \
+                (!defined(__mips16) || defined(__mips_mips16e2)) \
+            ) \
+        ) \
+    ) \
+)
+#    define XXH_FORCE_MEMORY_ACCESS 1
+#  endif
+#endif
+
+#ifndef XXH_FORCE_ALIGN_CHECK  /* can be defined externally */
+#  if defined(__i386)  || defined(__x86_64__) || defined(__aarch64__) \
+   || defined(_M_IX86) || defined(_M_X64)     || defined(_M_ARM64) /* visual */
+#    define XXH_FORCE_ALIGN_CHECK 0
+#  else
+#    define XXH_FORCE_ALIGN_CHECK 1
+#  endif
+#endif
+
+#ifndef XXH_NO_INLINE_HINTS
+#  if defined(__OPTIMIZE_SIZE__) /* -Os, -Oz */ \
+   || defined(__NO_INLINE__)     /* -O0, -fno-inline */
+#    define XXH_NO_INLINE_HINTS 1
+#  else
+#    define XXH_NO_INLINE_HINTS 0
+#  endif
+#endif
+
+#ifndef XXH32_ENDJMP
+/* generally preferable for performance */
+#  define XXH32_ENDJMP 0
+#endif
+
+/*!
+ * @defgroup impl Implementation
+ * @{
+ */
+
+
+/* *************************************
+*  Includes & Memory related functions
+***************************************/
+/* Modify the local functions below should you wish to use some other memory routines */
+/* for ZSTD_malloc(), ZSTD_free() */
+#define ZSTD_DEPS_NEED_MALLOC
+#include "zstd_deps.h"  /* size_t, ZSTD_malloc, ZSTD_free, ZSTD_memcpy */
+static void* XXH_malloc(size_t s) { return ZSTD_malloc(s); }
+static void  XXH_free  (void* p)  { ZSTD_free(p); }
+static void* XXH_memcpy(void* dest, const void* src, size_t size) { return ZSTD_memcpy(dest,src,size); }
+
+
+/* *************************************
+*  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
+
+
+
+/* *************************************
+*  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
+#  define XXH_ASSERT(c)   ((void)0)
+#endif
+
+/* note: use after variable declarations */
+#ifndef XXH_STATIC_ASSERT
+#  if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)    /* C11 */
+#    include <assert.h>
+#    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 (e.g. 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__ __volatile__("" : "+r" (var))
+#else
+#  define XXH_COMPILER_GUARD(var) ((void)0)
+#endif
+
+/* *************************************
+*  Basic Types
+***************************************/
+#if !defined (__VMS) \
+ && (defined (__cplusplus) \
+ || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+# include <stdint.h>
+  typedef uint8_t xxh_u8;
+#else
+  typedef unsigned char xxh_u8;
+#endif
+typedef XXH32_hash_t xxh_u32;
+
+#ifdef XXH_OLD_NAMES
+#  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))
+
+/*
+ * __pack instructions are safer but compiler specific, hence potentially
+ * problematic for some compilers.
+ *
+ * Currently only defined for GCC and ICC.
+ */
+#ifdef XXH_OLD_NAMES
+typedef union { xxh_u32 u32; } __attribute__((packed)) unalign;
+#endif
+static xxh_u32 XXH_read32(const void* ptr)
+{
+    typedef union { xxh_u32 u32; } __attribute__((packed)) xxh_unalign;
+    return ((const xxh_unalign*)ptr)->u32;
+}
+
+#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
+
+/*!
+ * @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
+ * @{
+ */
+ /* #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(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 autovectorizes it incorrectly
+     * and it is pointless writing a NEON implementation that is basically the
+     * same speed as scalar for XXH32.
+     */
+    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 h32 The hash to avalanche.
+ * @return The avalanched hash.
+ */
+static xxh_u32 XXH32_avalanche(xxh_u32 h32)
+{
+    h32 ^= h32 >> 15;
+    h32 *= XXH_PRIME32_2;
+    h32 ^= h32 >> 13;
+    h32 *= XXH_PRIME32_3;
+    h32 ^= h32 >> 16;
+    return(h32);
+}
+
+#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 h32 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.
+ */
+static xxh_u32
+XXH32_finalize(xxh_u32 h32, const xxh_u8* ptr, size_t len, XXH_alignment align)
+{
+#define XXH_PROCESS1 do {                           \
+    h32 += (*ptr++) * XXH_PRIME32_5;                \
+    h32 = XXH_rotl32(h32, 11) * XXH_PRIME32_1;      \
+} while (0)
+
+#define XXH_PROCESS4 do {                           \
+    h32 += XXH_get32bits(ptr) * XXH_PRIME32_3;      \
+    ptr += 4;                                   \
+    h32  = XXH_rotl32(h32, 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(h32);
+    } else {
+         switch(len&15) /* or switch(bEnd - p) */ {
+           case 12:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;
+           case 8:       XXH_PROCESS4;
+                         XXH_FALLTHROUGH;
+           case 4:       XXH_PROCESS4;
+                         return XXH32_avalanche(h32);
+
+           case 13:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;
+           case 9:       XXH_PROCESS4;
+                         XXH_FALLTHROUGH;
+           case 5:       XXH_PROCESS4;
+                         XXH_PROCESS1;
+                         return XXH32_avalanche(h32);
+
+           case 14:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;
+           case 10:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;
+           case 6:       XXH_PROCESS4;
+                         XXH_PROCESS1;
+                         XXH_PROCESS1;
+                         return XXH32_avalanche(h32);
+
+           case 15:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;
+           case 11:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;
+           case 7:       XXH_PROCESS4;
+                         XXH_FALLTHROUGH;
+           case 3:       XXH_PROCESS1;
+                         XXH_FALLTHROUGH;
+           case 2:       XXH_PROCESS1;
+                         XXH_FALLTHROUGH;
+           case 1:       XXH_PROCESS1;
+                         XXH_FALLTHROUGH;
+           case 0:       return XXH32_avalanche(h32);
+        }
+        XXH_ASSERT(0);
+        return h32;   /* 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_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 0
+    /* 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   *******/
+/*!
+ * @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);
+}
+
+
+/*******   Canonical representation   *******/
+
+/*!
+ * @ingroup xxh32_family
+ * The default return values from XXH functions are unsigned 32 and 64 bit
+ * integers.
+ *
+ * The canonical representation uses big endian convention, the same convention
+ * as human-readable numbers (large digits first).
+ *
+ * This way, hash values can be written into a file or buffer, remaining
+ * comparable across different systems.
+ *
+ * The following functions allow transformation of hash values to and from their
+ * canonical format.
+ */
+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))
+
+/*
+ * __pack instructions are safer, but compiler specific, hence potentially
+ * problematic for some compilers.
+ *
+ * Currently only defined for GCC and ICC.
+ */
+#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 union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) xxh_unalign64;
+    return ((const xxh_unalign64*)ptr)->u64;
+}
+
+#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
+ * @{
+ */
+/* #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
+
+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;
+    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;
+}
+
+static xxh_u64 XXH64_avalanche(xxh_u64 h64)
+{
+    h64 ^= h64 >> 33;
+    h64 *= XXH_PRIME64_2;
+    h64 ^= h64 >> 29;
+    h64 *= XXH_PRIME64_3;
+    h64 ^= h64 >> 32;
+    return h64;
+}
+
+
+#define XXH_get64bits(p) XXH_readLE64_align(p, align)
+
+static xxh_u64
+XXH64_finalize(xxh_u64 h64, 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;
+        h64 ^= k1;
+        h64  = XXH_rotl64(h64,27) * XXH_PRIME64_1 + XXH_PRIME64_4;
+        len -= 8;
+    }
+    if (len >= 4) {
+        h64 ^= (xxh_u64)(XXH_get32bits(ptr)) * XXH_PRIME64_1;
+        ptr += 4;
+        h64 = XXH_rotl64(h64, 23) * XXH_PRIME64_2 + XXH_PRIME64_3;
+        len -= 4;
+    }
+    while (len > 0) {
+        h64 ^= (*ptr++) * XXH_PRIME64_5;
+        h64 = XXH_rotl64(h64, 11) * XXH_PRIME64_1;
+        --len;
+    }
+    return  XXH64_avalanche(h64);
+}
+
+#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
+
+XXH_FORCE_INLINE 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 (const void* input, size_t len, XXH64_hash_t seed)
+{
+#if 0
+    /* 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   *******/
+
+/*! @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(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(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 (XXH64_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 += 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(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);
+}
+
+
+/******* Canonical representation   *******/
+
+/*! @ingroup xxh64_family */
+XXH_PUBLIC_API void XXH64_canonicalFromHash(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(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
+#else
+/* Note: it might be useful to define __restrict or __restrict__ for some C++ compilers */
+#  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
+
+#if defined(__GNUC__) || defined(__clang__)
+#  if defined(__ARM_NEON__) || defined(__ARM_NEON) \
+   || defined(__aarch64__)  || defined(_M_ARM) \
+   || defined(_M_ARM64)     || defined(_M_ARM64EC)
+#    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.
+ * @ref 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 and all AArch64 */
+    XXH_VSX    = 5,  /*!< VSX and ZVector for POWER8/z13 (64-bit) */
+};
+/*!
+ * @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
+#endif
+
+#ifndef XXH_VECTOR    /* can be defined on command line */
+#  if ( \
+        defined(__ARM_NEON__) || defined(__ARM_NEON) /* gcc */ \
+     || defined(_M_ARM) || defined(_M_ARM64) || defined(_M_ARM64EC) /* msvc */ \
+   ) && ( \
+        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
+
+/*
+ * 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
+#  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
+#else
+#  define XXH_SEC_ALIGN 8
+#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__) && !defined(__OPTIMIZE_SIZE__) /* respect -O0 and -Os */
+#  pragma GCC push_options
+#  pragma GCC optimize("-O2")
+#endif
+
+
+#if XXH_VECTOR == XXH_NEON
+/*
+ * NEON's setup for vmlal_u32 is a little more complicated than it is on
+ * SSE2, AVX2, and VSX.
+ *
+ * While PMULUDQ and VMULEUW both perform a mask, VMLAL.U32 performs an upcast.
+ *
+ * To do the same operation, the 128-bit 'Q' register needs to be split into
+ * two 64-bit 'D' registers, performing this operation::
+ *
+ *   [                a                 |                 b                ]
+ *            |              '---------. .--------'                |
+ *            |                         x                          |
+ *            |              .---------' '--------.                |
+ *   [ a & 0xFFFFFFFF | b & 0xFFFFFFFF ],[    a >> 32     |     b >> 32    ]
+ *
+ * Due to significant changes in aarch64, the fastest method for aarch64 is
+ * completely different than the fastest method for ARMv7-A.
+ *
+ * ARMv7-A treats D registers as unions overlaying Q registers, so modifying
+ * D11 will modify the high half of Q5. This is similar to how modifying AH
+ * will only affect bits 8-15 of AX on x86.
+ *
+ * VZIP takes two registers, and puts even lanes in one register and odd lanes
+ * in the other.
+ *
+ * On ARMv7-A, this strangely modifies both parameters in place instead of
+ * taking the usual 3-operand form.
+ *
+ * Therefore, if we want to do this, we can simply use a D-form VZIP.32 on the
+ * lower and upper halves of the Q register to end up with the high and low
+ * halves where we want - all in one instruction.
+ *
+ *   vzip.32   d10, d11       @ d10 = { d10[0], d11[0] }; d11 = { d10[1], d11[1] }
+ *
+ * Unfortunately we need inline assembly for this: Instructions modifying two
+ * registers at once is not possible in GCC or Clang's IR, and they have to
+ * create a copy.
+ *
+ * aarch64 requires a different approach.
+ *
+ * In order to make it easier to write a decent compiler for aarch64, many
+ * quirks were removed, such as conditional execution.
+ *
+ * NEON was also affected by this.
+ *
+ * aarch64 cannot access the high bits of a Q-form register, and writes to a
+ * D-form register zero the high bits, similar to how writes to W-form scalar
+ * registers (or DWORD registers on x86_64) work.
+ *
+ * The formerly free vget_high intrinsics now require a vext (with a few
+ * exceptions)
+ *
+ * Additionally, VZIP was replaced by ZIP1 and ZIP2, which are the equivalent
+ * of PUNPCKL* and PUNPCKH* in SSE, respectively, in order to only modify one
+ * operand.
+ *
+ * The equivalent of the VZIP.32 on the lower and upper halves would be this
+ * mess:
+ *
+ *   ext     v2.4s, v0.4s, v0.4s, #2 // v2 = { v0[2], v0[3], v0[0], v0[1] }
+ *   zip1    v1.2s, v0.2s, v2.2s     // v1 = { v0[0], v2[0] }
+ *   zip2    v0.2s, v0.2s, v1.2s     // v0 = { v0[1], v2[1] }
+ *
+ * Instead, we use a literal downcast, vmovn_u64 (XTN), and vshrn_n_u64 (SHRN):
+ *
+ *   shrn    v1.2s, v0.2d, #32  // v1 = (uint32x2_t)(v0 >> 32);
+ *   xtn     v0.2s, v0.2d       // v0 = (uint32x2_t)(v0 & 0xFFFFFFFF);
+ *
+ * This is available on ARMv7-A, but is less efficient than a single VZIP.32.
+ */
+
+/*!
+ * Function-like macro:
+ * void XXH_SPLIT_IN_PLACE(uint64x2_t &in, uint32x2_t &outLo, uint32x2_t &outHi)
+ * {
+ *     outLo = (uint32x2_t)(in & 0xFFFFFFFF);
+ *     outHi = (uint32x2_t)(in >> 32);
+ *     in = UNDEFINED;
+ * }
+ */
+# if !defined(XXH_NO_VZIP_HACK) /* define to disable */ \
+   && (defined(__GNUC__) || defined(__clang__)) \
+   && (defined(__arm__) || defined(__thumb__) || defined(_M_ARM))
+#  define XXH_SPLIT_IN_PLACE(in, outLo, outHi)                                              \
+    do {                                                                                    \
+      /* Undocumented GCC/Clang operand modifier: %e0 = lower D half, %f0 = upper D half */ \
+      /* https://github.com/gcc-mirror/gcc/blob/38cf91e5/gcc/config/arm/arm.c#L22486 */     \
+      /* https://github.com/llvm-mirror/llvm/blob/2c4ca683/lib/Target/ARM/ARMAsmPrinter.cpp#L399 */ \
+      __asm__("vzip.32  %e0, %f0" : "+w" (in));                                             \
+      (outLo) = vget_low_u32 (vreinterpretq_u32_u64(in));                                   \
+      (outHi) = vget_high_u32(vreinterpretq_u32_u64(in));                                   \
+   } while (0)
+# else
+#  define XXH_SPLIT_IN_PLACE(in, outLo, outHi)                                            \
+    do {                                                                                  \
+      (outLo) = vmovn_u64    (in);                                                        \
+      (outHi) = vshrn_n_u64  ((in), 32);                                                  \
+    } while (0)
+# endif
+
+/*!
+ * @ingroup tuning
+ * @brief Controls the NEON to scalar ratio for XXH3
+ *
+ * On AArch64 when not optimizing for size, XXH3 will run 6 lanes using NEON and
+ * 2 lanes on scalar by default.
+ *
+ * This can be set to 2, 4, 6, or 8. ARMv7 will default to all 8 NEON lanes, as the
+ * emulated 64-bit arithmetic is too slow.
+ *
+ * Modern ARM CPUs are _very_ sensitive to how their pipelines are used.
+ *
+ * For example, the Cortex-A73 can dispatch 3 micro-ops per cycle, but it can't
+ * have more than 2 NEON (F0/F1) micro-ops. 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 A76 which
+ * can dispatch 8 micro-ops per cycle, but still only 2 NEON micro-ops at once.
+ *
+ * Therefore, @ref XXH3_NEON_LANES lanes will be processed using NEON, and the
+ * remaining lanes will use scalar instructions. This improves the bandwidth
+ * and also gives the integer pipelines something to do besides twiddling loop
+ * counters and pointers.
+ *
+ * This change benefits CPUs with large micro-op buffers without negatively affecting
+ * 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% |
+ *
+ * It also seems to fix some bad codegen on GCC, making it almost as fast as clang.
+ *
+ * @see XXH3_accumulate_512_neon()
+ */
+# ifndef XXH3_NEON_LANES
+#  if (defined(__aarch64__) || defined(__arm64__) || defined(_M_ARM64) || defined(_M_ARM64EC)) \
+   && !defined(__OPTIMIZE_SIZE__)
+#   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
+#  if defined(__s390x__)
+#    include <s390intrin.h>
+#  else
+/* gcc's altivec.h can have the unwanted consequence to unconditionally
+ * #define bool, vector, and pixel keywords,
+ * with bad consequences for programs already using these keywords for other purposes.
+ * The paragraph defining these macros is skipped when __APPLE_ALTIVEC__ is defined.
+ * __APPLE_ALTIVEC__ is _generally_ defined automatically by the compiler,
+ * but it seems that, in some cases, it isn't.
+ * Force the build macro to be defined, so that keywords are not altered.
+ */
+#    if defined(__GNUC__) && !defined(__APPLE_ALTIVEC__)
+#      define __APPLE_ALTIVEC__
+#    endif
+#    include <altivec.h>
+#  endif
+
+typedef __vector unsigned long long xxh_u64x2;
+typedef __vector unsigned char xxh_u8x16;
+typedef __vector unsigned xxh_u32x4;
+
+# 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)
+/* Clang has a better way to control this, we can just use the builtin which doesn't swap. */
+#  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 */
+
+
+/* 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 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,
+};
+
+
+#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_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 *= 0x165667919E3779F9ULL;
+    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 *= 0x9FB21C651E98DF25ULL;
+    h64 ^= (h64 >> 35) + len ;
+    h64 *= 0x9FB21C651E98DF25ULL;
+    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 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 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 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 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 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 (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);
+
+        return XXH3_avalanche(acc);
+    }
+}
+
+#define XXH3_MIDSIZE_MAX 240
+
+XXH_NO_INLINE 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;
+        int const nbRounds = (int)len / 16;
+        int i;
+        for (i=0; i<8; i++) {
+            acc += XXH3_mix16B(input+(16*i), secret+(16*i), seed);
+        }
+        acc = XXH3_avalanche(acc);
+        XXH_ASSERT(nbRounds >= 8);
+#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++) {
+            acc += XXH3_mix16B(input+(16*i), secret+(16*(i-8)) + XXH3_MIDSIZE_STARTOFFSET, seed);
+        }
+        /* last bytes */
+        acc += XXH3_mix16B(input + len - 16, secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, seed);
+        return XXH3_avalanche(acc);
+    }
+}
+
+
+/* =======     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
+
+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_shuffle_epi32 (data_key, (_MM_PERM_ENUM)_MM_SHUFFLE(0, 3, 0, 1));
+        /* 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);
+    }
+}
+
+/*
+ * 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);
+        __m512i const data_vec    = _mm512_xor_si512     (acc_vec, shifted);
+        /* xacc[0] ^= secret; */
+        __m512i const key_vec     = _mm512_loadu_si512   (secret);
+        __m512i const data_key    = _mm512_xor_si512     (data_vec, key_vec);
+
+        /* xacc[0] *= XXH_PRIME32_1; */
+        __m512i const data_key_hi = _mm512_shuffle_epi32 (data_key, (_MM_PERM_ENUM)_MM_SHUFFLE(0, 3, 0, 1));
+        __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 = _mm512_mask_set1_epi64(_mm512_set1_epi64((xxh_i64)seed64), 0xAA, (xxh_i64)(0U - seed64));
+
+        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) {
+            /* GCC has a bug, _mm512_stream_load_si512 accepts 'void*', not 'void const*',
+             * this will warn "discards 'const' qualifier". */
+            union {
+                const __m512i* cp;
+                void* p;
+            } remote_const_void;
+            remote_const_void.cp = src + i;
+            dest[i] = _mm512_add_epi64(_mm512_stream_load_si512(remote_const_void.p), 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_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
+            /* 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 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_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
+            __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_stream_load_si256(src+0), seed);
+        dest[1] = _mm256_add_epi64(_mm256_stream_load_si256(src+1), seed);
+        dest[2] = _mm256_add_epi64(_mm256_stream_load_si256(src+2), seed);
+        dest[3] = _mm256_add_epi64(_mm256_stream_load_si256(src+3), seed);
+        dest[4] = _mm256_add_epi64(_mm256_stream_load_si256(src+4), seed);
+        dest[5] = _mm256_add_epi64(_mm256_stream_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 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.
+ *
+ * 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.
+ */
+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);
+    {
+        uint64x2_t* const xacc = (uint64x2_t *) acc;
+        /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. */
+        uint8_t const* const xinput = (const uint8_t *) input;
+        uint8_t const* const xsecret  = (const uint8_t *) secret;
+
+        size_t i;
+        /* NEON for the first few lanes (these loops are normally interleaved) */
+        for (i=0; i < XXH3_NEON_LANES / 2; i++) {
+            /* data_vec = xinput[i]; */
+            uint8x16_t data_vec    = vld1q_u8(xinput  + (i * 16));
+            /* key_vec  = xsecret[i];  */
+            uint8x16_t key_vec     = vld1q_u8(xsecret + (i * 16));
+            uint64x2_t data_key;
+            uint32x2_t data_key_lo, data_key_hi;
+            /* xacc[i] += swap(data_vec); */
+            uint64x2_t const data64  = vreinterpretq_u64_u8(data_vec);
+            uint64x2_t const swapped = vextq_u64(data64, data64, 1);
+            xacc[i] = vaddq_u64 (xacc[i], swapped);
+            /* data_key = data_vec ^ key_vec; */
+            data_key = vreinterpretq_u64_u8(veorq_u8(data_vec, key_vec));
+            /* data_key_lo = (uint32x2_t) (data_key & 0xFFFFFFFF);
+             * data_key_hi = (uint32x2_t) (data_key >> 32);
+             * data_key = UNDEFINED; */
+            XXH_SPLIT_IN_PLACE(data_key, data_key_lo, data_key_hi);
+            /* xacc[i] += (uint64x2_t) data_key_lo * (uint64x2_t) data_key_hi; */
+            xacc[i] = vmlal_u32 (xacc[i], data_key_lo, data_key_hi);
+
+        }
+        /* Scalar for the remainder. This may be a zero iteration loop. */
+        for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
+            XXH3_scalarRound(acc, input, secret, i);
+        }
+    }
+}
+
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_neon(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+
+    {   uint64x2_t* xacc       = (uint64x2_t*) acc;
+        uint8_t const* xsecret = (uint8_t const*) secret;
+        uint32x2_t prime       = vdup_n_u32 (XXH_PRIME32_1);
+
+        size_t i;
+        /* NEON for the first few lanes (these loops are normally interleaved) */
+        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]; */
+            uint8x16_t key_vec  = vld1q_u8    (xsecret + (i * 16));
+            uint64x2_t data_key = veorq_u64   (data_vec, vreinterpretq_u64_u8(key_vec));
+
+            /* xacc[i] *= XXH_PRIME32_1 */
+            uint32x2_t data_key_lo, data_key_hi;
+            /* data_key_lo = (uint32x2_t) (xacc[i] & 0xFFFFFFFF);
+             * data_key_hi = (uint32x2_t) (xacc[i] >> 32);
+             * xacc[i] = UNDEFINED; */
+            XXH_SPLIT_IN_PLACE(data_key, data_key_lo, data_key_hi);
+            {   /*
+                 * prod_hi = (data_key >> 32) * XXH_PRIME32_1;
+                 *
+                 * Avoid vmul_u32 + vshll_n_u32 since Clang 6 and 7 will
+                 * incorrectly "optimize" this:
+                 *   tmp     = vmul_u32(vmovn_u64(a), vmovn_u64(b));
+                 *   shifted = vshll_n_u32(tmp, 32);
+                 * to this:
+                 *   tmp     = "vmulq_u64"(a, b); // no such thing!
+                 *   shifted = vshlq_n_u64(tmp, 32);
+                 *
+                 * However, unlike SSE, Clang lacks a 64-bit multiply routine
+                 * for NEON, and it scalarizes two 64-bit multiplies instead.
+                 *
+                 * vmull_u32 has the same timing as vmul_u32, and it avoids
+                 * this bug completely.
+                 * See https://bugs.llvm.org/show_bug.cgi?id=39967
+                 */
+                uint64x2_t prod_hi = vmull_u32 (data_key_hi, prime);
+                /* xacc[i] = prod_hi << 32; */
+                xacc[i] = vshlq_n_u64(prod_hi, 32);
+                /* xacc[i] += (prod_hi & 0xFFFFFFFF) * XXH_PRIME32_1; */
+                xacc[i] = vmlal_u32(xacc[i], data_key_lo, prime);
+            }
+        }
+        /* Scalar for the remainder. This may be a zero iteration loop. */
+        for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
+            XXH3_scalarScrambleRound(acc, secret, i);
+        }
+    }
+}
+
+#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 */
+    unsigned int* const xacc = (unsigned int*) acc;
+    xxh_u64x2 const* const xinput   = (xxh_u64x2 const*) input;   /* no alignment restriction */
+    xxh_u64x2 const* const xsecret  = (xxh_u64x2 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 + i);
+        /* key_vec = xsecret[i]; */
+        xxh_u64x2 const key_vec  = XXH_vec_loadu(xsecret + 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        = (xxh_u64x2)vec_xl(0, xacc + 4 * 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; */
+        vec_xst((xxh_u32x4)acc_vec, 0, xacc + 4 * i);
+    }
+}
+
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_vsx(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+
+    {         xxh_u64x2* const xacc    =       (xxh_u64x2*) acc;
+        const xxh_u64x2* const xsecret = (const xxh_u64x2*) 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 + 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
+
+/* scalar variants - universal */
+
+/*!
+ * @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(data_key & 0xFFFFFFFF, data_key >> 32);
+    }
+}
+
+/*!
+ * @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;
+    for (i=0; i < XXH_ACC_NB; i++) {
+        XXH3_scalarRound(acc, input, secret, i);
+    }
+}
+
+/*!
+ * @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(__clang__) && defined(__aarch64__)
+    /*
+     * UGLY HACK:
+     * Clang generates 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 Clang 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
+    /*
+     * Note: in debug mode, this overrides the asm optimization
+     * and Clang will emit MOVK chains again.
+     */
+    XXH_ASSERT(kSecretPtr == XXH3_kSecret);
+
+    {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / 16;
+        int i;
+        for (i=0; i < nbRounds; i++) {
+            /*
+             * The asm hack causes Clang 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_512)(void* XXH_RESTRICT, const void*, const void*);
+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_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_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_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_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_scrambleAcc    XXH3_scrambleAcc_vsx
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+
+#else /* scalar */
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_scalar
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_scalar
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+
+#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 */
+
+/*
+ * XXH3_accumulate()
+ * Loops over XXH3_accumulate_512().
+ * Assumption: nbStripes will not overflow the secret size
+ */
+XXH_FORCE_INLINE void
+XXH3_accumulate(     xxh_u64* XXH_RESTRICT acc,
+                const xxh_u8* XXH_RESTRICT input,
+                const xxh_u8* XXH_RESTRICT secret,
+                      size_t nbStripes,
+                      XXH3_f_accumulate_512 f_acc512)
+{
+    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);
+        f_acc512(acc,
+                 in,
+                 secret + n*XXH_SECRET_CONSUME_RATE);
+    }
+}
+
+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_512 f_acc512,
+                            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++) {
+        XXH3_accumulate(acc, input + n*block_len, secret, nbStripesPerBlock, f_acc512);
+        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));
+        XXH3_accumulate(acc, input + nb_blocks*block_len, secret, nbStripes, f_acc512);
+
+        /* 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 */
+            f_acc512(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_512 f_acc512,
+                           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_acc512, 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.
+ */
+XXH_FORCE_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_512, 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 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_512, 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_512 f_acc512,
+                                    XXH3_f_scrambleAcc f_scramble,
+                                    XXH3_f_initCustomSecret f_initSec)
+{
+    if (seed == 0)
+        return XXH3_hashLong_64b_internal(input, len,
+                                          XXH3_kSecret, sizeof(XXH3_kSecret),
+                                          f_acc512, f_scramble);
+    {   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_acc512, 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* input, size_t len,
+                           XXH64_hash_t seed, const xxh_u8* secret, size_t secretLen)
+{
+    (void)secret; (void)secretLen;
+    return XXH3_hashLong_64b_withSeed_internal(input, len, seed,
+                XXH3_accumulate_512, 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(const void* input, size_t len)
+{
+    return XXH3_64bits_internal(input, len, 0, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_default);
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSecret(const void* input, size_t len, const void* secret, size_t secretSize)
+{
+    return XXH3_64bits_internal(input, len, 0, secret, secretSize, XXH3_hashLong_64b_withSecret);
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSeed(const void* input, size_t len, XXH64_hash_t seed)
+{
+    return XXH3_64bits_internal(input, len, seed, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_withSeed);
+}
+
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSecretandSeed(const void* input, size_t len, const void* secret, size_t secretSize, XXH64_hash_t seed)
+{
+    if (len <= XXH3_MIDSIZE_MAX)
+        return XXH3_64bits_internal(input, len, seed, XXH3_kSecret, sizeof(XXH3_kSecret), NULL);
+    return XXH3_hashLong_64b_withSecret(input, len, seed, (const xxh_u8*)secret, secretSize);
+}
+
+
+/* ===   XXH3 streaming   === */
+
+/*
+ * 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 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 */
+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 */
+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(XXH3_state_t* dst_state, 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(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(XXH3_state_t* statePtr, 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(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(XXH3_state_t* statePtr, 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;
+}
+
+/* Note : when XXH3_consumeStripes() is invoked,
+ * there must be a guarantee that at least one more byte must be consumed from input
+ * so that the function can blindly consume all stripes using the "normal" secret segment */
+XXH_FORCE_INLINE void
+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_512 f_acc512,
+                    XXH3_f_scrambleAcc f_scramble)
+{
+    XXH_ASSERT(nbStripes <= nbStripesPerBlock);  /* can handle max 1 scramble per invocation */
+    XXH_ASSERT(*nbStripesSoFarPtr < nbStripesPerBlock);
+    if (nbStripesPerBlock - *nbStripesSoFarPtr <= nbStripes) {
+        /* need a scrambling operation */
+        size_t const nbStripesToEndofBlock = nbStripesPerBlock - *nbStripesSoFarPtr;
+        size_t const nbStripesAfterBlock = nbStripes - nbStripesToEndofBlock;
+        XXH3_accumulate(acc, input, secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, nbStripesToEndofBlock, f_acc512);
+        f_scramble(acc, secret + secretLimit);
+        XXH3_accumulate(acc, input + nbStripesToEndofBlock * XXH_STRIPE_LEN, secret, nbStripesAfterBlock, f_acc512);
+        *nbStripesSoFarPtr = nbStripesAfterBlock;
+    } else {
+        XXH3_accumulate(acc, input, secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, nbStripes, f_acc512);
+        *nbStripesSoFarPtr += nbStripes;
+    }
+}
+
+#ifndef XXH3_STREAM_USE_STACK
+# ifndef __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_512 f_acc512,
+            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]; 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 (state->bufferedSize + len <= XXH3_INTERNALBUFFER_SIZE) {
+            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_acc512, f_scramble);
+            state->bufferedSize = 0;
+        }
+        XXH_ASSERT(input < bEnd);
+
+        /* large input to consume : ingest per full block */
+        if ((size_t)(bEnd - input) > state->nbStripesPerBlock * XXH_STRIPE_LEN) {
+            size_t nbStripes = (size_t)(bEnd - 1 - input) / XXH_STRIPE_LEN;
+            XXH_ASSERT(state->nbStripesPerBlock >= state->nbStripesSoFar);
+            /* join to current block's end */
+            {   size_t const nbStripesToEnd = state->nbStripesPerBlock - state->nbStripesSoFar;
+                XXH_ASSERT(nbStripesToEnd <= nbStripes);
+                XXH3_accumulate(acc, input, secret + state->nbStripesSoFar * XXH_SECRET_CONSUME_RATE, nbStripesToEnd, f_acc512);
+                f_scramble(acc, secret + state->secretLimit);
+                state->nbStripesSoFar = 0;
+                input += nbStripesToEnd * XXH_STRIPE_LEN;
+                nbStripes -= nbStripesToEnd;
+            }
+            /* consume per entire blocks */
+            while(nbStripes >= state->nbStripesPerBlock) {
+                XXH3_accumulate(acc, input, secret, state->nbStripesPerBlock, f_acc512);
+                f_scramble(acc, secret + state->secretLimit);
+                input += state->nbStripesPerBlock * XXH_STRIPE_LEN;
+                nbStripes -= state->nbStripesPerBlock;
+            }
+            /* consume last partial block */
+            XXH3_accumulate(acc, input, secret, nbStripes, f_acc512);
+            input += nbStripes * XXH_STRIPE_LEN;
+            XXH_ASSERT(input < bEnd);  /* at least some bytes left */
+            state->nbStripesSoFar = nbStripes;
+            /* buffer predecessor of last partial stripe */
+            XXH_memcpy(state->buffer + sizeof(state->buffer) - XXH_STRIPE_LEN, input - XXH_STRIPE_LEN, XXH_STRIPE_LEN);
+            XXH_ASSERT(bEnd - input <= XXH_STRIPE_LEN);
+        } else {
+            /* content to consume <= block size */
+            /* Consume input by a multiple of internal buffer size */
+            if (bEnd - input > XXH3_INTERNALBUFFER_SIZE) {
+                const xxh_u8* const limit = bEnd - XXH3_INTERNALBUFFER_SIZE;
+                do {
+                    XXH3_consumeStripes(acc,
+                                       &state->nbStripesSoFar, state->nbStripesPerBlock,
+                                        input, XXH3_INTERNALBUFFER_STRIPES,
+                                        secret, state->secretLimit,
+                                        f_acc512, f_scramble);
+                    input += XXH3_INTERNALBUFFER_SIZE;
+                } while (input<limit);
+                /* buffer predecessor of last partial stripe */
+                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 */
+        memcpy(state->acc, acc, sizeof(acc));
+#endif
+    }
+
+    return XXH_OK;
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_update(XXH3_state_t* state, const void* input, size_t len)
+{
+    return XXH3_update(state, (const xxh_u8*)input, len,
+                       XXH3_accumulate_512, XXH3_scrambleAcc);
+}
+
+
+XXH_FORCE_INLINE void
+XXH3_digest_long (XXH64_hash_t* acc,
+                  const XXH3_state_t* state,
+                  const unsigned char* secret)
+{
+    /*
+     * 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) {
+        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_512, XXH3_scrambleAcc);
+        /* last stripe */
+        XXH3_accumulate_512(acc,
+                            state->buffer + state->bufferedSize - XXH_STRIPE_LEN,
+                            secret + state->secretLimit - XXH_SECRET_LASTACC_START);
+    } else {  /* bufferedSize < XXH_STRIPE_LEN */
+        xxh_u8 lastStripe[XXH_STRIPE_LEN];
+        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);
+        XXH3_accumulate_512(acc,
+                            lastStripe,
+                            secret + state->secretLimit - XXH_SECRET_LASTACC_START);
+    }
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (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);
+}
+
+
+
+/* ==========================================
+ * 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 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 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  *= 0x9FB21C651E98DF25ULL;
+        m128.low64   = XXH_xorshift64(m128.low64, 28);
+        m128.high64  = XXH3_avalanche(m128.high64);
+        return m128;
+    }
+}
+
+XXH_FORCE_INLINE 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 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 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 (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);
+        {   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 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;
+        int const nbRounds = (int)len / 32;
+        int i;
+        acc.low64 = len * XXH_PRIME64_1;
+        acc.high64 = 0;
+        for (i=0; i<4; i++) {
+            acc = XXH128_mix32B(acc,
+                                input  + (32 * i),
+                                input  + (32 * i) + 16,
+                                secret + (32 * i),
+                                seed);
+        }
+        acc.low64 = XXH3_avalanche(acc.low64);
+        acc.high64 = XXH3_avalanche(acc.high64);
+        XXH_ASSERT(nbRounds >= 4);
+        for (i=4 ; i < nbRounds; i++) {
+            acc = XXH128_mix32B(acc,
+                                input + (32 * i),
+                                input + (32 * i) + 16,
+                                secret + XXH3_MIDSIZE_STARTOFFSET + (32 * (i - 4)),
+                                seed);
+        }
+        /* last bytes */
+        acc = XXH128_mix32B(acc,
+                            input + len - 16,
+                            input + len - 32,
+                            secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET - 16,
+                            0ULL - 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_512 f_acc512,
+                            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_acc512, 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 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_512, XXH3_scrambleAcc);
+}
+
+/*
+ * It's important for performance to pass @secretLen (when it's static)
+ * to the compiler, so that it can properly optimize the vectorized loop.
+ */
+XXH_FORCE_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_512, 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_512 f_acc512,
+                                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_acc512, 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_acc512, 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_512, 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(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(const void* input, size_t len, 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(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(const void* input, size_t len, 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(const void* input, size_t len, XXH64_hash_t seed)
+{
+    return XXH3_128bits_withSeed(input, len, seed);
+}
+
+
+/* ===   XXH3 128-bit streaming   === */
+
+/*
+ * 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(XXH3_state_t* statePtr)
+{
+    return XXH3_64bits_reset(statePtr);
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSecret(XXH3_state_t* statePtr, 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(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(XXH3_state_t* statePtr, 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(XXH3_state_t* state, const void* input, size_t len)
+{
+    return XXH3_update(state, (const xxh_u8*)input, len,
+                       XXH3_accumulate_512, XXH3_scrambleAcc);
+}
+
+/*! @ingroup xxh3_family */
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (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);
+}
+
+/* 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(const void* h128_1, 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(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(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(void* secretBuffer, size_t secretSize, 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(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__) && !defined(__OPTIMIZE_SIZE__) /* respect -O0 and -Os */
+#  pragma GCC pop_options
+#endif
+
+#endif  /* XXH_NO_LONG_LONG */
+
+#endif  /* XXH_NO_XXH3 */
+
+/*!
+ * @}
+ */
+#endif  /* XXH_IMPLEMENTATION */
 
 
 #if defined (__cplusplus)

src/borg/algorithms/zstd/lib/common/zstd_common.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -13,8 +13,7 @@
 /*-*************************************
 *  Dependencies
 ***************************************/
-#include <stdlib.h>      /* malloc, calloc, free */
-#include <string.h>      /* memset */
+#define ZSTD_DEPS_NEED_MALLOC
 #include "error_private.h"
 #include "zstd_internal.h"
 
@@ -47,37 +46,3 @@ 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); }
-
-
-
-/*=**************************************************************
-*  Custom allocator
-****************************************************************/
-void* ZSTD_malloc(size_t size, ZSTD_customMem customMem)
-{
-    if (customMem.customAlloc)
-        return customMem.customAlloc(customMem.opaque, size);
-    return malloc(size);
-}
-
-void* ZSTD_calloc(size_t size, ZSTD_customMem customMem)
-{
-    if (customMem.customAlloc) {
-        /* calloc implemented as malloc+memset;
-         * not as efficient as calloc, but next best guess for custom malloc */
-        void* const ptr = customMem.customAlloc(customMem.opaque, size);
-        memset(ptr, 0, size);
-        return ptr;
-    }
-    return calloc(1, size);
-}
-
-void ZSTD_free(void* ptr, ZSTD_customMem customMem)
-{
-    if (ptr!=NULL) {
-        if (customMem.customFree)
-            customMem.customFree(customMem.opaque, ptr);
-        else
-            free(ptr);
-    }
-}

src/borg/algorithms/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 */

src/borg/algorithms/zstd/lib/common/zstd_internal.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -19,10 +19,8 @@
 /*-*************************************
 *  Dependencies
 ***************************************/
-#ifdef __aarch64__
-#include <arm_neon.h>
-#endif
 #include "compiler.h"
+#include "cpu.h"
 #include "mem.h"
 #include "debug.h"                 /* assert, DEBUGLOG, RAWLOG, g_debuglevel */
 #include "error_private.h"
@@ -30,12 +28,16 @@
 #include "../zstd.h"
 #define FSE_STATIC_LINKING_ONLY
 #include "fse.h"
-#define HUF_STATIC_LINKING_ONLY
 #include "huf.h"
 #ifndef XXH_STATIC_LINKING_ONLY
 #  define XXH_STATIC_LINKING_ONLY  /* XXH64_state_t */
 #endif
 #include "xxhash.h"                /* XXH_reset, update, digest */
+#ifndef ZSTD_NO_TRACE
+#  include "zstd_trace.h"
+#else
+#  define ZSTD_TRACE 0
+#endif
 
 #if defined (__cplusplus)
 extern "C" {
@@ -55,81 +57,7 @@ extern "C" {
 #undef MAX
 #define MIN(a,b) ((a)<(b) ? (a) : (b))
 #define MAX(a,b) ((a)>(b) ? (a) : (b))
-
-/**
- * Ignore: this is an internal helper.
- *
- * This is a helper function to help force C99-correctness during compilation.
- * Under strict compilation modes, variadic macro arguments can't be empty.
- * However, variadic function arguments can be. Using a function therefore lets
- * us statically check that at least one (string) argument was passed,
- * independent of the compilation flags.
- */
-static INLINE_KEYWORD UNUSED_ATTR
-void _force_has_format_string(const char *format, ...) {
-  (void)format;
-}
-
-/**
- * Ignore: this is an internal helper.
- *
- * We want to force this function invocation to be syntactically correct, but
- * we don't want to force runtime evaluation of its arguments.
- */
-#define _FORCE_HAS_FORMAT_STRING(...) \
-  if (0) { \
-    _force_has_format_string(__VA_ARGS__); \
-  }
-
-/**
- * Return the specified error if the condition evaluates to true.
- *
- * In debug modes, prints additional information.
- * In order to do that (particularly, printing the conditional that failed),
- * this can't just wrap RETURN_ERROR().
- */
-#define RETURN_ERROR_IF(cond, err, ...) \
-  if (cond) { \
-    RAWLOG(3, "%s:%d: ERROR!: check %s failed, returning %s", \
-           __FILE__, __LINE__, ZSTD_QUOTE(cond), ZSTD_QUOTE(ERROR(err))); \
-    _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
-    RAWLOG(3, ": " __VA_ARGS__); \
-    RAWLOG(3, "\n"); \
-    return ERROR(err); \
-  }
-
-/**
- * Unconditionally return the specified error.
- *
- * In debug modes, prints additional information.
- */
-#define RETURN_ERROR(err, ...) \
-  do { \
-    RAWLOG(3, "%s:%d: ERROR!: unconditional check failed, returning %s", \
-           __FILE__, __LINE__, ZSTD_QUOTE(ERROR(err))); \
-    _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
-    RAWLOG(3, ": " __VA_ARGS__); \
-    RAWLOG(3, "\n"); \
-    return ERROR(err); \
-  } while(0);
-
-/**
- * If the provided expression evaluates to an error code, returns that error code.
- *
- * In debug modes, prints additional information.
- */
-#define FORWARD_IF_ERROR(err, ...) \
-  do { \
-    size_t const err_code = (err); \
-    if (ERR_isError(err_code)) { \
-      RAWLOG(3, "%s:%d: ERROR!: forwarding error in %s: %s", \
-             __FILE__, __LINE__, ZSTD_QUOTE(err), ERR_getErrorName(err_code)); \
-      _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
-      RAWLOG(3, ": " __VA_ARGS__); \
-      RAWLOG(3, "\n"); \
-      return err_code; \
-    } \
-  } while(0);
+#define BOUNDED(min,val,max) (MAX(min,MIN(val,max)))
 
 
 /*-*************************************
@@ -138,8 +66,7 @@ void _force_has_format_string(const char *format, ...) {
 #define ZSTD_OPT_NUM    (1<<12)
 
 #define ZSTD_REP_NUM      3                 /* number of repcodes */
-#define ZSTD_REP_MOVE     (ZSTD_REP_NUM-1)
-static const U32 repStartValue[ZSTD_REP_NUM] = { 1, 4, 8 };
+static UNUSED_ATTR const U32 repStartValue[ZSTD_REP_NUM] = { 1, 4, 8 };
 
 #define KB *(1 <<10)
 #define MB *(1 <<20)
@@ -153,21 +80,21 @@ static const U32 repStartValue[ZSTD_REP_NUM] = { 1, 4, 8 };
 #define BIT0   1
 
 #define ZSTD_WINDOWLOG_ABSOLUTEMIN 10
-static const size_t ZSTD_fcs_fieldSize[4] = { 0, 2, 4, 8 };
-static const size_t ZSTD_did_fieldSize[4] = { 0, 1, 2, 4 };
+static UNUSED_ATTR const size_t ZSTD_fcs_fieldSize[4] = { 0, 2, 4, 8 };
+static UNUSED_ATTR const size_t ZSTD_did_fieldSize[4] = { 0, 1, 2, 4 };
 
 #define ZSTD_FRAMEIDSIZE 4   /* magic number size */
 
 #define ZSTD_BLOCKHEADERSIZE 3   /* C standard doesn't allow `static const` variable to be init using another `static const` variable */
-static const size_t ZSTD_blockHeaderSize = ZSTD_BLOCKHEADERSIZE;
+static UNUSED_ATTR const size_t ZSTD_blockHeaderSize = ZSTD_BLOCKHEADERSIZE;
 typedef enum { bt_raw, bt_rle, bt_compressed, bt_reserved } blockType_e;
 
 #define ZSTD_FRAMECHECKSUMSIZE 4
 
 #define MIN_SEQUENCES_SIZE 1 /* nbSeq==0 */
-#define MIN_CBLOCK_SIZE (1 /*litCSize*/ + 1 /* RLE or RAW */ + MIN_SEQUENCES_SIZE /* nbSeq==0 */)   /* for a non-null block */
+#define MIN_CBLOCK_SIZE (1 /*litCSize*/ + 1 /* RLE or RAW */)   /* for a non-null block */
+#define MIN_LITERALS_FOR_4_STREAMS 6
 
-#define HufLog 12
 typedef enum { set_basic, set_rle, set_compressed, set_repeat } symbolEncodingType_e;
 
 #define LONGNBSEQ 0x7F00
@@ -175,6 +102,7 @@ typedef enum { set_basic, set_rle, set_compressed, set_repeat } symbolEncodingTy
 #define MINMATCH 3
 
 #define Litbits  8
+#define LitHufLog 11
 #define MaxLit ((1<<Litbits) - 1)
 #define MaxML   52
 #define MaxLL   35
@@ -185,62 +113,89 @@ typedef enum { set_basic, set_rle, set_compressed, set_repeat } symbolEncodingTy
 #define LLFSELog    9
 #define OffFSELog   8
 #define MaxFSELog  MAX(MAX(MLFSELog, LLFSELog), OffFSELog)
-
-static const U32 LL_bits[MaxLL+1] = { 0, 0, 0, 0, 0, 0, 0, 0,
-                                      0, 0, 0, 0, 0, 0, 0, 0,
-                                      1, 1, 1, 1, 2, 2, 3, 3,
-                                      4, 6, 7, 8, 9,10,11,12,
-                                     13,14,15,16 };
-static const S16 LL_defaultNorm[MaxLL+1] = { 4, 3, 2, 2, 2, 2, 2, 2,
-                                             2, 2, 2, 2, 2, 1, 1, 1,
-                                             2, 2, 2, 2, 2, 2, 2, 2,
-                                             2, 3, 2, 1, 1, 1, 1, 1,
-                                            -1,-1,-1,-1 };
+#define MaxMLBits 16
+#define MaxLLBits 16
+
+#define ZSTD_MAX_HUF_HEADER_SIZE 128 /* header + <= 127 byte tree description */
+/* Each table cannot take more than #symbols * FSELog bits */
+#define ZSTD_MAX_FSE_HEADERS_SIZE (((MaxML + 1) * MLFSELog + (MaxLL + 1) * LLFSELog + (MaxOff + 1) * OffFSELog + 7) / 8)
+
+static UNUSED_ATTR const U8 LL_bits[MaxLL+1] = {
+     0, 0, 0, 0, 0, 0, 0, 0,
+     0, 0, 0, 0, 0, 0, 0, 0,
+     1, 1, 1, 1, 2, 2, 3, 3,
+     4, 6, 7, 8, 9,10,11,12,
+    13,14,15,16
+};
+static UNUSED_ATTR const S16 LL_defaultNorm[MaxLL+1] = {
+     4, 3, 2, 2, 2, 2, 2, 2,
+     2, 2, 2, 2, 2, 1, 1, 1,
+     2, 2, 2, 2, 2, 2, 2, 2,
+     2, 3, 2, 1, 1, 1, 1, 1,
+    -1,-1,-1,-1
+};
 #define LL_DEFAULTNORMLOG 6  /* for static allocation */
-static const U32 LL_defaultNormLog = LL_DEFAULTNORMLOG;
-
-static const U32 ML_bits[MaxML+1] = { 0, 0, 0, 0, 0, 0, 0, 0,
-                                      0, 0, 0, 0, 0, 0, 0, 0,
-                                      0, 0, 0, 0, 0, 0, 0, 0,
-                                      0, 0, 0, 0, 0, 0, 0, 0,
-                                      1, 1, 1, 1, 2, 2, 3, 3,
-                                      4, 4, 5, 7, 8, 9,10,11,
-                                     12,13,14,15,16 };
-static const S16 ML_defaultNorm[MaxML+1] = { 1, 4, 3, 2, 2, 2, 2, 2,
-                                             2, 1, 1, 1, 1, 1, 1, 1,
-                                             1, 1, 1, 1, 1, 1, 1, 1,
-                                             1, 1, 1, 1, 1, 1, 1, 1,
-                                             1, 1, 1, 1, 1, 1, 1, 1,
-                                             1, 1, 1, 1, 1, 1,-1,-1,
-                                            -1,-1,-1,-1,-1 };
+static UNUSED_ATTR const U32 LL_defaultNormLog = LL_DEFAULTNORMLOG;
+
+static UNUSED_ATTR const U8 ML_bits[MaxML+1] = {
+     0, 0, 0, 0, 0, 0, 0, 0,
+     0, 0, 0, 0, 0, 0, 0, 0,
+     0, 0, 0, 0, 0, 0, 0, 0,
+     0, 0, 0, 0, 0, 0, 0, 0,
+     1, 1, 1, 1, 2, 2, 3, 3,
+     4, 4, 5, 7, 8, 9,10,11,
+    12,13,14,15,16
+};
+static UNUSED_ATTR const S16 ML_defaultNorm[MaxML+1] = {
+     1, 4, 3, 2, 2, 2, 2, 2,
+     2, 1, 1, 1, 1, 1, 1, 1,
+     1, 1, 1, 1, 1, 1, 1, 1,
+     1, 1, 1, 1, 1, 1, 1, 1,
+     1, 1, 1, 1, 1, 1, 1, 1,
+     1, 1, 1, 1, 1, 1,-1,-1,
+    -1,-1,-1,-1,-1
+};
 #define ML_DEFAULTNORMLOG 6  /* for static allocation */
-static const U32 ML_defaultNormLog = ML_DEFAULTNORMLOG;
-
-static const S16 OF_defaultNorm[DefaultMaxOff+1] = { 1, 1, 1, 1, 1, 1, 2, 2,
-                                                     2, 1, 1, 1, 1, 1, 1, 1,
-                                                     1, 1, 1, 1, 1, 1, 1, 1,
-                                                    -1,-1,-1,-1,-1 };
+static UNUSED_ATTR const U32 ML_defaultNormLog = ML_DEFAULTNORMLOG;
+
+static UNUSED_ATTR const S16 OF_defaultNorm[DefaultMaxOff+1] = {
+     1, 1, 1, 1, 1, 1, 2, 2,
+     2, 1, 1, 1, 1, 1, 1, 1,
+     1, 1, 1, 1, 1, 1, 1, 1,
+    -1,-1,-1,-1,-1
+};
 #define OF_DEFAULTNORMLOG 5  /* for static allocation */
-static const U32 OF_defaultNormLog = OF_DEFAULTNORMLOG;
+static UNUSED_ATTR const U32 OF_defaultNormLog = OF_DEFAULTNORMLOG;
 
 
 /*-*******************************************
 *  Shared functions to include for inlining
 *********************************************/
 static void ZSTD_copy8(void* dst, const void* src) {
-#ifdef __aarch64__
+#if defined(ZSTD_ARCH_ARM_NEON)
     vst1_u8((uint8_t*)dst, vld1_u8((const uint8_t*)src));
 #else
-    memcpy(dst, src, 8);
+    ZSTD_memcpy(dst, src, 8);
 #endif
 }
-
 #define COPY8(d,s) { ZSTD_copy8(d,s); d+=8; s+=8; }
+
+/* Need to use memmove here since the literal buffer can now be located within
+   the dst buffer. In circumstances where the op "catches up" to where the
+   literal buffer is, there can be partial overlaps in this call on the final
+   copy if the literal is being shifted by less than 16 bytes. */
 static void ZSTD_copy16(void* dst, const void* src) {
-#ifdef __aarch64__
+#if defined(ZSTD_ARCH_ARM_NEON)
     vst1q_u8((uint8_t*)dst, vld1q_u8((const uint8_t*)src));
+#elif defined(ZSTD_ARCH_X86_SSE2)
+    _mm_storeu_si128((__m128i*)dst, _mm_loadu_si128((const __m128i*)src));
+#elif defined(__clang__)
+    ZSTD_memmove(dst, src, 16);
 #else
-    memcpy(dst, src, 16);
+    /* ZSTD_memmove is not inlined properly by gcc */
+    BYTE copy16_buf[16];
+    ZSTD_memcpy(copy16_buf, src, 16);
+    ZSTD_memcpy(dst, copy16_buf, 16);
 #endif
 }
 #define COPY16(d,s) { ZSTD_copy16(d,s); d+=16; s+=16; }
@@ -255,13 +210,13 @@ typedef enum {
 } ZSTD_overlap_e;
 
 /*! ZSTD_wildcopy() :
- *  Custom version of memcpy(), can over read/write up to WILDCOPY_OVERLENGTH bytes (if length==0)
+ *  Custom version of ZSTD_memcpy(), can over read/write up to WILDCOPY_OVERLENGTH bytes (if length==0)
  *  @param ovtype controls the overlap detection
  *         - ZSTD_no_overlap: The source and destination are guaranteed to be at least WILDCOPY_VECLEN bytes apart.
  *         - ZSTD_overlap_src_before_dst: The src and dst may overlap, but they MUST be at least 8 bytes apart.
  *           The src buffer must be before the dst buffer.
  */
-MEM_STATIC FORCE_INLINE_ATTR 
+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;
@@ -269,8 +224,6 @@ void ZSTD_wildcopy(void* dst, const void* src, ptrdiff_t length, ZSTD_overlap_e
     BYTE* op = (BYTE*)dst;
     BYTE* const oend = op + length;
 
-    assert(diff >= 8 || (ovtype == ZSTD_no_overlap && diff <= -WILDCOPY_VECLEN));
-
     if (ovtype == ZSTD_overlap_src_before_dst && diff < WILDCOPY_VECLEN) {
         /* Handle short offset copies. */
         do {
@@ -284,20 +237,15 @@ void ZSTD_wildcopy(void* dst, const void* src, ptrdiff_t length, ZSTD_overlap_e
          * one COPY16() in the first call. Then, do two calls per loop since
          * at that point it is more likely to have a high trip count.
          */
-#ifndef __aarch64__
-        do {
-            COPY16(op, ip);
-        }
-        while (op < oend);
-#else
-        COPY16(op, ip);
-        if (op >= oend) return;
+        ZSTD_copy16(op, ip);
+        if (16 >= length) return;
+        op += 16;
+        ip += 16;
         do {
             COPY16(op, ip);
             COPY16(op, ip);
         }
         while (op < oend);
-#endif
     }
 }
 
@@ -305,7 +253,7 @@ MEM_STATIC size_t ZSTD_limitCopy(void* dst, size_t dstCapacity, const void* src,
 {
     size_t const length = MIN(dstCapacity, srcSize);
     if (length > 0) {
-        memcpy(dst, src, length);
+        ZSTD_memcpy(dst, src, length);
     }
     return length;
 }
@@ -320,28 +268,46 @@ MEM_STATIC size_t ZSTD_limitCopy(void* dst, size_t dstCapacity, const void* src,
  * In which case, resize it down to free some memory */
 #define ZSTD_WORKSPACETOOLARGE_MAXDURATION 128
 
+/* Controls whether the input/output buffer is buffered or stable. */
+typedef enum {
+    ZSTD_bm_buffered = 0,  /* Buffer the input/output */
+    ZSTD_bm_stable = 1     /* ZSTD_inBuffer/ZSTD_outBuffer is stable */
+} ZSTD_bufferMode_e;
+
 
 /*-*******************************************
 *  Private declarations
 *********************************************/
 typedef struct seqDef_s {
-    U32 offset;
+    U32 offBase;   /* offBase == Offset + ZSTD_REP_NUM, or repcode 1,2,3 */
     U16 litLength;
-    U16 matchLength;
+    U16 mlBase;    /* mlBase == matchLength - MINMATCH */
 } seqDef;
 
+/* Controls whether seqStore has a single "long" litLength or matchLength. See seqStore_t. */
+typedef enum {
+    ZSTD_llt_none = 0,             /* no longLengthType */
+    ZSTD_llt_literalLength = 1,    /* represents a long literal */
+    ZSTD_llt_matchLength = 2       /* represents a long match */
+} ZSTD_longLengthType_e;
+
 typedef struct {
     seqDef* sequencesStart;
-    seqDef* sequences;
-    BYTE* litStart;
-    BYTE* lit;
-    BYTE* llCode;
-    BYTE* mlCode;
-    BYTE* ofCode;
+    seqDef* sequences;      /* ptr to end of sequences */
+    BYTE*  litStart;
+    BYTE*  lit;             /* ptr to end of literals */
+    BYTE*  llCode;
+    BYTE*  mlCode;
+    BYTE*  ofCode;
     size_t maxNbSeq;
     size_t maxNbLit;
-    U32   longLengthID;   /* 0 == no longLength; 1 == Lit.longLength; 2 == Match.longLength; */
-    U32   longLengthPos;
+
+    /* longLengthPos and longLengthType to allow us to represent either a single litLength or matchLength
+     * in the seqStore that has a value larger than U16 (if it exists). To do so, we increment
+     * the existing value of the litLength or matchLength by 0x10000.
+     */
+    ZSTD_longLengthType_e longLengthType;
+    U32                   longLengthPos;  /* Index of the sequence to apply long length modification to */
 } seqStore_t;
 
 typedef struct {
@@ -351,19 +317,19 @@ typedef struct {
 
 /**
  * Returns the ZSTD_sequenceLength for the given sequences. It handles the decoding of long sequences
- * indicated by longLengthPos and longLengthID, and adds MINMATCH back to matchLength.
+ * indicated by longLengthPos and longLengthType, and adds MINMATCH back to matchLength.
  */
 MEM_STATIC ZSTD_sequenceLength ZSTD_getSequenceLength(seqStore_t const* seqStore, seqDef const* seq)
 {
     ZSTD_sequenceLength seqLen;
     seqLen.litLength = seq->litLength;
-    seqLen.matchLength = seq->matchLength + MINMATCH;
+    seqLen.matchLength = seq->mlBase + MINMATCH;
     if (seqStore->longLengthPos == (U32)(seq - seqStore->sequencesStart)) {
-        if (seqStore->longLengthID == 1) {
-            seqLen.litLength += 0xFFFF;
+        if (seqStore->longLengthType == ZSTD_llt_literalLength) {
+            seqLen.litLength += 0x10000;
         }
-        if (seqStore->longLengthID == 2) {
-            seqLen.matchLength += 0xFFFF;
+        if (seqStore->longLengthType == ZSTD_llt_matchLength) {
+            seqLen.matchLength += 0x10000;
         }
     }
     return seqLen;
@@ -376,42 +342,13 @@ MEM_STATIC ZSTD_sequenceLength ZSTD_getSequenceLength(seqStore_t const* seqStore
  *          `decompressedBound != ZSTD_CONTENTSIZE_ERROR`
  */
 typedef struct {
+    size_t nbBlocks;
     size_t compressedSize;
     unsigned long long decompressedBound;
 } ZSTD_frameSizeInfo;   /* decompress & legacy */
 
 const seqStore_t* ZSTD_getSeqStore(const ZSTD_CCtx* ctx);   /* compress & dictBuilder */
-void ZSTD_seqToCodes(const seqStore_t* seqStorePtr);   /* compress, dictBuilder, decodeCorpus (shouldn't get its definition from here) */
-
-/* custom memory allocation functions */
-void* ZSTD_malloc(size_t size, ZSTD_customMem customMem);
-void* ZSTD_calloc(size_t size, ZSTD_customMem customMem);
-void ZSTD_free(void* ptr, ZSTD_customMem customMem);
-
-
-MEM_STATIC U32 ZSTD_highbit32(U32 val)   /* compress, dictBuilder, decodeCorpus */
-{
-    assert(val != 0);
-    {
-#   if defined(_MSC_VER)   /* Visual */
-        unsigned long r=0;
-        return _BitScanReverse(&r, val) ? (unsigned)r : 0;
-#   elif defined(__GNUC__) && (__GNUC__ >= 3)   /* GCC Intrinsic */
-        return __builtin_clz (val) ^ 31;
-#   elif defined(__ICCARM__)    /* IAR Intrinsic */
-        return 31 - __CLZ(val);
-#   else   /* Software version */
-        static const U32 DeBruijnClz[32] = { 0, 9, 1, 10, 13, 21, 2, 29, 11, 14, 16, 18, 22, 25, 3, 30, 8, 12, 20, 28, 15, 17, 24, 7, 19, 27, 23, 6, 26, 5, 4, 31 };
-        U32 v = val;
-        v |= v >> 1;
-        v |= v >> 2;
-        v |= v >> 4;
-        v |= v >> 8;
-        v |= v >> 16;
-        return DeBruijnClz[(v * 0x07C4ACDDU) >> 27];
-#   endif
-    }
-}
+int ZSTD_seqToCodes(const seqStore_t* seqStorePtr);   /* compress, dictBuilder, decodeCorpus (shouldn't get its definition from here) */
 
 
 /* ZSTD_invalidateRepCodes() :
@@ -439,6 +376,14 @@ size_t ZSTD_getcBlockSize(const void* src, size_t srcSize,
 size_t ZSTD_decodeSeqHeaders(ZSTD_DCtx* dctx, int* nbSeqPtr,
                        const void* src, size_t srcSize);
 
+/**
+ * @returns true iff the CPU supports dynamic BMI2 dispatch.
+ */
+MEM_STATIC int ZSTD_cpuSupportsBmi2(void)
+{
+    ZSTD_cpuid_t cpuid = ZSTD_cpuid();
+    return ZSTD_cpuid_bmi1(cpuid) && ZSTD_cpuid_bmi2(cpuid);
+}
 
 #if defined (__cplusplus)
 }

src/borg/algorithms/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 */

src/borg/algorithms/zstd/lib/compress/clevels.h

@@ -0,0 +1,134 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_CLEVELS_H
+#define ZSTD_CLEVELS_H
+
+#define ZSTD_STATIC_LINKING_ONLY  /* ZSTD_compressionParameters  */
+#include "../zstd.h"
+
+/*-=====  Pre-defined compression levels  =====-*/
+
+#define ZSTD_MAX_CLEVEL     22
+
+#ifdef __GNUC__
+__attribute__((__unused__))
+#endif
+
+static const ZSTD_compressionParameters ZSTD_defaultCParameters[4][ZSTD_MAX_CLEVEL+1] = {
+{   /* "default" - for any srcSize > 256 KB */
+    /* W,  C,  H,  S,  L, TL, strat */
+    { 19, 12, 13,  1,  6,  1, ZSTD_fast    },  /* base for negative levels */
+    { 19, 13, 14,  1,  7,  0, ZSTD_fast    },  /* level  1 */
+    { 20, 15, 16,  1,  6,  0, ZSTD_fast    },  /* level  2 */
+    { 21, 16, 17,  1,  5,  0, ZSTD_dfast   },  /* level  3 */
+    { 21, 18, 18,  1,  5,  0, ZSTD_dfast   },  /* level  4 */
+    { 21, 18, 19,  3,  5,  2, ZSTD_greedy  },  /* level  5 */
+    { 21, 18, 19,  3,  5,  4, ZSTD_lazy    },  /* level  6 */
+    { 21, 19, 20,  4,  5,  8, ZSTD_lazy    },  /* level  7 */
+    { 21, 19, 20,  4,  5, 16, ZSTD_lazy2   },  /* level  8 */
+    { 22, 20, 21,  4,  5, 16, ZSTD_lazy2   },  /* level  9 */
+    { 22, 21, 22,  5,  5, 16, ZSTD_lazy2   },  /* level 10 */
+    { 22, 21, 22,  6,  5, 16, ZSTD_lazy2   },  /* level 11 */
+    { 22, 22, 23,  6,  5, 32, ZSTD_lazy2   },  /* level 12 */
+    { 22, 22, 22,  4,  5, 32, ZSTD_btlazy2 },  /* level 13 */
+    { 22, 22, 23,  5,  5, 32, ZSTD_btlazy2 },  /* level 14 */
+    { 22, 23, 23,  6,  5, 32, ZSTD_btlazy2 },  /* level 15 */
+    { 22, 22, 22,  5,  5, 48, ZSTD_btopt   },  /* level 16 */
+    { 23, 23, 22,  5,  4, 64, ZSTD_btopt   },  /* level 17 */
+    { 23, 23, 22,  6,  3, 64, ZSTD_btultra },  /* level 18 */
+    { 23, 24, 22,  7,  3,256, ZSTD_btultra2},  /* level 19 */
+    { 25, 25, 23,  7,  3,256, ZSTD_btultra2},  /* level 20 */
+    { 26, 26, 24,  7,  3,512, ZSTD_btultra2},  /* level 21 */
+    { 27, 27, 25,  9,  3,999, ZSTD_btultra2},  /* level 22 */
+},
+{   /* for srcSize <= 256 KB */
+    /* W,  C,  H,  S,  L,  T, strat */
+    { 18, 12, 13,  1,  5,  1, ZSTD_fast    },  /* base for negative levels */
+    { 18, 13, 14,  1,  6,  0, ZSTD_fast    },  /* level  1 */
+    { 18, 14, 14,  1,  5,  0, ZSTD_dfast   },  /* level  2 */
+    { 18, 16, 16,  1,  4,  0, ZSTD_dfast   },  /* level  3 */
+    { 18, 16, 17,  3,  5,  2, ZSTD_greedy  },  /* level  4.*/
+    { 18, 17, 18,  5,  5,  2, ZSTD_greedy  },  /* level  5.*/
+    { 18, 18, 19,  3,  5,  4, ZSTD_lazy    },  /* level  6.*/
+    { 18, 18, 19,  4,  4,  4, ZSTD_lazy    },  /* level  7 */
+    { 18, 18, 19,  4,  4,  8, ZSTD_lazy2   },  /* level  8 */
+    { 18, 18, 19,  5,  4,  8, ZSTD_lazy2   },  /* level  9 */
+    { 18, 18, 19,  6,  4,  8, ZSTD_lazy2   },  /* level 10 */
+    { 18, 18, 19,  5,  4, 12, ZSTD_btlazy2 },  /* level 11.*/
+    { 18, 19, 19,  7,  4, 12, ZSTD_btlazy2 },  /* level 12.*/
+    { 18, 18, 19,  4,  4, 16, ZSTD_btopt   },  /* level 13 */
+    { 18, 18, 19,  4,  3, 32, ZSTD_btopt   },  /* level 14.*/
+    { 18, 18, 19,  6,  3,128, ZSTD_btopt   },  /* level 15.*/
+    { 18, 19, 19,  6,  3,128, ZSTD_btultra },  /* level 16.*/
+    { 18, 19, 19,  8,  3,256, ZSTD_btultra },  /* level 17.*/
+    { 18, 19, 19,  6,  3,128, ZSTD_btultra2},  /* level 18.*/
+    { 18, 19, 19,  8,  3,256, ZSTD_btultra2},  /* level 19.*/
+    { 18, 19, 19, 10,  3,512, ZSTD_btultra2},  /* level 20.*/
+    { 18, 19, 19, 12,  3,512, ZSTD_btultra2},  /* level 21.*/
+    { 18, 19, 19, 13,  3,999, ZSTD_btultra2},  /* level 22.*/
+},
+{   /* for srcSize <= 128 KB */
+    /* W,  C,  H,  S,  L,  T, strat */
+    { 17, 12, 12,  1,  5,  1, ZSTD_fast    },  /* base for negative levels */
+    { 17, 12, 13,  1,  6,  0, ZSTD_fast    },  /* level  1 */
+    { 17, 13, 15,  1,  5,  0, ZSTD_fast    },  /* level  2 */
+    { 17, 15, 16,  2,  5,  0, ZSTD_dfast   },  /* level  3 */
+    { 17, 17, 17,  2,  4,  0, ZSTD_dfast   },  /* level  4 */
+    { 17, 16, 17,  3,  4,  2, ZSTD_greedy  },  /* level  5 */
+    { 17, 16, 17,  3,  4,  4, ZSTD_lazy    },  /* level  6 */
+    { 17, 16, 17,  3,  4,  8, ZSTD_lazy2   },  /* level  7 */
+    { 17, 16, 17,  4,  4,  8, ZSTD_lazy2   },  /* level  8 */
+    { 17, 16, 17,  5,  4,  8, ZSTD_lazy2   },  /* level  9 */
+    { 17, 16, 17,  6,  4,  8, ZSTD_lazy2   },  /* level 10 */
+    { 17, 17, 17,  5,  4,  8, ZSTD_btlazy2 },  /* level 11 */
+    { 17, 18, 17,  7,  4, 12, ZSTD_btlazy2 },  /* level 12 */
+    { 17, 18, 17,  3,  4, 12, ZSTD_btopt   },  /* level 13.*/
+    { 17, 18, 17,  4,  3, 32, ZSTD_btopt   },  /* level 14.*/
+    { 17, 18, 17,  6,  3,256, ZSTD_btopt   },  /* level 15.*/
+    { 17, 18, 17,  6,  3,128, ZSTD_btultra },  /* level 16.*/
+    { 17, 18, 17,  8,  3,256, ZSTD_btultra },  /* level 17.*/
+    { 17, 18, 17, 10,  3,512, ZSTD_btultra },  /* level 18.*/
+    { 17, 18, 17,  5,  3,256, ZSTD_btultra2},  /* level 19.*/
+    { 17, 18, 17,  7,  3,512, ZSTD_btultra2},  /* level 20.*/
+    { 17, 18, 17,  9,  3,512, ZSTD_btultra2},  /* level 21.*/
+    { 17, 18, 17, 11,  3,999, ZSTD_btultra2},  /* level 22.*/
+},
+{   /* for srcSize <= 16 KB */
+    /* W,  C,  H,  S,  L,  T, strat */
+    { 14, 12, 13,  1,  5,  1, ZSTD_fast    },  /* base for negative levels */
+    { 14, 14, 15,  1,  5,  0, ZSTD_fast    },  /* level  1 */
+    { 14, 14, 15,  1,  4,  0, ZSTD_fast    },  /* level  2 */
+    { 14, 14, 15,  2,  4,  0, ZSTD_dfast   },  /* level  3 */
+    { 14, 14, 14,  4,  4,  2, ZSTD_greedy  },  /* level  4 */
+    { 14, 14, 14,  3,  4,  4, ZSTD_lazy    },  /* level  5.*/
+    { 14, 14, 14,  4,  4,  8, ZSTD_lazy2   },  /* level  6 */
+    { 14, 14, 14,  6,  4,  8, ZSTD_lazy2   },  /* level  7 */
+    { 14, 14, 14,  8,  4,  8, ZSTD_lazy2   },  /* level  8.*/
+    { 14, 15, 14,  5,  4,  8, ZSTD_btlazy2 },  /* level  9.*/
+    { 14, 15, 14,  9,  4,  8, ZSTD_btlazy2 },  /* level 10.*/
+    { 14, 15, 14,  3,  4, 12, ZSTD_btopt   },  /* level 11.*/
+    { 14, 15, 14,  4,  3, 24, ZSTD_btopt   },  /* level 12.*/
+    { 14, 15, 14,  5,  3, 32, ZSTD_btultra },  /* level 13.*/
+    { 14, 15, 15,  6,  3, 64, ZSTD_btultra },  /* level 14.*/
+    { 14, 15, 15,  7,  3,256, ZSTD_btultra },  /* level 15.*/
+    { 14, 15, 15,  5,  3, 48, ZSTD_btultra2},  /* level 16.*/
+    { 14, 15, 15,  6,  3,128, ZSTD_btultra2},  /* level 17.*/
+    { 14, 15, 15,  7,  3,256, ZSTD_btultra2},  /* level 18.*/
+    { 14, 15, 15,  8,  3,256, ZSTD_btultra2},  /* level 19.*/
+    { 14, 15, 15,  8,  3,512, ZSTD_btultra2},  /* level 20.*/
+    { 14, 15, 15,  9,  3,512, ZSTD_btultra2},  /* level 21.*/
+    { 14, 15, 15, 10,  3,999, ZSTD_btultra2},  /* level 22.*/
+},
+};
+
+
+
+#endif  /* ZSTD_CLEVELS_H */

src/borg/algorithms/zstd/lib/compress/fse_compress.c

@@ -1,6 +1,6 @@
 /* ******************************************************************
  * FSE : Finite State Entropy encoder
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  *  You can contact the author at :
  *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -15,8 +15,6 @@
 /* **************************************************************
 *  Includes
 ****************************************************************/
-#include <stdlib.h>     /* malloc, free, qsort */
-#include <string.h>     /* memcpy, memset */
 #include "../common/compiler.h"
 #include "../common/mem.h"        /* U32, U16, etc. */
 #include "../common/debug.h"      /* assert, DEBUGLOG */
@@ -25,6 +23,10 @@
 #define FSE_STATIC_LINKING_ONLY
 #include "../common/fse.h"
 #include "../common/error_private.h"
+#define ZSTD_DEPS_NEED_MALLOC
+#define ZSTD_DEPS_NEED_MATH64
+#include "../common/zstd_deps.h"  /* ZSTD_malloc, ZSTD_free, ZSTD_memcpy, ZSTD_memset */
+#include "../common/bits.h" /* ZSTD_highbit32 */
 
 
 /* **************************************************************
@@ -74,41 +76,85 @@ size_t FSE_buildCTable_wksp(FSE_CTable* ct,
     void* const FSCT = ((U32*)ptr) + 1 /* header */ + (tableLog ? tableSize>>1 : 1) ;
     FSE_symbolCompressionTransform* const symbolTT = (FSE_symbolCompressionTransform*) (FSCT);
     U32 const step = FSE_TABLESTEP(tableSize);
-    U32 cumul[FSE_MAX_SYMBOL_VALUE+2];
+    U32 const maxSV1 = maxSymbolValue+1;
+
+    U16* cumul = (U16*)workSpace;   /* size = maxSV1 */
+    FSE_FUNCTION_TYPE* const tableSymbol = (FSE_FUNCTION_TYPE*)(cumul + (maxSV1+1));  /* size = tableSize */
 
-    FSE_FUNCTION_TYPE* const tableSymbol = (FSE_FUNCTION_TYPE*)workSpace;
     U32 highThreshold = tableSize-1;
 
+    assert(((size_t)workSpace & 1) == 0);  /* Must be 2 bytes-aligned */
+    if (FSE_BUILD_CTABLE_WORKSPACE_SIZE(maxSymbolValue, tableLog) > wkspSize) return ERROR(tableLog_tooLarge);
     /* CTable header */
-    if (((size_t)1 << tableLog) * sizeof(FSE_FUNCTION_TYPE) > wkspSize) return ERROR(tableLog_tooLarge);
     tableU16[-2] = (U16) tableLog;
     tableU16[-1] = (U16) maxSymbolValue;
     assert(tableLog < 16);   /* required for threshold strategy to work */
 
     /* For explanations on how to distribute symbol values over the table :
-     * http://fastcompression.blogspot.fr/2014/02/fse-distributing-symbol-values.html */
+     * https://fastcompression.blogspot.fr/2014/02/fse-distributing-symbol-values.html */
 
      #ifdef __clang_analyzer__
-     memset(tableSymbol, 0, sizeof(*tableSymbol) * tableSize);   /* useless initialization, just to keep scan-build happy */
+     ZSTD_memset(tableSymbol, 0, sizeof(*tableSymbol) * tableSize);   /* useless initialization, just to keep scan-build happy */
      #endif
 
     /* symbol start positions */
     {   U32 u;
         cumul[0] = 0;
-        for (u=1; u <= maxSymbolValue+1; u++) {
+        for (u=1; u <= maxSV1; u++) {
             if (normalizedCounter[u-1]==-1) {  /* Low proba symbol */
                 cumul[u] = cumul[u-1] + 1;
                 tableSymbol[highThreshold--] = (FSE_FUNCTION_TYPE)(u-1);
             } else {
-                cumul[u] = cumul[u-1] + normalizedCounter[u-1];
+                assert(normalizedCounter[u-1] >= 0);
+                cumul[u] = cumul[u-1] + (U16)normalizedCounter[u-1];
+                assert(cumul[u] >= cumul[u-1]);  /* no overflow */
         }   }
-        cumul[maxSymbolValue+1] = tableSize+1;
+        cumul[maxSV1] = (U16)(tableSize+1);
     }
 
     /* Spread symbols */
-    {   U32 position = 0;
+    if (highThreshold == tableSize - 1) {
+        /* Case for no low prob count symbols. Lay down 8 bytes at a time
+         * to reduce branch misses since we are operating on a small block
+         */
+        BYTE* const spread = tableSymbol + tableSize; /* size = tableSize + 8 (may write beyond tableSize) */
+        {   U64 const add = 0x0101010101010101ull;
+            size_t pos = 0;
+            U64 sv = 0;
+            U32 s;
+            for (s=0; s<maxSV1; ++s, sv += add) {
+                int i;
+                int const n = normalizedCounter[s];
+                MEM_write64(spread + pos, sv);
+                for (i = 8; i < n; i += 8) {
+                    MEM_write64(spread + pos + i, sv);
+                }
+                assert(n>=0);
+                pos += (size_t)n;
+            }
+        }
+        /* Spread symbols across the table. Lack of lowprob symbols means that
+         * we don't need variable sized inner loop, so we can unroll the loop and
+         * reduce branch misses.
+         */
+        {   size_t position = 0;
+            size_t s;
+            size_t const unroll = 2; /* Experimentally determined optimal unroll */
+            assert(tableSize % unroll == 0); /* FSE_MIN_TABLELOG is 5 */
+            for (s = 0; s < (size_t)tableSize; s += unroll) {
+                size_t u;
+                for (u = 0; u < unroll; ++u) {
+                    size_t const uPosition = (position + (u * step)) & tableMask;
+                    tableSymbol[uPosition] = spread[s + u];
+                }
+                position = (position + (unroll * step)) & tableMask;
+            }
+            assert(position == 0);   /* Must have initialized all positions */
+        }
+    } else {
+        U32 position = 0;
         U32 symbol;
-        for (symbol=0; symbol<=maxSymbolValue; symbol++) {
+        for (symbol=0; symbol<maxSV1; symbol++) {
             int nbOccurrences;
             int const freq = normalizedCounter[symbol];
             for (nbOccurrences=0; nbOccurrences<freq; nbOccurrences++) {
@@ -117,7 +163,6 @@ size_t FSE_buildCTable_wksp(FSE_CTable* ct,
                 while (position > highThreshold)
                     position = (position + step) & tableMask;   /* Low proba area */
         }   }
-
         assert(position==0);  /* Must have initialized all positions */
     }
 
@@ -141,16 +186,17 @@ size_t FSE_buildCTable_wksp(FSE_CTable* ct,
             case -1:
             case  1:
                 symbolTT[s].deltaNbBits = (tableLog << 16) - (1<<tableLog);
-                symbolTT[s].deltaFindState = total - 1;
+                assert(total <= INT_MAX);
+                symbolTT[s].deltaFindState = (int)(total - 1);
                 total ++;
                 break;
             default :
-                {
-                    U32 const maxBitsOut = tableLog - BIT_highbit32 (normalizedCounter[s]-1);
-                    U32 const minStatePlus = normalizedCounter[s] << maxBitsOut;
+                assert(normalizedCounter[s] > 1);
+                {   U32 const maxBitsOut = tableLog - ZSTD_highbit32 ((U32)normalizedCounter[s]-1);
+                    U32 const minStatePlus = (U32)normalizedCounter[s] << maxBitsOut;
                     symbolTT[s].deltaNbBits = (maxBitsOut << 16) - minStatePlus;
-                    symbolTT[s].deltaFindState = total - normalizedCounter[s];
-                    total +=  normalizedCounter[s];
+                    symbolTT[s].deltaFindState = (int)(total - (unsigned)normalizedCounter[s]);
+                    total +=  (unsigned)normalizedCounter[s];
     }   }   }   }
 
 #if 0  /* debug : symbol costs */
@@ -161,31 +207,26 @@ size_t FSE_buildCTable_wksp(FSE_CTable* ct,
                 symbol, normalizedCounter[symbol],
                 FSE_getMaxNbBits(symbolTT, symbol),
                 (double)FSE_bitCost(symbolTT, tableLog, symbol, 8) / 256);
-        }
-    }
+    }   }
 #endif
 
     return 0;
 }
 
 
-size_t FSE_buildCTable(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog)
-{
-    FSE_FUNCTION_TYPE tableSymbol[FSE_MAX_TABLESIZE];   /* memset() is not necessary, even if static analyzer complain about it */
-    return FSE_buildCTable_wksp(ct, normalizedCounter, maxSymbolValue, tableLog, tableSymbol, sizeof(tableSymbol));
-}
-
-
 
 #ifndef FSE_COMMONDEFS_ONLY
 
-
 /*-**************************************************************
 *  FSE NCount encoding
 ****************************************************************/
 size_t FSE_NCountWriteBound(unsigned maxSymbolValue, unsigned tableLog)
 {
-    size_t const maxHeaderSize = (((maxSymbolValue+1) * tableLog) >> 3) + 3;
+    size_t const maxHeaderSize = (((maxSymbolValue+1) * tableLog
+                                   + 4 /* bitCount initialized at 4 */
+                                   + 2 /* first two symbols may use one additional bit each */) / 8)
+                                    + 1 /* round up to whole nb bytes */
+                                    + 2 /* additional two bytes for bitstream flush */;
     return maxSymbolValue ? maxHeaderSize : FSE_NCOUNTBOUND;  /* maxSymbolValue==0 ? use default */
 }
 
@@ -302,21 +343,11 @@ size_t FSE_writeNCount (void* buffer, size_t bufferSize,
 *  FSE Compression Code
 ****************************************************************/
 
-FSE_CTable* FSE_createCTable (unsigned maxSymbolValue, unsigned tableLog)
-{
-    size_t size;
-    if (tableLog > FSE_TABLELOG_ABSOLUTE_MAX) tableLog = FSE_TABLELOG_ABSOLUTE_MAX;
-    size = FSE_CTABLE_SIZE_U32 (tableLog, maxSymbolValue) * sizeof(U32);
-    return (FSE_CTable*)malloc(size);
-}
-
-void FSE_freeCTable (FSE_CTable* ct) { free(ct); }
-
 /* provides the minimum logSize to safely represent a distribution */
 static unsigned FSE_minTableLog(size_t srcSize, unsigned maxSymbolValue)
 {
-    U32 minBitsSrc = BIT_highbit32((U32)(srcSize)) + 1;
-    U32 minBitsSymbols = BIT_highbit32(maxSymbolValue) + 2;
+    U32 minBitsSrc = ZSTD_highbit32((U32)(srcSize)) + 1;
+    U32 minBitsSymbols = ZSTD_highbit32(maxSymbolValue) + 2;
     U32 minBits = minBitsSrc < minBitsSymbols ? minBitsSrc : minBitsSymbols;
     assert(srcSize > 1); /* Not supported, RLE should be used instead */
     return minBits;
@@ -324,7 +355,7 @@ static unsigned FSE_minTableLog(size_t srcSize, unsigned maxSymbolValue)
 
 unsigned FSE_optimalTableLog_internal(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, unsigned minus)
 {
-    U32 maxBitsSrc = BIT_highbit32((U32)(srcSize - 1)) - minus;
+    U32 maxBitsSrc = ZSTD_highbit32((U32)(srcSize - 1)) - minus;
     U32 tableLog = maxTableLog;
     U32 minBits = FSE_minTableLog(srcSize, maxSymbolValue);
     assert(srcSize > 1); /* Not supported, RLE should be used instead */
@@ -341,11 +372,10 @@ unsigned FSE_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxS
     return FSE_optimalTableLog_internal(maxTableLog, srcSize, maxSymbolValue, 2);
 }
 
-
 /* Secondary normalization method.
    To be used when primary method fails. */
 
-static size_t FSE_normalizeM2(short* norm, U32 tableLog, const unsigned* count, size_t total, U32 maxSymbolValue)
+static size_t FSE_normalizeM2(short* norm, U32 tableLog, const unsigned* count, size_t total, U32 maxSymbolValue, short lowProbCount)
 {
     short const NOT_YET_ASSIGNED = -2;
     U32 s;
@@ -362,7 +392,7 @@ static size_t FSE_normalizeM2(short* norm, U32 tableLog, const unsigned* count,
             continue;
         }
         if (count[s] <= lowThreshold) {
-            norm[s] = -1;
+            norm[s] = lowProbCount;
             distributed++;
             total -= count[s];
             continue;
@@ -414,7 +444,7 @@ static size_t FSE_normalizeM2(short* norm, U32 tableLog, const unsigned* count,
 
     {   U64 const vStepLog = 62 - tableLog;
         U64 const mid = (1ULL << (vStepLog-1)) - 1;
-        U64 const rStep = ((((U64)1<<vStepLog) * ToDistribute) + mid) / total;   /* scale on remaining */
+        U64 const rStep = ZSTD_div64((((U64)1<<vStepLog) * ToDistribute) + mid, (U32)total);   /* scale on remaining */
         U64 tmpTotal = mid;
         for (s=0; s<=maxSymbolValue; s++) {
             if (norm[s]==NOT_YET_ASSIGNED) {
@@ -431,10 +461,9 @@ static size_t FSE_normalizeM2(short* norm, U32 tableLog, const unsigned* count,
     return 0;
 }
 
-
 size_t FSE_normalizeCount (short* normalizedCounter, unsigned tableLog,
                            const unsigned* count, size_t total,
-                           unsigned maxSymbolValue)
+                           unsigned maxSymbolValue, unsigned useLowProbCount)
 {
     /* Sanity checks */
     if (tableLog==0) tableLog = FSE_DEFAULT_TABLELOG;
@@ -443,8 +472,9 @@ size_t FSE_normalizeCount (short* normalizedCounter, unsigned tableLog,
     if (tableLog < FSE_minTableLog(total, maxSymbolValue)) return ERROR(GENERIC);   /* Too small tableLog, compression potentially impossible */
 
     {   static U32 const rtbTable[] = {     0, 473195, 504333, 520860, 550000, 700000, 750000, 830000 };
+        short const lowProbCount = useLowProbCount ? -1 : 1;
         U64 const scale = 62 - tableLog;
-        U64 const step = ((U64)1<<62) / total;   /* <== here, one division ! */
+        U64 const step = ZSTD_div64((U64)1<<62, (U32)total);   /* <== here, one division ! */
         U64 const vStep = 1ULL<<(scale-20);
         int stillToDistribute = 1<<tableLog;
         unsigned s;
@@ -456,7 +486,7 @@ size_t FSE_normalizeCount (short* normalizedCounter, unsigned tableLog,
             if (count[s] == total) return 0;   /* rle special case */
             if (count[s] == 0) { normalizedCounter[s]=0; continue; }
             if (count[s] <= lowThreshold) {
-                normalizedCounter[s] = -1;
+                normalizedCounter[s] = lowProbCount;
                 stillToDistribute--;
             } else {
                 short proba = (short)((count[s]*step) >> scale);
@@ -470,7 +500,7 @@ size_t FSE_normalizeCount (short* normalizedCounter, unsigned tableLog,
         }   }
         if (-stillToDistribute >= (normalizedCounter[largest] >> 1)) {
             /* corner case, need another normalization method */
-            size_t const errorCode = FSE_normalizeM2(normalizedCounter, tableLog, count, total, maxSymbolValue);
+            size_t const errorCode = FSE_normalizeM2(normalizedCounter, tableLog, count, total, maxSymbolValue, lowProbCount);
             if (FSE_isError(errorCode)) return errorCode;
         }
         else normalizedCounter[largest] += (short)stillToDistribute;
@@ -493,40 +523,6 @@ size_t FSE_normalizeCount (short* normalizedCounter, unsigned tableLog,
     return tableLog;
 }
 
-
-/* fake FSE_CTable, for raw (uncompressed) input */
-size_t FSE_buildCTable_raw (FSE_CTable* ct, unsigned nbBits)
-{
-    const unsigned tableSize = 1 << nbBits;
-    const unsigned tableMask = tableSize - 1;
-    const unsigned maxSymbolValue = tableMask;
-    void* const ptr = ct;
-    U16* const tableU16 = ( (U16*) ptr) + 2;
-    void* const FSCT = ((U32*)ptr) + 1 /* header */ + (tableSize>>1);   /* assumption : tableLog >= 1 */
-    FSE_symbolCompressionTransform* const symbolTT = (FSE_symbolCompressionTransform*) (FSCT);
-    unsigned s;
-
-    /* Sanity checks */
-    if (nbBits < 1) return ERROR(GENERIC);             /* min size */
-
-    /* header */
-    tableU16[-2] = (U16) nbBits;
-    tableU16[-1] = (U16) maxSymbolValue;
-
-    /* Build table */
-    for (s=0; s<tableSize; s++)
-        tableU16[s] = (U16)(tableSize + s);
-
-    /* Build Symbol Transformation Table */
-    {   const U32 deltaNbBits = (nbBits << 16) - (1 << nbBits);
-        for (s=0; s<=maxSymbolValue; s++) {
-            symbolTT[s].deltaNbBits = deltaNbBits;
-            symbolTT[s].deltaFindState = s-1;
-    }   }
-
-    return 0;
-}
-
 /* fake FSE_CTable, for rle input (always same symbol) */
 size_t FSE_buildCTable_rle (FSE_CTable* ct, BYTE symbolValue)
 {
@@ -625,74 +621,4 @@ size_t FSE_compress_usingCTable (void* dst, size_t dstSize,
 
 size_t FSE_compressBound(size_t size) { return FSE_COMPRESSBOUND(size); }
 
-/* FSE_compress_wksp() :
- * Same as FSE_compress2(), but using an externally allocated scratch buffer (`workSpace`).
- * `wkspSize` size must be `(1<<tableLog)`.
- */
-size_t FSE_compress_wksp (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize)
-{
-    BYTE* const ostart = (BYTE*) dst;
-    BYTE* op = ostart;
-    BYTE* const oend = ostart + dstSize;
-
-    unsigned count[FSE_MAX_SYMBOL_VALUE+1];
-    S16   norm[FSE_MAX_SYMBOL_VALUE+1];
-    FSE_CTable* CTable = (FSE_CTable*)workSpace;
-    size_t const CTableSize = FSE_CTABLE_SIZE_U32(tableLog, maxSymbolValue);
-    void* scratchBuffer = (void*)(CTable + CTableSize);
-    size_t const scratchBufferSize = wkspSize - (CTableSize * sizeof(FSE_CTable));
-
-    /* init conditions */
-    if (wkspSize < FSE_WKSP_SIZE_U32(tableLog, maxSymbolValue)) return ERROR(tableLog_tooLarge);
-    if (srcSize <= 1) return 0;  /* Not compressible */
-    if (!maxSymbolValue) maxSymbolValue = FSE_MAX_SYMBOL_VALUE;
-    if (!tableLog) tableLog = FSE_DEFAULT_TABLELOG;
-
-    /* Scan input and build symbol stats */
-    {   CHECK_V_F(maxCount, HIST_count_wksp(count, &maxSymbolValue, src, srcSize, scratchBuffer, scratchBufferSize) );
-        if (maxCount == srcSize) return 1;   /* only a single symbol in src : rle */
-        if (maxCount == 1) return 0;         /* each symbol present maximum once => not compressible */
-        if (maxCount < (srcSize >> 7)) return 0;   /* Heuristic : not compressible enough */
-    }
-
-    tableLog = FSE_optimalTableLog(tableLog, srcSize, maxSymbolValue);
-    CHECK_F( FSE_normalizeCount(norm, tableLog, count, srcSize, maxSymbolValue) );
-
-    /* Write table description header */
-    {   CHECK_V_F(nc_err, FSE_writeNCount(op, oend-op, norm, maxSymbolValue, tableLog) );
-        op += nc_err;
-    }
-
-    /* Compress */
-    CHECK_F( FSE_buildCTable_wksp(CTable, norm, maxSymbolValue, tableLog, scratchBuffer, scratchBufferSize) );
-    {   CHECK_V_F(cSize, FSE_compress_usingCTable(op, oend - op, src, srcSize, CTable) );
-        if (cSize == 0) return 0;   /* not enough space for compressed data */
-        op += cSize;
-    }
-
-    /* check compressibility */
-    if ( (size_t)(op-ostart) >= srcSize-1 ) return 0;
-
-    return op-ostart;
-}
-
-typedef struct {
-    FSE_CTable CTable_max[FSE_CTABLE_SIZE_U32(FSE_MAX_TABLELOG, FSE_MAX_SYMBOL_VALUE)];
-    BYTE scratchBuffer[1 << FSE_MAX_TABLELOG];
-} fseWkspMax_t;
-
-size_t FSE_compress2 (void* dst, size_t dstCapacity, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog)
-{
-    fseWkspMax_t scratchBuffer;
-    DEBUG_STATIC_ASSERT(sizeof(scratchBuffer) >= FSE_WKSP_SIZE_U32(FSE_MAX_TABLELOG, FSE_MAX_SYMBOL_VALUE));   /* compilation failures here means scratchBuffer is not large enough */
-    if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);
-    return FSE_compress_wksp(dst, dstCapacity, src, srcSize, maxSymbolValue, tableLog, &scratchBuffer, sizeof(scratchBuffer));
-}
-
-size_t FSE_compress (void* dst, size_t dstCapacity, const void* src, size_t srcSize)
-{
-    return FSE_compress2(dst, dstCapacity, src, srcSize, FSE_MAX_SYMBOL_VALUE, FSE_DEFAULT_TABLELOG);
-}
-
-
 #endif   /* FSE_COMMONDEFS_ONLY */

src/borg/algorithms/zstd/lib/compress/hist.c

@@ -1,7 +1,7 @@
 /* ******************************************************************
  * hist : Histogram functions
  * part of Finite State Entropy project
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  *  You can contact the author at :
  *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -34,7 +34,7 @@ unsigned HIST_count_simple(unsigned* count, unsigned* maxSymbolValuePtr,
     unsigned maxSymbolValue = *maxSymbolValuePtr;
     unsigned largestCount=0;
 
-    memset(count, 0, (maxSymbolValue+1) * sizeof(*count));
+    ZSTD_memset(count, 0, (maxSymbolValue+1) * sizeof(*count));
     if (srcSize==0) { *maxSymbolValuePtr = 0; return 0; }
 
     while (ip<end) {
@@ -60,9 +60,9 @@ typedef enum { trustInput, checkMaxSymbolValue } HIST_checkInput_e;
  * this design makes better use of OoO cpus,
  * and is noticeably faster when some values are heavily repeated.
  * But it needs some additional workspace for intermediate tables.
- * `workSpace` size must be a table of size >= HIST_WKSP_SIZE_U32.
+ * `workSpace` must be a U32 table of size >= HIST_WKSP_SIZE_U32.
  * @return : largest histogram frequency,
- *           or an error code (notably when histogram would be larger than *maxSymbolValuePtr). */
+ *           or an error code (notably when histogram's alphabet is larger than *maxSymbolValuePtr) */
 static size_t HIST_count_parallel_wksp(
                                 unsigned* count, unsigned* maxSymbolValuePtr,
                                 const void* source, size_t sourceSize,
@@ -71,22 +71,21 @@ static size_t HIST_count_parallel_wksp(
 {
     const BYTE* ip = (const BYTE*)source;
     const BYTE* const iend = ip+sourceSize;
-    unsigned maxSymbolValue = *maxSymbolValuePtr;
+    size_t const countSize = (*maxSymbolValuePtr + 1) * sizeof(*count);
     unsigned max=0;
     U32* const Counting1 = workSpace;
     U32* const Counting2 = Counting1 + 256;
     U32* const Counting3 = Counting2 + 256;
     U32* const Counting4 = Counting3 + 256;
 
-    memset(workSpace, 0, 4*256*sizeof(unsigned));
-
     /* safety checks */
+    assert(*maxSymbolValuePtr <= 255);
     if (!sourceSize) {
-        memset(count, 0, maxSymbolValue + 1);
+        ZSTD_memset(count, 0, countSize);
         *maxSymbolValuePtr = 0;
         return 0;
     }
-    if (!maxSymbolValue) maxSymbolValue = 255;            /* 0 == default */
+    ZSTD_memset(workSpace, 0, 4*256*sizeof(unsigned));
 
     /* by stripes of 16 bytes */
     {   U32 cached = MEM_read32(ip); ip += 4;
@@ -118,21 +117,18 @@ static size_t HIST_count_parallel_wksp(
     /* finish last symbols */
     while (ip<iend) Counting1[*ip++]++;
 
-    if (check) {   /* verify stats will fit into destination table */
-        U32 s; for (s=255; s>maxSymbolValue; s--) {
-            Counting1[s] += Counting2[s] + Counting3[s] + Counting4[s];
-            if (Counting1[s]) return ERROR(maxSymbolValue_tooSmall);
-    }   }
-
     {   U32 s;
-        if (maxSymbolValue > 255) maxSymbolValue = 255;
-        for (s=0; s<=maxSymbolValue; s++) {
-            count[s] = Counting1[s] + Counting2[s] + Counting3[s] + Counting4[s];
-            if (count[s] > max) max = count[s];
+        for (s=0; s<256; s++) {
+            Counting1[s] += Counting2[s] + Counting3[s] + Counting4[s];
+            if (Counting1[s] > max) max = Counting1[s];
     }   }
 
-    while (!count[maxSymbolValue]) maxSymbolValue--;
-    *maxSymbolValuePtr = maxSymbolValue;
+    {   unsigned maxSymbolValue = 255;
+        while (!Counting1[maxSymbolValue]) maxSymbolValue--;
+        if (check && maxSymbolValue > *maxSymbolValuePtr) return ERROR(maxSymbolValue_tooSmall);
+        *maxSymbolValuePtr = maxSymbolValue;
+        ZSTD_memmove(count, Counting1, countSize);   /* in case count & Counting1 are overlapping */
+    }
     return (size_t)max;
 }
 
@@ -152,14 +148,6 @@ size_t HIST_countFast_wksp(unsigned* count, unsigned* maxSymbolValuePtr,
     return HIST_count_parallel_wksp(count, maxSymbolValuePtr, source, sourceSize, trustInput, (U32*)workSpace);
 }
 
-/* fast variant (unsafe : won't check if src contains values beyond count[] limit) */
-size_t HIST_countFast(unsigned* count, unsigned* maxSymbolValuePtr,
-                     const void* source, size_t sourceSize)
-{
-    unsigned tmpCounters[HIST_WKSP_SIZE_U32];
-    return HIST_countFast_wksp(count, maxSymbolValuePtr, source, sourceSize, tmpCounters, sizeof(tmpCounters));
-}
-
 /* HIST_count_wksp() :
  * Same as HIST_count(), but using an externally provided scratch buffer.
  * `workSpace` size must be table of >= HIST_WKSP_SIZE_U32 unsigned */
@@ -175,9 +163,19 @@ size_t HIST_count_wksp(unsigned* count, unsigned* maxSymbolValuePtr,
     return HIST_countFast_wksp(count, maxSymbolValuePtr, source, sourceSize, workSpace, workSpaceSize);
 }
 
+#ifndef ZSTD_NO_UNUSED_FUNCTIONS
+/* fast variant (unsafe : won't check if src contains values beyond count[] limit) */
+size_t HIST_countFast(unsigned* count, unsigned* maxSymbolValuePtr,
+                     const void* source, size_t sourceSize)
+{
+    unsigned tmpCounters[HIST_WKSP_SIZE_U32];
+    return HIST_countFast_wksp(count, maxSymbolValuePtr, source, sourceSize, tmpCounters, sizeof(tmpCounters));
+}
+
 size_t HIST_count(unsigned* count, unsigned* maxSymbolValuePtr,
                  const void* src, size_t srcSize)
 {
     unsigned tmpCounters[HIST_WKSP_SIZE_U32];
     return HIST_count_wksp(count, maxSymbolValuePtr, src, srcSize, tmpCounters, sizeof(tmpCounters));
 }
+#endif

src/borg/algorithms/zstd/lib/compress/hist.h

@@ -1,7 +1,7 @@
 /* ******************************************************************
  * hist : Histogram functions
  * part of Finite State Entropy project
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  *  You can contact the author at :
  *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -14,7 +14,7 @@
 ****************************************************************** */
 
 /* --- dependencies --- */
-#include <stddef.h>   /* size_t */
+#include "../common/zstd_deps.h"   /* size_t */
 
 
 /* --- simple histogram functions --- */

src/borg/algorithms/zstd/lib/compress/zstd_compress_literals.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -13,11 +13,36 @@
  ***************************************/
 #include "zstd_compress_literals.h"
 
+
+/* **************************************************************
+*  Debug Traces
+****************************************************************/
+#if DEBUGLEVEL >= 2
+
+static size_t showHexa(const void* src, size_t srcSize)
+{
+    const BYTE* const ip = (const BYTE*)src;
+    size_t u;
+    for (u=0; u<srcSize; u++) {
+        RAWLOG(5, " %02X", ip[u]); (void)ip;
+    }
+    RAWLOG(5, " \n");
+    return srcSize;
+}
+
+#endif
+
+
+/* **************************************************************
+*  Literals compression - special cases
+****************************************************************/
 size_t ZSTD_noCompressLiterals (void* dst, size_t dstCapacity, const void* src, size_t srcSize)
 {
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     U32   const flSize = 1 + (srcSize>31) + (srcSize>4095);
 
+    DEBUGLOG(5, "ZSTD_noCompressLiterals: srcSize=%zu, dstCapacity=%zu", srcSize, dstCapacity);
+
     RETURN_ERROR_IF(srcSize + flSize > dstCapacity, dstSize_tooSmall, "");
 
     switch(flSize)
@@ -35,17 +60,31 @@ size_t ZSTD_noCompressLiterals (void* dst, size_t dstCapacity, const void* src,
             assert(0);
     }
 
-    memcpy(ostart + flSize, src, srcSize);
-    DEBUGLOG(5, "Raw literals: %u -> %u", (U32)srcSize, (U32)(srcSize + flSize));
+    ZSTD_memcpy(ostart + flSize, src, srcSize);
+    DEBUGLOG(5, "Raw (uncompressed) literals: %u -> %u", (U32)srcSize, (U32)(srcSize + flSize));
     return srcSize + flSize;
 }
 
+static int allBytesIdentical(const void* src, size_t srcSize)
+{
+    assert(srcSize >= 1);
+    assert(src != NULL);
+    {   const BYTE b = ((const BYTE*)src)[0];
+        size_t p;
+        for (p=1; p<srcSize; p++) {
+            if (((const BYTE*)src)[p] != b) return 0;
+        }
+        return 1;
+    }
+}
+
 size_t ZSTD_compressRleLiteralsBlock (void* dst, size_t dstCapacity, const void* src, size_t srcSize)
 {
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     U32   const flSize = 1 + (srcSize>31) + (srcSize>4095);
 
-    (void)dstCapacity;  /* dstCapacity already guaranteed to be >=4, hence large enough */
+    assert(dstCapacity >= 4); (void)dstCapacity;
+    assert(allBytesIdentical(src, srcSize));
 
     switch(flSize)
     {
@@ -63,68 +102,103 @@ size_t ZSTD_compressRleLiteralsBlock (void* dst, size_t dstCapacity, const void*
     }
 
     ostart[flSize] = *(const BYTE*)src;
-    DEBUGLOG(5, "RLE literals: %u -> %u", (U32)srcSize, (U32)flSize + 1);
+    DEBUGLOG(5, "RLE : Repeated Literal (%02X: %u times) -> %u bytes encoded", ((const BYTE*)src)[0], (U32)srcSize, (U32)flSize + 1);
     return flSize+1;
 }
 
-size_t ZSTD_compressLiterals (ZSTD_hufCTables_t const* prevHuf,
-                              ZSTD_hufCTables_t* nextHuf,
-                              ZSTD_strategy strategy, int disableLiteralCompression,
-                              void* dst, size_t dstCapacity,
-                        const void* src, size_t srcSize,
-                              void* entropyWorkspace, size_t entropyWorkspaceSize,
-                        const int bmi2)
+/* ZSTD_minLiteralsToCompress() :
+ * returns minimal amount of literals
+ * for literal compression to even be attempted.
+ * Minimum is made tighter as compression strategy increases.
+ */
+static size_t
+ZSTD_minLiteralsToCompress(ZSTD_strategy strategy, HUF_repeat huf_repeat)
+{
+    assert((int)strategy >= 0);
+    assert((int)strategy <= 9);
+    /* btultra2 : min 8 bytes;
+     * then 2x larger for each successive compression strategy
+     * max threshold 64 bytes */
+    {   int const shift = MIN(9-(int)strategy, 3);
+        size_t const mintc = (huf_repeat == HUF_repeat_valid) ? 6 : (size_t)8 << shift;
+        DEBUGLOG(7, "minLiteralsToCompress = %zu", mintc);
+        return mintc;
+    }
+}
+
+size_t ZSTD_compressLiterals (
+                  void* dst, size_t dstCapacity,
+            const void* src, size_t srcSize,
+                  void* entropyWorkspace, size_t entropyWorkspaceSize,
+            const ZSTD_hufCTables_t* prevHuf,
+                  ZSTD_hufCTables_t* nextHuf,
+                  ZSTD_strategy strategy,
+                  int disableLiteralCompression,
+                  int suspectUncompressible,
+                  int bmi2)
 {
-    size_t const minGain = ZSTD_minGain(srcSize, strategy);
     size_t const lhSize = 3 + (srcSize >= 1 KB) + (srcSize >= 16 KB);
     BYTE*  const ostart = (BYTE*)dst;
     U32 singleStream = srcSize < 256;
     symbolEncodingType_e hType = set_compressed;
     size_t cLitSize;
 
-    DEBUGLOG(5,"ZSTD_compressLiterals (disableLiteralCompression=%i srcSize=%u)",
-                disableLiteralCompression, (U32)srcSize);
+    DEBUGLOG(5,"ZSTD_compressLiterals (disableLiteralCompression=%i, srcSize=%u, dstCapacity=%zu)",
+                disableLiteralCompression, (U32)srcSize, dstCapacity);
+
+    DEBUGLOG(6, "Completed literals listing (%zu bytes)", showHexa(src, srcSize));
 
     /* Prepare nextEntropy assuming reusing the existing table */
-    memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
+    ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
 
     if (disableLiteralCompression)
         return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
 
-    /* small ? don't even attempt compression (speed opt) */
-#   define COMPRESS_LITERALS_SIZE_MIN 63
-    {   size_t const minLitSize = (prevHuf->repeatMode == HUF_repeat_valid) ? 6 : COMPRESS_LITERALS_SIZE_MIN;
-        if (srcSize <= minLitSize) return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
-    }
+    /* if too small, don't even attempt compression (speed opt) */
+    if (srcSize < ZSTD_minLiteralsToCompress(strategy, prevHuf->repeatMode))
+        return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
 
     RETURN_ERROR_IF(dstCapacity < lhSize+1, dstSize_tooSmall, "not enough space for compression");
     {   HUF_repeat repeat = prevHuf->repeatMode;
-        int const preferRepeat = strategy < ZSTD_lazy ? srcSize <= 1024 : 0;
+        int const flags = 0
+            | (bmi2 ? HUF_flags_bmi2 : 0)
+            | (strategy < ZSTD_lazy && srcSize <= 1024 ? HUF_flags_preferRepeat : 0)
+            | (strategy >= HUF_OPTIMAL_DEPTH_THRESHOLD ? HUF_flags_optimalDepth : 0)
+            | (suspectUncompressible ? HUF_flags_suspectUncompressible : 0);
+
+        typedef size_t (*huf_compress_f)(void*, size_t, const void*, size_t, unsigned, unsigned, void*, size_t, HUF_CElt*, HUF_repeat*, int);
+        huf_compress_f huf_compress;
         if (repeat == HUF_repeat_valid && lhSize == 3) singleStream = 1;
-        cLitSize = singleStream ?
-            HUF_compress1X_repeat(
-                ostart+lhSize, dstCapacity-lhSize, src, srcSize,
-                HUF_SYMBOLVALUE_MAX, HUF_TABLELOG_DEFAULT, entropyWorkspace, entropyWorkspaceSize,
-                (HUF_CElt*)nextHuf->CTable, &repeat, preferRepeat, bmi2) :
-            HUF_compress4X_repeat(
-                ostart+lhSize, dstCapacity-lhSize, src, srcSize,
-                HUF_SYMBOLVALUE_MAX, HUF_TABLELOG_DEFAULT, entropyWorkspace, entropyWorkspaceSize,
-                (HUF_CElt*)nextHuf->CTable, &repeat, preferRepeat, bmi2);
+        huf_compress = singleStream ? HUF_compress1X_repeat : HUF_compress4X_repeat;
+        cLitSize = huf_compress(ostart+lhSize, dstCapacity-lhSize,
+                                src, srcSize,
+                                HUF_SYMBOLVALUE_MAX, LitHufLog,
+                                entropyWorkspace, entropyWorkspaceSize,
+                                (HUF_CElt*)nextHuf->CTable,
+                                &repeat, flags);
+        DEBUGLOG(5, "%zu literals compressed into %zu bytes (before header)", srcSize, cLitSize);
         if (repeat != HUF_repeat_none) {
             /* reused the existing table */
-            DEBUGLOG(5, "Reusing previous huffman table");
+            DEBUGLOG(5, "reusing statistics from previous huffman block");
             hType = set_repeat;
         }
     }
 
-    if ((cLitSize==0) | (cLitSize >= srcSize - minGain) | ERR_isError(cLitSize)) {
-        memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
-        return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
-    }
+    {   size_t const minGain = ZSTD_minGain(srcSize, strategy);
+        if ((cLitSize==0) || (cLitSize >= srcSize - minGain) || ERR_isError(cLitSize)) {
+            ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
+            return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
+    }   }
     if (cLitSize==1) {
-        memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
-        return ZSTD_compressRleLiteralsBlock(dst, dstCapacity, src, srcSize);
-    }
+        /* A return value of 1 signals that the alphabet consists of a single symbol.
+         * However, in some rare circumstances, it could be the compressed size (a single byte).
+         * For that outcome to have a chance to happen, it's necessary that `srcSize < 8`.
+         * (it's also necessary to not generate statistics).
+         * Therefore, in such a case, actively check that all bytes are identical. */
+        if ((srcSize >= 8) || allBytesIdentical(src, srcSize)) {
+            ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
+            return ZSTD_compressRleLiteralsBlock(dst, dstCapacity, src, srcSize);
+    }   }
 
     if (hType == set_compressed) {
         /* using a newly constructed table */
@@ -135,16 +209,19 @@ size_t ZSTD_compressLiterals (ZSTD_hufCTables_t const* prevHuf,
     switch(lhSize)
     {
     case 3: /* 2 - 2 - 10 - 10 */
-        {   U32 const lhc = hType + ((!singleStream) << 2) + ((U32)srcSize<<4) + ((U32)cLitSize<<14);
+        if (!singleStream) assert(srcSize >= MIN_LITERALS_FOR_4_STREAMS);
+        {   U32 const lhc = hType + ((U32)(!singleStream) << 2) + ((U32)srcSize<<4) + ((U32)cLitSize<<14);
             MEM_writeLE24(ostart, lhc);
             break;
         }
     case 4: /* 2 - 2 - 14 - 14 */
+        assert(srcSize >= MIN_LITERALS_FOR_4_STREAMS);
         {   U32 const lhc = hType + (2 << 2) + ((U32)srcSize<<4) + ((U32)cLitSize<<18);
             MEM_writeLE32(ostart, lhc);
             break;
         }
     case 5: /* 2 - 2 - 18 - 18 */
+        assert(srcSize >= MIN_LITERALS_FOR_4_STREAMS);
         {   U32 const lhc = hType + (3 << 2) + ((U32)srcSize<<4) + ((U32)cLitSize<<22);
             MEM_writeLE32(ostart, lhc);
             ostart[4] = (BYTE)(cLitSize >> 10);

src/borg/algorithms/zstd/lib/compress/zstd_compress_literals.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -16,14 +16,24 @@
 
 size_t ZSTD_noCompressLiterals (void* dst, size_t dstCapacity, const void* src, size_t srcSize);
 
+/* ZSTD_compressRleLiteralsBlock() :
+ * Conditions :
+ * - All bytes in @src are identical
+ * - dstCapacity >= 4 */
 size_t ZSTD_compressRleLiteralsBlock (void* dst, size_t dstCapacity, const void* src, size_t srcSize);
 
-size_t ZSTD_compressLiterals (ZSTD_hufCTables_t const* prevHuf,
-                              ZSTD_hufCTables_t* nextHuf,
-                              ZSTD_strategy strategy, int disableLiteralCompression,
-                              void* dst, size_t dstCapacity,
+/* ZSTD_compressLiterals():
+ * @entropyWorkspace: must be aligned on 4-bytes boundaries
+ * @entropyWorkspaceSize : must be >= HUF_WORKSPACE_SIZE
+ * @suspectUncompressible: sampling checks, to potentially skip huffman coding
+ */
+size_t ZSTD_compressLiterals (void* dst, size_t dstCapacity,
                         const void* src, size_t srcSize,
                               void* entropyWorkspace, size_t entropyWorkspaceSize,
-                        const int bmi2);
+                        const ZSTD_hufCTables_t* prevHuf,
+                              ZSTD_hufCTables_t* nextHuf,
+                              ZSTD_strategy strategy, int disableLiteralCompression,
+                              int suspectUncompressible,
+                              int bmi2);
 
 #endif /* ZSTD_COMPRESS_LITERALS_H */

src/borg/algorithms/zstd/lib/compress/zstd_compress_sequences.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -50,6 +50,19 @@ static unsigned ZSTD_getFSEMaxSymbolValue(FSE_CTable const* ctable) {
   return maxSymbolValue;
 }
 
+/**
+ * Returns true if we should use ncount=-1 else we should
+ * use ncount=1 for low probability symbols instead.
+ */
+static unsigned ZSTD_useLowProbCount(size_t const nbSeq)
+{
+    /* Heuristic: This should cover most blocks <= 16K and
+     * start to fade out after 16K to about 32K depending on
+     * compressibility.
+     */
+    return nbSeq >= 2048;
+}
+
 /**
  * Returns the cost in bytes of encoding the normalized count header.
  * Returns an error if any of the helper functions return an error.
@@ -60,7 +73,7 @@ static size_t ZSTD_NCountCost(unsigned const* count, unsigned const max,
     BYTE wksp[FSE_NCOUNTBOUND];
     S16 norm[MaxSeq + 1];
     const U32 tableLog = FSE_optimalTableLog(FSELog, nbSeq, max);
-    FORWARD_IF_ERROR(FSE_normalizeCount(norm, tableLog, count, nbSeq, max), "");
+    FORWARD_IF_ERROR(FSE_normalizeCount(norm, tableLog, count, nbSeq, max, ZSTD_useLowProbCount(nbSeq)), "");
     return FSE_writeNCount(wksp, sizeof(wksp), norm, max, tableLog);
 }
 
@@ -72,6 +85,8 @@ static size_t ZSTD_entropyCost(unsigned const* count, unsigned const max, size_t
 {
     unsigned cost = 0;
     unsigned s;
+
+    assert(total > 0);
     for (s = 0; s <= max; ++s) {
         unsigned norm = (unsigned)((256 * count[s]) / total);
         if (count[s] != 0 && norm == 0)
@@ -151,7 +166,7 @@ ZSTD_selectEncodingType(
     if (mostFrequent == nbSeq) {
         *repeatMode = FSE_repeat_none;
         if (isDefaultAllowed && nbSeq <= 2) {
-            /* Prefer set_basic over set_rle when there are 2 or less symbols,
+            /* Prefer set_basic over set_rle when there are 2 or fewer symbols,
              * since RLE uses 1 byte, but set_basic uses 5-6 bits per symbol.
              * If basic encoding isn't possible, always choose RLE.
              */
@@ -219,6 +234,11 @@ ZSTD_selectEncodingType(
     return set_compressed;
 }
 
+typedef struct {
+    S16 norm[MaxSeq + 1];
+    U32 wksp[FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(MaxSeq, MaxFSELog)];
+} ZSTD_BuildCTableWksp;
+
 size_t
 ZSTD_buildCTable(void* dst, size_t dstCapacity,
                 FSE_CTable* nextCTable, U32 FSELog, symbolEncodingType_e type,
@@ -239,13 +259,13 @@ ZSTD_buildCTable(void* dst, size_t dstCapacity,
         *op = codeTable[0];
         return 1;
     case set_repeat:
-        memcpy(nextCTable, prevCTable, prevCTableSize);
+        ZSTD_memcpy(nextCTable, prevCTable, prevCTableSize);
         return 0;
     case set_basic:
         FORWARD_IF_ERROR(FSE_buildCTable_wksp(nextCTable, defaultNorm, defaultMax, defaultNormLog, entropyWorkspace, entropyWorkspaceSize), "");  /* note : could be pre-calculated */
         return 0;
     case set_compressed: {
-        S16 norm[MaxSeq + 1];
+        ZSTD_BuildCTableWksp* wksp = (ZSTD_BuildCTableWksp*)entropyWorkspace;
         size_t nbSeq_1 = nbSeq;
         const U32 tableLog = FSE_optimalTableLog(FSELog, nbSeq, max);
         if (count[codeTable[nbSeq-1]] > 1) {
@@ -253,10 +273,13 @@ ZSTD_buildCTable(void* dst, size_t dstCapacity,
             nbSeq_1--;
         }
         assert(nbSeq_1 > 1);
-        FORWARD_IF_ERROR(FSE_normalizeCount(norm, tableLog, count, nbSeq_1, max), "");
-        {   size_t const NCountSize = FSE_writeNCount(op, oend - op, norm, max, tableLog);   /* overflow protected */
+        assert(entropyWorkspaceSize >= sizeof(ZSTD_BuildCTableWksp));
+        (void)entropyWorkspaceSize;
+        FORWARD_IF_ERROR(FSE_normalizeCount(wksp->norm, tableLog, count, nbSeq_1, max, ZSTD_useLowProbCount(nbSeq_1)), "FSE_normalizeCount failed");
+        assert(oend >= op);
+        {   size_t const NCountSize = FSE_writeNCount(op, (size_t)(oend - op), wksp->norm, max, tableLog);   /* overflow protected */
             FORWARD_IF_ERROR(NCountSize, "FSE_writeNCount failed");
-            FORWARD_IF_ERROR(FSE_buildCTable_wksp(nextCTable, norm, max, tableLog, entropyWorkspace, entropyWorkspaceSize), "");
+            FORWARD_IF_ERROR(FSE_buildCTable_wksp(nextCTable, wksp->norm, max, tableLog, wksp->wksp, sizeof(wksp->wksp)), "FSE_buildCTable_wksp failed");
             return NCountSize;
         }
     }
@@ -290,19 +313,19 @@ ZSTD_encodeSequences_body(
     FSE_initCState2(&stateLitLength,   CTable_LitLength,   llCodeTable[nbSeq-1]);
     BIT_addBits(&blockStream, sequences[nbSeq-1].litLength, LL_bits[llCodeTable[nbSeq-1]]);
     if (MEM_32bits()) BIT_flushBits(&blockStream);
-    BIT_addBits(&blockStream, sequences[nbSeq-1].matchLength, ML_bits[mlCodeTable[nbSeq-1]]);
+    BIT_addBits(&blockStream, sequences[nbSeq-1].mlBase, ML_bits[mlCodeTable[nbSeq-1]]);
     if (MEM_32bits()) BIT_flushBits(&blockStream);
     if (longOffsets) {
         U32 const ofBits = ofCodeTable[nbSeq-1];
         unsigned const extraBits = ofBits - MIN(ofBits, STREAM_ACCUMULATOR_MIN-1);
         if (extraBits) {
-            BIT_addBits(&blockStream, sequences[nbSeq-1].offset, extraBits);
+            BIT_addBits(&blockStream, sequences[nbSeq-1].offBase, extraBits);
             BIT_flushBits(&blockStream);
         }
-        BIT_addBits(&blockStream, sequences[nbSeq-1].offset >> extraBits,
+        BIT_addBits(&blockStream, sequences[nbSeq-1].offBase >> extraBits,
                     ofBits - extraBits);
     } else {
-        BIT_addBits(&blockStream, sequences[nbSeq-1].offset, ofCodeTable[nbSeq-1]);
+        BIT_addBits(&blockStream, sequences[nbSeq-1].offBase, ofCodeTable[nbSeq-1]);
     }
     BIT_flushBits(&blockStream);
 
@@ -316,8 +339,8 @@ ZSTD_encodeSequences_body(
             U32  const mlBits = ML_bits[mlCode];
             DEBUGLOG(6, "encoding: litlen:%2u - matchlen:%2u - offCode:%7u",
                         (unsigned)sequences[n].litLength,
-                        (unsigned)sequences[n].matchLength + MINMATCH,
-                        (unsigned)sequences[n].offset);
+                        (unsigned)sequences[n].mlBase + MINMATCH,
+                        (unsigned)sequences[n].offBase);
                                                                             /* 32b*/  /* 64b*/
                                                                             /* (7)*/  /* (7)*/
             FSE_encodeSymbol(&blockStream, &stateOffsetBits, ofCode);       /* 15 */  /* 15 */
@@ -328,18 +351,18 @@ ZSTD_encodeSequences_body(
                 BIT_flushBits(&blockStream);                                /* (7)*/
             BIT_addBits(&blockStream, sequences[n].litLength, llBits);
             if (MEM_32bits() && ((llBits+mlBits)>24)) BIT_flushBits(&blockStream);
-            BIT_addBits(&blockStream, sequences[n].matchLength, mlBits);
+            BIT_addBits(&blockStream, sequences[n].mlBase, mlBits);
             if (MEM_32bits() || (ofBits+mlBits+llBits > 56)) BIT_flushBits(&blockStream);
             if (longOffsets) {
                 unsigned const extraBits = ofBits - MIN(ofBits, STREAM_ACCUMULATOR_MIN-1);
                 if (extraBits) {
-                    BIT_addBits(&blockStream, sequences[n].offset, extraBits);
+                    BIT_addBits(&blockStream, sequences[n].offBase, extraBits);
                     BIT_flushBits(&blockStream);                            /* (7)*/
                 }
-                BIT_addBits(&blockStream, sequences[n].offset >> extraBits,
+                BIT_addBits(&blockStream, sequences[n].offBase >> extraBits,
                             ofBits - extraBits);                            /* 31 */
             } else {
-                BIT_addBits(&blockStream, sequences[n].offset, ofBits);     /* 31 */
+                BIT_addBits(&blockStream, sequences[n].offBase, ofBits);     /* 31 */
             }
             BIT_flushBits(&blockStream);                                    /* (7)*/
             DEBUGLOG(7, "remaining space : %i", (int)(blockStream.endPtr - blockStream.ptr));
@@ -376,7 +399,7 @@ ZSTD_encodeSequences_default(
 
 #if DYNAMIC_BMI2
 
-static TARGET_ATTRIBUTE("bmi2") size_t
+static BMI2_TARGET_ATTRIBUTE size_t
 ZSTD_encodeSequences_bmi2(
             void* dst, size_t dstCapacity,
             FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,

src/borg/algorithms/zstd/lib/compress/zstd_compress_sequences.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the

src/borg/algorithms/zstd/lib/compress/zstd_compress_superblock.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -15,288 +15,10 @@
 
 #include "../common/zstd_internal.h"  /* ZSTD_getSequenceLength */
 #include "hist.h"                     /* HIST_countFast_wksp */
-#include "zstd_compress_internal.h"
+#include "zstd_compress_internal.h"   /* ZSTD_[huf|fse|entropy]CTablesMetadata_t */
 #include "zstd_compress_sequences.h"
 #include "zstd_compress_literals.h"
 
-/*-*************************************
-*  Superblock entropy buffer structs
-***************************************/
-/** ZSTD_hufCTablesMetadata_t :
- *  Stores Literals Block Type for a super-block in hType, and
- *  huffman tree description in hufDesBuffer.
- *  hufDesSize refers to the size of huffman tree description in bytes.
- *  This metadata is populated in ZSTD_buildSuperBlockEntropy_literal() */
-typedef struct {
-    symbolEncodingType_e hType;
-    BYTE hufDesBuffer[500]; /* TODO give name to this value */
-    size_t hufDesSize;
-} ZSTD_hufCTablesMetadata_t;
-
-/** ZSTD_fseCTablesMetadata_t :
- *  Stores symbol compression modes for a super-block in {ll, ol, ml}Type, and
- *  fse tables in fseTablesBuffer.
- *  fseTablesSize refers to the size of fse tables in bytes.
- *  This metadata is populated in ZSTD_buildSuperBlockEntropy_sequences() */
-typedef struct {
-    symbolEncodingType_e llType;
-    symbolEncodingType_e ofType;
-    symbolEncodingType_e mlType;
-    BYTE fseTablesBuffer[500]; /* TODO give name to this value */
-    size_t fseTablesSize;
-    size_t lastCountSize; /* This is to account for bug in 1.3.4. More detail in ZSTD_compressSubBlock_sequences() */
-} ZSTD_fseCTablesMetadata_t;
-
-typedef struct {
-    ZSTD_hufCTablesMetadata_t hufMetadata;
-    ZSTD_fseCTablesMetadata_t fseMetadata;
-} ZSTD_entropyCTablesMetadata_t;
-
-
-/** ZSTD_buildSuperBlockEntropy_literal() :
- *  Builds entropy for the super-block literals.
- *  Stores literals block type (raw, rle, compressed, repeat) and
- *  huffman description table to hufMetadata.
- *  @return : size of huffman description table or error code */
-static size_t ZSTD_buildSuperBlockEntropy_literal(void* const src, size_t srcSize,
-                                            const ZSTD_hufCTables_t* prevHuf,
-                                                  ZSTD_hufCTables_t* nextHuf,
-                                                  ZSTD_hufCTablesMetadata_t* hufMetadata,
-                                                  const int disableLiteralsCompression,
-                                                  void* workspace, size_t wkspSize)
-{
-    BYTE* const wkspStart = (BYTE*)workspace;
-    BYTE* const wkspEnd = wkspStart + wkspSize;
-    BYTE* const countWkspStart = wkspStart;
-    unsigned* const countWksp = (unsigned*)workspace;
-    const size_t countWkspSize = (HUF_SYMBOLVALUE_MAX + 1) * sizeof(unsigned);
-    BYTE* const nodeWksp = countWkspStart + countWkspSize;
-    const size_t nodeWkspSize = wkspEnd-nodeWksp;
-    unsigned maxSymbolValue = 255;
-    unsigned huffLog = HUF_TABLELOG_DEFAULT;
-    HUF_repeat repeat = prevHuf->repeatMode;
-
-    DEBUGLOG(5, "ZSTD_buildSuperBlockEntropy_literal (srcSize=%zu)", srcSize);
-
-    /* Prepare nextEntropy assuming reusing the existing table */
-    memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
-
-    if (disableLiteralsCompression) {
-        DEBUGLOG(5, "set_basic - disabled");
-        hufMetadata->hType = set_basic;
-        return 0;
-    }
-
-    /* small ? don't even attempt compression (speed opt) */
-#   define COMPRESS_LITERALS_SIZE_MIN 63
-    {   size_t const minLitSize = (prevHuf->repeatMode == HUF_repeat_valid) ? 6 : COMPRESS_LITERALS_SIZE_MIN;
-        if (srcSize <= minLitSize) {
-            DEBUGLOG(5, "set_basic - too small");
-            hufMetadata->hType = set_basic;
-            return 0;
-        }
-    }
-
-    /* Scan input and build symbol stats */
-    {   size_t const largest = HIST_count_wksp (countWksp, &maxSymbolValue, (const BYTE*)src, srcSize, workspace, wkspSize);
-        FORWARD_IF_ERROR(largest, "HIST_count_wksp failed");
-        if (largest == srcSize) {
-            DEBUGLOG(5, "set_rle");
-            hufMetadata->hType = set_rle;
-            return 0;
-        }
-        if (largest <= (srcSize >> 7)+4) {
-            DEBUGLOG(5, "set_basic - no gain");
-            hufMetadata->hType = set_basic;
-            return 0;
-        }
-    }
-
-    /* Validate the previous Huffman table */
-    if (repeat == HUF_repeat_check && !HUF_validateCTable((HUF_CElt const*)prevHuf->CTable, countWksp, maxSymbolValue)) {
-        repeat = HUF_repeat_none;
-    }
-
-    /* Build Huffman Tree */
-    memset(nextHuf->CTable, 0, sizeof(nextHuf->CTable));
-    huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue);
-    {   size_t const maxBits = HUF_buildCTable_wksp((HUF_CElt*)nextHuf->CTable, countWksp,
-                                                    maxSymbolValue, huffLog,
-                                                    nodeWksp, nodeWkspSize);
-        FORWARD_IF_ERROR(maxBits, "HUF_buildCTable_wksp");
-        huffLog = (U32)maxBits;
-        {   /* Build and write the CTable */
-            size_t const newCSize = HUF_estimateCompressedSize(
-                    (HUF_CElt*)nextHuf->CTable, countWksp, maxSymbolValue);
-            size_t const hSize = HUF_writeCTable(
-                    hufMetadata->hufDesBuffer, sizeof(hufMetadata->hufDesBuffer),
-                    (HUF_CElt*)nextHuf->CTable, maxSymbolValue, huffLog);
-            /* Check against repeating the previous CTable */
-            if (repeat != HUF_repeat_none) {
-                size_t const oldCSize = HUF_estimateCompressedSize(
-                        (HUF_CElt const*)prevHuf->CTable, countWksp, maxSymbolValue);
-                if (oldCSize < srcSize && (oldCSize <= hSize + newCSize || hSize + 12 >= srcSize)) {
-                    DEBUGLOG(5, "set_repeat - smaller");
-                    memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
-                    hufMetadata->hType = set_repeat;
-                    return 0;
-                }
-            }
-            if (newCSize + hSize >= srcSize) {
-                DEBUGLOG(5, "set_basic - no gains");
-                memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
-                hufMetadata->hType = set_basic;
-                return 0;
-            }
-            DEBUGLOG(5, "set_compressed (hSize=%u)", (U32)hSize);
-            hufMetadata->hType = set_compressed;
-            nextHuf->repeatMode = HUF_repeat_check;
-            return hSize;
-        }
-    }
-}
-
-/** ZSTD_buildSuperBlockEntropy_sequences() :
- *  Builds entropy for the super-block sequences.
- *  Stores symbol compression modes and fse table to fseMetadata.
- *  @return : size of fse tables or error code */
-static size_t ZSTD_buildSuperBlockEntropy_sequences(seqStore_t* seqStorePtr,
-                                              const ZSTD_fseCTables_t* prevEntropy,
-                                                    ZSTD_fseCTables_t* nextEntropy,
-                                              const ZSTD_CCtx_params* cctxParams,
-                                                    ZSTD_fseCTablesMetadata_t* fseMetadata,
-                                                    void* workspace, size_t wkspSize)
-{
-    BYTE* const wkspStart = (BYTE*)workspace;
-    BYTE* const wkspEnd = wkspStart + wkspSize;
-    BYTE* const countWkspStart = wkspStart;
-    unsigned* const countWksp = (unsigned*)workspace;
-    const size_t countWkspSize = (MaxSeq + 1) * sizeof(unsigned);
-    BYTE* const cTableWksp = countWkspStart + countWkspSize;
-    const size_t cTableWkspSize = wkspEnd-cTableWksp;
-    ZSTD_strategy const strategy = cctxParams->cParams.strategy;
-    FSE_CTable* CTable_LitLength = nextEntropy->litlengthCTable;
-    FSE_CTable* CTable_OffsetBits = nextEntropy->offcodeCTable;
-    FSE_CTable* CTable_MatchLength = nextEntropy->matchlengthCTable;
-    const BYTE* const ofCodeTable = seqStorePtr->ofCode;
-    const BYTE* const llCodeTable = seqStorePtr->llCode;
-    const BYTE* const mlCodeTable = seqStorePtr->mlCode;
-    size_t const nbSeq = seqStorePtr->sequences - seqStorePtr->sequencesStart;
-    BYTE* const ostart = fseMetadata->fseTablesBuffer;
-    BYTE* const oend = ostart + sizeof(fseMetadata->fseTablesBuffer);
-    BYTE* op = ostart;
-
-    assert(cTableWkspSize >= (1 << MaxFSELog) * sizeof(FSE_FUNCTION_TYPE));
-    DEBUGLOG(5, "ZSTD_buildSuperBlockEntropy_sequences (nbSeq=%zu)", nbSeq);
-    memset(workspace, 0, wkspSize);
-
-    fseMetadata->lastCountSize = 0;
-    /* convert length/distances into codes */
-    ZSTD_seqToCodes(seqStorePtr);
-    /* build CTable for Literal Lengths */
-    {   U32 LLtype;
-        unsigned max = MaxLL;
-        size_t const mostFrequent = HIST_countFast_wksp(countWksp, &max, llCodeTable, nbSeq, workspace, wkspSize);  /* can't fail */
-        DEBUGLOG(5, "Building LL table");
-        nextEntropy->litlength_repeatMode = prevEntropy->litlength_repeatMode;
-        LLtype = ZSTD_selectEncodingType(&nextEntropy->litlength_repeatMode,
-                                        countWksp, max, mostFrequent, nbSeq,
-                                        LLFSELog, prevEntropy->litlengthCTable,
-                                        LL_defaultNorm, LL_defaultNormLog,
-                                        ZSTD_defaultAllowed, strategy);
-        assert(set_basic < set_compressed && set_rle < set_compressed);
-        assert(!(LLtype < set_compressed && nextEntropy->litlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */
-        {   size_t const countSize = ZSTD_buildCTable(op, oend - op, CTable_LitLength, LLFSELog, (symbolEncodingType_e)LLtype,
-                                                    countWksp, max, llCodeTable, nbSeq, LL_defaultNorm, LL_defaultNormLog, MaxLL,
-                                                    prevEntropy->litlengthCTable, sizeof(prevEntropy->litlengthCTable),
-                                                    cTableWksp, cTableWkspSize);
-            FORWARD_IF_ERROR(countSize, "ZSTD_buildCTable for LitLens failed");
-            if (LLtype == set_compressed)
-                fseMetadata->lastCountSize = countSize;
-            op += countSize;
-            fseMetadata->llType = (symbolEncodingType_e) LLtype;
-    }   }
-    /* build CTable for Offsets */
-    {   U32 Offtype;
-        unsigned max = MaxOff;
-        size_t const mostFrequent = HIST_countFast_wksp(countWksp, &max, ofCodeTable, nbSeq, workspace, wkspSize);  /* can't fail */
-        /* We can only use the basic table if max <= DefaultMaxOff, otherwise the offsets are too large */
-        ZSTD_defaultPolicy_e const defaultPolicy = (max <= DefaultMaxOff) ? ZSTD_defaultAllowed : ZSTD_defaultDisallowed;
-        DEBUGLOG(5, "Building OF table");
-        nextEntropy->offcode_repeatMode = prevEntropy->offcode_repeatMode;
-        Offtype = ZSTD_selectEncodingType(&nextEntropy->offcode_repeatMode,
-                                        countWksp, max, mostFrequent, nbSeq,
-                                        OffFSELog, prevEntropy->offcodeCTable,
-                                        OF_defaultNorm, OF_defaultNormLog,
-                                        defaultPolicy, strategy);
-        assert(!(Offtype < set_compressed && nextEntropy->offcode_repeatMode != FSE_repeat_none)); /* We don't copy tables */
-        {   size_t const countSize = ZSTD_buildCTable(op, oend - op, CTable_OffsetBits, OffFSELog, (symbolEncodingType_e)Offtype,
-                                                    countWksp, max, ofCodeTable, nbSeq, OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff,
-                                                    prevEntropy->offcodeCTable, sizeof(prevEntropy->offcodeCTable),
-                                                    cTableWksp, cTableWkspSize);
-            FORWARD_IF_ERROR(countSize, "ZSTD_buildCTable for Offsets failed");
-            if (Offtype == set_compressed)
-                fseMetadata->lastCountSize = countSize;
-            op += countSize;
-            fseMetadata->ofType = (symbolEncodingType_e) Offtype;
-    }   }
-    /* build CTable for MatchLengths */
-    {   U32 MLtype;
-        unsigned max = MaxML;
-        size_t const mostFrequent = HIST_countFast_wksp(countWksp, &max, mlCodeTable, nbSeq, workspace, wkspSize);   /* can't fail */
-        DEBUGLOG(5, "Building ML table (remaining space : %i)", (int)(oend-op));
-        nextEntropy->matchlength_repeatMode = prevEntropy->matchlength_repeatMode;
-        MLtype = ZSTD_selectEncodingType(&nextEntropy->matchlength_repeatMode,
-                                        countWksp, max, mostFrequent, nbSeq,
-                                        MLFSELog, prevEntropy->matchlengthCTable,
-                                        ML_defaultNorm, ML_defaultNormLog,
-                                        ZSTD_defaultAllowed, strategy);
-        assert(!(MLtype < set_compressed && nextEntropy->matchlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */
-        {   size_t const countSize = ZSTD_buildCTable(op, oend - op, CTable_MatchLength, MLFSELog, (symbolEncodingType_e)MLtype,
-                                                    countWksp, max, mlCodeTable, nbSeq, ML_defaultNorm, ML_defaultNormLog, MaxML,
-                                                    prevEntropy->matchlengthCTable, sizeof(prevEntropy->matchlengthCTable),
-                                                    cTableWksp, cTableWkspSize);
-            FORWARD_IF_ERROR(countSize, "ZSTD_buildCTable for MatchLengths failed");
-            if (MLtype == set_compressed)
-                fseMetadata->lastCountSize = countSize;
-            op += countSize;
-            fseMetadata->mlType = (symbolEncodingType_e) MLtype;
-    }   }
-    assert((size_t) (op-ostart) <= sizeof(fseMetadata->fseTablesBuffer));
-    return op-ostart;
-}
-
-
-/** ZSTD_buildSuperBlockEntropy() :
- *  Builds entropy for the super-block.
- *  @return : 0 on success or error code */
-static size_t
-ZSTD_buildSuperBlockEntropy(seqStore_t* seqStorePtr,
-                      const ZSTD_entropyCTables_t* prevEntropy,
-                            ZSTD_entropyCTables_t* nextEntropy,
-                      const ZSTD_CCtx_params* cctxParams,
-                            ZSTD_entropyCTablesMetadata_t* entropyMetadata,
-                            void* workspace, size_t wkspSize)
-{
-    size_t const litSize = seqStorePtr->lit - seqStorePtr->litStart;
-    DEBUGLOG(5, "ZSTD_buildSuperBlockEntropy");
-    entropyMetadata->hufMetadata.hufDesSize =
-        ZSTD_buildSuperBlockEntropy_literal(seqStorePtr->litStart, litSize,
-                                            &prevEntropy->huf, &nextEntropy->huf,
-                                            &entropyMetadata->hufMetadata,
-                                            ZSTD_disableLiteralsCompression(cctxParams),
-                                            workspace, wkspSize);
-    FORWARD_IF_ERROR(entropyMetadata->hufMetadata.hufDesSize, "ZSTD_buildSuperBlockEntropy_literal failed");
-    entropyMetadata->fseMetadata.fseTablesSize =
-        ZSTD_buildSuperBlockEntropy_sequences(seqStorePtr,
-                                              &prevEntropy->fse, &nextEntropy->fse,
-                                              cctxParams,
-                                              &entropyMetadata->fseMetadata,
-                                              workspace, wkspSize);
-    FORWARD_IF_ERROR(entropyMetadata->fseMetadata.fseTablesSize, "ZSTD_buildSuperBlockEntropy_sequences failed");
-    return 0;
-}
-
 /** ZSTD_compressSubBlock_literal() :
  *  Compresses literals section for a sub-block.
  *  When we have to write the Huffman table we will sometimes choose a header
@@ -304,7 +26,7 @@ ZSTD_buildSuperBlockEntropy(seqStore_t* seqStorePtr,
  *  before we know the table size + compressed size, so we have a bound on the
  *  table size. If we guessed incorrectly, we fall back to uncompressed literals.
  *
- *  We write the header when writeEntropy=1 and set entropyWrriten=1 when we succeeded
+ *  We write the header when writeEntropy=1 and set entropyWritten=1 when we succeeded
  *  in writing the header, otherwise it is set to 0.
  *
  *  hufMetadata->hType has literals block type info.
@@ -314,13 +36,14 @@ ZSTD_buildSuperBlockEntropy(seqStore_t* seqStorePtr,
  *      If it is set_compressed, first sub-block's literals section will be Treeless_Literals_Block
  *      and the following sub-blocks' literals sections will be Treeless_Literals_Block.
  *  @return : compressed size of literals section of a sub-block
- *            Or 0 if it unable to compress.
+ *            Or 0 if unable to compress.
  *            Or error code */
-static size_t ZSTD_compressSubBlock_literal(const HUF_CElt* hufTable,
-                                    const ZSTD_hufCTablesMetadata_t* hufMetadata,
-                                    const BYTE* literals, size_t litSize,
-                                    void* dst, size_t dstSize,
-                                    const int bmi2, int writeEntropy, int* entropyWritten)
+static size_t
+ZSTD_compressSubBlock_literal(const HUF_CElt* hufTable,
+                              const ZSTD_hufCTablesMetadata_t* hufMetadata,
+                              const BYTE* literals, size_t litSize,
+                              void* dst, size_t dstSize,
+                              const int bmi2, int writeEntropy, int* entropyWritten)
 {
     size_t const header = writeEntropy ? 200 : 0;
     size_t const lhSize = 3 + (litSize >= (1 KB - header)) + (litSize >= (16 KB - header));
@@ -331,8 +54,6 @@ static size_t ZSTD_compressSubBlock_literal(const HUF_CElt* hufTable,
     symbolEncodingType_e hType = writeEntropy ? hufMetadata->hType : set_repeat;
     size_t cLitSize = 0;
 
-    (void)bmi2; /* TODO bmi2... */
-
     DEBUGLOG(5, "ZSTD_compressSubBlock_literal (litSize=%zu, lhSize=%zu, writeEntropy=%d)", litSize, lhSize, writeEntropy);
 
     *entropyWritten = 0;
@@ -348,15 +69,15 @@ static size_t ZSTD_compressSubBlock_literal(const HUF_CElt* hufTable,
     assert(hufMetadata->hType == set_compressed || hufMetadata->hType == set_repeat);
 
     if (writeEntropy && hufMetadata->hType == set_compressed) {
-        memcpy(op, hufMetadata->hufDesBuffer, hufMetadata->hufDesSize);
+        ZSTD_memcpy(op, hufMetadata->hufDesBuffer, hufMetadata->hufDesSize);
         op += hufMetadata->hufDesSize;
         cLitSize += hufMetadata->hufDesSize;
         DEBUGLOG(5, "ZSTD_compressSubBlock_literal (hSize=%zu)", hufMetadata->hufDesSize);
     }
 
-    /* TODO bmi2 */
-    {   const size_t cSize = singleStream ? HUF_compress1X_usingCTable(op, oend-op, literals, litSize, hufTable)
-                                          : HUF_compress4X_usingCTable(op, oend-op, literals, litSize, hufTable);
+    {   int const flags = bmi2 ? HUF_flags_bmi2 : 0;
+        const size_t cSize = singleStream ? HUF_compress1X_usingCTable(op, oend-op, literals, litSize, hufTable, flags)
+                                          : HUF_compress4X_usingCTable(op, oend-op, literals, litSize, hufTable, flags);
         op += cSize;
         cLitSize += cSize;
         if (cSize == 0 || ERR_isError(cSize)) {
@@ -404,12 +125,17 @@ static size_t ZSTD_compressSubBlock_literal(const HUF_CElt* hufTable,
     return op-ostart;
 }
 
-static size_t ZSTD_seqDecompressedSize(seqStore_t const* seqStore, const seqDef* sequences, size_t nbSeq, size_t litSize, int lastSequence) {
+static size_t
+ZSTD_seqDecompressedSize(seqStore_t const* seqStore,
+                   const seqDef* sequences, size_t nbSeq,
+                         size_t litSize, int lastSequence)
+{
     const seqDef* const sstart = sequences;
     const seqDef* const send = sequences + nbSeq;
     const seqDef* sp = sstart;
     size_t matchLengthSum = 0;
     size_t litLengthSum = 0;
+    (void)(litLengthSum); /* suppress unused variable warning on some environments */
     while (send-sp > 0) {
         ZSTD_sequenceLength const seqLen = ZSTD_getSequenceLength(seqStore, sp);
         litLengthSum += seqLen.litLength;
@@ -433,13 +159,14 @@ static size_t ZSTD_seqDecompressedSize(seqStore_t const* seqStore, const seqDef*
  *  @return : compressed size of sequences section of a sub-block
  *            Or 0 if it is unable to compress
  *            Or error code. */
-static size_t ZSTD_compressSubBlock_sequences(const ZSTD_fseCTables_t* fseTables,
-                                              const ZSTD_fseCTablesMetadata_t* fseMetadata,
-                                              const seqDef* sequences, size_t nbSeq,
-                                              const BYTE* llCode, const BYTE* mlCode, const BYTE* ofCode,
-                                              const ZSTD_CCtx_params* cctxParams,
-                                              void* dst, size_t dstCapacity,
-                                              const int bmi2, int writeEntropy, int* entropyWritten)
+static size_t
+ZSTD_compressSubBlock_sequences(const ZSTD_fseCTables_t* fseTables,
+                                const ZSTD_fseCTablesMetadata_t* fseMetadata,
+                                const seqDef* sequences, size_t nbSeq,
+                                const BYTE* llCode, const BYTE* mlCode, const BYTE* ofCode,
+                                const ZSTD_CCtx_params* cctxParams,
+                                void* dst, size_t dstCapacity,
+                                const int bmi2, int writeEntropy, int* entropyWritten)
 {
     const int longOffsets = cctxParams->cParams.windowLog > STREAM_ACCUMULATOR_MIN;
     BYTE* const ostart = (BYTE*)dst;
@@ -474,7 +201,7 @@ static size_t ZSTD_compressSubBlock_sequences(const ZSTD_fseCTables_t* fseTables
         const U32 MLtype = fseMetadata->mlType;
         DEBUGLOG(5, "ZSTD_compressSubBlock_sequences (fseTablesSize=%zu)", fseMetadata->fseTablesSize);
         *seqHead = (BYTE)((LLtype<<6) + (Offtype<<4) + (MLtype<<2));
-        memcpy(op, fseMetadata->fseTablesBuffer, fseMetadata->fseTablesSize);
+        ZSTD_memcpy(op, fseMetadata->fseTablesBuffer, fseMetadata->fseTablesSize);
         op += fseMetadata->fseTablesSize;
     } else {
         const U32 repeat = set_repeat;
@@ -602,8 +329,8 @@ static size_t ZSTD_estimateSubBlockSize_literal(const BYTE* literals, size_t lit
 static size_t ZSTD_estimateSubBlockSize_symbolType(symbolEncodingType_e type,
                         const BYTE* codeTable, unsigned maxCode,
                         size_t nbSeq, const FSE_CTable* fseCTable,
-                        const U32* additionalBits,
-                        short const* defaultNorm, U32 defaultNormLog,
+                        const U8* additionalBits,
+                        short const* defaultNorm, U32 defaultNormLog, U32 defaultMax,
                         void* workspace, size_t wkspSize)
 {
     unsigned* const countWksp = (unsigned*)workspace;
@@ -615,7 +342,11 @@ static size_t ZSTD_estimateSubBlockSize_symbolType(symbolEncodingType_e type,
 
     HIST_countFast_wksp(countWksp, &max, codeTable, nbSeq, workspace, wkspSize);  /* can't fail */
     if (type == set_basic) {
-        cSymbolTypeSizeEstimateInBits = ZSTD_crossEntropyCost(defaultNorm, defaultNormLog, countWksp, max);
+        /* We selected this encoding type, so it must be valid. */
+        assert(max <= defaultMax);
+        cSymbolTypeSizeEstimateInBits = max <= defaultMax
+                ? ZSTD_crossEntropyCost(defaultNorm, defaultNormLog, countWksp, max)
+                : ERROR(GENERIC);
     } else if (type == set_rle) {
         cSymbolTypeSizeEstimateInBits = 0;
     } else if (type == set_compressed || type == set_repeat) {
@@ -639,19 +370,20 @@ static size_t ZSTD_estimateSubBlockSize_sequences(const BYTE* ofCodeTable,
                                                   void* workspace, size_t wkspSize,
                                                   int writeEntropy)
 {
-    size_t sequencesSectionHeaderSize = 3; /* Use hard coded size of 3 bytes */
+    size_t const sequencesSectionHeaderSize = 3; /* Use hard coded size of 3 bytes */
     size_t cSeqSizeEstimate = 0;
+    if (nbSeq == 0) return sequencesSectionHeaderSize;
     cSeqSizeEstimate += ZSTD_estimateSubBlockSize_symbolType(fseMetadata->ofType, ofCodeTable, MaxOff,
                                          nbSeq, fseTables->offcodeCTable, NULL,
-                                         OF_defaultNorm, OF_defaultNormLog,
+                                         OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff,
                                          workspace, wkspSize);
     cSeqSizeEstimate += ZSTD_estimateSubBlockSize_symbolType(fseMetadata->llType, llCodeTable, MaxLL,
                                          nbSeq, fseTables->litlengthCTable, LL_bits,
-                                         LL_defaultNorm, LL_defaultNormLog,
+                                         LL_defaultNorm, LL_defaultNormLog, MaxLL,
                                          workspace, wkspSize);
     cSeqSizeEstimate += ZSTD_estimateSubBlockSize_symbolType(fseMetadata->mlType, mlCodeTable, MaxML,
                                          nbSeq, fseTables->matchlengthCTable, ML_bits,
-                                         ML_defaultNorm, ML_defaultNormLog,
+                                         ML_defaultNorm, ML_defaultNormLog, MaxML,
                                          workspace, wkspSize);
     if (writeEntropy) cSeqSizeEstimate += fseMetadata->fseTablesSize;
     return cSeqSizeEstimate + sequencesSectionHeaderSize;
@@ -747,7 +479,7 @@ static size_t ZSTD_compressSubBlock_multi(const seqStore_t* seqStorePtr,
         /* I think there is an optimization opportunity here.
          * Calling ZSTD_estimateSubBlockSize for every sequence can be wasteful
          * since it recalculates estimate from scratch.
-         * For example, it would recount literal distribution and symbol codes everytime.
+         * For example, it would recount literal distribution and symbol codes every time.
          */
         cBlockSizeEstimate = ZSTD_estimateSubBlockSize(lp, litSize, ofCodePtr, llCodePtr, mlCodePtr, seqCount,
                                                        &nextCBlock->entropy, entropyMetadata,
@@ -790,7 +522,7 @@ static size_t ZSTD_compressSubBlock_multi(const seqStore_t* seqStorePtr,
     } while (!lastSequence);
     if (writeLitEntropy) {
         DEBUGLOG(5, "ZSTD_compressSubBlock_multi has literal entropy tables unwritten");
-        memcpy(&nextCBlock->entropy.huf, &prevCBlock->entropy.huf, sizeof(prevCBlock->entropy.huf));
+        ZSTD_memcpy(&nextCBlock->entropy.huf, &prevCBlock->entropy.huf, sizeof(prevCBlock->entropy.huf));
     }
     if (writeSeqEntropy && ZSTD_needSequenceEntropyTables(&entropyMetadata->fseMetadata)) {
         /* If we haven't written our entropy tables, then we've violated our contract and
@@ -809,11 +541,11 @@ static size_t ZSTD_compressSubBlock_multi(const seqStore_t* seqStorePtr,
         if (sp < send) {
             seqDef const* seq;
             repcodes_t rep;
-            memcpy(&rep, prevCBlock->rep, sizeof(rep)); 
+            ZSTD_memcpy(&rep, prevCBlock->rep, sizeof(rep));
             for (seq = sstart; seq < sp; ++seq) {
-                rep = ZSTD_updateRep(rep.rep, seq->offset - 1, ZSTD_getSequenceLength(seqStorePtr, seq).litLength == 0);
+                ZSTD_updateRep(rep.rep, seq->offBase, ZSTD_getSequenceLength(seqStorePtr, seq).litLength == 0);
             }
-            memcpy(nextCBlock->rep, &rep, sizeof(rep));
+            ZSTD_memcpy(nextCBlock->rep, &rep, sizeof(rep));
         }
     }
     DEBUGLOG(5, "ZSTD_compressSubBlock_multi compressed");
@@ -826,12 +558,12 @@ size_t ZSTD_compressSuperBlock(ZSTD_CCtx* zc,
                                unsigned lastBlock) {
     ZSTD_entropyCTablesMetadata_t entropyMetadata;
 
-    FORWARD_IF_ERROR(ZSTD_buildSuperBlockEntropy(&zc->seqStore,
+    FORWARD_IF_ERROR(ZSTD_buildBlockEntropyStats(&zc->seqStore,
           &zc->blockState.prevCBlock->entropy,
           &zc->blockState.nextCBlock->entropy,
           &zc->appliedParams,
           &entropyMetadata,
-          zc->entropyWorkspace, HUF_WORKSPACE_SIZE /* statically allocated in resetCCtx */), "");
+          zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */), "");
 
     return ZSTD_compressSubBlock_multi(&zc->seqStore,
             zc->blockState.prevCBlock,
@@ -841,5 +573,5 @@ size_t ZSTD_compressSuperBlock(ZSTD_CCtx* zc,
             dst, dstCapacity,
             src, srcSize,
             zc->bmi2, lastBlock,
-            zc->entropyWorkspace, HUF_WORKSPACE_SIZE /* statically allocated in resetCCtx */);
+            zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */);
 }

src/borg/algorithms/zstd/lib/compress/zstd_compress_superblock.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the

src/borg/algorithms/zstd/lib/compress/zstd_cwksp.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -14,7 +14,9 @@
 /*-*************************************
 *  Dependencies
 ***************************************/
+#include "../common/allocations.h"  /* ZSTD_customMalloc, ZSTD_customFree */
 #include "../common/zstd_internal.h"
+#include "../common/portability_macros.h"
 
 #if defined (__cplusplus)
 extern "C" {
@@ -35,15 +37,30 @@ extern "C" {
 #define ZSTD_CWKSP_ASAN_REDZONE_SIZE 128
 #endif
 
+
+/* Set our tables and aligneds to align by 64 bytes */
+#define ZSTD_CWKSP_ALIGNMENT_BYTES 64
+
 /*-*************************************
 *  Structures
 ***************************************/
 typedef enum {
     ZSTD_cwksp_alloc_objects,
-    ZSTD_cwksp_alloc_buffers,
-    ZSTD_cwksp_alloc_aligned
+    ZSTD_cwksp_alloc_aligned_init_once,
+    ZSTD_cwksp_alloc_aligned,
+    ZSTD_cwksp_alloc_buffers
 } ZSTD_cwksp_alloc_phase_e;
 
+/**
+ * Used to describe whether the workspace is statically allocated (and will not
+ * necessarily ever be freed), or if it's dynamically allocated and we can
+ * expect a well-formed caller to free this.
+ */
+typedef enum {
+    ZSTD_cwksp_dynamic_alloc,
+    ZSTD_cwksp_static_alloc
+} ZSTD_cwksp_static_alloc_e;
+
 /**
  * Zstd fits all its internal datastructures into a single continuous buffer,
  * so that it only needs to perform a single OS allocation (or so that a buffer
@@ -84,15 +101,15 @@ typedef enum {
  *
  * Workspace Layout:
  *
- * [                        ... workspace ...                         ]
- * [objects][tables ... ->] free space [<- ... aligned][<- ... buffers]
+ * [                        ... workspace ...                           ]
+ * [objects][tables ->] free space [<- buffers][<- aligned][<- init once]
  *
  * The various objects that live in the workspace are divided into the
  * following categories, and are allocated separately:
  *
  * - Static objects: this is optionally the enclosing ZSTD_CCtx or ZSTD_CDict,
  *   so that literally everything fits in a single buffer. Note: if present,
- *   this must be the first object in the workspace, since ZSTD_free{CCtx,
+ *   this must be the first object in the workspace, since ZSTD_customFree{CCtx,
  *   CDict}() rely on a pointer comparison to see whether one or two frees are
  *   required.
  *
@@ -107,10 +124,20 @@ typedef enum {
  * - Tables: these are any of several different datastructures (hash tables,
  *   chain tables, binary trees) that all respect a common format: they are
  *   uint32_t arrays, all of whose values are between 0 and (nextSrc - base).
- *   Their sizes depend on the cparams.
+ *   Their sizes depend on the cparams. These tables are 64-byte aligned.
  *
- * - Aligned: these buffers are used for various purposes that require 4 byte
- *   alignment, but don't require any initialization before they're used.
+ * - Init once: these buffers require to be initialized at least once before
+ *   use. They should be used when we want to skip memory initialization
+ *   while not triggering memory checkers (like Valgrind) when reading from
+ *   from this memory without writing to it first.
+ *   These buffers should be used carefully as they might contain data
+ *   from previous compressions.
+ *   Buffers are aligned to 64 bytes.
+ *
+ * - Aligned: these buffers don't require any initialization before they're
+ *   used. The user of the buffer should make sure they write into a buffer
+ *   location before reading from it.
+ *   Buffers are aligned to 64 bytes.
  *
  * - Buffers: these buffers are used for various purposes that don't require
  *   any alignment or initialization before they're used. This means they can
@@ -122,9 +149,9 @@ typedef enum {
  * correctly packed into the workspace buffer. That order is:
  *
  * 1. Objects
- * 2. Buffers
- * 3. Aligned
- * 4. Tables
+ * 2. Init once / Tables
+ * 3. Aligned / Tables
+ * 4. Buffers / Tables
  *
  * Attempts to reserve objects of different types out of order will fail.
  */
@@ -136,10 +163,12 @@ typedef struct {
     void* tableEnd;
     void* tableValidEnd;
     void* allocStart;
+    void* initOnceStart;
 
-    int allocFailed;
+    BYTE allocFailed;
     int workspaceOversizedDuration;
     ZSTD_cwksp_alloc_phase_e phase;
+    ZSTD_cwksp_static_alloc_e isStatic;
 } ZSTD_cwksp;
 
 /*-*************************************
@@ -147,6 +176,7 @@ typedef struct {
 ***************************************/
 
 MEM_STATIC size_t ZSTD_cwksp_available_space(ZSTD_cwksp* ws);
+MEM_STATIC void*  ZSTD_cwksp_initialAllocStart(ZSTD_cwksp* ws);
 
 MEM_STATIC void ZSTD_cwksp_assert_internal_consistency(ZSTD_cwksp* ws) {
     (void)ws;
@@ -156,6 +186,20 @@ MEM_STATIC void ZSTD_cwksp_assert_internal_consistency(ZSTD_cwksp* ws) {
     assert(ws->tableEnd <= ws->allocStart);
     assert(ws->tableValidEnd <= ws->allocStart);
     assert(ws->allocStart <= ws->workspaceEnd);
+    assert(ws->initOnceStart <= ZSTD_cwksp_initialAllocStart(ws));
+    assert(ws->workspace <= ws->initOnceStart);
+#if ZSTD_MEMORY_SANITIZER
+    {
+        intptr_t const offset = __msan_test_shadow(ws->initOnceStart,
+            (U8*)ZSTD_cwksp_initialAllocStart(ws) - (U8*)ws->initOnceStart);
+#if defined(ZSTD_MSAN_PRINT)
+        if(offset!=-1) {
+            __msan_print_shadow((U8*)ws->initOnceStart + offset - 8, 32);
+        }
+#endif
+        assert(offset==-1);
+    };
+#endif
 }
 
 /**
@@ -176,63 +220,72 @@ MEM_STATIC size_t ZSTD_cwksp_align(size_t size, size_t const align) {
  * Since tables aren't currently redzoned, you don't need to call through this
  * to figure out how much space you need for the matchState tables. Everything
  * else is though.
+ *
+ * Do not use for sizing aligned buffers. Instead, use ZSTD_cwksp_aligned_alloc_size().
  */
 MEM_STATIC size_t ZSTD_cwksp_alloc_size(size_t size) {
-#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
+    if (size == 0)
+        return 0;
+#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
     return size + 2 * ZSTD_CWKSP_ASAN_REDZONE_SIZE;
 #else
     return size;
 #endif
 }
 
-MEM_STATIC void ZSTD_cwksp_internal_advance_phase(
-        ZSTD_cwksp* ws, ZSTD_cwksp_alloc_phase_e phase) {
-    assert(phase >= ws->phase);
-    if (phase > ws->phase) {
-        if (ws->phase < ZSTD_cwksp_alloc_buffers &&
-                phase >= ZSTD_cwksp_alloc_buffers) {
-            ws->tableValidEnd = ws->objectEnd;
-        }
-        if (ws->phase < ZSTD_cwksp_alloc_aligned &&
-                phase >= ZSTD_cwksp_alloc_aligned) {
-            /* If unaligned allocations down from a too-large top have left us
-             * unaligned, we need to realign our alloc ptr. Technically, this
-             * can consume space that is unaccounted for in the neededSpace
-             * calculation. However, I believe this can only happen when the
-             * workspace is too large, and specifically when it is too large
-             * by a larger margin than the space that will be consumed. */
-            /* TODO: cleaner, compiler warning friendly way to do this??? */
-            ws->allocStart = (BYTE*)ws->allocStart - ((size_t)ws->allocStart & (sizeof(U32)-1));
-            if (ws->allocStart < ws->tableValidEnd) {
-                ws->tableValidEnd = ws->allocStart;
-            }
-        }
-        ws->phase = phase;
-    }
+/**
+ * Returns an adjusted alloc size that is the nearest larger multiple of 64 bytes.
+ * Used to determine the number of bytes required for a given "aligned".
+ */
+MEM_STATIC size_t ZSTD_cwksp_aligned_alloc_size(size_t size) {
+    return ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(size, ZSTD_CWKSP_ALIGNMENT_BYTES));
 }
 
 /**
- * Returns whether this object/buffer/etc was allocated in this workspace.
+ * Returns the amount of additional space the cwksp must allocate
+ * for internal purposes (currently only alignment).
  */
-MEM_STATIC int ZSTD_cwksp_owns_buffer(const ZSTD_cwksp* ws, const void* ptr) {
-    return (ptr != NULL) && (ws->workspace <= ptr) && (ptr <= ws->workspaceEnd);
+MEM_STATIC size_t ZSTD_cwksp_slack_space_required(void) {
+    /* For alignment, the wksp will always allocate an additional 2*ZSTD_CWKSP_ALIGNMENT_BYTES
+     * bytes to align the beginning of tables section and end of buffers;
+     */
+    size_t const slackSpace = ZSTD_CWKSP_ALIGNMENT_BYTES * 2;
+    return slackSpace;
 }
 
+
 /**
- * Internal function. Do not use directly.
+ * Return the number of additional bytes required to align a pointer to the given number of bytes.
+ * alignBytes must be a power of two.
  */
-MEM_STATIC void* ZSTD_cwksp_reserve_internal(
-        ZSTD_cwksp* ws, size_t bytes, ZSTD_cwksp_alloc_phase_e phase) {
-    void* alloc;
-    void* bottom = ws->tableEnd;
-    ZSTD_cwksp_internal_advance_phase(ws, phase);
-    alloc = (BYTE *)ws->allocStart - bytes;
+MEM_STATIC size_t ZSTD_cwksp_bytes_to_align_ptr(void* ptr, const size_t alignBytes) {
+    size_t const alignBytesMask = alignBytes - 1;
+    size_t const bytes = (alignBytes - ((size_t)ptr & (alignBytesMask))) & alignBytesMask;
+    assert((alignBytes & alignBytesMask) == 0);
+    assert(bytes < alignBytes);
+    return bytes;
+}
 
-#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
-    /* over-reserve space */
-    alloc = (BYTE *)alloc - 2 * ZSTD_CWKSP_ASAN_REDZONE_SIZE;
-#endif
+/**
+ * Returns the initial value for allocStart which is used to determine the position from
+ * which we can allocate from the end of the workspace.
+ */
+MEM_STATIC void*  ZSTD_cwksp_initialAllocStart(ZSTD_cwksp* ws) {
+    return (void*)((size_t)ws->workspaceEnd & ~(ZSTD_CWKSP_ALIGNMENT_BYTES-1));
+}
 
+/**
+ * Internal function. Do not use directly.
+ * Reserves the given number of bytes within the aligned/buffer segment of the wksp,
+ * which counts from the end of the wksp (as opposed to the object/table segment).
+ *
+ * Returns a pointer to the beginning of that space.
+ */
+MEM_STATIC void*
+ZSTD_cwksp_reserve_internal_buffer_space(ZSTD_cwksp* ws, size_t const bytes)
+{
+    void* const alloc = (BYTE*)ws->allocStart - bytes;
+    void* const bottom = ws->tableEnd;
     DEBUGLOG(5, "cwksp: reserving %p %zd bytes, %zd bytes remaining",
         alloc, bytes, ZSTD_cwksp_available_space(ws) - bytes);
     ZSTD_cwksp_assert_internal_consistency(ws);
@@ -242,16 +295,88 @@ MEM_STATIC void* ZSTD_cwksp_reserve_internal(
         ws->allocFailed = 1;
         return NULL;
     }
+    /* the area is reserved from the end of wksp.
+     * If it overlaps with tableValidEnd, it voids guarantees on values' range */
     if (alloc < ws->tableValidEnd) {
         ws->tableValidEnd = alloc;
     }
     ws->allocStart = alloc;
+    return alloc;
+}
+
+/**
+ * Moves the cwksp to the next phase, and does any necessary allocations.
+ * cwksp initialization must necessarily go through each phase in order.
+ * Returns a 0 on success, or zstd error
+ */
+MEM_STATIC size_t
+ZSTD_cwksp_internal_advance_phase(ZSTD_cwksp* ws, ZSTD_cwksp_alloc_phase_e phase)
+{
+    assert(phase >= ws->phase);
+    if (phase > ws->phase) {
+        /* Going from allocating objects to allocating initOnce / tables */
+        if (ws->phase < ZSTD_cwksp_alloc_aligned_init_once &&
+            phase >= ZSTD_cwksp_alloc_aligned_init_once) {
+            ws->tableValidEnd = ws->objectEnd;
+            ws->initOnceStart = ZSTD_cwksp_initialAllocStart(ws);
+
+            {   /* Align the start of the tables to 64 bytes. Use [0, 63] bytes */
+                void *const alloc = ws->objectEnd;
+                size_t const bytesToAlign = ZSTD_cwksp_bytes_to_align_ptr(alloc, ZSTD_CWKSP_ALIGNMENT_BYTES);
+                void *const objectEnd = (BYTE *) alloc + bytesToAlign;
+                DEBUGLOG(5, "reserving table alignment addtl space: %zu", bytesToAlign);
+                RETURN_ERROR_IF(objectEnd > ws->workspaceEnd, memory_allocation,
+                                "table phase - alignment initial allocation failed!");
+                ws->objectEnd = objectEnd;
+                ws->tableEnd = objectEnd;  /* table area starts being empty */
+                if (ws->tableValidEnd < ws->tableEnd) {
+                    ws->tableValidEnd = ws->tableEnd;
+                }
+            }
+        }
+        ws->phase = phase;
+        ZSTD_cwksp_assert_internal_consistency(ws);
+    }
+    return 0;
+}
+
+/**
+ * Returns whether this object/buffer/etc was allocated in this workspace.
+ */
+MEM_STATIC int ZSTD_cwksp_owns_buffer(const ZSTD_cwksp* ws, const void* ptr)
+{
+    return (ptr != NULL) && (ws->workspace <= ptr) && (ptr < ws->workspaceEnd);
+}
 
-#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
+/**
+ * Internal function. Do not use directly.
+ */
+MEM_STATIC void*
+ZSTD_cwksp_reserve_internal(ZSTD_cwksp* ws, size_t bytes, ZSTD_cwksp_alloc_phase_e phase)
+{
+    void* alloc;
+    if (ZSTD_isError(ZSTD_cwksp_internal_advance_phase(ws, phase)) || bytes == 0) {
+        return NULL;
+    }
+
+#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
+    /* over-reserve space */
+    bytes += 2 * ZSTD_CWKSP_ASAN_REDZONE_SIZE;
+#endif
+
+    alloc = ZSTD_cwksp_reserve_internal_buffer_space(ws, bytes);
+
+#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
     /* Move alloc so there's ZSTD_CWKSP_ASAN_REDZONE_SIZE unused space on
      * either size. */
-    alloc = (BYTE *)alloc + ZSTD_CWKSP_ASAN_REDZONE_SIZE;
-    __asan_unpoison_memory_region(alloc, bytes);
+    if (alloc) {
+        alloc = (BYTE *)alloc + ZSTD_CWKSP_ASAN_REDZONE_SIZE;
+        if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) {
+            /* We need to keep the redzone poisoned while unpoisoning the bytes that
+             * are actually allocated. */
+            __asan_unpoison_memory_region(alloc, bytes - 2 * ZSTD_CWKSP_ASAN_REDZONE_SIZE);
+        }
+    }
 #endif
 
     return alloc;
@@ -260,33 +385,78 @@ MEM_STATIC void* ZSTD_cwksp_reserve_internal(
 /**
  * Reserves and returns unaligned memory.
  */
-MEM_STATIC BYTE* ZSTD_cwksp_reserve_buffer(ZSTD_cwksp* ws, size_t bytes) {
+MEM_STATIC BYTE* ZSTD_cwksp_reserve_buffer(ZSTD_cwksp* ws, size_t bytes)
+{
     return (BYTE*)ZSTD_cwksp_reserve_internal(ws, bytes, ZSTD_cwksp_alloc_buffers);
 }
 
 /**
- * Reserves and returns memory sized on and aligned on sizeof(unsigned).
+ * Reserves and returns memory sized on and aligned on ZSTD_CWKSP_ALIGNMENT_BYTES (64 bytes).
+ * This memory has been initialized at least once in the past.
+ * This doesn't mean it has been initialized this time, and it might contain data from previous
+ * operations.
+ * The main usage is for algorithms that might need read access into uninitialized memory.
+ * The algorithm must maintain safety under these conditions and must make sure it doesn't
+ * leak any of the past data (directly or in side channels).
  */
-MEM_STATIC void* ZSTD_cwksp_reserve_aligned(ZSTD_cwksp* ws, size_t bytes) {
-    assert((bytes & (sizeof(U32)-1)) == 0);
-    return ZSTD_cwksp_reserve_internal(ws, ZSTD_cwksp_align(bytes, sizeof(U32)), ZSTD_cwksp_alloc_aligned);
+MEM_STATIC void* ZSTD_cwksp_reserve_aligned_init_once(ZSTD_cwksp* ws, size_t bytes)
+{
+    size_t const alignedBytes = ZSTD_cwksp_align(bytes, ZSTD_CWKSP_ALIGNMENT_BYTES);
+    void* ptr = ZSTD_cwksp_reserve_internal(ws, alignedBytes, ZSTD_cwksp_alloc_aligned_init_once);
+    assert(((size_t)ptr & (ZSTD_CWKSP_ALIGNMENT_BYTES-1))== 0);
+    if(ptr && ptr < ws->initOnceStart) {
+        /* We assume the memory following the current allocation is either:
+         * 1. Not usable as initOnce memory (end of workspace)
+         * 2. Another initOnce buffer that has been allocated before (and so was previously memset)
+         * 3. An ASAN redzone, in which case we don't want to write on it
+         * For these reasons it should be fine to not explicitly zero every byte up to ws->initOnceStart.
+         * Note that we assume here that MSAN and ASAN cannot run in the same time. */
+        ZSTD_memset(ptr, 0, MIN((size_t)((U8*)ws->initOnceStart - (U8*)ptr), alignedBytes));
+        ws->initOnceStart = ptr;
+    }
+#if ZSTD_MEMORY_SANITIZER
+    assert(__msan_test_shadow(ptr, bytes) == -1);
+#endif
+    return ptr;
 }
 
 /**
- * Aligned on sizeof(unsigned). These buffers have the special property that
+ * Reserves and returns memory sized on and aligned on ZSTD_CWKSP_ALIGNMENT_BYTES (64 bytes).
+ */
+MEM_STATIC void* ZSTD_cwksp_reserve_aligned(ZSTD_cwksp* ws, size_t bytes)
+{
+    void* ptr = ZSTD_cwksp_reserve_internal(ws, ZSTD_cwksp_align(bytes, ZSTD_CWKSP_ALIGNMENT_BYTES),
+                                            ZSTD_cwksp_alloc_aligned);
+    assert(((size_t)ptr & (ZSTD_CWKSP_ALIGNMENT_BYTES-1))== 0);
+    return ptr;
+}
+
+/**
+ * Aligned on 64 bytes. These buffers have the special property that
  * their values remain constrained, allowing us to re-use them without
  * memset()-ing them.
  */
-MEM_STATIC void* ZSTD_cwksp_reserve_table(ZSTD_cwksp* ws, size_t bytes) {
-    const ZSTD_cwksp_alloc_phase_e phase = ZSTD_cwksp_alloc_aligned;
-    void* alloc = ws->tableEnd;
-    void* end = (BYTE *)alloc + bytes;
-    void* top = ws->allocStart;
+MEM_STATIC void* ZSTD_cwksp_reserve_table(ZSTD_cwksp* ws, size_t bytes)
+{
+    const ZSTD_cwksp_alloc_phase_e phase = ZSTD_cwksp_alloc_aligned_init_once;
+    void* alloc;
+    void* end;
+    void* top;
+
+    /* We can only start allocating tables after we are done reserving space for objects at the
+     * start of the workspace */
+    if(ws->phase < phase) {
+        if (ZSTD_isError(ZSTD_cwksp_internal_advance_phase(ws, phase))) {
+            return NULL;
+        }
+    }
+    alloc = ws->tableEnd;
+    end = (BYTE *)alloc + bytes;
+    top = ws->allocStart;
 
     DEBUGLOG(5, "cwksp: reserving %p table %zd bytes, %zd bytes remaining",
         alloc, bytes, ZSTD_cwksp_available_space(ws) - bytes);
     assert((bytes & (sizeof(U32)-1)) == 0);
-    ZSTD_cwksp_internal_advance_phase(ws, phase);
     ZSTD_cwksp_assert_internal_consistency(ws);
     assert(end <= top);
     if (end > top) {
@@ -296,35 +466,41 @@ MEM_STATIC void* ZSTD_cwksp_reserve_table(ZSTD_cwksp* ws, size_t bytes) {
     }
     ws->tableEnd = end;
 
-#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
-    __asan_unpoison_memory_region(alloc, bytes);
+#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
+    if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) {
+        __asan_unpoison_memory_region(alloc, bytes);
+    }
 #endif
 
+    assert((bytes & (ZSTD_CWKSP_ALIGNMENT_BYTES-1)) == 0);
+    assert(((size_t)alloc & (ZSTD_CWKSP_ALIGNMENT_BYTES-1))== 0);
     return alloc;
 }
 
 /**
  * Aligned on sizeof(void*).
+ * Note : should happen only once, at workspace first initialization
  */
-MEM_STATIC void* ZSTD_cwksp_reserve_object(ZSTD_cwksp* ws, size_t bytes) {
-    size_t roundedBytes = ZSTD_cwksp_align(bytes, sizeof(void*));
+MEM_STATIC void* ZSTD_cwksp_reserve_object(ZSTD_cwksp* ws, size_t bytes)
+{
+    size_t const roundedBytes = ZSTD_cwksp_align(bytes, sizeof(void*));
     void* alloc = ws->objectEnd;
     void* end = (BYTE*)alloc + roundedBytes;
 
-#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
+#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
     /* over-reserve space */
     end = (BYTE *)end + 2 * ZSTD_CWKSP_ASAN_REDZONE_SIZE;
 #endif
 
-    DEBUGLOG(5,
+    DEBUGLOG(4,
         "cwksp: reserving %p object %zd bytes (rounded to %zd), %zd bytes remaining",
         alloc, bytes, roundedBytes, ZSTD_cwksp_available_space(ws) - roundedBytes);
-    assert(((size_t)alloc & (sizeof(void*)-1)) == 0);
-    assert((bytes & (sizeof(void*)-1)) == 0);
+    assert((size_t)alloc % ZSTD_ALIGNOF(void*) == 0);
+    assert(bytes % ZSTD_ALIGNOF(void*) == 0);
     ZSTD_cwksp_assert_internal_consistency(ws);
     /* we must be in the first phase, no advance is possible */
     if (ws->phase != ZSTD_cwksp_alloc_objects || end > ws->workspaceEnd) {
-        DEBUGLOG(4, "cwksp: object alloc failed!");
+        DEBUGLOG(3, "cwksp: object alloc failed!");
         ws->allocFailed = 1;
         return NULL;
     }
@@ -332,27 +508,38 @@ MEM_STATIC void* ZSTD_cwksp_reserve_object(ZSTD_cwksp* ws, size_t bytes) {
     ws->tableEnd = end;
     ws->tableValidEnd = end;
 
-#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
+#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
     /* Move alloc so there's ZSTD_CWKSP_ASAN_REDZONE_SIZE unused space on
      * either size. */
-    alloc = (BYTE *)alloc + ZSTD_CWKSP_ASAN_REDZONE_SIZE;
-    __asan_unpoison_memory_region(alloc, bytes);
+    alloc = (BYTE*)alloc + ZSTD_CWKSP_ASAN_REDZONE_SIZE;
+    if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) {
+        __asan_unpoison_memory_region(alloc, bytes);
+    }
 #endif
 
     return alloc;
 }
 
-MEM_STATIC void ZSTD_cwksp_mark_tables_dirty(ZSTD_cwksp* ws) {
+MEM_STATIC void ZSTD_cwksp_mark_tables_dirty(ZSTD_cwksp* ws)
+{
     DEBUGLOG(4, "cwksp: ZSTD_cwksp_mark_tables_dirty");
 
-#if defined (MEMORY_SANITIZER) && !defined (ZSTD_MSAN_DONT_POISON_WORKSPACE)
+#if ZSTD_MEMORY_SANITIZER && !defined (ZSTD_MSAN_DONT_POISON_WORKSPACE)
     /* To validate that the table re-use logic is sound, and that we don't
      * access table space that we haven't cleaned, we re-"poison" the table
-     * space every time we mark it dirty. */
+     * space every time we mark it dirty.
+     * Since tableValidEnd space and initOnce space may overlap we don't poison
+     * the initOnce portion as it break its promise. This means that this poisoning
+     * check isn't always applied fully. */
     {
         size_t size = (BYTE*)ws->tableValidEnd - (BYTE*)ws->objectEnd;
         assert(__msan_test_shadow(ws->objectEnd, size) == -1);
-        __msan_poison(ws->objectEnd, size);
+        if((BYTE*)ws->tableValidEnd < (BYTE*)ws->initOnceStart) {
+            __msan_poison(ws->objectEnd, size);
+        } else {
+            assert(ws->initOnceStart >= ws->objectEnd);
+            __msan_poison(ws->objectEnd, (BYTE*)ws->initOnceStart - (BYTE*)ws->objectEnd);
+        }
     }
 #endif
 
@@ -380,7 +567,7 @@ MEM_STATIC void ZSTD_cwksp_clean_tables(ZSTD_cwksp* ws) {
     assert(ws->tableValidEnd >= ws->objectEnd);
     assert(ws->tableValidEnd <= ws->allocStart);
     if (ws->tableValidEnd < ws->tableEnd) {
-        memset(ws->tableValidEnd, 0, (BYTE*)ws->tableEnd - (BYTE*)ws->tableValidEnd);
+        ZSTD_memset(ws->tableValidEnd, 0, (size_t)((BYTE*)ws->tableEnd - (BYTE*)ws->tableValidEnd));
     }
     ZSTD_cwksp_mark_tables_clean(ws);
 }
@@ -392,8 +579,12 @@ MEM_STATIC void ZSTD_cwksp_clean_tables(ZSTD_cwksp* ws) {
 MEM_STATIC void ZSTD_cwksp_clear_tables(ZSTD_cwksp* ws) {
     DEBUGLOG(4, "cwksp: clearing tables!");
 
-#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
-    {
+#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
+    /* We don't do this when the workspace is statically allocated, because
+     * when that is the case, we have no capability to hook into the end of the
+     * workspace's lifecycle to unpoison the memory.
+     */
+    if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) {
         size_t size = (BYTE*)ws->tableValidEnd - (BYTE*)ws->objectEnd;
         __asan_poison_memory_region(ws->objectEnd, size);
     }
@@ -410,29 +601,36 @@ MEM_STATIC void ZSTD_cwksp_clear_tables(ZSTD_cwksp* ws) {
 MEM_STATIC void ZSTD_cwksp_clear(ZSTD_cwksp* ws) {
     DEBUGLOG(4, "cwksp: clearing!");
 
-#if defined (MEMORY_SANITIZER) && !defined (ZSTD_MSAN_DONT_POISON_WORKSPACE)
+#if ZSTD_MEMORY_SANITIZER && !defined (ZSTD_MSAN_DONT_POISON_WORKSPACE)
     /* To validate that the context re-use logic is sound, and that we don't
      * access stuff that this compression hasn't initialized, we re-"poison"
-     * the workspace (or at least the non-static, non-table parts of it)
-     * every time we start a new compression. */
+     * the workspace except for the areas in which we expect memory re-use
+     * without initialization (objects, valid tables area and init once
+     * memory). */
     {
-        size_t size = (BYTE*)ws->workspaceEnd - (BYTE*)ws->tableValidEnd;
-        __msan_poison(ws->tableValidEnd, size);
+        if((BYTE*)ws->tableValidEnd < (BYTE*)ws->initOnceStart) {
+            size_t size = (BYTE*)ws->initOnceStart - (BYTE*)ws->tableValidEnd;
+            __msan_poison(ws->tableValidEnd, size);
+        }
     }
 #endif
 
-#if defined (ADDRESS_SANITIZER) && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
-    {
+#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
+    /* We don't do this when the workspace is statically allocated, because
+     * when that is the case, we have no capability to hook into the end of the
+     * workspace's lifecycle to unpoison the memory.
+     */
+    if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) {
         size_t size = (BYTE*)ws->workspaceEnd - (BYTE*)ws->objectEnd;
         __asan_poison_memory_region(ws->objectEnd, size);
     }
 #endif
 
     ws->tableEnd = ws->objectEnd;
-    ws->allocStart = ws->workspaceEnd;
+    ws->allocStart = ZSTD_cwksp_initialAllocStart(ws);
     ws->allocFailed = 0;
-    if (ws->phase > ZSTD_cwksp_alloc_buffers) {
-        ws->phase = ZSTD_cwksp_alloc_buffers;
+    if (ws->phase > ZSTD_cwksp_alloc_aligned_init_once) {
+        ws->phase = ZSTD_cwksp_alloc_aligned_init_once;
     }
     ZSTD_cwksp_assert_internal_consistency(ws);
 }
@@ -442,47 +640,54 @@ MEM_STATIC void ZSTD_cwksp_clear(ZSTD_cwksp* ws) {
  * Any existing values in the workspace are ignored (the previously managed
  * buffer, if present, must be separately freed).
  */
-MEM_STATIC void ZSTD_cwksp_init(ZSTD_cwksp* ws, void* start, size_t size) {
+MEM_STATIC void ZSTD_cwksp_init(ZSTD_cwksp* ws, void* start, size_t size, ZSTD_cwksp_static_alloc_e isStatic) {
     DEBUGLOG(4, "cwksp: init'ing workspace with %zd bytes", size);
     assert(((size_t)start & (sizeof(void*)-1)) == 0); /* ensure correct alignment */
     ws->workspace = start;
     ws->workspaceEnd = (BYTE*)start + size;
     ws->objectEnd = ws->workspace;
     ws->tableValidEnd = ws->objectEnd;
+    ws->initOnceStart = ZSTD_cwksp_initialAllocStart(ws);
     ws->phase = ZSTD_cwksp_alloc_objects;
+    ws->isStatic = isStatic;
     ZSTD_cwksp_clear(ws);
     ws->workspaceOversizedDuration = 0;
     ZSTD_cwksp_assert_internal_consistency(ws);
 }
 
 MEM_STATIC size_t ZSTD_cwksp_create(ZSTD_cwksp* ws, size_t size, ZSTD_customMem customMem) {
-    void* workspace = ZSTD_malloc(size, customMem);
+    void* workspace = ZSTD_customMalloc(size, customMem);
     DEBUGLOG(4, "cwksp: creating new workspace with %zd bytes", size);
     RETURN_ERROR_IF(workspace == NULL, memory_allocation, "NULL pointer!");
-    ZSTD_cwksp_init(ws, workspace, size);
+    ZSTD_cwksp_init(ws, workspace, size, ZSTD_cwksp_dynamic_alloc);
     return 0;
 }
 
 MEM_STATIC void ZSTD_cwksp_free(ZSTD_cwksp* ws, ZSTD_customMem customMem) {
     void *ptr = ws->workspace;
     DEBUGLOG(4, "cwksp: freeing workspace");
-    memset(ws, 0, sizeof(ZSTD_cwksp));
-    ZSTD_free(ptr, customMem);
+    ZSTD_memset(ws, 0, sizeof(ZSTD_cwksp));
+    ZSTD_customFree(ptr, customMem);
 }
 
 /**
  * Moves the management of a workspace from one cwksp to another. The src cwksp
- * is left in an invalid state (src must be re-init()'ed before its used again).
+ * is left in an invalid state (src must be re-init()'ed before it's used again).
  */
 MEM_STATIC void ZSTD_cwksp_move(ZSTD_cwksp* dst, ZSTD_cwksp* src) {
     *dst = *src;
-    memset(src, 0, sizeof(ZSTD_cwksp));
+    ZSTD_memset(src, 0, sizeof(ZSTD_cwksp));
 }
 
 MEM_STATIC size_t ZSTD_cwksp_sizeof(const ZSTD_cwksp* ws) {
     return (size_t)((BYTE*)ws->workspaceEnd - (BYTE*)ws->workspace);
 }
 
+MEM_STATIC size_t ZSTD_cwksp_used(const ZSTD_cwksp* ws) {
+    return (size_t)((BYTE*)ws->tableEnd - (BYTE*)ws->workspace)
+         + (size_t)((BYTE*)ws->workspaceEnd - (BYTE*)ws->allocStart);
+}
+
 MEM_STATIC int ZSTD_cwksp_reserve_failed(const ZSTD_cwksp* ws) {
     return ws->allocFailed;
 }
@@ -491,6 +696,18 @@ MEM_STATIC int ZSTD_cwksp_reserve_failed(const ZSTD_cwksp* ws) {
 *  Functions Checking Free Space
 ***************************************/
 
+/* ZSTD_alignmentSpaceWithinBounds() :
+ * Returns if the estimated space needed for a wksp is within an acceptable limit of the
+ * actual amount of space used.
+ */
+MEM_STATIC int ZSTD_cwksp_estimated_space_within_bounds(const ZSTD_cwksp *const ws, size_t const estimatedSpace) {
+    /* We have an alignment space between objects and tables between tables and buffers, so we can have up to twice
+     * the alignment bytes difference between estimation and actual usage */
+    return (estimatedSpace - ZSTD_cwksp_slack_space_required()) <= ZSTD_cwksp_used(ws) &&
+           ZSTD_cwksp_used(ws) <= estimatedSpace;
+}
+
+
 MEM_STATIC size_t ZSTD_cwksp_available_space(ZSTD_cwksp* ws) {
     return (size_t)((BYTE*)ws->allocStart - (BYTE*)ws->tableEnd);
 }

src/borg/algorithms/zstd/lib/compress/zstd_double_fast.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -11,8 +11,43 @@
 #include "zstd_compress_internal.h"
 #include "zstd_double_fast.h"
 
+static void ZSTD_fillDoubleHashTableForCDict(ZSTD_matchState_t* ms,
+                              void const* end, ZSTD_dictTableLoadMethod_e dtlm)
+{
+    const ZSTD_compressionParameters* const cParams = &ms->cParams;
+    U32* const hashLarge = ms->hashTable;
+    U32  const hBitsL = cParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS;
+    U32  const mls = cParams->minMatch;
+    U32* const hashSmall = ms->chainTable;
+    U32  const hBitsS = cParams->chainLog + ZSTD_SHORT_CACHE_TAG_BITS;
+    const BYTE* const base = ms->window.base;
+    const BYTE* ip = base + ms->nextToUpdate;
+    const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE;
+    const U32 fastHashFillStep = 3;
 
-void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms,
+    /* Always insert every fastHashFillStep position into the hash tables.
+     * Insert the other positions into the large hash table if their entry
+     * is empty.
+     */
+    for (; ip + fastHashFillStep - 1 <= iend; ip += fastHashFillStep) {
+        U32 const curr = (U32)(ip - base);
+        U32 i;
+        for (i = 0; i < fastHashFillStep; ++i) {
+            size_t const smHashAndTag = ZSTD_hashPtr(ip + i, hBitsS, mls);
+            size_t const lgHashAndTag = ZSTD_hashPtr(ip + i, hBitsL, 8);
+            if (i == 0) {
+                ZSTD_writeTaggedIndex(hashSmall, smHashAndTag, curr + i);
+            }
+            if (i == 0 || hashLarge[lgHashAndTag >> ZSTD_SHORT_CACHE_TAG_BITS] == 0) {
+                ZSTD_writeTaggedIndex(hashLarge, lgHashAndTag, curr + i);
+            }
+            /* Only load extra positions for ZSTD_dtlm_full */
+            if (dtlm == ZSTD_dtlm_fast)
+                break;
+    }   }
+}
+
+static void ZSTD_fillDoubleHashTableForCCtx(ZSTD_matchState_t* ms,
                               void const* end, ZSTD_dictTableLoadMethod_e dtlm)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
@@ -31,27 +66,249 @@ void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms,
      * is empty.
      */
     for (; ip + fastHashFillStep - 1 <= iend; ip += fastHashFillStep) {
-        U32 const current = (U32)(ip - base);
+        U32 const curr = (U32)(ip - base);
         U32 i;
         for (i = 0; i < fastHashFillStep; ++i) {
             size_t const smHash = ZSTD_hashPtr(ip + i, hBitsS, mls);
             size_t const lgHash = ZSTD_hashPtr(ip + i, hBitsL, 8);
             if (i == 0)
-                hashSmall[smHash] = current + i;
+                hashSmall[smHash] = curr + i;
             if (i == 0 || hashLarge[lgHash] == 0)
-                hashLarge[lgHash] = current + i;
+                hashLarge[lgHash] = curr + i;
             /* Only load extra positions for ZSTD_dtlm_full */
             if (dtlm == ZSTD_dtlm_fast)
                 break;
-    }   }
+        }   }
+}
+
+void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms,
+                        const void* const end,
+                        ZSTD_dictTableLoadMethod_e dtlm,
+                        ZSTD_tableFillPurpose_e tfp)
+{
+    if (tfp == ZSTD_tfp_forCDict) {
+        ZSTD_fillDoubleHashTableForCDict(ms, end, dtlm);
+    } else {
+        ZSTD_fillDoubleHashTableForCCtx(ms, end, dtlm);
+    }
+}
+
+
+FORCE_INLINE_TEMPLATE
+size_t ZSTD_compressBlock_doubleFast_noDict_generic(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize, U32 const mls /* template */)
+{
+    ZSTD_compressionParameters const* cParams = &ms->cParams;
+    U32* const hashLong = ms->hashTable;
+    const U32 hBitsL = cParams->hashLog;
+    U32* const hashSmall = ms->chainTable;
+    const U32 hBitsS = cParams->chainLog;
+    const BYTE* const base = ms->window.base;
+    const BYTE* const istart = (const BYTE*)src;
+    const BYTE* anchor = istart;
+    const U32 endIndex = (U32)((size_t)(istart - base) + srcSize);
+    /* presumes that, if there is a dictionary, it must be using Attach mode */
+    const U32 prefixLowestIndex = ZSTD_getLowestPrefixIndex(ms, endIndex, cParams->windowLog);
+    const BYTE* const prefixLowest = base + prefixLowestIndex;
+    const BYTE* const iend = istart + srcSize;
+    const BYTE* const ilimit = iend - HASH_READ_SIZE;
+    U32 offset_1=rep[0], offset_2=rep[1];
+    U32 offsetSaved1 = 0, offsetSaved2 = 0;
+
+    size_t mLength;
+    U32 offset;
+    U32 curr;
+
+    /* how many positions to search before increasing step size */
+    const size_t kStepIncr = 1 << kSearchStrength;
+    /* the position at which to increment the step size if no match is found */
+    const BYTE* nextStep;
+    size_t step; /* the current step size */
+
+    size_t hl0; /* the long hash at ip */
+    size_t hl1; /* the long hash at ip1 */
+
+    U32 idxl0; /* the long match index for ip */
+    U32 idxl1; /* the long match index for ip1 */
+
+    const BYTE* matchl0; /* the long match for ip */
+    const BYTE* matchs0; /* the short match for ip */
+    const BYTE* matchl1; /* the long match for ip1 */
+
+    const BYTE* ip = istart; /* the current position */
+    const BYTE* ip1; /* the next position */
+
+    DEBUGLOG(5, "ZSTD_compressBlock_doubleFast_noDict_generic");
+
+    /* init */
+    ip += ((ip - prefixLowest) == 0);
+    {
+        U32 const current = (U32)(ip - base);
+        U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, current, cParams->windowLog);
+        U32 const maxRep = current - windowLow;
+        if (offset_2 > maxRep) offsetSaved2 = offset_2, offset_2 = 0;
+        if (offset_1 > maxRep) offsetSaved1 = offset_1, offset_1 = 0;
+    }
+
+    /* Outer Loop: one iteration per match found and stored */
+    while (1) {
+        step = 1;
+        nextStep = ip + kStepIncr;
+        ip1 = ip + step;
+
+        if (ip1 > ilimit) {
+            goto _cleanup;
+        }
+
+        hl0 = ZSTD_hashPtr(ip, hBitsL, 8);
+        idxl0 = hashLong[hl0];
+        matchl0 = base + idxl0;
+
+        /* Inner Loop: one iteration per search / position */
+        do {
+            const size_t hs0 = ZSTD_hashPtr(ip, hBitsS, mls);
+            const U32 idxs0 = hashSmall[hs0];
+            curr = (U32)(ip-base);
+            matchs0 = base + idxs0;
+
+            hashLong[hl0] = hashSmall[hs0] = curr;   /* update hash tables */
+
+            /* check noDict repcode */
+            if ((offset_1 > 0) & (MEM_read32(ip+1-offset_1) == MEM_read32(ip+1))) {
+                mLength = ZSTD_count(ip+1+4, ip+1+4-offset_1, iend) + 4;
+                ip++;
+                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength);
+                goto _match_stored;
+            }
+
+            hl1 = ZSTD_hashPtr(ip1, hBitsL, 8);
+
+            if (idxl0 > prefixLowestIndex) {
+                /* check prefix long match */
+                if (MEM_read64(matchl0) == MEM_read64(ip)) {
+                    mLength = ZSTD_count(ip+8, matchl0+8, iend) + 8;
+                    offset = (U32)(ip-matchl0);
+                    while (((ip>anchor) & (matchl0>prefixLowest)) && (ip[-1] == matchl0[-1])) { ip--; matchl0--; mLength++; } /* catch up */
+                    goto _match_found;
+                }
+            }
+
+            idxl1 = hashLong[hl1];
+            matchl1 = base + idxl1;
+
+            if (idxs0 > prefixLowestIndex) {
+                /* check prefix short match */
+                if (MEM_read32(matchs0) == MEM_read32(ip)) {
+                    goto _search_next_long;
+                }
+            }
+
+            if (ip1 >= nextStep) {
+                PREFETCH_L1(ip1 + 64);
+                PREFETCH_L1(ip1 + 128);
+                step++;
+                nextStep += kStepIncr;
+            }
+            ip = ip1;
+            ip1 += step;
+
+            hl0 = hl1;
+            idxl0 = idxl1;
+            matchl0 = matchl1;
+    #if defined(__aarch64__)
+            PREFETCH_L1(ip+256);
+    #endif
+        } while (ip1 <= ilimit);
+
+_cleanup:
+        /* If offset_1 started invalid (offsetSaved1 != 0) and became valid (offset_1 != 0),
+         * rotate saved offsets. See comment in ZSTD_compressBlock_fast_noDict for more context. */
+        offsetSaved2 = ((offsetSaved1 != 0) && (offset_1 != 0)) ? offsetSaved1 : offsetSaved2;
+
+        /* save reps for next block */
+        rep[0] = offset_1 ? offset_1 : offsetSaved1;
+        rep[1] = offset_2 ? offset_2 : offsetSaved2;
+
+        /* Return the last literals size */
+        return (size_t)(iend - anchor);
+
+_search_next_long:
+
+        /* check prefix long +1 match */
+        if (idxl1 > prefixLowestIndex) {
+            if (MEM_read64(matchl1) == MEM_read64(ip1)) {
+                ip = ip1;
+                mLength = ZSTD_count(ip+8, matchl1+8, iend) + 8;
+                offset = (U32)(ip-matchl1);
+                while (((ip>anchor) & (matchl1>prefixLowest)) && (ip[-1] == matchl1[-1])) { ip--; matchl1--; mLength++; } /* catch up */
+                goto _match_found;
+            }
+        }
+
+        /* if no long +1 match, explore the short match we found */
+        mLength = ZSTD_count(ip+4, matchs0+4, iend) + 4;
+        offset = (U32)(ip - matchs0);
+        while (((ip>anchor) & (matchs0>prefixLowest)) && (ip[-1] == matchs0[-1])) { ip--; matchs0--; mLength++; } /* catch up */
+
+        /* fall-through */
+
+_match_found: /* requires ip, offset, mLength */
+        offset_2 = offset_1;
+        offset_1 = offset;
+
+        if (step < 4) {
+            /* It is unsafe to write this value back to the hashtable when ip1 is
+             * greater than or equal to the new ip we will have after we're done
+             * processing this match. Rather than perform that test directly
+             * (ip1 >= ip + mLength), which costs speed in practice, we do a simpler
+             * more predictable test. The minmatch even if we take a short match is
+             * 4 bytes, so as long as step, the distance between ip and ip1
+             * (initially) is less than 4, we know ip1 < new ip. */
+            hashLong[hl1] = (U32)(ip1 - base);
+        }
+
+        ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
+
+_match_stored:
+        /* match found */
+        ip += mLength;
+        anchor = ip;
+
+        if (ip <= ilimit) {
+            /* Complementary insertion */
+            /* done after iLimit test, as candidates could be > iend-8 */
+            {   U32 const indexToInsert = curr+2;
+                hashLong[ZSTD_hashPtr(base+indexToInsert, hBitsL, 8)] = indexToInsert;
+                hashLong[ZSTD_hashPtr(ip-2, hBitsL, 8)] = (U32)(ip-2-base);
+                hashSmall[ZSTD_hashPtr(base+indexToInsert, hBitsS, mls)] = indexToInsert;
+                hashSmall[ZSTD_hashPtr(ip-1, hBitsS, mls)] = (U32)(ip-1-base);
+            }
+
+            /* check immediate repcode */
+            while ( (ip <= ilimit)
+                 && ( (offset_2>0)
+                    & (MEM_read32(ip) == MEM_read32(ip - offset_2)) )) {
+                /* store sequence */
+                size_t const rLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4;
+                U32 const tmpOff = offset_2; offset_2 = offset_1; offset_1 = tmpOff;  /* swap offset_2 <=> offset_1 */
+                hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = (U32)(ip-base);
+                hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = (U32)(ip-base);
+                ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, rLength);
+                ip += rLength;
+                anchor = ip;
+                continue;   /* faster when present ... (?) */
+            }
+        }
+    }
 }
 
 
 FORCE_INLINE_TEMPLATE
-size_t ZSTD_compressBlock_doubleFast_generic(
+size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize,
-        U32 const mls /* template */, ZSTD_dictMode_e const dictMode)
+        U32 const mls /* template */)
 {
     ZSTD_compressionParameters const* cParams = &ms->cParams;
     U32* const hashLong = ms->hashTable;
@@ -69,57 +326,39 @@ size_t ZSTD_compressBlock_doubleFast_generic(
     const BYTE* const iend = istart + srcSize;
     const BYTE* const ilimit = iend - HASH_READ_SIZE;
     U32 offset_1=rep[0], offset_2=rep[1];
-    U32 offsetSaved = 0;
 
     const ZSTD_matchState_t* const dms = ms->dictMatchState;
-    const ZSTD_compressionParameters* const dictCParams =
-                                     dictMode == ZSTD_dictMatchState ?
-                                     &dms->cParams : NULL;
-    const U32* const dictHashLong  = dictMode == ZSTD_dictMatchState ?
-                                     dms->hashTable : NULL;
-    const U32* const dictHashSmall = dictMode == ZSTD_dictMatchState ?
-                                     dms->chainTable : NULL;
-    const U32 dictStartIndex       = dictMode == ZSTD_dictMatchState ?
-                                     dms->window.dictLimit : 0;
-    const BYTE* const dictBase     = dictMode == ZSTD_dictMatchState ?
-                                     dms->window.base : NULL;
-    const BYTE* const dictStart    = dictMode == ZSTD_dictMatchState ?
-                                     dictBase + dictStartIndex : NULL;
-    const BYTE* const dictEnd      = dictMode == ZSTD_dictMatchState ?
-                                     dms->window.nextSrc : NULL;
-    const U32 dictIndexDelta       = dictMode == ZSTD_dictMatchState ?
-                                     prefixLowestIndex - (U32)(dictEnd - dictBase) :
-                                     0;
-    const U32 dictHBitsL           = dictMode == ZSTD_dictMatchState ?
-                                     dictCParams->hashLog : hBitsL;
-    const U32 dictHBitsS           = dictMode == ZSTD_dictMatchState ?
-                                     dictCParams->chainLog : hBitsS;
+    const ZSTD_compressionParameters* const dictCParams = &dms->cParams;
+    const U32* const dictHashLong  = dms->hashTable;
+    const U32* const dictHashSmall = dms->chainTable;
+    const U32 dictStartIndex       = dms->window.dictLimit;
+    const BYTE* const dictBase     = dms->window.base;
+    const BYTE* const dictStart    = dictBase + dictStartIndex;
+    const BYTE* const dictEnd      = dms->window.nextSrc;
+    const U32 dictIndexDelta       = prefixLowestIndex - (U32)(dictEnd - dictBase);
+    const U32 dictHBitsL           = dictCParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS;
+    const U32 dictHBitsS           = dictCParams->chainLog + ZSTD_SHORT_CACHE_TAG_BITS;
     const U32 dictAndPrefixLength  = (U32)((ip - prefixLowest) + (dictEnd - dictStart));
 
-    DEBUGLOG(5, "ZSTD_compressBlock_doubleFast_generic");
-
-    assert(dictMode == ZSTD_noDict || dictMode == ZSTD_dictMatchState);
+    DEBUGLOG(5, "ZSTD_compressBlock_doubleFast_dictMatchState_generic");
 
     /* if a dictionary is attached, it must be within window range */
-    if (dictMode == ZSTD_dictMatchState) {
-        assert(ms->window.dictLimit + (1U << cParams->windowLog) >= endIndex);
+    assert(ms->window.dictLimit + (1U << cParams->windowLog) >= endIndex);
+
+    if (ms->prefetchCDictTables) {
+        size_t const hashTableBytes = (((size_t)1) << dictCParams->hashLog) * sizeof(U32);
+        size_t const chainTableBytes = (((size_t)1) << dictCParams->chainLog) * sizeof(U32);
+        PREFETCH_AREA(dictHashLong, hashTableBytes)
+        PREFETCH_AREA(dictHashSmall, chainTableBytes)
     }
 
     /* init */
     ip += (dictAndPrefixLength == 0);
-    if (dictMode == ZSTD_noDict) {
-        U32 const current = (U32)(ip - base);
-        U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, current, cParams->windowLog);
-        U32 const maxRep = current - windowLow;
-        if (offset_2 > maxRep) offsetSaved = offset_2, offset_2 = 0;
-        if (offset_1 > maxRep) offsetSaved = offset_1, offset_1 = 0;
-    }
-    if (dictMode == ZSTD_dictMatchState) {
-        /* dictMatchState repCode checks don't currently handle repCode == 0
-         * disabling. */
-        assert(offset_1 <= dictAndPrefixLength);
-        assert(offset_2 <= dictAndPrefixLength);
-    }
+
+    /* dictMatchState repCode checks don't currently handle repCode == 0
+     * disabling. */
+    assert(offset_1 <= dictAndPrefixLength);
+    assert(offset_2 <= dictAndPrefixLength);
 
     /* Main Search Loop */
     while (ip < ilimit) {   /* < instead of <=, because repcode check at (ip+1) */
@@ -127,37 +366,30 @@ size_t ZSTD_compressBlock_doubleFast_generic(
         U32 offset;
         size_t const h2 = ZSTD_hashPtr(ip, hBitsL, 8);
         size_t const h = ZSTD_hashPtr(ip, hBitsS, mls);
-        size_t const dictHL = ZSTD_hashPtr(ip, dictHBitsL, 8);
-        size_t const dictHS = ZSTD_hashPtr(ip, dictHBitsS, mls);
-        U32 const current = (U32)(ip-base);
+        size_t const dictHashAndTagL = ZSTD_hashPtr(ip, dictHBitsL, 8);
+        size_t const dictHashAndTagS = ZSTD_hashPtr(ip, dictHBitsS, mls);
+        U32 const dictMatchIndexAndTagL = dictHashLong[dictHashAndTagL >> ZSTD_SHORT_CACHE_TAG_BITS];
+        U32 const dictMatchIndexAndTagS = dictHashSmall[dictHashAndTagS >> ZSTD_SHORT_CACHE_TAG_BITS];
+        int const dictTagsMatchL = ZSTD_comparePackedTags(dictMatchIndexAndTagL, dictHashAndTagL);
+        int const dictTagsMatchS = ZSTD_comparePackedTags(dictMatchIndexAndTagS, dictHashAndTagS);
+        U32 const curr = (U32)(ip-base);
         U32 const matchIndexL = hashLong[h2];
         U32 matchIndexS = hashSmall[h];
         const BYTE* matchLong = base + matchIndexL;
         const BYTE* match = base + matchIndexS;
-        const U32 repIndex = current + 1 - offset_1;
-        const BYTE* repMatch = (dictMode == ZSTD_dictMatchState
-                            && repIndex < prefixLowestIndex) ?
+        const U32 repIndex = curr + 1 - offset_1;
+        const BYTE* repMatch = (repIndex < prefixLowestIndex) ?
                                dictBase + (repIndex - dictIndexDelta) :
                                base + repIndex;
-        hashLong[h2] = hashSmall[h] = current;   /* update hash tables */
+        hashLong[h2] = hashSmall[h] = curr;   /* update hash tables */
 
-        /* check dictMatchState repcode */
-        if (dictMode == ZSTD_dictMatchState
-            && ((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
+        /* check repcode */
+        if (((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
             && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
             const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
             mLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
             ip++;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, 0, mLength-MINMATCH);
-            goto _match_stored;
-        }
-
-        /* check noDict repcode */
-        if ( dictMode == ZSTD_noDict
-          && ((offset_1 > 0) & (MEM_read32(ip+1-offset_1) == MEM_read32(ip+1)))) {
-            mLength = ZSTD_count(ip+1+4, ip+1+4-offset_1, iend) + 4;
-            ip++;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, 0, mLength-MINMATCH);
+            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength);
             goto _match_stored;
         }
 
@@ -169,15 +401,15 @@ size_t ZSTD_compressBlock_doubleFast_generic(
                 while (((ip>anchor) & (matchLong>prefixLowest)) && (ip[-1] == matchLong[-1])) { ip--; matchLong--; mLength++; } /* catch up */
                 goto _match_found;
             }
-        } else if (dictMode == ZSTD_dictMatchState) {
+        } else if (dictTagsMatchL) {
             /* check dictMatchState long match */
-            U32 const dictMatchIndexL = dictHashLong[dictHL];
+            U32 const dictMatchIndexL = dictMatchIndexAndTagL >> ZSTD_SHORT_CACHE_TAG_BITS;
             const BYTE* dictMatchL = dictBase + dictMatchIndexL;
             assert(dictMatchL < dictEnd);
 
             if (dictMatchL > dictStart && MEM_read64(dictMatchL) == MEM_read64(ip)) {
                 mLength = ZSTD_count_2segments(ip+8, dictMatchL+8, iend, dictEnd, prefixLowest) + 8;
-                offset = (U32)(current - dictMatchIndexL - dictIndexDelta);
+                offset = (U32)(curr - dictMatchIndexL - dictIndexDelta);
                 while (((ip>anchor) & (dictMatchL>dictStart)) && (ip[-1] == dictMatchL[-1])) { ip--; dictMatchL--; mLength++; } /* catch up */
                 goto _match_found;
         }   }
@@ -187,9 +419,9 @@ size_t ZSTD_compressBlock_doubleFast_generic(
             if (MEM_read32(match) == MEM_read32(ip)) {
                 goto _search_next_long;
             }
-        } else if (dictMode == ZSTD_dictMatchState) {
+        } else if (dictTagsMatchS) {
             /* check dictMatchState short match */
-            U32 const dictMatchIndexS = dictHashSmall[dictHS];
+            U32 const dictMatchIndexS = dictMatchIndexAndTagS >> ZSTD_SHORT_CACHE_TAG_BITS;
             match = dictBase + dictMatchIndexS;
             matchIndexS = dictMatchIndexS + dictIndexDelta;
 
@@ -204,12 +436,13 @@ size_t ZSTD_compressBlock_doubleFast_generic(
         continue;
 
 _search_next_long:
-
         {   size_t const hl3 = ZSTD_hashPtr(ip+1, hBitsL, 8);
-            size_t const dictHLNext = ZSTD_hashPtr(ip+1, dictHBitsL, 8);
+            size_t const dictHashAndTagL3 = ZSTD_hashPtr(ip+1, dictHBitsL, 8);
             U32 const matchIndexL3 = hashLong[hl3];
+            U32 const dictMatchIndexAndTagL3 = dictHashLong[dictHashAndTagL3 >> ZSTD_SHORT_CACHE_TAG_BITS];
+            int const dictTagsMatchL3 = ZSTD_comparePackedTags(dictMatchIndexAndTagL3, dictHashAndTagL3);
             const BYTE* matchL3 = base + matchIndexL3;
-            hashLong[hl3] = current + 1;
+            hashLong[hl3] = curr + 1;
 
             /* check prefix long +1 match */
             if (matchIndexL3 > prefixLowestIndex) {
@@ -220,23 +453,23 @@ _search_next_long:
                     while (((ip>anchor) & (matchL3>prefixLowest)) && (ip[-1] == matchL3[-1])) { ip--; matchL3--; mLength++; } /* catch up */
                     goto _match_found;
                 }
-            } else if (dictMode == ZSTD_dictMatchState) {
+            } else if (dictTagsMatchL3) {
                 /* check dict long +1 match */
-                U32 const dictMatchIndexL3 = dictHashLong[dictHLNext];
+                U32 const dictMatchIndexL3 = dictMatchIndexAndTagL3 >> ZSTD_SHORT_CACHE_TAG_BITS;
                 const BYTE* dictMatchL3 = dictBase + dictMatchIndexL3;
                 assert(dictMatchL3 < dictEnd);
                 if (dictMatchL3 > dictStart && MEM_read64(dictMatchL3) == MEM_read64(ip+1)) {
                     mLength = ZSTD_count_2segments(ip+1+8, dictMatchL3+8, iend, dictEnd, prefixLowest) + 8;
                     ip++;
-                    offset = (U32)(current + 1 - dictMatchIndexL3 - dictIndexDelta);
+                    offset = (U32)(curr + 1 - dictMatchIndexL3 - dictIndexDelta);
                     while (((ip>anchor) & (dictMatchL3>dictStart)) && (ip[-1] == dictMatchL3[-1])) { ip--; dictMatchL3--; mLength++; } /* catch up */
                     goto _match_found;
         }   }   }
 
         /* if no long +1 match, explore the short match we found */
-        if (dictMode == ZSTD_dictMatchState && matchIndexS < prefixLowestIndex) {
+        if (matchIndexS < prefixLowestIndex) {
             mLength = ZSTD_count_2segments(ip+4, match+4, iend, dictEnd, prefixLowest) + 4;
-            offset = (U32)(current - matchIndexS);
+            offset = (U32)(curr - matchIndexS);
             while (((ip>anchor) & (match>dictStart)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; } /* catch up */
         } else {
             mLength = ZSTD_count(ip+4, match+4, iend) + 4;
@@ -244,13 +477,11 @@ _search_next_long:
             while (((ip>anchor) & (match>prefixLowest)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; } /* catch up */
         }
 
-        /* fall-through */
-
 _match_found:
         offset_2 = offset_1;
         offset_1 = offset;
 
-        ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH);
+        ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
 
 _match_stored:
         /* match found */
@@ -260,7 +491,7 @@ _match_stored:
         if (ip <= ilimit) {
             /* Complementary insertion */
             /* done after iLimit test, as candidates could be > iend-8 */
-            {   U32 const indexToInsert = current+2;
+            {   U32 const indexToInsert = curr+2;
                 hashLong[ZSTD_hashPtr(base+indexToInsert, hBitsL, 8)] = indexToInsert;
                 hashLong[ZSTD_hashPtr(ip-2, hBitsL, 8)] = (U32)(ip-2-base);
                 hashSmall[ZSTD_hashPtr(base+indexToInsert, hBitsS, mls)] = indexToInsert;
@@ -268,53 +499,55 @@ _match_stored:
             }
 
             /* check immediate repcode */
-            if (dictMode == ZSTD_dictMatchState) {
-                while (ip <= ilimit) {
-                    U32 const current2 = (U32)(ip-base);
-                    U32 const repIndex2 = current2 - offset_2;
-                    const BYTE* repMatch2 = dictMode == ZSTD_dictMatchState
-                        && repIndex2 < prefixLowestIndex ?
-                            dictBase + repIndex2 - dictIndexDelta :
-                            base + repIndex2;
-                    if ( ((U32)((prefixLowestIndex-1) - (U32)repIndex2) >= 3 /* intentional overflow */)
-                       && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
-                        const BYTE* const repEnd2 = repIndex2 < prefixLowestIndex ? dictEnd : iend;
-                        size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixLowest) + 4;
-                        U32 tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset;   /* swap offset_2 <=> offset_1 */
-                        ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, repLength2-MINMATCH);
-                        hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = current2;
-                        hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = current2;
-                        ip += repLength2;
-                        anchor = ip;
-                        continue;
-                    }
-                    break;
-            }   }
-
-            if (dictMode == ZSTD_noDict) {
-                while ( (ip <= ilimit)
-                     && ( (offset_2>0)
-                        & (MEM_read32(ip) == MEM_read32(ip - offset_2)) )) {
-                    /* store sequence */
-                    size_t const rLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4;
-                    U32 const tmpOff = offset_2; offset_2 = offset_1; offset_1 = tmpOff;  /* swap offset_2 <=> offset_1 */
-                    hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = (U32)(ip-base);
-                    hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = (U32)(ip-base);
-                    ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, rLength-MINMATCH);
-                    ip += rLength;
+            while (ip <= ilimit) {
+                U32 const current2 = (U32)(ip-base);
+                U32 const repIndex2 = current2 - offset_2;
+                const BYTE* repMatch2 = repIndex2 < prefixLowestIndex ?
+                        dictBase + repIndex2 - dictIndexDelta :
+                        base + repIndex2;
+                if ( ((U32)((prefixLowestIndex-1) - (U32)repIndex2) >= 3 /* intentional overflow */)
+                   && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
+                    const BYTE* const repEnd2 = repIndex2 < prefixLowestIndex ? dictEnd : iend;
+                    size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixLowest) + 4;
+                    U32 tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset;   /* swap offset_2 <=> offset_1 */
+                    ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
+                    hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = current2;
+                    hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = current2;
+                    ip += repLength2;
                     anchor = ip;
-                    continue;   /* faster when present ... (?) */
-        }   }   }
+                    continue;
+                }
+                break;
+            }
+        }
     }   /* while (ip < ilimit) */
 
     /* save reps for next block */
-    rep[0] = offset_1 ? offset_1 : offsetSaved;
-    rep[1] = offset_2 ? offset_2 : offsetSaved;
+    rep[0] = offset_1;
+    rep[1] = offset_2;
 
     /* Return the last literals size */
     return (size_t)(iend - anchor);
 }
 
+#define ZSTD_GEN_DFAST_FN(dictMode, mls)                                                                 \
+    static size_t ZSTD_compressBlock_doubleFast_##dictMode##_##mls(                                      \
+            ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],                          \
+            void const* src, size_t srcSize)                                                             \
+    {                                                                                                    \
+        return ZSTD_compressBlock_doubleFast_##dictMode##_generic(ms, seqStore, rep, src, srcSize, mls); \
+    }
+
+ZSTD_GEN_DFAST_FN(noDict, 4)
+ZSTD_GEN_DFAST_FN(noDict, 5)
+ZSTD_GEN_DFAST_FN(noDict, 6)
+ZSTD_GEN_DFAST_FN(noDict, 7)
+
+ZSTD_GEN_DFAST_FN(dictMatchState, 4)
+ZSTD_GEN_DFAST_FN(dictMatchState, 5)
+ZSTD_GEN_DFAST_FN(dictMatchState, 6)
+ZSTD_GEN_DFAST_FN(dictMatchState, 7)
+
 
 size_t ZSTD_compressBlock_doubleFast(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
@@ -325,13 +558,13 @@ size_t ZSTD_compressBlock_doubleFast(
     {
     default: /* includes case 3 */
     case 4 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 4, ZSTD_noDict);
+        return ZSTD_compressBlock_doubleFast_noDict_4(ms, seqStore, rep, src, srcSize);
     case 5 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 5, ZSTD_noDict);
+        return ZSTD_compressBlock_doubleFast_noDict_5(ms, seqStore, rep, src, srcSize);
     case 6 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 6, ZSTD_noDict);
+        return ZSTD_compressBlock_doubleFast_noDict_6(ms, seqStore, rep, src, srcSize);
     case 7 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 7, ZSTD_noDict);
+        return ZSTD_compressBlock_doubleFast_noDict_7(ms, seqStore, rep, src, srcSize);
     }
 }
 
@@ -345,13 +578,13 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState(
     {
     default: /* includes case 3 */
     case 4 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 4, ZSTD_dictMatchState);
+        return ZSTD_compressBlock_doubleFast_dictMatchState_4(ms, seqStore, rep, src, srcSize);
     case 5 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 5, ZSTD_dictMatchState);
+        return ZSTD_compressBlock_doubleFast_dictMatchState_5(ms, seqStore, rep, src, srcSize);
     case 6 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 6, ZSTD_dictMatchState);
+        return ZSTD_compressBlock_doubleFast_dictMatchState_6(ms, seqStore, rep, src, srcSize);
     case 7 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 7, ZSTD_dictMatchState);
+        return ZSTD_compressBlock_doubleFast_dictMatchState_7(ms, seqStore, rep, src, srcSize);
     }
 }
 
@@ -387,7 +620,7 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
 
     /* if extDict is invalidated due to maxDistance, switch to "regular" variant */
     if (prefixStartIndex == dictStartIndex)
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, mls, ZSTD_noDict);
+        return ZSTD_compressBlock_doubleFast(ms, seqStore, rep, src, srcSize);
 
     /* Search Loop */
     while (ip < ilimit) {  /* < instead of <=, because (ip+1) */
@@ -401,31 +634,31 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
         const BYTE* const matchLongBase = matchLongIndex < prefixStartIndex ? dictBase : base;
         const BYTE* matchLong = matchLongBase + matchLongIndex;
 
-        const U32 current = (U32)(ip-base);
-        const U32 repIndex = current + 1 - offset_1;   /* offset_1 expected <= current +1 */
+        const U32 curr = (U32)(ip-base);
+        const U32 repIndex = curr + 1 - offset_1;   /* offset_1 expected <= curr +1 */
         const BYTE* const repBase = repIndex < prefixStartIndex ? dictBase : base;
         const BYTE* const repMatch = repBase + repIndex;
         size_t mLength;
-        hashSmall[hSmall] = hashLong[hLong] = current;   /* update hash table */
+        hashSmall[hSmall] = hashLong[hLong] = curr;   /* update hash table */
 
         if ((((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow : ensure repIndex doesn't overlap dict + prefix */
-            & (repIndex > dictStartIndex))
+            & (offset_1 <= curr+1 - dictStartIndex)) /* note: we are searching at curr+1 */
           && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
             const BYTE* repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
             mLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixStart) + 4;
             ip++;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, 0, mLength-MINMATCH);
+            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength);
         } else {
             if ((matchLongIndex > dictStartIndex) && (MEM_read64(matchLong) == MEM_read64(ip))) {
                 const BYTE* const matchEnd = matchLongIndex < prefixStartIndex ? dictEnd : iend;
                 const BYTE* const lowMatchPtr = matchLongIndex < prefixStartIndex ? dictStart : prefixStart;
                 U32 offset;
                 mLength = ZSTD_count_2segments(ip+8, matchLong+8, iend, matchEnd, prefixStart) + 8;
-                offset = current - matchLongIndex;
+                offset = curr - matchLongIndex;
                 while (((ip>anchor) & (matchLong>lowMatchPtr)) && (ip[-1] == matchLong[-1])) { ip--; matchLong--; mLength++; }   /* catch up */
                 offset_2 = offset_1;
                 offset_1 = offset;
-                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH);
+                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
 
             } else if ((matchIndex > dictStartIndex) && (MEM_read32(match) == MEM_read32(ip))) {
                 size_t const h3 = ZSTD_hashPtr(ip+1, hBitsL, 8);
@@ -433,24 +666,24 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
                 const BYTE* const match3Base = matchIndex3 < prefixStartIndex ? dictBase : base;
                 const BYTE* match3 = match3Base + matchIndex3;
                 U32 offset;
-                hashLong[h3] = current + 1;
+                hashLong[h3] = curr + 1;
                 if ( (matchIndex3 > dictStartIndex) && (MEM_read64(match3) == MEM_read64(ip+1)) ) {
                     const BYTE* const matchEnd = matchIndex3 < prefixStartIndex ? dictEnd : iend;
                     const BYTE* const lowMatchPtr = matchIndex3 < prefixStartIndex ? dictStart : prefixStart;
                     mLength = ZSTD_count_2segments(ip+9, match3+8, iend, matchEnd, prefixStart) + 8;
                     ip++;
-                    offset = current+1 - matchIndex3;
+                    offset = curr+1 - matchIndex3;
                     while (((ip>anchor) & (match3>lowMatchPtr)) && (ip[-1] == match3[-1])) { ip--; match3--; mLength++; } /* catch up */
                 } else {
                     const BYTE* const matchEnd = matchIndex < prefixStartIndex ? dictEnd : iend;
                     const BYTE* const lowMatchPtr = matchIndex < prefixStartIndex ? dictStart : prefixStart;
                     mLength = ZSTD_count_2segments(ip+4, match+4, iend, matchEnd, prefixStart) + 4;
-                    offset = current - matchIndex;
+                    offset = curr - matchIndex;
                     while (((ip>anchor) & (match>lowMatchPtr)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; }   /* catch up */
                 }
                 offset_2 = offset_1;
                 offset_1 = offset;
-                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH);
+                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
 
             } else {
                 ip += ((ip-anchor) >> kSearchStrength) + 1;
@@ -464,7 +697,7 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
         if (ip <= ilimit) {
             /* Complementary insertion */
             /* done after iLimit test, as candidates could be > iend-8 */
-            {   U32 const indexToInsert = current+2;
+            {   U32 const indexToInsert = curr+2;
                 hashLong[ZSTD_hashPtr(base+indexToInsert, hBitsL, 8)] = indexToInsert;
                 hashLong[ZSTD_hashPtr(ip-2, hBitsL, 8)] = (U32)(ip-2-base);
                 hashSmall[ZSTD_hashPtr(base+indexToInsert, hBitsS, mls)] = indexToInsert;
@@ -477,12 +710,12 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
                 U32 const repIndex2 = current2 - offset_2;
                 const BYTE* repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2;
                 if ( (((U32)((prefixStartIndex-1) - repIndex2) >= 3)   /* intentional overflow : ensure repIndex2 doesn't overlap dict + prefix */
-                    & (repIndex2 > dictStartIndex))
+                    & (offset_2 <= current2 - dictStartIndex))
                   && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
                     const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
                     size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
                     U32 const tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset;   /* swap offset_2 <=> offset_1 */
-                    ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, repLength2-MINMATCH);
+                    ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
                     hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = current2;
                     hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = current2;
                     ip += repLength2;
@@ -500,6 +733,10 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
     return (size_t)(iend - anchor);
 }
 
+ZSTD_GEN_DFAST_FN(extDict, 4)
+ZSTD_GEN_DFAST_FN(extDict, 5)
+ZSTD_GEN_DFAST_FN(extDict, 6)
+ZSTD_GEN_DFAST_FN(extDict, 7)
 
 size_t ZSTD_compressBlock_doubleFast_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
@@ -510,12 +747,12 @@ size_t ZSTD_compressBlock_doubleFast_extDict(
     {
     default: /* includes case 3 */
     case 4 :
-        return ZSTD_compressBlock_doubleFast_extDict_generic(ms, seqStore, rep, src, srcSize, 4);
+        return ZSTD_compressBlock_doubleFast_extDict_4(ms, seqStore, rep, src, srcSize);
     case 5 :
-        return ZSTD_compressBlock_doubleFast_extDict_generic(ms, seqStore, rep, src, srcSize, 5);
+        return ZSTD_compressBlock_doubleFast_extDict_5(ms, seqStore, rep, src, srcSize);
     case 6 :
-        return ZSTD_compressBlock_doubleFast_extDict_generic(ms, seqStore, rep, src, srcSize, 6);
+        return ZSTD_compressBlock_doubleFast_extDict_6(ms, seqStore, rep, src, srcSize);
     case 7 :
-        return ZSTD_compressBlock_doubleFast_extDict_generic(ms, seqStore, rep, src, srcSize, 7);
+        return ZSTD_compressBlock_doubleFast_extDict_7(ms, seqStore, rep, src, srcSize);
     }
 }

src/borg/algorithms/zstd/lib/compress/zstd_double_fast.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -19,7 +19,8 @@ extern "C" {
 #include "zstd_compress_internal.h"     /* ZSTD_CCtx, size_t */
 
 void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms,
-                              void const* end, ZSTD_dictTableLoadMethod_e dtlm);
+                              void const* end, ZSTD_dictTableLoadMethod_e dtlm,
+                              ZSTD_tableFillPurpose_e tfp);
 size_t ZSTD_compressBlock_doubleFast(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);

src/borg/algorithms/zstd/lib/compress/zstd_fast.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -11,8 +11,42 @@
 #include "zstd_compress_internal.h"  /* ZSTD_hashPtr, ZSTD_count, ZSTD_storeSeq */
 #include "zstd_fast.h"
 
+static void ZSTD_fillHashTableForCDict(ZSTD_matchState_t* ms,
+                        const void* const end,
+                        ZSTD_dictTableLoadMethod_e dtlm)
+{
+    const ZSTD_compressionParameters* const cParams = &ms->cParams;
+    U32* const hashTable = ms->hashTable;
+    U32  const hBits = cParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS;
+    U32  const mls = cParams->minMatch;
+    const BYTE* const base = ms->window.base;
+    const BYTE* ip = base + ms->nextToUpdate;
+    const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE;
+    const U32 fastHashFillStep = 3;
 
-void ZSTD_fillHashTable(ZSTD_matchState_t* ms,
+    /* Currently, we always use ZSTD_dtlm_full for filling CDict tables.
+     * Feel free to remove this assert if there's a good reason! */
+    assert(dtlm == ZSTD_dtlm_full);
+
+    /* Always insert every fastHashFillStep position into the hash table.
+     * Insert the other positions if their hash entry is empty.
+     */
+    for ( ; ip + fastHashFillStep < iend + 2; ip += fastHashFillStep) {
+        U32 const curr = (U32)(ip - base);
+        {   size_t const hashAndTag = ZSTD_hashPtr(ip, hBits, mls);
+            ZSTD_writeTaggedIndex(hashTable, hashAndTag, curr);   }
+
+        if (dtlm == ZSTD_dtlm_fast) continue;
+        /* Only load extra positions for ZSTD_dtlm_full */
+        {   U32 p;
+            for (p = 1; p < fastHashFillStep; ++p) {
+                size_t const hashAndTag = ZSTD_hashPtr(ip + p, hBits, mls);
+                if (hashTable[hashAndTag >> ZSTD_SHORT_CACHE_TAG_BITS] == 0) {  /* not yet filled */
+                    ZSTD_writeTaggedIndex(hashTable, hashAndTag, curr + p);
+                }   }   }   }
+}
+
+static void ZSTD_fillHashTableForCCtx(ZSTD_matchState_t* ms,
                         const void* const end,
                         ZSTD_dictTableLoadMethod_e dtlm)
 {
@@ -25,163 +59,365 @@ void ZSTD_fillHashTable(ZSTD_matchState_t* ms,
     const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE;
     const U32 fastHashFillStep = 3;
 
+    /* Currently, we always use ZSTD_dtlm_fast for filling CCtx tables.
+     * Feel free to remove this assert if there's a good reason! */
+    assert(dtlm == ZSTD_dtlm_fast);
+
     /* Always insert every fastHashFillStep position into the hash table.
      * Insert the other positions if their hash entry is empty.
      */
     for ( ; ip + fastHashFillStep < iend + 2; ip += fastHashFillStep) {
-        U32 const current = (U32)(ip - base);
+        U32 const curr = (U32)(ip - base);
         size_t const hash0 = ZSTD_hashPtr(ip, hBits, mls);
-        hashTable[hash0] = current;
+        hashTable[hash0] = curr;
         if (dtlm == ZSTD_dtlm_fast) continue;
         /* Only load extra positions for ZSTD_dtlm_full */
         {   U32 p;
             for (p = 1; p < fastHashFillStep; ++p) {
                 size_t const hash = ZSTD_hashPtr(ip + p, hBits, mls);
                 if (hashTable[hash] == 0) {  /* not yet filled */
-                    hashTable[hash] = current + p;
+                    hashTable[hash] = curr + p;
     }   }   }   }
 }
 
+void ZSTD_fillHashTable(ZSTD_matchState_t* ms,
+                        const void* const end,
+                        ZSTD_dictTableLoadMethod_e dtlm,
+                        ZSTD_tableFillPurpose_e tfp)
+{
+    if (tfp == ZSTD_tfp_forCDict) {
+        ZSTD_fillHashTableForCDict(ms, end, dtlm);
+    } else {
+        ZSTD_fillHashTableForCCtx(ms, end, dtlm);
+    }
+}
+
 
+/**
+ * If you squint hard enough (and ignore repcodes), the search operation at any
+ * given position is broken into 4 stages:
+ *
+ * 1. Hash   (map position to hash value via input read)
+ * 2. Lookup (map hash val to index via hashtable read)
+ * 3. Load   (map index to value at that position via input read)
+ * 4. Compare
+ *
+ * Each of these steps involves a memory read at an address which is computed
+ * from the previous step. This means these steps must be sequenced and their
+ * latencies are cumulative.
+ *
+ * Rather than do 1->2->3->4 sequentially for a single position before moving
+ * onto the next, this implementation interleaves these operations across the
+ * next few positions:
+ *
+ * R = Repcode Read & Compare
+ * H = Hash
+ * T = Table Lookup
+ * M = Match Read & Compare
+ *
+ * Pos | Time -->
+ * ----+-------------------
+ * N   | ... M
+ * N+1 | ...   TM
+ * N+2 |    R H   T M
+ * N+3 |         H    TM
+ * N+4 |           R H   T M
+ * N+5 |                H   ...
+ * N+6 |                  R ...
+ *
+ * This is very much analogous to the pipelining of execution in a CPU. And just
+ * like a CPU, we have to dump the pipeline when we find a match (i.e., take a
+ * branch).
+ *
+ * When this happens, we throw away our current state, and do the following prep
+ * to re-enter the loop:
+ *
+ * Pos | Time -->
+ * ----+-------------------
+ * N   | H T
+ * N+1 |  H
+ *
+ * This is also the work we do at the beginning to enter the loop initially.
+ */
 FORCE_INLINE_TEMPLATE size_t
-ZSTD_compressBlock_fast_generic(
+ZSTD_compressBlock_fast_noDict_generic(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize,
-        U32 const mls)
+        U32 const mls, U32 const hasStep)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
     U32* const hashTable = ms->hashTable;
     U32 const hlog = cParams->hashLog;
     /* support stepSize of 0 */
-    size_t const stepSize = cParams->targetLength + !(cParams->targetLength) + 1;
+    size_t const stepSize = hasStep ? (cParams->targetLength + !(cParams->targetLength) + 1) : 2;
     const BYTE* const base = ms->window.base;
     const BYTE* const istart = (const BYTE*)src;
-    /* We check ip0 (ip + 0) and ip1 (ip + 1) each loop */
-    const BYTE* ip0 = istart;
-    const BYTE* ip1;
-    const BYTE* anchor = istart;
     const U32   endIndex = (U32)((size_t)(istart - base) + srcSize);
     const U32   prefixStartIndex = ZSTD_getLowestPrefixIndex(ms, endIndex, cParams->windowLog);
     const BYTE* const prefixStart = base + prefixStartIndex;
     const BYTE* const iend = istart + srcSize;
     const BYTE* const ilimit = iend - HASH_READ_SIZE;
-    U32 offset_1=rep[0], offset_2=rep[1];
-    U32 offsetSaved = 0;
 
-    /* init */
+    const BYTE* anchor = istart;
+    const BYTE* ip0 = istart;
+    const BYTE* ip1;
+    const BYTE* ip2;
+    const BYTE* ip3;
+    U32 current0;
+
+    U32 rep_offset1 = rep[0];
+    U32 rep_offset2 = rep[1];
+    U32 offsetSaved1 = 0, offsetSaved2 = 0;
+
+    size_t hash0; /* hash for ip0 */
+    size_t hash1; /* hash for ip1 */
+    U32 idx; /* match idx for ip0 */
+    U32 mval; /* src value at match idx */
+
+    U32 offcode;
+    const BYTE* match0;
+    size_t mLength;
+
+    /* ip0 and ip1 are always adjacent. The targetLength skipping and
+     * uncompressibility acceleration is applied to every other position,
+     * matching the behavior of #1562. step therefore represents the gap
+     * between pairs of positions, from ip0 to ip2 or ip1 to ip3. */
+    size_t step;
+    const BYTE* nextStep;
+    const size_t kStepIncr = (1 << (kSearchStrength - 1));
+
     DEBUGLOG(5, "ZSTD_compressBlock_fast_generic");
     ip0 += (ip0 == prefixStart);
+    {   U32 const curr = (U32)(ip0 - base);
+        U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, curr, cParams->windowLog);
+        U32 const maxRep = curr - windowLow;
+        if (rep_offset2 > maxRep) offsetSaved2 = rep_offset2, rep_offset2 = 0;
+        if (rep_offset1 > maxRep) offsetSaved1 = rep_offset1, rep_offset1 = 0;
+    }
+
+    /* start each op */
+_start: /* Requires: ip0 */
+
+    step = stepSize;
+    nextStep = ip0 + kStepIncr;
+
+    /* calculate positions, ip0 - anchor == 0, so we skip step calc */
     ip1 = ip0 + 1;
-    {   U32 const current = (U32)(ip0 - base);
-        U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, current, cParams->windowLog);
-        U32 const maxRep = current - windowLow;
-        if (offset_2 > maxRep) offsetSaved = offset_2, offset_2 = 0;
-        if (offset_1 > maxRep) offsetSaved = offset_1, offset_1 = 0;
+    ip2 = ip0 + step;
+    ip3 = ip2 + 1;
+
+    if (ip3 >= ilimit) {
+        goto _cleanup;
     }
 
-    /* Main Search Loop */
-#ifdef __INTEL_COMPILER
-    /* From intel 'The vector pragma indicates that the loop should be
-     * vectorized if it is legal to do so'. Can be used together with
-     * #pragma ivdep (but have opted to exclude that because intel
-     * warns against using it).*/
-    #pragma vector always
-#endif
-    while (ip1 < ilimit) {   /* < instead of <=, because check at ip0+2 */
-        size_t mLength;
-        BYTE const* ip2 = ip0 + 2;
-        size_t const h0 = ZSTD_hashPtr(ip0, hlog, mls);
-        U32 const val0 = MEM_read32(ip0);
-        size_t const h1 = ZSTD_hashPtr(ip1, hlog, mls);
-        U32 const val1 = MEM_read32(ip1);
-        U32 const current0 = (U32)(ip0-base);
-        U32 const current1 = (U32)(ip1-base);
-        U32 const matchIndex0 = hashTable[h0];
-        U32 const matchIndex1 = hashTable[h1];
-        BYTE const* repMatch = ip2 - offset_1;
-        const BYTE* match0 = base + matchIndex0;
-        const BYTE* match1 = base + matchIndex1;
-        U32 offcode;
-
-#if defined(__aarch64__)
-        PREFETCH_L1(ip0+256);
-#endif
-
-        hashTable[h0] = current0;   /* update hash table */
-        hashTable[h1] = current1;   /* update hash table */
-
-        assert(ip0 + 1 == ip1);
-
-        if ((offset_1 > 0) & (MEM_read32(repMatch) == MEM_read32(ip2))) {
-            mLength = (ip2[-1] == repMatch[-1]) ? 1 : 0;
-            ip0 = ip2 - mLength;
-            match0 = repMatch - mLength;
+    hash0 = ZSTD_hashPtr(ip0, hlog, mls);
+    hash1 = ZSTD_hashPtr(ip1, hlog, mls);
+
+    idx = hashTable[hash0];
+
+    do {
+        /* load repcode match for ip[2]*/
+        const U32 rval = MEM_read32(ip2 - rep_offset1);
+
+        /* write back hash table entry */
+        current0 = (U32)(ip0 - base);
+        hashTable[hash0] = current0;
+
+        /* check repcode at ip[2] */
+        if ((MEM_read32(ip2) == rval) & (rep_offset1 > 0)) {
+            ip0 = ip2;
+            match0 = ip0 - rep_offset1;
+            mLength = ip0[-1] == match0[-1];
+            ip0 -= mLength;
+            match0 -= mLength;
+            offcode = REPCODE1_TO_OFFBASE;
             mLength += 4;
-            offcode = 0;
+
+            /* First write next hash table entry; we've already calculated it.
+             * This write is known to be safe because the ip1 is before the
+             * repcode (ip2). */
+            hashTable[hash1] = (U32)(ip1 - base);
+
             goto _match;
         }
-        if ((matchIndex0 > prefixStartIndex) && MEM_read32(match0) == val0) {
-            /* found a regular match */
-            goto _offset;
+
+        /* load match for ip[0] */
+        if (idx >= prefixStartIndex) {
+            mval = MEM_read32(base + idx);
+        } else {
+            mval = MEM_read32(ip0) ^ 1; /* guaranteed to not match. */
         }
-        if ((matchIndex1 > prefixStartIndex) && MEM_read32(match1) == val1) {
-            /* found a regular match after one literal */
-            ip0 = ip1;
-            match0 = match1;
+
+        /* check match at ip[0] */
+        if (MEM_read32(ip0) == mval) {
+            /* found a match! */
+
+            /* First write next hash table entry; we've already calculated it.
+             * This write is known to be safe because the ip1 == ip0 + 1, so
+             * we know we will resume searching after ip1 */
+            hashTable[hash1] = (U32)(ip1 - base);
+
             goto _offset;
         }
-        {   size_t const step = ((size_t)(ip0-anchor) >> (kSearchStrength - 1)) + stepSize;
-            assert(step >= 2);
-            ip0 += step;
-            ip1 += step;
-            continue;
+
+        /* lookup ip[1] */
+        idx = hashTable[hash1];
+
+        /* hash ip[2] */
+        hash0 = hash1;
+        hash1 = ZSTD_hashPtr(ip2, hlog, mls);
+
+        /* advance to next positions */
+        ip0 = ip1;
+        ip1 = ip2;
+        ip2 = ip3;
+
+        /* write back hash table entry */
+        current0 = (U32)(ip0 - base);
+        hashTable[hash0] = current0;
+
+        /* load match for ip[0] */
+        if (idx >= prefixStartIndex) {
+            mval = MEM_read32(base + idx);
+        } else {
+            mval = MEM_read32(ip0) ^ 1; /* guaranteed to not match. */
         }
-_offset: /* Requires: ip0, match0 */
-        /* Compute the offset code */
-        offset_2 = offset_1;
-        offset_1 = (U32)(ip0-match0);
-        offcode = offset_1 + ZSTD_REP_MOVE;
-        mLength = 4;
-        /* Count the backwards match length */
-        while (((ip0>anchor) & (match0>prefixStart))
-             && (ip0[-1] == match0[-1])) { ip0--; match0--; mLength++; } /* catch up */
 
-_match: /* Requires: ip0, match0, offcode */
-        /* Count the forward length */
-        mLength += ZSTD_count(ip0+mLength, match0+mLength, iend);
-        ZSTD_storeSeq(seqStore, (size_t)(ip0-anchor), anchor, iend, offcode, mLength-MINMATCH);
-        /* match found */
-        ip0 += mLength;
-        anchor = ip0;
+        /* check match at ip[0] */
+        if (MEM_read32(ip0) == mval) {
+            /* found a match! */
+
+            /* first write next hash table entry; we've already calculated it */
+            if (step <= 4) {
+                /* We need to avoid writing an index into the hash table >= the
+                 * position at which we will pick up our searching after we've
+                 * taken this match.
+                 *
+                 * The minimum possible match has length 4, so the earliest ip0
+                 * can be after we take this match will be the current ip0 + 4.
+                 * ip1 is ip0 + step - 1. If ip1 is >= ip0 + 4, we can't safely
+                 * write this position.
+                 */
+                hashTable[hash1] = (U32)(ip1 - base);
+            }
 
-        if (ip0 <= ilimit) {
-            /* Fill Table */
-            assert(base+current0+2 > istart);  /* check base overflow */
-            hashTable[ZSTD_hashPtr(base+current0+2, hlog, mls)] = current0+2;  /* here because current+2 could be > iend-8 */
-            hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base);
+            goto _offset;
+        }
 
-            if (offset_2 > 0) { /* offset_2==0 means offset_2 is invalidated */
-                while ( (ip0 <= ilimit) && (MEM_read32(ip0) == MEM_read32(ip0 - offset_2)) ) {
-                    /* store sequence */
-                    size_t const rLength = ZSTD_count(ip0+4, ip0+4-offset_2, iend) + 4;
-                    { U32 const tmpOff = offset_2; offset_2 = offset_1; offset_1 = tmpOff; } /* swap offset_2 <=> offset_1 */
-                    hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = (U32)(ip0-base);
-                    ip0 += rLength;
-                    ZSTD_storeSeq(seqStore, 0 /*litLen*/, anchor, iend, 0 /*offCode*/, rLength-MINMATCH);
-                    anchor = ip0;
-                    continue;   /* faster when present (confirmed on gcc-8) ... (?) */
-        }   }   }
-        ip1 = ip0 + 1;
-    }
+        /* lookup ip[1] */
+        idx = hashTable[hash1];
+
+        /* hash ip[2] */
+        hash0 = hash1;
+        hash1 = ZSTD_hashPtr(ip2, hlog, mls);
+
+        /* advance to next positions */
+        ip0 = ip1;
+        ip1 = ip2;
+        ip2 = ip0 + step;
+        ip3 = ip1 + step;
+
+        /* calculate step */
+        if (ip2 >= nextStep) {
+            step++;
+            PREFETCH_L1(ip1 + 64);
+            PREFETCH_L1(ip1 + 128);
+            nextStep += kStepIncr;
+        }
+    } while (ip3 < ilimit);
+
+_cleanup:
+    /* Note that there are probably still a couple positions we could search.
+     * However, it seems to be a meaningful performance hit to try to search
+     * them. So let's not. */
+
+    /* When the repcodes are outside of the prefix, we set them to zero before the loop.
+     * When the offsets are still zero, we need to restore them after the block to have a correct
+     * repcode history. If only one offset was invalid, it is easy. The tricky case is when both
+     * offsets were invalid. We need to figure out which offset to refill with.
+     *     - If both offsets are zero they are in the same order.
+     *     - If both offsets are non-zero, we won't restore the offsets from `offsetSaved[12]`.
+     *     - If only one is zero, we need to decide which offset to restore.
+     *         - If rep_offset1 is non-zero, then rep_offset2 must be offsetSaved1.
+     *         - It is impossible for rep_offset2 to be non-zero.
+     *
+     * So if rep_offset1 started invalid (offsetSaved1 != 0) and became valid (rep_offset1 != 0), then
+     * set rep[0] = rep_offset1 and rep[1] = offsetSaved1.
+     */
+    offsetSaved2 = ((offsetSaved1 != 0) && (rep_offset1 != 0)) ? offsetSaved1 : offsetSaved2;
 
     /* save reps for next block */
-    rep[0] = offset_1 ? offset_1 : offsetSaved;
-    rep[1] = offset_2 ? offset_2 : offsetSaved;
+    rep[0] = rep_offset1 ? rep_offset1 : offsetSaved1;
+    rep[1] = rep_offset2 ? rep_offset2 : offsetSaved2;
 
     /* Return the last literals size */
     return (size_t)(iend - anchor);
+
+_offset: /* Requires: ip0, idx */
+
+    /* Compute the offset code. */
+    match0 = base + idx;
+    rep_offset2 = rep_offset1;
+    rep_offset1 = (U32)(ip0-match0);
+    offcode = OFFSET_TO_OFFBASE(rep_offset1);
+    mLength = 4;
+
+    /* Count the backwards match length. */
+    while (((ip0>anchor) & (match0>prefixStart)) && (ip0[-1] == match0[-1])) {
+        ip0--;
+        match0--;
+        mLength++;
+    }
+
+_match: /* Requires: ip0, match0, offcode */
+
+    /* Count the forward length. */
+    mLength += ZSTD_count(ip0 + mLength, match0 + mLength, iend);
+
+    ZSTD_storeSeq(seqStore, (size_t)(ip0 - anchor), anchor, iend, offcode, mLength);
+
+    ip0 += mLength;
+    anchor = ip0;
+
+    /* Fill table and check for immediate repcode. */
+    if (ip0 <= ilimit) {
+        /* Fill Table */
+        assert(base+current0+2 > istart);  /* check base overflow */
+        hashTable[ZSTD_hashPtr(base+current0+2, hlog, mls)] = current0+2;  /* here because current+2 could be > iend-8 */
+        hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base);
+
+        if (rep_offset2 > 0) { /* rep_offset2==0 means rep_offset2 is invalidated */
+            while ( (ip0 <= ilimit) && (MEM_read32(ip0) == MEM_read32(ip0 - rep_offset2)) ) {
+                /* store sequence */
+                size_t const rLength = ZSTD_count(ip0+4, ip0+4-rep_offset2, iend) + 4;
+                { U32 const tmpOff = rep_offset2; rep_offset2 = rep_offset1; rep_offset1 = tmpOff; } /* swap rep_offset2 <=> rep_offset1 */
+                hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = (U32)(ip0-base);
+                ip0 += rLength;
+                ZSTD_storeSeq(seqStore, 0 /*litLen*/, anchor, iend, REPCODE1_TO_OFFBASE, rLength);
+                anchor = ip0;
+                continue;   /* faster when present (confirmed on gcc-8) ... (?) */
+    }   }   }
+
+    goto _start;
 }
 
+#define ZSTD_GEN_FAST_FN(dictMode, mls, step)                                                            \
+    static size_t ZSTD_compressBlock_fast_##dictMode##_##mls##_##step(                                      \
+            ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],                    \
+            void const* src, size_t srcSize)                                                       \
+    {                                                                                              \
+        return ZSTD_compressBlock_fast_##dictMode##_generic(ms, seqStore, rep, src, srcSize, mls, step); \
+    }
+
+ZSTD_GEN_FAST_FN(noDict, 4, 1)
+ZSTD_GEN_FAST_FN(noDict, 5, 1)
+ZSTD_GEN_FAST_FN(noDict, 6, 1)
+ZSTD_GEN_FAST_FN(noDict, 7, 1)
+
+ZSTD_GEN_FAST_FN(noDict, 4, 0)
+ZSTD_GEN_FAST_FN(noDict, 5, 0)
+ZSTD_GEN_FAST_FN(noDict, 6, 0)
+ZSTD_GEN_FAST_FN(noDict, 7, 0)
 
 size_t ZSTD_compressBlock_fast(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
@@ -189,24 +425,40 @@ size_t ZSTD_compressBlock_fast(
 {
     U32 const mls = ms->cParams.minMatch;
     assert(ms->dictMatchState == NULL);
-    switch(mls)
-    {
-    default: /* includes case 3 */
-    case 4 :
-        return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, 4);
-    case 5 :
-        return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, 5);
-    case 6 :
-        return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, 6);
-    case 7 :
-        return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, 7);
+    if (ms->cParams.targetLength > 1) {
+        switch(mls)
+        {
+        default: /* includes case 3 */
+        case 4 :
+            return ZSTD_compressBlock_fast_noDict_4_1(ms, seqStore, rep, src, srcSize);
+        case 5 :
+            return ZSTD_compressBlock_fast_noDict_5_1(ms, seqStore, rep, src, srcSize);
+        case 6 :
+            return ZSTD_compressBlock_fast_noDict_6_1(ms, seqStore, rep, src, srcSize);
+        case 7 :
+            return ZSTD_compressBlock_fast_noDict_7_1(ms, seqStore, rep, src, srcSize);
+        }
+    } else {
+        switch(mls)
+        {
+        default: /* includes case 3 */
+        case 4 :
+            return ZSTD_compressBlock_fast_noDict_4_0(ms, seqStore, rep, src, srcSize);
+        case 5 :
+            return ZSTD_compressBlock_fast_noDict_5_0(ms, seqStore, rep, src, srcSize);
+        case 6 :
+            return ZSTD_compressBlock_fast_noDict_6_0(ms, seqStore, rep, src, srcSize);
+        case 7 :
+            return ZSTD_compressBlock_fast_noDict_7_0(ms, seqStore, rep, src, srcSize);
+        }
+
     }
 }
 
 FORCE_INLINE_TEMPLATE
 size_t ZSTD_compressBlock_fast_dictMatchState_generic(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize, U32 const mls)
+        void const* src, size_t srcSize, U32 const mls, U32 const hasStep)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
     U32* const hashTable = ms->hashTable;
@@ -215,14 +467,14 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic(
     U32 const stepSize = cParams->targetLength + !(cParams->targetLength);
     const BYTE* const base = ms->window.base;
     const BYTE* const istart = (const BYTE*)src;
-    const BYTE* ip = istart;
+    const BYTE* ip0 = istart;
+    const BYTE* ip1 = ip0 + stepSize; /* we assert below that stepSize >= 1 */
     const BYTE* anchor = istart;
     const U32   prefixStartIndex = ms->window.dictLimit;
     const BYTE* const prefixStart = base + prefixStartIndex;
     const BYTE* const iend = istart + srcSize;
     const BYTE* const ilimit = iend - HASH_READ_SIZE;
     U32 offset_1=rep[0], offset_2=rep[1];
-    U32 offsetSaved = 0;
 
     const ZSTD_matchState_t* const dms = ms->dictMatchState;
     const ZSTD_compressionParameters* const dictCParams = &dms->cParams ;
@@ -232,125 +484,182 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic(
     const BYTE* const dictStart    = dictBase + dictStartIndex;
     const BYTE* const dictEnd      = dms->window.nextSrc;
     const U32 dictIndexDelta       = prefixStartIndex - (U32)(dictEnd - dictBase);
-    const U32 dictAndPrefixLength  = (U32)(ip - prefixStart + dictEnd - dictStart);
-    const U32 dictHLog             = dictCParams->hashLog;
+    const U32 dictAndPrefixLength  = (U32)(istart - prefixStart + dictEnd - dictStart);
+    const U32 dictHBits            = dictCParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS;
 
     /* if a dictionary is still attached, it necessarily means that
      * it is within window size. So we just check it. */
     const U32 maxDistance = 1U << cParams->windowLog;
-    const U32 endIndex = (U32)((size_t)(ip - base) + srcSize);
+    const U32 endIndex = (U32)((size_t)(istart - base) + srcSize);
     assert(endIndex - prefixStartIndex <= maxDistance);
     (void)maxDistance; (void)endIndex;   /* these variables are not used when assert() is disabled */
 
-    /* ensure there will be no no underflow
+    (void)hasStep; /* not currently specialized on whether it's accelerated */
+
+    /* ensure there will be no underflow
      * when translating a dict index into a local index */
     assert(prefixStartIndex >= (U32)(dictEnd - dictBase));
 
+    if (ms->prefetchCDictTables) {
+        size_t const hashTableBytes = (((size_t)1) << dictCParams->hashLog) * sizeof(U32);
+        PREFETCH_AREA(dictHashTable, hashTableBytes)
+    }
+
     /* init */
     DEBUGLOG(5, "ZSTD_compressBlock_fast_dictMatchState_generic");
-    ip += (dictAndPrefixLength == 0);
+    ip0 += (dictAndPrefixLength == 0);
     /* dictMatchState repCode checks don't currently handle repCode == 0
      * disabling. */
     assert(offset_1 <= dictAndPrefixLength);
     assert(offset_2 <= dictAndPrefixLength);
 
-    /* Main Search Loop */
-    while (ip < ilimit) {   /* < instead of <=, because repcode check at (ip+1) */
+    /* Outer search loop */
+    assert(stepSize >= 1);
+    while (ip1 <= ilimit) {   /* repcode check at (ip0 + 1) is safe because ip0 < ip1 */
         size_t mLength;
-        size_t const h = ZSTD_hashPtr(ip, hlog, mls);
-        U32 const current = (U32)(ip-base);
-        U32 const matchIndex = hashTable[h];
-        const BYTE* match = base + matchIndex;
-        const U32 repIndex = current + 1 - offset_1;
-        const BYTE* repMatch = (repIndex < prefixStartIndex) ?
-                               dictBase + (repIndex - dictIndexDelta) :
-                               base + repIndex;
-        hashTable[h] = current;   /* update hash table */
-
-        if ( ((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow : ensure repIndex isn't overlapping dict + prefix */
-          && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
-            const BYTE* const repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
-            mLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixStart) + 4;
-            ip++;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, 0, mLength-MINMATCH);
-        } else if ( (matchIndex <= prefixStartIndex) ) {
-            size_t const dictHash = ZSTD_hashPtr(ip, dictHLog, mls);
-            U32 const dictMatchIndex = dictHashTable[dictHash];
-            const BYTE* dictMatch = dictBase + dictMatchIndex;
-            if (dictMatchIndex <= dictStartIndex ||
-                MEM_read32(dictMatch) != MEM_read32(ip)) {
-                assert(stepSize >= 1);
-                ip += ((ip-anchor) >> kSearchStrength) + stepSize;
-                continue;
-            } else {
-                /* found a dict match */
-                U32 const offset = (U32)(current-dictMatchIndex-dictIndexDelta);
-                mLength = ZSTD_count_2segments(ip+4, dictMatch+4, iend, dictEnd, prefixStart) + 4;
-                while (((ip>anchor) & (dictMatch>dictStart))
-                     && (ip[-1] == dictMatch[-1])) {
-                    ip--; dictMatch--; mLength++;
+        size_t hash0 = ZSTD_hashPtr(ip0, hlog, mls);
+
+        size_t const dictHashAndTag0 = ZSTD_hashPtr(ip0, dictHBits, mls);
+        U32 dictMatchIndexAndTag = dictHashTable[dictHashAndTag0 >> ZSTD_SHORT_CACHE_TAG_BITS];
+        int dictTagsMatch = ZSTD_comparePackedTags(dictMatchIndexAndTag, dictHashAndTag0);
+
+        U32 matchIndex = hashTable[hash0];
+        U32 curr = (U32)(ip0 - base);
+        size_t step = stepSize;
+        const size_t kStepIncr = 1 << kSearchStrength;
+        const BYTE* nextStep = ip0 + kStepIncr;
+
+        /* Inner search loop */
+        while (1) {
+            const BYTE* match = base + matchIndex;
+            const U32 repIndex = curr + 1 - offset_1;
+            const BYTE* repMatch = (repIndex < prefixStartIndex) ?
+                                   dictBase + (repIndex - dictIndexDelta) :
+                                   base + repIndex;
+            const size_t hash1 = ZSTD_hashPtr(ip1, hlog, mls);
+            size_t const dictHashAndTag1 = ZSTD_hashPtr(ip1, dictHBits, mls);
+            hashTable[hash0] = curr;   /* update hash table */
+
+            if (((U32) ((prefixStartIndex - 1) - repIndex) >=
+                 3) /* intentional underflow : ensure repIndex isn't overlapping dict + prefix */
+                && (MEM_read32(repMatch) == MEM_read32(ip0 + 1))) {
+                const BYTE* const repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
+                mLength = ZSTD_count_2segments(ip0 + 1 + 4, repMatch + 4, iend, repMatchEnd, prefixStart) + 4;
+                ip0++;
+                ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength);
+                break;
+            }
+
+            if (dictTagsMatch) {
+                /* Found a possible dict match */
+                const U32 dictMatchIndex = dictMatchIndexAndTag >> ZSTD_SHORT_CACHE_TAG_BITS;
+                const BYTE* dictMatch = dictBase + dictMatchIndex;
+                if (dictMatchIndex > dictStartIndex &&
+                    MEM_read32(dictMatch) == MEM_read32(ip0)) {
+                    /* To replicate extDict parse behavior, we only use dict matches when the normal matchIndex is invalid */
+                    if (matchIndex <= prefixStartIndex) {
+                        U32 const offset = (U32) (curr - dictMatchIndex - dictIndexDelta);
+                        mLength = ZSTD_count_2segments(ip0 + 4, dictMatch + 4, iend, dictEnd, prefixStart) + 4;
+                        while (((ip0 > anchor) & (dictMatch > dictStart))
+                            && (ip0[-1] == dictMatch[-1])) {
+                            ip0--;
+                            dictMatch--;
+                            mLength++;
+                        } /* catch up */
+                        offset_2 = offset_1;
+                        offset_1 = offset;
+                        ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
+                        break;
+                    }
+                }
+            }
+
+            if (matchIndex > prefixStartIndex && MEM_read32(match) == MEM_read32(ip0)) {
+                /* found a regular match */
+                U32 const offset = (U32) (ip0 - match);
+                mLength = ZSTD_count(ip0 + 4, match + 4, iend) + 4;
+                while (((ip0 > anchor) & (match > prefixStart))
+                       && (ip0[-1] == match[-1])) {
+                    ip0--;
+                    match--;
+                    mLength++;
                 } /* catch up */
                 offset_2 = offset_1;
                 offset_1 = offset;
-                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH);
+                ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
+                break;
             }
-        } else if (MEM_read32(match) != MEM_read32(ip)) {
-            /* it's not a match, and we're not going to check the dictionary */
-            assert(stepSize >= 1);
-            ip += ((ip-anchor) >> kSearchStrength) + stepSize;
-            continue;
-        } else {
-            /* found a regular match */
-            U32 const offset = (U32)(ip-match);
-            mLength = ZSTD_count(ip+4, match+4, iend) + 4;
-            while (((ip>anchor) & (match>prefixStart))
-                 && (ip[-1] == match[-1])) { ip--; match--; mLength++; } /* catch up */
-            offset_2 = offset_1;
-            offset_1 = offset;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH);
-        }
+
+            /* Prepare for next iteration */
+            dictMatchIndexAndTag = dictHashTable[dictHashAndTag1 >> ZSTD_SHORT_CACHE_TAG_BITS];
+            dictTagsMatch = ZSTD_comparePackedTags(dictMatchIndexAndTag, dictHashAndTag1);
+            matchIndex = hashTable[hash1];
+
+            if (ip1 >= nextStep) {
+                step++;
+                nextStep += kStepIncr;
+            }
+            ip0 = ip1;
+            ip1 = ip1 + step;
+            if (ip1 > ilimit) goto _cleanup;
+
+            curr = (U32)(ip0 - base);
+            hash0 = hash1;
+        }   /* end inner search loop */
 
         /* match found */
-        ip += mLength;
-        anchor = ip;
+        assert(mLength);
+        ip0 += mLength;
+        anchor = ip0;
 
-        if (ip <= ilimit) {
+        if (ip0 <= ilimit) {
             /* Fill Table */
-            assert(base+current+2 > istart);  /* check base overflow */
-            hashTable[ZSTD_hashPtr(base+current+2, hlog, mls)] = current+2;  /* here because current+2 could be > iend-8 */
-            hashTable[ZSTD_hashPtr(ip-2, hlog, mls)] = (U32)(ip-2-base);
+            assert(base+curr+2 > istart);  /* check base overflow */
+            hashTable[ZSTD_hashPtr(base+curr+2, hlog, mls)] = curr+2;  /* here because curr+2 could be > iend-8 */
+            hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base);
 
             /* check immediate repcode */
-            while (ip <= ilimit) {
-                U32 const current2 = (U32)(ip-base);
+            while (ip0 <= ilimit) {
+                U32 const current2 = (U32)(ip0-base);
                 U32 const repIndex2 = current2 - offset_2;
                 const BYTE* repMatch2 = repIndex2 < prefixStartIndex ?
                         dictBase - dictIndexDelta + repIndex2 :
                         base + repIndex2;
                 if ( ((U32)((prefixStartIndex-1) - (U32)repIndex2) >= 3 /* intentional overflow */)
-                   && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
+                   && (MEM_read32(repMatch2) == MEM_read32(ip0))) {
                     const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
-                    size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
+                    size_t const repLength2 = ZSTD_count_2segments(ip0+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
                     U32 tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset;   /* swap offset_2 <=> offset_1 */
-                    ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, repLength2-MINMATCH);
-                    hashTable[ZSTD_hashPtr(ip, hlog, mls)] = current2;
-                    ip += repLength2;
-                    anchor = ip;
+                    ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
+                    hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = current2;
+                    ip0 += repLength2;
+                    anchor = ip0;
                     continue;
                 }
                 break;
             }
         }
+
+        /* Prepare for next iteration */
+        assert(ip0 == anchor);
+        ip1 = ip0 + stepSize;
     }
 
+_cleanup:
     /* save reps for next block */
-    rep[0] = offset_1 ? offset_1 : offsetSaved;
-    rep[1] = offset_2 ? offset_2 : offsetSaved;
+    rep[0] = offset_1;
+    rep[1] = offset_2;
 
     /* Return the last literals size */
     return (size_t)(iend - anchor);
 }
 
+
+ZSTD_GEN_FAST_FN(dictMatchState, 4, 0)
+ZSTD_GEN_FAST_FN(dictMatchState, 5, 0)
+ZSTD_GEN_FAST_FN(dictMatchState, 6, 0)
+ZSTD_GEN_FAST_FN(dictMatchState, 7, 0)
+
 size_t ZSTD_compressBlock_fast_dictMatchState(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
@@ -361,30 +670,29 @@ size_t ZSTD_compressBlock_fast_dictMatchState(
     {
     default: /* includes case 3 */
     case 4 :
-        return ZSTD_compressBlock_fast_dictMatchState_generic(ms, seqStore, rep, src, srcSize, 4);
+        return ZSTD_compressBlock_fast_dictMatchState_4_0(ms, seqStore, rep, src, srcSize);
     case 5 :
-        return ZSTD_compressBlock_fast_dictMatchState_generic(ms, seqStore, rep, src, srcSize, 5);
+        return ZSTD_compressBlock_fast_dictMatchState_5_0(ms, seqStore, rep, src, srcSize);
     case 6 :
-        return ZSTD_compressBlock_fast_dictMatchState_generic(ms, seqStore, rep, src, srcSize, 6);
+        return ZSTD_compressBlock_fast_dictMatchState_6_0(ms, seqStore, rep, src, srcSize);
     case 7 :
-        return ZSTD_compressBlock_fast_dictMatchState_generic(ms, seqStore, rep, src, srcSize, 7);
+        return ZSTD_compressBlock_fast_dictMatchState_7_0(ms, seqStore, rep, src, srcSize);
     }
 }
 
 
 static size_t ZSTD_compressBlock_fast_extDict_generic(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize, U32 const mls)
+        void const* src, size_t srcSize, U32 const mls, U32 const hasStep)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
     U32* const hashTable = ms->hashTable;
     U32 const hlog = cParams->hashLog;
     /* support stepSize of 0 */
-    U32 const stepSize = cParams->targetLength + !(cParams->targetLength);
+    size_t const stepSize = cParams->targetLength + !(cParams->targetLength) + 1;
     const BYTE* const base = ms->window.base;
     const BYTE* const dictBase = ms->window.dictBase;
     const BYTE* const istart = (const BYTE*)src;
-    const BYTE* ip = istart;
     const BYTE* anchor = istart;
     const U32   endIndex = (U32)((size_t)(istart - base) + srcSize);
     const U32   lowLimit = ZSTD_getLowestMatchIndex(ms, endIndex, cParams->windowLog);
@@ -397,100 +705,256 @@ static size_t ZSTD_compressBlock_fast_extDict_generic(
     const BYTE* const iend = istart + srcSize;
     const BYTE* const ilimit = iend - 8;
     U32 offset_1=rep[0], offset_2=rep[1];
+    U32 offsetSaved1 = 0, offsetSaved2 = 0;
+
+    const BYTE* ip0 = istart;
+    const BYTE* ip1;
+    const BYTE* ip2;
+    const BYTE* ip3;
+    U32 current0;
+
+
+    size_t hash0; /* hash for ip0 */
+    size_t hash1; /* hash for ip1 */
+    U32 idx; /* match idx for ip0 */
+    const BYTE* idxBase; /* base pointer for idx */
+
+    U32 offcode;
+    const BYTE* match0;
+    size_t mLength;
+    const BYTE* matchEnd = 0; /* initialize to avoid warning, assert != 0 later */
+
+    size_t step;
+    const BYTE* nextStep;
+    const size_t kStepIncr = (1 << (kSearchStrength - 1));
+
+    (void)hasStep; /* not currently specialized on whether it's accelerated */
 
     DEBUGLOG(5, "ZSTD_compressBlock_fast_extDict_generic (offset_1=%u)", offset_1);
 
     /* switch to "regular" variant if extDict is invalidated due to maxDistance */
     if (prefixStartIndex == dictStartIndex)
-        return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, mls);
-
-    /* Search Loop */
-    while (ip < ilimit) {  /* < instead of <=, because (ip+1) */
-        const size_t h = ZSTD_hashPtr(ip, hlog, mls);
-        const U32    matchIndex = hashTable[h];
-        const BYTE* const matchBase = matchIndex < prefixStartIndex ? dictBase : base;
-        const BYTE*  match = matchBase + matchIndex;
-        const U32    current = (U32)(ip-base);
-        const U32    repIndex = current + 1 - offset_1;
-        const BYTE* const repBase = repIndex < prefixStartIndex ? dictBase : base;
-        const BYTE* const repMatch = repBase + repIndex;
-        hashTable[h] = current;   /* update hash table */
-        DEBUGLOG(7, "offset_1 = %u , current = %u", offset_1, current);
-        assert(offset_1 <= current +1);   /* check repIndex */
-
-        if ( (((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow */ & (repIndex > dictStartIndex))
-           && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
-            const BYTE* const repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
-            size_t const rLength = ZSTD_count_2segments(ip+1 +4, repMatch +4, iend, repMatchEnd, prefixStart) + 4;
-            ip++;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, 0, rLength-MINMATCH);
-            ip += rLength;
-            anchor = ip;
-        } else {
-            if ( (matchIndex < dictStartIndex) ||
-                 (MEM_read32(match) != MEM_read32(ip)) ) {
-                assert(stepSize >= 1);
-                ip += ((ip-anchor) >> kSearchStrength) + stepSize;
-                continue;
+        return ZSTD_compressBlock_fast(ms, seqStore, rep, src, srcSize);
+
+    {   U32 const curr = (U32)(ip0 - base);
+        U32 const maxRep = curr - dictStartIndex;
+        if (offset_2 >= maxRep) offsetSaved2 = offset_2, offset_2 = 0;
+        if (offset_1 >= maxRep) offsetSaved1 = offset_1, offset_1 = 0;
+    }
+
+    /* start each op */
+_start: /* Requires: ip0 */
+
+    step = stepSize;
+    nextStep = ip0 + kStepIncr;
+
+    /* calculate positions, ip0 - anchor == 0, so we skip step calc */
+    ip1 = ip0 + 1;
+    ip2 = ip0 + step;
+    ip3 = ip2 + 1;
+
+    if (ip3 >= ilimit) {
+        goto _cleanup;
+    }
+
+    hash0 = ZSTD_hashPtr(ip0, hlog, mls);
+    hash1 = ZSTD_hashPtr(ip1, hlog, mls);
+
+    idx = hashTable[hash0];
+    idxBase = idx < prefixStartIndex ? dictBase : base;
+
+    do {
+        {   /* load repcode match for ip[2] */
+            U32 const current2 = (U32)(ip2 - base);
+            U32 const repIndex = current2 - offset_1;
+            const BYTE* const repBase = repIndex < prefixStartIndex ? dictBase : base;
+            U32 rval;
+            if ( ((U32)(prefixStartIndex - repIndex) >= 4) /* intentional underflow */
+                 & (offset_1 > 0) ) {
+                rval = MEM_read32(repBase + repIndex);
+            } else {
+                rval = MEM_read32(ip2) ^ 1; /* guaranteed to not match. */
             }
-            {   const BYTE* const matchEnd = matchIndex < prefixStartIndex ? dictEnd : iend;
-                const BYTE* const lowMatchPtr = matchIndex < prefixStartIndex ? dictStart : prefixStart;
-                U32 const offset = current - matchIndex;
-                size_t mLength = ZSTD_count_2segments(ip+4, match+4, iend, matchEnd, prefixStart) + 4;
-                while (((ip>anchor) & (match>lowMatchPtr)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; }   /* catch up */
-                offset_2 = offset_1; offset_1 = offset;  /* update offset history */
-                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH);
-                ip += mLength;
-                anchor = ip;
+
+            /* write back hash table entry */
+            current0 = (U32)(ip0 - base);
+            hashTable[hash0] = current0;
+
+            /* check repcode at ip[2] */
+            if (MEM_read32(ip2) == rval) {
+                ip0 = ip2;
+                match0 = repBase + repIndex;
+                matchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
+                assert((match0 != prefixStart) & (match0 != dictStart));
+                mLength = ip0[-1] == match0[-1];
+                ip0 -= mLength;
+                match0 -= mLength;
+                offcode = REPCODE1_TO_OFFBASE;
+                mLength += 4;
+                goto _match;
         }   }
 
-        if (ip <= ilimit) {
-            /* Fill Table */
-            hashTable[ZSTD_hashPtr(base+current+2, hlog, mls)] = current+2;
-            hashTable[ZSTD_hashPtr(ip-2, hlog, mls)] = (U32)(ip-2-base);
-            /* check immediate repcode */
-            while (ip <= ilimit) {
-                U32 const current2 = (U32)(ip-base);
-                U32 const repIndex2 = current2 - offset_2;
-                const BYTE* const repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2;
-                if ( (((U32)((prefixStartIndex-1) - repIndex2) >= 3) & (repIndex2 > dictStartIndex))  /* intentional overflow */
-                   && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
-                    const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
-                    size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
-                    { U32 const tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; }  /* swap offset_2 <=> offset_1 */
-                    ZSTD_storeSeq(seqStore, 0 /*litlen*/, anchor, iend, 0 /*offcode*/, repLength2-MINMATCH);
-                    hashTable[ZSTD_hashPtr(ip, hlog, mls)] = current2;
-                    ip += repLength2;
-                    anchor = ip;
-                    continue;
-                }
-                break;
-    }   }   }
+        {   /* load match for ip[0] */
+            U32 const mval = idx >= dictStartIndex ?
+                    MEM_read32(idxBase + idx) :
+                    MEM_read32(ip0) ^ 1; /* guaranteed not to match */
+
+            /* check match at ip[0] */
+            if (MEM_read32(ip0) == mval) {
+                /* found a match! */
+                goto _offset;
+        }   }
+
+        /* lookup ip[1] */
+        idx = hashTable[hash1];
+        idxBase = idx < prefixStartIndex ? dictBase : base;
+
+        /* hash ip[2] */
+        hash0 = hash1;
+        hash1 = ZSTD_hashPtr(ip2, hlog, mls);
+
+        /* advance to next positions */
+        ip0 = ip1;
+        ip1 = ip2;
+        ip2 = ip3;
+
+        /* write back hash table entry */
+        current0 = (U32)(ip0 - base);
+        hashTable[hash0] = current0;
+
+        {   /* load match for ip[0] */
+            U32 const mval = idx >= dictStartIndex ?
+                    MEM_read32(idxBase + idx) :
+                    MEM_read32(ip0) ^ 1; /* guaranteed not to match */
+
+            /* check match at ip[0] */
+            if (MEM_read32(ip0) == mval) {
+                /* found a match! */
+                goto _offset;
+        }   }
+
+        /* lookup ip[1] */
+        idx = hashTable[hash1];
+        idxBase = idx < prefixStartIndex ? dictBase : base;
+
+        /* hash ip[2] */
+        hash0 = hash1;
+        hash1 = ZSTD_hashPtr(ip2, hlog, mls);
+
+        /* advance to next positions */
+        ip0 = ip1;
+        ip1 = ip2;
+        ip2 = ip0 + step;
+        ip3 = ip1 + step;
+
+        /* calculate step */
+        if (ip2 >= nextStep) {
+            step++;
+            PREFETCH_L1(ip1 + 64);
+            PREFETCH_L1(ip1 + 128);
+            nextStep += kStepIncr;
+        }
+    } while (ip3 < ilimit);
+
+_cleanup:
+    /* Note that there are probably still a couple positions we could search.
+     * However, it seems to be a meaningful performance hit to try to search
+     * them. So let's not. */
+
+    /* If offset_1 started invalid (offsetSaved1 != 0) and became valid (offset_1 != 0),
+     * rotate saved offsets. See comment in ZSTD_compressBlock_fast_noDict for more context. */
+    offsetSaved2 = ((offsetSaved1 != 0) && (offset_1 != 0)) ? offsetSaved1 : offsetSaved2;
 
     /* save reps for next block */
-    rep[0] = offset_1;
-    rep[1] = offset_2;
+    rep[0] = offset_1 ? offset_1 : offsetSaved1;
+    rep[1] = offset_2 ? offset_2 : offsetSaved2;
 
     /* Return the last literals size */
     return (size_t)(iend - anchor);
+
+_offset: /* Requires: ip0, idx, idxBase */
+
+    /* Compute the offset code. */
+    {   U32 const offset = current0 - idx;
+        const BYTE* const lowMatchPtr = idx < prefixStartIndex ? dictStart : prefixStart;
+        matchEnd = idx < prefixStartIndex ? dictEnd : iend;
+        match0 = idxBase + idx;
+        offset_2 = offset_1;
+        offset_1 = offset;
+        offcode = OFFSET_TO_OFFBASE(offset);
+        mLength = 4;
+
+        /* Count the backwards match length. */
+        while (((ip0>anchor) & (match0>lowMatchPtr)) && (ip0[-1] == match0[-1])) {
+            ip0--;
+            match0--;
+            mLength++;
+    }   }
+
+_match: /* Requires: ip0, match0, offcode, matchEnd */
+
+    /* Count the forward length. */
+    assert(matchEnd != 0);
+    mLength += ZSTD_count_2segments(ip0 + mLength, match0 + mLength, iend, matchEnd, prefixStart);
+
+    ZSTD_storeSeq(seqStore, (size_t)(ip0 - anchor), anchor, iend, offcode, mLength);
+
+    ip0 += mLength;
+    anchor = ip0;
+
+    /* write next hash table entry */
+    if (ip1 < ip0) {
+        hashTable[hash1] = (U32)(ip1 - base);
+    }
+
+    /* Fill table and check for immediate repcode. */
+    if (ip0 <= ilimit) {
+        /* Fill Table */
+        assert(base+current0+2 > istart);  /* check base overflow */
+        hashTable[ZSTD_hashPtr(base+current0+2, hlog, mls)] = current0+2;  /* here because current+2 could be > iend-8 */
+        hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base);
+
+        while (ip0 <= ilimit) {
+            U32 const repIndex2 = (U32)(ip0-base) - offset_2;
+            const BYTE* const repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2;
+            if ( (((U32)((prefixStartIndex-1) - repIndex2) >= 3) & (offset_2 > 0))  /* intentional underflow */
+                 && (MEM_read32(repMatch2) == MEM_read32(ip0)) ) {
+                const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
+                size_t const repLength2 = ZSTD_count_2segments(ip0+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
+                { U32 const tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; }  /* swap offset_2 <=> offset_1 */
+                ZSTD_storeSeq(seqStore, 0 /*litlen*/, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
+                hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = (U32)(ip0-base);
+                ip0 += repLength2;
+                anchor = ip0;
+                continue;
+            }
+            break;
+    }   }
+
+    goto _start;
 }
 
+ZSTD_GEN_FAST_FN(extDict, 4, 0)
+ZSTD_GEN_FAST_FN(extDict, 5, 0)
+ZSTD_GEN_FAST_FN(extDict, 6, 0)
+ZSTD_GEN_FAST_FN(extDict, 7, 0)
 
 size_t ZSTD_compressBlock_fast_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
 {
     U32 const mls = ms->cParams.minMatch;
+    assert(ms->dictMatchState == NULL);
     switch(mls)
     {
     default: /* includes case 3 */
     case 4 :
-        return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, 4);
+        return ZSTD_compressBlock_fast_extDict_4_0(ms, seqStore, rep, src, srcSize);
     case 5 :
-        return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, 5);
+        return ZSTD_compressBlock_fast_extDict_5_0(ms, seqStore, rep, src, srcSize);
     case 6 :
-        return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, 6);
+        return ZSTD_compressBlock_fast_extDict_6_0(ms, seqStore, rep, src, srcSize);
     case 7 :
-        return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, 7);
+        return ZSTD_compressBlock_fast_extDict_7_0(ms, seqStore, rep, src, srcSize);
     }
 }

src/borg/algorithms/zstd/lib/compress/zstd_fast.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -19,7 +19,8 @@ extern "C" {
 #include "zstd_compress_internal.h"
 
 void ZSTD_fillHashTable(ZSTD_matchState_t* ms,
-                        void const* end, ZSTD_dictTableLoadMethod_e dtlm);
+                        void const* end, ZSTD_dictTableLoadMethod_e dtlm,
+                        ZSTD_tableFillPurpose_e tfp);
 size_t ZSTD_compressBlock_fast(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);

src/borg/algorithms/zstd/lib/compress/zstd_lazy.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -17,7 +17,20 @@ extern "C" {
 
 #include "zstd_compress_internal.h"
 
+/**
+ * Dedicated Dictionary Search Structure bucket log. In the
+ * ZSTD_dedicatedDictSearch mode, the hashTable has
+ * 2 ** ZSTD_LAZY_DDSS_BUCKET_LOG entries in each bucket, rather than just
+ * one.
+ */
+#define ZSTD_LAZY_DDSS_BUCKET_LOG 2
+
+#define ZSTD_ROW_HASH_TAG_BITS 8        /* nb bits to use for the tag */
+
 U32 ZSTD_insertAndFindFirstIndex(ZSTD_matchState_t* ms, const BYTE* ip);
+void ZSTD_row_update(ZSTD_matchState_t* const ms, const BYTE* ip);
+
+void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t* ms, const BYTE* const ip);
 
 void ZSTD_preserveUnsortedMark (U32* const table, U32 const size, U32 const reducerValue);  /*! used in ZSTD_reduceIndex(). preemptively increase value of ZSTD_DUBT_UNSORTED_MARK */
 
@@ -33,6 +46,15 @@ size_t ZSTD_compressBlock_lazy(
 size_t ZSTD_compressBlock_greedy(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy2_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_greedy_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
 
 size_t ZSTD_compressBlock_btlazy2_dictMatchState(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
@@ -46,6 +68,34 @@ size_t ZSTD_compressBlock_lazy_dictMatchState(
 size_t ZSTD_compressBlock_greedy_dictMatchState(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy2_dictMatchState_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy_dictMatchState_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_greedy_dictMatchState_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+
+size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy_dedicatedDictSearch(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_greedy_dedicatedDictSearch(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy_dedicatedDictSearch_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_greedy_dedicatedDictSearch_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
 
 size_t ZSTD_compressBlock_greedy_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
@@ -56,10 +106,20 @@ size_t ZSTD_compressBlock_lazy_extDict(
 size_t ZSTD_compressBlock_lazy2_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_greedy_extDict_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy_extDict_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy2_extDict_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
 size_t ZSTD_compressBlock_btlazy2_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
 
+
 #if defined (__cplusplus)
 }
 #endif

src/borg/algorithms/zstd/lib/compress/zstd_ldm.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -11,13 +11,126 @@
 #include "zstd_ldm.h"
 
 #include "../common/debug.h"
+#include "../common/xxhash.h"
 #include "zstd_fast.h"          /* ZSTD_fillHashTable() */
 #include "zstd_double_fast.h"   /* ZSTD_fillDoubleHashTable() */
+#include "zstd_ldm_geartab.h"
 
 #define LDM_BUCKET_SIZE_LOG 3
 #define LDM_MIN_MATCH_LENGTH 64
 #define LDM_HASH_RLOG 7
-#define LDM_HASH_CHAR_OFFSET 10
+
+typedef struct {
+    U64 rolling;
+    U64 stopMask;
+} ldmRollingHashState_t;
+
+/** ZSTD_ldm_gear_init():
+ *
+ * Initializes the rolling hash state such that it will honor the
+ * settings in params. */
+static void ZSTD_ldm_gear_init(ldmRollingHashState_t* state, ldmParams_t const* params)
+{
+    unsigned maxBitsInMask = MIN(params->minMatchLength, 64);
+    unsigned hashRateLog = params->hashRateLog;
+
+    state->rolling = ~(U32)0;
+
+    /* The choice of the splitting criterion is subject to two conditions:
+     *   1. it has to trigger on average every 2^(hashRateLog) bytes;
+     *   2. ideally, it has to depend on a window of minMatchLength bytes.
+     *
+     * In the gear hash algorithm, bit n depends on the last n bytes;
+     * so in order to obtain a good quality splitting criterion it is
+     * preferable to use bits with high weight.
+     *
+     * To match condition 1 we use a mask with hashRateLog bits set
+     * and, because of the previous remark, we make sure these bits
+     * have the highest possible weight while still respecting
+     * condition 2.
+     */
+    if (hashRateLog > 0 && hashRateLog <= maxBitsInMask) {
+        state->stopMask = (((U64)1 << hashRateLog) - 1) << (maxBitsInMask - hashRateLog);
+    } else {
+        /* In this degenerate case we simply honor the hash rate. */
+        state->stopMask = ((U64)1 << hashRateLog) - 1;
+    }
+}
+
+/** ZSTD_ldm_gear_reset()
+ * Feeds [data, data + minMatchLength) into the hash without registering any
+ * splits. This effectively resets the hash state. This is used when skipping
+ * over data, either at the beginning of a block, or skipping sections.
+ */
+static void ZSTD_ldm_gear_reset(ldmRollingHashState_t* state,
+                                BYTE const* data, size_t minMatchLength)
+{
+    U64 hash = state->rolling;
+    size_t n = 0;
+
+#define GEAR_ITER_ONCE() do {                                  \
+        hash = (hash << 1) + ZSTD_ldm_gearTab[data[n] & 0xff]; \
+        n += 1;                                                \
+    } while (0)
+    while (n + 3 < minMatchLength) {
+        GEAR_ITER_ONCE();
+        GEAR_ITER_ONCE();
+        GEAR_ITER_ONCE();
+        GEAR_ITER_ONCE();
+    }
+    while (n < minMatchLength) {
+        GEAR_ITER_ONCE();
+    }
+#undef GEAR_ITER_ONCE
+}
+
+/** ZSTD_ldm_gear_feed():
+ *
+ * Registers in the splits array all the split points found in the first
+ * size bytes following the data pointer. This function terminates when
+ * either all the data has been processed or LDM_BATCH_SIZE splits are
+ * present in the splits array.
+ *
+ * Precondition: The splits array must not be full.
+ * Returns: The number of bytes processed. */
+static size_t ZSTD_ldm_gear_feed(ldmRollingHashState_t* state,
+                                 BYTE const* data, size_t size,
+                                 size_t* splits, unsigned* numSplits)
+{
+    size_t n;
+    U64 hash, mask;
+
+    hash = state->rolling;
+    mask = state->stopMask;
+    n = 0;
+
+#define GEAR_ITER_ONCE() do { \
+        hash = (hash << 1) + ZSTD_ldm_gearTab[data[n] & 0xff]; \
+        n += 1; \
+        if (UNLIKELY((hash & mask) == 0)) { \
+            splits[*numSplits] = n; \
+            *numSplits += 1; \
+            if (*numSplits == LDM_BATCH_SIZE) \
+                goto done; \
+        } \
+    } while (0)
+
+    while (n + 3 < size) {
+        GEAR_ITER_ONCE();
+        GEAR_ITER_ONCE();
+        GEAR_ITER_ONCE();
+        GEAR_ITER_ONCE();
+    }
+    while (n < size) {
+        GEAR_ITER_ONCE();
+    }
+
+#undef GEAR_ITER_ONCE
+
+done:
+    state->rolling = hash;
+    return n;
+}
 
 void ZSTD_ldm_adjustParameters(ldmParams_t* params,
                                ZSTD_compressionParameters const* cParams)
@@ -27,13 +140,6 @@ void ZSTD_ldm_adjustParameters(ldmParams_t* params,
     DEBUGLOG(4, "ZSTD_ldm_adjustParameters");
     if (!params->bucketSizeLog) params->bucketSizeLog = LDM_BUCKET_SIZE_LOG;
     if (!params->minMatchLength) params->minMatchLength = LDM_MIN_MATCH_LENGTH;
-    if (cParams->strategy >= ZSTD_btopt) {
-      /* Get out of the way of the optimal parser */
-      U32 const minMatch = MAX(cParams->targetLength, params->minMatchLength);
-      assert(minMatch >= ZSTD_LDM_MINMATCH_MIN);
-      assert(minMatch <= ZSTD_LDM_MINMATCH_MAX);
-      params->minMatchLength = minMatch;
-    }
     if (params->hashLog == 0) {
         params->hashLog = MAX(ZSTD_HASHLOG_MIN, params->windowLog - LDM_HASH_RLOG);
         assert(params->hashLog <= ZSTD_HASHLOG_MAX);
@@ -53,47 +159,12 @@ size_t ZSTD_ldm_getTableSize(ldmParams_t params)
     size_t const ldmBucketSize = ((size_t)1) << (params.hashLog - ldmBucketSizeLog);
     size_t const totalSize = ZSTD_cwksp_alloc_size(ldmBucketSize)
                            + ZSTD_cwksp_alloc_size(ldmHSize * sizeof(ldmEntry_t));
-    return params.enableLdm ? totalSize : 0;
+    return params.enableLdm == ZSTD_ps_enable ? totalSize : 0;
 }
 
 size_t ZSTD_ldm_getMaxNbSeq(ldmParams_t params, size_t maxChunkSize)
 {
-    return params.enableLdm ? (maxChunkSize / params.minMatchLength) : 0;
-}
-
-/** ZSTD_ldm_getSmallHash() :
- *  numBits should be <= 32
- *  If numBits==0, returns 0.
- *  @return : the most significant numBits of value. */
-static U32 ZSTD_ldm_getSmallHash(U64 value, U32 numBits)
-{
-    assert(numBits <= 32);
-    return numBits == 0 ? 0 : (U32)(value >> (64 - numBits));
-}
-
-/** ZSTD_ldm_getChecksum() :
- *  numBitsToDiscard should be <= 32
- *  @return : the next most significant 32 bits after numBitsToDiscard */
-static U32 ZSTD_ldm_getChecksum(U64 hash, U32 numBitsToDiscard)
-{
-    assert(numBitsToDiscard <= 32);
-    return (hash >> (64 - 32 - numBitsToDiscard)) & 0xFFFFFFFF;
-}
-
-/** ZSTD_ldm_getTag() ;
- *  Given the hash, returns the most significant numTagBits bits
- *  after (32 + hbits) bits.
- *
- *  If there are not enough bits remaining, return the last
- *  numTagBits bits. */
-static U32 ZSTD_ldm_getTag(U64 hash, U32 hbits, U32 numTagBits)
-{
-    assert(numTagBits < 32 && hbits <= 32);
-    if (32 - hbits < numTagBits) {
-        return hash & (((U32)1 << numTagBits) - 1);
-    } else {
-        return (hash >> (32 - hbits - numTagBits)) & (((U32)1 << numTagBits) - 1);
-    }
+    return params.enableLdm == ZSTD_ps_enable ? (maxChunkSize / params.minMatchLength) : 0;
 }
 
 /** ZSTD_ldm_getBucket() :
@@ -110,38 +181,12 @@ static void ZSTD_ldm_insertEntry(ldmState_t* ldmState,
                                  size_t const hash, const ldmEntry_t entry,
                                  ldmParams_t const ldmParams)
 {
-    BYTE* const bucketOffsets = ldmState->bucketOffsets;
-    *(ZSTD_ldm_getBucket(ldmState, hash, ldmParams) + bucketOffsets[hash]) = entry;
-    bucketOffsets[hash]++;
-    bucketOffsets[hash] &= ((U32)1 << ldmParams.bucketSizeLog) - 1;
-}
+    BYTE* const pOffset = ldmState->bucketOffsets + hash;
+    unsigned const offset = *pOffset;
+
+    *(ZSTD_ldm_getBucket(ldmState, hash, ldmParams) + offset) = entry;
+    *pOffset = (BYTE)((offset + 1) & ((1u << ldmParams.bucketSizeLog) - 1));
 
-/** ZSTD_ldm_makeEntryAndInsertByTag() :
- *
- *  Gets the small hash, checksum, and tag from the rollingHash.
- *
- *  If the tag matches (1 << ldmParams.hashRateLog)-1, then
- *  creates an ldmEntry from the offset, and inserts it into the hash table.
- *
- *  hBits is the length of the small hash, which is the most significant hBits
- *  of rollingHash. The checksum is the next 32 most significant bits, followed
- *  by ldmParams.hashRateLog bits that make up the tag. */
-static void ZSTD_ldm_makeEntryAndInsertByTag(ldmState_t* ldmState,
-                                             U64 const rollingHash,
-                                             U32 const hBits,
-                                             U32 const offset,
-                                             ldmParams_t const ldmParams)
-{
-    U32 const tag = ZSTD_ldm_getTag(rollingHash, hBits, ldmParams.hashRateLog);
-    U32 const tagMask = ((U32)1 << ldmParams.hashRateLog) - 1;
-    if (tag == tagMask) {
-        U32 const hash = ZSTD_ldm_getSmallHash(rollingHash, hBits);
-        U32 const checksum = ZSTD_ldm_getChecksum(rollingHash, hBits);
-        ldmEntry_t entry;
-        entry.offset = offset;
-        entry.checksum = checksum;
-        ZSTD_ldm_insertEntry(ldmState, hash, entry, ldmParams);
-    }
 }
 
 /** ZSTD_ldm_countBackwardsMatch() :
@@ -150,10 +195,10 @@ static void ZSTD_ldm_makeEntryAndInsertByTag(ldmState_t* ldmState,
  *  We count only bytes where pMatch >= pBase and pIn >= pAnchor. */
 static size_t ZSTD_ldm_countBackwardsMatch(
             const BYTE* pIn, const BYTE* pAnchor,
-            const BYTE* pMatch, const BYTE* pBase)
+            const BYTE* pMatch, const BYTE* pMatchBase)
 {
     size_t matchLength = 0;
-    while (pIn > pAnchor && pMatch > pBase && pIn[-1] == pMatch[-1]) {
+    while (pIn > pAnchor && pMatch > pMatchBase && pIn[-1] == pMatch[-1]) {
         pIn--;
         pMatch--;
         matchLength++;
@@ -161,6 +206,27 @@ static size_t ZSTD_ldm_countBackwardsMatch(
     return matchLength;
 }
 
+/** ZSTD_ldm_countBackwardsMatch_2segments() :
+ *  Returns the number of bytes that match backwards from pMatch,
+ *  even with the backwards match spanning 2 different segments.
+ *
+ *  On reaching `pMatchBase`, start counting from mEnd */
+static size_t ZSTD_ldm_countBackwardsMatch_2segments(
+                    const BYTE* pIn, const BYTE* pAnchor,
+                    const BYTE* pMatch, const BYTE* pMatchBase,
+                    const BYTE* pExtDictStart, const BYTE* pExtDictEnd)
+{
+    size_t matchLength = ZSTD_ldm_countBackwardsMatch(pIn, pAnchor, pMatch, pMatchBase);
+    if (pMatch - matchLength != pMatchBase || pMatchBase == pExtDictStart) {
+        /* If backwards match is entirely in the extDict or prefix, immediately return */
+        return matchLength;
+    }
+    DEBUGLOG(7, "ZSTD_ldm_countBackwardsMatch_2segments: found 2-parts backwards match (length in prefix==%zu)", matchLength);
+    matchLength += ZSTD_ldm_countBackwardsMatch(pIn - matchLength, pAnchor, pExtDictEnd, pExtDictStart);
+    DEBUGLOG(7, "final backwards match length = %zu", matchLength);
+    return matchLength;
+}
+
 /** ZSTD_ldm_fillFastTables() :
  *
  *  Fills the relevant tables for the ZSTD_fast and ZSTD_dfast strategies.
@@ -176,11 +242,11 @@ static size_t ZSTD_ldm_fillFastTables(ZSTD_matchState_t* ms,
     switch(ms->cParams.strategy)
     {
     case ZSTD_fast:
-        ZSTD_fillHashTable(ms, iend, ZSTD_dtlm_fast);
+        ZSTD_fillHashTable(ms, iend, ZSTD_dtlm_fast, ZSTD_tfp_forCCtx);
         break;
 
     case ZSTD_dfast:
-        ZSTD_fillDoubleHashTable(ms, iend, ZSTD_dtlm_fast);
+        ZSTD_fillDoubleHashTable(ms, iend, ZSTD_dtlm_fast, ZSTD_tfp_forCCtx);
         break;
 
     case ZSTD_greedy:
@@ -198,43 +264,42 @@ static size_t ZSTD_ldm_fillFastTables(ZSTD_matchState_t* ms,
     return 0;
 }
 
-/** ZSTD_ldm_fillLdmHashTable() :
- *
- *  Fills hashTable from (lastHashed + 1) to iend (non-inclusive).
- *  lastHash is the rolling hash that corresponds to lastHashed.
- *
- *  Returns the rolling hash corresponding to position iend-1. */
-static U64 ZSTD_ldm_fillLdmHashTable(ldmState_t* state,
-                                     U64 lastHash, const BYTE* lastHashed,
-                                     const BYTE* iend, const BYTE* base,
-                                     U32 hBits, ldmParams_t const ldmParams)
-{
-    U64 rollingHash = lastHash;
-    const BYTE* cur = lastHashed + 1;
-
-    while (cur < iend) {
-        rollingHash = ZSTD_rollingHash_rotate(rollingHash, cur[-1],
-                                              cur[ldmParams.minMatchLength-1],
-                                              state->hashPower);
-        ZSTD_ldm_makeEntryAndInsertByTag(state,
-                                         rollingHash, hBits,
-                                         (U32)(cur - base), ldmParams);
-        ++cur;
-    }
-    return rollingHash;
-}
-
 void ZSTD_ldm_fillHashTable(
-            ldmState_t* state, const BYTE* ip,
+            ldmState_t* ldmState, const BYTE* ip,
             const BYTE* iend, ldmParams_t const* params)
 {
+    U32 const minMatchLength = params->minMatchLength;
+    U32 const hBits = params->hashLog - params->bucketSizeLog;
+    BYTE const* const base = ldmState->window.base;
+    BYTE const* const istart = ip;
+    ldmRollingHashState_t hashState;
+    size_t* const splits = ldmState->splitIndices;
+    unsigned numSplits;
+
     DEBUGLOG(5, "ZSTD_ldm_fillHashTable");
-    if ((size_t)(iend - ip) >= params->minMatchLength) {
-        U64 startingHash = ZSTD_rollingHash_compute(ip, params->minMatchLength);
-        ZSTD_ldm_fillLdmHashTable(
-            state, startingHash, ip, iend - params->minMatchLength, state->window.base,
-            params->hashLog - params->bucketSizeLog,
-            *params);
+
+    ZSTD_ldm_gear_init(&hashState, params);
+    while (ip < iend) {
+        size_t hashed;
+        unsigned n;
+
+        numSplits = 0;
+        hashed = ZSTD_ldm_gear_feed(&hashState, ip, iend - ip, splits, &numSplits);
+
+        for (n = 0; n < numSplits; n++) {
+            if (ip + splits[n] >= istart + minMatchLength) {
+                BYTE const* const split = ip + splits[n] - minMatchLength;
+                U64 const xxhash = XXH64(split, minMatchLength, 0);
+                U32 const hash = (U32)(xxhash & (((U32)1 << hBits) - 1));
+                ldmEntry_t entry;
+
+                entry.offset = (U32)(split - base);
+                entry.checksum = (U32)(xxhash >> 32);
+                ZSTD_ldm_insertEntry(ldmState, hash, entry, *params);
+            }
+        }
+
+        ip += hashed;
     }
 }
 
@@ -246,10 +311,10 @@ void ZSTD_ldm_fillHashTable(
  *  (after a long match, only update tables a limited amount). */
 static void ZSTD_ldm_limitTableUpdate(ZSTD_matchState_t* ms, const BYTE* anchor)
 {
-    U32 const current = (U32)(anchor - ms->window.base);
-    if (current > ms->nextToUpdate + 1024) {
+    U32 const curr = (U32)(anchor - ms->window.base);
+    if (curr > ms->nextToUpdate + 1024) {
         ms->nextToUpdate =
-            current - MIN(512, current - ms->nextToUpdate - 1024);
+            curr - MIN(512, curr - ms->nextToUpdate - 1024);
     }
 }
 
@@ -260,11 +325,8 @@ static size_t ZSTD_ldm_generateSequences_internal(
     /* LDM parameters */
     int const extDict = ZSTD_window_hasExtDict(ldmState->window);
     U32 const minMatchLength = params->minMatchLength;
-    U64 const hashPower = ldmState->hashPower;
+    U32 const entsPerBucket = 1U << params->bucketSizeLog;
     U32 const hBits = params->hashLog - params->bucketSizeLog;
-    U32 const ldmBucketSize = 1U << params->bucketSizeLog;
-    U32 const hashRateLog = params->hashRateLog;
-    U32 const ldmTagMask = (1U << params->hashRateLog) - 1;
     /* Prefix and extDict parameters */
     U32 const dictLimit = ldmState->window.dictLimit;
     U32 const lowestIndex = extDict ? ldmState->window.lowLimit : dictLimit;
@@ -276,45 +338,69 @@ static size_t ZSTD_ldm_generateSequences_internal(
     /* Input bounds */
     BYTE const* const istart = (BYTE const*)src;
     BYTE const* const iend = istart + srcSize;
-    BYTE const* const ilimit = iend - MAX(minMatchLength, HASH_READ_SIZE);
+    BYTE const* const ilimit = iend - HASH_READ_SIZE;
     /* Input positions */
     BYTE const* anchor = istart;
     BYTE const* ip = istart;
-    /* Rolling hash */
-    BYTE const* lastHashed = NULL;
-    U64 rollingHash = 0;
-
-    while (ip <= ilimit) {
-        size_t mLength;
-        U32 const current = (U32)(ip - base);
-        size_t forwardMatchLength = 0, backwardMatchLength = 0;
-        ldmEntry_t* bestEntry = NULL;
-        if (ip != istart) {
-            rollingHash = ZSTD_rollingHash_rotate(rollingHash, lastHashed[0],
-                                                  lastHashed[minMatchLength],
-                                                  hashPower);
-        } else {
-            rollingHash = ZSTD_rollingHash_compute(ip, minMatchLength);
+    /* Rolling hash state */
+    ldmRollingHashState_t hashState;
+    /* Arrays for staged-processing */
+    size_t* const splits = ldmState->splitIndices;
+    ldmMatchCandidate_t* const candidates = ldmState->matchCandidates;
+    unsigned numSplits;
+
+    if (srcSize < minMatchLength)
+        return iend - anchor;
+
+    /* Initialize the rolling hash state with the first minMatchLength bytes */
+    ZSTD_ldm_gear_init(&hashState, params);
+    ZSTD_ldm_gear_reset(&hashState, ip, minMatchLength);
+    ip += minMatchLength;
+
+    while (ip < ilimit) {
+        size_t hashed;
+        unsigned n;
+
+        numSplits = 0;
+        hashed = ZSTD_ldm_gear_feed(&hashState, ip, ilimit - ip,
+                                    splits, &numSplits);
+
+        for (n = 0; n < numSplits; n++) {
+            BYTE const* const split = ip + splits[n] - minMatchLength;
+            U64 const xxhash = XXH64(split, minMatchLength, 0);
+            U32 const hash = (U32)(xxhash & (((U32)1 << hBits) - 1));
+
+            candidates[n].split = split;
+            candidates[n].hash = hash;
+            candidates[n].checksum = (U32)(xxhash >> 32);
+            candidates[n].bucket = ZSTD_ldm_getBucket(ldmState, hash, *params);
+            PREFETCH_L1(candidates[n].bucket);
         }
-        lastHashed = ip;
 
-        /* Do not insert and do not look for a match */
-        if (ZSTD_ldm_getTag(rollingHash, hBits, hashRateLog) != ldmTagMask) {
-           ip++;
-           continue;
-        }
+        for (n = 0; n < numSplits; n++) {
+            size_t forwardMatchLength = 0, backwardMatchLength = 0,
+                   bestMatchLength = 0, mLength;
+            U32 offset;
+            BYTE const* const split = candidates[n].split;
+            U32 const checksum = candidates[n].checksum;
+            U32 const hash = candidates[n].hash;
+            ldmEntry_t* const bucket = candidates[n].bucket;
+            ldmEntry_t const* cur;
+            ldmEntry_t const* bestEntry = NULL;
+            ldmEntry_t newEntry;
+
+            newEntry.offset = (U32)(split - base);
+            newEntry.checksum = checksum;
+
+            /* If a split point would generate a sequence overlapping with
+             * the previous one, we merely register it in the hash table and
+             * move on */
+            if (split < anchor) {
+                ZSTD_ldm_insertEntry(ldmState, hash, newEntry, *params);
+                continue;
+            }
 
-        /* Get the best entry and compute the match lengths */
-        {
-            ldmEntry_t* const bucket =
-                ZSTD_ldm_getBucket(ldmState,
-                                   ZSTD_ldm_getSmallHash(rollingHash, hBits),
-                                   *params);
-            ldmEntry_t* cur;
-            size_t bestMatchLength = 0;
-            U32 const checksum = ZSTD_ldm_getChecksum(rollingHash, hBits);
-
-            for (cur = bucket; cur < bucket + ldmBucketSize; ++cur) {
+            for (cur = bucket; cur < bucket + entsPerBucket; cur++) {
                 size_t curForwardMatchLength, curBackwardMatchLength,
                        curTotalMatchLength;
                 if (cur->checksum != checksum || cur->offset <= lowestIndex) {
@@ -328,30 +414,23 @@ static size_t ZSTD_ldm_generateSequences_internal(
                         cur->offset < dictLimit ? dictEnd : iend;
                     BYTE const* const lowMatchPtr =
                         cur->offset < dictLimit ? dictStart : lowPrefixPtr;
-
-                    curForwardMatchLength = ZSTD_count_2segments(
-                                                ip, pMatch, iend,
-                                                matchEnd, lowPrefixPtr);
+                    curForwardMatchLength =
+                        ZSTD_count_2segments(split, pMatch, iend, matchEnd, lowPrefixPtr);
                     if (curForwardMatchLength < minMatchLength) {
                         continue;
                     }
-                    curBackwardMatchLength =
-                        ZSTD_ldm_countBackwardsMatch(ip, anchor, pMatch,
-                                                     lowMatchPtr);
-                    curTotalMatchLength = curForwardMatchLength +
-                                          curBackwardMatchLength;
+                    curBackwardMatchLength = ZSTD_ldm_countBackwardsMatch_2segments(
+                            split, anchor, pMatch, lowMatchPtr, dictStart, dictEnd);
                 } else { /* !extDict */
                     BYTE const* const pMatch = base + cur->offset;
-                    curForwardMatchLength = ZSTD_count(ip, pMatch, iend);
+                    curForwardMatchLength = ZSTD_count(split, pMatch, iend);
                     if (curForwardMatchLength < minMatchLength) {
                         continue;
                     }
                     curBackwardMatchLength =
-                        ZSTD_ldm_countBackwardsMatch(ip, anchor, pMatch,
-                                                     lowPrefixPtr);
-                    curTotalMatchLength = curForwardMatchLength +
-                                          curBackwardMatchLength;
+                        ZSTD_ldm_countBackwardsMatch(split, anchor, pMatch, lowPrefixPtr);
                 }
+                curTotalMatchLength = curForwardMatchLength + curBackwardMatchLength;
 
                 if (curTotalMatchLength > bestMatchLength) {
                     bestMatchLength = curTotalMatchLength;
@@ -360,57 +439,54 @@ static size_t ZSTD_ldm_generateSequences_internal(
                     bestEntry = cur;
                 }
             }
-        }
-
-        /* No match found -- continue searching */
-        if (bestEntry == NULL) {
-            ZSTD_ldm_makeEntryAndInsertByTag(ldmState, rollingHash,
-                                             hBits, current,
-                                             *params);
-            ip++;
-            continue;
-        }
 
-        /* Match found */
-        mLength = forwardMatchLength + backwardMatchLength;
-        ip -= backwardMatchLength;
+            /* No match found -- insert an entry into the hash table
+             * and process the next candidate match */
+            if (bestEntry == NULL) {
+                ZSTD_ldm_insertEntry(ldmState, hash, newEntry, *params);
+                continue;
+            }
 
-        {
-            /* Store the sequence:
-             * ip = current - backwardMatchLength
-             * The match is at (bestEntry->offset - backwardMatchLength)
-             */
-            U32 const matchIndex = bestEntry->offset;
-            U32 const offset = current - matchIndex;
-            rawSeq* const seq = rawSeqStore->seq + rawSeqStore->size;
-
-            /* Out of sequence storage */
-            if (rawSeqStore->size == rawSeqStore->capacity)
-                return ERROR(dstSize_tooSmall);
-            seq->litLength = (U32)(ip - anchor);
-            seq->matchLength = (U32)mLength;
-            seq->offset = offset;
-            rawSeqStore->size++;
-        }
+            /* Match found */
+            offset = (U32)(split - base) - bestEntry->offset;
+            mLength = forwardMatchLength + backwardMatchLength;
+            {
+                rawSeq* const seq = rawSeqStore->seq + rawSeqStore->size;
+
+                /* Out of sequence storage */
+                if (rawSeqStore->size == rawSeqStore->capacity)
+                    return ERROR(dstSize_tooSmall);
+                seq->litLength = (U32)(split - backwardMatchLength - anchor);
+                seq->matchLength = (U32)mLength;
+                seq->offset = offset;
+                rawSeqStore->size++;
+            }
 
-        /* Insert the current entry into the hash table */
-        ZSTD_ldm_makeEntryAndInsertByTag(ldmState, rollingHash, hBits,
-                                         (U32)(lastHashed - base),
-                                         *params);
+            /* Insert the current entry into the hash table --- it must be
+             * done after the previous block to avoid clobbering bestEntry */
+            ZSTD_ldm_insertEntry(ldmState, hash, newEntry, *params);
 
-        assert(ip + backwardMatchLength == lastHashed);
+            anchor = split + forwardMatchLength;
 
-        /* Fill the hash table from lastHashed+1 to ip+mLength*/
-        /* Heuristic: don't need to fill the entire table at end of block */
-        if (ip + mLength <= ilimit) {
-            rollingHash = ZSTD_ldm_fillLdmHashTable(
-                              ldmState, rollingHash, lastHashed,
-                              ip + mLength, base, hBits, *params);
-            lastHashed = ip + mLength - 1;
+            /* If we find a match that ends after the data that we've hashed
+             * then we have a repeating, overlapping, pattern. E.g. all zeros.
+             * If one repetition of the pattern matches our `stopMask` then all
+             * repetitions will. We don't need to insert them all into out table,
+             * only the first one. So skip over overlapping matches.
+             * This is a major speed boost (20x) for compressing a single byte
+             * repeated, when that byte ends up in the table.
+             */
+            if (anchor > ip + hashed) {
+                ZSTD_ldm_gear_reset(&hashState, anchor - minMatchLength, minMatchLength);
+                /* Continue the outer loop at anchor (ip + hashed == anchor). */
+                ip = anchor - hashed;
+                break;
+            }
         }
-        ip += mLength;
-        anchor = ip;
+
+        ip += hashed;
     }
+
     return iend - anchor;
 }
 
@@ -459,7 +535,7 @@ size_t ZSTD_ldm_generateSequences(
 
         assert(chunkStart < iend);
         /* 1. Perform overflow correction if necessary. */
-        if (ZSTD_window_needOverflowCorrection(ldmState->window, chunkEnd)) {
+        if (ZSTD_window_needOverflowCorrection(ldmState->window, 0, maxDist, ldmState->loadedDictEnd, chunkStart, chunkEnd)) {
             U32 const ldmHSize = 1U << params->hashLog;
             U32 const correction = ZSTD_window_correctOverflow(
                 &ldmState->window, /* cycleLog */ 0, maxDist, chunkStart);
@@ -473,7 +549,7 @@ size_t ZSTD_ldm_generateSequences(
          * the window through early invalidation.
          * TODO: * Test the chunk size.
          *       * Try invalidation after the sequence generation and test the
-         *         the offset against maxDist directly.
+         *         offset against maxDist directly.
          *
          * NOTE: Because of dictionaries + sequence splitting we MUST make sure
          * that any offset used is valid at the END of the sequence, since it may
@@ -503,7 +579,9 @@ size_t ZSTD_ldm_generateSequences(
     return 0;
 }
 
-void ZSTD_ldm_skipSequences(rawSeqStore_t* rawSeqStore, size_t srcSize, U32 const minMatch) {
+void
+ZSTD_ldm_skipSequences(rawSeqStore_t* rawSeqStore, size_t srcSize, U32 const minMatch)
+{
     while (srcSize > 0 && rawSeqStore->pos < rawSeqStore->size) {
         rawSeq* seq = rawSeqStore->seq + rawSeqStore->pos;
         if (srcSize <= seq->litLength) {
@@ -562,14 +640,32 @@ static rawSeq maybeSplitSequence(rawSeqStore_t* rawSeqStore,
     return sequence;
 }
 
+void ZSTD_ldm_skipRawSeqStoreBytes(rawSeqStore_t* rawSeqStore, size_t nbBytes) {
+    U32 currPos = (U32)(rawSeqStore->posInSequence + nbBytes);
+    while (currPos && rawSeqStore->pos < rawSeqStore->size) {
+        rawSeq currSeq = rawSeqStore->seq[rawSeqStore->pos];
+        if (currPos >= currSeq.litLength + currSeq.matchLength) {
+            currPos -= currSeq.litLength + currSeq.matchLength;
+            rawSeqStore->pos++;
+        } else {
+            rawSeqStore->posInSequence = currPos;
+            break;
+        }
+    }
+    if (currPos == 0 || rawSeqStore->pos == rawSeqStore->size) {
+        rawSeqStore->posInSequence = 0;
+    }
+}
+
 size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore,
     ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+    ZSTD_paramSwitch_e useRowMatchFinder,
     void const* src, size_t srcSize)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
     unsigned const minMatch = cParams->minMatch;
     ZSTD_blockCompressor const blockCompressor =
-        ZSTD_selectBlockCompressor(cParams->strategy, ZSTD_matchState_dictMode(ms));
+        ZSTD_selectBlockCompressor(cParams->strategy, useRowMatchFinder, ZSTD_matchState_dictMode(ms));
     /* Input bounds */
     BYTE const* const istart = (BYTE const*)src;
     BYTE const* const iend = istart + srcSize;
@@ -577,9 +673,18 @@ size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore,
     BYTE const* ip = istart;
 
     DEBUGLOG(5, "ZSTD_ldm_blockCompress: srcSize=%zu", srcSize);
+    /* If using opt parser, use LDMs only as candidates rather than always accepting them */
+    if (cParams->strategy >= ZSTD_btopt) {
+        size_t lastLLSize;
+        ms->ldmSeqStore = rawSeqStore;
+        lastLLSize = blockCompressor(ms, seqStore, rep, src, srcSize);
+        ZSTD_ldm_skipRawSeqStoreBytes(rawSeqStore, srcSize);
+        return lastLLSize;
+    }
+
     assert(rawSeqStore->pos <= rawSeqStore->size);
     assert(rawSeqStore->size <= rawSeqStore->capacity);
-    /* Loop through each sequence and apply the block compressor to the lits */
+    /* Loop through each sequence and apply the block compressor to the literals */
     while (rawSeqStore->pos < rawSeqStore->size && ip < iend) {
         /* maybeSplitSequence updates rawSeqStore->pos */
         rawSeq const sequence = maybeSplitSequence(rawSeqStore,
@@ -606,8 +711,8 @@ size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore,
             rep[0] = sequence.offset;
             /* Store the sequence */
             ZSTD_storeSeq(seqStore, newLitLength, ip - newLitLength, iend,
-                          sequence.offset + ZSTD_REP_MOVE,
-                          sequence.matchLength - MINMATCH);
+                          OFFSET_TO_OFFBASE(sequence.offset),
+                          sequence.matchLength);
             ip += sequence.matchLength;
         }
     }

src/borg/algorithms/zstd/lib/compress/zstd_ldm.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -66,6 +66,7 @@ size_t ZSTD_ldm_generateSequences(
  */
 size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore,
             ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+            ZSTD_paramSwitch_e useRowMatchFinder,
             void const* src, size_t srcSize);
 
 /**
@@ -73,11 +74,17 @@ size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore,
  *
  * Skip past `srcSize` bytes worth of sequences in `rawSeqStore`.
  * Avoids emitting matches less than `minMatch` bytes.
- * Must be called for data with is not passed to ZSTD_ldm_blockCompress().
+ * Must be called for data that is not passed to ZSTD_ldm_blockCompress().
  */
 void ZSTD_ldm_skipSequences(rawSeqStore_t* rawSeqStore, size_t srcSize,
     U32 const minMatch);
 
+/* ZSTD_ldm_skipRawSeqStoreBytes():
+ * Moves forward in rawSeqStore by nbBytes, updating fields 'pos' and 'posInSequence'.
+ * Not to be used in conjunction with ZSTD_ldm_skipSequences().
+ * Must be called for data with is not passed to ZSTD_ldm_blockCompress().
+ */
+void ZSTD_ldm_skipRawSeqStoreBytes(rawSeqStore_t* rawSeqStore, size_t nbBytes);
 
 /** ZSTD_ldm_getTableSize() :
  *  Estimate the space needed for long distance matching tables or 0 if LDM is

src/borg/algorithms/zstd/lib/compress/zstd_ldm_geartab.h

@@ -0,0 +1,106 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_LDM_GEARTAB_H
+#define ZSTD_LDM_GEARTAB_H
+
+#include "../common/compiler.h" /* UNUSED_ATTR */
+#include "../common/mem.h"      /* U64 */
+
+static UNUSED_ATTR const U64 ZSTD_ldm_gearTab[256] = {
+    0xf5b8f72c5f77775c, 0x84935f266b7ac412, 0xb647ada9ca730ccc,
+    0xb065bb4b114fb1de, 0x34584e7e8c3a9fd0, 0x4e97e17c6ae26b05,
+    0x3a03d743bc99a604, 0xcecd042422c4044f, 0x76de76c58524259e,
+    0x9c8528f65badeaca, 0x86563706e2097529, 0x2902475fa375d889,
+    0xafb32a9739a5ebe6, 0xce2714da3883e639, 0x21eaf821722e69e,
+    0x37b628620b628,    0x49a8d455d88caf5,  0x8556d711e6958140,
+    0x4f7ae74fc605c1f,  0x829f0c3468bd3a20, 0x4ffdc885c625179e,
+    0x8473de048a3daf1b, 0x51008822b05646b2, 0x69d75d12b2d1cc5f,
+    0x8c9d4a19159154bc, 0xc3cc10f4abbd4003, 0xd06ddc1cecb97391,
+    0xbe48e6e7ed80302e, 0x3481db31cee03547, 0xacc3f67cdaa1d210,
+    0x65cb771d8c7f96cc, 0x8eb27177055723dd, 0xc789950d44cd94be,
+    0x934feadc3700b12b, 0x5e485f11edbdf182, 0x1e2e2a46fd64767a,
+    0x2969ca71d82efa7c, 0x9d46e9935ebbba2e, 0xe056b67e05e6822b,
+    0x94d73f55739d03a0, 0xcd7010bdb69b5a03, 0x455ef9fcd79b82f4,
+    0x869cb54a8749c161, 0x38d1a4fa6185d225, 0xb475166f94bbe9bb,
+    0xa4143548720959f1, 0x7aed4780ba6b26ba, 0xd0ce264439e02312,
+    0x84366d746078d508, 0xa8ce973c72ed17be, 0x21c323a29a430b01,
+    0x9962d617e3af80ee, 0xab0ce91d9c8cf75b, 0x530e8ee6d19a4dbc,
+    0x2ef68c0cf53f5d72, 0xc03a681640a85506, 0x496e4e9f9c310967,
+    0x78580472b59b14a0, 0x273824c23b388577, 0x66bf923ad45cb553,
+    0x47ae1a5a2492ba86, 0x35e304569e229659, 0x4765182a46870b6f,
+    0x6cbab625e9099412, 0xddac9a2e598522c1, 0x7172086e666624f2,
+    0xdf5003ca503b7837, 0x88c0c1db78563d09, 0x58d51865acfc289d,
+    0x177671aec65224f1, 0xfb79d8a241e967d7, 0x2be1e101cad9a49a,
+    0x6625682f6e29186b, 0x399553457ac06e50, 0x35dffb4c23abb74,
+    0x429db2591f54aade, 0xc52802a8037d1009, 0x6acb27381f0b25f3,
+    0xf45e2551ee4f823b, 0x8b0ea2d99580c2f7, 0x3bed519cbcb4e1e1,
+    0xff452823dbb010a,  0x9d42ed614f3dd267, 0x5b9313c06257c57b,
+    0xa114b8008b5e1442, 0xc1fe311c11c13d4b, 0x66e8763ea34c5568,
+    0x8b982af1c262f05d, 0xee8876faaa75fbb7, 0x8a62a4d0d172bb2a,
+    0xc13d94a3b7449a97, 0x6dbbba9dc15d037c, 0xc786101f1d92e0f1,
+    0xd78681a907a0b79b, 0xf61aaf2962c9abb9, 0x2cfd16fcd3cb7ad9,
+    0x868c5b6744624d21, 0x25e650899c74ddd7, 0xba042af4a7c37463,
+    0x4eb1a539465a3eca, 0xbe09dbf03b05d5ca, 0x774e5a362b5472ba,
+    0x47a1221229d183cd, 0x504b0ca18ef5a2df, 0xdffbdfbde2456eb9,
+    0x46cd2b2fbee34634, 0xf2aef8fe819d98c3, 0x357f5276d4599d61,
+    0x24a5483879c453e3, 0x88026889192b4b9,  0x28da96671782dbec,
+    0x4ef37c40588e9aaa, 0x8837b90651bc9fb3, 0xc164f741d3f0e5d6,
+    0xbc135a0a704b70ba, 0x69cd868f7622ada,  0xbc37ba89e0b9c0ab,
+    0x47c14a01323552f6, 0x4f00794bacee98bb, 0x7107de7d637a69d5,
+    0x88af793bb6f2255e, 0xf3c6466b8799b598, 0xc288c616aa7f3b59,
+    0x81ca63cf42fca3fd, 0x88d85ace36a2674b, 0xd056bd3792389e7,
+    0xe55c396c4e9dd32d, 0xbefb504571e6c0a6, 0x96ab32115e91e8cc,
+    0xbf8acb18de8f38d1, 0x66dae58801672606, 0x833b6017872317fb,
+    0xb87c16f2d1c92864, 0xdb766a74e58b669c, 0x89659f85c61417be,
+    0xc8daad856011ea0c, 0x76a4b565b6fe7eae, 0xa469d085f6237312,
+    0xaaf0365683a3e96c, 0x4dbb746f8424f7b8, 0x638755af4e4acc1,
+    0x3d7807f5bde64486, 0x17be6d8f5bbb7639, 0x903f0cd44dc35dc,
+    0x67b672eafdf1196c, 0xa676ff93ed4c82f1, 0x521d1004c5053d9d,
+    0x37ba9ad09ccc9202, 0x84e54d297aacfb51, 0xa0b4b776a143445,
+    0x820d471e20b348e,  0x1874383cb83d46dc, 0x97edeec7a1efe11c,
+    0xb330e50b1bdc42aa, 0x1dd91955ce70e032, 0xa514cdb88f2939d5,
+    0x2791233fd90db9d3, 0x7b670a4cc50f7a9b, 0x77c07d2a05c6dfa5,
+    0xe3778b6646d0a6fa, 0xb39c8eda47b56749, 0x933ed448addbef28,
+    0xaf846af6ab7d0bf4, 0xe5af208eb666e49,  0x5e6622f73534cd6a,
+    0x297daeca42ef5b6e, 0x862daef3d35539a6, 0xe68722498f8e1ea9,
+    0x981c53093dc0d572, 0xfa09b0bfbf86fbf5, 0x30b1e96166219f15,
+    0x70e7d466bdc4fb83, 0x5a66736e35f2a8e9, 0xcddb59d2b7c1baef,
+    0xd6c7d247d26d8996, 0xea4e39eac8de1ba3, 0x539c8bb19fa3aff2,
+    0x9f90e4c5fd508d8,  0xa34e5956fbaf3385, 0x2e2f8e151d3ef375,
+    0x173691e9b83faec1, 0xb85a8d56bf016379, 0x8382381267408ae3,
+    0xb90f901bbdc0096d, 0x7c6ad32933bcec65, 0x76bb5e2f2c8ad595,
+    0x390f851a6cf46d28, 0xc3e6064da1c2da72, 0xc52a0c101cfa5389,
+    0xd78eaf84a3fbc530, 0x3781b9e2288b997e, 0x73c2f6dea83d05c4,
+    0x4228e364c5b5ed7,  0x9d7a3edf0da43911, 0x8edcfeda24686756,
+    0x5e7667a7b7a9b3a1, 0x4c4f389fa143791d, 0xb08bc1023da7cddc,
+    0x7ab4be3ae529b1cc, 0x754e6132dbe74ff9, 0x71635442a839df45,
+    0x2f6fb1643fbe52de, 0x961e0a42cf7a8177, 0xf3b45d83d89ef2ea,
+    0xee3de4cf4a6e3e9b, 0xcd6848542c3295e7, 0xe4cee1664c78662f,
+    0x9947548b474c68c4, 0x25d73777a5ed8b0b, 0xc915b1d636b7fc,
+    0x21c2ba75d9b0d2da, 0x5f6b5dcf608a64a1, 0xdcf333255ff9570c,
+    0x633b922418ced4ee, 0xc136dde0b004b34a, 0x58cc83b05d4b2f5a,
+    0x5eb424dda28e42d2, 0x62df47369739cd98, 0xb4e0b42485e4ce17,
+    0x16e1f0c1f9a8d1e7, 0x8ec3916707560ebf, 0x62ba6e2df2cc9db3,
+    0xcbf9f4ff77d83a16, 0x78d9d7d07d2bbcc4, 0xef554ce1e02c41f4,
+    0x8d7581127eccf94d, 0xa9b53336cb3c8a05, 0x38c42c0bf45c4f91,
+    0x640893cdf4488863, 0x80ec34bc575ea568, 0x39f324f5b48eaa40,
+    0xe9d9ed1f8eff527f, 0x9224fc058cc5a214, 0xbaba00b04cfe7741,
+    0x309a9f120fcf52af, 0xa558f3ec65626212, 0x424bec8b7adabe2f,
+    0x41622513a6aea433, 0xb88da2d5324ca798, 0xd287733b245528a4,
+    0x9a44697e6d68aec3, 0x7b1093be2f49bb28, 0x50bbec632e3d8aad,
+    0x6cd90723e1ea8283, 0x897b9e7431b02bf3, 0x219efdcb338a7047,
+    0x3b0311f0a27c0656, 0xdb17bf91c0db96e7, 0x8cd4fd6b4e85a5b2,
+    0xfab071054ba6409d, 0x40d6fe831fa9dfd9, 0xaf358debad7d791e,
+    0xeb8d0e25a65e3e58, 0xbbcbd3df14e08580, 0xcf751f27ecdab2b,
+    0x2b4da14f2613d8f4
+};
+
+#endif /* ZSTD_LDM_GEARTAB_H */

src/borg/algorithms/zstd/lib/compress/zstd_opt.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the

src/borg/algorithms/zstd/lib/compress/zstdmt_compress.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -19,113 +19,60 @@
 /* Note : This is an internal API.
  *        These APIs used to be exposed with ZSTDLIB_API,
  *        because it used to be the only way to invoke MT compression.
- *        Now, it's recommended to use ZSTD_compress2 and ZSTD_compressStream2()
- *        instead.
- *
- *        If you depend on these APIs and can't switch, then define
- *        ZSTD_LEGACY_MULTITHREADED_API when making the dynamic library.
- *        However, we may completely remove these functions in a future
- *        release, so please switch soon.
+ *        Now, you must use ZSTD_compress2 and ZSTD_compressStream2() instead.
  *
  *        This API requires ZSTD_MULTITHREAD to be defined during compilation,
  *        otherwise ZSTDMT_createCCtx*() will fail.
  */
 
-#ifdef ZSTD_LEGACY_MULTITHREADED_API
-#  define ZSTDMT_API ZSTDLIB_API
-#else
-#  define ZSTDMT_API
-#endif
-
 /* ===   Dependencies   === */
-#include <stddef.h>                /* size_t */
+#include "../common/zstd_deps.h"   /* size_t */
 #define ZSTD_STATIC_LINKING_ONLY   /* ZSTD_parameters */
 #include "../zstd.h"            /* ZSTD_inBuffer, ZSTD_outBuffer, ZSTDLIB_API */
 
 
 /* ===   Constants   === */
-#ifndef ZSTDMT_NBWORKERS_MAX
-#  define ZSTDMT_NBWORKERS_MAX 200
+#ifndef ZSTDMT_NBWORKERS_MAX /* a different value can be selected at compile time */
+#  define ZSTDMT_NBWORKERS_MAX ((sizeof(void*)==4) /*32-bit*/ ? 64 : 256)
 #endif
-#ifndef ZSTDMT_JOBSIZE_MIN
-#  define ZSTDMT_JOBSIZE_MIN (1 MB)
+#ifndef ZSTDMT_JOBSIZE_MIN   /* a different value can be selected at compile time */
+#  define ZSTDMT_JOBSIZE_MIN (512 KB)
 #endif
 #define ZSTDMT_JOBLOG_MAX   (MEM_32bits() ? 29 : 30)
 #define ZSTDMT_JOBSIZE_MAX  (MEM_32bits() ? (512 MB) : (1024 MB))
 
 
+/* ========================================================
+ * ===  Private interface, for use by ZSTD_compress.c   ===
+ * ===  Not exposed in libzstd. Never invoke directly   ===
+ * ======================================================== */
+
 /* ===   Memory management   === */
 typedef struct ZSTDMT_CCtx_s ZSTDMT_CCtx;
 /* Requires ZSTD_MULTITHREAD to be defined during compilation, otherwise it will return NULL. */
-ZSTDMT_API ZSTDMT_CCtx* ZSTDMT_createCCtx(unsigned nbWorkers);
-/* Requires ZSTD_MULTITHREAD to be defined during compilation, otherwise it will return NULL. */
-ZSTDMT_API ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced(unsigned nbWorkers,
-                                                    ZSTD_customMem cMem);
-ZSTDMT_API size_t ZSTDMT_freeCCtx(ZSTDMT_CCtx* mtctx);
-
-ZSTDMT_API size_t ZSTDMT_sizeof_CCtx(ZSTDMT_CCtx* mtctx);
-
-
-/* ===   Simple one-pass compression function   === */
-
-ZSTDMT_API size_t ZSTDMT_compressCCtx(ZSTDMT_CCtx* mtctx,
-                                       void* dst, size_t dstCapacity,
-                                 const void* src, size_t srcSize,
-                                       int compressionLevel);
-
+ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced(unsigned nbWorkers,
+                                        ZSTD_customMem cMem,
+					ZSTD_threadPool *pool);
+size_t ZSTDMT_freeCCtx(ZSTDMT_CCtx* mtctx);
 
+size_t ZSTDMT_sizeof_CCtx(ZSTDMT_CCtx* mtctx);
 
 /* ===   Streaming functions   === */
 
-ZSTDMT_API size_t ZSTDMT_initCStream(ZSTDMT_CCtx* mtctx, int compressionLevel);
-ZSTDMT_API size_t ZSTDMT_resetCStream(ZSTDMT_CCtx* mtctx, unsigned long long pledgedSrcSize);  /**< if srcSize is not known at reset time, use ZSTD_CONTENTSIZE_UNKNOWN. Note: for compatibility with older programs, 0 means the same as ZSTD_CONTENTSIZE_UNKNOWN, but it will change in the future to mean "empty" */
-
-ZSTDMT_API size_t ZSTDMT_nextInputSizeHint(const ZSTDMT_CCtx* mtctx);
-ZSTDMT_API size_t ZSTDMT_compressStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output, ZSTD_inBuffer* input);
-
-ZSTDMT_API size_t ZSTDMT_flushStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output);   /**< @return : 0 == all flushed; >0 : still some data to be flushed; or an error code (ZSTD_isError()) */
-ZSTDMT_API size_t ZSTDMT_endStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output);     /**< @return : 0 == all flushed; >0 : still some data to be flushed; or an error code (ZSTD_isError()) */
-
-
-/* ===   Advanced functions and parameters  === */
-
-ZSTDMT_API size_t ZSTDMT_compress_advanced(ZSTDMT_CCtx* mtctx,
-                                          void* dst, size_t dstCapacity,
-                                    const void* src, size_t srcSize,
-                                    const ZSTD_CDict* cdict,
-                                          ZSTD_parameters params,
-                                          int overlapLog);
-
-ZSTDMT_API size_t ZSTDMT_initCStream_advanced(ZSTDMT_CCtx* mtctx,
-                                        const void* dict, size_t dictSize,   /* dict can be released after init, a local copy is preserved within zcs */
-                                        ZSTD_parameters params,
-                                        unsigned long long pledgedSrcSize);  /* pledgedSrcSize is optional and can be zero == unknown */
-
-ZSTDMT_API size_t ZSTDMT_initCStream_usingCDict(ZSTDMT_CCtx* mtctx,
-                                        const ZSTD_CDict* cdict,
-                                        ZSTD_frameParameters fparams,
-                                        unsigned long long pledgedSrcSize);  /* note : zero means empty */
-
-/* ZSTDMT_parameter :
- * List of parameters that can be set using ZSTDMT_setMTCtxParameter() */
-typedef enum {
-    ZSTDMT_p_jobSize,     /* Each job is compressed in parallel. By default, this value is dynamically determined depending on compression parameters. Can be set explicitly here. */
-    ZSTDMT_p_overlapLog,  /* Each job may reload a part of previous job to enhance compression ratio; 0 == no overlap, 6(default) == use 1/8th of window, >=9 == use full window. This is a "sticky" parameter : its value will be re-used on next compression job */
-    ZSTDMT_p_rsyncable    /* Enables rsyncable mode. */
-} ZSTDMT_parameter;
-
-/* ZSTDMT_setMTCtxParameter() :
- * allow setting individual parameters, one at a time, among a list of enums defined in ZSTDMT_parameter.
- * The function must be called typically after ZSTD_createCCtx() but __before ZSTDMT_init*() !__
- * Parameters not explicitly reset by ZSTDMT_init*() remain the same in consecutive compression sessions.
- * @return : 0, or an error code (which can be tested using ZSTD_isError()) */
-ZSTDMT_API size_t ZSTDMT_setMTCtxParameter(ZSTDMT_CCtx* mtctx, ZSTDMT_parameter parameter, int value);
-
-/* ZSTDMT_getMTCtxParameter() :
- * Query the ZSTDMT_CCtx for a parameter value.
- * @return : 0, or an error code (which can be tested using ZSTD_isError()) */
-ZSTDMT_API size_t ZSTDMT_getMTCtxParameter(ZSTDMT_CCtx* mtctx, ZSTDMT_parameter parameter, int* value);
+size_t ZSTDMT_nextInputSizeHint(const ZSTDMT_CCtx* mtctx);
 
+/*! ZSTDMT_initCStream_internal() :
+ *  Private use only. Init streaming operation.
+ *  expects params to be valid.
+ *  must receive dict, or cdict, or none, but not both.
+ *  mtctx can be freshly constructed or reused from a prior compression.
+ *  If mtctx is reused, memory allocations from the prior compression may not be freed,
+ *  even if they are not needed for the current compression.
+ *  @return : 0, or an error code */
+size_t ZSTDMT_initCStream_internal(ZSTDMT_CCtx* mtctx,
+                    const void* dict, size_t dictSize, ZSTD_dictContentType_e dictContentType,
+                    const ZSTD_CDict* cdict,
+                    ZSTD_CCtx_params params, unsigned long long pledgedSrcSize);
 
 /*! ZSTDMT_compressStream_generic() :
  *  Combines ZSTDMT_compressStream() with optional ZSTDMT_flushStream() or ZSTDMT_endStream()
@@ -134,16 +81,10 @@ ZSTDMT_API size_t ZSTDMT_getMTCtxParameter(ZSTDMT_CCtx* mtctx, ZSTDMT_parameter
  *           0 if fully flushed
  *           or an error code
  *  note : needs to be init using any ZSTD_initCStream*() variant */
-ZSTDMT_API size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx,
-                                                ZSTD_outBuffer* output,
-                                                ZSTD_inBuffer* input,
-                                                ZSTD_EndDirective endOp);
-
-
-/* ========================================================
- * ===  Private interface, for use by ZSTD_compress.c   ===
- * ===  Not exposed in libzstd. Never invoke directly   ===
- * ======================================================== */
+size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx,
+                                     ZSTD_outBuffer* output,
+                                     ZSTD_inBuffer* input,
+                                     ZSTD_EndDirective endOp);
 
  /*! ZSTDMT_toFlushNow()
   *  Tell how many bytes are ready to be flushed immediately.
@@ -153,15 +94,6 @@ ZSTDMT_API size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx,
   *  therefore flushing is limited by speed of oldest job. */
 size_t ZSTDMT_toFlushNow(ZSTDMT_CCtx* mtctx);
 
-/*! ZSTDMT_CCtxParam_setMTCtxParameter()
- *  like ZSTDMT_setMTCtxParameter(), but into a ZSTD_CCtx_Params */
-size_t ZSTDMT_CCtxParam_setMTCtxParameter(ZSTD_CCtx_params* params, ZSTDMT_parameter parameter, int value);
-
-/*! ZSTDMT_CCtxParam_setNbWorkers()
- *  Set nbWorkers, and clamp it.
- *  Also reset jobSize and overlapLog */
-size_t ZSTDMT_CCtxParam_setNbWorkers(ZSTD_CCtx_params* params, unsigned nbWorkers);
-
 /*! ZSTDMT_updateCParams_whileCompressing() :
  *  Updates only a selected set of compression parameters, to remain compatible with current frame.
  *  New parameters will be applied to next compression job. */
@@ -174,17 +106,6 @@ void ZSTDMT_updateCParams_whileCompressing(ZSTDMT_CCtx* mtctx, const ZSTD_CCtx_p
 ZSTD_frameProgression ZSTDMT_getFrameProgression(ZSTDMT_CCtx* mtctx);
 
 
-/*! ZSTDMT_initCStream_internal() :
- *  Private use only. Init streaming operation.
- *  expects params to be valid.
- *  must receive dict, or cdict, or none, but not both.
- *  @return : 0, or an error code */
-size_t ZSTDMT_initCStream_internal(ZSTDMT_CCtx* zcs,
-                    const void* dict, size_t dictSize, ZSTD_dictContentType_e dictContentType,
-                    const ZSTD_CDict* cdict,
-                    ZSTD_CCtx_params params, unsigned long long pledgedSrcSize);
-
-
 #if defined (__cplusplus)
 }
 #endif

src/borg/algorithms/zstd/lib/decompress/huf_decompress_amd64.S

@@ -0,0 +1,576 @@
+/*
+ * 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.
+ */
+
+#include "../common/portability_macros.h"
+
+/* Stack marking
+ * ref: https://wiki.gentoo.org/wiki/Hardened/GNU_stack_quickstart
+ */
+#if defined(__ELF__) && defined(__GNUC__)
+.section .note.GNU-stack,"",%progbits
+#endif
+
+#if ZSTD_ENABLE_ASM_X86_64_BMI2
+
+/* Calling convention:
+ *
+ * %rdi contains the first argument: HUF_DecompressAsmArgs*.
+ * %rbp isn't maintained (no frame pointer).
+ * %rsp contains the stack pointer that grows down.
+ *      No red-zone is assumed, only addresses >= %rsp are used.
+ * All register contents are preserved.
+ *
+ * TODO: Support Windows calling convention.
+ */
+
+ZSTD_HIDE_ASM_FUNCTION(HUF_decompress4X1_usingDTable_internal_fast_asm_loop)
+ZSTD_HIDE_ASM_FUNCTION(HUF_decompress4X2_usingDTable_internal_fast_asm_loop)
+ZSTD_HIDE_ASM_FUNCTION(_HUF_decompress4X2_usingDTable_internal_fast_asm_loop)
+ZSTD_HIDE_ASM_FUNCTION(_HUF_decompress4X1_usingDTable_internal_fast_asm_loop)
+.global HUF_decompress4X1_usingDTable_internal_fast_asm_loop
+.global HUF_decompress4X2_usingDTable_internal_fast_asm_loop
+.global _HUF_decompress4X1_usingDTable_internal_fast_asm_loop
+.global _HUF_decompress4X2_usingDTable_internal_fast_asm_loop
+.text
+
+/* Sets up register mappings for clarity.
+ * op[], bits[], dtable & ip[0] each get their own register.
+ * ip[1,2,3] & olimit alias var[].
+ * %rax is a scratch register.
+ */
+
+#define op0    rsi
+#define op1    rbx
+#define op2    rcx
+#define op3    rdi
+
+#define ip0    r8
+#define ip1    r9
+#define ip2    r10
+#define ip3    r11
+
+#define bits0  rbp
+#define bits1  rdx
+#define bits2  r12
+#define bits3  r13
+#define dtable r14
+#define olimit r15
+
+/* var[] aliases ip[1,2,3] & olimit
+ * ip[1,2,3] are saved every iteration.
+ * olimit is only used in compute_olimit.
+ */
+#define var0   r15
+#define var1   r9
+#define var2   r10
+#define var3   r11
+
+/* 32-bit var registers */
+#define vard0  r15d
+#define vard1  r9d
+#define vard2  r10d
+#define vard3  r11d
+
+/* Calls X(N) for each stream 0, 1, 2, 3. */
+#define FOR_EACH_STREAM(X) \
+    X(0);                  \
+    X(1);                  \
+    X(2);                  \
+    X(3)
+
+/* Calls X(N, idx) for each stream 0, 1, 2, 3. */
+#define FOR_EACH_STREAM_WITH_INDEX(X, idx) \
+    X(0, idx);                             \
+    X(1, idx);                             \
+    X(2, idx);                             \
+    X(3, idx)
+
+/* Define both _HUF_* & HUF_* symbols because MacOS
+ * C symbols are prefixed with '_' & Linux symbols aren't.
+ */
+_HUF_decompress4X1_usingDTable_internal_fast_asm_loop:
+HUF_decompress4X1_usingDTable_internal_fast_asm_loop:
+    ZSTD_CET_ENDBRANCH
+    /* Save all registers - even if they are callee saved for simplicity. */
+    push %rax
+    push %rbx
+    push %rcx
+    push %rdx
+    push %rbp
+    push %rsi
+    push %rdi
+    push %r8
+    push %r9
+    push %r10
+    push %r11
+    push %r12
+    push %r13
+    push %r14
+    push %r15
+
+    /* Read HUF_DecompressAsmArgs* args from %rax */
+    movq %rdi, %rax
+    movq  0(%rax), %ip0
+    movq  8(%rax), %ip1
+    movq 16(%rax), %ip2
+    movq 24(%rax), %ip3
+    movq 32(%rax), %op0
+    movq 40(%rax), %op1
+    movq 48(%rax), %op2
+    movq 56(%rax), %op3
+    movq 64(%rax), %bits0
+    movq 72(%rax), %bits1
+    movq 80(%rax), %bits2
+    movq 88(%rax), %bits3
+    movq 96(%rax), %dtable
+    push %rax      /* argument */
+    push 104(%rax) /* ilimit */
+    push 112(%rax) /* oend */
+    push %olimit   /* olimit space */
+
+    subq $24, %rsp
+
+.L_4X1_compute_olimit:
+    /* Computes how many iterations we can do safely
+     * %r15, %rax may be clobbered
+     * rbx, rdx must be saved
+     * op3 & ip0 mustn't be clobbered
+     */
+    movq %rbx, 0(%rsp)
+    movq %rdx, 8(%rsp)
+
+    movq 32(%rsp), %rax /* rax = oend */
+    subq %op3,    %rax  /* rax = oend - op3 */
+
+    /* r15 = (oend - op3) / 5 */
+    movabsq $-3689348814741910323, %rdx
+    mulq %rdx
+    movq %rdx, %r15
+    shrq $2, %r15
+
+    movq %ip0,     %rax /* rax = ip0 */
+    movq 40(%rsp), %rdx /* rdx = ilimit */
+    subq %rdx,     %rax /* rax = ip0 - ilimit */
+    movq %rax,     %rbx /* rbx = ip0 - ilimit */
+
+    /* rdx = (ip0 - ilimit) / 7 */
+    movabsq $2635249153387078803, %rdx
+    mulq %rdx
+    subq %rdx, %rbx
+    shrq %rbx
+    addq %rbx, %rdx
+    shrq $2, %rdx
+
+    /* r15 = min(%rdx, %r15) */
+    cmpq %rdx, %r15
+    cmova %rdx, %r15
+
+    /* r15 = r15 * 5 */
+    leaq (%r15, %r15, 4), %r15
+
+    /* olimit = op3 + r15 */
+    addq %op3, %olimit
+
+    movq 8(%rsp), %rdx
+    movq 0(%rsp), %rbx
+
+    /* If (op3 + 20 > olimit) */
+    movq %op3, %rax    /* rax = op3 */
+    addq $20,  %rax    /* rax = op3 + 20 */
+    cmpq %rax, %olimit /* op3 + 20 > olimit */
+    jb .L_4X1_exit
+
+    /* If (ip1 < ip0) go to exit */
+    cmpq %ip0, %ip1
+    jb .L_4X1_exit
+
+    /* If (ip2 < ip1) go to exit */
+    cmpq %ip1, %ip2
+    jb .L_4X1_exit
+
+    /* If (ip3 < ip2) go to exit */
+    cmpq %ip2, %ip3
+    jb .L_4X1_exit
+
+/* Reads top 11 bits from bits[n]
+ * Loads dt[bits[n]] into var[n]
+ */
+#define GET_NEXT_DELT(n)                \
+    movq $53, %var##n;                  \
+    shrxq %var##n, %bits##n, %var##n;   \
+    movzwl (%dtable,%var##n,2),%vard##n
+
+/* var[n] must contain the DTable entry computed with GET_NEXT_DELT
+ * Moves var[n] to %rax
+ * bits[n] <<= var[n] & 63
+ * op[n][idx] = %rax >> 8
+ * %ah is a way to access bits [8, 16) of %rax
+ */
+#define DECODE_FROM_DELT(n, idx)       \
+    movq %var##n, %rax;                \
+    shlxq %var##n, %bits##n, %bits##n; \
+    movb %ah, idx(%op##n)
+
+/* Assumes GET_NEXT_DELT has been called.
+ * Calls DECODE_FROM_DELT then GET_NEXT_DELT
+ */
+#define DECODE_AND_GET_NEXT(n, idx) \
+    DECODE_FROM_DELT(n, idx);       \
+    GET_NEXT_DELT(n)                \
+
+/* // ctz & nbBytes is stored in bits[n]
+ * // nbBits is stored in %rax
+ * ctz  = CTZ[bits[n]]
+ * nbBits  = ctz & 7
+ * nbBytes = ctz >> 3
+ * op[n]  += 5
+ * ip[n]  -= nbBytes
+ * // Note: x86-64 is little-endian ==> no bswap
+ * bits[n] = MEM_readST(ip[n]) | 1
+ * bits[n] <<= nbBits
+ */
+#define RELOAD_BITS(n)             \
+    bsfq %bits##n, %bits##n;       \
+    movq %bits##n, %rax;           \
+    andq $7, %rax;                 \
+    shrq $3, %bits##n;             \
+    leaq 5(%op##n), %op##n;        \
+    subq %bits##n, %ip##n;         \
+    movq (%ip##n), %bits##n;       \
+    orq $1, %bits##n;              \
+    shlx %rax, %bits##n, %bits##n
+
+    /* Store clobbered variables on the stack */
+    movq %olimit, 24(%rsp)
+    movq %ip1, 0(%rsp)
+    movq %ip2, 8(%rsp)
+    movq %ip3, 16(%rsp)
+
+    /* Call GET_NEXT_DELT for each stream */
+    FOR_EACH_STREAM(GET_NEXT_DELT)
+
+    .p2align 6
+
+.L_4X1_loop_body:
+    /* Decode 5 symbols in each of the 4 streams (20 total)
+     * Must have called GET_NEXT_DELT for each stream
+     */
+    FOR_EACH_STREAM_WITH_INDEX(DECODE_AND_GET_NEXT, 0)
+    FOR_EACH_STREAM_WITH_INDEX(DECODE_AND_GET_NEXT, 1)
+    FOR_EACH_STREAM_WITH_INDEX(DECODE_AND_GET_NEXT, 2)
+    FOR_EACH_STREAM_WITH_INDEX(DECODE_AND_GET_NEXT, 3)
+    FOR_EACH_STREAM_WITH_INDEX(DECODE_FROM_DELT, 4)
+
+    /* Load ip[1,2,3] from stack (var[] aliases them)
+     * ip[] is needed for RELOAD_BITS
+     * Each will be stored back to the stack after RELOAD
+     */
+    movq 0(%rsp), %ip1
+    movq 8(%rsp), %ip2
+    movq 16(%rsp), %ip3
+
+    /* Reload each stream & fetch the next table entry
+     * to prepare for the next iteration
+     */
+    RELOAD_BITS(0)
+    GET_NEXT_DELT(0)
+
+    RELOAD_BITS(1)
+    movq %ip1, 0(%rsp)
+    GET_NEXT_DELT(1)
+
+    RELOAD_BITS(2)
+    movq %ip2, 8(%rsp)
+    GET_NEXT_DELT(2)
+
+    RELOAD_BITS(3)
+    movq %ip3, 16(%rsp)
+    GET_NEXT_DELT(3)
+
+    /* If op3 < olimit: continue the loop */
+    cmp %op3, 24(%rsp)
+    ja .L_4X1_loop_body
+
+    /* Reload ip[1,2,3] from stack */
+    movq 0(%rsp), %ip1
+    movq 8(%rsp), %ip2
+    movq 16(%rsp), %ip3
+
+    /* Re-compute olimit */
+    jmp .L_4X1_compute_olimit
+
+#undef GET_NEXT_DELT
+#undef DECODE_FROM_DELT
+#undef DECODE
+#undef RELOAD_BITS
+.L_4X1_exit:
+    addq $24, %rsp
+
+    /* Restore stack (oend & olimit) */
+    pop %rax /* olimit */
+    pop %rax /* oend */
+    pop %rax /* ilimit */
+    pop %rax /* arg */
+
+    /* Save ip / op / bits */
+    movq %ip0,  0(%rax)
+    movq %ip1,  8(%rax)
+    movq %ip2, 16(%rax)
+    movq %ip3, 24(%rax)
+    movq %op0, 32(%rax)
+    movq %op1, 40(%rax)
+    movq %op2, 48(%rax)
+    movq %op3, 56(%rax)
+    movq %bits0, 64(%rax)
+    movq %bits1, 72(%rax)
+    movq %bits2, 80(%rax)
+    movq %bits3, 88(%rax)
+
+    /* Restore registers */
+    pop %r15
+    pop %r14
+    pop %r13
+    pop %r12
+    pop %r11
+    pop %r10
+    pop %r9
+    pop %r8
+    pop %rdi
+    pop %rsi
+    pop %rbp
+    pop %rdx
+    pop %rcx
+    pop %rbx
+    pop %rax
+    ret
+
+_HUF_decompress4X2_usingDTable_internal_fast_asm_loop:
+HUF_decompress4X2_usingDTable_internal_fast_asm_loop:
+    ZSTD_CET_ENDBRANCH
+    /* Save all registers - even if they are callee saved for simplicity. */
+    push %rax
+    push %rbx
+    push %rcx
+    push %rdx
+    push %rbp
+    push %rsi
+    push %rdi
+    push %r8
+    push %r9
+    push %r10
+    push %r11
+    push %r12
+    push %r13
+    push %r14
+    push %r15
+
+    movq %rdi, %rax
+    movq  0(%rax), %ip0
+    movq  8(%rax), %ip1
+    movq 16(%rax), %ip2
+    movq 24(%rax), %ip3
+    movq 32(%rax), %op0
+    movq 40(%rax), %op1
+    movq 48(%rax), %op2
+    movq 56(%rax), %op3
+    movq 64(%rax), %bits0
+    movq 72(%rax), %bits1
+    movq 80(%rax), %bits2
+    movq 88(%rax), %bits3
+    movq 96(%rax), %dtable
+    push %rax      /* argument */
+    push %rax      /* olimit */
+    push 104(%rax) /* ilimit */
+
+    movq 112(%rax), %rax
+    push %rax /* oend3 */
+
+    movq %op3, %rax
+    push %rax /* oend2 */
+
+    movq %op2, %rax
+    push %rax /* oend1 */
+
+    movq %op1, %rax
+    push %rax /* oend0 */
+
+    /* Scratch space */
+    subq $8, %rsp
+
+.L_4X2_compute_olimit:
+    /* Computes how many iterations we can do safely
+     * %r15, %rax may be clobbered
+     * rdx must be saved
+     * op[1,2,3,4] & ip0 mustn't be clobbered
+     */
+    movq %rdx, 0(%rsp)
+
+    /* We can consume up to 7 input bytes each iteration. */
+    movq %ip0,     %rax  /* rax = ip0 */
+    movq 40(%rsp), %rdx  /* rdx = ilimit */
+    subq %rdx,     %rax  /* rax = ip0 - ilimit */
+    movq %rax,    %r15   /* r15 = ip0 - ilimit */
+
+    /* rdx = rax / 7 */
+    movabsq $2635249153387078803, %rdx
+    mulq %rdx
+    subq %rdx, %r15
+    shrq %r15
+    addq %r15, %rdx
+    shrq $2, %rdx
+
+    /* r15 = (ip0 - ilimit) / 7 */
+    movq %rdx, %r15
+
+    /* r15 = min(r15, min(oend0 - op0, oend1 - op1, oend2 - op2, oend3 - op3) / 10) */
+    movq 8(%rsp),  %rax /* rax = oend0 */
+    subq %op0,     %rax /* rax = oend0 - op0 */
+    movq 16(%rsp), %rdx /* rdx = oend1 */
+    subq %op1,     %rdx /* rdx = oend1 - op1 */
+
+    cmpq  %rax,    %rdx
+    cmova %rax,    %rdx /* rdx = min(%rdx, %rax) */
+
+    movq 24(%rsp), %rax /* rax = oend2 */
+    subq %op2,     %rax /* rax = oend2 - op2 */
+
+    cmpq  %rax,    %rdx
+    cmova %rax,    %rdx /* rdx = min(%rdx, %rax) */
+
+    movq 32(%rsp), %rax /* rax = oend3 */
+    subq %op3,     %rax /* rax = oend3 - op3 */
+
+    cmpq  %rax,    %rdx
+    cmova %rax,    %rdx /* rdx = min(%rdx, %rax) */
+
+    movabsq $-3689348814741910323, %rax
+    mulq %rdx
+    shrq $3,       %rdx /* rdx = rdx / 10 */
+
+    /* r15 = min(%rdx, %r15) */
+    cmpq  %rdx, %r15
+    cmova %rdx, %r15
+
+    /* olimit = op3 + 5 * r15 */
+    movq %r15, %rax
+    leaq (%op3, %rax, 4), %olimit
+    addq %rax, %olimit
+
+    movq 0(%rsp), %rdx
+
+    /* If (op3 + 10 > olimit) */
+    movq %op3, %rax    /* rax = op3 */
+    addq $10,  %rax    /* rax = op3 + 10 */
+    cmpq %rax, %olimit /* op3 + 10 > olimit */
+    jb .L_4X2_exit
+
+    /* If (ip1 < ip0) go to exit */
+    cmpq %ip0, %ip1
+    jb .L_4X2_exit
+
+    /* If (ip2 < ip1) go to exit */
+    cmpq %ip1, %ip2
+    jb .L_4X2_exit
+
+    /* If (ip3 < ip2) go to exit */
+    cmpq %ip2, %ip3
+    jb .L_4X2_exit
+
+#define DECODE(n, idx)              \
+    movq %bits##n, %rax;            \
+    shrq $53, %rax;                 \
+    movzwl 0(%dtable,%rax,4),%r8d;  \
+    movzbl 2(%dtable,%rax,4),%r15d; \
+    movzbl 3(%dtable,%rax,4),%eax;  \
+    movw %r8w, (%op##n);            \
+    shlxq %r15, %bits##n, %bits##n; \
+    addq %rax, %op##n
+
+#define RELOAD_BITS(n)              \
+    bsfq %bits##n, %bits##n;        \
+    movq %bits##n, %rax;            \
+    shrq $3, %bits##n;              \
+    andq $7, %rax;                  \
+    subq %bits##n, %ip##n;          \
+    movq (%ip##n), %bits##n;        \
+    orq $1, %bits##n;               \
+    shlxq %rax, %bits##n, %bits##n
+
+
+    movq %olimit, 48(%rsp)
+
+    .p2align 6
+
+.L_4X2_loop_body:
+    /* We clobber r8, so store it on the stack */
+    movq %r8, 0(%rsp)
+
+    /* Decode 5 symbols from each of the 4 streams (20 symbols total). */
+    FOR_EACH_STREAM_WITH_INDEX(DECODE, 0)
+    FOR_EACH_STREAM_WITH_INDEX(DECODE, 1)
+    FOR_EACH_STREAM_WITH_INDEX(DECODE, 2)
+    FOR_EACH_STREAM_WITH_INDEX(DECODE, 3)
+    FOR_EACH_STREAM_WITH_INDEX(DECODE, 4)
+
+    /* Reload r8 */
+    movq 0(%rsp), %r8
+
+    FOR_EACH_STREAM(RELOAD_BITS)
+
+    cmp %op3, 48(%rsp)
+    ja .L_4X2_loop_body
+    jmp .L_4X2_compute_olimit
+
+#undef DECODE
+#undef RELOAD_BITS
+.L_4X2_exit:
+    addq $8, %rsp
+    /* Restore stack (oend & olimit) */
+    pop %rax /* oend0 */
+    pop %rax /* oend1 */
+    pop %rax /* oend2 */
+    pop %rax /* oend3 */
+    pop %rax /* ilimit */
+    pop %rax /* olimit */
+    pop %rax /* arg */
+
+    /* Save ip / op / bits */
+    movq %ip0,  0(%rax)
+    movq %ip1,  8(%rax)
+    movq %ip2, 16(%rax)
+    movq %ip3, 24(%rax)
+    movq %op0, 32(%rax)
+    movq %op1, 40(%rax)
+    movq %op2, 48(%rax)
+    movq %op3, 56(%rax)
+    movq %bits0, 64(%rax)
+    movq %bits1, 72(%rax)
+    movq %bits2, 80(%rax)
+    movq %bits3, 88(%rax)
+
+    /* Restore registers */
+    pop %r15
+    pop %r14
+    pop %r13
+    pop %r12
+    pop %r11
+    pop %r10
+    pop %r9
+    pop %r8
+    pop %rdi
+    pop %rsi
+    pop %rbp
+    pop %rdx
+    pop %rcx
+    pop %rbx
+    pop %rax
+    ret
+
+#endif

src/borg/algorithms/zstd/lib/decompress/zstd_ddict.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -14,12 +14,12 @@
 /*-*******************************************************
 *  Dependencies
 *********************************************************/
-#include <string.h>      /* memcpy, memmove, memset */
+#include "../common/allocations.h"  /* ZSTD_customMalloc, ZSTD_customFree */
+#include "../common/zstd_deps.h"   /* ZSTD_memcpy, ZSTD_memmove, ZSTD_memset */
 #include "../common/cpu.h"         /* bmi2 */
 #include "../common/mem.h"         /* low level memory routines */
 #define FSE_STATIC_LINKING_ONLY
 #include "../common/fse.h"
-#define HUF_STATIC_LINKING_ONLY
 #include "../common/huf.h"
 #include "zstd_decompress_internal.h"
 #include "zstd_ddict.h"
@@ -127,14 +127,14 @@ static size_t ZSTD_initDDict_internal(ZSTD_DDict* ddict,
         ddict->dictContent = dict;
         if (!dict) dictSize = 0;
     } else {
-        void* const internalBuffer = ZSTD_malloc(dictSize, ddict->cMem);
+        void* const internalBuffer = ZSTD_customMalloc(dictSize, ddict->cMem);
         ddict->dictBuffer = internalBuffer;
         ddict->dictContent = internalBuffer;
         if (!internalBuffer) return ERROR(memory_allocation);
-        memcpy(internalBuffer, dict, dictSize);
+        ZSTD_memcpy(internalBuffer, dict, dictSize);
     }
     ddict->dictSize = dictSize;
-    ddict->entropy.hufTable[0] = (HUF_DTable)((HufLog)*0x1000001);  /* cover both little and big endian */
+    ddict->entropy.hufTable[0] = (HUF_DTable)((ZSTD_HUFFDTABLE_CAPACITY_LOG)*0x1000001);  /* cover both little and big endian */
 
     /* parse dictionary content */
     FORWARD_IF_ERROR( ZSTD_loadEntropy_intoDDict(ddict, dictContentType) , "");
@@ -147,9 +147,9 @@ ZSTD_DDict* ZSTD_createDDict_advanced(const void* dict, size_t dictSize,
                                       ZSTD_dictContentType_e dictContentType,
                                       ZSTD_customMem customMem)
 {
-    if (!customMem.customAlloc ^ !customMem.customFree) return NULL;
+    if ((!customMem.customAlloc) ^ (!customMem.customFree)) return NULL;
 
-    {   ZSTD_DDict* const ddict = (ZSTD_DDict*) ZSTD_malloc(sizeof(ZSTD_DDict), customMem);
+    {   ZSTD_DDict* const ddict = (ZSTD_DDict*) ZSTD_customMalloc(sizeof(ZSTD_DDict), customMem);
         if (ddict == NULL) return NULL;
         ddict->cMem = customMem;
         {   size_t const initResult = ZSTD_initDDict_internal(ddict,
@@ -198,7 +198,7 @@ const ZSTD_DDict* ZSTD_initStaticDDict(
     if ((size_t)sBuffer & 7) return NULL;   /* 8-aligned */
     if (sBufferSize < neededSpace) return NULL;
     if (dictLoadMethod == ZSTD_dlm_byCopy) {
-        memcpy(ddict+1, dict, dictSize);  /* local copy */
+        ZSTD_memcpy(ddict+1, dict, dictSize);  /* local copy */
         dict = ddict+1;
     }
     if (ZSTD_isError( ZSTD_initDDict_internal(ddict,
@@ -213,8 +213,8 @@ size_t ZSTD_freeDDict(ZSTD_DDict* ddict)
 {
     if (ddict==NULL) return 0;   /* support free on NULL */
     {   ZSTD_customMem const cMem = ddict->cMem;
-        ZSTD_free(ddict->dictBuffer, cMem);
-        ZSTD_free(ddict, cMem);
+        ZSTD_customFree(ddict->dictBuffer, cMem);
+        ZSTD_customFree(ddict, cMem);
         return 0;
     }
 }
@@ -240,5 +240,5 @@ size_t ZSTD_sizeof_DDict(const ZSTD_DDict* ddict)
 unsigned ZSTD_getDictID_fromDDict(const ZSTD_DDict* ddict)
 {
     if (ddict==NULL) return 0;
-    return ZSTD_getDictID_fromDict(ddict->dictContent, ddict->dictSize);
+    return ddict->dictID;
 }

src/borg/algorithms/zstd/lib/decompress/zstd_ddict.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -15,7 +15,7 @@
 /*-*******************************************************
  *  Dependencies
  *********************************************************/
-#include <stddef.h>   /* size_t */
+#include "../common/zstd_deps.h"   /* size_t */
 #include "../zstd.h"     /* ZSTD_DDict, and several public functions */
 
 

src/borg/algorithms/zstd/lib/decompress/zstd_decompress_block.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -15,7 +15,7 @@
 /*-*******************************************************
  *  Dependencies
  *********************************************************/
-#include <stddef.h>   /* size_t */
+#include "../common/zstd_deps.h"   /* size_t */
 #include "../zstd.h"    /* DCtx, and some public functions */
 #include "../common/zstd_internal.h"  /* blockProperties_t, and some public functions */
 #include "zstd_decompress_internal.h"  /* ZSTD_seqSymbol */
@@ -33,6 +33,12 @@
  */
 
 
+ /* Streaming state is used to inform allocation of the literal buffer */
+typedef enum {
+    not_streaming = 0,
+    is_streaming = 1
+} streaming_operation;
+
 /* ZSTD_decompressBlock_internal() :
  * decompress block, starting at `src`,
  * into destination buffer `dst`.
@@ -41,19 +47,27 @@
  */
 size_t ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
                                void* dst, size_t dstCapacity,
-                         const void* src, size_t srcSize, const int frame);
+                         const void* src, size_t srcSize, const int frame, const streaming_operation streaming);
 
 /* ZSTD_buildFSETable() :
  * generate FSE decoding table for one symbol (ll, ml or off)
  * this function must be called with valid parameters only
  * (dt is large enough, normalizedCounter distribution total is a power of 2, max is within range, etc.)
  * in which case it cannot fail.
+ * The workspace must be 4-byte aligned and at least ZSTD_BUILD_FSE_TABLE_WKSP_SIZE bytes, which is
+ * defined in zstd_decompress_internal.h.
  * Internal use only.
  */
 void ZSTD_buildFSETable(ZSTD_seqSymbol* dt,
              const short* normalizedCounter, unsigned maxSymbolValue,
-             const U32* baseValue, const U32* nbAdditionalBits,
-                   unsigned tableLog);
+             const U32* baseValue, const U8* nbAdditionalBits,
+                   unsigned tableLog, void* wksp, size_t wkspSize,
+                   int bmi2);
+
+/* Internal definition of ZSTD_decompressBlock() to avoid deprecation warnings. */
+size_t ZSTD_decompressBlock_deprecated(ZSTD_DCtx* dctx,
+                            void* dst, size_t dstCapacity,
+                      const void* src, size_t srcSize);
 
 
 #endif /* ZSTD_DEC_BLOCK_H */

src/borg/algorithms/zstd/lib/decompress/zstd_decompress_internal.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -20,33 +20,33 @@
  *  Dependencies
  *********************************************************/
 #include "../common/mem.h"             /* BYTE, U16, U32 */
-#include "../common/zstd_internal.h"   /* ZSTD_seqSymbol */
+#include "../common/zstd_internal.h"   /* constants : MaxLL, MaxML, MaxOff, LLFSELog, etc. */
 
 
 
 /*-*******************************************************
  *  Constants
  *********************************************************/
-static const U32 LL_base[MaxLL+1] = {
+static UNUSED_ATTR const U32 LL_base[MaxLL+1] = {
                  0,    1,    2,     3,     4,     5,     6,      7,
                  8,    9,   10,    11,    12,    13,    14,     15,
                 16,   18,   20,    22,    24,    28,    32,     40,
                 48,   64, 0x80, 0x100, 0x200, 0x400, 0x800, 0x1000,
                 0x2000, 0x4000, 0x8000, 0x10000 };
 
-static const U32 OF_base[MaxOff+1] = {
+static UNUSED_ATTR const U32 OF_base[MaxOff+1] = {
                  0,        1,       1,       5,     0xD,     0x1D,     0x3D,     0x7D,
                  0xFD,   0x1FD,   0x3FD,   0x7FD,   0xFFD,   0x1FFD,   0x3FFD,   0x7FFD,
                  0xFFFD, 0x1FFFD, 0x3FFFD, 0x7FFFD, 0xFFFFD, 0x1FFFFD, 0x3FFFFD, 0x7FFFFD,
                  0xFFFFFD, 0x1FFFFFD, 0x3FFFFFD, 0x7FFFFFD, 0xFFFFFFD, 0x1FFFFFFD, 0x3FFFFFFD, 0x7FFFFFFD };
 
-static const U32 OF_bits[MaxOff+1] = {
+static UNUSED_ATTR const U8 OF_bits[MaxOff+1] = {
                      0,  1,  2,  3,  4,  5,  6,  7,
                      8,  9, 10, 11, 12, 13, 14, 15,
                     16, 17, 18, 19, 20, 21, 22, 23,
                     24, 25, 26, 27, 28, 29, 30, 31 };
 
-static const U32 ML_base[MaxML+1] = {
+static UNUSED_ATTR const U32 ML_base[MaxML+1] = {
                      3,  4,  5,    6,     7,     8,     9,    10,
                     11, 12, 13,   14,    15,    16,    17,    18,
                     19, 20, 21,   22,    23,    24,    25,    26,
@@ -73,12 +73,17 @@ static const U32 ML_base[MaxML+1] = {
 
  #define SEQSYMBOL_TABLE_SIZE(log)   (1 + (1 << (log)))
 
+#define ZSTD_BUILD_FSE_TABLE_WKSP_SIZE (sizeof(S16) * (MaxSeq + 1) + (1u << MaxFSELog) + sizeof(U64))
+#define ZSTD_BUILD_FSE_TABLE_WKSP_SIZE_U32 ((ZSTD_BUILD_FSE_TABLE_WKSP_SIZE + sizeof(U32) - 1) / sizeof(U32))
+#define ZSTD_HUFFDTABLE_CAPACITY_LOG 12
+
 typedef struct {
     ZSTD_seqSymbol LLTable[SEQSYMBOL_TABLE_SIZE(LLFSELog)];    /* Note : Space reserved for FSE Tables */
     ZSTD_seqSymbol OFTable[SEQSYMBOL_TABLE_SIZE(OffFSELog)];   /* is also used as temporary workspace while building hufTable during DDict creation */
     ZSTD_seqSymbol MLTable[SEQSYMBOL_TABLE_SIZE(MLFSELog)];    /* and therefore must be at least HUF_DECOMPRESS_WORKSPACE_SIZE large */
-    HUF_DTable hufTable[HUF_DTABLE_SIZE(HufLog)];  /* can accommodate HUF_decompress4X */
+    HUF_DTable hufTable[HUF_DTABLE_SIZE(ZSTD_HUFFDTABLE_CAPACITY_LOG)];  /* can accommodate HUF_decompress4X */
     U32 rep[ZSTD_REP_NUM];
+    U32 workspace[ZSTD_BUILD_FSE_TABLE_WKSP_SIZE_U32];
 } ZSTD_entropyDTables_t;
 
 typedef enum { ZSTDds_getFrameHeaderSize, ZSTDds_decodeFrameHeader,
@@ -95,10 +100,28 @@ typedef enum {
     ZSTD_use_once = 1            /* Use the dictionary once and set to ZSTD_dont_use */
 } ZSTD_dictUses_e;
 
+/* Hashset for storing references to multiple ZSTD_DDict within ZSTD_DCtx */
+typedef struct {
+    const ZSTD_DDict** ddictPtrTable;
+    size_t ddictPtrTableSize;
+    size_t ddictPtrCount;
+} ZSTD_DDictHashSet;
+
+#ifndef ZSTD_DECODER_INTERNAL_BUFFER
+#  define ZSTD_DECODER_INTERNAL_BUFFER  (1 << 16)
+#endif
+
+#define ZSTD_LBMIN 64
+#define ZSTD_LBMAX (128 << 10)
+
+/* extra buffer, compensates when dst is not large enough to store litBuffer */
+#define ZSTD_LITBUFFEREXTRASIZE  BOUNDED(ZSTD_LBMIN, ZSTD_DECODER_INTERNAL_BUFFER, ZSTD_LBMAX)
+
 typedef enum {
-    ZSTD_obm_buffered = 0,  /* Buffer the output */
-    ZSTD_obm_stable = 1     /* ZSTD_outBuffer is stable */
-} ZSTD_outBufferMode_e;
+    ZSTD_not_in_dst = 0,  /* Stored entirely within litExtraBuffer */
+    ZSTD_in_dst = 1,           /* Stored entirely within dst (in memory after current output write) */
+    ZSTD_split = 2            /* Split between litExtraBuffer and dst */
+} ZSTD_litLocation_e;
 
 struct ZSTD_DCtx_s
 {
@@ -114,6 +137,7 @@ struct ZSTD_DCtx_s
     const void* dictEnd;          /* end of previous segment */
     size_t expected;
     ZSTD_frameHeader fParams;
+    U64 processedCSize;
     U64 decodedSize;
     blockType_e bType;            /* used in ZSTD_decompressContinue(), store blockType between block header decoding and block decompression stages */
     ZSTD_dStage stage;
@@ -122,12 +146,16 @@ struct ZSTD_DCtx_s
     XXH64_state_t xxhState;
     size_t headerSize;
     ZSTD_format_e format;
+    ZSTD_forceIgnoreChecksum_e forceIgnoreChecksum;   /* User specified: if == 1, will ignore checksums in compressed frame. Default == 0 */
+    U32 validateChecksum;         /* if == 1, will validate checksum. Is == 1 if (fParams.checksumFlag == 1) and (forceIgnoreChecksum == 0). */
     const BYTE* litPtr;
     ZSTD_customMem customMem;
     size_t litSize;
     size_t rleSize;
     size_t staticSize;
+#if DYNAMIC_BMI2 != 0
     int bmi2;                     /* == 1 if the CPU supports BMI2 and 0 otherwise. CPU support is determined dynamically once per context lifetime. */
+#endif
 
     /* dictionary */
     ZSTD_DDict* ddictLocal;
@@ -135,6 +163,9 @@ struct ZSTD_DCtx_s
     U32 dictID;
     int ddictIsCold;             /* if == 1 : dictionary is "new" for working context, and presumed "cold" (not in cpu cache) */
     ZSTD_dictUses_e dictUses;
+    ZSTD_DDictHashSet* ddictSet;                    /* Hash set for multiple ddicts */
+    ZSTD_refMultipleDDicts_e refMultipleDDicts;     /* User specified: if == 1, will allow references to multiple DDicts. Default == 0 (disabled) */
+    int disableHufAsm;
 
     /* streaming */
     ZSTD_dStreamStage streamStage;
@@ -147,16 +178,21 @@ struct ZSTD_DCtx_s
     size_t outStart;
     size_t outEnd;
     size_t lhSize;
+#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
     void* legacyContext;
     U32 previousLegacyVersion;
     U32 legacyVersion;
+#endif
     U32 hostageByte;
     int noForwardProgress;
-    ZSTD_outBufferMode_e outBufferMode;
+    ZSTD_bufferMode_e outBufferMode;
     ZSTD_outBuffer expectedOutBuffer;
 
     /* workspace */
-    BYTE litBuffer[ZSTD_BLOCKSIZE_MAX + WILDCOPY_OVERLENGTH];
+    BYTE* litBuffer;
+    const BYTE* litBufferEnd;
+    ZSTD_litLocation_e litBufferLocation;
+    BYTE litExtraBuffer[ZSTD_LITBUFFEREXTRASIZE + WILDCOPY_OVERLENGTH]; /* literal buffer can be split between storage within dst and within this scratch buffer */
     BYTE headerBuffer[ZSTD_FRAMEHEADERSIZE_MAX];
 
     size_t oversizedDuration;
@@ -165,8 +201,21 @@ struct ZSTD_DCtx_s
     void const* dictContentBeginForFuzzing;
     void const* dictContentEndForFuzzing;
 #endif
+
+    /* Tracing */
+#if ZSTD_TRACE
+    ZSTD_TraceCtx traceCtx;
+#endif
 };  /* typedef'd to ZSTD_DCtx within "zstd.h" */
 
+MEM_STATIC int ZSTD_DCtx_get_bmi2(const struct ZSTD_DCtx_s *dctx) {
+#if DYNAMIC_BMI2 != 0
+	return dctx->bmi2;
+#else
+    (void)dctx;
+	return 0;
+#endif
+}
 
 /*-*******************************************************
  *  Shared internal functions
@@ -183,7 +232,7 @@ size_t ZSTD_loadDEntropy(ZSTD_entropyDTables_t* entropy,
  *  If yes, do nothing (continue on current segment).
  *  If not, classify previous segment as "external dictionary", and start a new segment.
  *  This function cannot fail. */
-void ZSTD_checkContinuity(ZSTD_DCtx* dctx, const void* dst);
+void ZSTD_checkContinuity(ZSTD_DCtx* dctx, const void* dst, size_t dstSize);
 
 
 #endif /* ZSTD_DECOMPRESS_INTERNAL_H */

src/borg/algorithms/zstd/lib/deprecated/zbuff.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the

src/borg/algorithms/zstd/lib/deprecated/zbuff_common.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the

src/borg/algorithms/zstd/lib/deprecated/zbuff_compress.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -15,6 +15,7 @@
 ***************************************/
 #define ZBUFF_STATIC_LINKING_ONLY
 #include "zbuff.h"
+#include "../common/error_private.h"
 
 
 /*-***********************************************************
@@ -73,13 +74,32 @@ size_t ZBUFF_compressInit_advanced(ZBUFF_CCtx* zbc,
                                    ZSTD_parameters params, unsigned long long pledgedSrcSize)
 {
     if (pledgedSrcSize==0) pledgedSrcSize = ZSTD_CONTENTSIZE_UNKNOWN;  /* preserve "0 == unknown" behavior */
-    return ZSTD_initCStream_advanced(zbc, dict, dictSize, params, pledgedSrcSize);
+    FORWARD_IF_ERROR(ZSTD_CCtx_reset(zbc, ZSTD_reset_session_only), "");
+    FORWARD_IF_ERROR(ZSTD_CCtx_setPledgedSrcSize(zbc, pledgedSrcSize), "");
+
+    FORWARD_IF_ERROR(ZSTD_checkCParams(params.cParams), "");
+    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(zbc, ZSTD_c_windowLog, params.cParams.windowLog), "");
+    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(zbc, ZSTD_c_hashLog, params.cParams.hashLog), "");
+    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(zbc, ZSTD_c_chainLog, params.cParams.chainLog), "");
+    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(zbc, ZSTD_c_searchLog, params.cParams.searchLog), "");
+    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(zbc, ZSTD_c_minMatch, params.cParams.minMatch), "");
+    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(zbc, ZSTD_c_targetLength, params.cParams.targetLength), "");
+    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(zbc, ZSTD_c_strategy, params.cParams.strategy), "");
+
+    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(zbc, ZSTD_c_contentSizeFlag, params.fParams.contentSizeFlag), "");
+    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(zbc, ZSTD_c_checksumFlag, params.fParams.checksumFlag), "");
+    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(zbc, ZSTD_c_dictIDFlag, params.fParams.noDictIDFlag), "");
+
+    FORWARD_IF_ERROR(ZSTD_CCtx_loadDictionary(zbc, dict, dictSize), "");
+    return 0;
 }
 
-
 size_t ZBUFF_compressInitDictionary(ZBUFF_CCtx* zbc, const void* dict, size_t dictSize, int compressionLevel)
 {
-    return ZSTD_initCStream_usingDict(zbc, dict, dictSize, compressionLevel);
+    FORWARD_IF_ERROR(ZSTD_CCtx_reset(zbc, ZSTD_reset_session_only), "");
+    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(zbc, ZSTD_c_compressionLevel, compressionLevel), "");
+    FORWARD_IF_ERROR(ZSTD_CCtx_loadDictionary(zbc, dict, dictSize), "");
+    return 0;
 }
 
 size_t ZBUFF_compressInit(ZBUFF_CCtx* zbc, int compressionLevel)

src/borg/algorithms/zstd/lib/deprecated/zbuff_decompress.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -13,6 +13,8 @@
 /* *************************************
 *  Dependencies
 ***************************************/
+#define ZSTD_DISABLE_DEPRECATE_WARNINGS  /* suppress warning on ZSTD_initDStream_usingDict */
+#include "../zstd.h"        /* ZSTD_CStream, ZSTD_DStream, ZSTDLIB_API */
 #define ZBUFF_STATIC_LINKING_ONLY
 #include "zbuff.h"
 

src/borg/algorithms/zstd/lib/dictBuilder/cover.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -26,47 +26,65 @@
 #include <string.h> /* memset */
 #include <time.h>   /* clock */
 
+#ifndef ZDICT_STATIC_LINKING_ONLY
+#  define ZDICT_STATIC_LINKING_ONLY
+#endif
+
 #include "../common/mem.h" /* read */
 #include "../common/pool.h"
 #include "../common/threading.h"
-#include "cover.h"
 #include "../common/zstd_internal.h" /* includes zstd.h */
-#ifndef ZDICT_STATIC_LINKING_ONLY
-#define ZDICT_STATIC_LINKING_ONLY
-#endif
-#include "zdict.h"
+#include "../common/bits.h" /* ZSTD_highbit32 */
+#include "../zdict.h"
+#include "cover.h"
 
 /*-*************************************
 *  Constants
 ***************************************/
+/**
+* There are 32bit indexes used to ref samples, so limit samples size to 4GB
+* on 64bit builds.
+* For 32bit builds we choose 1 GB.
+* Most 32bit platforms have 2GB user-mode addressable space and we allocate a large
+* contiguous buffer, so 1GB is already a high limit.
+*/
 #define COVER_MAX_SAMPLES_SIZE (sizeof(size_t) == 8 ? ((unsigned)-1) : ((unsigned)1 GB))
-#define DEFAULT_SPLITPOINT 1.0
+#define COVER_DEFAULT_SPLITPOINT 1.0
 
 /*-*************************************
 *  Console display
 ***************************************/
-static int g_displayLevel = 2;
+#ifndef LOCALDISPLAYLEVEL
+static int g_displayLevel = 0;
+#endif
+#undef  DISPLAY
 #define DISPLAY(...)                                                           \
   {                                                                            \
     fprintf(stderr, __VA_ARGS__);                                              \
     fflush(stderr);                                                            \
   }
+#undef  LOCALDISPLAYLEVEL
 #define LOCALDISPLAYLEVEL(displayLevel, l, ...)                                \
   if (displayLevel >= l) {                                                     \
     DISPLAY(__VA_ARGS__);                                                      \
   } /* 0 : no display;   1: errors;   2: default;  3: details;  4: debug */
+#undef  DISPLAYLEVEL
 #define DISPLAYLEVEL(l, ...) LOCALDISPLAYLEVEL(g_displayLevel, l, __VA_ARGS__)
 
+#ifndef LOCALDISPLAYUPDATE
+static const clock_t g_refreshRate = CLOCKS_PER_SEC * 15 / 100;
+static clock_t g_time = 0;
+#endif
+#undef  LOCALDISPLAYUPDATE
 #define LOCALDISPLAYUPDATE(displayLevel, l, ...)                               \
   if (displayLevel >= l) {                                                     \
-    if ((clock() - g_time > refreshRate) || (displayLevel >= 4)) {             \
+    if ((clock() - g_time > g_refreshRate) || (displayLevel >= 4)) {             \
       g_time = clock();                                                        \
       DISPLAY(__VA_ARGS__);                                                    \
     }                                                                          \
   }
+#undef  DISPLAYUPDATE
 #define DISPLAYUPDATE(l, ...) LOCALDISPLAYUPDATE(g_displayLevel, l, __VA_ARGS__)
-static const clock_t refreshRate = CLOCKS_PER_SEC * 15 / 100;
-static clock_t g_time = 0;
 
 /*-*************************************
 * Hash table
@@ -120,9 +138,9 @@ static int COVER_map_init(COVER_map_t *map, U32 size) {
 /**
  * Internal hash function
  */
-static const U32 prime4bytes = 2654435761U;
+static const U32 COVER_prime4bytes = 2654435761U;
 static U32 COVER_map_hash(COVER_map_t *map, U32 key) {
-  return (key * prime4bytes) >> (32 - map->sizeLog);
+  return (key * COVER_prime4bytes) >> (32 - map->sizeLog);
 }
 
 /**
@@ -215,7 +233,7 @@ typedef struct {
 } COVER_ctx_t;
 
 /* We need a global context for qsort... */
-static COVER_ctx_t *g_ctx = NULL;
+static COVER_ctx_t *g_coverCtx = NULL;
 
 /*-*************************************
 *  Helper functions
@@ -258,11 +276,11 @@ static int COVER_cmp8(COVER_ctx_t *ctx, const void *lp, const void *rp) {
 
 /**
  * Same as COVER_cmp() except ties are broken by pointer value
- * NOTE: g_ctx must be set to call this function.  A global is required because
+ * NOTE: g_coverCtx must be set to call this function.  A global is required because
  * qsort doesn't take an opaque pointer.
  */
-static int COVER_strict_cmp(const void *lp, const void *rp) {
-  int result = COVER_cmp(g_ctx, lp, rp);
+static int WIN_CDECL COVER_strict_cmp(const void *lp, const void *rp) {
+  int result = COVER_cmp(g_coverCtx, lp, rp);
   if (result == 0) {
     result = lp < rp ? -1 : 1;
   }
@@ -271,8 +289,8 @@ static int COVER_strict_cmp(const void *lp, const void *rp) {
 /**
  * Faster version for d <= 8.
  */
-static int COVER_strict_cmp8(const void *lp, const void *rp) {
-  int result = COVER_cmp8(g_ctx, lp, rp);
+static int WIN_CDECL COVER_strict_cmp8(const void *lp, const void *rp) {
+  int result = COVER_cmp8(g_coverCtx, lp, rp);
   if (result == 0) {
     result = lp < rp ? -1 : 1;
   }
@@ -524,7 +542,7 @@ static void COVER_ctx_destroy(COVER_ctx_t *ctx) {
 
 /**
  * Prepare a context for dictionary building.
- * The context is only dependent on the parameter `d` and can used multiple
+ * The context is only dependent on the parameter `d` and can be used multiple
  * times.
  * Returns 0 on success or error code on error.
  * The context must be destroyed with `COVER_ctx_destroy()`.
@@ -603,7 +621,7 @@ static size_t COVER_ctx_init(COVER_ctx_t *ctx, const void *samplesBuffer,
     /* qsort doesn't take an opaque pointer, so pass as a global.
      * On OpenBSD qsort() is not guaranteed to be stable, their mergesort() is.
      */
-    g_ctx = ctx;
+    g_coverCtx = ctx;
 #if defined(__OpenBSD__)
     mergesort(ctx->suffix, ctx->suffixSize, sizeof(U32),
           (ctx->d <= 8 ? &COVER_strict_cmp8 : &COVER_strict_cmp));
@@ -629,7 +647,7 @@ static size_t COVER_ctx_init(COVER_ctx_t *ctx, const void *samplesBuffer,
 
 void COVER_warnOnSmallCorpus(size_t maxDictSize, size_t nbDmers, int displayLevel)
 {
-  const double ratio = (double)nbDmers / maxDictSize;
+  const double ratio = (double)nbDmers / (double)maxDictSize;
   if (ratio >= 10) {
       return;
   }
@@ -725,7 +743,7 @@ ZDICTLIB_API size_t ZDICT_trainFromBuffer_cover(
   COVER_map_t activeDmers;
   parameters.splitPoint = 1.0;
   /* Initialize global data */
-  g_displayLevel = parameters.zParams.notificationLevel;
+  g_displayLevel = (int)parameters.zParams.notificationLevel;
   /* Checks */
   if (!COVER_checkParameters(parameters, dictBufferCapacity)) {
     DISPLAYLEVEL(1, "Cover parameters incorrect\n");
@@ -933,9 +951,17 @@ void COVER_best_finish(COVER_best_t *best, ZDICT_cover_params_t parameters,
   }
 }
 
+static COVER_dictSelection_t setDictSelection(BYTE* buf, size_t s, size_t csz)
+{
+    COVER_dictSelection_t ds;
+    ds.dictContent = buf;
+    ds.dictSize = s;
+    ds.totalCompressedSize = csz;
+    return ds;
+}
+
 COVER_dictSelection_t COVER_dictSelectionError(size_t error) {
-    COVER_dictSelection_t selection = { NULL, 0, error };
-    return selection;
+    return setDictSelection(NULL, 0, error);
 }
 
 unsigned COVER_dictSelectionIsError(COVER_dictSelection_t selection) {
@@ -946,7 +972,7 @@ void COVER_dictSelectionFree(COVER_dictSelection_t selection){
   free(selection.dictContent);
 }
 
-COVER_dictSelection_t COVER_selectDict(BYTE* customDictContent,
+COVER_dictSelection_t COVER_selectDict(BYTE* customDictContent, size_t dictBufferCapacity,
         size_t dictContentSize, const BYTE* samplesBuffer, const size_t* samplesSizes, unsigned nbFinalizeSamples,
         size_t nbCheckSamples, size_t nbSamples, ZDICT_cover_params_t params, size_t* offsets, size_t totalCompressedSize) {
 
@@ -954,8 +980,8 @@ COVER_dictSelection_t COVER_selectDict(BYTE* customDictContent,
   size_t largestCompressed = 0;
   BYTE* customDictContentEnd = customDictContent + dictContentSize;
 
-  BYTE * largestDictbuffer = (BYTE *)malloc(dictContentSize);
-  BYTE * candidateDictBuffer = (BYTE *)malloc(dictContentSize);
+  BYTE * largestDictbuffer = (BYTE *)malloc(dictBufferCapacity);
+  BYTE * candidateDictBuffer = (BYTE *)malloc(dictBufferCapacity);
   double regressionTolerance = ((double)params.shrinkDictMaxRegression / 100.0) + 1.00;
 
   if (!largestDictbuffer || !candidateDictBuffer) {
@@ -967,7 +993,7 @@ COVER_dictSelection_t COVER_selectDict(BYTE* customDictContent,
   /* Initial dictionary size and compressed size */
   memcpy(largestDictbuffer, customDictContent, dictContentSize);
   dictContentSize = ZDICT_finalizeDictionary(
-    largestDictbuffer, dictContentSize, customDictContent, dictContentSize,
+    largestDictbuffer, dictBufferCapacity, customDictContent, dictContentSize,
     samplesBuffer, samplesSizes, nbFinalizeSamples, params.zParams);
 
   if (ZDICT_isError(dictContentSize)) {
@@ -988,9 +1014,8 @@ COVER_dictSelection_t COVER_selectDict(BYTE* customDictContent,
   }
 
   if (params.shrinkDict == 0) {
-    COVER_dictSelection_t selection = { largestDictbuffer, dictContentSize, totalCompressedSize };
     free(candidateDictBuffer);
-    return selection;
+    return setDictSelection(largestDictbuffer, dictContentSize, totalCompressedSize);
   }
 
   largestDict = dictContentSize;
@@ -1001,7 +1026,7 @@ COVER_dictSelection_t COVER_selectDict(BYTE* customDictContent,
   while (dictContentSize < largestDict) {
     memcpy(candidateDictBuffer, largestDictbuffer, largestDict);
     dictContentSize = ZDICT_finalizeDictionary(
-      candidateDictBuffer, dictContentSize, customDictContentEnd - dictContentSize, dictContentSize,
+      candidateDictBuffer, dictBufferCapacity, customDictContentEnd - dictContentSize, dictContentSize,
       samplesBuffer, samplesSizes, nbFinalizeSamples, params.zParams);
 
     if (ZDICT_isError(dictContentSize)) {
@@ -1022,20 +1047,16 @@ COVER_dictSelection_t COVER_selectDict(BYTE* customDictContent,
       return COVER_dictSelectionError(totalCompressedSize);
     }
 
-    if (totalCompressedSize <= largestCompressed * regressionTolerance) {
-      COVER_dictSelection_t selection = { candidateDictBuffer, dictContentSize, totalCompressedSize };
+    if ((double)totalCompressedSize <= (double)largestCompressed * regressionTolerance) {
       free(largestDictbuffer);
-      return selection;
+      return setDictSelection( candidateDictBuffer, dictContentSize, totalCompressedSize );
     }
     dictContentSize *= 2;
   }
   dictContentSize = largestDict;
   totalCompressedSize = largestCompressed;
-  {
-    COVER_dictSelection_t selection = { largestDictbuffer, dictContentSize, totalCompressedSize };
-    free(candidateDictBuffer);
-    return selection;
-  }
+  free(candidateDictBuffer);
+  return setDictSelection( largestDictbuffer, dictContentSize, totalCompressedSize );
 }
 
 /**
@@ -1053,18 +1074,19 @@ typedef struct COVER_tryParameters_data_s {
  * This function is thread safe if zstd is compiled with multithreaded support.
  * It takes its parameters as an *OWNING* opaque pointer to support threading.
  */
-static void COVER_tryParameters(void *opaque) {
+static void COVER_tryParameters(void *opaque)
+{
   /* Save parameters as local variables */
-  COVER_tryParameters_data_t *const data = (COVER_tryParameters_data_t *)opaque;
+  COVER_tryParameters_data_t *const data = (COVER_tryParameters_data_t*)opaque;
   const COVER_ctx_t *const ctx = data->ctx;
   const ZDICT_cover_params_t parameters = data->parameters;
   size_t dictBufferCapacity = data->dictBufferCapacity;
   size_t totalCompressedSize = ERROR(GENERIC);
   /* Allocate space for hash table, dict, and freqs */
   COVER_map_t activeDmers;
-  BYTE *const dict = (BYTE * const)malloc(dictBufferCapacity);
+  BYTE* const dict = (BYTE*)malloc(dictBufferCapacity);
   COVER_dictSelection_t selection = COVER_dictSelectionError(ERROR(GENERIC));
-  U32 *freqs = (U32 *)malloc(ctx->suffixSize * sizeof(U32));
+  U32* const freqs = (U32*)malloc(ctx->suffixSize * sizeof(U32));
   if (!COVER_map_init(&activeDmers, parameters.k - parameters.d + 1)) {
     DISPLAYLEVEL(1, "Failed to allocate dmer map: out of memory\n");
     goto _cleanup;
@@ -1079,7 +1101,7 @@ static void COVER_tryParameters(void *opaque) {
   {
     const size_t tail = COVER_buildDictionary(ctx, freqs, &activeDmers, dict,
                                               dictBufferCapacity, parameters);
-    selection = COVER_selectDict(dict + tail, dictBufferCapacity - tail,
+    selection = COVER_selectDict(dict + tail, dictBufferCapacity, dictBufferCapacity - tail,
         ctx->samples, ctx->samplesSizes, (unsigned)ctx->nbTrainSamples, ctx->nbTrainSamples, ctx->nbSamples, parameters, ctx->offsets,
         totalCompressedSize);
 
@@ -1094,19 +1116,18 @@ _cleanup:
   free(data);
   COVER_map_destroy(&activeDmers);
   COVER_dictSelectionFree(selection);
-  if (freqs) {
-    free(freqs);
-  }
+  free(freqs);
 }
 
 ZDICTLIB_API size_t ZDICT_optimizeTrainFromBuffer_cover(
-    void *dictBuffer, size_t dictBufferCapacity, const void *samplesBuffer,
-    const size_t *samplesSizes, unsigned nbSamples,
-    ZDICT_cover_params_t *parameters) {
+    void* dictBuffer, size_t dictBufferCapacity, const void* samplesBuffer,
+    const size_t* samplesSizes, unsigned nbSamples,
+    ZDICT_cover_params_t* parameters)
+{
   /* constants */
   const unsigned nbThreads = parameters->nbThreads;
   const double splitPoint =
-      parameters->splitPoint <= 0.0 ? DEFAULT_SPLITPOINT : parameters->splitPoint;
+      parameters->splitPoint <= 0.0 ? COVER_DEFAULT_SPLITPOINT : parameters->splitPoint;
   const unsigned kMinD = parameters->d == 0 ? 6 : parameters->d;
   const unsigned kMaxD = parameters->d == 0 ? 8 : parameters->d;
   const unsigned kMinK = parameters->k == 0 ? 50 : parameters->k;

src/borg/algorithms/zstd/lib/dictBuilder/cover.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -8,6 +8,10 @@
  * You may select, at your option, one of the above-listed licenses.
  */
 
+#ifndef ZDICT_STATIC_LINKING_ONLY
+#  define ZDICT_STATIC_LINKING_ONLY
+#endif
+
 #include <stdio.h>  /* fprintf */
 #include <stdlib.h> /* malloc, free, qsort */
 #include <string.h> /* memset */
@@ -16,10 +20,7 @@
 #include "../common/pool.h"
 #include "../common/threading.h"
 #include "../common/zstd_internal.h" /* includes zstd.h */
-#ifndef ZDICT_STATIC_LINKING_ONLY
-#define ZDICT_STATIC_LINKING_ONLY
-#endif
-#include "zdict.h"
+#include "../zdict.h"
 
 /**
  * COVER_best_t is used for two purposes:
@@ -152,6 +153,6 @@ void COVER_dictSelectionFree(COVER_dictSelection_t selection);
  * smallest dictionary within a specified regression of the compressed size
  * from the largest dictionary.
  */
- COVER_dictSelection_t COVER_selectDict(BYTE* customDictContent,
+ COVER_dictSelection_t COVER_selectDict(BYTE* customDictContent, size_t dictBufferCapacity,
                        size_t dictContentSize, const BYTE* samplesBuffer, const size_t* samplesSizes, unsigned nbFinalizeSamples,
                        size_t nbCheckSamples, size_t nbSamples, ZDICT_cover_params_t params, size_t* offsets, size_t totalCompressedSize);

src/borg/algorithms/zstd/lib/dictBuilder/divsufsort.c

@@ -1576,7 +1576,7 @@ note:
     /* Construct the inverse suffix array of type B* suffixes using trsort. */
     trsort(ISAb, SA, m, 1);
 
-    /* Set the sorted order of tyoe B* suffixes. */
+    /* Set the sorted order of type B* suffixes. */
     for(i = n - 1, j = m, c0 = T[n - 1]; 0 <= i;) {
       for(--i, c1 = c0; (0 <= i) && ((c0 = T[i]) >= c1); --i, c1 = c0) { }
       if(0 <= i) {

src/borg/algorithms/zstd/lib/dictBuilder/fastcover.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2020, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -16,24 +16,33 @@
 #include <string.h> /* memset */
 #include <time.h>   /* clock */
 
+#ifndef ZDICT_STATIC_LINKING_ONLY
+#  define ZDICT_STATIC_LINKING_ONLY
+#endif
+
 #include "../common/mem.h" /* read */
 #include "../common/pool.h"
 #include "../common/threading.h"
-#include "cover.h"
 #include "../common/zstd_internal.h" /* includes zstd.h */
-#ifndef ZDICT_STATIC_LINKING_ONLY
-#define ZDICT_STATIC_LINKING_ONLY
-#endif
-#include "zdict.h"
+#include "../compress/zstd_compress_internal.h" /* ZSTD_hash*() */
+#include "../zdict.h"
+#include "cover.h"
 
 
 /*-*************************************
 *  Constants
 ***************************************/
+/**
+* There are 32bit indexes used to ref samples, so limit samples size to 4GB
+* on 64bit builds.
+* For 32bit builds we choose 1 GB.
+* Most 32bit platforms have 2GB user-mode addressable space and we allocate a large
+* contiguous buffer, so 1GB is already a high limit.
+*/
 #define FASTCOVER_MAX_SAMPLES_SIZE (sizeof(size_t) == 8 ? ((unsigned)-1) : ((unsigned)1 GB))
 #define FASTCOVER_MAX_F 31
 #define FASTCOVER_MAX_ACCEL 10
-#define DEFAULT_SPLITPOINT 0.75
+#define FASTCOVER_DEFAULT_SPLITPOINT 0.75
 #define DEFAULT_F 20
 #define DEFAULT_ACCEL 1
 
@@ -41,50 +50,50 @@
 /*-*************************************
 *  Console display
 ***************************************/
-static int g_displayLevel = 2;
+#ifndef LOCALDISPLAYLEVEL
+static int g_displayLevel = 0;
+#endif
+#undef  DISPLAY
 #define DISPLAY(...)                                                           \
   {                                                                            \
     fprintf(stderr, __VA_ARGS__);                                              \
     fflush(stderr);                                                            \
   }
+#undef  LOCALDISPLAYLEVEL
 #define LOCALDISPLAYLEVEL(displayLevel, l, ...)                                \
   if (displayLevel >= l) {                                                     \
     DISPLAY(__VA_ARGS__);                                                      \
   } /* 0 : no display;   1: errors;   2: default;  3: details;  4: debug */
+#undef  DISPLAYLEVEL
 #define DISPLAYLEVEL(l, ...) LOCALDISPLAYLEVEL(g_displayLevel, l, __VA_ARGS__)
 
+#ifndef LOCALDISPLAYUPDATE
+static const clock_t g_refreshRate = CLOCKS_PER_SEC * 15 / 100;
+static clock_t g_time = 0;
+#endif
+#undef  LOCALDISPLAYUPDATE
 #define LOCALDISPLAYUPDATE(displayLevel, l, ...)                               \
   if (displayLevel >= l) {                                                     \
-    if ((clock() - g_time > refreshRate) || (displayLevel >= 4)) {             \
+    if ((clock() - g_time > g_refreshRate) || (displayLevel >= 4)) {             \
       g_time = clock();                                                        \
       DISPLAY(__VA_ARGS__);                                                    \
     }                                                                          \
   }
+#undef  DISPLAYUPDATE
 #define DISPLAYUPDATE(l, ...) LOCALDISPLAYUPDATE(g_displayLevel, l, __VA_ARGS__)
-static const clock_t refreshRate = CLOCKS_PER_SEC * 15 / 100;
-static clock_t g_time = 0;
 
 
 /*-*************************************
 * Hash Functions
 ***************************************/
-static const U64 prime6bytes = 227718039650203ULL;
-static size_t ZSTD_hash6(U64 u, U32 h) { return (size_t)(((u  << (64-48)) * prime6bytes) >> (64-h)) ; }
-static size_t ZSTD_hash6Ptr(const void* p, U32 h) { return ZSTD_hash6(MEM_readLE64(p), h); }
-
-static const U64 prime8bytes = 0xCF1BBCDCB7A56463ULL;
-static size_t ZSTD_hash8(U64 u, U32 h) { return (size_t)(((u) * prime8bytes) >> (64-h)) ; }
-static size_t ZSTD_hash8Ptr(const void* p, U32 h) { return ZSTD_hash8(MEM_readLE64(p), h); }
-
-
 /**
- * Hash the d-byte value pointed to by p and mod 2^f
+ * Hash the d-byte value pointed to by p and mod 2^f into the frequency vector
  */
-static size_t FASTCOVER_hashPtrToIndex(const void* p, U32 h, unsigned d) {
+static size_t FASTCOVER_hashPtrToIndex(const void* p, U32 f, unsigned d) {
   if (d == 6) {
-    return ZSTD_hash6Ptr(p, h) & ((1 << h) - 1);
+    return ZSTD_hash6Ptr(p, f);
   }
-  return ZSTD_hash8Ptr(p, h) & ((1 << h) - 1);
+  return ZSTD_hash8Ptr(p, f);
 }
 
 
@@ -295,7 +304,7 @@ FASTCOVER_computeFrequency(U32* freqs, const FASTCOVER_ctx_t* ctx)
 
 /**
  * Prepare a context for dictionary building.
- * The context is only dependent on the parameter `d` and can used multiple
+ * The context is only dependent on the parameter `d` and can be used multiple
  * times.
  * Returns 0 on success or error code on error.
  * The context must be destroyed with `FASTCOVER_ctx_destroy()`.
@@ -461,20 +470,20 @@ typedef struct FASTCOVER_tryParameters_data_s {
  * This function is thread safe if zstd is compiled with multithreaded support.
  * It takes its parameters as an *OWNING* opaque pointer to support threading.
  */
-static void FASTCOVER_tryParameters(void *opaque)
+static void FASTCOVER_tryParameters(void* opaque)
 {
   /* Save parameters as local variables */
-  FASTCOVER_tryParameters_data_t *const data = (FASTCOVER_tryParameters_data_t *)opaque;
+  FASTCOVER_tryParameters_data_t *const data = (FASTCOVER_tryParameters_data_t*)opaque;
   const FASTCOVER_ctx_t *const ctx = data->ctx;
   const ZDICT_cover_params_t parameters = data->parameters;
   size_t dictBufferCapacity = data->dictBufferCapacity;
   size_t totalCompressedSize = ERROR(GENERIC);
   /* Initialize array to keep track of frequency of dmer within activeSegment */
-  U16* segmentFreqs = (U16 *)calloc(((U64)1 << ctx->f), sizeof(U16));
+  U16* segmentFreqs = (U16*)calloc(((U64)1 << ctx->f), sizeof(U16));
   /* Allocate space for hash table, dict, and freqs */
-  BYTE *const dict = (BYTE * const)malloc(dictBufferCapacity);
+  BYTE *const dict = (BYTE*)malloc(dictBufferCapacity);
   COVER_dictSelection_t selection = COVER_dictSelectionError(ERROR(GENERIC));
-  U32 *freqs = (U32*) malloc(((U64)1 << ctx->f) * sizeof(U32));
+  U32* freqs = (U32*) malloc(((U64)1 << ctx->f) * sizeof(U32));
   if (!segmentFreqs || !dict || !freqs) {
     DISPLAYLEVEL(1, "Failed to allocate buffers: out of memory\n");
     goto _cleanup;
@@ -486,7 +495,7 @@ static void FASTCOVER_tryParameters(void *opaque)
                                                     parameters, segmentFreqs);
 
     const unsigned nbFinalizeSamples = (unsigned)(ctx->nbTrainSamples * ctx->accelParams.finalize / 100);
-    selection = COVER_selectDict(dict + tail, dictBufferCapacity - tail,
+    selection = COVER_selectDict(dict + tail, dictBufferCapacity, dictBufferCapacity - tail,
          ctx->samples, ctx->samplesSizes, nbFinalizeSamples, ctx->nbTrainSamples, ctx->nbSamples, parameters, ctx->offsets,
          totalCompressedSize);
 
@@ -547,7 +556,7 @@ ZDICT_trainFromBuffer_fastCover(void* dictBuffer, size_t dictBufferCapacity,
     ZDICT_cover_params_t coverParams;
     FASTCOVER_accel_t accelParams;
     /* Initialize global data */
-    g_displayLevel = parameters.zParams.notificationLevel;
+    g_displayLevel = (int)parameters.zParams.notificationLevel;
     /* Assign splitPoint and f if not provided */
     parameters.splitPoint = 1.0;
     parameters.f = parameters.f == 0 ? DEFAULT_F : parameters.f;
@@ -617,7 +626,7 @@ ZDICT_optimizeTrainFromBuffer_fastCover(
     /* constants */
     const unsigned nbThreads = parameters->nbThreads;
     const double splitPoint =
-        parameters->splitPoint <= 0.0 ? DEFAULT_SPLITPOINT : parameters->splitPoint;
+        parameters->splitPoint <= 0.0 ? FASTCOVER_DEFAULT_SPLITPOINT : parameters->splitPoint;
     const unsigned kMinD = parameters->d == 0 ? 6 : parameters->d;
     const unsigned kMaxD = parameters->d == 0 ? 8 : parameters->d;
     const unsigned kMinK = parameters->k == 0 ? 50 : parameters->k;
@@ -630,7 +639,7 @@ ZDICT_optimizeTrainFromBuffer_fastCover(
     const unsigned accel = parameters->accel == 0 ? DEFAULT_ACCEL : parameters->accel;
     const unsigned shrinkDict = 0;
     /* Local variables */
-    const int displayLevel = parameters->zParams.notificationLevel;
+    const int displayLevel = (int)parameters->zParams.notificationLevel;
     unsigned iteration = 1;
     unsigned d;
     unsigned k;
@@ -714,7 +723,7 @@ ZDICT_optimizeTrainFromBuffer_fastCover(
         data->parameters.splitPoint = splitPoint;
         data->parameters.steps = kSteps;
         data->parameters.shrinkDict = shrinkDict;
-        data->parameters.zParams.notificationLevel = g_displayLevel;
+        data->parameters.zParams.notificationLevel = (unsigned)g_displayLevel;
         /* Check the parameters */
         if (!FASTCOVER_checkParameters(data->parameters, dictBufferCapacity,
                                        data->ctx->f, accel)) {

src/borg/algorithms/zstd/lib/dictBuilder/zdict.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -23,9 +23,13 @@
 /* Unix Large Files support (>4GB) */
 #define _FILE_OFFSET_BITS 64
 #if (defined(__sun__) && (!defined(__LP64__)))   /* Sun Solaris 32-bits requires specific definitions */
+#  ifndef _LARGEFILE_SOURCE
 #  define _LARGEFILE_SOURCE
+#  endif
 #elif ! defined(__LP64__)                        /* No point defining Large file for 64 bit */
+#  ifndef _LARGEFILE64_SOURCE
 #  define _LARGEFILE64_SOURCE
+#  endif
 #endif
 
 
@@ -37,18 +41,19 @@
 #include <stdio.h>         /* fprintf, fopen, ftello64 */
 #include <time.h>          /* clock */
 
+#ifndef ZDICT_STATIC_LINKING_ONLY
+#  define ZDICT_STATIC_LINKING_ONLY
+#endif
+
 #include "../common/mem.h"           /* read */
 #include "../common/fse.h"           /* FSE_normalizeCount, FSE_writeNCount */
-#define HUF_STATIC_LINKING_ONLY
 #include "../common/huf.h"           /* HUF_buildCTable, HUF_writeCTable */
 #include "../common/zstd_internal.h" /* includes zstd.h */
 #include "../common/xxhash.h"        /* XXH64 */
-#include "divsufsort.h"
-#ifndef ZDICT_STATIC_LINKING_ONLY
-#  define ZDICT_STATIC_LINKING_ONLY
-#endif
-#include "zdict.h"
 #include "../compress/zstd_compress_internal.h" /* ZSTD_loadCEntropy() */
+#include "../zdict.h"
+#include "divsufsort.h"
+#include "../common/bits.h"          /* ZSTD_NbCommonBytes */
 
 
 /*-*************************************
@@ -62,14 +67,15 @@
 
 #define NOISELENGTH 32
 
-static const int g_compressionLevel_default = 3;
 static const U32 g_selectivity_default = 9;
 
 
 /*-*************************************
 *  Console display
 ***************************************/
+#undef  DISPLAY
 #define DISPLAY(...)         { fprintf(stderr, __VA_ARGS__); fflush( stderr ); }
+#undef  DISPLAYLEVEL
 #define DISPLAYLEVEL(l, ...) if (notificationLevel>=l) { DISPLAY(__VA_ARGS__); }    /* 0 : no display;   1: errors;   2: default;  3: details;  4: debug */
 
 static clock_t ZDICT_clockSpan(clock_t nPrevious) { return clock() - nPrevious; }
@@ -105,20 +111,17 @@ size_t ZDICT_getDictHeaderSize(const void* dictBuffer, size_t dictSize)
     size_t headerSize;
     if (dictSize <= 8 || MEM_readLE32(dictBuffer) != ZSTD_MAGIC_DICTIONARY) return ERROR(dictionary_corrupted);
 
-    {   unsigned offcodeMaxValue = MaxOff;
-        ZSTD_compressedBlockState_t* bs = (ZSTD_compressedBlockState_t*)malloc(sizeof(ZSTD_compressedBlockState_t));
+    {   ZSTD_compressedBlockState_t* bs = (ZSTD_compressedBlockState_t*)malloc(sizeof(ZSTD_compressedBlockState_t));
         U32* wksp = (U32*)malloc(HUF_WORKSPACE_SIZE);
-        short* offcodeNCount = (short*)malloc((MaxOff+1)*sizeof(short));
-        if (!bs || !wksp || !offcodeNCount) {
+        if (!bs || !wksp) {
             headerSize = ERROR(memory_allocation);
         } else {
             ZSTD_reset_compressedBlockState(bs);
-            headerSize = ZSTD_loadCEntropy(bs, wksp, offcodeNCount, &offcodeMaxValue, dictBuffer, dictSize);
+            headerSize = ZSTD_loadCEntropy(bs, wksp, dictBuffer, dictSize);
         }
 
         free(bs);
         free(wksp);
-        free(offcodeNCount);
     }
 
     return headerSize;
@@ -127,65 +130,6 @@ size_t ZDICT_getDictHeaderSize(const void* dictBuffer, size_t dictSize)
 /*-********************************************************
 *  Dictionary training functions
 **********************************************************/
-static unsigned ZDICT_NbCommonBytes (size_t val)
-{
-    if (MEM_isLittleEndian()) {
-        if (MEM_64bits()) {
-#       if defined(_MSC_VER) && defined(_WIN64)
-            unsigned long r = 0;
-            _BitScanForward64( &r, (U64)val );
-            return (unsigned)(r>>3);
-#       elif defined(__GNUC__) && (__GNUC__ >= 3)
-            return (__builtin_ctzll((U64)val) >> 3);
-#       else
-            static const int DeBruijnBytePos[64] = { 0, 0, 0, 0, 0, 1, 1, 2, 0, 3, 1, 3, 1, 4, 2, 7, 0, 2, 3, 6, 1, 5, 3, 5, 1, 3, 4, 4, 2, 5, 6, 7, 7, 0, 1, 2, 3, 3, 4, 6, 2, 6, 5, 5, 3, 4, 5, 6, 7, 1, 2, 4, 6, 4, 4, 5, 7, 2, 6, 5, 7, 6, 7, 7 };
-            return DeBruijnBytePos[((U64)((val & -(long long)val) * 0x0218A392CDABBD3FULL)) >> 58];
-#       endif
-        } else { /* 32 bits */
-#       if defined(_MSC_VER)
-            unsigned long r=0;
-            _BitScanForward( &r, (U32)val );
-            return (unsigned)(r>>3);
-#       elif defined(__GNUC__) && (__GNUC__ >= 3)
-            return (__builtin_ctz((U32)val) >> 3);
-#       else
-            static const int DeBruijnBytePos[32] = { 0, 0, 3, 0, 3, 1, 3, 0, 3, 2, 2, 1, 3, 2, 0, 1, 3, 3, 1, 2, 2, 2, 2, 0, 3, 1, 2, 0, 1, 0, 1, 1 };
-            return DeBruijnBytePos[((U32)((val & -(S32)val) * 0x077CB531U)) >> 27];
-#       endif
-        }
-    } else {  /* Big Endian CPU */
-        if (MEM_64bits()) {
-#       if defined(_MSC_VER) && defined(_WIN64)
-            unsigned long r = 0;
-            _BitScanReverse64( &r, val );
-            return (unsigned)(r>>3);
-#       elif defined(__GNUC__) && (__GNUC__ >= 3)
-            return (__builtin_clzll(val) >> 3);
-#       else
-            unsigned r;
-            const unsigned n32 = sizeof(size_t)*4;   /* calculate this way due to compiler complaining in 32-bits mode */
-            if (!(val>>n32)) { r=4; } else { r=0; val>>=n32; }
-            if (!(val>>16)) { r+=2; val>>=8; } else { val>>=24; }
-            r += (!val);
-            return r;
-#       endif
-        } else { /* 32 bits */
-#       if defined(_MSC_VER)
-            unsigned long r = 0;
-            _BitScanReverse( &r, (unsigned long)val );
-            return (unsigned)(r>>3);
-#       elif defined(__GNUC__) && (__GNUC__ >= 3)
-            return (__builtin_clz((U32)val) >> 3);
-#       else
-            unsigned r;
-            if (!(val>>16)) { r=2; val>>=8; } else { r=0; val>>=24; }
-            r += (!val);
-            return r;
-#       endif
-    }   }
-}
-
-
 /*! ZDICT_count() :
     Count the nb of common bytes between 2 pointers.
     Note : this function presumes end of buffer followed by noisy guard band.
@@ -200,7 +144,7 @@ static size_t ZDICT_count(const void* pIn, const void* pMatch)
             pMatch = (const char*)pMatch+sizeof(size_t);
             continue;
         }
-        pIn = (const char*)pIn+ZDICT_NbCommonBytes(diff);
+        pIn = (const char*)pIn+ZSTD_NbCommonBytes(diff);
         return (size_t)((const char*)pIn - pStart);
     }
 }
@@ -232,7 +176,7 @@ static dictItem ZDICT_analyzePos(
     U32 savings[LLIMIT] = {0};
     const BYTE* b = (const BYTE*)buffer;
     size_t maxLength = LLIMIT;
-    size_t pos = suffix[start];
+    size_t pos = (size_t)suffix[start];
     U32 end = start;
     dictItem solution;
 
@@ -366,7 +310,7 @@ static dictItem ZDICT_analyzePos(
             savings[i] = savings[i-1] + (lengthList[i] * (i-3));
 
         DISPLAYLEVEL(4, "Selected dict at position %u, of length %u : saves %u (ratio: %.2f)  \n",
-                     (unsigned)pos, (unsigned)maxLength, (unsigned)savings[maxLength], (double)savings[maxLength] / maxLength);
+                     (unsigned)pos, (unsigned)maxLength, (unsigned)savings[maxLength], (double)savings[maxLength] / (double)maxLength);
 
         solution.pos = (U32)pos;
         solution.length = (U32)maxLength;
@@ -376,7 +320,7 @@ static dictItem ZDICT_analyzePos(
         {   U32 id;
             for (id=start; id<end; id++) {
                 U32 p, pEnd, length;
-                U32 const testedPos = suffix[id];
+                U32 const testedPos = (U32)suffix[id];
                 if (testedPos == pos)
                     length = solution.length;
                 else {
@@ -428,7 +372,7 @@ static U32 ZDICT_tryMerge(dictItem* table, dictItem elt, U32 eltNbToSkip, const
             elt = table[u];
             /* sort : improve rank */
             while ((u>1) && (table[u-1].savings < elt.savings))
-            table[u] = table[u-1], u--;
+                table[u] = table[u-1], u--;
             table[u] = elt;
             return u;
     }   }
@@ -439,7 +383,7 @@ static U32 ZDICT_tryMerge(dictItem* table, dictItem elt, U32 eltNbToSkip, const
 
         if ((table[u].pos + table[u].length >= elt.pos) && (table[u].pos < elt.pos)) {  /* overlap, existing < new */
             /* append */
-            int const addedLength = (int)eltEnd - (table[u].pos + table[u].length);
+            int const addedLength = (int)eltEnd - (int)(table[u].pos + table[u].length);
             table[u].savings += elt.length / 8;    /* rough approx bonus */
             if (addedLength > 0) {   /* otherwise, elt fully included into existing */
                 table[u].length += addedLength;
@@ -532,6 +476,7 @@ static size_t ZDICT_trainBuffer_legacy(dictItem* dictList, U32 dictListSize,
     clock_t displayClock = 0;
     clock_t const refreshRate = CLOCKS_PER_SEC * 3 / 10;
 
+#   undef  DISPLAYUPDATE
 #   define DISPLAYUPDATE(l, ...) if (notificationLevel>=l) { \
             if (ZDICT_clockSpan(displayClock) > refreshRate)  \
             { displayClock = clock(); DISPLAY(__VA_ARGS__); \
@@ -578,7 +523,7 @@ static size_t ZDICT_trainBuffer_legacy(dictItem* dictList, U32 dictListSize,
             if (solution.length==0) { cursor++; continue; }
             ZDICT_insertDictItem(dictList, dictListSize, solution, buffer);
             cursor += solution.length;
-            DISPLAYUPDATE(2, "\r%4.2f %% \r", (double)cursor / bufferSize * 100);
+            DISPLAYUPDATE(2, "\r%4.2f %% \r", (double)cursor / (double)bufferSize * 100.0);
     }   }
 
 _cleanup:
@@ -621,11 +566,11 @@ static void ZDICT_countEStats(EStats_ress_t esr, const ZSTD_parameters* params,
     size_t cSize;
 
     if (srcSize > blockSizeMax) srcSize = blockSizeMax;   /* protection vs large samples */
-    {   size_t const errorCode = ZSTD_compressBegin_usingCDict(esr.zc, esr.dict);
+    {   size_t const errorCode = ZSTD_compressBegin_usingCDict_deprecated(esr.zc, esr.dict);
         if (ZSTD_isError(errorCode)) { DISPLAYLEVEL(1, "warning : ZSTD_compressBegin_usingCDict failed \n"); return; }
 
     }
-    cSize = ZSTD_compressBlock(esr.zc, esr.workPlace, ZSTD_BLOCKSIZE_MAX, src, srcSize);
+    cSize = ZSTD_compressBlock_deprecated(esr.zc, esr.workPlace, ZSTD_BLOCKSIZE_MAX, src, srcSize);
     if (ZSTD_isError(cSize)) { DISPLAYLEVEL(3, "warning : could not compress sample size %u \n", (unsigned)srcSize); return; }
 
     if (cSize) {  /* if == 0; block is not compressible */
@@ -658,8 +603,8 @@ static void ZDICT_countEStats(EStats_ress_t esr, const ZSTD_parameters* params,
 
             if (nbSeq >= 2) { /* rep offsets */
                 const seqDef* const seq = seqStorePtr->sequencesStart;
-                U32 offset1 = seq[0].offset - 3;
-                U32 offset2 = seq[1].offset - 3;
+                U32 offset1 = seq[0].offBase - ZSTD_REP_NUM;
+                U32 offset2 = seq[1].offBase - ZSTD_REP_NUM;
                 if (offset1 >= MAXREPOFFSET) offset1 = 0;
                 if (offset2 >= MAXREPOFFSET) offset2 = 0;
                 repOffsets[offset1] += 3;
@@ -706,7 +651,7 @@ static void ZDICT_flatLit(unsigned* countLit)
 
 #define OFFCODE_MAX 30  /* only applicable to first block */
 static size_t ZDICT_analyzeEntropy(void*  dstBuffer, size_t maxDstSize,
-                                   unsigned compressionLevel,
+                                   int compressionLevel,
                              const void*  srcBuffer, const size_t* fileSizes, unsigned nbFiles,
                              const void* dictBuffer, size_t  dictBufferSize,
                                    unsigned notificationLevel)
@@ -730,6 +675,7 @@ static size_t ZDICT_analyzeEntropy(void*  dstBuffer, size_t maxDstSize,
     size_t const totalSrcSize = ZDICT_totalSampleSize(fileSizes, nbFiles);
     size_t const averageSampleSize = totalSrcSize / (nbFiles + !nbFiles);
     BYTE* dstPtr = (BYTE*)dstBuffer;
+    U32 wksp[HUF_CTABLE_WORKSPACE_SIZE_U32];
 
     /* init */
     DEBUGLOG(4, "ZDICT_analyzeEntropy");
@@ -741,7 +687,7 @@ static size_t ZDICT_analyzeEntropy(void*  dstBuffer, size_t maxDstSize,
     memset(repOffset, 0, sizeof(repOffset));
     repOffset[1] = repOffset[4] = repOffset[8] = 1;
     memset(bestRepOffset, 0, sizeof(bestRepOffset));
-    if (compressionLevel==0) compressionLevel = g_compressionLevel_default;
+    if (compressionLevel==0) compressionLevel = ZSTD_CLEVEL_DEFAULT;
     params = ZSTD_getParams(compressionLevel, averageSampleSize, dictBufferSize);
 
     esr.dict = ZSTD_createCDict_advanced(dictBuffer, dictBufferSize, ZSTD_dlm_byRef, ZSTD_dct_rawContent, params.cParams, ZSTD_defaultCMem);
@@ -762,8 +708,15 @@ static size_t ZDICT_analyzeEntropy(void*  dstBuffer, size_t maxDstSize,
         pos += fileSizes[u];
     }
 
+    if (notificationLevel >= 4) {
+        /* writeStats */
+        DISPLAYLEVEL(4, "Offset Code Frequencies : \n");
+        for (u=0; u<=offcodeMax; u++) {
+            DISPLAYLEVEL(4, "%2u :%7u \n", u, offcodeCount[u]);
+    }   }
+
     /* analyze, build stats, starting with literals */
-    {   size_t maxNbBits = HUF_buildCTable (hufTable, countLit, 255, huffLog);
+    {   size_t maxNbBits = HUF_buildCTable_wksp(hufTable, countLit, 255, huffLog, wksp, sizeof(wksp));
         if (HUF_isError(maxNbBits)) {
             eSize = maxNbBits;
             DISPLAYLEVEL(1, " HUF_buildCTable error \n");
@@ -772,7 +725,7 @@ static size_t ZDICT_analyzeEntropy(void*  dstBuffer, size_t maxDstSize,
         if (maxNbBits==8) {  /* not compressible : will fail on HUF_writeCTable() */
             DISPLAYLEVEL(2, "warning : pathological dataset : literals are not compressible : samples are noisy or too regular \n");
             ZDICT_flatLit(countLit);  /* replace distribution by a fake "mostly flat but still compressible" distribution, that HUF_writeCTable() can encode */
-            maxNbBits = HUF_buildCTable (hufTable, countLit, 255, huffLog);
+            maxNbBits = HUF_buildCTable_wksp(hufTable, countLit, 255, huffLog, wksp, sizeof(wksp));
             assert(maxNbBits==9);
         }
         huffLog = (U32)maxNbBits;
@@ -786,7 +739,7 @@ static size_t ZDICT_analyzeEntropy(void*  dstBuffer, size_t maxDstSize,
     /* note : the result of this phase should be used to better appreciate the impact on statistics */
 
     total=0; for (u=0; u<=offcodeMax; u++) total+=offcodeCount[u];
-    errorCode = FSE_normalizeCount(offcodeNCount, Offlog, offcodeCount, total, offcodeMax);
+    errorCode = FSE_normalizeCount(offcodeNCount, Offlog, offcodeCount, total, offcodeMax, /* useLowProbCount */ 1);
     if (FSE_isError(errorCode)) {
         eSize = errorCode;
         DISPLAYLEVEL(1, "FSE_normalizeCount error with offcodeCount \n");
@@ -795,7 +748,7 @@ static size_t ZDICT_analyzeEntropy(void*  dstBuffer, size_t maxDstSize,
     Offlog = (U32)errorCode;
 
     total=0; for (u=0; u<=MaxML; u++) total+=matchLengthCount[u];
-    errorCode = FSE_normalizeCount(matchLengthNCount, mlLog, matchLengthCount, total, MaxML);
+    errorCode = FSE_normalizeCount(matchLengthNCount, mlLog, matchLengthCount, total, MaxML, /* useLowProbCount */ 1);
     if (FSE_isError(errorCode)) {
         eSize = errorCode;
         DISPLAYLEVEL(1, "FSE_normalizeCount error with matchLengthCount \n");
@@ -804,7 +757,7 @@ static size_t ZDICT_analyzeEntropy(void*  dstBuffer, size_t maxDstSize,
     mlLog = (U32)errorCode;
 
     total=0; for (u=0; u<=MaxLL; u++) total+=litLengthCount[u];
-    errorCode = FSE_normalizeCount(litLengthNCount, llLog, litLengthCount, total, MaxLL);
+    errorCode = FSE_normalizeCount(litLengthNCount, llLog, litLengthCount, total, MaxLL, /* useLowProbCount */ 1);
     if (FSE_isError(errorCode)) {
         eSize = errorCode;
         DISPLAYLEVEL(1, "FSE_normalizeCount error with litLengthCount \n");
@@ -813,7 +766,7 @@ static size_t ZDICT_analyzeEntropy(void*  dstBuffer, size_t maxDstSize,
     llLog = (U32)errorCode;
 
     /* write result to buffer */
-    {   size_t const hhSize = HUF_writeCTable(dstPtr, maxDstSize, hufTable, 255, huffLog);
+    {   size_t const hhSize = HUF_writeCTable_wksp(dstPtr, maxDstSize, hufTable, 255, huffLog, wksp, sizeof(wksp));
         if (HUF_isError(hhSize)) {
             eSize = hhSize;
             DISPLAYLEVEL(1, "HUF_writeCTable error \n");
@@ -868,7 +821,7 @@ static size_t ZDICT_analyzeEntropy(void*  dstBuffer, size_t maxDstSize,
     MEM_writeLE32(dstPtr+8, bestRepOffset[2].offset);
 #else
     /* at this stage, we don't use the result of "most common first offset",
-       as the impact of statistics is not properly evaluated */
+     * as the impact of statistics is not properly evaluated */
     MEM_writeLE32(dstPtr+0, repStartValue[0]);
     MEM_writeLE32(dstPtr+4, repStartValue[1]);
     MEM_writeLE32(dstPtr+8, repStartValue[2]);
@@ -884,6 +837,17 @@ _cleanup:
 }
 
 
+/**
+ * @returns the maximum repcode value
+ */
+static U32 ZDICT_maxRep(U32 const reps[ZSTD_REP_NUM])
+{
+    U32 maxRep = reps[0];
+    int r;
+    for (r = 1; r < ZSTD_REP_NUM; ++r)
+        maxRep = MAX(maxRep, reps[r]);
+    return maxRep;
+}
 
 size_t ZDICT_finalizeDictionary(void* dictBuffer, size_t dictBufferCapacity,
                           const void* customDictContent, size_t dictContentSize,
@@ -893,13 +857,15 @@ size_t ZDICT_finalizeDictionary(void* dictBuffer, size_t dictBufferCapacity,
     size_t hSize;
 #define HBUFFSIZE 256   /* should prove large enough for all entropy headers */
     BYTE header[HBUFFSIZE];
-    int const compressionLevel = (params.compressionLevel == 0) ? g_compressionLevel_default : params.compressionLevel;
+    int const compressionLevel = (params.compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT : params.compressionLevel;
     U32 const notificationLevel = params.notificationLevel;
+    /* The final dictionary content must be at least as large as the largest repcode */
+    size_t const minContentSize = (size_t)ZDICT_maxRep(repStartValue);
+    size_t paddingSize;
 
     /* check conditions */
     DEBUGLOG(4, "ZDICT_finalizeDictionary");
     if (dictBufferCapacity < dictContentSize) return ERROR(dstSize_tooSmall);
-    if (dictContentSize < ZDICT_CONTENTSIZE_MIN) return ERROR(srcSize_wrong);
     if (dictBufferCapacity < ZDICT_DICTSIZE_MIN) return ERROR(dstSize_tooSmall);
 
     /* dictionary header */
@@ -923,12 +889,43 @@ size_t ZDICT_finalizeDictionary(void* dictBuffer, size_t dictBufferCapacity,
         hSize += eSize;
     }
 
-    /* copy elements in final buffer ; note : src and dst buffer can overlap */
-    if (hSize + dictContentSize > dictBufferCapacity) dictContentSize = dictBufferCapacity - hSize;
-    {   size_t const dictSize = hSize + dictContentSize;
-        char* dictEnd = (char*)dictBuffer + dictSize;
-        memmove(dictEnd - dictContentSize, customDictContent, dictContentSize);
-        memcpy(dictBuffer, header, hSize);
+    /* Shrink the content size if it doesn't fit in the buffer */
+    if (hSize + dictContentSize > dictBufferCapacity) {
+        dictContentSize = dictBufferCapacity - hSize;
+    }
+
+    /* Pad the dictionary content with zeros if it is too small */
+    if (dictContentSize < minContentSize) {
+        RETURN_ERROR_IF(hSize + minContentSize > dictBufferCapacity, dstSize_tooSmall,
+                        "dictBufferCapacity too small to fit max repcode");
+        paddingSize = minContentSize - dictContentSize;
+    } else {
+        paddingSize = 0;
+    }
+
+    {
+        size_t const dictSize = hSize + paddingSize + dictContentSize;
+
+        /* The dictionary consists of the header, optional padding, and the content.
+         * The padding comes before the content because the "best" position in the
+         * dictionary is the last byte.
+         */
+        BYTE* const outDictHeader = (BYTE*)dictBuffer;
+        BYTE* const outDictPadding = outDictHeader + hSize;
+        BYTE* const outDictContent = outDictPadding + paddingSize;
+
+        assert(dictSize <= dictBufferCapacity);
+        assert(outDictContent + dictContentSize == (BYTE*)dictBuffer + dictSize);
+
+        /* First copy the customDictContent into its final location.
+         * `customDictContent` and `dictBuffer` may overlap, so we must
+         * do this before any other writes into the output buffer.
+         * Then copy the header & padding into the output buffer.
+         */
+        memmove(outDictContent, customDictContent, dictContentSize);
+        memcpy(outDictHeader, header, hSize);
+        memset(outDictPadding, 0, paddingSize);
+
         return dictSize;
     }
 }
@@ -939,7 +936,7 @@ static size_t ZDICT_addEntropyTablesFromBuffer_advanced(
         const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
         ZDICT_params_t params)
 {
-    int const compressionLevel = (params.compressionLevel == 0) ? g_compressionLevel_default : params.compressionLevel;
+    int const compressionLevel = (params.compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT : params.compressionLevel;
     U32 const notificationLevel = params.notificationLevel;
     size_t hSize = 8;
 
@@ -968,16 +965,11 @@ static size_t ZDICT_addEntropyTablesFromBuffer_advanced(
     return MIN(dictBufferCapacity, hSize+dictContentSize);
 }
 
-/* Hidden declaration for dbio.c */
-size_t ZDICT_trainFromBuffer_unsafe_legacy(
-                            void* dictBuffer, size_t maxDictSize,
-                            const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
-                            ZDICT_legacy_params_t params);
 /*! ZDICT_trainFromBuffer_unsafe_legacy() :
-*   Warning : `samplesBuffer` must be followed by noisy guard band.
+*   Warning : `samplesBuffer` must be followed by noisy guard band !!!
 *   @return : size of dictionary, or an error code which can be tested with ZDICT_isError()
 */
-size_t ZDICT_trainFromBuffer_unsafe_legacy(
+static size_t ZDICT_trainFromBuffer_unsafe_legacy(
                             void* dictBuffer, size_t maxDictSize,
                             const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
                             ZDICT_legacy_params_t params)
@@ -1114,8 +1106,8 @@ size_t ZDICT_trainFromBuffer(void* dictBuffer, size_t dictBufferCapacity,
     memset(&params, 0, sizeof(params));
     params.d = 8;
     params.steps = 4;
-    /* Default to level 6 since no compression level information is available */
-    params.zParams.compressionLevel = 3;
+    /* Use default level since no compression level information is available */
+    params.zParams.compressionLevel = ZSTD_CLEVEL_DEFAULT;
 #if defined(DEBUGLEVEL) && (DEBUGLEVEL>=1)
     params.zParams.notificationLevel = DEBUGLEVEL;
 #endif

src/borg/algorithms/zstd/lib/legacy/zstd_legacy.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -242,6 +242,13 @@ MEM_STATIC ZSTD_frameSizeInfo ZSTD_findFrameSizeInfoLegacy(const void *src, size
         frameSizeInfo.compressedSize = ERROR(srcSize_wrong);
         frameSizeInfo.decompressedBound = ZSTD_CONTENTSIZE_ERROR;
     }
+    /* In all cases, decompressedBound == nbBlocks * ZSTD_BLOCKSIZE_MAX.
+     * So we can compute nbBlocks without having to change every function.
+     */
+    if (frameSizeInfo.decompressedBound != ZSTD_CONTENTSIZE_ERROR) {
+        assert((frameSizeInfo.decompressedBound & (ZSTD_BLOCKSIZE_MAX - 1)) == 0);
+        frameSizeInfo.nbBlocks = (size_t)(frameSizeInfo.decompressedBound / ZSTD_BLOCKSIZE_MAX);
+    }
     return frameSizeInfo;
 }
 

src/borg/algorithms/zstd/lib/legacy/zstd_v01.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -190,28 +190,6 @@ typedef   signed long long  S64;
 /****************************************************************
 *  Memory I/O
 *****************************************************************/
-/* FSE_FORCE_MEMORY_ACCESS
- * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
- * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
- * The below switch allow to select different access method for improved performance.
- * Method 0 (default) : use `memcpy()`. Safe and portable.
- * Method 1 : `__packed` statement. It depends on compiler extension (ie, not portable).
- *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
- * Method 2 : direct access. This method is portable but violate C standard.
- *            It can generate buggy code on targets generating assembly depending on alignment.
- *            But in some circumstances, it's the only known way to get the most performance (ie GCC + ARMv6)
- * See http://fastcompression.blogspot.fr/2015/08/accessing-unaligned-memory.html for details.
- * Prefer these methods in priority order (0 > 1 > 2)
- */
-#ifndef FSE_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
-#  if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
-#    define FSE_FORCE_MEMORY_ACCESS 2
-#  elif (defined(__INTEL_COMPILER) && !defined(WIN32)) || \
-  (defined(__GNUC__) && ( defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) || defined(__ARM_ARCH_7S__) ))
-#    define FSE_FORCE_MEMORY_ACCESS 1
-#  endif
-#endif
-
 
 static unsigned FSE_32bits(void)
 {
@@ -224,24 +202,6 @@ static unsigned FSE_isLittleEndian(void)
     return one.c[0];
 }
 
-#if defined(FSE_FORCE_MEMORY_ACCESS) && (FSE_FORCE_MEMORY_ACCESS==2)
-
-static U16 FSE_read16(const void* memPtr) { return *(const U16*) memPtr; }
-static U32 FSE_read32(const void* memPtr) { return *(const U32*) memPtr; }
-static U64 FSE_read64(const void* memPtr) { return *(const U64*) memPtr; }
-
-#elif defined(FSE_FORCE_MEMORY_ACCESS) && (FSE_FORCE_MEMORY_ACCESS==1)
-
-/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
-/* currently only defined for gcc and icc */
-typedef union { U16 u16; U32 u32; U64 u64; } __attribute__((packed)) unalign;
-
-static U16 FSE_read16(const void* ptr) { return ((const unalign*)ptr)->u16; }
-static U32 FSE_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
-static U64 FSE_read64(const void* ptr) { return ((const unalign*)ptr)->u64; }
-
-#else
-
 static U16 FSE_read16(const void* memPtr)
 {
     U16 val; memcpy(&val, memPtr, sizeof(val)); return val;
@@ -257,8 +217,6 @@ static U64 FSE_read64(const void* memPtr)
     U64 val; memcpy(&val, memPtr, sizeof(val)); return val;
 }
 
-#endif /* FSE_FORCE_MEMORY_ACCESS */
-
 static U16 FSE_readLE16(const void* memPtr)
 {
     if (FSE_isLittleEndian())
@@ -343,8 +301,7 @@ FORCE_INLINE unsigned FSE_highbit32 (U32 val)
 {
 #   if defined(_MSC_VER)   /* Visual */
     unsigned long r;
-    _BitScanReverse ( &r, val );
-    return (unsigned) r;
+    return _BitScanReverse(&r, val) ? (unsigned)r : 0;
 #   elif defined(__GNUC__) && (GCC_VERSION >= 304)   /* GCC Intrinsic */
     return __builtin_clz (val) ^ 31;
 #   else   /* Software version */
@@ -1194,7 +1151,7 @@ static size_t HUF_decompress (void* dst, size_t maxDstSize, const void* cSrc, si
     zstd - standard compression library
     Copyright (C) 2014-2015, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are
@@ -1280,7 +1237,11 @@ static size_t HUF_decompress (void* dst, size_t maxDstSize, const void* cSrc, si
 *  Basic Types
 *********************************************************/
 #if defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* C99 */
-# include <stdint.h>
+# if defined(_AIX)
+#  include <inttypes.h>
+# else
+#  include <stdint.h> /* intptr_t */
+# endif
 typedef  uint8_t BYTE;
 typedef uint16_t U16;
 typedef  int16_t S16;
@@ -1759,20 +1720,26 @@ static size_t ZSTD_execSequence(BYTE* op,
     static const int dec32table[] = {0, 1, 2, 1, 4, 4, 4, 4};   /* added */
     static const int dec64table[] = {8, 8, 8, 7, 8, 9,10,11};   /* subtracted */
     const BYTE* const ostart = op;
+    BYTE* const oLitEnd = op + sequence.litLength;
     const size_t litLength = sequence.litLength;
     BYTE* const endMatch = op + litLength + sequence.matchLength;    /* risk : address space overflow (32-bits) */
     const BYTE* const litEnd = *litPtr + litLength;
 
-    /* check */
+    /* checks */
+    size_t const seqLength = sequence.litLength + sequence.matchLength;
+
+    if (seqLength > (size_t)(oend - op)) return ERROR(dstSize_tooSmall);
+    if (sequence.litLength > (size_t)(litLimit - *litPtr)) return ERROR(corruption_detected);
+    /* Now we know there are no overflow in literal nor match lengths, can use pointer checks */
+    if (sequence.offset > (U32)(oLitEnd - base)) return ERROR(corruption_detected);
+
     if (endMatch > oend) return ERROR(dstSize_tooSmall);   /* overwrite beyond dst buffer */
-    if (litEnd > litLimit) return ERROR(corruption_detected);
-    if (sequence.matchLength > (size_t)(*litPtr-op))  return ERROR(dstSize_tooSmall);    /* overwrite literal segment */
+    if (litEnd > litLimit) return ERROR(corruption_detected);   /* overRead beyond lit buffer */
+    if (sequence.matchLength > (size_t)(*litPtr-op)) return ERROR(dstSize_tooSmall);  /* overwrite literal segment */
 
     /* copy Literals */
-    if (((size_t)(*litPtr - op) < 8) || ((size_t)(oend-litEnd) < 8) || (op+litLength > oend-8))
-        memmove(op, *litPtr, litLength);   /* overwrite risk */
-    else
-        ZSTD_wildcopy(op, *litPtr, litLength);
+    ZSTD_memmove(op, *litPtr, sequence.litLength);   /* note : v0.1 seems to allow scenarios where output or input are close to end of buffer */
+
     op += litLength;
     *litPtr = litEnd;   /* update for next sequence */
 

src/borg/algorithms/zstd/lib/legacy/zstd_v01.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the

src/borg/algorithms/zstd/lib/legacy/zstd_v02.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -28,7 +28,7 @@
    low-level memory access routines
    Copyright (C) 2013-2015, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -89,7 +89,11 @@ extern "C" {
 *  Basic Types
 *****************************************************************/
 #if defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
-# include <stdint.h>
+# if defined(_AIX)
+#  include <inttypes.h>
+# else
+#  include <stdint.h> /* intptr_t */
+# endif
   typedef  uint8_t BYTE;
   typedef uint16_t U16;
   typedef  int16_t S16;
@@ -111,27 +115,6 @@ extern "C" {
 /****************************************************************
 *  Memory I/O
 *****************************************************************/
-/* MEM_FORCE_MEMORY_ACCESS
- * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
- * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
- * The below switch allow to select different access method for improved performance.
- * Method 0 (default) : use `memcpy()`. Safe and portable.
- * Method 1 : `__packed` statement. It depends on compiler extension (ie, not portable).
- *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
- * Method 2 : direct access. This method is portable but violate C standard.
- *            It can generate buggy code on targets generating assembly depending on alignment.
- *            But in some circumstances, it's the only known way to get the most performance (ie GCC + ARMv6)
- * See http://fastcompression.blogspot.fr/2015/08/accessing-unaligned-memory.html for details.
- * Prefer these methods in priority order (0 > 1 > 2)
- */
-#ifndef MEM_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
-#  if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
-#    define MEM_FORCE_MEMORY_ACCESS 2
-#  elif (defined(__INTEL_COMPILER) && !defined(WIN32)) || \
-  (defined(__GNUC__) && ( defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) || defined(__ARM_ARCH_7S__) ))
-#    define MEM_FORCE_MEMORY_ACCESS 1
-#  endif
-#endif
 
 MEM_STATIC unsigned MEM_32bits(void) { return sizeof(void*)==4; }
 MEM_STATIC unsigned MEM_64bits(void) { return sizeof(void*)==8; }
@@ -142,33 +125,6 @@ MEM_STATIC unsigned MEM_isLittleEndian(void)
     return one.c[0];
 }
 
-#if defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==2)
-
-/* violates C standard 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 void MEM_write16(void* memPtr, U16 value) { *(U16*)memPtr = value; }
-
-#elif defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==1)
-
-/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
-/* currently only defined for gcc and icc */
-typedef union { U16 u16; U32 u32; U64 u64; } __attribute__((packed)) unalign;
-
-MEM_STATIC U16 MEM_read16(const void* ptr) { return ((const unalign*)ptr)->u16; }
-MEM_STATIC U32 MEM_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
-MEM_STATIC U64 MEM_read64(const void* ptr) { return ((const unalign*)ptr)->u64; }
-
-MEM_STATIC void MEM_write16(void* memPtr, U16 value) { ((unalign*)memPtr)->u16 = value; }
-
-#else
-
-/* default method, safe and standard.
-   can sometimes prove slower */
-
 MEM_STATIC U16 MEM_read16(const void* memPtr)
 {
     U16 val; memcpy(&val, memPtr, sizeof(val)); return val;
@@ -189,9 +145,6 @@ MEM_STATIC void MEM_write16(void* memPtr, U16 value)
     memcpy(memPtr, &value, sizeof(value));
 }
 
-#endif /* MEM_FORCE_MEMORY_ACCESS */
-
-
 MEM_STATIC U16 MEM_readLE16(const void* memPtr)
 {
     if (MEM_isLittleEndian())
@@ -268,7 +221,7 @@ MEM_STATIC size_t MEM_readLEST(const void* memPtr)
    header file (to include)
    Copyright (C) 2013-2015, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -349,9 +302,8 @@ MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, unsigned nbBits);
 MEM_STATIC unsigned BIT_highbit32 (U32 val)
 {
 #   if defined(_MSC_VER)   /* Visual */
-    unsigned long r=0;
-    _BitScanReverse ( &r, val );
-    return (unsigned) r;
+    unsigned long r;
+    return _BitScanReverse(&r, val) ? (unsigned)r : 0;
 #   elif defined(__GNUC__) && (__GNUC__ >= 3)   /* Use GCC Intrinsic */
     return __builtin_clz (val) ^ 31;
 #   else   /* Software version */
@@ -433,7 +385,7 @@ MEM_STATIC size_t BIT_lookBits(BIT_DStream_t* bitD, U32 nbBits)
 }
 
 /*! BIT_lookBitsFast :
-*   unsafe version; only works only if nbBits >= 1 */
+*   unsafe version; only works if nbBits >= 1 */
 MEM_STATIC size_t BIT_lookBitsFast(BIT_DStream_t* bitD, U32 nbBits)
 {
     const U32 bitMask = sizeof(bitD->bitContainer)*8 - 1;
@@ -453,7 +405,7 @@ MEM_STATIC size_t BIT_readBits(BIT_DStream_t* bitD, U32 nbBits)
 }
 
 /*!BIT_readBitsFast :
-*  unsafe version; only works only if nbBits >= 1 */
+*  unsafe version; only works if nbBits >= 1 */
 MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, U32 nbBits)
 {
     size_t value = BIT_lookBitsFast(bitD, nbBits);
@@ -510,7 +462,7 @@ MEM_STATIC unsigned BIT_endOfDStream(const BIT_DStream_t* DStream)
    Error codes and messages
    Copyright (C) 2013-2015, Yann Collet
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -609,7 +561,7 @@ typedef unsigned FSE_DTable;   /* don't allocate that. It's just a way to be mor
    header file for static linking (only)
    Copyright (C) 2013-2015, Yann Collet
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -753,7 +705,7 @@ MEM_STATIC unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr)
    header file for static linking (only)
    Copyright (C) 2013-2015, Yann Collet
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -822,7 +774,7 @@ static size_t HUF_decompress4X6 (void* dst, size_t dstSize, const void* cSrc, si
     Header File
     Copyright (C) 2014-2015, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are
@@ -882,7 +834,7 @@ typedef struct ZSTD_CCtx_s ZSTD_CCtx;   /* incomplete type */
     Header File for static linking only
     Copyright (C) 2014-2015, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are
@@ -946,7 +898,7 @@ typedef struct ZSTD_DCtx_s ZSTD_DCtx;
    FSE : Finite State Entropy coder
    Copyright (C) 2013-2015, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -1450,7 +1402,7 @@ static size_t FSE_decompress(void* dst, size_t maxDstSize, const void* cSrc, siz
    Huff0 : Huffman coder, part of New Generation Entropy library
    Copyright (C) 2013-2015, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -2609,7 +2561,7 @@ static size_t HUF_decompress (void* dst, size_t dstSize, const void* cSrc, size_
     zstd - standard compression library
     Copyright (C) 2014-2015, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are
@@ -3114,12 +3066,19 @@ static size_t ZSTD_execSequence(BYTE* op,
     const BYTE* const litEnd = *litPtr + sequence.litLength;
 
     /* checks */
-    if (oLitEnd > oend_8) return ERROR(dstSize_tooSmall);   /* last match must start at a minimum distance of 8 from oend */
+    size_t const seqLength = sequence.litLength + sequence.matchLength;
+
+    if (seqLength > (size_t)(oend - op)) return ERROR(dstSize_tooSmall);
+    if (sequence.litLength > (size_t)(litLimit - *litPtr)) return ERROR(corruption_detected);
+    /* Now we know there are no overflow in literal nor match lengths, can use the pointer check */
+    if (oLitEnd > oend_8) return ERROR(dstSize_tooSmall);
+    if (sequence.offset > (U32)(oLitEnd - base)) return ERROR(corruption_detected);
+
     if (oMatchEnd > oend) return ERROR(dstSize_tooSmall);   /* overwrite beyond dst buffer */
     if (litEnd > litLimit) return ERROR(corruption_detected);   /* overRead beyond lit buffer */
 
     /* copy Literals */
-    ZSTD_wildcopy(op, *litPtr, sequence.litLength);   /* note : oLitEnd <= oend-8 : no risk of overwrite beyond oend */
+    ZSTD_wildcopy(op, *litPtr, (ptrdiff_t)sequence.litLength);   /* note : oLitEnd <= oend-8 : no risk of overwrite beyond oend */
     op = oLitEnd;
     *litPtr = litEnd;   /* update for next sequence */
 

src/borg/algorithms/zstd/lib/legacy/zstd_v02.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the

src/borg/algorithms/zstd/lib/legacy/zstd_v03.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -29,7 +29,7 @@
    low-level memory access routines
    Copyright (C) 2013-2015, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -90,7 +90,11 @@ extern "C" {
 *  Basic Types
 *****************************************************************/
 #if defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
-# include <stdint.h>
+# if defined(_AIX)
+#  include <inttypes.h>
+# else
+#  include <stdint.h> /* intptr_t */
+# endif
   typedef  uint8_t BYTE;
   typedef uint16_t U16;
   typedef  int16_t S16;
@@ -112,27 +116,6 @@ extern "C" {
 /****************************************************************
 *  Memory I/O
 *****************************************************************/
-/* MEM_FORCE_MEMORY_ACCESS
- * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
- * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
- * The below switch allow to select different access method for improved performance.
- * Method 0 (default) : use `memcpy()`. Safe and portable.
- * Method 1 : `__packed` statement. It depends on compiler extension (ie, not portable).
- *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
- * Method 2 : direct access. This method is portable but violate C standard.
- *            It can generate buggy code on targets generating assembly depending on alignment.
- *            But in some circumstances, it's the only known way to get the most performance (ie GCC + ARMv6)
- * See http://fastcompression.blogspot.fr/2015/08/accessing-unaligned-memory.html for details.
- * Prefer these methods in priority order (0 > 1 > 2)
- */
-#ifndef MEM_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
-#  if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
-#    define MEM_FORCE_MEMORY_ACCESS 2
-#  elif (defined(__INTEL_COMPILER) && !defined(WIN32)) || \
-  (defined(__GNUC__) && ( defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) || defined(__ARM_ARCH_7S__) ))
-#    define MEM_FORCE_MEMORY_ACCESS 1
-#  endif
-#endif
 
 MEM_STATIC unsigned MEM_32bits(void) { return sizeof(void*)==4; }
 MEM_STATIC unsigned MEM_64bits(void) { return sizeof(void*)==8; }
@@ -143,33 +126,6 @@ MEM_STATIC unsigned MEM_isLittleEndian(void)
     return one.c[0];
 }
 
-#if defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==2)
-
-/* violates C standard 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 void MEM_write16(void* memPtr, U16 value) { *(U16*)memPtr = value; }
-
-#elif defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==1)
-
-/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
-/* currently only defined for gcc and icc */
-typedef union { U16 u16; U32 u32; U64 u64; } __attribute__((packed)) unalign;
-
-MEM_STATIC U16 MEM_read16(const void* ptr) { return ((const unalign*)ptr)->u16; }
-MEM_STATIC U32 MEM_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
-MEM_STATIC U64 MEM_read64(const void* ptr) { return ((const unalign*)ptr)->u64; }
-
-MEM_STATIC void MEM_write16(void* memPtr, U16 value) { ((unalign*)memPtr)->u16 = value; }
-
-#else
-
-/* default method, safe and standard.
-   can sometimes prove slower */
-
 MEM_STATIC U16 MEM_read16(const void* memPtr)
 {
     U16 val; memcpy(&val, memPtr, sizeof(val)); return val;
@@ -190,10 +146,6 @@ MEM_STATIC void MEM_write16(void* memPtr, U16 value)
     memcpy(memPtr, &value, sizeof(value));
 }
 
-
-#endif /* MEM_FORCE_MEMORY_ACCESS */
-
-
 MEM_STATIC U16 MEM_readLE16(const void* memPtr)
 {
     if (MEM_isLittleEndian())
@@ -270,7 +222,7 @@ MEM_STATIC size_t MEM_readLEST(const void* memPtr)
    header file (to include)
    Copyright (C) 2013-2015, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -352,9 +304,8 @@ MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, unsigned nbBits);
 MEM_STATIC unsigned BIT_highbit32 (U32 val)
 {
 #   if defined(_MSC_VER)   /* Visual */
-    unsigned long r=0;
-    _BitScanReverse ( &r, val );
-    return (unsigned) r;
+    unsigned long r;
+    return _BitScanReverse(&r, val) ? (unsigned)r : 0;
 #   elif defined(__GNUC__) && (__GNUC__ >= 3)   /* Use GCC Intrinsic */
     return __builtin_clz (val) ^ 31;
 #   else   /* Software version */
@@ -435,7 +386,7 @@ MEM_STATIC size_t BIT_lookBits(BIT_DStream_t* bitD, U32 nbBits)
 }
 
 /*! BIT_lookBitsFast :
-*   unsafe version; only works only if nbBits >= 1 */
+*   unsafe version; only works if nbBits >= 1 */
 MEM_STATIC size_t BIT_lookBitsFast(BIT_DStream_t* bitD, U32 nbBits)
 {
     const U32 bitMask = sizeof(bitD->bitContainer)*8 - 1;
@@ -455,7 +406,7 @@ MEM_STATIC size_t BIT_readBits(BIT_DStream_t* bitD, U32 nbBits)
 }
 
 /*!BIT_readBitsFast :
-*  unsafe version; only works only if nbBits >= 1 */
+*  unsafe version; only works if nbBits >= 1 */
 MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, U32 nbBits)
 {
     size_t value = BIT_lookBitsFast(bitD, nbBits);
@@ -512,7 +463,7 @@ MEM_STATIC unsigned BIT_endOfDStream(const BIT_DStream_t* DStream)
    Error codes and messages
    Copyright (C) 2013-2015, Yann Collet
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -611,7 +562,7 @@ typedef unsigned FSE_DTable;   /* don't allocate that. It's just a way to be mor
    header file for static linking (only)
    Copyright (C) 2013-2015, Yann Collet
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -755,7 +706,7 @@ MEM_STATIC unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr)
    header file for static linking (only)
    Copyright (C) 2013-2015, Yann Collet
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -823,7 +774,7 @@ static size_t HUF_decompress4X4 (void* dst, size_t dstSize, const void* cSrc, si
     Header File
     Copyright (C) 2014-2015, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are
@@ -883,7 +834,7 @@ typedef struct ZSTD_CCtx_s ZSTD_CCtx;   /* incomplete type */
     Header File for static linking only
     Copyright (C) 2014-2015, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are
@@ -947,7 +898,7 @@ typedef struct ZSTD_DCtx_s ZSTD_DCtx;
    FSE : Finite State Entropy coder
    Copyright (C) 2013-2015, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -1451,7 +1402,7 @@ static size_t FSE_decompress(void* dst, size_t maxDstSize, const void* cSrc, siz
    Huff0 : Huffman coder, part of New Generation Entropy library
    Copyright (C) 2013-2015, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -2248,7 +2199,7 @@ static size_t HUF_decompress (void* dst, size_t dstSize, const void* cSrc, size_
     zstd - standard compression library
     Copyright (C) 2014-2015, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are
@@ -2755,18 +2706,24 @@ static size_t ZSTD_execSequence(BYTE* op,
     const BYTE* const litEnd = *litPtr + sequence.litLength;
 
     /* checks */
-    if (oLitEnd > oend_8) return ERROR(dstSize_tooSmall);   /* last match must start at a minimum distance of 8 from oend */
+    size_t const seqLength = sequence.litLength + sequence.matchLength;
+
+    if (seqLength > (size_t)(oend - op)) return ERROR(dstSize_tooSmall);
+    if (sequence.litLength > (size_t)(litLimit - *litPtr)) return ERROR(corruption_detected);
+    /* Now we know there are no overflow in literal nor match lengths, can use pointer checks */
+    if (oLitEnd > oend_8) return ERROR(dstSize_tooSmall);
+    if (sequence.offset > (U32)(oLitEnd - base)) return ERROR(corruption_detected);
+
     if (oMatchEnd > oend) return ERROR(dstSize_tooSmall);   /* overwrite beyond dst buffer */
     if (litEnd > litLimit) return ERROR(corruption_detected);   /* overRead beyond lit buffer */
 
     /* copy Literals */
-    ZSTD_wildcopy(op, *litPtr, sequence.litLength);   /* note : oLitEnd <= oend-8 : no risk of overwrite beyond oend */
+    ZSTD_wildcopy(op, *litPtr, (ptrdiff_t)sequence.litLength);   /* note : oLitEnd <= oend-8 : no risk of overwrite beyond oend */
     op = oLitEnd;
     *litPtr = litEnd;   /* update for next sequence */
 
     /* copy Match */
-    {
-        const BYTE* match = op - sequence.offset;
+    {   const BYTE* match = op - sequence.offset;
 
         /* check */
         if (sequence.offset > (size_t)op) return ERROR(corruption_detected);   /* address space overflow test (this test seems kept by clang optimizer) */

src/borg/algorithms/zstd/lib/legacy/zstd_v03.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the

src/borg/algorithms/zstd/lib/legacy/zstd_v04.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -52,7 +52,11 @@ extern "C" {
 *  Basic Types
 *****************************************************************/
 #if defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
-# include <stdint.h>
+# if defined(_AIX)
+#  include <inttypes.h>
+# else
+#  include <stdint.h> /* intptr_t */
+# endif
   typedef  uint8_t BYTE;
   typedef uint16_t U16;
   typedef  int16_t S16;
@@ -74,7 +78,7 @@ extern "C" {
 /*-*************************************
 *  Debug
 ***************************************/
-#include "debug.h"
+#include "../common/debug.h"
 #ifndef assert
 #  define assert(condition) ((void)0)
 #endif
@@ -83,27 +87,6 @@ extern "C" {
 /****************************************************************
 *  Memory I/O
 *****************************************************************/
-/* MEM_FORCE_MEMORY_ACCESS
- * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
- * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
- * The below switch allow to select different access method for improved performance.
- * Method 0 (default) : use `memcpy()`. Safe and portable.
- * Method 1 : `__packed` statement. It depends on compiler extension (ie, not portable).
- *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
- * Method 2 : direct access. This method is portable but violate C standard.
- *            It can generate buggy code on targets generating assembly depending on alignment.
- *            But in some circumstances, it's the only known way to get the most performance (ie GCC + ARMv6)
- * See http://fastcompression.blogspot.fr/2015/08/accessing-unaligned-memory.html for details.
- * Prefer these methods in priority order (0 > 1 > 2)
- */
-#ifndef MEM_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
-#  if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
-#    define MEM_FORCE_MEMORY_ACCESS 2
-#  elif (defined(__INTEL_COMPILER) && !defined(WIN32)) || \
-  (defined(__GNUC__) && ( defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) || defined(__ARM_ARCH_7S__) ))
-#    define MEM_FORCE_MEMORY_ACCESS 1
-#  endif
-#endif
 
 MEM_STATIC unsigned MEM_32bits(void) { return sizeof(void*)==4; }
 MEM_STATIC unsigned MEM_64bits(void) { return sizeof(void*)==8; }
@@ -114,33 +97,6 @@ MEM_STATIC unsigned MEM_isLittleEndian(void)
     return one.c[0];
 }
 
-#if defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==2)
-
-/* violates C standard 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 void MEM_write16(void* memPtr, U16 value) { *(U16*)memPtr = value; }
-
-#elif defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==1)
-
-/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
-/* currently only defined for gcc and icc */
-typedef union { U16 u16; U32 u32; U64 u64; } __attribute__((packed)) unalign;
-
-MEM_STATIC U16 MEM_read16(const void* ptr) { return ((const unalign*)ptr)->u16; }
-MEM_STATIC U32 MEM_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
-MEM_STATIC U64 MEM_read64(const void* ptr) { return ((const unalign*)ptr)->u64; }
-
-MEM_STATIC void MEM_write16(void* memPtr, U16 value) { ((unalign*)memPtr)->u16 = value; }
-
-#else
-
-/* default method, safe and standard.
-   can sometimes prove slower */
-
 MEM_STATIC U16 MEM_read16(const void* memPtr)
 {
     U16 val; memcpy(&val, memPtr, sizeof(val)); return val;
@@ -161,9 +117,6 @@ MEM_STATIC void MEM_write16(void* memPtr, U16 value)
     memcpy(memPtr, &value, sizeof(value));
 }
 
-#endif /* MEM_FORCE_MEMORY_ACCESS */
-
-
 MEM_STATIC U16 MEM_readLE16(const void* memPtr)
 {
     if (MEM_isLittleEndian())
@@ -541,7 +494,7 @@ If there is an error, the function will return an error code, which can be teste
    header file (to include)
    Copyright (C) 2013-2015, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -623,9 +576,8 @@ MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, unsigned nbBits);
 MEM_STATIC unsigned BIT_highbit32 (U32 val)
 {
 #   if defined(_MSC_VER)   /* Visual */
-    unsigned long r=0;
-    _BitScanReverse ( &r, val );
-    return (unsigned) r;
+    unsigned long r;
+    return _BitScanReverse(&r, val) ? (unsigned)r : 0;
 #   elif defined(__GNUC__) && (__GNUC__ >= 3)   /* Use GCC Intrinsic */
     return __builtin_clz (val) ^ 31;
 #   else   /* Software version */
@@ -700,7 +652,7 @@ MEM_STATIC size_t BIT_lookBits(BIT_DStream_t* bitD, U32 nbBits)
 }
 
 /*! BIT_lookBitsFast :
-*   unsafe version; only works only if nbBits >= 1 */
+*   unsafe version; only works if nbBits >= 1 */
 MEM_STATIC size_t BIT_lookBitsFast(BIT_DStream_t* bitD, U32 nbBits)
 {
     const U32 bitMask = sizeof(bitD->bitContainer)*8 - 1;
@@ -720,7 +672,7 @@ MEM_STATIC size_t BIT_readBits(BIT_DStream_t* bitD, U32 nbBits)
 }
 
 /*!BIT_readBitsFast :
-*  unsafe version; only works only if nbBits >= 1 */
+*  unsafe version; only works if nbBits >= 1 */
 MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, U32 nbBits)
 {
     size_t value = BIT_lookBitsFast(bitD, nbBits);
@@ -781,7 +733,7 @@ MEM_STATIC unsigned BIT_endOfDStream(const BIT_DStream_t* DStream)
    header file for static linking (only)
    Copyright (C) 2013-2015, Yann Collet
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -930,7 +882,7 @@ MEM_STATIC unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr)
    FSE : Finite State Entropy coder
    Copyright (C) 2013-2015, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -1436,7 +1388,7 @@ static size_t FSE_decompress(void* dst, size_t maxDstSize, const void* cSrc, siz
    header file
    Copyright (C) 2013-2015, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -1514,7 +1466,7 @@ static unsigned    HUF_isError(size_t code);        /* tells if a return value i
    header file for static linking (only)
    Copyright (C) 2013-2015, Yann Collet
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -1601,7 +1553,7 @@ static size_t HUF_decompress4X4_usingDTable(void* dst, size_t maxDstSize, const
    Huff0 : Huffman coder, part of New Generation Entropy library
    Copyright (C) 2013-2015, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -2401,7 +2353,7 @@ static size_t HUF_decompress (void* dst, size_t dstSize, const void* cSrc, size_
     zstd - decompression module fo v0.4 legacy format
     Copyright (C) 2015-2016, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are
@@ -2876,13 +2828,19 @@ static size_t ZSTD_execSequence(BYTE* op,
     const BYTE* const litEnd = *litPtr + sequence.litLength;
     const BYTE* match = oLitEnd - sequence.offset;
 
-    /* check */
-    if (oLitEnd > oend_8) return ERROR(dstSize_tooSmall);   /* last match must start at a minimum distance of 8 from oend */
+    /* checks */
+    size_t const seqLength = sequence.litLength + sequence.matchLength;
+
+    if (seqLength > (size_t)(oend - op)) return ERROR(dstSize_tooSmall);
+    if (sequence.litLength > (size_t)(litLimit - *litPtr)) return ERROR(corruption_detected);
+    /* Now we know there are no overflow in literal nor match lengths, can use pointer checks */
+    if (oLitEnd > oend_8) return ERROR(dstSize_tooSmall);
+
     if (oMatchEnd > oend) return ERROR(dstSize_tooSmall);   /* overwrite beyond dst buffer */
-    if (litEnd > litLimit) return ERROR(corruption_detected);   /* risk read beyond lit buffer */
+    if (litEnd > litLimit) return ERROR(corruption_detected);   /* overRead beyond lit buffer */
 
     /* copy Literals */
-    ZSTD_wildcopy(op, *litPtr, sequence.litLength);   /* note : oLitEnd <= oend-8 : no risk of overwrite beyond oend */
+    ZSTD_wildcopy(op, *litPtr, (ptrdiff_t)sequence.litLength);   /* note : oLitEnd <= oend-8 : no risk of overwrite beyond oend */
     op = oLitEnd;
     *litPtr = litEnd;   /* update for next sequence */
 
@@ -3283,7 +3241,7 @@ static void ZSTD_decompress_insertDictionary(ZSTD_DCtx* ctx, const void* dict, s
     Buffered version of Zstd compression library
     Copyright (C) 2015, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are

src/borg/algorithms/zstd/lib/legacy/zstd_v04.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the

src/borg/algorithms/zstd/lib/legacy/zstd_v05.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -19,7 +19,7 @@
    low-level memory access routines
    Copyright (C) 2013-2015, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -80,7 +80,11 @@ extern "C" {
 *  Basic Types
 *****************************************************************/
 #if defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
-# include <stdint.h>
+# if defined(_AIX)
+#  include <inttypes.h>
+# else
+#  include <stdint.h> /* intptr_t */
+# endif
   typedef  uint8_t BYTE;
   typedef uint16_t U16;
   typedef  int16_t S16;
@@ -102,27 +106,6 @@ extern "C" {
 /*-**************************************************************
 *  Memory I/O
 *****************************************************************/
-/* MEM_FORCE_MEMORY_ACCESS :
- * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
- * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
- * The below switch allow to select different access method for improved performance.
- * Method 0 (default) : use `memcpy()`. Safe and portable.
- * Method 1 : `__packed` statement. It depends on compiler extension (ie, not portable).
- *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
- * Method 2 : direct access. This method is portable but violate C standard.
- *            It can generate buggy code on targets depending on alignment.
- *            In some circumstances, it's the only known way to get the most performance (ie GCC + ARMv6)
- * See http://fastcompression.blogspot.fr/2015/08/accessing-unaligned-memory.html for details.
- * Prefer these methods in priority order (0 > 1 > 2)
- */
-#ifndef MEM_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
-#  if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
-#    define MEM_FORCE_MEMORY_ACCESS 2
-#  elif (defined(__INTEL_COMPILER) && !defined(WIN32)) || \
-  (defined(__GNUC__) && ( defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) || defined(__ARM_ARCH_7S__) ))
-#    define MEM_FORCE_MEMORY_ACCESS 1
-#  endif
-#endif
 
 MEM_STATIC unsigned MEM_32bits(void) { return sizeof(void*)==4; }
 MEM_STATIC unsigned MEM_64bits(void) { return sizeof(void*)==8; }
@@ -133,37 +116,6 @@ MEM_STATIC unsigned MEM_isLittleEndian(void)
     return one.c[0];
 }
 
-#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 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)
-
-/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
-/* currently only defined for gcc and icc */
-typedef union { U16 u16; U32 u32; U64 u64; size_t st; } __attribute__((packed)) unalign;
-
-MEM_STATIC U16 MEM_read16(const void* ptr) { return ((const unalign*)ptr)->u16; }
-MEM_STATIC U32 MEM_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
-MEM_STATIC U64 MEM_read64(const void* ptr) { return ((const unalign*)ptr)->u64; }
-
-MEM_STATIC void MEM_write16(void* memPtr, U16 value) { ((unalign*)memPtr)->u16 = value; }
-MEM_STATIC void MEM_write32(void* memPtr, U32 value) { ((unalign*)memPtr)->u32 = value; }
-MEM_STATIC void MEM_write64(void* memPtr, U64 value) { ((unalign*)memPtr)->u64 = value; }
-
-#else
-
-/* default method, safe and standard.
-   can sometimes prove slower */
-
 MEM_STATIC U16 MEM_read16(const void* memPtr)
 {
     U16 val; memcpy(&val, memPtr, sizeof(val)); return val;
@@ -194,9 +146,6 @@ MEM_STATIC void MEM_write64(void* memPtr, U64 value)
     memcpy(memPtr, &value, sizeof(value));
 }
 
-#endif /* MEM_FORCE_MEMORY_ACCESS */
-
-
 MEM_STATIC U16 MEM_readLE16(const void* memPtr)
 {
     if (MEM_isLittleEndian())
@@ -261,7 +210,7 @@ MEM_STATIC size_t MEM_readLEST(const void* memPtr)
     Header File for static linking only
     Copyright (C) 2014-2016, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are
@@ -285,7 +234,7 @@ MEM_STATIC size_t MEM_readLEST(const void* memPtr)
     OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
     You can contact the author at :
-    - zstd homepage : http://www.zstd.net
+    - zstd homepage : https://facebook.github.io/zstd
 */
 #ifndef ZSTD_STATIC_H
 #define ZSTD_STATIC_H
@@ -397,7 +346,7 @@ size_t ZSTDv05_decompressBlock(ZSTDv05_DCtx* dctx, void* dst, size_t dstCapacity
     Header File for include
     Copyright (C) 2014-2016, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are
@@ -484,7 +433,7 @@ static const size_t ZSTDv05_frameHeaderSize_min = 5;
 #define FSEv05_ENCODING_DYNAMIC 3
 
 
-#define HufLog 12
+#define ZSTD_HUFFDTABLE_CAPACITY_LOG 12
 
 #define MIN_SEQUENCES_SIZE 1 /* nbSeq==0 */
 #define MIN_CBLOCK_SIZE (1 /*litCSize*/ + 1 /* RLE or RAW */ + MIN_SEQUENCES_SIZE /* nbSeq==0 */)   /* for a non-null block */
@@ -552,7 +501,7 @@ typedef struct {
    header file
    Copyright (C) 2013-2015, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -671,7 +620,7 @@ size_t FSEv05_decompress_usingDTable(void* dst, size_t dstCapacity, const void*
    header file (to include)
    Copyright (C) 2013-2016, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -752,9 +701,8 @@ MEM_STATIC size_t BITv05_readBitsFast(BITv05_DStream_t* bitD, unsigned nbBits);
 MEM_STATIC unsigned BITv05_highbit32 (U32 val)
 {
 #   if defined(_MSC_VER)   /* Visual */
-    unsigned long r=0;
-    _BitScanReverse ( &r, val );
-    return (unsigned) r;
+    unsigned long r;
+    return _BitScanReverse(&r, val) ? (unsigned)r : 0;
 #   elif defined(__GNUC__) && (__GNUC__ >= 3)   /* Use GCC Intrinsic */
     return __builtin_clz (val) ^ 31;
 #   else   /* Software version */
@@ -826,7 +774,7 @@ MEM_STATIC size_t BITv05_lookBits(BITv05_DStream_t* bitD, U32 nbBits)
 }
 
 /*! BITv05_lookBitsFast :
-*   unsafe version; only works only if nbBits >= 1 */
+*   unsafe version; only works if nbBits >= 1 */
 MEM_STATIC size_t BITv05_lookBitsFast(BITv05_DStream_t* bitD, U32 nbBits)
 {
     const U32 bitMask = sizeof(bitD->bitContainer)*8 - 1;
@@ -846,7 +794,7 @@ MEM_STATIC size_t BITv05_readBits(BITv05_DStream_t* bitD, unsigned nbBits)
 }
 
 /*!BITv05_readBitsFast :
-*  unsafe version; only works only if nbBits >= 1 */
+*  unsafe version; only works if nbBits >= 1 */
 MEM_STATIC size_t BITv05_readBitsFast(BITv05_DStream_t* bitD, unsigned nbBits)
 {
     size_t value = BITv05_lookBitsFast(bitD, nbBits);
@@ -901,7 +849,7 @@ MEM_STATIC unsigned BITv05_endOfDStream(const BITv05_DStream_t* DStream)
    header file for static linking (only)
    Copyright (C) 2013-2015, Yann Collet
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -1051,7 +999,7 @@ MEM_STATIC unsigned FSEv05_endOfDState(const FSEv05_DState_t* DStatePtr)
    FSEv05 : Finite State Entropy coder
    Copyright (C) 2013-2015, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -1537,7 +1485,7 @@ size_t FSEv05_decompress(void* dst, size_t maxDstSize, const void* cSrc, size_t
    header file
    Copyright (C) 2013-2016, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -1610,7 +1558,7 @@ const char* HUFv05_getErrorName(size_t code);   /* provides error code string (u
    header file, for static linking only
    Copyright (C) 2013-2016, Yann Collet
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -1702,7 +1650,7 @@ size_t HUFv05_decompress1X4_usingDTable(void* dst, size_t maxDstSize, const void
    Huff0 : Huffman coder, part of New Generation Entropy library
    Copyright (C) 2013-2015, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -2547,7 +2495,7 @@ size_t HUFv05_decompress (void* dst, size_t dstSize, const void* cSrc, size_t cS
     zstd - standard compression library
     Copyright (C) 2014-2016, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are
@@ -2645,7 +2593,7 @@ struct ZSTDv05_DCtx_s
     FSEv05_DTable LLTable[FSEv05_DTABLE_SIZE_U32(LLFSEv05Log)];
     FSEv05_DTable OffTable[FSEv05_DTABLE_SIZE_U32(OffFSEv05Log)];
     FSEv05_DTable MLTable[FSEv05_DTABLE_SIZE_U32(MLFSEv05Log)];
-    unsigned   hufTableX4[HUFv05_DTABLE_SIZE(HufLog)];
+    unsigned   hufTableX4[HUFv05_DTABLE_SIZE(ZSTD_HUFFDTABLE_CAPACITY_LOG)];
     const void* previousDstEnd;
     const void* base;
     const void* vBase;
@@ -2673,7 +2621,7 @@ size_t ZSTDv05_decompressBegin(ZSTDv05_DCtx* dctx)
     dctx->base = NULL;
     dctx->vBase = NULL;
     dctx->dictEnd = NULL;
-    dctx->hufTableX4[0] = HufLog;
+    dctx->hufTableX4[0] = ZSTD_HUFFDTABLE_CAPACITY_LOG;
     dctx->flagStaticTables = 0;
     return 0;
 }
@@ -2829,7 +2777,7 @@ static size_t ZSTDv05_decodeFrameHeader_Part2(ZSTDv05_DCtx* zc, const void* src,
 
 static size_t ZSTDv05_getcBlockSize(const void* src, size_t srcSize, blockProperties_t* bpPtr)
 {
-    const BYTE* const in = (const BYTE* const)src;
+    const BYTE* const in = (const BYTE*)src;
     BYTE headerFlags;
     U32 cSize;
 
@@ -2998,7 +2946,7 @@ static size_t ZSTDv05_decodeSeqHeaders(int* nbSeq, const BYTE** dumpsPtr, size_t
                          FSEv05_DTable* DTableLL, FSEv05_DTable* DTableML, FSEv05_DTable* DTableOffb,
                          const void* src, size_t srcSize, U32 flagStaticTable)
 {
-    const BYTE* const istart = (const BYTE* const)src;
+    const BYTE* const istart = (const BYTE*)src;
     const BYTE* ip = istart;
     const BYTE* const iend = istart + srcSize;
     U32 LLtype, Offtype, MLtype;
@@ -3234,13 +3182,19 @@ static size_t ZSTDv05_execSequence(BYTE* op,
     const BYTE* const litEnd = *litPtr + sequence.litLength;
     const BYTE* match = oLitEnd - sequence.offset;
 
-    /* check */
-    if (oLitEnd > oend_8) return ERROR(dstSize_tooSmall);   /* last match must start at a minimum distance of 8 from oend */
+    /* checks */
+    size_t const seqLength = sequence.litLength + sequence.matchLength;
+
+    if (seqLength > (size_t)(oend - op)) return ERROR(dstSize_tooSmall);
+    if (sequence.litLength > (size_t)(litLimit - *litPtr)) return ERROR(corruption_detected);
+    /* Now we know there are no overflow in literal nor match lengths, can use pointer checks */
+    if (oLitEnd > oend_8) return ERROR(dstSize_tooSmall);
+
     if (oMatchEnd > oend) return ERROR(dstSize_tooSmall);   /* overwrite beyond dst buffer */
-    if (litEnd > litLimit) return ERROR(corruption_detected);   /* risk read beyond lit buffer */
+    if (litEnd > litLimit) return ERROR(corruption_detected);   /* overRead beyond lit buffer */
 
     /* copy Literals */
-    ZSTDv05_wildcopy(op, *litPtr, sequence.litLength);   /* note : oLitEnd <= oend-8 : no risk of overwrite beyond oend */
+    ZSTDv05_wildcopy(op, *litPtr, (ptrdiff_t)sequence.litLength);   /* note : oLitEnd <= oend-8 : no risk of overwrite beyond oend */
     op = oLitEnd;
     *litPtr = litEnd;   /* update for next sequence */
 
@@ -3306,7 +3260,7 @@ static size_t ZSTDv05_decompressSequences(
 {
     const BYTE* ip = (const BYTE*)seqStart;
     const BYTE* const iend = ip + seqSize;
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     BYTE* op = ostart;
     BYTE* const oend = ostart + maxDstSize;
     size_t errorCode, dumpsLength=0;
@@ -3419,7 +3373,7 @@ static size_t ZSTDv05_decompress_continueDCtx(ZSTDv05_DCtx* dctx,
 {
     const BYTE* ip = (const BYTE*)src;
     const BYTE* iend = ip + srcSize;
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     BYTE* op = ostart;
     BYTE* const oend = ostart + maxDstSize;
     size_t remainingSize = srcSize;
@@ -3746,7 +3700,7 @@ size_t ZSTDv05_decompressBegin_usingDict(ZSTDv05_DCtx* dctx, const void* dict, s
     Buffered version of Zstd compression library
     Copyright (C) 2015-2016, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are

src/borg/algorithms/zstd/lib/legacy/zstd_v05.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the

src/borg/algorithms/zstd/lib/legacy/zstd_v06.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -23,7 +23,7 @@
    low-level memory access routines
    Copyright (C) 2013-2015, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -82,7 +82,11 @@ extern "C" {
 *  Basic Types
 *****************************************************************/
 #if  !defined (__VMS) && (defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
-# include <stdint.h>
+# if defined(_AIX)
+#  include <inttypes.h>
+# else
+#  include <stdint.h> /* intptr_t */
+# endif
   typedef  uint8_t BYTE;
   typedef uint16_t U16;
   typedef  int16_t S16;
@@ -104,27 +108,6 @@ extern "C" {
 /*-**************************************************************
 *  Memory I/O
 *****************************************************************/
-/* MEM_FORCE_MEMORY_ACCESS :
- * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
- * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
- * The below switch allow to select different access method for improved performance.
- * Method 0 (default) : use `memcpy()`. Safe and portable.
- * Method 1 : `__packed` statement. It depends on compiler extension (ie, not portable).
- *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
- * Method 2 : direct access. This method is portable but violate C standard.
- *            It can generate buggy code on targets depending on alignment.
- *            In some circumstances, it's the only known way to get the most performance (ie GCC + ARMv6)
- * See http://fastcompression.blogspot.fr/2015/08/accessing-unaligned-memory.html for details.
- * Prefer these methods in priority order (0 > 1 > 2)
- */
-#ifndef MEM_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
-#  if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
-#    define MEM_FORCE_MEMORY_ACCESS 2
-#  elif (defined(__INTEL_COMPILER) && !defined(WIN32)) || \
-  (defined(__GNUC__) && ( defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) || defined(__ARM_ARCH_7S__) ))
-#    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; }
@@ -135,33 +118,6 @@ MEM_STATIC unsigned MEM_isLittleEndian(void)
     return one.c[0];
 }
 
-#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 void MEM_write16(void* memPtr, U16 value) { *(U16*)memPtr = value; }
-
-#elif defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==1)
-
-/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
-/* currently only defined for gcc and icc */
-typedef union { U16 u16; U32 u32; U64 u64; size_t st; } __attribute__((packed)) unalign;
-
-MEM_STATIC U16 MEM_read16(const void* ptr) { return ((const unalign*)ptr)->u16; }
-MEM_STATIC U32 MEM_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
-MEM_STATIC U64 MEM_read64(const void* ptr) { return ((const unalign*)ptr)->u64; }
-
-MEM_STATIC void MEM_write16(void* memPtr, U16 value) { ((unalign*)memPtr)->u16 = value; }
-
-#else
-
-/* default method, safe and standard.
-   can sometimes prove slower */
-
 MEM_STATIC U16 MEM_read16(const void* memPtr)
 {
     U16 val; memcpy(&val, memPtr, sizeof(val)); return val;
@@ -182,9 +138,6 @@ MEM_STATIC void MEM_write16(void* memPtr, U16 value)
     memcpy(memPtr, &value, sizeof(value));
 }
 
-
-#endif /* MEM_FORCE_MEMORY_ACCESS */
-
 MEM_STATIC U32 MEM_swap32(U32 in)
 {
 #if defined(_MSC_VER)     /* Visual Studio */
@@ -280,7 +233,7 @@ MEM_STATIC size_t MEM_readLEST(const void* memPtr)
     Header File for static linking only
     Copyright (C) 2014-2016, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are
@@ -304,7 +257,7 @@ MEM_STATIC size_t MEM_readLEST(const void* memPtr)
     OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
     You can contact the author at :
-    - zstd homepage : http://www.zstd.net
+    - zstd homepage : https://facebook.github.io/zstd
 */
 #ifndef ZSTDv06_STATIC_H
 #define ZSTDv06_STATIC_H
@@ -411,7 +364,7 @@ ZSTDLIBv06_API size_t ZSTDv06_decompressBlock(ZSTDv06_DCtx* dctx, void* dst, siz
     Header File for include
     Copyright (C) 2014-2016, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are
@@ -478,7 +431,7 @@ typedef enum { bt_compressed, bt_raw, bt_rle, bt_end } blockType_t;
 #define MIN_SEQUENCES_SIZE 1 /* nbSeq==0 */
 #define MIN_CBLOCK_SIZE (1 /*litCSize*/ + 1 /* RLE or RAW */ + MIN_SEQUENCES_SIZE /* nbSeq==0 */)   /* for a non-null block */
 
-#define HufLog 12
+#define ZSTD_HUFFDTABLE_CAPACITY_LOG 12
 
 #define IS_HUF 0
 #define IS_PCH 1
@@ -618,7 +571,7 @@ void ZSTDv06_seqToCodes(const seqStore_t* seqStorePtr, size_t const nbSeq);
    Public Prototypes declaration
    Copyright (C) 2013-2016, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -766,7 +719,7 @@ If there is an error, the function will return an error code, which can be teste
    header file (to include)
    Copyright (C) 2013-2016, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -856,9 +809,8 @@ MEM_STATIC size_t BITv06_readBitsFast(BITv06_DStream_t* bitD, unsigned nbBits);
 MEM_STATIC unsigned BITv06_highbit32 ( U32 val)
 {
 #   if defined(_MSC_VER)   /* Visual */
-    unsigned long r=0;
-    _BitScanReverse ( &r, val );
-    return (unsigned) r;
+    unsigned long r;
+    return _BitScanReverse(&r, val) ? (unsigned)r : 0;
 #   elif defined(__GNUC__) && (__GNUC__ >= 3)   /* Use GCC Intrinsic */
     return __builtin_clz (val) ^ 31;
 #   else   /* Software version */
@@ -928,7 +880,7 @@ MEM_STATIC size_t BITv06_initDStream(BITv06_DStream_t* bitD, const void* srcBuff
 }
 
 /*! BITv06_lookBitsFast() :
-*   unsafe version; only works only if nbBits >= 1 */
+*   unsafe version; only works if nbBits >= 1 */
 MEM_STATIC size_t BITv06_lookBitsFast(const BITv06_DStream_t* bitD, U32 nbBits)
 {
     U32 const bitMask = sizeof(bitD->bitContainer)*8 - 1;
@@ -948,7 +900,7 @@ MEM_STATIC size_t BITv06_readBits(BITv06_DStream_t* bitD, U32 nbBits)
 }
 
 /*! BITv06_readBitsFast() :
-*   unsafe version; only works only if nbBits >= 1 */
+*   unsafe version; only works if nbBits >= 1 */
 MEM_STATIC size_t BITv06_readBitsFast(BITv06_DStream_t* bitD, U32 nbBits)
 {
     size_t const value = BITv06_lookBitsFast(bitD, nbBits);
@@ -1002,7 +954,7 @@ MEM_STATIC unsigned BITv06_endOfDStream(const BITv06_DStream_t* DStream)
    header file for static linking (only)
    Copyright (C) 2013-2015, Yann Collet
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -1210,7 +1162,7 @@ MEM_STATIC BYTE FSEv06_decodeSymbolFast(FSEv06_DState_t* DStatePtr, BITv06_DStre
    Common functions of New Generation Entropy library
    Copyright (C) 2016, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -1355,7 +1307,7 @@ size_t FSEv06_readNCount (short* normalizedCounter, unsigned* maxSVPtr, unsigned
    FSE : Finite State Entropy decoder
    Copyright (C) 2013-2015, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -1679,7 +1631,7 @@ size_t FSEv06_decompress(void* dst, size_t maxDstSize, const void* cSrc, size_t
    header file
    Copyright (C) 2013-2016, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -1749,7 +1701,7 @@ size_t HUFv06_compressBound(size_t size);       /**< maximum compressed size */
    header file, for static linking only
    Copyright (C) 2013-2016, Yann Collet
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -1931,7 +1883,7 @@ MEM_STATIC size_t HUFv06_readStats(BYTE* huffWeight, size_t hwSize, U32* rankSta
    Huffman decoder, part of New Generation Entropy library
    Copyright (C) 2013-2016, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -2676,7 +2628,7 @@ size_t HUFv06_decompress (void* dst, size_t dstSize, const void* cSrc, size_t cS
     Common functions of Zstd compression library
     Copyright (C) 2015-2016, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are
@@ -2700,7 +2652,7 @@ size_t HUFv06_decompress (void* dst, size_t dstSize, const void* cSrc, size_t cS
     OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
     You can contact the author at :
-    - zstd homepage : http://www.zstd.net/
+    - zstd homepage : https://facebook.github.io/zstd/
 */
 
 
@@ -2730,7 +2682,7 @@ const char* ZBUFFv06_getErrorName(size_t errorCode) { return ERR_getErrorName(er
     zstd - standard compression library
     Copyright (C) 2014-2016, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are
@@ -2754,7 +2706,7 @@ const char* ZBUFFv06_getErrorName(size_t errorCode) { return ERR_getErrorName(er
     OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
     You can contact the author at :
-    - zstd homepage : http://www.zstd.net
+    - zstd homepage : https://facebook.github.io/zstd
 */
 
 /* ***************************************************************
@@ -2806,7 +2758,7 @@ struct ZSTDv06_DCtx_s
     FSEv06_DTable LLTable[FSEv06_DTABLE_SIZE_U32(LLFSELog)];
     FSEv06_DTable OffTable[FSEv06_DTABLE_SIZE_U32(OffFSELog)];
     FSEv06_DTable MLTable[FSEv06_DTABLE_SIZE_U32(MLFSELog)];
-    unsigned   hufTableX4[HUFv06_DTABLE_SIZE(HufLog)];
+    unsigned   hufTableX4[HUFv06_DTABLE_SIZE(ZSTD_HUFFDTABLE_CAPACITY_LOG)];
     const void* previousDstEnd;
     const void* base;
     const void* vBase;
@@ -2834,7 +2786,7 @@ size_t ZSTDv06_decompressBegin(ZSTDv06_DCtx* dctx)
     dctx->base = NULL;
     dctx->vBase = NULL;
     dctx->dictEnd = NULL;
-    dctx->hufTableX4[0] = HufLog;
+    dctx->hufTableX4[0] = ZSTD_HUFFDTABLE_CAPACITY_LOG;
     dctx->flagRepeatTable = 0;
     return 0;
 }
@@ -3025,7 +2977,7 @@ typedef struct
 *   Provides the size of compressed block from block header `src` */
 static size_t ZSTDv06_getcBlockSize(const void* src, size_t srcSize, blockProperties_t* bpPtr)
 {
-    const BYTE* const in = (const BYTE* const)src;
+    const BYTE* const in = (const BYTE*)src;
     U32 cSize;
 
     if (srcSize < ZSTDv06_blockHeaderSize) return ERROR(srcSize_wrong);
@@ -3219,7 +3171,7 @@ static size_t ZSTDv06_decodeSeqHeaders(int* nbSeqPtr,
                              FSEv06_DTable* DTableLL, FSEv06_DTable* DTableML, FSEv06_DTable* DTableOffb, U32 flagRepeatTable,
                              const void* src, size_t srcSize)
 {
-    const BYTE* const istart = (const BYTE* const)src;
+    const BYTE* const istart = (const BYTE*)src;
     const BYTE* const iend = istart + srcSize;
     const BYTE* ip = istart;
 
@@ -3370,13 +3322,19 @@ static size_t ZSTDv06_execSequence(BYTE* op,
     const BYTE* const iLitEnd = *litPtr + sequence.litLength;
     const BYTE* match = oLitEnd - sequence.offset;
 
-    /* check */
-    if (oLitEnd > oend_8) return ERROR(dstSize_tooSmall);   /* last match must start at a minimum distance of 8 from oend */
+    /* checks */
+    size_t const seqLength = sequence.litLength + sequence.matchLength;
+
+    if (seqLength > (size_t)(oend - op)) return ERROR(dstSize_tooSmall);
+    if (sequence.litLength > (size_t)(litLimit - *litPtr)) return ERROR(corruption_detected);
+    /* Now we know there are no overflow in literal nor match lengths, can use pointer checks */
+    if (oLitEnd > oend_8) return ERROR(dstSize_tooSmall);
+
     if (oMatchEnd > oend) return ERROR(dstSize_tooSmall);   /* overwrite beyond dst buffer */
-    if (iLitEnd > litLimit) return ERROR(corruption_detected);   /* over-read beyond lit buffer */
+    if (iLitEnd > litLimit) return ERROR(corruption_detected);   /* overRead beyond lit buffer */
 
     /* copy Literals */
-    ZSTDv06_wildcopy(op, *litPtr, sequence.litLength);   /* note : oLitEnd <= oend-8 : no risk of overwrite beyond oend */
+    ZSTDv06_wildcopy(op, *litPtr, (ptrdiff_t)sequence.litLength);   /* note : oLitEnd <= oend-8 : no risk of overwrite beyond oend */
     op = oLitEnd;
     *litPtr = iLitEnd;   /* update for next sequence */
 
@@ -3441,7 +3399,7 @@ static size_t ZSTDv06_decompressSequences(
 {
     const BYTE* ip = (const BYTE*)seqStart;
     const BYTE* const iend = ip + seqSize;
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     BYTE* const oend = ostart + maxDstSize;
     BYTE* op = ostart;
     const BYTE* litPtr = dctx->litPtr;
@@ -3557,7 +3515,7 @@ static size_t ZSTDv06_decompressFrame(ZSTDv06_DCtx* dctx,
 {
     const BYTE* ip = (const BYTE*)src;
     const BYTE* const iend = ip + srcSize;
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     BYTE* op = ostart;
     BYTE* const oend = ostart + dstCapacity;
     size_t remainingSize = srcSize;
@@ -3889,7 +3847,7 @@ size_t ZSTDv06_decompressBegin_usingDict(ZSTDv06_DCtx* dctx, const void* dict, s
     Buffered version of Zstd compression library
     Copyright (C) 2015-2016, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are
@@ -3913,7 +3871,7 @@ size_t ZSTDv06_decompressBegin_usingDict(ZSTDv06_DCtx* dctx, const void* dict, s
     OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
     You can contact the author at :
-    - zstd homepage : http://www.zstd.net/
+    - zstd homepage : https://facebook.github.io/zstd/
 */
 
 
@@ -4035,7 +3993,8 @@ size_t ZBUFFv06_decompressContinue(ZBUFFv06_DCtx* zbd,
                     size_t const toLoad = hSize - zbd->lhSize;   /* if hSize!=0, hSize > zbd->lhSize */
                     if (ZSTDv06_isError(hSize)) return hSize;
                     if (toLoad > (size_t)(iend-ip)) {   /* not enough input to load full header */
-                        memcpy(zbd->headerBuffer + zbd->lhSize, ip, iend-ip);
+                        if (ip != NULL)
+                            memcpy(zbd->headerBuffer + zbd->lhSize, ip, iend-ip);
                         zbd->lhSize += iend-ip;
                         *dstCapacityPtr = 0;
                         return (hSize - zbd->lhSize) + ZSTDv06_blockHeaderSize;   /* remaining header bytes + next block header */

src/borg/algorithms/zstd/lib/legacy/zstd_v06.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the

src/borg/algorithms/zstd/lib/legacy/zstd_v07.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -184,7 +184,7 @@ ZSTDLIBv07_API size_t ZSTDv07_insertBlock(ZSTDv07_DCtx* dctx, const void* blockS
    low-level memory access routines
    Copyright (C) 2013-2015, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -242,7 +242,11 @@ extern "C" {
 *  Basic Types
 *****************************************************************/
 #if  !defined (__VMS) && (defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
-# include <stdint.h>
+# if defined(_AIX)
+#  include <inttypes.h>
+# else
+#  include <stdint.h> /* intptr_t */
+# endif
   typedef  uint8_t BYTE;
   typedef uint16_t U16;
   typedef  int16_t S16;
@@ -264,27 +268,6 @@ extern "C" {
 /*-**************************************************************
 *  Memory I/O
 *****************************************************************/
-/* MEM_FORCE_MEMORY_ACCESS :
- * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
- * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
- * The below switch allow to select different access method for improved performance.
- * Method 0 (default) : use `memcpy()`. Safe and portable.
- * Method 1 : `__packed` statement. It depends on compiler extension (ie, not portable).
- *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
- * Method 2 : direct access. This method is portable but violate C standard.
- *            It can generate buggy code on targets depending on alignment.
- *            In some circumstances, it's the only known way to get the most performance (ie GCC + ARMv6)
- * See http://fastcompression.blogspot.fr/2015/08/accessing-unaligned-memory.html for details.
- * Prefer these methods in priority order (0 > 1 > 2)
- */
-#ifndef MEM_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
-#  if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
-#    define MEM_FORCE_MEMORY_ACCESS 2
-#  elif (defined(__INTEL_COMPILER) && !defined(WIN32)) || \
-  (defined(__GNUC__) && ( defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) || defined(__ARM_ARCH_7S__) ))
-#    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; }
@@ -295,33 +278,6 @@ MEM_STATIC unsigned MEM_isLittleEndian(void)
     return one.c[0];
 }
 
-#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 void MEM_write16(void* memPtr, U16 value) { *(U16*)memPtr = value; }
-
-#elif defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==1)
-
-/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
-/* currently only defined for gcc and icc */
-typedef union { U16 u16; U32 u32; U64 u64; size_t st; } __attribute__((packed)) unalign;
-
-MEM_STATIC U16 MEM_read16(const void* ptr) { return ((const unalign*)ptr)->u16; }
-MEM_STATIC U32 MEM_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
-MEM_STATIC U64 MEM_read64(const void* ptr) { return ((const unalign*)ptr)->u64; }
-
-MEM_STATIC void MEM_write16(void* memPtr, U16 value) { ((unalign*)memPtr)->u16 = value; }
-
-#else
-
-/* default method, safe and standard.
-   can sometimes prove slower */
-
 MEM_STATIC U16 MEM_read16(const void* memPtr)
 {
     U16 val; memcpy(&val, memPtr, sizeof(val)); return val;
@@ -342,8 +298,6 @@ MEM_STATIC void MEM_write16(void* memPtr, U16 value)
     memcpy(memPtr, &value, sizeof(value));
 }
 
-#endif /* MEM_FORCE_MEMORY_ACCESS */
-
 MEM_STATIC U32 MEM_swap32(U32 in)
 {
 #if defined(_MSC_VER)     /* Visual Studio */
@@ -438,7 +392,7 @@ MEM_STATIC size_t MEM_readLEST(const void* memPtr)
    header file (to include)
    Copyright (C) 2013-2016, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -526,9 +480,8 @@ MEM_STATIC size_t BITv07_readBitsFast(BITv07_DStream_t* bitD, unsigned nbBits);
 MEM_STATIC unsigned BITv07_highbit32 (U32 val)
 {
 #   if defined(_MSC_VER)   /* Visual */
-    unsigned long r=0;
-    _BitScanReverse ( &r, val );
-    return (unsigned) r;
+    unsigned long r;
+    return _BitScanReverse(&r, val) ? (unsigned)r : 0;
 #   elif defined(__GNUC__) && (__GNUC__ >= 3)   /* Use GCC Intrinsic */
     return __builtin_clz (val) ^ 31;
 #   else   /* Software version */
@@ -596,7 +549,7 @@ MEM_STATIC size_t BITv07_initDStream(BITv07_DStream_t* bitD, const void* srcBuff
 }
 
 /*! BITv07_lookBitsFast() :
-*   unsafe version; only works only if nbBits >= 1 */
+*   unsafe version; only works if nbBits >= 1 */
 MEM_STATIC size_t BITv07_lookBitsFast(const BITv07_DStream_t* bitD, U32 nbBits)
 {
     U32 const bitMask = sizeof(bitD->bitContainer)*8 - 1;
@@ -616,7 +569,7 @@ MEM_STATIC size_t BITv07_readBits(BITv07_DStream_t* bitD, U32 nbBits)
 }
 
 /*! BITv07_readBitsFast() :
-*   unsafe version; only works only if nbBits >= 1 */
+*   unsafe version; only works if nbBits >= 1 */
 MEM_STATIC size_t BITv07_readBitsFast(BITv07_DStream_t* bitD, U32 nbBits)
 {
     size_t const value = BITv07_lookBitsFast(bitD, nbBits);
@@ -670,7 +623,7 @@ MEM_STATIC unsigned BITv07_endOfDStream(const BITv07_DStream_t* DStream)
    Public Prototypes declaration
    Copyright (C) 2013-2016, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -978,7 +931,7 @@ MEM_STATIC BYTE FSEv07_decodeSymbolFast(FSEv07_DState_t* DStatePtr, BITv07_DStre
    header file
    Copyright (C) 2013-2016, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -1151,7 +1104,7 @@ size_t HUFv07_decompress1X4_usingDTable(void* dst, size_t maxDstSize, const void
    Common functions of New Generation Entropy library
    Copyright (C) 2016, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -1375,7 +1328,7 @@ size_t HUFv07_readStats(BYTE* huffWeight, size_t hwSize, U32* rankStats,
    FSE : Finite State Entropy decoder
    Copyright (C) 2013-2015, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -1699,7 +1652,7 @@ size_t FSEv07_decompress(void* dst, size_t maxDstSize, const void* cSrc, size_t
    Huffman decoder, part of New Generation Entropy library
    Copyright (C) 2013-2016, Yann Collet.
 
-   BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+   BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are
@@ -2577,7 +2530,7 @@ size_t HUFv07_decompress1X_DCtx (HUFv07_DTable* dctx, void* dst, size_t dstSize,
     Common functions of Zstd compression library
     Copyright (C) 2015-2016, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are
@@ -2601,7 +2554,7 @@ size_t HUFv07_decompress1X_DCtx (HUFv07_DTable* dctx, void* dst, size_t dstSize,
     OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
     You can contact the author at :
-    - zstd homepage : http://www.zstd.net/
+    - zstd homepage : https://facebook.github.io/zstd/
 */
 
 
@@ -2647,7 +2600,7 @@ static void ZSTDv07_defaultFreeFunction(void* opaque, void* address)
     Header File for include
     Copyright (C) 2014-2016, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are
@@ -2717,7 +2670,7 @@ typedef enum { bt_compressed, bt_raw, bt_rle, bt_end } blockType_t;
 #define MIN_SEQUENCES_SIZE 1 /* nbSeq==0 */
 #define MIN_CBLOCK_SIZE (1 /*litCSize*/ + 1 /* RLE or RAW */ + MIN_SEQUENCES_SIZE /* nbSeq==0 */)   /* for a non-null block */
 
-#define HufLog 12
+#define ZSTD_HUFFDTABLE_CAPACITY_LOG 12
 typedef enum { lbt_huffman, lbt_repeat, lbt_raw, lbt_rle } litBlockType_t;
 
 #define LONGNBSEQ 0x7F00
@@ -2854,7 +2807,7 @@ static const ZSTDv07_customMem defaultCustomMem = { ZSTDv07_defaultAllocFunction
     zstd - standard compression library
     Copyright (C) 2014-2016, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are
@@ -2878,7 +2831,7 @@ static const ZSTDv07_customMem defaultCustomMem = { ZSTDv07_defaultAllocFunction
     OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
     You can contact the author at :
-    - zstd homepage : http://www.zstd.net
+    - zstd homepage : https://facebook.github.io/zstd
 */
 
 /* ***************************************************************
@@ -2931,7 +2884,7 @@ struct ZSTDv07_DCtx_s
     FSEv07_DTable LLTable[FSEv07_DTABLE_SIZE_U32(LLFSELog)];
     FSEv07_DTable OffTable[FSEv07_DTABLE_SIZE_U32(OffFSELog)];
     FSEv07_DTable MLTable[FSEv07_DTABLE_SIZE_U32(MLFSELog)];
-    HUFv07_DTable hufTable[HUFv07_DTABLE_SIZE(HufLog)];  /* can accommodate HUFv07_decompress4X */
+    HUFv07_DTable hufTable[HUFv07_DTABLE_SIZE(ZSTD_HUFFDTABLE_CAPACITY_LOG)];  /* can accommodate HUFv07_decompress4X */
     const void* previousDstEnd;
     const void* base;
     const void* vBase;
@@ -2967,7 +2920,7 @@ size_t ZSTDv07_decompressBegin(ZSTDv07_DCtx* dctx)
     dctx->base = NULL;
     dctx->vBase = NULL;
     dctx->dictEnd = NULL;
-    dctx->hufTable[0] = (HUFv07_DTable)((HufLog)*0x1000001);
+    dctx->hufTable[0] = (HUFv07_DTable)((ZSTD_HUFFDTABLE_CAPACITY_LOG)*0x1000001);
     dctx->litEntropy = dctx->fseEntropy = 0;
     dctx->dictID = 0;
     { int i; for (i=0; i<ZSTDv07_REP_NUM; i++) dctx->rep[i] = repStartValue[i]; }
@@ -3254,7 +3207,7 @@ typedef struct
 *   Provides the size of compressed block from block header `src` */
 static size_t ZSTDv07_getcBlockSize(const void* src, size_t srcSize, blockProperties_t* bpPtr)
 {
-    const BYTE* const in = (const BYTE* const)src;
+    const BYTE* const in = (const BYTE*)src;
     U32 cSize;
 
     if (srcSize < ZSTDv07_blockHeaderSize) return ERROR(srcSize_wrong);
@@ -3449,7 +3402,7 @@ static size_t ZSTDv07_decodeSeqHeaders(int* nbSeqPtr,
                              FSEv07_DTable* DTableLL, FSEv07_DTable* DTableML, FSEv07_DTable* DTableOffb, U32 flagRepeatTable,
                              const void* src, size_t srcSize)
 {
-    const BYTE* const istart = (const BYTE* const)src;
+    const BYTE* const istart = (const BYTE*)src;
     const BYTE* const iend = istart + srcSize;
     const BYTE* ip = istart;
 
@@ -3599,11 +3552,14 @@ size_t ZSTDv07_execSequence(BYTE* op,
     const BYTE* match = oLitEnd - sequence.offset;
 
     /* check */
-    if ((oLitEnd>oend_w) | (oMatchEnd>oend)) return ERROR(dstSize_tooSmall); /* last match must start at a minimum distance of WILDCOPY_OVERLENGTH from oend */
-    if (iLitEnd > litLimit) return ERROR(corruption_detected);   /* over-read beyond lit buffer */
+    assert(oend >= op);
+    if (sequence.litLength + WILDCOPY_OVERLENGTH > (size_t)(oend - op)) return ERROR(dstSize_tooSmall);
+    if (sequenceLength > (size_t)(oend - op)) return ERROR(dstSize_tooSmall);
+    assert(litLimit >= *litPtr);
+    if (sequence.litLength > (size_t)(litLimit - *litPtr)) return ERROR(corruption_detected);;
 
     /* copy Literals */
-    ZSTDv07_wildcopy(op, *litPtr, sequence.litLength);   /* note : since oLitEnd <= oend-WILDCOPY_OVERLENGTH, no risk of overwrite beyond oend */
+    ZSTDv07_wildcopy(op, *litPtr, (ptrdiff_t)sequence.litLength);   /* note : since oLitEnd <= oend-WILDCOPY_OVERLENGTH, no risk of overwrite beyond oend */
     op = oLitEnd;
     *litPtr = iLitEnd;   /* update for next sequence */
 
@@ -3617,7 +3573,7 @@ size_t ZSTDv07_execSequence(BYTE* op,
             return sequenceLength;
         }
         /* span extDict & currentPrefixSegment */
-        {   size_t const length1 = dictEnd - match;
+        {   size_t const length1 = (size_t)(dictEnd - match);
             memmove(oLitEnd, match, length1);
             op = oLitEnd + length1;
             sequence.matchLength -= length1;
@@ -3668,7 +3624,7 @@ static size_t ZSTDv07_decompressSequences(
 {
     const BYTE* ip = (const BYTE*)seqStart;
     const BYTE* const iend = ip + seqSize;
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     BYTE* const oend = ostart + maxDstSize;
     BYTE* op = ostart;
     const BYTE* litPtr = dctx->litPtr;
@@ -3795,7 +3751,7 @@ static size_t ZSTDv07_decompressFrame(ZSTDv07_DCtx* dctx,
 {
     const BYTE* ip = (const BYTE*)src;
     const BYTE* const iend = ip + srcSize;
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     BYTE* const oend = ostart + dstCapacity;
     BYTE* op = ostart;
     size_t remainingSize = srcSize;
@@ -4253,7 +4209,7 @@ ZSTDLIBv07_API size_t ZSTDv07_decompress_usingDDict(ZSTDv07_DCtx* dctx,
     Buffered version of Zstd compression library
     Copyright (C) 2015-2016, Yann Collet.
 
-    BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
+    BSD 2-Clause License (https://opensource.org/licenses/bsd-license.php)
 
     Redistribution and use in source and binary forms, with or without
     modification, are permitted provided that the following conditions are
@@ -4277,7 +4233,7 @@ ZSTDLIBv07_API size_t ZSTDv07_decompress_usingDDict(ZSTDv07_DCtx* dctx,
     OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
     You can contact the author at :
-    - zstd homepage : http://www.zstd.net/
+    - zstd homepage : https://facebook.github.io/zstd/
 */
 
 
@@ -4417,7 +4373,8 @@ size_t ZBUFFv07_decompressContinue(ZBUFFv07_DCtx* zbd,
                 if (hSize != 0) {
                     size_t const toLoad = hSize - zbd->lhSize;   /* if hSize!=0, hSize > zbd->lhSize */
                     if (toLoad > (size_t)(iend-ip)) {   /* not enough input to load full header */
-                        memcpy(zbd->headerBuffer + zbd->lhSize, ip, iend-ip);
+                        if (ip != NULL)
+                            memcpy(zbd->headerBuffer + zbd->lhSize, ip, iend-ip);
                         zbd->lhSize += iend-ip;
                         *dstCapacityPtr = 0;
                         return (hSize - zbd->lhSize) + ZSTDv07_blockHeaderSize;   /* remaining header bytes + next block header */

src/borg/algorithms/zstd/lib/legacy/zstd_v07.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the

src/borg/algorithms/zstd/lib/dictBuilder/zdict.h → src/borg/algorithms/zstd/lib/zdict.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -8,34 +8,184 @@
  * You may select, at your option, one of the above-listed licenses.
  */
 
-#ifndef DICTBUILDER_H_001
-#define DICTBUILDER_H_001
-
 #if defined (__cplusplus)
 extern "C" {
 #endif
 
+#ifndef ZSTD_ZDICT_H
+#define ZSTD_ZDICT_H
 
 /*======  Dependencies  ======*/
 #include <stddef.h>  /* size_t */
 
 
 /* =====   ZDICTLIB_API : control library symbols visibility   ===== */
-#ifndef ZDICTLIB_VISIBILITY
-#  if defined(__GNUC__) && (__GNUC__ >= 4)
-#    define ZDICTLIB_VISIBILITY __attribute__ ((visibility ("default")))
+#ifndef ZDICTLIB_VISIBLE
+   /* Backwards compatibility with old macro name */
+#  ifdef ZDICTLIB_VISIBILITY
+#    define ZDICTLIB_VISIBLE ZDICTLIB_VISIBILITY
+#  elif defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
+#    define ZDICTLIB_VISIBLE __attribute__ ((visibility ("default")))
+#  else
+#    define ZDICTLIB_VISIBLE
+#  endif
+#endif
+
+#ifndef ZDICTLIB_HIDDEN
+#  if defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
+#    define ZDICTLIB_HIDDEN __attribute__ ((visibility ("hidden")))
 #  else
-#    define ZDICTLIB_VISIBILITY
+#    define ZDICTLIB_HIDDEN
 #  endif
 #endif
+
 #if defined(ZSTD_DLL_EXPORT) && (ZSTD_DLL_EXPORT==1)
-#  define ZDICTLIB_API __declspec(dllexport) ZDICTLIB_VISIBILITY
+#  define ZDICTLIB_API __declspec(dllexport) ZDICTLIB_VISIBLE
 #elif defined(ZSTD_DLL_IMPORT) && (ZSTD_DLL_IMPORT==1)
-#  define ZDICTLIB_API __declspec(dllimport) ZDICTLIB_VISIBILITY /* It isn't required but allows to generate better code, saving a function pointer load from the IAT and an indirect jump.*/
+#  define ZDICTLIB_API __declspec(dllimport) ZDICTLIB_VISIBLE /* It isn't required but allows to generate better code, saving a function pointer load from the IAT and an indirect jump.*/
 #else
-#  define ZDICTLIB_API ZDICTLIB_VISIBILITY
+#  define ZDICTLIB_API ZDICTLIB_VISIBLE
 #endif
 
+/*******************************************************************************
+ * Zstd dictionary builder
+ *
+ * FAQ
+ * ===
+ * Why should I use a dictionary?
+ * ------------------------------
+ *
+ * Zstd can use dictionaries to improve compression ratio of small data.
+ * Traditionally small files don't compress well because there is very little
+ * repetition in a single sample, since it is small. But, if you are compressing
+ * many similar files, like a bunch of JSON records that share the same
+ * structure, you can train a dictionary on ahead of time on some samples of
+ * these files. Then, zstd can use the dictionary to find repetitions that are
+ * present across samples. This can vastly improve compression ratio.
+ *
+ * When is a dictionary useful?
+ * ----------------------------
+ *
+ * Dictionaries are useful when compressing many small files that are similar.
+ * The larger a file is, the less benefit a dictionary will have. Generally,
+ * we don't expect dictionary compression to be effective past 100KB. And the
+ * smaller a file is, the more we would expect the dictionary to help.
+ *
+ * How do I use a dictionary?
+ * --------------------------
+ *
+ * Simply pass the dictionary to the zstd compressor with
+ * `ZSTD_CCtx_loadDictionary()`. The same dictionary must then be passed to
+ * the decompressor, using `ZSTD_DCtx_loadDictionary()`. There are other
+ * more advanced functions that allow selecting some options, see zstd.h for
+ * complete documentation.
+ *
+ * What is a zstd dictionary?
+ * --------------------------
+ *
+ * A zstd dictionary has two pieces: Its header, and its content. The header
+ * contains a magic number, the dictionary ID, and entropy tables. These
+ * entropy tables allow zstd to save on header costs in the compressed file,
+ * which really matters for small data. The content is just bytes, which are
+ * repeated content that is common across many samples.
+ *
+ * What is a raw content dictionary?
+ * ---------------------------------
+ *
+ * A raw content dictionary is just bytes. It doesn't have a zstd dictionary
+ * header, a dictionary ID, or entropy tables. Any buffer is a valid raw
+ * content dictionary.
+ *
+ * How do I train a dictionary?
+ * ----------------------------
+ *
+ * Gather samples from your use case. These samples should be similar to each
+ * other. If you have several use cases, you could try to train one dictionary
+ * per use case.
+ *
+ * Pass those samples to `ZDICT_trainFromBuffer()` and that will train your
+ * dictionary. There are a few advanced versions of this function, but this
+ * is a great starting point. If you want to further tune your dictionary
+ * you could try `ZDICT_optimizeTrainFromBuffer_cover()`. If that is too slow
+ * you can try `ZDICT_optimizeTrainFromBuffer_fastCover()`.
+ *
+ * If the dictionary training function fails, that is likely because you
+ * either passed too few samples, or a dictionary would not be effective
+ * for your data. Look at the messages that the dictionary trainer printed,
+ * if it doesn't say too few samples, then a dictionary would not be effective.
+ *
+ * How large should my dictionary be?
+ * ----------------------------------
+ *
+ * A reasonable dictionary size, the `dictBufferCapacity`, is about 100KB.
+ * The zstd CLI defaults to a 110KB dictionary. You likely don't need a
+ * dictionary larger than that. But, most use cases can get away with a
+ * smaller dictionary. The advanced dictionary builders can automatically
+ * shrink the dictionary for you, and select the smallest size that doesn't
+ * hurt compression ratio too much. See the `shrinkDict` parameter.
+ * A smaller dictionary can save memory, and potentially speed up
+ * compression.
+ *
+ * How many samples should I provide to the dictionary builder?
+ * ------------------------------------------------------------
+ *
+ * We generally recommend passing ~100x the size of the dictionary
+ * in samples. A few thousand should suffice. Having too few samples
+ * can hurt the dictionaries effectiveness. Having more samples will
+ * only improve the dictionaries effectiveness. But having too many
+ * samples can slow down the dictionary builder.
+ *
+ * How do I determine if a dictionary will be effective?
+ * -----------------------------------------------------
+ *
+ * Simply train a dictionary and try it out. You can use zstd's built in
+ * benchmarking tool to test the dictionary effectiveness.
+ *
+ *   # Benchmark levels 1-3 without a dictionary
+ *   zstd -b1e3 -r /path/to/my/files
+ *   # Benchmark levels 1-3 with a dictionary
+ *   zstd -b1e3 -r /path/to/my/files -D /path/to/my/dictionary
+ *
+ * When should I retrain a dictionary?
+ * -----------------------------------
+ *
+ * You should retrain a dictionary when its effectiveness drops. Dictionary
+ * effectiveness drops as the data you are compressing changes. Generally, we do
+ * expect dictionaries to "decay" over time, as your data changes, but the rate
+ * at which they decay depends on your use case. Internally, we regularly
+ * retrain dictionaries, and if the new dictionary performs significantly
+ * better than the old dictionary, we will ship the new dictionary.
+ *
+ * I have a raw content dictionary, how do I turn it into a zstd dictionary?
+ * -------------------------------------------------------------------------
+ *
+ * If you have a raw content dictionary, e.g. by manually constructing it, or
+ * using a third-party dictionary builder, you can turn it into a zstd
+ * dictionary by using `ZDICT_finalizeDictionary()`. You'll also have to
+ * provide some samples of the data. It will add the zstd header to the
+ * raw content, which contains a dictionary ID and entropy tables, which
+ * will improve compression ratio, and allow zstd to write the dictionary ID
+ * into the frame, if you so choose.
+ *
+ * Do I have to use zstd's dictionary builder?
+ * -------------------------------------------
+ *
+ * No! You can construct dictionary content however you please, it is just
+ * bytes. It will always be valid as a raw content dictionary. If you want
+ * a zstd dictionary, which can improve compression ratio, use
+ * `ZDICT_finalizeDictionary()`.
+ *
+ * What is the attack surface of a zstd dictionary?
+ * ------------------------------------------------
+ *
+ * Zstd is heavily fuzz tested, including loading fuzzed dictionaries, so
+ * zstd should never crash, or access out-of-bounds memory no matter what
+ * the dictionary is. However, if an attacker can control the dictionary
+ * during decompression, they can cause zstd to generate arbitrary bytes,
+ * just like if they controlled the compressed data.
+ *
+ ******************************************************************************/
+
 
 /*! ZDICT_trainFromBuffer():
  *  Train a dictionary from an array of samples.
@@ -62,9 +212,16 @@ ZDICTLIB_API size_t ZDICT_trainFromBuffer(void* dictBuffer, size_t dictBufferCap
                                     const size_t* samplesSizes, unsigned nbSamples);
 
 typedef struct {
-    int      compressionLevel;   /*< optimize for a specific zstd compression level; 0 means default */
-    unsigned notificationLevel;  /*< Write log to stderr; 0 = none (default); 1 = errors; 2 = progression; 3 = details; 4 = debug; */
-    unsigned dictID;             /*< force dictID value; 0 means auto mode (32-bits random value) */
+    int      compressionLevel;   /**< optimize for a specific zstd compression level; 0 means default */
+    unsigned notificationLevel;  /**< Write log to stderr; 0 = none (default); 1 = errors; 2 = progression; 3 = details; 4 = debug; */
+    unsigned dictID;             /**< force dictID value; 0 means auto mode (32-bits random value)
+                                  *   NOTE: The zstd format reserves some dictionary IDs for future use.
+                                  *         You may use them in private settings, but be warned that they
+                                  *         may be used by zstd in a public dictionary registry in the future.
+                                  *         These dictionary IDs are:
+                                  *           - low range  : <= 32767
+                                  *           - high range : >= (2^31)
+                                  */
 } ZDICT_params_t;
 
 /*! ZDICT_finalizeDictionary():
@@ -91,7 +248,6 @@ typedef struct {
  * is presumed that the most profitable content is at the end of the dictionary,
  * since that is the cheapest to reference.
  *
- * `dictContentSize` must be >= ZDICT_CONTENTSIZE_MIN bytes.
  * `maxDictSize` must be >= max(dictContentSize, ZSTD_DICTSIZE_MIN).
  *
  * @return: size of dictionary stored into `dstDictBuffer` (<= `maxDictSize`),
@@ -115,9 +271,21 @@ ZDICTLIB_API size_t ZDICT_getDictHeaderSize(const void* dictBuffer, size_t dictS
 ZDICTLIB_API unsigned ZDICT_isError(size_t errorCode);
 ZDICTLIB_API const char* ZDICT_getErrorName(size_t errorCode);
 
+#endif   /* ZSTD_ZDICT_H */
 
+#if defined(ZDICT_STATIC_LINKING_ONLY) && !defined(ZSTD_ZDICT_H_STATIC)
+#define ZSTD_ZDICT_H_STATIC
 
-#ifdef ZDICT_STATIC_LINKING_ONLY
+/* This can be overridden externally to hide static symbols. */
+#ifndef ZDICTLIB_STATIC_API
+#  if defined(ZSTD_DLL_EXPORT) && (ZSTD_DLL_EXPORT==1)
+#    define ZDICTLIB_STATIC_API __declspec(dllexport) ZDICTLIB_VISIBLE
+#  elif defined(ZSTD_DLL_IMPORT) && (ZSTD_DLL_IMPORT==1)
+#    define ZDICTLIB_STATIC_API __declspec(dllimport) ZDICTLIB_VISIBLE
+#  else
+#    define ZDICTLIB_STATIC_API ZDICTLIB_VISIBLE
+#  endif
+#endif
 
 /* ====================================================================================
  * The definitions in this section are considered experimental.
@@ -126,8 +294,9 @@ ZDICTLIB_API const char* ZDICT_getErrorName(size_t errorCode);
  * Use them only in association with static linking.
  * ==================================================================================== */
 
-#define ZDICT_CONTENTSIZE_MIN 128
 #define ZDICT_DICTSIZE_MIN    256
+/* Deprecated: Remove in v1.6.0 */
+#define ZDICT_CONTENTSIZE_MIN 128
 
 /*! ZDICT_cover_params_t:
  *  k and d are the only required parameters.
@@ -172,7 +341,7 @@ typedef struct {
  *        In general, it's recommended to provide a few thousands samples, though this can vary a lot.
  *        It's recommended that total size of all samples be about ~x100 times the target size of dictionary.
  */
-ZDICTLIB_API size_t ZDICT_trainFromBuffer_cover(
+ZDICTLIB_STATIC_API size_t ZDICT_trainFromBuffer_cover(
           void *dictBuffer, size_t dictBufferCapacity,
     const void *samplesBuffer, const size_t *samplesSizes, unsigned nbSamples,
           ZDICT_cover_params_t parameters);
@@ -194,7 +363,7 @@ ZDICTLIB_API size_t ZDICT_trainFromBuffer_cover(
  *          See ZDICT_trainFromBuffer() for details on failure modes.
  * Note: ZDICT_optimizeTrainFromBuffer_cover() requires about 8 bytes of memory for each input byte and additionally another 5 bytes of memory for each byte of memory for each thread.
  */
-ZDICTLIB_API size_t ZDICT_optimizeTrainFromBuffer_cover(
+ZDICTLIB_STATIC_API size_t ZDICT_optimizeTrainFromBuffer_cover(
           void* dictBuffer, size_t dictBufferCapacity,
     const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
           ZDICT_cover_params_t* parameters);
@@ -215,7 +384,7 @@ ZDICTLIB_API size_t ZDICT_optimizeTrainFromBuffer_cover(
  *        In general, it's recommended to provide a few thousands samples, though this can vary a lot.
  *        It's recommended that total size of all samples be about ~x100 times the target size of dictionary.
  */
-ZDICTLIB_API size_t ZDICT_trainFromBuffer_fastCover(void *dictBuffer,
+ZDICTLIB_STATIC_API size_t ZDICT_trainFromBuffer_fastCover(void *dictBuffer,
                     size_t dictBufferCapacity, const void *samplesBuffer,
                     const size_t *samplesSizes, unsigned nbSamples,
                     ZDICT_fastCover_params_t parameters);
@@ -238,7 +407,7 @@ ZDICTLIB_API size_t ZDICT_trainFromBuffer_fastCover(void *dictBuffer,
  *          See ZDICT_trainFromBuffer() for details on failure modes.
  * Note: ZDICT_optimizeTrainFromBuffer_fastCover() requires about 6 * 2^f bytes of memory for each thread.
  */
-ZDICTLIB_API size_t ZDICT_optimizeTrainFromBuffer_fastCover(void* dictBuffer,
+ZDICTLIB_STATIC_API size_t ZDICT_optimizeTrainFromBuffer_fastCover(void* dictBuffer,
                     size_t dictBufferCapacity, const void* samplesBuffer,
                     const size_t* samplesSizes, unsigned nbSamples,
                     ZDICT_fastCover_params_t* parameters);
@@ -263,43 +432,43 @@ typedef struct {
  *        It's recommended that total size of all samples be about ~x100 times the target size of dictionary.
  *  Note: ZDICT_trainFromBuffer_legacy() will send notifications into stderr if instructed to, using notificationLevel>0.
  */
-ZDICTLIB_API size_t ZDICT_trainFromBuffer_legacy(
-    void *dictBuffer, size_t dictBufferCapacity,
-    const void *samplesBuffer, const size_t *samplesSizes, unsigned nbSamples,
+ZDICTLIB_STATIC_API size_t ZDICT_trainFromBuffer_legacy(
+    void* dictBuffer, size_t dictBufferCapacity,
+    const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
     ZDICT_legacy_params_t parameters);
 
+
 /* Deprecation warnings */
 /* It is generally possible to disable deprecation warnings from compiler,
    for example with -Wno-deprecated-declarations for gcc
    or _CRT_SECURE_NO_WARNINGS in Visual.
    Otherwise, it's also possible to manually define ZDICT_DISABLE_DEPRECATE_WARNINGS */
 #ifdef ZDICT_DISABLE_DEPRECATE_WARNINGS
-#  define ZDICT_DEPRECATED(message) ZDICTLIB_API   /* disable deprecation warnings */
+#  define ZDICT_DEPRECATED(message) /* disable deprecation warnings */
 #else
 #  define ZDICT_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
 #  if defined (__cplusplus) && (__cplusplus >= 201402) /* C++14 or greater */
-#    define ZDICT_DEPRECATED(message) [[deprecated(message)]] ZDICTLIB_API
-#  elif (ZDICT_GCC_VERSION >= 405) || defined(__clang__)
-#    define ZDICT_DEPRECATED(message) ZDICTLIB_API __attribute__((deprecated(message)))
+#    define ZDICT_DEPRECATED(message) [[deprecated(message)]]
+#  elif defined(__clang__) || (ZDICT_GCC_VERSION >= 405)
+#    define ZDICT_DEPRECATED(message) __attribute__((deprecated(message)))
 #  elif (ZDICT_GCC_VERSION >= 301)
-#    define ZDICT_DEPRECATED(message) ZDICTLIB_API __attribute__((deprecated))
+#    define ZDICT_DEPRECATED(message) __attribute__((deprecated))
 #  elif defined(_MSC_VER)
-#    define ZDICT_DEPRECATED(message) ZDICTLIB_API __declspec(deprecated(message))
+#    define ZDICT_DEPRECATED(message) __declspec(deprecated(message))
 #  else
 #    pragma message("WARNING: You need to implement ZDICT_DEPRECATED for this compiler")
-#    define ZDICT_DEPRECATED(message) ZDICTLIB_API
+#    define ZDICT_DEPRECATED(message)
 #  endif
 #endif /* ZDICT_DISABLE_DEPRECATE_WARNINGS */
 
 ZDICT_DEPRECATED("use ZDICT_finalizeDictionary() instead")
+ZDICTLIB_STATIC_API
 size_t ZDICT_addEntropyTablesFromBuffer(void* dictBuffer, size_t dictContentSize, size_t dictBufferCapacity,
                                   const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples);
 
 
-#endif   /* ZDICT_STATIC_LINKING_ONLY */
+#endif   /* ZSTD_ZDICT_H_STATIC */
 
 #if defined (__cplusplus)
 }
 #endif
-
-#endif   /* DICTBUILDER_H_001 */

src/borg/algorithms/zstd/lib/common/zstd_errors.h → src/borg/algorithms/zstd/lib/zstd_errors.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -20,19 +20,31 @@ extern "C" {
 
 
 /* =====   ZSTDERRORLIB_API : control library symbols visibility   ===== */
-#ifndef ZSTDERRORLIB_VISIBILITY
-#  if defined(__GNUC__) && (__GNUC__ >= 4)
-#    define ZSTDERRORLIB_VISIBILITY __attribute__ ((visibility ("default")))
+#ifndef ZSTDERRORLIB_VISIBLE
+   /* Backwards compatibility with old macro name */
+#  ifdef ZSTDERRORLIB_VISIBILITY
+#    define ZSTDERRORLIB_VISIBLE ZSTDERRORLIB_VISIBILITY
+#  elif defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
+#    define ZSTDERRORLIB_VISIBLE __attribute__ ((visibility ("default")))
 #  else
-#    define ZSTDERRORLIB_VISIBILITY
+#    define ZSTDERRORLIB_VISIBLE
 #  endif
 #endif
+
+#ifndef ZSTDERRORLIB_HIDDEN
+#  if defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
+#    define ZSTDERRORLIB_HIDDEN __attribute__ ((visibility ("hidden")))
+#  else
+#    define ZSTDERRORLIB_HIDDEN
+#  endif
+#endif
+
 #if defined(ZSTD_DLL_EXPORT) && (ZSTD_DLL_EXPORT==1)
-#  define ZSTDERRORLIB_API __declspec(dllexport) ZSTDERRORLIB_VISIBILITY
+#  define ZSTDERRORLIB_API __declspec(dllexport) ZSTDERRORLIB_VISIBLE
 #elif defined(ZSTD_DLL_IMPORT) && (ZSTD_DLL_IMPORT==1)
-#  define ZSTDERRORLIB_API __declspec(dllimport) ZSTDERRORLIB_VISIBILITY /* It isn't required but allows to generate better code, saving a function pointer load from the IAT and an indirect jump.*/
+#  define ZSTDERRORLIB_API __declspec(dllimport) ZSTDERRORLIB_VISIBLE /* It isn't required but allows to generate better code, saving a function pointer load from the IAT and an indirect jump.*/
 #else
-#  define ZSTDERRORLIB_API ZSTDERRORLIB_VISIBILITY
+#  define ZSTDERRORLIB_API ZSTDERRORLIB_VISIBLE
 #endif
 
 /*-*********************************************
@@ -58,14 +70,17 @@ typedef enum {
   ZSTD_error_frameParameter_windowTooLarge = 16,
   ZSTD_error_corruption_detected = 20,
   ZSTD_error_checksum_wrong      = 22,
+  ZSTD_error_literals_headerWrong = 24,
   ZSTD_error_dictionary_corrupted      = 30,
   ZSTD_error_dictionary_wrong          = 32,
   ZSTD_error_dictionaryCreation_failed = 34,
   ZSTD_error_parameter_unsupported   = 40,
+  ZSTD_error_parameter_combination_unsupported = 41,
   ZSTD_error_parameter_outOfBound    = 42,
   ZSTD_error_tableLog_tooLarge       = 44,
   ZSTD_error_maxSymbolValue_tooLarge = 46,
   ZSTD_error_maxSymbolValue_tooSmall = 48,
+  ZSTD_error_stabilityCondition_notRespected = 50,
   ZSTD_error_stage_wrong       = 60,
   ZSTD_error_init_missing      = 62,
   ZSTD_error_memory_allocation = 64,
@@ -73,10 +88,15 @@ typedef enum {
   ZSTD_error_dstSize_tooSmall = 70,
   ZSTD_error_srcSize_wrong    = 72,
   ZSTD_error_dstBuffer_null   = 74,
+  ZSTD_error_noForwardProgress_destFull = 80,
+  ZSTD_error_noForwardProgress_inputEmpty = 82,
   /* following error codes are __NOT STABLE__, they can be removed or changed in future versions */
   ZSTD_error_frameIndex_tooLarge = 100,
   ZSTD_error_seekableIO          = 102,
   ZSTD_error_dstBuffer_wrong     = 104,
+  ZSTD_error_srcBuffer_wrong     = 105,
+  ZSTD_error_sequenceProducer_failed = 106,
+  ZSTD_error_externalSequences_invalid = 107,
   ZSTD_error_maxCode = 120  /* never EVER use this value directly, it can change in future versions! Use ZSTD_isError() instead */
 } ZSTD_ErrorCode;