Update zstd to 1.5.4.

zstd/common/bits.h

@@ -0,0 +1,175 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_BITS_H
+#define ZSTD_BITS_H
+
+#include "mem.h"
+
+MEM_STATIC unsigned ZSTD_countTrailingZeros32_fallback(U32 val)
+{
+    assert(val != 0);
+    {
+        static const int DeBruijnBytePos[32] = {0, 1, 28, 2, 29, 14, 24, 3,
+                                                30, 22, 20, 15, 25, 17, 4, 8,
+                                                31, 27, 13, 23, 21, 19, 16, 7,
+                                                26, 12, 18, 6, 11, 5, 10, 9};
+        return DeBruijnBytePos[((U32) ((val & -(S32) val) * 0x077CB531U)) >> 27];
+    }
+}
+
+MEM_STATIC unsigned ZSTD_countTrailingZeros32(U32 val)
+{
+    assert(val != 0);
+#   if defined(_MSC_VER)
+#       if STATIC_BMI2 == 1
+            return _tzcnt_u32(val);
+#       else
+            if (val != 0) {
+                unsigned long r;
+                _BitScanForward(&r, val);
+                return (unsigned)r;
+            } else {
+                /* Should not reach this code path */
+                __assume(0);
+            }
+#       endif
+#   elif defined(__GNUC__) && (__GNUC__ >= 4)
+        return (unsigned)__builtin_ctz(val);
+#   else
+        return ZSTD_countTrailingZeros32_fallback(val);
+#   endif
+}
+
+MEM_STATIC unsigned ZSTD_countLeadingZeros32_fallback(U32 val) {
+    assert(val != 0);
+    {
+        static const U32 DeBruijnClz[32] = {0, 9, 1, 10, 13, 21, 2, 29,
+                                            11, 14, 16, 18, 22, 25, 3, 30,
+                                            8, 12, 20, 28, 15, 17, 24, 7,
+                                            19, 27, 23, 6, 26, 5, 4, 31};
+        val |= val >> 1;
+        val |= val >> 2;
+        val |= val >> 4;
+        val |= val >> 8;
+        val |= val >> 16;
+        return 31 - DeBruijnClz[(val * 0x07C4ACDDU) >> 27];
+    }
+}
+
+MEM_STATIC unsigned ZSTD_countLeadingZeros32(U32 val)
+{
+    assert(val != 0);
+#   if defined(_MSC_VER)
+#       if STATIC_BMI2 == 1
+            return _lzcnt_u32(val);
+#       else
+            if (val != 0) {
+                unsigned long r;
+                _BitScanReverse(&r, val);
+                return (unsigned)(31 - r);
+            } else {
+                /* Should not reach this code path */
+                __assume(0);
+            }
+#       endif
+#   elif defined(__GNUC__) && (__GNUC__ >= 4)
+        return (unsigned)__builtin_clz(val);
+#   else
+        return ZSTD_countLeadingZeros32_fallback(val);
+#   endif
+}
+
+MEM_STATIC unsigned ZSTD_countTrailingZeros64(U64 val)
+{
+    assert(val != 0);
+#   if defined(_MSC_VER) && defined(_WIN64)
+#       if STATIC_BMI2 == 1
+            return _tzcnt_u64(val);
+#       else
+            if (val != 0) {
+                unsigned long r;
+                _BitScanForward64(&r, val);
+                return (unsigned)r;
+            } else {
+                /* Should not reach this code path */
+                __assume(0);
+            }
+#       endif
+#   elif defined(__GNUC__) && (__GNUC__ >= 4) && defined(__LP64__)
+        return (unsigned)__builtin_ctzll(val);
+#   else
+        {
+            U32 mostSignificantWord = (U32)(val >> 32);
+            U32 leastSignificantWord = (U32)val;
+            if (leastSignificantWord == 0) {
+                return 32 + ZSTD_countTrailingZeros32(mostSignificantWord);
+            } else {
+                return ZSTD_countTrailingZeros32(leastSignificantWord);
+            }
+        }
+#   endif
+}
+
+MEM_STATIC unsigned ZSTD_countLeadingZeros64(U64 val)
+{
+    assert(val != 0);
+#   if defined(_MSC_VER) && defined(_WIN64)
+#       if STATIC_BMI2 == 1
+            return _lzcnt_u64(val);
+#       else
+            if (val != 0) {
+                unsigned long r;
+                _BitScanReverse64(&r, val);
+                return (unsigned)(63 - r);
+            } else {
+                /* Should not reach this code path */
+                __assume(0);
+            }
+#       endif
+#   elif defined(__GNUC__) && (__GNUC__ >= 4)
+        return (unsigned)(__builtin_clzll(val));
+#   else
+        {
+            U32 mostSignificantWord = (U32)(val >> 32);
+            U32 leastSignificantWord = (U32)val;
+            if (mostSignificantWord == 0) {
+                return 32 + ZSTD_countLeadingZeros32(leastSignificantWord);
+            } else {
+                return ZSTD_countLeadingZeros32(mostSignificantWord);
+            }
+        }
+#   endif
+}
+
+MEM_STATIC unsigned ZSTD_NbCommonBytes(size_t val)
+{
+    if (MEM_isLittleEndian()) {
+        if (MEM_64bits()) {
+            return ZSTD_countTrailingZeros64((U64)val) >> 3;
+        } else {
+            return ZSTD_countTrailingZeros32((U32)val) >> 3;
+        }
+    } else {  /* Big Endian CPU */
+        if (MEM_64bits()) {
+            return ZSTD_countLeadingZeros64((U64)val) >> 3;
+        } else {
+            return ZSTD_countLeadingZeros32((U32)val) >> 3;
+        }
+    }
+}
+
+MEM_STATIC unsigned ZSTD_highbit32(U32 val)   /* compress, dictBuilder, decodeCorpus */
+{
+    assert(val != 0);
+    return 31 - ZSTD_countLeadingZeros32(val);
+}
+
+#endif /* ZSTD_BITS_H */

zstd/common/bitstream.h

@@ -1,7 +1,7 @@
 /* ******************************************************************
  * bitstream
  * Part of FSE library
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -30,14 +30,15 @@ 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
 =========================================*/
 #ifndef ZSTD_NO_INTRINSICS
-#  if defined(__BMI__) && defined(__GNUC__)
+#  if (defined(__BMI__) || defined(__BMI2__)) && defined(__GNUC__)
-#    include <immintrin.h>   /* support for bextr (experimental) */
+#    include <immintrin.h>   /* support for bextr (experimental)/bzhi */
 #  elif defined(__ICCARM__)
 #    include <intrinsics.h>
 #  endif
@@ -132,48 +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 */
-#       if STATIC_BMI2 == 1
-            return _lzcnt_u32(val) ^ 31;
-#       else
-            if (val != 0) {
-                unsigned long r;
-                _BitScanReverse(&r, val);
-                return (unsigned)r;
-            } else {
-                /* Should not reach this code path */
-                __assume(0);
-            }
-#       endif
-#   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,
@@ -203,6 +162,16 @@ 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 ! */
@@ -212,7 +181,7 @@ MEM_STATIC void BIT_addBits(BIT_CStream_t* bitC,
     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;
 }
 
@@ -291,7 +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;
@@ -319,7 +288,7 @@ MEM_STATIC size_t BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, si
         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;
@@ -350,16 +319,6 @@ MEM_STATIC FORCE_INLINE_ATTR size_t BIT_getMiddleBits(size_t bitContainer, U32 c
 #endif
 }
 
-MEM_STATIC FORCE_INLINE_ATTR size_t BIT_getLowerBits(size_t bitContainer, U32 const nbBits)
-{
-#if defined(STATIC_BMI2) && STATIC_BMI2 == 1
-	return  _bzhi_u64(bitContainer, nbBits);
-#else
-    assert(nbBits < BIT_MASK_SIZE);
-    return bitContainer & BIT_mask[nbBits];
-#endif
-}
-
 /*! BIT_lookBits() :
  *  Provides next n bits from local register.
  *  local register is not modified.
@@ -406,7 +365,7 @@ MEM_STATIC FORCE_INLINE_ATTR size_t BIT_readBits(BIT_DStream_t* bitD, unsigned n
 }
 
 /*! 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);

zstd/common/compiler.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -165,6 +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 */
@@ -181,6 +187,8 @@
 #    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
 
@@ -273,7 +281,18 @@
 *  Sanitizer
 *****************************************************************/
 
-#if ZSTD_MEMORY_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... */
@@ -294,7 +313,7 @@ void __msan_poison(const volatile void *a, size_t size);
 intptr_t __msan_test_shadow(const volatile void *x, size_t size);
 #endif
 
-#if ZSTD_ADDRESS_SANITIZER
+#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... */

zstd/common/cpu.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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

zstd/common/debug.c

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

zstd/common/debug.h

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

zstd/common/entropy_common.c

@@ -1,6 +1,6 @@
 /* ******************************************************************
  * Common functions of New Generation Entropy library
- * Copyright (c) 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   ===*/
@@ -38,34 +38,6 @@ const char* HUF_getErrorName(size_t code) { return ERR_getErrorName(code); }
 /*-**************************************************************
 *  FSE NCount encoding-decoding
 ****************************************************************/
-static U32 FSE_ctz(U32 val)
-{
-    assert(val != 0);
-    {
-#   if defined(_MSC_VER)   /* Visual */
-        if (val != 0) {
-            unsigned long r;
-            _BitScanForward(&r, val);
-            return (unsigned)r;
-        } else {
-            /* Should not reach this code path */
-            __assume(0);
-        }
-#   elif defined(__GNUC__) && (__GNUC__ >= 3)   /* GCC Intrinsic */
-        return __builtin_ctz(val);
-#   elif defined(__ICCARM__)    /* IAR Intrinsic */
-        return __CTZ(val);
-#   else   /* Software version */
-        U32 count = 0;
-        while ((val & 1) == 0) {
-            val >>= 1;
-            ++count;
-        }
-        return count;
-#   endif
-    }
-}
-
 FORCE_INLINE_TEMPLATE
 size_t FSE_readNCount_body(short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
                            const void* headerBuffer, size_t hbSize)
@@ -113,7 +85,7 @@ size_t FSE_readNCount_body(short* normalizedCounter, unsigned* maxSVPtr, unsigne
              * repeat.
              * Avoid UB by setting the high bit to 1.
              */
-            int repeats = FSE_ctz(~bitStream | 0x80000000) >> 1;
+            int repeats = ZSTD_countTrailingZeros32(~bitStream | 0x80000000) >> 1;
             while (repeats >= 12) {
                 charnum += 3 * 12;
                 if (LIKELY(ip <= iend-7)) {
@@ -124,7 +96,7 @@ size_t FSE_readNCount_body(short* normalizedCounter, unsigned* maxSVPtr, unsigne
                     ip = iend - 4;
                 }
                 bitStream = MEM_readLE32(ip) >> bitCount;
-                repeats = FSE_ctz(~bitStream | 0x80000000) >> 1;
+                repeats = ZSTD_countTrailingZeros32(~bitStream | 0x80000000) >> 1;
             }
             charnum += 3 * repeats;
             bitStream >>= 2 * repeats;
@@ -189,7 +161,7 @@ size_t FSE_readNCount_body(short* normalizedCounter, unsigned* maxSVPtr, unsigne
                  * know that threshold > 1.
                  */
                 if (remaining <= 1) break;
-                nbBits = BIT_highbit32(remaining) + 1;
+                nbBits = ZSTD_highbit32(remaining) + 1;
                 threshold = 1 << (nbBits - 1);
             }
             if (charnum >= maxSV1) break;
@@ -264,7 +236,7 @@ size_t HUF_readStats(BYTE* huffWeight, size_t hwSize, U32* rankStats,
                      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), /* bmi2 */ 0);
+    return HUF_readStats_wksp(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, wksp, sizeof(wksp), /* flags */ 0);
 }
 
 FORCE_INLINE_TEMPLATE size_t
@@ -312,14 +284,14 @@ HUF_readStats_body(BYTE* huffWeight, size_t hwSize, U32* rankStats,
     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 verif = 1 << ZSTD_highbit32(rest);
-            U32 const lastWeight = BIT_highbit32(rest) + 1;
+            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]++;
@@ -356,13 +328,13 @@ 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 bmi2)
+                     int flags)
 {
 #if DYNAMIC_BMI2
-    if (bmi2) {
+    if (flags & HUF_flags_bmi2) {
         return HUF_readStats_body_bmi2(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize);
     }
 #endif
-    (void)bmi2;
+    (void)flags;
     return HUF_readStats_body_default(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize);
 }

zstd/common/error_private.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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";
@@ -38,17 +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;
     }

zstd/common/error_private.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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

zstd/common/fse.h

@@ -1,7 +1,7 @@
 /* ******************************************************************
  * FSE : Finite State Entropy codec
  * Public Prototypes declaration
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -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
 ******************************************/
@@ -164,8 +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().
@@ -241,23 +197,7 @@ FSE_PUBLIC_API size_t FSE_readNCount_bmi2(short* normalizedCounter,
                            unsigned* maxSymbolValuePtr, unsigned* tableLogPtr,
                            const void* rBuffer, size_t rBuffSize, int bmi2);
 
-/*! Constructor and Destructor of FSE_DTable.
-    Note that its size depends on 'tableLog' */
 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 :
@@ -320,16 +260,6 @@ 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_COMPRESS_WKSP_SIZE_U32() provides the minimum size required for `workSpace` as a table of FSE_CTable.
- */
-#define FSE_COMPRESS_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 */
 
@@ -347,19 +277,11 @@ size_t FSE_buildCTable_wksp(FSE_CTable* ct, const short* normalizedCounter, unsi
 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_raw (FSE_DTable* dt, unsigned nbBits);
+#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)
-/**< build a fake FSE_DTable, designed to read a flat distribution where each symbol uses nbBits */
-
-size_t FSE_buildDTable_rle (FSE_DTable* dt, unsigned char symbolValue);
-/**< build a fake FSE_DTable, designed to always generate the same symbolValue */
-
-#define FSE_DECOMPRESS_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) (FSE_DTABLE_SIZE_U32(maxTableLog) + 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(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize);
-/**< same as FSE_decompress(), using an externally allocated `workSpace` produced with `FSE_DECOMPRESS_WKSP_SIZE_U32(maxLog, maxSymbolValue)` */
-
 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_wksp() but with dynamic BMI2 support. Pass 1 if your CPU supports BMI2 or 0 if it doesn't. */
+/**< 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 */
@@ -555,7 +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)

zstd/common/fse_decompress.c

@@ -1,6 +1,6 @@
 /* ******************************************************************
  * FSE : Finite State Entropy decoder
- * Copyright (c) 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
@@ -24,6 +24,7 @@
 #include "error_private.h"
 #define ZSTD_DEPS_NEED_MALLOC
 #include "zstd_deps.h"
+#include "bits.h"       /* ZSTD_highbit32 */
 
 
 /* **************************************************************
@@ -55,19 +56,6 @@
 #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*)ZSTD_malloc( FSE_DTABLE_SIZE_U32(tableLog) * sizeof (U32) );
-}
-
-void FSE_freeDTable (FSE_DTable* dt)
-{
-    ZSTD_free(dt);
-}
-
 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 */
@@ -127,10 +115,10 @@ static size_t FSE_buildDTable_internal(FSE_DTable* dt, const short* normalizedCo
             }
         }
         /* Now we spread those positions across the table.
-         * The benefit of doing it in two stages is that we avoid the the
+         * 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 emperically worked best.
+         * We unroll the loop twice, since that is what empirically worked best.
          */
         {
             size_t position = 0;
@@ -166,7 +154,7 @@ static size_t FSE_buildDTable_internal(FSE_DTable* dt, const short* normalizedCo
         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);
     }   }
 
@@ -184,49 +172,6 @@ size_t FSE_buildDTable_wksp(FSE_DTable* dt, const short* normalizedCounter, unsi
 /*-*******************************************************
 *  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,
@@ -290,26 +235,6 @@ FORCE_INLINE_TEMPLATE size_t FSE_decompress_usingDTable_generic(
     return op-ostart;
 }
 
-
-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, unsigned maxLog, void* workSpace, size_t wkspSize)
-{
-    return FSE_decompress_wksp_bmi2(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize, /* bmi2 */ 0);
-}
-
 typedef struct {
     short ncount[FSE_MAX_SYMBOL_VALUE + 1];
     FSE_DTable dtable[1]; /* Dynamically sized */
@@ -342,7 +267,8 @@ FORCE_INLINE_TEMPLATE size_t FSE_decompress_wksp_body(
     }
 
     if (FSE_DECOMPRESS_WKSP_SIZE(tableLog, maxSymbolValue) > wkspSize) return ERROR(tableLog_tooLarge);
-    workSpace = wksp->dtable + FSE_DTABLE_SIZE_U32(tableLog);
+    assert(sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog) <= wkspSize);
+    workSpace = (BYTE*)workSpace + sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog);
     wkspSize -= sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog);
 
     CHECK_F( FSE_buildDTable_internal(wksp->dtable, wksp->ncount, maxSymbolValue, tableLog, workSpace, wkspSize) );
@@ -382,22 +308,4 @@ size_t FSE_decompress_wksp_bmi2(void* dst, size_t dstCapacity, const void* cSrc,
     return FSE_decompress_wksp_body_default(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize);
 }
 
-
-typedef FSE_DTable DTable_max_t[FSE_DTABLE_SIZE_U32(FSE_MAX_TABLELOG)];
-
-#ifndef ZSTD_NO_UNUSED_FUNCTIONS
-size_t FSE_buildDTable(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog) {
-    U32 wksp[FSE_BUILD_DTABLE_WKSP_SIZE_U32(FSE_TABLELOG_ABSOLUTE_MAX, FSE_MAX_SYMBOL_VALUE)];
-    return FSE_buildDTable_wksp(dt, normalizedCounter, maxSymbolValue, tableLog, wksp, sizeof(wksp));
-}
-
-size_t FSE_decompress(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize)
-{
-    /* Static analyzer seems unable to understand this table will be properly initialized later */
-    U32 wksp[FSE_DECOMPRESS_WKSP_SIZE_U32(FSE_MAX_TABLELOG, FSE_MAX_SYMBOL_VALUE)];
-    return FSE_decompress_wksp(dst, dstCapacity, cSrc, cSrcSize, FSE_MAX_TABLELOG, wksp, sizeof(wksp));
-}
-#endif
-
-
 #endif   /* FSE_COMMONDEFS_ONLY */

zstd/common/huf.h

@@ -1,7 +1,7 @@
 /* ******************************************************************
  * huff0 huffman codec,
  * part of Finite State Entropy library
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -21,100 +21,23 @@ extern "C" {
 
 /* *** Dependencies *** */
 #include "zstd_deps.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);
-
-
-/* ***   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) */
-
-/* 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 be at least as large as HUF_WORKSPACE_SIZE */
-#define HUF_WORKSPACE_SIZE ((8 << 10) + 512 /* sorting scratch space */)
-#define HUF_WORKSPACE_SIZE_U64 (HUF_WORKSPACE_SIZE / sizeof(U64))
-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);
-
-#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 FSE_STATIC_LINKING_ONLY
 #include "fse.h"
 
 
+/* ***   Tool functions *** */
+#define HUF_BLOCKSIZE_MAX (128 * 1024)   /**< maximum input size for a single block compressed with HUF_compress */
+size_t HUF_compressBound(size_t size);   /**< maximum compressed size (worst case) */
+
+/* Error Management */
+unsigned    HUF_isError(size_t code);       /**< tells if a return value is an error code */
+const char* HUF_getErrorName(size_t code);  /**< provides error code string (useful for debugging) */
+
+
+#define HUF_WORKSPACE_SIZE ((8 << 10) + 512 /* sorting scratch space */)
+#define HUF_WORKSPACE_SIZE_U64 (HUF_WORKSPACE_SIZE / sizeof(U64))
+
 /* *** Constants *** */
 #define HUF_TABLELOG_MAX      12      /* max runtime value of tableLog (due to static allocation); can be modified up to HUF_TABLELOG_ABSOLUTEMAX */
 #define HUF_TABLELOG_DEFAULT  11      /* default tableLog value when none specified */
@@ -154,25 +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 */
+ * Huffman flags bitset.
-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 */
+ * For all flags, 0 is the default value.
-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 */
+typedef enum {
-#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 */
+     * If compiled with DYNAMIC_BMI2: Set flag only if the CPU supports BMI2 at runtime.
-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 */
+     * Otherwise: Ignored.
-#endif
+     */
+    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")
@@ -185,12 +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);
+unsigned HUF_minTableLog(unsigned symbolCardinality);
-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 */
+unsigned HUF_cardinality(const unsigned* count, unsigned maxSymbolValue);
-size_t HUF_writeCTable (void* dst, size_t maxDstSize, const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog);
+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);
+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_compress4X_usingCTable_bmi2(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int bmi2);
 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);
 
@@ -199,6 +146,7 @@ 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.
@@ -209,13 +157,13 @@ 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, unsigned suspectUncompressible);
+                       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,
@@ -241,7 +189,7 @@ 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 bmi2);
+                          int flags);
 
 /** HUF_readCTable() :
  *  Loading a CTable saved with HUF_writeCTable() */
@@ -279,32 +227,12 @@ U32 HUF_selectDecoder (size_t dstSize, size_t cSrcSize);
 #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_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int flags);
-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_U64 U64 */
-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_bmi2(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int bmi2);
 /** 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.
@@ -315,49 +243,30 @@ 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, unsigned suspectUncompressible);
+                       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);
-#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 */
-#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 */
-#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);
-#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);
 #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);
+size_t HUF_decompress1X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int flags);
 #ifndef HUF_FORCE_DECOMPRESS_X2
-size_t HUF_decompress1X1_DCtx_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int bmi2);
+size_t HUF_decompress1X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags);
 #endif
-size_t HUF_decompress4X_usingDTable_bmi2(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int bmi2);
+size_t HUF_decompress4X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int flags);
-size_t HUF_decompress4X_hufOnly_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int bmi2);
+size_t HUF_decompress4X_hufOnly_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags);
 #ifndef HUF_FORCE_DECOMPRESS_X2
-size_t HUF_readDTableX1_wksp_bmi2(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize, int bmi2);
+size_t HUF_readDTableX1_wksp(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize, int flags);
 #endif
 #ifndef HUF_FORCE_DECOMPRESS_X1
-size_t HUF_readDTableX2_wksp_bmi2(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize, int bmi2);
+size_t HUF_readDTableX2_wksp(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize, int flags);
 #endif
 
-#endif /* HUF_STATIC_LINKING_ONLY */
+#endif   /* HUF_H_298734234 */
 
 #if defined (__cplusplus)
 }

zstd/common/mem.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -133,21 +133,15 @@ MEM_STATIC size_t MEM_swapST(size_t in);
 /*-**************************************************************
 *  Memory I/O Implementation
 *****************************************************************/
-/* MEM_FORCE_MEMORY_ACCESS :
+/* MEM_FORCE_MEMORY_ACCESS : For accessing unaligned memory:
- * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
+ * Method 0 : always use `memcpy()`. Safe and portable.
- * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
+ * Method 1 : Use compiler extension to set unaligned access.
- * 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`.
  * 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)
+ * Default  : method 1 if supported, else method 0
- * 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(__INTEL_COMPILER) || defined(__GNUC__) || defined(__ICCARM__)
+#  ifdef __GNUC__
 #    define MEM_FORCE_MEMORY_ACCESS 1
 #  endif
 #endif
@@ -190,30 +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 */
+typedef __attribute__((aligned(1))) U16 unalign16;
-/* currently only defined for gcc and icc */
+typedef __attribute__((aligned(1))) U32 unalign32;
-#if defined(_MSC_VER) || (defined(__INTEL_COMPILER) && defined(WIN32))
+typedef __attribute__((aligned(1))) U64 unalign64;
-    __pragma( pack(push, 1) )
+typedef __attribute__((aligned(1))) size_t unalignArch;
-    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
 
-MEM_STATIC U16 MEM_read16(const void* ptr) { return ((const unalign16*)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)->v; }
+MEM_STATIC U32 MEM_read32(const void* ptr) { return *(const unalign32*)ptr; }
-MEM_STATIC U64 MEM_read64(const void* ptr) { return ((const unalign64*)ptr)->v; }
+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)->v; }
+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_write16(void* memPtr, U16 value) { *(unalign16*)memPtr = value; }
-MEM_STATIC void MEM_write32(void* memPtr, U32 value) { ((unalign32*)memPtr)->v = value; }
+MEM_STATIC void MEM_write32(void* memPtr, U32 value) { *(unalign32*)memPtr = value; }
-MEM_STATIC void MEM_write64(void* memPtr, U64 value) { ((unalign64*)memPtr)->v = value; }
+MEM_STATIC void MEM_write64(void* memPtr, U64 value) { *(unalign64*)memPtr = value; }
 
 #else
 
@@ -257,6 +240,14 @@ MEM_STATIC void MEM_write64(void* memPtr, U64 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 */
@@ -265,13 +256,22 @@ MEM_STATIC U32 MEM_swap32(U32 in)
   || (defined(__clang__) && __has_builtin(__builtin_bswap32))
     return __builtin_bswap32(in);
 #else
-    return  ((in << 24) & 0xff000000 ) |
+    return MEM_swap32_fallback(in);
-            ((in <<  8) & 0x00ff0000 ) |
-            ((in >>  8) & 0x0000ff00 ) |
-            ((in >> 24) & 0x000000ff );
 #endif
 }
 
+MEM_STATIC U64 MEM_swap64_fallback(U64 in)
+{
+     return  ((in << 56) & 0xff00000000000000ULL) |
+            ((in << 40) & 0x00ff000000000000ULL) |
+            ((in << 24) & 0x0000ff0000000000ULL) |
+            ((in << 8)  & 0x000000ff00000000ULL) |
+            ((in >> 8)  & 0x00000000ff000000ULL) |
+            ((in >> 24) & 0x0000000000ff0000ULL) |
+            ((in >> 40) & 0x000000000000ff00ULL) |
+            ((in >> 56) & 0x00000000000000ffULL);
+}
+
 MEM_STATIC U64 MEM_swap64(U64 in)
 {
 #if defined(_MSC_VER)     /* Visual Studio */
@@ -280,14 +280,7 @@ MEM_STATIC U64 MEM_swap64(U64 in)
   || (defined(__clang__) && __has_builtin(__builtin_bswap64))
     return __builtin_bswap64(in);
 #else
-    return  ((in << 56) & 0xff00000000000000ULL) |
+    return MEM_swap64_fallback(in);
-            ((in << 40) & 0x00ff000000000000ULL) |
-            ((in << 24) & 0x0000ff0000000000ULL) |
-            ((in << 8)  & 0x000000ff00000000ULL) |
-            ((in >> 8)  & 0x00000000ff000000ULL) |
-            ((in >> 24) & 0x0000000000ff0000ULL) |
-            ((in >> 40) & 0x000000000000ff00ULL) |
-            ((in >> 56) & 0x00000000000000ffULL);
 #endif
 }
 

zstd/common/pool.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -12,7 +12,7 @@
 /* ======   Dependencies   ======= */
 #include "zstd_deps.h" /* size_t */
 #include "debug.h"     /* assert */
-#include "zstd_internal.h"  /* ZSTD_customMalloc, ZSTD_customFree */
+#include "zstd_internal.h"  /* ZSTD_customCalloc, ZSTD_customFree */
 #include "pool.h"
 
 /* ======   Compiler specifics   ====== */
@@ -96,9 +96,7 @@ 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_mutex_unlock(&ctx->queueMutex);
         }
     }  /* for (;;) */
@@ -128,7 +126,7 @@ POOL_ctx* POOL_create_advanced(size_t numThreads, size_t queueSize,
      * empty and full queues.
      */
     ctx->queueSize = queueSize + 1;
-    ctx->queue = (POOL_job*)ZSTD_customMalloc(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;
@@ -142,7 +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_customMalloc(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 */
@@ -175,7 +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 */
     }   }
 }
 
@@ -190,6 +188,17 @@ void POOL_free(POOL_ctx *ctx) {
     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);
 }
@@ -211,7 +220,7 @@ static int POOL_resize_internal(POOL_ctx* ctx, size_t numThreads)
         return 0;
     }
     /* numThreads > threadCapacity */
-    {   ZSTD_pthread_t* const threadPool = (ZSTD_pthread_t*)ZSTD_customMalloc(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 */
         ZSTD_memcpy(threadPool, ctx->threads, ctx->threadCapacity * sizeof(*threadPool));
@@ -262,7 +271,9 @@ static int isQueueFull(POOL_ctx const* ctx) {
 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;
 
@@ -330,6 +341,11 @@ void POOL_free(POOL_ctx* ctx) {
     (void)ctx;
 }
 
+void POOL_joinJobs(POOL_ctx* ctx){
+    assert(!ctx || ctx == &g_poolCtx);
+    (void)ctx;
+}
+
 int POOL_resize(POOL_ctx* ctx, size_t numThreads) {
     (void)ctx; (void)numThreads;
     return 0;

zstd/common/pool.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -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,

zstd/common/portability_macros.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -12,7 +12,7 @@
 #define ZSTD_PORTABILITY_MACROS_H
 
 /**
- * This header file contains macro defintions to support portability.
+ * 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.
  *
@@ -88,7 +88,7 @@
 #endif
 
 /**
- * Only enable assembly for GNUC comptabile compilers,
+ * 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
@@ -134,4 +134,23 @@
 # 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 */

zstd/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 :
+ * Windows minimalist Pthread Wrapper
- * http://www.cse.wustl.edu/~schmidt/win32-cv-1.html
  */
 
 
@@ -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;
+    void* (*start_routine)(void*);
-    thread->arg = thread->start_routine(thread->arg);
+    void* thread_arg;
+
+    /* Inialized 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
+    }
+
+    /* 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;

zstd/common/threading.h

@@ -23,8 +23,7 @@ extern "C" {
 #if defined(ZSTD_MULTITHREAD) && defined(_WIN32)
 
 /**
- * Windows minimalist Pthread Wrapper, based on :
+ * Windows minimalist Pthread Wrapper
- * http://www.cse.wustl.edu/~schmidt/win32-cv-1.html
  */
 #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 {
+typedef HANDLE ZSTD_pthread_t;
-    HANDLE handle;
-    void* (*start_routine)(void*);
-    void* arg;
-} ZSTD_pthread_t;
 
 int ZSTD_pthread_create(ZSTD_pthread_t* thread, const void* unused,
                    void* (*start_routine) (void*), void* arg);
 
-int ZSTD_pthread_join(ZSTD_pthread_t thread, void** value_ptr);
+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
 

zstd/common/xxhash.c

@@ -1,9 +1,9 @@
 /*
  *  xxHash - Fast Hash algorithm
- *  Copyright (c) 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

zstd/common/xxhash.h

@@ -1,9 +1,9 @@
 /*
  *  xxHash - Fast Hash algorithm
- *  Copyright (c) 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
@@ -1314,7 +1314,7 @@ XXH3_128bits_reset_withSecretandSeed(XXH3_state_t* statePtr,
  *   care, as what works on one compiler/platform/optimization level may cause
  *   another to read garbage data or even crash.
  *
- * See http://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html for details.
+ * See https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html for details.
  *
  * Prefer these methods in priority order (0 > 3 > 1 > 2)
  */
@@ -1534,7 +1534,7 @@ static void* XXH_memcpy(void* dest, const void* src, size_t size) { return ZSTD_
  * @brief Used to prevent unwanted optimizations for @p var.
  *
  * It uses an empty GCC inline assembly statement with a register constraint
- * which forces @p var into a general purpose register (eg eax, ebx, ecx
+ * 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.
@@ -1655,7 +1655,7 @@ static xxh_u32 XXH_read32(const void* ptr)
 
 /*
  * Portable and safe solution. Generally efficient.
- * see: http://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
+ * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
  */
 static xxh_u32 XXH_read32(const void* memPtr)
 {
@@ -2296,7 +2296,7 @@ static xxh_u64 XXH_read64(const void* ptr)
 
 /*
  * Portable and safe solution. Generally efficient.
- * see: http://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
+ * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
  */
 static xxh_u64 XXH_read64(const void* memPtr)
 {
@@ -2809,7 +2809,7 @@ enum XXH_VECTOR_TYPE /* fake enum */ {
  * @ingroup tuning
  * @brief Selects the minimum alignment for XXH3's accumulators.
  *
- * When using SIMD, this should match the alignment reqired for said vector
+ * When using SIMD, this should match the alignment required for said vector
  * type, so, for example, 32 for AVX2.
  *
  * Default: Auto detected.
@@ -3026,7 +3026,7 @@ enum XXH_VECTOR_TYPE /* fake enum */ {
  * 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 noticable on the more advanced cores like the A76 which
+ * 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

zstd/common/zstd_common.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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

zstd/common/zstd_deps.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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

zstd/common/zstd_internal.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -28,7 +28,6 @@
 #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 */
@@ -93,9 +92,9 @@ 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
@@ -103,6 +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
@@ -113,6 +113,8 @@ typedef enum { set_basic, set_rle, set_compressed, set_repeat } symbolEncodingTy
 #define LLFSELog    9
 #define OffFSELog   8
 #define MaxFSELog  MAX(MAX(MLFSELog, LLFSELog), OffFSELog)
+#define MaxMLBits 16
+#define MaxLLBits 16
 
 #define ZSTD_MAX_HUF_HEADER_SIZE 128 /* header + <= 127 byte tree description */
 /* Each table cannot take more than #symbols * FSELog bits */
@@ -235,12 +237,6 @@ 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.
          */
-#ifdef __aarch64__
-        do {
-            COPY16(op, ip);
-        }
-        while (op < oend);
-#else
         ZSTD_copy16(op, ip);
         if (16 >= length) return;
         op += 16;
@@ -250,7 +246,6 @@ void ZSTD_wildcopy(void* dst, const void* src, ptrdiff_t length, ZSTD_overlap_e
             COPY16(op, ip);
         }
         while (op < oend);
-#endif
     }
 }
 
@@ -331,10 +326,10 @@ MEM_STATIC ZSTD_sequenceLength ZSTD_getSequenceLength(seqStore_t const* seqStore
     seqLen.matchLength = seq->mlBase + MINMATCH;
     if (seqStore->longLengthPos == (U32)(seq - seqStore->sequencesStart)) {
         if (seqStore->longLengthType == ZSTD_llt_literalLength) {
-            seqLen.litLength += 0xFFFF;
+            seqLen.litLength += 0x10000;
         }
         if (seqStore->longLengthType == ZSTD_llt_matchLength) {
-            seqLen.matchLength += 0xFFFF;
+            seqLen.matchLength += 0x10000;
         }
     }
     return seqLen;
@@ -347,12 +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) */
+int ZSTD_seqToCodes(const seqStore_t* seqStorePtr);   /* compress, dictBuilder, decodeCorpus (shouldn't get its definition from here) */
 
 /* custom memory allocation functions */
 void* ZSTD_customMalloc(size_t size, ZSTD_customMem customMem);
@@ -360,98 +356,6 @@ void* ZSTD_customCalloc(size_t size, ZSTD_customMem customMem);
 void ZSTD_customFree(void* ptr, ZSTD_customMem customMem);
 
 
-MEM_STATIC U32 ZSTD_highbit32(U32 val)   /* compress, dictBuilder, decodeCorpus */
-{
-    assert(val != 0);
-    {
-#   if defined(_MSC_VER)   /* Visual */
-#       if STATIC_BMI2 == 1
-            return _lzcnt_u32(val)^31;
-#       else
-            if (val != 0) {
-                unsigned long r;
-                _BitScanReverse(&r, val);
-                return (unsigned)r;
-            } else {
-                /* Should not reach this code path */
-                __assume(0);
-            }
-#       endif
-#   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
-    }
-}
-
-/**
- * Counts the number of trailing zeros of a `size_t`.
- * Most compilers should support CTZ as a builtin. A backup
- * implementation is provided if the builtin isn't supported, but
- * it may not be terribly efficient.
- */
-MEM_STATIC unsigned ZSTD_countTrailingZeros(size_t val)
-{
-    if (MEM_64bits()) {
-#       if defined(_MSC_VER) && defined(_WIN64)
-#           if STATIC_BMI2
-                return _tzcnt_u64(val);
-#           else
-                if (val != 0) {
-                    unsigned long r;
-                    _BitScanForward64(&r, (U64)val);
-                    return (unsigned)r;
-                } else {
-                    /* Should not reach this code path */
-                    __assume(0);
-                }
-#           endif
-#       elif defined(__GNUC__) && (__GNUC__ >= 4)
-            return __builtin_ctzll((U64)val);
-#       else
-            static const int DeBruijnBytePos[64] = {  0,  1,  2,  7,  3, 13,  8, 19,
-                                                      4, 25, 14, 28,  9, 34, 20, 56,
-                                                      5, 17, 26, 54, 15, 41, 29, 43,
-                                                      10, 31, 38, 35, 21, 45, 49, 57,
-                                                      63,  6, 12, 18, 24, 27, 33, 55,
-                                                      16, 53, 40, 42, 30, 37, 44, 48,
-                                                      62, 11, 23, 32, 52, 39, 36, 47,
-                                                      61, 22, 51, 46, 60, 50, 59, 58 };
-            return DeBruijnBytePos[((U64)((val & -(long long)val) * 0x0218A392CDABBD3FULL)) >> 58];
-#       endif
-    } else { /* 32 bits */
-#       if defined(_MSC_VER)
-            if (val != 0) {
-                unsigned long r;
-                _BitScanForward(&r, (U32)val);
-                return (unsigned)r;
-            } else {
-                /* Should not reach this code path */
-                __assume(0);
-            }
-#       elif defined(__GNUC__) && (__GNUC__ >= 3)
-            return __builtin_ctz((U32)val);
-#       else
-            static const int DeBruijnBytePos[32] = {  0,  1, 28,  2, 29, 14, 24,  3,
-                                                     30, 22, 20, 15, 25, 17,  4,  8,
-                                                     31, 27, 13, 23, 21, 19, 16,  7,
-                                                     26, 12, 18,  6, 11,  5, 10,  9 };
-            return DeBruijnBytePos[((U32)((val & -(S32)val) * 0x077CB531U)) >> 27];
-#       endif
-    }
-}
-
-
 /* ZSTD_invalidateRepCodes() :
  * ensures next compression will not use repcodes from previous block.
  * Note : only works with regular variant;

zstd/common/zstd_trace.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -21,13 +21,13 @@ extern "C" {
  * For now, enable conservatively:
  * - Only GNUC
  * - Only ELF
- * - Only x86-64 and i386
+ * - 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(__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

zstd/compress/clevels.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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

zstd/compress/fse_compress.c

@@ -1,6 +1,6 @@
 /* ******************************************************************
  * FSE : Finite State Entropy encoder
- * Copyright (c) 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
@@ -26,6 +26,7 @@
 #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 */
 
 
 /* **************************************************************
@@ -90,7 +91,7 @@ size_t FSE_buildCTable_wksp(FSE_CTable* ct,
     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__
      ZSTD_memset(tableSymbol, 0, sizeof(*tableSymbol) * tableSize);   /* useless initialization, just to keep scan-build happy */
@@ -191,7 +192,7 @@ size_t FSE_buildCTable_wksp(FSE_CTable* ct,
                 break;
             default :
                 assert(normalizedCounter[s] > 1);
-                {   U32 const maxBitsOut = tableLog - BIT_highbit32 ((U32)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 = (int)(total - (unsigned)normalizedCounter[s]);
@@ -342,21 +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*)ZSTD_malloc(size);
-}
-
-void FSE_freeCTable (FSE_CTable* ct) { ZSTD_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 minBitsSrc = ZSTD_highbit32((U32)(srcSize)) + 1;
-    U32 minBitsSymbols = BIT_highbit32(maxSymbolValue) + 2;
+    U32 minBitsSymbols = ZSTD_highbit32(maxSymbolValue) + 2;
     U32 minBits = minBitsSrc < minBitsSymbols ? minBitsSrc : minBitsSymbols;
     assert(srcSize > 1); /* Not supported, RLE should be used instead */
     return minBits;
@@ -364,7 +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 */
@@ -532,40 +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)
 {
@@ -664,78 +621,4 @@ size_t FSE_compress_usingCTable (void* dst, size_t dstSize,
 
 size_t FSE_compressBound(size_t size) { return FSE_COMPRESSBOUND(size); }
 
-#ifndef ZSTD_NO_UNUSED_FUNCTIONS
-/* 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_COMPRESS_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, /* useLowProbCount */ srcSize >= 2048) );
-
-    /* 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)];
-    union {
-      U32 hist_wksp[HIST_WKSP_SIZE_U32];
-      BYTE scratchBuffer[1 << FSE_MAX_TABLELOG];
-    } workspace;
-} 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_COMPRESS_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
-
 #endif   /* FSE_COMMONDEFS_ONLY */

zstd/compress/hist.c

@@ -1,7 +1,7 @@
 /* ******************************************************************
  * hist : Histogram functions
  * part of Finite State Entropy project
- * Copyright (c) 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

zstd/compress/hist.h

@@ -1,7 +1,7 @@
 /* ******************************************************************
  * hist : Histogram functions
  * part of Finite State Entropy project
- * Copyright (c) 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

zstd/compress/huf_compress.c

@@ -1,6 +1,6 @@
 /* ******************************************************************
  * Huffman encoder, part of New Generation Entropy library
- * Copyright (c) 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
@@ -29,9 +29,9 @@
 #include "hist.h"
 #define FSE_STATIC_LINKING_ONLY   /* FSE_optimalTableLog_internal */
 #include "../common/fse.h"        /* header compression */
-#define HUF_STATIC_LINKING_ONLY
 #include "../common/huf.h"
 #include "../common/error_private.h"
+#include "../common/bits.h"       /* ZSTD_highbit32 */
 
 
 /* **************************************************************
@@ -42,13 +42,67 @@
 
 
 /* **************************************************************
-*  Utils
+*  Required declarations
 ****************************************************************/
-unsigned HUF_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue)
+typedef struct nodeElt_s {
+    U32 count;
+    U16 parent;
+    BYTE byte;
+    BYTE nbBits;
+} nodeElt;
+
+
+/* **************************************************************
+*  Debug Traces
+****************************************************************/
+
+#if DEBUGLEVEL >= 2
+
+static size_t showU32(const U32* arr, size_t size)
 {
-    return FSE_optimalTableLog_internal(maxTableLog, srcSize, maxSymbolValue, 1);
+    size_t u;
+    for (u=0; u<size; u++) {
+        RAWLOG(6, " %u", arr[u]); (void)arr;
+    }
+    RAWLOG(6, " \n");
+    return size;
 }
 
+static size_t HUF_getNbBits(HUF_CElt elt);
+
+static size_t showCTableBits(const HUF_CElt* ctable, size_t size)
+{
+    size_t u;
+    for (u=0; u<size; u++) {
+        RAWLOG(6, " %zu", HUF_getNbBits(ctable[u])); (void)ctable;
+    }
+    RAWLOG(6, " \n");
+    return size;
+
+}
+
+static size_t showHNodeSymbols(const nodeElt* hnode, size_t size)
+{
+    size_t u;
+    for (u=0; u<size; u++) {
+        RAWLOG(6, " %u", hnode[u].byte); (void)hnode;
+    }
+    RAWLOG(6, " \n");
+    return size;
+}
+
+static size_t showHNodeBits(const nodeElt* hnode, size_t size)
+{
+    size_t u;
+    for (u=0; u<size; u++) {
+        RAWLOG(6, " %u", hnode[u].nbBits); (void)hnode;
+    }
+    RAWLOG(6, " \n");
+    return size;
+}
+
+#endif
+
 
 /* *******************************************************
 *  HUF : Huffman block compression
@@ -89,7 +143,10 @@ typedef struct {
     S16 norm[HUF_TABLELOG_MAX+1];
 } HUF_CompressWeightsWksp;
 
-static size_t HUF_compressWeights(void* dst, size_t dstSize, const void* weightTable, size_t wtSize, void* workspace, size_t workspaceSize)
+static size_t
+HUF_compressWeights(void* dst, size_t dstSize,
+              const void* weightTable, size_t wtSize,
+                    void* workspace, size_t workspaceSize)
 {
     BYTE* const ostart = (BYTE*) dst;
     BYTE* op = ostart;
@@ -140,7 +197,7 @@ static size_t HUF_getNbBitsFast(HUF_CElt elt)
 
 static size_t HUF_getValue(HUF_CElt elt)
 {
-    return elt & ~0xFF;
+    return elt & ~(size_t)0xFF;
 }
 
 static size_t HUF_getValueFast(HUF_CElt elt)
@@ -178,6 +235,8 @@ size_t HUF_writeCTable_wksp(void* dst, size_t maxDstSize,
     U32 n;
     HUF_WriteCTableWksp* wksp = (HUF_WriteCTableWksp*)HUF_alignUpWorkspace(workspace, &workspaceSize, ZSTD_ALIGNOF(U32));
 
+    HUF_STATIC_ASSERT(HUF_CTABLE_WORKSPACE_SIZE >= sizeof(HUF_WriteCTableWksp));
+
     /* check conditions */
     if (workspaceSize < sizeof(HUF_WriteCTableWksp)) return ERROR(GENERIC);
     if (maxSymbolValue > HUF_SYMBOLVALUE_MAX) return ERROR(maxSymbolValue_tooLarge);
@@ -207,16 +266,6 @@ size_t HUF_writeCTable_wksp(void* dst, size_t maxDstSize,
     return ((maxSymbolValue+1)/2) + 1;
 }
 
-/*! HUF_writeCTable() :
-    `CTable` : Huffman tree to save, using huf representation.
-    @return : size of saved CTable */
-size_t HUF_writeCTable (void* dst, size_t maxDstSize,
-                        const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog)
-{
-    HUF_WriteCTableWksp wksp;
-    return HUF_writeCTable_wksp(dst, maxDstSize, CTable, maxSymbolValue, huffLog, &wksp, sizeof(wksp));
-}
-
 
 size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void* src, size_t srcSize, unsigned* hasZeroWeights)
 {
@@ -272,68 +321,64 @@ size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void
 
 U32 HUF_getNbBitsFromCTable(HUF_CElt const* CTable, U32 symbolValue)
 {
-    const HUF_CElt* ct = CTable + 1;
+    const HUF_CElt* const ct = CTable + 1;
     assert(symbolValue <= HUF_SYMBOLVALUE_MAX);
     return (U32)HUF_getNbBits(ct[symbolValue]);
 }
 
 
-typedef struct nodeElt_s {
-    U32 count;
-    U16 parent;
-    BYTE byte;
-    BYTE nbBits;
-} nodeElt;
-
 /**
  * HUF_setMaxHeight():
- * Enforces maxNbBits on the Huffman tree described in huffNode.
+ * Try to enforce @targetNbBits on the Huffman tree described in @huffNode.
  *
- * It sets all nodes with nbBits > maxNbBits to be maxNbBits. Then it adjusts
+ * It attempts to convert all nodes with nbBits > @targetNbBits
- * the tree to so that it is a valid canonical Huffman tree.
+ * to employ @targetNbBits instead. Then it adjusts the tree
+ * so that it remains a valid canonical Huffman tree.
  *
  * @pre               The sum of the ranks of each symbol == 2^largestBits,
  *                    where largestBits == huffNode[lastNonNull].nbBits.
  * @post              The sum of the ranks of each symbol == 2^largestBits,
- *                    where largestBits is the return value <= maxNbBits.
+ *                    where largestBits is the return value (expected <= targetNbBits).
  *
- * @param huffNode    The Huffman tree modified in place to enforce maxNbBits.
+ * @param huffNode    The Huffman tree modified in place to enforce targetNbBits.
+ *                    It's presumed sorted, from most frequent to rarest symbol.
  * @param lastNonNull The symbol with the lowest count in the Huffman tree.
- * @param maxNbBits   The maximum allowed number of bits, which the Huffman tree
+ * @param targetNbBits  The allowed number of bits, which the Huffman tree
  *                    may not respect. After this function the Huffman tree will
- *                    respect maxNbBits.
+ *                    respect targetNbBits.
- * @return            The maximum number of bits of the Huffman tree after adjustment,
+ * @return            The maximum number of bits of the Huffman tree after adjustment.
- *                    necessarily no more than maxNbBits.
  */
-static U32 HUF_setMaxHeight(nodeElt* huffNode, U32 lastNonNull, U32 maxNbBits)
+static U32 HUF_setMaxHeight(nodeElt* huffNode, U32 lastNonNull, U32 targetNbBits)
 {
     const U32 largestBits = huffNode[lastNonNull].nbBits;
-    /* early exit : no elt > maxNbBits, so the tree is already valid. */
+    /* early exit : no elt > targetNbBits, so the tree is already valid. */
-    if (largestBits <= maxNbBits) return largestBits;
+    if (largestBits <= targetNbBits) return largestBits;
+
+    DEBUGLOG(5, "HUF_setMaxHeight (targetNbBits = %u)", targetNbBits);
 
     /* there are several too large elements (at least >= 2) */
     {   int totalCost = 0;
-        const U32 baseCost = 1 << (largestBits - maxNbBits);
+        const U32 baseCost = 1 << (largestBits - targetNbBits);
         int n = (int)lastNonNull;
 
-        /* Adjust any ranks > maxNbBits to maxNbBits.
+        /* Adjust any ranks > targetNbBits to targetNbBits.
          * Compute totalCost, which is how far the sum of the ranks is
          * we are over 2^largestBits after adjust the offending ranks.
          */
-        while (huffNode[n].nbBits > maxNbBits) {
+        while (huffNode[n].nbBits > targetNbBits) {
             totalCost += baseCost - (1 << (largestBits - huffNode[n].nbBits));
-            huffNode[n].nbBits = (BYTE)maxNbBits;
+            huffNode[n].nbBits = (BYTE)targetNbBits;
             n--;
         }
-        /* n stops at huffNode[n].nbBits <= maxNbBits */
+        /* n stops at huffNode[n].nbBits <= targetNbBits */
-        assert(huffNode[n].nbBits <= maxNbBits);
+        assert(huffNode[n].nbBits <= targetNbBits);
-        /* n end at index of smallest symbol using < maxNbBits */
+        /* n end at index of smallest symbol using < targetNbBits */
-        while (huffNode[n].nbBits == maxNbBits) --n;
+        while (huffNode[n].nbBits == targetNbBits) --n;
 
-        /* renorm totalCost from 2^largestBits to 2^maxNbBits
+        /* renorm totalCost from 2^largestBits to 2^targetNbBits
          * note : totalCost is necessarily a multiple of baseCost */
-        assert((totalCost & (baseCost - 1)) == 0);
+        assert(((U32)totalCost & (baseCost - 1)) == 0);
-        totalCost >>= (largestBits - maxNbBits);
+        totalCost >>= (largestBits - targetNbBits);
         assert(totalCost > 0);
 
         /* repay normalized cost */
@@ -342,19 +387,19 @@ static U32 HUF_setMaxHeight(nodeElt* huffNode, U32 lastNonNull, U32 maxNbBits)
 
             /* Get pos of last (smallest = lowest cum. count) symbol per rank */
             ZSTD_memset(rankLast, 0xF0, sizeof(rankLast));
-            {   U32 currentNbBits = maxNbBits;
+            {   U32 currentNbBits = targetNbBits;
                 int pos;
                 for (pos=n ; pos >= 0; pos--) {
                     if (huffNode[pos].nbBits >= currentNbBits) continue;
-                    currentNbBits = huffNode[pos].nbBits;   /* < maxNbBits */
+                    currentNbBits = huffNode[pos].nbBits;   /* < targetNbBits */
-                    rankLast[maxNbBits-currentNbBits] = (U32)pos;
+                    rankLast[targetNbBits-currentNbBits] = (U32)pos;
             }   }
 
             while (totalCost > 0) {
                 /* Try to reduce the next power of 2 above totalCost because we
                  * gain back half the rank.
                  */
-                U32 nBitsToDecrease = BIT_highbit32((U32)totalCost) + 1;
+                U32 nBitsToDecrease = ZSTD_highbit32((U32)totalCost) + 1;
                 for ( ; nBitsToDecrease > 1; nBitsToDecrease--) {
                     U32 const highPos = rankLast[nBitsToDecrease];
                     U32 const lowPos = rankLast[nBitsToDecrease-1];
@@ -394,7 +439,7 @@ static U32 HUF_setMaxHeight(nodeElt* huffNode, U32 lastNonNull, U32 maxNbBits)
                     rankLast[nBitsToDecrease] = noSymbol;
                 else {
                     rankLast[nBitsToDecrease]--;
-                    if (huffNode[rankLast[nBitsToDecrease]].nbBits != maxNbBits-nBitsToDecrease)
+                    if (huffNode[rankLast[nBitsToDecrease]].nbBits != targetNbBits-nBitsToDecrease)
                         rankLast[nBitsToDecrease] = noSymbol;   /* this rank is now empty */
                 }
             }   /* while (totalCost > 0) */
@@ -406,11 +451,11 @@ static U32 HUF_setMaxHeight(nodeElt* huffNode, U32 lastNonNull, U32 maxNbBits)
              * TODO.
              */
             while (totalCost < 0) {  /* Sometimes, cost correction overshoot */
-                /* special case : no rank 1 symbol (using maxNbBits-1);
+                /* special case : no rank 1 symbol (using targetNbBits-1);
-                 * let's create one from largest rank 0 (using maxNbBits).
+                 * let's create one from largest rank 0 (using targetNbBits).
                  */
                 if (rankLast[1] == noSymbol) {
-                    while (huffNode[n].nbBits == maxNbBits) n--;
+                    while (huffNode[n].nbBits == targetNbBits) n--;
                     huffNode[n+1].nbBits--;
                     assert(n >= 0);
                     rankLast[1] = (U32)(n+1);
@@ -424,7 +469,7 @@ static U32 HUF_setMaxHeight(nodeElt* huffNode, U32 lastNonNull, U32 maxNbBits)
         }   /* repay normalized cost */
     }   /* there are several too large elements (at least >= 2) */
 
-    return maxNbBits;
+    return targetNbBits;
 }
 
 typedef struct {
@@ -432,7 +477,7 @@ typedef struct {
     U16 curr;
 } rankPos;
 
-typedef nodeElt huffNodeTable[HUF_CTABLE_WORKSPACE_SIZE_U32];
+typedef nodeElt huffNodeTable[2 * (HUF_SYMBOLVALUE_MAX + 1)];
 
 /* Number of buckets available for HUF_sort() */
 #define RANK_POSITION_TABLE_SIZE 192
@@ -451,8 +496,8 @@ typedef struct {
  * Let buckets 166 to 192 represent all remaining counts up to RANK_POSITION_MAX_COUNT_LOG using log2 bucketing.
  */
 #define RANK_POSITION_MAX_COUNT_LOG 32
-#define RANK_POSITION_LOG_BUCKETS_BEGIN (RANK_POSITION_TABLE_SIZE - 1) - RANK_POSITION_MAX_COUNT_LOG - 1 /* == 158 */
+#define RANK_POSITION_LOG_BUCKETS_BEGIN ((RANK_POSITION_TABLE_SIZE - 1) - RANK_POSITION_MAX_COUNT_LOG - 1 /* == 158 */)
-#define RANK_POSITION_DISTINCT_COUNT_CUTOFF RANK_POSITION_LOG_BUCKETS_BEGIN + BIT_highbit32(RANK_POSITION_LOG_BUCKETS_BEGIN) /* == 166 */
+#define RANK_POSITION_DISTINCT_COUNT_CUTOFF (RANK_POSITION_LOG_BUCKETS_BEGIN + ZSTD_highbit32(RANK_POSITION_LOG_BUCKETS_BEGIN) /* == 166 */)
 
 /* Return the appropriate bucket index for a given count. See definition of
  * RANK_POSITION_DISTINCT_COUNT_CUTOFF for explanation of bucketing strategy.
@@ -460,7 +505,7 @@ typedef struct {
 static U32 HUF_getIndex(U32 const count) {
     return (count < RANK_POSITION_DISTINCT_COUNT_CUTOFF)
         ? count
-        : BIT_highbit32(count) + RANK_POSITION_LOG_BUCKETS_BEGIN;
+        : ZSTD_highbit32(count) + RANK_POSITION_LOG_BUCKETS_BEGIN;
 }
 
 /* Helper swap function for HUF_quickSortPartition() */
@@ -583,7 +628,7 @@ static void HUF_sort(nodeElt huffNode[], const unsigned count[], U32 const maxSy
 
     /* Sort each bucket. */
     for (n = RANK_POSITION_DISTINCT_COUNT_CUTOFF; n < RANK_POSITION_TABLE_SIZE - 1; ++n) {
-        U32 const bucketSize = rankPosition[n].curr-rankPosition[n].base;
+        int const bucketSize = rankPosition[n].curr - rankPosition[n].base;
         U32 const bucketStartIdx = rankPosition[n].base;
         if (bucketSize > 1) {
             assert(bucketStartIdx < maxSymbolValue1);
@@ -594,6 +639,7 @@ static void HUF_sort(nodeElt huffNode[], const unsigned count[], U32 const maxSy
     assert(HUF_isSorted(huffNode, maxSymbolValue1));
 }
 
+
 /** HUF_buildCTable_wksp() :
  *  Same as HUF_buildCTable(), but using externally allocated scratch buffer.
  *  `workSpace` must be aligned on 4-bytes boundaries, and be at least as large as sizeof(HUF_buildCTable_wksp_tables).
@@ -614,6 +660,7 @@ static int HUF_buildTree(nodeElt* huffNode, U32 maxSymbolValue)
     int lowS, lowN;
     int nodeNb = STARTNODE;
     int n, nodeRoot;
+    DEBUGLOG(5, "HUF_buildTree (alphabet size = %u)", maxSymbolValue + 1);
     /* init for parents */
     nonNullRank = (int)maxSymbolValue;
     while(huffNode[nonNullRank].count == 0) nonNullRank--;
@@ -640,6 +687,8 @@ static int HUF_buildTree(nodeElt* huffNode, U32 maxSymbolValue)
     for (n=0; n<=nonNullRank; n++)
         huffNode[n].nbBits = huffNode[ huffNode[n].parent ].nbBits + 1;
 
+    DEBUGLOG(6, "Initial distribution of bits completed (%zu sorted symbols)", showHNodeBits(huffNode, maxSymbolValue+1));
+
     return nonNullRank;
 }
 
@@ -677,13 +726,20 @@ static void HUF_buildCTableFromTree(HUF_CElt* CTable, nodeElt const* huffNode, i
     CTable[0] = maxNbBits;
 }
 
-size_t HUF_buildCTable_wksp (HUF_CElt* CTable, const unsigned* count, U32 maxSymbolValue, U32 maxNbBits, void* workSpace, size_t wkspSize)
+size_t
+HUF_buildCTable_wksp(HUF_CElt* CTable, const unsigned* count, U32 maxSymbolValue, U32 maxNbBits,
+                     void* workSpace, size_t wkspSize)
 {
-    HUF_buildCTable_wksp_tables* const wksp_tables = (HUF_buildCTable_wksp_tables*)HUF_alignUpWorkspace(workSpace, &wkspSize, ZSTD_ALIGNOF(U32));
+    HUF_buildCTable_wksp_tables* const wksp_tables =
+        (HUF_buildCTable_wksp_tables*)HUF_alignUpWorkspace(workSpace, &wkspSize, ZSTD_ALIGNOF(U32));
     nodeElt* const huffNode0 = wksp_tables->huffNodeTbl;
     nodeElt* const huffNode = huffNode0+1;
     int nonNullRank;
 
+    HUF_STATIC_ASSERT(HUF_CTABLE_WORKSPACE_SIZE == sizeof(HUF_buildCTable_wksp_tables));
+
+    DEBUGLOG(5, "HUF_buildCTable_wksp (alphabet size = %u)", maxSymbolValue+1);
+
     /* safety checks */
     if (wkspSize < sizeof(HUF_buildCTable_wksp_tables))
         return ERROR(workSpace_tooSmall);
@@ -694,11 +750,12 @@ size_t HUF_buildCTable_wksp (HUF_CElt* CTable, const unsigned* count, U32 maxSym
 
     /* sort, decreasing order */
     HUF_sort(huffNode, count, maxSymbolValue, wksp_tables->rankPosition);
+    DEBUGLOG(6, "sorted symbols completed (%zu symbols)", showHNodeSymbols(huffNode, maxSymbolValue+1));
 
     /* build tree */
     nonNullRank = HUF_buildTree(huffNode, maxSymbolValue);
 
-    /* enforce maxTableLog */
+    /* determine and enforce maxTableLog */
     maxNbBits = HUF_setMaxHeight(huffNode, (U32)nonNullRank, maxNbBits);
     if (maxNbBits > HUF_TABLELOG_MAX) return ERROR(GENERIC);   /* check fit into table */
 
@@ -807,7 +864,7 @@ FORCE_INLINE_TEMPLATE void HUF_addBits(HUF_CStream_t* bitC, HUF_CElt elt, int id
 #if DEBUGLEVEL >= 1
     {
         size_t const nbBits = HUF_getNbBits(elt);
-        size_t const dirtyBits = nbBits == 0 ? 0 : BIT_highbit32((U32)nbBits) + 1;
+        size_t const dirtyBits = nbBits == 0 ? 0 : ZSTD_highbit32((U32)nbBits) + 1;
         (void)dirtyBits;
         /* Middle bits are 0. */
         assert(((elt >> dirtyBits) << (dirtyBits + nbBits)) == 0);
@@ -887,7 +944,7 @@ static size_t HUF_closeCStream(HUF_CStream_t* bitC)
     {
         size_t const nbBits = bitC->bitPos[0] & 0xFF;
         if (bitC->ptr >= bitC->endPtr) return 0; /* overflow detected */
-        return (bitC->ptr - bitC->startPtr) + (nbBits > 0);
+        return (size_t)(bitC->ptr - bitC->startPtr) + (nbBits > 0);
     }
 }
 
@@ -1048,9 +1105,9 @@ HUF_compress1X_usingCTable_internal_default(void* dst, size_t dstSize,
 static size_t
 HUF_compress1X_usingCTable_internal(void* dst, size_t dstSize,
                               const void* src, size_t srcSize,
-                              const HUF_CElt* CTable, const int bmi2)
+                              const HUF_CElt* CTable, const int flags)
 {
-    if (bmi2) {
+    if (flags & HUF_flags_bmi2) {
         return HUF_compress1X_usingCTable_internal_bmi2(dst, dstSize, src, srcSize, CTable);
     }
     return HUF_compress1X_usingCTable_internal_default(dst, dstSize, src, srcSize, CTable);
@@ -1061,28 +1118,23 @@ HUF_compress1X_usingCTable_internal(void* dst, size_t dstSize,
 static size_t
 HUF_compress1X_usingCTable_internal(void* dst, size_t dstSize,
                               const void* src, size_t srcSize,
-                              const HUF_CElt* CTable, const int bmi2)
+                              const HUF_CElt* CTable, const int flags)
 {
-    (void)bmi2;
+    (void)flags;
     return HUF_compress1X_usingCTable_internal_body(dst, dstSize, src, srcSize, CTable);
 }
 
 #endif
 
-size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable)
+size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int flags)
 {
-    return HUF_compress1X_usingCTable_bmi2(dst, dstSize, src, srcSize, CTable, /* bmi2 */ 0);
+    return HUF_compress1X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, flags);
-}
-
-size_t HUF_compress1X_usingCTable_bmi2(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int bmi2)
-{
-    return HUF_compress1X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, bmi2);
 }
 
 static size_t
 HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize,
                               const void* src, size_t srcSize,
-                              const HUF_CElt* CTable, int bmi2)
+                              const HUF_CElt* CTable, int flags)
 {
     size_t const segmentSize = (srcSize+3)/4;   /* first 3 segments */
     const BYTE* ip = (const BYTE*) src;
@@ -1096,7 +1148,7 @@ HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize,
     op += 6;   /* jumpTable */
 
     assert(op <= oend);
-    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, bmi2) );
+    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, flags) );
         if (cSize == 0 || cSize > 65535) return 0;
         MEM_writeLE16(ostart, (U16)cSize);
         op += cSize;
@@ -1104,7 +1156,7 @@ HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize,
 
     ip += segmentSize;
     assert(op <= oend);
-    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, bmi2) );
+    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, flags) );
         if (cSize == 0 || cSize > 65535) return 0;
         MEM_writeLE16(ostart+2, (U16)cSize);
         op += cSize;
@@ -1112,7 +1164,7 @@ HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize,
 
     ip += segmentSize;
     assert(op <= oend);
-    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, bmi2) );
+    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, flags) );
         if (cSize == 0 || cSize > 65535) return 0;
         MEM_writeLE16(ostart+4, (U16)cSize);
         op += cSize;
@@ -1121,7 +1173,7 @@ HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize,
     ip += segmentSize;
     assert(op <= oend);
     assert(ip <= iend);
-    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, (size_t)(iend-ip), CTable, bmi2) );
+    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, (size_t)(iend-ip), CTable, flags) );
         if (cSize == 0 || cSize > 65535) return 0;
         op += cSize;
     }
@@ -1129,14 +1181,9 @@ HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize,
     return (size_t)(op-ostart);
 }
 
-size_t HUF_compress4X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable)
+size_t HUF_compress4X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int flags)
 {
-    return HUF_compress4X_usingCTable_bmi2(dst, dstSize, src, srcSize, CTable, /* bmi2 */ 0);
+    return HUF_compress4X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, flags);
-}
-
-size_t HUF_compress4X_usingCTable_bmi2(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int bmi2)
-{
-    return HUF_compress4X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, bmi2);
 }
 
 typedef enum { HUF_singleStream, HUF_fourStreams } HUF_nbStreams_e;
@@ -1144,11 +1191,11 @@ typedef enum { HUF_singleStream, HUF_fourStreams } HUF_nbStreams_e;
 static size_t HUF_compressCTable_internal(
                 BYTE* const ostart, BYTE* op, BYTE* const oend,
                 const void* src, size_t srcSize,
-                HUF_nbStreams_e nbStreams, const HUF_CElt* CTable, const int bmi2)
+                HUF_nbStreams_e nbStreams, const HUF_CElt* CTable, const int flags)
 {
     size_t const cSize = (nbStreams==HUF_singleStream) ?
-                         HUF_compress1X_usingCTable_internal(op, (size_t)(oend - op), src, srcSize, CTable, bmi2) :
+                         HUF_compress1X_usingCTable_internal(op, (size_t)(oend - op), src, srcSize, CTable, flags) :
-                         HUF_compress4X_usingCTable_internal(op, (size_t)(oend - op), src, srcSize, CTable, bmi2);
+                         HUF_compress4X_usingCTable_internal(op, (size_t)(oend - op), src, srcSize, CTable, flags);
     if (HUF_isError(cSize)) { return cSize; }
     if (cSize==0) { return 0; }   /* uncompressible */
     op += cSize;
@@ -1171,6 +1218,79 @@ typedef struct {
 #define SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE 4096
 #define SUSPECT_INCOMPRESSIBLE_SAMPLE_RATIO 10  /* Must be >= 2 */
 
+unsigned HUF_cardinality(const unsigned* count, unsigned maxSymbolValue)
+{
+    unsigned cardinality = 0;
+    unsigned i;
+
+    for (i = 0; i < maxSymbolValue + 1; i++) {
+        if (count[i] != 0) cardinality += 1;
+    }
+
+    return cardinality;
+}
+
+unsigned HUF_minTableLog(unsigned symbolCardinality)
+{
+    U32 minBitsSymbols = ZSTD_highbit32(symbolCardinality) + 1;
+    return minBitsSymbols;
+}
+
+unsigned HUF_optimalTableLog(
+            unsigned maxTableLog,
+            size_t srcSize,
+            unsigned maxSymbolValue,
+            void* workSpace, size_t wkspSize,
+            HUF_CElt* table,
+      const unsigned* count,
+            int flags)
+{
+    assert(srcSize > 1); /* Not supported, RLE should be used instead */
+    assert(wkspSize >= sizeof(HUF_buildCTable_wksp_tables));
+
+    if (!(flags & HUF_flags_optimalDepth)) {
+        /* cheap evaluation, based on FSE */
+        return FSE_optimalTableLog_internal(maxTableLog, srcSize, maxSymbolValue, 1);
+    }
+
+    {   BYTE* dst = (BYTE*)workSpace + sizeof(HUF_WriteCTableWksp);
+        size_t dstSize = wkspSize - sizeof(HUF_WriteCTableWksp);
+        size_t maxBits, hSize, newSize;
+        const unsigned symbolCardinality = HUF_cardinality(count, maxSymbolValue);
+        const unsigned minTableLog = HUF_minTableLog(symbolCardinality);
+        size_t optSize = ((size_t) ~0) - 1;
+        unsigned optLog = maxTableLog, optLogGuess;
+
+        DEBUGLOG(6, "HUF_optimalTableLog: probing huf depth (srcSize=%zu)", srcSize);
+
+        /* Search until size increases */
+        for (optLogGuess = minTableLog; optLogGuess <= maxTableLog; optLogGuess++) {
+            DEBUGLOG(7, "checking for huffLog=%u", optLogGuess);
+            maxBits = HUF_buildCTable_wksp(table, count, maxSymbolValue, optLogGuess, workSpace, wkspSize);
+            if (ERR_isError(maxBits)) continue;
+
+            if (maxBits < optLogGuess && optLogGuess > minTableLog) break;
+
+            hSize = HUF_writeCTable_wksp(dst, dstSize, table, maxSymbolValue, (U32)maxBits, workSpace, wkspSize);
+
+            if (ERR_isError(hSize)) continue;
+
+            newSize = HUF_estimateCompressedSize(table, count, maxSymbolValue) + hSize;
+
+            if (newSize > optSize + 1) {
+                break;
+            }
+
+            if (newSize < optSize) {
+                optSize = newSize;
+                optLog = optLogGuess;
+            }
+        }
+        assert(optLog <= HUF_TABLELOG_MAX);
+        return optLog;
+    }
+}
+
 /* HUF_compress_internal() :
  * `workSpace_align4` must be aligned on 4-bytes boundaries,
  * and occupies the same space as a table of HUF_WORKSPACE_SIZE_U64 unsigned */
@@ -1180,14 +1300,14 @@ HUF_compress_internal (void* dst, size_t dstSize,
                        unsigned maxSymbolValue, unsigned huffLog,
                        HUF_nbStreams_e nbStreams,
                        void* workSpace, size_t wkspSize,
-                       HUF_CElt* oldHufTable, HUF_repeat* repeat, int preferRepeat,
+                       HUF_CElt* oldHufTable, HUF_repeat* repeat, int flags)
-                 const int bmi2, unsigned suspectUncompressible)
 {
     HUF_compress_tables_t* const table = (HUF_compress_tables_t*)HUF_alignUpWorkspace(workSpace, &wkspSize, ZSTD_ALIGNOF(size_t));
     BYTE* const ostart = (BYTE*)dst;
     BYTE* const oend = ostart + dstSize;
     BYTE* op = ostart;
 
+    DEBUGLOG(5, "HUF_compress_internal (srcSize=%zu)", srcSize);
     HUF_STATIC_ASSERT(sizeof(*table) + HUF_WORKSPACE_MAX_ALIGNMENT <= HUF_WORKSPACE_SIZE);
 
     /* checks & inits */
@@ -1201,16 +1321,17 @@ HUF_compress_internal (void* dst, size_t dstSize,
     if (!huffLog) huffLog = HUF_TABLELOG_DEFAULT;
 
     /* Heuristic : If old table is valid, use it for small inputs */
-    if (preferRepeat && repeat && *repeat == HUF_repeat_valid) {
+    if ((flags & HUF_flags_preferRepeat) && repeat && *repeat == HUF_repeat_valid) {
         return HUF_compressCTable_internal(ostart, op, oend,
                                            src, srcSize,
-                                           nbStreams, oldHufTable, bmi2);
+                                           nbStreams, oldHufTable, flags);
     }
 
     /* If uncompressible data is suspected, do a smaller sampling first */
     DEBUG_STATIC_ASSERT(SUSPECT_INCOMPRESSIBLE_SAMPLE_RATIO >= 2);
-    if (suspectUncompressible && srcSize >= (SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE * SUSPECT_INCOMPRESSIBLE_SAMPLE_RATIO)) {
+    if ((flags & HUF_flags_suspectUncompressible) && srcSize >= (SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE * SUSPECT_INCOMPRESSIBLE_SAMPLE_RATIO)) {
         size_t largestTotal = 0;
+        DEBUGLOG(5, "input suspected incompressible : sampling to check");
         {   unsigned maxSymbolValueBegin = maxSymbolValue;
             CHECK_V_F(largestBegin, HIST_count_simple (table->count, &maxSymbolValueBegin, (const BYTE*)src, SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE) );
             largestTotal += largestBegin;
@@ -1227,6 +1348,7 @@ HUF_compress_internal (void* dst, size_t dstSize,
         if (largest == srcSize) { *ostart = ((const BYTE*)src)[0]; return 1; }   /* single symbol, rle */
         if (largest <= (srcSize >> 7)+4) return 0;   /* heuristic : probably not compressible enough */
     }
+    DEBUGLOG(6, "histogram detail completed (%zu symbols)", showU32(table->count, maxSymbolValue+1));
 
     /* Check validity of previous table */
     if ( repeat
@@ -1235,19 +1357,20 @@ HUF_compress_internal (void* dst, size_t dstSize,
         *repeat = HUF_repeat_none;
     }
     /* Heuristic : use existing table for small inputs */
-    if (preferRepeat && repeat && *repeat != HUF_repeat_none) {
+    if ((flags & HUF_flags_preferRepeat) && repeat && *repeat != HUF_repeat_none) {
         return HUF_compressCTable_internal(ostart, op, oend,
                                            src, srcSize,
-                                           nbStreams, oldHufTable, bmi2);
+                                           nbStreams, oldHufTable, flags);
     }
 
     /* Build Huffman Tree */
-    huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue);
+    huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue, &table->wksps, sizeof(table->wksps), table->CTable, table->count, flags);
     {   size_t const maxBits = HUF_buildCTable_wksp(table->CTable, table->count,
                                             maxSymbolValue, huffLog,
                                             &table->wksps.buildCTable_wksp, sizeof(table->wksps.buildCTable_wksp));
         CHECK_F(maxBits);
         huffLog = (U32)maxBits;
+        DEBUGLOG(6, "bit distribution completed (%zu symbols)", showCTableBits(table->CTable + 1, maxSymbolValue+1));
     }
     /* Zero unused symbols in CTable, so we can check it for validity */
     {
@@ -1266,7 +1389,7 @@ HUF_compress_internal (void* dst, size_t dstSize,
             if (oldSize <= hSize + newSize || hSize + 12 >= srcSize) {
                 return HUF_compressCTable_internal(ostart, op, oend,
                                                    src, srcSize,
-                                                   nbStreams, oldHufTable, bmi2);
+                                                   nbStreams, oldHufTable, flags);
         }   }
 
         /* Use the new huffman table */
@@ -1278,46 +1401,20 @@ HUF_compress_internal (void* dst, size_t dstSize,
     }
     return HUF_compressCTable_internal(ostart, op, oend,
                                        src, srcSize,
-                                       nbStreams, table->CTable, bmi2);
+                                       nbStreams, table->CTable, flags);
-}
-
-
-size_t HUF_compress1X_wksp (void* dst, size_t dstSize,
-                      const void* src, size_t srcSize,
-                      unsigned maxSymbolValue, unsigned huffLog,
-                      void* workSpace, size_t wkspSize)
-{
-    return HUF_compress_internal(dst, dstSize, src, srcSize,
-                                 maxSymbolValue, huffLog, HUF_singleStream,
-                                 workSpace, wkspSize,
-                                 NULL, NULL, 0, 0 /*bmi2*/, 0);
 }
 
 size_t HUF_compress1X_repeat (void* dst, size_t dstSize,
                       const void* src, size_t srcSize,
                       unsigned maxSymbolValue, unsigned huffLog,
                       void* workSpace, size_t wkspSize,
-                      HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat,
+                      HUF_CElt* hufTable, HUF_repeat* repeat, int flags)
-                      int bmi2, unsigned suspectUncompressible)
 {
+    DEBUGLOG(5, "HUF_compress1X_repeat (srcSize = %zu)", srcSize);
     return HUF_compress_internal(dst, dstSize, src, srcSize,
                                  maxSymbolValue, huffLog, HUF_singleStream,
                                  workSpace, wkspSize, hufTable,
-                                 repeat, preferRepeat, bmi2, suspectUncompressible);
+                                 repeat, flags);
-}
-
-/* HUF_compress4X_repeat():
- * compress input using 4 streams.
- * provide workspace to generate compression tables */
-size_t HUF_compress4X_wksp (void* dst, size_t dstSize,
-                      const void* src, size_t srcSize,
-                      unsigned maxSymbolValue, unsigned huffLog,
-                      void* workSpace, size_t wkspSize)
-{
-    return HUF_compress_internal(dst, dstSize, src, srcSize,
-                                 maxSymbolValue, huffLog, HUF_fourStreams,
-                                 workSpace, wkspSize,
-                                 NULL, NULL, 0, 0 /*bmi2*/, 0);
 }
 
 /* HUF_compress4X_repeat():
@@ -1328,43 +1425,11 @@ size_t HUF_compress4X_repeat (void* dst, size_t dstSize,
                       const void* src, size_t srcSize,
                       unsigned maxSymbolValue, unsigned huffLog,
                       void* workSpace, size_t wkspSize,
-                      HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2, unsigned suspectUncompressible)
+                      HUF_CElt* hufTable, HUF_repeat* repeat, int flags)
 {
+    DEBUGLOG(5, "HUF_compress4X_repeat (srcSize = %zu)", srcSize);
     return HUF_compress_internal(dst, dstSize, src, srcSize,
                                  maxSymbolValue, huffLog, HUF_fourStreams,
                                  workSpace, wkspSize,
-                                 hufTable, repeat, preferRepeat, bmi2, suspectUncompressible);
+                                 hufTable, repeat, flags);
 }
-
-#ifndef ZSTD_NO_UNUSED_FUNCTIONS
-/** HUF_buildCTable() :
- * @return : maxNbBits
- *  Note : count is used before tree is written, so they can safely overlap
- */
-size_t HUF_buildCTable (HUF_CElt* tree, const unsigned* count, unsigned maxSymbolValue, unsigned maxNbBits)
-{
-    HUF_buildCTable_wksp_tables workspace;
-    return HUF_buildCTable_wksp(tree, count, maxSymbolValue, maxNbBits, &workspace, sizeof(workspace));
-}
-
-size_t HUF_compress1X (void* dst, size_t dstSize,
-                 const void* src, size_t srcSize,
-                 unsigned maxSymbolValue, unsigned huffLog)
-{
-    U64 workSpace[HUF_WORKSPACE_SIZE_U64];
-    return HUF_compress1X_wksp(dst, dstSize, src, srcSize, maxSymbolValue, huffLog, workSpace, sizeof(workSpace));
-}
-
-size_t HUF_compress2 (void* dst, size_t dstSize,
-                const void* src, size_t srcSize,
-                unsigned maxSymbolValue, unsigned huffLog)
-{
-    U64 workSpace[HUF_WORKSPACE_SIZE_U64];
-    return HUF_compress4X_wksp(dst, dstSize, src, srcSize, maxSymbolValue, huffLog, workSpace, sizeof(workSpace));
-}
-
-size_t HUF_compress (void* dst, size_t maxDstSize, const void* src, size_t srcSize)
-{
-    return HUF_compress2(dst, maxDstSize, src, srcSize, 255, HUF_TABLELOG_DEFAULT);
-}
-#endif

zstd/compress/zstd_compress_internal.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -23,6 +23,7 @@
 #ifdef ZSTD_MULTITHREAD
 #  include "zstdmt_compress.h"
 #endif
+#include "../common/bits.h" /* ZSTD_highbit32, ZSTD_NbCommonBytes */
 
 #if defined (__cplusplus)
 extern "C" {
@@ -117,7 +118,8 @@ typedef struct {
 /** ZSTD_buildBlockEntropyStats() :
  *  Builds entropy for the block.
  *  @return : 0 on success or error code */
-size_t ZSTD_buildBlockEntropyStats(seqStore_t* seqStorePtr,
+size_t ZSTD_buildBlockEntropyStats(
+                    const seqStore_t* seqStorePtr,
                     const ZSTD_entropyCTables_t* prevEntropy,
                           ZSTD_entropyCTables_t* nextEntropy,
                     const ZSTD_CCtx_params* cctxParams,
@@ -148,6 +150,12 @@ typedef struct {
   size_t capacity;      /* The capacity starting from `seq` pointer */
 } rawSeqStore_t;
 
+typedef struct {
+    U32 idx;            /* Index in array of ZSTD_Sequence */
+    U32 posInSequence;  /* Position within sequence at idx */
+    size_t posInSrc;    /* Number of bytes given by sequences provided so far */
+} ZSTD_sequencePosition;
+
 UNUSED_ATTR static const rawSeqStore_t kNullRawSeqStore = {NULL, 0, 0, 0, 0};
 
 typedef struct {
@@ -234,6 +242,11 @@ struct ZSTD_matchState_t {
     const ZSTD_matchState_t* dictMatchState;
     ZSTD_compressionParameters cParams;
     const rawSeqStore_t* ldmSeqStore;
+
+    /* Controls prefetching in some dictMatchState matchfinders.
+     * This behavior is controlled from the cctx ms.
+     * This parameter has no effect in the cdict ms. */
+    int prefetchCDictTables;
 };
 
 typedef struct {
@@ -330,6 +343,24 @@ struct ZSTD_CCtx_params_s {
 
     /* Internal use, for createCCtxParams() and freeCCtxParams() only */
     ZSTD_customMem customMem;
+
+    /* Controls prefetching in some dictMatchState matchfinders */
+    ZSTD_paramSwitch_e prefetchCDictTables;
+
+    /* Controls whether zstd will fall back to an internal matchfinder
+     * if the external matchfinder returns an error code. */
+    int enableMatchFinderFallback;
+
+    /* Indicates whether an external matchfinder has been referenced.
+     * Users can't set this externally.
+     * It is set internally in ZSTD_registerSequenceProducer(). */
+    int useSequenceProducer;
+
+    /* Adjust the max block size*/
+    size_t maxBlockSize;
+
+    /* Controls repcode search in external sequence parsing */
+    ZSTD_paramSwitch_e searchForExternalRepcodes;
 };  /* typedef'd to ZSTD_CCtx_params within "zstd.h" */
 
 #define COMPRESS_SEQUENCES_WORKSPACE_SIZE (sizeof(unsigned) * (MaxSeq + 2))
@@ -361,6 +392,14 @@ typedef struct {
     ZSTD_entropyCTablesMetadata_t entropyMetadata;
 } ZSTD_blockSplitCtx;
 
+/* Context for block-level external matchfinder API */
+typedef struct {
+  void* mState;
+  ZSTD_sequenceProducer_F* mFinder;
+  ZSTD_Sequence* seqBuffer;
+  size_t seqBufferCapacity;
+} ZSTD_externalMatchCtx;
+
 struct ZSTD_CCtx_s {
     ZSTD_compressionStage_e stage;
     int cParamsChanged;                  /* == 1 if cParams(except wlog) or compression level are changed in requestedParams. Triggers transmission of new params to ZSTDMT (if available) then reset to 0. */
@@ -410,6 +449,7 @@ struct ZSTD_CCtx_s {
 
     /* Stable in/out buffer verification */
     ZSTD_inBuffer expectedInBuffer;
+    size_t stableIn_notConsumed; /* nb bytes within stable input buffer that are said to be consumed but are not */
     size_t expectedOutBufferSize;
 
     /* Dictionary */
@@ -429,9 +469,13 @@ struct ZSTD_CCtx_s {
 
     /* Workspace for block splitter */
     ZSTD_blockSplitCtx blockSplitCtx;
+
+    /* Workspace for external matchfinder */
+    ZSTD_externalMatchCtx externalMatchCtx;
 };
 
 typedef enum { ZSTD_dtlm_fast, ZSTD_dtlm_full } ZSTD_dictTableLoadMethod_e;
+typedef enum { ZSTD_tfp_forCCtx, ZSTD_tfp_forCDict } ZSTD_tableFillPurpose_e;
 
 typedef enum {
     ZSTD_noDict = 0,
@@ -453,7 +497,7 @@ typedef enum {
                                  * In this mode we take both the source size and the dictionary size
                                  * into account when selecting and adjusting the parameters.
                                  */
-    ZSTD_cpm_unknown = 3,       /* ZSTD_getCParams, ZSTD_getParams, ZSTD_adjustParams.
+    ZSTD_cpm_unknown = 3        /* ZSTD_getCParams, ZSTD_getParams, ZSTD_adjustParams.
                                  * We don't know what these parameters are for. We default to the legacy
                                  * behavior of taking both the source size and the dict size into account
                                  * when selecting and adjusting parameters.
@@ -512,9 +556,11 @@ MEM_STATIC int ZSTD_cParam_withinBounds(ZSTD_cParameter cParam, int value)
 /* ZSTD_noCompressBlock() :
  * Writes uncompressed block to dst buffer from given src.
  * Returns the size of the block */
-MEM_STATIC size_t ZSTD_noCompressBlock (void* dst, size_t dstCapacity, const void* src, size_t srcSize, U32 lastBlock)
+MEM_STATIC size_t
+ZSTD_noCompressBlock(void* dst, size_t dstCapacity, const void* src, size_t srcSize, U32 lastBlock)
 {
     U32 const cBlockHeader24 = lastBlock + (((U32)bt_raw)<<1) + (U32)(srcSize << 3);
+    DEBUGLOG(5, "ZSTD_noCompressBlock (srcSize=%zu, dstCapacity=%zu)", srcSize, dstCapacity);
     RETURN_ERROR_IF(srcSize + ZSTD_blockHeaderSize > dstCapacity,
                     dstSize_tooSmall, "dst buf too small for uncompressed block");
     MEM_writeLE24(dst, cBlockHeader24);
@@ -522,7 +568,8 @@ MEM_STATIC size_t ZSTD_noCompressBlock (void* dst, size_t dstCapacity, const voi
     return ZSTD_blockHeaderSize + srcSize;
 }
 
-MEM_STATIC size_t ZSTD_rleCompressBlock (void* dst, size_t dstCapacity, BYTE src, size_t srcSize, U32 lastBlock)
+MEM_STATIC size_t
+ZSTD_rleCompressBlock(void* dst, size_t dstCapacity, BYTE src, size_t srcSize, U32 lastBlock)
 {
     BYTE* const op = (BYTE*)dst;
     U32 const cBlockHeader = lastBlock + (((U32)bt_rle)<<1) + (U32)(srcSize << 3);
@@ -541,7 +588,7 @@ MEM_STATIC size_t ZSTD_minGain(size_t srcSize, ZSTD_strategy strat)
 {
     U32 const minlog = (strat>=ZSTD_btultra) ? (U32)(strat) - 1 : 6;
     ZSTD_STATIC_ASSERT(ZSTD_btultra == 8);
-    assert(ZSTD_cParam_withinBounds(ZSTD_c_strategy, strat));
+    assert(ZSTD_cParam_withinBounds(ZSTD_c_strategy, (int)strat));
     return (srcSize >> minlog) + 2;
 }
 
@@ -577,29 +624,27 @@ ZSTD_safecopyLiterals(BYTE* op, BYTE const* ip, BYTE const* const iend, BYTE con
     while (ip < iend) *op++ = *ip++;
 }
 
-#define ZSTD_REP_MOVE     (ZSTD_REP_NUM-1)
+
-#define STORE_REPCODE_1 STORE_REPCODE(1)
+#define REPCODE1_TO_OFFBASE REPCODE_TO_OFFBASE(1)
-#define STORE_REPCODE_2 STORE_REPCODE(2)
+#define REPCODE2_TO_OFFBASE REPCODE_TO_OFFBASE(2)
-#define STORE_REPCODE_3 STORE_REPCODE(3)
+#define REPCODE3_TO_OFFBASE REPCODE_TO_OFFBASE(3)
-#define STORE_REPCODE(r) (assert((r)>=1), assert((r)<=3), (r)-1)
+#define REPCODE_TO_OFFBASE(r) (assert((r)>=1), assert((r)<=ZSTD_REP_NUM), (r)) /* accepts IDs 1,2,3 */
-#define STORE_OFFSET(o)  (assert((o)>0), o + ZSTD_REP_MOVE)
+#define OFFSET_TO_OFFBASE(o)  (assert((o)>0), o + ZSTD_REP_NUM)
-#define STORED_IS_OFFSET(o)  ((o) > ZSTD_REP_MOVE)
+#define OFFBASE_IS_OFFSET(o)  ((o) > ZSTD_REP_NUM)
-#define STORED_IS_REPCODE(o) ((o) <= ZSTD_REP_MOVE)
+#define OFFBASE_IS_REPCODE(o) ( 1 <= (o) && (o) <= ZSTD_REP_NUM)
-#define STORED_OFFSET(o)  (assert(STORED_IS_OFFSET(o)), (o)-ZSTD_REP_MOVE)
+#define OFFBASE_TO_OFFSET(o)  (assert(OFFBASE_IS_OFFSET(o)), (o) - ZSTD_REP_NUM)
-#define STORED_REPCODE(o) (assert(STORED_IS_REPCODE(o)), (o)+1)  /* returns ID 1,2,3 */
+#define OFFBASE_TO_REPCODE(o) (assert(OFFBASE_IS_REPCODE(o)), (o))  /* returns ID 1,2,3 */
-#define STORED_TO_OFFBASE(o) ((o)+1)
-#define OFFBASE_TO_STORED(o) ((o)-1)
 
 /*! ZSTD_storeSeq() :
- *  Store a sequence (litlen, litPtr, offCode and matchLength) into seqStore_t.
+ *  Store a sequence (litlen, litPtr, offBase and matchLength) into seqStore_t.
- *  @offBase_minus1 : Users should use employ macros STORE_REPCODE_X and STORE_OFFSET().
+ *  @offBase : Users should employ macros REPCODE_TO_OFFBASE() and OFFSET_TO_OFFBASE().
  *  @matchLength : must be >= MINMATCH
- *  Allowed to overread literals up to litLimit.
+ *  Allowed to over-read literals up to litLimit.
 */
 HINT_INLINE UNUSED_ATTR void
 ZSTD_storeSeq(seqStore_t* seqStorePtr,
               size_t litLength, const BYTE* literals, const BYTE* litLimit,
-              U32 offBase_minus1,
+              U32 offBase,
               size_t matchLength)
 {
     BYTE const* const litLimit_w = litLimit - WILDCOPY_OVERLENGTH;
@@ -608,8 +653,8 @@ ZSTD_storeSeq(seqStore_t* seqStorePtr,
     static const BYTE* g_start = NULL;
     if (g_start==NULL) g_start = (const BYTE*)literals;  /* note : index only works for compression within a single segment */
     {   U32 const pos = (U32)((const BYTE*)literals - g_start);
-        DEBUGLOG(6, "Cpos%7u :%3u literals, match%4u bytes at offCode%7u",
+        DEBUGLOG(6, "Cpos%7u :%3u literals, match%4u bytes at offBase%7u",
-               pos, (U32)litLength, (U32)matchLength, (U32)offBase_minus1);
+               pos, (U32)litLength, (U32)matchLength, (U32)offBase);
     }
 #endif
     assert((size_t)(seqStorePtr->sequences - seqStorePtr->sequencesStart) < seqStorePtr->maxNbSeq);
@@ -621,7 +666,7 @@ ZSTD_storeSeq(seqStore_t* seqStorePtr,
         /* Common case we can use wildcopy.
          * First copy 16 bytes, because literals are likely short.
          */
-        assert(WILDCOPY_OVERLENGTH >= 16);
+        ZSTD_STATIC_ASSERT(WILDCOPY_OVERLENGTH >= 16);
         ZSTD_copy16(seqStorePtr->lit, literals);
         if (litLength > 16) {
             ZSTD_wildcopy(seqStorePtr->lit+16, literals+16, (ptrdiff_t)litLength-16, ZSTD_no_overlap);
@@ -640,7 +685,7 @@ ZSTD_storeSeq(seqStore_t* seqStorePtr,
     seqStorePtr->sequences[0].litLength = (U16)litLength;
 
     /* match offset */
-    seqStorePtr->sequences[0].offBase = STORED_TO_OFFBASE(offBase_minus1);
+    seqStorePtr->sequences[0].offBase = offBase;
 
     /* match Length */
     assert(matchLength >= MINMATCH);
@@ -658,17 +703,17 @@ ZSTD_storeSeq(seqStore_t* seqStorePtr,
 
 /* ZSTD_updateRep() :
  * updates in-place @rep (array of repeat offsets)
- * @offBase_minus1 : sum-type, with same numeric representation as ZSTD_storeSeq()
+ * @offBase : sum-type, using numeric representation of ZSTD_storeSeq()
  */
 MEM_STATIC void
-ZSTD_updateRep(U32 rep[ZSTD_REP_NUM], U32 const offBase_minus1, U32 const ll0)
+ZSTD_updateRep(U32 rep[ZSTD_REP_NUM], U32 const offBase, U32 const ll0)
 {
-    if (STORED_IS_OFFSET(offBase_minus1)) {  /* full offset */
+    if (OFFBASE_IS_OFFSET(offBase)) {  /* full offset */
         rep[2] = rep[1];
         rep[1] = rep[0];
-        rep[0] = STORED_OFFSET(offBase_minus1);
+        rep[0] = OFFBASE_TO_OFFSET(offBase);
     } else {   /* repcode */
-        U32 const repCode = STORED_REPCODE(offBase_minus1) - 1 + ll0;
+        U32 const repCode = OFFBASE_TO_REPCODE(offBase) - 1 + ll0;
         if (repCode > 0) {  /* note : if repCode==0, no change */
             U32 const currentOffset = (repCode==ZSTD_REP_NUM) ? (rep[0] - 1) : rep[repCode];
             rep[2] = (repCode >= 2) ? rep[1] : rep[2];
@@ -685,11 +730,11 @@ typedef struct repcodes_s {
 } repcodes_t;
 
 MEM_STATIC repcodes_t
-ZSTD_newRep(U32 const rep[ZSTD_REP_NUM], U32 const offBase_minus1, U32 const ll0)
+ZSTD_newRep(U32 const rep[ZSTD_REP_NUM], U32 const offBase, U32 const ll0)
 {
     repcodes_t newReps;
     ZSTD_memcpy(&newReps, rep, sizeof(newReps));
-    ZSTD_updateRep(newReps.rep, offBase_minus1, ll0);
+    ZSTD_updateRep(newReps.rep, offBase, ll0);
     return newReps;
 }
 
@@ -697,103 +742,6 @@ ZSTD_newRep(U32 const rep[ZSTD_REP_NUM], U32 const offBase_minus1, U32 const ll0
 /*-*************************************
 *  Match length counter
 ***************************************/
-static unsigned ZSTD_NbCommonBytes (size_t val)
-{
-    if (MEM_isLittleEndian()) {
-        if (MEM_64bits()) {
-#       if defined(_MSC_VER) && defined(_WIN64)
-#           if STATIC_BMI2
-                return _tzcnt_u64(val) >> 3;
-#           else
-                if (val != 0) {
-                    unsigned long r;
-                    _BitScanForward64(&r, (U64)val);
-                    return (unsigned)(r >> 3);
-                } else {
-                    /* Should not reach this code path */
-                    __assume(0);
-                }
-#           endif
-#       elif defined(__GNUC__) && (__GNUC__ >= 4)
-            return (__builtin_ctzll((U64)val) >> 3);
-#       else
-            static const int DeBruijnBytePos[64] = { 0, 0, 0, 0, 0, 1, 1, 2,
-                                                     0, 3, 1, 3, 1, 4, 2, 7,
-                                                     0, 2, 3, 6, 1, 5, 3, 5,
-                                                     1, 3, 4, 4, 2, 5, 6, 7,
-                                                     7, 0, 1, 2, 3, 3, 4, 6,
-                                                     2, 6, 5, 5, 3, 4, 5, 6,
-                                                     7, 1, 2, 4, 6, 4, 4, 5,
-                                                     7, 2, 6, 5, 7, 6, 7, 7 };
-            return DeBruijnBytePos[((U64)((val & -(long long)val) * 0x0218A392CDABBD3FULL)) >> 58];
-#       endif
-        } else { /* 32 bits */
-#       if defined(_MSC_VER)
-            if (val != 0) {
-                unsigned long r;
-                _BitScanForward(&r, (U32)val);
-                return (unsigned)(r >> 3);
-            } else {
-                /* Should not reach this code path */
-                __assume(0);
-            }
-#       elif defined(__GNUC__) && (__GNUC__ >= 3)
-            return (__builtin_ctz((U32)val) >> 3);
-#       else
-            static const int DeBruijnBytePos[32] = { 0, 0, 3, 0, 3, 1, 3, 0,
-                                                     3, 2, 2, 1, 3, 2, 0, 1,
-                                                     3, 3, 1, 2, 2, 2, 2, 0,
-                                                     3, 1, 2, 0, 1, 0, 1, 1 };
-            return DeBruijnBytePos[((U32)((val & -(S32)val) * 0x077CB531U)) >> 27];
-#       endif
-        }
-    } else {  /* Big Endian CPU */
-        if (MEM_64bits()) {
-#       if defined(_MSC_VER) && defined(_WIN64)
-#           if STATIC_BMI2
-			    return _lzcnt_u64(val) >> 3;
-#           else
-                if (val != 0) {
-                    unsigned long r;
-                    _BitScanReverse64(&r, (U64)val);
-                    return (unsigned)(r >> 3);
-                } else {
-                    /* Should not reach this code path */
-                    __assume(0);
-                }
-#           endif
-#       elif defined(__GNUC__) && (__GNUC__ >= 4)
-            return (__builtin_clzll(val) >> 3);
-#       else
-            unsigned r;
-            const unsigned n32 = sizeof(size_t)*4;   /* calculate this way due to compiler complaining in 32-bits mode */
-            if (!(val>>n32)) { r=4; } else { r=0; val>>=n32; }
-            if (!(val>>16)) { r+=2; val>>=8; } else { val>>=24; }
-            r += (!val);
-            return r;
-#       endif
-        } else { /* 32 bits */
-#       if defined(_MSC_VER)
-            if (val != 0) {
-                unsigned long r;
-                _BitScanReverse(&r, (unsigned long)val);
-                return (unsigned)(r >> 3);
-            } else {
-                /* Should not reach this code path */
-                __assume(0);
-            }
-#       elif defined(__GNUC__) && (__GNUC__ >= 3)
-            return (__builtin_clz((U32)val) >> 3);
-#       else
-            unsigned r;
-            if (!(val>>16)) { r=2; val>>=8; } else { r=0; val>>=24; }
-            r += (!val);
-            return r;
-#       endif
-    }   }
-}
-
-
 MEM_STATIC size_t ZSTD_count(const BYTE* pIn, const BYTE* pMatch, const BYTE* const pInLimit)
 {
     const BYTE* const pStart = pIn;
@@ -839,32 +787,36 @@ ZSTD_count_2segments(const BYTE* ip, const BYTE* match,
  *  Hashes
  ***************************************/
 static const U32 prime3bytes = 506832829U;
-static U32    ZSTD_hash3(U32 u, U32 h) { return ((u << (32-24)) * prime3bytes)  >> (32-h) ; }
+static U32    ZSTD_hash3(U32 u, U32 h) { assert(h <= 32); return ((u << (32-24)) * prime3bytes)  >> (32-h) ; }
 MEM_STATIC size_t ZSTD_hash3Ptr(const void* ptr, U32 h) { return ZSTD_hash3(MEM_readLE32(ptr), h); } /* only in zstd_opt.h */
 
 static const U32 prime4bytes = 2654435761U;
-static U32    ZSTD_hash4(U32 u, U32 h) { return (u * prime4bytes) >> (32-h) ; }
+static U32    ZSTD_hash4(U32 u, U32 h) { assert(h <= 32); return (u * prime4bytes) >> (32-h) ; }
-static size_t ZSTD_hash4Ptr(const void* ptr, U32 h) { return ZSTD_hash4(MEM_read32(ptr), h); }
+static size_t ZSTD_hash4Ptr(const void* ptr, U32 h) { return ZSTD_hash4(MEM_readLE32(ptr), h); }
 
 static const U64 prime5bytes = 889523592379ULL;
-static size_t ZSTD_hash5(U64 u, U32 h) { return (size_t)(((u  << (64-40)) * prime5bytes) >> (64-h)) ; }
+static size_t ZSTD_hash5(U64 u, U32 h) { assert(h <= 64); return (size_t)(((u  << (64-40)) * prime5bytes) >> (64-h)) ; }
 static size_t ZSTD_hash5Ptr(const void* p, U32 h) { return ZSTD_hash5(MEM_readLE64(p), h); }
 
 static const U64 prime6bytes = 227718039650203ULL;
-static size_t ZSTD_hash6(U64 u, U32 h) { return (size_t)(((u  << (64-48)) * prime6bytes) >> (64-h)) ; }
+static size_t ZSTD_hash6(U64 u, U32 h) { assert(h <= 64); return (size_t)(((u  << (64-48)) * prime6bytes) >> (64-h)) ; }
 static size_t ZSTD_hash6Ptr(const void* p, U32 h) { return ZSTD_hash6(MEM_readLE64(p), h); }
 
 static const U64 prime7bytes = 58295818150454627ULL;
-static size_t ZSTD_hash7(U64 u, U32 h) { return (size_t)(((u  << (64-56)) * prime7bytes) >> (64-h)) ; }
+static size_t ZSTD_hash7(U64 u, U32 h) { assert(h <= 64); return (size_t)(((u  << (64-56)) * prime7bytes) >> (64-h)) ; }
 static size_t ZSTD_hash7Ptr(const void* p, U32 h) { return ZSTD_hash7(MEM_readLE64(p), h); }
 
 static const U64 prime8bytes = 0xCF1BBCDCB7A56463ULL;
-static size_t ZSTD_hash8(U64 u, U32 h) { return (size_t)(((u) * prime8bytes) >> (64-h)) ; }
+static size_t ZSTD_hash8(U64 u, U32 h) { assert(h <= 64); return (size_t)(((u) * prime8bytes) >> (64-h)) ; }
 static size_t ZSTD_hash8Ptr(const void* p, U32 h) { return ZSTD_hash8(MEM_readLE64(p), h); }
 
 MEM_STATIC FORCE_INLINE_ATTR
 size_t ZSTD_hashPtr(const void* p, U32 hBits, U32 mls)
 {
+    /* Although some of these hashes do support hBits up to 64, some do not.
+     * To be on the safe side, always avoid hBits > 32. */
+    assert(hBits <= 32);
+
     switch(mls)
     {
     default:
@@ -1223,10 +1175,15 @@ ZSTD_checkDictValidity(const ZSTD_window_t* window,
                     (unsigned)blockEndIdx, (unsigned)maxDist, (unsigned)loadedDictEnd);
         assert(blockEndIdx >= loadedDictEnd);
 
-        if (blockEndIdx > loadedDictEnd + maxDist) {
+        if (blockEndIdx > loadedDictEnd + maxDist || loadedDictEnd != window->dictLimit) {
             /* On reaching window size, dictionaries are invalidated.
              * For simplification, if window size is reached anywhere within next block,
              * the dictionary is invalidated for the full block.
+             *
+             * We also have to invalidate the dictionary if ZSTD_window_update() has detected
+             * non-contiguous segments, which means that loadedDictEnd != window->dictLimit.
+             * loadedDictEnd may be 0, if forceWindow is true, but in that case we never use
+             * dictMatchState, so setting it to NULL is not a problem.
              */
             DEBUGLOG(6, "invalidating dictionary for current block (distance > windowSize)");
             *loadedDictEndPtr = 0;
@@ -1358,6 +1315,42 @@ MEM_STATIC void ZSTD_debugTable(const U32* table, U32 max)
 
 #endif
 
+/* Short Cache */
+
+/* Normally, zstd matchfinders follow this flow:
+ *     1. Compute hash at ip
+ *     2. Load index from hashTable[hash]
+ *     3. Check if *ip == *(base + index)
+ * In dictionary compression, loading *(base + index) is often an L2 or even L3 miss.
+ *
+ * Short cache is an optimization which allows us to avoid step 3 most of the time
+ * when the data doesn't actually match. With short cache, the flow becomes:
+ *     1. Compute (hash, currentTag) at ip. currentTag is an 8-bit independent hash at ip.
+ *     2. Load (index, matchTag) from hashTable[hash]. See ZSTD_writeTaggedIndex to understand how this works.
+ *     3. Only if currentTag == matchTag, check *ip == *(base + index). Otherwise, continue.
+ *
+ * Currently, short cache is only implemented in CDict hashtables. Thus, its use is limited to
+ * dictMatchState matchfinders.
+ */
+#define ZSTD_SHORT_CACHE_TAG_BITS 8
+#define ZSTD_SHORT_CACHE_TAG_MASK ((1u << ZSTD_SHORT_CACHE_TAG_BITS) - 1)
+
+/* Helper function for ZSTD_fillHashTable and ZSTD_fillDoubleHashTable.
+ * Unpacks hashAndTag into (hash, tag), then packs (index, tag) into hashTable[hash]. */
+MEM_STATIC void ZSTD_writeTaggedIndex(U32* const hashTable, size_t hashAndTag, U32 index) {
+    size_t const hash = hashAndTag >> ZSTD_SHORT_CACHE_TAG_BITS;
+    U32 const tag = (U32)(hashAndTag & ZSTD_SHORT_CACHE_TAG_MASK);
+    assert(index >> (32 - ZSTD_SHORT_CACHE_TAG_BITS) == 0);
+    hashTable[hash] = (index << ZSTD_SHORT_CACHE_TAG_BITS) | tag;
+}
+
+/* Helper function for short cache matchfinders.
+ * Unpacks tag1 and tag2 from lower bits of packedTag1 and packedTag2, then checks if the tags match. */
+MEM_STATIC int ZSTD_comparePackedTags(size_t packedTag1, size_t packedTag2) {
+    U32 const tag1 = packedTag1 & ZSTD_SHORT_CACHE_TAG_MASK;
+    U32 const tag2 = packedTag2 & ZSTD_SHORT_CACHE_TAG_MASK;
+    return tag1 == tag2;
+}
 
 #if defined (__cplusplus)
 }
@@ -1455,4 +1448,31 @@ U32 ZSTD_cycleLog(U32 hashLog, ZSTD_strategy strat);
  */
 void ZSTD_CCtx_trace(ZSTD_CCtx* cctx, size_t extraCSize);
 
+/* Returns 0 on success, and a ZSTD_error otherwise. This function scans through an array of
+ * ZSTD_Sequence, storing the sequences it finds, until it reaches a block delimiter.
+ * Note that the block delimiter must include the last literals of the block.
+ */
+size_t
+ZSTD_copySequencesToSeqStoreExplicitBlockDelim(ZSTD_CCtx* cctx,
+                                              ZSTD_sequencePosition* seqPos,
+                                        const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
+                                        const void* src, size_t blockSize, ZSTD_paramSwitch_e externalRepSearch);
+
+/* Returns the number of bytes to move the current read position back by.
+ * Only non-zero if we ended up splitting a sequence.
+ * Otherwise, it may return a ZSTD error if something went wrong.
+ *
+ * This function will attempt to scan through blockSize bytes
+ * represented by the sequences in @inSeqs,
+ * storing any (partial) sequences.
+ *
+ * Occasionally, we may want to change the actual number of bytes we consumed from inSeqs to
+ * avoid splitting a match, or to avoid splitting a match such that it would produce a match
+ * smaller than MINMATCH. In this case, we return the number of bytes that we didn't read from this block.
+ */
+size_t
+ZSTD_copySequencesToSeqStoreNoBlockDelim(ZSTD_CCtx* cctx, ZSTD_sequencePosition* seqPos,
+                                   const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
+                                   const void* src, size_t blockSize, ZSTD_paramSwitch_e externalRepSearch);
+
 #endif /* ZSTD_COMPRESS_H */

zstd/compress/zstd_compress_literals.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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*)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)
@@ -36,16 +61,30 @@ size_t ZSTD_noCompressLiterals (void* dst, size_t dstCapacity, const void* src,
     }
 
     ZSTD_memcpy(ostart + flSize, src, srcSize);
-    DEBUGLOG(5, "Raw literals: %u -> %u", (U32)srcSize, (U32)(srcSize + flSize));
+    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*)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)
     {
@@ -63,28 +102,51 @@ 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_minLiteralsToCompress() :
-                              ZSTD_hufCTables_t* nextHuf,
+ * returns minimal amount of literals
-                              ZSTD_strategy strategy, int disableLiteralCompression,
+ * 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 int bmi2,
+            const ZSTD_hufCTables_t* prevHuf,
-                        unsigned suspectUncompressible)
+                  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)",
+    DEBUGLOG(5,"ZSTD_compressLiterals (disableLiteralCompression=%i, srcSize=%u, dstCapacity=%zu)",
-                disableLiteralCompression, (U32)srcSize);
+                disableLiteralCompression, (U32)srcSize, dstCapacity);
+
+    DEBUGLOG(6, "Completed literals listing (%zu bytes)", showHexa(src, srcSize));
 
     /* Prepare nextEntropy assuming reusing the existing table */
     ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
@@ -92,40 +154,51 @@ size_t ZSTD_compressLiterals (ZSTD_hufCTables_t const* prevHuf,
     if (disableLiteralCompression)
         return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
 
-    /* small ? don't even attempt compression (speed opt) */
+    /* if too small, don't even attempt compression (speed opt) */
-#   define COMPRESS_LITERALS_SIZE_MIN 63
+    if (srcSize < ZSTD_minLiteralsToCompress(strategy, prevHuf->repeatMode))
-    {   size_t const minLitSize = (prevHuf->repeatMode == HUF_repeat_valid) ? 6 : COMPRESS_LITERALS_SIZE_MIN;
+        return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
-        if (srcSize <= minLitSize) 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_compress = singleStream ? HUF_compress1X_repeat : HUF_compress4X_repeat;
-            HUF_compress1X_repeat(
+        cLitSize = huf_compress(ostart+lhSize, dstCapacity-lhSize,
-                ostart+lhSize, dstCapacity-lhSize, src, srcSize,
+                                src, srcSize,
-                HUF_SYMBOLVALUE_MAX, HUF_TABLELOG_DEFAULT, entropyWorkspace, entropyWorkspaceSize,
+                                HUF_SYMBOLVALUE_MAX, LitHufLog,
-                (HUF_CElt*)nextHuf->CTable, &repeat, preferRepeat, bmi2, suspectUncompressible) :
+                                entropyWorkspace, entropyWorkspaceSize,
-            HUF_compress4X_repeat(
+                                (HUF_CElt*)nextHuf->CTable,
-                ostart+lhSize, dstCapacity-lhSize, src, srcSize,
+                                &repeat, flags);
-                HUF_SYMBOLVALUE_MAX, HUF_TABLELOG_DEFAULT, entropyWorkspace, entropyWorkspaceSize,
+        DEBUGLOG(5, "%zu literals compressed into %zu bytes (before header)", srcSize, cLitSize);
-                (HUF_CElt*)nextHuf->CTable, &repeat, preferRepeat, bmi2, suspectUncompressible);
         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;
         }
     }
 
+    {   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) {
+        /* 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 */
@@ -136,16 +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);

zstd/compress/zstd_compress_literals.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -16,16 +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);
 
-/* If suspectUncompressible then some sampling checks will be run to potentially skip huffman coding */
+/* ZSTD_compressLiterals():
-size_t ZSTD_compressLiterals (ZSTD_hufCTables_t const* prevHuf,
+ * @entropyWorkspace: must be aligned on 4-bytes boundaries
-                              ZSTD_hufCTables_t* nextHuf,
+ * @entropyWorkspaceSize : must be >= HUF_WORKSPACE_SIZE
-                              ZSTD_strategy strategy, int disableLiteralCompression,
+ * @suspectUncompressible: sampling checks, to potentially skip huffman coding
-                              void* dst, size_t dstCapacity,
+ */
+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,
-                        unsigned suspectUncompressible);
+                              ZSTD_hufCTables_t* nextHuf,
+                              ZSTD_strategy strategy, int disableLiteralCompression,
+                              int suspectUncompressible,
+                              int bmi2);
 
 #endif /* ZSTD_COMPRESS_LITERALS_H */

zstd/compress/zstd_compress_sequences.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -58,7 +58,7 @@ 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
-     * comprssibility.
+     * compressibility.
      */
     return nbSeq >= 2048;
 }
@@ -166,7 +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.
              */

zstd/compress/zstd_compress_sequences.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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

zstd/compress/zstd_compress_superblock.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -36,9 +36,10 @@
  *      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,
+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,
@@ -53,8 +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;
@@ -76,9 +75,9 @@ static size_t ZSTD_compressSubBlock_literal(const HUF_CElt* hufTable,
         DEBUGLOG(5, "ZSTD_compressSubBlock_literal (hSize=%zu)", hufMetadata->hufDesSize);
     }
 
-    /* TODO bmi2 */
+    {   int const flags = bmi2 ? HUF_flags_bmi2 : 0;
-    {   const size_t cSize = singleStream ? HUF_compress1X_usingCTable(op, oend-op, literals, litSize, hufTable)
+        const size_t cSize = singleStream ? HUF_compress1X_usingCTable(op, oend-op, literals, litSize, hufTable, flags)
-                                          : HUF_compress4X_usingCTable(op, oend-op, literals, litSize, hufTable);
+                                          : HUF_compress4X_usingCTable(op, oend-op, literals, litSize, hufTable, flags);
         op += cSize;
         cLitSize += cSize;
         if (cSize == 0 || ERR_isError(cSize)) {
@@ -126,7 +125,11 @@ 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;
@@ -156,7 +159,8 @@ 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,
+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,
@@ -539,7 +543,7 @@ static size_t ZSTD_compressSubBlock_multi(const seqStore_t* seqStorePtr,
             repcodes_t rep;
             ZSTD_memcpy(&rep, prevCBlock->rep, sizeof(rep));
             for (seq = sstart; seq < sp; ++seq) {
-                ZSTD_updateRep(rep.rep, seq->offBase - 1, ZSTD_getSequenceLength(seqStorePtr, seq).litLength == 0);
+                ZSTD_updateRep(rep.rep, seq->offBase, ZSTD_getSequenceLength(seqStorePtr, seq).litLength == 0);
             }
             ZSTD_memcpy(nextCBlock->rep, &rep, sizeof(rep));
         }

zstd/compress/zstd_compress_superblock.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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

zstd/compress/zstd_cwksp.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -348,7 +348,9 @@ ZSTD_cwksp_reserve_internal(ZSTD_cwksp* ws, size_t bytes, ZSTD_cwksp_alloc_phase
     if (alloc) {
         alloc = (BYTE *)alloc + ZSTD_CWKSP_ASAN_REDZONE_SIZE;
         if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) {
-            __asan_unpoison_memory_region(alloc, bytes);
+            /* 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
@@ -499,7 +501,7 @@ MEM_STATIC void ZSTD_cwksp_clean_tables(ZSTD_cwksp* ws) {
     assert(ws->tableValidEnd >= ws->objectEnd);
     assert(ws->tableValidEnd <= ws->allocStart);
     if (ws->tableValidEnd < ws->tableEnd) {
-        ZSTD_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);
 }

zstd/compress/zstd_double_fast.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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;
@@ -46,6 +81,18 @@ void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms,
         }   }
 }
 
+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(
@@ -67,7 +114,7 @@ size_t ZSTD_compressBlock_doubleFast_noDict_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;
+    U32 offsetSaved1 = 0, offsetSaved2 = 0;
 
     size_t mLength;
     U32 offset;
@@ -100,8 +147,8 @@ size_t ZSTD_compressBlock_doubleFast_noDict_generic(
         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_2 > maxRep) offsetSaved2 = offset_2, offset_2 = 0;
-        if (offset_1 > maxRep) offsetSaved = offset_1, offset_1 = 0;
+        if (offset_1 > maxRep) offsetSaved1 = offset_1, offset_1 = 0;
     }
 
     /* Outer Loop: one iteration per match found and stored */
@@ -131,7 +178,7 @@ size_t ZSTD_compressBlock_doubleFast_noDict_generic(
             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, STORE_REPCODE_1, mLength);
+                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength);
                 goto _match_stored;
             }
 
@@ -175,9 +222,13 @@ size_t ZSTD_compressBlock_doubleFast_noDict_generic(
         } 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 : offsetSaved;
+        rep[0] = offset_1 ? offset_1 : offsetSaved1;
-        rep[1] = offset_2 ? offset_2 : offsetSaved;
+        rep[1] = offset_2 ? offset_2 : offsetSaved2;
 
         /* Return the last literals size */
         return (size_t)(iend - anchor);
@@ -217,7 +268,7 @@ _match_found: /* requires ip, offset, mLength */
             hashLong[hl1] = (U32)(ip1 - base);
         }
 
-        ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength);
+        ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
 
 _match_stored:
         /* match found */
@@ -243,7 +294,7 @@ _match_stored:
                 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, STORE_REPCODE_1, rLength);
+                ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, rLength);
                 ip += rLength;
                 anchor = ip;
                 continue;   /* faster when present ... (?) */
@@ -275,7 +326,6 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState_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 = &dms->cParams;
@@ -286,8 +336,8 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic(
     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;
+    const U32 dictHBitsL           = dictCParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS;
-    const U32 dictHBitsS           = dictCParams->chainLog;
+    const U32 dictHBitsS           = dictCParams->chainLog + ZSTD_SHORT_CACHE_TAG_BITS;
     const U32 dictAndPrefixLength  = (U32)((ip - prefixLowest) + (dictEnd - dictStart));
 
     DEBUGLOG(5, "ZSTD_compressBlock_doubleFast_dictMatchState_generic");
@@ -295,6 +345,13 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic(
     /* if a dictionary is attached, it must be within window range */
     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);
 
@@ -309,8 +366,12 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState_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 dictHashAndTagL = ZSTD_hashPtr(ip, dictHBitsL, 8);
-        size_t const dictHS = ZSTD_hashPtr(ip, dictHBitsS, mls);
+        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];
@@ -328,7 +389,7 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic(
             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, STORE_REPCODE_1, mLength);
+            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength);
             goto _match_stored;
         }
 
@@ -340,9 +401,9 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic(
                 while (((ip>anchor) & (matchLong>prefixLowest)) && (ip[-1] == matchLong[-1])) { ip--; matchLong--; mLength++; } /* catch up */
                 goto _match_found;
             }
-        } else {
+        } 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);
 
@@ -358,9 +419,9 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic(
             if (MEM_read32(match) == MEM_read32(ip)) {
                 goto _search_next_long;
             }
-        } else {
+        } 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;
 
@@ -375,10 +436,11 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState_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] = curr + 1;
 
@@ -391,9 +453,9 @@ _search_next_long:
                     while (((ip>anchor) & (matchL3>prefixLowest)) && (ip[-1] == matchL3[-1])) { ip--; matchL3--; mLength++; } /* catch up */
                     goto _match_found;
                 }
-            } else {
+            } 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)) {
@@ -419,7 +481,7 @@ _match_found:
         offset_2 = offset_1;
         offset_1 = offset;
 
-        ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength);
+        ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
 
 _match_stored:
         /* match found */
@@ -448,7 +510,7 @@ _match_stored:
                     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, STORE_REPCODE_1, repLength2);
+                    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;
@@ -461,8 +523,8 @@ _match_stored:
     }   /* while (ip < ilimit) */
 
     /* save reps for next block */
-    rep[0] = offset_1 ? offset_1 : offsetSaved;
+    rep[0] = offset_1;
-    rep[1] = offset_2 ? offset_2 : offsetSaved;
+    rep[1] = offset_2;
 
     /* Return the last literals size */
     return (size_t)(iend - anchor);
@@ -585,7 +647,7 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
             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, STORE_REPCODE_1, mLength);
+            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;
@@ -596,7 +658,7 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
                 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, STORE_OFFSET(offset), mLength);
+                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);
@@ -621,7 +683,7 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
                 }
                 offset_2 = offset_1;
                 offset_1 = offset;
-                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength);
+                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
 
             } else {
                 ip += ((ip-anchor) >> kSearchStrength) + 1;
@@ -653,7 +715,7 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
                     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, STORE_REPCODE_1, repLength2);
+                    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;

zstd/compress/zstd_double_fast.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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);

zstd/compress/zstd_fast.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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)
 {
@@ -25,6 +59,10 @@ 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.
      */
@@ -42,6 +80,18 @@ void ZSTD_fillHashTable(ZSTD_matchState_t* ms,
     }   }   }   }
 }
 
+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
@@ -117,7 +167,7 @@ ZSTD_compressBlock_fast_noDict_generic(
 
     U32 rep_offset1 = rep[0];
     U32 rep_offset2 = rep[1];
-    U32 offsetSaved = 0;
+    U32 offsetSaved1 = 0, offsetSaved2 = 0;
 
     size_t hash0; /* hash for ip0 */
     size_t hash1; /* hash for ip1 */
@@ -141,8 +191,8 @@ ZSTD_compressBlock_fast_noDict_generic(
     {   U32 const curr = (U32)(ip0 - base);
         U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, curr, cParams->windowLog);
         U32 const maxRep = curr - windowLow;
-        if (rep_offset2 > maxRep) offsetSaved = rep_offset2, rep_offset2 = 0;
+        if (rep_offset2 > maxRep) offsetSaved2 = rep_offset2, rep_offset2 = 0;
-        if (rep_offset1 > maxRep) offsetSaved = rep_offset1, rep_offset1 = 0;
+        if (rep_offset1 > maxRep) offsetSaved1 = rep_offset1, rep_offset1 = 0;
     }
 
     /* start each op */
@@ -180,8 +230,14 @@ _start: /* Requires: ip0 */
             mLength = ip0[-1] == match0[-1];
             ip0 -= mLength;
             match0 -= mLength;
-            offcode = STORE_REPCODE_1;
+            offcode = REPCODE1_TO_OFFBASE;
             mLength += 4;
+
+            /* 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;
         }
 
@@ -195,6 +251,12 @@ _start: /* Requires: 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.
+             * 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;
         }
 
@@ -224,6 +286,21 @@ _start: /* Requires: 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);
+            }
+
             goto _offset;
         }
 
@@ -254,9 +331,24 @@ _cleanup:
      * 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] = rep_offset1 ? rep_offset1 : offsetSaved;
+    rep[0] = rep_offset1 ? rep_offset1 : offsetSaved1;
-    rep[1] = rep_offset2 ? rep_offset2 : offsetSaved;
+    rep[1] = rep_offset2 ? rep_offset2 : offsetSaved2;
 
     /* Return the last literals size */
     return (size_t)(iend - anchor);
@@ -267,7 +359,7 @@ _offset: /* Requires: ip0, idx */
     match0 = base + idx;
     rep_offset2 = rep_offset1;
     rep_offset1 = (U32)(ip0-match0);
-    offcode = STORE_OFFSET(rep_offset1);
+    offcode = OFFSET_TO_OFFBASE(rep_offset1);
     mLength = 4;
 
     /* Count the backwards match length. */
@@ -287,11 +379,6 @@ _match: /* Requires: ip0, match0, offcode */
     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 */
@@ -306,7 +393,7 @@ _match: /* Requires: ip0, match0, offcode */
                 { 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, STORE_REPCODE_1, rLength);
+                ZSTD_storeSeq(seqStore, 0 /*litLen*/, anchor, iend, REPCODE1_TO_OFFBASE, rLength);
                 anchor = ip0;
                 continue;   /* faster when present (confirmed on gcc-8) ... (?) */
     }   }   }
@@ -380,14 +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 ;
@@ -397,13 +484,13 @@ 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 dictAndPrefixLength  = (U32)(istart - prefixStart + dictEnd - dictStart);
-    const U32 dictHLog             = dictCParams->hashLog;
+    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 */
 
@@ -413,106 +500,155 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic(
      * 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 */
+    /* Outer search loop */
-    while (ip < ilimit) {   /* < instead of <=, because repcode check at (ip+1) */
+    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);
+        size_t hash0 = ZSTD_hashPtr(ip0, hlog, mls);
-        U32 const curr = (U32)(ip-base);
+
-        U32 const matchIndex = hashTable[h];
+        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;
-        hashTable[h] = curr;   /* update hash table */
+            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 */
+            if (((U32) ((prefixStartIndex - 1) - repIndex) >=
-          && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
+                 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(ip+1+4, repMatch+4, iend, repMatchEnd, prefixStart) + 4;
+                mLength = ZSTD_count_2segments(ip0 + 1 + 4, repMatch + 4, iend, repMatchEnd, prefixStart) + 4;
-            ip++;
+                ip0++;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_REPCODE_1, mLength);
+                ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength);
-        } else if ( (matchIndex <= prefixStartIndex) ) {
+                break;
-            size_t const dictHash = ZSTD_hashPtr(ip, dictHLog, mls);
+            }
-            U32 const dictMatchIndex = dictHashTable[dictHash];
+
+            if (dictTagsMatch) {
+                /* Found a possible dict match */
+                const U32 dictMatchIndex = dictMatchIndexAndTag >> ZSTD_SHORT_CACHE_TAG_BITS;
                 const BYTE* dictMatch = dictBase + dictMatchIndex;
-            if (dictMatchIndex <= dictStartIndex ||
+                if (dictMatchIndex > dictStartIndex &&
-                MEM_read32(dictMatch) != MEM_read32(ip)) {
+                    MEM_read32(dictMatch) == MEM_read32(ip0)) {
-                assert(stepSize >= 1);
+                    /* To replicate extDict parse behavior, we only use dict matches when the normal matchIndex is invalid */
-                ip += ((ip-anchor) >> kSearchStrength) + stepSize;
+                    if (matchIndex <= prefixStartIndex) {
-                continue;
+                        U32 const offset = (U32) (curr - dictMatchIndex - dictIndexDelta);
-            } else {
+                        mLength = ZSTD_count_2segments(ip0 + 4, dictMatch + 4, iend, dictEnd, prefixStart) + 4;
-                /* found a dict match */
+                        while (((ip0 > anchor) & (dictMatch > dictStart))
-                U32 const offset = (U32)(curr-dictMatchIndex-dictIndexDelta);
+                            && (ip0[-1] == dictMatch[-1])) {
-                mLength = ZSTD_count_2segments(ip+4, dictMatch+4, iend, dictEnd, prefixStart) + 4;
+                            ip0--;
-                while (((ip>anchor) & (dictMatch>dictStart))
+                            dictMatch--;
-                     && (ip[-1] == dictMatch[-1])) {
+                            mLength++;
-                    ip--; dictMatch--; mLength++;
                         } /* catch up */
                         offset_2 = offset_1;
                         offset_1 = offset;
-                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength);
+                        ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
-            }
-        } 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, STORE_OFFSET(offset), mLength);
-        }
-
-        /* match found */
-        ip += mLength;
-        anchor = ip;
-
-        if (ip <= ilimit) {
-            /* Fill Table */
-            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(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* repMatch2 = repIndex2 < prefixStartIndex ?
-                        dictBase - dictIndexDelta + repIndex2 :
-                        base + repIndex2;
-                if ( ((U32)((prefixStartIndex-1) - (U32)repIndex2) >= 3 /* 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 tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset;   /* swap offset_2 <=> offset_1 */
-                    ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, repLength2);
-                    hashTable[ZSTD_hashPtr(ip, hlog, mls)] = current2;
-                    ip += repLength2;
-                    anchor = ip;
-                    continue;
-                }
                         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) (ip0 - anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
+                break;
+            }
+
+            /* 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 */
+        assert(mLength);
+        ip0 += mLength;
+        anchor = ip0;
+
+        if (ip0 <= ilimit) {
+            /* Fill Table */
+            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 (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(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 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);
+                    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[0] = offset_1;
-    rep[1] = offset_2 ? offset_2 : offsetSaved;
+    rep[1] = offset_2;
 
     /* Return the last literals size */
     return (size_t)(iend - anchor);
@@ -553,11 +689,10 @@ static size_t ZSTD_compressBlock_fast_extDict_generic(
     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);
@@ -570,6 +705,28 @@ 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 */
 
@@ -579,75 +736,202 @@ static size_t ZSTD_compressBlock_fast_extDict_generic(
     if (prefixStartIndex == dictStartIndex)
         return ZSTD_compressBlock_fast(ms, seqStore, rep, src, srcSize);
 
-    /* Search Loop */
+    {   U32 const curr = (U32)(ip0 - base);
-    while (ip < ilimit) {  /* < instead of <=, because (ip+1) */
+        U32 const maxRep = curr - dictStartIndex;
-        const size_t h = ZSTD_hashPtr(ip, hlog, mls);
+        if (offset_2 >= maxRep) offsetSaved2 = offset_2, offset_2 = 0;
-        const U32    matchIndex = hashTable[h];
+        if (offset_1 >= maxRep) offsetSaved1 = offset_1, offset_1 = 0;
-        const BYTE* const matchBase = matchIndex < prefixStartIndex ? dictBase : base;
-        const BYTE*  match = matchBase + matchIndex;
-        const U32    curr = (U32)(ip-base);
-        const U32    repIndex = curr + 1 - offset_1;
-        const BYTE* const repBase = repIndex < prefixStartIndex ? dictBase : base;
-        const BYTE* const repMatch = repBase + repIndex;
-        hashTable[h] = curr;   /* update hash table */
-        DEBUGLOG(7, "offset_1 = %u , curr = %u", offset_1, curr);
-
-        if ( ( ((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow */
-             & (offset_1 <= curr+1 - dictStartIndex) ) /* note: we are searching at curr+1 */
-           && (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, STORE_REPCODE_1, rLength);
-            ip += rLength;
-            anchor = ip;
-        } else {
-            if ( (matchIndex < dictStartIndex) ||
-                 (MEM_read32(match) != MEM_read32(ip)) ) {
-                assert(stepSize >= 1);
-                ip += ((ip-anchor) >> kSearchStrength) + stepSize;
-                continue;
     }
-            {   const BYTE* const matchEnd = matchIndex < prefixStartIndex ? dictEnd : iend;
+
-                const BYTE* const lowMatchPtr = matchIndex < prefixStartIndex ? dictStart : prefixStart;
+    /* start each op */
-                U32 const offset = curr - matchIndex;
+_start: /* Requires: ip0 */
-                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 */
+    step = stepSize;
-                offset_2 = offset_1; offset_1 = offset;  /* update offset history */
+    nextStep = ip0 + kStepIncr;
-                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength);
+
-                ip += mLength;
+    /* calculate positions, ip0 - anchor == 0, so we skip step calc */
-                anchor = ip;
+    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. */
+            }
+
+            /* 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) {
+        {   /* load match for ip[0] */
-            /* Fill Table */
+            U32 const mval = idx >= dictStartIndex ?
-            hashTable[ZSTD_hashPtr(base+curr+2, hlog, mls)] = curr+2;
+                    MEM_read32(idxBase + idx) :
-            hashTable[ZSTD_hashPtr(ip-2, hlog, mls)] = (U32)(ip-2-base);
+                    MEM_read32(ip0) ^ 1; /* guaranteed not to match */
-            /* check immediate repcode */
+
-            while (ip <= ilimit) {
+            /* check match at ip[0] */
-                U32 const current2 = (U32)(ip-base);
+            if (MEM_read32(ip0) == mval) {
-                U32 const repIndex2 = current2 - offset_2;
+                /* found a match! */
-                const BYTE* const repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2;
+                goto _offset;
-                if ( (((U32)((prefixStartIndex-1) - repIndex2) >= 3) & (offset_2 <= curr - dictStartIndex))  /* intentional overflow */
+        }   }
-                   && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
+
-                    const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
+        /* lookup ip[1] */
-                    size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
+        idx = hashTable[hash1];
-                    { U32 const tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; }  /* swap offset_2 <=> offset_1 */
+        idxBase = idx < prefixStartIndex ? dictBase : base;
-                    ZSTD_storeSeq(seqStore, 0 /*litlen*/, anchor, iend, STORE_REPCODE_1, repLength2);
+
-                    hashTable[ZSTD_hashPtr(ip, hlog, mls)] = current2;
+        /* hash ip[2] */
-                    ip += repLength2;
+        hash0 = hash1;
-                    anchor = ip;
+        hash1 = ZSTD_hashPtr(ip2, hlog, mls);
-                    continue;
+
+        /* 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;
         }
-                break;
+    } 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[0] = offset_1 ? offset_1 : offsetSaved1;
-    rep[1] = offset_2;
+    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)
@@ -660,6 +944,7 @@ size_t ZSTD_compressBlock_fast_extDict(
         void const* src, size_t srcSize)
 {
     U32 const mls = ms->cParams.minMatch;
+    assert(ms->dictMatchState == NULL);
     switch(mls)
     {
     default: /* includes case 3 */

zstd/compress/zstd_fast.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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);

zstd/compress/zstd_lazy.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -10,6 +10,7 @@
 
 #include "zstd_compress_internal.h"
 #include "zstd_lazy.h"
+#include "../common/bits.h" /* ZSTD_countTrailingZeros64 */
 
 
 /*-*************************************
@@ -197,8 +198,8 @@ ZSTD_DUBT_findBetterDictMatch (
             U32 matchIndex = dictMatchIndex + dictIndexDelta;
             if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(curr-matchIndex+1) - ZSTD_highbit32((U32)offsetPtr[0]+1)) ) {
                 DEBUGLOG(9, "ZSTD_DUBT_findBetterDictMatch(%u) : found better match length %u -> %u and offsetCode %u -> %u (dictMatchIndex %u, matchIndex %u)",
-                    curr, (U32)bestLength, (U32)matchLength, (U32)*offsetPtr, STORE_OFFSET(curr - matchIndex), dictMatchIndex, matchIndex);
+                    curr, (U32)bestLength, (U32)matchLength, (U32)*offsetPtr, OFFSET_TO_OFFBASE(curr - matchIndex), dictMatchIndex, matchIndex);
-                bestLength = matchLength, *offsetPtr = STORE_OFFSET(curr - matchIndex);
+                bestLength = matchLength, *offsetPtr = OFFSET_TO_OFFBASE(curr - matchIndex);
             }
             if (ip+matchLength == iend) {   /* reached end of input : ip[matchLength] is not valid, no way to know if it's larger or smaller than match */
                 break;   /* drop, to guarantee consistency (miss a little bit of compression) */
@@ -218,7 +219,7 @@ ZSTD_DUBT_findBetterDictMatch (
     }
 
     if (bestLength >= MINMATCH) {
-        U32 const mIndex = curr - (U32)STORED_OFFSET(*offsetPtr); (void)mIndex;
+        U32 const mIndex = curr - (U32)OFFBASE_TO_OFFSET(*offsetPtr); (void)mIndex;
         DEBUGLOG(8, "ZSTD_DUBT_findBetterDictMatch(%u) : found match of length %u and offsetCode %u (pos %u)",
                     curr, (U32)bestLength, (U32)*offsetPtr, mIndex);
     }
@@ -230,7 +231,7 @@ ZSTD_DUBT_findBetterDictMatch (
 static size_t
 ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms,
                         const BYTE* const ip, const BYTE* const iend,
-                        size_t* offsetPtr,
+                        size_t* offBasePtr,
                         U32 const mls,
                         const ZSTD_dictMode_e dictMode)
 {
@@ -327,8 +328,8 @@ ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms,
             if (matchLength > bestLength) {
                 if (matchLength > matchEndIdx - matchIndex)
                     matchEndIdx = matchIndex + (U32)matchLength;
-                if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(curr-matchIndex+1) - ZSTD_highbit32((U32)offsetPtr[0]+1)) )
+                if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(curr - matchIndex + 1) - ZSTD_highbit32((U32)*offBasePtr)) )
-                    bestLength = matchLength, *offsetPtr = STORE_OFFSET(curr - matchIndex);
+                    bestLength = matchLength, *offBasePtr = OFFSET_TO_OFFBASE(curr - matchIndex);
                 if (ip+matchLength == iend) {   /* equal : no way to know if inf or sup */
                     if (dictMode == ZSTD_dictMatchState) {
                         nbCompares = 0; /* in addition to avoiding checking any
@@ -361,16 +362,16 @@ ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms,
         if (dictMode == ZSTD_dictMatchState && nbCompares) {
             bestLength = ZSTD_DUBT_findBetterDictMatch(
                     ms, ip, iend,
-                    offsetPtr, bestLength, nbCompares,
+                    offBasePtr, bestLength, nbCompares,
                     mls, dictMode);
         }
 
         assert(matchEndIdx > curr+8); /* ensure nextToUpdate is increased */
         ms->nextToUpdate = matchEndIdx - 8;   /* skip repetitive patterns */
         if (bestLength >= MINMATCH) {
-            U32 const mIndex = curr - (U32)STORED_OFFSET(*offsetPtr); (void)mIndex;
+            U32 const mIndex = curr - (U32)OFFBASE_TO_OFFSET(*offBasePtr); (void)mIndex;
             DEBUGLOG(8, "ZSTD_DUBT_findBestMatch(%u) : found match of length %u and offsetCode %u (pos %u)",
-                        curr, (U32)bestLength, (U32)*offsetPtr, mIndex);
+                        curr, (U32)bestLength, (U32)*offBasePtr, mIndex);
         }
         return bestLength;
     }
@@ -381,14 +382,14 @@ ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms,
 FORCE_INLINE_TEMPLATE size_t
 ZSTD_BtFindBestMatch( ZSTD_matchState_t* ms,
                 const BYTE* const ip, const BYTE* const iLimit,
-                      size_t* offsetPtr,
+                      size_t* offBasePtr,
                 const U32 mls /* template */,
                 const ZSTD_dictMode_e dictMode)
 {
     DEBUGLOG(7, "ZSTD_BtFindBestMatch");
     if (ip < ms->window.base + ms->nextToUpdate) return 0;   /* skipped area */
     ZSTD_updateDUBT(ms, ip, iLimit, mls);
-    return ZSTD_DUBT_findBestMatch(ms, ip, iLimit, offsetPtr, mls, dictMode);
+    return ZSTD_DUBT_findBestMatch(ms, ip, iLimit, offBasePtr, mls, dictMode);
 }
 
 /***********************************
@@ -561,7 +562,7 @@ size_t ZSTD_dedicatedDictSearch_lazy_search(size_t* offsetPtr, size_t ml, U32 nb
         /* save best solution */
         if (currentMl > ml) {
             ml = currentMl;
-            *offsetPtr = STORE_OFFSET(curr - (matchIndex + ddsIndexDelta));
+            *offsetPtr = OFFSET_TO_OFFBASE(curr - (matchIndex + ddsIndexDelta));
             if (ip+currentMl == iLimit) {
                 /* best possible, avoids read overflow on next attempt */
                 return ml;
@@ -598,7 +599,7 @@ size_t ZSTD_dedicatedDictSearch_lazy_search(size_t* offsetPtr, size_t ml, U32 nb
             /* save best solution */
             if (currentMl > ml) {
                 ml = currentMl;
-                *offsetPtr = STORE_OFFSET(curr - (matchIndex + ddsIndexDelta));
+                *offsetPtr = OFFSET_TO_OFFBASE(curr - (matchIndex + ddsIndexDelta));
                 if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
             }
         }
@@ -691,7 +692,8 @@ size_t ZSTD_HcFindBestMatch(
         if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) {
             const BYTE* const match = base + matchIndex;
             assert(matchIndex >= dictLimit);   /* ensures this is true if dictMode != ZSTD_extDict */
-            if (match[ml] == ip[ml])   /* potentially better */
+            /* read 4B starting from (match + ml + 1 - sizeof(U32)) */
+            if (MEM_read32(match + ml - 3) == MEM_read32(ip + ml - 3))   /* potentially better */
                 currentMl = ZSTD_count(ip, match, iLimit);
         } else {
             const BYTE* const match = dictBase + matchIndex;
@@ -703,7 +705,7 @@ size_t ZSTD_HcFindBestMatch(
         /* save best solution */
         if (currentMl > ml) {
             ml = currentMl;
-            *offsetPtr = STORE_OFFSET(curr - matchIndex);
+            *offsetPtr = OFFSET_TO_OFFBASE(curr - matchIndex);
             if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
         }
 
@@ -739,7 +741,7 @@ size_t ZSTD_HcFindBestMatch(
             if (currentMl > ml) {
                 ml = currentMl;
                 assert(curr > matchIndex + dmsIndexDelta);
-                *offsetPtr = STORE_OFFSET(curr - (matchIndex + dmsIndexDelta));
+                *offsetPtr = OFFSET_TO_OFFBASE(curr - (matchIndex + dmsIndexDelta));
                 if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
             }
 
@@ -757,7 +759,6 @@ size_t ZSTD_HcFindBestMatch(
 ***********************************/
 /* Constants for row-based hash */
 #define ZSTD_ROW_HASH_TAG_OFFSET 16     /* byte offset of hashes in the match state's tagTable from the beginning of a row */
-#define ZSTD_ROW_HASH_TAG_BITS 8        /* nb bits to use for the tag */
 #define ZSTD_ROW_HASH_TAG_MASK ((1u << ZSTD_ROW_HASH_TAG_BITS) - 1)
 #define ZSTD_ROW_HASH_MAX_ENTRIES 64    /* absolute maximum number of entries per row, for all configurations */
 
@@ -769,38 +770,8 @@ typedef U64 ZSTD_VecMask;   /* Clarifies when we are interacting with a U64 repr
  * Starting from the LSB, returns the idx of the next non-zero bit.
  * Basically counting the nb of trailing zeroes.
  */
-static U32 ZSTD_VecMask_next(ZSTD_VecMask val) {
+MEM_STATIC U32 ZSTD_VecMask_next(ZSTD_VecMask val) {
-    assert(val != 0);
+    return ZSTD_countTrailingZeros64(val);
-#   if defined(_MSC_VER) && defined(_WIN64)
-        if (val != 0) {
-            unsigned long r;
-            _BitScanForward64(&r, val);
-            return (U32)(r);
-        } else {
-            /* Should not reach this code path */
-            __assume(0);
-        }
-#   elif (defined(__GNUC__) && ((__GNUC__ > 3) || ((__GNUC__ == 3) && (__GNUC_MINOR__ >= 4))))
-    if (sizeof(size_t) == 4) {
-        U32 mostSignificantWord = (U32)(val >> 32);
-        U32 leastSignificantWord = (U32)val;
-        if (leastSignificantWord == 0) {
-            return 32 + (U32)__builtin_ctz(mostSignificantWord);
-        } else {
-            return (U32)__builtin_ctz(leastSignificantWord);
-        }
-    } else {
-        return (U32)__builtin_ctzll(val);
-    }
-#   else
-    /* Software ctz version: http://aggregate.org/MAGIC/#Trailing%20Zero%20Count
-     * and: https://stackoverflow.com/questions/2709430/count-number-of-bits-in-a-64-bit-long-big-integer
-     */
-    val = ~val & (val - 1ULL); /* Lowest set bit mask */
-    val = val - ((val >> 1) & 0x5555555555555555);
-    val = (val & 0x3333333333333333ULL) + ((val >> 2) & 0x3333333333333333ULL);
-    return (U32)((((val + (val >> 4)) & 0xF0F0F0F0F0F0F0FULL) * 0x101010101010101ULL) >> 56);
-#   endif
 }
 
 /* ZSTD_rotateRight_*():
@@ -980,7 +951,35 @@ void ZSTD_row_update(ZSTD_matchState_t* const ms, const BYTE* ip) {
     const U32 mls = MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */);
 
     DEBUGLOG(5, "ZSTD_row_update(), rowLog=%u", rowLog);
-    ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 0 /* dont use cache */);
+    ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 0 /* don't use cache */);
+}
+
+/* Returns the mask width of bits group of which will be set to 1. Given not all
+ * architectures have easy movemask instruction, this helps to iterate over
+ * groups of bits easier and faster.
+ */
+FORCE_INLINE_TEMPLATE U32
+ZSTD_row_matchMaskGroupWidth(const U32 rowEntries)
+{
+    assert((rowEntries == 16) || (rowEntries == 32) || rowEntries == 64);
+    assert(rowEntries <= ZSTD_ROW_HASH_MAX_ENTRIES);
+    (void)rowEntries;
+#if defined(ZSTD_ARCH_ARM_NEON)
+    /* NEON path only works for little endian */
+    if (!MEM_isLittleEndian()) {
+        return 1;
+    }
+    if (rowEntries == 16) {
+        return 4;
+    }
+    if (rowEntries == 32) {
+        return 2;
+    }
+    if (rowEntries == 64) {
+        return 1;
+    }
+#endif
+    return 1;
 }
 
 #if defined(ZSTD_ARCH_X86_SSE2)
@@ -1003,51 +1002,35 @@ ZSTD_row_getSSEMask(int nbChunks, const BYTE* const src, const BYTE tag, const U
 }
 #endif
 
-/* Returns a ZSTD_VecMask (U32) that has the nth bit set to 1 if the newly-computed "tag" matches
+#if defined(ZSTD_ARCH_ARM_NEON)
- * the hash at the nth position in a row of the tagTable.
- * Each row is a circular buffer beginning at the value of "head". So we must rotate the "matches" bitfield
- * to match up with the actual layout of the entries within the hashTable */
 FORCE_INLINE_TEMPLATE ZSTD_VecMask
-ZSTD_row_getMatchMask(const BYTE* const tagRow, const BYTE tag, const U32 head, const U32 rowEntries)
+ZSTD_row_getNEONMask(const U32 rowEntries, const BYTE* const src, const BYTE tag, const U32 headGrouped)
 {
-    const BYTE* const src = tagRow + ZSTD_ROW_HASH_TAG_OFFSET;
     assert((rowEntries == 16) || (rowEntries == 32) || rowEntries == 64);
-    assert(rowEntries <= ZSTD_ROW_HASH_MAX_ENTRIES);
-
-#if defined(ZSTD_ARCH_X86_SSE2)
-
-    return ZSTD_row_getSSEMask(rowEntries / 16, src, tag, head);
-
-#else /* SW or NEON-LE */
-
-# if defined(ZSTD_ARCH_ARM_NEON)
-  /* This NEON path only works for little endian - otherwise use SWAR below */
-    if (MEM_isLittleEndian()) {
     if (rowEntries == 16) {
+        /* vshrn_n_u16 shifts by 4 every u16 and narrows to 8 lower bits.
+         * After that groups of 4 bits represent the equalMask. We lower
+         * all bits except the highest in these groups by doing AND with
+         * 0x88 = 0b10001000.
+         */
         const uint8x16_t chunk = vld1q_u8(src);
         const uint16x8_t equalMask = vreinterpretq_u16_u8(vceqq_u8(chunk, vdupq_n_u8(tag)));
-            const uint16x8_t t0 = vshlq_n_u16(equalMask, 7);
+        const uint8x8_t res = vshrn_n_u16(equalMask, 4);
-            const uint32x4_t t1 = vreinterpretq_u32_u16(vsriq_n_u16(t0, t0, 14));
+        const U64 matches = vget_lane_u64(vreinterpret_u64_u8(res), 0);
-            const uint64x2_t t2 = vreinterpretq_u64_u32(vshrq_n_u32(t1, 14));
+        return ZSTD_rotateRight_U64(matches, headGrouped) & 0x8888888888888888ull;
-            const uint8x16_t t3 = vreinterpretq_u8_u64(vsraq_n_u64(t2, t2, 28));
-            const U16 hi = (U16)vgetq_lane_u8(t3, 8);
-            const U16 lo = (U16)vgetq_lane_u8(t3, 0);
-            return ZSTD_rotateRight_U16((hi << 8) | lo, head);
     } else if (rowEntries == 32) {
-            const uint16x8x2_t chunk = vld2q_u16((const U16*)(const void*)src);
+        /* Same idea as with rowEntries == 16 but doing AND with
+         * 0x55 = 0b01010101.
+         */
+        const uint16x8x2_t chunk = vld2q_u16((const uint16_t*)(const void*)src);
         const uint8x16_t chunk0 = vreinterpretq_u8_u16(chunk.val[0]);
         const uint8x16_t chunk1 = vreinterpretq_u8_u16(chunk.val[1]);
-            const uint8x16_t equalMask0 = vceqq_u8(chunk0, vdupq_n_u8(tag));
+        const uint8x16_t dup = vdupq_n_u8(tag);
-            const uint8x16_t equalMask1 = vceqq_u8(chunk1, vdupq_n_u8(tag));
+        const uint8x8_t t0 = vshrn_n_u16(vreinterpretq_u16_u8(vceqq_u8(chunk0, dup)), 6);
-            const int8x8_t pack0 = vqmovn_s16(vreinterpretq_s16_u8(equalMask0));
+        const uint8x8_t t1 = vshrn_n_u16(vreinterpretq_u16_u8(vceqq_u8(chunk1, dup)), 6);
-            const int8x8_t pack1 = vqmovn_s16(vreinterpretq_s16_u8(equalMask1));
+        const uint8x8_t res = vsli_n_u8(t0, t1, 4);
-            const uint8x8_t t0 = vreinterpret_u8_s8(pack0);
+        const U64 matches = vget_lane_u64(vreinterpret_u64_u8(res), 0) ;
-            const uint8x8_t t1 = vreinterpret_u8_s8(pack1);
+        return ZSTD_rotateRight_U64(matches, headGrouped) & 0x5555555555555555ull;
-            const uint8x8_t t2 = vsri_n_u8(t1, t0, 2);
-            const uint8x8x2_t t3 = vuzp_u8(t2, t0);
-            const uint8x8_t t4 = vsri_n_u8(t3.val[1], t3.val[0], 4);
-            const U32 matches = vget_lane_u32(vreinterpret_u32_u8(t4), 0);
-            return ZSTD_rotateRight_U32(matches, head);
     } else { /* rowEntries == 64 */
         const uint8x16x4_t chunk = vld4q_u8(src);
         const uint8x16_t dup = vdupq_n_u8(tag);
@@ -1062,12 +1045,39 @@ ZSTD_row_getMatchMask(const BYTE* const tagRow, const BYTE tag, const U32 head,
         const uint8x16_t t3 = vsriq_n_u8(t2, t2, 4);
         const uint8x8_t t4 = vshrn_n_u16(vreinterpretq_u16_u8(t3), 4);
         const U64 matches = vget_lane_u64(vreinterpret_u64_u8(t4), 0);
-            return ZSTD_rotateRight_U64(matches, head);
+        return ZSTD_rotateRight_U64(matches, headGrouped);
     }
+}
+#endif
+
+/* Returns a ZSTD_VecMask (U64) that has the nth group (determined by
+ * ZSTD_row_matchMaskGroupWidth) of bits set to 1 if the newly-computed "tag"
+ * matches the hash at the nth position in a row of the tagTable.
+ * Each row is a circular buffer beginning at the value of "headGrouped". So we
+ * must rotate the "matches" bitfield to match up with the actual layout of the
+ * entries within the hashTable */
+FORCE_INLINE_TEMPLATE ZSTD_VecMask
+ZSTD_row_getMatchMask(const BYTE* const tagRow, const BYTE tag, const U32 headGrouped, const U32 rowEntries)
+{
+    const BYTE* const src = tagRow + ZSTD_ROW_HASH_TAG_OFFSET;
+    assert((rowEntries == 16) || (rowEntries == 32) || rowEntries == 64);
+    assert(rowEntries <= ZSTD_ROW_HASH_MAX_ENTRIES);
+    assert(ZSTD_row_matchMaskGroupWidth(rowEntries) * rowEntries <= sizeof(ZSTD_VecMask) * 8);
+
+#if defined(ZSTD_ARCH_X86_SSE2)
+
+    return ZSTD_row_getSSEMask(rowEntries / 16, src, tag, headGrouped);
+
+#else /* SW or NEON-LE */
+
+# if defined(ZSTD_ARCH_ARM_NEON)
+  /* This NEON path only works for little endian - otherwise use SWAR below */
+    if (MEM_isLittleEndian()) {
+        return ZSTD_row_getNEONMask(rowEntries, src, tag, headGrouped);
     }
 # endif /* ZSTD_ARCH_ARM_NEON */
     /* SWAR */
-    {   const size_t chunkSize = sizeof(size_t);
+    {   const int chunkSize = sizeof(size_t);
         const size_t shiftAmount = ((chunkSize * 8) - chunkSize);
         const size_t xFF = ~((size_t)0);
         const size_t x01 = xFF / 0xFF;
@@ -1100,11 +1110,11 @@ ZSTD_row_getMatchMask(const BYTE* const tagRow, const BYTE tag, const U32 head,
         }
         matches = ~matches;
         if (rowEntries == 16) {
-            return ZSTD_rotateRight_U16((U16)matches, head);
+            return ZSTD_rotateRight_U16((U16)matches, headGrouped);
         } else if (rowEntries == 32) {
-            return ZSTD_rotateRight_U32((U32)matches, head);
+            return ZSTD_rotateRight_U32((U32)matches, headGrouped);
         } else {
-            return ZSTD_rotateRight_U64((U64)matches, head);
+            return ZSTD_rotateRight_U64((U64)matches, headGrouped);
         }
     }
 #endif
@@ -1152,6 +1162,7 @@ size_t ZSTD_RowFindBestMatch(
     const U32 rowEntries = (1U << rowLog);
     const U32 rowMask = rowEntries - 1;
     const U32 cappedSearchLog = MIN(cParams->searchLog, rowLog); /* nb of searches is capped at nb entries per row */
+    const U32 groupWidth = ZSTD_row_matchMaskGroupWidth(rowEntries);
     U32 nbAttempts = 1U << cappedSearchLog;
     size_t ml=4-1;
 
@@ -1194,15 +1205,15 @@ size_t ZSTD_RowFindBestMatch(
         U32 const tag = hash & ZSTD_ROW_HASH_TAG_MASK;
         U32* const row = hashTable + relRow;
         BYTE* tagRow = (BYTE*)(tagTable + relRow);
-        U32 const head = *tagRow & rowMask;
+        U32 const headGrouped = (*tagRow & rowMask) * groupWidth;
         U32 matchBuffer[ZSTD_ROW_HASH_MAX_ENTRIES];
         size_t numMatches = 0;
         size_t currMatch = 0;
-        ZSTD_VecMask matches = ZSTD_row_getMatchMask(tagRow, (BYTE)tag, head, rowEntries);
+        ZSTD_VecMask matches = ZSTD_row_getMatchMask(tagRow, (BYTE)tag, headGrouped, rowEntries);
 
         /* Cycle through the matches and prefetch */
         for (; (matches > 0) && (nbAttempts > 0); --nbAttempts, matches &= (matches - 1)) {
-            U32 const matchPos = (head + ZSTD_VecMask_next(matches)) & rowMask;
+            U32 const matchPos = ((headGrouped + ZSTD_VecMask_next(matches)) / groupWidth) & rowMask;
             U32 const matchIndex = row[matchPos];
             assert(numMatches < rowEntries);
             if (matchIndex < lowLimit)
@@ -1233,7 +1244,8 @@ size_t ZSTD_RowFindBestMatch(
             if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) {
                 const BYTE* const match = base + matchIndex;
                 assert(matchIndex >= dictLimit);   /* ensures this is true if dictMode != ZSTD_extDict */
-                if (match[ml] == ip[ml])   /* potentially better */
+                /* read 4B starting from (match + ml + 1 - sizeof(U32)) */
+                if (MEM_read32(match + ml - 3) == MEM_read32(ip + ml - 3))   /* potentially better */
                     currentMl = ZSTD_count(ip, match, iLimit);
             } else {
                 const BYTE* const match = dictBase + matchIndex;
@@ -1245,7 +1257,7 @@ size_t ZSTD_RowFindBestMatch(
             /* Save best solution */
             if (currentMl > ml) {
                 ml = currentMl;
-                *offsetPtr = STORE_OFFSET(curr - matchIndex);
+                *offsetPtr = OFFSET_TO_OFFBASE(curr - matchIndex);
                 if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
             }
         }
@@ -1263,14 +1275,14 @@ size_t ZSTD_RowFindBestMatch(
         const U32 dmsSize              = (U32)(dmsEnd - dmsBase);
         const U32 dmsIndexDelta        = dictLimit - dmsSize;
 
-        {   U32 const head = *dmsTagRow & rowMask;
+        {   U32 const headGrouped = (*dmsTagRow & rowMask) * groupWidth;
             U32 matchBuffer[ZSTD_ROW_HASH_MAX_ENTRIES];
             size_t numMatches = 0;
             size_t currMatch = 0;
-            ZSTD_VecMask matches = ZSTD_row_getMatchMask(dmsTagRow, (BYTE)dmsTag, head, rowEntries);
+            ZSTD_VecMask matches = ZSTD_row_getMatchMask(dmsTagRow, (BYTE)dmsTag, headGrouped, rowEntries);
 
             for (; (matches > 0) && (nbAttempts > 0); --nbAttempts, matches &= (matches - 1)) {
-                U32 const matchPos = (head + ZSTD_VecMask_next(matches)) & rowMask;
+                U32 const matchPos = ((headGrouped + ZSTD_VecMask_next(matches)) / groupWidth) & rowMask;
                 U32 const matchIndex = dmsRow[matchPos];
                 if (matchIndex < dmsLowestIndex)
                     break;
@@ -1294,7 +1306,7 @@ size_t ZSTD_RowFindBestMatch(
                 if (currentMl > ml) {
                     ml = currentMl;
                     assert(curr > matchIndex + dmsIndexDelta);
-                    *offsetPtr = STORE_OFFSET(curr - (matchIndex + dmsIndexDelta));
+                    *offsetPtr = OFFSET_TO_OFFBASE(curr - (matchIndex + dmsIndexDelta));
                     if (ip+currentMl == iLimit) break;
                 }
             }
@@ -1304,14 +1316,10 @@ size_t ZSTD_RowFindBestMatch(
 }
 
 
-typedef size_t (*searchMax_f)(
-                    ZSTD_matchState_t* ms,
-                    const BYTE* ip, const BYTE* iLimit, size_t* offsetPtr);
-
 /**
- * This struct contains the functions necessary for lazy to search.
+ * Generate search functions templated on (dictMode, mls, rowLog).
- * Currently, that is only searchMax. However, it is still valuable to have the
+ * These functions are outlined for code size & compilation time.
- * VTable because this makes it easier to add more functions to the VTable later.
+ * ZSTD_searchMax() dispatches to the correct implementation function.
  *
  * TODO: The start of the search function involves loading and calculating a
  * bunch of constants from the ZSTD_matchState_t. These computations could be
@@ -1329,25 +1337,25 @@ typedef size_t (*searchMax_f)(
  * the single segment loop. It should go in searchMax instead of its own
  * function to avoid having multiple virtual function calls per search.
  */
-typedef struct {
-    searchMax_f searchMax;
-} ZSTD_LazyVTable;
 
-#define GEN_ZSTD_BT_VTABLE(dictMode, mls)                                             \
+#define ZSTD_BT_SEARCH_FN(dictMode, mls) ZSTD_BtFindBestMatch_##dictMode##_##mls
-    static size_t ZSTD_BtFindBestMatch_##dictMode##_##mls(                            \
+#define ZSTD_HC_SEARCH_FN(dictMode, mls) ZSTD_HcFindBestMatch_##dictMode##_##mls
+#define ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog) ZSTD_RowFindBestMatch_##dictMode##_##mls##_##rowLog
+
+#define ZSTD_SEARCH_FN_ATTRS FORCE_NOINLINE
+
+#define GEN_ZSTD_BT_SEARCH_FN(dictMode, mls)                                           \
+    ZSTD_SEARCH_FN_ATTRS size_t ZSTD_BT_SEARCH_FN(dictMode, mls)(                      \
             ZSTD_matchState_t* ms,                                                     \
             const BYTE* ip, const BYTE* const iLimit,                                  \
-            size_t* offsetPtr)                                                        \
+            size_t* offBasePtr)                                                        \
     {                                                                                  \
         assert(MAX(4, MIN(6, ms->cParams.minMatch)) == mls);                           \
-        return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, mls, ZSTD_##dictMode); \
+        return ZSTD_BtFindBestMatch(ms, ip, iLimit, offBasePtr, mls, ZSTD_##dictMode); \
     }                                                                                  \
-    static const ZSTD_LazyVTable ZSTD_BtVTable_##dictMode##_##mls = {                 \
-        ZSTD_BtFindBestMatch_##dictMode##_##mls                                       \
-    };
 
-#define GEN_ZSTD_HC_VTABLE(dictMode, mls)                                             \
+#define GEN_ZSTD_HC_SEARCH_FN(dictMode, mls)                                          \
-    static size_t ZSTD_HcFindBestMatch_##dictMode##_##mls(                            \
+    ZSTD_SEARCH_FN_ATTRS size_t ZSTD_HC_SEARCH_FN(dictMode, mls)(                     \
             ZSTD_matchState_t* ms,                                                    \
             const BYTE* ip, const BYTE* const iLimit,                                 \
             size_t* offsetPtr)                                                        \
@@ -1355,12 +1363,9 @@ typedef struct {
         assert(MAX(4, MIN(6, ms->cParams.minMatch)) == mls);                          \
         return ZSTD_HcFindBestMatch(ms, ip, iLimit, offsetPtr, mls, ZSTD_##dictMode); \
     }                                                                                 \
-    static const ZSTD_LazyVTable ZSTD_HcVTable_##dictMode##_##mls = {                 \
-        ZSTD_HcFindBestMatch_##dictMode##_##mls                                       \
-    };
 
-#define GEN_ZSTD_ROW_VTABLE(dictMode, mls, rowLog)                                             \
+#define GEN_ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog)                                          \
-    static size_t ZSTD_RowFindBestMatch_##dictMode##_##mls##_##rowLog(                         \
+    ZSTD_SEARCH_FN_ATTRS size_t ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog)(                     \
             ZSTD_matchState_t* ms,                                                             \
             const BYTE* ip, const BYTE* const iLimit,                                          \
             size_t* offsetPtr)                                                                 \
@@ -1369,9 +1374,6 @@ typedef struct {
         assert(MAX(4, MIN(6, ms->cParams.searchLog)) == rowLog);                               \
         return ZSTD_RowFindBestMatch(ms, ip, iLimit, offsetPtr, mls, ZSTD_##dictMode, rowLog); \
     }                                                                                          \
-    static const ZSTD_LazyVTable ZSTD_RowVTable_##dictMode##_##mls##_##rowLog = {              \
-        ZSTD_RowFindBestMatch_##dictMode##_##mls##_##rowLog                                    \
-    };
 
 #define ZSTD_FOR_EACH_ROWLOG(X, dictMode, mls) \
     X(dictMode, mls, 4)                        \
@@ -1394,83 +1396,102 @@ typedef struct {
     X(__VA_ARGS__, dictMatchState)      \
     X(__VA_ARGS__, dedicatedDictSearch)
 
-/* Generate Row VTables for each combination of (dictMode, mls, rowLog) */
+/* Generate row search fns for each combination of (dictMode, mls, rowLog) */
-ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS_ROWLOG, GEN_ZSTD_ROW_VTABLE)
+ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS_ROWLOG, GEN_ZSTD_ROW_SEARCH_FN)
-/* Generate Binary Tree VTables for each combination of (dictMode, mls) */
+/* Generate binary Tree search fns for each combination of (dictMode, mls) */
-ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS, GEN_ZSTD_BT_VTABLE)
+ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS, GEN_ZSTD_BT_SEARCH_FN)
-/* Generate Hash Chain VTables for each combination of (dictMode, mls) */
+/* Generate hash chain search fns for each combination of (dictMode, mls) */
-ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS, GEN_ZSTD_HC_VTABLE)
+ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS, GEN_ZSTD_HC_SEARCH_FN)
 
-#define GEN_ZSTD_BT_VTABLE_ARRAY(dictMode) \
+typedef enum { search_hashChain=0, search_binaryTree=1, search_rowHash=2 } searchMethod_e;
-    {                                      \
+
-        &ZSTD_BtVTable_##dictMode##_4,     \
+#define GEN_ZSTD_CALL_BT_SEARCH_FN(dictMode, mls)                         \
-        &ZSTD_BtVTable_##dictMode##_5,     \
+    case mls:                                                             \
-        &ZSTD_BtVTable_##dictMode##_6      \
+        return ZSTD_BT_SEARCH_FN(dictMode, mls)(ms, ip, iend, offsetPtr);
+#define GEN_ZSTD_CALL_HC_SEARCH_FN(dictMode, mls)                         \
+    case mls:                                                             \
+        return ZSTD_HC_SEARCH_FN(dictMode, mls)(ms, ip, iend, offsetPtr);
+#define GEN_ZSTD_CALL_ROW_SEARCH_FN(dictMode, mls, rowLog)                         \
+    case rowLog:                                                                   \
+        return ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog)(ms, ip, iend, offsetPtr);
+
+#define ZSTD_SWITCH_MLS(X, dictMode)   \
+    switch (mls) {                     \
+        ZSTD_FOR_EACH_MLS(X, dictMode) \
     }
 
-#define GEN_ZSTD_HC_VTABLE_ARRAY(dictMode) \
+#define ZSTD_SWITCH_ROWLOG(dictMode, mls)                                    \
-    {                                      \
+    case mls:                                                                \
-        &ZSTD_HcVTable_##dictMode##_4,     \
+        switch (rowLog) {                                                    \
-        &ZSTD_HcVTable_##dictMode##_5,     \
+            ZSTD_FOR_EACH_ROWLOG(GEN_ZSTD_CALL_ROW_SEARCH_FN, dictMode, mls) \
-        &ZSTD_HcVTable_##dictMode##_6      \
+        }                                                                    \
-    }
+        ZSTD_UNREACHABLE;                                                    \
+        break;
 
-#define GEN_ZSTD_ROW_VTABLE_ARRAY_(dictMode, mls) \
+#define ZSTD_SWITCH_SEARCH_METHOD(dictMode)                       \
-    {                                             \
+    switch (searchMethod) {                                       \
-        &ZSTD_RowVTable_##dictMode##_##mls##_4,   \
+        case search_hashChain:                                    \
-        &ZSTD_RowVTable_##dictMode##_##mls##_5,   \
+            ZSTD_SWITCH_MLS(GEN_ZSTD_CALL_HC_SEARCH_FN, dictMode) \
-        &ZSTD_RowVTable_##dictMode##_##mls##_6    \
+            break;                                                \
-    }
+        case search_binaryTree:                                   \
+            ZSTD_SWITCH_MLS(GEN_ZSTD_CALL_BT_SEARCH_FN, dictMode) \
+            break;                                                \
+        case search_rowHash:                                      \
+            ZSTD_SWITCH_MLS(ZSTD_SWITCH_ROWLOG, dictMode)         \
+            break;                                                \
+    }                                                             \
+    ZSTD_UNREACHABLE;
 
-#define GEN_ZSTD_ROW_VTABLE_ARRAY(dictMode)      \
+/**
-    {                                            \
+ * Searches for the longest match at @p ip.
-        GEN_ZSTD_ROW_VTABLE_ARRAY_(dictMode, 4), \
+ * Dispatches to the correct implementation function based on the
-        GEN_ZSTD_ROW_VTABLE_ARRAY_(dictMode, 5), \
+ * (searchMethod, dictMode, mls, rowLog). We use switch statements
-        GEN_ZSTD_ROW_VTABLE_ARRAY_(dictMode, 6)  \
+ * here instead of using an indirect function call through a function
-    }
+ * pointer because after Spectre and Meltdown mitigations, indirect
-
+ * function calls can be very costly, especially in the kernel.
-#define GEN_ZSTD_VTABLE_ARRAY(X) \
+ *
-    {                            \
+ * NOTE: dictMode and searchMethod should be templated, so those switch
-        X(noDict),               \
+ * statements should be optimized out. Only the mls & rowLog switches
-        X(extDict),              \
+ * should be left.
-        X(dictMatchState),       \
+ *
-        X(dedicatedDictSearch)   \
+ * @param ms The match state.
+ * @param ip The position to search at.
+ * @param iend The end of the input data.
+ * @param[out] offsetPtr Stores the match offset into this pointer.
+ * @param mls The minimum search length, in the range [4, 6].
+ * @param rowLog The row log (if applicable), in the range [4, 6].
+ * @param searchMethod The search method to use (templated).
+ * @param dictMode The dictMode (templated).
+ *
+ * @returns The length of the longest match found, or < mls if no match is found.
+ * If a match is found its offset is stored in @p offsetPtr.
+ */
+FORCE_INLINE_TEMPLATE size_t ZSTD_searchMax(
+    ZSTD_matchState_t* ms,
+    const BYTE* ip,
+    const BYTE* iend,
+    size_t* offsetPtr,
+    U32 const mls,
+    U32 const rowLog,
+    searchMethod_e const searchMethod,
+    ZSTD_dictMode_e const dictMode)
+{
+    if (dictMode == ZSTD_noDict) {
+        ZSTD_SWITCH_SEARCH_METHOD(noDict)
+    } else if (dictMode == ZSTD_extDict) {
+        ZSTD_SWITCH_SEARCH_METHOD(extDict)
+    } else if (dictMode == ZSTD_dictMatchState) {
+        ZSTD_SWITCH_SEARCH_METHOD(dictMatchState)
+    } else if (dictMode == ZSTD_dedicatedDictSearch) {
+        ZSTD_SWITCH_SEARCH_METHOD(dedicatedDictSearch)
     }
+    ZSTD_UNREACHABLE;
+    return 0;
+}
 
 /* *******************************
 *  Common parser - lazy strategy
 *********************************/
-typedef enum { search_hashChain=0, search_binaryTree=1, search_rowHash=2 } searchMethod_e;
-
-/**
- * This table is indexed first by the four ZSTD_dictMode_e values, and then
- * by the two searchMethod_e values. NULLs are placed for configurations
- * that should never occur (extDict modes go to the other implementation
- * below and there is no DDSS for binary tree search yet).
- */
-
-static ZSTD_LazyVTable const*
-ZSTD_selectLazyVTable(ZSTD_matchState_t const* ms, searchMethod_e searchMethod, ZSTD_dictMode_e dictMode)
-{
-    /* Fill the Hc/Bt VTable arrays with the right functions for the (dictMode, mls) combination. */
-    ZSTD_LazyVTable const* const hcVTables[4][3] = GEN_ZSTD_VTABLE_ARRAY(GEN_ZSTD_HC_VTABLE_ARRAY);
-    ZSTD_LazyVTable const* const btVTables[4][3] = GEN_ZSTD_VTABLE_ARRAY(GEN_ZSTD_BT_VTABLE_ARRAY);
-    /* Fill the Row VTable array with the right functions for the (dictMode, mls, rowLog) combination. */
-    ZSTD_LazyVTable const* const rowVTables[4][3][3] = GEN_ZSTD_VTABLE_ARRAY(GEN_ZSTD_ROW_VTABLE_ARRAY);
-
-    U32 const mls = MAX(4, MIN(6, ms->cParams.minMatch));
-    U32 const rowLog = MAX(4, MIN(6, ms->cParams.searchLog));
-    switch (searchMethod) {
-        case search_hashChain:
-            return hcVTables[dictMode][mls - 4];
-        case search_binaryTree:
-            return btVTables[dictMode][mls - 4];
-        case search_rowHash:
-            return rowVTables[dictMode][mls - 4][rowLog - 4];
-        default:
-            return NULL;
-    }
-}
 
 FORCE_INLINE_TEMPLATE size_t
 ZSTD_compressBlock_lazy_generic(
@@ -1488,9 +1509,11 @@ ZSTD_compressBlock_lazy_generic(
     const BYTE* const base = ms->window.base;
     const U32 prefixLowestIndex = ms->window.dictLimit;
     const BYTE* const prefixLowest = base + prefixLowestIndex;
+    const U32 mls = BOUNDED(4, ms->cParams.minMatch, 6);
+    const U32 rowLog = BOUNDED(4, ms->cParams.searchLog, 6);
 
-    searchMax_f const searchMax = ZSTD_selectLazyVTable(ms, searchMethod, dictMode)->searchMax;
+    U32 offset_1 = rep[0], offset_2 = rep[1];
-    U32 offset_1 = rep[0], offset_2 = rep[1], savedOffset=0;
+    U32 offsetSaved1 = 0, offsetSaved2 = 0;
 
     const int isDMS = dictMode == ZSTD_dictMatchState;
     const int isDDS = dictMode == ZSTD_dedicatedDictSearch;
@@ -1505,16 +1528,14 @@ ZSTD_compressBlock_lazy_generic(
                                      0;
     const U32 dictAndPrefixLength = (U32)((ip - prefixLowest) + (dictEnd - dictLowest));
 
-    assert(searchMax != NULL);
-
     DEBUGLOG(5, "ZSTD_compressBlock_lazy_generic (dictMode=%u) (searchFunc=%u)", (U32)dictMode, (U32)searchMethod);
     ip += (dictAndPrefixLength == 0);
     if (dictMode == ZSTD_noDict) {
         U32 const curr = (U32)(ip - base);
         U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, curr, ms->cParams.windowLog);
         U32 const maxRep = curr - windowLow;
-        if (offset_2 > maxRep) savedOffset = offset_2, offset_2 = 0;
+        if (offset_2 > maxRep) offsetSaved2 = offset_2, offset_2 = 0;
-        if (offset_1 > maxRep) savedOffset = offset_1, offset_1 = 0;
+        if (offset_1 > maxRep) offsetSaved1 = offset_1, offset_1 = 0;
     }
     if (isDxS) {
         /* dictMatchState repCode checks don't currently handle repCode == 0
@@ -1524,7 +1545,6 @@ ZSTD_compressBlock_lazy_generic(
     }
 
     if (searchMethod == search_rowHash) {
-        const U32 rowLog = MAX(4, MIN(6, ms->cParams.searchLog));
         ZSTD_row_fillHashCache(ms, base, rowLog,
                             MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */),
                             ms->nextToUpdate, ilimit);
@@ -1539,7 +1559,7 @@ ZSTD_compressBlock_lazy_generic(
 #endif
     while (ip < ilimit) {
         size_t matchLength=0;
-        size_t offcode=STORE_REPCODE_1;
+        size_t offBase = REPCODE1_TO_OFFBASE;
         const BYTE* start=ip+1;
         DEBUGLOG(7, "search baseline (depth 0)");
 
@@ -1564,10 +1584,10 @@ ZSTD_compressBlock_lazy_generic(
         }
 
         /* first search (depth 0) */
-        {   size_t offsetFound = 999999999;
+        {   size_t offbaseFound = 999999999;
-            size_t const ml2 = searchMax(ms, ip, iend, &offsetFound);
+            size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &offbaseFound, mls, rowLog, searchMethod, dictMode);
             if (ml2 > matchLength)
-                matchLength = ml2, start = ip, offcode=offsetFound;
+                matchLength = ml2, start = ip, offBase = offbaseFound;
         }
 
         if (matchLength < 4) {
@@ -1581,12 +1601,12 @@ ZSTD_compressBlock_lazy_generic(
             DEBUGLOG(7, "search depth 1");
             ip ++;
             if ( (dictMode == ZSTD_noDict)
-              && (offcode) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
+              && (offBase) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
                 size_t const mlRep = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4;
                 int const gain2 = (int)(mlRep * 3);
-                int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1);
+                int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offBase) + 1);
                 if ((mlRep >= 4) && (gain2 > gain1))
-                    matchLength = mlRep, offcode = STORE_REPCODE_1, start = ip;
+                    matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
             }
             if (isDxS) {
                 const U32 repIndex = (U32)(ip - base) - offset_1;
@@ -1598,17 +1618,17 @@ ZSTD_compressBlock_lazy_generic(
                     const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
                     size_t const mlRep = ZSTD_count_2segments(ip+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
                     int const gain2 = (int)(mlRep * 3);
-                    int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1);
+                    int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offBase) + 1);
                     if ((mlRep >= 4) && (gain2 > gain1))
-                        matchLength = mlRep, offcode = STORE_REPCODE_1, start = ip;
+                        matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
                 }
             }
-            {   size_t offset2=999999999;
+            {   size_t ofbCandidate=999999999;
-                size_t const ml2 = searchMax(ms, ip, iend, &offset2);
+                size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, dictMode);
-                int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offset2)));   /* raw approx */
+                int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */
-                int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 4);
+                int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 4);
                 if ((ml2 >= 4) && (gain2 > gain1)) {
-                    matchLength = ml2, offcode = offset2, start = ip;
+                    matchLength = ml2, offBase = ofbCandidate, start = ip;
                     continue;   /* search a better one */
             }   }
 
@@ -1617,12 +1637,12 @@ ZSTD_compressBlock_lazy_generic(
                 DEBUGLOG(7, "search depth 2");
                 ip ++;
                 if ( (dictMode == ZSTD_noDict)
-                  && (offcode) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
+                  && (offBase) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
                     size_t const mlRep = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4;
                     int const gain2 = (int)(mlRep * 4);
-                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1);
+                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 1);
                     if ((mlRep >= 4) && (gain2 > gain1))
-                        matchLength = mlRep, offcode = STORE_REPCODE_1, start = ip;
+                        matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
                 }
                 if (isDxS) {
                     const U32 repIndex = (U32)(ip - base) - offset_1;
@@ -1634,17 +1654,17 @@ ZSTD_compressBlock_lazy_generic(
                         const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
                         size_t const mlRep = ZSTD_count_2segments(ip+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
                         int const gain2 = (int)(mlRep * 4);
-                        int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1);
+                        int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 1);
                         if ((mlRep >= 4) && (gain2 > gain1))
-                            matchLength = mlRep, offcode = STORE_REPCODE_1, start = ip;
+                            matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
                     }
                 }
-                {   size_t offset2=999999999;
+                {   size_t ofbCandidate=999999999;
-                    size_t const ml2 = searchMax(ms, ip, iend, &offset2);
+                    size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, dictMode);
-                    int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offset2)));   /* raw approx */
+                    int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */
-                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 7);
+                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 7);
                     if ((ml2 >= 4) && (gain2 > gain1)) {
-                        matchLength = ml2, offcode = offset2, start = ip;
+                        matchLength = ml2, offBase = ofbCandidate, start = ip;
                         continue;
             }   }   }
             break;  /* nothing found : store previous solution */
@@ -1655,24 +1675,24 @@ ZSTD_compressBlock_lazy_generic(
          * notably if `value` is unsigned, resulting in a large positive `-value`.
          */
         /* catch up */
-        if (STORED_IS_OFFSET(offcode)) {
+        if (OFFBASE_IS_OFFSET(offBase)) {
             if (dictMode == ZSTD_noDict) {
-                while ( ((start > anchor) & (start - STORED_OFFSET(offcode) > prefixLowest))
+                while ( ((start > anchor) & (start - OFFBASE_TO_OFFSET(offBase) > prefixLowest))
-                     && (start[-1] == (start-STORED_OFFSET(offcode))[-1]) )  /* only search for offset within prefix */
+                     && (start[-1] == (start-OFFBASE_TO_OFFSET(offBase))[-1]) )  /* only search for offset within prefix */
                     { start--; matchLength++; }
             }
             if (isDxS) {
-                U32 const matchIndex = (U32)((size_t)(start-base) - STORED_OFFSET(offcode));
+                U32 const matchIndex = (U32)((size_t)(start-base) - OFFBASE_TO_OFFSET(offBase));
                 const BYTE* match = (matchIndex < prefixLowestIndex) ? dictBase + matchIndex - dictIndexDelta : base + matchIndex;
                 const BYTE* const mStart = (matchIndex < prefixLowestIndex) ? dictLowest : prefixLowest;
                 while ((start>anchor) && (match>mStart) && (start[-1] == match[-1])) { start--; match--; matchLength++; }  /* catch up */
             }
-            offset_2 = offset_1; offset_1 = (U32)STORED_OFFSET(offcode);
+            offset_2 = offset_1; offset_1 = (U32)OFFBASE_TO_OFFSET(offBase);
         }
         /* store sequence */
 _storeSequence:
         {   size_t const litLength = (size_t)(start - anchor);
-            ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offcode, matchLength);
+            ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offBase, matchLength);
             anchor = ip = start + matchLength;
         }
 
@@ -1688,8 +1708,8 @@ _storeSequence:
                    && (MEM_read32(repMatch) == MEM_read32(ip)) ) {
                     const BYTE* const repEnd2 = repIndex < prefixLowestIndex ? dictEnd : iend;
                     matchLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd2, prefixLowest) + 4;
-                    offcode = offset_2; offset_2 = offset_1; offset_1 = (U32)offcode;   /* swap offset_2 <=> offset_1 */
+                    offBase = offset_2; offset_2 = offset_1; offset_1 = (U32)offBase;   /* swap offset_2 <=> offset_1 */
-                    ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, matchLength);
+                    ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, matchLength);
                     ip += matchLength;
                     anchor = ip;
                     continue;
@@ -1703,16 +1723,20 @@ _storeSequence:
                  && (MEM_read32(ip) == MEM_read32(ip - offset_2)) ) {
                 /* store sequence */
                 matchLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4;
-                offcode = offset_2; offset_2 = offset_1; offset_1 = (U32)offcode; /* swap repcodes */
+                offBase = offset_2; offset_2 = offset_1; offset_1 = (U32)offBase; /* swap repcodes */
-                ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, matchLength);
+                ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, matchLength);
                 ip += matchLength;
                 anchor = ip;
                 continue;   /* faster when present ... (?) */
     }   }   }
 
-    /* Save reps for next block */
+    /* If offset_1 started invalid (offsetSaved1 != 0) and became valid (offset_1 != 0),
-    rep[0] = offset_1 ? offset_1 : savedOffset;
+     * rotate saved offsets. See comment in ZSTD_compressBlock_fast_noDict for more context. */
-    rep[1] = offset_2 ? offset_2 : savedOffset;
+    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);
@@ -1881,9 +1905,9 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
     const BYTE* const dictEnd  = dictBase + dictLimit;
     const BYTE* const dictStart  = dictBase + ms->window.lowLimit;
     const U32 windowLog = ms->cParams.windowLog;
-    const U32 rowLog = ms->cParams.searchLog < 5 ? 4 : 5;
+    const U32 mls = BOUNDED(4, ms->cParams.minMatch, 6);
+    const U32 rowLog = BOUNDED(4, ms->cParams.searchLog, 6);
 
-    searchMax_f const searchMax = ZSTD_selectLazyVTable(ms, searchMethod, ZSTD_extDict)->searchMax;
     U32 offset_1 = rep[0], offset_2 = rep[1];
 
     DEBUGLOG(5, "ZSTD_compressBlock_lazy_extDict_generic (searchFunc=%u)", (U32)searchMethod);
@@ -1905,7 +1929,7 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
 #endif
     while (ip < ilimit) {
         size_t matchLength=0;
-        size_t offcode=STORE_REPCODE_1;
+        size_t offBase = REPCODE1_TO_OFFBASE;
         const BYTE* start=ip+1;
         U32 curr = (U32)(ip-base);
 
@@ -1924,10 +1948,10 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
         }   }
 
         /* first search (depth 0) */
-        {   size_t offsetFound = 999999999;
+        {   size_t ofbCandidate = 999999999;
-            size_t const ml2 = searchMax(ms, ip, iend, &offsetFound);
+            size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, ZSTD_extDict);
             if (ml2 > matchLength)
-                matchLength = ml2, start = ip, offcode=offsetFound;
+                matchLength = ml2, start = ip, offBase = ofbCandidate;
         }
 
         if (matchLength < 4) {
@@ -1941,7 +1965,7 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
             ip ++;
             curr++;
             /* check repCode */
-            if (offcode) {
+            if (offBase) {
                 const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr, windowLog);
                 const U32 repIndex = (U32)(curr - offset_1);
                 const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
@@ -1953,18 +1977,18 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
                     const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
                     size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
                     int const gain2 = (int)(repLength * 3);
-                    int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1);
+                    int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offBase) + 1);
                     if ((repLength >= 4) && (gain2 > gain1))
-                        matchLength = repLength, offcode = STORE_REPCODE_1, start = ip;
+                        matchLength = repLength, offBase = REPCODE1_TO_OFFBASE, start = ip;
             }   }
 
             /* search match, depth 1 */
-            {   size_t offset2=999999999;
+            {   size_t ofbCandidate = 999999999;
-                size_t const ml2 = searchMax(ms, ip, iend, &offset2);
+                size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, ZSTD_extDict);
-                int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offset2)));   /* raw approx */
+                int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */
-                int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 4);
+                int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 4);
                 if ((ml2 >= 4) && (gain2 > gain1)) {
-                    matchLength = ml2, offcode = offset2, start = ip;
+                    matchLength = ml2, offBase = ofbCandidate, start = ip;
                     continue;   /* search a better one */
             }   }
 
@@ -1973,7 +1997,7 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
                 ip ++;
                 curr++;
                 /* check repCode */
-                if (offcode) {
+                if (offBase) {
                     const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr, windowLog);
                     const U32 repIndex = (U32)(curr - offset_1);
                     const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
@@ -1985,36 +2009,36 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
                         const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
                         size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
                         int const gain2 = (int)(repLength * 4);
-                        int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 1);
+                        int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 1);
                         if ((repLength >= 4) && (gain2 > gain1))
-                            matchLength = repLength, offcode = STORE_REPCODE_1, start = ip;
+                            matchLength = repLength, offBase = REPCODE1_TO_OFFBASE, start = ip;
                 }   }
 
                 /* search match, depth 2 */
-                {   size_t offset2=999999999;
+                {   size_t ofbCandidate = 999999999;
-                    size_t const ml2 = searchMax(ms, ip, iend, &offset2);
+                    size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, ZSTD_extDict);
-                    int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offset2)));   /* raw approx */
+                    int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */
-                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)STORED_TO_OFFBASE(offcode)) + 7);
+                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 7);
                     if ((ml2 >= 4) && (gain2 > gain1)) {
-                        matchLength = ml2, offcode = offset2, start = ip;
+                        matchLength = ml2, offBase = ofbCandidate, start = ip;
                         continue;
             }   }   }
             break;  /* nothing found : store previous solution */
         }
 
         /* catch up */
-        if (STORED_IS_OFFSET(offcode)) {
+        if (OFFBASE_IS_OFFSET(offBase)) {
-            U32 const matchIndex = (U32)((size_t)(start-base) - STORED_OFFSET(offcode));
+            U32 const matchIndex = (U32)((size_t)(start-base) - OFFBASE_TO_OFFSET(offBase));
             const BYTE* match = (matchIndex < dictLimit) ? dictBase + matchIndex : base + matchIndex;
             const BYTE* const mStart = (matchIndex < dictLimit) ? dictStart : prefixStart;
             while ((start>anchor) && (match>mStart) && (start[-1] == match[-1])) { start--; match--; matchLength++; }  /* catch up */
-            offset_2 = offset_1; offset_1 = (U32)STORED_OFFSET(offcode);
+            offset_2 = offset_1; offset_1 = (U32)OFFBASE_TO_OFFSET(offBase);
         }
 
         /* store sequence */
 _storeSequence:
         {   size_t const litLength = (size_t)(start - anchor);
-            ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offcode, matchLength);
+            ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offBase, matchLength);
             anchor = ip = start + matchLength;
         }
 
@@ -2031,8 +2055,8 @@ _storeSequence:
                 /* repcode detected we should take it */
                 const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
                 matchLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
-                offcode = offset_2; offset_2 = offset_1; offset_1 = (U32)offcode;   /* swap offset history */
+                offBase = offset_2; offset_2 = offset_1; offset_1 = (U32)offBase;   /* swap offset history */
-                ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, matchLength);
+                ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, matchLength);
                 ip += matchLength;
                 anchor = ip;
                 continue;   /* faster when present ... (?) */
@@ -2098,7 +2122,6 @@ size_t ZSTD_compressBlock_lazy_extDict_row(
 size_t ZSTD_compressBlock_lazy2_extDict_row(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
-
 {
     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2);
 }

zstd/compress/zstd_lazy.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -25,6 +25,8 @@ extern "C" {
  */
 #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);
 

zstd/compress/zstd_ldm.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -242,11 +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:
@@ -549,7 +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
@@ -711,7 +711,7 @@ size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore,
             rep[0] = sequence.offset;
             /* Store the sequence */
             ZSTD_storeSeq(seqStore, newLitLength, ip - newLitLength, iend,
-                          STORE_OFFSET(sequence.offset),
+                          OFFSET_TO_OFFBASE(sequence.offset),
                           sequence.matchLength);
             ip += sequence.matchLength;
         }

zstd/compress/zstd_ldm.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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

zstd/compress/zstd_ldm_geartab.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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

zstd/compress/zstd_opt.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) Przemyslaw Skibinski, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -16,7 +16,7 @@
 #define ZSTD_LITFREQ_ADD    2   /* scaling factor for litFreq, so that frequencies adapt faster to new stats */
 #define ZSTD_MAX_PRICE     (1<<30)
 
-#define ZSTD_PREDEF_THRESHOLD 1024   /* if srcSize < ZSTD_PREDEF_THRESHOLD, symbols' cost is assumed static, directly determined by pre-defined distributions */
+#define ZSTD_PREDEF_THRESHOLD 8   /* if srcSize < ZSTD_PREDEF_THRESHOLD, symbols' cost is assumed static, directly determined by pre-defined distributions */
 
 
 /*-*************************************
@@ -26,27 +26,35 @@
 #if 0    /* approximation at bit level (for tests) */
 #  define BITCOST_ACCURACY 0
 #  define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY)
-#  define WEIGHT(stat, opt) ((void)opt, ZSTD_bitWeight(stat))
+#  define WEIGHT(stat, opt) ((void)(opt), ZSTD_bitWeight(stat))
 #elif 0  /* fractional bit accuracy (for tests) */
 #  define BITCOST_ACCURACY 8
 #  define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY)
-#  define WEIGHT(stat,opt) ((void)opt, ZSTD_fracWeight(stat))
+#  define WEIGHT(stat,opt) ((void)(opt), ZSTD_fracWeight(stat))
 #else    /* opt==approx, ultra==accurate */
 #  define BITCOST_ACCURACY 8
 #  define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY)
-#  define WEIGHT(stat,opt) (opt ? ZSTD_fracWeight(stat) : ZSTD_bitWeight(stat))
+#  define WEIGHT(stat,opt) ((opt) ? ZSTD_fracWeight(stat) : ZSTD_bitWeight(stat))
 #endif
 
+/* ZSTD_bitWeight() :
+ * provide estimated "cost" of a stat in full bits only */
 MEM_STATIC U32 ZSTD_bitWeight(U32 stat)
 {
     return (ZSTD_highbit32(stat+1) * BITCOST_MULTIPLIER);
 }
 
+/* ZSTD_fracWeight() :
+ * provide fractional-bit "cost" of a stat,
+ * using linear interpolation approximation */
 MEM_STATIC U32 ZSTD_fracWeight(U32 rawStat)
 {
     U32 const stat = rawStat + 1;
     U32 const hb = ZSTD_highbit32(stat);
     U32 const BWeight = hb * BITCOST_MULTIPLIER;
+    /* Fweight was meant for "Fractional weight"
+     * but it's effectively a value between 1 and 2
+     * using fixed point arithmetic */
     U32 const FWeight = (stat << BITCOST_ACCURACY) >> hb;
     U32 const weight = BWeight + FWeight;
     assert(hb + BITCOST_ACCURACY < 31);
@@ -57,7 +65,7 @@ MEM_STATIC U32 ZSTD_fracWeight(U32 rawStat)
 /* debugging function,
  * @return price in bytes as fractional value
  * for debug messages only */
-MEM_STATIC double ZSTD_fCost(U32 price)
+MEM_STATIC double ZSTD_fCost(int price)
 {
     return (double)price / (BITCOST_MULTIPLIER*8);
 }
@@ -88,20 +96,26 @@ static U32 sum_u32(const unsigned table[], size_t nbElts)
     return total;
 }
 
-static U32 ZSTD_downscaleStats(unsigned* table, U32 lastEltIndex, U32 shift)
+typedef enum { base_0possible=0, base_1guaranteed=1 } base_directive_e;
+
+static U32
+ZSTD_downscaleStats(unsigned* table, U32 lastEltIndex, U32 shift, base_directive_e base1)
 {
     U32 s, sum=0;
-    DEBUGLOG(5, "ZSTD_downscaleStats (nbElts=%u, shift=%u)", (unsigned)lastEltIndex+1, (unsigned)shift);
+    DEBUGLOG(5, "ZSTD_downscaleStats (nbElts=%u, shift=%u)",
+            (unsigned)lastEltIndex+1, (unsigned)shift );
     assert(shift < 30);
     for (s=0; s<lastEltIndex+1; s++) {
-        table[s] = 1 + (table[s] >> shift);
+        unsigned const base = base1 ? 1 : (table[s]>0);
-        sum += table[s];
+        unsigned const newStat = base + (table[s] >> shift);
+        sum += newStat;
+        table[s] = newStat;
     }
     return sum;
 }
 
 /* ZSTD_scaleStats() :
- * reduce all elements in table is sum too large
+ * reduce all elt frequencies in table if sum too large
  * return the resulting sum of elements */
 static U32 ZSTD_scaleStats(unsigned* table, U32 lastEltIndex, U32 logTarget)
 {
@@ -110,7 +124,7 @@ static U32 ZSTD_scaleStats(unsigned* table, U32 lastEltIndex, U32 logTarget)
     DEBUGLOG(5, "ZSTD_scaleStats (nbElts=%u, target=%u)", (unsigned)lastEltIndex+1, (unsigned)logTarget);
     assert(logTarget < 30);
     if (factor <= 1) return prevsum;
-    return ZSTD_downscaleStats(table, lastEltIndex, ZSTD_highbit32(factor));
+    return ZSTD_downscaleStats(table, lastEltIndex, ZSTD_highbit32(factor), base_1guaranteed);
 }
 
 /* ZSTD_rescaleFreqs() :
@@ -129,18 +143,22 @@ ZSTD_rescaleFreqs(optState_t* const optPtr,
     DEBUGLOG(5, "ZSTD_rescaleFreqs (srcSize=%u)", (unsigned)srcSize);
     optPtr->priceType = zop_dynamic;
 
-    if (optPtr->litLengthSum == 0) {  /* first block : init */
+    if (optPtr->litLengthSum == 0) {  /* no literals stats collected -> first block assumed -> init */
-        if (srcSize <= ZSTD_PREDEF_THRESHOLD) {  /* heuristic */
+
-            DEBUGLOG(5, "(srcSize <= ZSTD_PREDEF_THRESHOLD) => zop_predef");
+        /* heuristic: use pre-defined stats for too small inputs */
+        if (srcSize <= ZSTD_PREDEF_THRESHOLD) {
+            DEBUGLOG(5, "srcSize <= %i : use predefined stats", ZSTD_PREDEF_THRESHOLD);
             optPtr->priceType = zop_predef;
         }
 
         assert(optPtr->symbolCosts != NULL);
         if (optPtr->symbolCosts->huf.repeatMode == HUF_repeat_valid) {
-            /* huffman table presumed generated by dictionary */
+
+            /* huffman stats covering the full value set : table presumed generated by dictionary */
             optPtr->priceType = zop_dynamic;
 
             if (compressedLiterals) {
+                /* generate literals statistics from huffman table */
                 unsigned lit;
                 assert(optPtr->litFreq != NULL);
                 optPtr->litSum = 0;
@@ -188,13 +206,14 @@ ZSTD_rescaleFreqs(optState_t* const optPtr,
                     optPtr->offCodeSum += optPtr->offCodeFreq[of];
             }   }
 
-        } else {  /* not a dictionary */
+        } else {  /* first block, no dictionary */
 
             assert(optPtr->litFreq != NULL);
             if (compressedLiterals) {
+                /* base initial cost of literals on direct frequency within src */
                 unsigned lit = MaxLit;
                 HIST_count_simple(optPtr->litFreq, &lit, src, srcSize);   /* use raw first block to init statistics */
-                optPtr->litSum = ZSTD_downscaleStats(optPtr->litFreq, MaxLit, 8);
+                optPtr->litSum = ZSTD_downscaleStats(optPtr->litFreq, MaxLit, 8, base_0possible);
             }
 
             {   unsigned const baseLLfreqs[MaxLL+1] = {
@@ -224,10 +243,9 @@ ZSTD_rescaleFreqs(optState_t* const optPtr,
                 optPtr->offCodeSum = sum_u32(baseOFCfreqs, MaxOff+1);
             }
 
-
         }
 
-    } else {   /* new block : re-use previous statistics, scaled down */
+    } else {   /* new block : scale down accumulated statistics */
 
         if (compressedLiterals)
             optPtr->litSum = ZSTD_scaleStats(optPtr->litFreq, MaxLit, 12);
@@ -255,11 +273,14 @@ static U32 ZSTD_rawLiteralsCost(const BYTE* const literals, U32 const litLength,
         return (litLength*6) * BITCOST_MULTIPLIER;  /* 6 bit per literal - no statistic used */
 
     /* dynamic statistics */
-    {   U32 price = litLength * optPtr->litSumBasePrice;
+    {   U32 price = optPtr->litSumBasePrice * litLength;
+        U32 const litPriceMax = optPtr->litSumBasePrice - BITCOST_MULTIPLIER;
         U32 u;
+        assert(optPtr->litSumBasePrice >= BITCOST_MULTIPLIER);
         for (u=0; u < litLength; u++) {
-            assert(WEIGHT(optPtr->litFreq[literals[u]], optLevel) <= optPtr->litSumBasePrice);   /* literal cost should never be negative */
+            U32 litPrice = WEIGHT(optPtr->litFreq[literals[u]], optLevel);
-            price -= WEIGHT(optPtr->litFreq[literals[u]], optLevel);
+            if (UNLIKELY(litPrice > litPriceMax)) litPrice = litPriceMax;
+            price -= litPrice;
         }
         return price;
     }
@@ -272,10 +293,11 @@ static U32 ZSTD_litLengthPrice(U32 const litLength, const optState_t* const optP
     assert(litLength <= ZSTD_BLOCKSIZE_MAX);
     if (optPtr->priceType == zop_predef)
         return WEIGHT(litLength, optLevel);
-    /* We can't compute the litLength price for sizes >= ZSTD_BLOCKSIZE_MAX
+
-     * because it isn't representable in the zstd format. So instead just
+    /* ZSTD_LLcode() can't compute litLength price for sizes >= ZSTD_BLOCKSIZE_MAX
-     * call it 1 bit more than ZSTD_BLOCKSIZE_MAX - 1. In this case the block
+     * because it isn't representable in the zstd format.
-     * would be all literals.
+     * So instead just pretend it would cost 1 bit more than ZSTD_BLOCKSIZE_MAX - 1.
+     * In such a case, the block would be all literals.
      */
     if (litLength == ZSTD_BLOCKSIZE_MAX)
         return BITCOST_MULTIPLIER + ZSTD_litLengthPrice(ZSTD_BLOCKSIZE_MAX - 1, optPtr, optLevel);
@@ -289,24 +311,25 @@ static U32 ZSTD_litLengthPrice(U32 const litLength, const optState_t* const optP
 }
 
 /* ZSTD_getMatchPrice() :
- * Provides the cost of the match part (offset + matchLength) of a sequence
+ * Provides the cost of the match part (offset + matchLength) of a sequence.
  * Must be combined with ZSTD_fullLiteralsCost() to get the full cost of a sequence.
- * @offcode : expects a scale where 0,1,2 are repcodes 1-3, and 3+ are real_offsets+2
+ * @offBase : sumtype, representing an offset or a repcode, and using numeric representation of ZSTD_storeSeq()
  * @optLevel: when <2, favors small offset for decompression speed (improved cache efficiency)
  */
 FORCE_INLINE_TEMPLATE U32
-ZSTD_getMatchPrice(U32 const offcode,
+ZSTD_getMatchPrice(U32 const offBase,
                    U32 const matchLength,
              const optState_t* const optPtr,
                    int const optLevel)
 {
     U32 price;
-    U32 const offCode = ZSTD_highbit32(STORED_TO_OFFBASE(offcode));
+    U32 const offCode = ZSTD_highbit32(offBase);
     U32 const mlBase = matchLength - MINMATCH;
     assert(matchLength >= MINMATCH);
 
-    if (optPtr->priceType == zop_predef)  /* fixed scheme, do not use statistics */
+    if (optPtr->priceType == zop_predef)  /* fixed scheme, does not use statistics */
-        return WEIGHT(mlBase, optLevel) + ((16 + offCode) * BITCOST_MULTIPLIER);
+        return WEIGHT(mlBase, optLevel)
+             + ((16 + offCode) * BITCOST_MULTIPLIER); /* emulated offset cost */
 
     /* dynamic statistics */
     price = (offCode * BITCOST_MULTIPLIER) + (optPtr->offCodeSumBasePrice - WEIGHT(optPtr->offCodeFreq[offCode], optLevel));
@@ -325,10 +348,10 @@ ZSTD_getMatchPrice(U32 const offcode,
 }
 
 /* ZSTD_updateStats() :
- * assumption : literals + litLengtn <= iend */
+ * assumption : literals + litLength <= iend */
 static void ZSTD_updateStats(optState_t* const optPtr,
                              U32 litLength, const BYTE* literals,
-                             U32 offsetCode, U32 matchLength)
+                             U32 offBase, U32 matchLength)
 {
     /* literals */
     if (ZSTD_compressedLiterals(optPtr)) {
@@ -344,8 +367,8 @@ static void ZSTD_updateStats(optState_t* const optPtr,
         optPtr->litLengthSum++;
     }
 
-    /* offset code : expected to follow storeSeq() numeric representation */
+    /* offset code : follows storeSeq() numeric representation */
-    {   U32 const offCode = ZSTD_highbit32(STORED_TO_OFFBASE(offsetCode));
+    {   U32 const offCode = ZSTD_highbit32(offBase);
         assert(offCode <= MaxOff);
         optPtr->offCodeFreq[offCode]++;
         optPtr->offCodeSum++;
@@ -552,16 +575,17 @@ void ZSTD_updateTree(ZSTD_matchState_t* ms, const BYTE* ip, const BYTE* iend) {
     ZSTD_updateTree_internal(ms, ip, iend, ms->cParams.minMatch, ZSTD_noDict);
 }
 
-FORCE_INLINE_TEMPLATE
+FORCE_INLINE_TEMPLATE U32
-U32 ZSTD_insertBtAndGetAllMatches (
+ZSTD_insertBtAndGetAllMatches (
                 ZSTD_match_t* matches,  /* store result (found matches) in this table (presumed large enough) */
                 ZSTD_matchState_t* ms,
                 U32* nextToUpdate3,
-                    const BYTE* const ip, const BYTE* const iLimit, const ZSTD_dictMode_e dictMode,
+                const BYTE* const ip, const BYTE* const iLimit,
+                const ZSTD_dictMode_e dictMode,
                 const U32 rep[ZSTD_REP_NUM],
-                    U32 const ll0,   /* tells if associated literal length is 0 or not. This value must be 0 or 1 */
+                const U32 ll0,  /* tells if associated literal length is 0 or not. This value must be 0 or 1 */
                 const U32 lengthToBeat,
-                    U32 const mls /* template */)
+                const U32 mls /* template */)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
     U32 const sufficient_len = MIN(cParams->targetLength, ZSTD_OPT_NUM -1);
@@ -644,7 +668,7 @@ U32 ZSTD_insertBtAndGetAllMatches (
                 DEBUGLOG(8, "found repCode %u (ll0:%u, offset:%u) of length %u",
                             repCode, ll0, repOffset, repLen);
                 bestLength = repLen;
-                matches[mnum].off = STORE_REPCODE(repCode - ll0 + 1);  /* expect value between 1 and 3 */
+                matches[mnum].off = REPCODE_TO_OFFBASE(repCode - ll0 + 1);  /* expect value between 1 and 3 */
                 matches[mnum].len = (U32)repLen;
                 mnum++;
                 if ( (repLen > sufficient_len)
@@ -673,7 +697,7 @@ U32 ZSTD_insertBtAndGetAllMatches (
                 bestLength = mlen;
                 assert(curr > matchIndex3);
                 assert(mnum==0);  /* no prior solution */
-                matches[0].off = STORE_OFFSET(curr - matchIndex3);
+                matches[0].off = OFFSET_TO_OFFBASE(curr - matchIndex3);
                 matches[0].len = (U32)mlen;
                 mnum = 1;
                 if ( (mlen > sufficient_len) |
@@ -706,13 +730,13 @@ U32 ZSTD_insertBtAndGetAllMatches (
         }
 
         if (matchLength > bestLength) {
-            DEBUGLOG(8, "found match of length %u at distance %u (offCode=%u)",
+            DEBUGLOG(8, "found match of length %u at distance %u (offBase=%u)",
-                    (U32)matchLength, curr - matchIndex, STORE_OFFSET(curr - matchIndex));
+                    (U32)matchLength, curr - matchIndex, OFFSET_TO_OFFBASE(curr - matchIndex));
             assert(matchEndIdx > matchIndex);
             if (matchLength > matchEndIdx - matchIndex)
                 matchEndIdx = matchIndex + (U32)matchLength;
             bestLength = matchLength;
-            matches[mnum].off = STORE_OFFSET(curr - matchIndex);
+            matches[mnum].off = OFFSET_TO_OFFBASE(curr - matchIndex);
             matches[mnum].len = (U32)matchLength;
             mnum++;
             if ( (matchLength > ZSTD_OPT_NUM)
@@ -754,12 +778,12 @@ U32 ZSTD_insertBtAndGetAllMatches (
 
             if (matchLength > bestLength) {
                 matchIndex = dictMatchIndex + dmsIndexDelta;
-                DEBUGLOG(8, "found dms match of length %u at distance %u (offCode=%u)",
+                DEBUGLOG(8, "found dms match of length %u at distance %u (offBase=%u)",
-                        (U32)matchLength, curr - matchIndex, STORE_OFFSET(curr - matchIndex));
+                        (U32)matchLength, curr - matchIndex, OFFSET_TO_OFFBASE(curr - matchIndex));
                 if (matchLength > matchEndIdx - matchIndex)
                     matchEndIdx = matchIndex + (U32)matchLength;
                 bestLength = matchLength;
-                matches[mnum].off = STORE_OFFSET(curr - matchIndex);
+                matches[mnum].off = OFFSET_TO_OFFBASE(curr - matchIndex);
                 matches[mnum].len = (U32)matchLength;
                 mnum++;
                 if ( (matchLength > ZSTD_OPT_NUM)
@@ -960,7 +984,7 @@ static void ZSTD_optLdm_maybeAddMatch(ZSTD_match_t* matches, U32* nbMatches,
                                       const ZSTD_optLdm_t* optLdm, U32 currPosInBlock)
 {
     U32 const posDiff = currPosInBlock - optLdm->startPosInBlock;
-    /* Note: ZSTD_match_t actually contains offCode and matchLength (before subtracting MINMATCH) */
+    /* Note: ZSTD_match_t actually contains offBase and matchLength (before subtracting MINMATCH) */
     U32 const candidateMatchLength = optLdm->endPosInBlock - optLdm->startPosInBlock - posDiff;
 
     /* Ensure that current block position is not outside of the match */
@@ -971,11 +995,11 @@ static void ZSTD_optLdm_maybeAddMatch(ZSTD_match_t* matches, U32* nbMatches,
     }
 
     if (*nbMatches == 0 || ((candidateMatchLength > matches[*nbMatches-1].len) && *nbMatches < ZSTD_OPT_NUM)) {
-        U32 const candidateOffCode = STORE_OFFSET(optLdm->offset);
+        U32 const candidateOffBase = OFFSET_TO_OFFBASE(optLdm->offset);
-        DEBUGLOG(6, "ZSTD_optLdm_maybeAddMatch(): Adding ldm candidate match (offCode: %u matchLength %u) at block position=%u",
+        DEBUGLOG(6, "ZSTD_optLdm_maybeAddMatch(): Adding ldm candidate match (offBase: %u matchLength %u) at block position=%u",
-                 candidateOffCode, candidateMatchLength, currPosInBlock);
+                 candidateOffBase, candidateMatchLength, currPosInBlock);
         matches[*nbMatches].len = candidateMatchLength;
-        matches[*nbMatches].off = candidateOffCode;
+        matches[*nbMatches].off = candidateOffBase;
         (*nbMatches)++;
     }
 }
@@ -1098,14 +1122,14 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
 
             /* large match -> immediate encoding */
             {   U32 const maxML = matches[nbMatches-1].len;
-                U32 const maxOffcode = matches[nbMatches-1].off;
+                U32 const maxOffBase = matches[nbMatches-1].off;
-                DEBUGLOG(6, "found %u matches of maxLength=%u and maxOffCode=%u at cPos=%u => start new series",
+                DEBUGLOG(6, "found %u matches of maxLength=%u and maxOffBase=%u at cPos=%u => start new series",
-                            nbMatches, maxML, maxOffcode, (U32)(ip-prefixStart));
+                            nbMatches, maxML, maxOffBase, (U32)(ip-prefixStart));
 
                 if (maxML > sufficient_len) {
                     lastSequence.litlen = litlen;
                     lastSequence.mlen = maxML;
-                    lastSequence.off = maxOffcode;
+                    lastSequence.off = maxOffBase;
                     DEBUGLOG(6, "large match (%u>%u), immediate encoding",
                                 maxML, sufficient_len);
                     cur = 0;
@@ -1122,15 +1146,15 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
                     opt[pos].price = ZSTD_MAX_PRICE;   /* mlen, litlen and price will be fixed during forward scanning */
                 }
                 for (matchNb = 0; matchNb < nbMatches; matchNb++) {
-                    U32 const offcode = matches[matchNb].off;
+                    U32 const offBase = matches[matchNb].off;
                     U32 const end = matches[matchNb].len;
                     for ( ; pos <= end ; pos++ ) {
-                        U32 const matchPrice = ZSTD_getMatchPrice(offcode, pos, optStatePtr, optLevel);
+                        U32 const matchPrice = ZSTD_getMatchPrice(offBase, pos, optStatePtr, optLevel);
                         U32 const sequencePrice = literalsPrice + matchPrice;
                         DEBUGLOG(7, "rPos:%u => set initial price : %.2f",
-                                    pos, ZSTD_fCost(sequencePrice));
+                                    pos, ZSTD_fCost((int)sequencePrice));
                         opt[pos].mlen = pos;
-                        opt[pos].off = offcode;
+                        opt[pos].off = offBase;
                         opt[pos].litlen = litlen;
                         opt[pos].price = (int)sequencePrice;
                 }   }
@@ -1230,7 +1254,7 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
                     U32 const startML = (matchNb>0) ? matches[matchNb-1].len+1 : minMatch;
                     U32 mlen;
 
-                    DEBUGLOG(7, "testing match %u => offCode=%4u, mlen=%2u, llen=%2u",
+                    DEBUGLOG(7, "testing match %u => offBase=%4u, mlen=%2u, llen=%2u",
                                 matchNb, matches[matchNb].off, lastML, litlen);
 
                     for (mlen = lastML; mlen >= startML; mlen--) {  /* scan downward */
@@ -1296,7 +1320,7 @@ _shortestPath:   /* cur, last_pos, best_mlen, best_off have to be set */
                 for (storePos=storeStart; storePos <= storeEnd; storePos++) {
                     U32 const llen = opt[storePos].litlen;
                     U32 const mlen = opt[storePos].mlen;
-                    U32 const offCode = opt[storePos].off;
+                    U32 const offBase = opt[storePos].off;
                     U32 const advance = llen + mlen;
                     DEBUGLOG(6, "considering seq starting at %zi, llen=%u, mlen=%u",
                                 anchor - istart, (unsigned)llen, (unsigned)mlen);
@@ -1308,8 +1332,8 @@ _shortestPath:   /* cur, last_pos, best_mlen, best_off have to be set */
                     }
 
                     assert(anchor + llen <= iend);
-                    ZSTD_updateStats(optStatePtr, llen, anchor, offCode, mlen);
+                    ZSTD_updateStats(optStatePtr, llen, anchor, offBase, mlen);
-                    ZSTD_storeSeq(seqStore, llen, anchor, iend, offCode, mlen);
+                    ZSTD_storeSeq(seqStore, llen, anchor, iend, offBase, mlen);
                     anchor += advance;
                     ip = anchor;
             }   }
@@ -1349,7 +1373,7 @@ size_t ZSTD_compressBlock_btopt(
 /* ZSTD_initStats_ultra():
  * make a first compression pass, just to seed stats with more accurate starting values.
  * only works on first block, with no dictionary and no ldm.
- * this function cannot error, hence its contract must be respected.
+ * this function cannot error out, its narrow contract must be respected.
  */
 static void
 ZSTD_initStats_ultra(ZSTD_matchState_t* ms,
@@ -1368,7 +1392,7 @@ ZSTD_initStats_ultra(ZSTD_matchState_t* ms,
 
     ZSTD_compressBlock_opt2(ms, seqStore, tmpRep, src, srcSize, ZSTD_noDict);   /* generate stats into ms->opt*/
 
-    /* invalidate first scan from history */
+    /* invalidate first scan from history, only keep entropy stats */
     ZSTD_resetSeqStore(seqStore);
     ms->window.base -= srcSize;
     ms->window.dictLimit += (U32)srcSize;
@@ -1392,20 +1416,20 @@ size_t ZSTD_compressBlock_btultra2(
     U32 const curr = (U32)((const BYTE*)src - ms->window.base);
     DEBUGLOG(5, "ZSTD_compressBlock_btultra2 (srcSize=%zu)", srcSize);
 
-    /* 2-pass strategy:
+    /* 2-passes strategy:
      * this strategy makes a first pass over first block to collect statistics
-     * and seed next round's statistics with it.
+     * in order to seed next round's statistics with it.
-     * After 1st pass, function forgets everything, and starts a new block.
+     * After 1st pass, function forgets history, and starts a new block.
      * Consequently, this can only work if no data has been previously loaded in tables,
      * aka, no dictionary, no prefix, no ldm preprocessing.
      * The compression ratio gain is generally small (~0.5% on first block),
-     * the cost is 2x cpu time on first block. */
+    ** the cost is 2x cpu time on first block. */
     assert(srcSize <= ZSTD_BLOCKSIZE_MAX);
     if ( (ms->opt.litLengthSum==0)   /* first block */
       && (seqStore->sequences == seqStore->sequencesStart)  /* no ldm */
       && (ms->window.dictLimit == ms->window.lowLimit)   /* no dictionary */
       && (curr == ms->window.dictLimit)    /* start of frame, nothing already loaded nor skipped */
-      && (srcSize > ZSTD_PREDEF_THRESHOLD)
+      && (srcSize > ZSTD_PREDEF_THRESHOLD) /* input large enough to not employ default stats */
       ) {
         ZSTD_initStats_ultra(ms, seqStore, rep, src, srcSize);
     }

zstd/compress/zstd_opt.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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

zstd/compress/zstdmt_compress.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -266,11 +266,11 @@ static void ZSTDMT_releaseBuffer(ZSTDMT_bufferPool* bufPool, buffer_t buf)
  *   1 buffer for input loading
  *   1 buffer for "next input" when submitting current one
  *   1 buffer stuck in queue */
-#define BUF_POOL_MAX_NB_BUFFERS(nbWorkers) 2*nbWorkers + 3
+#define BUF_POOL_MAX_NB_BUFFERS(nbWorkers) (2*(nbWorkers) + 3)
 
 /* After a worker releases its rawSeqStore, it is immediately ready for reuse.
  * So we only need one seq buffer per worker. */
-#define SEQ_POOL_MAX_NB_BUFFERS(nbWorkers) nbWorkers
+#define SEQ_POOL_MAX_NB_BUFFERS(nbWorkers) (nbWorkers)
 
 /* =====   Seq Pool Wrapper   ====== */
 
@@ -1734,7 +1734,7 @@ findSynchronizationPoint(ZSTDMT_CCtx const* mtctx, ZSTD_inBuffer const input)
         }
     } else {
         /* We have enough bytes buffered to initialize the hash,
-         * and are have processed enough bytes to find a sync point.
+         * and have processed enough bytes to find a sync point.
          * Start scanning at the beginning of the input.
          */
         assert(mtctx->inBuff.filled >= RSYNC_MIN_BLOCK_SIZE);
@@ -1761,17 +1761,24 @@ findSynchronizationPoint(ZSTDMT_CCtx const* mtctx, ZSTD_inBuffer const input)
      * then a block will be emitted anyways, but this is okay, since if we
      * are already synchronized we will remain synchronized.
      */
+    assert(pos < RSYNC_LENGTH || ZSTD_rollingHash_compute(istart + pos - RSYNC_LENGTH, RSYNC_LENGTH) == hash);
     for (; pos < syncPoint.toLoad; ++pos) {
         BYTE const toRemove = pos < RSYNC_LENGTH ? prev[pos] : istart[pos - RSYNC_LENGTH];
-        assert(pos < RSYNC_LENGTH || ZSTD_rollingHash_compute(istart + pos - RSYNC_LENGTH, RSYNC_LENGTH) == hash);
+        /* This assert is very expensive, and Debian compiles with asserts enabled.
+         * So disable it for now. We can get similar coverage by checking it at the
+         * beginning & end of the loop.
+         * assert(pos < RSYNC_LENGTH || ZSTD_rollingHash_compute(istart + pos - RSYNC_LENGTH, RSYNC_LENGTH) == hash);
+         */
         hash = ZSTD_rollingHash_rotate(hash, toRemove, istart[pos], primePower);
         assert(mtctx->inBuff.filled + pos >= RSYNC_MIN_BLOCK_SIZE);
         if ((hash & hitMask) == hitMask) {
             syncPoint.toLoad = pos + 1;
             syncPoint.flush = 1;
+            ++pos; /* for assert */
             break;
         }
     }
+    assert(pos < RSYNC_LENGTH || ZSTD_rollingHash_compute(istart + pos - RSYNC_LENGTH, RSYNC_LENGTH) == hash);
     return syncPoint;
 }
 

zstd/compress/zstdmt_compress.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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

zstd/decompress/huf_decompress_amd64.S

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -30,14 +30,14 @@
  * TODO: Support Windows calling convention.
  */
 
-ZSTD_HIDE_ASM_FUNCTION(HUF_decompress4X1_usingDTable_internal_bmi2_asm_loop)
+ZSTD_HIDE_ASM_FUNCTION(HUF_decompress4X1_usingDTable_internal_fast_asm_loop)
-ZSTD_HIDE_ASM_FUNCTION(HUF_decompress4X2_usingDTable_internal_bmi2_asm_loop)
+ZSTD_HIDE_ASM_FUNCTION(HUF_decompress4X2_usingDTable_internal_fast_asm_loop)
-ZSTD_HIDE_ASM_FUNCTION(_HUF_decompress4X2_usingDTable_internal_bmi2_asm_loop)
+ZSTD_HIDE_ASM_FUNCTION(_HUF_decompress4X2_usingDTable_internal_fast_asm_loop)
-ZSTD_HIDE_ASM_FUNCTION(_HUF_decompress4X1_usingDTable_internal_bmi2_asm_loop)
+ZSTD_HIDE_ASM_FUNCTION(_HUF_decompress4X1_usingDTable_internal_fast_asm_loop)
-.global HUF_decompress4X1_usingDTable_internal_bmi2_asm_loop
+.global HUF_decompress4X1_usingDTable_internal_fast_asm_loop
-.global HUF_decompress4X2_usingDTable_internal_bmi2_asm_loop
+.global HUF_decompress4X2_usingDTable_internal_fast_asm_loop
-.global _HUF_decompress4X1_usingDTable_internal_bmi2_asm_loop
+.global _HUF_decompress4X1_usingDTable_internal_fast_asm_loop
-.global _HUF_decompress4X2_usingDTable_internal_bmi2_asm_loop
+.global _HUF_decompress4X2_usingDTable_internal_fast_asm_loop
 .text
 
 /* Sets up register mappings for clarity.
@@ -95,8 +95,9 @@ ZSTD_HIDE_ASM_FUNCTION(_HUF_decompress4X1_usingDTable_internal_bmi2_asm_loop)
 /* Define both _HUF_* & HUF_* symbols because MacOS
  * C symbols are prefixed with '_' & Linux symbols aren't.
  */
-_HUF_decompress4X1_usingDTable_internal_bmi2_asm_loop:
+_HUF_decompress4X1_usingDTable_internal_fast_asm_loop:
-HUF_decompress4X1_usingDTable_internal_bmi2_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
@@ -350,8 +351,9 @@ HUF_decompress4X1_usingDTable_internal_bmi2_asm_loop:
     pop %rax
     ret
 
-_HUF_decompress4X2_usingDTable_internal_bmi2_asm_loop:
+_HUF_decompress4X2_usingDTable_internal_fast_asm_loop:
-HUF_decompress4X2_usingDTable_internal_bmi2_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
@@ -427,41 +429,30 @@ HUF_decompress4X2_usingDTable_internal_bmi2_asm_loop:
     /* r15 = (ip0 - ilimit) / 7 */
     movq %rdx, %r15
 
-    movabsq $-3689348814741910323, %rdx
+    /* 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 */
-    mulq %rdx
+    movq 16(%rsp), %rdx /* rdx = oend1 */
-    shrq $3,      %rdx /* rdx = rax / 10 */
+    subq %op1,     %rdx /* rdx = oend1 - op1 */
 
-    /* r15 = min(%rdx, %r15) */
+    cmpq  %rax,    %rdx
-    cmpq  %rdx, %r15
+    cmova %rax,    %rdx /* rdx = min(%rdx, %rax) */
-    cmova %rdx, %r15
 
-    movabsq $-3689348814741910323, %rdx
-    movq 16(%rsp), %rax /* rax = oend1 */
-    subq %op1,     %rax /* rax = oend1 - op1 */
-    mulq %rdx
-    shrq $3,       %rdx /* rdx = rax / 10 */
-
-    /* r15 = min(%rdx, %r15) */
-    cmpq  %rdx, %r15
-    cmova %rdx, %r15
-
-    movabsq $-3689348814741910323, %rdx
     movq 24(%rsp), %rax /* rax = oend2 */
     subq %op2,     %rax /* rax = oend2 - op2 */
-    mulq %rdx
-    shrq $3,       %rdx /* rdx = rax / 10 */
 
-    /* r15 = min(%rdx, %r15) */
+    cmpq  %rax,    %rdx
-    cmpq  %rdx, %r15
+    cmova %rax,    %rdx /* rdx = min(%rdx, %rax) */
-    cmova %rdx, %r15
 
-    movabsq $-3689348814741910323, %rdx
     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 = rax / 10 */
+    shrq $3,       %rdx /* rdx = rdx / 10 */
 
     /* r15 = min(%rdx, %r15) */
     cmpq  %rdx, %r15

zstd/decompress/zstd_ddict.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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 +19,6 @@
 #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"
@@ -134,7 +133,7 @@ static size_t ZSTD_initDDict_internal(ZSTD_DDict* ddict,
         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) , "");
@@ -240,5 +239,5 @@ size_t ZSTD_sizeof_DDict(const ZSTD_DDict* ddict)
 unsigned ZSTD_getDictID_fromDDict(const ZSTD_DDict* ddict)
 {
     if (ddict==NULL) return 0;
-    return ZSTD_getDictID_fromDict(ddict->dictContent, ddict->dictSize);
+    return ddict->dictID;
 }

zstd/decompress/zstd_ddict.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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

zstd/decompress/zstd_decompress.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -59,13 +59,13 @@
 #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 "../common/xxhash.h" /* XXH64_reset, XXH64_update, XXH64_digest, XXH64 */
 #include "../common/zstd_internal.h"  /* blockProperties_t */
 #include "zstd_decompress_internal.h"   /* ZSTD_DCtx */
 #include "zstd_ddict.h"  /* ZSTD_DDictDictContent */
 #include "zstd_decompress_block.h"   /* ZSTD_decompressBlock_internal */
+#include "../common/bits.h"  /* ZSTD_highbit32 */
 
 #if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
 #  include "../legacy/zstd_legacy.h"
@@ -243,6 +243,7 @@ static void ZSTD_DCtx_resetParameters(ZSTD_DCtx* dctx)
     dctx->outBufferMode = ZSTD_bm_buffered;
     dctx->forceIgnoreChecksum = ZSTD_d_validateChecksum;
     dctx->refMultipleDDicts = ZSTD_rmd_refSingleDDict;
+    dctx->disableHufAsm = 0;
 }
 
 static void ZSTD_initDCtx_internal(ZSTD_DCtx* dctx)
@@ -438,16 +439,40 @@ size_t ZSTD_frameHeaderSize(const void* src, size_t srcSize)
  *  note : only works for formats ZSTD_f_zstd1 and ZSTD_f_zstd1_magicless
  * @return : 0, `zfhPtr` is correctly filled,
  *          >0, `srcSize` is too small, value is wanted `srcSize` amount,
- *           or an error code, which can be tested using ZSTD_isError() */
+**           or an error code, which can be tested using ZSTD_isError() */
 size_t ZSTD_getFrameHeader_advanced(ZSTD_frameHeader* zfhPtr, const void* src, size_t srcSize, ZSTD_format_e format)
 {
     const BYTE* ip = (const BYTE*)src;
     size_t const minInputSize = ZSTD_startingInputLength(format);
 
-    ZSTD_memset(zfhPtr, 0, sizeof(*zfhPtr));   /* not strictly necessary, but static analyzer do not understand that zfhPtr is only going to be read only if return value is zero, since they are 2 different signals */
+    DEBUGLOG(5, "ZSTD_getFrameHeader_advanced: minInputSize = %zu, srcSize = %zu", minInputSize, srcSize);
-    if (srcSize < minInputSize) return minInputSize;
-    RETURN_ERROR_IF(src==NULL, GENERIC, "invalid parameter");
 
+    if (srcSize > 0) {
+        /* note : technically could be considered an assert(), since it's an invalid entry */
+        RETURN_ERROR_IF(src==NULL, GENERIC, "invalid parameter : src==NULL, but srcSize>0");
+    }
+    if (srcSize < minInputSize) {
+        if (srcSize > 0 && format != ZSTD_f_zstd1_magicless) {
+            /* when receiving less than @minInputSize bytes,
+             * control these bytes at least correspond to a supported magic number
+             * in order to error out early if they don't.
+            **/
+            size_t const toCopy = MIN(4, srcSize);
+            unsigned char hbuf[4]; MEM_writeLE32(hbuf, ZSTD_MAGICNUMBER);
+            assert(src != NULL);
+            ZSTD_memcpy(hbuf, src, toCopy);
+            if ( MEM_readLE32(hbuf) != ZSTD_MAGICNUMBER ) {
+                /* not a zstd frame : let's check if it's a skippable frame */
+                MEM_writeLE32(hbuf, ZSTD_MAGIC_SKIPPABLE_START);
+                ZSTD_memcpy(hbuf, src, toCopy);
+                if ((MEM_readLE32(hbuf) & ZSTD_MAGIC_SKIPPABLE_MASK) != ZSTD_MAGIC_SKIPPABLE_START) {
+                    RETURN_ERROR(prefix_unknown,
+                                "first bytes don't correspond to any supported magic number");
+        }   }   }
+        return minInputSize;
+    }
+
+    ZSTD_memset(zfhPtr, 0, sizeof(*zfhPtr));   /* not strictly necessary, but static analyzers may not understand that zfhPtr will be read only if return value is zero, since they are 2 different signals */
     if ( (format != ZSTD_f_zstd1_magicless)
       && (MEM_readLE32(src) != ZSTD_MAGICNUMBER) ) {
         if ((MEM_readLE32(src) & ZSTD_MAGIC_SKIPPABLE_MASK) == ZSTD_MAGIC_SKIPPABLE_START) {
@@ -757,10 +782,11 @@ static ZSTD_frameSizeInfo ZSTD_findFrameSizeInfo(const void* src, size_t srcSize
             ip += 4;
         }
 
+        frameSizeInfo.nbBlocks = nbBlocks;
         frameSizeInfo.compressedSize = (size_t)(ip - ipstart);
         frameSizeInfo.decompressedBound = (zfh.frameContentSize != ZSTD_CONTENTSIZE_UNKNOWN)
                                         ? zfh.frameContentSize
-                                        : nbBlocks * zfh.blockSizeMax;
+                                        : (unsigned long long)nbBlocks * zfh.blockSizeMax;
         return frameSizeInfo;
     }
 }
@@ -800,6 +826,48 @@ unsigned long long ZSTD_decompressBound(const void* src, size_t srcSize)
     return bound;
 }
 
+size_t ZSTD_decompressionMargin(void const* src, size_t srcSize)
+{
+    size_t margin = 0;
+    unsigned maxBlockSize = 0;
+
+    /* Iterate over each frame */
+    while (srcSize > 0) {
+        ZSTD_frameSizeInfo const frameSizeInfo = ZSTD_findFrameSizeInfo(src, srcSize);
+        size_t const compressedSize = frameSizeInfo.compressedSize;
+        unsigned long long const decompressedBound = frameSizeInfo.decompressedBound;
+        ZSTD_frameHeader zfh;
+
+        FORWARD_IF_ERROR(ZSTD_getFrameHeader(&zfh, src, srcSize), "");
+        if (ZSTD_isError(compressedSize) || decompressedBound == ZSTD_CONTENTSIZE_ERROR)
+            return ERROR(corruption_detected);
+
+        if (zfh.frameType == ZSTD_frame) {
+            /* Add the frame header to our margin */
+            margin += zfh.headerSize;
+            /* Add the checksum to our margin */
+            margin += zfh.checksumFlag ? 4 : 0;
+            /* Add 3 bytes per block */
+            margin += 3 * frameSizeInfo.nbBlocks;
+
+            /* Compute the max block size */
+            maxBlockSize = MAX(maxBlockSize, zfh.blockSizeMax);
+        } else {
+            assert(zfh.frameType == ZSTD_skippableFrame);
+            /* Add the entire skippable frame size to our margin. */
+            margin += compressedSize;
+        }
+
+        assert(srcSize >= compressedSize);
+        src = (const BYTE*)src + compressedSize;
+        srcSize -= compressedSize;
+    }
+
+    /* Add the max block size back to the margin. */
+    margin += maxBlockSize;
+
+    return margin;
+}
 
 /*-*************************************************************
  *   Frame decoding
@@ -825,7 +893,7 @@ static size_t ZSTD_copyRawBlock(void* dst, size_t dstCapacity,
         if (srcSize == 0) return 0;
         RETURN_ERROR(dstBuffer_null, "");
     }
-    ZSTD_memcpy(dst, src, srcSize);
+    ZSTD_memmove(dst, src, srcSize);
     return srcSize;
 }
 
@@ -903,6 +971,7 @@ static size_t ZSTD_decompressFrame(ZSTD_DCtx* dctx,
 
     /* Loop on each block */
     while (1) {
+        BYTE* oBlockEnd = oend;
         size_t decodedSize;
         blockProperties_t blockProperties;
         size_t const cBlockSize = ZSTD_getcBlockSize(ip, remainingSrcSize, &blockProperties);
@@ -912,16 +981,34 @@ static size_t ZSTD_decompressFrame(ZSTD_DCtx* dctx,
         remainingSrcSize -= ZSTD_blockHeaderSize;
         RETURN_ERROR_IF(cBlockSize > remainingSrcSize, srcSize_wrong, "");
 
+        if (ip >= op && ip < oBlockEnd) {
+            /* We are decompressing in-place. Limit the output pointer so that we
+             * don't overwrite the block that we are currently reading. This will
+             * fail decompression if the input & output pointers aren't spaced
+             * far enough apart.
+             *
+             * This is important to set, even when the pointers are far enough
+             * apart, because ZSTD_decompressBlock_internal() can decide to store
+             * literals in the output buffer, after the block it is decompressing.
+             * Since we don't want anything to overwrite our input, we have to tell
+             * ZSTD_decompressBlock_internal to never write past ip.
+             *
+             * See ZSTD_allocateLiteralsBuffer() for reference.
+             */
+            oBlockEnd = op + (ip - op);
+        }
+
         switch(blockProperties.blockType)
         {
         case bt_compressed:
-            decodedSize = ZSTD_decompressBlock_internal(dctx, op, (size_t)(oend-op), ip, cBlockSize, /* frame */ 1, not_streaming);
+            decodedSize = ZSTD_decompressBlock_internal(dctx, op, (size_t)(oBlockEnd-op), ip, cBlockSize, /* frame */ 1, not_streaming);
             break;
         case bt_raw :
+            /* Use oend instead of oBlockEnd because this function is safe to overlap. It uses memmove. */
             decodedSize = ZSTD_copyRawBlock(op, (size_t)(oend-op), ip, cBlockSize);
             break;
         case bt_rle :
-            decodedSize = ZSTD_setRleBlock(op, (size_t)(oend-op), *ip, blockProperties.origSize);
+            decodedSize = ZSTD_setRleBlock(op, (size_t)(oBlockEnd-op), *ip, blockProperties.origSize);
             break;
         case bt_reserved :
         default:
@@ -956,6 +1043,7 @@ static size_t ZSTD_decompressFrame(ZSTD_DCtx* dctx,
     }
     ZSTD_DCtx_trace_end(dctx, (U64)(op-ostart), (U64)(ip-istart), /* streaming */ 0);
     /* Allow caller to get size read */
+    DEBUGLOG(4, "ZSTD_decompressFrame: decompressed frame of size %zi, consuming %zi bytes of input", op-ostart, ip - (const BYTE*)*srcPtr);
     *srcPtr = ip;
     *srcSizePtr = remainingSrcSize;
     return (size_t)(op-ostart);
@@ -1108,8 +1196,8 @@ size_t ZSTD_decompress(void* dst, size_t dstCapacity, const void* src, size_t sr
 size_t ZSTD_nextSrcSizeToDecompress(ZSTD_DCtx* dctx) { return dctx->expected; }
 
 /**
- * Similar to ZSTD_nextSrcSizeToDecompress(), but when when a block input can be streamed,
+ * Similar to ZSTD_nextSrcSizeToDecompress(), but when a block input can be streamed, we
- * we allow taking a partial block as the input. Currently only raw uncompressed blocks can
+ * allow taking a partial block as the input. Currently only raw uncompressed blocks can
  * be streamed.
  *
  * For blocks that can be streamed, this allows us to reduce the latency until we produce
@@ -1309,7 +1397,7 @@ size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, c
 
     default:
         assert(0);   /* impossible */
-        RETURN_ERROR(GENERIC, "impossible to reach");   /* some compiler require default to do something */
+        RETURN_ERROR(GENERIC, "impossible to reach");   /* some compilers require default to do something */
     }
 }
 
@@ -1350,11 +1438,11 @@ ZSTD_loadDEntropy(ZSTD_entropyDTables_t* entropy,
         /* in minimal huffman, we always use X1 variants */
         size_t const hSize = HUF_readDTableX1_wksp(entropy->hufTable,
                                                 dictPtr, dictEnd - dictPtr,
-                                                workspace, workspaceSize);
+                                                workspace, workspaceSize, /* flags */ 0);
 #else
         size_t const hSize = HUF_readDTableX2_wksp(entropy->hufTable,
                                                 dictPtr, (size_t)(dictEnd - dictPtr),
-                                                workspace, workspaceSize);
+                                                workspace, workspaceSize, /* flags */ 0);
 #endif
         RETURN_ERROR_IF(HUF_isError(hSize), dictionary_corrupted, "");
         dictPtr += hSize;
@@ -1453,7 +1541,7 @@ size_t ZSTD_decompressBegin(ZSTD_DCtx* dctx)
     dctx->prefixStart = NULL;
     dctx->virtualStart = NULL;
     dctx->dictEnd = NULL;
-    dctx->entropy.hufTable[0] = (HUF_DTable)((HufLog)*0x1000001);  /* cover both little and big endian */
+    dctx->entropy.hufTable[0] = (HUF_DTable)((ZSTD_HUFFDTABLE_CAPACITY_LOG)*0x1000001);  /* cover both little and big endian */
     dctx->litEntropy = dctx->fseEntropy = 0;
     dctx->dictID = 0;
     dctx->bType = bt_reserved;
@@ -1515,7 +1603,7 @@ unsigned ZSTD_getDictID_fromDict(const void* dict, size_t dictSize)
  *  This could for one of the following reasons :
  *  - The frame does not require a dictionary (most common case).
  *  - The frame was built with dictID intentionally removed.
- *    Needed dictionary is a hidden information.
+ *    Needed dictionary is a hidden piece of information.
  *    Note : this use case also happens when using a non-conformant dictionary.
  *  - `srcSize` is too small, and as a result, frame header could not be decoded.
  *    Note : possible if `srcSize < ZSTD_FRAMEHEADERSIZE_MAX`.
@@ -1524,7 +1612,7 @@ unsigned ZSTD_getDictID_fromDict(const void* dict, size_t dictSize)
  *  ZSTD_getFrameHeader(), which will provide a more precise error code. */
 unsigned ZSTD_getDictID_fromFrame(const void* src, size_t srcSize)
 {
-    ZSTD_frameHeader zfp = { 0, 0, 0, ZSTD_frame, 0, 0, 0 };
+    ZSTD_frameHeader zfp = { 0, 0, 0, ZSTD_frame, 0, 0, 0, 0, 0 };
     size_t const hError = ZSTD_getFrameHeader(&zfp, src, srcSize);
     if (ZSTD_isError(hError)) return 0;
     return zfp.dictID;
@@ -1631,7 +1719,9 @@ size_t ZSTD_initDStream_usingDict(ZSTD_DStream* zds, const void* dict, size_t di
 size_t ZSTD_initDStream(ZSTD_DStream* zds)
 {
     DEBUGLOG(4, "ZSTD_initDStream");
-    return ZSTD_initDStream_usingDDict(zds, NULL);
+    FORWARD_IF_ERROR(ZSTD_DCtx_reset(zds, ZSTD_reset_session_only), "");
+    FORWARD_IF_ERROR(ZSTD_DCtx_refDDict(zds, NULL), "");
+    return ZSTD_startingInputLength(zds->format);
 }
 
 /* ZSTD_initDStream_usingDDict() :
@@ -1639,6 +1729,7 @@ size_t ZSTD_initDStream(ZSTD_DStream* zds)
  * this function cannot fail */
 size_t ZSTD_initDStream_usingDDict(ZSTD_DStream* dctx, const ZSTD_DDict* ddict)
 {
+    DEBUGLOG(4, "ZSTD_initDStream_usingDDict");
     FORWARD_IF_ERROR( ZSTD_DCtx_reset(dctx, ZSTD_reset_session_only) , "");
     FORWARD_IF_ERROR( ZSTD_DCtx_refDDict(dctx, ddict) , "");
     return ZSTD_startingInputLength(dctx->format);
@@ -1649,6 +1740,7 @@ size_t ZSTD_initDStream_usingDDict(ZSTD_DStream* dctx, const ZSTD_DDict* ddict)
  * this function cannot fail */
 size_t ZSTD_resetDStream(ZSTD_DStream* dctx)
 {
+    DEBUGLOG(4, "ZSTD_resetDStream");
     FORWARD_IF_ERROR(ZSTD_DCtx_reset(dctx, ZSTD_reset_session_only), "");
     return ZSTD_startingInputLength(dctx->format);
 }
@@ -1720,6 +1812,11 @@ ZSTD_bounds ZSTD_dParam_getBounds(ZSTD_dParameter dParam)
             bounds.lowerBound = (int)ZSTD_rmd_refSingleDDict;
             bounds.upperBound = (int)ZSTD_rmd_refMultipleDDicts;
             return bounds;
+        case ZSTD_d_disableHuffmanAssembly:
+            bounds.lowerBound = 0;
+            bounds.upperBound = 1;
+            return bounds;
+
         default:;
     }
     bounds.error = ERROR(parameter_unsupported);
@@ -1760,6 +1857,9 @@ size_t ZSTD_DCtx_getParameter(ZSTD_DCtx* dctx, ZSTD_dParameter param, int* value
         case ZSTD_d_refMultipleDDicts:
             *value = (int)dctx->refMultipleDDicts;
             return 0;
+        case ZSTD_d_disableHuffmanAssembly:
+            *value = (int)dctx->disableHufAsm;
+            return 0;
         default:;
     }
     RETURN_ERROR(parameter_unsupported, "");
@@ -1793,6 +1893,10 @@ size_t ZSTD_DCtx_setParameter(ZSTD_DCtx* dctx, ZSTD_dParameter dParam, int value
             }
             dctx->refMultipleDDicts = (ZSTD_refMultipleDDicts_e)value;
             return 0;
+        case ZSTD_d_disableHuffmanAssembly:
+            CHECK_DBOUNDS(ZSTD_d_disableHuffmanAssembly, value);
+            dctx->disableHufAsm = value != 0;
+            return 0;
         default:;
     }
     RETURN_ERROR(parameter_unsupported, "");
@@ -1980,7 +2084,6 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
                 if (zds->refMultipleDDicts && zds->ddictSet) {
                     ZSTD_DCtx_selectFrameDDict(zds);
                 }
-                DEBUGLOG(5, "header size : %u", (U32)hSize);
                 if (ZSTD_isError(hSize)) {
 #if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
                     U32 const legacyVersion = ZSTD_isLegacy(istart, iend-istart);
@@ -2012,6 +2115,11 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
                             zds->lhSize += remainingInput;
                         }
                         input->pos = input->size;
+                        /* check first few bytes */
+                        FORWARD_IF_ERROR(
+                            ZSTD_getFrameHeader_advanced(&zds->fParams, zds->headerBuffer, zds->lhSize, zds->format),
+                            "First few bytes detected incorrect" );
+                        /* return hint input size */
                         return (MAX((size_t)ZSTD_FRAMEHEADERSIZE_MIN(zds->format), hSize) - zds->lhSize) + ZSTD_blockHeaderSize;   /* remaining header bytes + next block header */
                     }
                     assert(ip != NULL);
@@ -2029,8 +2137,9 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
                     size_t const decompressedSize = ZSTD_decompress_usingDDict(zds, op, (size_t)(oend-op), istart, cSize, ZSTD_getDDict(zds));
                     if (ZSTD_isError(decompressedSize)) return decompressedSize;
                     DEBUGLOG(4, "shortcut to single-pass ZSTD_decompress_usingDDict()")
+                    assert(istart != NULL);
                     ip = istart + cSize;
-                    op += decompressedSize;
+                    op = op ? op + decompressedSize : op; /* can occur if frameContentSize = 0 (empty frame) */
                     zds->expected = 0;
                     zds->streamStage = zdss_init;
                     someMoreWork = 0;
@@ -2114,6 +2223,7 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
                 }
                 if ((size_t)(iend-ip) >= neededInSize) {  /* decode directly from src */
                     FORWARD_IF_ERROR(ZSTD_decompressContinueStream(zds, &op, oend, ip, neededInSize), "");
+                    assert(ip != NULL);
                     ip += neededInSize;
                     /* Function modifies the stage so we must break */
                     break;
@@ -2128,7 +2238,7 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
                 int const isSkipFrame = ZSTD_isSkipFrame(zds);
                 size_t loadedSize;
                 /* At this point we shouldn't be decompressing a block that we can stream. */
-                assert(neededInSize == ZSTD_nextSrcSizeToDecompressWithInputSize(zds, iend - ip));
+                assert(neededInSize == ZSTD_nextSrcSizeToDecompressWithInputSize(zds, (size_t)(iend - ip)));
                 if (isSkipFrame) {
                     loadedSize = MIN(toLoad, (size_t)(iend-ip));
                 } else {
@@ -2137,8 +2247,11 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
                                     "should never happen");
                     loadedSize = ZSTD_limitCopy(zds->inBuff + zds->inPos, toLoad, ip, (size_t)(iend-ip));
                 }
+                if (loadedSize != 0) {
+                    /* ip may be NULL */
                     ip += loadedSize;
                     zds->inPos += loadedSize;
+                }
                 if (loadedSize < toLoad) { someMoreWork = 0; break; }   /* not enough input, wait for more */
 
                 /* decode loaded input */
@@ -2148,9 +2261,12 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
                 break;
             }
         case zdss_flush:
-            {   size_t const toFlushSize = zds->outEnd - zds->outStart;
+            {
+                size_t const toFlushSize = zds->outEnd - zds->outStart;
                 size_t const flushedSize = ZSTD_limitCopy(op, (size_t)(oend-op), zds->outBuff + zds->outStart, toFlushSize);
-                op += flushedSize;
+
+                op = op ? op + flushedSize : op;
+
                 zds->outStart += flushedSize;
                 if (flushedSize == toFlushSize) {  /* flush completed */
                     zds->streamStage = zdss_read;
@@ -2169,7 +2285,7 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
 
         default:
             assert(0);    /* impossible */
-            RETURN_ERROR(GENERIC, "impossible to reach");   /* some compiler require default to do something */
+            RETURN_ERROR(GENERIC, "impossible to reach");   /* some compilers require default to do something */
     }   }
 
     /* result */
@@ -2182,8 +2298,8 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
     if ((ip==istart) && (op==ostart)) {  /* no forward progress */
         zds->noForwardProgress ++;
         if (zds->noForwardProgress >= ZSTD_NO_FORWARD_PROGRESS_MAX) {
-            RETURN_ERROR_IF(op==oend, dstSize_tooSmall, "");
+            RETURN_ERROR_IF(op==oend, noForwardProgress_destFull, "");
-            RETURN_ERROR_IF(ip==iend, srcSize_wrong, "");
+            RETURN_ERROR_IF(ip==iend, noForwardProgress_inputEmpty, "");
             assert(0);
         }
     } else {
@@ -2220,11 +2336,17 @@ size_t ZSTD_decompressStream_simpleArgs (
                             void* dst, size_t dstCapacity, size_t* dstPos,
                       const void* src, size_t srcSize, size_t* srcPos)
 {
-    ZSTD_outBuffer output = { dst, dstCapacity, *dstPos };
+    ZSTD_outBuffer output;
-    ZSTD_inBuffer  input  = { src, srcSize, *srcPos };
+    ZSTD_inBuffer  input;
-    /* ZSTD_compress_generic() will check validity of dstPos and srcPos */
+    output.dst = dst;
-    size_t const cErr = ZSTD_decompressStream(dctx, &output, &input);
+    output.size = dstCapacity;
+    output.pos = *dstPos;
+    input.src = src;
+    input.size = srcSize;
+    input.pos = *srcPos;
+    {   size_t const cErr = ZSTD_decompressStream(dctx, &output, &input);
         *dstPos = output.pos;
         *srcPos = input.pos;
         return cErr;
+    }
 }

zstd/decompress/zstd_decompress_block.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -20,12 +20,12 @@
 #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 "../common/zstd_internal.h"
 #include "zstd_decompress_internal.h"   /* ZSTD_DCtx */
 #include "zstd_ddict.h"  /* ZSTD_DDictDictContent */
 #include "zstd_decompress_block.h"
+#include "../common/bits.h"  /* ZSTD_highbit32 */
 
 /*_*******************************************************
 *  Macros
@@ -89,7 +89,7 @@ static void ZSTD_allocateLiteralsBuffer(ZSTD_DCtx* dctx, void* const dst, const
             dctx->litBufferEnd = dctx->litBuffer + litSize - ZSTD_LITBUFFEREXTRASIZE;
         }
         else {
-            /* initially this will be stored entirely in dst during huffman decoding, it will partially shifted to litExtraBuffer after */
+            /* initially this will be stored entirely in dst during huffman decoding, it will partially be shifted to litExtraBuffer after */
             dctx->litBuffer = (BYTE*)dst + expectedWriteSize - litSize;
             dctx->litBufferEnd = (BYTE*)dst + expectedWriteSize;
         }
@@ -134,13 +134,16 @@ size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
             ZSTD_FALLTHROUGH;
 
         case set_compressed:
-            RETURN_ERROR_IF(srcSize < 5, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 3; here we need up to 5 for case 3");
+            RETURN_ERROR_IF(srcSize < 5, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need up to 5 for case 3");
             {   size_t lhSize, litSize, litCSize;
                 U32 singleStream=0;
                 U32 const lhlCode = (istart[0] >> 2) & 3;
                 U32 const lhc = MEM_readLE32(istart);
                 size_t hufSuccess;
                 size_t expectedWriteSize = MIN(ZSTD_BLOCKSIZE_MAX, dstCapacity);
+                int const flags = 0
+                    | (ZSTD_DCtx_get_bmi2(dctx) ? HUF_flags_bmi2 : 0)
+                    | (dctx->disableHufAsm ? HUF_flags_disableAsm : 0);
                 switch(lhlCode)
                 {
                 case 0: case 1: default:   /* note : default is impossible, since lhlCode into [0..3] */
@@ -165,6 +168,10 @@ size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
                 }
                 RETURN_ERROR_IF(litSize > 0 && dst == NULL, dstSize_tooSmall, "NULL not handled");
                 RETURN_ERROR_IF(litSize > ZSTD_BLOCKSIZE_MAX, corruption_detected, "");
+                if (!singleStream)
+                    RETURN_ERROR_IF(litSize < MIN_LITERALS_FOR_4_STREAMS, literals_headerWrong,
+                        "Not enough literals (%zu) for the 4-streams mode (min %u)",
+                        litSize, MIN_LITERALS_FOR_4_STREAMS);
                 RETURN_ERROR_IF(litCSize + lhSize > srcSize, corruption_detected, "");
                 RETURN_ERROR_IF(expectedWriteSize < litSize , dstSize_tooSmall, "");
                 ZSTD_allocateLiteralsBuffer(dctx, dst, dstCapacity, litSize, streaming, expectedWriteSize, 0);
@@ -176,13 +183,14 @@ size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
 
                 if (litEncType==set_repeat) {
                     if (singleStream) {
-                        hufSuccess = HUF_decompress1X_usingDTable_bmi2(
+                        hufSuccess = HUF_decompress1X_usingDTable(
                             dctx->litBuffer, litSize, istart+lhSize, litCSize,
-                            dctx->HUFptr, ZSTD_DCtx_get_bmi2(dctx));
+                            dctx->HUFptr, flags);
                     } else {
-                        hufSuccess = HUF_decompress4X_usingDTable_bmi2(
+                        assert(litSize >= MIN_LITERALS_FOR_4_STREAMS);
+                        hufSuccess = HUF_decompress4X_usingDTable(
                             dctx->litBuffer, litSize, istart+lhSize, litCSize,
-                            dctx->HUFptr, ZSTD_DCtx_get_bmi2(dctx));
+                            dctx->HUFptr, flags);
                     }
                 } else {
                     if (singleStream) {
@@ -190,18 +198,18 @@ size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
                         hufSuccess = HUF_decompress1X_DCtx_wksp(
                             dctx->entropy.hufTable, dctx->litBuffer, litSize,
                             istart+lhSize, litCSize, dctx->workspace,
-                            sizeof(dctx->workspace));
+                            sizeof(dctx->workspace), flags);
 #else
-                        hufSuccess = HUF_decompress1X1_DCtx_wksp_bmi2(
+                        hufSuccess = HUF_decompress1X1_DCtx_wksp(
                             dctx->entropy.hufTable, dctx->litBuffer, litSize,
                             istart+lhSize, litCSize, dctx->workspace,
-                            sizeof(dctx->workspace), ZSTD_DCtx_get_bmi2(dctx));
+                            sizeof(dctx->workspace), flags);
 #endif
                     } else {
-                        hufSuccess = HUF_decompress4X_hufOnly_wksp_bmi2(
+                        hufSuccess = HUF_decompress4X_hufOnly_wksp(
                             dctx->entropy.hufTable, dctx->litBuffer, litSize,
                             istart+lhSize, litCSize, dctx->workspace,
-                            sizeof(dctx->workspace), ZSTD_DCtx_get_bmi2(dctx));
+                            sizeof(dctx->workspace), flags);
                     }
                 }
                 if (dctx->litBufferLocation == ZSTD_split)
@@ -237,6 +245,7 @@ size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
                     break;
                 case 3:
                     lhSize = 3;
+                    RETURN_ERROR_IF(srcSize<3, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need lhSize = 3");
                     litSize = MEM_readLE24(istart) >> 4;
                     break;
                 }
@@ -279,12 +288,13 @@ size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
                     break;
                 case 1:
                     lhSize = 2;
+                    RETURN_ERROR_IF(srcSize<3, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need lhSize+1 = 3");
                     litSize = MEM_readLE16(istart) >> 4;
                     break;
                 case 3:
                     lhSize = 3;
+                    RETURN_ERROR_IF(srcSize<4, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need lhSize+1 = 4");
                     litSize = MEM_readLE24(istart) >> 4;
-                    RETURN_ERROR_IF(srcSize<4, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 3; here we need lhSize+1 = 4");
                     break;
                 }
                 RETURN_ERROR_IF(litSize > 0 && dst == NULL, dstSize_tooSmall, "NULL not handled");
@@ -506,14 +516,15 @@ void ZSTD_buildFSETable_body(ZSTD_seqSymbol* dt,
                 for (i = 8; i < n; i += 8) {
                     MEM_write64(spread + pos + i, sv);
                 }
-                pos += n;
+                assert(n>=0);
+                pos += (size_t)n;
             }
         }
         /* Now we spread those positions across the table.
-         * The benefit of doing it in two stages is that we avoid the the
+         * 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 emperically worked best.
+         * We unroll the loop twice, since that is what empirically worked best.
          */
         {
             size_t position = 0;
@@ -540,7 +551,7 @@ void ZSTD_buildFSETable_body(ZSTD_seqSymbol* dt,
             for (i=0; i<n; i++) {
                 tableDecode[position].baseValue = s;
                 position = (position + step) & tableMask;
-                while (position > highThreshold) position = (position + step) & tableMask;   /* lowprob area */
+                while (UNLIKELY(position > highThreshold)) position = (position + step) & tableMask;   /* lowprob area */
         }   }
         assert(position == 0); /* position must reach all cells once, otherwise normalizedCounter is incorrect */
     }
@@ -551,7 +562,7 @@ void ZSTD_buildFSETable_body(ZSTD_seqSymbol* dt,
         for (u=0; u<tableSize; u++) {
             U32 const symbol = tableDecode[u].baseValue;
             U32 const nextState = symbolNext[symbol]++;
-            tableDecode[u].nbBits = (BYTE) (tableLog - BIT_highbit32(nextState) );
+            tableDecode[u].nbBits = (BYTE) (tableLog - ZSTD_highbit32(nextState) );
             tableDecode[u].nextState = (U16) ( (nextState << tableDecode[u].nbBits) - tableSize);
             assert(nbAdditionalBits[symbol] < 255);
             tableDecode[u].nbAdditionalBits = nbAdditionalBits[symbol];
@@ -964,6 +975,11 @@ size_t ZSTD_execSequence(BYTE* op,
 
     assert(op != NULL /* Precondition */);
     assert(oend_w < oend /* No underflow */);
+
+#if defined(__aarch64__)
+    /* prefetch sequence starting from match that will be used for copy later */
+    PREFETCH_L1(match);
+#endif
     /* Handle edge cases in a slow path:
      *   - Read beyond end of literals
      *   - Match end is within WILDCOPY_OVERLIMIT of oend
@@ -1154,7 +1170,7 @@ ZSTD_updateFseStateWithDInfo(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD, U16
 }
 
 /* We need to add at most (ZSTD_WINDOWLOG_MAX_32 - 1) bits to read the maximum
- * offset bits. But we can only read at most (STREAM_ACCUMULATOR_MIN_32 - 1)
+ * offset bits. But we can only read at most STREAM_ACCUMULATOR_MIN_32
  * bits before reloading. This value is the maximum number of bytes we read
  * after reloading when we are decoding long offsets.
  */
@@ -1169,9 +1185,27 @@ FORCE_INLINE_TEMPLATE seq_t
 ZSTD_decodeSequence(seqState_t* seqState, const ZSTD_longOffset_e longOffsets)
 {
     seq_t seq;
+    /*
+     * ZSTD_seqSymbol is a structure with a total of 64 bits wide. So it can be
+     * loaded in one operation and extracted its fields by simply shifting or
+     * bit-extracting on aarch64.
+     * GCC doesn't recognize this and generates more unnecessary ldr/ldrb/ldrh
+     * operations that cause performance drop. This can be avoided by using this
+     * ZSTD_memcpy hack.
+     */
+#if defined(__aarch64__) && (defined(__GNUC__) && !defined(__clang__))
+    ZSTD_seqSymbol llDInfoS, mlDInfoS, ofDInfoS;
+    ZSTD_seqSymbol* const llDInfo = &llDInfoS;
+    ZSTD_seqSymbol* const mlDInfo = &mlDInfoS;
+    ZSTD_seqSymbol* const ofDInfo = &ofDInfoS;
+    ZSTD_memcpy(llDInfo, seqState->stateLL.table + seqState->stateLL.state, sizeof(ZSTD_seqSymbol));
+    ZSTD_memcpy(mlDInfo, seqState->stateML.table + seqState->stateML.state, sizeof(ZSTD_seqSymbol));
+    ZSTD_memcpy(ofDInfo, seqState->stateOffb.table + seqState->stateOffb.state, sizeof(ZSTD_seqSymbol));
+#else
     const ZSTD_seqSymbol* const llDInfo = seqState->stateLL.table + seqState->stateLL.state;
     const ZSTD_seqSymbol* const mlDInfo = seqState->stateML.table + seqState->stateML.state;
     const ZSTD_seqSymbol* const ofDInfo = seqState->stateOffb.table + seqState->stateOffb.state;
+#endif
     seq.matchLength = mlDInfo->baseValue;
     seq.litLength = llDInfo->baseValue;
     {   U32 const ofBase = ofDInfo->baseValue;
@@ -1186,9 +1220,13 @@ ZSTD_decodeSequence(seqState_t* seqState, const ZSTD_longOffset_e longOffsets)
         U32 const llnbBits = llDInfo->nbBits;
         U32 const mlnbBits = mlDInfo->nbBits;
         U32 const ofnbBits = ofDInfo->nbBits;
+
+        assert(llBits <= MaxLLBits);
+        assert(mlBits <= MaxMLBits);
+        assert(ofBits <= MaxOff);
         /*
          * As gcc has better branch and block analyzers, sometimes it is only
-         * valuable to mark likelyness for clang, it gives around 3-4% of
+         * valuable to mark likeliness for clang, it gives around 3-4% of
          * performance.
          */
 
@@ -1201,13 +1239,16 @@ ZSTD_decodeSequence(seqState_t* seqState, const ZSTD_longOffset_e longOffsets)
     #endif
                 ZSTD_STATIC_ASSERT(ZSTD_lo_isLongOffset == 1);
                 ZSTD_STATIC_ASSERT(LONG_OFFSETS_MAX_EXTRA_BITS_32 == 5);
-                assert(ofBits <= MaxOff);
+                ZSTD_STATIC_ASSERT(STREAM_ACCUMULATOR_MIN_32 > LONG_OFFSETS_MAX_EXTRA_BITS_32);
+                ZSTD_STATIC_ASSERT(STREAM_ACCUMULATOR_MIN_32 - LONG_OFFSETS_MAX_EXTRA_BITS_32 >= MaxMLBits);
                 if (MEM_32bits() && longOffsets && (ofBits >= STREAM_ACCUMULATOR_MIN_32)) {
-                    U32 const extraBits = ofBits - MIN(ofBits, 32 - seqState->DStream.bitsConsumed);
+                    /* Always read extra bits, this keeps the logic simple,
+                     * avoids branches, and avoids accidentally reading 0 bits.
+                     */
+                    U32 const extraBits = LONG_OFFSETS_MAX_EXTRA_BITS_32;
                     offset = ofBase + (BIT_readBitsFast(&seqState->DStream, ofBits - extraBits) << extraBits);
                     BIT_reloadDStream(&seqState->DStream);
-                    if (extraBits) offset += BIT_readBitsFast(&seqState->DStream, extraBits);
+                    offset += BIT_readBitsFast(&seqState->DStream, extraBits);
-                    assert(extraBits <= LONG_OFFSETS_MAX_EXTRA_BITS_32);   /* to avoid another reload */
                 } else {
                     offset = ofBase + BIT_readBitsFast(&seqState->DStream, ofBits/*>0*/);   /* <=  (ZSTD_WINDOWLOG_MAX-1) bits */
                     if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream);
@@ -1552,7 +1593,7 @@ ZSTD_decompressSequences_body(ZSTD_DCtx* dctx,
     const BYTE* const prefixStart = (const BYTE*)(dctx->prefixStart);
     const BYTE* const vBase = (const BYTE*)(dctx->virtualStart);
     const BYTE* const dictEnd = (const BYTE*)(dctx->dictEnd);
-    DEBUGLOG(5, "ZSTD_decompressSequences_body");
+    DEBUGLOG(5, "ZSTD_decompressSequences_body: nbSeq = %d", nbSeq);
     (void)frame;
 
     /* Regen sequences */
@@ -1945,34 +1986,79 @@ ZSTD_decompressSequencesLong(ZSTD_DCtx* dctx,
 #endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT */
 
 
+/**
+ * @returns The total size of the history referencable by zstd, including
+ * both the prefix and the extDict. At @p op any offset larger than this
+ * is invalid.
+ */
+static size_t ZSTD_totalHistorySize(BYTE* op, BYTE const* virtualStart)
+{
+    return (size_t)(op - virtualStart);
+}
 
-#if !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \
+typedef struct {
-    !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG)
+    unsigned longOffsetShare;
-/* ZSTD_getLongOffsetsShare() :
+    unsigned maxNbAdditionalBits;
+} ZSTD_OffsetInfo;
+
+/* ZSTD_getOffsetInfo() :
  * condition : offTable must be valid
  * @return : "share" of long offsets (arbitrarily defined as > (1<<23))
- *           compared to maximum possible of (1<<OffFSELog) */
+ *           compared to maximum possible of (1<<OffFSELog),
-static unsigned
+ *           as well as the maximum number additional bits required.
-ZSTD_getLongOffsetsShare(const ZSTD_seqSymbol* offTable)
+ */
+static ZSTD_OffsetInfo
+ZSTD_getOffsetInfo(const ZSTD_seqSymbol* offTable, int nbSeq)
 {
+    ZSTD_OffsetInfo info = {0, 0};
+    /* If nbSeq == 0, then the offTable is uninitialized, but we have
+     * no sequences, so both values should be 0.
+     */
+    if (nbSeq != 0) {
         const void* ptr = offTable;
         U32 const tableLog = ((const ZSTD_seqSymbol_header*)ptr)[0].tableLog;
         const ZSTD_seqSymbol* table = offTable + 1;
         U32 const max = 1 << tableLog;
-    U32 u, total = 0;
+        U32 u;
         DEBUGLOG(5, "ZSTD_getLongOffsetsShare: (tableLog=%u)", tableLog);
 
         assert(max <= (1 << OffFSELog));  /* max not too large */
         for (u=0; u<max; u++) {
-        if (table[u].nbAdditionalBits > 22) total += 1;
+            info.maxNbAdditionalBits = MAX(info.maxNbAdditionalBits, table[u].nbAdditionalBits);
+            if (table[u].nbAdditionalBits > 22) info.longOffsetShare += 1;
         }
 
         assert(tableLog <= OffFSELog);
-    total <<= (OffFSELog - tableLog);  /* scale to OffFSELog */
+        info.longOffsetShare <<= (OffFSELog - tableLog);  /* scale to OffFSELog */
+    }
 
-    return total;
+    return info;
+}
+
+/**
+ * @returns The maximum offset we can decode in one read of our bitstream, without
+ * reloading more bits in the middle of the offset bits read. Any offsets larger
+ * than this must use the long offset decoder.
+ */
+static size_t ZSTD_maxShortOffset(void)
+{
+    if (MEM_64bits()) {
+        /* We can decode any offset without reloading bits.
+         * This might change if the max window size grows.
+         */
+        ZSTD_STATIC_ASSERT(ZSTD_WINDOWLOG_MAX <= 31);
+        return (size_t)-1;
+    } else {
+        /* The maximum offBase is (1 << (STREAM_ACCUMULATOR_MIN + 1)) - 1.
+         * This offBase would require STREAM_ACCUMULATOR_MIN extra bits.
+         * Then we have to subtract ZSTD_REP_NUM to get the maximum possible offset.
+         */
+        size_t const maxOffbase = ((size_t)1 << (STREAM_ACCUMULATOR_MIN + 1)) - 1;
+        size_t const maxOffset = maxOffbase - ZSTD_REP_NUM;
+        assert(ZSTD_highbit32((U32)maxOffbase) == STREAM_ACCUMULATOR_MIN);
+        return maxOffset;
+    }
 }
-#endif
 
 size_t
 ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
@@ -1980,20 +2066,21 @@ ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
                         const void* src, size_t srcSize, const int frame, const streaming_operation streaming)
 {   /* blockType == blockCompressed */
     const BYTE* ip = (const BYTE*)src;
-    /* isLongOffset must be true if there are long offsets.
-     * Offsets are long if they are larger than 2^STREAM_ACCUMULATOR_MIN.
-     * We don't expect that to be the case in 64-bit mode.
-     * In block mode, window size is not known, so we have to be conservative.
-     * (note: but it could be evaluated from current-lowLimit)
-     */
-    ZSTD_longOffset_e const isLongOffset = (ZSTD_longOffset_e)(MEM_32bits() && (!frame || (dctx->fParams.windowSize > (1ULL << STREAM_ACCUMULATOR_MIN))));
     DEBUGLOG(5, "ZSTD_decompressBlock_internal (size : %u)", (U32)srcSize);
 
-    RETURN_ERROR_IF(srcSize >= ZSTD_BLOCKSIZE_MAX, srcSize_wrong, "");
+    /* Note : the wording of the specification
+     * allows compressed block to be sized exactly ZSTD_BLOCKSIZE_MAX.
+     * This generally does not happen, as it makes little sense,
+     * since an uncompressed block would feature same size and have no decompression cost.
+     * Also, note that decoder from reference libzstd before < v1.5.4
+     * would consider this edge case as an error.
+     * As a consequence, avoid generating compressed blocks of size ZSTD_BLOCKSIZE_MAX
+     * for broader compatibility with the deployed ecosystem of zstd decoders */
+    RETURN_ERROR_IF(srcSize > ZSTD_BLOCKSIZE_MAX, srcSize_wrong, "");
 
     /* Decode literals section */
     {   size_t const litCSize = ZSTD_decodeLiteralsBlock(dctx, src, srcSize, dst, dstCapacity, streaming);
-        DEBUGLOG(5, "ZSTD_decodeLiteralsBlock : %u", (U32)litCSize);
+        DEBUGLOG(5, "ZSTD_decodeLiteralsBlock : cSize=%u, nbLiterals=%zu", (U32)litCSize, dctx->litSize);
         if (ZSTD_isError(litCSize)) return litCSize;
         ip += litCSize;
         srcSize -= litCSize;
@@ -2001,6 +2088,23 @@ ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
 
     /* Build Decoding Tables */
     {
+        /* Compute the maximum block size, which must also work when !frame and fParams are unset.
+         * Additionally, take the min with dstCapacity to ensure that the totalHistorySize fits in a size_t.
+         */
+        size_t const blockSizeMax = MIN(dstCapacity, (frame ? dctx->fParams.blockSizeMax : ZSTD_BLOCKSIZE_MAX));
+        size_t const totalHistorySize = ZSTD_totalHistorySize((BYTE*)dst + blockSizeMax, (BYTE const*)dctx->virtualStart);
+        /* isLongOffset must be true if there are long offsets.
+         * Offsets are long if they are larger than ZSTD_maxShortOffset().
+         * We don't expect that to be the case in 64-bit mode.
+         *
+         * We check here to see if our history is large enough to allow long offsets.
+         * If it isn't, then we can't possible have (valid) long offsets. If the offset
+         * is invalid, then it is okay to read it incorrectly.
+         *
+         * If isLongOffsets is true, then we will later check our decoding table to see
+         * if it is even possible to generate long offsets.
+         */
+        ZSTD_longOffset_e isLongOffset = (ZSTD_longOffset_e)(MEM_32bits() && (totalHistorySize > ZSTD_maxShortOffset()));
         /* These macros control at build-time which decompressor implementation
          * we use. If neither is defined, we do some inspection and dispatch at
          * runtime.
@@ -2008,6 +2112,11 @@ ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
 #if !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \
     !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG)
         int usePrefetchDecoder = dctx->ddictIsCold;
+#else
+        /* Set to 1 to avoid computing offset info if we don't need to.
+         * Otherwise this value is ignored.
+         */
+        int usePrefetchDecoder = 1;
 #endif
         int nbSeq;
         size_t const seqHSize = ZSTD_decodeSeqHeaders(dctx, &nbSeq, ip, srcSize);
@@ -2017,26 +2126,38 @@ ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
 
         RETURN_ERROR_IF(dst == NULL && nbSeq > 0, dstSize_tooSmall, "NULL not handled");
 
-#if !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \
+        /* If we could potentially have long offsets, or we might want to use the prefetch decoder,
-    !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG)
+         * compute information about the share of long offsets, and the maximum nbAdditionalBits.
-        if ( !usePrefetchDecoder
+         * NOTE: could probably use a larger nbSeq limit
-          && (!frame || (dctx->fParams.windowSize > (1<<24)))
+         */
-          && (nbSeq>ADVANCED_SEQS) ) {  /* could probably use a larger nbSeq limit */
+        if (isLongOffset || (!usePrefetchDecoder && (totalHistorySize > (1u << 24)) && (nbSeq > 8))) {
-            U32 const shareLongOffsets = ZSTD_getLongOffsetsShare(dctx->OFTptr);
+            ZSTD_OffsetInfo const info = ZSTD_getOffsetInfo(dctx->OFTptr, nbSeq);
-            U32 const minShare = MEM_64bits() ? 7 : 20; /* heuristic values, correspond to 2.73% and 7.81% */
+            if (isLongOffset && info.maxNbAdditionalBits <= STREAM_ACCUMULATOR_MIN) {
-            usePrefetchDecoder = (shareLongOffsets >= minShare);
+                /* If isLongOffset, but the maximum number of additional bits that we see in our table is small
+                 * enough, then we know it is impossible to have too long an offset in this block, so we can
+                 * use the regular offset decoder.
+                 */
+                isLongOffset = ZSTD_lo_isRegularOffset;
+            }
+            if (!usePrefetchDecoder) {
+                U32 const minShare = MEM_64bits() ? 7 : 20; /* heuristic values, correspond to 2.73% and 7.81% */
+                usePrefetchDecoder = (info.longOffsetShare >= minShare);
+            }
         }
-#endif
 
         dctx->ddictIsCold = 0;
 
 #if !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \
     !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG)
-        if (usePrefetchDecoder)
+        if (usePrefetchDecoder) {
+#else
+        (void)usePrefetchDecoder;
+        {
 #endif
 #ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT
             return ZSTD_decompressSequencesLong(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset, frame);
 #endif
+        }
 
 #ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG
         /* else */

zstd/decompress/zstd_decompress_block.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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

zstd/decompress/zstd_decompress_internal.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -75,12 +75,13 @@ static UNUSED_ATTR const U32 ML_base[MaxML+1] = {
 
 #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;
@@ -164,6 +165,7 @@ struct ZSTD_DCtx_s
     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;

zstd/dictBuilder/cover.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -34,6 +34,7 @@
 #include "../common/pool.h"
 #include "../common/threading.h"
 #include "../common/zstd_internal.h" /* includes zstd.h */
+#include "../common/bits.h" /* ZSTD_highbit32 */
 #include "../zdict.h"
 #include "cover.h"
 
@@ -541,7 +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()`.
@@ -646,7 +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;
   }
@@ -950,9 +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 setDictSelection(NULL, 0, error);
-    return selection;
 }
 
 unsigned COVER_dictSelectionIsError(COVER_dictSelection_t selection) {
@@ -1005,9 +1014,8 @@ COVER_dictSelection_t COVER_selectDict(BYTE* customDictContent, size_t dictBuffe
   }
 
   if (params.shrinkDict == 0) {
-    COVER_dictSelection_t selection = { largestDictbuffer, dictContentSize, totalCompressedSize };
     free(candidateDictBuffer);
-    return selection;
+    return setDictSelection(largestDictbuffer, dictContentSize, totalCompressedSize);
   }
 
   largestDict = dictContentSize;
@@ -1039,20 +1047,16 @@ COVER_dictSelection_t COVER_selectDict(BYTE* customDictContent, size_t dictBuffe
       return COVER_dictSelectionError(totalCompressedSize);
     }
 
-    if (totalCompressedSize <= largestCompressed * regressionTolerance) {
+    if ((double)totalCompressedSize <= (double)largestCompressed * regressionTolerance) {
-      COVER_dictSelection_t selection = { candidateDictBuffer, dictContentSize, totalCompressedSize };
       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;
+  return setDictSelection( largestDictbuffer, dictContentSize, totalCompressedSize );
-  }
 }
 
 /**

zstd/dictBuilder/cover.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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

zstd/dictBuilder/fastcover.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -304,7 +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()`.

zstd/dictBuilder/zdict.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -44,7 +44,6 @@
 #ifndef ZDICT_STATIC_LINKING_ONLY
 #  define ZDICT_STATIC_LINKING_ONLY
 #endif
-#define HUF_STATIC_LINKING_ONLY
 
 #include "../common/mem.h"           /* read */
 #include "../common/fse.h"           /* FSE_normalizeCount, FSE_writeNCount */
@@ -54,6 +53,7 @@
 #include "../compress/zstd_compress_internal.h" /* ZSTD_loadCEntropy() */
 #include "../zdict.h"
 #include "divsufsort.h"
+#include "../common/bits.h"          /* ZSTD_NbCommonBytes */
 
 
 /*-*************************************
@@ -130,85 +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)
-            if (val != 0) {
-                unsigned long r;
-                _BitScanForward64(&r, (U64)val);
-                return (unsigned)(r >> 3);
-            } else {
-                /* Should not reach this code path */
-                __assume(0);
-            }
-#       elif defined(__GNUC__) && (__GNUC__ >= 3)
-            return (unsigned)(__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)
-            if (val != 0) {
-                unsigned long r;
-                _BitScanForward(&r, (U32)val);
-                return (unsigned)(r >> 3);
-            } else {
-                /* Should not reach this code path */
-                __assume(0);
-            }
-#       elif defined(__GNUC__) && (__GNUC__ >= 3)
-            return (unsigned)(__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)
-            if (val != 0) {
-                unsigned long r;
-                _BitScanReverse64(&r, val);
-                return (unsigned)(r >> 3);
-            } else {
-                /* Should not reach this code path */
-                __assume(0);
-            }
-#       elif defined(__GNUC__) && (__GNUC__ >= 3)
-            return (unsigned)(__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)
-            if (val != 0) {
-                unsigned long r;
-                _BitScanReverse(&r, (unsigned long)val);
-                return (unsigned)(r >> 3);
-            } else {
-                /* Should not reach this code path */
-                __assume(0);
-            }
-#       elif defined(__GNUC__) && (__GNUC__ >= 3)
-            return (unsigned)(__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.
@@ -223,7 +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);
     }
 }
@@ -602,7 +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:
@@ -754,6 +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");
@@ -794,7 +716,7 @@ static size_t ZDICT_analyzeEntropy(void*  dstBuffer, size_t maxDstSize,
     }   }
 
     /* 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");
@@ -803,7 +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;
@@ -844,7 +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");

zstd/zdict.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -8,32 +8,43 @@
  * 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
+#ifndef ZDICTLIB_VISIBLE
-#  if defined(__GNUC__) && (__GNUC__ >= 4)
+   /* Backwards compatibility with old macro name */
-#    define ZDICTLIB_VISIBILITY __attribute__ ((visibility ("default")))
+#  ifdef ZDICTLIB_VISIBILITY
+#    define ZDICTLIB_VISIBLE ZDICTLIB_VISIBILITY
+#  elif defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
+#    define ZDICTLIB_VISIBLE __attribute__ ((visibility ("default")))
 #  else
-#    define ZDICTLIB_VISIBILITY
+#    define ZDICTLIB_VISIBLE
 #  endif
 #endif
+
+#ifndef ZDICTLIB_HIDDEN
+#  if defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
+#    define ZDICTLIB_HIDDEN __attribute__ ((visibility ("hidden")))
+#  else
+#    define ZDICTLIB_HIDDEN
+#  endif
+#endif
+
 #if defined(ZSTD_DLL_EXPORT) && (ZSTD_DLL_EXPORT==1)
-#  define ZDICTLIB_API __declspec(dllexport) ZDICTLIB_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
 
 /*******************************************************************************
@@ -110,8 +121,8 @@ extern "C" {
  * 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 a the smallest size that
+ * shrink the dictionary for you, and select the smallest size that doesn't
- * doesn't hurt compression ratio too much. See the `shrinkDict` parameter.
+ * hurt compression ratio too much. See the `shrinkDict` parameter.
  * A smaller dictionary can save memory, and potentially speed up
  * compression.
  *
@@ -201,9 +212,9 @@ 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 */
+    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 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)
+    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.
@@ -260,9 +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.
@@ -318,7 +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);
@@ -340,7 +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);
@@ -361,7 +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);
@@ -384,7 +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);
@@ -409,7 +432,7 @@ 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(
+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);
@@ -421,32 +444,31 @@ ZDICTLIB_API size_t ZDICT_trainFromBuffer_legacy(
    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
+#    define ZDICT_DEPRECATED(message) [[deprecated(message)]]
 #  elif defined(__clang__) || (ZDICT_GCC_VERSION >= 405)
-#    define ZDICT_DEPRECATED(message) ZDICTLIB_API __attribute__((deprecated(message)))
+#    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 */

zstd/zstd.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
@@ -14,21 +14,31 @@ extern "C" {
 #ifndef ZSTD_H_235446
 #define ZSTD_H_235446
 
-/* ======   Dependency   ======*/
+/* ======   Dependencies   ======*/
 #include <limits.h>   /* INT_MAX */
 #include <stddef.h>   /* size_t */
 
 
 /* =====   ZSTDLIB_API : control library symbols visibility   ===== */
 #ifndef ZSTDLIB_VISIBLE
-#  if defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
+   /* Backwards compatibility with old macro name */
+#  ifdef ZSTDLIB_VISIBILITY
+#    define ZSTDLIB_VISIBLE ZSTDLIB_VISIBILITY
+#  elif defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
 #    define ZSTDLIB_VISIBLE __attribute__ ((visibility ("default")))
-#    define ZSTDLIB_HIDDEN __attribute__ ((visibility ("hidden")))
 #  else
 #    define ZSTDLIB_VISIBLE
+#  endif
+#endif
+
+#ifndef ZSTDLIB_HIDDEN
+#  if defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
+#    define ZSTDLIB_HIDDEN __attribute__ ((visibility ("hidden")))
+#  else
 #    define ZSTDLIB_HIDDEN
 #  endif
 #endif
+
 #if defined(ZSTD_DLL_EXPORT) && (ZSTD_DLL_EXPORT==1)
 #  define ZSTDLIB_API __declspec(dllexport) ZSTDLIB_VISIBLE
 #elif defined(ZSTD_DLL_IMPORT) && (ZSTD_DLL_IMPORT==1)
@@ -37,6 +47,28 @@ extern "C" {
 #  define ZSTDLIB_API ZSTDLIB_VISIBLE
 #endif
 
+/* Deprecation warnings :
+ * Should these warnings be a problem, it is generally possible to disable them,
+ * typically with -Wno-deprecated-declarations for gcc or _CRT_SECURE_NO_WARNINGS in Visual.
+ * Otherwise, it's also possible to define ZSTD_DISABLE_DEPRECATE_WARNINGS.
+ */
+#ifdef ZSTD_DISABLE_DEPRECATE_WARNINGS
+#  define ZSTD_DEPRECATED(message) /* disable deprecation warnings */
+#else
+#  if defined (__cplusplus) && (__cplusplus >= 201402) /* C++14 or greater */
+#    define ZSTD_DEPRECATED(message) [[deprecated(message)]]
+#  elif (defined(GNUC) && (GNUC > 4 || (GNUC == 4 && GNUC_MINOR >= 5))) || defined(__clang__)
+#    define ZSTD_DEPRECATED(message) __attribute__((deprecated(message)))
+#  elif defined(__GNUC__) && (__GNUC__ >= 3)
+#    define ZSTD_DEPRECATED(message) __attribute__((deprecated))
+#  elif defined(_MSC_VER)
+#    define ZSTD_DEPRECATED(message) __declspec(deprecated(message))
+#  else
+#    pragma message("WARNING: You need to implement ZSTD_DEPRECATED for this compiler")
+#    define ZSTD_DEPRECATED(message)
+#  endif
+#endif /* ZSTD_DISABLE_DEPRECATE_WARNINGS */
+
 
 /*******************************************************************************
   Introduction
@@ -74,7 +106,7 @@ extern "C" {
 /*------   Version   ------*/
 #define ZSTD_VERSION_MAJOR    1
 #define ZSTD_VERSION_MINOR    5
-#define ZSTD_VERSION_RELEASE  2
+#define ZSTD_VERSION_RELEASE  4
 #define ZSTD_VERSION_NUMBER  (ZSTD_VERSION_MAJOR *100*100 + ZSTD_VERSION_MINOR *100 + ZSTD_VERSION_RELEASE)
 
 /*! ZSTD_versionNumber() :
@@ -165,7 +197,9 @@ ZSTDLIB_API unsigned long long ZSTD_getFrameContentSize(const void *src, size_t
  *  "empty", "unknown" and "error" results to the same return value (0),
  *  while ZSTD_getFrameContentSize() gives them separate return values.
  * @return : decompressed size of `src` frame content _if known and not empty_, 0 otherwise. */
-ZSTDLIB_API unsigned long long ZSTD_getDecompressedSize(const void* src, size_t srcSize);
+ZSTD_DEPRECATED("Replaced by ZSTD_getFrameContentSize")
+ZSTDLIB_API
+unsigned long long ZSTD_getDecompressedSize(const void* src, size_t srcSize);
 
 /*! ZSTD_findFrameCompressedSize() : Requires v1.4.0+
  * `src` should point to the start of a ZSTD frame or skippable frame.
@@ -177,8 +211,30 @@ ZSTDLIB_API size_t ZSTD_findFrameCompressedSize(const void* src, size_t srcSize)
 
 
 /*======  Helper functions  ======*/
-#define ZSTD_COMPRESSBOUND(srcSize)   ((srcSize) + ((srcSize)>>8) + (((srcSize) < (128<<10)) ? (((128<<10) - (srcSize)) >> 11) /* margin, from 64 to 0 */ : 0))  /* this formula ensures that bound(A) + bound(B) <= bound(A+B) as long as A and B >= 128 KB */
+/* ZSTD_compressBound() :
+ * maximum compressed size in worst case single-pass scenario.
+ * When invoking `ZSTD_compress()` or any other one-pass compression function,
+ * it's recommended to provide @dstCapacity >= ZSTD_compressBound(srcSize)
+ * as it eliminates one potential failure scenario,
+ * aka not enough room in dst buffer to write the compressed frame.
+ * Note : ZSTD_compressBound() itself can fail, if @srcSize > ZSTD_MAX_INPUT_SIZE .
+ *        In which case, ZSTD_compressBound() will return an error code
+ *        which can be tested using ZSTD_isError().
+ *
+ * ZSTD_COMPRESSBOUND() :
+ * same as ZSTD_compressBound(), but as a macro.
+ * It can be used to produce constants, which can be useful for static allocation,
+ * for example to size a static array on stack.
+ * Will produce constant value 0 if srcSize too large.
+ */
+#define ZSTD_MAX_INPUT_SIZE ((sizeof(size_t)==8) ? 0xFF00FF00FF00FF00LLU : 0xFF00FF00U)
+#define ZSTD_COMPRESSBOUND(srcSize)   (((size_t)(srcSize) >= ZSTD_MAX_INPUT_SIZE) ? 0 : (srcSize) + ((srcSize)>>8) + (((srcSize) < (128<<10)) ? (((128<<10) - (srcSize)) >> 11) /* margin, from 64 to 0 */ : 0))  /* this formula ensures that bound(A) + bound(B) <= bound(A+B) as long as A and B >= 128 KB */
 ZSTDLIB_API size_t ZSTD_compressBound(size_t srcSize); /*!< maximum compressed size in worst case single-pass scenario */
+/* ZSTD_isError() :
+ * Most ZSTD_* functions returning a size_t value can be tested for error,
+ * using ZSTD_isError().
+ * @return 1 if error, 0 otherwise
+ */
 ZSTDLIB_API unsigned    ZSTD_isError(size_t code);          /*!< tells if a `size_t` function result is an error code */
 ZSTDLIB_API const char* ZSTD_getErrorName(size_t code);     /*!< provides readable string from an error code */
 ZSTDLIB_API int         ZSTD_minCLevel(void);               /*!< minimum negative compression level allowed, requires v1.4.0+ */
@@ -421,6 +477,9 @@ typedef enum {
      * ZSTD_c_validateSequences
      * ZSTD_c_useBlockSplitter
      * ZSTD_c_useRowMatchFinder
+     * ZSTD_c_prefetchCDictTables
+     * ZSTD_c_enableSeqProducerFallback
+     * ZSTD_c_maxBlockSize
      * Because they are not stable, it's necessary to define ZSTD_STATIC_LINKING_ONLY to access them.
      * note : never ever use experimentalParam? names directly;
      *        also, the enums values themselves are unstable and can still change.
@@ -439,7 +498,11 @@ typedef enum {
      ZSTD_c_experimentalParam12=1009,
      ZSTD_c_experimentalParam13=1010,
      ZSTD_c_experimentalParam14=1011,
-     ZSTD_c_experimentalParam15=1012
+     ZSTD_c_experimentalParam15=1012,
+     ZSTD_c_experimentalParam16=1013,
+     ZSTD_c_experimentalParam17=1014,
+     ZSTD_c_experimentalParam18=1015,
+     ZSTD_c_experimentalParam19=1016
 } ZSTD_cParameter;
 
 typedef struct {
@@ -502,7 +565,7 @@ typedef enum {
  *                  They will be used to compress next frame.
  *                  Resetting session never fails.
  *  - The parameters : changes all parameters back to "default".
- *                  This removes any reference to any dictionary too.
+ *                  This also removes any reference to any dictionary or external sequence producer.
  *                  Parameters can only be changed between 2 sessions (i.e. no compression is currently ongoing)
  *                  otherwise the reset fails, and function returns an error value (which can be tested using ZSTD_isError())
  *  - Both : similar to resetting the session, followed by resetting parameters.
@@ -552,13 +615,15 @@ typedef enum {
      * ZSTD_d_stableOutBuffer
      * ZSTD_d_forceIgnoreChecksum
      * ZSTD_d_refMultipleDDicts
+     * ZSTD_d_disableHuffmanAssembly
      * Because they are not stable, it's necessary to define ZSTD_STATIC_LINKING_ONLY to access them.
      * note : never ever use experimentalParam? names directly
      */
      ZSTD_d_experimentalParam1=1000,
      ZSTD_d_experimentalParam2=1001,
      ZSTD_d_experimentalParam3=1002,
-     ZSTD_d_experimentalParam4=1003
+     ZSTD_d_experimentalParam4=1003,
+     ZSTD_d_experimentalParam5=1004
 
 } ZSTD_dParameter;
 
@@ -737,8 +802,6 @@ ZSTDLIB_API size_t ZSTD_CStreamOutSize(void);   /**< recommended size for output
  * This following is a legacy streaming API, available since v1.0+ .
  * It can be replaced by ZSTD_CCtx_reset() and ZSTD_compressStream2().
  * It is redundant, but remains fully supported.
- * Streaming in combination with advanced parameters and dictionary compression
- * can only be used through the new API.
  ******************************************************************************/
 
 /*!
@@ -747,6 +810,9 @@ ZSTDLIB_API size_t ZSTD_CStreamOutSize(void);   /**< recommended size for output
  *     ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only);
  *     ZSTD_CCtx_refCDict(zcs, NULL); // clear the dictionary (if any)
  *     ZSTD_CCtx_setParameter(zcs, ZSTD_c_compressionLevel, compressionLevel);
+ *
+ * Note that ZSTD_initCStream() clears any previously set dictionary. Use the new API
+ * to compress with a dictionary.
  */
 ZSTDLIB_API size_t ZSTD_initCStream(ZSTD_CStream* zcs, int compressionLevel);
 /*!
@@ -797,13 +863,31 @@ ZSTDLIB_API size_t ZSTD_freeDStream(ZSTD_DStream* zds);  /* accept NULL pointer
 
 /*===== Streaming decompression functions =====*/
 
-/* This function is redundant with the advanced API and equivalent to:
+/*! ZSTD_initDStream() :
+ * Initialize/reset DStream state for new decompression operation.
+ * Call before new decompression operation using same DStream.
  *
+ * Note : This function is redundant with the advanced API and equivalent to:
  *     ZSTD_DCtx_reset(zds, ZSTD_reset_session_only);
  *     ZSTD_DCtx_refDDict(zds, NULL);
  */
 ZSTDLIB_API size_t ZSTD_initDStream(ZSTD_DStream* zds);
 
+/*! ZSTD_decompressStream() :
+ * Streaming decompression function.
+ * Call repetitively to consume full input updating it as necessary.
+ * Function will update both input and output `pos` fields exposing current state via these fields:
+ * - `input.pos < input.size`, some input remaining and caller should provide remaining input
+ *   on the next call.
+ * - `output.pos < output.size`, decoder finished and flushed all remaining buffers.
+ * - `output.pos == output.size`, potentially uncflushed data present in the internal buffers,
+ *   call ZSTD_decompressStream() again to flush remaining data to output.
+ * Note : with no additional input, amount of data flushed <= ZSTD_BLOCKSIZE_MAX.
+ *
+ * @return : 0 when a frame is completely decoded and fully flushed,
+ *           or an error code, which can be tested using ZSTD_isError(),
+ *           or any other value > 0, which means there is some decoding or flushing to do to complete current frame.
+ */
 ZSTDLIB_API size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inBuffer* input);
 
 ZSTDLIB_API size_t ZSTD_DStreamInSize(void);    /*!< recommended size for input buffer */
@@ -922,7 +1006,7 @@ ZSTDLIB_API unsigned ZSTD_getDictID_fromDDict(const ZSTD_DDict* ddict);
  *  If @return == 0, the dictID could not be decoded.
  *  This could for one of the following reasons :
  *  - The frame does not require a dictionary to be decoded (most common case).
- *  - The frame was built with dictID intentionally removed. Whatever dictionary is necessary is a hidden information.
+ *  - The frame was built with dictID intentionally removed. Whatever dictionary is necessary is a hidden piece of information.
  *    Note : this use case also happens when using a non-conformant dictionary.
  *  - `srcSize` is too small, and as a result, the frame header could not be decoded (only possible if `srcSize < ZSTD_FRAMEHEADERSIZE_MAX`).
  *  - This is not a Zstandard frame.
@@ -946,8 +1030,9 @@ ZSTDLIB_API unsigned ZSTD_getDictID_fromFrame(const void* src, size_t srcSize);
  * @result : 0, or an error code (which can be tested with ZSTD_isError()).
  *  Special: Loading a NULL (or 0-size) dictionary invalidates previous dictionary,
  *           meaning "return to no-dictionary mode".
- *  Note 1 : Dictionary is sticky, it will be used for all future compressed frames.
+ *  Note 1 : Dictionary is sticky, it will be used for all future compressed frames,
- *           To return to "no-dictionary" situation, load a NULL dictionary (or reset parameters).
+ *           until parameters are reset, a new dictionary is loaded, or the dictionary
+ *           is explicitly invalidated by loading a NULL dictionary.
  *  Note 2 : Loading a dictionary involves building tables.
  *           It's also a CPU consuming operation, with non-negligible impact on latency.
  *           Tables are dependent on compression parameters, and for this reason,
@@ -960,7 +1045,7 @@ ZSTDLIB_API unsigned ZSTD_getDictID_fromFrame(const void* src, size_t srcSize);
 ZSTDLIB_API size_t ZSTD_CCtx_loadDictionary(ZSTD_CCtx* cctx, const void* dict, size_t dictSize);
 
 /*! ZSTD_CCtx_refCDict() : Requires v1.4.0+
- *  Reference a prepared dictionary, to be used for all next compressed frames.
+ *  Reference a prepared dictionary, to be used for all future compressed frames.
  *  Note that compression parameters are enforced from within CDict,
  *  and supersede any compression parameter previously set within CCtx.
  *  The parameters ignored are labelled as "superseded-by-cdict" in the ZSTD_cParameter enum docs.
@@ -995,9 +1080,9 @@ ZSTDLIB_API size_t ZSTD_CCtx_refPrefix(ZSTD_CCtx* cctx,
                                  const void* prefix, size_t prefixSize);
 
 /*! ZSTD_DCtx_loadDictionary() : Requires v1.4.0+
- *  Create an internal DDict from dict buffer,
+ *  Create an internal DDict from dict buffer, to be used to decompress all future frames.
- *  to be used to decompress next frames.
+ *  The dictionary remains valid for all future frames, until explicitly invalidated, or
- *  The dictionary remains valid for all future frames, until explicitly invalidated.
+ *  a new dictionary is loaded.
  * @result : 0, or an error code (which can be tested with ZSTD_isError()).
  *  Special : Adding a NULL (or 0-size) dictionary invalidates any previous dictionary,
  *            meaning "return to no-dictionary mode".
@@ -1021,9 +1106,10 @@ ZSTDLIB_API size_t ZSTD_DCtx_loadDictionary(ZSTD_DCtx* dctx, const void* dict, s
  *  The memory for the table is allocated on the first call to refDDict, and can be
  *  freed with ZSTD_freeDCtx().
  *
+ *  If called with ZSTD_d_refMultipleDDicts disabled (the default), only one dictionary
+ *  will be managed, and referencing a dictionary effectively "discards" any previous one.
+ *
  * @result : 0, or an error code (which can be tested with ZSTD_isError()).
- *  Note 1 : Currently, only one dictionary can be managed.
- *           Referencing a new dictionary effectively "discards" any previous one.
  *  Special: referencing a NULL DDict means "return to no-dictionary mode".
  *  Note 2 : DDict is just referenced, its lifetime must outlive its usage from DCtx.
  */
@@ -1086,28 +1172,6 @@ ZSTDLIB_API size_t ZSTD_sizeof_DDict(const ZSTD_DDict* ddict);
 #  endif
 #endif
 
-/* Deprecation warnings :
- * Should these warnings be a problem, it is generally possible to disable them,
- * typically with -Wno-deprecated-declarations for gcc or _CRT_SECURE_NO_WARNINGS in Visual.
- * Otherwise, it's also possible to define ZSTD_DISABLE_DEPRECATE_WARNINGS.
- */
-#ifdef ZSTD_DISABLE_DEPRECATE_WARNINGS
-#  define ZSTD_DEPRECATED(message) ZSTDLIB_STATIC_API  /* disable deprecation warnings */
-#else
-#  if defined (__cplusplus) && (__cplusplus >= 201402) /* C++14 or greater */
-#    define ZSTD_DEPRECATED(message) [[deprecated(message)]] ZSTDLIB_STATIC_API
-#  elif (defined(GNUC) && (GNUC > 4 || (GNUC == 4 && GNUC_MINOR >= 5))) || defined(__clang__)
-#    define ZSTD_DEPRECATED(message) ZSTDLIB_STATIC_API __attribute__((deprecated(message)))
-#  elif defined(__GNUC__) && (__GNUC__ >= 3)
-#    define ZSTD_DEPRECATED(message) ZSTDLIB_STATIC_API __attribute__((deprecated))
-#  elif defined(_MSC_VER)
-#    define ZSTD_DEPRECATED(message) ZSTDLIB_STATIC_API __declspec(deprecated(message))
-#  else
-#    pragma message("WARNING: You need to implement ZSTD_DEPRECATED for this compiler")
-#    define ZSTD_DEPRECATED(message) ZSTDLIB_STATIC_API
-#  endif
-#endif /* ZSTD_DISABLE_DEPRECATE_WARNINGS */
-
 /****************************************************************************************
  *   experimental API (static linking only)
  ****************************************************************************************
@@ -1142,6 +1206,7 @@ ZSTDLIB_API size_t ZSTD_sizeof_DDict(const ZSTD_DDict* ddict);
 #define ZSTD_TARGETLENGTH_MIN     0   /* note : comparing this constant to an unsigned results in a tautological test */
 #define ZSTD_STRATEGY_MIN        ZSTD_fast
 #define ZSTD_STRATEGY_MAX        ZSTD_btultra2
+#define ZSTD_BLOCKSIZE_MAX_MIN (1 << 10) /* The minimum valid max blocksize. Maximum blocksizes smaller than this make compressBound() inaccurate. */
 
 
 #define ZSTD_OVERLAPLOG_MIN       0
@@ -1369,33 +1434,89 @@ ZSTDLIB_STATIC_API unsigned long long ZSTD_decompressBound(const void* src, size
  *           or an error code (if srcSize is too small) */
 ZSTDLIB_STATIC_API size_t ZSTD_frameHeaderSize(const void* src, size_t srcSize);
 
+/*! ZSTD_decompressionMargin() :
+ * Zstd supports in-place decompression, where the input and output buffers overlap.
+ * In this case, the output buffer must be at least (Margin + Output_Size) bytes large,
+ * and the input buffer must be at the end of the output buffer.
+ *
+ *  _______________________ Output Buffer ________________________
+ * |                                                              |
+ * |                                        ____ Input Buffer ____|
+ * |                                       |                      |
+ * v                                       v                      v
+ * |---------------------------------------|-----------|----------|
+ * ^                                                   ^          ^
+ * |___________________ Output_Size ___________________|_ Margin _|
+ *
+ * NOTE: See also ZSTD_DECOMPRESSION_MARGIN().
+ * NOTE: This applies only to single-pass decompression through ZSTD_decompress() or
+ * ZSTD_decompressDCtx().
+ * NOTE: This function supports multi-frame input.
+ *
+ * @param src The compressed frame(s)
+ * @param srcSize The size of the compressed frame(s)
+ * @returns The decompression margin or an error that can be checked with ZSTD_isError().
+ */
+ZSTDLIB_STATIC_API size_t ZSTD_decompressionMargin(const void* src, size_t srcSize);
+
+/*! ZSTD_DECOMPRESS_MARGIN() :
+ * Similar to ZSTD_decompressionMargin(), but instead of computing the margin from
+ * the compressed frame, compute it from the original size and the blockSizeLog.
+ * See ZSTD_decompressionMargin() for details.
+ *
+ * WARNING: This macro does not support multi-frame input, the input must be a single
+ * zstd frame. If you need that support use the function, or implement it yourself.
+ *
+ * @param originalSize The original uncompressed size of the data.
+ * @param blockSize    The block size == MIN(windowSize, ZSTD_BLOCKSIZE_MAX).
+ *                     Unless you explicitly set the windowLog smaller than
+ *                     ZSTD_BLOCKSIZELOG_MAX you can just use ZSTD_BLOCKSIZE_MAX.
+ */
+#define ZSTD_DECOMPRESSION_MARGIN(originalSize, blockSize) ((size_t)(                                              \
+        ZSTD_FRAMEHEADERSIZE_MAX                                                              /* Frame header */ + \
+        4                                                                                         /* checksum */ + \
+        ((originalSize) == 0 ? 0 : 3 * (((originalSize) + (blockSize) - 1) / blockSize)) /* 3 bytes per block */ + \
+        (blockSize)                                                                    /* One block of margin */   \
+    ))
+
 typedef enum {
   ZSTD_sf_noBlockDelimiters = 0,         /* Representation of ZSTD_Sequence has no block delimiters, sequences only */
   ZSTD_sf_explicitBlockDelimiters = 1    /* Representation of ZSTD_Sequence contains explicit block delimiters */
 } ZSTD_sequenceFormat_e;
 
+/*! ZSTD_sequenceBound() :
+ * `srcSize` : size of the input buffer
+ *  @return : upper-bound for the number of sequences that can be generated
+ *            from a buffer of srcSize bytes
+ *
+ *  note : returns number of sequences - to get bytes, multiply by sizeof(ZSTD_Sequence).
+ */
+ZSTDLIB_STATIC_API size_t ZSTD_sequenceBound(size_t srcSize);
+
 /*! ZSTD_generateSequences() :
- * Generate sequences using ZSTD_compress2, given a source buffer.
+ * Generate sequences using ZSTD_compress2(), given a source buffer.
  *
  * Each block will end with a dummy sequence
  * with offset == 0, matchLength == 0, and litLength == length of last literals.
  * litLength may be == 0, and if so, then the sequence of (of: 0 ml: 0 ll: 0)
  * simply acts as a block delimiter.
  *
- * zc can be used to insert custom compression params.
+ * @zc can be used to insert custom compression params.
- * This function invokes ZSTD_compress2
+ * This function invokes ZSTD_compress2().
  *
  * The output of this function can be fed into ZSTD_compressSequences() with CCtx
  * setting of ZSTD_c_blockDelimiters as ZSTD_sf_explicitBlockDelimiters
  * @return : number of sequences generated
  */
 
-ZSTDLIB_STATIC_API size_t ZSTD_generateSequences(ZSTD_CCtx* zc, ZSTD_Sequence* outSeqs,
+ZSTDLIB_STATIC_API size_t
-                                          size_t outSeqsSize, const void* src, size_t srcSize);
+ZSTD_generateSequences( ZSTD_CCtx* zc,
+                        ZSTD_Sequence* outSeqs, size_t outSeqsSize,
+                        const void* src, size_t srcSize);
 
 /*! ZSTD_mergeBlockDelimiters() :
  * Given an array of ZSTD_Sequence, remove all sequences that represent block delimiters/last literals
- * by merging them into into the literals of the next sequence.
+ * by merging them into the literals of the next sequence.
  *
  * As such, the final generated result has no explicit representation of block boundaries,
  * and the final last literals segment is not represented in the sequences.
@@ -1407,7 +1528,9 @@ ZSTDLIB_STATIC_API size_t ZSTD_generateSequences(ZSTD_CCtx* zc, ZSTD_Sequence* o
 ZSTDLIB_STATIC_API size_t ZSTD_mergeBlockDelimiters(ZSTD_Sequence* sequences, size_t seqsSize);
 
 /*! ZSTD_compressSequences() :
- * Compress an array of ZSTD_Sequence, generated from the original source buffer, into dst.
+ * Compress an array of ZSTD_Sequence, associated with @src buffer, into dst.
+ * @src contains the entire input (not just the literals).
+ * If @srcSize > sum(sequence.length), the remaining bytes are considered all literals
  * If a dictionary is included, then the cctx should reference the dict. (see: ZSTD_CCtx_refCDict(), ZSTD_CCtx_loadDictionary(), etc.)
  * The entire source is compressed into a single frame.
  *
@@ -1432,9 +1555,10 @@ ZSTDLIB_STATIC_API size_t ZSTD_mergeBlockDelimiters(ZSTD_Sequence* sequences, si
  * Note: Repcodes are, as of now, always re-calculated within this function, so ZSTD_Sequence::rep is unused.
  * Note 2: Once we integrate ability to ingest repcodes, the explicit block delims mode must respect those repcodes exactly,
  *         and cannot emit an RLE block that disagrees with the repcode history
- * @return : final compressed size or a ZSTD error.
+ * @return : final compressed size, or a ZSTD error code.
  */
-ZSTDLIB_STATIC_API size_t ZSTD_compressSequences(ZSTD_CCtx* const cctx, void* dst, size_t dstSize,
+ZSTDLIB_STATIC_API size_t
+ZSTD_compressSequences( ZSTD_CCtx* cctx, void* dst, size_t dstSize,
                         const ZSTD_Sequence* inSeqs, size_t inSeqsSize,
                         const void* src, size_t srcSize);
 
@@ -1442,7 +1566,7 @@ ZSTDLIB_STATIC_API size_t ZSTD_compressSequences(ZSTD_CCtx* const cctx, void* ds
 /*! ZSTD_writeSkippableFrame() :
  * Generates a zstd skippable frame containing data given by src, and writes it to dst buffer.
  *
- * Skippable frames begin with a a 4-byte magic number. There are 16 possible choices of magic number,
+ * Skippable frames begin with a 4-byte magic number. There are 16 possible choices of magic number,
  * ranging from ZSTD_MAGIC_SKIPPABLE_START to ZSTD_MAGIC_SKIPPABLE_START+15.
  * As such, the parameter magicVariant controls the exact skippable frame magic number variant used, so
  * the magic number used will be ZSTD_MAGIC_SKIPPABLE_START + magicVariant.
@@ -1500,8 +1624,11 @@ ZSTDLIB_API unsigned ZSTD_isSkippableFrame(const void* buffer, size_t size);
  *  and ZSTD_estimateCCtxSize_usingCCtxParams(), which can be used in tandem with ZSTD_CCtxParams_setParameter().
  *  Both can be used to estimate memory using custom compression parameters and arbitrary srcSize limits.
  *
- *  Note 2 : only single-threaded compression is supported.
+ *  Note : only single-threaded compression is supported.
  *  ZSTD_estimateCCtxSize_usingCCtxParams() will return an error code if ZSTD_c_nbWorkers is >= 1.
+ *
+ *  Note 2 : ZSTD_estimateCCtxSize* functions are not compatible with the Block-Level Sequence Producer API at this time.
+ *  Size estimates assume that no external sequence producer is registered.
  */
 ZSTDLIB_STATIC_API size_t ZSTD_estimateCCtxSize(int compressionLevel);
 ZSTDLIB_STATIC_API size_t ZSTD_estimateCCtxSize_usingCParams(ZSTD_compressionParameters cParams);
@@ -1520,7 +1647,12 @@ ZSTDLIB_STATIC_API size_t ZSTD_estimateDCtxSize(void);
  *  or deducted from a valid frame Header, using ZSTD_estimateDStreamSize_fromFrame();
  *  Note : if streaming is init with function ZSTD_init?Stream_usingDict(),
  *         an internal ?Dict will be created, which additional size is not estimated here.
- *         In this case, get total size by adding ZSTD_estimate?DictSize */
+ *         In this case, get total size by adding ZSTD_estimate?DictSize
+ *  Note 2 : only single-threaded compression is supported.
+ *  ZSTD_estimateCStreamSize_usingCCtxParams() will return an error code if ZSTD_c_nbWorkers is >= 1.
+ *  Note 3 : ZSTD_estimateCStreamSize* functions are not compatible with the Block-Level Sequence Producer API at this time.
+ *  Size estimates assume that no external sequence producer is registered.
+ */
 ZSTDLIB_STATIC_API size_t ZSTD_estimateCStreamSize(int compressionLevel);
 ZSTDLIB_STATIC_API size_t ZSTD_estimateCStreamSize_usingCParams(ZSTD_compressionParameters cParams);
 ZSTDLIB_STATIC_API size_t ZSTD_estimateCStreamSize_usingCCtxParams(const ZSTD_CCtx_params* params);
@@ -1670,11 +1802,19 @@ ZSTDLIB_STATIC_API size_t ZSTD_checkCParams(ZSTD_compressionParameters params);
  *  This function never fails (wide contract) */
 ZSTDLIB_STATIC_API ZSTD_compressionParameters ZSTD_adjustCParams(ZSTD_compressionParameters cPar, unsigned long long srcSize, size_t dictSize);
 
+/*! ZSTD_CCtx_setCParams() :
+ *  Set all parameters provided within @cparams into the working @cctx.
+ *  Note : if modifying parameters during compression (MT mode only),
+ *         note that changes to the .windowLog parameter will be ignored.
+ * @return 0 on success, or an error code (can be checked with ZSTD_isError()) */
+ZSTDLIB_STATIC_API size_t ZSTD_CCtx_setCParams(ZSTD_CCtx* cctx, ZSTD_compressionParameters cparams);
+
 /*! ZSTD_compress_advanced() :
  *  Note : this function is now DEPRECATED.
  *         It can be replaced by ZSTD_compress2(), in combination with ZSTD_CCtx_setParameter() and other parameter setters.
  *  This prototype will generate compilation warnings. */
 ZSTD_DEPRECATED("use ZSTD_compress2")
+ZSTDLIB_STATIC_API
 size_t ZSTD_compress_advanced(ZSTD_CCtx* cctx,
                               void* dst, size_t dstCapacity,
                         const void* src, size_t srcSize,
@@ -1686,6 +1826,7 @@ size_t ZSTD_compress_advanced(ZSTD_CCtx* cctx,
  *         It can be replaced by ZSTD_compress2(), in combination with ZSTD_CCtx_loadDictionary() and other parameter setters.
  *  This prototype will generate compilation warnings. */
 ZSTD_DEPRECATED("use ZSTD_compress2 with ZSTD_CCtx_loadDictionary")
+ZSTDLIB_STATIC_API
 size_t ZSTD_compress_usingCDict_advanced(ZSTD_CCtx* cctx,
                                               void* dst, size_t dstCapacity,
                                         const void* src, size_t srcSize,
@@ -1829,13 +1970,16 @@ ZSTDLIB_STATIC_API size_t ZSTD_CCtx_refPrefix_advanced(ZSTD_CCtx* cctx, const vo
  * Experimental parameter.
  * Default is 0 == disabled. Set to 1 to enable.
  *
- * Tells the compressor that the ZSTD_inBuffer will ALWAYS be the same
+ * Tells the compressor that input data presented with ZSTD_inBuffer
- * between calls, except for the modifications that zstd makes to pos (the
+ * will ALWAYS be the same between calls.
- * caller must not modify pos). This is checked by the compressor, and
+ * Technically, the @src pointer must never be changed,
- * compression will fail if it ever changes. This means the only flush
+ * and the @pos field can only be updated by zstd.
- * mode that makes sense is ZSTD_e_end, so zstd will error if ZSTD_e_end
+ * However, it's possible to increase the @size field,
- * is not used. The data in the ZSTD_inBuffer in the range [src, src + pos)
+ * allowing scenarios where more data can be appended after compressions starts.
- * MUST not be modified during compression or you will get data corruption.
+ * These conditions are checked by the compressor,
+ * and compression will fail if they are not respected.
+ * Also, data in the ZSTD_inBuffer within the range [src, src + pos)
+ * MUST not be modified during compression or it will result in data corruption.
  *
  * When this flag is enabled zstd won't allocate an input window buffer,
  * because the user guarantees it can reference the ZSTD_inBuffer until
@@ -1843,18 +1987,15 @@ ZSTDLIB_STATIC_API size_t ZSTD_CCtx_refPrefix_advanced(ZSTD_CCtx* cctx, const vo
  * large enough to fit a block (see ZSTD_c_stableOutBuffer). This will also
  * avoid the memcpy() from the input buffer to the input window buffer.
  *
- * NOTE: ZSTD_compressStream2() will error if ZSTD_e_end is not used.
- * That means this flag cannot be used with ZSTD_compressStream().
- *
  * NOTE: So long as the ZSTD_inBuffer always points to valid memory, using
  * this flag is ALWAYS memory safe, and will never access out-of-bounds
- * memory. However, compression WILL fail if you violate the preconditions.
+ * memory. However, compression WILL fail if conditions are not respected.
  *
- * WARNING: The data in the ZSTD_inBuffer in the range [dst, dst + pos) MUST
+ * WARNING: The data in the ZSTD_inBuffer in the range [src, src + pos) MUST
- * not be modified during compression or you will get data corruption. This
+ * not be modified during compression or it will result in data corruption.
- * is because zstd needs to reference data in the ZSTD_inBuffer to find
+ * This is because zstd needs to reference data in the ZSTD_inBuffer to find
  * matches. Normally zstd maintains its own window buffer for this purpose,
- * but passing this flag tells zstd to use the user provided buffer.
+ * but passing this flag tells zstd to rely on user provided buffer instead.
  */
 #define ZSTD_c_stableInBuffer ZSTD_c_experimentalParam9
 
@@ -1899,7 +2040,7 @@ ZSTDLIB_STATIC_API size_t ZSTD_CCtx_refPrefix_advanced(ZSTD_CCtx* cctx, const vo
  * Without validation, providing a sequence that does not conform to the zstd spec will cause
  * undefined behavior, and may produce a corrupted block.
  *
- * With validation enabled, a if sequence is invalid (see doc/zstd_compression_format.md for
+ * With validation enabled, if sequence is invalid (see doc/zstd_compression_format.md for
  * specifics regarding offset/matchlength requirements) then the function will bail out and
  * return an error.
  *
@@ -1949,6 +2090,79 @@ ZSTDLIB_STATIC_API size_t ZSTD_CCtx_refPrefix_advanced(ZSTD_CCtx* cctx, const vo
  */
 #define ZSTD_c_deterministicRefPrefix ZSTD_c_experimentalParam15
 
+/* ZSTD_c_prefetchCDictTables
+ * Controlled with ZSTD_paramSwitch_e enum. Default is ZSTD_ps_auto.
+ *
+ * In some situations, zstd uses CDict tables in-place rather than copying them
+ * into the working context. (See docs on ZSTD_dictAttachPref_e above for details).
+ * In such situations, compression speed is seriously impacted when CDict tables are
+ * "cold" (outside CPU cache). This parameter instructs zstd to prefetch CDict tables
+ * when they are used in-place.
+ *
+ * For sufficiently small inputs, the cost of the prefetch will outweigh the benefit.
+ * For sufficiently large inputs, zstd will by default memcpy() CDict tables
+ * into the working context, so there is no need to prefetch. This parameter is
+ * targeted at a middle range of input sizes, where a prefetch is cheap enough to be
+ * useful but memcpy() is too expensive. The exact range of input sizes where this
+ * makes sense is best determined by careful experimentation.
+ *
+ * Note: for this parameter, ZSTD_ps_auto is currently equivalent to ZSTD_ps_disable,
+ * but in the future zstd may conditionally enable this feature via an auto-detection
+ * heuristic for cold CDicts.
+ * Use ZSTD_ps_disable to opt out of prefetching under any circumstances.
+ */
+#define ZSTD_c_prefetchCDictTables ZSTD_c_experimentalParam16
+
+/* ZSTD_c_enableSeqProducerFallback
+ * Allowed values are 0 (disable) and 1 (enable). The default setting is 0.
+ *
+ * Controls whether zstd will fall back to an internal sequence producer if an
+ * external sequence producer is registered and returns an error code. This fallback
+ * is block-by-block: the internal sequence producer will only be called for blocks
+ * where the external sequence producer returns an error code. Fallback parsing will
+ * follow any other cParam settings, such as compression level, the same as in a
+ * normal (fully-internal) compression operation.
+ *
+ * The user is strongly encouraged to read the full Block-Level Sequence Producer API
+ * documentation (below) before setting this parameter. */
+#define ZSTD_c_enableSeqProducerFallback ZSTD_c_experimentalParam17
+
+/* ZSTD_c_maxBlockSize
+ * Allowed values are between 1KB and ZSTD_BLOCKSIZE_MAX (128KB).
+ * The default is ZSTD_BLOCKSIZE_MAX, and setting to 0 will set to the default.
+ *
+ * This parameter can be used to set an upper bound on the blocksize
+ * that overrides the default ZSTD_BLOCKSIZE_MAX. It cannot be used to set upper
+ * bounds greater than ZSTD_BLOCKSIZE_MAX or bounds lower than 1KB (will make
+ * compressBound() innacurate). Only currently meant to be used for testing.
+ *
+ */
+#define ZSTD_c_maxBlockSize ZSTD_c_experimentalParam18
+
+/* ZSTD_c_searchForExternalRepcodes
+ * This parameter affects how zstd parses external sequences, such as sequences
+ * provided through the compressSequences() API or from an external block-level
+ * sequence producer.
+ *
+ * If set to ZSTD_ps_enable, the library will check for repeated offsets in
+ * external sequences, even if those repcodes are not explicitly indicated in
+ * the "rep" field. Note that this is the only way to exploit repcode matches
+ * while using compressSequences() or an external sequence producer, since zstd
+ * currently ignores the "rep" field of external sequences.
+ *
+ * If set to ZSTD_ps_disable, the library will not exploit repeated offsets in
+ * external sequences, regardless of whether the "rep" field has been set. This
+ * reduces sequence compression overhead by about 25% while sacrificing some
+ * compression ratio.
+ *
+ * The default value is ZSTD_ps_auto, for which the library will enable/disable
+ * based on compression level.
+ *
+ * Note: for now, this param only has an effect if ZSTD_c_blockDelimiters is
+ * set to ZSTD_sf_explicitBlockDelimiters. That may change in the future.
+ */
+#define ZSTD_c_searchForExternalRepcodes ZSTD_c_experimentalParam19
+
 /*! ZSTD_CCtx_getParameter() :
  *  Get the requested compression parameter value, selected by enum ZSTD_cParameter,
  *  and store it into int* value.
@@ -2105,7 +2319,7 @@ ZSTDLIB_STATIC_API size_t ZSTD_DCtx_getParameter(ZSTD_DCtx* dctx, ZSTD_dParamete
  * in the range [dst, dst + pos) MUST not be modified during decompression
  * or you will get data corruption.
  *
- * When this flags is enabled zstd won't allocate an output buffer, because
+ * When this flag is enabled zstd won't allocate an output buffer, because
  * it can write directly to the ZSTD_outBuffer, but it will still allocate
  * an input buffer large enough to fit any compressed block. This will also
  * avoid the memcpy() from the internal output buffer to the ZSTD_outBuffer.
@@ -2158,6 +2372,17 @@ ZSTDLIB_STATIC_API size_t ZSTD_DCtx_getParameter(ZSTD_DCtx* dctx, ZSTD_dParamete
  */
 #define ZSTD_d_refMultipleDDicts ZSTD_d_experimentalParam4
 
+/* ZSTD_d_disableHuffmanAssembly
+ * Set to 1 to disable the Huffman assembly implementation.
+ * The default value is 0, which allows zstd to use the Huffman assembly
+ * implementation if available.
+ *
+ * This parameter can be used to disable Huffman assembly at runtime.
+ * If you want to disable it at compile time you can define the macro
+ * ZSTD_DISABLE_ASM.
+ */
+#define ZSTD_d_disableHuffmanAssembly ZSTD_d_experimentalParam5
+
 
 /*! ZSTD_DCtx_setFormat() :
  *  This function is REDUNDANT. Prefer ZSTD_DCtx_setParameter().
@@ -2166,6 +2391,7 @@ ZSTDLIB_STATIC_API size_t ZSTD_DCtx_getParameter(ZSTD_DCtx* dctx, ZSTD_dParamete
  *  such ZSTD_f_zstd1_magicless for example.
  * @return : 0, or an error code (which can be tested using ZSTD_isError()). */
 ZSTD_DEPRECATED("use ZSTD_DCtx_setParameter() instead")
+ZSTDLIB_STATIC_API
 size_t ZSTD_DCtx_setFormat(ZSTD_DCtx* dctx, ZSTD_format_e format);
 
 /*! ZSTD_decompressStream_simpleArgs() :
@@ -2202,6 +2428,7 @@ ZSTDLIB_STATIC_API size_t ZSTD_decompressStream_simpleArgs (
  * This prototype will generate compilation warnings.
  */
 ZSTD_DEPRECATED("use ZSTD_CCtx_reset, see zstd.h for detailed instructions")
+ZSTDLIB_STATIC_API
 size_t ZSTD_initCStream_srcSize(ZSTD_CStream* zcs,
                          int compressionLevel,
                          unsigned long long pledgedSrcSize);
@@ -2219,6 +2446,7 @@ size_t ZSTD_initCStream_srcSize(ZSTD_CStream* zcs,
  * This prototype will generate compilation warnings.
  */
 ZSTD_DEPRECATED("use ZSTD_CCtx_reset, see zstd.h for detailed instructions")
+ZSTDLIB_STATIC_API
 size_t ZSTD_initCStream_usingDict(ZSTD_CStream* zcs,
                      const void* dict, size_t dictSize,
                            int compressionLevel);
@@ -2239,6 +2467,7 @@ size_t ZSTD_initCStream_usingDict(ZSTD_CStream* zcs,
  * This prototype will generate compilation warnings.
  */
 ZSTD_DEPRECATED("use ZSTD_CCtx_reset, see zstd.h for detailed instructions")
+ZSTDLIB_STATIC_API
 size_t ZSTD_initCStream_advanced(ZSTD_CStream* zcs,
                     const void* dict, size_t dictSize,
                           ZSTD_parameters params,
@@ -2253,6 +2482,7 @@ size_t ZSTD_initCStream_advanced(ZSTD_CStream* zcs,
  * This prototype will generate compilation warnings.
  */
 ZSTD_DEPRECATED("use ZSTD_CCtx_reset and ZSTD_CCtx_refCDict, see zstd.h for detailed instructions")
+ZSTDLIB_STATIC_API
 size_t ZSTD_initCStream_usingCDict(ZSTD_CStream* zcs, const ZSTD_CDict* cdict);
 
 /*! ZSTD_initCStream_usingCDict_advanced() :
@@ -2271,6 +2501,7 @@ size_t ZSTD_initCStream_usingCDict(ZSTD_CStream* zcs, const ZSTD_CDict* cdict);
  * This prototype will generate compilation warnings.
  */
 ZSTD_DEPRECATED("use ZSTD_CCtx_reset and ZSTD_CCtx_refCDict, see zstd.h for detailed instructions")
+ZSTDLIB_STATIC_API
 size_t ZSTD_initCStream_usingCDict_advanced(ZSTD_CStream* zcs,
                                const ZSTD_CDict* cdict,
                                      ZSTD_frameParameters fParams,
@@ -2295,6 +2526,7 @@ size_t ZSTD_initCStream_usingCDict_advanced(ZSTD_CStream* zcs,
  *  This prototype will generate compilation warnings.
  */
 ZSTD_DEPRECATED("use ZSTD_CCtx_reset, see zstd.h for detailed instructions")
+ZSTDLIB_STATIC_API
 size_t ZSTD_resetCStream(ZSTD_CStream* zcs, unsigned long long pledgedSrcSize);
 
 
@@ -2340,8 +2572,8 @@ ZSTDLIB_STATIC_API size_t ZSTD_toFlushNow(ZSTD_CCtx* cctx);
  *     ZSTD_DCtx_loadDictionary(zds, dict, dictSize);
  *
  * note: no dictionary will be used if dict == NULL or dictSize < 8
- * Note : this prototype will be marked as deprecated and generate compilation warnings on reaching v1.5.x
  */
+ZSTD_DEPRECATED("use ZSTD_DCtx_reset + ZSTD_DCtx_loadDictionary, see zstd.h for detailed instructions")
 ZSTDLIB_STATIC_API size_t ZSTD_initDStream_usingDict(ZSTD_DStream* zds, const void* dict, size_t dictSize);
 
 /*!
@@ -2351,8 +2583,8 @@ ZSTDLIB_STATIC_API size_t ZSTD_initDStream_usingDict(ZSTD_DStream* zds, const vo
  *     ZSTD_DCtx_refDDict(zds, ddict);
  *
  * note : ddict is referenced, it must outlive decompression session
- * Note : this prototype will be marked as deprecated and generate compilation warnings on reaching v1.5.x
  */
+ZSTD_DEPRECATED("use ZSTD_DCtx_reset + ZSTD_DCtx_refDDict, see zstd.h for detailed instructions")
 ZSTDLIB_STATIC_API size_t ZSTD_initDStream_usingDDict(ZSTD_DStream* zds, const ZSTD_DDict* ddict);
 
 /*!
@@ -2361,8 +2593,8 @@ ZSTDLIB_STATIC_API size_t ZSTD_initDStream_usingDDict(ZSTD_DStream* zds, const Z
  *     ZSTD_DCtx_reset(zds, ZSTD_reset_session_only);
  *
  * re-use decompression parameters from previous init; saves dictionary loading
- * Note : this prototype will be marked as deprecated and generate compilation warnings on reaching v1.5.x
  */
+ZSTD_DEPRECATED("use ZSTD_DCtx_reset, see zstd.h for detailed instructions")
 ZSTDLIB_STATIC_API size_t ZSTD_resetDStream(ZSTD_DStream* zds);
 
 
@@ -2383,7 +2615,6 @@ ZSTDLIB_STATIC_API size_t ZSTD_resetDStream(ZSTD_DStream* zds);
 
   Start by initializing a context.
   Use ZSTD_compressBegin(), or ZSTD_compressBegin_usingDict() for dictionary compression.
-  It's also possible to duplicate a reference context which has already been initialized, using ZSTD_copyCCtx()
 
   Then, consume your input using ZSTD_compressContinue().
   There are some important considerations to keep in mind when using this advanced function :
@@ -2408,15 +2639,20 @@ ZSTDLIB_STATIC_API size_t ZSTD_resetDStream(ZSTD_DStream* zds);
 ZSTDLIB_STATIC_API size_t ZSTD_compressBegin(ZSTD_CCtx* cctx, int compressionLevel);
 ZSTDLIB_STATIC_API size_t ZSTD_compressBegin_usingDict(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, int compressionLevel);
 ZSTDLIB_STATIC_API size_t ZSTD_compressBegin_usingCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict); /**< note: fails if cdict==NULL */
-ZSTDLIB_STATIC_API size_t ZSTD_copyCCtx(ZSTD_CCtx* cctx, const ZSTD_CCtx* preparedCCtx, unsigned long long pledgedSrcSize); /**<  note: if pledgedSrcSize is not known, use ZSTD_CONTENTSIZE_UNKNOWN */
+
+ZSTD_DEPRECATED("This function will likely be removed in a future release. It is misleading and has very limited utility.")
+ZSTDLIB_STATIC_API
+size_t ZSTD_copyCCtx(ZSTD_CCtx* cctx, const ZSTD_CCtx* preparedCCtx, unsigned long long pledgedSrcSize); /**<  note: if pledgedSrcSize is not known, use ZSTD_CONTENTSIZE_UNKNOWN */
 
 ZSTDLIB_STATIC_API size_t ZSTD_compressContinue(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
 ZSTDLIB_STATIC_API size_t ZSTD_compressEnd(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
 
 /* The ZSTD_compressBegin_advanced() and ZSTD_compressBegin_usingCDict_advanced() are now DEPRECATED and will generate a compiler warning */
 ZSTD_DEPRECATED("use advanced API to access custom parameters")
+ZSTDLIB_STATIC_API
 size_t ZSTD_compressBegin_advanced(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, ZSTD_parameters params, unsigned long long pledgedSrcSize); /**< pledgedSrcSize : If srcSize is not known at init time, use ZSTD_CONTENTSIZE_UNKNOWN */
 ZSTD_DEPRECATED("use advanced API to access custom parameters")
+ZSTDLIB_STATIC_API
 size_t ZSTD_compressBegin_usingCDict_advanced(ZSTD_CCtx* const cctx, const ZSTD_CDict* const cdict, ZSTD_frameParameters const fParams, unsigned long long const pledgedSrcSize);   /* compression parameters are already set within cdict. pledgedSrcSize must be correct. If srcSize is not known, use macro ZSTD_CONTENTSIZE_UNKNOWN */
 /**
   Buffer-less streaming decompression (synchronous mode)
@@ -2429,8 +2665,8 @@ size_t ZSTD_compressBegin_usingCDict_advanced(ZSTD_CCtx* const cctx, const ZSTD_
   Frame header is extracted from the beginning of compressed frame, so providing only the frame's beginning is enough.
   Data fragment must be large enough to ensure successful decoding.
  `ZSTD_frameHeaderSize_max` bytes is guaranteed to always be large enough.
-  @result : 0 : successful decoding, the `ZSTD_frameHeader` structure is correctly filled.
+  result  : 0 : successful decoding, the `ZSTD_frameHeader` structure is correctly filled.
-           >0 : `srcSize` is too small, please provide at least @result bytes on next attempt.
+           >0 : `srcSize` is too small, please provide at least result bytes on next attempt.
            errorCode, which can be tested using ZSTD_isError().
 
   It fills a ZSTD_frameHeader structure with important information to correctly decode the frame,
@@ -2449,7 +2685,7 @@ size_t ZSTD_compressBegin_usingCDict_advanced(ZSTD_CCtx* const cctx, const ZSTD_
 
   The most memory efficient way is to use a round buffer of sufficient size.
   Sufficient size is determined by invoking ZSTD_decodingBufferSize_min(),
-  which can @return an error code if required value is too large for current system (in 32-bits mode).
+  which can return an error code if required value is too large for current system (in 32-bits mode).
   In a round buffer methodology, ZSTD_decompressContinue() decompresses each block next to previous one,
   up to the moment there is not enough room left in the buffer to guarantee decoding another full block,
   which maximum size is provided in `ZSTD_frameHeader` structure, field `blockSizeMax`.
@@ -2469,7 +2705,7 @@ size_t ZSTD_compressBegin_usingCDict_advanced(ZSTD_CCtx* const cctx, const ZSTD_
   ZSTD_nextSrcSizeToDecompress() tells how many bytes to provide as 'srcSize' to ZSTD_decompressContinue().
   ZSTD_decompressContinue() requires this _exact_ amount of bytes, or it will fail.
 
- @result of ZSTD_decompressContinue() is the number of bytes regenerated within 'dst' (necessarily <= dstCapacity).
+  result of ZSTD_decompressContinue() is the number of bytes regenerated within 'dst' (necessarily <= dstCapacity).
   It can be zero : it just means ZSTD_decompressContinue() has decoded some metadata item.
   It can also be an error code, which can be tested with ZSTD_isError().
 
@@ -2501,6 +2737,8 @@ typedef struct {
     unsigned headerSize;
     unsigned dictID;
     unsigned checksumFlag;
+    unsigned _reserved1;
+    unsigned _reserved2;
 } ZSTD_frameHeader;
 
 /*! ZSTD_getFrameHeader() :
@@ -2523,6 +2761,7 @@ ZSTDLIB_STATIC_API size_t ZSTD_nextSrcSizeToDecompress(ZSTD_DCtx* dctx);
 ZSTDLIB_STATIC_API size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
 
 /* misc */
+ZSTD_DEPRECATED("This function will likely be removed in the next minor release. It is misleading and has very limited utility.")
 ZSTDLIB_STATIC_API void   ZSTD_copyDCtx(ZSTD_DCtx* dctx, const ZSTD_DCtx* preparedDCtx);
 typedef enum { ZSTDnit_frameHeader, ZSTDnit_blockHeader, ZSTDnit_block, ZSTDnit_lastBlock, ZSTDnit_checksum, ZSTDnit_skippableFrame } ZSTD_nextInputType_e;
 ZSTDLIB_STATIC_API ZSTD_nextInputType_e ZSTD_nextInputType(ZSTD_DCtx* dctx);
@@ -2545,7 +2784,6 @@ ZSTDLIB_STATIC_API ZSTD_nextInputType_e ZSTD_nextInputType(ZSTD_DCtx* dctx);
     - It is necessary to init context before starting
       + compression : any ZSTD_compressBegin*() variant, including with dictionary
       + decompression : any ZSTD_decompressBegin*() variant, including with dictionary
-      + copyCCtx() and copyDCtx() can be used too
     - Block size is limited, it must be <= ZSTD_getBlockSize() <= ZSTD_BLOCKSIZE_MAX == 128 KB
       + If input is larger than a block size, it's necessary to split input data into multiple blocks
       + For inputs larger than a single block, consider using regular ZSTD_compress() instead.
@@ -2568,6 +2806,167 @@ ZSTDLIB_STATIC_API size_t ZSTD_decompressBlock(ZSTD_DCtx* dctx, void* dst, size_
 ZSTDLIB_STATIC_API size_t ZSTD_insertBlock    (ZSTD_DCtx* dctx, const void* blockStart, size_t blockSize);  /**< insert uncompressed block into `dctx` history. Useful for multi-blocks decompression. */
 
 
+/* ********************* BLOCK-LEVEL SEQUENCE PRODUCER API *********************
+ *
+ * *** OVERVIEW ***
+ * The Block-Level Sequence Producer API allows users to provide their own custom
+ * sequence producer which libzstd invokes to process each block. The produced list
+ * of sequences (literals and matches) is then post-processed by libzstd to produce
+ * valid compressed blocks.
+ *
+ * This block-level offload API is a more granular complement of the existing
+ * frame-level offload API compressSequences() (introduced in v1.5.1). It offers
+ * an easier migration story for applications already integrated with libzstd: the
+ * user application continues to invoke the same compression functions
+ * ZSTD_compress2() or ZSTD_compressStream2() as usual, and transparently benefits
+ * from the specific advantages of the external sequence producer. For example,
+ * the sequence producer could be tuned to take advantage of known characteristics
+ * of the input, to offer better speed / ratio, or could leverage hardware
+ * acceleration not available within libzstd itself.
+ *
+ * See contrib/externalSequenceProducer for an example program employing the
+ * Block-Level Sequence Producer API.
+ *
+ * *** USAGE ***
+ * The user is responsible for implementing a function of type
+ * ZSTD_sequenceProducer_F. For each block, zstd will pass the following
+ * arguments to the user-provided function:
+ *
+ *   - sequenceProducerState: a pointer to a user-managed state for the sequence
+ *     producer.
+ *
+ *   - outSeqs, outSeqsCapacity: an output buffer for the sequence producer.
+ *     outSeqsCapacity is guaranteed >= ZSTD_sequenceBound(srcSize). The memory
+ *     backing outSeqs is managed by the CCtx.
+ *
+ *   - src, srcSize: an input buffer for the sequence producer to parse.
+ *     srcSize is guaranteed to be <= ZSTD_BLOCKSIZE_MAX.
+ *
+ *   - dict, dictSize: a history buffer, which may be empty, which the sequence
+ *     producer may reference as it parses the src buffer. Currently, zstd will
+ *     always pass dictSize == 0 into external sequence producers, but this will
+ *     change in the future.
+ *
+ *   - compressionLevel: a signed integer representing the zstd compression level
+ *     set by the user for the current operation. The sequence producer may choose
+ *     to use this information to change its compression strategy and speed/ratio
+ *     tradeoff. Note: the compression level does not reflect zstd parameters set
+ *     through the advanced API.
+ *
+ *   - windowSize: a size_t representing the maximum allowed offset for external
+ *     sequences. Note that sequence offsets are sometimes allowed to exceed the
+ *     windowSize if a dictionary is present, see doc/zstd_compression_format.md
+ *     for details.
+ *
+ * The user-provided function shall return a size_t representing the number of
+ * sequences written to outSeqs. This return value will be treated as an error
+ * code if it is greater than outSeqsCapacity. The return value must be non-zero
+ * if srcSize is non-zero. The ZSTD_SEQUENCE_PRODUCER_ERROR macro is provided
+ * for convenience, but any value greater than outSeqsCapacity will be treated as
+ * an error code.
+ *
+ * If the user-provided function does not return an error code, the sequences
+ * written to outSeqs must be a valid parse of the src buffer. Data corruption may
+ * occur if the parse is not valid. A parse is defined to be valid if the
+ * following conditions hold:
+ *   - The sum of matchLengths and literalLengths must equal srcSize.
+ *   - All sequences in the parse, except for the final sequence, must have
+ *     matchLength >= ZSTD_MINMATCH_MIN. The final sequence must have
+ *     matchLength >= ZSTD_MINMATCH_MIN or matchLength == 0.
+ *   - All offsets must respect the windowSize parameter as specified in
+ *     doc/zstd_compression_format.md.
+ *   - If the final sequence has matchLength == 0, it must also have offset == 0.
+ *
+ * zstd will only validate these conditions (and fail compression if they do not
+ * hold) if the ZSTD_c_validateSequences cParam is enabled. Note that sequence
+ * validation has a performance cost.
+ *
+ * If the user-provided function returns an error, zstd will either fall back
+ * to an internal sequence producer or fail the compression operation. The user can
+ * choose between the two behaviors by setting the ZSTD_c_enableSeqProducerFallback
+ * cParam. Fallback compression will follow any other cParam settings, such as
+ * compression level, the same as in a normal compression operation.
+ *
+ * The user shall instruct zstd to use a particular ZSTD_sequenceProducer_F
+ * function by calling
+ *         ZSTD_registerSequenceProducer(cctx,
+ *                                       sequenceProducerState,
+ *                                       sequenceProducer)
+ * This setting will persist until the next parameter reset of the CCtx.
+ *
+ * The sequenceProducerState must be initialized by the user before calling
+ * ZSTD_registerSequenceProducer(). The user is responsible for destroying the
+ * sequenceProducerState.
+ *
+ * *** LIMITATIONS ***
+ * This API is compatible with all zstd compression APIs which respect advanced parameters.
+ * However, there are three limitations:
+ *
+ * First, the ZSTD_c_enableLongDistanceMatching cParam is not currently supported.
+ * COMPRESSION WILL FAIL if it is enabled and the user tries to compress with a block-level
+ * external sequence producer.
+ *   - Note that ZSTD_c_enableLongDistanceMatching is auto-enabled by default in some
+ *     cases (see its documentation for details). Users must explicitly set
+ *     ZSTD_c_enableLongDistanceMatching to ZSTD_ps_disable in such cases if an external
+ *     sequence producer is registered.
+ *   - As of this writing, ZSTD_c_enableLongDistanceMatching is disabled by default
+ *     whenever ZSTD_c_windowLog < 128MB, but that's subject to change. Users should
+ *     check the docs on ZSTD_c_enableLongDistanceMatching whenever the Block-Level Sequence
+ *     Producer API is used in conjunction with advanced settings (like ZSTD_c_windowLog).
+ *
+ * Second, history buffers are not currently supported. Concretely, zstd will always pass
+ * dictSize == 0 to the external sequence producer (for now). This has two implications:
+ *   - Dictionaries are not currently supported. Compression will *not* fail if the user
+ *     references a dictionary, but the dictionary won't have any effect.
+ *   - Stream history is not currently supported. All advanced compression APIs, including
+ *     streaming APIs, work with external sequence producers, but each block is treated as
+ *     an independent chunk without history from previous blocks.
+ *
+ * Third, multi-threading within a single compression is not currently supported. In other words,
+ * COMPRESSION WILL FAIL if ZSTD_c_nbWorkers > 0 and an external sequence producer is registered.
+ * Multi-threading across compressions is fine: simply create one CCtx per thread.
+ *
+ * Long-term, we plan to overcome all three limitations. There is no technical blocker to
+ * overcoming them. It is purely a question of engineering effort.
+ */
+
+#define ZSTD_SEQUENCE_PRODUCER_ERROR ((size_t)(-1))
+
+typedef size_t ZSTD_sequenceProducer_F (
+  void* sequenceProducerState,
+  ZSTD_Sequence* outSeqs, size_t outSeqsCapacity,
+  const void* src, size_t srcSize,
+  const void* dict, size_t dictSize,
+  int compressionLevel,
+  size_t windowSize
+);
+
+/*! ZSTD_registerSequenceProducer() :
+ * Instruct zstd to use a block-level external sequence producer function.
+ *
+ * The sequenceProducerState must be initialized by the caller, and the caller is
+ * responsible for managing its lifetime. This parameter is sticky across
+ * compressions. It will remain set until the user explicitly resets compression
+ * parameters.
+ *
+ * Sequence producer registration is considered to be an "advanced parameter",
+ * part of the "advanced API". This means it will only have an effect on compression
+ * APIs which respect advanced parameters, such as compress2() and compressStream2().
+ * Older compression APIs such as compressCCtx(), which predate the introduction of
+ * "advanced parameters", will ignore any external sequence producer setting.
+ *
+ * The sequence producer can be "cleared" by registering a NULL function pointer. This
+ * removes all limitations described above in the "LIMITATIONS" section of the API docs.
+ *
+ * The user is strongly encouraged to read the full API documentation (above) before
+ * calling this function. */
+ZSTDLIB_STATIC_API void
+ZSTD_registerSequenceProducer(
+  ZSTD_CCtx* cctx,
+  void* sequenceProducerState,
+  ZSTD_sequenceProducer_F* sequenceProducer
+);
+
 #endif   /* ZSTD_H_ZSTD_STATIC_LINKING_ONLY */
 
 #if defined (__cplusplus)

zstd/zstd_errors.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 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
+#ifndef ZSTDERRORLIB_VISIBLE
-#  if defined(__GNUC__) && (__GNUC__ >= 4)
+   /* Backwards compatibility with old macro name */
-#    define ZSTDERRORLIB_VISIBILITY __attribute__ ((visibility ("default")))
+#  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,
@@ -73,11 +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;