Bump zstd to 1.5.1.

capture/build/win32/capture.vcxproj

@@ -165,11 +165,13 @@
     <ClInclude Include="..\..\..\zstd\common\huf.h" />
     <ClInclude Include="..\..\..\zstd\common\mem.h" />
     <ClInclude Include="..\..\..\zstd\common\pool.h" />
+    <ClInclude Include="..\..\..\zstd\common\portability_macros.h" />
     <ClInclude Include="..\..\..\zstd\common\threading.h" />
     <ClInclude Include="..\..\..\zstd\common\xxhash.h" />
     <ClInclude Include="..\..\..\zstd\common\zstd_deps.h" />
     <ClInclude Include="..\..\..\zstd\common\zstd_internal.h" />
     <ClInclude Include="..\..\..\zstd\common\zstd_trace.h" />
+    <ClInclude Include="..\..\..\zstd\compress\clevels.h" />
     <ClInclude Include="..\..\..\zstd\compress\hist.h" />
     <ClInclude Include="..\..\..\zstd\compress\zstdmt_compress.h" />
     <ClInclude Include="..\..\..\zstd\compress\zstd_compress_internal.h" />
@@ -192,6 +194,9 @@
     <ClInclude Include="..\..\..\zstd\zstd.h" />
     <ClInclude Include="..\..\..\zstd\zstd_errors.h" />
   </ItemGroup>
+  <ItemGroup>
+    <None Include="..\..\..\zstd\decompress\huf_decompress_amd64.S" />
+  </ItemGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">
   </ImportGroup>

capture/build/win32/capture.vcxproj.filters

@@ -344,5 +344,16 @@
     <ClInclude Include="..\..\..\common\TracyStackFrames.hpp">
       <Filter>common</Filter>
     </ClInclude>
+    <ClInclude Include="..\..\..\zstd\common\portability_macros.h">
+      <Filter>zstd\common</Filter>
+    </ClInclude>
+    <ClInclude Include="..\..\..\zstd\compress\clevels.h">
+      <Filter>zstd\compress</Filter>
+    </ClInclude>
+  </ItemGroup>
+  <ItemGroup>
+    <None Include="..\..\..\zstd\decompress\huf_decompress_amd64.S">
+      <Filter>zstd\decompress</Filter>
+    </None>
   </ItemGroup>
 </Project>

csvexport/build/win32/csvexport.vcxproj

@@ -165,11 +165,13 @@
     <ClInclude Include="..\..\..\zstd\common\huf.h" />
     <ClInclude Include="..\..\..\zstd\common\mem.h" />
     <ClInclude Include="..\..\..\zstd\common\pool.h" />
+    <ClInclude Include="..\..\..\zstd\common\portability_macros.h" />
     <ClInclude Include="..\..\..\zstd\common\threading.h" />
     <ClInclude Include="..\..\..\zstd\common\xxhash.h" />
     <ClInclude Include="..\..\..\zstd\common\zstd_deps.h" />
     <ClInclude Include="..\..\..\zstd\common\zstd_internal.h" />
     <ClInclude Include="..\..\..\zstd\common\zstd_trace.h" />
+    <ClInclude Include="..\..\..\zstd\compress\clevels.h" />
     <ClInclude Include="..\..\..\zstd\compress\hist.h" />
     <ClInclude Include="..\..\..\zstd\compress\zstdmt_compress.h" />
     <ClInclude Include="..\..\..\zstd\compress\zstd_compress_internal.h" />
@@ -192,6 +194,9 @@
     <ClInclude Include="..\..\..\zstd\zstd.h" />
     <ClInclude Include="..\..\..\zstd\zstd_errors.h" />
   </ItemGroup>
+  <ItemGroup>
+    <None Include="..\..\..\zstd\decompress\huf_decompress_amd64.S" />
+  </ItemGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">
   </ImportGroup>

csvexport/build/win32/csvexport.vcxproj.filters

@@ -344,5 +344,16 @@
     <ClInclude Include="..\..\..\common\TracyStackFrames.hpp">
       <Filter>common</Filter>
     </ClInclude>
+    <ClInclude Include="..\..\..\zstd\common\portability_macros.h">
+      <Filter>zstd\common</Filter>
+    </ClInclude>
+    <ClInclude Include="..\..\..\zstd\compress\clevels.h">
+      <Filter>zstd\compress</Filter>
+    </ClInclude>
+  </ItemGroup>
+  <ItemGroup>
+    <None Include="..\..\..\zstd\decompress\huf_decompress_amd64.S">
+      <Filter>zstd\decompress</Filter>
+    </None>
   </ItemGroup>
 </Project>

import-chrome/build/win32/import-chrome.vcxproj

@@ -161,11 +161,13 @@
     <ClInclude Include="..\..\..\zstd\common\huf.h" />
     <ClInclude Include="..\..\..\zstd\common\mem.h" />
     <ClInclude Include="..\..\..\zstd\common\pool.h" />
+    <ClInclude Include="..\..\..\zstd\common\portability_macros.h" />
     <ClInclude Include="..\..\..\zstd\common\threading.h" />
     <ClInclude Include="..\..\..\zstd\common\xxhash.h" />
     <ClInclude Include="..\..\..\zstd\common\zstd_deps.h" />
     <ClInclude Include="..\..\..\zstd\common\zstd_internal.h" />
     <ClInclude Include="..\..\..\zstd\common\zstd_trace.h" />
+    <ClInclude Include="..\..\..\zstd\compress\clevels.h" />
     <ClInclude Include="..\..\..\zstd\compress\hist.h" />
     <ClInclude Include="..\..\..\zstd\compress\zstdmt_compress.h" />
     <ClInclude Include="..\..\..\zstd\compress\zstd_compress_internal.h" />
@@ -188,6 +190,9 @@
     <ClInclude Include="..\..\..\zstd\zstd.h" />
     <ClInclude Include="..\..\..\zstd\zstd_errors.h" />
   </ItemGroup>
+  <ItemGroup>
+    <None Include="..\..\..\zstd\decompress\huf_decompress_amd64.S" />
+  </ItemGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">
   </ImportGroup>

import-chrome/build/win32/import-chrome.vcxproj.filters

@@ -329,5 +329,16 @@
     <ClInclude Include="..\..\..\common\TracyStackFrames.hpp">
       <Filter>common</Filter>
     </ClInclude>
+    <ClInclude Include="..\..\..\zstd\common\portability_macros.h">
+      <Filter>zstd\common</Filter>
+    </ClInclude>
+    <ClInclude Include="..\..\..\zstd\compress\clevels.h">
+      <Filter>zstd\compress</Filter>
+    </ClInclude>
+  </ItemGroup>
+  <ItemGroup>
+    <None Include="..\..\..\zstd\decompress\huf_decompress_amd64.S">
+      <Filter>zstd\decompress</Filter>
+    </None>
   </ItemGroup>
 </Project>

profiler/build/win32/Tracy.vcxproj

@@ -250,11 +250,13 @@
     <ClInclude Include="..\..\..\zstd\common\huf.h" />
     <ClInclude Include="..\..\..\zstd\common\mem.h" />
     <ClInclude Include="..\..\..\zstd\common\pool.h" />
+    <ClInclude Include="..\..\..\zstd\common\portability_macros.h" />
     <ClInclude Include="..\..\..\zstd\common\threading.h" />
     <ClInclude Include="..\..\..\zstd\common\xxhash.h" />
     <ClInclude Include="..\..\..\zstd\common\zstd_deps.h" />
     <ClInclude Include="..\..\..\zstd\common\zstd_internal.h" />
     <ClInclude Include="..\..\..\zstd\common\zstd_trace.h" />
+    <ClInclude Include="..\..\..\zstd\compress\clevels.h" />
     <ClInclude Include="..\..\..\zstd\compress\hist.h" />
     <ClInclude Include="..\..\..\zstd\compress\zstdmt_compress.h" />
     <ClInclude Include="..\..\..\zstd\compress\zstd_compress_internal.h" />
@@ -298,5 +300,8 @@
   <ItemGroup>
     <Manifest Include="Tracy.manifest" />
   </ItemGroup>
+  <ItemGroup>
+    <None Include="..\..\..\zstd\decompress\huf_decompress_amd64.S" />
+  </ItemGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
 </Project>

profiler/build/win32/Tracy.vcxproj.filters

@@ -578,6 +578,12 @@
     <ClInclude Include="..\..\src\FiraCodeRetina.hpp">
       <Filter>src</Filter>
     </ClInclude>
+    <ClInclude Include="..\..\..\zstd\common\portability_macros.h">
+      <Filter>zstd\common</Filter>
+    </ClInclude>
+    <ClInclude Include="..\..\..\zstd\compress\clevels.h">
+      <Filter>zstd\compress</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <Natvis Include="DebugVis.natvis" />
@@ -588,4 +594,9 @@
   <ItemGroup>
     <Manifest Include="Tracy.manifest" />
   </ItemGroup>
+  <ItemGroup>
+    <None Include="..\..\..\zstd\decompress\huf_decompress_amd64.S">
+      <Filter>zstd\decompress</Filter>
+    </None>
+  </ItemGroup>
 </Project>

update/build/win32/update.vcxproj

@@ -165,11 +165,13 @@
     <ClInclude Include="..\..\..\zstd\common\huf.h" />
     <ClInclude Include="..\..\..\zstd\common\mem.h" />
     <ClInclude Include="..\..\..\zstd\common\pool.h" />
+    <ClInclude Include="..\..\..\zstd\common\portability_macros.h" />
     <ClInclude Include="..\..\..\zstd\common\threading.h" />
     <ClInclude Include="..\..\..\zstd\common\xxhash.h" />
     <ClInclude Include="..\..\..\zstd\common\zstd_deps.h" />
     <ClInclude Include="..\..\..\zstd\common\zstd_internal.h" />
     <ClInclude Include="..\..\..\zstd\common\zstd_trace.h" />
+    <ClInclude Include="..\..\..\zstd\compress\clevels.h" />
     <ClInclude Include="..\..\..\zstd\compress\hist.h" />
     <ClInclude Include="..\..\..\zstd\compress\zstdmt_compress.h" />
     <ClInclude Include="..\..\..\zstd\compress\zstd_compress_internal.h" />
@@ -192,6 +194,9 @@
     <ClInclude Include="..\..\..\zstd\zstd.h" />
     <ClInclude Include="..\..\..\zstd\zstd_errors.h" />
   </ItemGroup>
+  <ItemGroup>
+    <None Include="..\..\..\zstd\decompress\huf_decompress_amd64.S" />
+  </ItemGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">
   </ImportGroup>

update/build/win32/update.vcxproj.filters

@@ -344,5 +344,16 @@
     <ClInclude Include="..\..\..\common\TracyStackFrames.hpp">
       <Filter>common</Filter>
     </ClInclude>
+    <ClInclude Include="..\..\..\zstd\common\portability_macros.h">
+      <Filter>zstd\common</Filter>
+    </ClInclude>
+    <ClInclude Include="..\..\..\zstd\compress\clevels.h">
+      <Filter>zstd\compress</Filter>
+    </ClInclude>
+  </ItemGroup>
+  <ItemGroup>
+    <None Include="..\..\..\zstd\decompress\huf_decompress_amd64.S">
+      <Filter>zstd\decompress</Filter>
+    </None>
   </ItemGroup>
 </Project>

zstd/common/bitstream.h

@@ -145,8 +145,14 @@ MEM_STATIC unsigned BIT_highbit32 (U32 val)
 #       if STATIC_BMI2 == 1
             return _lzcnt_u32(val) ^ 31;
 #       else
-		unsigned long r = 0;
+            if (val != 0) {
-		return _BitScanReverse(&r, val) ? (unsigned)r : 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;
@@ -293,22 +299,22 @@ MEM_STATIC size_t BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, si
         switch(srcSize)
         {
         case 7: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[6]) << (sizeof(bitD->bitContainer)*8 - 16);
-                /* fall-through */
+                ZSTD_FALLTHROUGH;
 
         case 6: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[5]) << (sizeof(bitD->bitContainer)*8 - 24);
-                /* fall-through */
+                ZSTD_FALLTHROUGH;
 
         case 5: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[4]) << (sizeof(bitD->bitContainer)*8 - 32);
-                /* fall-through */
+                ZSTD_FALLTHROUGH;
 
         case 4: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[3]) << 24;
-                /* fall-through */
+                ZSTD_FALLTHROUGH;
 
         case 3: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[2]) << 16;
-                /* fall-through */
+                ZSTD_FALLTHROUGH;
 
         case 2: bitD->bitContainer += (size_t)(((const BYTE*)(srcBuffer))[1]) <<  8;
-                /* fall-through */
+                ZSTD_FALLTHROUGH;
 
         default: break;
         }
@@ -332,7 +338,16 @@ MEM_STATIC FORCE_INLINE_ATTR size_t BIT_getMiddleBits(size_t bitContainer, U32 c
     U32 const regMask = sizeof(bitContainer)*8 - 1;
     /* if start > regMask, bitstream is corrupted, and result is undefined */
     assert(nbBits < BIT_MASK_SIZE);
+    /* x86 transform & ((1 << nbBits) - 1) to bzhi instruction, it is better
+     * than accessing memory. When bmi2 instruction is not present, we consider
+     * such cpus old (pre-Haswell, 2013) and their performance is not of that
+     * importance.
+     */
+#if defined(__x86_64__) || defined(_M_X86)
+    return (bitContainer >> (start & regMask)) & ((((U64)1) << nbBits) - 1);
+#else
     return (bitContainer >> (start & regMask)) & BIT_mask[nbBits];
+#endif
 }
 
 MEM_STATIC FORCE_INLINE_ATTR size_t BIT_getLowerBits(size_t bitContainer, U32 const nbBits)

zstd/common/compiler.h

@@ -11,6 +11,8 @@
 #ifndef ZSTD_COMPILER_H
 #define ZSTD_COMPILER_H
 
+#include "portability_macros.h"
+
 /*-*******************************************************
 *  Compiler specifics
 *********************************************************/
@@ -40,7 +42,7 @@
 
 /**
   On MSVC qsort requires that functions passed into it use the __cdecl calling conversion(CC).
-  This explictly marks such functions as __cdecl so that the code will still compile
+  This explicitly marks such functions as __cdecl so that the code will still compile
   if a CC other than __cdecl has been made the default.
 */
 #if  defined(_MSC_VER)
@@ -92,29 +94,17 @@
 
 
 /* target attribute */
-#ifndef __has_attribute
-  #define __has_attribute(x) 0  /* Compatibility with non-clang compilers. */
-#endif
 #if defined(__GNUC__) || defined(__ICCARM__)
 #  define TARGET_ATTRIBUTE(target) __attribute__((__target__(target)))
 #else
 #  define TARGET_ATTRIBUTE(target)
 #endif
 
-/* Enable runtime BMI2 dispatch based on the CPU.
+/* Target attribute for BMI2 dynamic dispatch.
- * Enabled for clang & gcc >=4.8 on x86 when BMI2 isn't enabled by default.
+ * Enable lzcnt, bmi, and bmi2.
+ * We test for bmi1 & bmi2. lzcnt is included in bmi1.
  */
-#ifndef DYNAMIC_BMI2
+#define BMI2_TARGET_ATTRIBUTE TARGET_ATTRIBUTE("lzcnt,bmi,bmi2")
-  #if ((defined(__clang__) && __has_attribute(__target__)) \
-      || (defined(__GNUC__) \
-          && (__GNUC__ >= 5 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8)))) \
-      && (defined(__x86_64__) || defined(_M_X86)) \
-      && !defined(__BMI2__)
-  #  define DYNAMIC_BMI2 1
-  #else
-  #  define DYNAMIC_BMI2 0
-  #endif
-#endif
 
 /* prefetch
  * can be disabled, by declaring NO_PREFETCH build macro */
@@ -150,8 +140,9 @@
 }
 
 /* vectorization
- * older GCC (pre gcc-4.3 picked as the cutoff) uses a different syntax */
+ * older GCC (pre gcc-4.3 picked as the cutoff) uses a different syntax,
-#if !defined(__INTEL_COMPILER) && !defined(__clang__) && defined(__GNUC__)
+ * and some compilers, like Intel ICC and MCST LCC, do not support it at all. */
+#if !defined(__INTEL_COMPILER) && !defined(__clang__) && defined(__GNUC__) && !defined(__LCC__)
 #  if (__GNUC__ == 4 && __GNUC_MINOR__ > 3) || (__GNUC__ >= 5)
 #    define DONT_VECTORIZE __attribute__((optimize("no-tree-vectorize")))
 #  else
@@ -197,25 +188,91 @@
     #define STATIC_BMI2 0
 #endif
 
-/* compat. with non-clang compilers */
+/* compile time determination of SIMD support */
-#ifndef __has_builtin
+#if !defined(ZSTD_NO_INTRINSICS)
-#  define __has_builtin(x) 0
+#  if defined(__SSE2__) || defined(_M_AMD64) || (defined (_M_IX86) && defined(_M_IX86_FP) && (_M_IX86_FP >= 2))
+#    define ZSTD_ARCH_X86_SSE2
+#  endif
+#  if defined(__ARM_NEON) || defined(_M_ARM64)
+#    define ZSTD_ARCH_ARM_NEON
+#  endif
+#
+#  if defined(ZSTD_ARCH_X86_SSE2)
+#    include <emmintrin.h>
+#  elif defined(ZSTD_ARCH_ARM_NEON)
+#    include <arm_neon.h>
+#  endif
 #endif
 
-/* compat. with non-clang compilers */
+/* C-language Attributes are added in C23. */
-#ifndef __has_feature
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ > 201710L) && defined(__has_c_attribute)
-#  define __has_feature(x) 0
+# define ZSTD_HAS_C_ATTRIBUTE(x) __has_c_attribute(x)
-#endif
-
-/* detects whether we are being compiled under msan */
-#ifndef ZSTD_MEMORY_SANITIZER
-#  if __has_feature(memory_sanitizer)
-#    define ZSTD_MEMORY_SANITIZER 1
 #else
-#    define ZSTD_MEMORY_SANITIZER 0
+# define ZSTD_HAS_C_ATTRIBUTE(x) 0
+#endif
+
+/* Only use C++ attributes in C++. Some compilers report support for C++
+ * attributes when compiling with C.
+ */
+#if defined(__cplusplus) && defined(__has_cpp_attribute)
+# define ZSTD_HAS_CPP_ATTRIBUTE(x) __has_cpp_attribute(x)
+#else
+# define ZSTD_HAS_CPP_ATTRIBUTE(x) 0
+#endif
+
+/* Define ZSTD_FALLTHROUGH macro for annotating switch case with the 'fallthrough' attribute.
+ * - C23: https://en.cppreference.com/w/c/language/attributes/fallthrough
+ * - CPP17: https://en.cppreference.com/w/cpp/language/attributes/fallthrough
+ * - Else: __attribute__((__fallthrough__))
+ */
+#ifndef ZSTD_FALLTHROUGH
+# if ZSTD_HAS_C_ATTRIBUTE(fallthrough)
+#  define ZSTD_FALLTHROUGH [[fallthrough]]
+# elif ZSTD_HAS_CPP_ATTRIBUTE(fallthrough)
+#  define ZSTD_FALLTHROUGH [[fallthrough]]
+# elif __has_attribute(__fallthrough__)
+/* Leading semicolon is to satisfy gcc-11 with -pedantic. Without the semicolon
+ * gcc complains about: a label can only be part of a statement and a declaration is not a statement.
+ */
+#  define ZSTD_FALLTHROUGH ; __attribute__((__fallthrough__))
+# else
+#  define ZSTD_FALLTHROUGH
 # endif
 #endif
 
+/*-**************************************************************
+*  Alignment check
+*****************************************************************/
+
+/* this test was initially positioned in mem.h,
+ * but this file is removed (or replaced) for linux kernel
+ * so it's now hosted in compiler.h,
+ * which remains valid for both user & kernel spaces.
+ */
+
+#ifndef ZSTD_ALIGNOF
+# if defined(__GNUC__) || defined(_MSC_VER)
+/* covers gcc, clang & MSVC */
+/* note : this section must come first, before C11,
+ * due to a limitation in the kernel source generator */
+#  define ZSTD_ALIGNOF(T) __alignof(T)
+
+# elif defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)
+/* C11 support */
+#  include <stdalign.h>
+#  define ZSTD_ALIGNOF(T) alignof(T)
+
+# else
+/* No known support for alignof() - imperfect backup */
+#  define ZSTD_ALIGNOF(T) (sizeof(void*) < sizeof(T) ? sizeof(void*) : sizeof(T))
+
+# endif
+#endif /* ZSTD_ALIGNOF */
+
+/*-**************************************************************
+*  Sanitizer
+*****************************************************************/
+
 #if ZSTD_MEMORY_SANITIZER
 /* Not all platforms that support msan provide sanitizers/msan_interface.h.
  * We therefore declare the functions we need ourselves, rather than trying to
@@ -237,17 +294,6 @@ void __msan_poison(const volatile void *a, size_t size);
 intptr_t __msan_test_shadow(const volatile void *x, size_t size);
 #endif
 
-/* detects whether we are being compiled under asan */
-#ifndef ZSTD_ADDRESS_SANITIZER
-#  if __has_feature(address_sanitizer)
-#    define ZSTD_ADDRESS_SANITIZER 1
-#  elif defined(__SANITIZE_ADDRESS__)
-#    define ZSTD_ADDRESS_SANITIZER 1
-#  else
-#    define ZSTD_ADDRESS_SANITIZER 0
-#  endif
-#endif
-
 #if ZSTD_ADDRESS_SANITIZER
 /* Not all platforms that support asan provide sanitizers/asan_interface.h.
  * We therefore declare the functions we need ourselves, rather than trying to

zstd/common/entropy_common.c

@@ -43,8 +43,14 @@ static U32 FSE_ctz(U32 val)
     assert(val != 0);
     {
 #   if defined(_MSC_VER)   /* Visual */
-        unsigned long r=0;
+        if (val != 0) {
-        return _BitScanForward(&r, val) ? (unsigned)r : 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 */
@@ -217,7 +223,7 @@ static size_t FSE_readNCount_body_default(
 }
 
 #if DYNAMIC_BMI2
-TARGET_ATTRIBUTE("bmi2") static size_t FSE_readNCount_body_bmi2(
+BMI2_TARGET_ATTRIBUTE static size_t FSE_readNCount_body_bmi2(
         short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
         const void* headerBuffer, size_t hbSize)
 {
@@ -299,7 +305,7 @@ HUF_readStats_body(BYTE* huffWeight, size_t hwSize, U32* rankStats,
     ZSTD_memset(rankStats, 0, (HUF_TABLELOG_MAX + 1) * sizeof(U32));
     weightTotal = 0;
     {   U32 n; for (n=0; n<oSize; n++) {
-            if (huffWeight[n] >= HUF_TABLELOG_MAX) return ERROR(corruption_detected);
+            if (huffWeight[n] > HUF_TABLELOG_MAX) return ERROR(corruption_detected);
             rankStats[huffWeight[n]]++;
             weightTotal += (1 << huffWeight[n]) >> 1;
     }   }
@@ -337,7 +343,7 @@ static size_t HUF_readStats_body_default(BYTE* huffWeight, size_t hwSize, U32* r
 }
 
 #if DYNAMIC_BMI2
-static TARGET_ATTRIBUTE("bmi2") size_t HUF_readStats_body_bmi2(BYTE* huffWeight, size_t hwSize, U32* rankStats,
+static BMI2_TARGET_ATTRIBUTE size_t HUF_readStats_body_bmi2(BYTE* huffWeight, size_t hwSize, U32* rankStats,
                      U32* nbSymbolsPtr, U32* tableLogPtr,
                      const void* src, size_t srcSize,
                      void* workSpace, size_t wkspSize)

zstd/common/error_private.h

@@ -22,6 +22,8 @@ extern "C" {
 *  Dependencies
 ******************************************/
 #include "../zstd_errors.h"  /* enum list */
+#include "compiler.h"
+#include "debug.h"
 #include "zstd_deps.h"       /* size_t */
 
 
@@ -73,6 +75,83 @@ ERR_STATIC const char* ERR_getErrorName(size_t code)
     return ERR_getErrorString(ERR_getErrorCode(code));
 }
 
+/**
+ * Ignore: this is an internal helper.
+ *
+ * This is a helper function to help force C99-correctness during compilation.
+ * Under strict compilation modes, variadic macro arguments can't be empty.
+ * However, variadic function arguments can be. Using a function therefore lets
+ * us statically check that at least one (string) argument was passed,
+ * independent of the compilation flags.
+ */
+static INLINE_KEYWORD UNUSED_ATTR
+void _force_has_format_string(const char *format, ...) {
+  (void)format;
+}
+
+/**
+ * Ignore: this is an internal helper.
+ *
+ * We want to force this function invocation to be syntactically correct, but
+ * we don't want to force runtime evaluation of its arguments.
+ */
+#define _FORCE_HAS_FORMAT_STRING(...) \
+  if (0) { \
+    _force_has_format_string(__VA_ARGS__); \
+  }
+
+#define ERR_QUOTE(str) #str
+
+/**
+ * Return the specified error if the condition evaluates to true.
+ *
+ * In debug modes, prints additional information.
+ * In order to do that (particularly, printing the conditional that failed),
+ * this can't just wrap RETURN_ERROR().
+ */
+#define RETURN_ERROR_IF(cond, err, ...) \
+  if (cond) { \
+    RAWLOG(3, "%s:%d: ERROR!: check %s failed, returning %s", \
+           __FILE__, __LINE__, ERR_QUOTE(cond), ERR_QUOTE(ERROR(err))); \
+    _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
+    RAWLOG(3, ": " __VA_ARGS__); \
+    RAWLOG(3, "\n"); \
+    return ERROR(err); \
+  }
+
+/**
+ * Unconditionally return the specified error.
+ *
+ * In debug modes, prints additional information.
+ */
+#define RETURN_ERROR(err, ...) \
+  do { \
+    RAWLOG(3, "%s:%d: ERROR!: unconditional check failed, returning %s", \
+           __FILE__, __LINE__, ERR_QUOTE(ERROR(err))); \
+    _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
+    RAWLOG(3, ": " __VA_ARGS__); \
+    RAWLOG(3, "\n"); \
+    return ERROR(err); \
+  } while(0);
+
+/**
+ * If the provided expression evaluates to an error code, returns that error code.
+ *
+ * In debug modes, prints additional information.
+ */
+#define FORWARD_IF_ERROR(err, ...) \
+  do { \
+    size_t const err_code = (err); \
+    if (ERR_isError(err_code)) { \
+      RAWLOG(3, "%s:%d: ERROR!: forwarding error in %s: %s", \
+             __FILE__, __LINE__, ERR_QUOTE(err), ERR_getErrorName(err_code)); \
+      _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
+      RAWLOG(3, ": " __VA_ARGS__); \
+      RAWLOG(3, "\n"); \
+      return err_code; \
+    } \
+  } while(0);
+
 #if defined (__cplusplus)
 }
 #endif

zstd/common/fse.h

@@ -336,8 +336,9 @@ size_t FSE_buildCTable_rle (FSE_CTable* ct, unsigned char symbolValue);
 /* FSE_buildCTable_wksp() :
  * Same as FSE_buildCTable(), but using an externally allocated scratch buffer (`workSpace`).
  * `wkspSize` must be >= `FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(maxSymbolValue, tableLog)` of `unsigned`.
+ * See FSE_buildCTable_wksp() for breakdown of workspace usage.
  */
-#define FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(maxSymbolValue, tableLog) (maxSymbolValue + 2 + (1ull << (tableLog - 2)))
+#define FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(maxSymbolValue, tableLog) (((maxSymbolValue + 2) + (1ull << (tableLog)))/2 + sizeof(U64)/sizeof(U32) /* additional 8 bytes for potential table overwrite */)
 #define FSE_BUILD_CTABLE_WORKSPACE_SIZE(maxSymbolValue, tableLog) (sizeof(unsigned) * FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(maxSymbolValue, tableLog))
 size_t FSE_buildCTable_wksp(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);
 

zstd/common/fse_decompress.c

@@ -365,7 +365,7 @@ static size_t FSE_decompress_wksp_body_default(void* dst, size_t dstCapacity, co
 }
 
 #if DYNAMIC_BMI2
-TARGET_ATTRIBUTE("bmi2") static size_t FSE_decompress_wksp_body_bmi2(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize)
+BMI2_TARGET_ATTRIBUTE static size_t FSE_decompress_wksp_body_bmi2(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize)
 {
     return FSE_decompress_wksp_body(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize, 1);
 }

zstd/common/huf.h

@@ -89,9 +89,9 @@ HUF_PUBLIC_API size_t HUF_compress2 (void* dst, size_t dstCapacity,
 
 /** HUF_compress4X_wksp() :
  *  Same as HUF_compress2(), but uses externally allocated `workSpace`.
- * `workspace` must have minimum alignment of 4, and be at least as large as HUF_WORKSPACE_SIZE */
+ * `workspace` must be at least as large as HUF_WORKSPACE_SIZE */
-#define HUF_WORKSPACE_SIZE ((6 << 10) + 256)
+#define HUF_WORKSPACE_SIZE ((8 << 10) + 512 /* sorting scratch space */)
-#define HUF_WORKSPACE_SIZE_U32 (HUF_WORKSPACE_SIZE / sizeof(U32))
+#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,
@@ -116,11 +116,11 @@ HUF_PUBLIC_API size_t HUF_compress4X_wksp (void* dst, size_t dstCapacity,
 
 
 /* *** Constants *** */
-#define HUF_TABLELOG_MAX      12      /* max runtime value of tableLog (due to static allocation); can be modified up to HUF_ABSOLUTEMAX_TABLELOG */
+#define HUF_TABLELOG_MAX      12      /* max runtime value of tableLog (due to static allocation); can be modified up to HUF_TABLELOG_ABSOLUTEMAX */
 #define HUF_TABLELOG_DEFAULT  11      /* default tableLog value when none specified */
 #define HUF_SYMBOLVALUE_MAX  255
 
-#define HUF_TABLELOG_ABSOLUTEMAX  15  /* absolute limit of HUF_MAX_TABLELOG. Beyond that value, code does not work */
+#define HUF_TABLELOG_ABSOLUTEMAX  12  /* absolute limit of HUF_MAX_TABLELOG. Beyond that value, code does not work */
 #if (HUF_TABLELOG_MAX > HUF_TABLELOG_ABSOLUTEMAX)
 #  error "HUF_TABLELOG_MAX is too large !"
 #endif
@@ -136,15 +136,11 @@ HUF_PUBLIC_API size_t HUF_compress4X_wksp (void* dst, size_t dstCapacity,
 
 /* static allocation of HUF's Compression Table */
 /* this is a private definition, just exposed for allocation and strict aliasing purpose. never EVER access its members directly */
-struct HUF_CElt_s {
+typedef size_t HUF_CElt;   /* consider it an incomplete type */
-  U16  val;
+#define HUF_CTABLE_SIZE_ST(maxSymbolValue)   ((maxSymbolValue)+2)   /* Use tables of size_t, for proper alignment */
-  BYTE nbBits;
+#define HUF_CTABLE_SIZE(maxSymbolValue)       (HUF_CTABLE_SIZE_ST(maxSymbolValue) * sizeof(size_t))
-};   /* typedef'd to HUF_CElt */
-typedef struct HUF_CElt_s HUF_CElt;   /* consider it an incomplete type */
-#define HUF_CTABLE_SIZE_U32(maxSymbolValue)   ((maxSymbolValue)+1)   /* Use tables of U32, for proper alignment */
-#define HUF_CTABLE_SIZE(maxSymbolValue)       (HUF_CTABLE_SIZE_U32(maxSymbolValue) * sizeof(U32))
 #define HUF_CREATE_STATIC_CTABLE(name, maxSymbolValue) \
-    HUF_CElt name[HUF_CTABLE_SIZE_U32(maxSymbolValue)] /* no final ; */
+    HUF_CElt name[HUF_CTABLE_SIZE_ST(maxSymbolValue)] /* no final ; */
 
 /* static allocation of HUF's DTable */
 typedef U32 HUF_DTable;
@@ -194,6 +190,7 @@ size_t HUF_buildCTable (HUF_CElt* CTable, const unsigned* count, unsigned maxSym
 size_t HUF_writeCTable (void* dst, size_t maxDstSize, const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog);
 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_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);
 
@@ -206,12 +203,13 @@ typedef enum {
  *  Same as HUF_compress4X_wksp(), but considers using hufTable if *repeat != HUF_repeat_none.
  *  If it uses hufTable it does not modify hufTable or repeat.
  *  If it doesn't, it sets *repeat = HUF_repeat_none, and it sets hufTable to the table used.
- *  If preferRepeat then the old table will always be used if valid. */
+ *  If preferRepeat then the old table will always be used if valid.
+ *  If suspectUncompressible then some sampling checks will be run to potentially skip huffman coding */
 size_t HUF_compress4X_repeat(void* dst, size_t dstSize,
                        const void* src, size_t srcSize,
                        unsigned maxSymbolValue, unsigned tableLog,
                        void* workSpace, size_t wkspSize,    /**< `workSpace` must be aligned on 4-bytes boundaries, `wkspSize` must be >= HUF_WORKSPACE_SIZE */
-                       HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2);
+                       HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2, unsigned suspectUncompressible);
 
 /** HUF_buildCTable_wksp() :
  *  Same as HUF_buildCTable(), but using externally allocated scratch buffer.
@@ -249,11 +247,10 @@ size_t HUF_readStats_wksp(BYTE* huffWeight, size_t hwSize,
  *  Loading a CTable saved with HUF_writeCTable() */
 size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void* src, size_t srcSize, unsigned *hasZeroWeights);
 
-/** HUF_getNbBits() :
+/** HUF_getNbBitsFromCTable() :
  *  Read nbBits from CTable symbolTable, for symbol `symbolValue` presumed <= HUF_SYMBOLVALUE_MAX
- *  Note 1 : is not inlined, as HUF_CElt definition is private
+ *  Note 1 : is not inlined, as HUF_CElt definition is private */
- *  Note 2 : const void* used, so that it can provide a statically allocated table as argument (which uses type U32) */
+U32 HUF_getNbBitsFromCTable(const HUF_CElt* symbolTable, U32 symbolValue);
-U32 HUF_getNbBits(const void* symbolTable, U32 symbolValue);
 
 /*
  * HUF_decompress() does the following:
@@ -305,18 +302,20 @@ size_t HUF_decompress4X2_usingDTable(void* dst, size_t maxDstSize, const void* c
 /* ====================== */
 
 size_t HUF_compress1X (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog);
-size_t HUF_compress1X_wksp (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);  /**< `workSpace` must be a table of at least HUF_WORKSPACE_SIZE_U32 unsigned */
+size_t HUF_compress1X_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.
  *  If it doesn't, it sets *repeat = HUF_repeat_none, and it sets hufTable to the table used.
- *  If preferRepeat then the old table will always be used if valid. */
+ *  If preferRepeat then the old table will always be used if valid.
+ *  If suspectUncompressible then some sampling checks will be run to potentially skip huffman coding */
 size_t HUF_compress1X_repeat(void* dst, size_t dstSize,
                        const void* src, size_t srcSize,
                        unsigned maxSymbolValue, unsigned tableLog,
                        void* workSpace, size_t wkspSize,   /**< `workSpace` must be aligned on 4-bytes boundaries, `wkspSize` must be >= HUF_WORKSPACE_SIZE */
-                       HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2);
+                       HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2, unsigned suspectUncompressible);
 
 size_t HUF_decompress1X1 (void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize);   /* single-symbol decoder */
 #ifndef HUF_FORCE_DECOMPRESS_X1
@@ -354,6 +353,9 @@ size_t HUF_decompress4X_hufOnly_wksp_bmi2(HUF_DTable* dctx, void* dst, size_t ds
 #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);
 #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);
+#endif
 
 #endif /* HUF_STATIC_LINKING_ONLY */
 

zstd/common/mem.h

@@ -51,6 +51,8 @@ extern "C" {
 #    include <stdint.h> /* intptr_t */
 #  endif
   typedef   uint8_t BYTE;
+  typedef   uint8_t U8;
+  typedef    int8_t S8;
   typedef  uint16_t U16;
   typedef   int16_t S16;
   typedef  uint32_t U32;
@@ -63,6 +65,8 @@ extern "C" {
 #  error "this implementation requires char to be exactly 8-bit type"
 #endif
   typedef unsigned char      BYTE;
+  typedef unsigned char      U8;
+  typedef   signed char      S8;
 #if USHRT_MAX != 65535
 #  error "this implementation requires short to be exactly 16-bit type"
 #endif
@@ -153,8 +157,22 @@ MEM_STATIC unsigned MEM_64bits(void) { return sizeof(size_t)==8; }
 
 MEM_STATIC unsigned MEM_isLittleEndian(void)
 {
+#if defined(__BYTE_ORDER__) && defined(__ORDER_LITTLE_ENDIAN__) && (__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
+    return 1;
+#elif defined(__BYTE_ORDER__) && defined(__ORDER_BIG_ENDIAN__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+    return 0;
+#elif defined(__clang__) && __LITTLE_ENDIAN__
+    return 1;
+#elif defined(__clang__) && __BIG_ENDIAN__
+    return 0;
+#elif defined(_MSC_VER) && (_M_AMD64 || _M_IX86)
+    return 1;
+#elif defined(__DMC__) && defined(_M_IX86)
+    return 1;
+#else
     const union { U32 u; BYTE c[4]; } one = { 1 };   /* don't use static : performance detrimental  */
     return one.c[0];
+#endif
 }
 
 #if defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==2)

zstd/common/portability_macros.h

@@ -0,0 +1,131 @@
+/*
+ * Copyright (c) Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under both the BSD-style license (found in the
+ * LICENSE file in the root directory of this source tree) and the GPLv2 (found
+ * in the COPYING file in the root directory of this source tree).
+ * You may select, at your option, one of the above-listed licenses.
+ */
+
+#ifndef ZSTD_PORTABILITY_MACROS_H
+#define ZSTD_PORTABILITY_MACROS_H
+
+/**
+ * This header file contains macro defintions to support portability.
+ * This header is shared between C and ASM code, so it MUST only
+ * contain macro definitions. It MUST not contain any C code.
+ *
+ * This header ONLY defines macros to detect platforms/feature support.
+ *
+ */
+
+
+/* compat. with non-clang compilers */
+#ifndef __has_attribute
+  #define __has_attribute(x) 0
+#endif
+
+/* compat. with non-clang compilers */
+#ifndef __has_builtin
+#  define __has_builtin(x) 0
+#endif
+
+/* compat. with non-clang compilers */
+#ifndef __has_feature
+#  define __has_feature(x) 0
+#endif
+
+/* detects whether we are being compiled under msan */
+#ifndef ZSTD_MEMORY_SANITIZER
+#  if __has_feature(memory_sanitizer)
+#    define ZSTD_MEMORY_SANITIZER 1
+#  else
+#    define ZSTD_MEMORY_SANITIZER 0
+#  endif
+#endif
+
+/* detects whether we are being compiled under asan */
+#ifndef ZSTD_ADDRESS_SANITIZER
+#  if __has_feature(address_sanitizer)
+#    define ZSTD_ADDRESS_SANITIZER 1
+#  elif defined(__SANITIZE_ADDRESS__)
+#    define ZSTD_ADDRESS_SANITIZER 1
+#  else
+#    define ZSTD_ADDRESS_SANITIZER 0
+#  endif
+#endif
+
+/* detects whether we are being compiled under dfsan */
+#ifndef ZSTD_DATAFLOW_SANITIZER
+# if __has_feature(dataflow_sanitizer)
+#  define ZSTD_DATAFLOW_SANITIZER 1
+# else
+#  define ZSTD_DATAFLOW_SANITIZER 0
+# endif
+#endif
+
+
+/* Enable runtime BMI2 dispatch based on the CPU.
+ * Enabled for clang & gcc >=4.8 on x86 when BMI2 isn't enabled by default.
+ */
+#ifndef DYNAMIC_BMI2
+  #if ((defined(__clang__) && __has_attribute(__target__)) \
+      || (defined(__GNUC__) \
+          && (__GNUC__ >= 5 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8)))) \
+      && (defined(__x86_64__) || defined(_M_X64)) \
+      && !defined(__BMI2__)
+  #  define DYNAMIC_BMI2 1
+  #else
+  #  define DYNAMIC_BMI2 0
+  #endif
+#endif
+
+/**
+ * Only enable assembly for GNUC comptabile compilers,
+ * because other platforms may not support GAS assembly syntax.
+ *
+ * Only enable assembly for Linux / MacOS, other platforms may
+ * work, but they haven't been tested. This could likely be
+ * extended to BSD systems.
+ *
+ * Disable assembly when MSAN is enabled, because MSAN requires
+ * 100% of code to be instrumented to work.
+ */
+#if defined(__GNUC__)
+#  if defined(__linux__) || defined(__linux) || defined(__APPLE__)
+#    if ZSTD_MEMORY_SANITIZER
+#      define ZSTD_ASM_SUPPORTED 0
+#    elif ZSTD_DATAFLOW_SANITIZER
+#      define ZSTD_ASM_SUPPORTED 0
+#    else
+#      define ZSTD_ASM_SUPPORTED 1
+#    endif
+#  else
+#    define ZSTD_ASM_SUPPORTED 0
+#  endif
+#else
+#  define ZSTD_ASM_SUPPORTED 0
+#endif
+
+/**
+ * Determines whether we should enable assembly for x86-64
+ * with BMI2.
+ *
+ * Enable if all of the following conditions hold:
+ * - ASM hasn't been explicitly disabled by defining ZSTD_DISABLE_ASM
+ * - Assembly is supported
+ * - We are compiling for x86-64 and either:
+ *   - DYNAMIC_BMI2 is enabled
+ *   - BMI2 is supported at compile time
+ */
+#if !defined(ZSTD_DISABLE_ASM) &&                                 \
+    ZSTD_ASM_SUPPORTED &&                                         \
+    defined(__x86_64__) &&                                        \
+    (DYNAMIC_BMI2 || defined(__BMI2__))
+# define ZSTD_ENABLE_ASM_X86_64_BMI2 1
+#else
+# define ZSTD_ENABLE_ASM_X86_64_BMI2 0
+#endif
+
+#endif /* ZSTD_PORTABILITY_MACROS_H */

zstd/common/xxhash.c

@@ -13,812 +13,12 @@
 */
 
 
-/* *************************************
+
-*  Tuning parameters
+/*
-***************************************/
+ * xxhash.c instantiates functions defined in xxhash.h
-/*!XXH_FORCE_MEMORY_ACCESS :
- * By default, access to unaligned memory is controlled by `memcpy()`, which is safe and portable.
- * Unfortunately, on some target/compiler combinations, the generated assembly is sub-optimal.
- * The below switch allow to select different access method for improved performance.
- * Method 0 (default) : use `memcpy()`. Safe and portable.
- * Method 1 : `__packed` statement. It depends on compiler extension (ie, not portable).
- *            This method is safe if your compiler supports it, and *generally* as fast or faster than `memcpy`.
- * Method 2 : direct access. This method doesn't depend on compiler but violate C standard.
- *            It can generate buggy code on targets which do not support unaligned memory accesses.
- *            But in some circumstances, it's the only known way to get the most performance (ie GCC + ARMv6)
- * See http://stackoverflow.com/a/32095106/646947 for details.
- * Prefer these methods in priority order (0 > 1 > 2)
  */
-#ifndef XXH_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
-#  if (defined(__INTEL_COMPILER) && !defined(WIN32)) || \
-  (defined(__GNUC__) && ( defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) || defined(__ARM_ARCH_7S__) )) || \
-  defined(__ICCARM__)
-#    define XXH_FORCE_MEMORY_ACCESS 1
-#  endif
-#endif
 
-/*!XXH_ACCEPT_NULL_INPUT_POINTER :
+#define XXH_STATIC_LINKING_ONLY   /* access advanced declarations */
- * If the input pointer is a null pointer, xxHash default behavior is to trigger a memory access error, since it is a bad pointer.
+#define XXH_IMPLEMENTATION   /* access definitions */
- * When this option is enabled, xxHash output for null input pointers will be the same as a null-length input.
- * By default, this option is disabled. To enable it, uncomment below define :
- */
-/* #define XXH_ACCEPT_NULL_INPUT_POINTER 1 */
 
-/*!XXH_FORCE_NATIVE_FORMAT :
- * By default, xxHash library provides endian-independent Hash values, based on little-endian convention.
- * Results are therefore identical for little-endian and big-endian CPU.
- * This comes at a performance cost for big-endian CPU, since some swapping is required to emulate little-endian format.
- * Should endian-independence be of no importance for your application, you may set the #define below to 1,
- * to improve speed for Big-endian CPU.
- * This option has no impact on Little_Endian CPU.
- */
-#ifndef XXH_FORCE_NATIVE_FORMAT   /* can be defined externally */
-#  define XXH_FORCE_NATIVE_FORMAT 0
-#endif
-
-/*!XXH_FORCE_ALIGN_CHECK :
- * This is a minor performance trick, only useful with lots of very small keys.
- * It means : check for aligned/unaligned input.
- * The check costs one initial branch per hash; set to 0 when the input data
- * is guaranteed to be aligned.
- */
-#ifndef XXH_FORCE_ALIGN_CHECK /* can be defined externally */
-#  if defined(__i386) || defined(_M_IX86) || defined(__x86_64__) || defined(_M_X64)
-#    define XXH_FORCE_ALIGN_CHECK 0
-#  else
-#    define XXH_FORCE_ALIGN_CHECK 1
-#  endif
-#endif
-
-
-/* *************************************
-*  Includes & Memory related functions
-***************************************/
-/* Modify the local functions below should you wish to use some other memory routines */
-/* for ZSTD_malloc(), ZSTD_free() */
-#define ZSTD_DEPS_NEED_MALLOC
-#include "zstd_deps.h"  /* size_t, ZSTD_malloc, ZSTD_free, ZSTD_memcpy */
-static void* XXH_malloc(size_t s) { return ZSTD_malloc(s); }
-static void  XXH_free  (void* p)  { ZSTD_free(p); }
-static void* XXH_memcpy(void* dest, const void* src, size_t size) { return ZSTD_memcpy(dest,src,size); }
-
-#ifndef XXH_STATIC_LINKING_ONLY
-#  define XXH_STATIC_LINKING_ONLY
-#endif
 #include "xxhash.h"
-
-
-/* *************************************
-*  Compiler Specific Options
-***************************************/
-#include "compiler.h"
-
-
-/* *************************************
-*  Basic Types
-***************************************/
-#include "mem.h"  /* BYTE, U32, U64, size_t */
-
-#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
-
-/* Force direct memory access. Only works on CPU which support unaligned memory access in hardware */
-static U32 XXH_read32(const void* memPtr) { return *(const U32*) memPtr; }
-static U64 XXH_read64(const void* memPtr) { return *(const U64*) memPtr; }
-
-#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
-
-/* __pack instructions are safer, but compiler specific, hence potentially problematic for some compilers */
-/* currently only defined for gcc and icc */
-typedef union { U32 u32; U64 u64; } __attribute__((packed)) unalign;
-
-static U32 XXH_read32(const void* ptr) { return ((const unalign*)ptr)->u32; }
-static U64 XXH_read64(const void* ptr) { return ((const unalign*)ptr)->u64; }
-
-#else
-
-/* portable and safe solution. Generally efficient.
- * see : http://stackoverflow.com/a/32095106/646947
- */
-
-static U32 XXH_read32(const void* memPtr)
-{
-    U32 val;
-    ZSTD_memcpy(&val, memPtr, sizeof(val));
-    return val;
-}
-
-static U64 XXH_read64(const void* memPtr)
-{
-    U64 val;
-    ZSTD_memcpy(&val, memPtr, sizeof(val));
-    return val;
-}
-
-#endif   /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
-
-
-/* ****************************************
-*  Compiler-specific Functions and Macros
-******************************************/
-#define GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
-
-/* Note : although _rotl exists for minGW (GCC under windows), performance seems poor */
-#if defined(_MSC_VER)
-#  define XXH_rotl32(x,r) _rotl(x,r)
-#  define XXH_rotl64(x,r) _rotl64(x,r)
-#else
-#if defined(__ICCARM__)
-#  include <intrinsics.h>
-#  define XXH_rotl32(x,r) __ROR(x,(32 - r))
-#else
-#  define XXH_rotl32(x,r) ((x << r) | (x >> (32 - r)))
-#endif
-#  define XXH_rotl64(x,r) ((x << r) | (x >> (64 - r)))
-#endif
-
-#if defined(_MSC_VER)     /* Visual Studio */
-#  define XXH_swap32 _byteswap_ulong
-#  define XXH_swap64 _byteswap_uint64
-#elif GCC_VERSION >= 403
-#  define XXH_swap32 __builtin_bswap32
-#  define XXH_swap64 __builtin_bswap64
-#else
-static U32 XXH_swap32 (U32 x)
-{
-    return  ((x << 24) & 0xff000000 ) |
-            ((x <<  8) & 0x00ff0000 ) |
-            ((x >>  8) & 0x0000ff00 ) |
-            ((x >> 24) & 0x000000ff );
-}
-static U64 XXH_swap64 (U64 x)
-{
-    return  ((x << 56) & 0xff00000000000000ULL) |
-            ((x << 40) & 0x00ff000000000000ULL) |
-            ((x << 24) & 0x0000ff0000000000ULL) |
-            ((x << 8)  & 0x000000ff00000000ULL) |
-            ((x >> 8)  & 0x00000000ff000000ULL) |
-            ((x >> 24) & 0x0000000000ff0000ULL) |
-            ((x >> 40) & 0x000000000000ff00ULL) |
-            ((x >> 56) & 0x00000000000000ffULL);
-}
-#endif
-
-
-/* *************************************
-*  Architecture Macros
-***************************************/
-typedef enum { XXH_bigEndian=0, XXH_littleEndian=1 } XXH_endianess;
-
-/* XXH_CPU_LITTLE_ENDIAN can be defined externally, for example on the compiler command line */
-#ifndef XXH_CPU_LITTLE_ENDIAN
-    static const int g_one = 1;
-#   define XXH_CPU_LITTLE_ENDIAN   (*(const char*)(&g_one))
-#endif
-
-
-/* ***************************
-*  Memory reads
-*****************************/
-typedef enum { XXH_aligned, XXH_unaligned } XXH_alignment;
-
-FORCE_INLINE_TEMPLATE U32 XXH_readLE32_align(const void* ptr, XXH_endianess endian, XXH_alignment align)
-{
-    if (align==XXH_unaligned)
-        return endian==XXH_littleEndian ? XXH_read32(ptr) : XXH_swap32(XXH_read32(ptr));
-    else
-        return endian==XXH_littleEndian ? *(const U32*)ptr : XXH_swap32(*(const U32*)ptr);
-}
-
-FORCE_INLINE_TEMPLATE U32 XXH_readLE32(const void* ptr, XXH_endianess endian)
-{
-    return XXH_readLE32_align(ptr, endian, XXH_unaligned);
-}
-
-static U32 XXH_readBE32(const void* ptr)
-{
-    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap32(XXH_read32(ptr)) : XXH_read32(ptr);
-}
-
-FORCE_INLINE_TEMPLATE U64 XXH_readLE64_align(const void* ptr, XXH_endianess endian, XXH_alignment align)
-{
-    if (align==XXH_unaligned)
-        return endian==XXH_littleEndian ? XXH_read64(ptr) : XXH_swap64(XXH_read64(ptr));
-    else
-        return endian==XXH_littleEndian ? *(const U64*)ptr : XXH_swap64(*(const U64*)ptr);
-}
-
-FORCE_INLINE_TEMPLATE U64 XXH_readLE64(const void* ptr, XXH_endianess endian)
-{
-    return XXH_readLE64_align(ptr, endian, XXH_unaligned);
-}
-
-static U64 XXH_readBE64(const void* ptr)
-{
-    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap64(XXH_read64(ptr)) : XXH_read64(ptr);
-}
-
-
-/* *************************************
-*  Macros
-***************************************/
-#define XXH_STATIC_ASSERT(c)   { enum { XXH_static_assert = 1/(int)(!!(c)) }; }    /* use only *after* variable declarations */
-
-
-/* *************************************
-*  Constants
-***************************************/
-static const U32 PRIME32_1 = 2654435761U;
-static const U32 PRIME32_2 = 2246822519U;
-static const U32 PRIME32_3 = 3266489917U;
-static const U32 PRIME32_4 =  668265263U;
-static const U32 PRIME32_5 =  374761393U;
-
-static const U64 PRIME64_1 = 11400714785074694791ULL;
-static const U64 PRIME64_2 = 14029467366897019727ULL;
-static const U64 PRIME64_3 =  1609587929392839161ULL;
-static const U64 PRIME64_4 =  9650029242287828579ULL;
-static const U64 PRIME64_5 =  2870177450012600261ULL;
-
-XXH_PUBLIC_API unsigned XXH_versionNumber (void) { return XXH_VERSION_NUMBER; }
-
-
-/* **************************
-*  Utils
-****************************/
-XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* restrict dstState, const XXH32_state_t* restrict srcState)
-{
-    ZSTD_memcpy(dstState, srcState, sizeof(*dstState));
-}
-
-XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t* restrict dstState, const XXH64_state_t* restrict srcState)
-{
-    ZSTD_memcpy(dstState, srcState, sizeof(*dstState));
-}
-
-
-/* ***************************
-*  Simple Hash Functions
-*****************************/
-
-static U32 XXH32_round(U32 seed, U32 input)
-{
-    seed += input * PRIME32_2;
-    seed  = XXH_rotl32(seed, 13);
-    seed *= PRIME32_1;
-    return seed;
-}
-
-FORCE_INLINE_TEMPLATE U32 XXH32_endian_align(const void* input, size_t len, U32 seed, XXH_endianess endian, XXH_alignment align)
-{
-    const BYTE* p = (const BYTE*)input;
-    const BYTE* bEnd = p + len;
-    U32 h32;
-#define XXH_get32bits(p) XXH_readLE32_align(p, endian, align)
-
-#ifdef XXH_ACCEPT_NULL_INPUT_POINTER
-    if (p==NULL) {
-        len=0;
-        bEnd=p=(const BYTE*)(size_t)16;
-    }
-#endif
-
-    if (len>=16) {
-        const BYTE* const limit = bEnd - 16;
-        U32 v1 = seed + PRIME32_1 + PRIME32_2;
-        U32 v2 = seed + PRIME32_2;
-        U32 v3 = seed + 0;
-        U32 v4 = seed - PRIME32_1;
-
-        do {
-            v1 = XXH32_round(v1, XXH_get32bits(p)); p+=4;
-            v2 = XXH32_round(v2, XXH_get32bits(p)); p+=4;
-            v3 = XXH32_round(v3, XXH_get32bits(p)); p+=4;
-            v4 = XXH32_round(v4, XXH_get32bits(p)); p+=4;
-        } while (p<=limit);
-
-        h32 = XXH_rotl32(v1, 1) + XXH_rotl32(v2, 7) + XXH_rotl32(v3, 12) + XXH_rotl32(v4, 18);
-    } else {
-        h32  = seed + PRIME32_5;
-    }
-
-    h32 += (U32) len;
-
-    while (p+4<=bEnd) {
-        h32 += XXH_get32bits(p) * PRIME32_3;
-        h32  = XXH_rotl32(h32, 17) * PRIME32_4 ;
-        p+=4;
-    }
-
-    while (p<bEnd) {
-        h32 += (*p) * PRIME32_5;
-        h32 = XXH_rotl32(h32, 11) * PRIME32_1 ;
-        p++;
-    }
-
-    h32 ^= h32 >> 15;
-    h32 *= PRIME32_2;
-    h32 ^= h32 >> 13;
-    h32 *= PRIME32_3;
-    h32 ^= h32 >> 16;
-
-    return h32;
-}
-
-
-XXH_PUBLIC_API unsigned int XXH32 (const void* input, size_t len, unsigned int seed)
-{
-#if 0
-    /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
-    XXH32_CREATESTATE_STATIC(state);
-    XXH32_reset(state, seed);
-    XXH32_update(state, input, len);
-    return XXH32_digest(state);
-#else
-    XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN;
-
-    if (XXH_FORCE_ALIGN_CHECK) {
-        if ((((size_t)input) & 3) == 0) {   /* Input is 4-bytes aligned, leverage the speed benefit */
-            if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
-                return XXH32_endian_align(input, len, seed, XXH_littleEndian, XXH_aligned);
-            else
-                return XXH32_endian_align(input, len, seed, XXH_bigEndian, XXH_aligned);
-    }   }
-
-    if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
-        return XXH32_endian_align(input, len, seed, XXH_littleEndian, XXH_unaligned);
-    else
-        return XXH32_endian_align(input, len, seed, XXH_bigEndian, XXH_unaligned);
-#endif
-}
-
-
-static U64 XXH64_round(U64 acc, U64 input)
-{
-    acc += input * PRIME64_2;
-    acc  = XXH_rotl64(acc, 31);
-    acc *= PRIME64_1;
-    return acc;
-}
-
-static U64 XXH64_mergeRound(U64 acc, U64 val)
-{
-    val  = XXH64_round(0, val);
-    acc ^= val;
-    acc  = acc * PRIME64_1 + PRIME64_4;
-    return acc;
-}
-
-FORCE_INLINE_TEMPLATE U64 XXH64_endian_align(const void* input, size_t len, U64 seed, XXH_endianess endian, XXH_alignment align)
-{
-    const BYTE* p = (const BYTE*)input;
-    const BYTE* const bEnd = p + len;
-    U64 h64;
-#define XXH_get64bits(p) XXH_readLE64_align(p, endian, align)
-
-#ifdef XXH_ACCEPT_NULL_INPUT_POINTER
-    if (p==NULL) {
-        len=0;
-        bEnd=p=(const BYTE*)(size_t)32;
-    }
-#endif
-
-    if (len>=32) {
-        const BYTE* const limit = bEnd - 32;
-        U64 v1 = seed + PRIME64_1 + PRIME64_2;
-        U64 v2 = seed + PRIME64_2;
-        U64 v3 = seed + 0;
-        U64 v4 = seed - PRIME64_1;
-
-        do {
-            v1 = XXH64_round(v1, XXH_get64bits(p)); p+=8;
-            v2 = XXH64_round(v2, XXH_get64bits(p)); p+=8;
-            v3 = XXH64_round(v3, XXH_get64bits(p)); p+=8;
-            v4 = XXH64_round(v4, XXH_get64bits(p)); p+=8;
-        } while (p<=limit);
-
-        h64 = XXH_rotl64(v1, 1) + XXH_rotl64(v2, 7) + XXH_rotl64(v3, 12) + XXH_rotl64(v4, 18);
-        h64 = XXH64_mergeRound(h64, v1);
-        h64 = XXH64_mergeRound(h64, v2);
-        h64 = XXH64_mergeRound(h64, v3);
-        h64 = XXH64_mergeRound(h64, v4);
-
-    } else {
-        h64  = seed + PRIME64_5;
-    }
-
-    h64 += (U64) len;
-
-    while (p+8<=bEnd) {
-        U64 const k1 = XXH64_round(0, XXH_get64bits(p));
-        h64 ^= k1;
-        h64  = XXH_rotl64(h64,27) * PRIME64_1 + PRIME64_4;
-        p+=8;
-    }
-
-    if (p+4<=bEnd) {
-        h64 ^= (U64)(XXH_get32bits(p)) * PRIME64_1;
-        h64 = XXH_rotl64(h64, 23) * PRIME64_2 + PRIME64_3;
-        p+=4;
-    }
-
-    while (p<bEnd) {
-        h64 ^= (*p) * PRIME64_5;
-        h64 = XXH_rotl64(h64, 11) * PRIME64_1;
-        p++;
-    }
-
-    h64 ^= h64 >> 33;
-    h64 *= PRIME64_2;
-    h64 ^= h64 >> 29;
-    h64 *= PRIME64_3;
-    h64 ^= h64 >> 32;
-
-    return h64;
-}
-
-
-XXH_PUBLIC_API unsigned long long XXH64 (const void* input, size_t len, unsigned long long seed)
-{
-#if 0
-    /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
-    XXH64_CREATESTATE_STATIC(state);
-    XXH64_reset(state, seed);
-    XXH64_update(state, input, len);
-    return XXH64_digest(state);
-#else
-    XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN;
-
-    if (XXH_FORCE_ALIGN_CHECK) {
-        if ((((size_t)input) & 7)==0) {  /* Input is aligned, let's leverage the speed advantage */
-            if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
-                return XXH64_endian_align(input, len, seed, XXH_littleEndian, XXH_aligned);
-            else
-                return XXH64_endian_align(input, len, seed, XXH_bigEndian, XXH_aligned);
-    }   }
-
-    if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
-        return XXH64_endian_align(input, len, seed, XXH_littleEndian, XXH_unaligned);
-    else
-        return XXH64_endian_align(input, len, seed, XXH_bigEndian, XXH_unaligned);
-#endif
-}
-
-
-/* **************************************************
-*  Advanced Hash Functions
-****************************************************/
-
-XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void)
-{
-    return (XXH32_state_t*)XXH_malloc(sizeof(XXH32_state_t));
-}
-XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr)
-{
-    XXH_free(statePtr);
-    return XXH_OK;
-}
-
-XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void)
-{
-    return (XXH64_state_t*)XXH_malloc(sizeof(XXH64_state_t));
-}
-XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr)
-{
-    XXH_free(statePtr);
-    return XXH_OK;
-}
-
-
-/*** Hash feed ***/
-
-XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, unsigned int seed)
-{
-    XXH32_state_t state;   /* using a local state to memcpy() in order to avoid strict-aliasing warnings */
-    ZSTD_memset(&state, 0, sizeof(state)-4);   /* do not write into reserved, for future removal */
-    state.v1 = seed + PRIME32_1 + PRIME32_2;
-    state.v2 = seed + PRIME32_2;
-    state.v3 = seed + 0;
-    state.v4 = seed - PRIME32_1;
-    ZSTD_memcpy(statePtr, &state, sizeof(state));
-    return XXH_OK;
-}
-
-
-XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH64_state_t* statePtr, unsigned long long seed)
-{
-    XXH64_state_t state;   /* using a local state to memcpy() in order to avoid strict-aliasing warnings */
-    ZSTD_memset(&state, 0, sizeof(state)-8);   /* do not write into reserved, for future removal */
-    state.v1 = seed + PRIME64_1 + PRIME64_2;
-    state.v2 = seed + PRIME64_2;
-    state.v3 = seed + 0;
-    state.v4 = seed - PRIME64_1;
-    ZSTD_memcpy(statePtr, &state, sizeof(state));
-    return XXH_OK;
-}
-
-
-FORCE_INLINE_TEMPLATE XXH_errorcode XXH32_update_endian (XXH32_state_t* state, const void* input, size_t len, XXH_endianess endian)
-{
-    const BYTE* p = (const BYTE*)input;
-    const BYTE* const bEnd = p + len;
-
-#ifdef XXH_ACCEPT_NULL_INPUT_POINTER
-    if (input==NULL) return XXH_ERROR;
-#endif
-
-    state->total_len_32 += (unsigned)len;
-    state->large_len |= (len>=16) | (state->total_len_32>=16);
-
-    if (state->memsize + len < 16)  {   /* fill in tmp buffer */
-        XXH_memcpy((BYTE*)(state->mem32) + state->memsize, input, len);
-        state->memsize += (unsigned)len;
-        return XXH_OK;
-    }
-
-    if (state->memsize) {   /* some data left from previous update */
-        XXH_memcpy((BYTE*)(state->mem32) + state->memsize, input, 16-state->memsize);
-        {   const U32* p32 = state->mem32;
-            state->v1 = XXH32_round(state->v1, XXH_readLE32(p32, endian)); p32++;
-            state->v2 = XXH32_round(state->v2, XXH_readLE32(p32, endian)); p32++;
-            state->v3 = XXH32_round(state->v3, XXH_readLE32(p32, endian)); p32++;
-            state->v4 = XXH32_round(state->v4, XXH_readLE32(p32, endian)); p32++;
-        }
-        p += 16-state->memsize;
-        state->memsize = 0;
-    }
-
-    if (p <= bEnd-16) {
-        const BYTE* const limit = bEnd - 16;
-        U32 v1 = state->v1;
-        U32 v2 = state->v2;
-        U32 v3 = state->v3;
-        U32 v4 = state->v4;
-
-        do {
-            v1 = XXH32_round(v1, XXH_readLE32(p, endian)); p+=4;
-            v2 = XXH32_round(v2, XXH_readLE32(p, endian)); p+=4;
-            v3 = XXH32_round(v3, XXH_readLE32(p, endian)); p+=4;
-            v4 = XXH32_round(v4, XXH_readLE32(p, endian)); p+=4;
-        } while (p<=limit);
-
-        state->v1 = v1;
-        state->v2 = v2;
-        state->v3 = v3;
-        state->v4 = v4;
-    }
-
-    if (p < bEnd) {
-        XXH_memcpy(state->mem32, p, (size_t)(bEnd-p));
-        state->memsize = (unsigned)(bEnd-p);
-    }
-
-    return XXH_OK;
-}
-
-XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* state_in, const void* input, size_t len)
-{
-    XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN;
-
-    if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
-        return XXH32_update_endian(state_in, input, len, XXH_littleEndian);
-    else
-        return XXH32_update_endian(state_in, input, len, XXH_bigEndian);
-}
-
-
-
-FORCE_INLINE_TEMPLATE U32 XXH32_digest_endian (const XXH32_state_t* state, XXH_endianess endian)
-{
-    const BYTE * p = (const BYTE*)state->mem32;
-    const BYTE* const bEnd = (const BYTE*)(state->mem32) + state->memsize;
-    U32 h32;
-
-    if (state->large_len) {
-        h32 = XXH_rotl32(state->v1, 1) + XXH_rotl32(state->v2, 7) + XXH_rotl32(state->v3, 12) + XXH_rotl32(state->v4, 18);
-    } else {
-        h32 = state->v3 /* == seed */ + PRIME32_5;
-    }
-
-    h32 += state->total_len_32;
-
-    while (p+4<=bEnd) {
-        h32 += XXH_readLE32(p, endian) * PRIME32_3;
-        h32  = XXH_rotl32(h32, 17) * PRIME32_4;
-        p+=4;
-    }
-
-    while (p<bEnd) {
-        h32 += (*p) * PRIME32_5;
-        h32  = XXH_rotl32(h32, 11) * PRIME32_1;
-        p++;
-    }
-
-    h32 ^= h32 >> 15;
-    h32 *= PRIME32_2;
-    h32 ^= h32 >> 13;
-    h32 *= PRIME32_3;
-    h32 ^= h32 >> 16;
-
-    return h32;
-}
-
-
-XXH_PUBLIC_API unsigned int XXH32_digest (const XXH32_state_t* state_in)
-{
-    XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN;
-
-    if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
-        return XXH32_digest_endian(state_in, XXH_littleEndian);
-    else
-        return XXH32_digest_endian(state_in, XXH_bigEndian);
-}
-
-
-
-/* **** XXH64 **** */
-
-FORCE_INLINE_TEMPLATE XXH_errorcode XXH64_update_endian (XXH64_state_t* state, const void* input, size_t len, XXH_endianess endian)
-{
-    const BYTE* p = (const BYTE*)input;
-    const BYTE* const bEnd = p + len;
-
-#ifdef XXH_ACCEPT_NULL_INPUT_POINTER
-    if (input==NULL) return XXH_ERROR;
-#endif
-
-    state->total_len += len;
-
-    if (state->memsize + len < 32) {  /* fill in tmp buffer */
-        if (input != NULL) {
-            XXH_memcpy(((BYTE*)state->mem64) + state->memsize, input, len);
-        }
-        state->memsize += (U32)len;
-        return XXH_OK;
-    }
-
-    if (state->memsize) {   /* tmp buffer is full */
-        XXH_memcpy(((BYTE*)state->mem64) + state->memsize, input, 32-state->memsize);
-        state->v1 = XXH64_round(state->v1, XXH_readLE64(state->mem64+0, endian));
-        state->v2 = XXH64_round(state->v2, XXH_readLE64(state->mem64+1, endian));
-        state->v3 = XXH64_round(state->v3, XXH_readLE64(state->mem64+2, endian));
-        state->v4 = XXH64_round(state->v4, XXH_readLE64(state->mem64+3, endian));
-        p += 32-state->memsize;
-        state->memsize = 0;
-    }
-
-    if (p+32 <= bEnd) {
-        const BYTE* const limit = bEnd - 32;
-        U64 v1 = state->v1;
-        U64 v2 = state->v2;
-        U64 v3 = state->v3;
-        U64 v4 = state->v4;
-
-        do {
-            v1 = XXH64_round(v1, XXH_readLE64(p, endian)); p+=8;
-            v2 = XXH64_round(v2, XXH_readLE64(p, endian)); p+=8;
-            v3 = XXH64_round(v3, XXH_readLE64(p, endian)); p+=8;
-            v4 = XXH64_round(v4, XXH_readLE64(p, endian)); p+=8;
-        } while (p<=limit);
-
-        state->v1 = v1;
-        state->v2 = v2;
-        state->v3 = v3;
-        state->v4 = v4;
-    }
-
-    if (p < bEnd) {
-        XXH_memcpy(state->mem64, p, (size_t)(bEnd-p));
-        state->memsize = (unsigned)(bEnd-p);
-    }
-
-    return XXH_OK;
-}
-
-XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH64_state_t* state_in, const void* input, size_t len)
-{
-    XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN;
-
-    if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
-        return XXH64_update_endian(state_in, input, len, XXH_littleEndian);
-    else
-        return XXH64_update_endian(state_in, input, len, XXH_bigEndian);
-}
-
-
-
-FORCE_INLINE_TEMPLATE U64 XXH64_digest_endian (const XXH64_state_t* state, XXH_endianess endian)
-{
-    const BYTE * p = (const BYTE*)state->mem64;
-    const BYTE* const bEnd = (const BYTE*)state->mem64 + state->memsize;
-    U64 h64;
-
-    if (state->total_len >= 32) {
-        U64 const v1 = state->v1;
-        U64 const v2 = state->v2;
-        U64 const v3 = state->v3;
-        U64 const v4 = state->v4;
-
-        h64 = XXH_rotl64(v1, 1) + XXH_rotl64(v2, 7) + XXH_rotl64(v3, 12) + XXH_rotl64(v4, 18);
-        h64 = XXH64_mergeRound(h64, v1);
-        h64 = XXH64_mergeRound(h64, v2);
-        h64 = XXH64_mergeRound(h64, v3);
-        h64 = XXH64_mergeRound(h64, v4);
-    } else {
-        h64  = state->v3 + PRIME64_5;
-    }
-
-    h64 += (U64) state->total_len;
-
-    while (p+8<=bEnd) {
-        U64 const k1 = XXH64_round(0, XXH_readLE64(p, endian));
-        h64 ^= k1;
-        h64  = XXH_rotl64(h64,27) * PRIME64_1 + PRIME64_4;
-        p+=8;
-    }
-
-    if (p+4<=bEnd) {
-        h64 ^= (U64)(XXH_readLE32(p, endian)) * PRIME64_1;
-        h64  = XXH_rotl64(h64, 23) * PRIME64_2 + PRIME64_3;
-        p+=4;
-    }
-
-    while (p<bEnd) {
-        h64 ^= (*p) * PRIME64_5;
-        h64  = XXH_rotl64(h64, 11) * PRIME64_1;
-        p++;
-    }
-
-    h64 ^= h64 >> 33;
-    h64 *= PRIME64_2;
-    h64 ^= h64 >> 29;
-    h64 *= PRIME64_3;
-    h64 ^= h64 >> 32;
-
-    return h64;
-}
-
-
-XXH_PUBLIC_API unsigned long long XXH64_digest (const XXH64_state_t* state_in)
-{
-    XXH_endianess endian_detected = (XXH_endianess)XXH_CPU_LITTLE_ENDIAN;
-
-    if ((endian_detected==XXH_littleEndian) || XXH_FORCE_NATIVE_FORMAT)
-        return XXH64_digest_endian(state_in, XXH_littleEndian);
-    else
-        return XXH64_digest_endian(state_in, XXH_bigEndian);
-}
-
-
-/* **************************
-*  Canonical representation
-****************************/
-
-/*! Default XXH result types are basic unsigned 32 and 64 bits.
-*   The canonical representation follows human-readable write convention, aka big-endian (large digits first).
-*   These functions allow transformation of hash result into and from its canonical format.
-*   This way, hash values can be written into a file or buffer, and remain comparable across different systems and programs.
-*/
-
-XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash)
-{
-    XXH_STATIC_ASSERT(sizeof(XXH32_canonical_t) == sizeof(XXH32_hash_t));
-    if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap32(hash);
-    ZSTD_memcpy(dst, &hash, sizeof(*dst));
-}
-
-XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t* dst, XXH64_hash_t hash)
-{
-    XXH_STATIC_ASSERT(sizeof(XXH64_canonical_t) == sizeof(XXH64_hash_t));
-    if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap64(hash);
-    ZSTD_memcpy(dst, &hash, sizeof(*dst));
-}
-
-XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src)
-{
-    return XXH_readBE32(src);
-}
-
-XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src)
-{
-    return XXH_readBE64(src);
-}

zstd/common/zstd_internal.h

@@ -19,10 +19,8 @@
 /*-*************************************
 *  Dependencies
 ***************************************/
-#if !defined(ZSTD_NO_INTRINSICS) && defined(__ARM_NEON)
-#include <arm_neon.h>
-#endif
 #include "compiler.h"
+#include "cpu.h"
 #include "mem.h"
 #include "debug.h"                 /* assert, DEBUGLOG, RAWLOG, g_debuglevel */
 #include "error_private.h"
@@ -60,81 +58,7 @@ extern "C" {
 #undef MAX
 #define MIN(a,b) ((a)<(b) ? (a) : (b))
 #define MAX(a,b) ((a)>(b) ? (a) : (b))
-
+#define BOUNDED(min,val,max) (MAX(min,MIN(val,max)))
-/**
- * Ignore: this is an internal helper.
- *
- * This is a helper function to help force C99-correctness during compilation.
- * Under strict compilation modes, variadic macro arguments can't be empty.
- * However, variadic function arguments can be. Using a function therefore lets
- * us statically check that at least one (string) argument was passed,
- * independent of the compilation flags.
- */
-static INLINE_KEYWORD UNUSED_ATTR
-void _force_has_format_string(const char *format, ...) {
-  (void)format;
-}
-
-/**
- * Ignore: this is an internal helper.
- *
- * We want to force this function invocation to be syntactically correct, but
- * we don't want to force runtime evaluation of its arguments.
- */
-#define _FORCE_HAS_FORMAT_STRING(...) \
-  if (0) { \
-    _force_has_format_string(__VA_ARGS__); \
-  }
-
-/**
- * Return the specified error if the condition evaluates to true.
- *
- * In debug modes, prints additional information.
- * In order to do that (particularly, printing the conditional that failed),
- * this can't just wrap RETURN_ERROR().
- */
-#define RETURN_ERROR_IF(cond, err, ...) \
-  if (cond) { \
-    RAWLOG(3, "%s:%d: ERROR!: check %s failed, returning %s", \
-           __FILE__, __LINE__, ZSTD_QUOTE(cond), ZSTD_QUOTE(ERROR(err))); \
-    _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
-    RAWLOG(3, ": " __VA_ARGS__); \
-    RAWLOG(3, "\n"); \
-    return ERROR(err); \
-  }
-
-/**
- * Unconditionally return the specified error.
- *
- * In debug modes, prints additional information.
- */
-#define RETURN_ERROR(err, ...) \
-  do { \
-    RAWLOG(3, "%s:%d: ERROR!: unconditional check failed, returning %s", \
-           __FILE__, __LINE__, ZSTD_QUOTE(ERROR(err))); \
-    _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
-    RAWLOG(3, ": " __VA_ARGS__); \
-    RAWLOG(3, "\n"); \
-    return ERROR(err); \
-  } while(0);
-
-/**
- * If the provided expression evaluates to an error code, returns that error code.
- *
- * In debug modes, prints additional information.
- */
-#define FORWARD_IF_ERROR(err, ...) \
-  do { \
-    size_t const err_code = (err); \
-    if (ERR_isError(err_code)) { \
-      RAWLOG(3, "%s:%d: ERROR!: forwarding error in %s: %s", \
-             __FILE__, __LINE__, ZSTD_QUOTE(err), ERR_getErrorName(err_code)); \
-      _FORCE_HAS_FORMAT_STRING(__VA_ARGS__); \
-      RAWLOG(3, ": " __VA_ARGS__); \
-      RAWLOG(3, "\n"); \
-      return err_code; \
-    } \
-  } while(0);
 
 
 /*-*************************************
@@ -195,7 +119,7 @@ typedef enum { set_basic, set_rle, set_compressed, set_repeat } symbolEncodingTy
 /* Each table cannot take more than #symbols * FSELog bits */
 #define ZSTD_MAX_FSE_HEADERS_SIZE (((MaxML + 1) * MLFSELog + (MaxLL + 1) * LLFSELog + (MaxOff + 1) * OffFSELog + 7) / 8)
 
-static UNUSED_ATTR const U32 LL_bits[MaxLL+1] = {
+static UNUSED_ATTR const U8 LL_bits[MaxLL+1] = {
      0, 0, 0, 0, 0, 0, 0, 0,
      0, 0, 0, 0, 0, 0, 0, 0,
      1, 1, 1, 1, 2, 2, 3, 3,
@@ -212,7 +136,7 @@ static UNUSED_ATTR const S16 LL_defaultNorm[MaxLL+1] = {
 #define LL_DEFAULTNORMLOG 6  /* for static allocation */
 static UNUSED_ATTR const U32 LL_defaultNormLog = LL_DEFAULTNORMLOG;
 
-static UNUSED_ATTR const U32 ML_bits[MaxML+1] = {
+static UNUSED_ATTR const U8 ML_bits[MaxML+1] = {
      0, 0, 0, 0, 0, 0, 0, 0,
      0, 0, 0, 0, 0, 0, 0, 0,
      0, 0, 0, 0, 0, 0, 0, 0,
@@ -247,19 +171,30 @@ static UNUSED_ATTR const U32 OF_defaultNormLog = OF_DEFAULTNORMLOG;
 *  Shared functions to include for inlining
 *********************************************/
 static void ZSTD_copy8(void* dst, const void* src) {
-#if !defined(ZSTD_NO_INTRINSICS) && defined(__ARM_NEON)
+#if defined(ZSTD_ARCH_ARM_NEON)
     vst1_u8((uint8_t*)dst, vld1_u8((const uint8_t*)src));
 #else
     ZSTD_memcpy(dst, src, 8);
 #endif
 }
-
 #define COPY8(d,s) { ZSTD_copy8(d,s); d+=8; s+=8; }
+
+/* Need to use memmove here since the literal buffer can now be located within
+   the dst buffer. In circumstances where the op "catches up" to where the
+   literal buffer is, there can be partial overlaps in this call on the final
+   copy if the literal is being shifted by less than 16 bytes. */
 static void ZSTD_copy16(void* dst, const void* src) {
-#if !defined(ZSTD_NO_INTRINSICS) && defined(__ARM_NEON)
+#if defined(ZSTD_ARCH_ARM_NEON)
     vst1q_u8((uint8_t*)dst, vld1q_u8((const uint8_t*)src));
+#elif defined(ZSTD_ARCH_X86_SSE2)
+    _mm_storeu_si128((__m128i*)dst, _mm_loadu_si128((const __m128i*)src));
+#elif defined(__clang__)
+    ZSTD_memmove(dst, src, 16);
 #else
-    ZSTD_memcpy(dst, src, 16);
+    /* ZSTD_memmove is not inlined properly by gcc */
+    BYTE copy16_buf[16];
+    ZSTD_memcpy(copy16_buf, src, 16);
+    ZSTD_memcpy(dst, copy16_buf, 16);
 #endif
 }
 #define COPY16(d,s) { ZSTD_copy16(d,s); d+=16; s+=16; }
@@ -288,8 +223,6 @@ void ZSTD_wildcopy(void* dst, const void* src, ptrdiff_t length, ZSTD_overlap_e
     BYTE* op = (BYTE*)dst;
     BYTE* const oend = op + length;
 
-    assert(diff >= 8 || (ovtype == ZSTD_no_overlap && diff <= -WILDCOPY_VECLEN));
-
     if (ovtype == ZSTD_overlap_src_before_dst && diff < WILDCOPY_VECLEN) {
         /* Handle short offset copies. */
         do {
@@ -436,8 +369,14 @@ MEM_STATIC U32 ZSTD_highbit32(U32 val)   /* compress, dictBuilder, decodeCorpus
 #       if STATIC_BMI2 == 1
             return _lzcnt_u32(val)^31;
 #       else
-            unsigned long r=0;
+            if (val != 0) {
-            return _BitScanReverse(&r, val) ? (unsigned)r : 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;
@@ -456,6 +395,63 @@ MEM_STATIC U32 ZSTD_highbit32(U32 val)   /* compress, dictBuilder, decodeCorpus
     }
 }
 
+/**
+ * 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.
@@ -482,6 +478,14 @@ size_t ZSTD_getcBlockSize(const void* src, size_t srcSize,
 size_t ZSTD_decodeSeqHeaders(ZSTD_DCtx* dctx, int* nbSeqPtr,
                        const void* src, size_t srcSize);
 
+/**
+ * @returns true iff the CPU supports dynamic BMI2 dispatch.
+ */
+MEM_STATIC int ZSTD_cpuSupportsBmi2(void)
+{
+    ZSTD_cpuid_t cpuid = ZSTD_cpuid();
+    return ZSTD_cpuid_bmi1(cpuid) && ZSTD_cpuid_bmi2(cpuid);
+}
 
 #if defined (__cplusplus)
 }

zstd/common/zstd_trace.h

@@ -17,10 +17,19 @@ extern "C" {
 
 #include <stddef.h>
 
-/* weak symbol support */
+/* weak symbol support
-#if !defined(ZSTD_HAVE_WEAK_SYMBOLS) && defined(__GNUC__) && \
+ * For now, enable conservatively:
+ * - Only GNUC
+ * - Only ELF
+ * - Only x86-64 and i386
+ * 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(__APPLE__) && !defined(_WIN32) && !defined(__MINGW32__) && \
-    !defined(__CYGWIN__)
+    !defined(__CYGWIN__) && !defined(_AIX)
 #  define ZSTD_HAVE_WEAK_SYMBOLS 1
 #else
 #  define ZSTD_HAVE_WEAK_SYMBOLS 0

zstd/compress/clevels.h

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

zstd/compress/fse_compress.c

@@ -75,13 +75,14 @@ size_t FSE_buildCTable_wksp(FSE_CTable* ct,
     void* const FSCT = ((U32*)ptr) + 1 /* header */ + (tableLog ? tableSize>>1 : 1) ;
     FSE_symbolCompressionTransform* const symbolTT = (FSE_symbolCompressionTransform*) (FSCT);
     U32 const step = FSE_TABLESTEP(tableSize);
+    U32 const maxSV1 = maxSymbolValue+1;
 
-    U32* cumul = (U32*)workSpace;
+    U16* cumul = (U16*)workSpace;   /* size = maxSV1 */
-    FSE_FUNCTION_TYPE* tableSymbol = (FSE_FUNCTION_TYPE*)(cumul + (maxSymbolValue + 2));
+    FSE_FUNCTION_TYPE* const tableSymbol = (FSE_FUNCTION_TYPE*)(cumul + (maxSV1+1));  /* size = tableSize */
 
     U32 highThreshold = tableSize-1;
 
-    if ((size_t)workSpace & 3) return ERROR(GENERIC); /* Must be 4 byte aligned */
+    assert(((size_t)workSpace & 1) == 0);  /* Must be 2 bytes-aligned */
     if (FSE_BUILD_CTABLE_WORKSPACE_SIZE(maxSymbolValue, tableLog) > wkspSize) return ERROR(tableLog_tooLarge);
     /* CTable header */
     tableU16[-2] = (U16) tableLog;
@@ -98,20 +99,61 @@ size_t FSE_buildCTable_wksp(FSE_CTable* ct,
     /* symbol start positions */
     {   U32 u;
         cumul[0] = 0;
-        for (u=1; u <= maxSymbolValue+1; u++) {
+        for (u=1; u <= maxSV1; u++) {
             if (normalizedCounter[u-1]==-1) {  /* Low proba symbol */
                 cumul[u] = cumul[u-1] + 1;
                 tableSymbol[highThreshold--] = (FSE_FUNCTION_TYPE)(u-1);
             } else {
-                cumul[u] = cumul[u-1] + normalizedCounter[u-1];
+                assert(normalizedCounter[u-1] >= 0);
+                cumul[u] = cumul[u-1] + (U16)normalizedCounter[u-1];
+                assert(cumul[u] >= cumul[u-1]);  /* no overflow */
         }   }
-        cumul[maxSymbolValue+1] = tableSize+1;
+        cumul[maxSV1] = (U16)(tableSize+1);
     }
 
     /* Spread symbols */
-    {   U32 position = 0;
+    if (highThreshold == tableSize - 1) {
+        /* Case for no low prob count symbols. Lay down 8 bytes at a time
+         * to reduce branch misses since we are operating on a small block
+         */
+        BYTE* const spread = tableSymbol + tableSize; /* size = tableSize + 8 (may write beyond tableSize) */
+        {   U64 const add = 0x0101010101010101ull;
+            size_t pos = 0;
+            U64 sv = 0;
+            U32 s;
+            for (s=0; s<maxSV1; ++s, sv += add) {
+                int i;
+                int const n = normalizedCounter[s];
+                MEM_write64(spread + pos, sv);
+                for (i = 8; i < n; i += 8) {
+                    MEM_write64(spread + pos + i, sv);
+                }
+                assert(n>=0);
+                pos += (size_t)n;
+            }
+        }
+        /* Spread symbols across the table. Lack of lowprob symbols means that
+         * we don't need variable sized inner loop, so we can unroll the loop and
+         * reduce branch misses.
+         */
+        {   size_t position = 0;
+            size_t s;
+            size_t const unroll = 2; /* Experimentally determined optimal unroll */
+            assert(tableSize % unroll == 0); /* FSE_MIN_TABLELOG is 5 */
+            for (s = 0; s < (size_t)tableSize; s += unroll) {
+                size_t u;
+                for (u = 0; u < unroll; ++u) {
+                    size_t const uPosition = (position + (u * step)) & tableMask;
+                    tableSymbol[uPosition] = spread[s + u];
+                }
+                position = (position + (unroll * step)) & tableMask;
+            }
+            assert(position == 0);   /* Must have initialized all positions */
+        }
+    } else {
+        U32 position = 0;
         U32 symbol;
-        for (symbol=0; symbol<=maxSymbolValue; symbol++) {
+        for (symbol=0; symbol<maxSV1; symbol++) {
             int nbOccurrences;
             int const freq = normalizedCounter[symbol];
             for (nbOccurrences=0; nbOccurrences<freq; nbOccurrences++) {
@@ -120,7 +162,6 @@ size_t FSE_buildCTable_wksp(FSE_CTable* ct,
                 while (position > highThreshold)
                     position = (position + step) & tableMask;   /* Low proba area */
         }   }
-
         assert(position==0);  /* Must have initialized all positions */
     }
 
@@ -144,16 +185,17 @@ size_t FSE_buildCTable_wksp(FSE_CTable* ct,
             case -1:
             case  1:
                 symbolTT[s].deltaNbBits = (tableLog << 16) - (1<<tableLog);
-                symbolTT[s].deltaFindState = total - 1;
+                assert(total <= INT_MAX);
+                symbolTT[s].deltaFindState = (int)(total - 1);
                 total ++;
                 break;
             default :
-                {
+                assert(normalizedCounter[s] > 1);
-                    U32 const maxBitsOut = tableLog - BIT_highbit32 (normalizedCounter[s]-1);
+                {   U32 const maxBitsOut = tableLog - BIT_highbit32 ((U32)normalizedCounter[s]-1);
-                    U32 const minStatePlus = normalizedCounter[s] << maxBitsOut;
+                    U32 const minStatePlus = (U32)normalizedCounter[s] << maxBitsOut;
                     symbolTT[s].deltaNbBits = (maxBitsOut << 16) - minStatePlus;
-                    symbolTT[s].deltaFindState = total - normalizedCounter[s];
+                    symbolTT[s].deltaFindState = (int)(total - (unsigned)normalizedCounter[s]);
-                    total +=  normalizedCounter[s];
+                    total +=  (unsigned)normalizedCounter[s];
     }   }   }   }
 
 #if 0  /* debug : symbol costs */
@@ -164,32 +206,26 @@ size_t FSE_buildCTable_wksp(FSE_CTable* ct,
                 symbol, normalizedCounter[symbol],
                 FSE_getMaxNbBits(symbolTT, symbol),
                 (double)FSE_bitCost(symbolTT, tableLog, symbol, 8) / 256);
-        }
+    }   }
-    }
 #endif
 
     return 0;
 }
 
-#ifndef ZSTD_NO_UNUSED_FUNCTIONS
-size_t FSE_buildCTable(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog)
-{
-    FSE_FUNCTION_TYPE tableSymbol[FSE_MAX_TABLESIZE];   /* memset() is not necessary, even if static analyzer complain about it */
-    return FSE_buildCTable_wksp(ct, normalizedCounter, maxSymbolValue, tableLog, tableSymbol, sizeof(tableSymbol));
-}
-#endif
-
 
 
 #ifndef FSE_COMMONDEFS_ONLY
 
-
 /*-**************************************************************
 *  FSE NCount encoding
 ****************************************************************/
 size_t FSE_NCountWriteBound(unsigned maxSymbolValue, unsigned tableLog)
 {
-    size_t const maxHeaderSize = (((maxSymbolValue+1) * tableLog) >> 3) + 3;
+    size_t const maxHeaderSize = (((maxSymbolValue+1) * tableLog
+                                   + 4 /* bitCount initialized at 4 */
+                                   + 2 /* first two symbols may use one additional bit each */) / 8)
+                                    + 1 /* round up to whole nb bytes */
+                                    + 2 /* additional two bytes for bitstream flush */;
     return maxSymbolValue ? maxHeaderSize : FSE_NCOUNTBOUND;  /* maxSymbolValue==0 ? use default */
 }
 

zstd/compress/huf_compress.c

@@ -53,6 +53,28 @@ unsigned HUF_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxS
 /* *******************************************************
 *  HUF : Huffman block compression
 *********************************************************/
+#define HUF_WORKSPACE_MAX_ALIGNMENT 8
+
+static void* HUF_alignUpWorkspace(void* workspace, size_t* workspaceSizePtr, size_t align)
+{
+    size_t const mask = align - 1;
+    size_t const rem = (size_t)workspace & mask;
+    size_t const add = (align - rem) & mask;
+    BYTE* const aligned = (BYTE*)workspace + add;
+    assert((align & (align - 1)) == 0); /* pow 2 */
+    assert(align <= HUF_WORKSPACE_MAX_ALIGNMENT);
+    if (*workspaceSizePtr >= add) {
+        assert(add < align);
+        assert(((size_t)aligned & mask) == 0);
+        *workspaceSizePtr -= add;
+        return aligned;
+    } else {
+        *workspaceSizePtr = 0;
+        return NULL;
+    }
+}
+
+
 /* HUF_compressWeights() :
  * Same as FSE_compress(), but dedicated to huff0's weights compression.
  * The use case needs much less stack memory.
@@ -75,7 +97,7 @@ static size_t HUF_compressWeights(void* dst, size_t dstSize, const void* weightT
 
     unsigned maxSymbolValue = HUF_TABLELOG_MAX;
     U32 tableLog = MAX_FSE_TABLELOG_FOR_HUFF_HEADER;
-    HUF_CompressWeightsWksp* wksp = (HUF_CompressWeightsWksp*)workspace;
+    HUF_CompressWeightsWksp* wksp = (HUF_CompressWeightsWksp*)HUF_alignUpWorkspace(workspace, &workspaceSize, ZSTD_ALIGNOF(U32));
 
     if (workspaceSize < sizeof(HUF_CompressWeightsWksp)) return ERROR(GENERIC);
 
@@ -106,6 +128,40 @@ static size_t HUF_compressWeights(void* dst, size_t dstSize, const void* weightT
     return (size_t)(op-ostart);
 }
 
+static size_t HUF_getNbBits(HUF_CElt elt)
+{
+    return elt & 0xFF;
+}
+
+static size_t HUF_getNbBitsFast(HUF_CElt elt)
+{
+    return elt;
+}
+
+static size_t HUF_getValue(HUF_CElt elt)
+{
+    return elt & ~0xFF;
+}
+
+static size_t HUF_getValueFast(HUF_CElt elt)
+{
+    return elt;
+}
+
+static void HUF_setNbBits(HUF_CElt* elt, size_t nbBits)
+{
+    assert(nbBits <= HUF_TABLELOG_ABSOLUTEMAX);
+    *elt = nbBits;
+}
+
+static void HUF_setValue(HUF_CElt* elt, size_t value)
+{
+    size_t const nbBits = HUF_getNbBits(*elt);
+    if (nbBits > 0) {
+        assert((value >> nbBits) == 0);
+        *elt |= value << (sizeof(HUF_CElt) * 8 - nbBits);
+    }
+}
 
 typedef struct {
     HUF_CompressWeightsWksp wksp;
@@ -117,9 +173,10 @@ size_t HUF_writeCTable_wksp(void* dst, size_t maxDstSize,
                             const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog,
                             void* workspace, size_t workspaceSize)
 {
+    HUF_CElt const* const ct = CTable + 1;
     BYTE* op = (BYTE*)dst;
     U32 n;
-    HUF_WriteCTableWksp* wksp = (HUF_WriteCTableWksp*)workspace;
+    HUF_WriteCTableWksp* wksp = (HUF_WriteCTableWksp*)HUF_alignUpWorkspace(workspace, &workspaceSize, ZSTD_ALIGNOF(U32));
 
     /* check conditions */
     if (workspaceSize < sizeof(HUF_WriteCTableWksp)) return ERROR(GENERIC);
@@ -130,9 +187,10 @@ size_t HUF_writeCTable_wksp(void* dst, size_t maxDstSize,
     for (n=1; n<huffLog+1; n++)
         wksp->bitsToWeight[n] = (BYTE)(huffLog + 1 - n);
     for (n=0; n<maxSymbolValue; n++)
-        wksp->huffWeight[n] = wksp->bitsToWeight[CTable[n].nbBits];
+        wksp->huffWeight[n] = wksp->bitsToWeight[HUF_getNbBits(ct[n])];
 
     /* attempt weights compression by FSE */
+    if (maxDstSize < 1) return ERROR(dstSize_tooSmall);
     {   CHECK_V_F(hSize, HUF_compressWeights(op+1, maxDstSize-1, wksp->huffWeight, maxSymbolValue, &wksp->wksp, sizeof(wksp->wksp)) );
         if ((hSize>1) & (hSize < maxSymbolValue/2)) {   /* FSE compressed */
             op[0] = (BYTE)hSize;
@@ -166,6 +224,7 @@ size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void
     U32 rankVal[HUF_TABLELOG_ABSOLUTEMAX + 1];   /* large enough for values from 0 to 16 */
     U32 tableLog = 0;
     U32 nbSymbols = 0;
+    HUF_CElt* const ct = CTable + 1;
 
     /* get symbol weights */
     CHECK_V_F(readSize, HUF_readStats(huffWeight, HUF_SYMBOLVALUE_MAX+1, rankVal, &nbSymbols, &tableLog, src, srcSize));
@@ -175,6 +234,8 @@ size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void
     if (tableLog > HUF_TABLELOG_MAX) return ERROR(tableLog_tooLarge);
     if (nbSymbols > *maxSymbolValuePtr+1) return ERROR(maxSymbolValue_tooSmall);
 
+    CTable[0] = tableLog;
+
     /* Prepare base value per rank */
     {   U32 n, nextRankStart = 0;
         for (n=1; n<=tableLog; n++) {
@@ -186,13 +247,13 @@ size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void
     /* fill nbBits */
     {   U32 n; for (n=0; n<nbSymbols; n++) {
             const U32 w = huffWeight[n];
-            CTable[n].nbBits = (BYTE)(tableLog + 1 - w) & -(w != 0);
+            HUF_setNbBits(ct + n, (BYTE)(tableLog + 1 - w) & -(w != 0));
     }   }
 
     /* fill val */
     {   U16 nbPerRank[HUF_TABLELOG_MAX+2]  = {0};  /* support w=0=>n=tableLog+1 */
         U16 valPerRank[HUF_TABLELOG_MAX+2] = {0};
-        { U32 n; for (n=0; n<nbSymbols; n++) nbPerRank[CTable[n].nbBits]++; }
+        { U32 n; for (n=0; n<nbSymbols; n++) nbPerRank[HUF_getNbBits(ct[n])]++; }
         /* determine stating value per rank */
         valPerRank[tableLog+1] = 0;   /* for w==0 */
         {   U16 min = 0;
@@ -202,18 +263,18 @@ size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void
                 min >>= 1;
         }   }
         /* assign value within rank, symbol order */
-        { U32 n; for (n=0; n<nbSymbols; n++) CTable[n].val = valPerRank[CTable[n].nbBits]++; }
+        { U32 n; for (n=0; n<nbSymbols; n++) HUF_setValue(ct + n, valPerRank[HUF_getNbBits(ct[n])]++); }
     }
 
     *maxSymbolValuePtr = nbSymbols - 1;
     return readSize;
 }
 
-U32 HUF_getNbBits(const void* symbolTable, U32 symbolValue)
+U32 HUF_getNbBitsFromCTable(HUF_CElt const* CTable, U32 symbolValue)
 {
-    const HUF_CElt* table = (const HUF_CElt*)symbolTable;
+    const HUF_CElt* ct = CTable + 1;
     assert(symbolValue <= HUF_SYMBOLVALUE_MAX);
-    return table[symbolValue].nbBits;
+    return (U32)HUF_getNbBits(ct[symbolValue]);
 }
 
 
@@ -367,22 +428,118 @@ static U32 HUF_setMaxHeight(nodeElt* huffNode, U32 lastNonNull, U32 maxNbBits)
 }
 
 typedef struct {
-    U32 base;
+    U16 base;
-    U32 curr;
+    U16 curr;
 } rankPos;
 
 typedef nodeElt huffNodeTable[HUF_CTABLE_WORKSPACE_SIZE_U32];
 
-#define RANK_POSITION_TABLE_SIZE 32
+/* Number of buckets available for HUF_sort() */
+#define RANK_POSITION_TABLE_SIZE 192
 
 typedef struct {
   huffNodeTable huffNodeTbl;
   rankPos rankPosition[RANK_POSITION_TABLE_SIZE];
 } HUF_buildCTable_wksp_tables;
 
+/* RANK_POSITION_DISTINCT_COUNT_CUTOFF == Cutoff point in HUF_sort() buckets for which we use log2 bucketing.
+ * Strategy is to use as many buckets as possible for representing distinct
+ * counts while using the remainder to represent all "large" counts.
+ *
+ * To satisfy this requirement for 192 buckets, we can do the following:
+ * Let buckets 0-166 represent distinct counts of [0, 166]
+ * 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_DISTINCT_COUNT_CUTOFF RANK_POSITION_LOG_BUCKETS_BEGIN + BIT_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.
+ */
+static U32 HUF_getIndex(U32 const count) {
+    return (count < RANK_POSITION_DISTINCT_COUNT_CUTOFF)
+        ? count
+        : BIT_highbit32(count) + RANK_POSITION_LOG_BUCKETS_BEGIN;
+}
+
+/* Helper swap function for HUF_quickSortPartition() */
+static void HUF_swapNodes(nodeElt* a, nodeElt* b) {
+	nodeElt tmp = *a;
+	*a = *b;
+	*b = tmp;
+}
+
+/* Returns 0 if the huffNode array is not sorted by descending count */
+MEM_STATIC int HUF_isSorted(nodeElt huffNode[], U32 const maxSymbolValue1) {
+    U32 i;
+    for (i = 1; i < maxSymbolValue1; ++i) {
+        if (huffNode[i].count > huffNode[i-1].count) {
+            return 0;
+        }
+    }
+    return 1;
+}
+
+/* Insertion sort by descending order */
+HINT_INLINE void HUF_insertionSort(nodeElt huffNode[], int const low, int const high) {
+    int i;
+    int const size = high-low+1;
+    huffNode += low;
+    for (i = 1; i < size; ++i) {
+        nodeElt const key = huffNode[i];
+        int j = i - 1;
+        while (j >= 0 && huffNode[j].count < key.count) {
+            huffNode[j + 1] = huffNode[j];
+            j--;
+        }
+        huffNode[j + 1] = key;
+    }
+}
+
+/* Pivot helper function for quicksort. */
+static int HUF_quickSortPartition(nodeElt arr[], int const low, int const high) {
+    /* Simply select rightmost element as pivot. "Better" selectors like
+     * median-of-three don't experimentally appear to have any benefit.
+     */
+    U32 const pivot = arr[high].count;
+    int i = low - 1;
+    int j = low;
+    for ( ; j < high; j++) {
+        if (arr[j].count > pivot) {
+            i++;
+            HUF_swapNodes(&arr[i], &arr[j]);
+        }
+    }
+    HUF_swapNodes(&arr[i + 1], &arr[high]);
+    return i + 1;
+}
+
+/* Classic quicksort by descending with partially iterative calls
+ * to reduce worst case callstack size.
+ */
+static void HUF_simpleQuickSort(nodeElt arr[], int low, int high) {
+    int const kInsertionSortThreshold = 8;
+    if (high - low < kInsertionSortThreshold) {
+        HUF_insertionSort(arr, low, high);
+        return;
+    }
+    while (low < high) {
+        int const idx = HUF_quickSortPartition(arr, low, high);
+        if (idx - low < high - idx) {
+            HUF_simpleQuickSort(arr, low, idx - 1);
+            low = idx + 1;
+        } else {
+            HUF_simpleQuickSort(arr, idx + 1, high);
+            high = idx - 1;
+        }
+    }
+}
+
 /**
  * HUF_sort():
  * Sorts the symbols [0, maxSymbolValue] by count[symbol] in decreasing order.
+ * This is a typical bucket sorting strategy that uses either quicksort or insertion sort to sort each bucket.
  *
  * @param[out] huffNode       Sorted symbols by decreasing count. Only members `.count` and `.byte` are filled.
  *                            Must have (maxSymbolValue + 1) entries.
@@ -390,44 +547,52 @@ typedef struct {
  * @param[in]  maxSymbolValue Maximum symbol value.
  * @param      rankPosition   This is a scratch workspace. Must have RANK_POSITION_TABLE_SIZE entries.
  */
-static void HUF_sort(nodeElt* huffNode, const unsigned* count, U32 maxSymbolValue, rankPos* rankPosition)
+static void HUF_sort(nodeElt huffNode[], const unsigned count[], U32 const maxSymbolValue, rankPos rankPosition[]) {
-{
+    U32 n;
-    int n;
+    U32 const maxSymbolValue1 = maxSymbolValue+1;
-    int const maxSymbolValue1 = (int)maxSymbolValue + 1;
 
     /* Compute base and set curr to base.
-     * For symbol s let lowerRank = BIT_highbit32(count[n]+1) and rank = lowerRank + 1.
+     * For symbol s let lowerRank = HUF_getIndex(count[n]) and rank = lowerRank + 1.
-     * Then 2^lowerRank <= count[n]+1 <= 2^rank.
+     * See HUF_getIndex to see bucketing strategy.
      * We attribute each symbol to lowerRank's base value, because we want to know where
      * each rank begins in the output, so for rank R we want to count ranks R+1 and above.
      */
     ZSTD_memset(rankPosition, 0, sizeof(*rankPosition) * RANK_POSITION_TABLE_SIZE);
     for (n = 0; n < maxSymbolValue1; ++n) {
-        U32 lowerRank = BIT_highbit32(count[n] + 1);
+        U32 lowerRank = HUF_getIndex(count[n]);
+        assert(lowerRank < RANK_POSITION_TABLE_SIZE - 1);
         rankPosition[lowerRank].base++;
     }
+
     assert(rankPosition[RANK_POSITION_TABLE_SIZE - 1].base == 0);
+    /* Set up the rankPosition table */
     for (n = RANK_POSITION_TABLE_SIZE - 1; n > 0; --n) {
         rankPosition[n-1].base += rankPosition[n].base;
         rankPosition[n-1].curr = rankPosition[n-1].base;
     }
-    /* Sort */
+
+    /* Insert each symbol into their appropriate bucket, setting up rankPosition table. */
     for (n = 0; n < maxSymbolValue1; ++n) {
         U32 const c = count[n];
-        U32 const r = BIT_highbit32(c+1) + 1;
+        U32 const r = HUF_getIndex(c) + 1;
-        U32 pos = rankPosition[r].curr++;
+        U32 const pos = rankPosition[r].curr++;
-        /* Insert into the correct position in the rank.
+        assert(pos < maxSymbolValue1);
-         * We have at most 256 symbols, so this insertion should be fine.
-         */
-        while ((pos > rankPosition[r].base) && (c > huffNode[pos-1].count)) {
-            huffNode[pos] = huffNode[pos-1];
-            pos--;
-        }
         huffNode[pos].count = c;
         huffNode[pos].byte  = (BYTE)n;
     }
+
+    /* 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;
+        U32 const bucketStartIdx = rankPosition[n].base;
+        if (bucketSize > 1) {
+            assert(bucketStartIdx < maxSymbolValue1);
+            HUF_simpleQuickSort(huffNode + bucketStartIdx, 0, bucketSize-1);
+        }
     }
 
+    assert(HUF_isSorted(huffNode, maxSymbolValue1));
+}
 
 /** HUF_buildCTable_wksp() :
  *  Same as HUF_buildCTable(), but using externally allocated scratch buffer.
@@ -490,6 +655,7 @@ static int HUF_buildTree(nodeElt* huffNode, U32 maxSymbolValue)
  */
 static void HUF_buildCTableFromTree(HUF_CElt* CTable, nodeElt const* huffNode, int nonNullRank, U32 maxSymbolValue, U32 maxNbBits)
 {
+    HUF_CElt* const ct = CTable + 1;
     /* fill result into ctable (val, nbBits) */
     int n;
     U16 nbPerRank[HUF_TABLELOG_MAX+1] = {0};
@@ -505,20 +671,20 @@ static void HUF_buildCTableFromTree(HUF_CElt* CTable, nodeElt const* huffNode, i
             min >>= 1;
     }   }
     for (n=0; n<alphabetSize; n++)
-        CTable[huffNode[n].byte].nbBits = huffNode[n].nbBits;   /* push nbBits per symbol, symbol order */
+        HUF_setNbBits(ct + huffNode[n].byte, huffNode[n].nbBits);   /* push nbBits per symbol, symbol order */
     for (n=0; n<alphabetSize; n++)
-        CTable[n].val = valPerRank[CTable[n].nbBits]++;   /* assign value within rank, symbol order */
+        HUF_setValue(ct + n, valPerRank[HUF_getNbBits(ct[n])]++);   /* assign value within rank, symbol order */
+    CTable[0] = maxNbBits;
 }
 
-size_t HUF_buildCTable_wksp (HUF_CElt* tree, 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*)workSpace;
+    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;
 
     /* safety checks */
-    if (((size_t)workSpace & 3) != 0) return ERROR(GENERIC);  /* must be aligned on 4-bytes boundaries */
     if (wkspSize < sizeof(HUF_buildCTable_wksp_tables))
       return ERROR(workSpace_tooSmall);
     if (maxNbBits == 0) maxNbBits = HUF_TABLELOG_DEFAULT;
@@ -536,96 +702,334 @@ size_t HUF_buildCTable_wksp (HUF_CElt* tree, const unsigned* count, U32 maxSymbo
     maxNbBits = HUF_setMaxHeight(huffNode, (U32)nonNullRank, maxNbBits);
     if (maxNbBits > HUF_TABLELOG_MAX) return ERROR(GENERIC);   /* check fit into table */
 
-    HUF_buildCTableFromTree(tree, huffNode, nonNullRank, maxSymbolValue, maxNbBits);
+    HUF_buildCTableFromTree(CTable, huffNode, nonNullRank, maxSymbolValue, maxNbBits);
 
     return maxNbBits;
 }
 
 size_t HUF_estimateCompressedSize(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue)
 {
+    HUF_CElt const* ct = CTable + 1;
     size_t nbBits = 0;
     int s;
     for (s = 0; s <= (int)maxSymbolValue; ++s) {
-        nbBits += CTable[s].nbBits * count[s];
+        nbBits += HUF_getNbBits(ct[s]) * count[s];
     }
     return nbBits >> 3;
 }
 
 int HUF_validateCTable(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue) {
+  HUF_CElt const* ct = CTable + 1;
   int bad = 0;
   int s;
   for (s = 0; s <= (int)maxSymbolValue; ++s) {
-    bad |= (count[s] != 0) & (CTable[s].nbBits == 0);
+    bad |= (count[s] != 0) & (HUF_getNbBits(ct[s]) == 0);
   }
   return !bad;
 }
 
 size_t HUF_compressBound(size_t size) { return HUF_COMPRESSBOUND(size); }
 
-FORCE_INLINE_TEMPLATE void
+/** HUF_CStream_t:
-HUF_encodeSymbol(BIT_CStream_t* bitCPtr, U32 symbol, const HUF_CElt* CTable)
+ * Huffman uses its own BIT_CStream_t implementation.
+ * There are three major differences from BIT_CStream_t:
+ *   1. HUF_addBits() takes a HUF_CElt (size_t) which is
+ *      the pair (nbBits, value) in the format:
+ *      format:
+ *        - Bits [0, 4)            = nbBits
+ *        - Bits [4, 64 - nbBits)  = 0
+ *        - Bits [64 - nbBits, 64) = value
+ *   2. The bitContainer is built from the upper bits and
+ *      right shifted. E.g. to add a new value of N bits
+ *      you right shift the bitContainer by N, then or in
+ *      the new value into the N upper bits.
+ *   3. The bitstream has two bit containers. You can add
+ *      bits to the second container and merge them into
+ *      the first container.
+ */
+
+#define HUF_BITS_IN_CONTAINER (sizeof(size_t) * 8)
+
+typedef struct {
+    size_t bitContainer[2];
+    size_t bitPos[2];
+
+    BYTE* startPtr;
+    BYTE* ptr;
+    BYTE* endPtr;
+} HUF_CStream_t;
+
+/**! HUF_initCStream():
+ * Initializes the bitstream.
+ * @returns 0 or an error code.
+ */
+static size_t HUF_initCStream(HUF_CStream_t* bitC,
+                                  void* startPtr, size_t dstCapacity)
 {
-    BIT_addBitsFast(bitCPtr, CTable[symbol].val, CTable[symbol].nbBits);
+    ZSTD_memset(bitC, 0, sizeof(*bitC));
+    bitC->startPtr = (BYTE*)startPtr;
+    bitC->ptr = bitC->startPtr;
+    bitC->endPtr = bitC->startPtr + dstCapacity - sizeof(bitC->bitContainer[0]);
+    if (dstCapacity <= sizeof(bitC->bitContainer[0])) return ERROR(dstSize_tooSmall);
+    return 0;
 }
 
-#define HUF_FLUSHBITS(s)  BIT_flushBits(s)
+/*! HUF_addBits():
+ * Adds the symbol stored in HUF_CElt elt to the bitstream.
+ *
+ * @param elt   The element we're adding. This is a (nbBits, value) pair.
+ *              See the HUF_CStream_t docs for the format.
+ * @param idx   Insert into the bitstream at this idx.
+ * @param kFast This is a template parameter. If the bitstream is guaranteed
+ *              to have at least 4 unused bits after this call it may be 1,
+ *              otherwise it must be 0. HUF_addBits() is faster when fast is set.
+ */
+FORCE_INLINE_TEMPLATE void HUF_addBits(HUF_CStream_t* bitC, HUF_CElt elt, int idx, int kFast)
+{
+    assert(idx <= 1);
+    assert(HUF_getNbBits(elt) <= HUF_TABLELOG_ABSOLUTEMAX);
+    /* This is efficient on x86-64 with BMI2 because shrx
+     * only reads the low 6 bits of the register. The compiler
+     * knows this and elides the mask. When fast is set,
+     * every operation can use the same value loaded from elt.
+     */
+    bitC->bitContainer[idx] >>= HUF_getNbBits(elt);
+    bitC->bitContainer[idx] |= kFast ? HUF_getValueFast(elt) : HUF_getValue(elt);
+    /* We only read the low 8 bits of bitC->bitPos[idx] so it
+     * doesn't matter that the high bits have noise from the value.
+     */
+    bitC->bitPos[idx] += HUF_getNbBitsFast(elt);
+    assert((bitC->bitPos[idx] & 0xFF) <= HUF_BITS_IN_CONTAINER);
+    /* The last 4-bits of elt are dirty if fast is set,
+     * so we must not be overwriting bits that have already been
+     * inserted into the bit container.
+     */
+#if DEBUGLEVEL >= 1
+    {
+        size_t const nbBits = HUF_getNbBits(elt);
+        size_t const dirtyBits = nbBits == 0 ? 0 : BIT_highbit32((U32)nbBits) + 1;
+        (void)dirtyBits;
+        /* Middle bits are 0. */
+        assert(((elt >> dirtyBits) << (dirtyBits + nbBits)) == 0);
+        /* We didn't overwrite any bits in the bit container. */
+        assert(!kFast || (bitC->bitPos[idx] & 0xFF) <= HUF_BITS_IN_CONTAINER);
+        (void)dirtyBits;
+    }
+#endif
+}
 
-#define HUF_FLUSHBITS_1(stream) \
+FORCE_INLINE_TEMPLATE void HUF_zeroIndex1(HUF_CStream_t* bitC)
-    if (sizeof((stream)->bitContainer)*8 < HUF_TABLELOG_MAX*2+7) HUF_FLUSHBITS(stream)
+{
+    bitC->bitContainer[1] = 0;
+    bitC->bitPos[1] = 0;
+}
+
+/*! HUF_mergeIndex1() :
+ * Merges the bit container @ index 1 into the bit container @ index 0
+ * and zeros the bit container @ index 1.
+ */
+FORCE_INLINE_TEMPLATE void HUF_mergeIndex1(HUF_CStream_t* bitC)
+{
+    assert((bitC->bitPos[1] & 0xFF) < HUF_BITS_IN_CONTAINER);
+    bitC->bitContainer[0] >>= (bitC->bitPos[1] & 0xFF);
+    bitC->bitContainer[0] |= bitC->bitContainer[1];
+    bitC->bitPos[0] += bitC->bitPos[1];
+    assert((bitC->bitPos[0] & 0xFF) <= HUF_BITS_IN_CONTAINER);
+}
+
+/*! HUF_flushBits() :
+* Flushes the bits in the bit container @ index 0.
+*
+* @post bitPos will be < 8.
+* @param kFast If kFast is set then we must know a-priori that
+*              the bit container will not overflow.
+*/
+FORCE_INLINE_TEMPLATE void HUF_flushBits(HUF_CStream_t* bitC, int kFast)
+{
+    /* The upper bits of bitPos are noisy, so we must mask by 0xFF. */
+    size_t const nbBits = bitC->bitPos[0] & 0xFF;
+    size_t const nbBytes = nbBits >> 3;
+    /* The top nbBits bits of bitContainer are the ones we need. */
+    size_t const bitContainer = bitC->bitContainer[0] >> (HUF_BITS_IN_CONTAINER - nbBits);
+    /* Mask bitPos to account for the bytes we consumed. */
+    bitC->bitPos[0] &= 7;
+    assert(nbBits > 0);
+    assert(nbBits <= sizeof(bitC->bitContainer[0]) * 8);
+    assert(bitC->ptr <= bitC->endPtr);
+    MEM_writeLEST(bitC->ptr, bitContainer);
+    bitC->ptr += nbBytes;
+    assert(!kFast || bitC->ptr <= bitC->endPtr);
+    if (!kFast && bitC->ptr > bitC->endPtr) bitC->ptr = bitC->endPtr;
+    /* bitContainer doesn't need to be modified because the leftover
+     * bits are already the top bitPos bits. And we don't care about
+     * noise in the lower values.
+     */
+}
+
+/*! HUF_endMark()
+ * @returns The Huffman stream end mark: A 1-bit value = 1.
+ */
+static HUF_CElt HUF_endMark(void)
+{
+    HUF_CElt endMark;
+    HUF_setNbBits(&endMark, 1);
+    HUF_setValue(&endMark, 1);
+    return endMark;
+}
+
+/*! HUF_closeCStream() :
+ *  @return Size of CStream, in bytes,
+ *          or 0 if it could not fit into dstBuffer */
+static size_t HUF_closeCStream(HUF_CStream_t* bitC)
+{
+    HUF_addBits(bitC, HUF_endMark(), /* idx */ 0, /* kFast */ 0);
+    HUF_flushBits(bitC, /* kFast */ 0);
+    {
+        size_t const nbBits = bitC->bitPos[0] & 0xFF;
+        if (bitC->ptr >= bitC->endPtr) return 0; /* overflow detected */
+        return (bitC->ptr - bitC->startPtr) + (nbBits > 0);
+    }
+}
+
+FORCE_INLINE_TEMPLATE void
+HUF_encodeSymbol(HUF_CStream_t* bitCPtr, U32 symbol, const HUF_CElt* CTable, int idx, int fast)
+{
+    HUF_addBits(bitCPtr, CTable[symbol], idx, fast);
+}
+
+FORCE_INLINE_TEMPLATE void
+HUF_compress1X_usingCTable_internal_body_loop(HUF_CStream_t* bitC,
+                                   const BYTE* ip, size_t srcSize,
+                                   const HUF_CElt* ct,
+                                   int kUnroll, int kFastFlush, int kLastFast)
+{
+    /* Join to kUnroll */
+    int n = (int)srcSize;
+    int rem = n % kUnroll;
+    if (rem > 0) {
+        for (; rem > 0; --rem) {
+            HUF_encodeSymbol(bitC, ip[--n], ct, 0, /* fast */ 0);
+        }
+        HUF_flushBits(bitC, kFastFlush);
+    }
+    assert(n % kUnroll == 0);
+
+    /* Join to 2 * kUnroll */
+    if (n % (2 * kUnroll)) {
+        int u;
+        for (u = 1; u < kUnroll; ++u) {
+            HUF_encodeSymbol(bitC, ip[n - u], ct, 0, 1);
+        }
+        HUF_encodeSymbol(bitC, ip[n - kUnroll], ct, 0, kLastFast);
+        HUF_flushBits(bitC, kFastFlush);
+        n -= kUnroll;
+    }
+    assert(n % (2 * kUnroll) == 0);
+
+    for (; n>0; n-= 2 * kUnroll) {
+        /* Encode kUnroll symbols into the bitstream @ index 0. */
+        int u;
+        for (u = 1; u < kUnroll; ++u) {
+            HUF_encodeSymbol(bitC, ip[n - u], ct, /* idx */ 0, /* fast */ 1);
+        }
+        HUF_encodeSymbol(bitC, ip[n - kUnroll], ct, /* idx */ 0, /* fast */ kLastFast);
+        HUF_flushBits(bitC, kFastFlush);
+        /* Encode kUnroll symbols into the bitstream @ index 1.
+         * This allows us to start filling the bit container
+         * without any data dependencies.
+         */
+        HUF_zeroIndex1(bitC);
+        for (u = 1; u < kUnroll; ++u) {
+            HUF_encodeSymbol(bitC, ip[n - kUnroll - u], ct, /* idx */ 1, /* fast */ 1);
+        }
+        HUF_encodeSymbol(bitC, ip[n - kUnroll - kUnroll], ct, /* idx */ 1, /* fast */ kLastFast);
+        /* Merge bitstream @ index 1 into the bitstream @ index 0 */
+        HUF_mergeIndex1(bitC);
+        HUF_flushBits(bitC, kFastFlush);
+    }
+    assert(n == 0);
+
+}
+
+/**
+ * Returns a tight upper bound on the output space needed by Huffman
+ * with 8 bytes buffer to handle over-writes. If the output is at least
+ * this large we don't need to do bounds checks during Huffman encoding.
+ */
+static size_t HUF_tightCompressBound(size_t srcSize, size_t tableLog)
+{
+    return ((srcSize * tableLog) >> 3) + 8;
+}
 
-#define HUF_FLUSHBITS_2(stream) \
-    if (sizeof((stream)->bitContainer)*8 < HUF_TABLELOG_MAX*4+7) HUF_FLUSHBITS(stream)
 
 FORCE_INLINE_TEMPLATE size_t
 HUF_compress1X_usingCTable_internal_body(void* dst, size_t dstSize,
                                    const void* src, size_t srcSize,
                                    const HUF_CElt* CTable)
 {
+    U32 const tableLog = (U32)CTable[0];
+    HUF_CElt const* ct = CTable + 1;
     const BYTE* ip = (const BYTE*) src;
     BYTE* const ostart = (BYTE*)dst;
     BYTE* const oend = ostart + dstSize;
     BYTE* op = ostart;
-    size_t n;
+    HUF_CStream_t bitC;
-    BIT_CStream_t bitC;
 
     /* init */
     if (dstSize < 8) return 0;   /* not enough space to compress */
-    { size_t const initErr = BIT_initCStream(&bitC, op, (size_t)(oend-op));
+    { size_t const initErr = HUF_initCStream(&bitC, op, (size_t)(oend-op));
       if (HUF_isError(initErr)) return 0; }
 
-    n = srcSize & ~3;  /* join to mod 4 */
+    if (dstSize < HUF_tightCompressBound(srcSize, (size_t)tableLog) || tableLog > 11)
-    switch (srcSize & 3)
+        HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ MEM_32bits() ? 2 : 4, /* kFast */ 0, /* kLastFast */ 0);
-    {
+    else {
-        case 3 : HUF_encodeSymbol(&bitC, ip[n+ 2], CTable);
+        if (MEM_32bits()) {
-                 HUF_FLUSHBITS_2(&bitC);
+            switch (tableLog) {
-		 /* fall-through */
+            case 11:
-        case 2 : HUF_encodeSymbol(&bitC, ip[n+ 1], CTable);
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 2, /* kFastFlush */ 1, /* kLastFast */ 0);
-                 HUF_FLUSHBITS_1(&bitC);
+                break;
-		 /* fall-through */
+            case 10: ZSTD_FALLTHROUGH;
-        case 1 : HUF_encodeSymbol(&bitC, ip[n+ 0], CTable);
+            case 9: ZSTD_FALLTHROUGH;
-                 HUF_FLUSHBITS(&bitC);
+            case 8:
-		 /* fall-through */
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 2, /* kFastFlush */ 1, /* kLastFast */ 1);
-        case 0 : /* fall-through */
+                break;
-        default: break;
+            case 7: ZSTD_FALLTHROUGH;
+            default:
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 3, /* kFastFlush */ 1, /* kLastFast */ 1);
+                break;
             }
-
+        } else {
-    for (; n>0; n-=4) {  /* note : n&3==0 at this stage */
+            switch (tableLog) {
-        HUF_encodeSymbol(&bitC, ip[n- 1], CTable);
+            case 11:
-        HUF_FLUSHBITS_1(&bitC);
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 5, /* kFastFlush */ 1, /* kLastFast */ 0);
-        HUF_encodeSymbol(&bitC, ip[n- 2], CTable);
+                break;
-        HUF_FLUSHBITS_2(&bitC);
+            case 10:
-        HUF_encodeSymbol(&bitC, ip[n- 3], CTable);
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 5, /* kFastFlush */ 1, /* kLastFast */ 1);
-        HUF_FLUSHBITS_1(&bitC);
+                break;
-        HUF_encodeSymbol(&bitC, ip[n- 4], CTable);
+            case 9:
-        HUF_FLUSHBITS(&bitC);
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 6, /* kFastFlush */ 1, /* kLastFast */ 0);
+                break;
+            case 8:
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 7, /* kFastFlush */ 1, /* kLastFast */ 0);
+                break;
+            case 7:
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 8, /* kFastFlush */ 1, /* kLastFast */ 0);
+                break;
+            case 6: ZSTD_FALLTHROUGH;
+            default:
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 9, /* kFastFlush */ 1, /* kLastFast */ 1);
+                break;
             }
+        }
+    }
+    assert(bitC.ptr <= bitC.endPtr);
 
-    return BIT_closeCStream(&bitC);
+    return HUF_closeCStream(&bitC);
 }
 
 #if DYNAMIC_BMI2
 
-static TARGET_ATTRIBUTE("bmi2") size_t
+static BMI2_TARGET_ATTRIBUTE size_t
 HUF_compress1X_usingCTable_internal_bmi2(void* dst, size_t dstSize,
                                    const void* src, size_t srcSize,
                                    const HUF_CElt* CTable)
@@ -667,9 +1071,13 @@ HUF_compress1X_usingCTable_internal(void* dst, size_t dstSize,
 
 size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable)
 {
-    return HUF_compress1X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, /* bmi2 */ 0);
+    return HUF_compress1X_usingCTable_bmi2(dst, dstSize, src, srcSize, CTable, /* bmi2 */ 0);
 }
 
+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,
@@ -689,8 +1097,7 @@ HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize,
 
     assert(op <= oend);
     {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, bmi2) );
-        if (cSize==0) return 0;
+        if (cSize == 0 || cSize > 65535) return 0;
-        assert(cSize <= 65535);
         MEM_writeLE16(ostart, (U16)cSize);
         op += cSize;
     }
@@ -698,8 +1105,7 @@ 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) );
-        if (cSize==0) return 0;
+        if (cSize == 0 || cSize > 65535) return 0;
-        assert(cSize <= 65535);
         MEM_writeLE16(ostart+2, (U16)cSize);
         op += cSize;
     }
@@ -707,8 +1113,7 @@ 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) );
-        if (cSize==0) return 0;
+        if (cSize == 0 || cSize > 65535) return 0;
-        assert(cSize <= 65535);
         MEM_writeLE16(ostart+4, (U16)cSize);
         op += cSize;
     }
@@ -717,7 +1122,7 @@ HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize,
     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) );
-        if (cSize==0) return 0;
+        if (cSize == 0 || cSize > 65535) return 0;
         op += cSize;
     }
 
@@ -726,7 +1131,12 @@ HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize,
 
 size_t HUF_compress4X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable)
 {
-    return HUF_compress4X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, /* bmi2 */ 0);
+    return HUF_compress4X_usingCTable_bmi2(dst, dstSize, src, srcSize, CTable, /* bmi2 */ 0);
+}
+
+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;
@@ -750,35 +1160,38 @@ static size_t HUF_compressCTable_internal(
 
 typedef struct {
     unsigned count[HUF_SYMBOLVALUE_MAX + 1];
-    HUF_CElt CTable[HUF_SYMBOLVALUE_MAX + 1];
+    HUF_CElt CTable[HUF_CTABLE_SIZE_ST(HUF_SYMBOLVALUE_MAX)];
     union {
         HUF_buildCTable_wksp_tables buildCTable_wksp;
         HUF_WriteCTableWksp writeCTable_wksp;
+        U32 hist_wksp[HIST_WKSP_SIZE_U32];
     } wksps;
 } HUF_compress_tables_t;
 
+#define SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE 4096
+#define SUSPECT_INCOMPRESSIBLE_SAMPLE_RATIO 10  /* Must be >= 2 */
+
 /* HUF_compress_internal() :
  * `workSpace_align4` must be aligned on 4-bytes boundaries,
- * and occupies the same space as a table of HUF_WORKSPACE_SIZE_U32 unsigned */
+ * and occupies the same space as a table of HUF_WORKSPACE_SIZE_U64 unsigned */
 static size_t
 HUF_compress_internal (void* dst, size_t dstSize,
                  const void* src, size_t srcSize,
                        unsigned maxSymbolValue, unsigned huffLog,
                        HUF_nbStreams_e nbStreams,
-                       void* workSpace_align4, size_t wkspSize,
+                       void* workSpace, size_t wkspSize,
                        HUF_CElt* oldHufTable, HUF_repeat* repeat, int preferRepeat,
-                 const int bmi2)
+                 const int bmi2, unsigned suspectUncompressible)
 {
-    HUF_compress_tables_t* const table = (HUF_compress_tables_t*)workSpace_align4;
+    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;
 
-    HUF_STATIC_ASSERT(sizeof(*table) <= HUF_WORKSPACE_SIZE);
+    HUF_STATIC_ASSERT(sizeof(*table) + HUF_WORKSPACE_MAX_ALIGNMENT <= HUF_WORKSPACE_SIZE);
-    assert(((size_t)workSpace_align4 & 3) == 0);   /* must be aligned on 4-bytes boundaries */
 
     /* checks & inits */
-    if (wkspSize < HUF_WORKSPACE_SIZE) return ERROR(workSpace_tooSmall);
+    if (wkspSize < sizeof(*table)) return ERROR(workSpace_tooSmall);
     if (!srcSize) return 0;  /* Uncompressed */
     if (!dstSize) return 0;  /* cannot fit anything within dst budget */
     if (srcSize > HUF_BLOCKSIZE_MAX) return ERROR(srcSize_wrong);   /* current block size limit */
@@ -794,8 +1207,23 @@ HUF_compress_internal (void* dst, size_t dstSize,
                                            nbStreams, oldHufTable, bmi2);
     }
 
+    /* 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)) {
+        size_t largestTotal = 0;
+        {   unsigned maxSymbolValueBegin = maxSymbolValue;
+            CHECK_V_F(largestBegin, HIST_count_simple (table->count, &maxSymbolValueBegin, (const BYTE*)src, SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE) );
+            largestTotal += largestBegin;
+        }
+        {   unsigned maxSymbolValueEnd = maxSymbolValue;
+            CHECK_V_F(largestEnd, HIST_count_simple (table->count, &maxSymbolValueEnd, (const BYTE*)src + srcSize - SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE, SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE) );
+            largestTotal += largestEnd;
+        }
+        if (largestTotal <= ((2 * SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE) >> 7)+4) return 0;   /* heuristic : probably not compressible enough */
+    }
+
     /* Scan input and build symbol stats */
-    {   CHECK_V_F(largest, HIST_count_wksp (table->count, &maxSymbolValue, (const BYTE*)src, srcSize, workSpace_align4, wkspSize) );
+    {   CHECK_V_F(largest, HIST_count_wksp (table->count, &maxSymbolValue, (const BYTE*)src, srcSize, table->wksps.hist_wksp, sizeof(table->wksps.hist_wksp)) );
         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 */
     }
@@ -820,9 +1248,12 @@ HUF_compress_internal (void* dst, size_t dstSize,
                                             &table->wksps.buildCTable_wksp, sizeof(table->wksps.buildCTable_wksp));
         CHECK_F(maxBits);
         huffLog = (U32)maxBits;
+    }
     /* Zero unused symbols in CTable, so we can check it for validity */
-        ZSTD_memset(table->CTable + (maxSymbolValue + 1), 0,
+    {
-               sizeof(table->CTable) - ((maxSymbolValue + 1) * sizeof(HUF_CElt)));
+        size_t const ctableSize = HUF_CTABLE_SIZE_ST(maxSymbolValue);
+        size_t const unusedSize = sizeof(table->CTable) - ctableSize * sizeof(HUF_CElt);
+        ZSTD_memset(table->CTable + ctableSize, 0, unusedSize);
     }
 
     /* Write table description header */
@@ -859,19 +1290,20 @@ size_t HUF_compress1X_wksp (void* dst, size_t dstSize,
     return HUF_compress_internal(dst, dstSize, src, srcSize,
                                  maxSymbolValue, huffLog, HUF_singleStream,
                                  workSpace, wkspSize,
-                                 NULL, NULL, 0, 0 /*bmi2*/);
+                                 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, int bmi2)
+                      HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat,
+                      int bmi2, unsigned suspectUncompressible)
 {
     return HUF_compress_internal(dst, dstSize, src, srcSize,
                                  maxSymbolValue, huffLog, HUF_singleStream,
                                  workSpace, wkspSize, hufTable,
-                                 repeat, preferRepeat, bmi2);
+                                 repeat, preferRepeat, bmi2, suspectUncompressible);
 }
 
 /* HUF_compress4X_repeat():
@@ -885,22 +1317,23 @@ size_t HUF_compress4X_wksp (void* dst, size_t dstSize,
     return HUF_compress_internal(dst, dstSize, src, srcSize,
                                  maxSymbolValue, huffLog, HUF_fourStreams,
                                  workSpace, wkspSize,
-                                 NULL, NULL, 0, 0 /*bmi2*/);
+                                 NULL, NULL, 0, 0 /*bmi2*/, 0);
 }
 
 /* HUF_compress4X_repeat():
  * compress input using 4 streams.
+ * consider skipping quickly
  * re-use an existing huffman compression table */
 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)
+                      HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2, unsigned suspectUncompressible)
 {
     return HUF_compress_internal(dst, dstSize, src, srcSize,
                                  maxSymbolValue, huffLog, HUF_fourStreams,
                                  workSpace, wkspSize,
-                                 hufTable, repeat, preferRepeat, bmi2);
+                                 hufTable, repeat, preferRepeat, bmi2, suspectUncompressible);
 }
 
 #ifndef ZSTD_NO_UNUSED_FUNCTIONS
@@ -918,7 +1351,7 @@ size_t HUF_compress1X (void* dst, size_t dstSize,
                  const void* src, size_t srcSize,
                  unsigned maxSymbolValue, unsigned huffLog)
 {
-    unsigned workSpace[HUF_WORKSPACE_SIZE_U32];
+    U64 workSpace[HUF_WORKSPACE_SIZE_U64];
     return HUF_compress1X_wksp(dst, dstSize, src, srcSize, maxSymbolValue, huffLog, workSpace, sizeof(workSpace));
 }
 
@@ -926,7 +1359,7 @@ size_t HUF_compress2 (void* dst, size_t dstSize,
                 const void* src, size_t srcSize,
                 unsigned maxSymbolValue, unsigned huffLog)
 {
-    unsigned workSpace[HUF_WORKSPACE_SIZE_U32];
+    U64 workSpace[HUF_WORKSPACE_SIZE_U64];
     return HUF_compress4X_wksp(dst, dstSize, src, srcSize, maxSymbolValue, huffLog, workSpace, sizeof(workSpace));
 }
 

zstd/compress/zstd_compress_internal.h

@@ -63,7 +63,7 @@ typedef struct {
 } ZSTD_localDict;
 
 typedef struct {
-    HUF_CElt CTable[HUF_CTABLE_SIZE_U32(255)];
+    HUF_CElt CTable[HUF_CTABLE_SIZE_ST(255)];
     HUF_repeat repeatMode;
 } ZSTD_hufCTables_t;
 
@@ -179,7 +179,7 @@ typedef struct {
     U32  offCodeSumBasePrice;    /* to compare to log2(offreq)  */
     ZSTD_OptPrice_e priceType;   /* prices can be determined dynamically, or follow a pre-defined cost structure */
     const ZSTD_entropyCTables_t* symbolCosts;  /* pre-calculated dictionary statistics */
-    ZSTD_literalCompressionMode_e literalCompressionMode;
+    ZSTD_paramSwitch_e literalCompressionMode;
 } optState_t;
 
 typedef struct {
@@ -199,6 +199,8 @@ typedef struct {
                                 */
 } ZSTD_window_t;
 
+#define ZSTD_WINDOW_START_INDEX 2
+
 typedef struct ZSTD_matchState_t ZSTD_matchState_t;
 
 #define ZSTD_ROW_HASH_CACHE_SIZE 8       /* Size of prefetching hash cache for row-based matchfinder */
@@ -264,7 +266,7 @@ typedef struct {
 } ldmState_t;
 
 typedef struct {
-    U32 enableLdm;          /* 1 if enable long distance matching */
+    ZSTD_paramSwitch_e enableLdm; /* ZSTD_ps_enable to enable LDM. ZSTD_ps_auto by default */
     U32 hashLog;            /* Log size of hashTable */
     U32 bucketSizeLog;      /* Log bucket size for collision resolution, at most 8 */
     U32 minMatchLength;     /* Minimum match length */
@@ -295,7 +297,7 @@ struct ZSTD_CCtx_params_s {
                                 * There is no guarantee that hint is close to actual source size */
 
     ZSTD_dictAttachPref_e attachDictPref;
-    ZSTD_literalCompressionMode_e literalCompressionMode;
+    ZSTD_paramSwitch_e literalCompressionMode;
 
     /* Multithreading: used to pass parameters to mtctx */
     int nbWorkers;
@@ -318,10 +320,10 @@ struct ZSTD_CCtx_params_s {
     int validateSequences;
 
     /* Block splitting */
-    int splitBlocks;
+    ZSTD_paramSwitch_e useBlockSplitter;
 
     /* Param for deciding whether to use row-based matchfinder */
-    ZSTD_useRowMatchFinderMode_e useRowMatchFinder;
+    ZSTD_paramSwitch_e useRowMatchFinder;
 
     /* Always load a dictionary in ext-dict mode (not prefix mode)? */
     int deterministicRefPrefix;
@@ -343,6 +345,22 @@ typedef enum {
     ZSTDb_buffered
 } ZSTD_buffered_policy_e;
 
+/**
+ * Struct that contains all elements of block splitter that should be allocated
+ * in a wksp.
+ */
+#define ZSTD_MAX_NB_BLOCK_SPLITS 196
+typedef struct {
+    seqStore_t fullSeqStoreChunk;
+    seqStore_t firstHalfSeqStore;
+    seqStore_t secondHalfSeqStore;
+    seqStore_t currSeqStore;
+    seqStore_t nextSeqStore;
+
+    U32 partitions[ZSTD_MAX_NB_BLOCK_SPLITS];
+    ZSTD_entropyCTablesMetadata_t entropyMetadata;
+} ZSTD_blockSplitCtx;
+
 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. */
@@ -374,7 +392,7 @@ struct ZSTD_CCtx_s {
     ZSTD_blockState_t blockState;
     U32* entropyWorkspace;  /* entropy workspace of ENTROPY_WORKSPACE_SIZE bytes */
 
-    /* Wether we are streaming or not */
+    /* Whether we are streaming or not */
     ZSTD_buffered_policy_e bufferedPolicy;
 
     /* streaming */
@@ -408,6 +426,9 @@ struct ZSTD_CCtx_s {
 #if ZSTD_TRACE
     ZSTD_TraceCtx traceCtx;
 #endif
+
+    /* Workspace for block splitter */
+    ZSTD_blockSplitCtx blockSplitCtx;
 };
 
 typedef enum { ZSTD_dtlm_fast, ZSTD_dtlm_full } ZSTD_dictTableLoadMethod_e;
@@ -442,7 +463,7 @@ typedef enum {
 typedef size_t (*ZSTD_blockCompressor) (
         ZSTD_matchState_t* bs, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
-ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_useRowMatchFinderMode_e rowMatchfinderMode, ZSTD_dictMode_e dictMode);
+ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_paramSwitch_e rowMatchfinderMode, ZSTD_dictMode_e dictMode);
 
 
 MEM_STATIC U32 ZSTD_LLcode(U32 litLength)
@@ -549,17 +570,17 @@ MEM_STATIC size_t ZSTD_minGain(size_t srcSize, ZSTD_strategy strat)
     return (srcSize >> minlog) + 2;
 }
 
-MEM_STATIC int ZSTD_disableLiteralsCompression(const ZSTD_CCtx_params* cctxParams)
+MEM_STATIC int ZSTD_literalsCompressionIsDisabled(const ZSTD_CCtx_params* cctxParams)
 {
     switch (cctxParams->literalCompressionMode) {
-    case ZSTD_lcm_huffman:
+    case ZSTD_ps_enable:
         return 0;
-    case ZSTD_lcm_uncompressed:
+    case ZSTD_ps_disable:
         return 1;
     default:
         assert(0 /* impossible: pre-validated */);
-        /* fall-through */
+        ZSTD_FALLTHROUGH;
-    case ZSTD_lcm_auto:
+    case ZSTD_ps_auto:
         return (cctxParams->cParams.strategy == ZSTD_fast) && (cctxParams->cParams.targetLength > 0);
     }
 }
@@ -651,8 +672,14 @@ static unsigned ZSTD_NbCommonBytes (size_t val)
 #           if STATIC_BMI2
                 return _tzcnt_u64(val) >> 3;
 #           else
-                unsigned long r = 0;
+                if (val != 0) {
-                return _BitScanForward64( &r, (U64)val ) ? (unsigned)(r >> 3) : 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);
@@ -669,8 +696,14 @@ static unsigned ZSTD_NbCommonBytes (size_t val)
 #       endif
         } else { /* 32 bits */
 #       if defined(_MSC_VER)
-            unsigned long r=0;
+            if (val != 0) {
-            return _BitScanForward( &r, (U32)val ) ? (unsigned)(r >> 3) : 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
@@ -687,8 +720,14 @@ static unsigned ZSTD_NbCommonBytes (size_t val)
 #           if STATIC_BMI2
 			    return _lzcnt_u64(val) >> 3;
 #           else
-			    unsigned long r = 0;
+                if (val != 0) {
-			    return _BitScanReverse64(&r, (U64)val) ? (unsigned)(r >> 3) : 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);
@@ -702,8 +741,14 @@ static unsigned ZSTD_NbCommonBytes (size_t val)
 #       endif
         } else { /* 32 bits */
 #       if defined(_MSC_VER)
-            unsigned long r = 0;
+            if (val != 0) {
-            return _BitScanReverse( &r, (unsigned long)val ) ? (unsigned)(r >> 3) : 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
@@ -884,9 +929,9 @@ MEM_STATIC void ZSTD_window_clear(ZSTD_window_t* window)
 
 MEM_STATIC U32 ZSTD_window_isEmpty(ZSTD_window_t const window)
 {
-    return window.dictLimit == 1 &&
+    return window.dictLimit == ZSTD_WINDOW_START_INDEX &&
-           window.lowLimit == 1 &&
+           window.lowLimit == ZSTD_WINDOW_START_INDEX &&
-           (window.nextSrc - window.base) == 1;
+           (window.nextSrc - window.base) == ZSTD_WINDOW_START_INDEX;
 }
 
 /**
@@ -937,7 +982,9 @@ MEM_STATIC U32 ZSTD_window_canOverflowCorrect(ZSTD_window_t const window,
 {
     U32 const cycleSize = 1u << cycleLog;
     U32 const curr = (U32)((BYTE const*)src - window.base);
-    U32 const minIndexToOverflowCorrect = cycleSize + MAX(maxDist, cycleSize);
+    U32 const minIndexToOverflowCorrect = cycleSize
+                                        + MAX(maxDist, cycleSize)
+                                        + ZSTD_WINDOW_START_INDEX;
 
     /* Adjust the min index to backoff the overflow correction frequency,
      * so we don't waste too much CPU in overflow correction. If this
@@ -1012,10 +1059,14 @@ MEM_STATIC U32 ZSTD_window_correctOverflow(ZSTD_window_t* window, U32 cycleLog,
     U32 const cycleSize = 1u << cycleLog;
     U32 const cycleMask = cycleSize - 1;
     U32 const curr = (U32)((BYTE const*)src - window->base);
-    U32 const currentCycle0 = curr & cycleMask;
+    U32 const currentCycle = curr & cycleMask;
-    /* Exclude zero so that newCurrent - maxDist >= 1. */
+    /* Ensure newCurrent - maxDist >= ZSTD_WINDOW_START_INDEX. */
-    U32 const currentCycle1 = currentCycle0 == 0 ? cycleSize : currentCycle0;
+    U32 const currentCycleCorrection = currentCycle < ZSTD_WINDOW_START_INDEX
-    U32 const newCurrent = currentCycle1 + MAX(maxDist, cycleSize);
+                                     ? MAX(cycleSize, ZSTD_WINDOW_START_INDEX)
+                                     : 0;
+    U32 const newCurrent = currentCycle
+                         + currentCycleCorrection
+                         + MAX(maxDist, cycleSize);
     U32 const correction = curr - newCurrent;
     /* maxDist must be a power of two so that:
      *   (newCurrent & cycleMask) == (curr & cycleMask)
@@ -1031,14 +1082,20 @@ MEM_STATIC U32 ZSTD_window_correctOverflow(ZSTD_window_t* window, U32 cycleLog,
 
     window->base += correction;
     window->dictBase += correction;
-    if (window->lowLimit <= correction) window->lowLimit = 1;
+    if (window->lowLimit < correction + ZSTD_WINDOW_START_INDEX) {
-    else window->lowLimit -= correction;
+        window->lowLimit = ZSTD_WINDOW_START_INDEX;
-    if (window->dictLimit <= correction) window->dictLimit = 1;
+    } else {
-    else window->dictLimit -= correction;
+        window->lowLimit -= correction;
+    }
+    if (window->dictLimit < correction + ZSTD_WINDOW_START_INDEX) {
+        window->dictLimit = ZSTD_WINDOW_START_INDEX;
+    } else {
+        window->dictLimit -= correction;
+    }
 
     /* Ensure we can still reference the full window. */
     assert(newCurrent >= maxDist);
-    assert(newCurrent - maxDist >= 1);
+    assert(newCurrent - maxDist >= ZSTD_WINDOW_START_INDEX);
     /* Ensure that lowLimit and dictLimit didn't underflow. */
     assert(window->lowLimit <= newCurrent);
     assert(window->dictLimit <= newCurrent);
@@ -1151,9 +1208,10 @@ MEM_STATIC void ZSTD_window_init(ZSTD_window_t* window) {
     ZSTD_memset(window, 0, sizeof(*window));
     window->base = (BYTE const*)" ";
     window->dictBase = (BYTE const*)" ";
-    window->dictLimit = 1;    /* start from 1, so that 1st position is valid */
+    ZSTD_STATIC_ASSERT(ZSTD_DUBT_UNSORTED_MARK < ZSTD_WINDOW_START_INDEX); /* Start above ZSTD_DUBT_UNSORTED_MARK */
-    window->lowLimit = 1;     /* it ensures first and later CCtx usages compress the same */
+    window->dictLimit = ZSTD_WINDOW_START_INDEX;    /* start from >0, so that 1st position is valid */
-    window->nextSrc = window->base + 1;   /* see issue #1241 */
+    window->lowLimit = ZSTD_WINDOW_START_INDEX;     /* it ensures first and later CCtx usages compress the same */
+    window->nextSrc = window->base + ZSTD_WINDOW_START_INDEX;   /* see issue #1241 */
     window->nbOverflowCorrections = 0;
 }
 

zstd/compress/zstd_compress_literals.c

@@ -73,7 +73,8 @@ size_t ZSTD_compressLiterals (ZSTD_hufCTables_t const* prevHuf,
                               void* dst, size_t dstCapacity,
                         const void* src, size_t srcSize,
                               void* entropyWorkspace, size_t entropyWorkspaceSize,
-                        const int bmi2)
+                        const int bmi2,
+                        unsigned suspectUncompressible)
 {
     size_t const minGain = ZSTD_minGain(srcSize, strategy);
     size_t const lhSize = 3 + (srcSize >= 1 KB) + (srcSize >= 16 KB);
@@ -105,11 +106,11 @@ size_t ZSTD_compressLiterals (ZSTD_hufCTables_t const* prevHuf,
             HUF_compress1X_repeat(
                 ostart+lhSize, dstCapacity-lhSize, src, srcSize,
                 HUF_SYMBOLVALUE_MAX, HUF_TABLELOG_DEFAULT, entropyWorkspace, entropyWorkspaceSize,
-                (HUF_CElt*)nextHuf->CTable, &repeat, preferRepeat, bmi2) :
+                (HUF_CElt*)nextHuf->CTable, &repeat, preferRepeat, bmi2, suspectUncompressible) :
             HUF_compress4X_repeat(
                 ostart+lhSize, dstCapacity-lhSize, src, srcSize,
                 HUF_SYMBOLVALUE_MAX, HUF_TABLELOG_DEFAULT, entropyWorkspace, entropyWorkspaceSize,
-                (HUF_CElt*)nextHuf->CTable, &repeat, preferRepeat, bmi2);
+                (HUF_CElt*)nextHuf->CTable, &repeat, preferRepeat, bmi2, suspectUncompressible);
         if (repeat != HUF_repeat_none) {
             /* reused the existing table */
             DEBUGLOG(5, "Reusing previous huffman table");

zstd/compress/zstd_compress_literals.h

@@ -18,12 +18,14 @@ size_t ZSTD_noCompressLiterals (void* dst, size_t dstCapacity, const void* src,
 
 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 */
 size_t ZSTD_compressLiterals (ZSTD_hufCTables_t const* prevHuf,
                               ZSTD_hufCTables_t* nextHuf,
                               ZSTD_strategy strategy, int disableLiteralCompression,
                               void* dst, size_t dstCapacity,
                         const void* src, size_t srcSize,
                               void* entropyWorkspace, size_t entropyWorkspaceSize,
-                        const int bmi2);
+                        const int bmi2,
+                        unsigned suspectUncompressible);
 
 #endif /* ZSTD_COMPRESS_LITERALS_H */

zstd/compress/zstd_compress_sequences.c

@@ -275,10 +275,11 @@ ZSTD_buildCTable(void* dst, size_t dstCapacity,
         assert(nbSeq_1 > 1);
         assert(entropyWorkspaceSize >= sizeof(ZSTD_BuildCTableWksp));
         (void)entropyWorkspaceSize;
-        FORWARD_IF_ERROR(FSE_normalizeCount(wksp->norm, tableLog, count, nbSeq_1, max, ZSTD_useLowProbCount(nbSeq_1)), "");
+        FORWARD_IF_ERROR(FSE_normalizeCount(wksp->norm, tableLog, count, nbSeq_1, max, ZSTD_useLowProbCount(nbSeq_1)), "FSE_normalizeCount failed");
-        {   size_t const NCountSize = FSE_writeNCount(op, oend - op, wksp->norm, max, tableLog);   /* overflow protected */
+        assert(oend >= op);
+        {   size_t const NCountSize = FSE_writeNCount(op, (size_t)(oend - op), wksp->norm, max, tableLog);   /* overflow protected */
             FORWARD_IF_ERROR(NCountSize, "FSE_writeNCount failed");
-            FORWARD_IF_ERROR(FSE_buildCTable_wksp(nextCTable, wksp->norm, max, tableLog, wksp->wksp, sizeof(wksp->wksp)), "");
+            FORWARD_IF_ERROR(FSE_buildCTable_wksp(nextCTable, wksp->norm, max, tableLog, wksp->wksp, sizeof(wksp->wksp)), "FSE_buildCTable_wksp failed");
             return NCountSize;
         }
     }
@@ -398,7 +399,7 @@ ZSTD_encodeSequences_default(
 
 #if DYNAMIC_BMI2
 
-static TARGET_ATTRIBUTE("bmi2") size_t
+static BMI2_TARGET_ATTRIBUTE size_t
 ZSTD_encodeSequences_bmi2(
             void* dst, size_t dstCapacity,
             FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,

zstd/compress/zstd_compress_superblock.c

@@ -132,6 +132,7 @@ static size_t ZSTD_seqDecompressedSize(seqStore_t const* seqStore, const seqDef*
     const seqDef* sp = sstart;
     size_t matchLengthSum = 0;
     size_t litLengthSum = 0;
+    (void)(litLengthSum); /* suppress unused variable warning on some environments */
     while (send-sp > 0) {
         ZSTD_sequenceLength const seqLen = ZSTD_getSequenceLength(seqStore, sp);
         litLengthSum += seqLen.litLength;
@@ -324,7 +325,7 @@ static size_t ZSTD_estimateSubBlockSize_literal(const BYTE* literals, size_t lit
 static size_t ZSTD_estimateSubBlockSize_symbolType(symbolEncodingType_e type,
                         const BYTE* codeTable, unsigned maxCode,
                         size_t nbSeq, const FSE_CTable* fseCTable,
-                        const U32* additionalBits,
+                        const U8* additionalBits,
                         short const* defaultNorm, U32 defaultNormLog, U32 defaultMax,
                         void* workspace, size_t wkspSize)
 {

zstd/compress/zstd_cwksp.h

@@ -219,7 +219,7 @@ MEM_STATIC size_t ZSTD_cwksp_aligned_alloc_size(size_t size) {
 MEM_STATIC size_t ZSTD_cwksp_slack_space_required(void) {
     /* For alignment, the wksp will always allocate an additional n_1=[1, 64] bytes
      * to align the beginning of tables section, as well as another n_2=[0, 63] bytes
-     * to align the beginning of the aligned secion.
+     * to align the beginning of the aligned section.
      *
      * n_1 + n_2 == 64 bytes if the cwksp is freshly allocated, due to tables and
      * aligneds being sized in multiples of 64 bytes.
@@ -422,8 +422,8 @@ MEM_STATIC void* ZSTD_cwksp_reserve_object(ZSTD_cwksp* ws, size_t bytes) {
     DEBUGLOG(5,
         "cwksp: reserving %p object %zd bytes (rounded to %zd), %zd bytes remaining",
         alloc, bytes, roundedBytes, ZSTD_cwksp_available_space(ws) - roundedBytes);
-    assert(((size_t)alloc & (sizeof(void*)-1)) == 0);
+    assert((size_t)alloc % ZSTD_ALIGNOF(void*) == 0);
-    assert((bytes & (sizeof(void*)-1)) == 0);
+    assert(bytes % ZSTD_ALIGNOF(void*) == 0);
     ZSTD_cwksp_assert_internal_consistency(ws);
     /* we must be in the first phase, no advance is possible */
     if (ws->phase != ZSTD_cwksp_alloc_objects || end > ws->workspaceEnd) {

zstd/compress/zstd_double_fast.c

@@ -48,10 +48,216 @@ void ZSTD_fillDoubleHashTable(ZSTD_matchState_t* ms,
 
 
 FORCE_INLINE_TEMPLATE
-size_t ZSTD_compressBlock_doubleFast_generic(
+size_t ZSTD_compressBlock_doubleFast_noDict_generic(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize, U32 const mls /* template */)
+{
+    ZSTD_compressionParameters const* cParams = &ms->cParams;
+    U32* const hashLong = ms->hashTable;
+    const U32 hBitsL = cParams->hashLog;
+    U32* const hashSmall = ms->chainTable;
+    const U32 hBitsS = cParams->chainLog;
+    const BYTE* const base = ms->window.base;
+    const BYTE* const istart = (const BYTE*)src;
+    const BYTE* anchor = istart;
+    const U32 endIndex = (U32)((size_t)(istart - base) + srcSize);
+    /* presumes that, if there is a dictionary, it must be using Attach mode */
+    const U32 prefixLowestIndex = ZSTD_getLowestPrefixIndex(ms, endIndex, cParams->windowLog);
+    const BYTE* const prefixLowest = base + prefixLowestIndex;
+    const BYTE* const iend = istart + srcSize;
+    const BYTE* const ilimit = iend - HASH_READ_SIZE;
+    U32 offset_1=rep[0], offset_2=rep[1];
+    U32 offsetSaved = 0;
+
+    size_t mLength;
+    U32 offset;
+    U32 curr;
+
+    /* how many positions to search before increasing step size */
+    const size_t kStepIncr = 1 << kSearchStrength;
+    /* the position at which to increment the step size if no match is found */
+    const BYTE* nextStep;
+    size_t step; /* the current step size */
+
+    size_t hl0; /* the long hash at ip */
+    size_t hl1; /* the long hash at ip1 */
+
+    U32 idxl0; /* the long match index for ip */
+    U32 idxl1; /* the long match index for ip1 */
+
+    const BYTE* matchl0; /* the long match for ip */
+    const BYTE* matchs0; /* the short match for ip */
+    const BYTE* matchl1; /* the long match for ip1 */
+
+    const BYTE* ip = istart; /* the current position */
+    const BYTE* ip1; /* the next position */
+
+    DEBUGLOG(5, "ZSTD_compressBlock_doubleFast_noDict_generic");
+
+    /* init */
+    ip += ((ip - prefixLowest) == 0);
+    {
+        U32 const current = (U32)(ip - base);
+        U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, current, cParams->windowLog);
+        U32 const maxRep = current - windowLow;
+        if (offset_2 > maxRep) offsetSaved = offset_2, offset_2 = 0;
+        if (offset_1 > maxRep) offsetSaved = offset_1, offset_1 = 0;
+    }
+
+    /* Outer Loop: one iteration per match found and stored */
+    while (1) {
+        step = 1;
+        nextStep = ip + kStepIncr;
+        ip1 = ip + step;
+
+        if (ip1 > ilimit) {
+            goto _cleanup;
+        }
+
+        hl0 = ZSTD_hashPtr(ip, hBitsL, 8);
+        idxl0 = hashLong[hl0];
+        matchl0 = base + idxl0;
+
+        /* Inner Loop: one iteration per search / position */
+        do {
+            const size_t hs0 = ZSTD_hashPtr(ip, hBitsS, mls);
+            const U32 idxs0 = hashSmall[hs0];
+            curr = (U32)(ip-base);
+            matchs0 = base + idxs0;
+
+            hashLong[hl0] = hashSmall[hs0] = curr;   /* update hash tables */
+
+            /* check noDict repcode */
+            if ((offset_1 > 0) & (MEM_read32(ip+1-offset_1) == MEM_read32(ip+1))) {
+                mLength = ZSTD_count(ip+1+4, ip+1+4-offset_1, iend) + 4;
+                ip++;
+                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, 0, mLength-MINMATCH);
+                goto _match_stored;
+            }
+
+            hl1 = ZSTD_hashPtr(ip1, hBitsL, 8);
+
+            if (idxl0 > prefixLowestIndex) {
+                /* check prefix long match */
+                if (MEM_read64(matchl0) == MEM_read64(ip)) {
+                    mLength = ZSTD_count(ip+8, matchl0+8, iend) + 8;
+                    offset = (U32)(ip-matchl0);
+                    while (((ip>anchor) & (matchl0>prefixLowest)) && (ip[-1] == matchl0[-1])) { ip--; matchl0--; mLength++; } /* catch up */
+                    goto _match_found;
+                }
+            }
+
+            idxl1 = hashLong[hl1];
+            matchl1 = base + idxl1;
+
+            if (idxs0 > prefixLowestIndex) {
+                /* check prefix short match */
+                if (MEM_read32(matchs0) == MEM_read32(ip)) {
+                    goto _search_next_long;
+                }
+            }
+
+            if (ip1 >= nextStep) {
+                PREFETCH_L1(ip1 + 64);
+                PREFETCH_L1(ip1 + 128);
+                step++;
+                nextStep += kStepIncr;
+            }
+            ip = ip1;
+            ip1 += step;
+
+            hl0 = hl1;
+            idxl0 = idxl1;
+            matchl0 = matchl1;
+    #if defined(__aarch64__)
+            PREFETCH_L1(ip+256);
+    #endif
+        } while (ip1 <= ilimit);
+
+_cleanup:
+        /* save reps for next block */
+        rep[0] = offset_1 ? offset_1 : offsetSaved;
+        rep[1] = offset_2 ? offset_2 : offsetSaved;
+
+        /* Return the last literals size */
+        return (size_t)(iend - anchor);
+
+_search_next_long:
+
+        /* check prefix long +1 match */
+        if (idxl1 > prefixLowestIndex) {
+            if (MEM_read64(matchl1) == MEM_read64(ip1)) {
+                ip = ip1;
+                mLength = ZSTD_count(ip+8, matchl1+8, iend) + 8;
+                offset = (U32)(ip-matchl1);
+                while (((ip>anchor) & (matchl1>prefixLowest)) && (ip[-1] == matchl1[-1])) { ip--; matchl1--; mLength++; } /* catch up */
+                goto _match_found;
+            }
+        }
+
+        /* if no long +1 match, explore the short match we found */
+        mLength = ZSTD_count(ip+4, matchs0+4, iend) + 4;
+        offset = (U32)(ip - matchs0);
+        while (((ip>anchor) & (matchs0>prefixLowest)) && (ip[-1] == matchs0[-1])) { ip--; matchs0--; mLength++; } /* catch up */
+
+        /* fall-through */
+
+_match_found: /* requires ip, offset, mLength */
+        offset_2 = offset_1;
+        offset_1 = offset;
+
+        if (step < 4) {
+            /* It is unsafe to write this value back to the hashtable when ip1 is
+             * greater than or equal to the new ip we will have after we're done
+             * processing this match. Rather than perform that test directly
+             * (ip1 >= ip + mLength), which costs speed in practice, we do a simpler
+             * more predictable test. The minmatch even if we take a short match is
+             * 4 bytes, so as long as step, the distance between ip and ip1
+             * (initially) is less than 4, we know ip1 < new ip. */
+            hashLong[hl1] = (U32)(ip1 - base);
+        }
+
+        ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH);
+
+_match_stored:
+        /* match found */
+        ip += mLength;
+        anchor = ip;
+
+        if (ip <= ilimit) {
+            /* Complementary insertion */
+            /* done after iLimit test, as candidates could be > iend-8 */
+            {   U32 const indexToInsert = curr+2;
+                hashLong[ZSTD_hashPtr(base+indexToInsert, hBitsL, 8)] = indexToInsert;
+                hashLong[ZSTD_hashPtr(ip-2, hBitsL, 8)] = (U32)(ip-2-base);
+                hashSmall[ZSTD_hashPtr(base+indexToInsert, hBitsS, mls)] = indexToInsert;
+                hashSmall[ZSTD_hashPtr(ip-1, hBitsS, mls)] = (U32)(ip-1-base);
+            }
+
+            /* check immediate repcode */
+            while ( (ip <= ilimit)
+                 && ( (offset_2>0)
+                    & (MEM_read32(ip) == MEM_read32(ip - offset_2)) )) {
+                /* store sequence */
+                size_t const rLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4;
+                U32 const tmpOff = offset_2; offset_2 = offset_1; offset_1 = tmpOff;  /* swap offset_2 <=> offset_1 */
+                hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = (U32)(ip-base);
+                hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = (U32)(ip-base);
+                ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, rLength-MINMATCH);
+                ip += rLength;
+                anchor = ip;
+                continue;   /* faster when present ... (?) */
+            }
+        }
+    }
+}
+
+
+FORCE_INLINE_TEMPLATE
+size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize,
-        U32 const mls /* template */, ZSTD_dictMode_e const dictMode)
+        U32 const mls /* template */)
 {
     ZSTD_compressionParameters const* cParams = &ms->cParams;
     U32* const hashLong = ms->hashTable;
@@ -72,54 +278,30 @@ size_t ZSTD_compressBlock_doubleFast_generic(
     U32 offsetSaved = 0;
 
     const ZSTD_matchState_t* const dms = ms->dictMatchState;
-    const ZSTD_compressionParameters* const dictCParams =
+    const ZSTD_compressionParameters* const dictCParams = &dms->cParams;
-                                     dictMode == ZSTD_dictMatchState ?
+    const U32* const dictHashLong  = dms->hashTable;
-                                     &dms->cParams : NULL;
+    const U32* const dictHashSmall = dms->chainTable;
-    const U32* const dictHashLong  = dictMode == ZSTD_dictMatchState ?
+    const U32 dictStartIndex       = dms->window.dictLimit;
-                                     dms->hashTable : NULL;
+    const BYTE* const dictBase     = dms->window.base;
-    const U32* const dictHashSmall = dictMode == ZSTD_dictMatchState ?
+    const BYTE* const dictStart    = dictBase + dictStartIndex;
-                                     dms->chainTable : NULL;
+    const BYTE* const dictEnd      = dms->window.nextSrc;
-    const U32 dictStartIndex       = dictMode == ZSTD_dictMatchState ?
+    const U32 dictIndexDelta       = prefixLowestIndex - (U32)(dictEnd - dictBase);
-                                     dms->window.dictLimit : 0;
+    const U32 dictHBitsL           = dictCParams->hashLog;
-    const BYTE* const dictBase     = dictMode == ZSTD_dictMatchState ?
+    const U32 dictHBitsS           = dictCParams->chainLog;
-                                     dms->window.base : NULL;
-    const BYTE* const dictStart    = dictMode == ZSTD_dictMatchState ?
-                                     dictBase + dictStartIndex : NULL;
-    const BYTE* const dictEnd      = dictMode == ZSTD_dictMatchState ?
-                                     dms->window.nextSrc : NULL;
-    const U32 dictIndexDelta       = dictMode == ZSTD_dictMatchState ?
-                                     prefixLowestIndex - (U32)(dictEnd - dictBase) :
-                                     0;
-    const U32 dictHBitsL           = dictMode == ZSTD_dictMatchState ?
-                                     dictCParams->hashLog : hBitsL;
-    const U32 dictHBitsS           = dictMode == ZSTD_dictMatchState ?
-                                     dictCParams->chainLog : hBitsS;
     const U32 dictAndPrefixLength  = (U32)((ip - prefixLowest) + (dictEnd - dictStart));
 
-    DEBUGLOG(5, "ZSTD_compressBlock_doubleFast_generic");
+    DEBUGLOG(5, "ZSTD_compressBlock_doubleFast_dictMatchState_generic");
-
-    assert(dictMode == ZSTD_noDict || dictMode == ZSTD_dictMatchState);
 
     /* if a dictionary is attached, it must be within window range */
-    if (dictMode == ZSTD_dictMatchState) {
     assert(ms->window.dictLimit + (1U << cParams->windowLog) >= endIndex);
-    }
 
     /* init */
     ip += (dictAndPrefixLength == 0);
-    if (dictMode == ZSTD_noDict) {
+
-        U32 const curr = (U32)(ip - base);
-        U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, curr, cParams->windowLog);
-        U32 const maxRep = curr - windowLow;
-        if (offset_2 > maxRep) offsetSaved = offset_2, offset_2 = 0;
-        if (offset_1 > maxRep) offsetSaved = offset_1, offset_1 = 0;
-    }
-    if (dictMode == ZSTD_dictMatchState) {
     /* dictMatchState repCode checks don't currently handle repCode == 0
      * disabling. */
     assert(offset_1 <= dictAndPrefixLength);
     assert(offset_2 <= dictAndPrefixLength);
-    }
 
     /* Main Search Loop */
     while (ip < ilimit) {   /* < instead of <=, because repcode check at (ip+1) */
@@ -135,15 +317,13 @@ size_t ZSTD_compressBlock_doubleFast_generic(
         const BYTE* matchLong = base + matchIndexL;
         const BYTE* match = base + matchIndexS;
         const U32 repIndex = curr + 1 - offset_1;
-        const BYTE* repMatch = (dictMode == ZSTD_dictMatchState
+        const BYTE* repMatch = (repIndex < prefixLowestIndex) ?
-                            && repIndex < prefixLowestIndex) ?
                                dictBase + (repIndex - dictIndexDelta) :
                                base + repIndex;
         hashLong[h2] = hashSmall[h] = curr;   /* update hash tables */
 
-        /* check dictMatchState repcode */
+        /* check repcode */
-        if (dictMode == ZSTD_dictMatchState
+        if (((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
-            && ((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
             && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
             const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
             mLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
@@ -152,15 +332,6 @@ size_t ZSTD_compressBlock_doubleFast_generic(
             goto _match_stored;
         }
 
-        /* check noDict repcode */
-        if ( dictMode == ZSTD_noDict
-          && ((offset_1 > 0) & (MEM_read32(ip+1-offset_1) == MEM_read32(ip+1)))) {
-            mLength = ZSTD_count(ip+1+4, ip+1+4-offset_1, iend) + 4;
-            ip++;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, 0, mLength-MINMATCH);
-            goto _match_stored;
-        }
-
         if (matchIndexL > prefixLowestIndex) {
             /* check prefix long match */
             if (MEM_read64(matchLong) == MEM_read64(ip)) {
@@ -169,7 +340,7 @@ size_t ZSTD_compressBlock_doubleFast_generic(
                 while (((ip>anchor) & (matchLong>prefixLowest)) && (ip[-1] == matchLong[-1])) { ip--; matchLong--; mLength++; } /* catch up */
                 goto _match_found;
             }
-        } else if (dictMode == ZSTD_dictMatchState) {
+        } else {
             /* check dictMatchState long match */
             U32 const dictMatchIndexL = dictHashLong[dictHL];
             const BYTE* dictMatchL = dictBase + dictMatchIndexL;
@@ -187,7 +358,7 @@ size_t ZSTD_compressBlock_doubleFast_generic(
             if (MEM_read32(match) == MEM_read32(ip)) {
                 goto _search_next_long;
             }
-        } else if (dictMode == ZSTD_dictMatchState) {
+        } else {
             /* check dictMatchState short match */
             U32 const dictMatchIndexS = dictHashSmall[dictHS];
             match = dictBase + dictMatchIndexS;
@@ -220,7 +391,7 @@ _search_next_long:
                     while (((ip>anchor) & (matchL3>prefixLowest)) && (ip[-1] == matchL3[-1])) { ip--; matchL3--; mLength++; } /* catch up */
                     goto _match_found;
                 }
-            } else if (dictMode == ZSTD_dictMatchState) {
+            } else {
                 /* check dict long +1 match */
                 U32 const dictMatchIndexL3 = dictHashLong[dictHLNext];
                 const BYTE* dictMatchL3 = dictBase + dictMatchIndexL3;
@@ -234,7 +405,7 @@ _search_next_long:
         }   }   }
 
         /* if no long +1 match, explore the short match we found */
-        if (dictMode == ZSTD_dictMatchState && matchIndexS < prefixLowestIndex) {
+        if (matchIndexS < prefixLowestIndex) {
             mLength = ZSTD_count_2segments(ip+4, match+4, iend, dictEnd, prefixLowest) + 4;
             offset = (U32)(curr - matchIndexS);
             while (((ip>anchor) & (match>dictStart)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; } /* catch up */
@@ -244,8 +415,6 @@ _search_next_long:
             while (((ip>anchor) & (match>prefixLowest)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; } /* catch up */
         }
 
-        /* fall-through */
-
 _match_found:
         offset_2 = offset_1;
         offset_1 = offset;
@@ -268,12 +437,10 @@ _match_stored:
             }
 
             /* check immediate repcode */
-            if (dictMode == ZSTD_dictMatchState) {
             while (ip <= ilimit) {
                 U32 const current2 = (U32)(ip-base);
                 U32 const repIndex2 = current2 - offset_2;
-                    const BYTE* repMatch2 = dictMode == ZSTD_dictMatchState
+                const BYTE* repMatch2 = repIndex2 < prefixLowestIndex ?
-                        && repIndex2 < prefixLowestIndex ?
                         dictBase + repIndex2 - dictIndexDelta :
                         base + repIndex2;
                 if ( ((U32)((prefixLowestIndex-1) - (U32)repIndex2) >= 3 /* intentional overflow */)
@@ -289,22 +456,8 @@ _match_stored:
                     continue;
                 }
                 break;
-            }   }
+            }
-
+        }
-            if (dictMode == ZSTD_noDict) {
-                while ( (ip <= ilimit)
-                     && ( (offset_2>0)
-                        & (MEM_read32(ip) == MEM_read32(ip - offset_2)) )) {
-                    /* store sequence */
-                    size_t const rLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4;
-                    U32 const tmpOff = offset_2; offset_2 = offset_1; offset_1 = tmpOff;  /* swap offset_2 <=> offset_1 */
-                    hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = (U32)(ip-base);
-                    hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = (U32)(ip-base);
-                    ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, rLength-MINMATCH);
-                    ip += rLength;
-                    anchor = ip;
-                    continue;   /* faster when present ... (?) */
-        }   }   }
     }   /* while (ip < ilimit) */
 
     /* save reps for next block */
@@ -315,6 +468,24 @@ _match_stored:
     return (size_t)(iend - anchor);
 }
 
+#define ZSTD_GEN_DFAST_FN(dictMode, mls)                                                                 \
+    static size_t ZSTD_compressBlock_doubleFast_##dictMode##_##mls(                                      \
+            ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],                          \
+            void const* src, size_t srcSize)                                                             \
+    {                                                                                                    \
+        return ZSTD_compressBlock_doubleFast_##dictMode##_generic(ms, seqStore, rep, src, srcSize, mls); \
+    }
+
+ZSTD_GEN_DFAST_FN(noDict, 4)
+ZSTD_GEN_DFAST_FN(noDict, 5)
+ZSTD_GEN_DFAST_FN(noDict, 6)
+ZSTD_GEN_DFAST_FN(noDict, 7)
+
+ZSTD_GEN_DFAST_FN(dictMatchState, 4)
+ZSTD_GEN_DFAST_FN(dictMatchState, 5)
+ZSTD_GEN_DFAST_FN(dictMatchState, 6)
+ZSTD_GEN_DFAST_FN(dictMatchState, 7)
+
 
 size_t ZSTD_compressBlock_doubleFast(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
@@ -325,13 +496,13 @@ size_t ZSTD_compressBlock_doubleFast(
     {
     default: /* includes case 3 */
     case 4 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 4, ZSTD_noDict);
+        return ZSTD_compressBlock_doubleFast_noDict_4(ms, seqStore, rep, src, srcSize);
     case 5 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 5, ZSTD_noDict);
+        return ZSTD_compressBlock_doubleFast_noDict_5(ms, seqStore, rep, src, srcSize);
     case 6 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 6, ZSTD_noDict);
+        return ZSTD_compressBlock_doubleFast_noDict_6(ms, seqStore, rep, src, srcSize);
     case 7 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 7, ZSTD_noDict);
+        return ZSTD_compressBlock_doubleFast_noDict_7(ms, seqStore, rep, src, srcSize);
     }
 }
 
@@ -345,13 +516,13 @@ size_t ZSTD_compressBlock_doubleFast_dictMatchState(
     {
     default: /* includes case 3 */
     case 4 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 4, ZSTD_dictMatchState);
+        return ZSTD_compressBlock_doubleFast_dictMatchState_4(ms, seqStore, rep, src, srcSize);
     case 5 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 5, ZSTD_dictMatchState);
+        return ZSTD_compressBlock_doubleFast_dictMatchState_5(ms, seqStore, rep, src, srcSize);
     case 6 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 6, ZSTD_dictMatchState);
+        return ZSTD_compressBlock_doubleFast_dictMatchState_6(ms, seqStore, rep, src, srcSize);
     case 7 :
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, 7, ZSTD_dictMatchState);
+        return ZSTD_compressBlock_doubleFast_dictMatchState_7(ms, seqStore, rep, src, srcSize);
     }
 }
 
@@ -387,7 +558,7 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
 
     /* if extDict is invalidated due to maxDistance, switch to "regular" variant */
     if (prefixStartIndex == dictStartIndex)
-        return ZSTD_compressBlock_doubleFast_generic(ms, seqStore, rep, src, srcSize, mls, ZSTD_noDict);
+        return ZSTD_compressBlock_doubleFast(ms, seqStore, rep, src, srcSize);
 
     /* Search Loop */
     while (ip < ilimit) {  /* < instead of <=, because (ip+1) */
@@ -409,7 +580,7 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
         hashSmall[hSmall] = hashLong[hLong] = curr;   /* update hash table */
 
         if ((((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow : ensure repIndex doesn't overlap dict + prefix */
-            & (offset_1 < curr+1 - dictStartIndex)) /* note: we are searching at curr+1 */
+            & (offset_1 <= curr+1 - dictStartIndex)) /* note: we are searching at curr+1 */
           && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
             const BYTE* repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
             mLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixStart) + 4;
@@ -477,7 +648,7 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
                 U32 const repIndex2 = current2 - offset_2;
                 const BYTE* repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2;
                 if ( (((U32)((prefixStartIndex-1) - repIndex2) >= 3)   /* intentional overflow : ensure repIndex2 doesn't overlap dict + prefix */
-                    & (offset_2 < current2 - dictStartIndex))
+                    & (offset_2 <= current2 - dictStartIndex))
                   && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
                     const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
                     size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
@@ -500,6 +671,10 @@ static size_t ZSTD_compressBlock_doubleFast_extDict_generic(
     return (size_t)(iend - anchor);
 }
 
+ZSTD_GEN_DFAST_FN(extDict, 4)
+ZSTD_GEN_DFAST_FN(extDict, 5)
+ZSTD_GEN_DFAST_FN(extDict, 6)
+ZSTD_GEN_DFAST_FN(extDict, 7)
 
 size_t ZSTD_compressBlock_doubleFast_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
@@ -510,12 +685,12 @@ size_t ZSTD_compressBlock_doubleFast_extDict(
     {
     default: /* includes case 3 */
     case 4 :
-        return ZSTD_compressBlock_doubleFast_extDict_generic(ms, seqStore, rep, src, srcSize, 4);
+        return ZSTD_compressBlock_doubleFast_extDict_4(ms, seqStore, rep, src, srcSize);
     case 5 :
-        return ZSTD_compressBlock_doubleFast_extDict_generic(ms, seqStore, rep, src, srcSize, 5);
+        return ZSTD_compressBlock_doubleFast_extDict_5(ms, seqStore, rep, src, srcSize);
     case 6 :
-        return ZSTD_compressBlock_doubleFast_extDict_generic(ms, seqStore, rep, src, srcSize, 6);
+        return ZSTD_compressBlock_doubleFast_extDict_6(ms, seqStore, rep, src, srcSize);
     case 7 :
-        return ZSTD_compressBlock_doubleFast_extDict_generic(ms, seqStore, rep, src, srcSize, 7);
+        return ZSTD_compressBlock_doubleFast_extDict_7(ms, seqStore, rep, src, srcSize);
     }
 }

zstd/compress/zstd_fast.c

@@ -43,145 +43,294 @@ void ZSTD_fillHashTable(ZSTD_matchState_t* ms,
 }
 
 
+/**
+ * If you squint hard enough (and ignore repcodes), the search operation at any
+ * given position is broken into 4 stages:
+ *
+ * 1. Hash   (map position to hash value via input read)
+ * 2. Lookup (map hash val to index via hashtable read)
+ * 3. Load   (map index to value at that position via input read)
+ * 4. Compare
+ *
+ * Each of these steps involves a memory read at an address which is computed
+ * from the previous step. This means these steps must be sequenced and their
+ * latencies are cumulative.
+ *
+ * Rather than do 1->2->3->4 sequentially for a single position before moving
+ * onto the next, this implementation interleaves these operations across the
+ * next few positions:
+ *
+ * R = Repcode Read & Compare
+ * H = Hash
+ * T = Table Lookup
+ * M = Match Read & Compare
+ *
+ * Pos | Time -->
+ * ----+-------------------
+ * N   | ... M
+ * N+1 | ...   TM
+ * N+2 |    R H   T M
+ * N+3 |         H    TM
+ * N+4 |           R H   T M
+ * N+5 |                H   ...
+ * N+6 |                  R ...
+ *
+ * This is very much analogous to the pipelining of execution in a CPU. And just
+ * like a CPU, we have to dump the pipeline when we find a match (i.e., take a
+ * branch).
+ *
+ * When this happens, we throw away our current state, and do the following prep
+ * to re-enter the loop:
+ *
+ * Pos | Time -->
+ * ----+-------------------
+ * N   | H T
+ * N+1 |  H
+ *
+ * This is also the work we do at the beginning to enter the loop initially.
+ */
 FORCE_INLINE_TEMPLATE size_t
-ZSTD_compressBlock_fast_generic(
+ZSTD_compressBlock_fast_noDict_generic(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize,
-        U32 const mls)
+        U32 const mls, U32 const hasStep)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
     U32* const hashTable = ms->hashTable;
     U32 const hlog = cParams->hashLog;
     /* support stepSize of 0 */
-    size_t const stepSize = cParams->targetLength + !(cParams->targetLength) + 1;
+    size_t const stepSize = hasStep ? (cParams->targetLength + !(cParams->targetLength) + 1) : 2;
     const BYTE* const base = ms->window.base;
     const BYTE* const istart = (const BYTE*)src;
-    /* We check ip0 (ip + 0) and ip1 (ip + 1) each loop */
-    const BYTE* ip0 = istart;
-    const BYTE* ip1;
-    const BYTE* anchor = istart;
     const U32   endIndex = (U32)((size_t)(istart - base) + srcSize);
     const U32   prefixStartIndex = ZSTD_getLowestPrefixIndex(ms, endIndex, cParams->windowLog);
     const BYTE* const prefixStart = base + prefixStartIndex;
     const BYTE* const iend = istart + srcSize;
     const BYTE* const ilimit = iend - HASH_READ_SIZE;
-    U32 offset_1=rep[0], offset_2=rep[1];
+
+    const BYTE* anchor = istart;
+    const BYTE* ip0 = istart;
+    const BYTE* ip1;
+    const BYTE* ip2;
+    const BYTE* ip3;
+    U32 current0;
+
+    U32 rep_offset1 = rep[0];
+    U32 rep_offset2 = rep[1];
     U32 offsetSaved = 0;
 
-    /* init */
+    size_t hash0; /* hash for ip0 */
+    size_t hash1; /* hash for ip1 */
+    U32 idx; /* match idx for ip0 */
+    U32 mval; /* src value at match idx */
+
+    U32 offcode;
+    const BYTE* match0;
+    size_t mLength;
+
+    /* ip0 and ip1 are always adjacent. The targetLength skipping and
+     * uncompressibility acceleration is applied to every other position,
+     * matching the behavior of #1562. step therefore represents the gap
+     * between pairs of positions, from ip0 to ip2 or ip1 to ip3. */
+    size_t step;
+    const BYTE* nextStep;
+    const size_t kStepIncr = (1 << (kSearchStrength - 1));
+
     DEBUGLOG(5, "ZSTD_compressBlock_fast_generic");
     ip0 += (ip0 == prefixStart);
-    ip1 = ip0 + 1;
     {   U32 const curr = (U32)(ip0 - base);
         U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, curr, cParams->windowLog);
         U32 const maxRep = curr - windowLow;
-        if (offset_2 > maxRep) offsetSaved = offset_2, offset_2 = 0;
+        if (rep_offset2 > maxRep) offsetSaved = rep_offset2, rep_offset2 = 0;
-        if (offset_1 > maxRep) offsetSaved = offset_1, offset_1 = 0;
+        if (rep_offset1 > maxRep) offsetSaved = rep_offset1, rep_offset1 = 0;
     }
 
-    /* Main Search Loop */
+    /* start each op */
-#ifdef __INTEL_COMPILER
+_start: /* Requires: ip0 */
-    /* From intel 'The vector pragma indicates that the loop should be
-     * vectorized if it is legal to do so'. Can be used together with
-     * #pragma ivdep (but have opted to exclude that because intel
-     * warns against using it).*/
-    #pragma vector always
-#endif
-    while (ip1 < ilimit) {   /* < instead of <=, because check at ip0+2 */
-        size_t mLength;
-        BYTE const* ip2 = ip0 + 2;
-        size_t const h0 = ZSTD_hashPtr(ip0, hlog, mls);
-        U32 const val0 = MEM_read32(ip0);
-        size_t const h1 = ZSTD_hashPtr(ip1, hlog, mls);
-        U32 const val1 = MEM_read32(ip1);
-        U32 const current0 = (U32)(ip0-base);
-        U32 const current1 = (U32)(ip1-base);
-        U32 const matchIndex0 = hashTable[h0];
-        U32 const matchIndex1 = hashTable[h1];
-        BYTE const* repMatch = ip2 - offset_1;
-        const BYTE* match0 = base + matchIndex0;
-        const BYTE* match1 = base + matchIndex1;
-        U32 offcode;
 
-#if defined(__aarch64__)
+    step = stepSize;
-        PREFETCH_L1(ip0+256);
+    nextStep = ip0 + kStepIncr;
-#endif
 
-        hashTable[h0] = current0;   /* update hash table */
+    /* calculate positions, ip0 - anchor == 0, so we skip step calc */
-        hashTable[h1] = current1;   /* update hash table */
+    ip1 = ip0 + 1;
+    ip2 = ip0 + step;
+    ip3 = ip2 + 1;
 
-        assert(ip0 + 1 == ip1);
+    if (ip3 >= ilimit) {
+        goto _cleanup;
+    }
 
-        if ((offset_1 > 0) & (MEM_read32(repMatch) == MEM_read32(ip2))) {
+    hash0 = ZSTD_hashPtr(ip0, hlog, mls);
-            mLength = (ip2[-1] == repMatch[-1]) ? 1 : 0;
+    hash1 = ZSTD_hashPtr(ip1, hlog, mls);
-            ip0 = ip2 - mLength;
+
-            match0 = repMatch - mLength;
+    idx = hashTable[hash0];
-            mLength += 4;
+
+    do {
+        /* load repcode match for ip[2]*/
+        const U32 rval = MEM_read32(ip2 - rep_offset1);
+
+        /* write back hash table entry */
+        current0 = (U32)(ip0 - base);
+        hashTable[hash0] = current0;
+
+        /* check repcode at ip[2] */
+        if ((MEM_read32(ip2) == rval) & (rep_offset1 > 0)) {
+            ip0 = ip2;
+            match0 = ip0 - rep_offset1;
+            mLength = ip0[-1] == match0[-1];
+            ip0 -= mLength;
+            match0 -= mLength;
             offcode = 0;
+            mLength += 4;
             goto _match;
         }
-        if ((matchIndex0 > prefixStartIndex) && MEM_read32(match0) == val0) {
+
-            /* found a regular match */
+        /* load match for ip[0] */
+        if (idx >= prefixStartIndex) {
+            mval = MEM_read32(base + idx);
+        } else {
+            mval = MEM_read32(ip0) ^ 1; /* guaranteed to not match. */
+        }
+
+        /* check match at ip[0] */
+        if (MEM_read32(ip0) == mval) {
+            /* found a match! */
             goto _offset;
         }
-        if ((matchIndex1 > prefixStartIndex) && MEM_read32(match1) == val1) {
+
-            /* found a regular match after one literal */
+        /* lookup ip[1] */
+        idx = hashTable[hash1];
+
+        /* hash ip[2] */
+        hash0 = hash1;
+        hash1 = ZSTD_hashPtr(ip2, hlog, mls);
+
+        /* advance to next positions */
         ip0 = ip1;
-            match0 = match1;
+        ip1 = ip2;
+        ip2 = ip3;
+
+        /* write back hash table entry */
+        current0 = (U32)(ip0 - base);
+        hashTable[hash0] = current0;
+
+        /* load match for ip[0] */
+        if (idx >= prefixStartIndex) {
+            mval = MEM_read32(base + idx);
+        } else {
+            mval = MEM_read32(ip0) ^ 1; /* guaranteed to not match. */
+        }
+
+        /* check match at ip[0] */
+        if (MEM_read32(ip0) == mval) {
+            /* found a match! */
             goto _offset;
         }
-        {   size_t const step = ((size_t)(ip0-anchor) >> (kSearchStrength - 1)) + stepSize;
+
-            assert(step >= 2);
+        /* lookup ip[1] */
-            ip0 += step;
+        idx = hashTable[hash1];
-            ip1 += step;
+
-            continue;
+        /* 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;
         }
-_offset: /* Requires: ip0, match0 */
+    } while (ip3 < ilimit);
-        /* Compute the offset code */
+
-        offset_2 = offset_1;
+_cleanup:
-        offset_1 = (U32)(ip0-match0);
+    /* Note that there are probably still a couple positions we could search.
-        offcode = offset_1 + ZSTD_REP_MOVE;
+     * However, it seems to be a meaningful performance hit to try to search
+     * them. So let's not. */
+
+    /* save reps for next block */
+    rep[0] = rep_offset1 ? rep_offset1 : offsetSaved;
+    rep[1] = rep_offset2 ? rep_offset2 : offsetSaved;
+
+    /* Return the last literals size */
+    return (size_t)(iend - anchor);
+
+_offset: /* Requires: ip0, idx */
+
+    /* Compute the offset code. */
+    match0 = base + idx;
+    rep_offset2 = rep_offset1;
+    rep_offset1 = (U32)(ip0-match0);
+    offcode = rep_offset1 + ZSTD_REP_MOVE;
     mLength = 4;
-        /* Count the backwards match length */
+
-        while (((ip0>anchor) & (match0>prefixStart))
+    /* Count the backwards match length. */
-             && (ip0[-1] == match0[-1])) { ip0--; match0--; mLength++; } /* catch up */
+    while (((ip0>anchor) & (match0>prefixStart)) && (ip0[-1] == match0[-1])) {
+        ip0--;
+        match0--;
+        mLength++;
+    }
 
 _match: /* Requires: ip0, match0, offcode */
-        /* Count the forward length */
+
+    /* Count the forward length. */
     mLength += ZSTD_count(ip0 + mLength, match0 + mLength, iend);
+
     ZSTD_storeSeq(seqStore, (size_t)(ip0 - anchor), anchor, iend, offcode, mLength - MINMATCH);
-        /* match found */
+
     ip0 += mLength;
     anchor = ip0;
 
+    /* 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);
 
-            if (offset_2 > 0) { /* offset_2==0 means offset_2 is invalidated */
+        if (rep_offset2 > 0) { /* rep_offset2==0 means rep_offset2 is invalidated */
-                while ( (ip0 <= ilimit) && (MEM_read32(ip0) == MEM_read32(ip0 - offset_2)) ) {
+            while ( (ip0 <= ilimit) && (MEM_read32(ip0) == MEM_read32(ip0 - rep_offset2)) ) {
                 /* store sequence */
-                    size_t const rLength = ZSTD_count(ip0+4, ip0+4-offset_2, iend) + 4;
+                size_t const rLength = ZSTD_count(ip0+4, ip0+4-rep_offset2, iend) + 4;
-                    { U32 const tmpOff = offset_2; offset_2 = offset_1; offset_1 = tmpOff; } /* swap offset_2 <=> offset_1 */
+                { 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, 0 /*offCode*/, rLength-MINMATCH);
                 anchor = ip0;
                 continue;   /* faster when present (confirmed on gcc-8) ... (?) */
     }   }   }
-        ip1 = ip0 + 1;
+
+    goto _start;
 }
 
-    /* save reps for next block */
+#define ZSTD_GEN_FAST_FN(dictMode, mls, step)                                                            \
-    rep[0] = offset_1 ? offset_1 : offsetSaved;
+    static size_t ZSTD_compressBlock_fast_##dictMode##_##mls##_##step(                                      \
-    rep[1] = offset_2 ? offset_2 : offsetSaved;
+            ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],                    \
-
+            void const* src, size_t srcSize)                                                       \
-    /* Return the last literals size */
+    {                                                                                              \
-    return (size_t)(iend - anchor);
+        return ZSTD_compressBlock_fast_##dictMode##_generic(ms, seqStore, rep, src, srcSize, mls, step); \
     }
 
+ZSTD_GEN_FAST_FN(noDict, 4, 1)
+ZSTD_GEN_FAST_FN(noDict, 5, 1)
+ZSTD_GEN_FAST_FN(noDict, 6, 1)
+ZSTD_GEN_FAST_FN(noDict, 7, 1)
+
+ZSTD_GEN_FAST_FN(noDict, 4, 0)
+ZSTD_GEN_FAST_FN(noDict, 5, 0)
+ZSTD_GEN_FAST_FN(noDict, 6, 0)
+ZSTD_GEN_FAST_FN(noDict, 7, 0)
 
 size_t ZSTD_compressBlock_fast(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
@@ -189,24 +338,40 @@ size_t ZSTD_compressBlock_fast(
 {
     U32 const mls = ms->cParams.minMatch;
     assert(ms->dictMatchState == NULL);
+    if (ms->cParams.targetLength > 1) {
         switch(mls)
         {
         default: /* includes case 3 */
         case 4 :
-        return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, 4);
+            return ZSTD_compressBlock_fast_noDict_4_1(ms, seqStore, rep, src, srcSize);
         case 5 :
-        return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, 5);
+            return ZSTD_compressBlock_fast_noDict_5_1(ms, seqStore, rep, src, srcSize);
         case 6 :
-        return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, 6);
+            return ZSTD_compressBlock_fast_noDict_6_1(ms, seqStore, rep, src, srcSize);
         case 7 :
-        return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, 7);
+            return ZSTD_compressBlock_fast_noDict_7_1(ms, seqStore, rep, src, srcSize);
+        }
+    } else {
+        switch(mls)
+        {
+        default: /* includes case 3 */
+        case 4 :
+            return ZSTD_compressBlock_fast_noDict_4_0(ms, seqStore, rep, src, srcSize);
+        case 5 :
+            return ZSTD_compressBlock_fast_noDict_5_0(ms, seqStore, rep, src, srcSize);
+        case 6 :
+            return ZSTD_compressBlock_fast_noDict_6_0(ms, seqStore, rep, src, srcSize);
+        case 7 :
+            return ZSTD_compressBlock_fast_noDict_7_0(ms, seqStore, rep, src, srcSize);
+        }
+
     }
 }
 
 FORCE_INLINE_TEMPLATE
 size_t ZSTD_compressBlock_fast_dictMatchState_generic(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize, U32 const mls)
+        void const* src, size_t srcSize, U32 const mls, U32 const hasStep)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
     U32* const hashTable = ms->hashTable;
@@ -242,6 +407,8 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic(
     assert(endIndex - prefixStartIndex <= maxDistance);
     (void)maxDistance; (void)endIndex;   /* these variables are not used when assert() is disabled */
 
+    (void)hasStep; /* not currently specialized on whether it's accelerated */
+
     /* ensure there will be no underflow
      * when translating a dict index into a local index */
     assert(prefixStartIndex >= (U32)(dictEnd - dictBase));
@@ -351,6 +518,12 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic(
     return (size_t)(iend - anchor);
 }
 
+
+ZSTD_GEN_FAST_FN(dictMatchState, 4, 0)
+ZSTD_GEN_FAST_FN(dictMatchState, 5, 0)
+ZSTD_GEN_FAST_FN(dictMatchState, 6, 0)
+ZSTD_GEN_FAST_FN(dictMatchState, 7, 0)
+
 size_t ZSTD_compressBlock_fast_dictMatchState(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
@@ -361,20 +534,20 @@ size_t ZSTD_compressBlock_fast_dictMatchState(
     {
     default: /* includes case 3 */
     case 4 :
-        return ZSTD_compressBlock_fast_dictMatchState_generic(ms, seqStore, rep, src, srcSize, 4);
+        return ZSTD_compressBlock_fast_dictMatchState_4_0(ms, seqStore, rep, src, srcSize);
     case 5 :
-        return ZSTD_compressBlock_fast_dictMatchState_generic(ms, seqStore, rep, src, srcSize, 5);
+        return ZSTD_compressBlock_fast_dictMatchState_5_0(ms, seqStore, rep, src, srcSize);
     case 6 :
-        return ZSTD_compressBlock_fast_dictMatchState_generic(ms, seqStore, rep, src, srcSize, 6);
+        return ZSTD_compressBlock_fast_dictMatchState_6_0(ms, seqStore, rep, src, srcSize);
     case 7 :
-        return ZSTD_compressBlock_fast_dictMatchState_generic(ms, seqStore, rep, src, srcSize, 7);
+        return ZSTD_compressBlock_fast_dictMatchState_7_0(ms, seqStore, rep, src, srcSize);
     }
 }
 
 
 static size_t ZSTD_compressBlock_fast_extDict_generic(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-        void const* src, size_t srcSize, U32 const mls)
+        void const* src, size_t srcSize, U32 const mls, U32 const hasStep)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
     U32* const hashTable = ms->hashTable;
@@ -398,11 +571,13 @@ static size_t ZSTD_compressBlock_fast_extDict_generic(
     const BYTE* const ilimit = iend - 8;
     U32 offset_1=rep[0], offset_2=rep[1];
 
+    (void)hasStep; /* not currently specialized on whether it's accelerated */
+
     DEBUGLOG(5, "ZSTD_compressBlock_fast_extDict_generic (offset_1=%u)", offset_1);
 
     /* switch to "regular" variant if extDict is invalidated due to maxDistance */
     if (prefixStartIndex == dictStartIndex)
-        return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, mls);
+        return ZSTD_compressBlock_fast(ms, seqStore, rep, src, srcSize);
 
     /* Search Loop */
     while (ip < ilimit) {  /* < instead of <=, because (ip+1) */
@@ -418,7 +593,7 @@ static size_t ZSTD_compressBlock_fast_extDict_generic(
         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 */
+             & (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;
@@ -453,7 +628,7 @@ static size_t ZSTD_compressBlock_fast_extDict_generic(
                 U32 const current2 = (U32)(ip-base);
                 U32 const repIndex2 = current2 - offset_2;
                 const BYTE* const repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2;
-                if ( (((U32)((prefixStartIndex-1) - repIndex2) >= 3) & (offset_2 < curr - dictStartIndex))  /* intentional overflow */
+                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;
                     size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
@@ -475,6 +650,10 @@ static size_t ZSTD_compressBlock_fast_extDict_generic(
     return (size_t)(iend - anchor);
 }
 
+ZSTD_GEN_FAST_FN(extDict, 4, 0)
+ZSTD_GEN_FAST_FN(extDict, 5, 0)
+ZSTD_GEN_FAST_FN(extDict, 6, 0)
+ZSTD_GEN_FAST_FN(extDict, 7, 0)
 
 size_t ZSTD_compressBlock_fast_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
@@ -485,12 +664,12 @@ size_t ZSTD_compressBlock_fast_extDict(
     {
     default: /* includes case 3 */
     case 4 :
-        return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, 4);
+        return ZSTD_compressBlock_fast_extDict_4_0(ms, seqStore, rep, src, srcSize);
     case 5 :
-        return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, 5);
+        return ZSTD_compressBlock_fast_extDict_5_0(ms, seqStore, rep, src, srcSize);
     case 6 :
-        return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, 6);
+        return ZSTD_compressBlock_fast_extDict_6_0(ms, seqStore, rep, src, srcSize);
     case 7 :
-        return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, 7);
+        return ZSTD_compressBlock_fast_extDict_7_0(ms, seqStore, rep, src, srcSize);
     }
 }

zstd/compress/zstd_ldm.c

@@ -159,12 +159,12 @@ size_t ZSTD_ldm_getTableSize(ldmParams_t params)
     size_t const ldmBucketSize = ((size_t)1) << (params.hashLog - ldmBucketSizeLog);
     size_t const totalSize = ZSTD_cwksp_alloc_size(ldmBucketSize)
                            + ZSTD_cwksp_alloc_size(ldmHSize * sizeof(ldmEntry_t));
-    return params.enableLdm ? totalSize : 0;
+    return params.enableLdm == ZSTD_ps_enable ? totalSize : 0;
 }
 
 size_t ZSTD_ldm_getMaxNbSeq(ldmParams_t params, size_t maxChunkSize)
 {
-    return params.enableLdm ? (maxChunkSize / params.minMatchLength) : 0;
+    return params.enableLdm == ZSTD_ps_enable ? (maxChunkSize / params.minMatchLength) : 0;
 }
 
 /** ZSTD_ldm_getBucket() :
@@ -478,7 +478,7 @@ static size_t ZSTD_ldm_generateSequences_internal(
              */
             if (anchor > ip + hashed) {
                 ZSTD_ldm_gear_reset(&hashState, anchor - minMatchLength, minMatchLength);
-                /* Continue the outter loop at anchor (ip + hashed == anchor). */
+                /* Continue the outer loop at anchor (ip + hashed == anchor). */
                 ip = anchor - hashed;
                 break;
             }
@@ -657,7 +657,7 @@ void ZSTD_ldm_skipRawSeqStoreBytes(rawSeqStore_t* rawSeqStore, size_t nbBytes) {
 
 size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore,
     ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-    ZSTD_useRowMatchFinderMode_e useRowMatchFinder,
+    ZSTD_paramSwitch_e useRowMatchFinder,
     void const* src, size_t srcSize)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;

zstd/compress/zstd_ldm.h

@@ -66,7 +66,7 @@ size_t ZSTD_ldm_generateSequences(
  */
 size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore,
             ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
-            ZSTD_useRowMatchFinderMode_e useRowMatchFinder,
+            ZSTD_paramSwitch_e useRowMatchFinder,
             void const* src, size_t srcSize);
 
 /**

zstd/compress/zstd_ldm_geartab.h

@@ -11,7 +11,10 @@
 #ifndef ZSTD_LDM_GEARTAB_H
 #define ZSTD_LDM_GEARTAB_H
 
-static U64 ZSTD_ldm_gearTab[256] = {
+#include "../common/compiler.h" /* UNUSED_ATTR */
+#include "../common/mem.h"      /* U64 */
+
+static UNUSED_ATTR const U64 ZSTD_ldm_gearTab[256] = {
     0xf5b8f72c5f77775c, 0x84935f266b7ac412, 0xb647ada9ca730ccc,
     0xb065bb4b114fb1de, 0x34584e7e8c3a9fd0, 0x4e97e17c6ae26b05,
     0x3a03d743bc99a604, 0xcecd042422c4044f, 0x76de76c58524259e,

zstd/compress/zstd_opt.c

@@ -14,7 +14,6 @@
 
 
 #define ZSTD_LITFREQ_ADD    2   /* scaling factor for litFreq, so that frequencies adapt faster to new stats */
-#define ZSTD_FREQ_DIV       4   /* log factor when using previous stats to init next 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 */
@@ -24,11 +23,11 @@
 *  Price functions for optimal parser
 ***************************************/
 
-#if 0    /* approximation at bit level */
+#if 0    /* approximation at bit level (for tests) */
 #  define BITCOST_ACCURACY 0
 #  define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY)
-#  define WEIGHT(stat)  ((void)opt, ZSTD_bitWeight(stat))
+#  define WEIGHT(stat, opt) ((void)opt, ZSTD_bitWeight(stat))
-#elif 0  /* fractional bit accuracy */
+#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))
@@ -66,7 +65,7 @@ MEM_STATIC double ZSTD_fCost(U32 price)
 
 static int ZSTD_compressedLiterals(optState_t const* const optPtr)
 {
-    return optPtr->literalCompressionMode != ZSTD_lcm_uncompressed;
+    return optPtr->literalCompressionMode != ZSTD_ps_disable;
 }
 
 static void ZSTD_setBasePrices(optState_t* optPtr, int optLevel)
@@ -79,25 +78,46 @@ static void ZSTD_setBasePrices(optState_t* optPtr, int optLevel)
 }
 
 
-/* ZSTD_downscaleStat() :
+static U32 sum_u32(const unsigned table[], size_t nbElts)
- * reduce all elements in table by a factor 2^(ZSTD_FREQ_DIV+malus)
+{
- * return the resulting sum of elements */
+    size_t n;
-static U32 ZSTD_downscaleStat(unsigned* table, U32 lastEltIndex, int malus)
+    U32 total = 0;
+    for (n=0; n<nbElts; n++) {
+        total += table[n];
+    }
+    return total;
+}
+
+static U32 ZSTD_downscaleStats(unsigned* table, U32 lastEltIndex, U32 shift)
 {
     U32 s, sum=0;
-    DEBUGLOG(5, "ZSTD_downscaleStat (nbElts=%u)", (unsigned)lastEltIndex+1);
+    DEBUGLOG(5, "ZSTD_downscaleStats (nbElts=%u, shift=%u)", (unsigned)lastEltIndex+1, (unsigned)shift);
-    assert(ZSTD_FREQ_DIV+malus > 0 && ZSTD_FREQ_DIV+malus < 31);
+    assert(shift < 30);
     for (s=0; s<lastEltIndex+1; s++) {
-        table[s] = 1 + (table[s] >> (ZSTD_FREQ_DIV+malus));
+        table[s] = 1 + (table[s] >> shift);
         sum += table[s];
     }
     return sum;
 }
 
+/* ZSTD_scaleStats() :
+ * reduce all elements in table is sum too large
+ * return the resulting sum of elements */
+static U32 ZSTD_scaleStats(unsigned* table, U32 lastEltIndex, U32 logTarget)
+{
+    U32 const prevsum = sum_u32(table, lastEltIndex+1);
+    U32 const factor = prevsum >> 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));
+}
+
 /* ZSTD_rescaleFreqs() :
  * if first block (detected by optPtr->litLengthSum == 0) : init statistics
  *    take hints from dictionary if there is one
- *    or init from zero, using src for literals stats, or flat 1 for match symbols
+ *    and init from zero if there is none,
+ *    using src for literals stats, and baseline stats for sequence symbols
  * otherwise downscale existing stats, to be used as seed for next block.
  */
 static void
@@ -126,7 +146,7 @@ ZSTD_rescaleFreqs(optState_t* const optPtr,
                 optPtr->litSum = 0;
                 for (lit=0; lit<=MaxLit; lit++) {
                     U32 const scaleLog = 11;   /* scale to 2K */
-                    U32 const bitCost = HUF_getNbBits(optPtr->symbolCosts->huf.CTable, lit);
+                    U32 const bitCost = HUF_getNbBitsFromCTable(optPtr->symbolCosts->huf.CTable, lit);
                     assert(bitCost <= scaleLog);
                     optPtr->litFreq[lit] = bitCost ? 1 << (scaleLog-bitCost) : 1 /*minimum to calculate cost*/;
                     optPtr->litSum += optPtr->litFreq[lit];
@@ -174,14 +194,18 @@ ZSTD_rescaleFreqs(optState_t* const optPtr,
             if (compressedLiterals) {
                 unsigned lit = MaxLit;
                 HIST_count_simple(optPtr->litFreq, &lit, src, srcSize);   /* use raw first block to init statistics */
-                optPtr->litSum = ZSTD_downscaleStat(optPtr->litFreq, MaxLit, 1);
+                optPtr->litSum = ZSTD_downscaleStats(optPtr->litFreq, MaxLit, 8);
             }
 
-            {   unsigned ll;
+            {   unsigned const baseLLfreqs[MaxLL+1] = {
-                for (ll=0; ll<=MaxLL; ll++)
+                    4, 2, 1, 1, 1, 1, 1, 1,
-                    optPtr->litLengthFreq[ll] = 1;
+                    1, 1, 1, 1, 1, 1, 1, 1,
+                    1, 1, 1, 1, 1, 1, 1, 1,
+                    1, 1, 1, 1, 1, 1, 1, 1,
+                    1, 1, 1, 1
+                };
+                ZSTD_memcpy(optPtr->litLengthFreq, baseLLfreqs, sizeof(baseLLfreqs)); optPtr->litLengthSum = sum_u32(baseLLfreqs, MaxLL+1);
             }
-            optPtr->litLengthSum = MaxLL+1;
 
             {   unsigned ml;
                 for (ml=0; ml<=MaxML; ml++)
@@ -189,21 +213,25 @@ ZSTD_rescaleFreqs(optState_t* const optPtr,
             }
             optPtr->matchLengthSum = MaxML+1;
 
-            {   unsigned of;
+            {   unsigned const baseOFCfreqs[MaxOff+1] = {
-                for (of=0; of<=MaxOff; of++)
+                    6, 2, 1, 1, 2, 3, 4, 4,
-                    optPtr->offCodeFreq[of] = 1;
+                    4, 3, 2, 1, 1, 1, 1, 1,
+                    1, 1, 1, 1, 1, 1, 1, 1,
+                    1, 1, 1, 1, 1, 1, 1, 1
+                };
+                ZSTD_memcpy(optPtr->offCodeFreq, baseOFCfreqs, sizeof(baseOFCfreqs)); optPtr->offCodeSum = sum_u32(baseOFCfreqs, MaxOff+1);
             }
-            optPtr->offCodeSum = MaxOff+1;
+
 
         }
 
     } else {   /* new block : re-use previous statistics, scaled down */
 
         if (compressedLiterals)
-            optPtr->litSum = ZSTD_downscaleStat(optPtr->litFreq, MaxLit, 1);
+            optPtr->litSum = ZSTD_scaleStats(optPtr->litFreq, MaxLit, 12);
-        optPtr->litLengthSum = ZSTD_downscaleStat(optPtr->litLengthFreq, MaxLL, 0);
+        optPtr->litLengthSum = ZSTD_scaleStats(optPtr->litLengthFreq, MaxLL, 11);
-        optPtr->matchLengthSum = ZSTD_downscaleStat(optPtr->matchLengthFreq, MaxML, 0);
+        optPtr->matchLengthSum = ZSTD_scaleStats(optPtr->matchLengthFreq, MaxML, 11);
-        optPtr->offCodeSum = ZSTD_downscaleStat(optPtr->offCodeFreq, MaxOff, 0);
+        optPtr->offCodeSum = ZSTD_scaleStats(optPtr->offCodeFreq, MaxOff, 11);
     }
 
     ZSTD_setBasePrices(optPtr, optLevel);
@@ -338,7 +366,7 @@ MEM_STATIC U32 ZSTD_readMINMATCH(const void* memPtr, U32 length)
 
 /* Update hashTable3 up to ip (excluded)
    Assumption : always within prefix (i.e. not within extDict) */
-static U32 ZSTD_insertAndFindFirstIndexHash3 (ZSTD_matchState_t* ms,
+static U32 ZSTD_insertAndFindFirstIndexHash3 (const ZSTD_matchState_t* ms,
                                               U32* nextToUpdate3,
                                               const BYTE* const ip)
 {
@@ -364,11 +392,13 @@ static U32 ZSTD_insertAndFindFirstIndexHash3 (ZSTD_matchState_t* ms,
 *  Binary Tree search
 ***************************************/
 /** ZSTD_insertBt1() : add one or multiple positions to tree.
- *  ip : assumed <= iend-8 .
+ * @param ip assumed <= iend-8 .
+ * @param target The target of ZSTD_updateTree_internal() - we are filling to this position
  * @return : nb of positions added */
 static U32 ZSTD_insertBt1(
-                ZSTD_matchState_t* ms,
+                const ZSTD_matchState_t* ms,
                 const BYTE* const ip, const BYTE* const iend,
+                U32 const target,
                 U32 const mls, const int extDict)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
@@ -391,7 +421,10 @@ static U32 ZSTD_insertBt1(
     U32* smallerPtr = bt + 2*(curr&btMask);
     U32* largerPtr  = smallerPtr + 1;
     U32 dummy32;   /* to be nullified at the end */
-    U32 const windowLow = ms->window.lowLimit;
+    /* windowLow is based on target because
+     * we only need positions that will be in the window at the end of the tree update.
+     */
+    U32 const windowLow = ZSTD_getLowestMatchIndex(ms, target, cParams->windowLog);
     U32 matchEndIdx = curr+8+1;
     size_t bestLength = 8;
     U32 nbCompares = 1U << cParams->searchLog;
@@ -404,11 +437,12 @@ static U32 ZSTD_insertBt1(
 
     DEBUGLOG(8, "ZSTD_insertBt1 (%u)", curr);
 
+    assert(curr <= target);
     assert(ip <= iend-8);   /* required for h calculation */
     hashTable[h] = curr;   /* Update Hash Table */
 
     assert(windowLow > 0);
-    while (nbCompares-- && (matchIndex >= windowLow)) {
+    for (; nbCompares && (matchIndex >= windowLow); --nbCompares) {
         U32* const nextPtr = bt + 2*(matchIndex & btMask);
         size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
         assert(matchIndex < curr);
@@ -492,7 +526,7 @@ void ZSTD_updateTree_internal(
                 idx, target, dictMode);
 
     while(idx < target) {
-        U32 const forward = ZSTD_insertBt1(ms, base+idx, iend, mls, dictMode == ZSTD_extDict);
+        U32 const forward = ZSTD_insertBt1(ms, base+idx, iend, target, mls, dictMode == ZSTD_extDict);
         assert(idx < (U32)(idx + forward));
         idx += forward;
     }
@@ -635,11 +669,11 @@ U32 ZSTD_insertBtAndGetAllMatches (
                     return 1;
         }   }   }
         /* no dictMatchState lookup: dicts don't have a populated HC3 table */
-    }
+    }  /* if (mls == 3) */
 
     hashTable[h] = curr;   /* Update Hash Table */
 
-    while (nbCompares-- && (matchIndex >= matchLow)) {
+    for (; nbCompares && (matchIndex >= matchLow); --nbCompares) {
         U32* const nextPtr = bt + 2*(matchIndex & btMask);
         const BYTE* match;
         size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
@@ -672,8 +706,7 @@ U32 ZSTD_insertBtAndGetAllMatches (
                | (ip+matchLength == iLimit) /* equal : no way to know if inf or sup */) {
                 if (dictMode == ZSTD_dictMatchState) nbCompares = 0; /* break should also skip searching dms */
                 break; /* drop, to preserve bt consistency (miss a little bit of compression) */
-            }
+        }   }
-        }
 
         if (match[matchLength] < ip[matchLength]) {
             /* match smaller than current */
@@ -692,12 +725,13 @@ U32 ZSTD_insertBtAndGetAllMatches (
 
     *smallerPtr = *largerPtr = 0;
 
+    assert(nbCompares <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */
     if (dictMode == ZSTD_dictMatchState && nbCompares) {
         size_t const dmsH = ZSTD_hashPtr(ip, dmsHashLog, mls);
         U32 dictMatchIndex = dms->hashTable[dmsH];
         const U32* const dmsBt = dms->chainTable;
         commonLengthSmaller = commonLengthLarger = 0;
-        while (nbCompares-- && (dictMatchIndex > dmsLowLimit)) {
+        for (; nbCompares && (dictMatchIndex > dmsLowLimit); --nbCompares) {
             const U32* const nextPtr = dmsBt + 2*(dictMatchIndex & dmsBtMask);
             size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
             const BYTE* match = dmsBase + dictMatchIndex;
@@ -718,8 +752,7 @@ U32 ZSTD_insertBtAndGetAllMatches (
                 if ( (matchLength > ZSTD_OPT_NUM)
                    | (ip+matchLength == iLimit) /* equal : no way to know if inf or sup */) {
                     break;   /* drop, to guarantee consistency (miss a little bit of compression) */
-                }
+            }   }
-            }
 
             if (dictMatchIndex <= dmsBtLow) { break; }   /* beyond tree size, stop the search */
             if (match[matchLength] < ip[matchLength]) {
@@ -729,39 +762,90 @@ U32 ZSTD_insertBtAndGetAllMatches (
                 /* match is larger than current */
                 commonLengthLarger = matchLength;
                 dictMatchIndex = nextPtr[0];
-            }
+    }   }   }  /* if (dictMode == ZSTD_dictMatchState) */
-        }
-    }
 
     assert(matchEndIdx > curr+8);
     ms->nextToUpdate = matchEndIdx - 8;  /* skip repetitive patterns */
     return mnum;
 }
 
-
+typedef U32 (*ZSTD_getAllMatchesFn)(
-FORCE_INLINE_TEMPLATE U32 ZSTD_BtGetAllMatches (
+    ZSTD_match_t*,
-                        ZSTD_match_t* matches,   /* store result (match found, increasing size) in this table */
+    ZSTD_matchState_t*,
-                        ZSTD_matchState_t* ms,
+    U32*,
-                        U32* nextToUpdate3,
+    const BYTE*,
-                        const BYTE* ip, const BYTE* const iHighLimit, const ZSTD_dictMode_e dictMode,
+    const BYTE*,
     const U32 rep[ZSTD_REP_NUM],
     U32 const ll0,
-                        U32 const lengthToBeat)
+    U32 const lengthToBeat);
+
+FORCE_INLINE_TEMPLATE U32 ZSTD_btGetAllMatches_internal(
+        ZSTD_match_t* matches,
+        ZSTD_matchState_t* ms,
+        U32* nextToUpdate3,
+        const BYTE* ip,
+        const BYTE* const iHighLimit,
+        const U32 rep[ZSTD_REP_NUM],
+        U32 const ll0,
+        U32 const lengthToBeat,
+        const ZSTD_dictMode_e dictMode,
+        const U32 mls)
 {
-    const ZSTD_compressionParameters* const cParams = &ms->cParams;
+    assert(BOUNDED(3, ms->cParams.minMatch, 6) == mls);
-    U32 const matchLengthSearch = cParams->minMatch;
+    DEBUGLOG(8, "ZSTD_BtGetAllMatches(dictMode=%d, mls=%u)", (int)dictMode, mls);
-    DEBUGLOG(8, "ZSTD_BtGetAllMatches");
+    if (ip < ms->window.base + ms->nextToUpdate)
-    if (ip < ms->window.base + ms->nextToUpdate) return 0;   /* skipped area */
+        return 0;   /* skipped area */
-    ZSTD_updateTree_internal(ms, ip, iHighLimit, matchLengthSearch, dictMode);
+    ZSTD_updateTree_internal(ms, ip, iHighLimit, mls, dictMode);
-    switch(matchLengthSearch)
+    return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, mls);
-    {
-    case 3 : return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, 3);
-    default :
-    case 4 : return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, 4);
-    case 5 : return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, 5);
-    case 7 :
-    case 6 : return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, 6);
 }
+
+#define ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, mls) ZSTD_btGetAllMatches_##dictMode##_##mls
+
+#define GEN_ZSTD_BT_GET_ALL_MATCHES_(dictMode, mls)            \
+    static U32 ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, mls)(      \
+            ZSTD_match_t* matches,                             \
+            ZSTD_matchState_t* ms,                             \
+            U32* nextToUpdate3,                                \
+            const BYTE* ip,                                    \
+            const BYTE* const iHighLimit,                      \
+            const U32 rep[ZSTD_REP_NUM],                       \
+            U32 const ll0,                                     \
+            U32 const lengthToBeat)                            \
+    {                                                          \
+        return ZSTD_btGetAllMatches_internal(                  \
+                matches, ms, nextToUpdate3, ip, iHighLimit,    \
+                rep, ll0, lengthToBeat, ZSTD_##dictMode, mls); \
+    }
+
+#define GEN_ZSTD_BT_GET_ALL_MATCHES(dictMode)  \
+    GEN_ZSTD_BT_GET_ALL_MATCHES_(dictMode, 3)  \
+    GEN_ZSTD_BT_GET_ALL_MATCHES_(dictMode, 4)  \
+    GEN_ZSTD_BT_GET_ALL_MATCHES_(dictMode, 5)  \
+    GEN_ZSTD_BT_GET_ALL_MATCHES_(dictMode, 6)
+
+GEN_ZSTD_BT_GET_ALL_MATCHES(noDict)
+GEN_ZSTD_BT_GET_ALL_MATCHES(extDict)
+GEN_ZSTD_BT_GET_ALL_MATCHES(dictMatchState)
+
+#define ZSTD_BT_GET_ALL_MATCHES_ARRAY(dictMode)  \
+    {                                            \
+        ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, 3), \
+        ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, 4), \
+        ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, 5), \
+        ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, 6)  \
+    }
+
+static ZSTD_getAllMatchesFn ZSTD_selectBtGetAllMatches(ZSTD_matchState_t const* ms, ZSTD_dictMode_e const dictMode)
+{
+    ZSTD_getAllMatchesFn const getAllMatchesFns[3][4] = {
+        ZSTD_BT_GET_ALL_MATCHES_ARRAY(noDict),
+        ZSTD_BT_GET_ALL_MATCHES_ARRAY(extDict),
+        ZSTD_BT_GET_ALL_MATCHES_ARRAY(dictMatchState)
+    };
+    U32 const mls = BOUNDED(3, ms->cParams.minMatch, 6);
+    assert((U32)dictMode < 3);
+    assert(mls - 3 < 4);
+    return getAllMatchesFns[(int)dictMode][mls - 3];
 }
 
 /*************************
@@ -899,11 +983,11 @@ static void ZSTD_optLdm_processMatchCandidate(ZSTD_optLdm_t* optLdm, ZSTD_match_
     ZSTD_optLdm_maybeAddMatch(matches, nbMatches, optLdm, currPosInBlock);
 }
 
+
 /*-*******************************
 *  Optimal parser
 *********************************/
 
-
 static U32 ZSTD_totalLen(ZSTD_optimal_t sol)
 {
     return sol.litlen + sol.mlen;
@@ -944,6 +1028,8 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
     const BYTE* const prefixStart = base + ms->window.dictLimit;
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
 
+    ZSTD_getAllMatchesFn getAllMatches = ZSTD_selectBtGetAllMatches(ms, dictMode);
+
     U32 const sufficient_len = MIN(cParams->targetLength, ZSTD_OPT_NUM -1);
     U32 const minMatch = (cParams->minMatch == 3) ? 3 : 4;
     U32 nextToUpdate3 = ms->nextToUpdate;
@@ -971,7 +1057,7 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
         /* find first match */
         {   U32 const litlen = (U32)(ip - anchor);
             U32 const ll0 = !litlen;
-            U32 nbMatches = ZSTD_BtGetAllMatches(matches, ms, &nextToUpdate3, ip, iend, dictMode, rep, ll0, minMatch);
+            U32 nbMatches = getAllMatches(matches, ms, &nextToUpdate3, ip, iend, rep, ll0, minMatch);
             ZSTD_optLdm_processMatchCandidate(&optLdm, matches, &nbMatches,
                                               (U32)(ip-istart), (U32)(iend - ip));
             if (!nbMatches) { ip++; continue; }
@@ -985,7 +1071,7 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
              * in every price. We include the literal length to avoid negative
              * prices when we subtract the previous literal length.
              */
-            opt[0].price = ZSTD_litLengthPrice(litlen, optStatePtr, optLevel);
+            opt[0].price = (int)ZSTD_litLengthPrice(litlen, optStatePtr, optLevel);
 
             /* large match -> immediate encoding */
             {   U32 const maxML = matches[nbMatches-1].len;
@@ -1005,7 +1091,8 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
             }   }
 
             /* set prices for first matches starting position == 0 */
-            {   U32 const literalsPrice = opt[0].price + ZSTD_litLengthPrice(0, optStatePtr, optLevel);
+            assert(opt[0].price >= 0);
+            {   U32 const literalsPrice = (U32)opt[0].price + ZSTD_litLengthPrice(0, optStatePtr, optLevel);
                 U32 pos;
                 U32 matchNb;
                 for (pos = 1; pos < minMatch; pos++) {
@@ -1022,7 +1109,7 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
                         opt[pos].mlen = pos;
                         opt[pos].off = offset;
                         opt[pos].litlen = litlen;
-                        opt[pos].price = sequencePrice;
+                        opt[pos].price = (int)sequencePrice;
                 }   }
                 last_pos = pos-1;
             }
@@ -1037,9 +1124,9 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
             /* Fix current position with one literal if cheaper */
             {   U32 const litlen = (opt[cur-1].mlen == 0) ? opt[cur-1].litlen + 1 : 1;
                 int const price = opt[cur-1].price
-                                + ZSTD_rawLiteralsCost(ip+cur-1, 1, optStatePtr, optLevel)
+                                + (int)ZSTD_rawLiteralsCost(ip+cur-1, 1, optStatePtr, optLevel)
-                                + ZSTD_litLengthPrice(litlen, optStatePtr, optLevel)
+                                + (int)ZSTD_litLengthPrice(litlen, optStatePtr, optLevel)
-                                - ZSTD_litLengthPrice(litlen-1, optStatePtr, optLevel);
+                                - (int)ZSTD_litLengthPrice(litlen-1, optStatePtr, optLevel);
                 assert(price < 1000000000); /* overflow check */
                 if (price <= opt[cur].price) {
                     DEBUGLOG(7, "cPos:%zi==rPos:%u : better price (%.2f<=%.2f) using literal (ll==%u) (hist:%u,%u,%u)",
@@ -1082,11 +1169,12 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
                 continue;  /* skip unpromising positions; about ~+6% speed, -0.01 ratio */
             }
 
+            assert(opt[cur].price >= 0);
             {   U32 const ll0 = (opt[cur].mlen != 0);
                 U32 const litlen = (opt[cur].mlen == 0) ? opt[cur].litlen : 0;
-                U32 const previousPrice = opt[cur].price;
+                U32 const previousPrice = (U32)opt[cur].price;
                 U32 const basePrice = previousPrice + ZSTD_litLengthPrice(0, optStatePtr, optLevel);
-                U32 nbMatches = ZSTD_BtGetAllMatches(matches, ms, &nextToUpdate3, inr, iend, dictMode, opt[cur].rep, ll0, minMatch);
+                U32 nbMatches = getAllMatches(matches, ms, &nextToUpdate3, inr, iend, opt[cur].rep, ll0, minMatch);
                 U32 matchNb;
 
                 ZSTD_optLdm_processMatchCandidate(&optLdm, matches, &nbMatches,
@@ -1124,7 +1212,7 @@ ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
 
                     for (mlen = lastML; mlen >= startML; mlen--) {  /* scan downward */
                         U32 const pos = cur + mlen;
-                        int const price = basePrice + ZSTD_getMatchPrice(offset, mlen, optStatePtr, optLevel);
+                        int const price = (int)basePrice + (int)ZSTD_getMatchPrice(offset, mlen, optStatePtr, optLevel);
 
                         if ((pos > last_pos) || (price < opt[pos].price)) {
                             DEBUGLOG(7, "rPos:%u (ml=%2u) => new better price (%.2f<%.2f)",
@@ -1210,38 +1298,30 @@ _shortestPath:   /* cur, last_pos, best_mlen, best_off have to be set */
     return (size_t)(iend - anchor);
 }
 
+static size_t ZSTD_compressBlock_opt0(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        const void* src, size_t srcSize, const ZSTD_dictMode_e dictMode)
+{
+    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /* optLevel */, dictMode);
+}
+
+static size_t ZSTD_compressBlock_opt2(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        const void* src, size_t srcSize, const ZSTD_dictMode_e dictMode)
+{
+    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /* optLevel */, dictMode);
+}
 
 size_t ZSTD_compressBlock_btopt(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         const void* src, size_t srcSize)
 {
     DEBUGLOG(5, "ZSTD_compressBlock_btopt");
-    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /*optLevel*/, ZSTD_noDict);
+    return ZSTD_compressBlock_opt0(ms, seqStore, rep, src, srcSize, ZSTD_noDict);
 }
 
 
-/* used in 2-pass strategy */
-static U32 ZSTD_upscaleStat(unsigned* table, U32 lastEltIndex, int bonus)
-{
-    U32 s, sum=0;
-    assert(ZSTD_FREQ_DIV+bonus >= 0);
-    for (s=0; s<lastEltIndex+1; s++) {
-        table[s] <<= ZSTD_FREQ_DIV+bonus;
-        table[s]--;
-        sum += table[s];
-    }
-    return sum;
-}
 
-/* used in 2-pass strategy */
-MEM_STATIC void ZSTD_upscaleStats(optState_t* optPtr)
-{
-    if (ZSTD_compressedLiterals(optPtr))
-        optPtr->litSum = ZSTD_upscaleStat(optPtr->litFreq, MaxLit, 0);
-    optPtr->litLengthSum = ZSTD_upscaleStat(optPtr->litLengthFreq, MaxLL, 0);
-    optPtr->matchLengthSum = ZSTD_upscaleStat(optPtr->matchLengthFreq, MaxML, 0);
-    optPtr->offCodeSum = ZSTD_upscaleStat(optPtr->offCodeFreq, MaxOff, 0);
-}
 
 /* ZSTD_initStats_ultra():
  * make a first compression pass, just to seed stats with more accurate starting values.
@@ -1263,7 +1343,7 @@ ZSTD_initStats_ultra(ZSTD_matchState_t* ms,
     assert(ms->window.dictLimit == ms->window.lowLimit);   /* no dictionary */
     assert(ms->window.dictLimit - ms->nextToUpdate <= 1);  /* no prefix (note: intentional overflow, defined as 2-complement) */
 
-    ZSTD_compressBlock_opt_generic(ms, seqStore, tmpRep, src, srcSize, 2 /*optLevel*/, ZSTD_noDict);   /* generate stats into ms->opt*/
+    ZSTD_compressBlock_opt2(ms, seqStore, tmpRep, src, srcSize, ZSTD_noDict);   /* generate stats into ms->opt*/
 
     /* invalidate first scan from history */
     ZSTD_resetSeqStore(seqStore);
@@ -1272,8 +1352,6 @@ ZSTD_initStats_ultra(ZSTD_matchState_t* ms,
     ms->window.lowLimit = ms->window.dictLimit;
     ms->nextToUpdate = ms->window.dictLimit;
 
-    /* re-inforce weight of collected statistics */
-    ZSTD_upscaleStats(&ms->opt);
 }
 
 size_t ZSTD_compressBlock_btultra(
@@ -1281,7 +1359,7 @@ size_t ZSTD_compressBlock_btultra(
         const void* src, size_t srcSize)
 {
     DEBUGLOG(5, "ZSTD_compressBlock_btultra (srcSize=%zu)", srcSize);
-    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_noDict);
+    return ZSTD_compressBlock_opt2(ms, seqStore, rep, src, srcSize, ZSTD_noDict);
 }
 
 size_t ZSTD_compressBlock_btultra2(
@@ -1309,35 +1387,35 @@ size_t ZSTD_compressBlock_btultra2(
         ZSTD_initStats_ultra(ms, seqStore, rep, src, srcSize);
     }
 
-    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_noDict);
+    return ZSTD_compressBlock_opt2(ms, seqStore, rep, src, srcSize, ZSTD_noDict);
 }
 
 size_t ZSTD_compressBlock_btopt_dictMatchState(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         const void* src, size_t srcSize)
 {
-    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /*optLevel*/, ZSTD_dictMatchState);
+    return ZSTD_compressBlock_opt0(ms, seqStore, rep, src, srcSize, ZSTD_dictMatchState);
 }
 
 size_t ZSTD_compressBlock_btultra_dictMatchState(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         const void* src, size_t srcSize)
 {
-    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_dictMatchState);
+    return ZSTD_compressBlock_opt2(ms, seqStore, rep, src, srcSize, ZSTD_dictMatchState);
 }
 
 size_t ZSTD_compressBlock_btopt_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         const void* src, size_t srcSize)
 {
-    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /*optLevel*/, ZSTD_extDict);
+    return ZSTD_compressBlock_opt0(ms, seqStore, rep, src, srcSize, ZSTD_extDict);
 }
 
 size_t ZSTD_compressBlock_btultra_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         const void* src, size_t srcSize)
 {
-    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /*optLevel*/, ZSTD_extDict);
+    return ZSTD_compressBlock_opt2(ms, seqStore, rep, src, srcSize, ZSTD_extDict);
 }
 
 /* note : no btultra2 variant for extDict nor dictMatchState,

zstd/compress/zstdmt_compress.c

@@ -467,7 +467,7 @@ ZSTDMT_serialState_reset(serialState_t* serialState,
                          ZSTD_dictContentType_e dictContentType)
 {
     /* Adjust parameters */
-    if (params.ldmParams.enableLdm) {
+    if (params.ldmParams.enableLdm == ZSTD_ps_enable) {
         DEBUGLOG(4, "LDM window size = %u KB", (1U << params.cParams.windowLog) >> 10);
         ZSTD_ldm_adjustParameters(&params.ldmParams, &params.cParams);
         assert(params.ldmParams.hashLog >= params.ldmParams.bucketSizeLog);
@@ -478,7 +478,7 @@ ZSTDMT_serialState_reset(serialState_t* serialState,
     serialState->nextJobID = 0;
     if (params.fParams.checksumFlag)
         XXH64_reset(&serialState->xxhState, 0);
-    if (params.ldmParams.enableLdm) {
+    if (params.ldmParams.enableLdm == ZSTD_ps_enable) {
         ZSTD_customMem cMem = params.customMem;
         unsigned const hashLog = params.ldmParams.hashLog;
         size_t const hashSize = ((size_t)1 << hashLog) * sizeof(ldmEntry_t);
@@ -564,7 +564,7 @@ static void ZSTDMT_serialState_update(serialState_t* serialState,
     /* A future job may error and skip our job */
     if (serialState->nextJobID == jobID) {
         /* It is now our turn, do any processing necessary */
-        if (serialState->params.ldmParams.enableLdm) {
+        if (serialState->params.ldmParams.enableLdm == ZSTD_ps_enable) {
             size_t error;
             assert(seqStore.seq != NULL && seqStore.pos == 0 &&
                    seqStore.size == 0 && seqStore.capacity > 0);
@@ -594,7 +594,7 @@ static void ZSTDMT_serialState_update(serialState_t* serialState,
     if (seqStore.size > 0) {
         size_t const err = ZSTD_referenceExternalSequences(
             jobCCtx, seqStore.seq, seqStore.size);
-        assert(serialState->params.ldmParams.enableLdm);
+        assert(serialState->params.ldmParams.enableLdm == ZSTD_ps_enable);
         assert(!ZSTD_isError(err));
         (void)err;
     }
@@ -672,7 +672,7 @@ static void ZSTDMT_compressionJob(void* jobDescription)
         if (dstBuff.start==NULL) JOB_ERROR(ERROR(memory_allocation));
         job->dstBuff = dstBuff;   /* this value can be read in ZSTDMT_flush, when it copies the whole job */
     }
-    if (jobParams.ldmParams.enableLdm && rawSeqStore.seq == NULL)
+    if (jobParams.ldmParams.enableLdm == ZSTD_ps_enable && rawSeqStore.seq == NULL)
         JOB_ERROR(ERROR(memory_allocation));
 
     /* Don't compute the checksum for chunks, since we compute it externally,
@@ -680,7 +680,7 @@ static void ZSTDMT_compressionJob(void* jobDescription)
      */
     if (job->jobID != 0) jobParams.fParams.checksumFlag = 0;
     /* Don't run LDM for the chunks, since we handle it externally */
-    jobParams.ldmParams.enableLdm = 0;
+    jobParams.ldmParams.enableLdm = ZSTD_ps_disable;
     /* Correct nbWorkers to 0. */
     jobParams.nbWorkers = 0;
 
@@ -807,6 +807,15 @@ typedef struct {
 static const roundBuff_t kNullRoundBuff = {NULL, 0, 0};
 
 #define RSYNC_LENGTH 32
+/* Don't create chunks smaller than the zstd block size.
+ * This stops us from regressing compression ratio too much,
+ * and ensures our output fits in ZSTD_compressBound().
+ *
+ * If this is shrunk < ZSTD_BLOCKSIZELOG_MIN then
+ * ZSTD_COMPRESSBOUND() will need to be updated.
+ */
+#define RSYNC_MIN_BLOCK_LOG ZSTD_BLOCKSIZELOG_MAX
+#define RSYNC_MIN_BLOCK_SIZE (1<<RSYNC_MIN_BLOCK_LOG)
 
 typedef struct {
   U64 hash;
@@ -1135,7 +1144,7 @@ size_t ZSTDMT_toFlushNow(ZSTDMT_CCtx* mtctx)
 static unsigned ZSTDMT_computeTargetJobLog(const ZSTD_CCtx_params* params)
 {
     unsigned jobLog;
-    if (params->ldmParams.enableLdm) {
+    if (params->ldmParams.enableLdm == ZSTD_ps_enable) {
         /* In Long Range Mode, the windowLog is typically oversized.
          * In which case, it's preferable to determine the jobSize
          * based on cycleLog instead. */
@@ -1179,7 +1188,7 @@ static size_t ZSTDMT_computeOverlapSize(const ZSTD_CCtx_params* params)
     int const overlapRLog = 9 - ZSTDMT_overlapLog(params->overlapLog, params->cParams.strategy);
     int ovLog = (overlapRLog >= 8) ? 0 : (params->cParams.windowLog - overlapRLog);
     assert(0 <= overlapRLog && overlapRLog <= 8);
-    if (params->ldmParams.enableLdm) {
+    if (params->ldmParams.enableLdm == ZSTD_ps_enable) {
         /* In Long Range Mode, the windowLog is typically oversized.
          * In which case, it's preferable to determine the jobSize
          * based on chainLog instead.
@@ -1252,6 +1261,9 @@ size_t ZSTDMT_initCStream_internal(
         /* Aim for the targetsectionSize as the average job size. */
         U32 const jobSizeKB = (U32)(mtctx->targetSectionSize >> 10);
         U32 const rsyncBits = (assert(jobSizeKB >= 1), ZSTD_highbit32(jobSizeKB) + 10);
+        /* We refuse to create jobs < RSYNC_MIN_BLOCK_SIZE bytes, so make sure our
+         * expected job size is at least 4x larger. */
+        assert(rsyncBits >= RSYNC_MIN_BLOCK_LOG + 2);
         DEBUGLOG(4, "rsyncLog = %u", rsyncBits);
         mtctx->rsync.hash = 0;
         mtctx->rsync.hitMask = (1ULL << rsyncBits) - 1;
@@ -1263,7 +1275,7 @@ size_t ZSTDMT_initCStream_internal(
     ZSTDMT_setBufferSize(mtctx->bufPool, ZSTD_compressBound(mtctx->targetSectionSize));
     {
         /* If ldm is enabled we need windowSize space. */
-        size_t const windowSize = mtctx->params.ldmParams.enableLdm ? (1U << mtctx->params.cParams.windowLog) : 0;
+        size_t const windowSize = mtctx->params.ldmParams.enableLdm == ZSTD_ps_enable ? (1U << mtctx->params.cParams.windowLog) : 0;
         /* Two buffers of slack, plus extra space for the overlap
          * This is the minimum slack that LDM works with. One extra because
          * flush might waste up to targetSectionSize-1 bytes. Another extra
@@ -1538,18 +1550,22 @@ static range_t ZSTDMT_getInputDataInUse(ZSTDMT_CCtx* mtctx)
 static int ZSTDMT_isOverlapped(buffer_t buffer, range_t range)
 {
     BYTE const* const bufferStart = (BYTE const*)buffer.start;
-    BYTE const* const bufferEnd = bufferStart + buffer.capacity;
     BYTE const* const rangeStart = (BYTE const*)range.start;
-    BYTE const* const rangeEnd = range.size != 0 ? rangeStart + range.size : rangeStart;
 
     if (rangeStart == NULL || bufferStart == NULL)
         return 0;
+
+    {
+        BYTE const* const bufferEnd = bufferStart + buffer.capacity;
+        BYTE const* const rangeEnd = rangeStart + range.size;
+
         /* Empty ranges cannot overlap */
         if (bufferStart == bufferEnd || rangeStart == rangeEnd)
             return 0;
 
         return bufferStart < rangeEnd && rangeStart < bufferEnd;
     }
+}
 
 static int ZSTDMT_doesOverlapWindow(buffer_t buffer, ZSTD_window_t window)
 {
@@ -1575,7 +1591,7 @@ static int ZSTDMT_doesOverlapWindow(buffer_t buffer, ZSTD_window_t window)
 
 static void ZSTDMT_waitForLdmComplete(ZSTDMT_CCtx* mtctx, buffer_t buffer)
 {
-    if (mtctx->params.ldmParams.enableLdm) {
+    if (mtctx->params.ldmParams.enableLdm == ZSTD_ps_enable) {
         ZSTD_pthread_mutex_t* mutex = &mtctx->serial.ldmWindowMutex;
         DEBUGLOG(5, "ZSTDMT_waitForLdmComplete");
         DEBUGLOG(5, "source  [0x%zx, 0x%zx)",
@@ -1678,6 +1694,11 @@ findSynchronizationPoint(ZSTDMT_CCtx const* mtctx, ZSTD_inBuffer const input)
     if (!mtctx->params.rsyncable)
         /* Rsync is disabled. */
         return syncPoint;
+    if (mtctx->inBuff.filled + input.size - input.pos < RSYNC_MIN_BLOCK_SIZE)
+        /* We don't emit synchronization points if it would produce too small blocks.
+         * We don't have enough input to find a synchronization point, so don't look.
+         */
+        return syncPoint;
     if (mtctx->inBuff.filled + syncPoint.toLoad < RSYNC_LENGTH)
         /* Not enough to compute the hash.
          * We will miss any synchronization points in this RSYNC_LENGTH byte
@@ -1688,10 +1709,28 @@ findSynchronizationPoint(ZSTDMT_CCtx const* mtctx, ZSTD_inBuffer const input)
          */
         return syncPoint;
     /* Initialize the loop variables. */
-    if (mtctx->inBuff.filled >= RSYNC_LENGTH) {
+    if (mtctx->inBuff.filled < RSYNC_MIN_BLOCK_SIZE) {
-        /* We have enough bytes buffered to initialize the hash.
+        /* We don't need to scan the first RSYNC_MIN_BLOCK_SIZE positions
+         * because they can't possibly be a sync point. So we can start
+         * part way through the input buffer.
+         */
+        pos = RSYNC_MIN_BLOCK_SIZE - mtctx->inBuff.filled;
+        if (pos >= RSYNC_LENGTH) {
+            prev = istart + pos - RSYNC_LENGTH;
+            hash = ZSTD_rollingHash_compute(prev, RSYNC_LENGTH);
+        } else {
+            assert(mtctx->inBuff.filled >= RSYNC_LENGTH);
+            prev = (BYTE const*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled - RSYNC_LENGTH;
+            hash = ZSTD_rollingHash_compute(prev + pos, (RSYNC_LENGTH - pos));
+            hash = ZSTD_rollingHash_append(hash, istart, pos);
+        }
+    } else {
+        /* We have enough bytes buffered to initialize the hash,
+         * and are 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);
+        assert(RSYNC_MIN_BLOCK_SIZE >= RSYNC_LENGTH);
         pos = 0;
         prev = (BYTE const*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled - RSYNC_LENGTH;
         hash = ZSTD_rollingHash_compute(prev, RSYNC_LENGTH);
@@ -1705,16 +1744,6 @@ findSynchronizationPoint(ZSTDMT_CCtx const* mtctx, ZSTD_inBuffer const input)
             syncPoint.flush = 1;
             return syncPoint;
         }
-    } else {
-        /* We don't have enough bytes buffered to initialize the hash, but
-         * we know we have at least RSYNC_LENGTH bytes total.
-         * Start scanning after the first RSYNC_LENGTH bytes less the bytes
-         * already buffered.
-         */
-        pos = RSYNC_LENGTH - mtctx->inBuff.filled;
-        prev = (BYTE const*)mtctx->inBuff.buffer.start - pos;
-        hash = ZSTD_rollingHash_compute(mtctx->inBuff.buffer.start, mtctx->inBuff.filled);
-        hash = ZSTD_rollingHash_append(hash, istart, pos);
     }
     /* Starting with the hash of the previous RSYNC_LENGTH bytes, roll
      * through the input. If we hit a synchronization point, then cut the
@@ -1726,8 +1755,9 @@ findSynchronizationPoint(ZSTDMT_CCtx const* mtctx, ZSTD_inBuffer const input)
      */
     for (; pos < syncPoint.toLoad; ++pos) {
         BYTE const toRemove = pos < RSYNC_LENGTH ? prev[pos] : istart[pos - RSYNC_LENGTH];
-        /* if (pos >= RSYNC_LENGTH) assert(ZSTD_rollingHash_compute(istart + pos - RSYNC_LENGTH, RSYNC_LENGTH) == hash); */
+        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;

zstd/decompress/huf_decompress_amd64.S

@@ -0,0 +1,571 @@
+#include "../common/portability_macros.h"
+
+#if ZSTD_ENABLE_ASM_X86_64_BMI2
+
+/* Stack marking
+ * ref: https://wiki.gentoo.org/wiki/Hardened/GNU_stack_quickstart
+ */
+#if defined(__linux__) && defined(__ELF__)
+.section .note.GNU-stack,"",%progbits
+#endif
+
+/* Calling convention:
+ *
+ * %rdi contains the first argument: HUF_DecompressAsmArgs*.
+ * %rbp isn't maintained (no frame pointer).
+ * %rsp contains the stack pointer that grows down.
+ *      No red-zone is assumed, only addresses >= %rsp are used.
+ * All register contents are preserved.
+ *
+ * TODO: Support Windows calling convention.
+ */
+
+.global HUF_decompress4X1_usingDTable_internal_bmi2_asm_loop
+.global HUF_decompress4X2_usingDTable_internal_bmi2_asm_loop
+.global _HUF_decompress4X1_usingDTable_internal_bmi2_asm_loop
+.global _HUF_decompress4X2_usingDTable_internal_bmi2_asm_loop
+.text
+
+/* Sets up register mappings for clarity.
+ * op[], bits[], dtable & ip[0] each get their own register.
+ * ip[1,2,3] & olimit alias var[].
+ * %rax is a scratch register.
+ */
+
+#define op0    rsi
+#define op1    rbx
+#define op2    rcx
+#define op3    rdi
+
+#define ip0    r8
+#define ip1    r9
+#define ip2    r10
+#define ip3    r11
+
+#define bits0  rbp
+#define bits1  rdx
+#define bits2  r12
+#define bits3  r13
+#define dtable r14
+#define olimit r15
+
+/* var[] aliases ip[1,2,3] & olimit
+ * ip[1,2,3] are saved every iteration.
+ * olimit is only used in compute_olimit.
+ */
+#define var0   r15
+#define var1   r9
+#define var2   r10
+#define var3   r11
+
+/* 32-bit var registers */
+#define vard0  r15d
+#define vard1  r9d
+#define vard2  r10d
+#define vard3  r11d
+
+/* Calls X(N) for each stream 0, 1, 2, 3. */
+#define FOR_EACH_STREAM(X) \
+    X(0);                  \
+    X(1);                  \
+    X(2);                  \
+    X(3)
+
+/* Calls X(N, idx) for each stream 0, 1, 2, 3. */
+#define FOR_EACH_STREAM_WITH_INDEX(X, idx) \
+    X(0, idx);                             \
+    X(1, idx);                             \
+    X(2, idx);                             \
+    X(3, idx)
+
+/* Define both _HUF_* & HUF_* symbols because MacOS
+ * C symbols are prefixed with '_' & Linux symbols aren't.
+ */
+_HUF_decompress4X1_usingDTable_internal_bmi2_asm_loop:
+HUF_decompress4X1_usingDTable_internal_bmi2_asm_loop:
+    /* Save all registers - even if they are callee saved for simplicity. */
+    push %rax
+    push %rbx
+    push %rcx
+    push %rdx
+    push %rbp
+    push %rsi
+    push %rdi
+    push %r8
+    push %r9
+    push %r10
+    push %r11
+    push %r12
+    push %r13
+    push %r14
+    push %r15
+
+    /* Read HUF_DecompressAsmArgs* args from %rax */
+    movq %rdi, %rax
+    movq  0(%rax), %ip0
+    movq  8(%rax), %ip1
+    movq 16(%rax), %ip2
+    movq 24(%rax), %ip3
+    movq 32(%rax), %op0
+    movq 40(%rax), %op1
+    movq 48(%rax), %op2
+    movq 56(%rax), %op3
+    movq 64(%rax), %bits0
+    movq 72(%rax), %bits1
+    movq 80(%rax), %bits2
+    movq 88(%rax), %bits3
+    movq 96(%rax), %dtable
+    push %rax      /* argument */
+    push 104(%rax) /* ilimit */
+    push 112(%rax) /* oend */
+    push %olimit   /* olimit space */
+
+    subq $24, %rsp
+
+.L_4X1_compute_olimit:
+    /* Computes how many iterations we can do safely
+     * %r15, %rax may be clobbered
+     * rbx, rdx must be saved
+     * op3 & ip0 mustn't be clobbered
+     */
+    movq %rbx, 0(%rsp)
+    movq %rdx, 8(%rsp)
+
+    movq 32(%rsp), %rax /* rax = oend */
+    subq %op3,    %rax  /* rax = oend - op3 */
+
+    /* r15 = (oend - op3) / 5 */
+    movabsq $-3689348814741910323, %rdx
+    mulq %rdx
+    movq %rdx, %r15
+    shrq $2, %r15
+
+    movq %ip0,     %rax /* rax = ip0 */
+    movq 40(%rsp), %rdx /* rdx = ilimit */
+    subq %rdx,     %rax /* rax = ip0 - ilimit */
+    movq %rax,     %rbx /* rbx = ip0 - ilimit */
+
+    /* rdx = (ip0 - ilimit) / 7 */
+    movabsq $2635249153387078803, %rdx
+    mulq %rdx
+    subq %rdx, %rbx
+    shrq %rbx
+    addq %rbx, %rdx
+    shrq $2, %rdx
+
+    /* r15 = min(%rdx, %r15) */
+    cmpq %rdx, %r15
+    cmova %rdx, %r15
+
+    /* r15 = r15 * 5 */
+    leaq (%r15, %r15, 4), %r15
+
+    /* olimit = op3 + r15 */
+    addq %op3, %olimit
+
+    movq 8(%rsp), %rdx
+    movq 0(%rsp), %rbx
+
+    /* If (op3 + 20 > olimit) */
+    movq %op3, %rax    /* rax = op3 */
+    addq $20,  %rax    /* rax = op3 + 20 */
+    cmpq %rax, %olimit /* op3 + 20 > olimit */
+    jb .L_4X1_exit
+
+    /* If (ip1 < ip0) go to exit */
+    cmpq %ip0, %ip1
+    jb .L_4X1_exit
+
+    /* If (ip2 < ip1) go to exit */
+    cmpq %ip1, %ip2
+    jb .L_4X1_exit
+
+    /* If (ip3 < ip2) go to exit */
+    cmpq %ip2, %ip3
+    jb .L_4X1_exit
+
+/* Reads top 11 bits from bits[n]
+ * Loads dt[bits[n]] into var[n]
+ */
+#define GET_NEXT_DELT(n)                \
+    movq $53, %var##n;                  \
+    shrxq %var##n, %bits##n, %var##n;   \
+    movzwl (%dtable,%var##n,2),%vard##n
+
+/* var[n] must contain the DTable entry computed with GET_NEXT_DELT
+ * Moves var[n] to %rax
+ * bits[n] <<= var[n] & 63
+ * op[n][idx] = %rax >> 8
+ * %ah is a way to access bits [8, 16) of %rax
+ */
+#define DECODE_FROM_DELT(n, idx)       \
+    movq %var##n, %rax;                \
+    shlxq %var##n, %bits##n, %bits##n; \
+    movb %ah, idx(%op##n)
+
+/* Assumes GET_NEXT_DELT has been called.
+ * Calls DECODE_FROM_DELT then GET_NEXT_DELT
+ */
+#define DECODE_AND_GET_NEXT(n, idx) \
+    DECODE_FROM_DELT(n, idx);       \
+    GET_NEXT_DELT(n)                \
+
+/* // ctz & nbBytes is stored in bits[n]
+ * // nbBits is stored in %rax
+ * ctz  = CTZ[bits[n]]
+ * nbBits  = ctz & 7
+ * nbBytes = ctz >> 3
+ * op[n]  += 5
+ * ip[n]  -= nbBytes
+ * // Note: x86-64 is little-endian ==> no bswap
+ * bits[n] = MEM_readST(ip[n]) | 1
+ * bits[n] <<= nbBits
+ */
+#define RELOAD_BITS(n)             \
+    bsfq %bits##n, %bits##n;       \
+    movq %bits##n, %rax;           \
+    andq $7, %rax;                 \
+    shrq $3, %bits##n;             \
+    leaq 5(%op##n), %op##n;        \
+    subq %bits##n, %ip##n;         \
+    movq (%ip##n), %bits##n;       \
+    orq $1, %bits##n;              \
+    shlx %rax, %bits##n, %bits##n
+
+    /* Store clobbered variables on the stack */
+    movq %olimit, 24(%rsp)
+    movq %ip1, 0(%rsp)
+    movq %ip2, 8(%rsp)
+    movq %ip3, 16(%rsp)
+
+    /* Call GET_NEXT_DELT for each stream */
+    FOR_EACH_STREAM(GET_NEXT_DELT)
+
+    .p2align 6
+
+.L_4X1_loop_body:
+    /* Decode 5 symbols in each of the 4 streams (20 total)
+     * Must have called GET_NEXT_DELT for each stream
+     */
+    FOR_EACH_STREAM_WITH_INDEX(DECODE_AND_GET_NEXT, 0)
+    FOR_EACH_STREAM_WITH_INDEX(DECODE_AND_GET_NEXT, 1)
+    FOR_EACH_STREAM_WITH_INDEX(DECODE_AND_GET_NEXT, 2)
+    FOR_EACH_STREAM_WITH_INDEX(DECODE_AND_GET_NEXT, 3)
+    FOR_EACH_STREAM_WITH_INDEX(DECODE_FROM_DELT, 4)
+
+    /* Load ip[1,2,3] from stack (var[] aliases them)
+     * ip[] is needed for RELOAD_BITS
+     * Each will be stored back to the stack after RELOAD
+     */
+    movq 0(%rsp), %ip1
+    movq 8(%rsp), %ip2
+    movq 16(%rsp), %ip3
+
+    /* Reload each stream & fetch the next table entry
+     * to prepare for the next iteration
+     */
+    RELOAD_BITS(0)
+    GET_NEXT_DELT(0)
+
+    RELOAD_BITS(1)
+    movq %ip1, 0(%rsp)
+    GET_NEXT_DELT(1)
+
+    RELOAD_BITS(2)
+    movq %ip2, 8(%rsp)
+    GET_NEXT_DELT(2)
+
+    RELOAD_BITS(3)
+    movq %ip3, 16(%rsp)
+    GET_NEXT_DELT(3)
+
+    /* If op3 < olimit: continue the loop */
+    cmp %op3, 24(%rsp)
+    ja .L_4X1_loop_body
+
+    /* Reload ip[1,2,3] from stack */
+    movq 0(%rsp), %ip1
+    movq 8(%rsp), %ip2
+    movq 16(%rsp), %ip3
+
+    /* Re-compute olimit */
+    jmp .L_4X1_compute_olimit
+
+#undef GET_NEXT_DELT
+#undef DECODE_FROM_DELT
+#undef DECODE
+#undef RELOAD_BITS
+.L_4X1_exit:
+    addq $24, %rsp
+
+    /* Restore stack (oend & olimit) */
+    pop %rax /* olimit */
+    pop %rax /* oend */
+    pop %rax /* ilimit */
+    pop %rax /* arg */
+
+    /* Save ip / op / bits */
+    movq %ip0,  0(%rax)
+    movq %ip1,  8(%rax)
+    movq %ip2, 16(%rax)
+    movq %ip3, 24(%rax)
+    movq %op0, 32(%rax)
+    movq %op1, 40(%rax)
+    movq %op2, 48(%rax)
+    movq %op3, 56(%rax)
+    movq %bits0, 64(%rax)
+    movq %bits1, 72(%rax)
+    movq %bits2, 80(%rax)
+    movq %bits3, 88(%rax)
+
+    /* Restore registers */
+    pop %r15
+    pop %r14
+    pop %r13
+    pop %r12
+    pop %r11
+    pop %r10
+    pop %r9
+    pop %r8
+    pop %rdi
+    pop %rsi
+    pop %rbp
+    pop %rdx
+    pop %rcx
+    pop %rbx
+    pop %rax
+    ret
+
+_HUF_decompress4X2_usingDTable_internal_bmi2_asm_loop:
+HUF_decompress4X2_usingDTable_internal_bmi2_asm_loop:
+    /* Save all registers - even if they are callee saved for simplicity. */
+    push %rax
+    push %rbx
+    push %rcx
+    push %rdx
+    push %rbp
+    push %rsi
+    push %rdi
+    push %r8
+    push %r9
+    push %r10
+    push %r11
+    push %r12
+    push %r13
+    push %r14
+    push %r15
+
+    movq %rdi, %rax
+    movq  0(%rax), %ip0
+    movq  8(%rax), %ip1
+    movq 16(%rax), %ip2
+    movq 24(%rax), %ip3
+    movq 32(%rax), %op0
+    movq 40(%rax), %op1
+    movq 48(%rax), %op2
+    movq 56(%rax), %op3
+    movq 64(%rax), %bits0
+    movq 72(%rax), %bits1
+    movq 80(%rax), %bits2
+    movq 88(%rax), %bits3
+    movq 96(%rax), %dtable
+    push %rax      /* argument */
+    push %rax      /* olimit */
+    push 104(%rax) /* ilimit */
+
+    movq 112(%rax), %rax
+    push %rax /* oend3 */
+
+    movq %op3, %rax
+    push %rax /* oend2 */
+
+    movq %op2, %rax
+    push %rax /* oend1 */
+
+    movq %op1, %rax
+    push %rax /* oend0 */
+
+    /* Scratch space */
+    subq $8, %rsp
+
+.L_4X2_compute_olimit:
+    /* Computes how many iterations we can do safely
+     * %r15, %rax may be clobbered
+     * rdx must be saved
+     * op[1,2,3,4] & ip0 mustn't be clobbered
+     */
+    movq %rdx, 0(%rsp)
+
+    /* We can consume up to 7 input bytes each iteration. */
+    movq %ip0,     %rax  /* rax = ip0 */
+    movq 40(%rsp), %rdx  /* rdx = ilimit */
+    subq %rdx,     %rax  /* rax = ip0 - ilimit */
+    movq %rax,    %r15   /* r15 = ip0 - ilimit */
+
+    /* rdx = rax / 7 */
+    movabsq $2635249153387078803, %rdx
+    mulq %rdx
+    subq %rdx, %r15
+    shrq %r15
+    addq %r15, %rdx
+    shrq $2, %rdx
+
+    /* r15 = (ip0 - ilimit) / 7 */
+    movq %rdx, %r15
+
+    movabsq $-3689348814741910323, %rdx
+    movq 8(%rsp), %rax /* rax = oend0 */
+    subq %op0,    %rax /* rax = oend0 - op0 */
+    mulq %rdx
+    shrq $3,      %rdx /* rdx = rax / 10 */
+
+    /* r15 = min(%rdx, %r15) */
+    cmpq  %rdx, %r15
+    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  %rdx, %r15
+    cmova %rdx, %r15
+
+    movabsq $-3689348814741910323, %rdx
+    movq 32(%rsp), %rax /* rax = oend3 */
+    subq %op3,     %rax /* rax = oend3 - op3 */
+    mulq %rdx
+    shrq $3,       %rdx /* rdx = rax / 10 */
+
+    /* r15 = min(%rdx, %r15) */
+    cmpq  %rdx, %r15
+    cmova %rdx, %r15
+
+    /* olimit = op3 + 5 * r15 */
+    movq %r15, %rax
+    leaq (%op3, %rax, 4), %olimit
+    addq %rax, %olimit
+
+    movq 0(%rsp), %rdx
+
+    /* If (op3 + 10 > olimit) */
+    movq %op3, %rax    /* rax = op3 */
+    addq $10,  %rax    /* rax = op3 + 10 */
+    cmpq %rax, %olimit /* op3 + 10 > olimit */
+    jb .L_4X2_exit
+
+    /* If (ip1 < ip0) go to exit */
+    cmpq %ip0, %ip1
+    jb .L_4X2_exit
+
+    /* If (ip2 < ip1) go to exit */
+    cmpq %ip1, %ip2
+    jb .L_4X2_exit
+
+    /* If (ip3 < ip2) go to exit */
+    cmpq %ip2, %ip3
+    jb .L_4X2_exit
+
+#define DECODE(n, idx)              \
+    movq %bits##n, %rax;            \
+    shrq $53, %rax;                 \
+    movzwl 0(%dtable,%rax,4),%r8d;  \
+    movzbl 2(%dtable,%rax,4),%r15d; \
+    movzbl 3(%dtable,%rax,4),%eax;  \
+    movw %r8w, (%op##n);            \
+    shlxq %r15, %bits##n, %bits##n; \
+    addq %rax, %op##n
+
+#define RELOAD_BITS(n)              \
+    bsfq %bits##n, %bits##n;        \
+    movq %bits##n, %rax;            \
+    shrq $3, %bits##n;              \
+    andq $7, %rax;                  \
+    subq %bits##n, %ip##n;          \
+    movq (%ip##n), %bits##n;        \
+    orq $1, %bits##n;               \
+    shlxq %rax, %bits##n, %bits##n
+
+
+    movq %olimit, 48(%rsp)
+
+    .p2align 6
+
+.L_4X2_loop_body:
+    /* We clobber r8, so store it on the stack */
+    movq %r8, 0(%rsp)
+
+    /* Decode 5 symbols from each of the 4 streams (20 symbols total). */
+    FOR_EACH_STREAM_WITH_INDEX(DECODE, 0)
+    FOR_EACH_STREAM_WITH_INDEX(DECODE, 1)
+    FOR_EACH_STREAM_WITH_INDEX(DECODE, 2)
+    FOR_EACH_STREAM_WITH_INDEX(DECODE, 3)
+    FOR_EACH_STREAM_WITH_INDEX(DECODE, 4)
+
+    /* Reload r8 */
+    movq 0(%rsp), %r8
+
+    FOR_EACH_STREAM(RELOAD_BITS)
+
+    cmp %op3, 48(%rsp)
+    ja .L_4X2_loop_body
+    jmp .L_4X2_compute_olimit
+
+#undef DECODE
+#undef RELOAD_BITS
+.L_4X2_exit:
+    addq $8, %rsp
+    /* Restore stack (oend & olimit) */
+    pop %rax /* oend0 */
+    pop %rax /* oend1 */
+    pop %rax /* oend2 */
+    pop %rax /* oend3 */
+    pop %rax /* ilimit */
+    pop %rax /* olimit */
+    pop %rax /* arg */
+
+    /* Save ip / op / bits */
+    movq %ip0,  0(%rax)
+    movq %ip1,  8(%rax)
+    movq %ip2, 16(%rax)
+    movq %ip3, 24(%rax)
+    movq %op0, 32(%rax)
+    movq %op1, 40(%rax)
+    movq %op2, 48(%rax)
+    movq %op3, 56(%rax)
+    movq %bits0, 64(%rax)
+    movq %bits1, 72(%rax)
+    movq %bits2, 80(%rax)
+    movq %bits3, 88(%rax)
+
+    /* Restore registers */
+    pop %r15
+    pop %r14
+    pop %r13
+    pop %r12
+    pop %r11
+    pop %r10
+    pop %r9
+    pop %r8
+    pop %rdi
+    pop %rsi
+    pop %rbp
+    pop %rdx
+    pop %rcx
+    pop %rbx
+    pop %rax
+    ret
+
+#endif

zstd/decompress/zstd_decompress.c

@@ -56,7 +56,6 @@
 *  Dependencies
 *********************************************************/
 #include "../common/zstd_deps.h"   /* ZSTD_memcpy, ZSTD_memmove, ZSTD_memset */
-#include "../common/cpu.h"         /* bmi2 */
 #include "../common/mem.h"         /* low level memory routines */
 #define FSE_STATIC_LINKING_ONLY
 #include "../common/fse.h"
@@ -177,12 +176,15 @@ static const ZSTD_DDict* ZSTD_DDictHashSet_getDDict(ZSTD_DDictHashSet* hashSet,
 static ZSTD_DDictHashSet* ZSTD_createDDictHashSet(ZSTD_customMem customMem) {
     ZSTD_DDictHashSet* ret = (ZSTD_DDictHashSet*)ZSTD_customMalloc(sizeof(ZSTD_DDictHashSet), customMem);
     DEBUGLOG(4, "Allocating new hash set");
+    if (!ret)
+        return NULL;
     ret->ddictPtrTable = (const ZSTD_DDict**)ZSTD_customCalloc(DDICT_HASHSET_TABLE_BASE_SIZE * sizeof(ZSTD_DDict*), customMem);
-    ret->ddictPtrTableSize = DDICT_HASHSET_TABLE_BASE_SIZE;
+    if (!ret->ddictPtrTable) {
-    ret->ddictPtrCount = 0;
+        ZSTD_customFree(ret, customMem);
-    if (!ret || !ret->ddictPtrTable) {
         return NULL;
     }
+    ret->ddictPtrTableSize = DDICT_HASHSET_TABLE_BASE_SIZE;
+    ret->ddictPtrCount = 0;
     return ret;
 }
 
@@ -255,11 +257,15 @@ static void ZSTD_initDCtx_internal(ZSTD_DCtx* dctx)
     dctx->inBuffSize  = 0;
     dctx->outBuffSize = 0;
     dctx->streamStage = zdss_init;
+#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
     dctx->legacyContext = NULL;
     dctx->previousLegacyVersion = 0;
+#endif
     dctx->noForwardProgress = 0;
     dctx->oversizedDuration = 0;
-    dctx->bmi2 = ZSTD_cpuid_bmi2(ZSTD_cpuid());
+#if DYNAMIC_BMI2
+    dctx->bmi2 = ZSTD_cpuSupportsBmi2();
+#endif
     dctx->ddictSet = NULL;
     ZSTD_DCtx_resetParameters(dctx);
 #ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
@@ -280,8 +286,7 @@ ZSTD_DCtx* ZSTD_initStaticDCtx(void *workspace, size_t workspaceSize)
     return dctx;
 }
 
-ZSTD_DCtx* ZSTD_createDCtx_advanced(ZSTD_customMem customMem)
+static ZSTD_DCtx* ZSTD_createDCtx_internal(ZSTD_customMem customMem) {
-{
     if ((!customMem.customAlloc) ^ (!customMem.customFree)) return NULL;
 
     {   ZSTD_DCtx* const dctx = (ZSTD_DCtx*)ZSTD_customMalloc(sizeof(*dctx), customMem);
@@ -292,10 +297,15 @@ ZSTD_DCtx* ZSTD_createDCtx_advanced(ZSTD_customMem customMem)
     }
 }
 
+ZSTD_DCtx* ZSTD_createDCtx_advanced(ZSTD_customMem customMem)
+{
+    return ZSTD_createDCtx_internal(customMem);
+}
+
 ZSTD_DCtx* ZSTD_createDCtx(void)
 {
     DEBUGLOG(3, "ZSTD_createDCtx");
-    return ZSTD_createDCtx_advanced(ZSTD_defaultCMem);
+    return ZSTD_createDCtx_internal(ZSTD_defaultCMem);
 }
 
 static void ZSTD_clearDict(ZSTD_DCtx* dctx)
@@ -380,6 +390,19 @@ unsigned ZSTD_isFrame(const void* buffer, size_t size)
     return 0;
 }
 
+/*! ZSTD_isSkippableFrame() :
+ *  Tells if the content of `buffer` starts with a valid Frame Identifier for a skippable frame.
+ *  Note : Frame Identifier is 4 bytes. If `size < 4`, @return will always be 0.
+ */
+unsigned ZSTD_isSkippableFrame(const void* buffer, size_t size)
+{
+    if (size < ZSTD_FRAMEIDSIZE) return 0;
+    {   U32 const magic = MEM_readLE32(buffer);
+        if ((magic & ZSTD_MAGIC_SKIPPABLE_MASK) == ZSTD_MAGIC_SKIPPABLE_START) return 1;
+    }
+    return 0;
+}
+
 /** ZSTD_frameHeaderSize_internal() :
  *  srcSize must be large enough to reach header size fields.
  *  note : only works for formats ZSTD_f_zstd1 and ZSTD_f_zstd1_magicless.
@@ -466,7 +489,9 @@ size_t ZSTD_getFrameHeader_advanced(ZSTD_frameHeader* zfhPtr, const void* src, s
         }
         switch(dictIDSizeCode)
         {
-            default: assert(0);  /* impossible */
+            default:
+                assert(0);  /* impossible */
+                ZSTD_FALLTHROUGH;
             case 0 : break;
             case 1 : dictID = ip[pos]; pos++; break;
             case 2 : dictID = MEM_readLE16(ip+pos); pos+=2; break;
@@ -474,7 +499,9 @@ size_t ZSTD_getFrameHeader_advanced(ZSTD_frameHeader* zfhPtr, const void* src, s
         }
         switch(fcsID)
         {
-            default: assert(0);  /* impossible */
+            default:
+                assert(0);  /* impossible */
+                ZSTD_FALLTHROUGH;
             case 0 : if (singleSegment) frameContentSize = ip[pos]; break;
             case 1 : frameContentSize = MEM_readLE16(ip+pos)+256; break;
             case 2 : frameContentSize = MEM_readLE32(ip+pos); break;
@@ -503,7 +530,6 @@ size_t ZSTD_getFrameHeader(ZSTD_frameHeader* zfhPtr, const void* src, size_t src
     return ZSTD_getFrameHeader_advanced(zfhPtr, src, srcSize, ZSTD_f_zstd1);
 }
 
-
 /** ZSTD_getFrameContentSize() :
  *  compatible with legacy mode
  * @return : decompressed size of the single frame pointed to be `src` if known, otherwise
@@ -544,6 +570,37 @@ static size_t readSkippableFrameSize(void const* src, size_t srcSize)
     }
 }
 
+/*! ZSTD_readSkippableFrame() :
+ * Retrieves a zstd skippable frame containing data given by src, and writes it to dst buffer.
+ *
+ * The parameter magicVariant will receive the magicVariant that was supplied when the frame was written,
+ * i.e. magicNumber - ZSTD_MAGIC_SKIPPABLE_START.  This can be NULL if the caller is not interested
+ * in the magicVariant.
+ *
+ * Returns an error if destination buffer is not large enough, or if the frame is not skippable.
+ *
+ * @return : number of bytes written or a ZSTD error.
+ */
+ZSTDLIB_API size_t ZSTD_readSkippableFrame(void* dst, size_t dstCapacity, unsigned* magicVariant,
+                                            const void* src, size_t srcSize)
+{
+    U32 const magicNumber = MEM_readLE32(src);
+    size_t skippableFrameSize = readSkippableFrameSize(src, srcSize);
+    size_t skippableContentSize = skippableFrameSize - ZSTD_SKIPPABLEHEADERSIZE;
+
+    /* check input validity */
+    RETURN_ERROR_IF(!ZSTD_isSkippableFrame(src, srcSize), frameParameter_unsupported, "");
+    RETURN_ERROR_IF(skippableFrameSize < ZSTD_SKIPPABLEHEADERSIZE || skippableFrameSize > srcSize, srcSize_wrong, "");
+    RETURN_ERROR_IF(skippableContentSize > dstCapacity, dstSize_tooSmall, "");
+
+    /* deliver payload */
+    if (skippableContentSize > 0  && dst != NULL)
+        ZSTD_memcpy(dst, (const BYTE *)src + ZSTD_SKIPPABLEHEADERSIZE, skippableContentSize);
+    if (magicVariant != NULL)
+        *magicVariant = magicNumber - ZSTD_MAGIC_SKIPPABLE_START;
+    return skippableContentSize;
+}
+
 /** ZSTD_findDecompressedSize() :
  *  compatible with legacy mode
  *  `srcSize` must be the exact length of some number of ZSTD compressed and/or
@@ -858,7 +915,7 @@ static size_t ZSTD_decompressFrame(ZSTD_DCtx* dctx,
         switch(blockProperties.blockType)
         {
         case bt_compressed:
-            decodedSize = ZSTD_decompressBlock_internal(dctx, op, (size_t)(oend-op), ip, cBlockSize, /* frame */ 1);
+            decodedSize = ZSTD_decompressBlock_internal(dctx, op, (size_t)(oend-op), ip, cBlockSize, /* frame */ 1, not_streaming);
             break;
         case bt_raw :
             decodedSize = ZSTD_copyRawBlock(op, (size_t)(oend-op), ip, cBlockSize);
@@ -1009,7 +1066,7 @@ static ZSTD_DDict const* ZSTD_getDDict(ZSTD_DCtx* dctx)
     switch (dctx->dictUses) {
     default:
         assert(0 /* Impossible */);
-        /* fall-through */
+        ZSTD_FALLTHROUGH;
     case ZSTD_dont_use:
         ZSTD_clearDict(dctx);
         return NULL;
@@ -1031,7 +1088,7 @@ size_t ZSTD_decompress(void* dst, size_t dstCapacity, const void* src, size_t sr
 {
 #if defined(ZSTD_HEAPMODE) && (ZSTD_HEAPMODE>=1)
     size_t regenSize;
-    ZSTD_DCtx* const dctx = ZSTD_createDCtx();
+    ZSTD_DCtx* const dctx =  ZSTD_createDCtx_internal(ZSTD_defaultCMem);
     RETURN_ERROR_IF(dctx==NULL, memory_allocation, "NULL pointer!");
     regenSize = ZSTD_decompressDCtx(dctx, dst, dstCapacity, src, srcSize);
     ZSTD_freeDCtx(dctx);
@@ -1065,7 +1122,7 @@ static size_t ZSTD_nextSrcSizeToDecompressWithInputSize(ZSTD_DCtx* dctx, size_t
         return dctx->expected;
     if (dctx->bType != bt_raw)
         return dctx->expected;
-    return MIN(MAX(inputSize, 1), dctx->expected);
+    return BOUNDED(1, inputSize, dctx->expected);
 }
 
 ZSTD_nextInputType_e ZSTD_nextInputType(ZSTD_DCtx* dctx) {
@@ -1073,7 +1130,9 @@ ZSTD_nextInputType_e ZSTD_nextInputType(ZSTD_DCtx* dctx) {
     {
     default:   /* should not happen */
         assert(0);
+        ZSTD_FALLTHROUGH;
     case ZSTDds_getFrameHeaderSize:
+        ZSTD_FALLTHROUGH;
     case ZSTDds_decodeFrameHeader:
         return ZSTDnit_frameHeader;
     case ZSTDds_decodeBlockHeader:
@@ -1085,6 +1144,7 @@ ZSTD_nextInputType_e ZSTD_nextInputType(ZSTD_DCtx* dctx) {
     case ZSTDds_checkChecksum:
         return ZSTDnit_checksum;
     case ZSTDds_decodeSkippableHeader:
+        ZSTD_FALLTHROUGH;
     case ZSTDds_skipFrame:
         return ZSTDnit_skippableFrame;
     }
@@ -1168,7 +1228,7 @@ size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, c
             {
             case bt_compressed:
                 DEBUGLOG(5, "ZSTD_decompressContinue: case bt_compressed");
-                rSize = ZSTD_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize, /* frame */ 1);
+                rSize = ZSTD_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize, /* frame */ 1, is_streaming);
                 dctx->expected = 0;  /* Streaming not supported */
                 break;
             case bt_raw :
@@ -1493,7 +1553,7 @@ size_t ZSTD_decompress_usingDDict(ZSTD_DCtx* dctx,
 ZSTD_DStream* ZSTD_createDStream(void)
 {
     DEBUGLOG(3, "ZSTD_createDStream");
-    return ZSTD_createDStream_advanced(ZSTD_defaultCMem);
+    return ZSTD_createDCtx_internal(ZSTD_defaultCMem);
 }
 
 ZSTD_DStream* ZSTD_initStaticDStream(void *workspace, size_t workspaceSize)
@@ -1503,7 +1563,7 @@ ZSTD_DStream* ZSTD_initStaticDStream(void *workspace, size_t workspaceSize)
 
 ZSTD_DStream* ZSTD_createDStream_advanced(ZSTD_customMem customMem)
 {
-    return ZSTD_createDCtx_advanced(customMem);
+    return ZSTD_createDCtx_internal(customMem);
 }
 
 size_t ZSTD_freeDStream(ZSTD_DStream* zds)
@@ -1763,7 +1823,8 @@ size_t ZSTD_sizeof_DStream(const ZSTD_DStream* dctx)
 size_t ZSTD_decodingBufferSize_min(unsigned long long windowSize, unsigned long long frameContentSize)
 {
     size_t const blockSize = (size_t) MIN(windowSize, ZSTD_BLOCKSIZE_MAX);
-    unsigned long long const neededRBSize = windowSize + blockSize + (WILDCOPY_OVERLENGTH * 2);
+    /* space is needed to store the litbuffer after the output of a given block without stomping the extDict of a previous run, as well as to cover both windows against wildcopy*/
+    unsigned long long const neededRBSize = windowSize + blockSize + ZSTD_BLOCKSIZE_MAX + (WILDCOPY_OVERLENGTH * 2);
     unsigned long long const neededSize = MIN(frameContentSize, neededRBSize);
     size_t const minRBSize = (size_t) neededSize;
     RETURN_ERROR_IF((unsigned long long)minRBSize != neededSize,
@@ -1897,10 +1958,12 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
             DEBUGLOG(5, "stage zdss_init => transparent reset ");
             zds->streamStage = zdss_loadHeader;
             zds->lhSize = zds->inPos = zds->outStart = zds->outEnd = 0;
+#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
             zds->legacyVersion = 0;
+#endif
             zds->hostageByte = 0;
             zds->expectedOutBuffer = *output;
-            /* fall-through */
+            ZSTD_FALLTHROUGH;
 
         case zdss_loadHeader :
             DEBUGLOG(5, "stage zdss_loadHeader (srcSize : %u)", (U32)(iend - ip));
@@ -2038,7 +2101,7 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
                         zds->outBuffSize = neededOutBuffSize;
             }   }   }
             zds->streamStage = zdss_read;
-            /* fall-through */
+            ZSTD_FALLTHROUGH;
 
         case zdss_read:
             DEBUGLOG(5, "stage zdss_read");
@@ -2057,7 +2120,7 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
             }   }
             if (ip==iend) { someMoreWork = 0; break; }   /* no more input */
             zds->streamStage = zdss_load;
-            /* fall-through */
+            ZSTD_FALLTHROUGH;
 
         case zdss_load:
             {   size_t const neededInSize = ZSTD_nextSrcSizeToDecompress(zds);

zstd/decompress/zstd_decompress_block.h

@@ -33,6 +33,12 @@
  */
 
 
+ /* Streaming state is used to inform allocation of the literal buffer */
+typedef enum {
+    not_streaming = 0,
+    is_streaming = 1
+} streaming_operation;
+
 /* ZSTD_decompressBlock_internal() :
  * decompress block, starting at `src`,
  * into destination buffer `dst`.
@@ -41,7 +47,7 @@
  */
 size_t ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
                                void* dst, size_t dstCapacity,
-                         const void* src, size_t srcSize, const int frame);
+                         const void* src, size_t srcSize, const int frame, const streaming_operation streaming);
 
 /* ZSTD_buildFSETable() :
  * generate FSE decoding table for one symbol (ll, ml or off)
@@ -54,7 +60,7 @@ size_t ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
  */
 void ZSTD_buildFSETable(ZSTD_seqSymbol* dt,
              const short* normalizedCounter, unsigned maxSymbolValue,
-             const U32* baseValue, const U32* nbAdditionalBits,
+             const U32* baseValue, const U8* nbAdditionalBits,
                    unsigned tableLog, void* wksp, size_t wkspSize,
                    int bmi2);
 

zstd/decompress/zstd_decompress_internal.h

@@ -20,7 +20,7 @@
  *  Dependencies
  *********************************************************/
 #include "../common/mem.h"             /* BYTE, U16, U32 */
-#include "../common/zstd_internal.h"   /* ZSTD_seqSymbol */
+#include "../common/zstd_internal.h"   /* constants : MaxLL, MaxML, MaxOff, LLFSELog, etc. */
 
 
 
@@ -40,7 +40,7 @@ static UNUSED_ATTR const U32 OF_base[MaxOff+1] = {
                  0xFFFD, 0x1FFFD, 0x3FFFD, 0x7FFFD, 0xFFFFD, 0x1FFFFD, 0x3FFFFD, 0x7FFFFD,
                  0xFFFFFD, 0x1FFFFFD, 0x3FFFFFD, 0x7FFFFFD, 0xFFFFFFD, 0x1FFFFFFD, 0x3FFFFFFD, 0x7FFFFFFD };
 
-static UNUSED_ATTR const U32 OF_bits[MaxOff+1] = {
+static UNUSED_ATTR const U8 OF_bits[MaxOff+1] = {
                      0,  1,  2,  3,  4,  5,  6,  7,
                      8,  9, 10, 11, 12, 13, 14, 15,
                     16, 17, 18, 19, 20, 21, 22, 23,
@@ -106,6 +106,22 @@ typedef struct {
     size_t ddictPtrCount;
 } ZSTD_DDictHashSet;
 
+#ifndef ZSTD_DECODER_INTERNAL_BUFFER
+#  define ZSTD_DECODER_INTERNAL_BUFFER  (1 << 16)
+#endif
+
+#define ZSTD_LBMIN 64
+#define ZSTD_LBMAX (128 << 10)
+
+/* extra buffer, compensates when dst is not large enough to store litBuffer */
+#define ZSTD_LITBUFFEREXTRASIZE  BOUNDED(ZSTD_LBMIN, ZSTD_DECODER_INTERNAL_BUFFER, ZSTD_LBMAX)
+
+typedef enum {
+    ZSTD_not_in_dst = 0,  /* Stored entirely within litExtraBuffer */
+    ZSTD_in_dst = 1,           /* Stored entirely within dst (in memory after current output write) */
+    ZSTD_split = 2            /* Split between litExtraBuffer and dst */
+} ZSTD_litLocation_e;
+
 struct ZSTD_DCtx_s
 {
     const ZSTD_seqSymbol* LLTptr;
@@ -136,7 +152,9 @@ struct ZSTD_DCtx_s
     size_t litSize;
     size_t rleSize;
     size_t staticSize;
+#if DYNAMIC_BMI2 != 0
     int bmi2;                     /* == 1 if the CPU supports BMI2 and 0 otherwise. CPU support is determined dynamically once per context lifetime. */
+#endif
 
     /* dictionary */
     ZSTD_DDict* ddictLocal;
@@ -158,16 +176,21 @@ struct ZSTD_DCtx_s
     size_t outStart;
     size_t outEnd;
     size_t lhSize;
+#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
     void* legacyContext;
     U32 previousLegacyVersion;
     U32 legacyVersion;
+#endif
     U32 hostageByte;
     int noForwardProgress;
     ZSTD_bufferMode_e outBufferMode;
     ZSTD_outBuffer expectedOutBuffer;
 
     /* workspace */
-    BYTE litBuffer[ZSTD_BLOCKSIZE_MAX + WILDCOPY_OVERLENGTH];
+    BYTE* litBuffer;
+    const BYTE* litBufferEnd;
+    ZSTD_litLocation_e litBufferLocation;
+    BYTE litExtraBuffer[ZSTD_LITBUFFEREXTRASIZE + WILDCOPY_OVERLENGTH]; /* literal buffer can be split between storage within dst and within this scratch buffer */
     BYTE headerBuffer[ZSTD_FRAMEHEADERSIZE_MAX];
 
     size_t oversizedDuration;
@@ -183,6 +206,14 @@ struct ZSTD_DCtx_s
 #endif
 };  /* typedef'd to ZSTD_DCtx within "zstd.h" */
 
+MEM_STATIC int ZSTD_DCtx_get_bmi2(const struct ZSTD_DCtx_s *dctx) {
+#if DYNAMIC_BMI2 != 0
+	return dctx->bmi2;
+#else
+    (void)dctx;
+	return 0;
+#endif
+}
 
 /*-*******************************************************
  *  Shared internal functions

zstd/dictBuilder/cover.c

@@ -40,6 +40,13 @@
 /*-*************************************
 *  Constants
 ***************************************/
+/**
+* There are 32bit indexes used to ref samples, so limit samples size to 4GB
+* on 64bit builds.
+* For 32bit builds we choose 1 GB.
+* Most 32bit platforms have 2GB user-mode addressable space and we allocate a large
+* contiguous buffer, so 1GB is already a high limit.
+*/
 #define COVER_MAX_SAMPLES_SIZE (sizeof(size_t) == 8 ? ((unsigned)-1) : ((unsigned)1 GB))
 #define COVER_DEFAULT_SPLITPOINT 1.0
 
@@ -47,7 +54,7 @@
 *  Console display
 ***************************************/
 #ifndef LOCALDISPLAYLEVEL
-static int g_displayLevel = 2;
+static int g_displayLevel = 0;
 #endif
 #undef  DISPLAY
 #define DISPLAY(...)                                                           \
@@ -735,7 +742,7 @@ ZDICTLIB_API size_t ZDICT_trainFromBuffer_cover(
   COVER_map_t activeDmers;
   parameters.splitPoint = 1.0;
   /* Initialize global data */
-  g_displayLevel = parameters.zParams.notificationLevel;
+  g_displayLevel = (int)parameters.zParams.notificationLevel;
   /* Checks */
   if (!COVER_checkParameters(parameters, dictBufferCapacity)) {
     DISPLAYLEVEL(1, "Cover parameters incorrect\n");

zstd/dictBuilder/fastcover.c

@@ -32,6 +32,13 @@
 /*-*************************************
 *  Constants
 ***************************************/
+/**
+* There are 32bit indexes used to ref samples, so limit samples size to 4GB
+* on 64bit builds.
+* For 32bit builds we choose 1 GB.
+* Most 32bit platforms have 2GB user-mode addressable space and we allocate a large
+* contiguous buffer, so 1GB is already a high limit.
+*/
 #define FASTCOVER_MAX_SAMPLES_SIZE (sizeof(size_t) == 8 ? ((unsigned)-1) : ((unsigned)1 GB))
 #define FASTCOVER_MAX_F 31
 #define FASTCOVER_MAX_ACCEL 10
@@ -44,7 +51,7 @@
 *  Console display
 ***************************************/
 #ifndef LOCALDISPLAYLEVEL
-static int g_displayLevel = 2;
+static int g_displayLevel = 0;
 #endif
 #undef  DISPLAY
 #define DISPLAY(...)                                                           \
@@ -549,7 +556,7 @@ ZDICT_trainFromBuffer_fastCover(void* dictBuffer, size_t dictBufferCapacity,
     ZDICT_cover_params_t coverParams;
     FASTCOVER_accel_t accelParams;
     /* Initialize global data */
-    g_displayLevel = parameters.zParams.notificationLevel;
+    g_displayLevel = (int)parameters.zParams.notificationLevel;
     /* Assign splitPoint and f if not provided */
     parameters.splitPoint = 1.0;
     parameters.f = parameters.f == 0 ? DEFAULT_F : parameters.f;
@@ -632,7 +639,7 @@ ZDICT_optimizeTrainFromBuffer_fastCover(
     const unsigned accel = parameters->accel == 0 ? DEFAULT_ACCEL : parameters->accel;
     const unsigned shrinkDict = 0;
     /* Local variables */
-    const int displayLevel = parameters->zParams.notificationLevel;
+    const int displayLevel = (int)parameters->zParams.notificationLevel;
     unsigned iteration = 1;
     unsigned d;
     unsigned k;
@@ -716,7 +723,7 @@ ZDICT_optimizeTrainFromBuffer_fastCover(
         data->parameters.splitPoint = splitPoint;
         data->parameters.steps = kSteps;
         data->parameters.shrinkDict = shrinkDict;
-        data->parameters.zParams.notificationLevel = g_displayLevel;
+        data->parameters.zParams.notificationLevel = (unsigned)g_displayLevel;
         /* Check the parameters */
         if (!FASTCOVER_checkParameters(data->parameters, dictBufferCapacity,
                                        data->ctx->f, accel)) {

zstd/dictBuilder/zdict.c

@@ -135,22 +135,32 @@ static unsigned ZDICT_NbCommonBytes (size_t val)
     if (MEM_isLittleEndian()) {
         if (MEM_64bits()) {
 #       if defined(_MSC_VER) && defined(_WIN64)
-            unsigned long r = 0;
+            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 (__builtin_ctzll((U64)val) >> 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)
-            unsigned long r=0;
+            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);
+            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];
@@ -159,11 +169,16 @@ static unsigned ZDICT_NbCommonBytes (size_t val)
     } else {  /* Big Endian CPU */
         if (MEM_64bits()) {
 #       if defined(_MSC_VER) && defined(_WIN64)
-            unsigned long r = 0;
+            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 (__builtin_clzll(val) >> 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 */
@@ -174,11 +189,16 @@ static unsigned ZDICT_NbCommonBytes (size_t val)
 #       endif
         } else { /* 32 bits */
 #       if defined(_MSC_VER)
-            unsigned long r = 0;
+            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);
+            return (unsigned)(__builtin_clz((U32)val) >> 3);
 #       else
             unsigned r;
             if (!(val>>16)) { r=2; val>>=8; } else { r=0; val>>=24; }
@@ -235,7 +255,7 @@ static dictItem ZDICT_analyzePos(
     U32 savings[LLIMIT] = {0};
     const BYTE* b = (const BYTE*)buffer;
     size_t maxLength = LLIMIT;
-    size_t pos = suffix[start];
+    size_t pos = (size_t)suffix[start];
     U32 end = start;
     dictItem solution;
 
@@ -369,7 +389,7 @@ static dictItem ZDICT_analyzePos(
             savings[i] = savings[i-1] + (lengthList[i] * (i-3));
 
         DISPLAYLEVEL(4, "Selected dict at position %u, of length %u : saves %u (ratio: %.2f)  \n",
-                     (unsigned)pos, (unsigned)maxLength, (unsigned)savings[maxLength], (double)savings[maxLength] / maxLength);
+                     (unsigned)pos, (unsigned)maxLength, (unsigned)savings[maxLength], (double)savings[maxLength] / (double)maxLength);
 
         solution.pos = (U32)pos;
         solution.length = (U32)maxLength;
@@ -379,7 +399,7 @@ static dictItem ZDICT_analyzePos(
         {   U32 id;
             for (id=start; id<end; id++) {
                 U32 p, pEnd, length;
-                U32 const testedPos = suffix[id];
+                U32 const testedPos = (U32)suffix[id];
                 if (testedPos == pos)
                     length = solution.length;
                 else {
@@ -442,7 +462,7 @@ static U32 ZDICT_tryMerge(dictItem* table, dictItem elt, U32 eltNbToSkip, const
 
         if ((table[u].pos + table[u].length >= elt.pos) && (table[u].pos < elt.pos)) {  /* overlap, existing < new */
             /* append */
-            int const addedLength = (int)eltEnd - (table[u].pos + table[u].length);
+            int const addedLength = (int)eltEnd - (int)(table[u].pos + table[u].length);
             table[u].savings += elt.length / 8;    /* rough approx bonus */
             if (addedLength > 0) {   /* otherwise, elt fully included into existing */
                 table[u].length += addedLength;
@@ -766,6 +786,13 @@ static size_t ZDICT_analyzeEntropy(void*  dstBuffer, size_t maxDstSize,
         pos += fileSizes[u];
     }
 
+    if (notificationLevel >= 4) {
+        /* writeStats */
+        DISPLAYLEVEL(4, "Offset Code Frequencies : \n");
+        for (u=0; u<=offcodeMax; u++) {
+            DISPLAYLEVEL(4, "%2u :%7u \n", u, offcodeCount[u]);
+    }   }
+
     /* analyze, build stats, starting with literals */
     {   size_t maxNbBits = HUF_buildCTable (hufTable, countLit, 255, huffLog);
         if (HUF_isError(maxNbBits)) {
@@ -872,7 +899,7 @@ static size_t ZDICT_analyzeEntropy(void*  dstBuffer, size_t maxDstSize,
     MEM_writeLE32(dstPtr+8, bestRepOffset[2].offset);
 #else
     /* at this stage, we don't use the result of "most common first offset",
-       as the impact of statistics is not properly evaluated */
+     * as the impact of statistics is not properly evaluated */
     MEM_writeLE32(dstPtr+0, repStartValue[0]);
     MEM_writeLE32(dstPtr+4, repStartValue[1]);
     MEM_writeLE32(dstPtr+8, repStartValue[2]);
@@ -888,6 +915,17 @@ _cleanup:
 }
 
 
+/**
+ * @returns the maximum repcode value
+ */
+static U32 ZDICT_maxRep(U32 const reps[ZSTD_REP_NUM])
+{
+    U32 maxRep = reps[0];
+    int r;
+    for (r = 1; r < ZSTD_REP_NUM; ++r)
+        maxRep = MAX(maxRep, reps[r]);
+    return maxRep;
+}
 
 size_t ZDICT_finalizeDictionary(void* dictBuffer, size_t dictBufferCapacity,
                           const void* customDictContent, size_t dictContentSize,
@@ -899,11 +937,13 @@ size_t ZDICT_finalizeDictionary(void* dictBuffer, size_t dictBufferCapacity,
     BYTE header[HBUFFSIZE];
     int const compressionLevel = (params.compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT : params.compressionLevel;
     U32 const notificationLevel = params.notificationLevel;
+    /* The final dictionary content must be at least as large as the largest repcode */
+    size_t const minContentSize = (size_t)ZDICT_maxRep(repStartValue);
+    size_t paddingSize;
 
     /* check conditions */
     DEBUGLOG(4, "ZDICT_finalizeDictionary");
     if (dictBufferCapacity < dictContentSize) return ERROR(dstSize_tooSmall);
-    if (dictContentSize < ZDICT_CONTENTSIZE_MIN) return ERROR(srcSize_wrong);
     if (dictBufferCapacity < ZDICT_DICTSIZE_MIN) return ERROR(dstSize_tooSmall);
 
     /* dictionary header */
@@ -927,12 +967,43 @@ size_t ZDICT_finalizeDictionary(void* dictBuffer, size_t dictBufferCapacity,
         hSize += eSize;
     }
 
-    /* copy elements in final buffer ; note : src and dst buffer can overlap */
+    /* Shrink the content size if it doesn't fit in the buffer */
-    if (hSize + dictContentSize > dictBufferCapacity) dictContentSize = dictBufferCapacity - hSize;
+    if (hSize + dictContentSize > dictBufferCapacity) {
-    {   size_t const dictSize = hSize + dictContentSize;
+        dictContentSize = dictBufferCapacity - hSize;
-        char* dictEnd = (char*)dictBuffer + dictSize;
+    }
-        memmove(dictEnd - dictContentSize, customDictContent, dictContentSize);
+
-        memcpy(dictBuffer, header, hSize);
+    /* Pad the dictionary content with zeros if it is too small */
+    if (dictContentSize < minContentSize) {
+        RETURN_ERROR_IF(hSize + minContentSize > dictBufferCapacity, dstSize_tooSmall,
+                        "dictBufferCapacity too small to fit max repcode");
+        paddingSize = minContentSize - dictContentSize;
+    } else {
+        paddingSize = 0;
+    }
+
+    {
+        size_t const dictSize = hSize + paddingSize + dictContentSize;
+
+        /* The dictionary consists of the header, optional padding, and the content.
+         * The padding comes before the content because the "best" position in the
+         * dictionary is the last byte.
+         */
+        BYTE* const outDictHeader = (BYTE*)dictBuffer;
+        BYTE* const outDictPadding = outDictHeader + hSize;
+        BYTE* const outDictContent = outDictPadding + paddingSize;
+
+        assert(dictSize <= dictBufferCapacity);
+        assert(outDictContent + dictContentSize == (BYTE*)dictBuffer + dictSize);
+
+        /* First copy the customDictContent into its final location.
+         * `customDictContent` and `dictBuffer` may overlap, so we must
+         * do this before any other writes into the output buffer.
+         * Then copy the header & padding into the output buffer.
+         */
+        memmove(outDictContent, customDictContent, dictContentSize);
+        memcpy(outDictHeader, header, hSize);
+        memset(outDictPadding, 0, paddingSize);
+
         return dictSize;
     }
 }

zstd/zdict.h

@@ -46,7 +46,7 @@ extern "C" {
  *
  * Zstd can use dictionaries to improve compression ratio of small data.
  * Traditionally small files don't compress well because there is very little
- * repetion in a single sample, since it is small. But, if you are compressing
+ * repetition in a single sample, since it is small. But, if you are compressing
  * many similar files, like a bunch of JSON records that share the same
  * structure, you can train a dictionary on ahead of time on some samples of
  * these files. Then, zstd can use the dictionary to find repetitions that are
@@ -132,7 +132,7 @@ extern "C" {
  *
  *   # Benchmark levels 1-3 without a dictionary
  *   zstd -b1e3 -r /path/to/my/files
- *   # Benchmark levels 1-3 with a dictioanry
+ *   # Benchmark levels 1-3 with a dictionary
  *   zstd -b1e3 -r /path/to/my/files -D /path/to/my/dictionary
  *
  * When should I retrain a dictionary?
@@ -237,7 +237,6 @@ typedef struct {
  * is presumed that the most profitable content is at the end of the dictionary,
  * since that is the cheapest to reference.
  *
- * `dictContentSize` must be >= ZDICT_CONTENTSIZE_MIN bytes.
  * `maxDictSize` must be >= max(dictContentSize, ZSTD_DICTSIZE_MIN).
  *
  * @return: size of dictionary stored into `dstDictBuffer` (<= `maxDictSize`),
@@ -272,8 +271,9 @@ ZDICTLIB_API const char* ZDICT_getErrorName(size_t errorCode);
  * Use them only in association with static linking.
  * ==================================================================================== */
 
-#define ZDICT_CONTENTSIZE_MIN 128
 #define ZDICT_DICTSIZE_MIN    256
+/* Deprecated: Remove in v1.6.0 */
+#define ZDICT_CONTENTSIZE_MIN 128
 
 /*! ZDICT_cover_params_t:
  *  k and d are the only required parameters.

zstd/zstd.h

@@ -20,19 +20,21 @@ extern "C" {
 
 
 /* =====   ZSTDLIB_API : control library symbols visibility   ===== */
-#ifndef ZSTDLIB_VISIBILITY
+#ifndef ZSTDLIB_VISIBLE
-#  if defined(__GNUC__) && (__GNUC__ >= 4)
+#  if defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
-#    define ZSTDLIB_VISIBILITY __attribute__ ((visibility ("default")))
+#    define ZSTDLIB_VISIBLE __attribute__ ((visibility ("default")))
+#    define ZSTDLIB_HIDDEN __attribute__ ((visibility ("hidden")))
 #  else
-#    define ZSTDLIB_VISIBILITY
+#    define ZSTDLIB_VISIBLE
+#    define ZSTDLIB_HIDDEN
 #  endif
 #endif
 #if defined(ZSTD_DLL_EXPORT) && (ZSTD_DLL_EXPORT==1)
-#  define ZSTDLIB_API __declspec(dllexport) ZSTDLIB_VISIBILITY
+#  define ZSTDLIB_API __declspec(dllexport) ZSTDLIB_VISIBLE
 #elif defined(ZSTD_DLL_IMPORT) && (ZSTD_DLL_IMPORT==1)
-#  define ZSTDLIB_API __declspec(dllimport) ZSTDLIB_VISIBILITY /* It isn't required but allows to generate better code, saving a function pointer load from the IAT and an indirect jump.*/
+#  define ZSTDLIB_API __declspec(dllimport) ZSTDLIB_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 ZSTDLIB_API ZSTDLIB_VISIBILITY
+#  define ZSTDLIB_API ZSTDLIB_VISIBLE
 #endif
 
 
@@ -72,7 +74,7 @@ extern "C" {
 /*------   Version   ------*/
 #define ZSTD_VERSION_MAJOR    1
 #define ZSTD_VERSION_MINOR    5
-#define ZSTD_VERSION_RELEASE  0
+#define ZSTD_VERSION_RELEASE  1
 #define ZSTD_VERSION_NUMBER  (ZSTD_VERSION_MAJOR *100*100 + ZSTD_VERSION_MINOR *100 + ZSTD_VERSION_RELEASE)
 
 /*! ZSTD_versionNumber() :
@@ -247,7 +249,7 @@ ZSTDLIB_API size_t ZSTD_decompressDCtx(ZSTD_DCtx* dctx,
  *
  *   It's possible to reset all parameters to "default" using ZSTD_CCtx_reset().
  *
- *   This API supercedes all other "advanced" API entry points in the experimental section.
+ *   This API supersedes all other "advanced" API entry points in the experimental section.
  *   In the future, we expect to remove from experimental API entry points which are redundant with this API.
  */
 
@@ -417,7 +419,7 @@ typedef enum {
      * ZSTD_c_stableOutBuffer
      * ZSTD_c_blockDelimiters
      * ZSTD_c_validateSequences
-     * ZSTD_c_splitBlocks
+     * ZSTD_c_useBlockSplitter
      * ZSTD_c_useRowMatchFinder
      * Because they are not stable, it's necessary to define ZSTD_STATIC_LINKING_ONLY to access them.
      * note : never ever use experimentalParam? names directly;
@@ -932,7 +934,7 @@ ZSTDLIB_API unsigned ZSTD_getDictID_fromFrame(const void* src, size_t srcSize);
  * Advanced dictionary and prefix API (Requires v1.4.0+)
  *
  * This API allows dictionaries to be used with ZSTD_compress2(),
- * ZSTD_compressStream2(), and ZSTD_decompress(). Dictionaries are sticky, and
+ * ZSTD_compressStream2(), and ZSTD_decompressDCtx(). Dictionaries are sticky, and
  * only reset with the context is reset with ZSTD_reset_parameters or
  * ZSTD_reset_session_and_parameters. Prefixes are single-use.
  ******************************************************************************/
@@ -1073,25 +1075,36 @@ ZSTDLIB_API size_t ZSTD_sizeof_DDict(const ZSTD_DDict* ddict);
 #if defined(ZSTD_STATIC_LINKING_ONLY) && !defined(ZSTD_H_ZSTD_STATIC_LINKING_ONLY)
 #define ZSTD_H_ZSTD_STATIC_LINKING_ONLY
 
+/* This can be overridden externally to hide static symbols. */
+#ifndef ZSTDLIB_STATIC_API
+#  if defined(ZSTD_DLL_EXPORT) && (ZSTD_DLL_EXPORT==1)
+#    define ZSTDLIB_STATIC_API __declspec(dllexport) ZSTDLIB_VISIBLE
+#  elif defined(ZSTD_DLL_IMPORT) && (ZSTD_DLL_IMPORT==1)
+#    define ZSTDLIB_STATIC_API __declspec(dllimport) ZSTDLIB_VISIBLE
+#  else
+#    define ZSTDLIB_STATIC_API ZSTDLIB_VISIBLE
+#  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_API  /* disable deprecation 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_API
+#    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_API __attribute__((deprecated(message)))
+#    define ZSTD_DEPRECATED(message) ZSTDLIB_STATIC_API __attribute__((deprecated(message)))
 #  elif defined(__GNUC__) && (__GNUC__ >= 3)
-#    define ZSTD_DEPRECATED(message) ZSTDLIB_API __attribute__((deprecated))
+#    define ZSTD_DEPRECATED(message) ZSTDLIB_STATIC_API __attribute__((deprecated))
 #  elif defined(_MSC_VER)
-#    define ZSTD_DEPRECATED(message) ZSTDLIB_API __declspec(deprecated(message))
+#    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_API
+#    define ZSTD_DEPRECATED(message) ZSTDLIB_STATIC_API
 #  endif
 #endif /* ZSTD_DISABLE_DEPRECATE_WARNINGS */
 
@@ -1157,9 +1170,6 @@ ZSTDLIB_API size_t ZSTD_sizeof_DDict(const ZSTD_DDict* ddict);
 #define ZSTD_SRCSIZEHINT_MIN        0
 #define ZSTD_SRCSIZEHINT_MAX        INT_MAX
 
-/* internal */
-#define ZSTD_HASHLOG3_MAX           17
-
 
 /* ---  Advanced types  --- */
 
@@ -1302,10 +1312,14 @@ typedef enum {
 } ZSTD_literalCompressionMode_e;
 
 typedef enum {
-  ZSTD_urm_auto = 0,                   /* Automatically determine whether or not we use row matchfinder */
+  /* Note: This enum controls features which are conditionally beneficial. Zstd typically will make a final
-  ZSTD_urm_disableRowMatchFinder = 1,  /* Never use row matchfinder */
+   * decision on whether or not to enable the feature (ZSTD_ps_auto), but setting the switch to ZSTD_ps_enable
-  ZSTD_urm_enableRowMatchFinder = 2    /* Always use row matchfinder when applicable */
+   * or ZSTD_ps_disable allow for a force enable/disable the feature.
-} ZSTD_useRowMatchFinderMode_e;
+   */
+  ZSTD_ps_auto = 0,         /* Let the library automatically determine whether the feature shall be enabled */
+  ZSTD_ps_enable = 1,       /* Force-enable the feature */
+  ZSTD_ps_disable = 2       /* Do not use the feature */
+} ZSTD_paramSwitch_e;
 
 /***************************************
 *  Frame size functions
@@ -1332,7 +1346,7 @@ typedef enum {
  *   note 5 : ZSTD_findDecompressedSize handles multiple frames, and so it must traverse the input to
  *            read each contained frame header.  This is fast as most of the data is skipped,
  *            however it does mean that all frame data must be present and valid. */
-ZSTDLIB_API unsigned long long ZSTD_findDecompressedSize(const void* src, size_t srcSize);
+ZSTDLIB_STATIC_API unsigned long long ZSTD_findDecompressedSize(const void* src, size_t srcSize);
 
 /*! ZSTD_decompressBound() :
  *  `src` should point to the start of a series of ZSTD encoded and/or skippable frames
@@ -1347,13 +1361,13 @@ ZSTDLIB_API unsigned long long ZSTD_findDecompressedSize(const void* src, size_t
  *  note 3  : when the decompressed size field isn't available, the upper-bound for that frame is calculated by:
  *              upper-bound = # blocks * min(128 KB, Window_Size)
  */
-ZSTDLIB_API unsigned long long ZSTD_decompressBound(const void* src, size_t srcSize);
+ZSTDLIB_STATIC_API unsigned long long ZSTD_decompressBound(const void* src, size_t srcSize);
 
 /*! ZSTD_frameHeaderSize() :
  *  srcSize must be >= ZSTD_FRAMEHEADERSIZE_PREFIX.
  * @return : size of the Frame Header,
  *           or an error code (if srcSize is too small) */
-ZSTDLIB_API size_t ZSTD_frameHeaderSize(const void* src, size_t srcSize);
+ZSTDLIB_STATIC_API size_t ZSTD_frameHeaderSize(const void* src, size_t srcSize);
 
 typedef enum {
   ZSTD_sf_noBlockDelimiters = 0,         /* Representation of ZSTD_Sequence has no block delimiters, sequences only */
@@ -1376,7 +1390,7 @@ typedef enum {
  * @return : number of sequences generated
  */
 
-ZSTDLIB_API size_t ZSTD_generateSequences(ZSTD_CCtx* zc, ZSTD_Sequence* outSeqs,
+ZSTDLIB_STATIC_API size_t ZSTD_generateSequences(ZSTD_CCtx* zc, ZSTD_Sequence* outSeqs,
                                           size_t outSeqsSize, const void* src, size_t srcSize);
 
 /*! ZSTD_mergeBlockDelimiters() :
@@ -1390,7 +1404,7 @@ ZSTDLIB_API size_t ZSTD_generateSequences(ZSTD_CCtx* zc, ZSTD_Sequence* outSeqs,
  * setting of ZSTD_c_blockDelimiters as ZSTD_sf_noBlockDelimiters
  * @return : number of sequences left after merging
  */
-ZSTDLIB_API size_t ZSTD_mergeBlockDelimiters(ZSTD_Sequence* sequences, size_t seqsSize);
+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.
@@ -1420,7 +1434,7 @@ ZSTDLIB_API size_t ZSTD_mergeBlockDelimiters(ZSTD_Sequence* sequences, size_t se
  *         and cannot emit an RLE block that disagrees with the repcode history
  * @return : final compressed size or a ZSTD error.
  */
-ZSTDLIB_API size_t ZSTD_compressSequences(ZSTD_CCtx* const cctx, void* dst, size_t dstSize,
+ZSTDLIB_STATIC_API size_t ZSTD_compressSequences(ZSTD_CCtx* const cctx, void* dst, size_t dstSize,
                                   const ZSTD_Sequence* inSeqs, size_t inSeqsSize,
                                   const void* src, size_t srcSize);
 
@@ -1438,9 +1452,29 @@ ZSTDLIB_API size_t ZSTD_compressSequences(ZSTD_CCtx* const cctx, void* dst, size
  *
  * @return : number of bytes written or a ZSTD error.
  */
-ZSTDLIB_API size_t ZSTD_writeSkippableFrame(void* dst, size_t dstCapacity,
+ZSTDLIB_STATIC_API size_t ZSTD_writeSkippableFrame(void* dst, size_t dstCapacity,
                                             const void* src, size_t srcSize, unsigned magicVariant);
 
+/*! ZSTD_readSkippableFrame() :
+ * Retrieves a zstd skippable frame containing data given by src, and writes it to dst buffer.
+ *
+ * The parameter magicVariant will receive the magicVariant that was supplied when the frame was written,
+ * i.e. magicNumber - ZSTD_MAGIC_SKIPPABLE_START.  This can be NULL if the caller is not interested
+ * in the magicVariant.
+ *
+ * Returns an error if destination buffer is not large enough, or if the frame is not skippable.
+ *
+ * @return : number of bytes written or a ZSTD error.
+ */
+ZSTDLIB_API size_t ZSTD_readSkippableFrame(void* dst, size_t dstCapacity, unsigned* magicVariant,
+                                            const void* src, size_t srcSize);
+
+/*! ZSTD_isSkippableFrame() :
+ *  Tells if the content of `buffer` starts with a valid Frame Identifier for a skippable frame.
+ */
+ZSTDLIB_API unsigned ZSTD_isSkippableFrame(const void* buffer, size_t size);
+
+
 
 /***************************************
 *  Memory management
@@ -1469,10 +1503,10 @@ ZSTDLIB_API size_t ZSTD_writeSkippableFrame(void* dst, size_t dstCapacity,
  *  Note 2 : only single-threaded compression is supported.
  *  ZSTD_estimateCCtxSize_usingCCtxParams() will return an error code if ZSTD_c_nbWorkers is >= 1.
  */
-ZSTDLIB_API size_t ZSTD_estimateCCtxSize(int compressionLevel);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateCCtxSize(int compressionLevel);
-ZSTDLIB_API size_t ZSTD_estimateCCtxSize_usingCParams(ZSTD_compressionParameters cParams);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateCCtxSize_usingCParams(ZSTD_compressionParameters cParams);
-ZSTDLIB_API size_t ZSTD_estimateCCtxSize_usingCCtxParams(const ZSTD_CCtx_params* params);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateCCtxSize_usingCCtxParams(const ZSTD_CCtx_params* params);
-ZSTDLIB_API size_t ZSTD_estimateDCtxSize(void);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateDCtxSize(void);
 
 /*! ZSTD_estimateCStreamSize() :
  *  ZSTD_estimateCStreamSize() will provide a budget large enough for any compression level up to selected one.
@@ -1487,20 +1521,20 @@ ZSTDLIB_API size_t ZSTD_estimateDCtxSize(void);
  *  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 */
-ZSTDLIB_API size_t ZSTD_estimateCStreamSize(int compressionLevel);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateCStreamSize(int compressionLevel);
-ZSTDLIB_API size_t ZSTD_estimateCStreamSize_usingCParams(ZSTD_compressionParameters cParams);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateCStreamSize_usingCParams(ZSTD_compressionParameters cParams);
-ZSTDLIB_API size_t ZSTD_estimateCStreamSize_usingCCtxParams(const ZSTD_CCtx_params* params);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateCStreamSize_usingCCtxParams(const ZSTD_CCtx_params* params);
-ZSTDLIB_API size_t ZSTD_estimateDStreamSize(size_t windowSize);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateDStreamSize(size_t windowSize);
-ZSTDLIB_API size_t ZSTD_estimateDStreamSize_fromFrame(const void* src, size_t srcSize);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateDStreamSize_fromFrame(const void* src, size_t srcSize);
 
 /*! ZSTD_estimate?DictSize() :
  *  ZSTD_estimateCDictSize() will bet that src size is relatively "small", and content is copied, like ZSTD_createCDict().
  *  ZSTD_estimateCDictSize_advanced() makes it possible to control compression parameters precisely, like ZSTD_createCDict_advanced().
  *  Note : dictionaries created by reference (`ZSTD_dlm_byRef`) are logically smaller.
  */
-ZSTDLIB_API size_t ZSTD_estimateCDictSize(size_t dictSize, int compressionLevel);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateCDictSize(size_t dictSize, int compressionLevel);
-ZSTDLIB_API size_t ZSTD_estimateCDictSize_advanced(size_t dictSize, ZSTD_compressionParameters cParams, ZSTD_dictLoadMethod_e dictLoadMethod);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateCDictSize_advanced(size_t dictSize, ZSTD_compressionParameters cParams, ZSTD_dictLoadMethod_e dictLoadMethod);
-ZSTDLIB_API size_t ZSTD_estimateDDictSize(size_t dictSize, ZSTD_dictLoadMethod_e dictLoadMethod);
+ZSTDLIB_STATIC_API size_t ZSTD_estimateDDictSize(size_t dictSize, ZSTD_dictLoadMethod_e dictLoadMethod);
 
 /*! ZSTD_initStatic*() :
  *  Initialize an object using a pre-allocated fixed-size buffer.
@@ -1523,20 +1557,20 @@ ZSTDLIB_API size_t ZSTD_estimateDDictSize(size_t dictSize, ZSTD_dictLoadMethod_e
  *  Limitation 2 : static cctx currently not compatible with multi-threading.
  *  Limitation 3 : static dctx is incompatible with legacy support.
  */
-ZSTDLIB_API ZSTD_CCtx*    ZSTD_initStaticCCtx(void* workspace, size_t workspaceSize);
+ZSTDLIB_STATIC_API ZSTD_CCtx*    ZSTD_initStaticCCtx(void* workspace, size_t workspaceSize);
-ZSTDLIB_API ZSTD_CStream* ZSTD_initStaticCStream(void* workspace, size_t workspaceSize);    /**< same as ZSTD_initStaticCCtx() */
+ZSTDLIB_STATIC_API ZSTD_CStream* ZSTD_initStaticCStream(void* workspace, size_t workspaceSize);    /**< same as ZSTD_initStaticCCtx() */
 
-ZSTDLIB_API ZSTD_DCtx*    ZSTD_initStaticDCtx(void* workspace, size_t workspaceSize);
+ZSTDLIB_STATIC_API ZSTD_DCtx*    ZSTD_initStaticDCtx(void* workspace, size_t workspaceSize);
-ZSTDLIB_API ZSTD_DStream* ZSTD_initStaticDStream(void* workspace, size_t workspaceSize);    /**< same as ZSTD_initStaticDCtx() */
+ZSTDLIB_STATIC_API ZSTD_DStream* ZSTD_initStaticDStream(void* workspace, size_t workspaceSize);    /**< same as ZSTD_initStaticDCtx() */
 
-ZSTDLIB_API const ZSTD_CDict* ZSTD_initStaticCDict(
+ZSTDLIB_STATIC_API const ZSTD_CDict* ZSTD_initStaticCDict(
                                         void* workspace, size_t workspaceSize,
                                         const void* dict, size_t dictSize,
                                         ZSTD_dictLoadMethod_e dictLoadMethod,
                                         ZSTD_dictContentType_e dictContentType,
                                         ZSTD_compressionParameters cParams);
 
-ZSTDLIB_API const ZSTD_DDict* ZSTD_initStaticDDict(
+ZSTDLIB_STATIC_API const ZSTD_DDict* ZSTD_initStaticDDict(
                                         void* workspace, size_t workspaceSize,
                                         const void* dict, size_t dictSize,
                                         ZSTD_dictLoadMethod_e dictLoadMethod,
@@ -1557,12 +1591,12 @@ __attribute__((__unused__))
 #endif
 ZSTD_customMem const ZSTD_defaultCMem = { NULL, NULL, NULL };  /**< this constant defers to stdlib's functions */
 
-ZSTDLIB_API ZSTD_CCtx*    ZSTD_createCCtx_advanced(ZSTD_customMem customMem);
+ZSTDLIB_STATIC_API ZSTD_CCtx*    ZSTD_createCCtx_advanced(ZSTD_customMem customMem);
-ZSTDLIB_API ZSTD_CStream* ZSTD_createCStream_advanced(ZSTD_customMem customMem);
+ZSTDLIB_STATIC_API ZSTD_CStream* ZSTD_createCStream_advanced(ZSTD_customMem customMem);
-ZSTDLIB_API ZSTD_DCtx*    ZSTD_createDCtx_advanced(ZSTD_customMem customMem);
+ZSTDLIB_STATIC_API ZSTD_DCtx*    ZSTD_createDCtx_advanced(ZSTD_customMem customMem);
-ZSTDLIB_API ZSTD_DStream* ZSTD_createDStream_advanced(ZSTD_customMem customMem);
+ZSTDLIB_STATIC_API ZSTD_DStream* ZSTD_createDStream_advanced(ZSTD_customMem customMem);
 
-ZSTDLIB_API ZSTD_CDict* ZSTD_createCDict_advanced(const void* dict, size_t dictSize,
+ZSTDLIB_STATIC_API ZSTD_CDict* ZSTD_createCDict_advanced(const void* dict, size_t dictSize,
                                                   ZSTD_dictLoadMethod_e dictLoadMethod,
                                                   ZSTD_dictContentType_e dictContentType,
                                                   ZSTD_compressionParameters cParams,
@@ -1579,22 +1613,22 @@ ZSTDLIB_API ZSTD_CDict* ZSTD_createCDict_advanced(const void* dict, size_t dictS
  *  ZSTD_freeThreadPool frees a thread pool, accepts NULL pointer.
  */
 typedef struct POOL_ctx_s ZSTD_threadPool;
-ZSTDLIB_API ZSTD_threadPool* ZSTD_createThreadPool(size_t numThreads);
+ZSTDLIB_STATIC_API ZSTD_threadPool* ZSTD_createThreadPool(size_t numThreads);
-ZSTDLIB_API void ZSTD_freeThreadPool (ZSTD_threadPool* pool);  /* accept NULL pointer */
+ZSTDLIB_STATIC_API void ZSTD_freeThreadPool (ZSTD_threadPool* pool);  /* accept NULL pointer */
-ZSTDLIB_API size_t ZSTD_CCtx_refThreadPool(ZSTD_CCtx* cctx, ZSTD_threadPool* pool);
+ZSTDLIB_STATIC_API size_t ZSTD_CCtx_refThreadPool(ZSTD_CCtx* cctx, ZSTD_threadPool* pool);
 
 
 /*
  * This API is temporary and is expected to change or disappear in the future!
  */
-ZSTDLIB_API ZSTD_CDict* ZSTD_createCDict_advanced2(
+ZSTDLIB_STATIC_API ZSTD_CDict* ZSTD_createCDict_advanced2(
     const void* dict, size_t dictSize,
     ZSTD_dictLoadMethod_e dictLoadMethod,
     ZSTD_dictContentType_e dictContentType,
     const ZSTD_CCtx_params* cctxParams,
     ZSTD_customMem customMem);
 
-ZSTDLIB_API ZSTD_DDict* ZSTD_createDDict_advanced(
+ZSTDLIB_STATIC_API ZSTD_DDict* ZSTD_createDDict_advanced(
     const void* dict, size_t dictSize,
     ZSTD_dictLoadMethod_e dictLoadMethod,
     ZSTD_dictContentType_e dictContentType,
@@ -1611,22 +1645,22 @@ ZSTDLIB_API ZSTD_DDict* ZSTD_createDDict_advanced(
  *  As a consequence, `dictBuffer` **must** outlive CDict,
  *  and its content must remain unmodified throughout the lifetime of CDict.
  *  note: equivalent to ZSTD_createCDict_advanced(), with dictLoadMethod==ZSTD_dlm_byRef */
-ZSTDLIB_API ZSTD_CDict* ZSTD_createCDict_byReference(const void* dictBuffer, size_t dictSize, int compressionLevel);
+ZSTDLIB_STATIC_API ZSTD_CDict* ZSTD_createCDict_byReference(const void* dictBuffer, size_t dictSize, int compressionLevel);
 
 /*! ZSTD_getCParams() :
  * @return ZSTD_compressionParameters structure for a selected compression level and estimated srcSize.
  * `estimatedSrcSize` value is optional, select 0 if not known */
-ZSTDLIB_API ZSTD_compressionParameters ZSTD_getCParams(int compressionLevel, unsigned long long estimatedSrcSize, size_t dictSize);
+ZSTDLIB_STATIC_API ZSTD_compressionParameters ZSTD_getCParams(int compressionLevel, unsigned long long estimatedSrcSize, size_t dictSize);
 
 /*! ZSTD_getParams() :
  *  same as ZSTD_getCParams(), but @return a full `ZSTD_parameters` object instead of sub-component `ZSTD_compressionParameters`.
  *  All fields of `ZSTD_frameParameters` are set to default : contentSize=1, checksum=0, noDictID=0 */
-ZSTDLIB_API ZSTD_parameters ZSTD_getParams(int compressionLevel, unsigned long long estimatedSrcSize, size_t dictSize);
+ZSTDLIB_STATIC_API ZSTD_parameters ZSTD_getParams(int compressionLevel, unsigned long long estimatedSrcSize, size_t dictSize);
 
 /*! ZSTD_checkCParams() :
  *  Ensure param values remain within authorized range.
  * @return 0 on success, or an error code (can be checked with ZSTD_isError()) */
-ZSTDLIB_API size_t ZSTD_checkCParams(ZSTD_compressionParameters params);
+ZSTDLIB_STATIC_API size_t ZSTD_checkCParams(ZSTD_compressionParameters params);
 
 /*! ZSTD_adjustCParams() :
  *  optimize params for a given `srcSize` and `dictSize`.
@@ -1634,7 +1668,7 @@ ZSTDLIB_API size_t ZSTD_checkCParams(ZSTD_compressionParameters params);
  * `dictSize` must be `0` when there is no dictionary.
  *  cPar can be invalid : all parameters will be clamped within valid range in the @return struct.
  *  This function never fails (wide contract) */
-ZSTDLIB_API ZSTD_compressionParameters ZSTD_adjustCParams(ZSTD_compressionParameters cPar, unsigned long long srcSize, size_t dictSize);
+ZSTDLIB_STATIC_API ZSTD_compressionParameters ZSTD_adjustCParams(ZSTD_compressionParameters cPar, unsigned long long srcSize, size_t dictSize);
 
 /*! ZSTD_compress_advanced() :
  *  Note : this function is now DEPRECATED.
@@ -1662,18 +1696,18 @@ size_t ZSTD_compress_usingCDict_advanced(ZSTD_CCtx* cctx,
 /*! ZSTD_CCtx_loadDictionary_byReference() :
  *  Same as ZSTD_CCtx_loadDictionary(), but dictionary content is referenced, instead of being copied into CCtx.
  *  It saves some memory, but also requires that `dict` outlives its usage within `cctx` */
-ZSTDLIB_API size_t ZSTD_CCtx_loadDictionary_byReference(ZSTD_CCtx* cctx, const void* dict, size_t dictSize);
+ZSTDLIB_STATIC_API size_t ZSTD_CCtx_loadDictionary_byReference(ZSTD_CCtx* cctx, const void* dict, size_t dictSize);
 
 /*! ZSTD_CCtx_loadDictionary_advanced() :
  *  Same as ZSTD_CCtx_loadDictionary(), but gives finer control over
  *  how to load the dictionary (by copy ? by reference ?)
  *  and how to interpret it (automatic ? force raw mode ? full mode only ?) */
-ZSTDLIB_API size_t ZSTD_CCtx_loadDictionary_advanced(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, ZSTD_dictLoadMethod_e dictLoadMethod, ZSTD_dictContentType_e dictContentType);
+ZSTDLIB_STATIC_API size_t ZSTD_CCtx_loadDictionary_advanced(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, ZSTD_dictLoadMethod_e dictLoadMethod, ZSTD_dictContentType_e dictContentType);
 
 /*! ZSTD_CCtx_refPrefix_advanced() :
  *  Same as ZSTD_CCtx_refPrefix(), but gives finer control over
  *  how to interpret prefix content (automatic ? force raw mode (default) ? full mode only ?) */
-ZSTDLIB_API size_t ZSTD_CCtx_refPrefix_advanced(ZSTD_CCtx* cctx, const void* prefix, size_t prefixSize, ZSTD_dictContentType_e dictContentType);
+ZSTDLIB_STATIC_API size_t ZSTD_CCtx_refPrefix_advanced(ZSTD_CCtx* cctx, const void* prefix, size_t prefixSize, ZSTD_dictContentType_e dictContentType);
 
 /* ===   experimental parameters   === */
 /* these parameters can be used with ZSTD_setParameter()
@@ -1712,9 +1746,15 @@ ZSTDLIB_API size_t ZSTD_CCtx_refPrefix_advanced(ZSTD_CCtx* cctx, const void* pre
  * See the comments on that enum for an explanation of the feature. */
 #define ZSTD_c_forceAttachDict ZSTD_c_experimentalParam4
 
-/* Controls how the literals are compressed (default is auto).
+/* Controlled with ZSTD_paramSwitch_e enum.
- * The value must be of type ZSTD_literalCompressionMode_e.
+ * Default is ZSTD_ps_auto.
- * See ZSTD_literalCompressionMode_e enum definition for details.
+ * Set to ZSTD_ps_disable to never compress literals.
+ * Set to ZSTD_ps_enable to always compress literals. (Note: uncompressed literals
+ * may still be emitted if huffman is not beneficial to use.)
+ *
+ * By default, in ZSTD_ps_auto, the library will decide at runtime whether to use
+ * literals compression based on the compression parameters - specifically,
+ * negative compression levels do not use literal compression.
  */
 #define ZSTD_c_literalCompressionMode ZSTD_c_experimentalParam5
 
@@ -1777,7 +1817,7 @@ ZSTDLIB_API size_t ZSTD_CCtx_refPrefix_advanced(ZSTD_CCtx* cctx, const void* pre
  *
  * Note that this means that the CDict tables can no longer be copied into the
  * CCtx, so the dict attachment mode ZSTD_dictForceCopy will no longer be
- * useable. The dictionary can only be attached or reloaded.
+ * usable. The dictionary can only be attached or reloaded.
  *
  * In general, you should expect compression to be faster--sometimes very much
  * so--and CDict creation to be slightly slower. Eventually, we will probably
@@ -1866,23 +1906,26 @@ ZSTDLIB_API size_t ZSTD_CCtx_refPrefix_advanced(ZSTD_CCtx* cctx, const void* pre
  */
 #define ZSTD_c_validateSequences ZSTD_c_experimentalParam12
 
-/* ZSTD_c_splitBlocks
+/* ZSTD_c_useBlockSplitter
- * Default is 0 == disabled. Set to 1 to enable block splitting.
+ * Controlled with ZSTD_paramSwitch_e enum.
+ * Default is ZSTD_ps_auto.
+ * Set to ZSTD_ps_disable to never use block splitter.
+ * Set to ZSTD_ps_enable to always use block splitter.
  *
- * Will attempt to split blocks in order to improve compression ratio at the cost of speed.
+ * By default, in ZSTD_ps_auto, the library will decide at runtime whether to use
+ * block splitting based on the compression parameters.
  */
-#define ZSTD_c_splitBlocks ZSTD_c_experimentalParam13
+#define ZSTD_c_useBlockSplitter ZSTD_c_experimentalParam13
 
 /* ZSTD_c_useRowMatchFinder
- * Default is ZSTD_urm_auto.
+ * Controlled with ZSTD_paramSwitch_e enum.
- * Controlled with ZSTD_useRowMatchFinderMode_e enum.
+ * Default is ZSTD_ps_auto.
+ * Set to ZSTD_ps_disable to never use row-based matchfinder.
+ * Set to ZSTD_ps_enable to force usage of row-based matchfinder.
  *
- * By default, in ZSTD_urm_auto, when finalizing the compression parameters, the library
+ * By default, in ZSTD_ps_auto, the library will decide at runtime whether to use
- * will decide at runtime whether to use the row-based matchfinder based on support for SIMD
+ * the row-based matchfinder based on support for SIMD instructions and the window log.
- * instructions as well as the windowLog.
+ * Note that this only pertains to compression strategies: greedy, lazy, and lazy2
- *
- * Set to ZSTD_urm_disableRowMatchFinder to never use row-based matchfinder.
- * Set to ZSTD_urm_enableRowMatchFinder to force usage of row-based matchfinder.
  */
 #define ZSTD_c_useRowMatchFinder ZSTD_c_experimentalParam14
 
@@ -1911,7 +1954,7 @@ ZSTDLIB_API size_t ZSTD_CCtx_refPrefix_advanced(ZSTD_CCtx* cctx, const void* pre
  *  and store it into int* value.
  * @return : 0, or an error code (which can be tested with ZSTD_isError()).
  */
-ZSTDLIB_API size_t ZSTD_CCtx_getParameter(const ZSTD_CCtx* cctx, ZSTD_cParameter param, int* value);
+ZSTDLIB_STATIC_API size_t ZSTD_CCtx_getParameter(const ZSTD_CCtx* cctx, ZSTD_cParameter param, int* value);
 
 
 /*! ZSTD_CCtx_params :
@@ -1931,25 +1974,25 @@ ZSTDLIB_API size_t ZSTD_CCtx_getParameter(const ZSTD_CCtx* cctx, ZSTD_cParameter
  *  This can be used with ZSTD_estimateCCtxSize_advanced_usingCCtxParams()
  *  for static allocation of CCtx for single-threaded compression.
  */
-ZSTDLIB_API ZSTD_CCtx_params* ZSTD_createCCtxParams(void);
+ZSTDLIB_STATIC_API ZSTD_CCtx_params* ZSTD_createCCtxParams(void);
-ZSTDLIB_API size_t ZSTD_freeCCtxParams(ZSTD_CCtx_params* params);  /* accept NULL pointer */
+ZSTDLIB_STATIC_API size_t ZSTD_freeCCtxParams(ZSTD_CCtx_params* params);  /* accept NULL pointer */
 
 /*! ZSTD_CCtxParams_reset() :
  *  Reset params to default values.
  */
-ZSTDLIB_API size_t ZSTD_CCtxParams_reset(ZSTD_CCtx_params* params);
+ZSTDLIB_STATIC_API size_t ZSTD_CCtxParams_reset(ZSTD_CCtx_params* params);
 
 /*! ZSTD_CCtxParams_init() :
  *  Initializes the compression parameters of cctxParams according to
  *  compression level. All other parameters are reset to their default values.
  */
-ZSTDLIB_API size_t ZSTD_CCtxParams_init(ZSTD_CCtx_params* cctxParams, int compressionLevel);
+ZSTDLIB_STATIC_API size_t ZSTD_CCtxParams_init(ZSTD_CCtx_params* cctxParams, int compressionLevel);
 
 /*! ZSTD_CCtxParams_init_advanced() :
  *  Initializes the compression and frame parameters of cctxParams according to
  *  params. All other parameters are reset to their default values.
  */
-ZSTDLIB_API size_t ZSTD_CCtxParams_init_advanced(ZSTD_CCtx_params* cctxParams, ZSTD_parameters params);
+ZSTDLIB_STATIC_API size_t ZSTD_CCtxParams_init_advanced(ZSTD_CCtx_params* cctxParams, ZSTD_parameters params);
 
 /*! ZSTD_CCtxParams_setParameter() : Requires v1.4.0+
  *  Similar to ZSTD_CCtx_setParameter.
@@ -1959,14 +2002,14 @@ ZSTDLIB_API size_t ZSTD_CCtxParams_init_advanced(ZSTD_CCtx_params* cctxParams, Z
  * @result : a code representing success or failure (which can be tested with
  *           ZSTD_isError()).
  */
-ZSTDLIB_API size_t ZSTD_CCtxParams_setParameter(ZSTD_CCtx_params* params, ZSTD_cParameter param, int value);
+ZSTDLIB_STATIC_API size_t ZSTD_CCtxParams_setParameter(ZSTD_CCtx_params* params, ZSTD_cParameter param, int value);
 
 /*! ZSTD_CCtxParams_getParameter() :
  * Similar to ZSTD_CCtx_getParameter.
  * Get the requested value of one compression parameter, selected by enum ZSTD_cParameter.
  * @result : 0, or an error code (which can be tested with ZSTD_isError()).
  */
-ZSTDLIB_API size_t ZSTD_CCtxParams_getParameter(const ZSTD_CCtx_params* params, ZSTD_cParameter param, int* value);
+ZSTDLIB_STATIC_API size_t ZSTD_CCtxParams_getParameter(const ZSTD_CCtx_params* params, ZSTD_cParameter param, int* value);
 
 /*! ZSTD_CCtx_setParametersUsingCCtxParams() :
  *  Apply a set of ZSTD_CCtx_params to the compression context.
@@ -1975,7 +2018,7 @@ ZSTDLIB_API size_t ZSTD_CCtxParams_getParameter(const ZSTD_CCtx_params* params,
  *    if nbWorkers>=1, new parameters will be picked up at next job,
  *       with a few restrictions (windowLog, pledgedSrcSize, nbWorkers, jobSize, and overlapLog are not updated).
  */
-ZSTDLIB_API size_t ZSTD_CCtx_setParametersUsingCCtxParams(
+ZSTDLIB_STATIC_API size_t ZSTD_CCtx_setParametersUsingCCtxParams(
         ZSTD_CCtx* cctx, const ZSTD_CCtx_params* params);
 
 /*! ZSTD_compressStream2_simpleArgs() :
@@ -1984,7 +2027,7 @@ ZSTDLIB_API size_t ZSTD_CCtx_setParametersUsingCCtxParams(
  *  This variant might be helpful for binders from dynamic languages
  *  which have troubles handling structures containing memory pointers.
  */
-ZSTDLIB_API size_t ZSTD_compressStream2_simpleArgs (
+ZSTDLIB_STATIC_API size_t ZSTD_compressStream2_simpleArgs (
                             ZSTD_CCtx* cctx,
                             void* dst, size_t dstCapacity, size_t* dstPos,
                       const void* src, size_t srcSize, size_t* srcPos,
@@ -2000,33 +2043,33 @@ ZSTDLIB_API size_t ZSTD_compressStream2_simpleArgs (
  *  Note : Frame Identifier is 4 bytes. If `size < 4`, @return will always be 0.
  *  Note 2 : Legacy Frame Identifiers are considered valid only if Legacy Support is enabled.
  *  Note 3 : Skippable Frame Identifiers are considered valid. */
-ZSTDLIB_API unsigned ZSTD_isFrame(const void* buffer, size_t size);
+ZSTDLIB_STATIC_API unsigned ZSTD_isFrame(const void* buffer, size_t size);
 
 /*! ZSTD_createDDict_byReference() :
  *  Create a digested dictionary, ready to start decompression operation without startup delay.
  *  Dictionary content is referenced, and therefore stays in dictBuffer.
  *  It is important that dictBuffer outlives DDict,
  *  it must remain read accessible throughout the lifetime of DDict */
-ZSTDLIB_API ZSTD_DDict* ZSTD_createDDict_byReference(const void* dictBuffer, size_t dictSize);
+ZSTDLIB_STATIC_API ZSTD_DDict* ZSTD_createDDict_byReference(const void* dictBuffer, size_t dictSize);
 
 /*! ZSTD_DCtx_loadDictionary_byReference() :
  *  Same as ZSTD_DCtx_loadDictionary(),
  *  but references `dict` content instead of copying it into `dctx`.
  *  This saves memory if `dict` remains around.,
  *  However, it's imperative that `dict` remains accessible (and unmodified) while being used, so it must outlive decompression. */
-ZSTDLIB_API size_t ZSTD_DCtx_loadDictionary_byReference(ZSTD_DCtx* dctx, const void* dict, size_t dictSize);
+ZSTDLIB_STATIC_API size_t ZSTD_DCtx_loadDictionary_byReference(ZSTD_DCtx* dctx, const void* dict, size_t dictSize);
 
 /*! ZSTD_DCtx_loadDictionary_advanced() :
  *  Same as ZSTD_DCtx_loadDictionary(),
  *  but gives direct control over
  *  how to load the dictionary (by copy ? by reference ?)
  *  and how to interpret it (automatic ? force raw mode ? full mode only ?). */
-ZSTDLIB_API size_t ZSTD_DCtx_loadDictionary_advanced(ZSTD_DCtx* dctx, const void* dict, size_t dictSize, ZSTD_dictLoadMethod_e dictLoadMethod, ZSTD_dictContentType_e dictContentType);
+ZSTDLIB_STATIC_API size_t ZSTD_DCtx_loadDictionary_advanced(ZSTD_DCtx* dctx, const void* dict, size_t dictSize, ZSTD_dictLoadMethod_e dictLoadMethod, ZSTD_dictContentType_e dictContentType);
 
 /*! ZSTD_DCtx_refPrefix_advanced() :
  *  Same as ZSTD_DCtx_refPrefix(), but gives finer control over
  *  how to interpret prefix content (automatic ? force raw mode (default) ? full mode only ?) */
-ZSTDLIB_API size_t ZSTD_DCtx_refPrefix_advanced(ZSTD_DCtx* dctx, const void* prefix, size_t prefixSize, ZSTD_dictContentType_e dictContentType);
+ZSTDLIB_STATIC_API size_t ZSTD_DCtx_refPrefix_advanced(ZSTD_DCtx* dctx, const void* prefix, size_t prefixSize, ZSTD_dictContentType_e dictContentType);
 
 /*! ZSTD_DCtx_setMaxWindowSize() :
  *  Refuses allocating internal buffers for frames requiring a window size larger than provided limit.
@@ -2035,14 +2078,14 @@ ZSTDLIB_API size_t ZSTD_DCtx_refPrefix_advanced(ZSTD_DCtx* dctx, const void* pre
  *  By default, a decompression context accepts all window sizes <= (1 << ZSTD_WINDOWLOG_LIMIT_DEFAULT)
  * @return : 0, or an error code (which can be tested using ZSTD_isError()).
  */
-ZSTDLIB_API size_t ZSTD_DCtx_setMaxWindowSize(ZSTD_DCtx* dctx, size_t maxWindowSize);
+ZSTDLIB_STATIC_API size_t ZSTD_DCtx_setMaxWindowSize(ZSTD_DCtx* dctx, size_t maxWindowSize);
 
 /*! ZSTD_DCtx_getParameter() :
  *  Get the requested decompression parameter value, selected by enum ZSTD_dParameter,
  *  and store it into int* value.
  * @return : 0, or an error code (which can be tested with ZSTD_isError()).
  */
-ZSTDLIB_API size_t ZSTD_DCtx_getParameter(ZSTD_DCtx* dctx, ZSTD_dParameter param, int* value);
+ZSTDLIB_STATIC_API size_t ZSTD_DCtx_getParameter(ZSTD_DCtx* dctx, ZSTD_dParameter param, int* value);
 
 /* ZSTD_d_format
  * experimental parameter,
@@ -2131,7 +2174,7 @@ size_t ZSTD_DCtx_setFormat(ZSTD_DCtx* dctx, ZSTD_format_e format);
  *  This can be helpful for binders from dynamic languages
  *  which have troubles handling structures containing memory pointers.
  */
-ZSTDLIB_API size_t ZSTD_decompressStream_simpleArgs (
+ZSTDLIB_STATIC_API size_t ZSTD_decompressStream_simpleArgs (
                             ZSTD_DCtx* dctx,
                             void* dst, size_t dstCapacity, size_t* dstPos,
                       const void* src, size_t srcSize, size_t* srcPos);
@@ -2270,7 +2313,7 @@ typedef struct {
  * Note : (ingested - consumed) is amount of input data buffered internally, not yet compressed.
  * Aggregates progression inside active worker threads.
  */
-ZSTDLIB_API ZSTD_frameProgression ZSTD_getFrameProgression(const ZSTD_CCtx* cctx);
+ZSTDLIB_STATIC_API ZSTD_frameProgression ZSTD_getFrameProgression(const ZSTD_CCtx* cctx);
 
 /*! ZSTD_toFlushNow() :
  *  Tell how many bytes are ready to be flushed immediately.
@@ -2285,7 +2328,7 @@ ZSTDLIB_API ZSTD_frameProgression ZSTD_getFrameProgression(const ZSTD_CCtx* cctx
  *    therefore flush speed is limited by production speed of oldest job
  *    irrespective of the speed of concurrent (and newer) jobs.
  */
-ZSTDLIB_API size_t ZSTD_toFlushNow(ZSTD_CCtx* cctx);
+ZSTDLIB_STATIC_API size_t ZSTD_toFlushNow(ZSTD_CCtx* cctx);
 
 
 /*=====   Advanced Streaming decompression functions  =====*/
@@ -2299,7 +2342,7 @@ ZSTDLIB_API size_t ZSTD_toFlushNow(ZSTD_CCtx* cctx);
  * 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
  */
-ZSTDLIB_API size_t ZSTD_initDStream_usingDict(ZSTD_DStream* zds, const void* dict, size_t dictSize);
+ZSTDLIB_STATIC_API size_t ZSTD_initDStream_usingDict(ZSTD_DStream* zds, const void* dict, size_t dictSize);
 
 /*!
  * This function is deprecated, and is equivalent to:
@@ -2310,7 +2353,7 @@ ZSTDLIB_API size_t ZSTD_initDStream_usingDict(ZSTD_DStream* zds, const void* dic
  * 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
  */
-ZSTDLIB_API size_t ZSTD_initDStream_usingDDict(ZSTD_DStream* zds, const ZSTD_DDict* ddict);
+ZSTDLIB_STATIC_API size_t ZSTD_initDStream_usingDDict(ZSTD_DStream* zds, const ZSTD_DDict* ddict);
 
 /*!
  * This function is deprecated, and is equivalent to:
@@ -2320,7 +2363,7 @@ ZSTDLIB_API size_t ZSTD_initDStream_usingDDict(ZSTD_DStream* zds, const ZSTD_DDi
  * 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
  */
-ZSTDLIB_API size_t ZSTD_resetDStream(ZSTD_DStream* zds);
+ZSTDLIB_STATIC_API size_t ZSTD_resetDStream(ZSTD_DStream* zds);
 
 
 /*********************************************************************
@@ -2362,13 +2405,13 @@ ZSTDLIB_API size_t ZSTD_resetDStream(ZSTD_DStream* zds);
 */
 
 /*=====   Buffer-less streaming compression functions  =====*/
-ZSTDLIB_API size_t ZSTD_compressBegin(ZSTD_CCtx* cctx, int compressionLevel);
+ZSTDLIB_STATIC_API size_t ZSTD_compressBegin(ZSTD_CCtx* cctx, int compressionLevel);
-ZSTDLIB_API size_t ZSTD_compressBegin_usingDict(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, int compressionLevel);
+ZSTDLIB_STATIC_API size_t ZSTD_compressBegin_usingDict(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, int compressionLevel);
-ZSTDLIB_API size_t ZSTD_compressBegin_usingCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict); /**< note: fails if cdict==NULL */
+ZSTDLIB_STATIC_API size_t ZSTD_compressBegin_usingCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict); /**< note: fails if cdict==NULL */
-ZSTDLIB_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_copyCCtx(ZSTD_CCtx* cctx, const ZSTD_CCtx* preparedCCtx, unsigned long long pledgedSrcSize); /**<  note: if pledgedSrcSize is not known, use ZSTD_CONTENTSIZE_UNKNOWN */
 
-ZSTDLIB_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_compressContinue(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
-ZSTDLIB_API size_t ZSTD_compressEnd(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")
@@ -2465,24 +2508,24 @@ typedef struct {
  * @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() */
-ZSTDLIB_API size_t ZSTD_getFrameHeader(ZSTD_frameHeader* zfhPtr, const void* src, size_t srcSize);   /**< doesn't consume input */
+ZSTDLIB_STATIC_API size_t ZSTD_getFrameHeader(ZSTD_frameHeader* zfhPtr, const void* src, size_t srcSize);   /**< doesn't consume input */
 /*! ZSTD_getFrameHeader_advanced() :
  *  same as ZSTD_getFrameHeader(),
  *  with added capability to select a format (like ZSTD_f_zstd1_magicless) */
-ZSTDLIB_API size_t ZSTD_getFrameHeader_advanced(ZSTD_frameHeader* zfhPtr, const void* src, size_t srcSize, ZSTD_format_e format);
+ZSTDLIB_STATIC_API size_t ZSTD_getFrameHeader_advanced(ZSTD_frameHeader* zfhPtr, const void* src, size_t srcSize, ZSTD_format_e format);
-ZSTDLIB_API size_t ZSTD_decodingBufferSize_min(unsigned long long windowSize, unsigned long long frameContentSize);  /**< when frame content size is not known, pass in frameContentSize == ZSTD_CONTENTSIZE_UNKNOWN */
+ZSTDLIB_STATIC_API size_t ZSTD_decodingBufferSize_min(unsigned long long windowSize, unsigned long long frameContentSize);  /**< when frame content size is not known, pass in frameContentSize == ZSTD_CONTENTSIZE_UNKNOWN */
 
-ZSTDLIB_API size_t ZSTD_decompressBegin(ZSTD_DCtx* dctx);
+ZSTDLIB_STATIC_API size_t ZSTD_decompressBegin(ZSTD_DCtx* dctx);
-ZSTDLIB_API size_t ZSTD_decompressBegin_usingDict(ZSTD_DCtx* dctx, const void* dict, size_t dictSize);
+ZSTDLIB_STATIC_API size_t ZSTD_decompressBegin_usingDict(ZSTD_DCtx* dctx, const void* dict, size_t dictSize);
-ZSTDLIB_API size_t ZSTD_decompressBegin_usingDDict(ZSTD_DCtx* dctx, const ZSTD_DDict* ddict);
+ZSTDLIB_STATIC_API size_t ZSTD_decompressBegin_usingDDict(ZSTD_DCtx* dctx, const ZSTD_DDict* ddict);
 
-ZSTDLIB_API size_t ZSTD_nextSrcSizeToDecompress(ZSTD_DCtx* dctx);
+ZSTDLIB_STATIC_API size_t ZSTD_nextSrcSizeToDecompress(ZSTD_DCtx* dctx);
-ZSTDLIB_API size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+ZSTDLIB_STATIC_API size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
 
 /* misc */
-ZSTDLIB_API void   ZSTD_copyDCtx(ZSTD_DCtx* dctx, const ZSTD_DCtx* preparedDCtx);
+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_API ZSTD_nextInputType_e ZSTD_nextInputType(ZSTD_DCtx* dctx);
+ZSTDLIB_STATIC_API ZSTD_nextInputType_e ZSTD_nextInputType(ZSTD_DCtx* dctx);
 
 
 
@@ -2519,10 +2562,10 @@ ZSTDLIB_API ZSTD_nextInputType_e ZSTD_nextInputType(ZSTD_DCtx* dctx);
 */
 
 /*=====   Raw zstd block functions  =====*/
-ZSTDLIB_API size_t ZSTD_getBlockSize   (const ZSTD_CCtx* cctx);
+ZSTDLIB_STATIC_API size_t ZSTD_getBlockSize   (const ZSTD_CCtx* cctx);
-ZSTDLIB_API size_t ZSTD_compressBlock  (ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+ZSTDLIB_STATIC_API size_t ZSTD_compressBlock  (ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
-ZSTDLIB_API size_t ZSTD_decompressBlock(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+ZSTDLIB_STATIC_API size_t ZSTD_decompressBlock(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
-ZSTDLIB_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. */
+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. */
 
 
 #endif   /* ZSTD_H_ZSTD_STATIC_LINKING_ONLY */