diff --git a/README.md b/README.md
index f3b215a..f357409 100644
--- a/README.md
+++ b/README.md
@@ -6,8 +6,8 @@
 
 [C Zstd Homepage](https://github.com/facebook/zstd)
 
-The current headers and C files are from *v1.4.4* (Commit
-[10f0e699](https://github.com/facebook/zstd/releases/tag/v1.4.4)).
+The current headers and C files are from *v1.5.0* (Commit
+[10f0e699](https://github.com/facebook/zstd/releases/tag/v1.5.0)).
 
 ## Usage
 
diff --git a/bitstream.h b/bitstream.h
index d9a2730..48aad7f 100644
--- a/bitstream.h
+++ b/bitstream.h
@@ -1,7 +1,7 @@
 /* ******************************************************************
  * bitstream
  * Part of FSE library
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -332,7 +332,16 @@ 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)
diff --git a/compiler.h b/compiler.h
index 3e454f3..012ff02 100644
--- a/compiler.h
+++ b/compiler.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -90,6 +90,7 @@
 #  endif
 #endif
 
+
 /* target attribute */
 #ifndef __has_attribute
   #define __has_attribute(x) 0  /* Compatibility with non-clang compilers. */
@@ -149,8 +150,9 @@
 }
 
 /* vectorization
- * older GCC (pre gcc-4.3 picked as the cutoff) uses a different syntax */
-#if !defined(__INTEL_COMPILER) && !defined(__clang__) && defined(__GNUC__)
+ * older GCC (pre gcc-4.3 picked as the cutoff) uses a different syntax,
+ * and some compilers, like Intel ICC and MCST LCC, do not support it at all. */
+#if !defined(__INTEL_COMPILER) && !defined(__clang__) && defined(__GNUC__) && !defined(__LCC__)
 #  if (__GNUC__ == 4 && __GNUC_MINOR__ > 3) || (__GNUC__ >= 5)
 #    define DONT_VECTORIZE __attribute__((optimize("no-tree-vectorize")))
 #  else
@@ -196,6 +198,22 @@
     #define STATIC_BMI2 0
 #endif
 
+/* compile time determination of SIMD support */
+#if !defined(ZSTD_NO_INTRINSICS)
+#  if defined(__SSE2__) || defined(_M_AMD64) || (defined (_M_IX86) && defined(_M_IX86_FP) && (_M_IX86_FP >= 2))
+#    define ZSTD_ARCH_X86_SSE2
+#  endif
+#  if defined(__ARM_NEON) || defined(_M_ARM64)
+#    define ZSTD_ARCH_ARM_NEON
+#  endif
+#
+#  if defined(ZSTD_ARCH_X86_SSE2)
+#    include <emmintrin.h>
+#  elif defined(ZSTD_ARCH_ARM_NEON)
+#    include <arm_neon.h>
+#  endif
+#endif  
+
 /* compat. with non-clang compilers */
 #ifndef __has_builtin
 #  define __has_builtin(x) 0
diff --git a/cover.c b/cover.c
index 697f483..49d7d5f 100644
--- a/cover.c
+++ b/cover.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -26,15 +26,16 @@
 #include <string.h> /* memset */
 #include <time.h>   /* clock */
 
+#ifndef ZDICT_STATIC_LINKING_ONLY
+#  define ZDICT_STATIC_LINKING_ONLY
+#endif
+
 #include "mem.h" /* read */
 #include "pool.h"
 #include "threading.h"
-#include "cover.h"
 #include "zstd_internal.h" /* includes zstd.h */
-#ifndef ZDICT_STATIC_LINKING_ONLY
-#define ZDICT_STATIC_LINKING_ONLY
-#endif
 #include "zdict.h"
+#include "cover.h"
 
 /*-*************************************
 *  Constants
@@ -1062,18 +1063,19 @@ typedef struct COVER_tryParameters_data_s {
  * This function is thread safe if zstd is compiled with multithreaded support.
  * It takes its parameters as an *OWNING* opaque pointer to support threading.
  */
-static void COVER_tryParameters(void *opaque) {
+static void COVER_tryParameters(void *opaque)
+{
   /* Save parameters as local variables */
-  COVER_tryParameters_data_t *const data = (COVER_tryParameters_data_t *)opaque;
+  COVER_tryParameters_data_t *const data = (COVER_tryParameters_data_t*)opaque;
   const COVER_ctx_t *const ctx = data->ctx;
   const ZDICT_cover_params_t parameters = data->parameters;
   size_t dictBufferCapacity = data->dictBufferCapacity;
   size_t totalCompressedSize = ERROR(GENERIC);
   /* Allocate space for hash table, dict, and freqs */
   COVER_map_t activeDmers;
-  BYTE *const dict = (BYTE * const)malloc(dictBufferCapacity);
+  BYTE* const dict = (BYTE*)malloc(dictBufferCapacity);
   COVER_dictSelection_t selection = COVER_dictSelectionError(ERROR(GENERIC));
-  U32 *freqs = (U32 *)malloc(ctx->suffixSize * sizeof(U32));
+  U32* const freqs = (U32*)malloc(ctx->suffixSize * sizeof(U32));
   if (!COVER_map_init(&activeDmers, parameters.k - parameters.d + 1)) {
     DISPLAYLEVEL(1, "Failed to allocate dmer map: out of memory\n");
     goto _cleanup;
@@ -1103,15 +1105,14 @@ static void COVER_tryParameters(void *opaque) {
   free(data);
   COVER_map_destroy(&activeDmers);
   COVER_dictSelectionFree(selection);
-  if (freqs) {
-    free(freqs);
-  }
+  free(freqs);
 }
 
 ZDICTLIB_API size_t ZDICT_optimizeTrainFromBuffer_cover(
-    void *dictBuffer, size_t dictBufferCapacity, const void *samplesBuffer,
-    const size_t *samplesSizes, unsigned nbSamples,
-    ZDICT_cover_params_t *parameters) {
+    void* dictBuffer, size_t dictBufferCapacity, const void* samplesBuffer,
+    const size_t* samplesSizes, unsigned nbSamples,
+    ZDICT_cover_params_t* parameters)
+{
   /* constants */
   const unsigned nbThreads = parameters->nbThreads;
   const double splitPoint =
diff --git a/cover.h b/cover.h
index f2fbc53..c3df990 100644
--- a/cover.h
+++ b/cover.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020, Facebook, Inc.
+ * Copyright (c) Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -8,6 +8,10 @@
  * You may select, at your option, one of the above-listed licenses.
  */
 
+#ifndef ZDICT_STATIC_LINKING_ONLY
+#  define ZDICT_STATIC_LINKING_ONLY
+#endif
+
 #include <stdio.h>  /* fprintf */
 #include <stdlib.h> /* malloc, free, qsort */
 #include <string.h> /* memset */
@@ -16,9 +20,6 @@
 #include "pool.h"
 #include "threading.h"
 #include "zstd_internal.h" /* includes zstd.h */
-#ifndef ZDICT_STATIC_LINKING_ONLY
-#define ZDICT_STATIC_LINKING_ONLY
-#endif
 #include "zdict.h"
 
 /**
diff --git a/cpu.h b/cpu.h
index cb21059..8acd33b 100644
--- a/cpu.h
+++ b/cpu.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2020, Facebook, Inc.
+ * Copyright (c) Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/debug.c b/debug.c
index f303f4a..bb863c9 100644
--- a/debug.c
+++ b/debug.c
@@ -1,7 +1,7 @@
 /* ******************************************************************
  * debug
  * Part of FSE library
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
diff --git a/debug.h b/debug.h
index 8b57343..3b2a320 100644
--- a/debug.h
+++ b/debug.h
@@ -1,7 +1,7 @@
 /* ******************************************************************
  * debug
  * Part of FSE library
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
diff --git a/divsufsort.c b/divsufsort.c
index ead9220..a2870fb 100644
--- a/divsufsort.c
+++ b/divsufsort.c
@@ -1576,7 +1576,7 @@ sort_typeBstar(const unsigned char *T, int *SA,
     /* Construct the inverse suffix array of type B* suffixes using trsort. */
     trsort(ISAb, SA, m, 1);
 
-    /* Set the sorted order of tyoe B* suffixes. */
+    /* Set the sorted order of type B* suffixes. */
     for(i = n - 1, j = m, c0 = T[n - 1]; 0 <= i;) {
       for(--i, c1 = c0; (0 <= i) && ((c0 = T[i]) >= c1); --i, c1 = c0) { }
       if(0 <= i) {
diff --git a/entropy_common.c b/entropy_common.c
index f9fcb1a..6ba1f22 100644
--- a/entropy_common.c
+++ b/entropy_common.c
@@ -1,6 +1,6 @@
 /* ******************************************************************
  * Common functions of New Generation Entropy library
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  *
  *  You can contact the author at :
  *  - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -299,7 +299,7 @@ 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;
     }   }
diff --git a/error_private.c b/error_private.c
index 45bba53..6d1135f 100644
--- a/error_private.c
+++ b/error_private.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/error_private.h b/error_private.h
index 71b37b8..a8d33c8 100644
--- a/error_private.h
+++ b/error_private.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -21,8 +21,8 @@ extern "C" {
 /* ****************************************
 *  Dependencies
 ******************************************/
-#include "zstd_deps.h"    /* size_t */
 #include "zstd_errors.h"  /* enum list */
+#include "zstd_deps.h"       /* size_t */
 
 
 /* ****************************************
diff --git a/fastcover.c b/fastcover.c
index dad521c..13406cd 100644
--- a/fastcover.c
+++ b/fastcover.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2020, Facebook, Inc.
+ * Copyright (c) Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -16,16 +16,17 @@
 #include <string.h> /* memset */
 #include <time.h>   /* clock */
 
+#ifndef ZDICT_STATIC_LINKING_ONLY
+#  define ZDICT_STATIC_LINKING_ONLY
+#endif
+
 #include "mem.h" /* read */
 #include "pool.h"
 #include "threading.h"
-#include "cover.h"
 #include "zstd_internal.h" /* includes zstd.h */
 #include "zstd_compress_internal.h" /* ZSTD_hash*() */
-#ifndef ZDICT_STATIC_LINKING_ONLY
-#define ZDICT_STATIC_LINKING_ONLY
-#endif
 #include "zdict.h"
+#include "cover.h"
 
 
 /*-*************************************
@@ -462,20 +463,20 @@ typedef struct FASTCOVER_tryParameters_data_s {
  * This function is thread safe if zstd is compiled with multithreaded support.
  * It takes its parameters as an *OWNING* opaque pointer to support threading.
  */
-static void FASTCOVER_tryParameters(void *opaque)
+static void FASTCOVER_tryParameters(void* opaque)
 {
   /* Save parameters as local variables */
-  FASTCOVER_tryParameters_data_t *const data = (FASTCOVER_tryParameters_data_t *)opaque;
+  FASTCOVER_tryParameters_data_t *const data = (FASTCOVER_tryParameters_data_t*)opaque;
   const FASTCOVER_ctx_t *const ctx = data->ctx;
   const ZDICT_cover_params_t parameters = data->parameters;
   size_t dictBufferCapacity = data->dictBufferCapacity;
   size_t totalCompressedSize = ERROR(GENERIC);
   /* Initialize array to keep track of frequency of dmer within activeSegment */
-  U16* segmentFreqs = (U16 *)calloc(((U64)1 << ctx->f), sizeof(U16));
+  U16* segmentFreqs = (U16*)calloc(((U64)1 << ctx->f), sizeof(U16));
   /* Allocate space for hash table, dict, and freqs */
-  BYTE *const dict = (BYTE * const)malloc(dictBufferCapacity);
+  BYTE *const dict = (BYTE*)malloc(dictBufferCapacity);
   COVER_dictSelection_t selection = COVER_dictSelectionError(ERROR(GENERIC));
-  U32 *freqs = (U32*) malloc(((U64)1 << ctx->f) * sizeof(U32));
+  U32* freqs = (U32*) malloc(((U64)1 << ctx->f) * sizeof(U32));
   if (!segmentFreqs || !dict || !freqs) {
     DISPLAYLEVEL(1, "Failed to allocate buffers: out of memory\n");
     goto _cleanup;
diff --git a/fse.h b/fse.h
index dd5fc44..714bfd3 100644
--- a/fse.h
+++ b/fse.h
@@ -1,7 +1,7 @@
 /* ******************************************************************
  * FSE : Finite State Entropy codec
  * Public Prototypes declaration
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -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);
 
@@ -352,7 +353,7 @@ size_t FSE_buildDTable_raw (FSE_DTable* dt, unsigned nbBits);
 size_t FSE_buildDTable_rle (FSE_DTable* dt, unsigned char symbolValue);
 /**< build a fake FSE_DTable, designed to always generate the same symbolValue */
 
-#define FSE_DECOMPRESS_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) (FSE_DTABLE_SIZE_U32(maxTableLog) + FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxTableLog, maxSymbolValue))
+#define FSE_DECOMPRESS_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) (FSE_DTABLE_SIZE_U32(maxTableLog) + FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) + (FSE_MAX_SYMBOL_VALUE + 1) / 2 + 1)
 #define FSE_DECOMPRESS_WKSP_SIZE(maxTableLog, maxSymbolValue) (FSE_DECOMPRESS_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) * sizeof(unsigned))
 size_t FSE_decompress_wksp(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize);
 /**< same as FSE_decompress(), using an externally allocated `workSpace` produced with `FSE_DECOMPRESS_WKSP_SIZE_U32(maxLog, maxSymbolValue)` */
diff --git a/fse_compress.c b/fse_compress.c
index 89cb9df..4b40311 100644
--- a/fse_compress.c
+++ b/fse_compress.c
@@ -1,6 +1,6 @@
 /* ******************************************************************
  * FSE : Finite State Entropy encoder
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  *
  *  You can contact the author at :
  *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -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;
-    FSE_FUNCTION_TYPE* tableSymbol = (FSE_FUNCTION_TYPE*)(cumul + (maxSymbolValue + 2));
+    U16* cumul = (U16*)workSpace;   /* size = maxSV1 */
+    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 :
-                {
-                    U32 const maxBitsOut = tableLog - BIT_highbit32 (normalizedCounter[s]-1);
-                    U32 const minStatePlus = normalizedCounter[s] << maxBitsOut;
+                assert(normalizedCounter[s] > 1);
+                {   U32 const maxBitsOut = tableLog - BIT_highbit32 ((U32)normalizedCounter[s]-1);
+                    U32 const minStatePlus = (U32)normalizedCounter[s] << maxBitsOut;
                     symbolTT[s].deltaNbBits = (maxBitsOut << 16) - minStatePlus;
-                    symbolTT[s].deltaFindState = total - normalizedCounter[s];
-                    total +=  normalizedCounter[s];
+                    symbolTT[s].deltaFindState = (int)(total - (unsigned)normalizedCounter[s]);
+                    total +=  (unsigned)normalizedCounter[s];
     }   }   }   }
 
 #if 0  /* debug : symbol costs */
@@ -164,26 +206,16 @@ 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
 ****************************************************************/
diff --git a/fse_decompress.c b/fse_decompress.c
index c164430..f4ff58f 100644
--- a/fse_decompress.c
+++ b/fse_decompress.c
@@ -1,6 +1,6 @@
 /* ******************************************************************
  * FSE : Finite State Entropy decoder
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  *
  *  You can contact the author at :
  *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -310,6 +310,12 @@ size_t FSE_decompress_wksp(void* dst, size_t dstCapacity, const void* cSrc, size
     return FSE_decompress_wksp_bmi2(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize, /* bmi2 */ 0);
 }
 
+typedef struct {
+    short ncount[FSE_MAX_SYMBOL_VALUE + 1];
+    FSE_DTable dtable[1]; /* Dynamically sized */
+} FSE_DecompressWksp;
+
+
 FORCE_INLINE_TEMPLATE size_t FSE_decompress_wksp_body(
         void* dst, size_t dstCapacity,
         const void* cSrc, size_t cSrcSize,
@@ -318,33 +324,37 @@ FORCE_INLINE_TEMPLATE size_t FSE_decompress_wksp_body(
 {
     const BYTE* const istart = (const BYTE*)cSrc;
     const BYTE* ip = istart;
-    short counting[FSE_MAX_SYMBOL_VALUE+1];
     unsigned tableLog;
     unsigned maxSymbolValue = FSE_MAX_SYMBOL_VALUE;
-    FSE_DTable* const dtable = (FSE_DTable*)workSpace;
+    FSE_DecompressWksp* const wksp = (FSE_DecompressWksp*)workSpace;
+
+    DEBUG_STATIC_ASSERT((FSE_MAX_SYMBOL_VALUE + 1) % 2 == 0);
+    if (wkspSize < sizeof(*wksp)) return ERROR(GENERIC);
 
     /* normal FSE decoding mode */
-    size_t const NCountLength = FSE_readNCount_bmi2(counting, &maxSymbolValue, &tableLog, istart, cSrcSize, bmi2);
-    if (FSE_isError(NCountLength)) return NCountLength;
-    if (tableLog > maxLog) return ERROR(tableLog_tooLarge);
-    assert(NCountLength <= cSrcSize);
-    ip += NCountLength;
-    cSrcSize -= NCountLength;
+    {
+        size_t const NCountLength = FSE_readNCount_bmi2(wksp->ncount, &maxSymbolValue, &tableLog, istart, cSrcSize, bmi2);
+        if (FSE_isError(NCountLength)) return NCountLength;
+        if (tableLog > maxLog) return ERROR(tableLog_tooLarge);
+        assert(NCountLength <= cSrcSize);
+        ip += NCountLength;
+        cSrcSize -= NCountLength;
+    }
 
     if (FSE_DECOMPRESS_WKSP_SIZE(tableLog, maxSymbolValue) > wkspSize) return ERROR(tableLog_tooLarge);
-    workSpace = dtable + FSE_DTABLE_SIZE_U32(tableLog);
-    wkspSize -= FSE_DTABLE_SIZE(tableLog);
+    workSpace = wksp->dtable + FSE_DTABLE_SIZE_U32(tableLog);
+    wkspSize -= sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog);
 
-    CHECK_F( FSE_buildDTable_internal(dtable, counting, maxSymbolValue, tableLog, workSpace, wkspSize) );
+    CHECK_F( FSE_buildDTable_internal(wksp->dtable, wksp->ncount, maxSymbolValue, tableLog, workSpace, wkspSize) );
 
     {
-        const void* ptr = dtable;
+        const void* ptr = wksp->dtable;
         const FSE_DTableHeader* DTableH = (const FSE_DTableHeader*)ptr;
         const U32 fastMode = DTableH->fastMode;
 
         /* select fast mode (static) */
-        if (fastMode) return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, dtable, 1);
-        return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, dtable, 0);
+        if (fastMode) return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, wksp->dtable, 1);
+        return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, wksp->dtable, 0);
     }
 }
 
diff --git a/hist.c b/hist.c
index 8be9a30..4934aac 100644
--- a/hist.c
+++ b/hist.c
@@ -1,7 +1,7 @@
 /* ******************************************************************
  * hist : Histogram functions
  * part of Finite State Entropy project
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  *
  *  You can contact the author at :
  *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
diff --git a/hist.h b/hist.h
index 245be35..f68ff17 100644
--- a/hist.h
+++ b/hist.h
@@ -1,7 +1,7 @@
 /* ******************************************************************
  * hist : Histogram functions
  * part of Finite State Entropy project
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  *
  *  You can contact the author at :
  *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
diff --git a/huf.h b/huf.h
index 1afef90..a3ae06e 100644
--- a/huf.h
+++ b/huf.h
@@ -1,7 +1,7 @@
 /* ******************************************************************
  * huff0 huffman codec,
  * part of Finite State Entropy library
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  *
  * You can contact the author at :
  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -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 */
-#define HUF_WORKSPACE_SIZE ((6 << 10) + 256)
-#define HUF_WORKSPACE_SIZE_U32 (HUF_WORKSPACE_SIZE / sizeof(U32))
+ * `workspace` must be at least as large as HUF_WORKSPACE_SIZE */
+#define HUF_WORKSPACE_SIZE ((8 << 10) + 512 /* sorting scratch space */)
+#define HUF_WORKSPACE_SIZE_U64 (HUF_WORKSPACE_SIZE / sizeof(U64))
 HUF_PUBLIC_API size_t HUF_compress4X_wksp (void* dst, size_t dstCapacity,
                                      const void* src, size_t srcSize,
                                      unsigned maxSymbolValue, unsigned tableLog,
@@ -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 {
-  U16  val;
-  BYTE nbBits;
-};   /* 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))
+typedef size_t HUF_CElt;   /* consider it an incomplete type */
+#define HUF_CTABLE_SIZE_ST(maxSymbolValue)   ((maxSymbolValue)+2)   /* Use tables of size_t, for proper alignment */
+#define HUF_CTABLE_SIZE(maxSymbolValue)       (HUF_CTABLE_SIZE_ST(maxSymbolValue) * sizeof(size_t))
 #define HUF_CREATE_STATIC_CTABLE(name, maxSymbolValue) \
-    HUF_CElt name[HUF_CTABLE_SIZE_U32(maxSymbolValue)] /* no final ; */
+    HUF_CElt name[HUF_CTABLE_SIZE_ST(maxSymbolValue)] /* no final ; */
 
 /* static allocation of HUF's DTable */
 typedef U32 HUF_DTable;
@@ -192,7 +188,9 @@ size_t HUF_decompress4X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
 unsigned HUF_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue);
 size_t HUF_buildCTable (HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue, unsigned maxNbBits);   /* @return : maxNbBits; CTable and count can overlap. In which case, CTable will overwrite count content */
 size_t HUF_writeCTable (void* dst, size_t maxDstSize, const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog);
+size_t HUF_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);
 
@@ -205,12 +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.
@@ -248,11 +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 2 : const void* used, so that it can provide a statically allocated table as argument (which uses type U32) */
-U32 HUF_getNbBits(const void* symbolTable, U32 symbolValue);
+ *  Note 1 : is not inlined, as HUF_CElt definition is private */
+U32 HUF_getNbBitsFromCTable(const HUF_CElt* symbolTable, U32 symbolValue);
 
 /*
  * HUF_decompress() does the following:
@@ -278,7 +276,7 @@ U32 HUF_selectDecoder (size_t dstSize, size_t cSrcSize);
  *  a required workspace size greater than that specified in the following
  *  macro.
  */
-#define HUF_DECOMPRESS_WORKSPACE_SIZE (2 << 10)
+#define HUF_DECOMPRESS_WORKSPACE_SIZE ((2 << 10) + (1 << 9))
 #define HUF_DECOMPRESS_WORKSPACE_SIZE_U32 (HUF_DECOMPRESS_WORKSPACE_SIZE / sizeof(U32))
 
 #ifndef HUF_FORCE_DECOMPRESS_X2
@@ -304,18 +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
diff --git a/huf_compress.c b/huf_compress.c
index fed76ca..5614e21 100644
--- a/huf_compress.c
+++ b/huf_compress.c
@@ -1,6 +1,6 @@
 /* ******************************************************************
  * Huffman encoder, part of New Generation Entropy library
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  *
  *  You can contact the author at :
  *  - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -53,13 +53,43 @@ 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.
  * Note : all elements within weightTable are supposed to be <= HUF_TABLELOG_MAX.
  */
 #define MAX_FSE_TABLELOG_FOR_HUFF_HEADER 6
-static size_t HUF_compressWeights (void* dst, size_t dstSize, const void* weightTable, size_t wtSize)
+
+typedef struct {
+    FSE_CTable CTable[FSE_CTABLE_SIZE_U32(MAX_FSE_TABLELOG_FOR_HUFF_HEADER, HUF_TABLELOG_MAX)];
+    U32 scratchBuffer[FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(HUF_TABLELOG_MAX, MAX_FSE_TABLELOG_FOR_HUFF_HEADER)];
+    unsigned count[HUF_TABLELOG_MAX+1];
+    S16 norm[HUF_TABLELOG_MAX+1];
+} HUF_CompressWeightsWksp;
+
+static size_t HUF_compressWeights(void* dst, size_t dstSize, const void* weightTable, size_t wtSize, void* workspace, size_t workspaceSize)
 {
     BYTE* const ostart = (BYTE*) dst;
     BYTE* op = ostart;
@@ -67,33 +97,30 @@ static size_t HUF_compressWeights (void* dst, size_t dstSize, const void* weight
 
     unsigned maxSymbolValue = HUF_TABLELOG_MAX;
     U32 tableLog = MAX_FSE_TABLELOG_FOR_HUFF_HEADER;
+    HUF_CompressWeightsWksp* wksp = (HUF_CompressWeightsWksp*)HUF_alignUpWorkspace(workspace, &workspaceSize, sizeof(U32));
 
-    FSE_CTable CTable[FSE_CTABLE_SIZE_U32(MAX_FSE_TABLELOG_FOR_HUFF_HEADER, HUF_TABLELOG_MAX)];
-    U32 scratchBuffer[FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(HUF_TABLELOG_MAX, MAX_FSE_TABLELOG_FOR_HUFF_HEADER)];
-
-    unsigned count[HUF_TABLELOG_MAX+1];
-    S16 norm[HUF_TABLELOG_MAX+1];
+    if (workspaceSize < sizeof(HUF_CompressWeightsWksp)) return ERROR(GENERIC);
 
     /* init conditions */
     if (wtSize <= 1) return 0;  /* Not compressible */
 
     /* Scan input and build symbol stats */
-    {   unsigned const maxCount = HIST_count_simple(count, &maxSymbolValue, weightTable, wtSize);   /* never fails */
+    {   unsigned const maxCount = HIST_count_simple(wksp->count, &maxSymbolValue, weightTable, wtSize);   /* never fails */
         if (maxCount == wtSize) return 1;   /* only a single symbol in src : rle */
         if (maxCount == 1) return 0;        /* each symbol present maximum once => not compressible */
     }
 
     tableLog = FSE_optimalTableLog(tableLog, wtSize, maxSymbolValue);
-    CHECK_F( FSE_normalizeCount(norm, tableLog, count, wtSize, maxSymbolValue, /* useLowProbCount */ 0) );
+    CHECK_F( FSE_normalizeCount(wksp->norm, tableLog, wksp->count, wtSize, maxSymbolValue, /* useLowProbCount */ 0) );
 
     /* Write table description header */
-    {   CHECK_V_F(hSize, FSE_writeNCount(op, (size_t)(oend-op), norm, maxSymbolValue, tableLog) );
+    {   CHECK_V_F(hSize, FSE_writeNCount(op, (size_t)(oend-op), wksp->norm, maxSymbolValue, tableLog) );
         op += hSize;
     }
 
     /* Compress */
-    CHECK_F( FSE_buildCTable_wksp(CTable, norm, maxSymbolValue, tableLog, scratchBuffer, sizeof(scratchBuffer)) );
-    {   CHECK_V_F(cSize, FSE_compress_usingCTable(op, (size_t)(oend - op), weightTable, wtSize, CTable) );
+    CHECK_F( FSE_buildCTable_wksp(wksp->CTable, wksp->norm, maxSymbolValue, tableLog, wksp->scratchBuffer, sizeof(wksp->scratchBuffer)) );
+    {   CHECK_V_F(cSize, FSE_compress_usingCTable(op, (size_t)(oend - op), weightTable, wtSize, wksp->CTable) );
         if (cSize == 0) return 0;   /* not enough space for compressed data */
         op += cSize;
     }
@@ -101,30 +128,70 @@ static size_t HUF_compressWeights (void* dst, size_t dstSize, const void* weight
     return (size_t)(op-ostart);
 }
 
+static size_t HUF_getNbBits(HUF_CElt elt)
+{
+    return elt & 0xFF;
+}
 
-/*! HUF_writeCTable() :
-    `CTable` : Huffman tree to save, using huf representation.
-    @return : size of saved CTable */
-size_t HUF_writeCTable (void* dst, size_t maxDstSize,
-                        const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog)
+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;
     BYTE bitsToWeight[HUF_TABLELOG_MAX + 1];   /* precomputed conversion table */
     BYTE huffWeight[HUF_SYMBOLVALUE_MAX];
+} HUF_WriteCTableWksp;
+
+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*)HUF_alignUpWorkspace(workspace, &workspaceSize, sizeof(U32));
 
-     /* check conditions */
+    /* check conditions */
+    if (workspaceSize < sizeof(HUF_WriteCTableWksp)) return ERROR(GENERIC);
     if (maxSymbolValue > HUF_SYMBOLVALUE_MAX) return ERROR(maxSymbolValue_tooLarge);
 
     /* convert to weight */
-    bitsToWeight[0] = 0;
+    wksp->bitsToWeight[0] = 0;
     for (n=1; n<huffLog+1; n++)
-        bitsToWeight[n] = (BYTE)(huffLog + 1 - n);
+        wksp->bitsToWeight[n] = (BYTE)(huffLog + 1 - n);
     for (n=0; n<maxSymbolValue; n++)
-        huffWeight[n] = bitsToWeight[CTable[n].nbBits];
+        wksp->huffWeight[n] = wksp->bitsToWeight[HUF_getNbBits(ct[n])];
 
     /* attempt weights compression by FSE */
-    {   CHECK_V_F(hSize, HUF_compressWeights(op+1, maxDstSize-1, huffWeight, maxSymbolValue) );
+    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;
             return hSize+1;
@@ -134,12 +201,22 @@ size_t HUF_writeCTable (void* dst, size_t maxDstSize,
     if (maxSymbolValue > (256-128)) return ERROR(GENERIC);   /* should not happen : likely means source cannot be compressed */
     if (((maxSymbolValue+1)/2) + 1 > maxDstSize) return ERROR(dstSize_tooSmall);   /* not enough space within dst buffer */
     op[0] = (BYTE)(128 /*special case*/ + (maxSymbolValue-1));
-    huffWeight[maxSymbolValue] = 0;   /* to be sure it doesn't cause msan issue in final combination */
+    wksp->huffWeight[maxSymbolValue] = 0;   /* to be sure it doesn't cause msan issue in final combination */
     for (n=0; n<maxSymbolValue; n+=2)
-        op[(n/2)+1] = (BYTE)((huffWeight[n] << 4) + huffWeight[n+1]);
+        op[(n/2)+1] = (BYTE)((wksp->huffWeight[n] << 4) + wksp->huffWeight[n+1]);
     return ((maxSymbolValue+1)/2) + 1;
 }
 
+/*! HUF_writeCTable() :
+    `CTable` : Huffman tree to save, using huf representation.
+    @return : size of saved CTable */
+size_t HUF_writeCTable (void* dst, size_t maxDstSize,
+                        const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog)
+{
+    HUF_WriteCTableWksp wksp;
+    return HUF_writeCTable_wksp(dst, maxDstSize, CTable, maxSymbolValue, huffLog, &wksp, sizeof(wksp));
+}
+
 
 size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void* src, size_t srcSize, unsigned* hasZeroWeights)
 {
@@ -147,6 +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));
@@ -156,6 +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++) {
@@ -167,13 +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;
@@ -183,18 +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]);
 }
 
 
@@ -348,22 +428,122 @@ static U32 HUF_setMaxHeight(nodeElt* huffNode, U32 lastNonNull, U32 maxNbBits)
 }
 
 typedef struct {
-    U32 base;
-    U32 curr;
+    U16 base;
+    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 128
 
 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 counts up to HUF_BLOCKSIZE_MAX
+ * using log2 bucketing.
+ * 
+ * To satisfy this requirement for 128 buckets, we can do the following:
+ * Let buckets 0-114 represent distinct counts of [0, 114]
+ * Let buckets 115 to 126 represent counts of [115, HUF_BLOCKSIZE_MAX]. (the final bucket 127 must remain empty)
+ * 
+ * Note that we don't actually need 17 buckets (assuming 2^17 maxcount) for log2 bucketing since
+ * the first few buckets in the log2 bucketing representation are already covered by the distinct count bucketing.
+ */
+#define RANK_POSITION_LOG_BUCKETS_BEGIN (RANK_POSITION_TABLE_SIZE - 1) - BIT_highbit32(HUF_BLOCKSIZE_MAX) - 1
+#define RANK_POSITION_DISTINCT_COUNT_CUTOFF RANK_POSITION_LOG_BUCKETS_BEGIN + BIT_highbit32(RANK_POSITION_LOG_BUCKETS_BEGIN)
+
+/* 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 */
+UNUSED_ATTR
+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.
@@ -371,44 +551,52 @@ 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)
-{
-    int n;
-    int const maxSymbolValue1 = (int)maxSymbolValue + 1;
+static void HUF_sort(nodeElt huffNode[], const unsigned count[], U32 const maxSymbolValue, rankPos rankPosition[]) {
+    U32 n;
+    U32 const maxSymbolValue1 = maxSymbolValue+1;
 
     /* Compute base and set curr to base.
-     * For symbol s let lowerRank = BIT_highbit32(count[n]+1) and rank = lowerRank + 1.
-     * Then 2^lowerRank <= count[n]+1 <= 2^rank.
+     * For symbol s let lowerRank = HUF_getIndex(count[n]) and rank = lowerRank + 1.
+     * 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 pos = rankPosition[r].curr++;
-        /* Insert into the correct position in the rank.
-         * 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--;
-        }
+        U32 const r = HUF_getIndex(c) + 1;
+        U32 const pos = rankPosition[r].curr++;
+        assert(pos < maxSymbolValue1);
         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.
@@ -471,6 +659,7 @@ 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};
@@ -486,20 +675,20 @@ 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, sizeof(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;
@@ -517,91 +706,327 @@ 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); }
 
+/** HUF_CStream_t:
+ * 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 bistream.
+ * @returns 0 or an error code.
+ */
+static size_t HUF_initCStream(HUF_CStream_t* bitC,
+                                  void* startPtr, size_t dstCapacity)
+{
+    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;
+}
+
+/*! 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 bistream 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;
+        /* 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);
+    }
+#endif
+}
+
+FORCE_INLINE_TEMPLATE void HUF_zeroIndex1(HUF_CStream_t* bitC)
+{
+    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(BIT_CStream_t* bitCPtr, U32 symbol, const HUF_CElt* CTable)
+HUF_encodeSymbol(HUF_CStream_t* bitCPtr, U32 symbol, const HUF_CElt* CTable, int idx, int fast)
 {
-    BIT_addBitsFast(bitCPtr, CTable[symbol].val, CTable[symbol].nbBits);
+    HUF_addBits(bitCPtr, CTable[symbol], idx, fast);
 }
 
-#define HUF_FLUSHBITS(s)  BIT_flushBits(s)
+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);
 
-#define HUF_FLUSHBITS_1(stream) \
-    if (sizeof((stream)->bitContainer)*8 < HUF_TABLELOG_MAX*2+7) HUF_FLUSHBITS(stream)
+    /* 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;
-    BIT_CStream_t bitC;
+    HUF_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 */
-    switch (srcSize & 3)
-    {
-        case 3 : HUF_encodeSymbol(&bitC, ip[n+ 2], CTable);
-                 HUF_FLUSHBITS_2(&bitC);
-		 /* fall-through */
-        case 2 : HUF_encodeSymbol(&bitC, ip[n+ 1], CTable);
-                 HUF_FLUSHBITS_1(&bitC);
-		 /* fall-through */
-        case 1 : HUF_encodeSymbol(&bitC, ip[n+ 0], CTable);
-                 HUF_FLUSHBITS(&bitC);
-		 /* fall-through */
-        case 0 : /* fall-through */
-        default: break;
-    }
-
-    for (; n>0; n-=4) {  /* note : n&3==0 at this stage */
-        HUF_encodeSymbol(&bitC, ip[n- 1], CTable);
-        HUF_FLUSHBITS_1(&bitC);
-        HUF_encodeSymbol(&bitC, ip[n- 2], CTable);
-        HUF_FLUSHBITS_2(&bitC);
-        HUF_encodeSymbol(&bitC, ip[n- 3], CTable);
-        HUF_FLUSHBITS_1(&bitC);
-        HUF_encodeSymbol(&bitC, ip[n- 4], CTable);
-        HUF_FLUSHBITS(&bitC);
+    if (dstSize < HUF_tightCompressBound(srcSize, (size_t)tableLog) || tableLog > 11)
+        HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ MEM_32bits() ? 2 : 4, /* kFast */ 0, /* kLastFast */ 0);
+    else {
+        if (MEM_32bits()) {
+            switch (tableLog) {
+            case 11:
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 2, /* kFastFlush */ 1, /* kLastFast */ 0);
+                break;
+            case 10:
+            case 9:
+            case 8:
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 2, /* kFastFlush */ 1, /* kLastFast */ 1);
+                break;
+            case 7:
+            default:
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 3, /* kFastFlush */ 1, /* kLastFast */ 1);
+                break;
+            }
+        } else {
+            switch (tableLog) {
+            case 11:
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 5, /* kFastFlush */ 1, /* kLastFast */ 0);
+                break;
+            case 10:
+                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 5, /* kFastFlush */ 1, /* kLastFast */ 1);
+                break;
+            case 9:
+                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:
+            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
@@ -648,9 +1073,13 @@ 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,
@@ -670,8 +1099,7 @@ 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;
-        assert(cSize <= 65535);
+        if (cSize == 0 || cSize > 65535) return 0;
         MEM_writeLE16(ostart, (U16)cSize);
         op += cSize;
     }
@@ -679,8 +1107,7 @@ 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;
-        assert(cSize <= 65535);
+        if (cSize == 0 || cSize > 65535) return 0;
         MEM_writeLE16(ostart+2, (U16)cSize);
         op += cSize;
     }
@@ -688,8 +1115,7 @@ 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;
-        assert(cSize <= 65535);
+        if (cSize == 0 || cSize > 65535) return 0;
         MEM_writeLE16(ostart+4, (U16)cSize);
         op += cSize;
     }
@@ -698,7 +1124,7 @@ 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;
     }
 
@@ -707,7 +1133,12 @@ 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;
@@ -731,32 +1162,38 @@ static size_t HUF_compressCTable_internal(
 
 typedef struct {
     unsigned count[HUF_SYMBOLVALUE_MAX + 1];
-    HUF_CElt CTable[HUF_SYMBOLVALUE_MAX + 1];
-    HUF_buildCTable_wksp_tables buildCTable_wksp;
+    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, sizeof(size_t));
     BYTE* const ostart = (BYTE*)dst;
     BYTE* const oend = ostart + dstSize;
     BYTE* op = ostart;
 
-    HUF_STATIC_ASSERT(sizeof(*table) <= HUF_WORKSPACE_SIZE);
-    assert(((size_t)workSpace_align4 & 3) == 0);   /* must be aligned on 4-bytes boundaries */
+    HUF_STATIC_ASSERT(sizeof(*table) + HUF_WORKSPACE_MAX_ALIGNMENT <= HUF_WORKSPACE_SIZE);
 
     /* 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 */
@@ -772,8 +1209,23 @@ 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 */
     }
@@ -795,16 +1247,20 @@ HUF_compress_internal (void* dst, size_t dstSize,
     huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue);
     {   size_t const maxBits = HUF_buildCTable_wksp(table->CTable, table->count,
                                             maxSymbolValue, huffLog,
-                                            &table->buildCTable_wksp, sizeof(table->buildCTable_wksp));
+                                            &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)));
+    }
+    /* Zero unused symbols in CTable, so we can check it for validity */
+    {
+        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 */
-    {   CHECK_V_F(hSize, HUF_writeCTable (op, dstSize, table->CTable, maxSymbolValue, huffLog) );
+    {   CHECK_V_F(hSize, HUF_writeCTable_wksp(op, dstSize, table->CTable, maxSymbolValue, huffLog,
+                                              &table->wksps.writeCTable_wksp, sizeof(table->wksps.writeCTable_wksp)) );
         /* Check if using previous huffman table is beneficial */
         if (repeat && *repeat != HUF_repeat_none) {
             size_t const oldSize = HUF_estimateCompressedSize(oldHufTable, table->count, maxSymbolValue);
@@ -836,19 +1292,20 @@ 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():
@@ -862,22 +1319,23 @@ 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
@@ -895,7 +1353,7 @@ 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));
 }
 
@@ -903,7 +1361,7 @@ 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));
 }
 
diff --git a/huf_decompress.c b/huf_decompress.c
index 65e157e..eb292e0 100644
--- a/huf_decompress.c
+++ b/huf_decompress.c
@@ -1,7 +1,7 @@
 /* ******************************************************************
  * huff0 huffman decoder,
  * part of Finite State Entropy library
- * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  *
  *  You can contact the author at :
  *  - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
@@ -528,13 +528,15 @@ typedef rankValCol_t rankVal_t[HUF_TABLELOG_MAX];
 static void HUF_fillDTableX2Level2(HUF_DEltX2* DTable, U32 sizeLog, const U32 consumed,
                            const U32* rankValOrigin, const int minWeight,
                            const sortedSymbol_t* sortedSymbols, const U32 sortedListSize,
-                           U32 nbBitsBaseline, U16 baseSeq)
+                           U32 nbBitsBaseline, U16 baseSeq, U32* wksp, size_t wkspSize)
 {
     HUF_DEltX2 DElt;
-    U32 rankVal[HUF_TABLELOG_MAX + 1];
+    U32* rankVal = wksp;
 
+    assert(wkspSize >= HUF_TABLELOG_MAX + 1);
+    (void)wkspSize;
     /* get pre-calculated rankVal */
-    ZSTD_memcpy(rankVal, rankValOrigin, sizeof(rankVal));
+    ZSTD_memcpy(rankVal, rankValOrigin, sizeof(U32) * (HUF_TABLELOG_MAX + 1));
 
     /* fill skipped values */
     if (minWeight>1) {
@@ -569,14 +571,18 @@ static void HUF_fillDTableX2Level2(HUF_DEltX2* DTable, U32 sizeLog, const U32 co
 static void HUF_fillDTableX2(HUF_DEltX2* DTable, const U32 targetLog,
                            const sortedSymbol_t* sortedList, const U32 sortedListSize,
                            const U32* rankStart, rankVal_t rankValOrigin, const U32 maxWeight,
-                           const U32 nbBitsBaseline)
+                           const U32 nbBitsBaseline, U32* wksp, size_t wkspSize)
 {
-    U32 rankVal[HUF_TABLELOG_MAX + 1];
+    U32* rankVal = wksp;
     const int scaleLog = nbBitsBaseline - targetLog;   /* note : targetLog >= srcLog, hence scaleLog <= 1 */
     const U32 minBits  = nbBitsBaseline - maxWeight;
     U32 s;
 
-    ZSTD_memcpy(rankVal, rankValOrigin, sizeof(rankVal));
+    assert(wkspSize >= HUF_TABLELOG_MAX + 1);
+    wksp += HUF_TABLELOG_MAX + 1;
+    wkspSize -= HUF_TABLELOG_MAX + 1;
+
+    ZSTD_memcpy(rankVal, rankValOrigin, sizeof(U32) * (HUF_TABLELOG_MAX + 1));
 
     /* fill DTable */
     for (s=0; s<sortedListSize; s++) {
@@ -594,7 +600,7 @@ static void HUF_fillDTableX2(HUF_DEltX2* DTable, const U32 targetLog,
             HUF_fillDTableX2Level2(DTable+start, targetLog-nbBits, nbBits,
                            rankValOrigin[nbBits], minWeight,
                            sortedList+sortedRank, sortedListSize-sortedRank,
-                           nbBitsBaseline, symbol);
+                           nbBitsBaseline, symbol, wksp, wkspSize);
         } else {
             HUF_DEltX2 DElt;
             MEM_writeLE16(&(DElt.sequence), symbol);
@@ -608,6 +614,15 @@ static void HUF_fillDTableX2(HUF_DEltX2* DTable, const U32 targetLog,
     }
 }
 
+typedef struct {
+    rankValCol_t rankVal[HUF_TABLELOG_MAX];
+    U32 rankStats[HUF_TABLELOG_MAX + 1];
+    U32 rankStart0[HUF_TABLELOG_MAX + 2];
+    sortedSymbol_t sortedSymbol[HUF_SYMBOLVALUE_MAX + 1];
+    BYTE weightList[HUF_SYMBOLVALUE_MAX + 1];
+    U32 calleeWksp[HUF_READ_STATS_WORKSPACE_SIZE_U32];
+} HUF_ReadDTableX2_Workspace;
+
 size_t HUF_readDTableX2_wksp(HUF_DTable* DTable,
                        const void* src, size_t srcSize,
                              void* workSpace, size_t wkspSize)
@@ -620,47 +635,32 @@ size_t HUF_readDTableX2_wksp(HUF_DTable* DTable,
     HUF_DEltX2* const dt = (HUF_DEltX2*)dtPtr;
     U32 *rankStart;
 
-    rankValCol_t* rankVal;
-    U32* rankStats;
-    U32* rankStart0;
-    sortedSymbol_t* sortedSymbol;
-    BYTE* weightList;
-    size_t spaceUsed32 = 0;
-
-    rankVal = (rankValCol_t *)((U32 *)workSpace + spaceUsed32);
-    spaceUsed32 += (sizeof(rankValCol_t) * HUF_TABLELOG_MAX) >> 2;
-    rankStats = (U32 *)workSpace + spaceUsed32;
-    spaceUsed32 += HUF_TABLELOG_MAX + 1;
-    rankStart0 = (U32 *)workSpace + spaceUsed32;
-    spaceUsed32 += HUF_TABLELOG_MAX + 2;
-    sortedSymbol = (sortedSymbol_t *)workSpace + (spaceUsed32 * sizeof(U32)) / sizeof(sortedSymbol_t);
-    spaceUsed32 += HUF_ALIGN(sizeof(sortedSymbol_t) * (HUF_SYMBOLVALUE_MAX + 1), sizeof(U32)) >> 2;
-    weightList = (BYTE *)((U32 *)workSpace + spaceUsed32);
-    spaceUsed32 += HUF_ALIGN(HUF_SYMBOLVALUE_MAX + 1, sizeof(U32)) >> 2;
-
-    if ((spaceUsed32 << 2) > wkspSize) return ERROR(tableLog_tooLarge);
-
-    rankStart = rankStart0 + 1;
-    ZSTD_memset(rankStats, 0, sizeof(U32) * (2 * HUF_TABLELOG_MAX + 2 + 1));
+    HUF_ReadDTableX2_Workspace* const wksp = (HUF_ReadDTableX2_Workspace*)workSpace;
+
+    if (sizeof(*wksp) > wkspSize) return ERROR(GENERIC);
+
+    rankStart = wksp->rankStart0 + 1;
+    ZSTD_memset(wksp->rankStats, 0, sizeof(wksp->rankStats));
+    ZSTD_memset(wksp->rankStart0, 0, sizeof(wksp->rankStart0));
 
     DEBUG_STATIC_ASSERT(sizeof(HUF_DEltX2) == sizeof(HUF_DTable));   /* if compiler fails here, assertion is wrong */
     if (maxTableLog > HUF_TABLELOG_MAX) return ERROR(tableLog_tooLarge);
     /* ZSTD_memset(weightList, 0, sizeof(weightList)); */  /* is not necessary, even though some analyzer complain ... */
 
-    iSize = HUF_readStats(weightList, HUF_SYMBOLVALUE_MAX + 1, rankStats, &nbSymbols, &tableLog, src, srcSize);
+    iSize = HUF_readStats_wksp(wksp->weightList, HUF_SYMBOLVALUE_MAX + 1, wksp->rankStats, &nbSymbols, &tableLog, src, srcSize, wksp->calleeWksp, sizeof(wksp->calleeWksp), /* bmi2 */ 0);
     if (HUF_isError(iSize)) return iSize;
 
     /* check result */
     if (tableLog > maxTableLog) return ERROR(tableLog_tooLarge);   /* DTable can't fit code depth */
 
     /* find maxWeight */
-    for (maxW = tableLog; rankStats[maxW]==0; maxW--) {}  /* necessarily finds a solution before 0 */
+    for (maxW = tableLog; wksp->rankStats[maxW]==0; maxW--) {}  /* necessarily finds a solution before 0 */
 
     /* Get start index of each weight */
     {   U32 w, nextRankStart = 0;
         for (w=1; w<maxW+1; w++) {
             U32 curr = nextRankStart;
-            nextRankStart += rankStats[w];
+            nextRankStart += wksp->rankStats[w];
             rankStart[w] = curr;
         }
         rankStart[0] = nextRankStart;   /* put all 0w symbols at the end of sorted list*/
@@ -670,37 +670,38 @@ size_t HUF_readDTableX2_wksp(HUF_DTable* DTable,
     /* sort symbols by weight */
     {   U32 s;
         for (s=0; s<nbSymbols; s++) {
-            U32 const w = weightList[s];
+            U32 const w = wksp->weightList[s];
             U32 const r = rankStart[w]++;
-            sortedSymbol[r].symbol = (BYTE)s;
-            sortedSymbol[r].weight = (BYTE)w;
+            wksp->sortedSymbol[r].symbol = (BYTE)s;
+            wksp->sortedSymbol[r].weight = (BYTE)w;
         }
         rankStart[0] = 0;   /* forget 0w symbols; this is beginning of weight(1) */
     }
 
     /* Build rankVal */
-    {   U32* const rankVal0 = rankVal[0];
+    {   U32* const rankVal0 = wksp->rankVal[0];
         {   int const rescale = (maxTableLog-tableLog) - 1;   /* tableLog <= maxTableLog */
             U32 nextRankVal = 0;
             U32 w;
             for (w=1; w<maxW+1; w++) {
                 U32 curr = nextRankVal;
-                nextRankVal += rankStats[w] << (w+rescale);
+                nextRankVal += wksp->rankStats[w] << (w+rescale);
                 rankVal0[w] = curr;
         }   }
         {   U32 const minBits = tableLog+1 - maxW;
             U32 consumed;
             for (consumed = minBits; consumed < maxTableLog - minBits + 1; consumed++) {
-                U32* const rankValPtr = rankVal[consumed];
+                U32* const rankValPtr = wksp->rankVal[consumed];
                 U32 w;
                 for (w = 1; w < maxW+1; w++) {
                     rankValPtr[w] = rankVal0[w] >> consumed;
     }   }   }   }
 
     HUF_fillDTableX2(dt, maxTableLog,
-                   sortedSymbol, sizeOfSort,
-                   rankStart0, rankVal, maxW,
-                   tableLog+1);
+                   wksp->sortedSymbol, sizeOfSort,
+                   wksp->rankStart0, wksp->rankVal, maxW,
+                   tableLog+1,
+                   wksp->calleeWksp, sizeof(wksp->calleeWksp) / sizeof(U32));
 
     dtd.tableLog = (BYTE)maxTableLog;
     dtd.tableType = 1;
diff --git a/mem.h b/mem.h
index 4728ef7..1c61b7e 100644
--- a/mem.h
+++ b/mem.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -143,9 +143,7 @@ MEM_STATIC size_t MEM_swapST(size_t in);
  * Prefer these methods in priority order (0 > 1 > 2)
  */
 #ifndef MEM_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
-#  if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
-#    define MEM_FORCE_MEMORY_ACCESS 2
-#  elif defined(__INTEL_COMPILER) || defined(__GNUC__) || defined(__ICCARM__)
+#  if defined(__INTEL_COMPILER) || defined(__GNUC__) || defined(__ICCARM__)
 #    define MEM_FORCE_MEMORY_ACCESS 1
 #  endif
 #endif
@@ -155,8 +153,22 @@ 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)
@@ -308,7 +320,7 @@ MEM_STATIC void MEM_writeLE16(void* memPtr, U16 val)
 
 MEM_STATIC U32 MEM_readLE24(const void* memPtr)
 {
-    return MEM_readLE16(memPtr) + (((const BYTE*)memPtr)[2] << 16);
+    return (U32)MEM_readLE16(memPtr) + ((U32)(((const BYTE*)memPtr)[2]) << 16);
 }
 
 MEM_STATIC void MEM_writeLE24(void* memPtr, U32 val)
diff --git a/pool.c b/pool.c
index 4c1b833..ea70b8b 100644
--- a/pool.c
+++ b/pool.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/pool.h b/pool.h
index 816e25f..411e318 100644
--- a/pool.h
+++ b/pool.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/xxhash.c b/xxhash.c
index e708df3..926b336 100644
--- a/xxhash.c
+++ b/xxhash.c
@@ -1,6 +1,6 @@
 /*
  *  xxHash - Fast Hash algorithm
- *  Copyright (c) 2012-2020, Yann Collet, Facebook, Inc.
+ *  Copyright (c) Yann Collet, Facebook, Inc.
  *
  *  You can contact the author at :
  *  - xxHash homepage: http://www.xxhash.com
@@ -30,9 +30,7 @@
  * Prefer these methods in priority order (0 > 1 > 2)
  */
 #ifndef XXH_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
-#  if defined(__GNUC__) && ( defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) )
-#    define XXH_FORCE_MEMORY_ACCESS 2
-#  elif (defined(__INTEL_COMPILER) && !defined(WIN32)) || \
+#  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
diff --git a/xxhash.h b/xxhash.h
index eceb55d..16c1f16 100644
--- a/xxhash.h
+++ b/xxhash.h
@@ -1,7 +1,7 @@
 /*
  * xxHash - Extremely Fast Hash algorithm
  * Header File
- * Copyright (c) 2012-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  *
  * You can contact the author at :
  * - xxHash source repository : https://github.com/Cyan4973/xxHash
diff --git a/zbuff.h b/zbuff.h
deleted file mode 100644
index 7686f76..0000000
--- a/zbuff.h
+++ /dev/null
@@ -1,214 +0,0 @@
-/*
- * Copyright (c) 2016-2020, 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.
- */
-
-/* ***************************************************************
-*  NOTES/WARNINGS
-******************************************************************/
-/* The streaming API defined here is deprecated.
- * Consider migrating towards ZSTD_compressStream() API in `zstd.h`
- * See 'lib/README.md'.
- *****************************************************************/
-
-
-#if defined (__cplusplus)
-extern "C" {
-#endif
-
-#ifndef ZSTD_BUFFERED_H_23987
-#define ZSTD_BUFFERED_H_23987
-
-/* *************************************
-*  Dependencies
-***************************************/
-#include <stddef.h>      /* size_t */
-#include "zstd.h"        /* ZSTD_CStream, ZSTD_DStream, ZSTDLIB_API */
-
-
-/* ***************************************************************
-*  Compiler specifics
-*****************************************************************/
-/* 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 ZBUFF_DISABLE_DEPRECATE_WARNINGS
- */
-#ifdef ZBUFF_DISABLE_DEPRECATE_WARNINGS
-#  define ZBUFF_DEPRECATED(message) ZSTDLIB_API  /* disable deprecation warnings */
-#else
-#  if defined (__cplusplus) && (__cplusplus >= 201402) /* C++14 or greater */
-#    define ZBUFF_DEPRECATED(message) [[deprecated(message)]] ZSTDLIB_API
-#  elif (defined(GNUC) && (GNUC > 4 || (GNUC == 4 && GNUC_MINOR >= 5))) || defined(__clang__)
-#    define ZBUFF_DEPRECATED(message) ZSTDLIB_API __attribute__((deprecated(message)))
-#  elif defined(__GNUC__) && (__GNUC__ >= 3)
-#    define ZBUFF_DEPRECATED(message) ZSTDLIB_API __attribute__((deprecated))
-#  elif defined(_MSC_VER)
-#    define ZBUFF_DEPRECATED(message) ZSTDLIB_API __declspec(deprecated(message))
-#  else
-#    pragma message("WARNING: You need to implement ZBUFF_DEPRECATED for this compiler")
-#    define ZBUFF_DEPRECATED(message) ZSTDLIB_API
-#  endif
-#endif /* ZBUFF_DISABLE_DEPRECATE_WARNINGS */
-
-
-/* *************************************
-*  Streaming functions
-***************************************/
-/* This is the easier "buffered" streaming API,
-*  using an internal buffer to lift all restrictions on user-provided buffers
-*  which can be any size, any place, for both input and output.
-*  ZBUFF and ZSTD are 100% interoperable,
-*  frames created by one can be decoded by the other one */
-
-typedef ZSTD_CStream ZBUFF_CCtx;
-ZBUFF_DEPRECATED("use ZSTD_createCStream") ZBUFF_CCtx* ZBUFF_createCCtx(void);
-ZBUFF_DEPRECATED("use ZSTD_freeCStream")   size_t      ZBUFF_freeCCtx(ZBUFF_CCtx* cctx);
-
-ZBUFF_DEPRECATED("use ZSTD_initCStream")           size_t ZBUFF_compressInit(ZBUFF_CCtx* cctx, int compressionLevel);
-ZBUFF_DEPRECATED("use ZSTD_initCStream_usingDict") size_t ZBUFF_compressInitDictionary(ZBUFF_CCtx* cctx, const void* dict, size_t dictSize, int compressionLevel);
-
-ZBUFF_DEPRECATED("use ZSTD_compressStream") size_t ZBUFF_compressContinue(ZBUFF_CCtx* cctx, void* dst, size_t* dstCapacityPtr, const void* src, size_t* srcSizePtr);
-ZBUFF_DEPRECATED("use ZSTD_flushStream")    size_t ZBUFF_compressFlush(ZBUFF_CCtx* cctx, void* dst, size_t* dstCapacityPtr);
-ZBUFF_DEPRECATED("use ZSTD_endStream")      size_t ZBUFF_compressEnd(ZBUFF_CCtx* cctx, void* dst, size_t* dstCapacityPtr);
-
-/*-*************************************************
-*  Streaming compression - howto
-*
-*  A ZBUFF_CCtx object is required to track streaming operation.
-*  Use ZBUFF_createCCtx() and ZBUFF_freeCCtx() to create/release resources.
-*  ZBUFF_CCtx objects can be reused multiple times.
-*
-*  Start by initializing ZBUF_CCtx.
-*  Use ZBUFF_compressInit() to start a new compression operation.
-*  Use ZBUFF_compressInitDictionary() for a compression which requires a dictionary.
-*
-*  Use ZBUFF_compressContinue() repetitively to consume input stream.
-*  *srcSizePtr and *dstCapacityPtr can be any size.
-*  The function will report how many bytes were read or written within *srcSizePtr and *dstCapacityPtr.
-*  Note that it may not consume the entire input, in which case it's up to the caller to present again remaining data.
-*  The content of `dst` will be overwritten (up to *dstCapacityPtr) at each call, so save its content if it matters or change @dst .
-*  @return : a hint to preferred nb of bytes to use as input for next function call (it's just a hint, to improve latency)
-*            or an error code, which can be tested using ZBUFF_isError().
-*
-*  At any moment, it's possible to flush whatever data remains within buffer, using ZBUFF_compressFlush().
-*  The nb of bytes written into `dst` will be reported into *dstCapacityPtr.
-*  Note that the function cannot output more than *dstCapacityPtr,
-*  therefore, some content might still be left into internal buffer if *dstCapacityPtr is too small.
-*  @return : nb of bytes still present into internal buffer (0 if it's empty)
-*            or an error code, which can be tested using ZBUFF_isError().
-*
-*  ZBUFF_compressEnd() instructs to finish a frame.
-*  It will perform a flush and write frame epilogue.
-*  The epilogue is required for decoders to consider a frame completed.
-*  Similar to ZBUFF_compressFlush(), it may not be able to output the entire internal buffer content if *dstCapacityPtr is too small.
-*  In which case, call again ZBUFF_compressFlush() to complete the flush.
-*  @return : nb of bytes still present into internal buffer (0 if it's empty)
-*            or an error code, which can be tested using ZBUFF_isError().
-*
-*  Hint : _recommended buffer_ sizes (not compulsory) : ZBUFF_recommendedCInSize() / ZBUFF_recommendedCOutSize()
-*  input : ZBUFF_recommendedCInSize==128 KB block size is the internal unit, use this value to reduce intermediate stages (better latency)
-*  output : ZBUFF_recommendedCOutSize==ZSTD_compressBound(128 KB) + 3 + 3 : ensures it's always possible to write/flush/end a full block. Skip some buffering.
-*  By using both, it ensures that input will be entirely consumed, and output will always contain the result, reducing intermediate buffering.
-* **************************************************/
-
-
-typedef ZSTD_DStream ZBUFF_DCtx;
-ZBUFF_DEPRECATED("use ZSTD_createDStream") ZBUFF_DCtx* ZBUFF_createDCtx(void);
-ZBUFF_DEPRECATED("use ZSTD_freeDStream")   size_t      ZBUFF_freeDCtx(ZBUFF_DCtx* dctx);
-
-ZBUFF_DEPRECATED("use ZSTD_initDStream")           size_t ZBUFF_decompressInit(ZBUFF_DCtx* dctx);
-ZBUFF_DEPRECATED("use ZSTD_initDStream_usingDict") size_t ZBUFF_decompressInitDictionary(ZBUFF_DCtx* dctx, const void* dict, size_t dictSize);
-
-ZBUFF_DEPRECATED("use ZSTD_decompressStream") size_t ZBUFF_decompressContinue(ZBUFF_DCtx* dctx,
-                                            void* dst, size_t* dstCapacityPtr,
-                                      const void* src, size_t* srcSizePtr);
-
-/*-***************************************************************************
-*  Streaming decompression howto
-*
-*  A ZBUFF_DCtx object is required to track streaming operations.
-*  Use ZBUFF_createDCtx() and ZBUFF_freeDCtx() to create/release resources.
-*  Use ZBUFF_decompressInit() to start a new decompression operation,
-*   or ZBUFF_decompressInitDictionary() if decompression requires a dictionary.
-*  Note that ZBUFF_DCtx objects can be re-init multiple times.
-*
-*  Use ZBUFF_decompressContinue() repetitively to consume your input.
-*  *srcSizePtr and *dstCapacityPtr can be any size.
-*  The function will report how many bytes were read or written by modifying *srcSizePtr and *dstCapacityPtr.
-*  Note that it may not consume the entire input, in which case it's up to the caller to present remaining input again.
-*  The content of `dst` will be overwritten (up to *dstCapacityPtr) at each function call, so save its content if it matters, or change `dst`.
-*  @return : 0 when a frame is completely decoded and fully flushed,
-*            1 when there is still some data left within internal buffer to flush,
-*            >1 when more data is expected, with value being a suggested next input size (it's just a hint, which helps latency),
-*            or an error code, which can be tested using ZBUFF_isError().
-*
-*  Hint : recommended buffer sizes (not compulsory) : ZBUFF_recommendedDInSize() and ZBUFF_recommendedDOutSize()
-*  output : ZBUFF_recommendedDOutSize== 128 KB block size is the internal unit, it ensures it's always possible to write a full block when decoded.
-*  input  : ZBUFF_recommendedDInSize == 128KB + 3;
-*           just follow indications from ZBUFF_decompressContinue() to minimize latency. It should always be <= 128 KB + 3 .
-* *******************************************************************************/
-
-
-/* *************************************
-*  Tool functions
-***************************************/
-ZBUFF_DEPRECATED("use ZSTD_isError")      unsigned ZBUFF_isError(size_t errorCode);
-ZBUFF_DEPRECATED("use ZSTD_getErrorName") const char* ZBUFF_getErrorName(size_t errorCode);
-
-/** Functions below provide recommended buffer sizes for Compression or Decompression operations.
-*   These sizes are just hints, they tend to offer better latency */
-ZBUFF_DEPRECATED("use ZSTD_CStreamInSize")  size_t ZBUFF_recommendedCInSize(void);
-ZBUFF_DEPRECATED("use ZSTD_CStreamOutSize") size_t ZBUFF_recommendedCOutSize(void);
-ZBUFF_DEPRECATED("use ZSTD_DStreamInSize")  size_t ZBUFF_recommendedDInSize(void);
-ZBUFF_DEPRECATED("use ZSTD_DStreamOutSize") size_t ZBUFF_recommendedDOutSize(void);
-
-#endif  /* ZSTD_BUFFERED_H_23987 */
-
-
-#ifdef ZBUFF_STATIC_LINKING_ONLY
-#ifndef ZBUFF_STATIC_H_30298098432
-#define ZBUFF_STATIC_H_30298098432
-
-/* ====================================================================================
- * The definitions in this section are considered experimental.
- * They should never be used in association with a dynamic library, as they may change in the future.
- * They are provided for advanced usages.
- * Use them only in association with static linking.
- * ==================================================================================== */
-
-/*--- Dependency ---*/
-#define ZSTD_STATIC_LINKING_ONLY   /* ZSTD_parameters, ZSTD_customMem */
-#include "zstd.h"
-
-
-/*--- Custom memory allocator ---*/
-/*! ZBUFF_createCCtx_advanced() :
- *  Create a ZBUFF compression context using external alloc and free functions */
-ZBUFF_DEPRECATED("use ZSTD_createCStream_advanced") ZBUFF_CCtx* ZBUFF_createCCtx_advanced(ZSTD_customMem customMem);
-
-/*! ZBUFF_createDCtx_advanced() :
- *  Create a ZBUFF decompression context using external alloc and free functions */
-ZBUFF_DEPRECATED("use ZSTD_createDStream_advanced") ZBUFF_DCtx* ZBUFF_createDCtx_advanced(ZSTD_customMem customMem);
-
-
-/*--- Advanced Streaming Initialization ---*/
-ZBUFF_DEPRECATED("use ZSTD_initDStream_usingDict") size_t ZBUFF_compressInit_advanced(ZBUFF_CCtx* zbc,
-                                               const void* dict, size_t dictSize,
-                                               ZSTD_parameters params, unsigned long long pledgedSrcSize);
-
-
-#endif    /* ZBUFF_STATIC_H_30298098432 */
-#endif    /* ZBUFF_STATIC_LINKING_ONLY */
-
-
-#if defined (__cplusplus)
-}
-#endif
diff --git a/zbuff_common.c b/zbuff_common.c
deleted file mode 100644
index e2b9613..0000000
--- a/zbuff_common.c
+++ /dev/null
@@ -1,26 +0,0 @@
-/*
- * Copyright (c) 2016-2020, 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.
- */
-
-/*-*************************************
-*  Dependencies
-***************************************/
-#include "error_private.h"
-#include "zbuff.h"
-
-/*-****************************************
-*  ZBUFF Error Management  (deprecated)
-******************************************/
-
-/*! ZBUFF_isError() :
-*   tells if a return value is an error code */
-unsigned ZBUFF_isError(size_t errorCode) { return ERR_isError(errorCode); }
-/*! ZBUFF_getErrorName() :
-*   provides error code string from function result (useful for debugging) */
-const char* ZBUFF_getErrorName(size_t errorCode) { return ERR_getErrorName(errorCode); }
diff --git a/zbuff_compress.c b/zbuff_compress.c
deleted file mode 100644
index 2d20b13..0000000
--- a/zbuff_compress.c
+++ /dev/null
@@ -1,147 +0,0 @@
-/*
- * Copyright (c) 2016-2020, 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.
- */
-
-
-
-/* *************************************
-*  Dependencies
-***************************************/
-#define ZBUFF_STATIC_LINKING_ONLY
-#include "zbuff.h"
-
-
-/*-***********************************************************
-*  Streaming compression
-*
-*  A ZBUFF_CCtx object is required to track streaming operation.
-*  Use ZBUFF_createCCtx() and ZBUFF_freeCCtx() to create/release resources.
-*  Use ZBUFF_compressInit() to start a new compression operation.
-*  ZBUFF_CCtx objects can be reused multiple times.
-*
-*  Use ZBUFF_compressContinue() repetitively to consume your input.
-*  *srcSizePtr and *dstCapacityPtr can be any size.
-*  The function will report how many bytes were read or written by modifying *srcSizePtr and *dstCapacityPtr.
-*  Note that it may not consume the entire input, in which case it's up to the caller to call again the function with remaining input.
-*  The content of dst will be overwritten (up to *dstCapacityPtr) at each function call, so save its content if it matters or change dst .
-*  @return : a hint to preferred nb of bytes to use as input for next function call (it's only a hint, to improve latency)
-*            or an error code, which can be tested using ZBUFF_isError().
-*
-*  ZBUFF_compressFlush() can be used to instruct ZBUFF to compress and output whatever remains within its buffer.
-*  Note that it will not output more than *dstCapacityPtr.
-*  Therefore, some content might still be left into its internal buffer if dst buffer is too small.
-*  @return : nb of bytes still present into internal buffer (0 if it's empty)
-*            or an error code, which can be tested using ZBUFF_isError().
-*
-*  ZBUFF_compressEnd() instructs to finish a frame.
-*  It will perform a flush and write frame epilogue.
-*  Similar to ZBUFF_compressFlush(), it may not be able to output the entire internal buffer content if *dstCapacityPtr is too small.
-*  @return : nb of bytes still present into internal buffer (0 if it's empty)
-*            or an error code, which can be tested using ZBUFF_isError().
-*
-*  Hint : recommended buffer sizes (not compulsory)
-*  input : ZSTD_BLOCKSIZE_MAX (128 KB), internal unit size, it improves latency to use this value.
-*  output : ZSTD_compressBound(ZSTD_BLOCKSIZE_MAX) + ZSTD_blockHeaderSize + ZBUFF_endFrameSize : ensures it's always possible to write/flush/end a full block at best speed.
-* ***********************************************************/
-
-ZBUFF_CCtx* ZBUFF_createCCtx(void)
-{
-    return ZSTD_createCStream();
-}
-
-ZBUFF_CCtx* ZBUFF_createCCtx_advanced(ZSTD_customMem customMem)
-{
-    return ZSTD_createCStream_advanced(customMem);
-}
-
-size_t ZBUFF_freeCCtx(ZBUFF_CCtx* zbc)
-{
-    return ZSTD_freeCStream(zbc);
-}
-
-
-/* ======   Initialization   ====== */
-
-size_t ZBUFF_compressInit_advanced(ZBUFF_CCtx* zbc,
-                                   const void* dict, size_t dictSize,
-                                   ZSTD_parameters params, unsigned long long pledgedSrcSize)
-{
-    if (pledgedSrcSize==0) pledgedSrcSize = ZSTD_CONTENTSIZE_UNKNOWN;  /* preserve "0 == unknown" behavior */
-    return ZSTD_initCStream_advanced(zbc, dict, dictSize, params, pledgedSrcSize);
-}
-
-
-size_t ZBUFF_compressInitDictionary(ZBUFF_CCtx* zbc, const void* dict, size_t dictSize, int compressionLevel)
-{
-    return ZSTD_initCStream_usingDict(zbc, dict, dictSize, compressionLevel);
-}
-
-size_t ZBUFF_compressInit(ZBUFF_CCtx* zbc, int compressionLevel)
-{
-    return ZSTD_initCStream(zbc, compressionLevel);
-}
-
-/* ======   Compression   ====== */
-
-
-size_t ZBUFF_compressContinue(ZBUFF_CCtx* zbc,
-                              void* dst, size_t* dstCapacityPtr,
-                        const void* src, size_t* srcSizePtr)
-{
-    size_t result;
-    ZSTD_outBuffer outBuff;
-    ZSTD_inBuffer inBuff;
-    outBuff.dst = dst;
-    outBuff.pos = 0;
-    outBuff.size = *dstCapacityPtr;
-    inBuff.src = src;
-    inBuff.pos = 0;
-    inBuff.size = *srcSizePtr;
-    result = ZSTD_compressStream(zbc, &outBuff, &inBuff);
-    *dstCapacityPtr = outBuff.pos;
-    *srcSizePtr = inBuff.pos;
-    return result;
-}
-
-
-
-/* ======   Finalize   ====== */
-
-size_t ZBUFF_compressFlush(ZBUFF_CCtx* zbc, void* dst, size_t* dstCapacityPtr)
-{
-    size_t result;
-    ZSTD_outBuffer outBuff;
-    outBuff.dst = dst;
-    outBuff.pos = 0;
-    outBuff.size = *dstCapacityPtr;
-    result = ZSTD_flushStream(zbc, &outBuff);
-    *dstCapacityPtr = outBuff.pos;
-    return result;
-}
-
-
-size_t ZBUFF_compressEnd(ZBUFF_CCtx* zbc, void* dst, size_t* dstCapacityPtr)
-{
-    size_t result;
-    ZSTD_outBuffer outBuff;
-    outBuff.dst = dst;
-    outBuff.pos = 0;
-    outBuff.size = *dstCapacityPtr;
-    result = ZSTD_endStream(zbc, &outBuff);
-    *dstCapacityPtr = outBuff.pos;
-    return result;
-}
-
-
-
-/* *************************************
-*  Tool functions
-***************************************/
-size_t ZBUFF_recommendedCInSize(void)  { return ZSTD_CStreamInSize(); }
-size_t ZBUFF_recommendedCOutSize(void) { return ZSTD_CStreamOutSize(); }
diff --git a/zbuff_decompress.c b/zbuff_decompress.c
deleted file mode 100644
index d3c49e8..0000000
--- a/zbuff_decompress.c
+++ /dev/null
@@ -1,75 +0,0 @@
-/*
- * Copyright (c) 2016-2020, 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.
- */
-
-
-
-/* *************************************
-*  Dependencies
-***************************************/
-#define ZBUFF_STATIC_LINKING_ONLY
-#include "zbuff.h"
-
-
-ZBUFF_DCtx* ZBUFF_createDCtx(void)
-{
-    return ZSTD_createDStream();
-}
-
-ZBUFF_DCtx* ZBUFF_createDCtx_advanced(ZSTD_customMem customMem)
-{
-    return ZSTD_createDStream_advanced(customMem);
-}
-
-size_t ZBUFF_freeDCtx(ZBUFF_DCtx* zbd)
-{
-    return ZSTD_freeDStream(zbd);
-}
-
-
-/* *** Initialization *** */
-
-size_t ZBUFF_decompressInitDictionary(ZBUFF_DCtx* zbd, const void* dict, size_t dictSize)
-{
-    return ZSTD_initDStream_usingDict(zbd, dict, dictSize);
-}
-
-size_t ZBUFF_decompressInit(ZBUFF_DCtx* zbd)
-{
-    return ZSTD_initDStream(zbd);
-}
-
-
-/* *** Decompression *** */
-
-size_t ZBUFF_decompressContinue(ZBUFF_DCtx* zbd,
-                                void* dst, size_t* dstCapacityPtr,
-                          const void* src, size_t* srcSizePtr)
-{
-    ZSTD_outBuffer outBuff;
-    ZSTD_inBuffer inBuff;
-    size_t result;
-    outBuff.dst  = dst;
-    outBuff.pos  = 0;
-    outBuff.size = *dstCapacityPtr;
-    inBuff.src  = src;
-    inBuff.pos  = 0;
-    inBuff.size = *srcSizePtr;
-    result = ZSTD_decompressStream(zbd, &outBuff, &inBuff);
-    *dstCapacityPtr = outBuff.pos;
-    *srcSizePtr = inBuff.pos;
-    return result;
-}
-
-
-/* *************************************
-*  Tool functions
-***************************************/
-size_t ZBUFF_recommendedDInSize(void)  { return ZSTD_DStreamInSize(); }
-size_t ZBUFF_recommendedDOutSize(void) { return ZSTD_DStreamOutSize(); }
diff --git a/zdict.c b/zdict.c
index 7ac7c60..4fd6369 100644
--- a/zdict.c
+++ b/zdict.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -23,9 +23,13 @@
 /* Unix Large Files support (>4GB) */
 #define _FILE_OFFSET_BITS 64
 #if (defined(__sun__) && (!defined(__LP64__)))   /* Sun Solaris 32-bits requires specific definitions */
+#  ifndef _LARGEFILE_SOURCE
 #  define _LARGEFILE_SOURCE
+#  endif
 #elif ! defined(__LP64__)                        /* No point defining Large file for 64 bit */
+#  ifndef _LARGEFILE64_SOURCE
 #  define _LARGEFILE64_SOURCE
+#  endif
 #endif
 
 
@@ -37,18 +41,19 @@
 #include <stdio.h>         /* fprintf, fopen, ftello64 */
 #include <time.h>          /* clock */
 
+#ifndef ZDICT_STATIC_LINKING_ONLY
+#  define ZDICT_STATIC_LINKING_ONLY
+#endif
+#define HUF_STATIC_LINKING_ONLY
+
 #include "mem.h"           /* read */
 #include "fse.h"           /* FSE_normalizeCount, FSE_writeNCount */
-#define HUF_STATIC_LINKING_ONLY
 #include "huf.h"           /* HUF_buildCTable, HUF_writeCTable */
 #include "zstd_internal.h" /* includes zstd.h */
 #include "xxhash.h"        /* XXH64 */
-#include "divsufsort.h"
-#ifndef ZDICT_STATIC_LINKING_ONLY
-#  define ZDICT_STATIC_LINKING_ONLY
-#endif
-#include "zdict.h"
 #include "zstd_compress_internal.h" /* ZSTD_loadCEntropy() */
+#include "zdict.h"
+#include "divsufsort.h"
 
 
 /*-*************************************
@@ -967,16 +972,11 @@ static size_t ZDICT_addEntropyTablesFromBuffer_advanced(
     return MIN(dictBufferCapacity, hSize+dictContentSize);
 }
 
-/* Hidden declaration for dbio.c */
-size_t ZDICT_trainFromBuffer_unsafe_legacy(
-                            void* dictBuffer, size_t maxDictSize,
-                            const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
-                            ZDICT_legacy_params_t params);
 /*! ZDICT_trainFromBuffer_unsafe_legacy() :
-*   Warning : `samplesBuffer` must be followed by noisy guard band.
+*   Warning : `samplesBuffer` must be followed by noisy guard band !!!
 *   @return : size of dictionary, or an error code which can be tested with ZDICT_isError()
 */
-size_t ZDICT_trainFromBuffer_unsafe_legacy(
+static size_t ZDICT_trainFromBuffer_unsafe_legacy(
                             void* dictBuffer, size_t maxDictSize,
                             const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
                             ZDICT_legacy_params_t params)
diff --git a/zdict.h b/zdict.h
index b782993..75b05db 100644
--- a/zdict.h
+++ b/zdict.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -36,6 +36,145 @@ extern "C" {
 #  define ZDICTLIB_API ZDICTLIB_VISIBILITY
 #endif
 
+/*******************************************************************************
+ * Zstd dictionary builder
+ *
+ * FAQ
+ * ===
+ * Why should I use a dictionary?
+ * ------------------------------
+ *
+ * Zstd can use dictionaries to improve compression ratio of small data.
+ * Traditionally small files don't compress well because there is very little
+ * repetion in a single sample, since it is small. But, if you are compressing
+ * many similar files, like a bunch of JSON records that share the same
+ * structure, you can train a dictionary on ahead of time on some samples of
+ * these files. Then, zstd can use the dictionary to find repetitions that are
+ * present across samples. This can vastly improve compression ratio.
+ *
+ * When is a dictionary useful?
+ * ----------------------------
+ *
+ * Dictionaries are useful when compressing many small files that are similar.
+ * The larger a file is, the less benefit a dictionary will have. Generally,
+ * we don't expect dictionary compression to be effective past 100KB. And the
+ * smaller a file is, the more we would expect the dictionary to help.
+ *
+ * How do I use a dictionary?
+ * --------------------------
+ *
+ * Simply pass the dictionary to the zstd compressor with
+ * `ZSTD_CCtx_loadDictionary()`. The same dictionary must then be passed to
+ * the decompressor, using `ZSTD_DCtx_loadDictionary()`. There are other
+ * more advanced functions that allow selecting some options, see zstd.h for
+ * complete documentation.
+ *
+ * What is a zstd dictionary?
+ * --------------------------
+ *
+ * A zstd dictionary has two pieces: Its header, and its content. The header
+ * contains a magic number, the dictionary ID, and entropy tables. These
+ * entropy tables allow zstd to save on header costs in the compressed file,
+ * which really matters for small data. The content is just bytes, which are
+ * repeated content that is common across many samples.
+ *
+ * What is a raw content dictionary?
+ * ---------------------------------
+ *
+ * A raw content dictionary is just bytes. It doesn't have a zstd dictionary
+ * header, a dictionary ID, or entropy tables. Any buffer is a valid raw
+ * content dictionary.
+ *
+ * How do I train a dictionary?
+ * ----------------------------
+ *
+ * Gather samples from your use case. These samples should be similar to each
+ * other. If you have several use cases, you could try to train one dictionary
+ * per use case.
+ *
+ * Pass those samples to `ZDICT_trainFromBuffer()` and that will train your
+ * dictionary. There are a few advanced versions of this function, but this
+ * is a great starting point. If you want to further tune your dictionary
+ * you could try `ZDICT_optimizeTrainFromBuffer_cover()`. If that is too slow
+ * you can try `ZDICT_optimizeTrainFromBuffer_fastCover()`.
+ *
+ * If the dictionary training function fails, that is likely because you
+ * either passed too few samples, or a dictionary would not be effective
+ * for your data. Look at the messages that the dictionary trainer printed,
+ * if it doesn't say too few samples, then a dictionary would not be effective.
+ *
+ * How large should my dictionary be?
+ * ----------------------------------
+ *
+ * A reasonable dictionary size, the `dictBufferCapacity`, is about 100KB.
+ * The zstd CLI defaults to a 110KB dictionary. You likely don't need a
+ * dictionary larger than that. But, most use cases can get away with a
+ * smaller dictionary. The advanced dictionary builders can automatically
+ * shrink the dictionary for you, and select a the smallest size that
+ * doesn't hurt compression ratio too much. See the `shrinkDict` parameter.
+ * A smaller dictionary can save memory, and potentially speed up
+ * compression.
+ *
+ * How many samples should I provide to the dictionary builder?
+ * ------------------------------------------------------------
+ *
+ * We generally recommend passing ~100x the size of the dictionary
+ * in samples. A few thousand should suffice. Having too few samples
+ * can hurt the dictionaries effectiveness. Having more samples will
+ * only improve the dictionaries effectiveness. But having too many
+ * samples can slow down the dictionary builder.
+ *
+ * How do I determine if a dictionary will be effective?
+ * -----------------------------------------------------
+ *
+ * Simply train a dictionary and try it out. You can use zstd's built in
+ * benchmarking tool to test the dictionary effectiveness.
+ *
+ *   # Benchmark levels 1-3 without a dictionary
+ *   zstd -b1e3 -r /path/to/my/files
+ *   # Benchmark levels 1-3 with a dictioanry
+ *   zstd -b1e3 -r /path/to/my/files -D /path/to/my/dictionary
+ *
+ * When should I retrain a dictionary?
+ * -----------------------------------
+ *
+ * You should retrain a dictionary when its effectiveness drops. Dictionary
+ * effectiveness drops as the data you are compressing changes. Generally, we do
+ * expect dictionaries to "decay" over time, as your data changes, but the rate
+ * at which they decay depends on your use case. Internally, we regularly
+ * retrain dictionaries, and if the new dictionary performs significantly
+ * better than the old dictionary, we will ship the new dictionary.
+ *
+ * I have a raw content dictionary, how do I turn it into a zstd dictionary?
+ * -------------------------------------------------------------------------
+ *
+ * If you have a raw content dictionary, e.g. by manually constructing it, or
+ * using a third-party dictionary builder, you can turn it into a zstd
+ * dictionary by using `ZDICT_finalizeDictionary()`. You'll also have to
+ * provide some samples of the data. It will add the zstd header to the
+ * raw content, which contains a dictionary ID and entropy tables, which
+ * will improve compression ratio, and allow zstd to write the dictionary ID
+ * into the frame, if you so choose.
+ *
+ * Do I have to use zstd's dictionary builder?
+ * -------------------------------------------
+ *
+ * No! You can construct dictionary content however you please, it is just
+ * bytes. It will always be valid as a raw content dictionary. If you want
+ * a zstd dictionary, which can improve compression ratio, use
+ * `ZDICT_finalizeDictionary()`.
+ *
+ * What is the attack surface of a zstd dictionary?
+ * ------------------------------------------------
+ *
+ * Zstd is heavily fuzz tested, including loading fuzzed dictionaries, so
+ * zstd should never crash, or access out-of-bounds memory no matter what
+ * the dictionary is. However, if an attacker can control the dictionary
+ * during decompression, they can cause zstd to generate arbitrary bytes,
+ * just like if they controlled the compressed data.
+ *
+ ******************************************************************************/
+
 
 /*! ZDICT_trainFromBuffer():
  *  Train a dictionary from an array of samples.
@@ -64,7 +203,14 @@ ZDICTLIB_API size_t ZDICT_trainFromBuffer(void* dictBuffer, size_t dictBufferCap
 typedef struct {
     int      compressionLevel;   /*< optimize for a specific zstd compression level; 0 means default */
     unsigned notificationLevel;  /*< Write log to stderr; 0 = none (default); 1 = errors; 2 = progression; 3 = details; 4 = debug; */
-    unsigned dictID;             /*< force dictID value; 0 means auto mode (32-bits random value) */
+    unsigned dictID;             /*< force dictID value; 0 means auto mode (32-bits random value)
+                                  *   NOTE: The zstd format reserves some dictionary IDs for future use.
+                                  *         You may use them in private settings, but be warned that they
+                                  *         may be used by zstd in a public dictionary registry in the future.
+                                  *         These dictionary IDs are:
+                                  *           - low range  : <= 32767
+                                  *           - high range : >= (2^31)
+                                  */
 } ZDICT_params_t;
 
 /*! ZDICT_finalizeDictionary():
@@ -264,10 +410,11 @@ typedef struct {
  *  Note: ZDICT_trainFromBuffer_legacy() will send notifications into stderr if instructed to, using notificationLevel>0.
  */
 ZDICTLIB_API size_t ZDICT_trainFromBuffer_legacy(
-    void *dictBuffer, size_t dictBufferCapacity,
-    const void *samplesBuffer, const size_t *samplesSizes, unsigned nbSamples,
+    void* dictBuffer, size_t dictBufferCapacity,
+    const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
     ZDICT_legacy_params_t parameters);
 
+
 /* Deprecation warnings */
 /* It is generally possible to disable deprecation warnings from compiler,
    for example with -Wno-deprecated-declarations for gcc
diff --git a/zstd.h b/zstd.h
index b0ecdf5..ebe2e04 100644
--- a/zstd.h
+++ b/zstd.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -71,8 +71,8 @@ extern "C" {
 
 /*------   Version   ------*/
 #define ZSTD_VERSION_MAJOR    1
-#define ZSTD_VERSION_MINOR    4
-#define ZSTD_VERSION_RELEASE  8
+#define ZSTD_VERSION_MINOR    5
+#define ZSTD_VERSION_RELEASE  0
 #define ZSTD_VERSION_NUMBER  (ZSTD_VERSION_MAJOR *100*100 + ZSTD_VERSION_MINOR *100 + ZSTD_VERSION_RELEASE)
 
 /*! ZSTD_versionNumber() :
@@ -109,7 +109,6 @@ ZSTDLIB_API const char* ZSTD_versionString(void);
 #define ZSTD_BLOCKSIZE_MAX     (1<<ZSTD_BLOCKSIZELOG_MAX)
 
 
-
 /***************************************
 *  Simple API
 ***************************************/
@@ -166,7 +165,7 @@ ZSTDLIB_API unsigned long long ZSTD_getFrameContentSize(const void *src, size_t
  * @return : decompressed size of `src` frame content _if known and not empty_, 0 otherwise. */
 ZSTDLIB_API unsigned long long ZSTD_getDecompressedSize(const void* src, size_t srcSize);
 
-/*! ZSTD_findFrameCompressedSize() :
+/*! ZSTD_findFrameCompressedSize() : Requires v1.4.0+
  * `src` should point to the start of a ZSTD frame or skippable frame.
  * `srcSize` must be >= first frame size
  * @return : the compressed size of the first frame starting at `src`,
@@ -180,8 +179,9 @@ ZSTDLIB_API size_t ZSTD_findFrameCompressedSize(const void* src, size_t srcSize)
 ZSTDLIB_API size_t      ZSTD_compressBound(size_t srcSize); /*!< maximum compressed size in worst case single-pass scenario */
 ZSTDLIB_API unsigned    ZSTD_isError(size_t code);          /*!< tells if a `size_t` function result is an error code */
 ZSTDLIB_API const char* ZSTD_getErrorName(size_t code);     /*!< provides readable string from an error code */
-ZSTDLIB_API int         ZSTD_minCLevel(void);               /*!< minimum negative compression level allowed */
+ZSTDLIB_API int         ZSTD_minCLevel(void);               /*!< minimum negative compression level allowed, requires v1.4.0+ */
 ZSTDLIB_API int         ZSTD_maxCLevel(void);               /*!< maximum compression level available */
+ZSTDLIB_API int         ZSTD_defaultCLevel(void);           /*!< default compression level, specified by ZSTD_CLEVEL_DEFAULT, requires v1.5.0+ */
 
 
 /***************************************
@@ -199,7 +199,7 @@ ZSTDLIB_API int         ZSTD_maxCLevel(void);               /*!< maximum compres
  */
 typedef struct ZSTD_CCtx_s ZSTD_CCtx;
 ZSTDLIB_API ZSTD_CCtx* ZSTD_createCCtx(void);
-ZSTDLIB_API size_t     ZSTD_freeCCtx(ZSTD_CCtx* cctx);
+ZSTDLIB_API size_t     ZSTD_freeCCtx(ZSTD_CCtx* cctx);  /* accept NULL pointer */
 
 /*! ZSTD_compressCCtx() :
  *  Same as ZSTD_compress(), using an explicit ZSTD_CCtx.
@@ -222,7 +222,7 @@ ZSTDLIB_API size_t ZSTD_compressCCtx(ZSTD_CCtx* cctx,
  *  Use one context per thread for parallel execution. */
 typedef struct ZSTD_DCtx_s ZSTD_DCtx;
 ZSTDLIB_API ZSTD_DCtx* ZSTD_createDCtx(void);
-ZSTDLIB_API size_t     ZSTD_freeDCtx(ZSTD_DCtx* dctx);
+ZSTDLIB_API size_t     ZSTD_freeDCtx(ZSTD_DCtx* dctx);  /* accept NULL pointer */
 
 /*! ZSTD_decompressDCtx() :
  *  Same as ZSTD_decompress(),
@@ -234,9 +234,9 @@ ZSTDLIB_API size_t ZSTD_decompressDCtx(ZSTD_DCtx* dctx,
                                  const void* src, size_t srcSize);
 
 
-/***************************************
-*  Advanced compression API
-***************************************/
+/*********************************************
+*  Advanced compression API (Requires v1.4.0+)
+**********************************************/
 
 /* API design :
  *   Parameters are pushed one by one into an existing context,
@@ -266,7 +266,6 @@ typedef enum { ZSTD_fast=1,
                          Only the order (from fast to strong) is guaranteed */
 } ZSTD_strategy;
 
-
 typedef enum {
 
     /* compression parameters
@@ -332,7 +331,6 @@ typedef enum {
                               * The higher the value of selected strategy, the more complex it is,
                               * resulting in stronger and slower compression.
                               * Special: value 0 means "use default strategy". */
-
     /* LDM mode parameters */
     ZSTD_c_enableLongDistanceMatching=160, /* Enable long distance matching.
                                      * This parameter is designed to improve compression ratio
@@ -389,7 +387,7 @@ typedef enum {
     ZSTD_c_jobSize=401,      /* Size of a compression job. This value is enforced only when nbWorkers >= 1.
                               * Each compression job is completed in parallel, so this value can indirectly impact the nb of active threads.
                               * 0 means default, which is dynamically determined based on compression parameters.
-                              * Job size must be a minimum of overlap size, or 1 MB, whichever is largest.
+                              * Job size must be a minimum of overlap size, or ZSTDMT_JOBSIZE_MIN (= 512 KB), whichever is largest.
                               * The minimum size is automatically and transparently enforced. */
     ZSTD_c_overlapLog=402,   /* Control the overlap size, as a fraction of window size.
                               * The overlap size is an amount of data reloaded from previous job at the beginning of a new job.
@@ -419,6 +417,8 @@ typedef enum {
      * ZSTD_c_stableOutBuffer
      * ZSTD_c_blockDelimiters
      * ZSTD_c_validateSequences
+     * ZSTD_c_splitBlocks
+     * 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;
      *        also, the enums values themselves are unstable and can still change.
@@ -434,7 +434,10 @@ typedef enum {
      ZSTD_c_experimentalParam9=1006,
      ZSTD_c_experimentalParam10=1007,
      ZSTD_c_experimentalParam11=1008,
-     ZSTD_c_experimentalParam12=1009
+     ZSTD_c_experimentalParam12=1009,
+     ZSTD_c_experimentalParam13=1010,
+     ZSTD_c_experimentalParam14=1011,
+     ZSTD_c_experimentalParam15=1012
 } ZSTD_cParameter;
 
 typedef struct {
@@ -519,9 +522,9 @@ ZSTDLIB_API size_t ZSTD_compress2( ZSTD_CCtx* cctx,
                              const void* src, size_t srcSize);
 
 
-/***************************************
-*  Advanced decompression API
-***************************************/
+/***********************************************
+*  Advanced decompression API (Requires v1.4.0+)
+************************************************/
 
 /* The advanced API pushes parameters one by one into an existing DCtx context.
  * Parameters are sticky, and remain valid for all following frames
@@ -546,12 +549,14 @@ typedef enum {
      * ZSTD_d_format
      * ZSTD_d_stableOutBuffer
      * ZSTD_d_forceIgnoreChecksum
+     * ZSTD_d_refMultipleDDicts
      * Because they are not stable, it's necessary to define ZSTD_STATIC_LINKING_ONLY to access them.
      * note : never ever use experimentalParam? names directly
      */
      ZSTD_d_experimentalParam1=1000,
      ZSTD_d_experimentalParam2=1001,
-     ZSTD_d_experimentalParam3=1002
+     ZSTD_d_experimentalParam3=1002,
+     ZSTD_d_experimentalParam4=1003
 
 } ZSTD_dParameter;
 
@@ -665,7 +670,7 @@ typedef ZSTD_CCtx ZSTD_CStream;  /**< CCtx and CStream are now effectively same
                                  /* Continue to distinguish them for compatibility with older versions <= v1.2.0 */
 /*===== ZSTD_CStream management functions =====*/
 ZSTDLIB_API ZSTD_CStream* ZSTD_createCStream(void);
-ZSTDLIB_API size_t ZSTD_freeCStream(ZSTD_CStream* zcs);
+ZSTDLIB_API size_t ZSTD_freeCStream(ZSTD_CStream* zcs);  /* accept NULL pointer */
 
 /*===== Streaming compression functions =====*/
 typedef enum {
@@ -681,7 +686,7 @@ typedef enum {
                         : note : multithreaded compression will block to flush as much output as possible. */
 } ZSTD_EndDirective;
 
-/*! ZSTD_compressStream2() :
+/*! ZSTD_compressStream2() : Requires v1.4.0+
  *  Behaves about the same as ZSTD_compressStream, with additional control on end directive.
  *  - Compression parameters are pushed into CCtx before starting compression, using ZSTD_CCtx_set*()
  *  - Compression parameters cannot be changed once compression is started (save a list of exceptions in multi-threading mode)
@@ -727,11 +732,11 @@ ZSTDLIB_API size_t ZSTD_CStreamOutSize(void);   /**< recommended size for output
 
 
 /* *****************************************************************************
- * This following is a legacy streaming API.
+ * This following is a legacy streaming API, available since v1.0+ .
  * It can be replaced by ZSTD_CCtx_reset() and ZSTD_compressStream2().
  * It is redundant, but remains fully supported.
- * Advanced parameters and dictionary compression can only be used through the
- * new API.
+ * Streaming in combination with advanced parameters and dictionary compression
+ * can only be used through the new API.
  ******************************************************************************/
 
 /*!
@@ -786,7 +791,7 @@ typedef ZSTD_DCtx ZSTD_DStream;  /**< DCtx and DStream are now effectively same
                                  /* For compatibility with versions <= v1.2.0, prefer differentiating them. */
 /*===== ZSTD_DStream management functions =====*/
 ZSTDLIB_API ZSTD_DStream* ZSTD_createDStream(void);
-ZSTDLIB_API size_t ZSTD_freeDStream(ZSTD_DStream* zds);
+ZSTDLIB_API size_t ZSTD_freeDStream(ZSTD_DStream* zds);  /* accept NULL pointer */
 
 /*===== Streaming decompression functions =====*/
 
@@ -809,7 +814,7 @@ ZSTDLIB_API size_t ZSTD_DStreamOutSize(void);   /*!< recommended size for output
 /*! ZSTD_compress_usingDict() :
  *  Compression at an explicit compression level using a Dictionary.
  *  A dictionary can be any arbitrary data segment (also called a prefix),
- *  or a buffer with specified information (see dictBuilder/zdict.h).
+ *  or a buffer with specified information (see zdict.h).
  *  Note : This function loads the dictionary, resulting in significant startup delay.
  *         It's intended for a dictionary used only once.
  *  Note 2 : When `dict == NULL || dictSize < 8` no dictionary is used. */
@@ -852,7 +857,8 @@ ZSTDLIB_API ZSTD_CDict* ZSTD_createCDict(const void* dictBuffer, size_t dictSize
                                          int compressionLevel);
 
 /*! ZSTD_freeCDict() :
- *  Function frees memory allocated by ZSTD_createCDict(). */
+ *  Function frees memory allocated by ZSTD_createCDict().
+ *  If a NULL pointer is passed, no operation is performed. */
 ZSTDLIB_API size_t      ZSTD_freeCDict(ZSTD_CDict* CDict);
 
 /*! ZSTD_compress_usingCDict() :
@@ -874,7 +880,8 @@ typedef struct ZSTD_DDict_s ZSTD_DDict;
 ZSTDLIB_API ZSTD_DDict* ZSTD_createDDict(const void* dictBuffer, size_t dictSize);
 
 /*! ZSTD_freeDDict() :
- *  Function frees memory allocated with ZSTD_createDDict() */
+ *  Function frees memory allocated with ZSTD_createDDict()
+ *  If a NULL pointer is passed, no operation is performed. */
 ZSTDLIB_API size_t      ZSTD_freeDDict(ZSTD_DDict* ddict);
 
 /*! ZSTD_decompress_usingDDict() :
@@ -890,19 +897,25 @@ ZSTDLIB_API size_t ZSTD_decompress_usingDDict(ZSTD_DCtx* dctx,
  *  Dictionary helper functions
  *******************************/
 
-/*! ZSTD_getDictID_fromDict() :
+/*! ZSTD_getDictID_fromDict() : Requires v1.4.0+
  *  Provides the dictID stored within dictionary.
  *  if @return == 0, the dictionary is not conformant with Zstandard specification.
  *  It can still be loaded, but as a content-only dictionary. */
 ZSTDLIB_API unsigned ZSTD_getDictID_fromDict(const void* dict, size_t dictSize);
 
-/*! ZSTD_getDictID_fromDDict() :
+/*! ZSTD_getDictID_fromCDict() : Requires v1.5.0+
+ *  Provides the dictID of the dictionary loaded into `cdict`.
+ *  If @return == 0, the dictionary is not conformant to Zstandard specification, or empty.
+ *  Non-conformant dictionaries can still be loaded, but as content-only dictionaries. */
+ZSTDLIB_API unsigned ZSTD_getDictID_fromCDict(const ZSTD_CDict* cdict);
+
+/*! ZSTD_getDictID_fromDDict() : Requires v1.4.0+
  *  Provides the dictID of the dictionary loaded into `ddict`.
  *  If @return == 0, the dictionary is not conformant to Zstandard specification, or empty.
  *  Non-conformant dictionaries can still be loaded, but as content-only dictionaries. */
 ZSTDLIB_API unsigned ZSTD_getDictID_fromDDict(const ZSTD_DDict* ddict);
 
-/*! ZSTD_getDictID_fromFrame() :
+/*! ZSTD_getDictID_fromFrame() : Requires v1.4.0+
  *  Provides the dictID required to decompressed the frame stored within `src`.
  *  If @return == 0, the dictID could not be decoded.
  *  This could for one of the following reasons :
@@ -916,7 +929,7 @@ ZSTDLIB_API unsigned ZSTD_getDictID_fromFrame(const void* src, size_t srcSize);
 
 
 /*******************************************************************************
- * Advanced dictionary and prefix API
+ * 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
@@ -925,7 +938,7 @@ ZSTDLIB_API unsigned ZSTD_getDictID_fromFrame(const void* src, size_t srcSize);
  ******************************************************************************/
 
 
-/*! ZSTD_CCtx_loadDictionary() :
+/*! ZSTD_CCtx_loadDictionary() : Requires v1.4.0+
  *  Create an internal CDict from `dict` buffer.
  *  Decompression will have to use same dictionary.
  * @result : 0, or an error code (which can be tested with ZSTD_isError()).
@@ -944,11 +957,11 @@ ZSTDLIB_API unsigned ZSTD_getDictID_fromFrame(const void* src, size_t srcSize);
  *           to precisely select how dictionary content must be interpreted. */
 ZSTDLIB_API size_t ZSTD_CCtx_loadDictionary(ZSTD_CCtx* cctx, const void* dict, size_t dictSize);
 
-/*! ZSTD_CCtx_refCDict() :
+/*! ZSTD_CCtx_refCDict() : Requires v1.4.0+
  *  Reference a prepared dictionary, to be used for all next compressed frames.
  *  Note that compression parameters are enforced from within CDict,
  *  and supersede any compression parameter previously set within CCtx.
- *  The parameters ignored are labled as "superseded-by-cdict" in the ZSTD_cParameter enum docs.
+ *  The parameters ignored are labelled as "superseded-by-cdict" in the ZSTD_cParameter enum docs.
  *  The ignored parameters will be used again if the CCtx is returned to no-dictionary mode.
  *  The dictionary will remain valid for future compressed frames using same CCtx.
  * @result : 0, or an error code (which can be tested with ZSTD_isError()).
@@ -958,7 +971,7 @@ ZSTDLIB_API size_t ZSTD_CCtx_loadDictionary(ZSTD_CCtx* cctx, const void* dict, s
  *  Note 2 : CDict is just referenced, its lifetime must outlive its usage within CCtx. */
 ZSTDLIB_API size_t ZSTD_CCtx_refCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict);
 
-/*! ZSTD_CCtx_refPrefix() :
+/*! ZSTD_CCtx_refPrefix() : Requires v1.4.0+
  *  Reference a prefix (single-usage dictionary) for next compressed frame.
  *  A prefix is **only used once**. Tables are discarded at end of frame (ZSTD_e_end).
  *  Decompression will need same prefix to properly regenerate data.
@@ -979,7 +992,7 @@ ZSTDLIB_API size_t ZSTD_CCtx_refCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict);
 ZSTDLIB_API size_t ZSTD_CCtx_refPrefix(ZSTD_CCtx* cctx,
                                  const void* prefix, size_t prefixSize);
 
-/*! ZSTD_DCtx_loadDictionary() :
+/*! ZSTD_DCtx_loadDictionary() : Requires v1.4.0+
  *  Create an internal DDict from dict buffer,
  *  to be used to decompress next frames.
  *  The dictionary remains valid for all future frames, until explicitly invalidated.
@@ -996,9 +1009,16 @@ ZSTDLIB_API size_t ZSTD_CCtx_refPrefix(ZSTD_CCtx* cctx,
  */
 ZSTDLIB_API size_t ZSTD_DCtx_loadDictionary(ZSTD_DCtx* dctx, const void* dict, size_t dictSize);
 
-/*! ZSTD_DCtx_refDDict() :
+/*! ZSTD_DCtx_refDDict() : Requires v1.4.0+
  *  Reference a prepared dictionary, to be used to decompress next frames.
  *  The dictionary remains active for decompression of future frames using same DCtx.
+ *
+ *  If called with ZSTD_d_refMultipleDDicts enabled, repeated calls of this function
+ *  will store the DDict references in a table, and the DDict used for decompression
+ *  will be determined at decompression time, as per the dict ID in the frame.
+ *  The memory for the table is allocated on the first call to refDDict, and can be
+ *  freed with ZSTD_freeDCtx().
+ *
  * @result : 0, or an error code (which can be tested with ZSTD_isError()).
  *  Note 1 : Currently, only one dictionary can be managed.
  *           Referencing a new dictionary effectively "discards" any previous one.
@@ -1007,7 +1027,7 @@ ZSTDLIB_API size_t ZSTD_DCtx_loadDictionary(ZSTD_DCtx* dctx, const void* dict, s
  */
 ZSTDLIB_API size_t ZSTD_DCtx_refDDict(ZSTD_DCtx* dctx, const ZSTD_DDict* ddict);
 
-/*! ZSTD_DCtx_refPrefix() :
+/*! ZSTD_DCtx_refPrefix() : Requires v1.4.0+
  *  Reference a prefix (single-usage dictionary) to decompress next frame.
  *  This is the reverse operation of ZSTD_CCtx_refPrefix(),
  *  and must use the same prefix as the one used during compression.
@@ -1028,7 +1048,7 @@ ZSTDLIB_API size_t ZSTD_DCtx_refPrefix(ZSTD_DCtx* dctx,
 
 /* ===   Memory management   === */
 
-/*! ZSTD_sizeof_*() :
+/*! ZSTD_sizeof_*() : Requires v1.4.0+
  *  These functions give the _current_ memory usage of selected object.
  *  Note that object memory usage can evolve (increase or decrease) over time. */
 ZSTDLIB_API size_t ZSTD_sizeof_CCtx(const ZSTD_CCtx* cctx);
@@ -1053,6 +1073,28 @@ 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
 
+/* 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 */
+#else
+#  if defined (__cplusplus) && (__cplusplus >= 201402) /* C++14 or greater */
+#    define ZSTD_DEPRECATED(message) [[deprecated(message)]] ZSTDLIB_API
+#  elif (defined(GNUC) && (GNUC > 4 || (GNUC == 4 && GNUC_MINOR >= 5))) || defined(__clang__)
+#    define ZSTD_DEPRECATED(message) ZSTDLIB_API __attribute__((deprecated(message)))
+#  elif defined(__GNUC__) && (__GNUC__ >= 3)
+#    define ZSTD_DEPRECATED(message) ZSTDLIB_API __attribute__((deprecated))
+#  elif defined(_MSC_VER)
+#    define ZSTD_DEPRECATED(message) ZSTDLIB_API __declspec(deprecated(message))
+#  else
+#    pragma message("WARNING: You need to implement ZSTD_DEPRECATED for this compiler")
+#    define ZSTD_DEPRECATED(message) ZSTDLIB_API
+#  endif
+#endif /* ZSTD_DISABLE_DEPRECATE_WARNINGS */
+
 /****************************************************************************************
  *   experimental API (static linking only)
  ****************************************************************************************
@@ -1115,9 +1157,6 @@ 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  --- */
 
@@ -1205,6 +1244,12 @@ typedef enum {
     ZSTD_d_ignoreChecksum = 1
 } ZSTD_forceIgnoreChecksum_e;
 
+typedef enum {
+    /* Note: this enum controls ZSTD_d_refMultipleDDicts */
+    ZSTD_rmd_refSingleDDict = 0,
+    ZSTD_rmd_refMultipleDDicts = 1
+} ZSTD_refMultipleDDicts_e;
+
 typedef enum {
     /* Note: this enum and the behavior it controls are effectively internal
      * implementation details of the compressor. They are expected to continue
@@ -1253,6 +1298,11 @@ typedef enum {
   ZSTD_lcm_uncompressed = 2   /**< Always emit uncompressed literals. */
 } ZSTD_literalCompressionMode_e;
 
+typedef enum {
+  ZSTD_urm_auto = 0,                   /* Automatically determine whether or not we use row matchfinder */
+  ZSTD_urm_disableRowMatchFinder = 1,  /* Never use row matchfinder */
+  ZSTD_urm_enableRowMatchFinder = 2    /* Always use row matchfinder when applicable */
+} ZSTD_useRowMatchFinderMode_e;
 
 /***************************************
 *  Frame size functions
@@ -1286,7 +1336,7 @@ ZSTDLIB_API unsigned long long ZSTD_findDecompressedSize(const void* src, size_t
  *  `srcSize` must be the _exact_ size of this series
  *       (i.e. there should be a frame boundary at `src + srcSize`)
  *  @return : - upper-bound for the decompressed size of all data in all successive frames
- *            - if an error occured: ZSTD_CONTENTSIZE_ERROR
+ *            - if an error occurred: ZSTD_CONTENTSIZE_ERROR
  *
  *  note 1  : an error can occur if `src` contains an invalid or incorrectly formatted frame.
  *  note 2  : the upper-bound is exact when the decompressed size field is available in every ZSTD encoded frame of `src`.
@@ -1372,6 +1422,43 @@ ZSTDLIB_API size_t ZSTD_compressSequences(ZSTD_CCtx* const cctx, void* dst, size
                                   const void* src, size_t srcSize);
 
 
+/*! ZSTD_writeSkippableFrame() :
+ * Generates a zstd skippable frame containing data given by src, and writes it to dst buffer.
+ *
+ * Skippable frames begin with a a 4-byte magic number. There are 16 possible choices of magic number,
+ * ranging from ZSTD_MAGIC_SKIPPABLE_START to ZSTD_MAGIC_SKIPPABLE_START+15.
+ * As such, the parameter magicVariant controls the exact skippable frame magic number variant used, so
+ * the magic number used will be ZSTD_MAGIC_SKIPPABLE_START + magicVariant.
+ *
+ * Returns an error if destination buffer is not large enough, if the source size is not representable
+ * with a 4-byte unsigned int, or if the parameter magicVariant is greater than 15 (and therefore invalid).
+ *
+ * @return : number of bytes written or a ZSTD error.
+ */
+ZSTDLIB_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
 ***************************************/
@@ -1506,13 +1593,14 @@ ZSTDLIB_API ZSTD_CDict* ZSTD_createCDict_advanced(const void* dict, size_t dictS
  * Note that the lifetime of such pool must exist while being used.
  * ZSTD_CCtx_refThreadPool assigns a thread pool to a context (use NULL argument value
  * to use an internal thread pool).
- * ZSTD_freeThreadPool frees a thread pool.
+ * 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_API void ZSTD_freeThreadPool (ZSTD_threadPool* pool);
+ZSTDLIB_API void ZSTD_freeThreadPool (ZSTD_threadPool* pool);  /* accept NULL pointer */
 ZSTDLIB_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!
  */
@@ -1523,10 +1611,12 @@ ZSTDLIB_API ZSTD_CDict* ZSTD_createCDict_advanced2(
     const ZSTD_CCtx_params* cctxParams,
     ZSTD_customMem customMem);
 
-ZSTDLIB_API ZSTD_DDict* ZSTD_createDDict_advanced(const void* dict, size_t dictSize,
-                                                  ZSTD_dictLoadMethod_e dictLoadMethod,
-                                                  ZSTD_dictContentType_e dictContentType,
-                                                  ZSTD_customMem customMem);
+ZSTDLIB_API ZSTD_DDict* ZSTD_createDDict_advanced(
+    const void* dict, size_t dictSize,
+    ZSTD_dictLoadMethod_e dictLoadMethod,
+    ZSTD_dictContentType_e dictContentType,
+    ZSTD_customMem customMem);
+
 
 /***************************************
 *  Advanced compression functions
@@ -1540,12 +1630,6 @@ ZSTDLIB_API ZSTD_DDict* ZSTD_createDDict_advanced(const void* dict, size_t dictS
  *  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);
 
-/*! ZSTD_getDictID_fromCDict() :
- *  Provides the dictID of the dictionary loaded into `cdict`.
- *  If @return == 0, the dictionary is not conformant to Zstandard specification, or empty.
- *  Non-conformant dictionaries can still be loaded, but as content-only dictionaries. */
-ZSTDLIB_API unsigned ZSTD_getDictID_fromCDict(const ZSTD_CDict* cdict);
-
 /*! ZSTD_getCParams() :
  * @return ZSTD_compressionParameters structure for a selected compression level and estimated srcSize.
  * `estimatedSrcSize` value is optional, select 0 if not known */
@@ -1572,18 +1656,20 @@ ZSTDLIB_API ZSTD_compressionParameters ZSTD_adjustCParams(ZSTD_compressionParame
 /*! ZSTD_compress_advanced() :
  *  Note : this function is now DEPRECATED.
  *         It can be replaced by ZSTD_compress2(), in combination with ZSTD_CCtx_setParameter() and other parameter setters.
- *  This prototype will be marked as deprecated and generate compilation warning on reaching v1.5.x */
-ZSTDLIB_API size_t ZSTD_compress_advanced(ZSTD_CCtx* cctx,
+ *  This prototype will generate compilation warnings. */
+ZSTD_DEPRECATED("use ZSTD_compress2")
+size_t ZSTD_compress_advanced(ZSTD_CCtx* cctx,
                                           void* dst, size_t dstCapacity,
                                     const void* src, size_t srcSize,
                                     const void* dict,size_t dictSize,
                                           ZSTD_parameters params);
 
 /*! ZSTD_compress_usingCDict_advanced() :
- *  Note : this function is now REDUNDANT.
+ *  Note : this function is now DEPRECATED.
  *         It can be replaced by ZSTD_compress2(), in combination with ZSTD_CCtx_loadDictionary() and other parameter setters.
- *  This prototype will be marked as deprecated and generate compilation warning in some future version */
-ZSTDLIB_API size_t ZSTD_compress_usingCDict_advanced(ZSTD_CCtx* cctx,
+ *  This prototype will generate compilation warnings. */
+ZSTD_DEPRECATED("use ZSTD_compress2 with ZSTD_CCtx_loadDictionary")
+size_t ZSTD_compress_usingCDict_advanced(ZSTD_CCtx* cctx,
                                               void* dst, size_t dstCapacity,
                                         const void* src, size_t srcSize,
                                         const ZSTD_CDict* cdict,
@@ -1645,7 +1731,7 @@ ZSTDLIB_API size_t ZSTD_CCtx_refPrefix_advanced(ZSTD_CCtx* cctx, const void* pre
 
 /* Controls how the literals are compressed (default is auto).
  * The value must be of type ZSTD_literalCompressionMode_e.
- * See ZSTD_literalCompressionMode_t enum definition for details.
+ * See ZSTD_literalCompressionMode_e enum definition for details.
  */
 #define ZSTD_c_literalCompressionMode ZSTD_c_experimentalParam5
 
@@ -1797,12 +1883,52 @@ ZSTDLIB_API size_t ZSTD_CCtx_refPrefix_advanced(ZSTD_CCtx* cctx, const void* pre
  */
 #define ZSTD_c_validateSequences ZSTD_c_experimentalParam12
 
+/* ZSTD_c_splitBlocks
+ * Default is 0 == disabled. Set to 1 to enable block splitting.
+ *
+ * Will attempt to split blocks in order to improve compression ratio at the cost of speed.
+ */
+#define ZSTD_c_splitBlocks ZSTD_c_experimentalParam13
+
+/* ZSTD_c_useRowMatchFinder
+ * Default is ZSTD_urm_auto.
+ * Controlled with ZSTD_useRowMatchFinderMode_e enum.
+ *
+ * By default, in ZSTD_urm_auto, when finalizing the compression parameters, the library
+ * will decide at runtime whether to use the row-based matchfinder based on support for SIMD
+ * instructions as well as the windowLog.
+ *
+ * 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
+
+/* ZSTD_c_deterministicRefPrefix
+ * Default is 0 == disabled. Set to 1 to enable.
+ *
+ * Zstd produces different results for prefix compression when the prefix is
+ * directly adjacent to the data about to be compressed vs. when it isn't.
+ * This is because zstd detects that the two buffers are contiguous and it can
+ * use a more efficient match finding algorithm. However, this produces different
+ * results than when the two buffers are non-contiguous. This flag forces zstd
+ * to always load the prefix in non-contiguous mode, even if it happens to be
+ * adjacent to the data, to guarantee determinism.
+ *
+ * If you really care about determinism when using a dictionary or prefix,
+ * like when doing delta compression, you should select this option. It comes
+ * at a speed penalty of about ~2.5% if the dictionary and data happened to be
+ * contiguous, and is free if they weren't contiguous. We don't expect that
+ * intentionally making the dictionary and data contiguous will be worth the
+ * cost to memcpy() the data.
+ */
+#define ZSTD_c_deterministicRefPrefix ZSTD_c_experimentalParam15
+
 /*! ZSTD_CCtx_getParameter() :
  *  Get the requested compression parameter value, selected by enum ZSTD_cParameter,
  *  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(ZSTD_CCtx* cctx, ZSTD_cParameter param, int* value);
+ZSTDLIB_API size_t ZSTD_CCtx_getParameter(const ZSTD_CCtx* cctx, ZSTD_cParameter param, int* value);
 
 
 /*! ZSTD_CCtx_params :
@@ -1817,13 +1943,13 @@ ZSTDLIB_API size_t ZSTD_CCtx_getParameter(ZSTD_CCtx* cctx, ZSTD_cParameter param
  *                                    These parameters will be applied to
  *                                    all subsequent frames.
  *  - ZSTD_compressStream2() : Do compression using the CCtx.
- *  - ZSTD_freeCCtxParams() : Free the memory.
+ *  - ZSTD_freeCCtxParams() : Free the memory, accept NULL pointer.
  *
  *  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_API size_t ZSTD_freeCCtxParams(ZSTD_CCtx_params* params);
+ZSTDLIB_API size_t ZSTD_freeCCtxParams(ZSTD_CCtx_params* params);  /* accept NULL pointer */
 
 /*! ZSTD_CCtxParams_reset() :
  *  Reset params to default values.
@@ -1842,7 +1968,7 @@ ZSTDLIB_API size_t ZSTD_CCtxParams_init(ZSTD_CCtx_params* cctxParams, int compre
  */
 ZSTDLIB_API size_t ZSTD_CCtxParams_init_advanced(ZSTD_CCtx_params* cctxParams, ZSTD_parameters params);
 
-/*! ZSTD_CCtxParams_setParameter() :
+/*! ZSTD_CCtxParams_setParameter() : Requires v1.4.0+
  *  Similar to ZSTD_CCtx_setParameter.
  *  Set one compression parameter, selected by enum ZSTD_cParameter.
  *  Parameters must be applied to a ZSTD_CCtx using
@@ -1857,7 +1983,7 @@ ZSTDLIB_API size_t ZSTD_CCtxParams_setParameter(ZSTD_CCtx_params* params, ZSTD_c
  * 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(ZSTD_CCtx_params* params, ZSTD_cParameter param, int* value);
+ZSTDLIB_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.
@@ -1983,12 +2109,38 @@ ZSTDLIB_API size_t ZSTD_DCtx_getParameter(ZSTD_DCtx* dctx, ZSTD_dParameter param
  */
 #define ZSTD_d_forceIgnoreChecksum ZSTD_d_experimentalParam3
 
+/* ZSTD_d_refMultipleDDicts
+ * Experimental parameter.
+ * Default is 0 == disabled. Set to 1 to enable
+ *
+ * If enabled and dctx is allocated on the heap, then additional memory will be allocated
+ * to store references to multiple ZSTD_DDict. That is, multiple calls of ZSTD_refDDict()
+ * using a given ZSTD_DCtx, rather than overwriting the previous DDict reference, will instead
+ * store all references. At decompression time, the appropriate dictID is selected
+ * from the set of DDicts based on the dictID in the frame.
+ *
+ * Usage is simply calling ZSTD_refDDict() on multiple dict buffers.
+ *
+ * Param has values of byte ZSTD_refMultipleDDicts_e
+ *
+ * WARNING: Enabling this parameter and calling ZSTD_DCtx_refDDict(), will trigger memory
+ * allocation for the hash table. ZSTD_freeDCtx() also frees this memory.
+ * Memory is allocated as per ZSTD_DCtx::customMem.
+ *
+ * Although this function allocates memory for the table, the user is still responsible for
+ * memory management of the underlying ZSTD_DDict* themselves.
+ */
+#define ZSTD_d_refMultipleDDicts ZSTD_d_experimentalParam4
+
+
 /*! ZSTD_DCtx_setFormat() :
+ *  This function is REDUNDANT. Prefer ZSTD_DCtx_setParameter().
  *  Instruct the decoder context about what kind of data to decode next.
  *  This instruction is mandatory to decode data without a fully-formed header,
  *  such ZSTD_f_zstd1_magicless for example.
  * @return : 0, or an error code (which can be tested using ZSTD_isError()). */
-ZSTDLIB_API size_t ZSTD_DCtx_setFormat(ZSTD_DCtx* dctx, ZSTD_format_e format);
+ZSTD_DEPRECATED("use ZSTD_DCtx_setParameter() instead")
+size_t ZSTD_DCtx_setFormat(ZSTD_DCtx* dctx, ZSTD_format_e format);
 
 /*! ZSTD_decompressStream_simpleArgs() :
  *  Same as ZSTD_decompressStream(),
@@ -2012,7 +2164,7 @@ ZSTDLIB_API size_t ZSTD_decompressStream_simpleArgs (
 /*=====   Advanced Streaming compression functions  =====*/
 
 /*! ZSTD_initCStream_srcSize() :
- * This function is deprecated, and equivalent to:
+ * This function is DEPRECATED, and equivalent to:
  *     ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only);
  *     ZSTD_CCtx_refCDict(zcs, NULL); // clear the dictionary (if any)
  *     ZSTD_CCtx_setParameter(zcs, ZSTD_c_compressionLevel, compressionLevel);
@@ -2021,15 +2173,15 @@ ZSTDLIB_API size_t ZSTD_decompressStream_simpleArgs (
  * pledgedSrcSize must be correct. If it is not known at init time, use
  * ZSTD_CONTENTSIZE_UNKNOWN. Note that, for compatibility with older programs,
  * "0" also disables frame content size field. It may be enabled in the future.
- * Note : this prototype will be marked as deprecated and generate compilation warnings on reaching v1.5.x
+ * This prototype will generate compilation warnings.
  */
-ZSTDLIB_API size_t
-ZSTD_initCStream_srcSize(ZSTD_CStream* zcs,
+ZSTD_DEPRECATED("use ZSTD_CCtx_reset, see zstd.h for detailed instructions")
+size_t ZSTD_initCStream_srcSize(ZSTD_CStream* zcs,
                          int compressionLevel,
                          unsigned long long pledgedSrcSize);
 
 /*! ZSTD_initCStream_usingDict() :
- * This function is deprecated, and is equivalent to:
+ * This function is DEPRECATED, and is equivalent to:
  *     ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only);
  *     ZSTD_CCtx_setParameter(zcs, ZSTD_c_compressionLevel, compressionLevel);
  *     ZSTD_CCtx_loadDictionary(zcs, dict, dictSize);
@@ -2038,15 +2190,15 @@ ZSTD_initCStream_srcSize(ZSTD_CStream* zcs,
  * dict == NULL or dictSize < 8, in which case no dict is used.
  * Note: dict is loaded with ZSTD_dct_auto (treated as a full zstd dictionary if
  * it begins with ZSTD_MAGIC_DICTIONARY, else as raw content) and ZSTD_dlm_byCopy.
- * Note : this prototype will be marked as deprecated and generate compilation warnings on reaching v1.5.x
+ * This prototype will generate compilation warnings.
  */
-ZSTDLIB_API size_t
-ZSTD_initCStream_usingDict(ZSTD_CStream* zcs,
+ZSTD_DEPRECATED("use ZSTD_CCtx_reset, see zstd.h for detailed instructions")
+size_t ZSTD_initCStream_usingDict(ZSTD_CStream* zcs,
                      const void* dict, size_t dictSize,
                            int compressionLevel);
 
 /*! ZSTD_initCStream_advanced() :
- * This function is deprecated, and is approximately equivalent to:
+ * This function is DEPRECATED, and is approximately equivalent to:
  *     ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only);
  *     // Pseudocode: Set each zstd parameter and leave the rest as-is.
  *     for ((param, value) : params) {
@@ -2058,23 +2210,24 @@ ZSTD_initCStream_usingDict(ZSTD_CStream* zcs,
  * dict is loaded with ZSTD_dct_auto and ZSTD_dlm_byCopy.
  * pledgedSrcSize must be correct.
  * If srcSize is not known at init time, use value ZSTD_CONTENTSIZE_UNKNOWN.
- * Note : this prototype will be marked as deprecated and generate compilation warnings on reaching v1.5.x
+ * This prototype will generate compilation warnings.
  */
-ZSTDLIB_API size_t
-ZSTD_initCStream_advanced(ZSTD_CStream* zcs,
+ZSTD_DEPRECATED("use ZSTD_CCtx_reset, see zstd.h for detailed instructions")
+size_t ZSTD_initCStream_advanced(ZSTD_CStream* zcs,
                     const void* dict, size_t dictSize,
                           ZSTD_parameters params,
                           unsigned long long pledgedSrcSize);
 
 /*! ZSTD_initCStream_usingCDict() :
- * This function is deprecated, and equivalent to:
+ * This function is DEPRECATED, and equivalent to:
  *     ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only);
  *     ZSTD_CCtx_refCDict(zcs, cdict);
  *
  * note : cdict will just be referenced, and must outlive compression session
- * Note : this prototype will be marked as deprecated and generate compilation warnings on reaching v1.5.x
+ * This prototype will generate compilation warnings.
  */
-ZSTDLIB_API size_t ZSTD_initCStream_usingCDict(ZSTD_CStream* zcs, const ZSTD_CDict* cdict);
+ZSTD_DEPRECATED("use ZSTD_CCtx_reset and ZSTD_CCtx_refCDict, see zstd.h for detailed instructions")
+size_t ZSTD_initCStream_usingCDict(ZSTD_CStream* zcs, const ZSTD_CDict* cdict);
 
 /*! ZSTD_initCStream_usingCDict_advanced() :
  *   This function is DEPRECATED, and is approximately equivalent to:
@@ -2089,18 +2242,21 @@ ZSTDLIB_API size_t ZSTD_initCStream_usingCDict(ZSTD_CStream* zcs, const ZSTD_CDi
  * same as ZSTD_initCStream_usingCDict(), with control over frame parameters.
  * pledgedSrcSize must be correct. If srcSize is not known at init time, use
  * value ZSTD_CONTENTSIZE_UNKNOWN.
- * Note : this prototype will be marked as deprecated and generate compilation warnings on reaching v1.5.x
+ * This prototype will generate compilation warnings.
  */
-ZSTDLIB_API size_t
-ZSTD_initCStream_usingCDict_advanced(ZSTD_CStream* zcs,
+ZSTD_DEPRECATED("use ZSTD_CCtx_reset and ZSTD_CCtx_refCDict, see zstd.h for detailed instructions")
+size_t ZSTD_initCStream_usingCDict_advanced(ZSTD_CStream* zcs,
                                const ZSTD_CDict* cdict,
                                      ZSTD_frameParameters fParams,
                                      unsigned long long pledgedSrcSize);
 
 /*! ZSTD_resetCStream() :
- * This function is deprecated, and is equivalent to:
+ * This function is DEPRECATED, and is equivalent to:
  *     ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only);
  *     ZSTD_CCtx_setPledgedSrcSize(zcs, pledgedSrcSize);
+ * Note: ZSTD_resetCStream() interprets pledgedSrcSize == 0 as ZSTD_CONTENTSIZE_UNKNOWN, but
+ *       ZSTD_CCtx_setPledgedSrcSize() does not do the same, so ZSTD_CONTENTSIZE_UNKNOWN must be
+ *       explicitly specified.
  *
  *  start a new frame, using same parameters from previous frame.
  *  This is typically useful to skip dictionary loading stage, since it will re-use it in-place.
@@ -2110,9 +2266,10 @@ ZSTD_initCStream_usingCDict_advanced(ZSTD_CStream* zcs,
  *  For the time being, pledgedSrcSize==0 is interpreted as "srcSize unknown" for compatibility with older programs,
  *  but it will change to mean "empty" in future version, so use macro ZSTD_CONTENTSIZE_UNKNOWN instead.
  * @return : 0, or an error code (which can be tested using ZSTD_isError())
- *  Note : this prototype will be marked as deprecated and generate compilation warnings on reaching v1.5.x
+ *  This prototype will generate compilation warnings.
  */
-ZSTDLIB_API size_t ZSTD_resetCStream(ZSTD_CStream* zcs, unsigned long long pledgedSrcSize);
+ZSTD_DEPRECATED("use ZSTD_CCtx_reset, see zstd.h for detailed instructions")
+size_t ZSTD_resetCStream(ZSTD_CStream* zcs, unsigned long long pledgedSrcSize);
 
 
 typedef struct {
@@ -2199,8 +2356,7 @@ ZSTDLIB_API size_t ZSTD_resetDStream(ZSTD_DStream* zds);
   ZSTD_CCtx object can be re-used multiple times within successive compression operations.
 
   Start by initializing a context.
-  Use ZSTD_compressBegin(), or ZSTD_compressBegin_usingDict() for dictionary compression,
-  or ZSTD_compressBegin_advanced(), for finer parameter control.
+  Use ZSTD_compressBegin(), or ZSTD_compressBegin_usingDict() for dictionary compression.
   It's also possible to duplicate a reference context which has already been initialized, using ZSTD_copyCCtx()
 
   Then, consume your input using ZSTD_compressContinue().
@@ -2225,15 +2381,17 @@ 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_API size_t ZSTD_compressBegin_usingDict(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, int compressionLevel);
-ZSTDLIB_API size_t ZSTD_compressBegin_advanced(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, ZSTD_parameters params, unsigned long long pledgedSrcSize); /**< pledgedSrcSize : If srcSize is not known at init time, use ZSTD_CONTENTSIZE_UNKNOWN */
 ZSTDLIB_API size_t ZSTD_compressBegin_usingCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict); /**< note: fails if cdict==NULL */
-ZSTDLIB_API size_t ZSTD_compressBegin_usingCDict_advanced(ZSTD_CCtx* const cctx, const ZSTD_CDict* const cdict, ZSTD_frameParameters const fParams, unsigned long long const pledgedSrcSize);   /* compression parameters are already set within cdict. pledgedSrcSize must be correct. If srcSize is not known, use macro ZSTD_CONTENTSIZE_UNKNOWN */
 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_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);
 
-
+/* 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")
+size_t ZSTD_compressBegin_advanced(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, ZSTD_parameters params, unsigned long long pledgedSrcSize); /**< pledgedSrcSize : If srcSize is not known at init time, use ZSTD_CONTENTSIZE_UNKNOWN */
+ZSTD_DEPRECATED("use advanced API to access custom parameters")
+size_t ZSTD_compressBegin_usingCDict_advanced(ZSTD_CCtx* const cctx, const ZSTD_CDict* const cdict, ZSTD_frameParameters const fParams, unsigned long long const pledgedSrcSize);   /* compression parameters are already set within cdict. pledgedSrcSize must be correct. If srcSize is not known, use macro ZSTD_CONTENTSIZE_UNKNOWN */
 /**
   Buffer-less streaming decompression (synchronous mode)
 
diff --git a/zstd_common.c b/zstd_common.c
index 939e9f0..3d7e35b 100644
--- a/zstd_common.c
+++ b/zstd_common.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_compress.c b/zstd_compress.c
index dd6f68d..fc2d314 100644
--- a/zstd_compress.c
+++ b/zstd_compress.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -42,6 +42,18 @@
 #  define ZSTD_COMPRESS_HEAPMODE 0
 #endif
 
+/*!
+ * ZSTD_HASHLOG3_MAX :
+ * Maximum size of the hash table dedicated to find 3-bytes matches,
+ * in log format, aka 17 => 1 << 17 == 128Ki positions.
+ * This structure is only used in zstd_opt.
+ * Since allocation is centralized for all strategies, it has to be known here.
+ * The actual (selected) size of the hash table is then stored in ZSTD_matchState_t.hashLog3,
+ * so that zstd_opt.c doesn't need to know about this constant.
+ */
+#ifndef ZSTD_HASHLOG3_MAX
+#  define ZSTD_HASHLOG3_MAX 17
+#endif
 
 /*-*************************************
 *  Helper functions
@@ -72,6 +84,10 @@ struct ZSTD_CDict_s {
     ZSTD_customMem customMem;
     U32 dictID;
     int compressionLevel; /* 0 indicates that advanced API was used to select CDict params */
+    ZSTD_useRowMatchFinderMode_e useRowMatchFinder; /* Indicates whether the CDict was created with params that would use
+                                                     * row-based matchfinder. Unless the cdict is reloaded, we will use
+                                                     * the same greedy/lazy matchfinder at compression time.
+                                                     */
 };  /* typedef'd to ZSTD_CDict within "zstd.h" */
 
 ZSTD_CCtx* ZSTD_createCCtx(void)
@@ -202,6 +218,49 @@ size_t ZSTD_sizeof_CStream(const ZSTD_CStream* zcs)
 /* private API call, for dictBuilder only */
 const seqStore_t* ZSTD_getSeqStore(const ZSTD_CCtx* ctx) { return &(ctx->seqStore); }
 
+/* Returns true if the strategy supports using a row based matchfinder */
+static int ZSTD_rowMatchFinderSupported(const ZSTD_strategy strategy) {
+    return (strategy >= ZSTD_greedy && strategy <= ZSTD_lazy2);
+}
+
+/* Returns true if the strategy and useRowMatchFinder mode indicate that we will use the row based matchfinder
+ * for this compression.
+ */
+static int ZSTD_rowMatchFinderUsed(const ZSTD_strategy strategy, const ZSTD_useRowMatchFinderMode_e mode) {
+    assert(mode != ZSTD_urm_auto);
+    return ZSTD_rowMatchFinderSupported(strategy) && (mode == ZSTD_urm_enableRowMatchFinder);
+}
+
+/* Returns row matchfinder usage enum given an initial mode and cParams */
+static ZSTD_useRowMatchFinderMode_e ZSTD_resolveRowMatchFinderMode(ZSTD_useRowMatchFinderMode_e mode,
+                                                                   const ZSTD_compressionParameters* const cParams) {
+#if defined(ZSTD_ARCH_X86_SSE2) || defined(ZSTD_ARCH_ARM_NEON)
+    int const kHasSIMD128 = 1;
+#else
+    int const kHasSIMD128 = 0;
+#endif
+    if (mode != ZSTD_urm_auto) return mode; /* if requested enabled, but no SIMD, we still will use row matchfinder */
+    mode = ZSTD_urm_disableRowMatchFinder;
+    if (!ZSTD_rowMatchFinderSupported(cParams->strategy)) return mode;
+    if (kHasSIMD128) {
+        if (cParams->windowLog > 14) mode = ZSTD_urm_enableRowMatchFinder;
+    } else {
+        if (cParams->windowLog > 17) mode = ZSTD_urm_enableRowMatchFinder;
+    }
+    return mode;
+}
+
+/* Returns 1 if the arguments indicate that we should allocate a chainTable, 0 otherwise */
+static int ZSTD_allocateChainTable(const ZSTD_strategy strategy,
+                                   const ZSTD_useRowMatchFinderMode_e useRowMatchFinder,
+                                   const U32 forDDSDict) {
+    assert(useRowMatchFinder != ZSTD_urm_auto);
+    /* We always should allocate a chaintable if we are allocating a matchstate for a DDS dictionary matchstate.
+     * We do not allocate a chaintable if we are using ZSTD_fast, or are using the row-based matchfinder.
+     */
+    return forDDSDict || ((strategy != ZSTD_fast) && !ZSTD_rowMatchFinderUsed(strategy, useRowMatchFinder));
+}
+
 /* Returns 1 if compression parameters are such that we should
  * enable long distance matching (wlog >= 27, strategy >= btopt).
  * Returns 0 otherwise.
@@ -210,6 +269,14 @@ static U32 ZSTD_CParams_shouldEnableLdm(const ZSTD_compressionParameters* const
     return cParams->strategy >= ZSTD_btopt && cParams->windowLog >= 27;
 }
 
+/* Returns 1 if compression parameters are such that we should
+ * enable blockSplitter (wlog >= 17, strategy >= btopt).
+ * Returns 0 otherwise.
+ */
+static U32 ZSTD_CParams_useBlockSplitter(const ZSTD_compressionParameters* const cParams) {
+    return cParams->strategy >= ZSTD_btopt && cParams->windowLog >= 17;
+}
+
 static ZSTD_CCtx_params ZSTD_makeCCtxParamsFromCParams(
         ZSTD_compressionParameters cParams)
 {
@@ -218,6 +285,7 @@ static ZSTD_CCtx_params ZSTD_makeCCtxParamsFromCParams(
     ZSTD_CCtxParams_init(&cctxParams, ZSTD_CLEVEL_DEFAULT);
     cctxParams.cParams = cParams;
 
+    /* Adjust advanced params according to cParams */
     if (ZSTD_CParams_shouldEnableLdm(&cParams)) {
         DEBUGLOG(4, "ZSTD_makeCCtxParamsFromCParams(): Including LDM into cctx params");
         cctxParams.ldmParams.enableLdm = 1;
@@ -227,6 +295,12 @@ static ZSTD_CCtx_params ZSTD_makeCCtxParamsFromCParams(
         assert(cctxParams.ldmParams.hashRateLog < 32);
     }
 
+    if (ZSTD_CParams_useBlockSplitter(&cParams)) {
+        DEBUGLOG(4, "ZSTD_makeCCtxParamsFromCParams(): Including block splitting into cctx params");
+        cctxParams.splitBlocks = 1;
+    }
+
+    cctxParams.useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(cctxParams.useRowMatchFinder, &cParams);
     assert(!ZSTD_checkCParams(cParams));
     return cctxParams;
 }
@@ -269,29 +343,59 @@ size_t ZSTD_CCtxParams_init(ZSTD_CCtx_params* cctxParams, int compressionLevel)
     return 0;
 }
 
+#define ZSTD_NO_CLEVEL 0
+
+/**
+ * Initializes the cctxParams from params and compressionLevel.
+ * @param compressionLevel If params are derived from a compression level then that compression level, otherwise ZSTD_NO_CLEVEL.
+ */
+static void ZSTD_CCtxParams_init_internal(ZSTD_CCtx_params* cctxParams, ZSTD_parameters const* params, int compressionLevel)
+{
+    assert(!ZSTD_checkCParams(params->cParams));
+    ZSTD_memset(cctxParams, 0, sizeof(*cctxParams));
+    cctxParams->cParams = params->cParams;
+    cctxParams->fParams = params->fParams;
+    /* Should not matter, as all cParams are presumed properly defined.
+     * But, set it for tracing anyway.
+     */
+    cctxParams->compressionLevel = compressionLevel;
+    cctxParams->useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(cctxParams->useRowMatchFinder, &params->cParams);
+    DEBUGLOG(4, "ZSTD_CCtxParams_init_internal: useRowMatchFinder=%d", cctxParams->useRowMatchFinder);
+
+    if (ZSTD_CParams_shouldEnableLdm(&params->cParams)) {
+        /* Enable LDM by default for optimal parser and window size >= 128MB */
+        DEBUGLOG(4, "LDM enabled by default (window size >= 128MB, strategy >= btopt)");
+        cctxParams->ldmParams.enableLdm = 1;
+    }
+
+    if (ZSTD_CParams_useBlockSplitter(&params->cParams)) {
+        DEBUGLOG(4, "Block splitter enabled by default (window size >= 128K, strategy >= btopt)");
+        cctxParams->splitBlocks = 1;
+    }
+}
+
 size_t ZSTD_CCtxParams_init_advanced(ZSTD_CCtx_params* cctxParams, ZSTD_parameters params)
 {
     RETURN_ERROR_IF(!cctxParams, GENERIC, "NULL pointer!");
     FORWARD_IF_ERROR( ZSTD_checkCParams(params.cParams) , "");
-    ZSTD_memset(cctxParams, 0, sizeof(*cctxParams));
-    assert(!ZSTD_checkCParams(params.cParams));
-    cctxParams->cParams = params.cParams;
-    cctxParams->fParams = params.fParams;
-    cctxParams->compressionLevel = ZSTD_CLEVEL_DEFAULT;   /* should not matter, as all cParams are presumed properly defined */
+    ZSTD_CCtxParams_init_internal(cctxParams, &params, ZSTD_NO_CLEVEL);
     return 0;
 }
 
-/* ZSTD_assignParamsToCCtxParams() :
- * params is presumed valid at this stage */
-static ZSTD_CCtx_params ZSTD_assignParamsToCCtxParams(
-        const ZSTD_CCtx_params* cctxParams, const ZSTD_parameters* params)
+/**
+ * Sets cctxParams' cParams and fParams from params, but otherwise leaves them alone.
+ * @param param Validated zstd parameters.
+ */
+static void ZSTD_CCtxParams_setZstdParams(
+        ZSTD_CCtx_params* cctxParams, const ZSTD_parameters* params)
 {
-    ZSTD_CCtx_params ret = *cctxParams;
     assert(!ZSTD_checkCParams(params->cParams));
-    ret.cParams = params->cParams;
-    ret.fParams = params->fParams;
-    ret.compressionLevel = ZSTD_CLEVEL_DEFAULT;   /* should not matter, as all cParams are presumed properly defined */
-    return ret;
+    cctxParams->cParams = params->cParams;
+    cctxParams->fParams = params->fParams;
+    /* Should not matter, as all cParams are presumed properly defined.
+     * But, set it for tracing anyway.
+     */
+    cctxParams->compressionLevel = ZSTD_NO_CLEVEL;
 }
 
 ZSTD_bounds ZSTD_cParam_getBounds(ZSTD_cParameter param)
@@ -468,6 +572,21 @@ ZSTD_bounds ZSTD_cParam_getBounds(ZSTD_cParameter param)
         bounds.upperBound = 1;
         return bounds;
 
+    case ZSTD_c_splitBlocks:
+        bounds.lowerBound = 0;
+        bounds.upperBound = 1;
+        return bounds;
+
+    case ZSTD_c_useRowMatchFinder:
+        bounds.lowerBound = (int)ZSTD_urm_auto;
+        bounds.upperBound = (int)ZSTD_urm_enableRowMatchFinder;
+        return bounds;
+
+    case ZSTD_c_deterministicRefPrefix:
+        bounds.lowerBound = 0;
+        bounds.upperBound = 1;
+        return bounds;
+
     default:
         bounds.error = ERROR(parameter_unsupported);
         return bounds;
@@ -529,6 +648,9 @@ static int ZSTD_isUpdateAuthorized(ZSTD_cParameter param)
     case ZSTD_c_stableOutBuffer:
     case ZSTD_c_blockDelimiters:
     case ZSTD_c_validateSequences:
+    case ZSTD_c_splitBlocks:
+    case ZSTD_c_useRowMatchFinder:
+    case ZSTD_c_deterministicRefPrefix:
     default:
         return 0;
     }
@@ -581,6 +703,9 @@ size_t ZSTD_CCtx_setParameter(ZSTD_CCtx* cctx, ZSTD_cParameter param, int value)
     case ZSTD_c_stableOutBuffer:
     case ZSTD_c_blockDelimiters:
     case ZSTD_c_validateSequences:
+    case ZSTD_c_splitBlocks:
+    case ZSTD_c_useRowMatchFinder:
+    case ZSTD_c_deterministicRefPrefix:
         break;
 
     default: RETURN_ERROR(parameter_unsupported, "unknown parameter");
@@ -792,17 +917,32 @@ size_t ZSTD_CCtxParams_setParameter(ZSTD_CCtx_params* CCtxParams,
         CCtxParams->validateSequences = value;
         return CCtxParams->validateSequences;
 
+    case ZSTD_c_splitBlocks:
+        BOUNDCHECK(ZSTD_c_splitBlocks, value);
+        CCtxParams->splitBlocks = value;
+        return CCtxParams->splitBlocks;
+
+    case ZSTD_c_useRowMatchFinder:
+        BOUNDCHECK(ZSTD_c_useRowMatchFinder, value);
+        CCtxParams->useRowMatchFinder = (ZSTD_useRowMatchFinderMode_e)value;
+        return CCtxParams->useRowMatchFinder;
+
+    case ZSTD_c_deterministicRefPrefix:
+        BOUNDCHECK(ZSTD_c_deterministicRefPrefix, value);
+        CCtxParams->deterministicRefPrefix = !!value;
+        return CCtxParams->deterministicRefPrefix;
+
     default: RETURN_ERROR(parameter_unsupported, "unknown parameter");
     }
 }
 
-size_t ZSTD_CCtx_getParameter(ZSTD_CCtx* cctx, ZSTD_cParameter param, int* value)
+size_t ZSTD_CCtx_getParameter(ZSTD_CCtx const* cctx, ZSTD_cParameter param, int* value)
 {
     return ZSTD_CCtxParams_getParameter(&cctx->requestedParams, param, value);
 }
 
 size_t ZSTD_CCtxParams_getParameter(
-        ZSTD_CCtx_params* CCtxParams, ZSTD_cParameter param, int* value)
+        ZSTD_CCtx_params const* CCtxParams, ZSTD_cParameter param, int* value)
 {
     switch(param)
     {
@@ -915,6 +1055,15 @@ size_t ZSTD_CCtxParams_getParameter(
     case ZSTD_c_validateSequences :
         *value = (int)CCtxParams->validateSequences;
         break;
+    case ZSTD_c_splitBlocks :
+        *value = (int)CCtxParams->splitBlocks;
+        break;
+    case ZSTD_c_useRowMatchFinder :
+        *value = (int)CCtxParams->useRowMatchFinder;
+        break;
+    case ZSTD_c_deterministicRefPrefix:
+        *value = (int)CCtxParams->deterministicRefPrefix;
+        break;
     default: RETURN_ERROR(parameter_unsupported, "unknown parameter");
     }
     return 0;
@@ -1188,15 +1337,26 @@ ZSTD_adjustCParams_internal(ZSTD_compressionParameters cPar,
     const U64 maxWindowResize = 1ULL << (ZSTD_WINDOWLOG_MAX-1);
     assert(ZSTD_checkCParams(cPar)==0);
 
-    if (dictSize && srcSize == ZSTD_CONTENTSIZE_UNKNOWN)
-        srcSize = minSrcSize;
-
     switch (mode) {
-    case ZSTD_cpm_noAttachDict:
     case ZSTD_cpm_unknown:
+    case ZSTD_cpm_noAttachDict:
+        /* If we don't know the source size, don't make any
+         * assumptions about it. We will already have selected
+         * smaller parameters if a dictionary is in use.
+         */
+        break;
     case ZSTD_cpm_createCDict:
+        /* Assume a small source size when creating a dictionary
+         * with an unkown source size.
+         */
+        if (dictSize && srcSize == ZSTD_CONTENTSIZE_UNKNOWN)
+            srcSize = minSrcSize;
         break;
     case ZSTD_cpm_attachDict:
+        /* Dictionary has its own dedicated parameters which have
+         * already been selected. We are selecting parameters
+         * for only the source.
+         */
         dictSize = 0;
         break;
     default:
@@ -1213,7 +1373,8 @@ ZSTD_adjustCParams_internal(ZSTD_compressionParameters cPar,
                             ZSTD_highbit32(tSize-1) + 1;
         if (cPar.windowLog > srcLog) cPar.windowLog = srcLog;
     }
-    {   U32 const dictAndWindowLog = ZSTD_dictAndWindowLog(cPar.windowLog, (U64)srcSize, (U64)dictSize);
+    if (srcSize != ZSTD_CONTENTSIZE_UNKNOWN) {
+        U32 const dictAndWindowLog = ZSTD_dictAndWindowLog(cPar.windowLog, (U64)srcSize, (U64)dictSize);
         U32 const cycleLog = ZSTD_cycleLog(cPar.chainLog, cPar.strategy);
         if (cPar.hashLog > dictAndWindowLog+1) cPar.hashLog = dictAndWindowLog+1;
         if (cycleLog > dictAndWindowLog)
@@ -1269,9 +1430,14 @@ ZSTD_compressionParameters ZSTD_getCParamsFromCCtxParams(
 
 static size_t
 ZSTD_sizeof_matchState(const ZSTD_compressionParameters* const cParams,
+                       const ZSTD_useRowMatchFinderMode_e useRowMatchFinder,
+                       const U32 enableDedicatedDictSearch,
                        const U32 forCCtx)
 {
-    size_t const chainSize = (cParams->strategy == ZSTD_fast) ? 0 : ((size_t)1 << cParams->chainLog);
+    /* chain table size should be 0 for fast or row-hash strategies */
+    size_t const chainSize = ZSTD_allocateChainTable(cParams->strategy, useRowMatchFinder, enableDedicatedDictSearch && !forCCtx)
+                                ? ((size_t)1 << cParams->chainLog)
+                                : 0;
     size_t const hSize = ((size_t)1) << cParams->hashLog;
     U32    const hashLog3 = (forCCtx && cParams->minMatch==3) ? MIN(ZSTD_HASHLOG3_MAX, cParams->windowLog) : 0;
     size_t const h3Size = hashLog3 ? ((size_t)1) << hashLog3 : 0;
@@ -1281,24 +1447,34 @@ ZSTD_sizeof_matchState(const ZSTD_compressionParameters* const cParams,
                             + hSize * sizeof(U32)
                             + h3Size * sizeof(U32);
     size_t const optPotentialSpace =
-        ZSTD_cwksp_alloc_size((MaxML+1) * sizeof(U32))
-      + ZSTD_cwksp_alloc_size((MaxLL+1) * sizeof(U32))
-      + ZSTD_cwksp_alloc_size((MaxOff+1) * sizeof(U32))
-      + ZSTD_cwksp_alloc_size((1<<Litbits) * sizeof(U32))
-      + ZSTD_cwksp_alloc_size((ZSTD_OPT_NUM+1) * sizeof(ZSTD_match_t))
-      + ZSTD_cwksp_alloc_size((ZSTD_OPT_NUM+1) * sizeof(ZSTD_optimal_t));
+        ZSTD_cwksp_aligned_alloc_size((MaxML+1) * sizeof(U32))
+      + ZSTD_cwksp_aligned_alloc_size((MaxLL+1) * sizeof(U32))
+      + ZSTD_cwksp_aligned_alloc_size((MaxOff+1) * sizeof(U32))
+      + ZSTD_cwksp_aligned_alloc_size((1<<Litbits) * sizeof(U32))
+      + ZSTD_cwksp_aligned_alloc_size((ZSTD_OPT_NUM+1) * sizeof(ZSTD_match_t))
+      + ZSTD_cwksp_aligned_alloc_size((ZSTD_OPT_NUM+1) * sizeof(ZSTD_optimal_t));
+    size_t const lazyAdditionalSpace = ZSTD_rowMatchFinderUsed(cParams->strategy, useRowMatchFinder)
+                                            ? ZSTD_cwksp_aligned_alloc_size(hSize*sizeof(U16))
+                                            : 0;
     size_t const optSpace = (forCCtx && (cParams->strategy >= ZSTD_btopt))
                                 ? optPotentialSpace
                                 : 0;
+    size_t const slackSpace = ZSTD_cwksp_slack_space_required();
+
+    /* tables are guaranteed to be sized in multiples of 64 bytes (or 16 uint32_t) */
+    ZSTD_STATIC_ASSERT(ZSTD_HASHLOG_MIN >= 4 && ZSTD_WINDOWLOG_MIN >= 4 && ZSTD_CHAINLOG_MIN >= 4);
+    assert(useRowMatchFinder != ZSTD_urm_auto);
+
     DEBUGLOG(4, "chainSize: %u - hSize: %u - h3Size: %u",
                 (U32)chainSize, (U32)hSize, (U32)h3Size);
-    return tableSpace + optSpace;
+    return tableSpace + optSpace + slackSpace + lazyAdditionalSpace;
 }
 
 static size_t ZSTD_estimateCCtxSize_usingCCtxParams_internal(
         const ZSTD_compressionParameters* cParams,
         const ldmParams_t* ldmParams,
         const int isStatic,
+        const ZSTD_useRowMatchFinderMode_e useRowMatchFinder,
         const size_t buffInSize,
         const size_t buffOutSize,
         const U64 pledgedSrcSize)
@@ -1308,16 +1484,16 @@ static size_t ZSTD_estimateCCtxSize_usingCCtxParams_internal(
     U32    const divider = (cParams->minMatch==3) ? 3 : 4;
     size_t const maxNbSeq = blockSize / divider;
     size_t const tokenSpace = ZSTD_cwksp_alloc_size(WILDCOPY_OVERLENGTH + blockSize)
-                            + ZSTD_cwksp_alloc_size(maxNbSeq * sizeof(seqDef))
+                            + ZSTD_cwksp_aligned_alloc_size(maxNbSeq * sizeof(seqDef))
                             + 3 * ZSTD_cwksp_alloc_size(maxNbSeq * sizeof(BYTE));
     size_t const entropySpace = ZSTD_cwksp_alloc_size(ENTROPY_WORKSPACE_SIZE);
     size_t const blockStateSpace = 2 * ZSTD_cwksp_alloc_size(sizeof(ZSTD_compressedBlockState_t));
-    size_t const matchStateSize = ZSTD_sizeof_matchState(cParams, /* forCCtx */ 1);
+    size_t const matchStateSize = ZSTD_sizeof_matchState(cParams, useRowMatchFinder, /* enableDedicatedDictSearch */ 0, /* forCCtx */ 1);
 
     size_t const ldmSpace = ZSTD_ldm_getTableSize(*ldmParams);
     size_t const maxNbLdmSeq = ZSTD_ldm_getMaxNbSeq(*ldmParams, blockSize);
     size_t const ldmSeqSpace = ldmParams->enableLdm ?
-        ZSTD_cwksp_alloc_size(maxNbLdmSeq * sizeof(rawSeq)) : 0;
+        ZSTD_cwksp_aligned_alloc_size(maxNbLdmSeq * sizeof(rawSeq)) : 0;
 
 
     size_t const bufferSpace = ZSTD_cwksp_alloc_size(buffInSize)
@@ -1343,25 +1519,45 @@ size_t ZSTD_estimateCCtxSize_usingCCtxParams(const ZSTD_CCtx_params* params)
 {
     ZSTD_compressionParameters const cParams =
                 ZSTD_getCParamsFromCCtxParams(params, ZSTD_CONTENTSIZE_UNKNOWN, 0, ZSTD_cpm_noAttachDict);
+    ZSTD_useRowMatchFinderMode_e const useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(params->useRowMatchFinder,
+                                                                                         &cParams);
 
     RETURN_ERROR_IF(params->nbWorkers > 0, GENERIC, "Estimate CCtx size is supported for single-threaded compression only.");
     /* estimateCCtxSize is for one-shot compression. So no buffers should
      * be needed. However, we still allocate two 0-sized buffers, which can
      * take space under ASAN. */
     return ZSTD_estimateCCtxSize_usingCCtxParams_internal(
-        &cParams, &params->ldmParams, 1, 0, 0, ZSTD_CONTENTSIZE_UNKNOWN);
+        &cParams, &params->ldmParams, 1, useRowMatchFinder, 0, 0, ZSTD_CONTENTSIZE_UNKNOWN);
 }
 
 size_t ZSTD_estimateCCtxSize_usingCParams(ZSTD_compressionParameters cParams)
 {
-    ZSTD_CCtx_params const params = ZSTD_makeCCtxParamsFromCParams(cParams);
-    return ZSTD_estimateCCtxSize_usingCCtxParams(&params);
+    ZSTD_CCtx_params initialParams = ZSTD_makeCCtxParamsFromCParams(cParams);
+    if (ZSTD_rowMatchFinderSupported(cParams.strategy)) {
+        /* Pick bigger of not using and using row-based matchfinder for greedy and lazy strategies */
+        size_t noRowCCtxSize;
+        size_t rowCCtxSize;
+        initialParams.useRowMatchFinder = ZSTD_urm_disableRowMatchFinder;
+        noRowCCtxSize = ZSTD_estimateCCtxSize_usingCCtxParams(&initialParams);
+        initialParams.useRowMatchFinder = ZSTD_urm_enableRowMatchFinder;
+        rowCCtxSize = ZSTD_estimateCCtxSize_usingCCtxParams(&initialParams);
+        return MAX(noRowCCtxSize, rowCCtxSize);
+    } else {
+        return ZSTD_estimateCCtxSize_usingCCtxParams(&initialParams);
+    }
 }
 
 static size_t ZSTD_estimateCCtxSize_internal(int compressionLevel)
 {
-    ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, 0, ZSTD_cpm_noAttachDict);
-    return ZSTD_estimateCCtxSize_usingCParams(cParams);
+    int tier = 0;
+    size_t largestSize = 0;
+    static const unsigned long long srcSizeTiers[4] = {16 KB, 128 KB, 256 KB, ZSTD_CONTENTSIZE_UNKNOWN};
+    for (; tier < 4; ++tier) {
+        /* Choose the set of cParams for a given level across all srcSizes that give the largest cctxSize */
+        ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, srcSizeTiers[tier], 0, ZSTD_cpm_noAttachDict);
+        largestSize = MAX(ZSTD_estimateCCtxSize_usingCParams(cParams), largestSize);
+    }
+    return largestSize;
 }
 
 size_t ZSTD_estimateCCtxSize(int compressionLevel)
@@ -1369,6 +1565,7 @@ size_t ZSTD_estimateCCtxSize(int compressionLevel)
     int level;
     size_t memBudget = 0;
     for (level=MIN(compressionLevel, 1); level<=compressionLevel; level++) {
+        /* Ensure monotonically increasing memory usage as compression level increases */
         size_t const newMB = ZSTD_estimateCCtxSize_internal(level);
         if (newMB > memBudget) memBudget = newMB;
     }
@@ -1387,17 +1584,29 @@ size_t ZSTD_estimateCStreamSize_usingCCtxParams(const ZSTD_CCtx_params* params)
         size_t const outBuffSize = (params->outBufferMode == ZSTD_bm_buffered)
                 ? ZSTD_compressBound(blockSize) + 1
                 : 0;
+        ZSTD_useRowMatchFinderMode_e const useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(params->useRowMatchFinder, &params->cParams);
 
         return ZSTD_estimateCCtxSize_usingCCtxParams_internal(
-            &cParams, &params->ldmParams, 1, inBuffSize, outBuffSize,
+            &cParams, &params->ldmParams, 1, useRowMatchFinder, inBuffSize, outBuffSize,
             ZSTD_CONTENTSIZE_UNKNOWN);
     }
 }
 
 size_t ZSTD_estimateCStreamSize_usingCParams(ZSTD_compressionParameters cParams)
 {
-    ZSTD_CCtx_params const params = ZSTD_makeCCtxParamsFromCParams(cParams);
-    return ZSTD_estimateCStreamSize_usingCCtxParams(&params);
+    ZSTD_CCtx_params initialParams = ZSTD_makeCCtxParamsFromCParams(cParams);
+    if (ZSTD_rowMatchFinderSupported(cParams.strategy)) {
+        /* Pick bigger of not using and using row-based matchfinder for greedy and lazy strategies */
+        size_t noRowCCtxSize;
+        size_t rowCCtxSize;
+        initialParams.useRowMatchFinder = ZSTD_urm_disableRowMatchFinder;
+        noRowCCtxSize = ZSTD_estimateCStreamSize_usingCCtxParams(&initialParams);
+        initialParams.useRowMatchFinder = ZSTD_urm_enableRowMatchFinder;
+        rowCCtxSize = ZSTD_estimateCStreamSize_usingCCtxParams(&initialParams);
+        return MAX(noRowCCtxSize, rowCCtxSize);
+    } else {
+        return ZSTD_estimateCStreamSize_usingCCtxParams(&initialParams);
+    }
 }
 
 static size_t ZSTD_estimateCStreamSize_internal(int compressionLevel)
@@ -1522,20 +1731,27 @@ typedef enum {
     ZSTD_resetTarget_CCtx
 } ZSTD_resetTarget_e;
 
+
 static size_t
 ZSTD_reset_matchState(ZSTD_matchState_t* ms,
                       ZSTD_cwksp* ws,
                 const ZSTD_compressionParameters* cParams,
+                const ZSTD_useRowMatchFinderMode_e useRowMatchFinder,
                 const ZSTD_compResetPolicy_e crp,
                 const ZSTD_indexResetPolicy_e forceResetIndex,
                 const ZSTD_resetTarget_e forWho)
 {
-    size_t const chainSize = (cParams->strategy == ZSTD_fast) ? 0 : ((size_t)1 << cParams->chainLog);
+    /* disable chain table allocation for fast or row-based strategies */
+    size_t const chainSize = ZSTD_allocateChainTable(cParams->strategy, useRowMatchFinder,
+                                                     ms->dedicatedDictSearch && (forWho == ZSTD_resetTarget_CDict))
+                                ? ((size_t)1 << cParams->chainLog)
+                                : 0;
     size_t const hSize = ((size_t)1) << cParams->hashLog;
     U32    const hashLog3 = ((forWho == ZSTD_resetTarget_CCtx) && cParams->minMatch==3) ? MIN(ZSTD_HASHLOG3_MAX, cParams->windowLog) : 0;
     size_t const h3Size = hashLog3 ? ((size_t)1) << hashLog3 : 0;
 
     DEBUGLOG(4, "reset indices : %u", forceResetIndex == ZSTDirp_reset);
+    assert(useRowMatchFinder != ZSTD_urm_auto);
     if (forceResetIndex == ZSTDirp_reset) {
         ZSTD_window_init(&ms->window);
         ZSTD_cwksp_mark_tables_dirty(ws);
@@ -1574,11 +1790,23 @@ ZSTD_reset_matchState(ZSTD_matchState_t* ms,
         ms->opt.priceTable = (ZSTD_optimal_t*)ZSTD_cwksp_reserve_aligned(ws, (ZSTD_OPT_NUM+1) * sizeof(ZSTD_optimal_t));
     }
 
+    if (ZSTD_rowMatchFinderUsed(cParams->strategy, useRowMatchFinder)) {
+        {   /* Row match finder needs an additional table of hashes ("tags") */
+            size_t const tagTableSize = hSize*sizeof(U16);
+            ms->tagTable = (U16*)ZSTD_cwksp_reserve_aligned(ws, tagTableSize);
+            if (ms->tagTable) ZSTD_memset(ms->tagTable, 0, tagTableSize);
+        }
+        {   /* Switch to 32-entry rows if searchLog is 5 (or more) */
+            U32 const rowLog = MAX(MIN(cParams->searchLog, 6), 4);
+            assert(cParams->hashLog >= rowLog);
+            ms->rowHashLog = cParams->hashLog - rowLog;
+        }
+    }
+
     ms->cParams = *cParams;
 
     RETURN_ERROR_IF(ZSTD_cwksp_reserve_failed(ws), memory_allocation,
                     "failed a workspace allocation in ZSTD_reset_matchState");
-
     return 0;
 }
 
@@ -1595,62 +1823,85 @@ static int ZSTD_indexTooCloseToMax(ZSTD_window_t w)
     return (size_t)(w.nextSrc - w.base) > (ZSTD_CURRENT_MAX - ZSTD_INDEXOVERFLOW_MARGIN);
 }
 
+/** ZSTD_dictTooBig():
+ * When dictionaries are larger than ZSTD_CHUNKSIZE_MAX they can't be loaded in
+ * one go generically. So we ensure that in that case we reset the tables to zero,
+ * so that we can load as much of the dictionary as possible.
+ */
+static int ZSTD_dictTooBig(size_t const loadedDictSize)
+{
+    return loadedDictSize > ZSTD_CHUNKSIZE_MAX;
+}
+
 /*! ZSTD_resetCCtx_internal() :
-    note : `params` are assumed fully validated at this stage */
+ * @param loadedDictSize The size of the dictionary to be loaded
+ * into the context, if any. If no dictionary is used, or the
+ * dictionary is being attached / copied, then pass 0.
+ * note : `params` are assumed fully validated at this stage.
+ */
 static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc,
-                                      ZSTD_CCtx_params params,
+                                      ZSTD_CCtx_params const* params,
                                       U64 const pledgedSrcSize,
+                                      size_t const loadedDictSize,
                                       ZSTD_compResetPolicy_e const crp,
                                       ZSTD_buffered_policy_e const zbuff)
 {
     ZSTD_cwksp* const ws = &zc->workspace;
-    DEBUGLOG(4, "ZSTD_resetCCtx_internal: pledgedSrcSize=%u, wlog=%u",
-                (U32)pledgedSrcSize, params.cParams.windowLog);
-    assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams)));
+    DEBUGLOG(4, "ZSTD_resetCCtx_internal: pledgedSrcSize=%u, wlog=%u, useRowMatchFinder=%d",
+                (U32)pledgedSrcSize, params->cParams.windowLog, (int)params->useRowMatchFinder);
+    assert(!ZSTD_isError(ZSTD_checkCParams(params->cParams)));
 
     zc->isFirstBlock = 1;
 
-    if (params.ldmParams.enableLdm) {
+    /* Set applied params early so we can modify them for LDM,
+     * and point params at the applied params.
+     */
+    zc->appliedParams = *params;
+    params = &zc->appliedParams;
+
+    assert(params->useRowMatchFinder != ZSTD_urm_auto);
+    if (params->ldmParams.enableLdm) {
         /* Adjust long distance matching parameters */
-        ZSTD_ldm_adjustParameters(&params.ldmParams, &params.cParams);
-        assert(params.ldmParams.hashLog >= params.ldmParams.bucketSizeLog);
-        assert(params.ldmParams.hashRateLog < 32);
-        zc->ldmState.hashPower = ZSTD_rollingHash_primePower(params.ldmParams.minMatchLength);
+        ZSTD_ldm_adjustParameters(&zc->appliedParams.ldmParams, &params->cParams);
+        assert(params->ldmParams.hashLog >= params->ldmParams.bucketSizeLog);
+        assert(params->ldmParams.hashRateLog < 32);
     }
 
-    {   size_t const windowSize = MAX(1, (size_t)MIN(((U64)1 << params.cParams.windowLog), pledgedSrcSize));
+    {   size_t const windowSize = MAX(1, (size_t)MIN(((U64)1 << params->cParams.windowLog), pledgedSrcSize));
         size_t const blockSize = MIN(ZSTD_BLOCKSIZE_MAX, windowSize);
-        U32    const divider = (params.cParams.minMatch==3) ? 3 : 4;
+        U32    const divider = (params->cParams.minMatch==3) ? 3 : 4;
         size_t const maxNbSeq = blockSize / divider;
-        size_t const buffOutSize = (zbuff == ZSTDb_buffered && params.outBufferMode == ZSTD_bm_buffered)
+        size_t const buffOutSize = (zbuff == ZSTDb_buffered && params->outBufferMode == ZSTD_bm_buffered)
                 ? ZSTD_compressBound(blockSize) + 1
                 : 0;
-        size_t const buffInSize = (zbuff == ZSTDb_buffered && params.inBufferMode == ZSTD_bm_buffered)
+        size_t const buffInSize = (zbuff == ZSTDb_buffered && params->inBufferMode == ZSTD_bm_buffered)
                 ? windowSize + blockSize
                 : 0;
-        size_t const maxNbLdmSeq = ZSTD_ldm_getMaxNbSeq(params.ldmParams, blockSize);
+        size_t const maxNbLdmSeq = ZSTD_ldm_getMaxNbSeq(params->ldmParams, blockSize);
 
         int const indexTooClose = ZSTD_indexTooCloseToMax(zc->blockState.matchState.window);
+        int const dictTooBig = ZSTD_dictTooBig(loadedDictSize);
         ZSTD_indexResetPolicy_e needsIndexReset =
-            (!indexTooClose && zc->initialized) ? ZSTDirp_continue : ZSTDirp_reset;
+            (indexTooClose || dictTooBig || !zc->initialized) ? ZSTDirp_reset : ZSTDirp_continue;
 
         size_t const neededSpace =
             ZSTD_estimateCCtxSize_usingCCtxParams_internal(
-                &params.cParams, &params.ldmParams, zc->staticSize != 0,
+                &params->cParams, &params->ldmParams, zc->staticSize != 0, params->useRowMatchFinder,
                 buffInSize, buffOutSize, pledgedSrcSize);
+        int resizeWorkspace;
+
         FORWARD_IF_ERROR(neededSpace, "cctx size estimate failed!");
 
         if (!zc->staticSize) ZSTD_cwksp_bump_oversized_duration(ws, 0);
 
-        /* Check if workspace is large enough, alloc a new one if needed */
-        {
+        {   /* Check if workspace is large enough, alloc a new one if needed */
             int const workspaceTooSmall = ZSTD_cwksp_sizeof(ws) < neededSpace;
             int const workspaceWasteful = ZSTD_cwksp_check_wasteful(ws, neededSpace);
-
+            resizeWorkspace = workspaceTooSmall || workspaceWasteful;
             DEBUGLOG(4, "Need %zu B workspace", neededSpace);
             DEBUGLOG(4, "windowSize: %zu - blockSize: %zu", windowSize, blockSize);
 
-            if (workspaceTooSmall || workspaceWasteful) {
+            if (resizeWorkspace) {
                 DEBUGLOG(4, "Resize workspaceSize from %zuKB to %zuKB",
                             ZSTD_cwksp_sizeof(ws) >> 10,
                             neededSpace >> 10);
@@ -1678,8 +1929,7 @@ static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc,
         ZSTD_cwksp_clear(ws);
 
         /* init params */
-        zc->appliedParams = params;
-        zc->blockState.matchState.cParams = params.cParams;
+        zc->blockState.matchState.cParams = params->cParams;
         zc->pledgedSrcSizePlusOne = pledgedSrcSize+1;
         zc->consumedSrcSize = 0;
         zc->producedCSize = 0;
@@ -1692,6 +1942,7 @@ static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc,
         XXH64_reset(&zc->xxhState, 0);
         zc->stage = ZSTDcs_init;
         zc->dictID = 0;
+        zc->dictContentSize = 0;
 
         ZSTD_reset_compressedBlockState(zc->blockState.prevCBlock);
 
@@ -1709,13 +1960,13 @@ static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc,
         zc->outBuff = (char*)ZSTD_cwksp_reserve_buffer(ws, buffOutSize);
 
         /* ldm bucketOffsets table */
-        if (params.ldmParams.enableLdm) {
+        if (params->ldmParams.enableLdm) {
             /* TODO: avoid memset? */
-            size_t const ldmBucketSize =
-                  ((size_t)1) << (params.ldmParams.hashLog -
-                                  params.ldmParams.bucketSizeLog);
-            zc->ldmState.bucketOffsets = ZSTD_cwksp_reserve_buffer(ws, ldmBucketSize);
-            ZSTD_memset(zc->ldmState.bucketOffsets, 0, ldmBucketSize);
+            size_t const numBuckets =
+                  ((size_t)1) << (params->ldmParams.hashLog -
+                                  params->ldmParams.bucketSizeLog);
+            zc->ldmState.bucketOffsets = ZSTD_cwksp_reserve_buffer(ws, numBuckets);
+            ZSTD_memset(zc->ldmState.bucketOffsets, 0, numBuckets);
         }
 
         /* sequences storage */
@@ -1729,32 +1980,28 @@ static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc,
         FORWARD_IF_ERROR(ZSTD_reset_matchState(
             &zc->blockState.matchState,
             ws,
-            &params.cParams,
+            &params->cParams,
+            params->useRowMatchFinder,
             crp,
             needsIndexReset,
             ZSTD_resetTarget_CCtx), "");
 
         /* ldm hash table */
-        if (params.ldmParams.enableLdm) {
+        if (params->ldmParams.enableLdm) {
             /* TODO: avoid memset? */
-            size_t const ldmHSize = ((size_t)1) << params.ldmParams.hashLog;
+            size_t const ldmHSize = ((size_t)1) << params->ldmParams.hashLog;
             zc->ldmState.hashTable = (ldmEntry_t*)ZSTD_cwksp_reserve_aligned(ws, ldmHSize * sizeof(ldmEntry_t));
             ZSTD_memset(zc->ldmState.hashTable, 0, ldmHSize * sizeof(ldmEntry_t));
             zc->ldmSequences = (rawSeq*)ZSTD_cwksp_reserve_aligned(ws, maxNbLdmSeq * sizeof(rawSeq));
             zc->maxNbLdmSequences = maxNbLdmSeq;
 
             ZSTD_window_init(&zc->ldmState.window);
-            ZSTD_window_clear(&zc->ldmState.window);
             zc->ldmState.loadedDictEnd = 0;
         }
 
-        /* Due to alignment, when reusing a workspace, we can actually consume
-         * up to 3 extra bytes for alignment. See the comments in zstd_cwksp.h
-         */
-        assert(ZSTD_cwksp_used(ws) >= neededSpace &&
-               ZSTD_cwksp_used(ws) <= neededSpace + 3);
-
+        assert(ZSTD_cwksp_estimated_space_within_bounds(ws, neededSpace, resizeWorkspace));
         DEBUGLOG(3, "wksp: finished allocating, %zd bytes remain available", ZSTD_cwksp_available_space(ws));
+
         zc->initialized = 1;
 
         return 0;
@@ -1810,6 +2057,8 @@ ZSTD_resetCCtx_byAttachingCDict(ZSTD_CCtx* cctx,
                         U64 pledgedSrcSize,
                         ZSTD_buffered_policy_e zbuff)
 {
+    DEBUGLOG(4, "ZSTD_resetCCtx_byAttachingCDict() pledgedSrcSize=%llu",
+                (unsigned long long)pledgedSrcSize);
     {
         ZSTD_compressionParameters adjusted_cdict_cParams = cdict->matchState.cParams;
         unsigned const windowLog = params.cParams.windowLog;
@@ -1825,7 +2074,9 @@ ZSTD_resetCCtx_byAttachingCDict(ZSTD_CCtx* cctx,
         params.cParams = ZSTD_adjustCParams_internal(adjusted_cdict_cParams, pledgedSrcSize,
                                                      cdict->dictContentSize, ZSTD_cpm_attachDict);
         params.cParams.windowLog = windowLog;
-        FORWARD_IF_ERROR(ZSTD_resetCCtx_internal(cctx, params, pledgedSrcSize,
+        params.useRowMatchFinder = cdict->useRowMatchFinder;    /* cdict overrides */
+        FORWARD_IF_ERROR(ZSTD_resetCCtx_internal(cctx, &params, pledgedSrcSize,
+                                                 /* loadedDictSize */ 0,
                                                  ZSTDcrp_makeClean, zbuff), "");
         assert(cctx->appliedParams.cParams.strategy == adjusted_cdict_cParams.strategy);
     }
@@ -1852,6 +2103,7 @@ ZSTD_resetCCtx_byAttachingCDict(ZSTD_CCtx* cctx,
     }   }
 
     cctx->dictID = cdict->dictID;
+    cctx->dictContentSize = cdict->dictContentSize;
 
     /* copy block state */
     ZSTD_memcpy(cctx->blockState.prevCBlock, &cdict->cBlockState, sizeof(cdict->cBlockState));
@@ -1868,15 +2120,17 @@ static size_t ZSTD_resetCCtx_byCopyingCDict(ZSTD_CCtx* cctx,
     const ZSTD_compressionParameters *cdict_cParams = &cdict->matchState.cParams;
 
     assert(!cdict->matchState.dedicatedDictSearch);
-
-    DEBUGLOG(4, "copying dictionary into context");
+    DEBUGLOG(4, "ZSTD_resetCCtx_byCopyingCDict() pledgedSrcSize=%llu",
+                (unsigned long long)pledgedSrcSize);
 
     {   unsigned const windowLog = params.cParams.windowLog;
         assert(windowLog != 0);
         /* Copy only compression parameters related to tables. */
         params.cParams = *cdict_cParams;
         params.cParams.windowLog = windowLog;
-        FORWARD_IF_ERROR(ZSTD_resetCCtx_internal(cctx, params, pledgedSrcSize,
+        params.useRowMatchFinder = cdict->useRowMatchFinder;
+        FORWARD_IF_ERROR(ZSTD_resetCCtx_internal(cctx, &params, pledgedSrcSize,
+                                                 /* loadedDictSize */ 0,
                                                  ZSTDcrp_leaveDirty, zbuff), "");
         assert(cctx->appliedParams.cParams.strategy == cdict_cParams->strategy);
         assert(cctx->appliedParams.cParams.hashLog == cdict_cParams->hashLog);
@@ -1884,17 +2138,30 @@ static size_t ZSTD_resetCCtx_byCopyingCDict(ZSTD_CCtx* cctx,
     }
 
     ZSTD_cwksp_mark_tables_dirty(&cctx->workspace);
+    assert(params.useRowMatchFinder != ZSTD_urm_auto);
 
     /* copy tables */
-    {   size_t const chainSize = (cdict_cParams->strategy == ZSTD_fast) ? 0 : ((size_t)1 << cdict_cParams->chainLog);
+    {   size_t const chainSize = ZSTD_allocateChainTable(cdict_cParams->strategy, cdict->useRowMatchFinder, 0 /* DDS guaranteed disabled */)
+                                                            ? ((size_t)1 << cdict_cParams->chainLog)
+                                                            : 0;
         size_t const hSize =  (size_t)1 << cdict_cParams->hashLog;
 
         ZSTD_memcpy(cctx->blockState.matchState.hashTable,
                cdict->matchState.hashTable,
                hSize * sizeof(U32));
-        ZSTD_memcpy(cctx->blockState.matchState.chainTable,
+        /* Do not copy cdict's chainTable if cctx has parameters such that it would not use chainTable */
+        if (ZSTD_allocateChainTable(cctx->appliedParams.cParams.strategy, cctx->appliedParams.useRowMatchFinder, 0 /* forDDSDict */)) {
+            ZSTD_memcpy(cctx->blockState.matchState.chainTable,
                cdict->matchState.chainTable,
                chainSize * sizeof(U32));
+        }
+        /* copy tag table */
+        if (ZSTD_rowMatchFinderUsed(cdict_cParams->strategy, cdict->useRowMatchFinder)) {
+            size_t const tagTableSize = hSize*sizeof(U16);
+            ZSTD_memcpy(cctx->blockState.matchState.tagTable,
+                cdict->matchState.tagTable,
+                tagTableSize);
+        }
     }
 
     /* Zero the hashTable3, since the cdict never fills it */
@@ -1915,6 +2182,7 @@ static size_t ZSTD_resetCCtx_byCopyingCDict(ZSTD_CCtx* cctx,
     }
 
     cctx->dictID = cdict->dictID;
+    cctx->dictContentSize = cdict->dictContentSize;
 
     /* copy block state */
     ZSTD_memcpy(cctx->blockState.prevCBlock, &cdict->cBlockState, sizeof(cdict->cBlockState));
@@ -1957,16 +2225,18 @@ static size_t ZSTD_copyCCtx_internal(ZSTD_CCtx* dstCCtx,
                             U64 pledgedSrcSize,
                             ZSTD_buffered_policy_e zbuff)
 {
-    DEBUGLOG(5, "ZSTD_copyCCtx_internal");
     RETURN_ERROR_IF(srcCCtx->stage!=ZSTDcs_init, stage_wrong,
                     "Can't copy a ctx that's not in init stage.");
-
+    DEBUGLOG(5, "ZSTD_copyCCtx_internal");
     ZSTD_memcpy(&dstCCtx->customMem, &srcCCtx->customMem, sizeof(ZSTD_customMem));
     {   ZSTD_CCtx_params params = dstCCtx->requestedParams;
         /* Copy only compression parameters related to tables. */
         params.cParams = srcCCtx->appliedParams.cParams;
+        assert(srcCCtx->appliedParams.useRowMatchFinder != ZSTD_urm_auto);
+        params.useRowMatchFinder = srcCCtx->appliedParams.useRowMatchFinder;
         params.fParams = fParams;
-        ZSTD_resetCCtx_internal(dstCCtx, params, pledgedSrcSize,
+        ZSTD_resetCCtx_internal(dstCCtx, &params, pledgedSrcSize,
+                                /* loadedDictSize */ 0,
                                 ZSTDcrp_leaveDirty, zbuff);
         assert(dstCCtx->appliedParams.cParams.windowLog == srcCCtx->appliedParams.cParams.windowLog);
         assert(dstCCtx->appliedParams.cParams.strategy == srcCCtx->appliedParams.cParams.strategy);
@@ -1978,7 +2248,11 @@ static size_t ZSTD_copyCCtx_internal(ZSTD_CCtx* dstCCtx,
     ZSTD_cwksp_mark_tables_dirty(&dstCCtx->workspace);
 
     /* copy tables */
-    {   size_t const chainSize = (srcCCtx->appliedParams.cParams.strategy == ZSTD_fast) ? 0 : ((size_t)1 << srcCCtx->appliedParams.cParams.chainLog);
+    {   size_t const chainSize = ZSTD_allocateChainTable(srcCCtx->appliedParams.cParams.strategy,
+                                                         srcCCtx->appliedParams.useRowMatchFinder,
+                                                         0 /* forDDSDict */)
+                                    ? ((size_t)1 << srcCCtx->appliedParams.cParams.chainLog)
+                                    : 0;
         size_t const hSize =  (size_t)1 << srcCCtx->appliedParams.cParams.hashLog;
         int const h3log = srcCCtx->blockState.matchState.hashLog3;
         size_t const h3Size = h3log ? ((size_t)1 << h3log) : 0;
@@ -2005,6 +2279,7 @@ static size_t ZSTD_copyCCtx_internal(ZSTD_CCtx* dstCCtx,
         dstMatchState->loadedDictEnd= srcMatchState->loadedDictEnd;
     }
     dstCCtx->dictID = srcCCtx->dictID;
+    dstCCtx->dictContentSize = srcCCtx->dictContentSize;
 
     /* copy block state */
     ZSTD_memcpy(dstCCtx->blockState.prevCBlock, srcCCtx->blockState.prevCBlock, sizeof(*srcCCtx->blockState.prevCBlock));
@@ -2091,7 +2366,7 @@ static void ZSTD_reduceIndex (ZSTD_matchState_t* ms, ZSTD_CCtx_params const* par
         ZSTD_reduceTable(ms->hashTable, hSize, reducerValue);
     }
 
-    if (params->cParams.strategy != ZSTD_fast) {
+    if (ZSTD_allocateChainTable(params->cParams.strategy, params->useRowMatchFinder, (U32)ms->dedicatedDictSearch)) {
         U32 const chainSize = (U32)1 << params->cParams.chainLog;
         if (params->cParams.strategy == ZSTD_btlazy2)
             ZSTD_reduceTable_btlazy2(ms->chainTable, chainSize, reducerValue);
@@ -2128,9 +2403,9 @@ void ZSTD_seqToCodes(const seqStore_t* seqStorePtr)
         ofCodeTable[u] = (BYTE)ZSTD_highbit32(sequences[u].offset);
         mlCodeTable[u] = (BYTE)ZSTD_MLcode(mlv);
     }
-    if (seqStorePtr->longLengthID==1)
+    if (seqStorePtr->longLengthType==ZSTD_llt_literalLength)
         llCodeTable[seqStorePtr->longLengthPos] = MaxLL;
-    if (seqStorePtr->longLengthID==2)
+    if (seqStorePtr->longLengthType==ZSTD_llt_matchLength)
         mlCodeTable[seqStorePtr->longLengthPos] = MaxML;
 }
 
@@ -2144,10 +2419,159 @@ static int ZSTD_useTargetCBlockSize(const ZSTD_CCtx_params* cctxParams)
     return (cctxParams->targetCBlockSize != 0);
 }
 
-/* ZSTD_entropyCompressSequences_internal():
- * actually compresses both literals and sequences */
+/* ZSTD_blockSplitterEnabled():
+ * Returns if block splitting param is being used
+ * If used, compression will do best effort to split a block in order to improve compression ratio.
+ * Returns 1 if true, 0 otherwise. */
+static int ZSTD_blockSplitterEnabled(ZSTD_CCtx_params* cctxParams)
+{
+    DEBUGLOG(5, "ZSTD_blockSplitterEnabled(splitBlocks=%d)", cctxParams->splitBlocks);
+    return (cctxParams->splitBlocks != 0);
+}
+
+/* Type returned by ZSTD_buildSequencesStatistics containing finalized symbol encoding types
+ * and size of the sequences statistics
+ */
+typedef struct {
+    U32 LLtype;
+    U32 Offtype;
+    U32 MLtype;
+    size_t size;
+    size_t lastCountSize; /* Accounts for bug in 1.3.4. More detail in ZSTD_entropyCompressSeqStore_internal() */
+} ZSTD_symbolEncodingTypeStats_t;
+
+/* ZSTD_buildSequencesStatistics():
+ * Returns a ZSTD_symbolEncodingTypeStats_t, or a zstd error code in the `size` field.
+ * Modifies `nextEntropy` to have the appropriate values as a side effect.
+ * nbSeq must be greater than 0.
+ *
+ * entropyWkspSize must be of size at least ENTROPY_WORKSPACE_SIZE - (MaxSeq + 1)*sizeof(U32)
+ */
+static ZSTD_symbolEncodingTypeStats_t
+ZSTD_buildSequencesStatistics(seqStore_t* seqStorePtr, size_t nbSeq,
+                        const ZSTD_fseCTables_t* prevEntropy, ZSTD_fseCTables_t* nextEntropy,
+                              BYTE* dst, const BYTE* const dstEnd,
+                              ZSTD_strategy strategy, unsigned* countWorkspace,
+                              void* entropyWorkspace, size_t entropyWkspSize) {
+    BYTE* const ostart = dst;
+    const BYTE* const oend = dstEnd;
+    BYTE* op = ostart;
+    FSE_CTable* CTable_LitLength = nextEntropy->litlengthCTable;
+    FSE_CTable* CTable_OffsetBits = nextEntropy->offcodeCTable;
+    FSE_CTable* CTable_MatchLength = nextEntropy->matchlengthCTable;
+    const BYTE* const ofCodeTable = seqStorePtr->ofCode;
+    const BYTE* const llCodeTable = seqStorePtr->llCode;
+    const BYTE* const mlCodeTable = seqStorePtr->mlCode;
+    ZSTD_symbolEncodingTypeStats_t stats;
+
+    stats.lastCountSize = 0;
+    /* convert length/distances into codes */
+    ZSTD_seqToCodes(seqStorePtr);
+    assert(op <= oend);
+    assert(nbSeq != 0); /* ZSTD_selectEncodingType() divides by nbSeq */
+    /* build CTable for Literal Lengths */
+    {   unsigned max = MaxLL;
+        size_t const mostFrequent = HIST_countFast_wksp(countWorkspace, &max, llCodeTable, nbSeq, entropyWorkspace, entropyWkspSize);   /* can't fail */
+        DEBUGLOG(5, "Building LL table");
+        nextEntropy->litlength_repeatMode = prevEntropy->litlength_repeatMode;
+        stats.LLtype = ZSTD_selectEncodingType(&nextEntropy->litlength_repeatMode,
+                                        countWorkspace, max, mostFrequent, nbSeq,
+                                        LLFSELog, prevEntropy->litlengthCTable,
+                                        LL_defaultNorm, LL_defaultNormLog,
+                                        ZSTD_defaultAllowed, strategy);
+        assert(set_basic < set_compressed && set_rle < set_compressed);
+        assert(!(stats.LLtype < set_compressed && nextEntropy->litlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */
+        {   size_t const countSize = ZSTD_buildCTable(
+                op, (size_t)(oend - op),
+                CTable_LitLength, LLFSELog, (symbolEncodingType_e)stats.LLtype,
+                countWorkspace, max, llCodeTable, nbSeq,
+                LL_defaultNorm, LL_defaultNormLog, MaxLL,
+                prevEntropy->litlengthCTable,
+                sizeof(prevEntropy->litlengthCTable),
+                entropyWorkspace, entropyWkspSize);
+            if (ZSTD_isError(countSize)) {
+                DEBUGLOG(3, "ZSTD_buildCTable for LitLens failed");
+                stats.size = countSize;
+                return stats;
+            }
+            if (stats.LLtype == set_compressed)
+                stats.lastCountSize = countSize;
+            op += countSize;
+            assert(op <= oend);
+    }   }
+    /* build CTable for Offsets */
+    {   unsigned max = MaxOff;
+        size_t const mostFrequent = HIST_countFast_wksp(
+            countWorkspace, &max, ofCodeTable, nbSeq, entropyWorkspace, entropyWkspSize);  /* can't fail */
+        /* We can only use the basic table if max <= DefaultMaxOff, otherwise the offsets are too large */
+        ZSTD_defaultPolicy_e const defaultPolicy = (max <= DefaultMaxOff) ? ZSTD_defaultAllowed : ZSTD_defaultDisallowed;
+        DEBUGLOG(5, "Building OF table");
+        nextEntropy->offcode_repeatMode = prevEntropy->offcode_repeatMode;
+        stats.Offtype = ZSTD_selectEncodingType(&nextEntropy->offcode_repeatMode,
+                                        countWorkspace, max, mostFrequent, nbSeq,
+                                        OffFSELog, prevEntropy->offcodeCTable,
+                                        OF_defaultNorm, OF_defaultNormLog,
+                                        defaultPolicy, strategy);
+        assert(!(stats.Offtype < set_compressed && nextEntropy->offcode_repeatMode != FSE_repeat_none)); /* We don't copy tables */
+        {   size_t const countSize = ZSTD_buildCTable(
+                op, (size_t)(oend - op),
+                CTable_OffsetBits, OffFSELog, (symbolEncodingType_e)stats.Offtype,
+                countWorkspace, max, ofCodeTable, nbSeq,
+                OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff,
+                prevEntropy->offcodeCTable,
+                sizeof(prevEntropy->offcodeCTable),
+                entropyWorkspace, entropyWkspSize);
+            if (ZSTD_isError(countSize)) {
+                DEBUGLOG(3, "ZSTD_buildCTable for Offsets failed");
+                stats.size = countSize;
+                return stats;
+            }
+            if (stats.Offtype == set_compressed)
+                stats.lastCountSize = countSize;
+            op += countSize;
+            assert(op <= oend);
+    }   }
+    /* build CTable for MatchLengths */
+    {   unsigned max = MaxML;
+        size_t const mostFrequent = HIST_countFast_wksp(
+            countWorkspace, &max, mlCodeTable, nbSeq, entropyWorkspace, entropyWkspSize);   /* can't fail */
+        DEBUGLOG(5, "Building ML table (remaining space : %i)", (int)(oend-op));
+        nextEntropy->matchlength_repeatMode = prevEntropy->matchlength_repeatMode;
+        stats.MLtype = ZSTD_selectEncodingType(&nextEntropy->matchlength_repeatMode,
+                                        countWorkspace, max, mostFrequent, nbSeq,
+                                        MLFSELog, prevEntropy->matchlengthCTable,
+                                        ML_defaultNorm, ML_defaultNormLog,
+                                        ZSTD_defaultAllowed, strategy);
+        assert(!(stats.MLtype < set_compressed && nextEntropy->matchlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */
+        {   size_t const countSize = ZSTD_buildCTable(
+                op, (size_t)(oend - op),
+                CTable_MatchLength, MLFSELog, (symbolEncodingType_e)stats.MLtype,
+                countWorkspace, max, mlCodeTable, nbSeq,
+                ML_defaultNorm, ML_defaultNormLog, MaxML,
+                prevEntropy->matchlengthCTable,
+                sizeof(prevEntropy->matchlengthCTable),
+                entropyWorkspace, entropyWkspSize);
+            if (ZSTD_isError(countSize)) {
+                DEBUGLOG(3, "ZSTD_buildCTable for MatchLengths failed");
+                stats.size = countSize;
+                return stats;
+            }
+            if (stats.MLtype == set_compressed)
+                stats.lastCountSize = countSize;
+            op += countSize;
+            assert(op <= oend);
+    }   }
+    stats.size = (size_t)(op-ostart);
+    return stats;
+}
+
+/* ZSTD_entropyCompressSeqStore_internal():
+ * compresses both literals and sequences
+ * Returns compressed size of block, or a zstd error.
+ */
+#define SUSPECT_UNCOMPRESSIBLE_LITERAL_RATIO 20
 MEM_STATIC size_t
-ZSTD_entropyCompressSequences_internal(seqStore_t* seqStorePtr,
+ZSTD_entropyCompressSeqStore_internal(seqStore_t* seqStorePtr,
                           const ZSTD_entropyCTables_t* prevEntropy,
                                 ZSTD_entropyCTables_t* nextEntropy,
                           const ZSTD_CCtx_params* cctxParams,
@@ -2161,27 +2585,29 @@ ZSTD_entropyCompressSequences_internal(seqStore_t* seqStorePtr,
     FSE_CTable* CTable_LitLength = nextEntropy->fse.litlengthCTable;
     FSE_CTable* CTable_OffsetBits = nextEntropy->fse.offcodeCTable;
     FSE_CTable* CTable_MatchLength = nextEntropy->fse.matchlengthCTable;
-    U32 LLtype, Offtype, MLtype;   /* compressed, raw or rle */
     const seqDef* const sequences = seqStorePtr->sequencesStart;
+    const size_t nbSeq = seqStorePtr->sequences - seqStorePtr->sequencesStart;
     const BYTE* const ofCodeTable = seqStorePtr->ofCode;
     const BYTE* const llCodeTable = seqStorePtr->llCode;
     const BYTE* const mlCodeTable = seqStorePtr->mlCode;
     BYTE* const ostart = (BYTE*)dst;
     BYTE* const oend = ostart + dstCapacity;
     BYTE* op = ostart;
-    size_t const nbSeq = (size_t)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
-    BYTE* seqHead;
-    BYTE* lastNCount = NULL;
+    size_t lastCountSize;
 
     entropyWorkspace = count + (MaxSeq + 1);
     entropyWkspSize -= (MaxSeq + 1) * sizeof(*count);
 
-    DEBUGLOG(4, "ZSTD_entropyCompressSequences_internal (nbSeq=%zu)", nbSeq);
+    DEBUGLOG(4, "ZSTD_entropyCompressSeqStore_internal (nbSeq=%zu)", nbSeq);
     ZSTD_STATIC_ASSERT(HUF_WORKSPACE_SIZE >= (1<<MAX(MLFSELog,LLFSELog)));
     assert(entropyWkspSize >= HUF_WORKSPACE_SIZE);
 
     /* Compress literals */
     {   const BYTE* const literals = seqStorePtr->litStart;
+        size_t const numSequences = seqStorePtr->sequences - seqStorePtr->sequencesStart;
+        size_t const numLiterals = seqStorePtr->lit - seqStorePtr->litStart;
+        /* Base suspicion of uncompressibility on ratio of literals to sequences */
+        unsigned const suspectUncompressible = (numSequences == 0) || (numLiterals / numSequences >= SUSPECT_UNCOMPRESSIBLE_LITERAL_RATIO);
         size_t const litSize = (size_t)(seqStorePtr->lit - literals);
         size_t const cSize = ZSTD_compressLiterals(
                                     &prevEntropy->huf, &nextEntropy->huf,
@@ -2190,7 +2616,7 @@ ZSTD_entropyCompressSequences_internal(seqStore_t* seqStorePtr,
                                     op, dstCapacity,
                                     literals, litSize,
                                     entropyWorkspace, entropyWkspSize,
-                                    bmi2);
+                                    bmi2, suspectUncompressible);
         FORWARD_IF_ERROR(cSize, "ZSTD_compressLiterals failed");
         assert(cSize <= dstCapacity);
         op += cSize;
@@ -2216,95 +2642,20 @@ ZSTD_entropyCompressSequences_internal(seqStore_t* seqStorePtr,
         ZSTD_memcpy(&nextEntropy->fse, &prevEntropy->fse, sizeof(prevEntropy->fse));
         return (size_t)(op - ostart);
     }
-
-    /* seqHead : flags for FSE encoding type */
-    seqHead = op++;
-    assert(op <= oend);
-
-    /* convert length/distances into codes */
-    ZSTD_seqToCodes(seqStorePtr);
-    /* build CTable for Literal Lengths */
-    {   unsigned max = MaxLL;
-        size_t const mostFrequent = HIST_countFast_wksp(count, &max, llCodeTable, nbSeq, entropyWorkspace, entropyWkspSize);   /* can't fail */
-        DEBUGLOG(5, "Building LL table");
-        nextEntropy->fse.litlength_repeatMode = prevEntropy->fse.litlength_repeatMode;
-        LLtype = ZSTD_selectEncodingType(&nextEntropy->fse.litlength_repeatMode,
-                                        count, max, mostFrequent, nbSeq,
-                                        LLFSELog, prevEntropy->fse.litlengthCTable,
-                                        LL_defaultNorm, LL_defaultNormLog,
-                                        ZSTD_defaultAllowed, strategy);
-        assert(set_basic < set_compressed && set_rle < set_compressed);
-        assert(!(LLtype < set_compressed && nextEntropy->fse.litlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */
-        {   size_t const countSize = ZSTD_buildCTable(
-                op, (size_t)(oend - op),
-                CTable_LitLength, LLFSELog, (symbolEncodingType_e)LLtype,
-                count, max, llCodeTable, nbSeq,
-                LL_defaultNorm, LL_defaultNormLog, MaxLL,
-                prevEntropy->fse.litlengthCTable,
-                sizeof(prevEntropy->fse.litlengthCTable),
-                entropyWorkspace, entropyWkspSize);
-            FORWARD_IF_ERROR(countSize, "ZSTD_buildCTable for LitLens failed");
-            if (LLtype == set_compressed)
-                lastNCount = op;
-            op += countSize;
-            assert(op <= oend);
-    }   }
-    /* build CTable for Offsets */
-    {   unsigned max = MaxOff;
-        size_t const mostFrequent = HIST_countFast_wksp(
-            count, &max, ofCodeTable, nbSeq, entropyWorkspace, entropyWkspSize);  /* can't fail */
-        /* We can only use the basic table if max <= DefaultMaxOff, otherwise the offsets are too large */
-        ZSTD_defaultPolicy_e const defaultPolicy = (max <= DefaultMaxOff) ? ZSTD_defaultAllowed : ZSTD_defaultDisallowed;
-        DEBUGLOG(5, "Building OF table");
-        nextEntropy->fse.offcode_repeatMode = prevEntropy->fse.offcode_repeatMode;
-        Offtype = ZSTD_selectEncodingType(&nextEntropy->fse.offcode_repeatMode,
-                                        count, max, mostFrequent, nbSeq,
-                                        OffFSELog, prevEntropy->fse.offcodeCTable,
-                                        OF_defaultNorm, OF_defaultNormLog,
-                                        defaultPolicy, strategy);
-        assert(!(Offtype < set_compressed && nextEntropy->fse.offcode_repeatMode != FSE_repeat_none)); /* We don't copy tables */
-        {   size_t const countSize = ZSTD_buildCTable(
-                op, (size_t)(oend - op),
-                CTable_OffsetBits, OffFSELog, (symbolEncodingType_e)Offtype,
-                count, max, ofCodeTable, nbSeq,
-                OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff,
-                prevEntropy->fse.offcodeCTable,
-                sizeof(prevEntropy->fse.offcodeCTable),
-                entropyWorkspace, entropyWkspSize);
-            FORWARD_IF_ERROR(countSize, "ZSTD_buildCTable for Offsets failed");
-            if (Offtype == set_compressed)
-                lastNCount = op;
-            op += countSize;
-            assert(op <= oend);
-    }   }
-    /* build CTable for MatchLengths */
-    {   unsigned max = MaxML;
-        size_t const mostFrequent = HIST_countFast_wksp(
-            count, &max, mlCodeTable, nbSeq, entropyWorkspace, entropyWkspSize);   /* can't fail */
-        DEBUGLOG(5, "Building ML table (remaining space : %i)", (int)(oend-op));
-        nextEntropy->fse.matchlength_repeatMode = prevEntropy->fse.matchlength_repeatMode;
-        MLtype = ZSTD_selectEncodingType(&nextEntropy->fse.matchlength_repeatMode,
-                                        count, max, mostFrequent, nbSeq,
-                                        MLFSELog, prevEntropy->fse.matchlengthCTable,
-                                        ML_defaultNorm, ML_defaultNormLog,
-                                        ZSTD_defaultAllowed, strategy);
-        assert(!(MLtype < set_compressed && nextEntropy->fse.matchlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */
-        {   size_t const countSize = ZSTD_buildCTable(
-                op, (size_t)(oend - op),
-                CTable_MatchLength, MLFSELog, (symbolEncodingType_e)MLtype,
-                count, max, mlCodeTable, nbSeq,
-                ML_defaultNorm, ML_defaultNormLog, MaxML,
-                prevEntropy->fse.matchlengthCTable,
-                sizeof(prevEntropy->fse.matchlengthCTable),
-                entropyWorkspace, entropyWkspSize);
-            FORWARD_IF_ERROR(countSize, "ZSTD_buildCTable for MatchLengths failed");
-            if (MLtype == set_compressed)
-                lastNCount = op;
-            op += countSize;
-            assert(op <= oend);
-    }   }
-
-    *seqHead = (BYTE)((LLtype<<6) + (Offtype<<4) + (MLtype<<2));
+    {
+        ZSTD_symbolEncodingTypeStats_t stats;
+        BYTE* seqHead = op++;
+        /* build stats for sequences */
+        stats = ZSTD_buildSequencesStatistics(seqStorePtr, nbSeq,
+                                             &prevEntropy->fse, &nextEntropy->fse,
+                                              op, oend,
+                                              strategy, count,
+                                              entropyWorkspace, entropyWkspSize);
+        FORWARD_IF_ERROR(stats.size, "ZSTD_buildSequencesStatistics failed!");
+        *seqHead = (BYTE)((stats.LLtype<<6) + (stats.Offtype<<4) + (stats.MLtype<<2));
+        lastCountSize = stats.lastCountSize;
+        op += stats.size;
+    }
 
     {   size_t const bitstreamSize = ZSTD_encodeSequences(
                                         op, (size_t)(oend - op),
@@ -2324,9 +2675,9 @@ ZSTD_entropyCompressSequences_internal(seqStore_t* seqStorePtr,
          * In this exceedingly rare case, we will simply emit an uncompressed
          * block, since it isn't worth optimizing.
          */
-        if (lastNCount && (op - lastNCount) < 4) {
-            /* NCountSize >= 2 && bitstreamSize > 0 ==> lastCountSize == 3 */
-            assert(op - lastNCount == 3);
+        if (lastCountSize && (lastCountSize + bitstreamSize) < 4) {
+            /* lastCountSize >= 2 && bitstreamSize > 0 ==> lastCountSize == 3 */
+            assert(lastCountSize + bitstreamSize == 3);
             DEBUGLOG(5, "Avoiding bug in zstd decoder in versions <= 1.3.4 by "
                         "emitting an uncompressed block.");
             return 0;
@@ -2338,7 +2689,7 @@ ZSTD_entropyCompressSequences_internal(seqStore_t* seqStorePtr,
 }
 
 MEM_STATIC size_t
-ZSTD_entropyCompressSequences(seqStore_t* seqStorePtr,
+ZSTD_entropyCompressSeqStore(seqStore_t* seqStorePtr,
                        const ZSTD_entropyCTables_t* prevEntropy,
                              ZSTD_entropyCTables_t* nextEntropy,
                        const ZSTD_CCtx_params* cctxParams,
@@ -2347,7 +2698,7 @@ ZSTD_entropyCompressSequences(seqStore_t* seqStorePtr,
                              void* entropyWorkspace, size_t entropyWkspSize,
                              int bmi2)
 {
-    size_t const cSize = ZSTD_entropyCompressSequences_internal(
+    size_t const cSize = ZSTD_entropyCompressSeqStore_internal(
                             seqStorePtr, prevEntropy, nextEntropy, cctxParams,
                             dst, dstCapacity,
                             entropyWorkspace, entropyWkspSize, bmi2);
@@ -2357,20 +2708,20 @@ ZSTD_entropyCompressSequences(seqStore_t* seqStorePtr,
      */
     if ((cSize == ERROR(dstSize_tooSmall)) & (srcSize <= dstCapacity))
         return 0;  /* block not compressed */
-    FORWARD_IF_ERROR(cSize, "ZSTD_entropyCompressSequences_internal failed");
+    FORWARD_IF_ERROR(cSize, "ZSTD_entropyCompressSeqStore_internal failed");
 
     /* Check compressibility */
     {   size_t const maxCSize = srcSize - ZSTD_minGain(srcSize, cctxParams->cParams.strategy);
         if (cSize >= maxCSize) return 0;  /* block not compressed */
     }
-    DEBUGLOG(4, "ZSTD_entropyCompressSequences() cSize: %zu\n", cSize);
+    DEBUGLOG(4, "ZSTD_entropyCompressSeqStore() cSize: %zu", cSize);
     return cSize;
 }
 
 /* ZSTD_selectBlockCompressor() :
  * Not static, but internal use only (used by long distance matcher)
  * assumption : strat is a valid strategy */
-ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_dictMode_e dictMode)
+ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_useRowMatchFinderMode_e useRowMatchFinder, ZSTD_dictMode_e dictMode)
 {
     static const ZSTD_blockCompressor blockCompressor[4][ZSTD_STRATEGY_MAX+1] = {
         { ZSTD_compressBlock_fast  /* default for 0 */,
@@ -2418,7 +2769,28 @@ ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_dictMo
     ZSTD_STATIC_ASSERT((unsigned)ZSTD_fast == 1);
 
     assert(ZSTD_cParam_withinBounds(ZSTD_c_strategy, strat));
-    selectedCompressor = blockCompressor[(int)dictMode][(int)strat];
+    DEBUGLOG(4, "Selected block compressor: dictMode=%d strat=%d rowMatchfinder=%d", (int)dictMode, (int)strat, (int)useRowMatchFinder);
+    if (ZSTD_rowMatchFinderUsed(strat, useRowMatchFinder)) {
+        static const ZSTD_blockCompressor rowBasedBlockCompressors[4][3] = {
+            { ZSTD_compressBlock_greedy_row,
+            ZSTD_compressBlock_lazy_row,
+            ZSTD_compressBlock_lazy2_row },
+            { ZSTD_compressBlock_greedy_extDict_row,
+            ZSTD_compressBlock_lazy_extDict_row,
+            ZSTD_compressBlock_lazy2_extDict_row },
+            { ZSTD_compressBlock_greedy_dictMatchState_row,
+            ZSTD_compressBlock_lazy_dictMatchState_row,
+            ZSTD_compressBlock_lazy2_dictMatchState_row },
+            { ZSTD_compressBlock_greedy_dedicatedDictSearch_row,
+            ZSTD_compressBlock_lazy_dedicatedDictSearch_row,
+            ZSTD_compressBlock_lazy2_dedicatedDictSearch_row }
+        };
+        DEBUGLOG(4, "Selecting a row-based matchfinder");
+        assert(useRowMatchFinder != ZSTD_urm_auto);
+        selectedCompressor = rowBasedBlockCompressors[(int)dictMode][(int)strat - (int)ZSTD_greedy];
+    } else {
+        selectedCompressor = blockCompressor[(int)dictMode][(int)strat];
+    }
     assert(selectedCompressor != NULL);
     return selectedCompressor;
 }
@@ -2434,7 +2806,7 @@ void ZSTD_resetSeqStore(seqStore_t* ssPtr)
 {
     ssPtr->lit = ssPtr->litStart;
     ssPtr->sequences = ssPtr->sequencesStart;
-    ssPtr->longLengthID = 0;
+    ssPtr->longLengthType = ZSTD_llt_none;
 }
 
 typedef enum { ZSTDbss_compress, ZSTDbss_noCompress } ZSTD_buildSeqStore_e;
@@ -2487,6 +2859,7 @@ static size_t ZSTD_buildSeqStore(ZSTD_CCtx* zc, const void* src, size_t srcSize)
                 ZSTD_ldm_blockCompress(&zc->externSeqStore,
                                        ms, &zc->seqStore,
                                        zc->blockState.nextCBlock->rep,
+                                       zc->appliedParams.useRowMatchFinder,
                                        src, srcSize);
             assert(zc->externSeqStore.pos <= zc->externSeqStore.size);
         } else if (zc->appliedParams.ldmParams.enableLdm) {
@@ -2503,10 +2876,13 @@ static size_t ZSTD_buildSeqStore(ZSTD_CCtx* zc, const void* src, size_t srcSize)
                 ZSTD_ldm_blockCompress(&ldmSeqStore,
                                        ms, &zc->seqStore,
                                        zc->blockState.nextCBlock->rep,
+                                       zc->appliedParams.useRowMatchFinder,
                                        src, srcSize);
             assert(ldmSeqStore.pos == ldmSeqStore.size);
         } else {   /* not long range mode */
-            ZSTD_blockCompressor const blockCompressor = ZSTD_selectBlockCompressor(zc->appliedParams.cParams.strategy, dictMode);
+            ZSTD_blockCompressor const blockCompressor = ZSTD_selectBlockCompressor(zc->appliedParams.cParams.strategy,
+                                                                                    zc->appliedParams.useRowMatchFinder,
+                                                                                    dictMode);
             ms->ldmSeqStore = NULL;
             lastLLSize = blockCompressor(ms, &zc->seqStore, zc->blockState.nextCBlock->rep, src, srcSize);
         }
@@ -2540,9 +2916,9 @@ static void ZSTD_copyBlockSequences(ZSTD_CCtx* zc)
         outSeqs[i].rep = 0;
 
         if (i == seqStore->longLengthPos) {
-            if (seqStore->longLengthID == 1) {
+            if (seqStore->longLengthType == ZSTD_llt_literalLength) {
                 outSeqs[i].litLength += 0x10000;
-            } else if (seqStore->longLengthID == 2) {
+            } else if (seqStore->longLengthType == ZSTD_llt_matchLength) {
                 outSeqs[i].matchLength += 0x10000;
             }
         }
@@ -2653,13 +3029,730 @@ static int ZSTD_maybeRLE(seqStore_t const* seqStore)
     return nbSeqs < 4 && nbLits < 10;
 }
 
-static void ZSTD_confirmRepcodesAndEntropyTables(ZSTD_CCtx* zc)
+static void ZSTD_blockState_confirmRepcodesAndEntropyTables(ZSTD_blockState_t* const bs)
+{
+    ZSTD_compressedBlockState_t* const tmp = bs->prevCBlock;
+    bs->prevCBlock = bs->nextCBlock;
+    bs->nextCBlock = tmp;
+}
+
+/* Writes the block header */
+static void writeBlockHeader(void* op, size_t cSize, size_t blockSize, U32 lastBlock) {
+    U32 const cBlockHeader = cSize == 1 ?
+                        lastBlock + (((U32)bt_rle)<<1) + (U32)(blockSize << 3) :
+                        lastBlock + (((U32)bt_compressed)<<1) + (U32)(cSize << 3);
+    MEM_writeLE24(op, cBlockHeader);
+    DEBUGLOG(3, "writeBlockHeader: cSize: %zu blockSize: %zu lastBlock: %u", cSize, blockSize, lastBlock);
+}
+
+/** ZSTD_buildBlockEntropyStats_literals() :
+ *  Builds entropy for the literals.
+ *  Stores literals block type (raw, rle, compressed, repeat) and
+ *  huffman description table to hufMetadata.
+ *  Requires ENTROPY_WORKSPACE_SIZE workspace
+ *  @return : size of huffman description table or error code */
+static size_t ZSTD_buildBlockEntropyStats_literals(void* const src, size_t srcSize,
+                                            const ZSTD_hufCTables_t* prevHuf,
+                                                  ZSTD_hufCTables_t* nextHuf,
+                                                  ZSTD_hufCTablesMetadata_t* hufMetadata,
+                                                  const int disableLiteralsCompression,
+                                                  void* workspace, size_t wkspSize)
+{
+    BYTE* const wkspStart = (BYTE*)workspace;
+    BYTE* const wkspEnd = wkspStart + wkspSize;
+    BYTE* const countWkspStart = wkspStart;
+    unsigned* const countWksp = (unsigned*)workspace;
+    const size_t countWkspSize = (HUF_SYMBOLVALUE_MAX + 1) * sizeof(unsigned);
+    BYTE* const nodeWksp = countWkspStart + countWkspSize;
+    const size_t nodeWkspSize = wkspEnd-nodeWksp;
+    unsigned maxSymbolValue = HUF_SYMBOLVALUE_MAX;
+    unsigned huffLog = HUF_TABLELOG_DEFAULT;
+    HUF_repeat repeat = prevHuf->repeatMode;
+    DEBUGLOG(5, "ZSTD_buildBlockEntropyStats_literals (srcSize=%zu)", srcSize);
+
+    /* Prepare nextEntropy assuming reusing the existing table */
+    ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
+
+    if (disableLiteralsCompression) {
+        DEBUGLOG(5, "set_basic - disabled");
+        hufMetadata->hType = set_basic;
+        return 0;
+    }
+
+    /* small ? don't even attempt compression (speed opt) */
+#ifndef COMPRESS_LITERALS_SIZE_MIN
+#define COMPRESS_LITERALS_SIZE_MIN 63
+#endif
+    {   size_t const minLitSize = (prevHuf->repeatMode == HUF_repeat_valid) ? 6 : COMPRESS_LITERALS_SIZE_MIN;
+        if (srcSize <= minLitSize) {
+            DEBUGLOG(5, "set_basic - too small");
+            hufMetadata->hType = set_basic;
+            return 0;
+        }
+    }
+
+    /* Scan input and build symbol stats */
+    {   size_t const largest = HIST_count_wksp (countWksp, &maxSymbolValue, (const BYTE*)src, srcSize, workspace, wkspSize);
+        FORWARD_IF_ERROR(largest, "HIST_count_wksp failed");
+        if (largest == srcSize) {
+            DEBUGLOG(5, "set_rle");
+            hufMetadata->hType = set_rle;
+            return 0;
+        }
+        if (largest <= (srcSize >> 7)+4) {
+            DEBUGLOG(5, "set_basic - no gain");
+            hufMetadata->hType = set_basic;
+            return 0;
+        }
+    }
+
+    /* Validate the previous Huffman table */
+    if (repeat == HUF_repeat_check && !HUF_validateCTable((HUF_CElt const*)prevHuf->CTable, countWksp, maxSymbolValue)) {
+        repeat = HUF_repeat_none;
+    }
+
+    /* Build Huffman Tree */
+    ZSTD_memset(nextHuf->CTable, 0, sizeof(nextHuf->CTable));
+    huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue);
+    {   size_t const maxBits = HUF_buildCTable_wksp((HUF_CElt*)nextHuf->CTable, countWksp,
+                                                    maxSymbolValue, huffLog,
+                                                    nodeWksp, nodeWkspSize);
+        FORWARD_IF_ERROR(maxBits, "HUF_buildCTable_wksp");
+        huffLog = (U32)maxBits;
+        {   /* Build and write the CTable */
+            size_t const newCSize = HUF_estimateCompressedSize(
+                    (HUF_CElt*)nextHuf->CTable, countWksp, maxSymbolValue);
+            size_t const hSize = HUF_writeCTable_wksp(
+                    hufMetadata->hufDesBuffer, sizeof(hufMetadata->hufDesBuffer),
+                    (HUF_CElt*)nextHuf->CTable, maxSymbolValue, huffLog,
+                    nodeWksp, nodeWkspSize);
+            /* Check against repeating the previous CTable */
+            if (repeat != HUF_repeat_none) {
+                size_t const oldCSize = HUF_estimateCompressedSize(
+                        (HUF_CElt const*)prevHuf->CTable, countWksp, maxSymbolValue);
+                if (oldCSize < srcSize && (oldCSize <= hSize + newCSize || hSize + 12 >= srcSize)) {
+                    DEBUGLOG(5, "set_repeat - smaller");
+                    ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
+                    hufMetadata->hType = set_repeat;
+                    return 0;
+                }
+            }
+            if (newCSize + hSize >= srcSize) {
+                DEBUGLOG(5, "set_basic - no gains");
+                ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
+                hufMetadata->hType = set_basic;
+                return 0;
+            }
+            DEBUGLOG(5, "set_compressed (hSize=%u)", (U32)hSize);
+            hufMetadata->hType = set_compressed;
+            nextHuf->repeatMode = HUF_repeat_check;
+            return hSize;
+        }
+    }
+}
+
+
+/* ZSTD_buildDummySequencesStatistics():
+ * Returns a ZSTD_symbolEncodingTypeStats_t with all encoding types as set_basic,
+ * and updates nextEntropy to the appropriate repeatMode.
+ */
+static ZSTD_symbolEncodingTypeStats_t
+ZSTD_buildDummySequencesStatistics(ZSTD_fseCTables_t* nextEntropy) {
+    ZSTD_symbolEncodingTypeStats_t stats = {set_basic, set_basic, set_basic, 0, 0};
+    nextEntropy->litlength_repeatMode = FSE_repeat_none;
+    nextEntropy->offcode_repeatMode = FSE_repeat_none;
+    nextEntropy->matchlength_repeatMode = FSE_repeat_none;
+    return stats;
+}
+
+/** ZSTD_buildBlockEntropyStats_sequences() :
+ *  Builds entropy for the sequences.
+ *  Stores symbol compression modes and fse table to fseMetadata.
+ *  Requires ENTROPY_WORKSPACE_SIZE wksp.
+ *  @return : size of fse tables or error code */
+static size_t ZSTD_buildBlockEntropyStats_sequences(seqStore_t* seqStorePtr,
+                                              const ZSTD_fseCTables_t* prevEntropy,
+                                                    ZSTD_fseCTables_t* nextEntropy,
+                                              const ZSTD_CCtx_params* cctxParams,
+                                                    ZSTD_fseCTablesMetadata_t* fseMetadata,
+                                                    void* workspace, size_t wkspSize)
 {
-    ZSTD_compressedBlockState_t* const tmp = zc->blockState.prevCBlock;
-    zc->blockState.prevCBlock = zc->blockState.nextCBlock;
-    zc->blockState.nextCBlock = tmp;
+    ZSTD_strategy const strategy = cctxParams->cParams.strategy;
+    size_t const nbSeq = seqStorePtr->sequences - seqStorePtr->sequencesStart;
+    BYTE* const ostart = fseMetadata->fseTablesBuffer;
+    BYTE* const oend = ostart + sizeof(fseMetadata->fseTablesBuffer);
+    BYTE* op = ostart;
+    unsigned* countWorkspace = (unsigned*)workspace;
+    unsigned* entropyWorkspace = countWorkspace + (MaxSeq + 1);
+    size_t entropyWorkspaceSize = wkspSize - (MaxSeq + 1) * sizeof(*countWorkspace);
+    ZSTD_symbolEncodingTypeStats_t stats;
+
+    DEBUGLOG(5, "ZSTD_buildBlockEntropyStats_sequences (nbSeq=%zu)", nbSeq);
+    stats = nbSeq != 0 ? ZSTD_buildSequencesStatistics(seqStorePtr, nbSeq,
+                                          prevEntropy, nextEntropy, op, oend,
+                                          strategy, countWorkspace,
+                                          entropyWorkspace, entropyWorkspaceSize)
+                       : ZSTD_buildDummySequencesStatistics(nextEntropy);
+    FORWARD_IF_ERROR(stats.size, "ZSTD_buildSequencesStatistics failed!");
+    fseMetadata->llType = (symbolEncodingType_e) stats.LLtype;
+    fseMetadata->ofType = (symbolEncodingType_e) stats.Offtype;
+    fseMetadata->mlType = (symbolEncodingType_e) stats.MLtype;
+    fseMetadata->lastCountSize = stats.lastCountSize;
+    return stats.size;
+}
+
+
+/** ZSTD_buildBlockEntropyStats() :
+ *  Builds entropy for the block.
+ *  Requires workspace size ENTROPY_WORKSPACE_SIZE
+ *
+ *  @return : 0 on success or error code
+ */
+size_t ZSTD_buildBlockEntropyStats(seqStore_t* seqStorePtr,
+                             const ZSTD_entropyCTables_t* prevEntropy,
+                                   ZSTD_entropyCTables_t* nextEntropy,
+                             const ZSTD_CCtx_params* cctxParams,
+                                   ZSTD_entropyCTablesMetadata_t* entropyMetadata,
+                                   void* workspace, size_t wkspSize)
+{
+    size_t const litSize = seqStorePtr->lit - seqStorePtr->litStart;
+    entropyMetadata->hufMetadata.hufDesSize =
+        ZSTD_buildBlockEntropyStats_literals(seqStorePtr->litStart, litSize,
+                                            &prevEntropy->huf, &nextEntropy->huf,
+                                            &entropyMetadata->hufMetadata,
+                                            ZSTD_disableLiteralsCompression(cctxParams),
+                                            workspace, wkspSize);
+    FORWARD_IF_ERROR(entropyMetadata->hufMetadata.hufDesSize, "ZSTD_buildBlockEntropyStats_literals failed");
+    entropyMetadata->fseMetadata.fseTablesSize =
+        ZSTD_buildBlockEntropyStats_sequences(seqStorePtr,
+                                              &prevEntropy->fse, &nextEntropy->fse,
+                                              cctxParams,
+                                              &entropyMetadata->fseMetadata,
+                                              workspace, wkspSize);
+    FORWARD_IF_ERROR(entropyMetadata->fseMetadata.fseTablesSize, "ZSTD_buildBlockEntropyStats_sequences failed");
+    return 0;
+}
+
+/* Returns the size estimate for the literals section (header + content) of a block */
+static size_t ZSTD_estimateBlockSize_literal(const BYTE* literals, size_t litSize,
+                                                const ZSTD_hufCTables_t* huf,
+                                                const ZSTD_hufCTablesMetadata_t* hufMetadata,
+                                                void* workspace, size_t wkspSize,
+                                                int writeEntropy)
+{
+    unsigned* const countWksp = (unsigned*)workspace;
+    unsigned maxSymbolValue = HUF_SYMBOLVALUE_MAX;
+    size_t literalSectionHeaderSize = 3 + (litSize >= 1 KB) + (litSize >= 16 KB);
+    U32 singleStream = litSize < 256;
+
+    if (hufMetadata->hType == set_basic) return litSize;
+    else if (hufMetadata->hType == set_rle) return 1;
+    else if (hufMetadata->hType == set_compressed || hufMetadata->hType == set_repeat) {
+        size_t const largest = HIST_count_wksp (countWksp, &maxSymbolValue, (const BYTE*)literals, litSize, workspace, wkspSize);
+        if (ZSTD_isError(largest)) return litSize;
+        {   size_t cLitSizeEstimate = HUF_estimateCompressedSize((const HUF_CElt*)huf->CTable, countWksp, maxSymbolValue);
+            if (writeEntropy) cLitSizeEstimate += hufMetadata->hufDesSize;
+            if (!singleStream) cLitSizeEstimate += 6; /* multi-stream huffman uses 6-byte jump table */
+            return cLitSizeEstimate + literalSectionHeaderSize;
+    }   }
+    assert(0); /* impossible */
+    return 0;
+}
+
+/* Returns the size estimate for the FSE-compressed symbols (of, ml, ll) of a block */
+static size_t ZSTD_estimateBlockSize_symbolType(symbolEncodingType_e type,
+                        const BYTE* codeTable, size_t nbSeq, unsigned maxCode,
+                        const FSE_CTable* fseCTable,
+                        const U32* additionalBits,
+                        short const* defaultNorm, U32 defaultNormLog, U32 defaultMax,
+                        void* workspace, size_t wkspSize)
+{
+    unsigned* const countWksp = (unsigned*)workspace;
+    const BYTE* ctp = codeTable;
+    const BYTE* const ctStart = ctp;
+    const BYTE* const ctEnd = ctStart + nbSeq;
+    size_t cSymbolTypeSizeEstimateInBits = 0;
+    unsigned max = maxCode;
+
+    HIST_countFast_wksp(countWksp, &max, codeTable, nbSeq, workspace, wkspSize);  /* can't fail */
+    if (type == set_basic) {
+        /* We selected this encoding type, so it must be valid. */
+        assert(max <= defaultMax);
+        (void)defaultMax;
+        cSymbolTypeSizeEstimateInBits = ZSTD_crossEntropyCost(defaultNorm, defaultNormLog, countWksp, max);
+    } else if (type == set_rle) {
+        cSymbolTypeSizeEstimateInBits = 0;
+    } else if (type == set_compressed || type == set_repeat) {
+        cSymbolTypeSizeEstimateInBits = ZSTD_fseBitCost(fseCTable, countWksp, max);
+    }
+    if (ZSTD_isError(cSymbolTypeSizeEstimateInBits)) {
+        return nbSeq * 10;
+    }
+    while (ctp < ctEnd) {
+        if (additionalBits) cSymbolTypeSizeEstimateInBits += additionalBits[*ctp];
+        else cSymbolTypeSizeEstimateInBits += *ctp; /* for offset, offset code is also the number of additional bits */
+        ctp++;
+    }
+    return cSymbolTypeSizeEstimateInBits >> 3;
+}
+
+/* Returns the size estimate for the sequences section (header + content) of a block */
+static size_t ZSTD_estimateBlockSize_sequences(const BYTE* ofCodeTable,
+                                                  const BYTE* llCodeTable,
+                                                  const BYTE* mlCodeTable,
+                                                  size_t nbSeq,
+                                                  const ZSTD_fseCTables_t* fseTables,
+                                                  const ZSTD_fseCTablesMetadata_t* fseMetadata,
+                                                  void* workspace, size_t wkspSize,
+                                                  int writeEntropy)
+{
+    size_t sequencesSectionHeaderSize = 1 /* seqHead */ + 1 /* min seqSize size */ + (nbSeq >= 128) + (nbSeq >= LONGNBSEQ);
+    size_t cSeqSizeEstimate = 0;
+    cSeqSizeEstimate += ZSTD_estimateBlockSize_symbolType(fseMetadata->ofType, ofCodeTable, nbSeq, MaxOff,
+                                         fseTables->offcodeCTable, NULL,
+                                         OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff,
+                                         workspace, wkspSize);
+    cSeqSizeEstimate += ZSTD_estimateBlockSize_symbolType(fseMetadata->llType, llCodeTable, nbSeq, MaxLL,
+                                         fseTables->litlengthCTable, LL_bits,
+                                         LL_defaultNorm, LL_defaultNormLog, MaxLL,
+                                         workspace, wkspSize);
+    cSeqSizeEstimate += ZSTD_estimateBlockSize_symbolType(fseMetadata->mlType, mlCodeTable, nbSeq, MaxML,
+                                         fseTables->matchlengthCTable, ML_bits,
+                                         ML_defaultNorm, ML_defaultNormLog, MaxML,
+                                         workspace, wkspSize);
+    if (writeEntropy) cSeqSizeEstimate += fseMetadata->fseTablesSize;
+    return cSeqSizeEstimate + sequencesSectionHeaderSize;
+}
+
+/* Returns the size estimate for a given stream of literals, of, ll, ml */
+static size_t ZSTD_estimateBlockSize(const BYTE* literals, size_t litSize,
+                                     const BYTE* ofCodeTable,
+                                     const BYTE* llCodeTable,
+                                     const BYTE* mlCodeTable,
+                                     size_t nbSeq,
+                                     const ZSTD_entropyCTables_t* entropy,
+                                     const ZSTD_entropyCTablesMetadata_t* entropyMetadata,
+                                     void* workspace, size_t wkspSize,
+                                     int writeLitEntropy, int writeSeqEntropy) {
+    size_t const literalsSize = ZSTD_estimateBlockSize_literal(literals, litSize,
+                                                         &entropy->huf, &entropyMetadata->hufMetadata,
+                                                         workspace, wkspSize, writeLitEntropy);
+    size_t const seqSize = ZSTD_estimateBlockSize_sequences(ofCodeTable, llCodeTable, mlCodeTable,
+                                                         nbSeq, &entropy->fse, &entropyMetadata->fseMetadata,
+                                                         workspace, wkspSize, writeSeqEntropy);
+    return seqSize + literalsSize + ZSTD_blockHeaderSize;
+}
+
+/* Builds entropy statistics and uses them for blocksize estimation.
+ *
+ * Returns the estimated compressed size of the seqStore, or a zstd error.
+ */
+static size_t ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(seqStore_t* seqStore, const ZSTD_CCtx* zc) {
+    ZSTD_entropyCTablesMetadata_t entropyMetadata;
+    DEBUGLOG(6, "ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize()");
+    FORWARD_IF_ERROR(ZSTD_buildBlockEntropyStats(seqStore,
+                    &zc->blockState.prevCBlock->entropy,
+                    &zc->blockState.nextCBlock->entropy,
+                    &zc->appliedParams,
+                    &entropyMetadata,
+                    zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */), "");
+    return ZSTD_estimateBlockSize(seqStore->litStart, (size_t)(seqStore->lit - seqStore->litStart),
+                    seqStore->ofCode, seqStore->llCode, seqStore->mlCode,
+                    (size_t)(seqStore->sequences - seqStore->sequencesStart),
+                    &zc->blockState.nextCBlock->entropy, &entropyMetadata, zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE,
+                    (int)(entropyMetadata.hufMetadata.hType == set_compressed), 1);
+}
+
+/* Returns literals bytes represented in a seqStore */
+static size_t ZSTD_countSeqStoreLiteralsBytes(const seqStore_t* const seqStore) {
+    size_t literalsBytes = 0;
+    size_t const nbSeqs = seqStore->sequences - seqStore->sequencesStart;
+    size_t i;
+    for (i = 0; i < nbSeqs; ++i) {
+        seqDef seq = seqStore->sequencesStart[i];
+        literalsBytes += seq.litLength;
+        if (i == seqStore->longLengthPos && seqStore->longLengthType == ZSTD_llt_literalLength) {
+            literalsBytes += 0x10000;
+        }
+    }
+    return literalsBytes;
+}
+
+/* Returns match bytes represented in a seqStore */
+static size_t ZSTD_countSeqStoreMatchBytes(const seqStore_t* const seqStore) {
+    size_t matchBytes = 0;
+    size_t const nbSeqs = seqStore->sequences - seqStore->sequencesStart;
+    size_t i;
+    for (i = 0; i < nbSeqs; ++i) {
+        seqDef seq = seqStore->sequencesStart[i];
+        matchBytes += seq.matchLength + MINMATCH;
+        if (i == seqStore->longLengthPos && seqStore->longLengthType == ZSTD_llt_matchLength) {
+            matchBytes += 0x10000;
+        }
+    }
+    return matchBytes;
+}
+
+/* Derives the seqStore that is a chunk of the originalSeqStore from [startIdx, endIdx).
+ * Stores the result in resultSeqStore.
+ */
+static void ZSTD_deriveSeqStoreChunk(seqStore_t* resultSeqStore,
+                               const seqStore_t* originalSeqStore,
+                                     size_t startIdx, size_t endIdx) {
+    BYTE* const litEnd = originalSeqStore->lit;
+    size_t literalsBytes;
+    size_t literalsBytesPreceding = 0;
+
+    *resultSeqStore = *originalSeqStore;
+    if (startIdx > 0) {
+        resultSeqStore->sequences = originalSeqStore->sequencesStart + startIdx;
+        literalsBytesPreceding = ZSTD_countSeqStoreLiteralsBytes(resultSeqStore);
+    }
+
+    /* Move longLengthPos into the correct position if necessary */
+    if (originalSeqStore->longLengthType != ZSTD_llt_none) {
+        if (originalSeqStore->longLengthPos < startIdx || originalSeqStore->longLengthPos > endIdx) {
+            resultSeqStore->longLengthType = ZSTD_llt_none;
+        } else {
+            resultSeqStore->longLengthPos -= (U32)startIdx;
+        }
+    }
+    resultSeqStore->sequencesStart = originalSeqStore->sequencesStart + startIdx;
+    resultSeqStore->sequences = originalSeqStore->sequencesStart + endIdx;
+    literalsBytes = ZSTD_countSeqStoreLiteralsBytes(resultSeqStore);
+    resultSeqStore->litStart += literalsBytesPreceding;
+    if (endIdx == (size_t)(originalSeqStore->sequences - originalSeqStore->sequencesStart)) {
+        /* This accounts for possible last literals if the derived chunk reaches the end of the block */
+        resultSeqStore->lit = litEnd;
+    } else {
+        resultSeqStore->lit = resultSeqStore->litStart+literalsBytes;
+    }
+    resultSeqStore->llCode += startIdx;
+    resultSeqStore->mlCode += startIdx;
+    resultSeqStore->ofCode += startIdx;
+}
+
+/**
+ * Returns the raw offset represented by the combination of offCode, ll0, and repcode history.
+ * offCode must be an offCode representing a repcode, therefore in the range of [0, 2].
+ */
+static U32 ZSTD_resolveRepcodeToRawOffset(const U32 rep[ZSTD_REP_NUM], const U32 offCode, const U32 ll0) {
+    U32 const adjustedOffCode = offCode + ll0;
+    assert(offCode < ZSTD_REP_NUM);
+    if (adjustedOffCode == ZSTD_REP_NUM) {
+        /* litlength == 0 and offCode == 2 implies selection of first repcode - 1 */
+        assert(rep[0] > 0);
+        return rep[0] - 1;
+    }
+    return rep[adjustedOffCode];
+}
+
+/**
+ * ZSTD_seqStore_resolveOffCodes() reconciles any possible divergences in offset history that may arise
+ * due to emission of RLE/raw blocks that disturb the offset history, and replaces any repcodes within
+ * the seqStore that may be invalid.
+ *
+ * dRepcodes are updated as would be on the decompression side. cRepcodes are updated exactly in
+ * accordance with the seqStore.
+ */
+static void ZSTD_seqStore_resolveOffCodes(repcodes_t* const dRepcodes, repcodes_t* const cRepcodes,
+                                          seqStore_t* const seqStore, U32 const nbSeq) {
+    U32 idx = 0;
+    for (; idx < nbSeq; ++idx) {
+        seqDef* const seq = seqStore->sequencesStart + idx;
+        U32 const ll0 = (seq->litLength == 0);
+        U32 offCode = seq->offset - 1;
+        assert(seq->offset > 0);
+        if (offCode <= ZSTD_REP_MOVE) {
+            U32 const dRawOffset = ZSTD_resolveRepcodeToRawOffset(dRepcodes->rep, offCode, ll0);
+            U32 const cRawOffset = ZSTD_resolveRepcodeToRawOffset(cRepcodes->rep, offCode, ll0);
+            /* Adjust simulated decompression repcode history if we come across a mismatch. Replace
+             * the repcode with the offset it actually references, determined by the compression
+             * repcode history.
+             */
+            if (dRawOffset != cRawOffset) {
+                seq->offset = cRawOffset + ZSTD_REP_NUM;
+            }
+        }
+        /* Compression repcode history is always updated with values directly from the unmodified seqStore.
+         * Decompression repcode history may use modified seq->offset value taken from compression repcode history.
+         */
+        *dRepcodes = ZSTD_updateRep(dRepcodes->rep, seq->offset - 1, ll0);
+        *cRepcodes = ZSTD_updateRep(cRepcodes->rep, offCode, ll0);
+    }
+}
+
+/* ZSTD_compressSeqStore_singleBlock():
+ * Compresses a seqStore into a block with a block header, into the buffer dst.
+ *
+ * Returns the total size of that block (including header) or a ZSTD error code.
+ */
+static size_t ZSTD_compressSeqStore_singleBlock(ZSTD_CCtx* zc, seqStore_t* const seqStore,
+                                                repcodes_t* const dRep, repcodes_t* const cRep,
+                                                void* dst, size_t dstCapacity,
+                                                const void* src, size_t srcSize,
+                                                U32 lastBlock, U32 isPartition) {
+    const U32 rleMaxLength = 25;
+    BYTE* op = (BYTE*)dst;
+    const BYTE* ip = (const BYTE*)src;
+    size_t cSize;
+    size_t cSeqsSize;
+
+    /* In case of an RLE or raw block, the simulated decompression repcode history must be reset */
+    repcodes_t const dRepOriginal = *dRep;
+    if (isPartition)
+        ZSTD_seqStore_resolveOffCodes(dRep, cRep, seqStore, (U32)(seqStore->sequences - seqStore->sequencesStart));
+
+    RETURN_ERROR_IF(dstCapacity < ZSTD_blockHeaderSize, dstSize_tooSmall, "Block header doesn't fit");
+    cSeqsSize = ZSTD_entropyCompressSeqStore(seqStore,
+                &zc->blockState.prevCBlock->entropy, &zc->blockState.nextCBlock->entropy,
+                &zc->appliedParams,
+                op + ZSTD_blockHeaderSize, dstCapacity - ZSTD_blockHeaderSize,
+                srcSize,
+                zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */,
+                zc->bmi2);
+    FORWARD_IF_ERROR(cSeqsSize, "ZSTD_entropyCompressSeqStore failed!");
+
+    if (!zc->isFirstBlock &&
+        cSeqsSize < rleMaxLength &&
+        ZSTD_isRLE((BYTE const*)src, srcSize)) {
+        /* We don't want to emit our first block as a RLE even if it qualifies because
+        * doing so will cause the decoder (cli only) to throw a "should consume all input error."
+        * This is only an issue for zstd <= v1.4.3
+        */
+        cSeqsSize = 1;
+    }
+
+    if (zc->seqCollector.collectSequences) {
+        ZSTD_copyBlockSequences(zc);
+        ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState);
+        return 0;
+    }
+
+    if (cSeqsSize == 0) {
+        cSize = ZSTD_noCompressBlock(op, dstCapacity, ip, srcSize, lastBlock);
+        FORWARD_IF_ERROR(cSize, "Nocompress block failed");
+        DEBUGLOG(4, "Writing out nocompress block, size: %zu", cSize);
+        *dRep = dRepOriginal; /* reset simulated decompression repcode history */
+    } else if (cSeqsSize == 1) {
+        cSize = ZSTD_rleCompressBlock(op, dstCapacity, *ip, srcSize, lastBlock);
+        FORWARD_IF_ERROR(cSize, "RLE compress block failed");
+        DEBUGLOG(4, "Writing out RLE block, size: %zu", cSize);
+        *dRep = dRepOriginal; /* reset simulated decompression repcode history */
+    } else {
+        ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState);
+        writeBlockHeader(op, cSeqsSize, srcSize, lastBlock);
+        cSize = ZSTD_blockHeaderSize + cSeqsSize;
+        DEBUGLOG(4, "Writing out compressed block, size: %zu", cSize);
+    }
+
+    if (zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode == FSE_repeat_valid)
+        zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode = FSE_repeat_check;
+
+    return cSize;
+}
+
+/* Struct to keep track of where we are in our recursive calls. */
+typedef struct {
+    U32* splitLocations;    /* Array of split indices */
+    size_t idx;             /* The current index within splitLocations being worked on */
+} seqStoreSplits;
+
+#define MIN_SEQUENCES_BLOCK_SPLITTING 300
+#define MAX_NB_SPLITS 196
+
+/* Helper function to perform the recursive search for block splits.
+ * Estimates the cost of seqStore prior to split, and estimates the cost of splitting the sequences in half.
+ * If advantageous to split, then we recurse down the two sub-blocks. If not, or if an error occurred in estimation, then
+ * we do not recurse.
+ *
+ * Note: The recursion depth is capped by a heuristic minimum number of sequences, defined by MIN_SEQUENCES_BLOCK_SPLITTING.
+ * In theory, this means the absolute largest recursion depth is 10 == log2(maxNbSeqInBlock/MIN_SEQUENCES_BLOCK_SPLITTING).
+ * In practice, recursion depth usually doesn't go beyond 4.
+ *
+ * Furthermore, the number of splits is capped by MAX_NB_SPLITS. At MAX_NB_SPLITS == 196 with the current existing blockSize
+ * maximum of 128 KB, this value is actually impossible to reach.
+ */
+static void ZSTD_deriveBlockSplitsHelper(seqStoreSplits* splits, size_t startIdx, size_t endIdx,
+                                         const ZSTD_CCtx* zc, const seqStore_t* origSeqStore) {
+    seqStore_t fullSeqStoreChunk;
+    seqStore_t firstHalfSeqStore;
+    seqStore_t secondHalfSeqStore;
+    size_t estimatedOriginalSize;
+    size_t estimatedFirstHalfSize;
+    size_t estimatedSecondHalfSize;
+    size_t midIdx = (startIdx + endIdx)/2;
+
+    if (endIdx - startIdx < MIN_SEQUENCES_BLOCK_SPLITTING || splits->idx >= MAX_NB_SPLITS) {
+        DEBUGLOG(6, "ZSTD_deriveBlockSplitsHelper: Too few sequences");
+        return;
+    }
+    ZSTD_deriveSeqStoreChunk(&fullSeqStoreChunk, origSeqStore, startIdx, endIdx);
+    ZSTD_deriveSeqStoreChunk(&firstHalfSeqStore, origSeqStore, startIdx, midIdx);
+    ZSTD_deriveSeqStoreChunk(&secondHalfSeqStore, origSeqStore, midIdx, endIdx);
+    estimatedOriginalSize = ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(&fullSeqStoreChunk, zc);
+    estimatedFirstHalfSize = ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(&firstHalfSeqStore, zc);
+    estimatedSecondHalfSize = ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(&secondHalfSeqStore, zc);
+    DEBUGLOG(5, "Estimated original block size: %zu -- First half split: %zu -- Second half split: %zu",
+             estimatedOriginalSize, estimatedFirstHalfSize, estimatedSecondHalfSize);
+    if (ZSTD_isError(estimatedOriginalSize) || ZSTD_isError(estimatedFirstHalfSize) || ZSTD_isError(estimatedSecondHalfSize)) {
+        return;
+    }
+    if (estimatedFirstHalfSize + estimatedSecondHalfSize < estimatedOriginalSize) {
+        ZSTD_deriveBlockSplitsHelper(splits, startIdx, midIdx, zc, origSeqStore);
+        splits->splitLocations[splits->idx] = (U32)midIdx;
+        splits->idx++;
+        ZSTD_deriveBlockSplitsHelper(splits, midIdx, endIdx, zc, origSeqStore);
+    }
+}
+
+/* Base recursive function. Populates a table with intra-block partition indices that can improve compression ratio.
+ *
+ * Returns the number of splits made (which equals the size of the partition table - 1).
+ */
+static size_t ZSTD_deriveBlockSplits(ZSTD_CCtx* zc, U32 partitions[], U32 nbSeq) {
+    seqStoreSplits splits = {partitions, 0};
+    if (nbSeq <= 4) {
+        DEBUGLOG(4, "ZSTD_deriveBlockSplits: Too few sequences to split");
+        /* Refuse to try and split anything with less than 4 sequences */
+        return 0;
+    }
+    ZSTD_deriveBlockSplitsHelper(&splits, 0, nbSeq, zc, &zc->seqStore);
+    splits.splitLocations[splits.idx] = nbSeq;
+    DEBUGLOG(5, "ZSTD_deriveBlockSplits: final nb partitions: %zu", splits.idx+1);
+    return splits.idx;
+}
+
+/* ZSTD_compressBlock_splitBlock():
+ * Attempts to split a given block into multiple blocks to improve compression ratio.
+ *
+ * Returns combined size of all blocks (which includes headers), or a ZSTD error code.
+ */
+static size_t ZSTD_compressBlock_splitBlock_internal(ZSTD_CCtx* zc, void* dst, size_t dstCapacity,
+                                                     const void* src, size_t blockSize, U32 lastBlock, U32 nbSeq) {
+    size_t cSize = 0;
+    const BYTE* ip = (const BYTE*)src;
+    BYTE* op = (BYTE*)dst;
+    U32 partitions[MAX_NB_SPLITS];
+    size_t i = 0;
+    size_t srcBytesTotal = 0;
+    size_t numSplits = ZSTD_deriveBlockSplits(zc, partitions, nbSeq);
+    seqStore_t nextSeqStore;
+    seqStore_t currSeqStore;
+
+    /* If a block is split and some partitions are emitted as RLE/uncompressed, then repcode history
+     * may become invalid. In order to reconcile potentially invalid repcodes, we keep track of two
+     * separate repcode histories that simulate repcode history on compression and decompression side,
+     * and use the histories to determine whether we must replace a particular repcode with its raw offset.
+     *
+     * 1) cRep gets updated for each partition, regardless of whether the block was emitted as uncompressed
+     *    or RLE. This allows us to retrieve the offset value that an invalid repcode references within
+     *    a nocompress/RLE block.
+     * 2) dRep gets updated only for compressed partitions, and when a repcode gets replaced, will use
+     *    the replacement offset value rather than the original repcode to update the repcode history.
+     *    dRep also will be the final repcode history sent to the next block.
+     *
+     * See ZSTD_seqStore_resolveOffCodes() for more details.
+     */
+    repcodes_t dRep;
+    repcodes_t cRep;
+    ZSTD_memcpy(dRep.rep, zc->blockState.prevCBlock->rep, sizeof(repcodes_t));
+    ZSTD_memcpy(cRep.rep, zc->blockState.prevCBlock->rep, sizeof(repcodes_t));
+    ZSTD_memset(&nextSeqStore, 0, sizeof(seqStore_t));
+
+    DEBUGLOG(4, "ZSTD_compressBlock_splitBlock_internal (dstCapacity=%u, dictLimit=%u, nextToUpdate=%u)",
+                (unsigned)dstCapacity, (unsigned)zc->blockState.matchState.window.dictLimit,
+                (unsigned)zc->blockState.matchState.nextToUpdate);
+
+    if (numSplits == 0) {
+        size_t cSizeSingleBlock = ZSTD_compressSeqStore_singleBlock(zc, &zc->seqStore,
+                                                                   &dRep, &cRep,
+                                                                    op, dstCapacity,
+                                                                    ip, blockSize,
+                                                                    lastBlock, 0 /* isPartition */);
+        FORWARD_IF_ERROR(cSizeSingleBlock, "Compressing single block from splitBlock_internal() failed!");
+        DEBUGLOG(5, "ZSTD_compressBlock_splitBlock_internal: No splits");
+        assert(cSizeSingleBlock <= ZSTD_BLOCKSIZE_MAX + ZSTD_blockHeaderSize);
+        return cSizeSingleBlock;
+    }
+
+    ZSTD_deriveSeqStoreChunk(&currSeqStore, &zc->seqStore, 0, partitions[0]);
+    for (i = 0; i <= numSplits; ++i) {
+        size_t srcBytes;
+        size_t cSizeChunk;
+        U32 const lastPartition = (i == numSplits);
+        U32 lastBlockEntireSrc = 0;
+
+        srcBytes = ZSTD_countSeqStoreLiteralsBytes(&currSeqStore) + ZSTD_countSeqStoreMatchBytes(&currSeqStore);
+        srcBytesTotal += srcBytes;
+        if (lastPartition) {
+            /* This is the final partition, need to account for possible last literals */
+            srcBytes += blockSize - srcBytesTotal;
+            lastBlockEntireSrc = lastBlock;
+        } else {
+            ZSTD_deriveSeqStoreChunk(&nextSeqStore, &zc->seqStore, partitions[i], partitions[i+1]);
+        }
+
+        cSizeChunk = ZSTD_compressSeqStore_singleBlock(zc, &currSeqStore,
+                                                      &dRep, &cRep,
+                                                       op, dstCapacity,
+                                                       ip, srcBytes,
+                                                       lastBlockEntireSrc, 1 /* isPartition */);
+        DEBUGLOG(5, "Estimated size: %zu actual size: %zu", ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(&currSeqStore, zc), cSizeChunk);
+        FORWARD_IF_ERROR(cSizeChunk, "Compressing chunk failed!");
+
+        ip += srcBytes;
+        op += cSizeChunk;
+        dstCapacity -= cSizeChunk;
+        cSize += cSizeChunk;
+        currSeqStore = nextSeqStore;
+        assert(cSizeChunk <= ZSTD_BLOCKSIZE_MAX + ZSTD_blockHeaderSize);
+    }
+    /* cRep and dRep may have diverged during the compression. If so, we use the dRep repcodes
+     * for the next block.
+     */
+    ZSTD_memcpy(zc->blockState.prevCBlock->rep, dRep.rep, sizeof(repcodes_t));
+    return cSize;
 }
 
+static size_t ZSTD_compressBlock_splitBlock(ZSTD_CCtx* zc,
+                                        void* dst, size_t dstCapacity,
+                                        const void* src, size_t srcSize, U32 lastBlock) {
+    const BYTE* ip = (const BYTE*)src;
+    BYTE* op = (BYTE*)dst;
+    U32 nbSeq;
+    size_t cSize;
+    DEBUGLOG(4, "ZSTD_compressBlock_splitBlock");
+
+    {   const size_t bss = ZSTD_buildSeqStore(zc, src, srcSize);
+        FORWARD_IF_ERROR(bss, "ZSTD_buildSeqStore failed");
+        if (bss == ZSTDbss_noCompress) {
+            if (zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode == FSE_repeat_valid)
+                zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode = FSE_repeat_check;
+            cSize = ZSTD_noCompressBlock(op, dstCapacity, ip, srcSize, lastBlock);
+            FORWARD_IF_ERROR(cSize, "ZSTD_noCompressBlock failed");
+            DEBUGLOG(4, "ZSTD_compressBlock_splitBlock: Nocompress block");
+            return cSize;
+        }
+        nbSeq = (U32)(zc->seqStore.sequences - zc->seqStore.sequencesStart);
+    }
+
+    assert(zc->appliedParams.splitBlocks == 1);
+    cSize = ZSTD_compressBlock_splitBlock_internal(zc, dst, dstCapacity, src, srcSize, lastBlock, nbSeq);
+    FORWARD_IF_ERROR(cSize, "Splitting blocks failed!");
+    return cSize;
+}
+
+/* ZSTD_convertBlockSequencesToSeqStore()
+ * Converts an array of ZSTD_Sequence* with the corresponding original src buffer into
+ * the seqStore of a cctx.
+ *
+ * Returns 0 on success, ZSTD_error on failure.
+ */
+static UNUSED_ATTR size_t ZSTD_convertBlockSequencesToSeqStore(ZSTD_CCtx* cctx,
+                                  const ZSTD_Sequence* inSeqs, size_t inSeqsSize,
+                                  const void* src, size_t srcSize);
+
 static size_t ZSTD_compressBlock_internal(ZSTD_CCtx* zc,
                                         void* dst, size_t dstCapacity,
                                         const void* src, size_t srcSize, U32 frame)
@@ -2683,12 +3776,12 @@ static size_t ZSTD_compressBlock_internal(ZSTD_CCtx* zc,
 
     if (zc->seqCollector.collectSequences) {
         ZSTD_copyBlockSequences(zc);
-        ZSTD_confirmRepcodesAndEntropyTables(zc);
+        ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState);
         return 0;
     }
 
     /* encode sequences and literals */
-    cSize = ZSTD_entropyCompressSequences(&zc->seqStore,
+    cSize = ZSTD_entropyCompressSeqStore(&zc->seqStore,
             &zc->blockState.prevCBlock->entropy, &zc->blockState.nextCBlock->entropy,
             &zc->appliedParams,
             dst, dstCapacity,
@@ -2696,12 +3789,6 @@ static size_t ZSTD_compressBlock_internal(ZSTD_CCtx* zc,
             zc->entropyWorkspace, ENTROPY_WORKSPACE_SIZE /* statically allocated in resetCCtx */,
             zc->bmi2);
 
-    if (zc->seqCollector.collectSequences) {
-        ZSTD_copyBlockSequences(zc);
-        return 0;
-    }
-
-
     if (frame &&
         /* We don't want to emit our first block as a RLE even if it qualifies because
          * doing so will cause the decoder (cli only) to throw a "should consume all input error."
@@ -2717,7 +3804,7 @@ static size_t ZSTD_compressBlock_internal(ZSTD_CCtx* zc,
 
 out:
     if (!ZSTD_isError(cSize) && cSize > 1) {
-        ZSTD_confirmRepcodesAndEntropyTables(zc);
+        ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState);
     }
     /* We check that dictionaries have offset codes available for the first
      * block. After the first block, the offcode table might not have large
@@ -2770,7 +3857,7 @@ static size_t ZSTD_compressBlock_targetCBlockSize_body(ZSTD_CCtx* zc,
                 size_t const maxCSize = srcSize - ZSTD_minGain(srcSize, zc->appliedParams.cParams.strategy);
                 FORWARD_IF_ERROR(cSize, "ZSTD_compressSuperBlock failed");
                 if (cSize != 0 && cSize < maxCSize + ZSTD_blockHeaderSize) {
-                    ZSTD_confirmRepcodesAndEntropyTables(zc);
+                    ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState);
                     return cSize;
                 }
             }
@@ -2810,9 +3897,9 @@ static void ZSTD_overflowCorrectIfNeeded(ZSTD_matchState_t* ms,
                                          void const* ip,
                                          void const* iend)
 {
-    if (ZSTD_window_needOverflowCorrection(ms->window, iend)) {
-        U32 const maxDist = (U32)1 << params->cParams.windowLog;
-        U32 const cycleLog = ZSTD_cycleLog(params->cParams.chainLog, params->cParams.strategy);
+    U32 const cycleLog = ZSTD_cycleLog(params->cParams.chainLog, params->cParams.strategy);
+    U32 const maxDist = (U32)1 << params->cParams.windowLog;
+    if (ZSTD_window_needOverflowCorrection(ms->window, cycleLog, maxDist, ms->loadedDictEnd, ip, iend)) {
         U32 const correction = ZSTD_window_correctOverflow(&ms->window, cycleLog, maxDist, ip);
         ZSTD_STATIC_ASSERT(ZSTD_CHAINLOG_MAX <= 30);
         ZSTD_STATIC_ASSERT(ZSTD_WINDOWLOG_MAX_32 <= 30);
@@ -2835,7 +3922,7 @@ static void ZSTD_overflowCorrectIfNeeded(ZSTD_matchState_t* ms,
 *   Frame is supposed already started (header already produced)
 *   @return : compressed size, or an error code
 */
-static size_t ZSTD_compress_frameChunk (ZSTD_CCtx* cctx,
+static size_t ZSTD_compress_frameChunk(ZSTD_CCtx* cctx,
                                      void* dst, size_t dstCapacity,
                                const void* src, size_t srcSize,
                                      U32 lastFrameChunk)
@@ -2865,6 +3952,7 @@ static size_t ZSTD_compress_frameChunk (ZSTD_CCtx* cctx,
         ZSTD_overflowCorrectIfNeeded(
             ms, &cctx->workspace, &cctx->appliedParams, ip, ip + blockSize);
         ZSTD_checkDictValidity(&ms->window, ip + blockSize, maxDist, &ms->loadedDictEnd, &ms->dictMatchState);
+        ZSTD_window_enforceMaxDist(&ms->window, ip, maxDist, &ms->loadedDictEnd, &ms->dictMatchState);
 
         /* Ensure hash/chain table insertion resumes no sooner than lowlimit */
         if (ms->nextToUpdate < ms->window.lowLimit) ms->nextToUpdate = ms->window.lowLimit;
@@ -2875,6 +3963,10 @@ static size_t ZSTD_compress_frameChunk (ZSTD_CCtx* cctx,
                 FORWARD_IF_ERROR(cSize, "ZSTD_compressBlock_targetCBlockSize failed");
                 assert(cSize > 0);
                 assert(cSize <= blockSize + ZSTD_blockHeaderSize);
+            } else if (ZSTD_blockSplitterEnabled(&cctx->appliedParams)) {
+                cSize = ZSTD_compressBlock_splitBlock(cctx, op, dstCapacity, ip, blockSize, lastBlock);
+                FORWARD_IF_ERROR(cSize, "ZSTD_compressBlock_splitBlock failed");
+                assert(cSize > 0 || cctx->seqCollector.collectSequences == 1);
             } else {
                 cSize = ZSTD_compressBlock_internal(cctx,
                                         op+ZSTD_blockHeaderSize, dstCapacity-ZSTD_blockHeaderSize,
@@ -2954,6 +4046,26 @@ static size_t ZSTD_writeFrameHeader(void* dst, size_t dstCapacity,
     return pos;
 }
 
+/* ZSTD_writeSkippableFrame_advanced() :
+ * Writes out a skippable frame with the specified magic number variant (16 are supported),
+ * from ZSTD_MAGIC_SKIPPABLE_START to ZSTD_MAGIC_SKIPPABLE_START+15, and the desired source data.
+ *
+ * Returns the total number of bytes written, or a ZSTD error code.
+ */
+size_t ZSTD_writeSkippableFrame(void* dst, size_t dstCapacity,
+                                const void* src, size_t srcSize, unsigned magicVariant) {
+    BYTE* op = (BYTE*)dst;
+    RETURN_ERROR_IF(dstCapacity < srcSize + ZSTD_SKIPPABLEHEADERSIZE /* Skippable frame overhead */,
+                    dstSize_tooSmall, "Not enough room for skippable frame");
+    RETURN_ERROR_IF(srcSize > (unsigned)0xFFFFFFFF, srcSize_wrong, "Src size too large for skippable frame");
+    RETURN_ERROR_IF(magicVariant > 15, parameter_outOfBound, "Skippable frame magic number variant not supported");
+
+    MEM_writeLE32(op, (U32)(ZSTD_MAGIC_SKIPPABLE_START + magicVariant));
+    MEM_writeLE32(op+4, (U32)srcSize);
+    ZSTD_memcpy(op+8, src, srcSize);
+    return srcSize + ZSTD_SKIPPABLEHEADERSIZE;
+}
+
 /* ZSTD_writeLastEmptyBlock() :
  * output an empty Block with end-of-frame mark to complete a frame
  * @return : size of data written into `dst` (== ZSTD_blockHeaderSize (defined in zstd_internal.h))
@@ -3010,11 +4122,12 @@ static size_t ZSTD_compressContinue_internal (ZSTD_CCtx* cctx,
 
     if (!srcSize) return fhSize;  /* do not generate an empty block if no input */
 
-    if (!ZSTD_window_update(&ms->window, src, srcSize)) {
+    if (!ZSTD_window_update(&ms->window, src, srcSize, ms->forceNonContiguous)) {
+        ms->forceNonContiguous = 0;
         ms->nextToUpdate = ms->window.dictLimit;
     }
     if (cctx->appliedParams.ldmParams.enableLdm) {
-        ZSTD_window_update(&cctx->ldmState.window, src, srcSize);
+        ZSTD_window_update(&cctx->ldmState.window, src, srcSize, /* forceNonContiguous */ 0);
     }
 
     if (!frame) {
@@ -3082,63 +4195,86 @@ static size_t ZSTD_loadDictionaryContent(ZSTD_matchState_t* ms,
 {
     const BYTE* ip = (const BYTE*) src;
     const BYTE* const iend = ip + srcSize;
+    int const loadLdmDict = params->ldmParams.enableLdm && ls != NULL;
+
+    /* Assert that we the ms params match the params we're being given */
+    ZSTD_assertEqualCParams(params->cParams, ms->cParams);
+
+    if (srcSize > ZSTD_CHUNKSIZE_MAX) {
+        /* Allow the dictionary to set indices up to exactly ZSTD_CURRENT_MAX.
+         * Dictionaries right at the edge will immediately trigger overflow
+         * correction, but I don't want to insert extra constraints here.
+         */
+        U32 const maxDictSize = ZSTD_CURRENT_MAX - 1;
+        /* We must have cleared our windows when our source is this large. */
+        assert(ZSTD_window_isEmpty(ms->window));
+        if (loadLdmDict)
+            assert(ZSTD_window_isEmpty(ls->window));
+        /* If the dictionary is too large, only load the suffix of the dictionary. */
+        if (srcSize > maxDictSize) {
+            ip = iend - maxDictSize;
+            src = ip;
+            srcSize = maxDictSize;
+        }
+    }
 
-    ZSTD_window_update(&ms->window, src, srcSize);
+    DEBUGLOG(4, "ZSTD_loadDictionaryContent(): useRowMatchFinder=%d", (int)params->useRowMatchFinder);
+    ZSTD_window_update(&ms->window, src, srcSize, /* forceNonContiguous */ 0);
     ms->loadedDictEnd = params->forceWindow ? 0 : (U32)(iend - ms->window.base);
+    ms->forceNonContiguous = params->deterministicRefPrefix;
 
-    if (params->ldmParams.enableLdm && ls != NULL) {
-        ZSTD_window_update(&ls->window, src, srcSize);
+    if (loadLdmDict) {
+        ZSTD_window_update(&ls->window, src, srcSize, /* forceNonContiguous */ 0);
         ls->loadedDictEnd = params->forceWindow ? 0 : (U32)(iend - ls->window.base);
     }
 
-    /* Assert that we the ms params match the params we're being given */
-    ZSTD_assertEqualCParams(params->cParams, ms->cParams);
-
     if (srcSize <= HASH_READ_SIZE) return 0;
 
-    while (iend - ip > HASH_READ_SIZE) {
-        size_t const remaining = (size_t)(iend - ip);
-        size_t const chunk = MIN(remaining, ZSTD_CHUNKSIZE_MAX);
-        const BYTE* const ichunk = ip + chunk;
+    ZSTD_overflowCorrectIfNeeded(ms, ws, params, ip, iend);
 
-        ZSTD_overflowCorrectIfNeeded(ms, ws, params, ip, ichunk);
+    if (loadLdmDict)
+        ZSTD_ldm_fillHashTable(ls, ip, iend, &params->ldmParams);
 
-        if (params->ldmParams.enableLdm && ls != NULL)
-            ZSTD_ldm_fillHashTable(ls, (const BYTE*)src, (const BYTE*)src + srcSize, &params->ldmParams);
-
-        switch(params->cParams.strategy)
-        {
-        case ZSTD_fast:
-            ZSTD_fillHashTable(ms, ichunk, dtlm);
-            break;
-        case ZSTD_dfast:
-            ZSTD_fillDoubleHashTable(ms, ichunk, dtlm);
-            break;
+    switch(params->cParams.strategy)
+    {
+    case ZSTD_fast:
+        ZSTD_fillHashTable(ms, iend, dtlm);
+        break;
+    case ZSTD_dfast:
+        ZSTD_fillDoubleHashTable(ms, iend, dtlm);
+        break;
 
-        case ZSTD_greedy:
-        case ZSTD_lazy:
-        case ZSTD_lazy2:
-            if (chunk >= HASH_READ_SIZE && ms->dedicatedDictSearch) {
-                assert(chunk == remaining); /* must load everything in one go */
-                ZSTD_dedicatedDictSearch_lazy_loadDictionary(ms, ichunk-HASH_READ_SIZE);
-            } else if (chunk >= HASH_READ_SIZE) {
-                ZSTD_insertAndFindFirstIndex(ms, ichunk-HASH_READ_SIZE);
+    case ZSTD_greedy:
+    case ZSTD_lazy:
+    case ZSTD_lazy2:
+        assert(srcSize >= HASH_READ_SIZE);
+        if (ms->dedicatedDictSearch) {
+            assert(ms->chainTable != NULL);
+            ZSTD_dedicatedDictSearch_lazy_loadDictionary(ms, iend-HASH_READ_SIZE);
+        } else {
+            assert(params->useRowMatchFinder != ZSTD_urm_auto);
+            if (params->useRowMatchFinder == ZSTD_urm_enableRowMatchFinder) {
+                size_t const tagTableSize = ((size_t)1 << params->cParams.hashLog) * sizeof(U16);
+                ZSTD_memset(ms->tagTable, 0, tagTableSize);
+                ZSTD_row_update(ms, iend-HASH_READ_SIZE);
+                DEBUGLOG(4, "Using row-based hash table for lazy dict");
+            } else {
+                ZSTD_insertAndFindFirstIndex(ms, iend-HASH_READ_SIZE);
+                DEBUGLOG(4, "Using chain-based hash table for lazy dict");
             }
-            break;
-
-        case ZSTD_btlazy2:   /* we want the dictionary table fully sorted */
-        case ZSTD_btopt:
-        case ZSTD_btultra:
-        case ZSTD_btultra2:
-            if (chunk >= HASH_READ_SIZE)
-                ZSTD_updateTree(ms, ichunk-HASH_READ_SIZE, ichunk);
-            break;
-
-        default:
-            assert(0);  /* not possible : not a valid strategy id */
         }
+        break;
 
-        ip = ichunk;
+    case ZSTD_btlazy2:   /* we want the dictionary table fully sorted */
+    case ZSTD_btopt:
+    case ZSTD_btultra:
+    case ZSTD_btultra2:
+        assert(srcSize >= HASH_READ_SIZE);
+        ZSTD_updateTree(ms, iend-HASH_READ_SIZE, iend);
+        break;
+
+    default:
+        assert(0);  /* not possible : not a valid strategy id */
     }
 
     ms->nextToUpdate = (U32)(iend - ms->window.base);
@@ -3277,7 +4413,6 @@ static size_t ZSTD_loadZstdDictionary(ZSTD_compressedBlockState_t* bs,
     const BYTE* const dictEnd = dictPtr + dictSize;
     size_t dictID;
     size_t eSize;
-
     ZSTD_STATIC_ASSERT(HUF_WORKSPACE_SIZE >= (1<<MAX(MLFSELog,LLFSELog)));
     assert(dictSize >= 8);
     assert(MEM_readLE32(dictPtr) == ZSTD_MAGIC_DICTIONARY);
@@ -3348,6 +4483,10 @@ static size_t ZSTD_compressBegin_internal(ZSTD_CCtx* cctx,
                                     const ZSTD_CCtx_params* params, U64 pledgedSrcSize,
                                     ZSTD_buffered_policy_e zbuff)
 {
+    size_t const dictContentSize = cdict ? cdict->dictContentSize : dictSize;
+#if ZSTD_TRACE
+    cctx->traceCtx = (ZSTD_trace_compress_begin != NULL) ? ZSTD_trace_compress_begin(cctx) : 0;
+#endif
     DEBUGLOG(4, "ZSTD_compressBegin_internal: wlog=%u", params->cParams.windowLog);
     /* params are supposed to be fully validated at this point */
     assert(!ZSTD_isError(ZSTD_checkCParams(params->cParams)));
@@ -3362,7 +4501,8 @@ static size_t ZSTD_compressBegin_internal(ZSTD_CCtx* cctx,
         return ZSTD_resetCCtx_usingCDict(cctx, cdict, params, pledgedSrcSize, zbuff);
     }
 
-    FORWARD_IF_ERROR( ZSTD_resetCCtx_internal(cctx, *params, pledgedSrcSize,
+    FORWARD_IF_ERROR( ZSTD_resetCCtx_internal(cctx, params, pledgedSrcSize,
+                                     dictContentSize,
                                      ZSTDcrp_makeClean, zbuff) , "");
     {   size_t const dictID = cdict ?
                 ZSTD_compress_insertDictionary(
@@ -3377,6 +4517,7 @@ static size_t ZSTD_compressBegin_internal(ZSTD_CCtx* cctx,
         FORWARD_IF_ERROR(dictID, "ZSTD_compress_insertDictionary failed");
         assert(dictID <= UINT_MAX);
         cctx->dictID = (U32)dictID;
+        cctx->dictContentSize = dictContentSize;
     }
     return 0;
 }
@@ -3405,8 +4546,8 @@ size_t ZSTD_compressBegin_advanced(ZSTD_CCtx* cctx,
                              const void* dict, size_t dictSize,
                                    ZSTD_parameters params, unsigned long long pledgedSrcSize)
 {
-    ZSTD_CCtx_params const cctxParams =
-            ZSTD_assignParamsToCCtxParams(&cctx->requestedParams, &params);
+    ZSTD_CCtx_params cctxParams;
+    ZSTD_CCtxParams_init_internal(&cctxParams, &params, ZSTD_NO_CLEVEL);
     return ZSTD_compressBegin_advanced_internal(cctx,
                                             dict, dictSize, ZSTD_dct_auto, ZSTD_dtlm_fast,
                                             NULL /*cdict*/,
@@ -3415,9 +4556,11 @@ size_t ZSTD_compressBegin_advanced(ZSTD_CCtx* cctx,
 
 size_t ZSTD_compressBegin_usingDict(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, int compressionLevel)
 {
-    ZSTD_parameters const params = ZSTD_getParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_noAttachDict);
-    ZSTD_CCtx_params const cctxParams =
-            ZSTD_assignParamsToCCtxParams(&cctx->requestedParams, &params);
+    ZSTD_CCtx_params cctxParams;
+    {
+        ZSTD_parameters const params = ZSTD_getParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_noAttachDict);
+        ZSTD_CCtxParams_init_internal(&cctxParams, &params, (compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT : compressionLevel);
+    }
     DEBUGLOG(4, "ZSTD_compressBegin_usingDict (dictSize=%u)", (unsigned)dictSize);
     return ZSTD_compressBegin_internal(cctx, dict, dictSize, ZSTD_dct_auto, ZSTD_dtlm_fast, NULL,
                                        &cctxParams, ZSTD_CONTENTSIZE_UNKNOWN, ZSTDb_not_buffered);
@@ -3471,6 +4614,30 @@ static size_t ZSTD_writeEpilogue(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity)
     return op-ostart;
 }
 
+void ZSTD_CCtx_trace(ZSTD_CCtx* cctx, size_t extraCSize)
+{
+#if ZSTD_TRACE
+    if (cctx->traceCtx && ZSTD_trace_compress_end != NULL) {
+        int const streaming = cctx->inBuffSize > 0 || cctx->outBuffSize > 0 || cctx->appliedParams.nbWorkers > 0;
+        ZSTD_Trace trace;
+        ZSTD_memset(&trace, 0, sizeof(trace));
+        trace.version = ZSTD_VERSION_NUMBER;
+        trace.streaming = streaming;
+        trace.dictionaryID = cctx->dictID;
+        trace.dictionarySize = cctx->dictContentSize;
+        trace.uncompressedSize = cctx->consumedSrcSize;
+        trace.compressedSize = cctx->producedCSize + extraCSize;
+        trace.params = &cctx->appliedParams;
+        trace.cctx = cctx;
+        ZSTD_trace_compress_end(cctx->traceCtx, &trace);
+    }
+    cctx->traceCtx = 0;
+#else
+    (void)cctx;
+    (void)extraCSize;
+#endif
+}
+
 size_t ZSTD_compressEnd (ZSTD_CCtx* cctx,
                          void* dst, size_t dstCapacity,
                    const void* src, size_t srcSize)
@@ -3493,25 +4660,10 @@ size_t ZSTD_compressEnd (ZSTD_CCtx* cctx,
             (unsigned)cctx->pledgedSrcSizePlusOne-1,
             (unsigned)cctx->consumedSrcSize);
     }
+    ZSTD_CCtx_trace(cctx, endResult);
     return cSize + endResult;
 }
 
-static size_t ZSTD_compress_internal (ZSTD_CCtx* cctx,
-                                      void* dst, size_t dstCapacity,
-                                const void* src, size_t srcSize,
-                                const void* dict,size_t dictSize,
-                                const ZSTD_parameters* params)
-{
-    ZSTD_CCtx_params const cctxParams =
-            ZSTD_assignParamsToCCtxParams(&cctx->requestedParams, params);
-    DEBUGLOG(4, "ZSTD_compress_internal");
-    return ZSTD_compress_advanced_internal(cctx,
-                                           dst, dstCapacity,
-                                           src, srcSize,
-                                           dict, dictSize,
-                                           &cctxParams);
-}
-
 size_t ZSTD_compress_advanced (ZSTD_CCtx* cctx,
                                void* dst, size_t dstCapacity,
                          const void* src, size_t srcSize,
@@ -3520,11 +4672,12 @@ size_t ZSTD_compress_advanced (ZSTD_CCtx* cctx,
 {
     DEBUGLOG(4, "ZSTD_compress_advanced");
     FORWARD_IF_ERROR(ZSTD_checkCParams(params.cParams), "");
-    return ZSTD_compress_internal(cctx,
-                                  dst, dstCapacity,
-                                  src, srcSize,
-                                  dict, dictSize,
-                                  &params);
+    ZSTD_CCtxParams_init_internal(&cctx->simpleApiParams, &params, ZSTD_NO_CLEVEL);
+    return ZSTD_compress_advanced_internal(cctx,
+                                           dst, dstCapacity,
+                                           src, srcSize,
+                                           dict, dictSize,
+                                           &cctx->simpleApiParams);
 }
 
 /* Internal */
@@ -3548,11 +4701,13 @@ size_t ZSTD_compress_usingDict(ZSTD_CCtx* cctx,
                          const void* dict, size_t dictSize,
                                int compressionLevel)
 {
-    ZSTD_parameters const params = ZSTD_getParams_internal(compressionLevel, srcSize, dict ? dictSize : 0, ZSTD_cpm_noAttachDict);
-    ZSTD_CCtx_params cctxParams = ZSTD_assignParamsToCCtxParams(&cctx->requestedParams, &params);
+    {
+        ZSTD_parameters const params = ZSTD_getParams_internal(compressionLevel, srcSize, dict ? dictSize : 0, ZSTD_cpm_noAttachDict);
+        assert(params.fParams.contentSizeFlag == 1);
+        ZSTD_CCtxParams_init_internal(&cctx->simpleApiParams, &params, (compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT: compressionLevel);
+    }
     DEBUGLOG(4, "ZSTD_compress_usingDict (srcSize=%u)", (unsigned)srcSize);
-    assert(params.fParams.contentSizeFlag == 1);
-    return ZSTD_compress_advanced_internal(cctx, dst, dstCapacity, src, srcSize, dict, dictSize, &cctxParams);
+    return ZSTD_compress_advanced_internal(cctx, dst, dstCapacity, src, srcSize, dict, dictSize, &cctx->simpleApiParams);
 }
 
 size_t ZSTD_compressCCtx(ZSTD_CCtx* cctx,
@@ -3596,7 +4751,10 @@ size_t ZSTD_estimateCDictSize_advanced(
     DEBUGLOG(5, "sizeof(ZSTD_CDict) : %u", (unsigned)sizeof(ZSTD_CDict));
     return ZSTD_cwksp_alloc_size(sizeof(ZSTD_CDict))
          + ZSTD_cwksp_alloc_size(HUF_WORKSPACE_SIZE)
-         + ZSTD_sizeof_matchState(&cParams, /* forCCtx */ 0)
+         /* enableDedicatedDictSearch == 1 ensures that CDict estimation will not be too small
+          * in case we are using DDS with row-hash. */
+         + ZSTD_sizeof_matchState(&cParams, ZSTD_resolveRowMatchFinderMode(ZSTD_urm_auto, &cParams),
+                                  /* enableDedicatedDictSearch */ 1, /* forCCtx */ 0)
          + (dictLoadMethod == ZSTD_dlm_byRef ? 0
             : ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(dictSize, sizeof(void *))));
 }
@@ -3627,9 +4785,6 @@ static size_t ZSTD_initCDict_internal(
     assert(!ZSTD_checkCParams(params.cParams));
     cdict->matchState.cParams = params.cParams;
     cdict->matchState.dedicatedDictSearch = params.enableDedicatedDictSearch;
-    if (cdict->matchState.dedicatedDictSearch && dictSize > ZSTD_CHUNKSIZE_MAX) {
-        cdict->matchState.dedicatedDictSearch = 0;
-    }
     if ((dictLoadMethod == ZSTD_dlm_byRef) || (!dictBuffer) || (!dictSize)) {
         cdict->dictContent = dictBuffer;
     } else {
@@ -3650,6 +4805,7 @@ static size_t ZSTD_initCDict_internal(
         &cdict->matchState,
         &cdict->workspace,
         &params.cParams,
+        params.useRowMatchFinder,
         ZSTDcrp_makeClean,
         ZSTDirp_reset,
         ZSTD_resetTarget_CDict), "");
@@ -3673,14 +4829,17 @@ static size_t ZSTD_initCDict_internal(
 
 static ZSTD_CDict* ZSTD_createCDict_advanced_internal(size_t dictSize,
                                       ZSTD_dictLoadMethod_e dictLoadMethod,
-                                      ZSTD_compressionParameters cParams, ZSTD_customMem customMem)
+                                      ZSTD_compressionParameters cParams,
+                                      ZSTD_useRowMatchFinderMode_e useRowMatchFinder,
+                                      U32 enableDedicatedDictSearch,
+                                      ZSTD_customMem customMem)
 {
     if ((!customMem.customAlloc) ^ (!customMem.customFree)) return NULL;
 
     {   size_t const workspaceSize =
             ZSTD_cwksp_alloc_size(sizeof(ZSTD_CDict)) +
             ZSTD_cwksp_alloc_size(HUF_WORKSPACE_SIZE) +
-            ZSTD_sizeof_matchState(&cParams, /* forCCtx */ 0) +
+            ZSTD_sizeof_matchState(&cParams, useRowMatchFinder, enableDedicatedDictSearch, /* forCCtx */ 0) +
             (dictLoadMethod == ZSTD_dlm_byRef ? 0
              : ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(dictSize, sizeof(void*))));
         void* const workspace = ZSTD_customMalloc(workspaceSize, customMem);
@@ -3698,8 +4857,8 @@ static ZSTD_CDict* ZSTD_createCDict_advanced_internal(size_t dictSize,
         assert(cdict != NULL);
         ZSTD_cwksp_move(&cdict->workspace, &ws);
         cdict->customMem = customMem;
-        cdict->compressionLevel = 0; /* signals advanced API usage */
-
+        cdict->compressionLevel = ZSTD_NO_CLEVEL; /* signals advanced API usage */
+        cdict->useRowMatchFinder = useRowMatchFinder;
         return cdict;
     }
 }
@@ -3751,10 +4910,13 @@ ZSTDLIB_API ZSTD_CDict* ZSTD_createCDict_advanced2(
             &cctxParams, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_createCDict);
     }
 
+    DEBUGLOG(3, "ZSTD_createCDict_advanced2: DDS: %u", cctxParams.enableDedicatedDictSearch);
     cctxParams.cParams = cParams;
+    cctxParams.useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(cctxParams.useRowMatchFinder, &cParams);
 
     cdict = ZSTD_createCDict_advanced_internal(dictSize,
                         dictLoadMethod, cctxParams.cParams,
+                        cctxParams.useRowMatchFinder, cctxParams.enableDedicatedDictSearch,
                         customMem);
 
     if (ZSTD_isError( ZSTD_initCDict_internal(cdict,
@@ -3823,7 +4985,9 @@ const ZSTD_CDict* ZSTD_initStaticCDict(
                                  ZSTD_dictContentType_e dictContentType,
                                  ZSTD_compressionParameters cParams)
 {
-    size_t const matchStateSize = ZSTD_sizeof_matchState(&cParams, /* forCCtx */ 0);
+    ZSTD_useRowMatchFinderMode_e const useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(ZSTD_urm_auto, &cParams);
+    /* enableDedicatedDictSearch == 1 ensures matchstate is not too small in case this CDict will be used for DDS + row hash */
+    size_t const matchStateSize = ZSTD_sizeof_matchState(&cParams, useRowMatchFinder, /* enableDedicatedDictSearch */ 1, /* forCCtx */ 0);
     size_t const neededSize = ZSTD_cwksp_alloc_size(sizeof(ZSTD_CDict))
                             + (dictLoadMethod == ZSTD_dlm_byRef ? 0
                                : ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(dictSize, sizeof(void*))))
@@ -3848,6 +5012,8 @@ const ZSTD_CDict* ZSTD_initStaticCDict(
 
     ZSTD_CCtxParams_init(&params, 0);
     params.cParams = cParams;
+    params.useRowMatchFinder = useRowMatchFinder;
+    cdict->useRowMatchFinder = useRowMatchFinder;
 
     if (ZSTD_isError( ZSTD_initCDict_internal(cdict,
                                               dict, dictSize,
@@ -3874,62 +5040,88 @@ unsigned ZSTD_getDictID_fromCDict(const ZSTD_CDict* cdict)
     return cdict->dictID;
 }
 
-
-/* ZSTD_compressBegin_usingCDict_advanced() :
- * cdict must be != NULL */
-size_t ZSTD_compressBegin_usingCDict_advanced(
+/* ZSTD_compressBegin_usingCDict_internal() :
+ * Implementation of various ZSTD_compressBegin_usingCDict* functions.
+ */
+static size_t ZSTD_compressBegin_usingCDict_internal(
     ZSTD_CCtx* const cctx, const ZSTD_CDict* const cdict,
     ZSTD_frameParameters const fParams, unsigned long long const pledgedSrcSize)
 {
-    DEBUGLOG(4, "ZSTD_compressBegin_usingCDict_advanced");
+    ZSTD_CCtx_params cctxParams;
+    DEBUGLOG(4, "ZSTD_compressBegin_usingCDict_internal");
     RETURN_ERROR_IF(cdict==NULL, dictionary_wrong, "NULL pointer!");
-    {   ZSTD_CCtx_params params = cctx->requestedParams;
+    /* Initialize the cctxParams from the cdict */
+    {
+        ZSTD_parameters params;
+        params.fParams = fParams;
         params.cParams = ( pledgedSrcSize < ZSTD_USE_CDICT_PARAMS_SRCSIZE_CUTOFF
                         || pledgedSrcSize < cdict->dictContentSize * ZSTD_USE_CDICT_PARAMS_DICTSIZE_MULTIPLIER
                         || pledgedSrcSize == ZSTD_CONTENTSIZE_UNKNOWN
-                        || cdict->compressionLevel == 0 )
-                      && (params.attachDictPref != ZSTD_dictForceLoad) ?
+                        || cdict->compressionLevel == 0 ) ?
                 ZSTD_getCParamsFromCDict(cdict)
               : ZSTD_getCParams(cdict->compressionLevel,
                                 pledgedSrcSize,
                                 cdict->dictContentSize);
-        /* Increase window log to fit the entire dictionary and source if the
-         * source size is known. Limit the increase to 19, which is the
-         * window log for compression level 1 with the largest source size.
-         */
-        if (pledgedSrcSize != ZSTD_CONTENTSIZE_UNKNOWN) {
-            U32 const limitedSrcSize = (U32)MIN(pledgedSrcSize, 1U << 19);
-            U32 const limitedSrcLog = limitedSrcSize > 1 ? ZSTD_highbit32(limitedSrcSize - 1) + 1 : 1;
-            params.cParams.windowLog = MAX(params.cParams.windowLog, limitedSrcLog);
-        }
-        params.fParams = fParams;
-        return ZSTD_compressBegin_internal(cctx,
-                                           NULL, 0, ZSTD_dct_auto, ZSTD_dtlm_fast,
-                                           cdict,
-                                           &params, pledgedSrcSize,
-                                           ZSTDb_not_buffered);
+        ZSTD_CCtxParams_init_internal(&cctxParams, &params, cdict->compressionLevel);
+    }
+    /* Increase window log to fit the entire dictionary and source if the
+     * source size is known. Limit the increase to 19, which is the
+     * window log for compression level 1 with the largest source size.
+     */
+    if (pledgedSrcSize != ZSTD_CONTENTSIZE_UNKNOWN) {
+        U32 const limitedSrcSize = (U32)MIN(pledgedSrcSize, 1U << 19);
+        U32 const limitedSrcLog = limitedSrcSize > 1 ? ZSTD_highbit32(limitedSrcSize - 1) + 1 : 1;
+        cctxParams.cParams.windowLog = MAX(cctxParams.cParams.windowLog, limitedSrcLog);
     }
+    return ZSTD_compressBegin_internal(cctx,
+                                        NULL, 0, ZSTD_dct_auto, ZSTD_dtlm_fast,
+                                        cdict,
+                                        &cctxParams, pledgedSrcSize,
+                                        ZSTDb_not_buffered);
+}
+
+
+/* ZSTD_compressBegin_usingCDict_advanced() :
+ * This function is DEPRECATED.
+ * cdict must be != NULL */
+size_t ZSTD_compressBegin_usingCDict_advanced(
+    ZSTD_CCtx* const cctx, const ZSTD_CDict* const cdict,
+    ZSTD_frameParameters const fParams, unsigned long long const pledgedSrcSize)
+{
+    return ZSTD_compressBegin_usingCDict_internal(cctx, cdict, fParams, pledgedSrcSize);
 }
 
 /* ZSTD_compressBegin_usingCDict() :
- * pledgedSrcSize=0 means "unknown"
- * if pledgedSrcSize>0, it will enable contentSizeFlag */
+ * cdict must be != NULL */
 size_t ZSTD_compressBegin_usingCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict)
 {
     ZSTD_frameParameters const fParams = { 0 /*content*/, 0 /*checksum*/, 0 /*noDictID*/ };
-    DEBUGLOG(4, "ZSTD_compressBegin_usingCDict : dictIDFlag == %u", !fParams.noDictIDFlag);
-    return ZSTD_compressBegin_usingCDict_advanced(cctx, cdict, fParams, ZSTD_CONTENTSIZE_UNKNOWN);
+    return ZSTD_compressBegin_usingCDict_internal(cctx, cdict, fParams, ZSTD_CONTENTSIZE_UNKNOWN);
 }
 
-size_t ZSTD_compress_usingCDict_advanced(ZSTD_CCtx* cctx,
+/*! ZSTD_compress_usingCDict_internal():
+ * Implementation of various ZSTD_compress_usingCDict* functions.
+ */
+static size_t ZSTD_compress_usingCDict_internal(ZSTD_CCtx* cctx,
                                 void* dst, size_t dstCapacity,
                                 const void* src, size_t srcSize,
                                 const ZSTD_CDict* cdict, ZSTD_frameParameters fParams)
 {
-    FORWARD_IF_ERROR(ZSTD_compressBegin_usingCDict_advanced(cctx, cdict, fParams, srcSize), "");   /* will check if cdict != NULL */
+    FORWARD_IF_ERROR(ZSTD_compressBegin_usingCDict_internal(cctx, cdict, fParams, srcSize), ""); /* will check if cdict != NULL */
     return ZSTD_compressEnd(cctx, dst, dstCapacity, src, srcSize);
 }
 
+/*! ZSTD_compress_usingCDict_advanced():
+ * This function is DEPRECATED.
+ */
+size_t ZSTD_compress_usingCDict_advanced(ZSTD_CCtx* cctx,
+                                void* dst, size_t dstCapacity,
+                                const void* src, size_t srcSize,
+                                const ZSTD_CDict* cdict, ZSTD_frameParameters fParams)
+{
+    return ZSTD_compress_usingCDict_internal(cctx, dst, dstCapacity, src, srcSize, cdict, fParams);
+}
+
 /*! ZSTD_compress_usingCDict() :
  *  Compression using a digested Dictionary.
  *  Faster startup than ZSTD_compress_usingDict(), recommended when same dictionary is used multiple times.
@@ -3941,7 +5133,7 @@ size_t ZSTD_compress_usingCDict(ZSTD_CCtx* cctx,
                                 const ZSTD_CDict* cdict)
 {
     ZSTD_frameParameters const fParams = { 1 /*content*/, 0 /*checksum*/, 0 /*noDictID*/ };
-    return ZSTD_compress_usingCDict_advanced(cctx, dst, dstCapacity, src, srcSize, cdict, fParams);
+    return ZSTD_compress_usingCDict_internal(cctx, dst, dstCapacity, src, srcSize, cdict, fParams);
 }
 
 
@@ -4071,7 +5263,7 @@ size_t ZSTD_initCStream_advanced(ZSTD_CStream* zcs,
     FORWARD_IF_ERROR( ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only) , "");
     FORWARD_IF_ERROR( ZSTD_CCtx_setPledgedSrcSize(zcs, pledgedSrcSize) , "");
     FORWARD_IF_ERROR( ZSTD_checkCParams(params.cParams) , "");
-    zcs->requestedParams = ZSTD_assignParamsToCCtxParams(&zcs->requestedParams, &params);
+    ZSTD_CCtxParams_setZstdParams(&zcs->requestedParams, &params);
     FORWARD_IF_ERROR( ZSTD_CCtx_loadDictionary(zcs, dict, dictSize) , "");
     return 0;
 }
@@ -4351,8 +5543,13 @@ static size_t ZSTD_CCtx_init_compressStream2(ZSTD_CCtx* cctx,
     FORWARD_IF_ERROR( ZSTD_initLocalDict(cctx) , ""); /* Init the local dict if present. */
     ZSTD_memset(&cctx->prefixDict, 0, sizeof(cctx->prefixDict));   /* single usage */
     assert(prefixDict.dict==NULL || cctx->cdict==NULL);    /* only one can be set */
-    if (cctx->cdict)
-        params.compressionLevel = cctx->cdict->compressionLevel; /* let cdict take priority in terms of compression level */
+    if (cctx->cdict && !cctx->localDict.cdict) {
+        /* Let the cdict's compression level take priority over the requested params.
+         * But do not take the cdict's compression level if the "cdict" is actually a localDict
+         * generated from ZSTD_initLocalDict().
+         */
+        params.compressionLevel = cctx->cdict->compressionLevel;
+    }
     DEBUGLOG(4, "ZSTD_compressStream2 : transparent init stage");
     if (endOp == ZSTD_e_end) cctx->pledgedSrcSizePlusOne = inSize + 1;  /* auto-fix pledgedSrcSize */
     {
@@ -4371,11 +5568,21 @@ static size_t ZSTD_CCtx_init_compressStream2(ZSTD_CCtx* cctx,
         params.ldmParams.enableLdm = 1;
     }
 
+    if (ZSTD_CParams_useBlockSplitter(&params.cParams)) {
+        DEBUGLOG(4, "Block splitter enabled by default (window size >= 128K, strategy >= btopt)");
+        params.splitBlocks = 1;
+    }
+
+    params.useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(params.useRowMatchFinder, &params.cParams);
+
 #ifdef ZSTD_MULTITHREAD
     if ((cctx->pledgedSrcSizePlusOne-1) <= ZSTDMT_JOBSIZE_MIN) {
         params.nbWorkers = 0; /* do not invoke multi-threading when src size is too small */
     }
     if (params.nbWorkers > 0) {
+#if ZSTD_TRACE
+        cctx->traceCtx = (ZSTD_trace_compress_begin != NULL) ? ZSTD_trace_compress_begin(cctx) : 0;
+#endif
         /* mt context creation */
         if (cctx->mtctx == NULL) {
             DEBUGLOG(4, "ZSTD_compressStream2: creating new mtctx for nbWorkers=%u",
@@ -4389,6 +5596,10 @@ static size_t ZSTD_CCtx_init_compressStream2(ZSTD_CCtx* cctx,
                     cctx->mtctx,
                     prefixDict.dict, prefixDict.dictSize, prefixDict.dictContentType,
                     cctx->cdict, params, cctx->pledgedSrcSizePlusOne-1) , "");
+        cctx->dictID = cctx->cdict ? cctx->cdict->dictID : 0;
+        cctx->dictContentSize = cctx->cdict ? cctx->cdict->dictContentSize : prefixDict.dictSize;
+        cctx->consumedSrcSize = 0;
+        cctx->producedCSize = 0;
         cctx->streamStage = zcss_load;
         cctx->appliedParams = params;
     } else
@@ -4450,8 +5661,12 @@ size_t ZSTD_compressStream2( ZSTD_CCtx* cctx,
             size_t const ipos = input->pos;
             size_t const opos = output->pos;
             flushMin = ZSTDMT_compressStream_generic(cctx->mtctx, output, input, endOp);
+            cctx->consumedSrcSize += (U64)(input->pos - ipos);
+            cctx->producedCSize += (U64)(output->pos - opos);
             if ( ZSTD_isError(flushMin)
               || (endOp == ZSTD_e_end && flushMin == 0) ) { /* compression completed */
+                if (flushMin == 0)
+                    ZSTD_CCtx_trace(cctx, 0);
                 ZSTD_CCtx_reset(cctx, ZSTD_reset_session_only);
             }
             FORWARD_IF_ERROR(flushMin, "ZSTDMT_compressStream_generic failed");
@@ -4632,6 +5847,13 @@ static size_t ZSTD_copySequencesToSeqStoreExplicitBlockDelim(ZSTD_CCtx* cctx, ZS
     return 0;
 }
 
+static size_t ZSTD_convertBlockSequencesToSeqStore(ZSTD_CCtx* cctx,
+                                  const ZSTD_Sequence* inSeqs, size_t inSeqsSize,
+                                  const void* src, size_t srcSize) {
+    ZSTD_sequencePosition dummySeqPos = {0, 0, 0};
+    return ZSTD_copySequencesToSeqStoreExplicitBlockDelim(cctx, &dummySeqPos, inSeqs, inSeqsSize, src, srcSize);
+}
+
 /* Returns the number of bytes to move the current read position back by. Only non-zero
  * if we ended up splitting a sequence. Otherwise, it may return a ZSTD error if something
  * went wrong.
@@ -4834,7 +6056,7 @@ static size_t ZSTD_compressSequences_internal(ZSTD_CCtx* cctx,
             continue;
         }
 
-        compressedSeqsSize = ZSTD_entropyCompressSequences(&cctx->seqStore,
+        compressedSeqsSize = ZSTD_entropyCompressSeqStore(&cctx->seqStore,
                                 &cctx->blockState.prevCBlock->entropy, &cctx->blockState.nextCBlock->entropy,
                                 &cctx->appliedParams,
                                 op + ZSTD_blockHeaderSize /* Leave space for block header */, dstCapacity - ZSTD_blockHeaderSize,
@@ -4866,7 +6088,7 @@ static size_t ZSTD_compressSequences_internal(ZSTD_CCtx* cctx,
         } else {
             U32 cBlockHeader;
             /* Error checking and repcodes update */
-            ZSTD_confirmRepcodesAndEntropyTables(cctx);
+            ZSTD_blockState_confirmRepcodesAndEntropyTables(&cctx->blockState);
             if (cctx->blockState.prevCBlock->entropy.fse.offcode_repeatMode == FSE_repeat_valid)
                 cctx->blockState.prevCBlock->entropy.fse.offcode_repeatMode = FSE_repeat_check;
 
@@ -4967,6 +6189,7 @@ size_t ZSTD_endStream(ZSTD_CStream* zcs, ZSTD_outBuffer* output)
 #define ZSTD_MAX_CLEVEL     22
 int ZSTD_maxCLevel(void) { return ZSTD_MAX_CLEVEL; }
 int ZSTD_minCLevel(void) { return (int)-ZSTD_TARGETLENGTH_MAX; }
+int ZSTD_defaultCLevel(void) { return ZSTD_CLEVEL_DEFAULT; }
 
 static const ZSTD_compressionParameters ZSTD_defaultCParameters[4][ZSTD_MAX_CLEVEL+1] = {
 {   /* "default" - for any srcSize > 256 KB */
@@ -4976,15 +6199,15 @@ static const ZSTD_compressionParameters ZSTD_defaultCParameters[4][ZSTD_MAX_CLEV
     { 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,  2,  5,  2, ZSTD_greedy  },  /* level  5 */
-    { 21, 19, 19,  3,  5,  4, ZSTD_greedy  },  /* level  6 */
-    { 21, 19, 19,  3,  5,  8, ZSTD_lazy    },  /* level  7 */
-    { 21, 19, 19,  3,  5, 16, ZSTD_lazy2   },  /* level  8 */
-    { 21, 19, 20,  4,  5, 16, ZSTD_lazy2   },  /* level  9 */
-    { 22, 20, 21,  4,  5, 16, ZSTD_lazy2   },  /* level 10 */
-    { 22, 21, 22,  4,  5, 16, ZSTD_lazy2   },  /* level 11 */
-    { 22, 21, 22,  5,  5, 16, ZSTD_lazy2   },  /* level 12 */
-    { 22, 21, 22,  5,  5, 32, ZSTD_btlazy2 },  /* level 13 */
+    { 21, 18, 19,  4,  5,  2, ZSTD_greedy  },  /* level  5 */
+    { 21, 19, 20,  5,  5,  4, ZSTD_greedy  },  /* level  6 */
+    { 21, 19, 20,  4,  5,  8, ZSTD_lazy    },  /* level  7 */
+    { 21, 19, 20,  5,  5, 16, ZSTD_lazy    },  /* level  8 */
+    { 21, 20, 21,  4,  5, 16, ZSTD_lazy2   },  /* level  9 */
+    { 22, 21, 22,  4,  5, 16, ZSTD_lazy2   },  /* level 10 */
+    { 22, 21, 22,  5,  5, 16, ZSTD_lazy2   },  /* level 11 */
+    { 22, 21, 22,  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 */
@@ -5001,8 +6224,8 @@ static const ZSTD_compressionParameters ZSTD_defaultCParameters[4][ZSTD_MAX_CLEV
     { 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,  2,  5,  2, ZSTD_greedy  },  /* level  4.*/
-    { 18, 18, 18,  3,  5,  2, ZSTD_greedy  },  /* level  5.*/
+    { 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 */
@@ -5029,11 +6252,11 @@ static const ZSTD_compressionParameters ZSTD_defaultCParameters[4][ZSTD_MAX_CLEV
     { 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, 17, 17,  3,  4,  4, ZSTD_lazy    },  /* level  6 */
-    { 17, 17, 17,  3,  4,  8, ZSTD_lazy2   },  /* level  7 */
-    { 17, 17, 17,  4,  4,  8, ZSTD_lazy2   },  /* level  8 */
-    { 17, 17, 17,  5,  4,  8, ZSTD_lazy2   },  /* level  9 */
-    { 17, 17, 17,  6,  4,  8, ZSTD_lazy2   },  /* level 10 */
+    { 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.*/
@@ -5099,7 +6322,10 @@ static ZSTD_compressionParameters ZSTD_dedicatedDictSearch_getCParams(int const
 static int ZSTD_dedicatedDictSearch_isSupported(
         ZSTD_compressionParameters const* cParams)
 {
-    return (cParams->strategy >= ZSTD_greedy) && (cParams->strategy <= ZSTD_lazy2);
+    return (cParams->strategy >= ZSTD_greedy)
+        && (cParams->strategy <= ZSTD_lazy2)
+        && (cParams->hashLog > cParams->chainLog)
+        && (cParams->chainLog <= 24);
 }
 
 /**
@@ -5117,6 +6343,9 @@ static void ZSTD_dedicatedDictSearch_revertCParams(
         case ZSTD_lazy:
         case ZSTD_lazy2:
             cParams->hashLog -= ZSTD_LAZY_DDSS_BUCKET_LOG;
+            if (cParams->hashLog < ZSTD_HASHLOG_MIN) {
+                cParams->hashLog = ZSTD_HASHLOG_MIN;
+            }
             break;
         case ZSTD_btlazy2:
         case ZSTD_btopt:
@@ -5165,6 +6394,7 @@ static ZSTD_compressionParameters ZSTD_getCParams_internal(int compressionLevel,
     else row = compressionLevel;
 
     {   ZSTD_compressionParameters cp = ZSTD_defaultCParameters[tableID][row];
+        DEBUGLOG(5, "ZSTD_getCParams_internal selected tableID: %u row: %u strat: %u", tableID, row, (U32)cp.strategy);
         /* acceleration factor */
         if (compressionLevel < 0) {
             int const clampedCompressionLevel = MAX(ZSTD_minCLevel(), compressionLevel);
diff --git a/zstd_compress_internal.h b/zstd_compress_internal.h
index 80d150a..f24da24 100644
--- a/zstd_compress_internal.h
+++ b/zstd_compress_internal.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -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;
 
@@ -81,6 +81,53 @@ typedef struct {
     ZSTD_fseCTables_t fse;
 } ZSTD_entropyCTables_t;
 
+/***********************************************
+*  Entropy buffer statistics structs and funcs *
+***********************************************/
+/** ZSTD_hufCTablesMetadata_t :
+ *  Stores Literals Block Type for a super-block in hType, and
+ *  huffman tree description in hufDesBuffer.
+ *  hufDesSize refers to the size of huffman tree description in bytes.
+ *  This metadata is populated in ZSTD_buildBlockEntropyStats_literals() */
+typedef struct {
+    symbolEncodingType_e hType;
+    BYTE hufDesBuffer[ZSTD_MAX_HUF_HEADER_SIZE];
+    size_t hufDesSize;
+} ZSTD_hufCTablesMetadata_t;
+
+/** ZSTD_fseCTablesMetadata_t :
+ *  Stores symbol compression modes for a super-block in {ll, ol, ml}Type, and
+ *  fse tables in fseTablesBuffer.
+ *  fseTablesSize refers to the size of fse tables in bytes.
+ *  This metadata is populated in ZSTD_buildBlockEntropyStats_sequences() */
+typedef struct {
+    symbolEncodingType_e llType;
+    symbolEncodingType_e ofType;
+    symbolEncodingType_e mlType;
+    BYTE fseTablesBuffer[ZSTD_MAX_FSE_HEADERS_SIZE];
+    size_t fseTablesSize;
+    size_t lastCountSize; /* This is to account for bug in 1.3.4. More detail in ZSTD_entropyCompressSeqStore_internal() */
+} ZSTD_fseCTablesMetadata_t;
+
+typedef struct {
+    ZSTD_hufCTablesMetadata_t hufMetadata;
+    ZSTD_fseCTablesMetadata_t fseMetadata;
+} ZSTD_entropyCTablesMetadata_t;
+
+/** ZSTD_buildBlockEntropyStats() :
+ *  Builds entropy for the block.
+ *  @return : 0 on success or error code */
+size_t ZSTD_buildBlockEntropyStats(seqStore_t* seqStorePtr,
+                             const ZSTD_entropyCTables_t* prevEntropy,
+                                   ZSTD_entropyCTables_t* nextEntropy,
+                             const ZSTD_CCtx_params* cctxParams,
+                                   ZSTD_entropyCTablesMetadata_t* entropyMetadata,
+                                   void* workspace, size_t wkspSize);
+
+/*********************************
+*  Compression internals structs *
+*********************************/
+
 typedef struct {
     U32 off;            /* Offset code (offset + ZSTD_REP_MOVE) for the match */
     U32 len;            /* Raw length of match */
@@ -141,14 +188,23 @@ typedef struct {
 } ZSTD_compressedBlockState_t;
 
 typedef struct {
-    BYTE const* nextSrc;    /* next block here to continue on current prefix */
-    BYTE const* base;       /* All regular indexes relative to this position */
-    BYTE const* dictBase;   /* extDict indexes relative to this position */
-    U32 dictLimit;          /* below that point, need extDict */
-    U32 lowLimit;           /* below that point, no more valid data */
+    BYTE const* nextSrc;       /* next block here to continue on current prefix */
+    BYTE const* base;          /* All regular indexes relative to this position */
+    BYTE const* dictBase;      /* extDict indexes relative to this position */
+    U32 dictLimit;             /* below that point, need extDict */
+    U32 lowLimit;              /* below that point, no more valid data */
+    U32 nbOverflowCorrections; /* Number of times overflow correction has run since
+                                * ZSTD_window_init(). Useful for debugging coredumps
+                                * and for ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY.
+                                */
 } 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 */
+
 struct ZSTD_matchState_t {
     ZSTD_window_t window;   /* State for window round buffer management */
     U32 loadedDictEnd;      /* index of end of dictionary, within context's referential.
@@ -160,9 +216,17 @@ struct ZSTD_matchState_t {
                              */
     U32 nextToUpdate;       /* index from which to continue table update */
     U32 hashLog3;           /* dispatch table for matches of len==3 : larger == faster, more memory */
+
+    U32 rowHashLog;                          /* For row-based matchfinder: Hashlog based on nb of rows in the hashTable.*/
+    U16* tagTable;                           /* For row-based matchFinder: A row-based table containing the hashes and head index. */
+    U32 hashCache[ZSTD_ROW_HASH_CACHE_SIZE]; /* For row-based matchFinder: a cache of hashes to improve speed */
+
     U32* hashTable;
     U32* hashTable3;
     U32* chainTable;
+
+    U32 forceNonContiguous; /* Non-zero if we should force non-contiguous load for the next window update. */
+
     int dedicatedDictSearch;  /* Indicates whether this matchState is using the
                                * dedicated dictionary search structure.
                                */
@@ -183,13 +247,22 @@ typedef struct {
     U32 checksum;
 } ldmEntry_t;
 
+typedef struct {
+    BYTE const* split;
+    U32 hash;
+    U32 checksum;
+    ldmEntry_t* bucket;
+} ldmMatchCandidate_t;
+
+#define LDM_BATCH_SIZE 64
+
 typedef struct {
     ZSTD_window_t window;   /* State for the window round buffer management */
     ldmEntry_t* hashTable;
     U32 loadedDictEnd;
     BYTE* bucketOffsets;    /* Next position in bucket to insert entry */
-    U64 hashPower;          /* Used to compute the rolling hash.
-                             * Depends on ldmParams.minMatchLength */
+    size_t splitIndices[LDM_BATCH_SIZE];
+    ldmMatchCandidate_t matchCandidates[LDM_BATCH_SIZE];
 } ldmState_t;
 
 typedef struct {
@@ -246,6 +319,15 @@ struct ZSTD_CCtx_params_s {
     ZSTD_sequenceFormat_e blockDelimiters;
     int validateSequences;
 
+    /* Block splitting */
+    int splitBlocks;
+
+    /* Param for deciding whether to use row-based matchfinder */
+    ZSTD_useRowMatchFinderMode_e useRowMatchFinder;
+
+    /* Always load a dictionary in ext-dict mode (not prefix mode)? */
+    int deterministicRefPrefix;
+
     /* Internal use, for createCCtxParams() and freeCCtxParams() only */
     ZSTD_customMem customMem;
 };  /* typedef'd to ZSTD_CCtx_params within "zstd.h" */
@@ -269,7 +351,9 @@ struct ZSTD_CCtx_s {
     int bmi2;                            /* == 1 if the CPU supports BMI2 and 0 otherwise. CPU support is determined dynamically once per context lifetime. */
     ZSTD_CCtx_params requestedParams;
     ZSTD_CCtx_params appliedParams;
+    ZSTD_CCtx_params simpleApiParams;    /* Param storage used by the simple API - not sticky. Must only be used in top-level simple API functions for storage. */
     U32   dictID;
+    size_t dictContentSize;
 
     ZSTD_cwksp workspace; /* manages buffer for dynamic allocations */
     size_t blockSize;
@@ -321,6 +405,11 @@ struct ZSTD_CCtx_s {
 #ifdef ZSTD_MULTITHREAD
     ZSTDMT_CCtx* mtctx;
 #endif
+
+    /* Tracing */
+#if ZSTD_TRACE
+    ZSTD_TraceCtx traceCtx;
+#endif
 };
 
 typedef enum { ZSTD_dtlm_fast, ZSTD_dtlm_full } ZSTD_dictTableLoadMethod_e;
@@ -355,7 +444,7 @@ 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_dictMode_e dictMode);
+ZSTD_blockCompressor ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_useRowMatchFinderMode_e rowMatchfinderMode, ZSTD_dictMode_e dictMode);
 
 
 MEM_STATIC U32 ZSTD_LLcode(U32 litLength)
@@ -532,8 +621,8 @@ void ZSTD_storeSeq(seqStore_t* seqStorePtr, size_t litLength, const BYTE* litera
 
     /* literal Length */
     if (litLength>0xFFFF) {
-        assert(seqStorePtr->longLengthID == 0); /* there can only be a single long length */
-        seqStorePtr->longLengthID = 1;
+        assert(seqStorePtr->longLengthType == ZSTD_llt_none); /* there can only be a single long length */
+        seqStorePtr->longLengthType = ZSTD_llt_literalLength;
         seqStorePtr->longLengthPos = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
     }
     seqStorePtr->sequences[0].litLength = (U16)litLength;
@@ -543,8 +632,8 @@ void ZSTD_storeSeq(seqStore_t* seqStorePtr, size_t litLength, const BYTE* litera
 
     /* match Length */
     if (mlBase>0xFFFF) {
-        assert(seqStorePtr->longLengthID == 0); /* there can only be a single long length */
-        seqStorePtr->longLengthID = 2;
+        assert(seqStorePtr->longLengthType == ZSTD_llt_none); /* there can only be a single long length */
+        seqStorePtr->longLengthType = ZSTD_llt_matchLength;
         seqStorePtr->longLengthPos = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
     }
     seqStorePtr->sequences[0].matchLength = (U16)mlBase;
@@ -795,6 +884,13 @@ MEM_STATIC void ZSTD_window_clear(ZSTD_window_t* window)
     window->dictLimit = end;
 }
 
+MEM_STATIC U32 ZSTD_window_isEmpty(ZSTD_window_t const window)
+{
+    return window.dictLimit == ZSTD_WINDOW_START_INDEX &&
+           window.lowLimit == ZSTD_WINDOW_START_INDEX &&
+           (window.nextSrc - window.base) == ZSTD_WINDOW_START_INDEX;
+}
+
 /**
  * ZSTD_window_hasExtDict():
  * Returns non-zero if the window has a non-empty extDict.
@@ -818,15 +914,69 @@ MEM_STATIC ZSTD_dictMode_e ZSTD_matchState_dictMode(const ZSTD_matchState_t *ms)
             ZSTD_noDict;
 }
 
+/* Defining this macro to non-zero tells zstd to run the overflow correction
+ * code much more frequently. This is very inefficient, and should only be
+ * used for tests and fuzzers.
+ */
+#ifndef ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY
+#  ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
+#    define ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY 1
+#  else
+#    define ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY 0
+#  endif
+#endif
+
+/**
+ * ZSTD_window_canOverflowCorrect():
+ * Returns non-zero if the indices are large enough for overflow correction
+ * to work correctly without impacting compression ratio.
+ */
+MEM_STATIC U32 ZSTD_window_canOverflowCorrect(ZSTD_window_t const window,
+                                              U32 cycleLog,
+                                              U32 maxDist,
+                                              U32 loadedDictEnd,
+                                              void const* src)
+{
+    U32 const cycleSize = 1u << cycleLog;
+    U32 const curr = (U32)((BYTE const*)src - window.base);
+    U32 const minIndexToOverflowCorrect = cycleSize + MAX(maxDist, cycleSize);
+
+    /* Adjust the min index to backoff the overflow correction frequency,
+     * so we don't waste too much CPU in overflow correction. If this
+     * computation overflows we don't really care, we just need to make
+     * sure it is at least minIndexToOverflowCorrect.
+     */
+    U32 const adjustment = window.nbOverflowCorrections + 1;
+    U32 const adjustedIndex = MAX(minIndexToOverflowCorrect * adjustment,
+                                  minIndexToOverflowCorrect);
+    U32 const indexLargeEnough = curr > adjustedIndex;
+
+    /* Only overflow correct early if the dictionary is invalidated already,
+     * so we don't hurt compression ratio.
+     */
+    U32 const dictionaryInvalidated = curr > maxDist + loadedDictEnd;
+
+    return indexLargeEnough && dictionaryInvalidated;
+}
+
 /**
  * ZSTD_window_needOverflowCorrection():
  * Returns non-zero if the indices are getting too large and need overflow
  * protection.
  */
 MEM_STATIC U32 ZSTD_window_needOverflowCorrection(ZSTD_window_t const window,
+                                                  U32 cycleLog,
+                                                  U32 maxDist,
+                                                  U32 loadedDictEnd,
+                                                  void const* src,
                                                   void const* srcEnd)
 {
     U32 const curr = (U32)((BYTE const*)srcEnd - window.base);
+    if (ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY) {
+        if (ZSTD_window_canOverflowCorrect(window, cycleLog, maxDist, loadedDictEnd, src)) {
+            return 1;
+        }
+    }
     return curr > ZSTD_CURRENT_MAX;
 }
 
@@ -838,7 +988,6 @@ MEM_STATIC U32 ZSTD_window_needOverflowCorrection(ZSTD_window_t const window,
  *
  * The least significant cycleLog bits of the indices must remain the same,
  * which may be 0. Every index up to maxDist in the past must be valid.
- * NOTE: (maxDist & cycleMask) must be zero.
  */
 MEM_STATIC U32 ZSTD_window_correctOverflow(ZSTD_window_t* window, U32 cycleLog,
                                            U32 maxDist, void const* src)
@@ -862,17 +1011,25 @@ MEM_STATIC U32 ZSTD_window_correctOverflow(ZSTD_window_t* window, U32 cycleLog,
      * 3. (cctx->lowLimit + 1<<windowLog) < 1<<32:
      *    windowLog <= 31 ==> 3<<29 + 1<<windowLog < 7<<29 < 1<<32.
      */
-    U32 const cycleMask = (1U << cycleLog) - 1;
+    U32 const cycleSize = 1u << cycleLog;
+    U32 const cycleMask = cycleSize - 1;
     U32 const curr = (U32)((BYTE const*)src - window->base);
     U32 const currentCycle0 = curr & cycleMask;
     /* Exclude zero so that newCurrent - maxDist >= 1. */
-    U32 const currentCycle1 = currentCycle0 == 0 ? (1U << cycleLog) : currentCycle0;
-    U32 const newCurrent = currentCycle1 + maxDist;
+    U32 const currentCycle1 = currentCycle0 == 0 ? cycleSize : currentCycle0;
+    U32 const newCurrent = currentCycle1 + MAX(maxDist, cycleSize);
     U32 const correction = curr - newCurrent;
-    assert((maxDist & cycleMask) == 0);
+    /* maxDist must be a power of two so that:
+     *   (newCurrent & cycleMask) == (curr & cycleMask)
+     * This is required to not corrupt the chains / binary tree.
+     */
+    assert((maxDist & (maxDist - 1)) == 0);
+    assert((curr & cycleMask) == (newCurrent & cycleMask));
     assert(curr > newCurrent);
-    /* Loose bound, should be around 1<<29 (see above) */
-    assert(correction > 1<<28);
+    if (!ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY) {
+        /* Loose bound, should be around 1<<29 (see above) */
+        assert(correction > 1<<28);
+    }
 
     window->base += correction;
     window->dictBase += correction;
@@ -888,6 +1045,8 @@ MEM_STATIC U32 ZSTD_window_correctOverflow(ZSTD_window_t* window, U32 cycleLog,
     assert(window->lowLimit <= newCurrent);
     assert(window->dictLimit <= newCurrent);
 
+    ++window->nbOverflowCorrections;
+
     DEBUGLOG(4, "Correction of 0x%x bytes to lowLimit=0x%x", correction,
              window->lowLimit);
     return correction;
@@ -992,11 +1151,13 @@ ZSTD_checkDictValidity(const ZSTD_window_t* window,
 
 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 */
-    window->lowLimit = 1;     /* it ensures first and later CCtx usages compress the same */
-    window->nextSrc = window->base + 1;   /* see issue #1241 */
+    window->base = (BYTE const*)" ";
+    window->dictBase = (BYTE const*)" ";
+    ZSTD_STATIC_ASSERT(ZSTD_DUBT_UNSORTED_MARK < ZSTD_WINDOW_START_INDEX); /* Start above ZSTD_DUBT_UNSORTED_MARK */
+    window->dictLimit = ZSTD_WINDOW_START_INDEX;    /* start from >0, so that 1st position is valid */
+    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;
 }
 
 /**
@@ -1007,7 +1168,8 @@ MEM_STATIC void ZSTD_window_init(ZSTD_window_t* window) {
  * Returns non-zero if the segment is contiguous.
  */
 MEM_STATIC U32 ZSTD_window_update(ZSTD_window_t* window,
-                                  void const* src, size_t srcSize)
+                                  void const* src, size_t srcSize,
+                                  int forceNonContiguous)
 {
     BYTE const* const ip = (BYTE const*)src;
     U32 contiguous = 1;
@@ -1017,7 +1179,7 @@ MEM_STATIC U32 ZSTD_window_update(ZSTD_window_t* window,
     assert(window->base != NULL);
     assert(window->dictBase != NULL);
     /* Check if blocks follow each other */
-    if (src != window->nextSrc) {
+    if (src != window->nextSrc || forceNonContiguous) {
         /* not contiguous */
         size_t const distanceFromBase = (size_t)(window->nextSrc - window->base);
         DEBUGLOG(5, "Non contiguous blocks, new segment starts at %u", window->dictLimit);
@@ -1200,4 +1362,9 @@ size_t ZSTD_referenceExternalSequences(ZSTD_CCtx* cctx, rawSeq* seq, size_t nbSe
  *  condition for correct operation : hashLog > 1 */
 U32 ZSTD_cycleLog(U32 hashLog, ZSTD_strategy strat);
 
+/** ZSTD_CCtx_trace() :
+ *  Trace the end of a compression call.
+ */
+void ZSTD_CCtx_trace(ZSTD_CCtx* cctx, size_t extraCSize);
+
 #endif /* ZSTD_COMPRESS_H */
diff --git a/zstd_compress_literals.c b/zstd_compress_literals.c
index 6dd1c14..52b0a80 100644
--- a/zstd_compress_literals.c
+++ b/zstd_compress_literals.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -15,7 +15,7 @@
 
 size_t ZSTD_noCompressLiterals (void* dst, size_t dstCapacity, const void* src, size_t srcSize)
 {
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     U32   const flSize = 1 + (srcSize>31) + (srcSize>4095);
 
     RETURN_ERROR_IF(srcSize + flSize > dstCapacity, dstSize_tooSmall, "");
@@ -42,7 +42,7 @@ 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)
 {
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     U32   const flSize = 1 + (srcSize>31) + (srcSize>4095);
 
     (void)dstCapacity;  /* dstCapacity already guaranteed to be >=4, hence large enough */
@@ -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");
@@ -117,7 +118,7 @@ size_t ZSTD_compressLiterals (ZSTD_hufCTables_t const* prevHuf,
         }
     }
 
-    if ((cLitSize==0) | (cLitSize >= srcSize - minGain) | ERR_isError(cLitSize)) {
+    if ((cLitSize==0) || (cLitSize >= srcSize - minGain) || ERR_isError(cLitSize)) {
         ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
         return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
     }
diff --git a/zstd_compress_literals.h b/zstd_compress_literals.h
index 8b08705..9775fb9 100644
--- a/zstd_compress_literals.h
+++ b/zstd_compress_literals.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -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 */
diff --git a/zstd_compress_sequences.c b/zstd_compress_sequences.c
index be30c08..44e1728 100644
--- a/zstd_compress_sequences.c
+++ b/zstd_compress_sequences.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -85,6 +85,8 @@ static size_t ZSTD_entropyCost(unsigned const* count, unsigned const max, size_t
 {
     unsigned cost = 0;
     unsigned s;
+
+    assert(total > 0);
     for (s = 0; s <= max; ++s) {
         unsigned norm = (unsigned)((256 * count[s]) / total);
         if (count[s] != 0 && norm == 0)
@@ -232,6 +234,11 @@ ZSTD_selectEncodingType(
     return set_compressed;
 }
 
+typedef struct {
+    S16 norm[MaxSeq + 1];
+    U32 wksp[FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(MaxSeq, MaxFSELog)];
+} ZSTD_BuildCTableWksp;
+
 size_t
 ZSTD_buildCTable(void* dst, size_t dstCapacity,
                 FSE_CTable* nextCTable, U32 FSELog, symbolEncodingType_e type,
@@ -258,7 +265,7 @@ ZSTD_buildCTable(void* dst, size_t dstCapacity,
         FORWARD_IF_ERROR(FSE_buildCTable_wksp(nextCTable, defaultNorm, defaultMax, defaultNormLog, entropyWorkspace, entropyWorkspaceSize), "");  /* note : could be pre-calculated */
         return 0;
     case set_compressed: {
-        S16 norm[MaxSeq + 1];
+        ZSTD_BuildCTableWksp* wksp = (ZSTD_BuildCTableWksp*)entropyWorkspace;
         size_t nbSeq_1 = nbSeq;
         const U32 tableLog = FSE_optimalTableLog(FSELog, nbSeq, max);
         if (count[codeTable[nbSeq-1]] > 1) {
@@ -266,11 +273,13 @@ ZSTD_buildCTable(void* dst, size_t dstCapacity,
             nbSeq_1--;
         }
         assert(nbSeq_1 > 1);
-        assert(entropyWorkspaceSize >= FSE_BUILD_CTABLE_WORKSPACE_SIZE(MaxSeq, MaxFSELog));
-        FORWARD_IF_ERROR(FSE_normalizeCount(norm, tableLog, count, nbSeq_1, max, ZSTD_useLowProbCount(nbSeq_1)), "");
-        {   size_t const NCountSize = FSE_writeNCount(op, oend - op, norm, max, tableLog);   /* overflow protected */
+        assert(entropyWorkspaceSize >= sizeof(ZSTD_BuildCTableWksp));
+        (void)entropyWorkspaceSize;
+        FORWARD_IF_ERROR(FSE_normalizeCount(wksp->norm, tableLog, count, nbSeq_1, max, ZSTD_useLowProbCount(nbSeq_1)), "FSE_normalizeCount failed");
+        assert(oend >= op);
+        {   size_t const NCountSize = FSE_writeNCount(op, (size_t)(oend - op), wksp->norm, max, tableLog);   /* overflow protected */
             FORWARD_IF_ERROR(NCountSize, "FSE_writeNCount failed");
-            FORWARD_IF_ERROR(FSE_buildCTable_wksp(nextCTable, norm, max, tableLog, entropyWorkspace, entropyWorkspaceSize), "");
+            FORWARD_IF_ERROR(FSE_buildCTable_wksp(nextCTable, wksp->norm, max, tableLog, wksp->wksp, sizeof(wksp->wksp)), "FSE_buildCTable_wksp failed");
             return NCountSize;
         }
     }
diff --git a/zstd_compress_sequences.h b/zstd_compress_sequences.h
index b884e45..df84886 100644
--- a/zstd_compress_sequences.h
+++ b/zstd_compress_sequences.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_compress_superblock.c b/zstd_compress_superblock.c
index c227e4a..ec96db9 100644
--- a/zstd_compress_superblock.c
+++ b/zstd_compress_superblock.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -15,288 +15,10 @@
 
 #include "zstd_internal.h"  /* ZSTD_getSequenceLength */
 #include "hist.h"                     /* HIST_countFast_wksp */
-#include "zstd_compress_internal.h"
+#include "zstd_compress_internal.h"   /* ZSTD_[huf|fse|entropy]CTablesMetadata_t */
 #include "zstd_compress_sequences.h"
 #include "zstd_compress_literals.h"
 
-/*-*************************************
-*  Superblock entropy buffer structs
-***************************************/
-/** ZSTD_hufCTablesMetadata_t :
- *  Stores Literals Block Type for a super-block in hType, and
- *  huffman tree description in hufDesBuffer.
- *  hufDesSize refers to the size of huffman tree description in bytes.
- *  This metadata is populated in ZSTD_buildSuperBlockEntropy_literal() */
-typedef struct {
-    symbolEncodingType_e hType;
-    BYTE hufDesBuffer[ZSTD_MAX_HUF_HEADER_SIZE];
-    size_t hufDesSize;
-} ZSTD_hufCTablesMetadata_t;
-
-/** ZSTD_fseCTablesMetadata_t :
- *  Stores symbol compression modes for a super-block in {ll, ol, ml}Type, and
- *  fse tables in fseTablesBuffer.
- *  fseTablesSize refers to the size of fse tables in bytes.
- *  This metadata is populated in ZSTD_buildSuperBlockEntropy_sequences() */
-typedef struct {
-    symbolEncodingType_e llType;
-    symbolEncodingType_e ofType;
-    symbolEncodingType_e mlType;
-    BYTE fseTablesBuffer[ZSTD_MAX_FSE_HEADERS_SIZE];
-    size_t fseTablesSize;
-    size_t lastCountSize; /* This is to account for bug in 1.3.4. More detail in ZSTD_compressSubBlock_sequences() */
-} ZSTD_fseCTablesMetadata_t;
-
-typedef struct {
-    ZSTD_hufCTablesMetadata_t hufMetadata;
-    ZSTD_fseCTablesMetadata_t fseMetadata;
-} ZSTD_entropyCTablesMetadata_t;
-
-
-/** ZSTD_buildSuperBlockEntropy_literal() :
- *  Builds entropy for the super-block literals.
- *  Stores literals block type (raw, rle, compressed, repeat) and
- *  huffman description table to hufMetadata.
- *  @return : size of huffman description table or error code */
-static size_t ZSTD_buildSuperBlockEntropy_literal(void* const src, size_t srcSize,
-                                            const ZSTD_hufCTables_t* prevHuf,
-                                                  ZSTD_hufCTables_t* nextHuf,
-                                                  ZSTD_hufCTablesMetadata_t* hufMetadata,
-                                                  const int disableLiteralsCompression,
-                                                  void* workspace, size_t wkspSize)
-{
-    BYTE* const wkspStart = (BYTE*)workspace;
-    BYTE* const wkspEnd = wkspStart + wkspSize;
-    BYTE* const countWkspStart = wkspStart;
-    unsigned* const countWksp = (unsigned*)workspace;
-    const size_t countWkspSize = (HUF_SYMBOLVALUE_MAX + 1) * sizeof(unsigned);
-    BYTE* const nodeWksp = countWkspStart + countWkspSize;
-    const size_t nodeWkspSize = wkspEnd-nodeWksp;
-    unsigned maxSymbolValue = 255;
-    unsigned huffLog = HUF_TABLELOG_DEFAULT;
-    HUF_repeat repeat = prevHuf->repeatMode;
-
-    DEBUGLOG(5, "ZSTD_buildSuperBlockEntropy_literal (srcSize=%zu)", srcSize);
-
-    /* Prepare nextEntropy assuming reusing the existing table */
-    ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
-
-    if (disableLiteralsCompression) {
-        DEBUGLOG(5, "set_basic - disabled");
-        hufMetadata->hType = set_basic;
-        return 0;
-    }
-
-    /* small ? don't even attempt compression (speed opt) */
-#   define COMPRESS_LITERALS_SIZE_MIN 63
-    {   size_t const minLitSize = (prevHuf->repeatMode == HUF_repeat_valid) ? 6 : COMPRESS_LITERALS_SIZE_MIN;
-        if (srcSize <= minLitSize) {
-            DEBUGLOG(5, "set_basic - too small");
-            hufMetadata->hType = set_basic;
-            return 0;
-        }
-    }
-
-    /* Scan input and build symbol stats */
-    {   size_t const largest = HIST_count_wksp (countWksp, &maxSymbolValue, (const BYTE*)src, srcSize, workspace, wkspSize);
-        FORWARD_IF_ERROR(largest, "HIST_count_wksp failed");
-        if (largest == srcSize) {
-            DEBUGLOG(5, "set_rle");
-            hufMetadata->hType = set_rle;
-            return 0;
-        }
-        if (largest <= (srcSize >> 7)+4) {
-            DEBUGLOG(5, "set_basic - no gain");
-            hufMetadata->hType = set_basic;
-            return 0;
-        }
-    }
-
-    /* Validate the previous Huffman table */
-    if (repeat == HUF_repeat_check && !HUF_validateCTable((HUF_CElt const*)prevHuf->CTable, countWksp, maxSymbolValue)) {
-        repeat = HUF_repeat_none;
-    }
-
-    /* Build Huffman Tree */
-    ZSTD_memset(nextHuf->CTable, 0, sizeof(nextHuf->CTable));
-    huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue);
-    {   size_t const maxBits = HUF_buildCTable_wksp((HUF_CElt*)nextHuf->CTable, countWksp,
-                                                    maxSymbolValue, huffLog,
-                                                    nodeWksp, nodeWkspSize);
-        FORWARD_IF_ERROR(maxBits, "HUF_buildCTable_wksp");
-        huffLog = (U32)maxBits;
-        {   /* Build and write the CTable */
-            size_t const newCSize = HUF_estimateCompressedSize(
-                    (HUF_CElt*)nextHuf->CTable, countWksp, maxSymbolValue);
-            size_t const hSize = HUF_writeCTable(
-                    hufMetadata->hufDesBuffer, sizeof(hufMetadata->hufDesBuffer),
-                    (HUF_CElt*)nextHuf->CTable, maxSymbolValue, huffLog);
-            /* Check against repeating the previous CTable */
-            if (repeat != HUF_repeat_none) {
-                size_t const oldCSize = HUF_estimateCompressedSize(
-                        (HUF_CElt const*)prevHuf->CTable, countWksp, maxSymbolValue);
-                if (oldCSize < srcSize && (oldCSize <= hSize + newCSize || hSize + 12 >= srcSize)) {
-                    DEBUGLOG(5, "set_repeat - smaller");
-                    ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
-                    hufMetadata->hType = set_repeat;
-                    return 0;
-                }
-            }
-            if (newCSize + hSize >= srcSize) {
-                DEBUGLOG(5, "set_basic - no gains");
-                ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
-                hufMetadata->hType = set_basic;
-                return 0;
-            }
-            DEBUGLOG(5, "set_compressed (hSize=%u)", (U32)hSize);
-            hufMetadata->hType = set_compressed;
-            nextHuf->repeatMode = HUF_repeat_check;
-            return hSize;
-        }
-    }
-}
-
-/** ZSTD_buildSuperBlockEntropy_sequences() :
- *  Builds entropy for the super-block sequences.
- *  Stores symbol compression modes and fse table to fseMetadata.
- *  @return : size of fse tables or error code */
-static size_t ZSTD_buildSuperBlockEntropy_sequences(seqStore_t* seqStorePtr,
-                                              const ZSTD_fseCTables_t* prevEntropy,
-                                                    ZSTD_fseCTables_t* nextEntropy,
-                                              const ZSTD_CCtx_params* cctxParams,
-                                                    ZSTD_fseCTablesMetadata_t* fseMetadata,
-                                                    void* workspace, size_t wkspSize)
-{
-    BYTE* const wkspStart = (BYTE*)workspace;
-    BYTE* const wkspEnd = wkspStart + wkspSize;
-    BYTE* const countWkspStart = wkspStart;
-    unsigned* const countWksp = (unsigned*)workspace;
-    const size_t countWkspSize = (MaxSeq + 1) * sizeof(unsigned);
-    BYTE* const cTableWksp = countWkspStart + countWkspSize;
-    const size_t cTableWkspSize = wkspEnd-cTableWksp;
-    ZSTD_strategy const strategy = cctxParams->cParams.strategy;
-    FSE_CTable* CTable_LitLength = nextEntropy->litlengthCTable;
-    FSE_CTable* CTable_OffsetBits = nextEntropy->offcodeCTable;
-    FSE_CTable* CTable_MatchLength = nextEntropy->matchlengthCTable;
-    const BYTE* const ofCodeTable = seqStorePtr->ofCode;
-    const BYTE* const llCodeTable = seqStorePtr->llCode;
-    const BYTE* const mlCodeTable = seqStorePtr->mlCode;
-    size_t const nbSeq = seqStorePtr->sequences - seqStorePtr->sequencesStart;
-    BYTE* const ostart = fseMetadata->fseTablesBuffer;
-    BYTE* const oend = ostart + sizeof(fseMetadata->fseTablesBuffer);
-    BYTE* op = ostart;
-
-    assert(cTableWkspSize >= (1 << MaxFSELog) * sizeof(FSE_FUNCTION_TYPE));
-    DEBUGLOG(5, "ZSTD_buildSuperBlockEntropy_sequences (nbSeq=%zu)", nbSeq);
-    ZSTD_memset(workspace, 0, wkspSize);
-
-    fseMetadata->lastCountSize = 0;
-    /* convert length/distances into codes */
-    ZSTD_seqToCodes(seqStorePtr);
-    /* build CTable for Literal Lengths */
-    {   U32 LLtype;
-        unsigned max = MaxLL;
-        size_t const mostFrequent = HIST_countFast_wksp(countWksp, &max, llCodeTable, nbSeq, workspace, wkspSize);  /* can't fail */
-        DEBUGLOG(5, "Building LL table");
-        nextEntropy->litlength_repeatMode = prevEntropy->litlength_repeatMode;
-        LLtype = ZSTD_selectEncodingType(&nextEntropy->litlength_repeatMode,
-                                        countWksp, max, mostFrequent, nbSeq,
-                                        LLFSELog, prevEntropy->litlengthCTable,
-                                        LL_defaultNorm, LL_defaultNormLog,
-                                        ZSTD_defaultAllowed, strategy);
-        assert(set_basic < set_compressed && set_rle < set_compressed);
-        assert(!(LLtype < set_compressed && nextEntropy->litlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */
-        {   size_t const countSize = ZSTD_buildCTable(op, oend - op, CTable_LitLength, LLFSELog, (symbolEncodingType_e)LLtype,
-                                                    countWksp, max, llCodeTable, nbSeq, LL_defaultNorm, LL_defaultNormLog, MaxLL,
-                                                    prevEntropy->litlengthCTable, sizeof(prevEntropy->litlengthCTable),
-                                                    cTableWksp, cTableWkspSize);
-            FORWARD_IF_ERROR(countSize, "ZSTD_buildCTable for LitLens failed");
-            if (LLtype == set_compressed)
-                fseMetadata->lastCountSize = countSize;
-            op += countSize;
-            fseMetadata->llType = (symbolEncodingType_e) LLtype;
-    }   }
-    /* build CTable for Offsets */
-    {   U32 Offtype;
-        unsigned max = MaxOff;
-        size_t const mostFrequent = HIST_countFast_wksp(countWksp, &max, ofCodeTable, nbSeq, workspace, wkspSize);  /* can't fail */
-        /* We can only use the basic table if max <= DefaultMaxOff, otherwise the offsets are too large */
-        ZSTD_defaultPolicy_e const defaultPolicy = (max <= DefaultMaxOff) ? ZSTD_defaultAllowed : ZSTD_defaultDisallowed;
-        DEBUGLOG(5, "Building OF table");
-        nextEntropy->offcode_repeatMode = prevEntropy->offcode_repeatMode;
-        Offtype = ZSTD_selectEncodingType(&nextEntropy->offcode_repeatMode,
-                                        countWksp, max, mostFrequent, nbSeq,
-                                        OffFSELog, prevEntropy->offcodeCTable,
-                                        OF_defaultNorm, OF_defaultNormLog,
-                                        defaultPolicy, strategy);
-        assert(!(Offtype < set_compressed && nextEntropy->offcode_repeatMode != FSE_repeat_none)); /* We don't copy tables */
-        {   size_t const countSize = ZSTD_buildCTable(op, oend - op, CTable_OffsetBits, OffFSELog, (symbolEncodingType_e)Offtype,
-                                                    countWksp, max, ofCodeTable, nbSeq, OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff,
-                                                    prevEntropy->offcodeCTable, sizeof(prevEntropy->offcodeCTable),
-                                                    cTableWksp, cTableWkspSize);
-            FORWARD_IF_ERROR(countSize, "ZSTD_buildCTable for Offsets failed");
-            if (Offtype == set_compressed)
-                fseMetadata->lastCountSize = countSize;
-            op += countSize;
-            fseMetadata->ofType = (symbolEncodingType_e) Offtype;
-    }   }
-    /* build CTable for MatchLengths */
-    {   U32 MLtype;
-        unsigned max = MaxML;
-        size_t const mostFrequent = HIST_countFast_wksp(countWksp, &max, mlCodeTable, nbSeq, workspace, wkspSize);   /* can't fail */
-        DEBUGLOG(5, "Building ML table (remaining space : %i)", (int)(oend-op));
-        nextEntropy->matchlength_repeatMode = prevEntropy->matchlength_repeatMode;
-        MLtype = ZSTD_selectEncodingType(&nextEntropy->matchlength_repeatMode,
-                                        countWksp, max, mostFrequent, nbSeq,
-                                        MLFSELog, prevEntropy->matchlengthCTable,
-                                        ML_defaultNorm, ML_defaultNormLog,
-                                        ZSTD_defaultAllowed, strategy);
-        assert(!(MLtype < set_compressed && nextEntropy->matchlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */
-        {   size_t const countSize = ZSTD_buildCTable(op, oend - op, CTable_MatchLength, MLFSELog, (symbolEncodingType_e)MLtype,
-                                                    countWksp, max, mlCodeTable, nbSeq, ML_defaultNorm, ML_defaultNormLog, MaxML,
-                                                    prevEntropy->matchlengthCTable, sizeof(prevEntropy->matchlengthCTable),
-                                                    cTableWksp, cTableWkspSize);
-            FORWARD_IF_ERROR(countSize, "ZSTD_buildCTable for MatchLengths failed");
-            if (MLtype == set_compressed)
-                fseMetadata->lastCountSize = countSize;
-            op += countSize;
-            fseMetadata->mlType = (symbolEncodingType_e) MLtype;
-    }   }
-    assert((size_t) (op-ostart) <= sizeof(fseMetadata->fseTablesBuffer));
-    return op-ostart;
-}
-
-
-/** ZSTD_buildSuperBlockEntropy() :
- *  Builds entropy for the super-block.
- *  @return : 0 on success or error code */
-static size_t
-ZSTD_buildSuperBlockEntropy(seqStore_t* seqStorePtr,
-                      const ZSTD_entropyCTables_t* prevEntropy,
-                            ZSTD_entropyCTables_t* nextEntropy,
-                      const ZSTD_CCtx_params* cctxParams,
-                            ZSTD_entropyCTablesMetadata_t* entropyMetadata,
-                            void* workspace, size_t wkspSize)
-{
-    size_t const litSize = seqStorePtr->lit - seqStorePtr->litStart;
-    DEBUGLOG(5, "ZSTD_buildSuperBlockEntropy");
-    entropyMetadata->hufMetadata.hufDesSize =
-        ZSTD_buildSuperBlockEntropy_literal(seqStorePtr->litStart, litSize,
-                                            &prevEntropy->huf, &nextEntropy->huf,
-                                            &entropyMetadata->hufMetadata,
-                                            ZSTD_disableLiteralsCompression(cctxParams),
-                                            workspace, wkspSize);
-    FORWARD_IF_ERROR(entropyMetadata->hufMetadata.hufDesSize, "ZSTD_buildSuperBlockEntropy_literal failed");
-    entropyMetadata->fseMetadata.fseTablesSize =
-        ZSTD_buildSuperBlockEntropy_sequences(seqStorePtr,
-                                              &prevEntropy->fse, &nextEntropy->fse,
-                                              cctxParams,
-                                              &entropyMetadata->fseMetadata,
-                                              workspace, wkspSize);
-    FORWARD_IF_ERROR(entropyMetadata->fseMetadata.fseTablesSize, "ZSTD_buildSuperBlockEntropy_sequences failed");
-    return 0;
-}
-
 /** ZSTD_compressSubBlock_literal() :
  *  Compresses literals section for a sub-block.
  *  When we have to write the Huffman table we will sometimes choose a header
@@ -304,7 +26,7 @@ ZSTD_buildSuperBlockEntropy(seqStore_t* seqStorePtr,
  *  before we know the table size + compressed size, so we have a bound on the
  *  table size. If we guessed incorrectly, we fall back to uncompressed literals.
  *
- *  We write the header when writeEntropy=1 and set entropyWrriten=1 when we succeeded
+ *  We write the header when writeEntropy=1 and set entropyWritten=1 when we succeeded
  *  in writing the header, otherwise it is set to 0.
  *
  *  hufMetadata->hType has literals block type info.
@@ -643,8 +365,9 @@ static size_t ZSTD_estimateSubBlockSize_sequences(const BYTE* ofCodeTable,
                                                   void* workspace, size_t wkspSize,
                                                   int writeEntropy)
 {
-    size_t sequencesSectionHeaderSize = 3; /* Use hard coded size of 3 bytes */
+    size_t const sequencesSectionHeaderSize = 3; /* Use hard coded size of 3 bytes */
     size_t cSeqSizeEstimate = 0;
+    if (nbSeq == 0) return sequencesSectionHeaderSize;
     cSeqSizeEstimate += ZSTD_estimateSubBlockSize_symbolType(fseMetadata->ofType, ofCodeTable, MaxOff,
                                          nbSeq, fseTables->offcodeCTable, NULL,
                                          OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff,
@@ -830,7 +553,7 @@ size_t ZSTD_compressSuperBlock(ZSTD_CCtx* zc,
                                unsigned lastBlock) {
     ZSTD_entropyCTablesMetadata_t entropyMetadata;
 
-    FORWARD_IF_ERROR(ZSTD_buildSuperBlockEntropy(&zc->seqStore,
+    FORWARD_IF_ERROR(ZSTD_buildBlockEntropyStats(&zc->seqStore,
           &zc->blockState.prevCBlock->entropy,
           &zc->blockState.nextCBlock->entropy,
           &zc->appliedParams,
diff --git a/zstd_compress_superblock.h b/zstd_compress_superblock.h
index 35d2072..ab549c2 100644
--- a/zstd_compress_superblock.h
+++ b/zstd_compress_superblock.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_cwksp.h b/zstd_cwksp.h
index ffb3a8d..e02829e 100644
--- a/zstd_cwksp.h
+++ b/zstd_cwksp.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -35,6 +35,10 @@ extern "C" {
 #define ZSTD_CWKSP_ASAN_REDZONE_SIZE 128
 #endif
 
+
+/* Set our tables and aligneds to align by 64 bytes */
+#define ZSTD_CWKSP_ALIGNMENT_BYTES 64
+
 /*-*************************************
 *  Structures
 ***************************************/
@@ -117,10 +121,11 @@ typedef enum {
  * - Tables: these are any of several different datastructures (hash tables,
  *   chain tables, binary trees) that all respect a common format: they are
  *   uint32_t arrays, all of whose values are between 0 and (nextSrc - base).
- *   Their sizes depend on the cparams.
+ *   Their sizes depend on the cparams. These tables are 64-byte aligned.
  *
  * - Aligned: these buffers are used for various purposes that require 4 byte
- *   alignment, but don't require any initialization before they're used.
+ *   alignment, but don't require any initialization before they're used. These
+ *   buffers are each aligned to 64 bytes.
  *
  * - Buffers: these buffers are used for various purposes that don't require
  *   any alignment or initialization before they're used. This means they can
@@ -133,8 +138,7 @@ typedef enum {
  *
  * 1. Objects
  * 2. Buffers
- * 3. Aligned
- * 4. Tables
+ * 3. Aligned/Tables
  *
  * Attempts to reserve objects of different types out of order will fail.
  */
@@ -187,6 +191,8 @@ MEM_STATIC size_t ZSTD_cwksp_align(size_t size, size_t const align) {
  * Since tables aren't currently redzoned, you don't need to call through this
  * to figure out how much space you need for the matchState tables. Everything
  * else is though.
+ *
+ * Do not use for sizing aligned buffers. Instead, use ZSTD_cwksp_aligned_alloc_size().
  */
 MEM_STATIC size_t ZSTD_cwksp_alloc_size(size_t size) {
     if (size == 0)
@@ -198,30 +204,110 @@ MEM_STATIC size_t ZSTD_cwksp_alloc_size(size_t size) {
 #endif
 }
 
-MEM_STATIC void ZSTD_cwksp_internal_advance_phase(
+/**
+ * Returns an adjusted alloc size that is the nearest larger multiple of 64 bytes.
+ * Used to determine the number of bytes required for a given "aligned".
+ */
+MEM_STATIC size_t ZSTD_cwksp_aligned_alloc_size(size_t size) {
+    return ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(size, ZSTD_CWKSP_ALIGNMENT_BYTES));
+}
+
+/**
+ * Returns the amount of additional space the cwksp must allocate
+ * for internal purposes (currently only alignment).
+ */
+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.
+     *
+     * n_1 + n_2 == 64 bytes if the cwksp is freshly allocated, due to tables and
+     * aligneds being sized in multiples of 64 bytes.
+     */
+    size_t const slackSpace = ZSTD_CWKSP_ALIGNMENT_BYTES;
+    return slackSpace;
+}
+
+
+/**
+ * Return the number of additional bytes required to align a pointer to the given number of bytes.
+ * alignBytes must be a power of two.
+ */
+MEM_STATIC size_t ZSTD_cwksp_bytes_to_align_ptr(void* ptr, const size_t alignBytes) {
+    size_t const alignBytesMask = alignBytes - 1;
+    size_t const bytes = (alignBytes - ((size_t)ptr & (alignBytesMask))) & alignBytesMask;
+    assert((alignBytes & alignBytesMask) == 0);
+    assert(bytes != ZSTD_CWKSP_ALIGNMENT_BYTES);
+    return bytes;
+}
+
+/**
+ * Internal function. Do not use directly.
+ * Reserves the given number of bytes within the aligned/buffer segment of the wksp, which
+ * counts from the end of the wksp. (as opposed to the object/table segment)
+ *
+ * Returns a pointer to the beginning of that space.
+ */
+MEM_STATIC void* ZSTD_cwksp_reserve_internal_buffer_space(ZSTD_cwksp* ws, size_t const bytes) {
+    void* const alloc = (BYTE*)ws->allocStart - bytes;
+    void* const bottom = ws->tableEnd;
+    DEBUGLOG(5, "cwksp: reserving %p %zd bytes, %zd bytes remaining",
+        alloc, bytes, ZSTD_cwksp_available_space(ws) - bytes);
+    ZSTD_cwksp_assert_internal_consistency(ws);
+    assert(alloc >= bottom);
+    if (alloc < bottom) {
+        DEBUGLOG(4, "cwksp: alloc failed!");
+        ws->allocFailed = 1;
+        return NULL;
+    }
+    if (alloc < ws->tableValidEnd) {
+        ws->tableValidEnd = alloc;
+    }
+    ws->allocStart = alloc;
+    return alloc;
+}
+
+/**
+ * Moves the cwksp to the next phase, and does any necessary allocations.
+ * Returns a 0 on success, or zstd error
+ */
+MEM_STATIC size_t ZSTD_cwksp_internal_advance_phase(
         ZSTD_cwksp* ws, ZSTD_cwksp_alloc_phase_e phase) {
     assert(phase >= ws->phase);
     if (phase > ws->phase) {
+        /* Going from allocating objects to allocating buffers */
         if (ws->phase < ZSTD_cwksp_alloc_buffers &&
                 phase >= ZSTD_cwksp_alloc_buffers) {
             ws->tableValidEnd = ws->objectEnd;
         }
+
+        /* Going from allocating buffers to allocating aligneds/tables */
         if (ws->phase < ZSTD_cwksp_alloc_aligned &&
                 phase >= ZSTD_cwksp_alloc_aligned) {
-            /* If unaligned allocations down from a too-large top have left us
-             * unaligned, we need to realign our alloc ptr. Technically, this
-             * can consume space that is unaccounted for in the neededSpace
-             * calculation. However, I believe this can only happen when the
-             * workspace is too large, and specifically when it is too large
-             * by a larger margin than the space that will be consumed. */
-            /* TODO: cleaner, compiler warning friendly way to do this??? */
-            ws->allocStart = (BYTE*)ws->allocStart - ((size_t)ws->allocStart & (sizeof(U32)-1));
-            if (ws->allocStart < ws->tableValidEnd) {
-                ws->tableValidEnd = ws->allocStart;
+            {   /* Align the start of the "aligned" to 64 bytes. Use [1, 64] bytes. */
+                size_t const bytesToAlign =
+                    ZSTD_CWKSP_ALIGNMENT_BYTES - ZSTD_cwksp_bytes_to_align_ptr(ws->allocStart, ZSTD_CWKSP_ALIGNMENT_BYTES);
+                DEBUGLOG(5, "reserving aligned alignment addtl space: %zu", bytesToAlign);
+                ZSTD_STATIC_ASSERT((ZSTD_CWKSP_ALIGNMENT_BYTES & (ZSTD_CWKSP_ALIGNMENT_BYTES - 1)) == 0); /* power of 2 */
+                RETURN_ERROR_IF(!ZSTD_cwksp_reserve_internal_buffer_space(ws, bytesToAlign),
+                                memory_allocation, "aligned phase - alignment initial allocation failed!");
+            }
+            {   /* Align the start of the tables to 64 bytes. Use [0, 63] bytes */
+                void* const alloc = ws->objectEnd;
+                size_t const bytesToAlign = ZSTD_cwksp_bytes_to_align_ptr(alloc, ZSTD_CWKSP_ALIGNMENT_BYTES);
+                void* const end = (BYTE*)alloc + bytesToAlign;
+                DEBUGLOG(5, "reserving table alignment addtl space: %zu", bytesToAlign);
+                RETURN_ERROR_IF(end > ws->workspaceEnd, memory_allocation,
+                                "table phase - alignment initial allocation failed!");
+                ws->objectEnd = end;
+                ws->tableEnd = end;
+                ws->tableValidEnd = end;
             }
         }
         ws->phase = phase;
+        ZSTD_cwksp_assert_internal_consistency(ws);
     }
+    return 0;
 }
 
 /**
@@ -237,38 +323,25 @@ MEM_STATIC int ZSTD_cwksp_owns_buffer(const ZSTD_cwksp* ws, const void* ptr) {
 MEM_STATIC void* ZSTD_cwksp_reserve_internal(
         ZSTD_cwksp* ws, size_t bytes, ZSTD_cwksp_alloc_phase_e phase) {
     void* alloc;
-    void* bottom = ws->tableEnd;
-    ZSTD_cwksp_internal_advance_phase(ws, phase);
-    alloc = (BYTE *)ws->allocStart - bytes;
-
-    if (bytes == 0)
+    if (ZSTD_isError(ZSTD_cwksp_internal_advance_phase(ws, phase)) || bytes == 0) {
         return NULL;
+    }
 
 #if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
     /* over-reserve space */
-    alloc = (BYTE *)alloc - 2 * ZSTD_CWKSP_ASAN_REDZONE_SIZE;
+    bytes += 2 * ZSTD_CWKSP_ASAN_REDZONE_SIZE;
 #endif
 
-    DEBUGLOG(5, "cwksp: reserving %p %zd bytes, %zd bytes remaining",
-        alloc, bytes, ZSTD_cwksp_available_space(ws) - bytes);
-    ZSTD_cwksp_assert_internal_consistency(ws);
-    assert(alloc >= bottom);
-    if (alloc < bottom) {
-        DEBUGLOG(4, "cwksp: alloc failed!");
-        ws->allocFailed = 1;
-        return NULL;
-    }
-    if (alloc < ws->tableValidEnd) {
-        ws->tableValidEnd = alloc;
-    }
-    ws->allocStart = alloc;
+    alloc = ZSTD_cwksp_reserve_internal_buffer_space(ws, bytes);
 
 #if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
     /* Move alloc so there's ZSTD_CWKSP_ASAN_REDZONE_SIZE unused space on
      * either size. */
-    alloc = (BYTE *)alloc + ZSTD_CWKSP_ASAN_REDZONE_SIZE;
-    if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) {
-        __asan_unpoison_memory_region(alloc, bytes);
+    if (alloc) {
+        alloc = (BYTE *)alloc + ZSTD_CWKSP_ASAN_REDZONE_SIZE;
+        if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) {
+            __asan_unpoison_memory_region(alloc, bytes);
+        }
     }
 #endif
 
@@ -283,28 +356,36 @@ MEM_STATIC BYTE* ZSTD_cwksp_reserve_buffer(ZSTD_cwksp* ws, size_t bytes) {
 }
 
 /**
- * Reserves and returns memory sized on and aligned on sizeof(unsigned).
+ * Reserves and returns memory sized on and aligned on ZSTD_CWKSP_ALIGNMENT_BYTES (64 bytes).
  */
 MEM_STATIC void* ZSTD_cwksp_reserve_aligned(ZSTD_cwksp* ws, size_t bytes) {
-    assert((bytes & (sizeof(U32)-1)) == 0);
-    return ZSTD_cwksp_reserve_internal(ws, ZSTD_cwksp_align(bytes, sizeof(U32)), ZSTD_cwksp_alloc_aligned);
+    void* ptr = ZSTD_cwksp_reserve_internal(ws, ZSTD_cwksp_align(bytes, ZSTD_CWKSP_ALIGNMENT_BYTES),
+                                            ZSTD_cwksp_alloc_aligned);
+    assert(((size_t)ptr & (ZSTD_CWKSP_ALIGNMENT_BYTES-1))== 0);
+    return ptr;
 }
 
 /**
- * Aligned on sizeof(unsigned). These buffers have the special property that
+ * Aligned on 64 bytes. These buffers have the special property that
  * their values remain constrained, allowing us to re-use them without
  * memset()-ing them.
  */
 MEM_STATIC void* ZSTD_cwksp_reserve_table(ZSTD_cwksp* ws, size_t bytes) {
     const ZSTD_cwksp_alloc_phase_e phase = ZSTD_cwksp_alloc_aligned;
-    void* alloc = ws->tableEnd;
-    void* end = (BYTE *)alloc + bytes;
-    void* top = ws->allocStart;
+    void* alloc;
+    void* end;
+    void* top;
+
+    if (ZSTD_isError(ZSTD_cwksp_internal_advance_phase(ws, phase))) {
+        return NULL;
+    }
+    alloc = ws->tableEnd;
+    end = (BYTE *)alloc + bytes;
+    top = ws->allocStart;
 
     DEBUGLOG(5, "cwksp: reserving %p table %zd bytes, %zd bytes remaining",
         alloc, bytes, ZSTD_cwksp_available_space(ws) - bytes);
     assert((bytes & (sizeof(U32)-1)) == 0);
-    ZSTD_cwksp_internal_advance_phase(ws, phase);
     ZSTD_cwksp_assert_internal_consistency(ws);
     assert(end <= top);
     if (end > top) {
@@ -320,6 +401,8 @@ MEM_STATIC void* ZSTD_cwksp_reserve_table(ZSTD_cwksp* ws, size_t bytes) {
     }
 #endif
 
+    assert((bytes & (ZSTD_CWKSP_ALIGNMENT_BYTES-1)) == 0);
+    assert(((size_t)alloc & (ZSTD_CWKSP_ALIGNMENT_BYTES-1))== 0);
     return alloc;
 }
 
@@ -503,7 +586,7 @@ MEM_STATIC void ZSTD_cwksp_free(ZSTD_cwksp* ws, ZSTD_customMem customMem) {
 
 /**
  * Moves the management of a workspace from one cwksp to another. The src cwksp
- * is left in an invalid state (src must be re-init()'ed before its used again).
+ * is left in an invalid state (src must be re-init()'ed before it's used again).
  */
 MEM_STATIC void ZSTD_cwksp_move(ZSTD_cwksp* dst, ZSTD_cwksp* src) {
     *dst = *src;
@@ -527,6 +610,24 @@ MEM_STATIC int ZSTD_cwksp_reserve_failed(const ZSTD_cwksp* ws) {
 *  Functions Checking Free Space
 ***************************************/
 
+/* ZSTD_alignmentSpaceWithinBounds() :
+ * Returns if the estimated space needed for a wksp is within an acceptable limit of the
+ * actual amount of space used.
+ */
+MEM_STATIC int ZSTD_cwksp_estimated_space_within_bounds(const ZSTD_cwksp* const ws,
+                                                        size_t const estimatedSpace, int resizedWorkspace) {
+    if (resizedWorkspace) {
+        /* Resized/newly allocated wksp should have exact bounds */
+        return ZSTD_cwksp_used(ws) == estimatedSpace;
+    } else {
+        /* Due to alignment, when reusing a workspace, we can actually consume 63 fewer or more bytes
+         * than estimatedSpace. See the comments in zstd_cwksp.h for details.
+         */
+        return (ZSTD_cwksp_used(ws) >= estimatedSpace - 63) && (ZSTD_cwksp_used(ws) <= estimatedSpace + 63);
+    }
+}
+
+
 MEM_STATIC size_t ZSTD_cwksp_available_space(ZSTD_cwksp* ws) {
     return (size_t)((BYTE*)ws->allocStart - (BYTE*)ws->tableEnd);
 }
diff --git a/zstd_ddict.c b/zstd_ddict.c
index b90f27b..f1ca935 100644
--- a/zstd_ddict.c
+++ b/zstd_ddict.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_ddict.h b/zstd_ddict.h
index b427f9a..a6bec10 100644
--- a/zstd_ddict.h
+++ b/zstd_ddict.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_decompress.c b/zstd_decompress.c
index be77d07..f93a83f 100644
--- a/zstd_decompress.c
+++ b/zstd_decompress.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -62,6 +62,7 @@
 #include "fse.h"
 #define HUF_STATIC_LINKING_ONLY
 #include "huf.h"
+#include "xxhash.h" /* XXH64_reset, XXH64_update, XXH64_digest, XXH64 */
 #include "zstd_internal.h"  /* blockProperties_t */
 #include "zstd_decompress_internal.h"   /* ZSTD_DCtx */
 #include "zstd_ddict.h"  /* ZSTD_DDictDictContent */
@@ -72,6 +73,144 @@
 #endif
 
 
+
+/*************************************
+ * Multiple DDicts Hashset internals *
+ *************************************/
+
+#define DDICT_HASHSET_MAX_LOAD_FACTOR_COUNT_MULT 4
+#define DDICT_HASHSET_MAX_LOAD_FACTOR_SIZE_MULT 3   /* These two constants represent SIZE_MULT/COUNT_MULT load factor without using a float.
+                                                     * Currently, that means a 0.75 load factor.
+                                                     * So, if count * COUNT_MULT / size * SIZE_MULT != 0, then we've exceeded
+                                                     * the load factor of the ddict hash set.
+                                                     */
+
+#define DDICT_HASHSET_TABLE_BASE_SIZE 64
+#define DDICT_HASHSET_RESIZE_FACTOR 2
+
+/* Hash function to determine starting position of dict insertion within the table
+ * Returns an index between [0, hashSet->ddictPtrTableSize]
+ */
+static size_t ZSTD_DDictHashSet_getIndex(const ZSTD_DDictHashSet* hashSet, U32 dictID) {
+    const U64 hash = XXH64(&dictID, sizeof(U32), 0);
+    /* DDict ptr table size is a multiple of 2, use size - 1 as mask to get index within [0, hashSet->ddictPtrTableSize) */
+    return hash & (hashSet->ddictPtrTableSize - 1);
+}
+
+/* Adds DDict to a hashset without resizing it.
+ * If inserting a DDict with a dictID that already exists in the set, replaces the one in the set.
+ * Returns 0 if successful, or a zstd error code if something went wrong.
+ */
+static size_t ZSTD_DDictHashSet_emplaceDDict(ZSTD_DDictHashSet* hashSet, const ZSTD_DDict* ddict) {
+    const U32 dictID = ZSTD_getDictID_fromDDict(ddict);
+    size_t idx = ZSTD_DDictHashSet_getIndex(hashSet, dictID);
+    const size_t idxRangeMask = hashSet->ddictPtrTableSize - 1;
+    RETURN_ERROR_IF(hashSet->ddictPtrCount == hashSet->ddictPtrTableSize, GENERIC, "Hash set is full!");
+    DEBUGLOG(4, "Hashed index: for dictID: %u is %zu", dictID, idx);
+    while (hashSet->ddictPtrTable[idx] != NULL) {
+        /* Replace existing ddict if inserting ddict with same dictID */
+        if (ZSTD_getDictID_fromDDict(hashSet->ddictPtrTable[idx]) == dictID) {
+            DEBUGLOG(4, "DictID already exists, replacing rather than adding");
+            hashSet->ddictPtrTable[idx] = ddict;
+            return 0;
+        }
+        idx &= idxRangeMask;
+        idx++;
+    }
+    DEBUGLOG(4, "Final idx after probing for dictID %u is: %zu", dictID, idx);
+    hashSet->ddictPtrTable[idx] = ddict;
+    hashSet->ddictPtrCount++;
+    return 0;
+}
+
+/* Expands hash table by factor of DDICT_HASHSET_RESIZE_FACTOR and
+ * rehashes all values, allocates new table, frees old table.
+ * Returns 0 on success, otherwise a zstd error code.
+ */
+static size_t ZSTD_DDictHashSet_expand(ZSTD_DDictHashSet* hashSet, ZSTD_customMem customMem) {
+    size_t newTableSize = hashSet->ddictPtrTableSize * DDICT_HASHSET_RESIZE_FACTOR;
+    const ZSTD_DDict** newTable = (const ZSTD_DDict**)ZSTD_customCalloc(sizeof(ZSTD_DDict*) * newTableSize, customMem);
+    const ZSTD_DDict** oldTable = hashSet->ddictPtrTable;
+    size_t oldTableSize = hashSet->ddictPtrTableSize;
+    size_t i;
+
+    DEBUGLOG(4, "Expanding DDict hash table! Old size: %zu new size: %zu", oldTableSize, newTableSize);
+    RETURN_ERROR_IF(!newTable, memory_allocation, "Expanded hashset allocation failed!");
+    hashSet->ddictPtrTable = newTable;
+    hashSet->ddictPtrTableSize = newTableSize;
+    hashSet->ddictPtrCount = 0;
+    for (i = 0; i < oldTableSize; ++i) {
+        if (oldTable[i] != NULL) {
+            FORWARD_IF_ERROR(ZSTD_DDictHashSet_emplaceDDict(hashSet, oldTable[i]), "");
+        }
+    }
+    ZSTD_customFree((void*)oldTable, customMem);
+    DEBUGLOG(4, "Finished re-hash");
+    return 0;
+}
+
+/* Fetches a DDict with the given dictID
+ * Returns the ZSTD_DDict* with the requested dictID. If it doesn't exist, then returns NULL.
+ */
+static const ZSTD_DDict* ZSTD_DDictHashSet_getDDict(ZSTD_DDictHashSet* hashSet, U32 dictID) {
+    size_t idx = ZSTD_DDictHashSet_getIndex(hashSet, dictID);
+    const size_t idxRangeMask = hashSet->ddictPtrTableSize - 1;
+    DEBUGLOG(4, "Hashed index: for dictID: %u is %zu", dictID, idx);
+    for (;;) {
+        size_t currDictID = ZSTD_getDictID_fromDDict(hashSet->ddictPtrTable[idx]);
+        if (currDictID == dictID || currDictID == 0) {
+            /* currDictID == 0 implies a NULL ddict entry */
+            break;
+        } else {
+            idx &= idxRangeMask;    /* Goes to start of table when we reach the end */
+            idx++;
+        }
+    }
+    DEBUGLOG(4, "Final idx after probing for dictID %u is: %zu", dictID, idx);
+    return hashSet->ddictPtrTable[idx];
+}
+
+/* Allocates space for and returns a ddict hash set
+ * The hash set's ZSTD_DDict* table has all values automatically set to NULL to begin with.
+ * Returns NULL if allocation failed.
+ */
+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");
+    ret->ddictPtrTable = (const ZSTD_DDict**)ZSTD_customCalloc(DDICT_HASHSET_TABLE_BASE_SIZE * sizeof(ZSTD_DDict*), customMem);
+    ret->ddictPtrTableSize = DDICT_HASHSET_TABLE_BASE_SIZE;
+    ret->ddictPtrCount = 0;
+    if (!ret || !ret->ddictPtrTable) {
+        return NULL;
+    }
+    return ret;
+}
+
+/* Frees the table of ZSTD_DDict* within a hashset, then frees the hashset itself.
+ * Note: The ZSTD_DDict* within the table are NOT freed.
+ */
+static void ZSTD_freeDDictHashSet(ZSTD_DDictHashSet* hashSet, ZSTD_customMem customMem) {
+    DEBUGLOG(4, "Freeing ddict hash set");
+    if (hashSet && hashSet->ddictPtrTable) {
+        ZSTD_customFree((void*)hashSet->ddictPtrTable, customMem);
+    }
+    if (hashSet) {
+        ZSTD_customFree(hashSet, customMem);
+    }
+}
+
+/* Public function: Adds a DDict into the ZSTD_DDictHashSet, possibly triggering a resize of the hash set.
+ * Returns 0 on success, or a ZSTD error.
+ */
+static size_t ZSTD_DDictHashSet_addDDict(ZSTD_DDictHashSet* hashSet, const ZSTD_DDict* ddict, ZSTD_customMem customMem) {
+    DEBUGLOG(4, "Adding dict ID: %u to hashset with - Count: %zu Tablesize: %zu", ZSTD_getDictID_fromDDict(ddict), hashSet->ddictPtrCount, hashSet->ddictPtrTableSize);
+    if (hashSet->ddictPtrCount * DDICT_HASHSET_MAX_LOAD_FACTOR_COUNT_MULT / hashSet->ddictPtrTableSize * DDICT_HASHSET_MAX_LOAD_FACTOR_SIZE_MULT != 0) {
+        FORWARD_IF_ERROR(ZSTD_DDictHashSet_expand(hashSet, customMem), "");
+    }
+    FORWARD_IF_ERROR(ZSTD_DDictHashSet_emplaceDDict(hashSet, ddict), "");
+    return 0;
+}
+
 /*-*************************************************************
 *   Context management
 ***************************************************************/
@@ -101,6 +240,7 @@ static void ZSTD_DCtx_resetParameters(ZSTD_DCtx* dctx)
     dctx->maxWindowSize = ZSTD_MAXWINDOWSIZE_DEFAULT;
     dctx->outBufferMode = ZSTD_bm_buffered;
     dctx->forceIgnoreChecksum = ZSTD_d_validateChecksum;
+    dctx->refMultipleDDicts = ZSTD_rmd_refSingleDDict;
 }
 
 static void ZSTD_initDCtx_internal(ZSTD_DCtx* dctx)
@@ -120,8 +260,8 @@ static void ZSTD_initDCtx_internal(ZSTD_DCtx* dctx)
     dctx->noForwardProgress = 0;
     dctx->oversizedDuration = 0;
     dctx->bmi2 = ZSTD_cpuid_bmi2(ZSTD_cpuid());
+    dctx->ddictSet = NULL;
     ZSTD_DCtx_resetParameters(dctx);
-    dctx->validateChecksum = 1;
 #ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
     dctx->dictContentEndForFuzzing = NULL;
 #endif
@@ -178,6 +318,10 @@ size_t ZSTD_freeDCtx(ZSTD_DCtx* dctx)
         if (dctx->legacyContext)
             ZSTD_freeLegacyStreamContext(dctx->legacyContext, dctx->previousLegacyVersion);
 #endif
+        if (dctx->ddictSet) {
+            ZSTD_freeDDictHashSet(dctx->ddictSet, cMem);
+            dctx->ddictSet = NULL;
+        }
         ZSTD_customFree(dctx, cMem);
         return 0;
     }
@@ -190,6 +334,29 @@ void ZSTD_copyDCtx(ZSTD_DCtx* dstDCtx, const ZSTD_DCtx* srcDCtx)
     ZSTD_memcpy(dstDCtx, srcDCtx, toCopy);  /* no need to copy workspace */
 }
 
+/* Given a dctx with a digested frame params, re-selects the correct ZSTD_DDict based on
+ * the requested dict ID from the frame. If there exists a reference to the correct ZSTD_DDict, then
+ * accordingly sets the ddict to be used to decompress the frame.
+ *
+ * If no DDict is found, then no action is taken, and the ZSTD_DCtx::ddict remains as-is.
+ *
+ * ZSTD_d_refMultipleDDicts must be enabled for this function to be called.
+ */
+static void ZSTD_DCtx_selectFrameDDict(ZSTD_DCtx* dctx) {
+    assert(dctx->refMultipleDDicts && dctx->ddictSet);
+    DEBUGLOG(4, "Adjusting DDict based on requested dict ID from frame");
+    if (dctx->ddict) {
+        const ZSTD_DDict* frameDDict = ZSTD_DDictHashSet_getDDict(dctx->ddictSet, dctx->fParams.dictID);
+        if (frameDDict) {
+            DEBUGLOG(4, "DDict found!");
+            ZSTD_clearDict(dctx);
+            dctx->dictID = dctx->fParams.dictID;
+            dctx->ddict = frameDDict;
+            dctx->dictUses = ZSTD_use_indefinitely;
+        }
+    }
+}
+
 
 /*-*************************************************************
  *   Frame header decoding
@@ -213,6 +380,19 @@ 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.
@@ -336,7 +516,6 @@ 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
@@ -377,6 +556,37 @@ 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
@@ -441,12 +651,19 @@ unsigned long long ZSTD_getDecompressedSize(const void* src, size_t srcSize)
 
 /** ZSTD_decodeFrameHeader() :
  * `headerSize` must be the size provided by ZSTD_frameHeaderSize().
+ * If multiple DDict references are enabled, also will choose the correct DDict to use.
  * @return : 0 if success, or an error code, which can be tested using ZSTD_isError() */
 static size_t ZSTD_decodeFrameHeader(ZSTD_DCtx* dctx, const void* src, size_t headerSize)
 {
     size_t const result = ZSTD_getFrameHeader_advanced(&(dctx->fParams), src, headerSize, dctx->format);
     if (ZSTD_isError(result)) return result;    /* invalid header */
     RETURN_ERROR_IF(result>0, srcSize_wrong, "headerSize too small");
+
+    /* Reference DDict requested by frame if dctx references multiple ddicts */
+    if (dctx->refMultipleDDicts == ZSTD_rmd_refMultipleDDicts && dctx->ddictSet) {
+        ZSTD_DCtx_selectFrameDDict(dctx);
+    }
+
 #ifndef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
     /* Skip the dictID check in fuzzing mode, because it makes the search
      * harder.
@@ -456,6 +673,7 @@ static size_t ZSTD_decodeFrameHeader(ZSTD_DCtx* dctx, const void* src, size_t he
 #endif
     dctx->validateChecksum = (dctx->fParams.checksumFlag && !dctx->forceIgnoreChecksum) ? 1 : 0;
     if (dctx->validateChecksum) XXH64_reset(&dctx->xxhState, 0);
+    dctx->processedCSize += headerSize;
     return 0;
 }
 
@@ -578,7 +796,7 @@ unsigned long long ZSTD_decompressBound(const void* src, size_t srcSize)
 size_t ZSTD_insertBlock(ZSTD_DCtx* dctx, const void* blockStart, size_t blockSize)
 {
     DEBUGLOG(5, "ZSTD_insertBlock: %u bytes", (unsigned)blockSize);
-    ZSTD_checkContinuity(dctx, blockStart);
+    ZSTD_checkContinuity(dctx, blockStart, blockSize);
     dctx->previousDstEnd = (const char*)blockStart + blockSize;
     return blockSize;
 }
@@ -610,6 +828,32 @@ static size_t ZSTD_setRleBlock(void* dst, size_t dstCapacity,
     return regenSize;
 }
 
+static void ZSTD_DCtx_trace_end(ZSTD_DCtx const* dctx, U64 uncompressedSize, U64 compressedSize, unsigned streaming)
+{
+#if ZSTD_TRACE
+    if (dctx->traceCtx && ZSTD_trace_decompress_end != NULL) {
+        ZSTD_Trace trace;
+        ZSTD_memset(&trace, 0, sizeof(trace));
+        trace.version = ZSTD_VERSION_NUMBER;
+        trace.streaming = streaming;
+        if (dctx->ddict) {
+            trace.dictionaryID = ZSTD_getDictID_fromDDict(dctx->ddict);
+            trace.dictionarySize = ZSTD_DDict_dictSize(dctx->ddict);
+            trace.dictionaryIsCold = dctx->ddictIsCold;
+        }
+        trace.uncompressedSize = (size_t)uncompressedSize;
+        trace.compressedSize = (size_t)compressedSize;
+        trace.dctx = dctx;
+        ZSTD_trace_decompress_end(dctx->traceCtx, &trace);
+    }
+#else
+    (void)dctx;
+    (void)uncompressedSize;
+    (void)compressedSize;
+    (void)streaming;
+#endif
+}
+
 
 /*! ZSTD_decompressFrame() :
  * @dctx must be properly initialized
@@ -619,8 +863,9 @@ static size_t ZSTD_decompressFrame(ZSTD_DCtx* dctx,
                                    void* dst, size_t dstCapacity,
                              const void** srcPtr, size_t *srcSizePtr)
 {
-    const BYTE* ip = (const BYTE*)(*srcPtr);
-    BYTE* const ostart = (BYTE* const)dst;
+    const BYTE* const istart = (const BYTE*)(*srcPtr);
+    const BYTE* ip = istart;
+    BYTE* const ostart = (BYTE*)dst;
     BYTE* const oend = dstCapacity != 0 ? ostart + dstCapacity : ostart;
     BYTE* op = ostart;
     size_t remainingSrcSize = *srcSizePtr;
@@ -695,7 +940,7 @@ static size_t ZSTD_decompressFrame(ZSTD_DCtx* dctx,
         ip += 4;
         remainingSrcSize -= 4;
     }
-
+    ZSTD_DCtx_trace_end(dctx, (U64)(op-ostart), (U64)(ip-istart), /* streaming */ 0);
     /* Allow caller to get size read */
     *srcPtr = ip;
     *srcSizePtr = remainingSrcSize;
@@ -764,7 +1009,7 @@ static size_t ZSTD_decompressMultiFrame(ZSTD_DCtx* dctx,
              * use this in all cases but ddict */
             FORWARD_IF_ERROR(ZSTD_decompressBegin_usingDict(dctx, dict, dictSize), "");
         }
-        ZSTD_checkContinuity(dctx, dst);
+        ZSTD_checkContinuity(dctx, dst, dstCapacity);
 
         {   const size_t res = ZSTD_decompressFrame(dctx, dst, dstCapacity,
                                                     &src, &srcSize);
@@ -899,7 +1144,9 @@ size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, c
     DEBUGLOG(5, "ZSTD_decompressContinue (srcSize:%u)", (unsigned)srcSize);
     /* Sanity check */
     RETURN_ERROR_IF(srcSize != ZSTD_nextSrcSizeToDecompressWithInputSize(dctx, srcSize), srcSize_wrong, "not allowed");
-    if (dstCapacity) ZSTD_checkContinuity(dctx, dst);
+    ZSTD_checkContinuity(dctx, dst, dstCapacity);
+
+    dctx->processedCSize += srcSize;
 
     switch (dctx->stage)
     {
@@ -1004,6 +1251,7 @@ size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, c
                     dctx->expected = 4;
                     dctx->stage = ZSTDds_checkChecksum;
                 } else {
+                    ZSTD_DCtx_trace_end(dctx, dctx->decodedSize, dctx->processedCSize, /* streaming */ 1);
                     dctx->expected = 0;   /* ends here */
                     dctx->stage = ZSTDds_getFrameHeaderSize;
                 }
@@ -1023,6 +1271,7 @@ size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, c
                 DEBUGLOG(4, "ZSTD_decompressContinue: checksum : calculated %08X :: %08X read", (unsigned)h32, (unsigned)check32);
                 RETURN_ERROR_IF(check32 != h32, checksum_wrong, "");
             }
+            ZSTD_DCtx_trace_end(dctx, dctx->decodedSize, dctx->processedCSize, /* streaming */ 1);
             dctx->expected = 0;
             dctx->stage = ZSTDds_getFrameHeaderSize;
             return 0;
@@ -1176,8 +1425,12 @@ static size_t ZSTD_decompress_insertDictionary(ZSTD_DCtx* dctx, const void* dict
 size_t ZSTD_decompressBegin(ZSTD_DCtx* dctx)
 {
     assert(dctx != NULL);
+#if ZSTD_TRACE
+    dctx->traceCtx = (ZSTD_trace_decompress_begin != NULL) ? ZSTD_trace_decompress_begin(dctx) : 0;
+#endif
     dctx->expected = ZSTD_startingInputLength(dctx->format);  /* dctx->format must be properly set */
     dctx->stage = ZSTDds_getFrameHeaderSize;
+    dctx->processedCSize = 0;
     dctx->decodedSize = 0;
     dctx->previousDstEnd = NULL;
     dctx->prefixStart = NULL;
@@ -1391,6 +1644,16 @@ size_t ZSTD_DCtx_refDDict(ZSTD_DCtx* dctx, const ZSTD_DDict* ddict)
     if (ddict) {
         dctx->ddict = ddict;
         dctx->dictUses = ZSTD_use_indefinitely;
+        if (dctx->refMultipleDDicts == ZSTD_rmd_refMultipleDDicts) {
+            if (dctx->ddictSet == NULL) {
+                dctx->ddictSet = ZSTD_createDDictHashSet(dctx->customMem);
+                if (!dctx->ddictSet) {
+                    RETURN_ERROR(memory_allocation, "Failed to allocate memory for hash set!");
+                }
+            }
+            assert(!dctx->staticSize);  /* Impossible: ddictSet cannot have been allocated if static dctx */
+            FORWARD_IF_ERROR(ZSTD_DDictHashSet_addDDict(dctx->ddictSet, ddict, dctx->customMem), "");
+        }
     }
     return 0;
 }
@@ -1436,6 +1699,10 @@ ZSTD_bounds ZSTD_dParam_getBounds(ZSTD_dParameter dParam)
             bounds.lowerBound = (int)ZSTD_d_validateChecksum;
             bounds.upperBound = (int)ZSTD_d_ignoreChecksum;
             return bounds;
+        case ZSTD_d_refMultipleDDicts:
+            bounds.lowerBound = (int)ZSTD_rmd_refSingleDDict;
+            bounds.upperBound = (int)ZSTD_rmd_refMultipleDDicts;
+            return bounds;
         default:;
     }
     bounds.error = ERROR(parameter_unsupported);
@@ -1473,6 +1740,9 @@ size_t ZSTD_DCtx_getParameter(ZSTD_DCtx* dctx, ZSTD_dParameter param, int* value
         case ZSTD_d_forceIgnoreChecksum:
             *value = (int)dctx->forceIgnoreChecksum;
             return 0;
+        case ZSTD_d_refMultipleDDicts:
+            *value = (int)dctx->refMultipleDDicts;
+            return 0;
         default:;
     }
     RETURN_ERROR(parameter_unsupported, "");
@@ -1499,6 +1769,13 @@ size_t ZSTD_DCtx_setParameter(ZSTD_DCtx* dctx, ZSTD_dParameter dParam, int value
             CHECK_DBOUNDS(ZSTD_d_forceIgnoreChecksum, value);
             dctx->forceIgnoreChecksum = (ZSTD_forceIgnoreChecksum_e)value;
             return 0;
+        case ZSTD_d_refMultipleDDicts:
+            CHECK_DBOUNDS(ZSTD_d_refMultipleDDicts, value);
+            if (dctx->staticSize != 0) {
+                RETURN_ERROR(parameter_unsupported, "Static dctx does not support multiple DDicts!");
+            }
+            dctx->refMultipleDDicts = (ZSTD_refMultipleDDicts_e)value;
+            return 0;
         default:;
     }
     RETURN_ERROR(parameter_unsupported, "");
@@ -1680,6 +1957,9 @@ size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inB
             }   }
 #endif
             {   size_t const hSize = ZSTD_getFrameHeader_advanced(&zds->fParams, zds->headerBuffer, zds->lhSize, zds->format);
+                if (zds->refMultipleDDicts && zds->ddictSet) {
+                    ZSTD_DCtx_selectFrameDDict(zds);
+                }
                 DEBUGLOG(5, "header size : %u", (U32)hSize);
                 if (ZSTD_isError(hSize)) {
 #if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
diff --git a/zstd_decompress_block.c b/zstd_decompress_block.c
index 19ddb7a..8818dea 100644
--- a/zstd_decompress_block.c
+++ b/zstd_decompress_block.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -577,7 +577,7 @@ static size_t ZSTD_buildSeqTable(ZSTD_seqSymbol* DTableSpace, const ZSTD_seqSymb
 size_t ZSTD_decodeSeqHeaders(ZSTD_DCtx* dctx, int* nbSeqPtr,
                              const void* src, size_t srcSize)
 {
-    const BYTE* const istart = (const BYTE* const)src;
+    const BYTE* const istart = (const BYTE*)src;
     const BYTE* const iend = istart + srcSize;
     const BYTE* ip = istart;
     int nbSeq;
@@ -658,7 +658,6 @@ typedef struct {
     size_t litLength;
     size_t matchLength;
     size_t offset;
-    const BYTE* match;
 } seq_t;
 
 typedef struct {
@@ -672,9 +671,6 @@ typedef struct {
     ZSTD_fseState stateOffb;
     ZSTD_fseState stateML;
     size_t prevOffset[ZSTD_REP_NUM];
-    const BYTE* prefixStart;
-    const BYTE* dictEnd;
-    size_t pos;
 } seqState_t;
 
 /*! ZSTD_overlapCopy8() :
@@ -909,20 +905,10 @@ ZSTD_initFseState(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD, const ZSTD_seqS
 }
 
 FORCE_INLINE_TEMPLATE void
-ZSTD_updateFseState(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD)
+ZSTD_updateFseStateWithDInfo(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD, U16 nextState, U32 nbBits)
 {
-    ZSTD_seqSymbol const DInfo = DStatePtr->table[DStatePtr->state];
-    U32 const nbBits = DInfo.nbBits;
     size_t const lowBits = BIT_readBits(bitD, nbBits);
-    DStatePtr->state = DInfo.nextState + lowBits;
-}
-
-FORCE_INLINE_TEMPLATE void
-ZSTD_updateFseStateWithDInfo(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD, ZSTD_seqSymbol const DInfo)
-{
-    U32 const nbBits = DInfo.nbBits;
-    size_t const lowBits = BIT_readBits(bitD, nbBits);
-    DStatePtr->state = DInfo.nextState + lowBits;
+    DStatePtr->state = nextState + lowBits;
 }
 
 /* We need to add at most (ZSTD_WINDOWLOG_MAX_32 - 1) bits to read the maximum
@@ -936,116 +922,105 @@ ZSTD_updateFseStateWithDInfo(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD, ZSTD
         : 0)
 
 typedef enum { ZSTD_lo_isRegularOffset, ZSTD_lo_isLongOffset=1 } ZSTD_longOffset_e;
-typedef enum { ZSTD_p_noPrefetch=0, ZSTD_p_prefetch=1 } ZSTD_prefetch_e;
 
 FORCE_INLINE_TEMPLATE seq_t
-ZSTD_decodeSequence(seqState_t* seqState, const ZSTD_longOffset_e longOffsets, const ZSTD_prefetch_e prefetch)
+ZSTD_decodeSequence(seqState_t* seqState, const ZSTD_longOffset_e longOffsets)
 {
     seq_t seq;
-    ZSTD_seqSymbol const llDInfo = seqState->stateLL.table[seqState->stateLL.state];
-    ZSTD_seqSymbol const mlDInfo = seqState->stateML.table[seqState->stateML.state];
-    ZSTD_seqSymbol const ofDInfo = seqState->stateOffb.table[seqState->stateOffb.state];
-    U32 const llBase = llDInfo.baseValue;
-    U32 const mlBase = mlDInfo.baseValue;
-    U32 const ofBase = ofDInfo.baseValue;
-    BYTE const llBits = llDInfo.nbAdditionalBits;
-    BYTE const mlBits = mlDInfo.nbAdditionalBits;
-    BYTE const ofBits = ofDInfo.nbAdditionalBits;
-    BYTE const totalBits = llBits+mlBits+ofBits;
-
-    /* sequence */
-    {   size_t offset;
-        if (ofBits > 1) {
-            ZSTD_STATIC_ASSERT(ZSTD_lo_isLongOffset == 1);
-            ZSTD_STATIC_ASSERT(LONG_OFFSETS_MAX_EXTRA_BITS_32 == 5);
-            assert(ofBits <= MaxOff);
-            if (MEM_32bits() && longOffsets && (ofBits >= STREAM_ACCUMULATOR_MIN_32)) {
-                U32 const extraBits = ofBits - MIN(ofBits, 32 - seqState->DStream.bitsConsumed);
-                offset = ofBase + (BIT_readBitsFast(&seqState->DStream, ofBits - extraBits) << extraBits);
-                BIT_reloadDStream(&seqState->DStream);
-                if (extraBits) offset += BIT_readBitsFast(&seqState->DStream, extraBits);
-                assert(extraBits <= LONG_OFFSETS_MAX_EXTRA_BITS_32);   /* to avoid another reload */
-            } else {
-                offset = ofBase + BIT_readBitsFast(&seqState->DStream, ofBits/*>0*/);   /* <=  (ZSTD_WINDOWLOG_MAX-1) bits */
-                if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream);
-            }
-            seqState->prevOffset[2] = seqState->prevOffset[1];
-            seqState->prevOffset[1] = seqState->prevOffset[0];
-            seqState->prevOffset[0] = offset;
-        } else {
-            U32 const ll0 = (llBase == 0);
-            if (LIKELY((ofBits == 0))) {
-                if (LIKELY(!ll0))
-                    offset = seqState->prevOffset[0];
-                else {
-                    offset = seqState->prevOffset[1];
-                    seqState->prevOffset[1] = seqState->prevOffset[0];
-                    seqState->prevOffset[0] = offset;
+    const ZSTD_seqSymbol* const llDInfo = seqState->stateLL.table + seqState->stateLL.state;
+    const ZSTD_seqSymbol* const mlDInfo = seqState->stateML.table + seqState->stateML.state;
+    const ZSTD_seqSymbol* const ofDInfo = seqState->stateOffb.table + seqState->stateOffb.state;
+    seq.matchLength = mlDInfo->baseValue;
+    seq.litLength = llDInfo->baseValue;
+    {   U32 const ofBase = ofDInfo->baseValue;
+        BYTE const llBits = llDInfo->nbAdditionalBits;
+        BYTE const mlBits = mlDInfo->nbAdditionalBits;
+        BYTE const ofBits = ofDInfo->nbAdditionalBits;
+        BYTE const totalBits = llBits+mlBits+ofBits;
+
+        U16 const llNext = llDInfo->nextState;
+        U16 const mlNext = mlDInfo->nextState;
+        U16 const ofNext = ofDInfo->nextState;
+        U32 const llnbBits = llDInfo->nbBits;
+        U32 const mlnbBits = mlDInfo->nbBits;
+        U32 const ofnbBits = ofDInfo->nbBits;
+        /*
+         * As gcc has better branch and block analyzers, sometimes it is only
+         * valuable to mark likelyness for clang, it gives around 3-4% of
+         * performance.
+         */
+
+        /* sequence */
+        {   size_t offset;
+    #if defined(__clang__)
+            if (LIKELY(ofBits > 1)) {
+    #else
+            if (ofBits > 1) {
+    #endif
+                ZSTD_STATIC_ASSERT(ZSTD_lo_isLongOffset == 1);
+                ZSTD_STATIC_ASSERT(LONG_OFFSETS_MAX_EXTRA_BITS_32 == 5);
+                assert(ofBits <= MaxOff);
+                if (MEM_32bits() && longOffsets && (ofBits >= STREAM_ACCUMULATOR_MIN_32)) {
+                    U32 const extraBits = ofBits - MIN(ofBits, 32 - seqState->DStream.bitsConsumed);
+                    offset = ofBase + (BIT_readBitsFast(&seqState->DStream, ofBits - extraBits) << extraBits);
+                    BIT_reloadDStream(&seqState->DStream);
+                    if (extraBits) offset += BIT_readBitsFast(&seqState->DStream, extraBits);
+                    assert(extraBits <= LONG_OFFSETS_MAX_EXTRA_BITS_32);   /* to avoid another reload */
+                } else {
+                    offset = ofBase + BIT_readBitsFast(&seqState->DStream, ofBits/*>0*/);   /* <=  (ZSTD_WINDOWLOG_MAX-1) bits */
+                    if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream);
                 }
+                seqState->prevOffset[2] = seqState->prevOffset[1];
+                seqState->prevOffset[1] = seqState->prevOffset[0];
+                seqState->prevOffset[0] = offset;
             } else {
-                offset = ofBase + ll0 + BIT_readBitsFast(&seqState->DStream, 1);
-                {   size_t temp = (offset==3) ? seqState->prevOffset[0] - 1 : seqState->prevOffset[offset];
-                    temp += !temp;   /* 0 is not valid; input is corrupted; force offset to 1 */
-                    if (offset != 1) seqState->prevOffset[2] = seqState->prevOffset[1];
-                    seqState->prevOffset[1] = seqState->prevOffset[0];
-                    seqState->prevOffset[0] = offset = temp;
-        }   }   }
-        seq.offset = offset;
-    }
-
-    seq.matchLength = mlBase;
-    if (mlBits > 0)
-        seq.matchLength += BIT_readBitsFast(&seqState->DStream, mlBits/*>0*/);
-
-    if (MEM_32bits() && (mlBits+llBits >= STREAM_ACCUMULATOR_MIN_32-LONG_OFFSETS_MAX_EXTRA_BITS_32))
-        BIT_reloadDStream(&seqState->DStream);
-    if (MEM_64bits() && UNLIKELY(totalBits >= STREAM_ACCUMULATOR_MIN_64-(LLFSELog+MLFSELog+OffFSELog)))
-        BIT_reloadDStream(&seqState->DStream);
-    /* Ensure there are enough bits to read the rest of data in 64-bit mode. */
-    ZSTD_STATIC_ASSERT(16+LLFSELog+MLFSELog+OffFSELog < STREAM_ACCUMULATOR_MIN_64);
-
-    seq.litLength = llBase;
-    if (llBits > 0)
-        seq.litLength += BIT_readBitsFast(&seqState->DStream, llBits/*>0*/);
-
-    if (MEM_32bits())
-        BIT_reloadDStream(&seqState->DStream);
-
-    DEBUGLOG(6, "seq: litL=%u, matchL=%u, offset=%u",
-                (U32)seq.litLength, (U32)seq.matchLength, (U32)seq.offset);
-
-    if (prefetch == ZSTD_p_prefetch) {
-        size_t const pos = seqState->pos + seq.litLength;
-        const BYTE* const matchBase = (seq.offset > pos) ? seqState->dictEnd : seqState->prefixStart;
-        seq.match = matchBase + pos - seq.offset;  /* note : this operation can overflow when seq.offset is really too large, which can only happen when input is corrupted.
-                                                    * No consequence though : no memory access will occur, offset is only used for prefetching */
-        seqState->pos = pos + seq.matchLength;
-    }
-
-    /* ANS state update
-     * gcc-9.0.0 does 2.5% worse with ZSTD_updateFseStateWithDInfo().
-     * clang-9.2.0 does 7% worse with ZSTD_updateFseState().
-     * Naturally it seems like ZSTD_updateFseStateWithDInfo() should be the
-     * better option, so it is the default for other compilers. But, if you
-     * measure that it is worse, please put up a pull request.
-     */
-    {
-#if defined(__GNUC__) && !defined(__clang__)
-        const int kUseUpdateFseState = 1;
-#else
-        const int kUseUpdateFseState = 0;
-#endif
-        if (kUseUpdateFseState) {
-            ZSTD_updateFseState(&seqState->stateLL, &seqState->DStream);    /* <=  9 bits */
-            ZSTD_updateFseState(&seqState->stateML, &seqState->DStream);    /* <=  9 bits */
-            if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream);    /* <= 18 bits */
-            ZSTD_updateFseState(&seqState->stateOffb, &seqState->DStream);  /* <=  8 bits */
-        } else {
-            ZSTD_updateFseStateWithDInfo(&seqState->stateLL, &seqState->DStream, llDInfo);    /* <=  9 bits */
-            ZSTD_updateFseStateWithDInfo(&seqState->stateML, &seqState->DStream, mlDInfo);    /* <=  9 bits */
-            if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream);    /* <= 18 bits */
-            ZSTD_updateFseStateWithDInfo(&seqState->stateOffb, &seqState->DStream, ofDInfo);  /* <=  8 bits */
+                U32 const ll0 = (llDInfo->baseValue == 0);
+                if (LIKELY((ofBits == 0))) {
+                    offset = seqState->prevOffset[ll0];
+                    seqState->prevOffset[1] = seqState->prevOffset[!ll0];
+                    seqState->prevOffset[0] = offset;
+                } else {
+                    offset = ofBase + ll0 + BIT_readBitsFast(&seqState->DStream, 1);
+                    {   size_t temp = (offset==3) ? seqState->prevOffset[0] - 1 : seqState->prevOffset[offset];
+                        temp += !temp;   /* 0 is not valid; input is corrupted; force offset to 1 */
+                        if (offset != 1) seqState->prevOffset[2] = seqState->prevOffset[1];
+                        seqState->prevOffset[1] = seqState->prevOffset[0];
+                        seqState->prevOffset[0] = offset = temp;
+            }   }   }
+            seq.offset = offset;
         }
+
+    #if defined(__clang__)
+        if (UNLIKELY(mlBits > 0))
+    #else
+        if (mlBits > 0)
+    #endif
+            seq.matchLength += BIT_readBitsFast(&seqState->DStream, mlBits/*>0*/);
+
+        if (MEM_32bits() && (mlBits+llBits >= STREAM_ACCUMULATOR_MIN_32-LONG_OFFSETS_MAX_EXTRA_BITS_32))
+            BIT_reloadDStream(&seqState->DStream);
+        if (MEM_64bits() && UNLIKELY(totalBits >= STREAM_ACCUMULATOR_MIN_64-(LLFSELog+MLFSELog+OffFSELog)))
+            BIT_reloadDStream(&seqState->DStream);
+        /* Ensure there are enough bits to read the rest of data in 64-bit mode. */
+        ZSTD_STATIC_ASSERT(16+LLFSELog+MLFSELog+OffFSELog < STREAM_ACCUMULATOR_MIN_64);
+
+    #if defined(__clang__)
+        if (UNLIKELY(llBits > 0))
+    #else
+        if (llBits > 0)
+    #endif
+            seq.litLength += BIT_readBitsFast(&seqState->DStream, llBits/*>0*/);
+
+        if (MEM_32bits())
+            BIT_reloadDStream(&seqState->DStream);
+
+        DEBUGLOG(6, "seq: litL=%u, matchL=%u, offset=%u",
+                    (U32)seq.litLength, (U32)seq.matchLength, (U32)seq.offset);
+
+        ZSTD_updateFseStateWithDInfo(&seqState->stateLL, &seqState->DStream, llNext, llnbBits);    /* <=  9 bits */
+        ZSTD_updateFseStateWithDInfo(&seqState->stateML, &seqState->DStream, mlNext, mlnbBits);    /* <=  9 bits */
+        if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream);    /* <= 18 bits */
+        ZSTD_updateFseStateWithDInfo(&seqState->stateOffb, &seqState->DStream, ofNext, ofnbBits);  /* <=  8 bits */
     }
 
     return seq;
@@ -1108,7 +1083,7 @@ ZSTD_decompressSequences_body( ZSTD_DCtx* dctx,
 {
     const BYTE* ip = (const BYTE*)seqStart;
     const BYTE* const iend = ip + seqSize;
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     BYTE* const oend = ostart + maxDstSize;
     BYTE* op = ostart;
     const BYTE* litPtr = dctx->litPtr;
@@ -1122,7 +1097,6 @@ ZSTD_decompressSequences_body( ZSTD_DCtx* dctx,
     /* Regen sequences */
     if (nbSeq) {
         seqState_t seqState;
-        size_t error = 0;
         dctx->fseEntropy = 1;
         { U32 i; for (i=0; i<ZSTD_REP_NUM; i++) seqState.prevOffset[i] = dctx->entropy.rep[i]; }
         RETURN_ERROR_IF(
@@ -1156,13 +1130,14 @@ ZSTD_decompressSequences_body( ZSTD_DCtx* dctx,
          * If you see most cycles served out of the DSB you've hit the good case.
          * If it is pretty even then you may be in an okay case.
          *
-         * I've been able to reproduce this issue on the following CPUs:
+         * This issue has been reproduced on the following CPUs:
          *   - Kabylake: Macbook Pro (15-inch, 2019) 2.4 GHz Intel Core i9
          *               Use Instruments->Counters to get DSB/MITE cycles.
          *               I never got performance swings, but I was able to
          *               go from the good case of mostly DSB to half of the
          *               cycles served from MITE.
          *   - Coffeelake: Intel i9-9900k
+         *   - Coffeelake: Intel i7-9700k
          *
          * I haven't been able to reproduce the instability or DSB misses on any
          * of the following CPUS:
@@ -1175,33 +1150,35 @@ ZSTD_decompressSequences_body( ZSTD_DCtx* dctx,
          *
          *   https://gist.github.com/terrelln/9889fc06a423fd5ca6e99351564473f4
          */
+        __asm__(".p2align 6");
+        __asm__("nop");
         __asm__(".p2align 5");
         __asm__("nop");
+#  if __GNUC__ >= 9 && __GNUC__ < 11
+        /* better for gcc-9 and gcc-10, worse for clang and gcc-8, gcc-11 */
+        __asm__(".p2align 3");
+#  else
         __asm__(".p2align 4");
+#  endif
 #endif
         for ( ; ; ) {
-            seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset, ZSTD_p_noPrefetch);
+            seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
             size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequence, &litPtr, litEnd, prefixStart, vBase, dictEnd);
 #if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
             assert(!ZSTD_isError(oneSeqSize));
             if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
 #endif
+            if (UNLIKELY(ZSTD_isError(oneSeqSize)))
+                return oneSeqSize;
             DEBUGLOG(6, "regenerated sequence size : %u", (U32)oneSeqSize);
-            BIT_reloadDStream(&(seqState.DStream));
             op += oneSeqSize;
-            /* gcc and clang both don't like early returns in this loop.
-             * Instead break and check for an error at the end of the loop.
-             */
-            if (UNLIKELY(ZSTD_isError(oneSeqSize))) {
-                error = oneSeqSize;
+            if (UNLIKELY(!--nbSeq))
                 break;
-            }
-            if (UNLIKELY(!--nbSeq)) break;
+            BIT_reloadDStream(&(seqState.DStream));
         }
 
         /* check if reached exact end */
         DEBUGLOG(5, "ZSTD_decompressSequences_body: after decode loop, remaining nbSeq : %i", nbSeq);
-        if (ZSTD_isError(error)) return error;
         RETURN_ERROR_IF(nbSeq, corruption_detected, "");
         RETURN_ERROR_IF(BIT_reloadDStream(&seqState.DStream) < BIT_DStream_completed, corruption_detected, "");
         /* save reps for next block */
@@ -1232,6 +1209,24 @@ ZSTD_decompressSequences_default(ZSTD_DCtx* dctx,
 #endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG */
 
 #ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT
+
+FORCE_INLINE_TEMPLATE size_t
+ZSTD_prefetchMatch(size_t prefetchPos, seq_t const sequence,
+                   const BYTE* const prefixStart, const BYTE* const dictEnd)
+{
+    prefetchPos += sequence.litLength;
+    {   const BYTE* const matchBase = (sequence.offset > prefetchPos) ? dictEnd : prefixStart;
+        const BYTE* const match = matchBase + prefetchPos - sequence.offset; /* note : this operation can overflow when seq.offset is really too large, which can only happen when input is corrupted.
+                                                                              * No consequence though : memory address is only used for prefetching, not for dereferencing */
+        PREFETCH_L1(match); PREFETCH_L1(match+CACHELINE_SIZE);   /* note : it's safe to invoke PREFETCH() on any memory address, including invalid ones */
+    }
+    return prefetchPos + sequence.matchLength;
+}
+
+/* This decoding function employs prefetching
+ * to reduce latency impact of cache misses.
+ * It's generally employed when block contains a significant portion of long-distance matches
+ * or when coupled with a "cold" dictionary */
 FORCE_INLINE_TEMPLATE size_t
 ZSTD_decompressSequencesLong_body(
                                ZSTD_DCtx* dctx,
@@ -1242,7 +1237,7 @@ ZSTD_decompressSequencesLong_body(
 {
     const BYTE* ip = (const BYTE*)seqStart;
     const BYTE* const iend = ip + seqSize;
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     BYTE* const oend = ostart + maxDstSize;
     BYTE* op = ostart;
     const BYTE* litPtr = dctx->litPtr;
@@ -1254,18 +1249,17 @@ ZSTD_decompressSequencesLong_body(
 
     /* Regen sequences */
     if (nbSeq) {
-#define STORED_SEQS 4
+#define STORED_SEQS 8
 #define STORED_SEQS_MASK (STORED_SEQS-1)
-#define ADVANCED_SEQS 4
+#define ADVANCED_SEQS STORED_SEQS
         seq_t sequences[STORED_SEQS];
         int const seqAdvance = MIN(nbSeq, ADVANCED_SEQS);
         seqState_t seqState;
         int seqNb;
+        size_t prefetchPos = (size_t)(op-prefixStart); /* track position relative to prefixStart */
+
         dctx->fseEntropy = 1;
         { int i; for (i=0; i<ZSTD_REP_NUM; i++) seqState.prevOffset[i] = dctx->entropy.rep[i]; }
-        seqState.prefixStart = prefixStart;
-        seqState.pos = (size_t)(op-prefixStart);
-        seqState.dictEnd = dictEnd;
         assert(dst != NULL);
         assert(iend >= ip);
         RETURN_ERROR_IF(
@@ -1277,21 +1271,23 @@ ZSTD_decompressSequencesLong_body(
 
         /* prepare in advance */
         for (seqNb=0; (BIT_reloadDStream(&seqState.DStream) <= BIT_DStream_completed) && (seqNb<seqAdvance); seqNb++) {
-            sequences[seqNb] = ZSTD_decodeSequence(&seqState, isLongOffset, ZSTD_p_prefetch);
-            PREFETCH_L1(sequences[seqNb].match); PREFETCH_L1(sequences[seqNb].match + sequences[seqNb].matchLength - 1); /* note : it's safe to invoke PREFETCH() on any memory address, including invalid ones */
+            seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
+            prefetchPos = ZSTD_prefetchMatch(prefetchPos, sequence, prefixStart, dictEnd);
+            sequences[seqNb] = sequence;
         }
         RETURN_ERROR_IF(seqNb<seqAdvance, corruption_detected, "");
 
         /* decode and decompress */
         for ( ; (BIT_reloadDStream(&(seqState.DStream)) <= BIT_DStream_completed) && (seqNb<nbSeq) ; seqNb++) {
-            seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset, ZSTD_p_prefetch);
+            seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
             size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequences[(seqNb-ADVANCED_SEQS) & STORED_SEQS_MASK], &litPtr, litEnd, prefixStart, dictStart, dictEnd);
 #if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
             assert(!ZSTD_isError(oneSeqSize));
             if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequences[(seqNb-ADVANCED_SEQS) & STORED_SEQS_MASK], prefixStart, dictStart);
 #endif
             if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
-            PREFETCH_L1(sequence.match); PREFETCH_L1(sequence.match + sequence.matchLength - 1); /* note : it's safe to invoke PREFETCH() on any memory address, including invalid ones */
+
+            prefetchPos = ZSTD_prefetchMatch(prefetchPos, sequence, prefixStart, dictEnd);
             sequences[seqNb & STORED_SEQS_MASK] = sequence;
             op += oneSeqSize;
         }
@@ -1517,9 +1513,9 @@ ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
 }
 
 
-void ZSTD_checkContinuity(ZSTD_DCtx* dctx, const void* dst)
+void ZSTD_checkContinuity(ZSTD_DCtx* dctx, const void* dst, size_t dstSize)
 {
-    if (dst != dctx->previousDstEnd) {   /* not contiguous */
+    if (dst != dctx->previousDstEnd && dstSize > 0) {   /* not contiguous */
         dctx->dictEnd = dctx->previousDstEnd;
         dctx->virtualStart = (const char*)dst - ((const char*)(dctx->previousDstEnd) - (const char*)(dctx->prefixStart));
         dctx->prefixStart = dst;
@@ -1533,7 +1529,7 @@ size_t ZSTD_decompressBlock(ZSTD_DCtx* dctx,
                       const void* src, size_t srcSize)
 {
     size_t dSize;
-    ZSTD_checkContinuity(dctx, dst);
+    ZSTD_checkContinuity(dctx, dst, dstCapacity);
     dSize = ZSTD_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize, /* frame */ 0);
     dctx->previousDstEnd = (char*)dst + dSize;
     return dSize;
diff --git a/zstd_decompress_block.h b/zstd_decompress_block.h
index e6bd839..3bac9e9 100644
--- a/zstd_decompress_block.h
+++ b/zstd_decompress_block.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_decompress_internal.h b/zstd_decompress_internal.h
index b604b22..2215b09 100644
--- a/zstd_decompress_internal.h
+++ b/zstd_decompress_internal.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -99,6 +99,13 @@ typedef enum {
     ZSTD_use_once = 1            /* Use the dictionary once and set to ZSTD_dont_use */
 } ZSTD_dictUses_e;
 
+/* Hashset for storing references to multiple ZSTD_DDict within ZSTD_DCtx */
+typedef struct {
+    const ZSTD_DDict** ddictPtrTable;
+    size_t ddictPtrTableSize;
+    size_t ddictPtrCount;
+} ZSTD_DDictHashSet;
+
 struct ZSTD_DCtx_s
 {
     const ZSTD_seqSymbol* LLTptr;
@@ -113,6 +120,7 @@ struct ZSTD_DCtx_s
     const void* dictEnd;          /* end of previous segment */
     size_t expected;
     ZSTD_frameHeader fParams;
+    U64 processedCSize;
     U64 decodedSize;
     blockType_e bType;            /* used in ZSTD_decompressContinue(), store blockType between block header decoding and block decompression stages */
     ZSTD_dStage stage;
@@ -136,6 +144,8 @@ struct ZSTD_DCtx_s
     U32 dictID;
     int ddictIsCold;             /* if == 1 : dictionary is "new" for working context, and presumed "cold" (not in cpu cache) */
     ZSTD_dictUses_e dictUses;
+    ZSTD_DDictHashSet* ddictSet;                    /* Hash set for multiple ddicts */
+    ZSTD_refMultipleDDicts_e refMultipleDDicts;     /* User specified: if == 1, will allow references to multiple DDicts. Default == 0 (disabled) */
 
     /* streaming */
     ZSTD_dStreamStage streamStage;
@@ -166,6 +176,11 @@ struct ZSTD_DCtx_s
     void const* dictContentBeginForFuzzing;
     void const* dictContentEndForFuzzing;
 #endif
+
+    /* Tracing */
+#if ZSTD_TRACE
+    ZSTD_TraceCtx traceCtx;
+#endif
 };  /* typedef'd to ZSTD_DCtx within "zstd.h" */
 
 
@@ -184,7 +199,7 @@ size_t ZSTD_loadDEntropy(ZSTD_entropyDTables_t* entropy,
  *  If yes, do nothing (continue on current segment).
  *  If not, classify previous segment as "external dictionary", and start a new segment.
  *  This function cannot fail. */
-void ZSTD_checkContinuity(ZSTD_DCtx* dctx, const void* dst);
+void ZSTD_checkContinuity(ZSTD_DCtx* dctx, const void* dst, size_t dstSize);
 
 
 #endif /* ZSTD_DECOMPRESS_INTERNAL_H */
diff --git a/zstd_deps.h b/zstd_deps.h
index 0fb8b78..1421134 100644
--- a/zstd_deps.h
+++ b/zstd_deps.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Facebook, Inc.
+ * Copyright (c) Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_double_fast.c b/zstd_double_fast.c
index ef12a52..b9e3c5e 100644
--- a/zstd_double_fast.c
+++ b/zstd_double_fast.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -409,7 +409,7 @@ 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 */
-            & (repIndex > dictStartIndex))
+            & (offset_1 <= curr+1 - dictStartIndex)) /* note: we are searching at curr+1 */
           && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
             const BYTE* repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
             mLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixStart) + 4;
@@ -477,7 +477,7 @@ 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 */
-                    & (repIndex2 > dictStartIndex))
+                    & (offset_2 <= current2 - dictStartIndex))
                   && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
                     const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
                     size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
diff --git a/zstd_double_fast.h b/zstd_double_fast.h
index 73a2002..b4ee5ef 100644
--- a/zstd_double_fast.h
+++ b/zstd_double_fast.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_errors.h b/zstd_errors.h
index 6d0d003..fa3686b 100644
--- a/zstd_errors.h
+++ b/zstd_errors.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_fast.c b/zstd_fast.c
index db7ce83..b6a1a7e 100644
--- a/zstd_fast.c
+++ b/zstd_fast.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -242,7 +242,7 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic(
     assert(endIndex - prefixStartIndex <= maxDistance);
     (void)maxDistance; (void)endIndex;   /* these variables are not used when assert() is disabled */
 
-    /* ensure there will be no no underflow
+    /* ensure there will be no underflow
      * when translating a dict index into a local index */
     assert(prefixStartIndex >= (U32)(dictEnd - dictBase));
 
@@ -416,9 +416,9 @@ static size_t ZSTD_compressBlock_fast_extDict_generic(
         const BYTE* const repMatch = repBase + repIndex;
         hashTable[h] = curr;   /* update hash table */
         DEBUGLOG(7, "offset_1 = %u , curr = %u", offset_1, curr);
-        assert(offset_1 <= curr +1);   /* check repIndex */
 
-        if ( (((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow */ & (repIndex > dictStartIndex))
+        if ( ( ((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow */
+             & (offset_1 <= curr+1 - dictStartIndex) ) /* note: we are searching at curr+1 */
            && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
             const BYTE* const repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
             size_t const rLength = ZSTD_count_2segments(ip+1 +4, repMatch +4, iend, repMatchEnd, prefixStart) + 4;
@@ -453,7 +453,7 @@ 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) & (repIndex2 > 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;
diff --git a/zstd_fast.h b/zstd_fast.h
index 3e15cc8..9b56fb6 100644
--- a/zstd_fast.h
+++ b/zstd_fast.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_internal.h b/zstd_internal.h
index d94c592..3b56f1a 100644
--- a/zstd_internal.h
+++ b/zstd_internal.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -19,9 +19,6 @@
 /*-*************************************
 *  Dependencies
 ***************************************/
-#if !defined(ZSTD_NO_INTRINSICS) && defined(__ARM_NEON)
-#include <arm_neon.h>
-#endif
 #include "compiler.h"
 #include "mem.h"
 #include "debug.h"                 /* assert, DEBUGLOG, RAWLOG, g_debuglevel */
@@ -36,6 +33,11 @@
 #  define XXH_STATIC_LINKING_ONLY  /* XXH64_state_t */
 #endif
 #include "xxhash.h"                /* XXH_reset, update, digest */
+#ifndef ZSTD_NO_TRACE
+#  include "zstd_trace.h"
+#else
+#  define ZSTD_TRACE 0
+#endif
 
 #if defined (__cplusplus)
 extern "C" {
@@ -242,7 +244,7 @@ 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);
@@ -251,7 +253,7 @@ static void ZSTD_copy8(void* dst, const void* src) {
 
 #define COPY8(d,s) { ZSTD_copy8(d,s); d+=8; s+=8; }
 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));
 #else
     ZSTD_memcpy(dst, src, 16);
@@ -347,11 +349,18 @@ typedef enum {
 *  Private declarations
 *********************************************/
 typedef struct seqDef_s {
-    U32 offset;         /* Offset code of the sequence */
+    U32 offset;         /* offset == rawOffset + ZSTD_REP_NUM, or equivalently, offCode + 1 */
     U16 litLength;
     U16 matchLength;
 } seqDef;
 
+/* Controls whether seqStore has a single "long" litLength or matchLength. See seqStore_t. */
+typedef enum {
+    ZSTD_llt_none = 0,             /* no longLengthType */
+    ZSTD_llt_literalLength = 1,    /* represents a long literal */
+    ZSTD_llt_matchLength = 2       /* represents a long match */
+} ZSTD_longLengthType_e;
+
 typedef struct {
     seqDef* sequencesStart;
     seqDef* sequences;      /* ptr to end of sequences */
@@ -363,12 +372,12 @@ typedef struct {
     size_t maxNbSeq;
     size_t maxNbLit;
 
-    /* longLengthPos and longLengthID to allow us to represent either a single litLength or matchLength
+    /* longLengthPos and longLengthType to allow us to represent either a single litLength or matchLength
      * in the seqStore that has a value larger than U16 (if it exists). To do so, we increment
      * the existing value of the litLength or matchLength by 0x10000.
      */
-    U32   longLengthID;   /* 0 == no longLength; 1 == Represent the long literal; 2 == Represent the long match; */
-    U32   longLengthPos;  /* Index of the sequence to apply long length modification to */
+    ZSTD_longLengthType_e   longLengthType;
+    U32                     longLengthPos;  /* Index of the sequence to apply long length modification to */
 } seqStore_t;
 
 typedef struct {
@@ -378,7 +387,7 @@ typedef struct {
 
 /**
  * Returns the ZSTD_sequenceLength for the given sequences. It handles the decoding of long sequences
- * indicated by longLengthPos and longLengthID, and adds MINMATCH back to matchLength.
+ * indicated by longLengthPos and longLengthType, and adds MINMATCH back to matchLength.
  */
 MEM_STATIC ZSTD_sequenceLength ZSTD_getSequenceLength(seqStore_t const* seqStore, seqDef const* seq)
 {
@@ -386,10 +395,10 @@ MEM_STATIC ZSTD_sequenceLength ZSTD_getSequenceLength(seqStore_t const* seqStore
     seqLen.litLength = seq->litLength;
     seqLen.matchLength = seq->matchLength + MINMATCH;
     if (seqStore->longLengthPos == (U32)(seq - seqStore->sequencesStart)) {
-        if (seqStore->longLengthID == 1) {
+        if (seqStore->longLengthType == ZSTD_llt_literalLength) {
             seqLen.litLength += 0xFFFF;
         }
-        if (seqStore->longLengthID == 2) {
+        if (seqStore->longLengthType == ZSTD_llt_matchLength) {
             seqLen.matchLength += 0xFFFF;
         }
     }
diff --git a/zstd_lazy.c b/zstd_lazy.c
index 49ec1b0..70319d5 100644
--- a/zstd_lazy.c
+++ b/zstd_lazy.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -438,43 +438,9 @@ static size_t ZSTD_BtFindBestMatch_extDict_selectMLS (
     }
 }
 
-
-
-/* *********************************
-*  Hash Chain
+/***********************************
+* Dedicated dict search
 ***********************************/
-#define NEXT_IN_CHAIN(d, mask)   chainTable[(d) & (mask)]
-
-/* Update chains up to ip (excluded)
-   Assumption : always within prefix (i.e. not within extDict) */
-FORCE_INLINE_TEMPLATE U32 ZSTD_insertAndFindFirstIndex_internal(
-                        ZSTD_matchState_t* ms,
-                        const ZSTD_compressionParameters* const cParams,
-                        const BYTE* ip, U32 const mls)
-{
-    U32* const hashTable  = ms->hashTable;
-    const U32 hashLog = cParams->hashLog;
-    U32* const chainTable = ms->chainTable;
-    const U32 chainMask = (1 << cParams->chainLog) - 1;
-    const BYTE* const base = ms->window.base;
-    const U32 target = (U32)(ip - base);
-    U32 idx = ms->nextToUpdate;
-
-    while(idx < target) { /* catch up */
-        size_t const h = ZSTD_hashPtr(base+idx, hashLog, mls);
-        NEXT_IN_CHAIN(idx, chainMask) = hashTable[h];
-        hashTable[h] = idx;
-        idx++;
-    }
-
-    ms->nextToUpdate = target;
-    return hashTable[ZSTD_hashPtr(ip, hashLog, mls)];
-}
-
-U32 ZSTD_insertAndFindFirstIndex(ZSTD_matchState_t* ms, const BYTE* ip) {
-    const ZSTD_compressionParameters* const cParams = &ms->cParams;
-    return ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, ms->cParams.minMatch);
-}
 
 void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t* ms, const BYTE* const ip)
 {
@@ -484,7 +450,7 @@ void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t* ms, const B
     U32* const chainTable = ms->chainTable;
     U32 const chainSize = 1 << ms->cParams.chainLog;
     U32 idx = ms->nextToUpdate;
-    U32 const minChain = chainSize < target ? target - chainSize : idx;
+    U32 const minChain = chainSize < target - idx ? target - chainSize : idx;
     U32 const bucketSize = 1 << ZSTD_LAZY_DDSS_BUCKET_LOG;
     U32 const cacheSize = bucketSize - 1;
     U32 const chainAttempts = (1 << ms->cParams.searchLog) - cacheSize;
@@ -500,11 +466,10 @@ void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t* ms, const B
     U32* const tmpChainTable = hashTable + ((size_t)1 << hashLog);
     U32 const tmpChainSize = ((1 << ZSTD_LAZY_DDSS_BUCKET_LOG) - 1) << hashLog;
     U32 const tmpMinChain = tmpChainSize < target ? target - tmpChainSize : idx;
-
     U32 hashIdx;
 
     assert(ms->cParams.chainLog <= 24);
-    assert(ms->cParams.hashLog >= ms->cParams.chainLog);
+    assert(ms->cParams.hashLog > ms->cParams.chainLog);
     assert(idx != 0);
     assert(tmpMinChain <= minChain);
 
@@ -535,7 +500,7 @@ void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t* ms, const B
             if (count == cacheSize) {
                 for (count = 0; count < chainLimit;) {
                     if (i < minChain) {
-                        if (!i || countBeyondMinChain++ > cacheSize) {
+                        if (!i || ++countBeyondMinChain > cacheSize) {
                             /* only allow pulling `cacheSize` number of entries
                              * into the cache or chainTable beyond `minChain`,
                              * to replace the entries pulled out of the
@@ -591,6 +556,139 @@ void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t* ms, const B
     ms->nextToUpdate = target;
 }
 
+/* Returns the longest match length found in the dedicated dict search structure.
+ * If none are longer than the argument ml, then ml will be returned.
+ */
+FORCE_INLINE_TEMPLATE
+size_t ZSTD_dedicatedDictSearch_lazy_search(size_t* offsetPtr, size_t ml, U32 nbAttempts,
+                                            const ZSTD_matchState_t* const dms,
+                                            const BYTE* const ip, const BYTE* const iLimit,
+                                            const BYTE* const prefixStart, const U32 curr,
+                                            const U32 dictLimit, const size_t ddsIdx) {
+    const U32 ddsLowestIndex  = dms->window.dictLimit;
+    const BYTE* const ddsBase = dms->window.base;
+    const BYTE* const ddsEnd  = dms->window.nextSrc;
+    const U32 ddsSize         = (U32)(ddsEnd - ddsBase);
+    const U32 ddsIndexDelta   = dictLimit - ddsSize;
+    const U32 bucketSize      = (1 << ZSTD_LAZY_DDSS_BUCKET_LOG);
+    const U32 bucketLimit     = nbAttempts < bucketSize - 1 ? nbAttempts : bucketSize - 1;
+    U32 ddsAttempt;
+    U32 matchIndex;
+
+    for (ddsAttempt = 0; ddsAttempt < bucketSize - 1; ddsAttempt++) {
+        PREFETCH_L1(ddsBase + dms->hashTable[ddsIdx + ddsAttempt]);
+    }
+
+    {
+        U32 const chainPackedPointer = dms->hashTable[ddsIdx + bucketSize - 1];
+        U32 const chainIndex = chainPackedPointer >> 8;
+
+        PREFETCH_L1(&dms->chainTable[chainIndex]);
+    }
+
+    for (ddsAttempt = 0; ddsAttempt < bucketLimit; ddsAttempt++) {
+        size_t currentMl=0;
+        const BYTE* match;
+        matchIndex = dms->hashTable[ddsIdx + ddsAttempt];
+        match = ddsBase + matchIndex;
+
+        if (!matchIndex) {
+            return ml;
+        }
+
+        /* guaranteed by table construction */
+        (void)ddsLowestIndex;
+        assert(matchIndex >= ddsLowestIndex);
+        assert(match+4 <= ddsEnd);
+        if (MEM_read32(match) == MEM_read32(ip)) {
+            /* assumption : matchIndex <= dictLimit-4 (by table construction) */
+            currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, ddsEnd, prefixStart) + 4;
+        }
+
+        /* save best solution */
+        if (currentMl > ml) {
+            ml = currentMl;
+            *offsetPtr = curr - (matchIndex + ddsIndexDelta) + ZSTD_REP_MOVE;
+            if (ip+currentMl == iLimit) {
+                /* best possible, avoids read overflow on next attempt */
+                return ml;
+            }
+        }
+    }
+
+    {
+        U32 const chainPackedPointer = dms->hashTable[ddsIdx + bucketSize - 1];
+        U32 chainIndex = chainPackedPointer >> 8;
+        U32 const chainLength = chainPackedPointer & 0xFF;
+        U32 const chainAttempts = nbAttempts - ddsAttempt;
+        U32 const chainLimit = chainAttempts > chainLength ? chainLength : chainAttempts;
+        U32 chainAttempt;
+
+        for (chainAttempt = 0 ; chainAttempt < chainLimit; chainAttempt++) {
+            PREFETCH_L1(ddsBase + dms->chainTable[chainIndex + chainAttempt]);
+        }
+
+        for (chainAttempt = 0 ; chainAttempt < chainLimit; chainAttempt++, chainIndex++) {
+            size_t currentMl=0;
+            const BYTE* match;
+            matchIndex = dms->chainTable[chainIndex];
+            match = ddsBase + matchIndex;
+
+            /* guaranteed by table construction */
+            assert(matchIndex >= ddsLowestIndex);
+            assert(match+4 <= ddsEnd);
+            if (MEM_read32(match) == MEM_read32(ip)) {
+                /* assumption : matchIndex <= dictLimit-4 (by table construction) */
+                currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, ddsEnd, prefixStart) + 4;
+            }
+
+            /* save best solution */
+            if (currentMl > ml) {
+                ml = currentMl;
+                *offsetPtr = curr - (matchIndex + ddsIndexDelta) + ZSTD_REP_MOVE;
+                if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
+            }
+        }
+    }
+    return ml;
+}
+
+
+/* *********************************
+*  Hash Chain
+***********************************/
+#define NEXT_IN_CHAIN(d, mask)   chainTable[(d) & (mask)]
+
+/* Update chains up to ip (excluded)
+   Assumption : always within prefix (i.e. not within extDict) */
+FORCE_INLINE_TEMPLATE U32 ZSTD_insertAndFindFirstIndex_internal(
+                        ZSTD_matchState_t* ms,
+                        const ZSTD_compressionParameters* const cParams,
+                        const BYTE* ip, U32 const mls)
+{
+    U32* const hashTable  = ms->hashTable;
+    const U32 hashLog = cParams->hashLog;
+    U32* const chainTable = ms->chainTable;
+    const U32 chainMask = (1 << cParams->chainLog) - 1;
+    const BYTE* const base = ms->window.base;
+    const U32 target = (U32)(ip - base);
+    U32 idx = ms->nextToUpdate;
+
+    while(idx < target) { /* catch up */
+        size_t const h = ZSTD_hashPtr(base+idx, hashLog, mls);
+        NEXT_IN_CHAIN(idx, chainMask) = hashTable[h];
+        hashTable[h] = idx;
+        idx++;
+    }
+
+    ms->nextToUpdate = target;
+    return hashTable[ZSTD_hashPtr(ip, hashLog, mls)];
+}
+
+U32 ZSTD_insertAndFindFirstIndex(ZSTD_matchState_t* ms, const BYTE* ip) {
+    const ZSTD_compressionParameters* const cParams = &ms->cParams;
+    return ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, ms->cParams.minMatch);
+}
 
 /* inlining is important to hardwire a hot branch (template emulation) */
 FORCE_INLINE_TEMPLATE
@@ -661,90 +759,8 @@ size_t ZSTD_HcFindBestMatch_generic (
     }
 
     if (dictMode == ZSTD_dedicatedDictSearch) {
-        const U32 ddsLowestIndex  = dms->window.dictLimit;
-        const BYTE* const ddsBase = dms->window.base;
-        const BYTE* const ddsEnd  = dms->window.nextSrc;
-        const U32 ddsSize         = (U32)(ddsEnd - ddsBase);
-        const U32 ddsIndexDelta   = dictLimit - ddsSize;
-        const U32 bucketSize      = (1 << ZSTD_LAZY_DDSS_BUCKET_LOG);
-        const U32 bucketLimit     = nbAttempts < bucketSize - 1 ? nbAttempts : bucketSize - 1;
-        U32 ddsAttempt;
-
-        for (ddsAttempt = 0; ddsAttempt < bucketSize - 1; ddsAttempt++) {
-            PREFETCH_L1(ddsBase + dms->hashTable[ddsIdx + ddsAttempt]);
-        }
-
-        {
-            U32 const chainPackedPointer = dms->hashTable[ddsIdx + bucketSize - 1];
-            U32 const chainIndex = chainPackedPointer >> 8;
-
-            PREFETCH_L1(&dms->chainTable[chainIndex]);
-        }
-
-        for (ddsAttempt = 0; ddsAttempt < bucketLimit; ddsAttempt++) {
-            size_t currentMl=0;
-            const BYTE* match;
-            matchIndex = dms->hashTable[ddsIdx + ddsAttempt];
-            match = ddsBase + matchIndex;
-
-            if (!matchIndex) {
-                return ml;
-            }
-
-            /* guaranteed by table construction */
-            (void)ddsLowestIndex;
-            assert(matchIndex >= ddsLowestIndex);
-            assert(match+4 <= ddsEnd);
-            if (MEM_read32(match) == MEM_read32(ip)) {
-                /* assumption : matchIndex <= dictLimit-4 (by table construction) */
-                currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, ddsEnd, prefixStart) + 4;
-            }
-
-            /* save best solution */
-            if (currentMl > ml) {
-                ml = currentMl;
-                *offsetPtr = curr - (matchIndex + ddsIndexDelta) + ZSTD_REP_MOVE;
-                if (ip+currentMl == iLimit) {
-                    /* best possible, avoids read overflow on next attempt */
-                    return ml;
-                }
-            }
-        }
-
-        {
-            U32 const chainPackedPointer = dms->hashTable[ddsIdx + bucketSize - 1];
-            U32 chainIndex = chainPackedPointer >> 8;
-            U32 const chainLength = chainPackedPointer & 0xFF;
-            U32 const chainAttempts = nbAttempts - ddsAttempt;
-            U32 const chainLimit = chainAttempts > chainLength ? chainLength : chainAttempts;
-            U32 chainAttempt;
-
-            for (chainAttempt = 0 ; chainAttempt < chainLimit; chainAttempt++) {
-                PREFETCH_L1(ddsBase + dms->chainTable[chainIndex + chainAttempt]);
-            }
-
-            for (chainAttempt = 0 ; chainAttempt < chainLimit; chainAttempt++, chainIndex++) {
-                size_t currentMl=0;
-                const BYTE* match;
-                matchIndex = dms->chainTable[chainIndex];
-                match = ddsBase + matchIndex;
-
-                /* guaranteed by table construction */
-                assert(matchIndex >= ddsLowestIndex);
-                assert(match+4 <= ddsEnd);
-                if (MEM_read32(match) == MEM_read32(ip)) {
-                    /* assumption : matchIndex <= dictLimit-4 (by table construction) */
-                    currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, ddsEnd, prefixStart) + 4;
-                }
-
-                /* save best solution */
-                if (currentMl > ml) {
-                    ml = currentMl;
-                    *offsetPtr = curr - (matchIndex + ddsIndexDelta) + ZSTD_REP_MOVE;
-                    if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
-                }
-            }
-        }
+        ml = ZSTD_dedicatedDictSearch_lazy_search(offsetPtr, ml, nbAttempts, dms,
+                                                  ip, iLimit, prefixStart, curr, dictLimit, ddsIdx);
     } else if (dictMode == ZSTD_dictMatchState) {
         const U32* const dmsChainTable = dms->chainTable;
         const U32 dmsChainSize         = (1 << dms->cParams.chainLog);
@@ -845,11 +861,602 @@ FORCE_INLINE_TEMPLATE size_t ZSTD_HcFindBestMatch_extDict_selectMLS (
     }
 }
 
+/* *********************************
+* (SIMD) Row-based matchfinder
+***********************************/
+/* Constants for row-based hash */
+#define ZSTD_ROW_HASH_TAG_OFFSET 16     /* byte offset of hashes in the match state's tagTable from the beginning of a row */
+#define ZSTD_ROW_HASH_TAG_BITS 8        /* nb bits to use for the tag */
+#define ZSTD_ROW_HASH_TAG_MASK ((1u << ZSTD_ROW_HASH_TAG_BITS) - 1)
+#define ZSTD_ROW_HASH_MAX_ENTRIES 64    /* absolute maximum number of entries per row, for all configurations */               
+
+#define ZSTD_ROW_HASH_CACHE_MASK (ZSTD_ROW_HASH_CACHE_SIZE - 1)
+
+typedef U64 ZSTD_VecMask;   /* Clarifies when we are interacting with a U64 representing a mask of matches */
+
+/* ZSTD_VecMask_next():
+ * Starting from the LSB, returns the idx of the next non-zero bit.
+ * Basically counting the nb of trailing zeroes.
+ */
+static U32 ZSTD_VecMask_next(ZSTD_VecMask val) {
+    assert(val != 0);
+#   if defined(_MSC_VER) && defined(_WIN64)
+    unsigned long r=0;
+    return _BitScanForward64(&r, val) ? (U32)r : 0; /* _BitScanForward64 not defined outside of x86/64 */
+#   elif (defined(__GNUC__) && ((__GNUC__ > 3) || ((__GNUC__ == 3) && (__GNUC_MINOR__ >= 4))))
+    if (sizeof(size_t) == 4) {
+        U32 mostSignificantWord = (U32)(val >> 32);
+        U32 leastSignificantWord = (U32)val;
+        if (leastSignificantWord == 0) {
+            return 32 + (U32)__builtin_ctz(mostSignificantWord);
+        } else {
+            return (U32)__builtin_ctz(leastSignificantWord);
+        }
+    } else {
+        return (U32)__builtin_ctzll(val);
+    }
+#   else
+    /* Software ctz version: http://aggregate.org/MAGIC/#Trailing%20Zero%20Count
+     * and: https://stackoverflow.com/questions/2709430/count-number-of-bits-in-a-64-bit-long-big-integer
+     */
+    val = ~val & (val - 1ULL); /* Lowest set bit mask */
+    val = val - ((val >> 1) & 0x5555555555555555);
+    val = (val & 0x3333333333333333ULL) + ((val >> 2) & 0x3333333333333333ULL);
+    return (U32)((((val + (val >> 4)) & 0xF0F0F0F0F0F0F0FULL) * 0x101010101010101ULL) >> 56);
+#   endif
+}
+
+/* ZSTD_rotateRight_*():
+ * Rotates a bitfield to the right by "count" bits.
+ * https://en.wikipedia.org/w/index.php?title=Circular_shift&oldid=991635599#Implementing_circular_shifts
+ */
+FORCE_INLINE_TEMPLATE
+U64 ZSTD_rotateRight_U64(U64 const value, U32 count) {
+    assert(count < 64);
+    count &= 0x3F; /* for fickle pattern recognition */
+    return (value >> count) | (U64)(value << ((0U - count) & 0x3F));
+}
+
+FORCE_INLINE_TEMPLATE
+U32 ZSTD_rotateRight_U32(U32 const value, U32 count) {
+    assert(count < 32);
+    count &= 0x1F; /* for fickle pattern recognition */
+    return (value >> count) | (U32)(value << ((0U - count) & 0x1F));
+}
+
+FORCE_INLINE_TEMPLATE
+U16 ZSTD_rotateRight_U16(U16 const value, U32 count) {
+    assert(count < 16);
+    count &= 0x0F; /* for fickle pattern recognition */
+    return (value >> count) | (U16)(value << ((0U - count) & 0x0F));
+}
+
+/* ZSTD_row_nextIndex():
+ * Returns the next index to insert at within a tagTable row, and updates the "head"
+ * value to reflect the update. Essentially cycles backwards from [0, {entries per row})
+ */
+FORCE_INLINE_TEMPLATE U32 ZSTD_row_nextIndex(BYTE* const tagRow, U32 const rowMask) {
+  U32 const next = (*tagRow - 1) & rowMask;
+  *tagRow = (BYTE)next;
+  return next;
+}
+
+/* ZSTD_isAligned():
+ * Checks that a pointer is aligned to "align" bytes which must be a power of 2.
+ */
+MEM_STATIC int ZSTD_isAligned(void const* ptr, size_t align) {
+    assert((align & (align - 1)) == 0);
+    return (((size_t)ptr) & (align - 1)) == 0;
+}
+
+/* ZSTD_row_prefetch():
+ * Performs prefetching for the hashTable and tagTable at a given row.
+ */
+FORCE_INLINE_TEMPLATE void ZSTD_row_prefetch(U32 const* hashTable, U16 const* tagTable, U32 const relRow, U32 const rowLog) {
+    PREFETCH_L1(hashTable + relRow);
+    if (rowLog >= 5) {
+        PREFETCH_L1(hashTable + relRow + 16);
+        /* Note: prefetching more of the hash table does not appear to be beneficial for 128-entry rows */
+    }
+    PREFETCH_L1(tagTable + relRow);
+    if (rowLog == 6) {
+        PREFETCH_L1(tagTable + relRow + 32);
+    }
+    assert(rowLog == 4 || rowLog == 5 || rowLog == 6);
+    assert(ZSTD_isAligned(hashTable + relRow, 64));                 /* prefetched hash row always 64-byte aligned */
+    assert(ZSTD_isAligned(tagTable + relRow, (size_t)1 << rowLog)); /* prefetched tagRow sits on correct multiple of bytes (32,64,128) */
+}
+
+/* ZSTD_row_fillHashCache():
+ * Fill up the hash cache starting at idx, prefetching up to ZSTD_ROW_HASH_CACHE_SIZE entries,
+ * but not beyond iLimit.
+ */
+static void ZSTD_row_fillHashCache(ZSTD_matchState_t* ms, const BYTE* base,
+                                   U32 const rowLog, U32 const mls,
+                                   U32 idx, const BYTE* const iLimit)
+{
+    U32 const* const hashTable = ms->hashTable;
+    U16 const* const tagTable = ms->tagTable;
+    U32 const hashLog = ms->rowHashLog;
+    U32 const maxElemsToPrefetch = (base + idx) > iLimit ? 0 : (U32)(iLimit - (base + idx) + 1);
+    U32 const lim = idx + MIN(ZSTD_ROW_HASH_CACHE_SIZE, maxElemsToPrefetch);
+
+    for (; idx < lim; ++idx) {
+        U32 const hash = (U32)ZSTD_hashPtr(base + idx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls);
+        U32 const row = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
+        ZSTD_row_prefetch(hashTable, tagTable, row, rowLog);
+        ms->hashCache[idx & ZSTD_ROW_HASH_CACHE_MASK] = hash;
+    }
+
+    DEBUGLOG(6, "ZSTD_row_fillHashCache(): [%u %u %u %u %u %u %u %u]", ms->hashCache[0], ms->hashCache[1],
+                                                     ms->hashCache[2], ms->hashCache[3], ms->hashCache[4],
+                                                     ms->hashCache[5], ms->hashCache[6], ms->hashCache[7]);
+}
+
+/* ZSTD_row_nextCachedHash():
+ * Returns the hash of base + idx, and replaces the hash in the hash cache with the byte at
+ * base + idx + ZSTD_ROW_HASH_CACHE_SIZE. Also prefetches the appropriate rows from hashTable and tagTable.
+ */
+FORCE_INLINE_TEMPLATE U32 ZSTD_row_nextCachedHash(U32* cache, U32 const* hashTable,
+                                                  U16 const* tagTable, BYTE const* base,
+                                                  U32 idx, U32 const hashLog,
+                                                  U32 const rowLog, U32 const mls)
+{
+    U32 const newHash = (U32)ZSTD_hashPtr(base+idx+ZSTD_ROW_HASH_CACHE_SIZE, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls);
+    U32 const row = (newHash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
+    ZSTD_row_prefetch(hashTable, tagTable, row, rowLog);
+    {   U32 const hash = cache[idx & ZSTD_ROW_HASH_CACHE_MASK];
+        cache[idx & ZSTD_ROW_HASH_CACHE_MASK] = newHash;
+        return hash;
+    }
+}
+
+/* ZSTD_row_update_internal():
+ * Inserts the byte at ip into the appropriate position in the hash table.
+ * Determines the relative row, and the position within the {16, 32} entry row to insert at.
+ */
+FORCE_INLINE_TEMPLATE void ZSTD_row_update_internal(ZSTD_matchState_t* ms, const BYTE* ip,
+                                                    U32 const mls, U32 const rowLog,
+                                                    U32 const rowMask, U32 const useCache)
+{
+    U32* const hashTable = ms->hashTable;
+    U16* const tagTable = ms->tagTable;
+    U32 const hashLog = ms->rowHashLog;
+    const BYTE* const base = ms->window.base;
+    const U32 target = (U32)(ip - base);
+    U32 idx = ms->nextToUpdate;
+
+    DEBUGLOG(6, "ZSTD_row_update_internal(): nextToUpdate=%u, current=%u", idx, target);
+    for (; idx < target; ++idx) {
+        U32 const hash = useCache ? ZSTD_row_nextCachedHash(ms->hashCache, hashTable, tagTable, base, idx, hashLog, rowLog, mls)
+                                  : (U32)ZSTD_hashPtr(base + idx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls);
+        U32 const relRow = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
+        U32* const row = hashTable + relRow;
+        BYTE* tagRow = (BYTE*)(tagTable + relRow);  /* Though tagTable is laid out as a table of U16, each tag is only 1 byte.
+                                                       Explicit cast allows us to get exact desired position within each row */
+        U32 const pos = ZSTD_row_nextIndex(tagRow, rowMask);
+
+        assert(hash == ZSTD_hashPtr(base + idx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls));
+        ((BYTE*)tagRow)[pos + ZSTD_ROW_HASH_TAG_OFFSET] = hash & ZSTD_ROW_HASH_TAG_MASK;
+        row[pos] = idx;
+    }
+    ms->nextToUpdate = target;
+}
+
+/* ZSTD_row_update():
+ * External wrapper for ZSTD_row_update_internal(). Used for filling the hashtable during dictionary
+ * processing.
+ */
+void ZSTD_row_update(ZSTD_matchState_t* const ms, const BYTE* ip) {
+    const U32 rowLog = MAX(MIN(ms->cParams.searchLog, 6), 4);
+    const U32 rowMask = (1u << rowLog) - 1;
+    const U32 mls = MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */);
+
+    DEBUGLOG(5, "ZSTD_row_update(), rowLog=%u", rowLog);
+    ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 0 /* dont use cache */);
+}
+
+/* Returns a ZSTD_VecMask (U32) that has the nth bit set to 1 if the newly-computed "tag" matches
+ * the hash at the nth position in a row of the tagTable.
+ * Each row is a circular buffer beginning at the value of "head". So we must rotate the "matches" bitfield
+ * to match up with the actual layout of the entries within the hashTable */
+FORCE_INLINE_TEMPLATE
+ZSTD_VecMask ZSTD_row_getMatchMask(const BYTE* const tagRow, const BYTE tag, const U32 head, const U32 rowEntries) {
+    const BYTE* const src = tagRow + ZSTD_ROW_HASH_TAG_OFFSET;
+    assert((rowEntries == 16) || (rowEntries == 32) || rowEntries == 64);
+    assert(rowEntries <= ZSTD_ROW_HASH_MAX_ENTRIES);
+#if defined(ZSTD_ARCH_X86_SSE2)
+    if (rowEntries == 16) {
+        const __m128i chunk = _mm_loadu_si128((const __m128i*)(const void*)src);
+        const __m128i equalMask = _mm_cmpeq_epi8(chunk, _mm_set1_epi8(tag));
+        const U16 matches = (U16)_mm_movemask_epi8(equalMask);
+        return ZSTD_rotateRight_U16(matches, head);
+    } else if (rowEntries == 32) {
+        const __m128i chunk0 = _mm_loadu_si128((const __m128i*)(const void*)&src[0]);
+        const __m128i chunk1 = _mm_loadu_si128((const __m128i*)(const void*)&src[16]);
+        const __m128i equalMask0 = _mm_cmpeq_epi8(chunk0, _mm_set1_epi8(tag));
+        const __m128i equalMask1 = _mm_cmpeq_epi8(chunk1, _mm_set1_epi8(tag));
+        const U32 lo = (U32)_mm_movemask_epi8(equalMask0);
+        const U32 hi = (U32)_mm_movemask_epi8(equalMask1);
+        return ZSTD_rotateRight_U32((hi << 16) | lo, head);
+    } else {  /* rowEntries == 64 */
+        const __m128i chunk0 = _mm_loadu_si128((const __m128i*)(const void*)&src[0]);
+        const __m128i chunk1 = _mm_loadu_si128((const __m128i*)(const void*)&src[16]);
+        const __m128i chunk2 = _mm_loadu_si128((const __m128i*)(const void*)&src[32]);
+        const __m128i chunk3 = _mm_loadu_si128((const __m128i*)(const void*)&src[48]);
+        const __m128i comparisonMask = _mm_set1_epi8(tag);
+        const __m128i equalMask0 = _mm_cmpeq_epi8(chunk0, comparisonMask);
+        const __m128i equalMask1 = _mm_cmpeq_epi8(chunk1, comparisonMask);
+        const __m128i equalMask2 = _mm_cmpeq_epi8(chunk2, comparisonMask);
+        const __m128i equalMask3 = _mm_cmpeq_epi8(chunk3, comparisonMask);
+        const U64 mask0 = (U64)_mm_movemask_epi8(equalMask0);
+        const U64 mask1 = (U64)_mm_movemask_epi8(equalMask1);
+        const U64 mask2 = (U64)_mm_movemask_epi8(equalMask2);
+        const U64 mask3 = (U64)_mm_movemask_epi8(equalMask3);
+        return ZSTD_rotateRight_U64((mask3 << 48) | (mask2 << 32) | (mask1 << 16) | mask0, head);
+    }
+#else
+#  if defined(ZSTD_ARCH_ARM_NEON)
+    if (MEM_isLittleEndian()) {
+        if (rowEntries == 16) {
+            const uint8x16_t chunk = vld1q_u8(src);
+            const uint16x8_t equalMask = vreinterpretq_u16_u8(vceqq_u8(chunk, vdupq_n_u8(tag)));
+            const uint16x8_t t0 = vshlq_n_u16(equalMask, 7);
+            const uint32x4_t t1 = vreinterpretq_u32_u16(vsriq_n_u16(t0, t0, 14));
+            const uint64x2_t t2 = vreinterpretq_u64_u32(vshrq_n_u32(t1, 14));
+            const uint8x16_t t3 = vreinterpretq_u8_u64(vsraq_n_u64(t2, t2, 28));
+            const U16 hi = (U16)vgetq_lane_u8(t3, 8);
+            const U16 lo = (U16)vgetq_lane_u8(t3, 0);
+            return ZSTD_rotateRight_U16((hi << 8) | lo, head);
+        } else if (rowEntries == 32) {
+            const uint16x8x2_t chunk = vld2q_u16((const U16*)(const void*)src);
+            const uint8x16_t chunk0 = vreinterpretq_u8_u16(chunk.val[0]);
+            const uint8x16_t chunk1 = vreinterpretq_u8_u16(chunk.val[1]);
+            const uint8x16_t equalMask0 = vceqq_u8(chunk0, vdupq_n_u8(tag));
+            const uint8x16_t equalMask1 = vceqq_u8(chunk1, vdupq_n_u8(tag));
+            const int8x8_t pack0 = vqmovn_s16(vreinterpretq_s16_u8(equalMask0));
+            const int8x8_t pack1 = vqmovn_s16(vreinterpretq_s16_u8(equalMask1));
+            const uint8x8_t t0 = vreinterpret_u8_s8(pack0);
+            const uint8x8_t t1 = vreinterpret_u8_s8(pack1);
+            const uint8x8_t t2 = vsri_n_u8(t1, t0, 2);
+            const uint8x8x2_t t3 = vuzp_u8(t2, t0);
+            const uint8x8_t t4 = vsri_n_u8(t3.val[1], t3.val[0], 4);
+            const U32 matches = vget_lane_u32(vreinterpret_u32_u8(t4), 0);
+            return ZSTD_rotateRight_U32(matches, head);
+        } else { /* rowEntries == 64 */
+            const uint8x16x4_t chunk = vld4q_u8(src);
+            const uint8x16_t dup = vdupq_n_u8(tag);
+            const uint8x16_t cmp0 = vceqq_u8(chunk.val[0], dup);
+            const uint8x16_t cmp1 = vceqq_u8(chunk.val[1], dup);
+            const uint8x16_t cmp2 = vceqq_u8(chunk.val[2], dup);
+            const uint8x16_t cmp3 = vceqq_u8(chunk.val[3], dup);
+
+            const uint8x16_t t0 = vsriq_n_u8(cmp1, cmp0, 1);
+            const uint8x16_t t1 = vsriq_n_u8(cmp3, cmp2, 1);
+            const uint8x16_t t2 = vsriq_n_u8(t1, t0, 2);
+            const uint8x16_t t3 = vsriq_n_u8(t2, t2, 4);
+            const uint8x8_t t4 = vshrn_n_u16(vreinterpretq_u16_u8(t3), 4);
+            const U64 matches = vget_lane_u64(vreinterpret_u64_u8(t4), 0);
+            return ZSTD_rotateRight_U64(matches, head);
+        }
+    }
+#  endif
+    { /* SWAR */
+        const size_t chunkSize = sizeof(size_t);
+        const size_t shiftAmount = ((chunkSize * 8) - chunkSize);
+        const size_t xFF = ~((size_t)0);
+        const size_t x01 = xFF / 0xFF;
+        const size_t x80 = x01 << 7;
+        const size_t splatChar = tag * x01;
+        ZSTD_VecMask matches = 0;
+        int i = rowEntries - chunkSize;
+        assert((sizeof(size_t) == 4) || (sizeof(size_t) == 8));
+        if (MEM_isLittleEndian()) { /* runtime check so have two loops */
+            const size_t extractMagic = (xFF / 0x7F) >> chunkSize;
+            do {
+                size_t chunk = MEM_readST(&src[i]);
+                chunk ^= splatChar;
+                chunk = (((chunk | x80) - x01) | chunk) & x80;
+                matches <<= chunkSize;
+                matches |= (chunk * extractMagic) >> shiftAmount;
+                i -= chunkSize;
+            } while (i >= 0);
+        } else { /* big endian: reverse bits during extraction */
+            const size_t msb = xFF ^ (xFF >> 1);
+            const size_t extractMagic = (msb / 0x1FF) | msb;
+            do {
+                size_t chunk = MEM_readST(&src[i]);
+                chunk ^= splatChar;
+                chunk = (((chunk | x80) - x01) | chunk) & x80;
+                matches <<= chunkSize;
+                matches |= ((chunk >> 7) * extractMagic) >> shiftAmount;
+                i -= chunkSize;
+            } while (i >= 0);
+        }
+        matches = ~matches;
+        if (rowEntries == 16) {
+            return ZSTD_rotateRight_U16((U16)matches, head);
+        } else if (rowEntries == 32) {
+            return ZSTD_rotateRight_U32((U32)matches, head);
+        } else {
+            return ZSTD_rotateRight_U64((U64)matches, head);
+        }
+    }
+#endif
+}
+
+/* The high-level approach of the SIMD row based match finder is as follows:
+ * - Figure out where to insert the new entry:
+ *      - Generate a hash from a byte along with an additional 1-byte "short hash". The additional byte is our "tag"
+ *      - The hashTable is effectively split into groups or "rows" of 16 or 32 entries of U32, and the hash determines
+ *        which row to insert into.
+ *      - Determine the correct position within the row to insert the entry into. Each row of 16 or 32 can
+ *        be considered as a circular buffer with a "head" index that resides in the tagTable.
+ *      - Also insert the "tag" into the equivalent row and position in the tagTable.
+ *          - Note: The tagTable has 17 or 33 1-byte entries per row, due to 16 or 32 tags, and 1 "head" entry.
+ *                  The 17 or 33 entry rows are spaced out to occur every 32 or 64 bytes, respectively,
+ *                  for alignment/performance reasons, leaving some bytes unused.
+ * - Use SIMD to efficiently compare the tags in the tagTable to the 1-byte "short hash" and
+ *   generate a bitfield that we can cycle through to check the collisions in the hash table.
+ * - Pick the longest match.
+ */
+FORCE_INLINE_TEMPLATE
+size_t ZSTD_RowFindBestMatch_generic (
+                        ZSTD_matchState_t* ms,
+                        const BYTE* const ip, const BYTE* const iLimit,
+                        size_t* offsetPtr,
+                        const U32 mls, const ZSTD_dictMode_e dictMode,
+                        const U32 rowLog)
+{
+    U32* const hashTable = ms->hashTable;
+    U16* const tagTable = ms->tagTable;
+    U32* const hashCache = ms->hashCache;
+    const U32 hashLog = ms->rowHashLog;
+    const ZSTD_compressionParameters* const cParams = &ms->cParams;
+    const BYTE* const base = ms->window.base;
+    const BYTE* const dictBase = ms->window.dictBase;
+    const U32 dictLimit = ms->window.dictLimit;
+    const BYTE* const prefixStart = base + dictLimit;
+    const BYTE* const dictEnd = dictBase + dictLimit;
+    const U32 curr = (U32)(ip-base);
+    const U32 maxDistance = 1U << cParams->windowLog;
+    const U32 lowestValid = ms->window.lowLimit;
+    const U32 withinMaxDistance = (curr - lowestValid > maxDistance) ? curr - maxDistance : lowestValid;
+    const U32 isDictionary = (ms->loadedDictEnd != 0);
+    const U32 lowLimit = isDictionary ? lowestValid : withinMaxDistance;
+    const U32 rowEntries = (1U << rowLog);
+    const U32 rowMask = rowEntries - 1;
+    const U32 cappedSearchLog = MIN(cParams->searchLog, rowLog); /* nb of searches is capped at nb entries per row */
+    U32 nbAttempts = 1U << cappedSearchLog;
+    size_t ml=4-1;
+
+    /* DMS/DDS variables that may be referenced laster */
+    const ZSTD_matchState_t* const dms = ms->dictMatchState;
+
+    /* Initialize the following variables to satisfy static analyzer */
+    size_t ddsIdx = 0;
+    U32 ddsExtraAttempts = 0; /* cctx hash tables are limited in searches, but allow extra searches into DDS */
+    U32 dmsTag = 0;
+    U32* dmsRow = NULL;
+    BYTE* dmsTagRow = NULL;
+
+    if (dictMode == ZSTD_dedicatedDictSearch) {
+        const U32 ddsHashLog = dms->cParams.hashLog - ZSTD_LAZY_DDSS_BUCKET_LOG;
+        {   /* Prefetch DDS hashtable entry */
+            ddsIdx = ZSTD_hashPtr(ip, ddsHashLog, mls) << ZSTD_LAZY_DDSS_BUCKET_LOG;
+            PREFETCH_L1(&dms->hashTable[ddsIdx]);
+        }
+        ddsExtraAttempts = cParams->searchLog > rowLog ? 1U << (cParams->searchLog - rowLog) : 0;
+    }
+
+    if (dictMode == ZSTD_dictMatchState) {
+        /* Prefetch DMS rows */
+        U32* const dmsHashTable = dms->hashTable;
+        U16* const dmsTagTable = dms->tagTable;
+        U32 const dmsHash = (U32)ZSTD_hashPtr(ip, dms->rowHashLog + ZSTD_ROW_HASH_TAG_BITS, mls);
+        U32 const dmsRelRow = (dmsHash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
+        dmsTag = dmsHash & ZSTD_ROW_HASH_TAG_MASK;
+        dmsTagRow = (BYTE*)(dmsTagTable + dmsRelRow);
+        dmsRow = dmsHashTable + dmsRelRow;
+        ZSTD_row_prefetch(dmsHashTable, dmsTagTable, dmsRelRow, rowLog);
+    }
+
+    /* Update the hashTable and tagTable up to (but not including) ip */
+    ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 1 /* useCache */);
+    {   /* Get the hash for ip, compute the appropriate row */
+        U32 const hash = ZSTD_row_nextCachedHash(hashCache, hashTable, tagTable, base, curr, hashLog, rowLog, mls);
+        U32 const relRow = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
+        U32 const tag = hash & ZSTD_ROW_HASH_TAG_MASK;
+        U32* const row = hashTable + relRow;
+        BYTE* tagRow = (BYTE*)(tagTable + relRow);
+        U32 const head = *tagRow & rowMask;
+        U32 matchBuffer[ZSTD_ROW_HASH_MAX_ENTRIES];
+        size_t numMatches = 0;
+        size_t currMatch = 0;
+        ZSTD_VecMask matches = ZSTD_row_getMatchMask(tagRow, (BYTE)tag, head, rowEntries);
+
+        /* Cycle through the matches and prefetch */
+        for (; (matches > 0) && (nbAttempts > 0); --nbAttempts, matches &= (matches - 1)) {
+            U32 const matchPos = (head + ZSTD_VecMask_next(matches)) & rowMask;
+            U32 const matchIndex = row[matchPos];
+            assert(numMatches < rowEntries);
+            if (matchIndex < lowLimit)
+                break;
+            if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) {
+                PREFETCH_L1(base + matchIndex);
+            } else {
+                PREFETCH_L1(dictBase + matchIndex);
+            }
+            matchBuffer[numMatches++] = matchIndex;
+        }
+
+        /* Speed opt: insert current byte into hashtable too. This allows us to avoid one iteration of the loop
+           in ZSTD_row_update_internal() at the next search. */
+        {
+            U32 const pos = ZSTD_row_nextIndex(tagRow, rowMask);
+            tagRow[pos + ZSTD_ROW_HASH_TAG_OFFSET] = (BYTE)tag;
+            row[pos] = ms->nextToUpdate++;
+        }
+
+        /* Return the longest match */
+        for (; currMatch < numMatches; ++currMatch) {
+            U32 const matchIndex = matchBuffer[currMatch];
+            size_t currentMl=0;
+            assert(matchIndex < curr);
+            assert(matchIndex >= lowLimit);
+
+            if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) {
+                const BYTE* const match = base + matchIndex;
+                assert(matchIndex >= dictLimit);   /* ensures this is true if dictMode != ZSTD_extDict */
+                if (match[ml] == ip[ml])   /* potentially better */
+                    currentMl = ZSTD_count(ip, match, iLimit);
+            } else {
+                const BYTE* const match = dictBase + matchIndex;
+                assert(match+4 <= dictEnd);
+                if (MEM_read32(match) == MEM_read32(ip))   /* assumption : matchIndex <= dictLimit-4 (by table construction) */
+                    currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dictEnd, prefixStart) + 4;
+            }
+
+            /* Save best solution */
+            if (currentMl > ml) {
+                ml = currentMl;
+                *offsetPtr = curr - matchIndex + ZSTD_REP_MOVE;
+                if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
+            }
+        }
+    }
+
+    if (dictMode == ZSTD_dedicatedDictSearch) {
+        ml = ZSTD_dedicatedDictSearch_lazy_search(offsetPtr, ml, nbAttempts + ddsExtraAttempts, dms,
+                                                  ip, iLimit, prefixStart, curr, dictLimit, ddsIdx);
+    } else if (dictMode == ZSTD_dictMatchState) {
+        /* TODO: Measure and potentially add prefetching to DMS */
+        const U32 dmsLowestIndex       = dms->window.dictLimit;
+        const BYTE* const dmsBase      = dms->window.base;
+        const BYTE* const dmsEnd       = dms->window.nextSrc;
+        const U32 dmsSize              = (U32)(dmsEnd - dmsBase);
+        const U32 dmsIndexDelta        = dictLimit - dmsSize;
+
+        {   U32 const head = *dmsTagRow & rowMask;
+            U32 matchBuffer[ZSTD_ROW_HASH_MAX_ENTRIES];
+            size_t numMatches = 0;
+            size_t currMatch = 0;
+            ZSTD_VecMask matches = ZSTD_row_getMatchMask(dmsTagRow, (BYTE)dmsTag, head, rowEntries);
+
+            for (; (matches > 0) && (nbAttempts > 0); --nbAttempts, matches &= (matches - 1)) {
+                U32 const matchPos = (head + ZSTD_VecMask_next(matches)) & rowMask;
+                U32 const matchIndex = dmsRow[matchPos];
+                if (matchIndex < dmsLowestIndex)
+                    break;
+                PREFETCH_L1(dmsBase + matchIndex);
+                matchBuffer[numMatches++] = matchIndex;
+            }
+
+            /* Return the longest match */
+            for (; currMatch < numMatches; ++currMatch) {
+                U32 const matchIndex = matchBuffer[currMatch];
+                size_t currentMl=0;
+                assert(matchIndex >= dmsLowestIndex);
+                assert(matchIndex < curr);
+
+                {   const BYTE* const match = dmsBase + matchIndex;
+                    assert(match+4 <= dmsEnd);
+                    if (MEM_read32(match) == MEM_read32(ip))
+                        currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dmsEnd, prefixStart) + 4;
+                }
+
+                if (currentMl > ml) {
+                    ml = currentMl;
+                    *offsetPtr = curr - (matchIndex + dmsIndexDelta) + ZSTD_REP_MOVE;
+                    if (ip+currentMl == iLimit) break;
+                }
+            }
+        }
+    }
+    return ml;
+}
+
+/* Inlining is important to hardwire a hot branch (template emulation) */
+FORCE_INLINE_TEMPLATE size_t ZSTD_RowFindBestMatch_selectMLS (
+                        ZSTD_matchState_t* ms,
+                        const BYTE* ip, const BYTE* const iLimit,
+                        const ZSTD_dictMode_e dictMode, size_t* offsetPtr, const U32 rowLog)
+{
+    switch(ms->cParams.minMatch)
+    {
+    default : /* includes case 3 */
+    case 4 : return ZSTD_RowFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 4, dictMode, rowLog);
+    case 5 : return ZSTD_RowFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 5, dictMode, rowLog);
+    case 7 :
+    case 6 : return ZSTD_RowFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 6, dictMode, rowLog);
+    }
+}
+
+FORCE_INLINE_TEMPLATE size_t ZSTD_RowFindBestMatch_selectRowLog (
+                        ZSTD_matchState_t* ms,
+                        const BYTE* ip, const BYTE* const iLimit,
+                        size_t* offsetPtr)
+{
+    const U32 cappedSearchLog = MIN(ms->cParams.searchLog, 6);
+    switch(cappedSearchLog)
+    {
+    default :
+    case 4 : return ZSTD_RowFindBestMatch_selectMLS(ms, ip, iLimit, ZSTD_noDict, offsetPtr, 4);
+    case 5 : return ZSTD_RowFindBestMatch_selectMLS(ms, ip, iLimit, ZSTD_noDict, offsetPtr, 5);
+    case 6 : return ZSTD_RowFindBestMatch_selectMLS(ms, ip, iLimit, ZSTD_noDict, offsetPtr, 6);
+    }
+}
+
+FORCE_INLINE_TEMPLATE size_t ZSTD_RowFindBestMatch_dictMatchState_selectRowLog(
+                        ZSTD_matchState_t* ms,
+                        const BYTE* ip, const BYTE* const iLimit,
+                        size_t* offsetPtr)
+{
+    const U32 cappedSearchLog = MIN(ms->cParams.searchLog, 6);
+    switch(cappedSearchLog)
+    {
+    default :
+    case 4 : return ZSTD_RowFindBestMatch_selectMLS(ms, ip, iLimit, ZSTD_dictMatchState, offsetPtr, 4);
+    case 5 : return ZSTD_RowFindBestMatch_selectMLS(ms, ip, iLimit, ZSTD_dictMatchState, offsetPtr, 5);
+    case 6 : return ZSTD_RowFindBestMatch_selectMLS(ms, ip, iLimit, ZSTD_dictMatchState, offsetPtr, 6);
+    }
+}
+
+FORCE_INLINE_TEMPLATE size_t ZSTD_RowFindBestMatch_dedicatedDictSearch_selectRowLog(
+                        ZSTD_matchState_t* ms,
+                        const BYTE* ip, const BYTE* const iLimit,
+                        size_t* offsetPtr)
+{
+    const U32 cappedSearchLog = MIN(ms->cParams.searchLog, 6);
+    switch(cappedSearchLog)
+    {
+    default :
+    case 4 : return ZSTD_RowFindBestMatch_selectMLS(ms, ip, iLimit, ZSTD_dedicatedDictSearch, offsetPtr, 4);
+    case 5 : return ZSTD_RowFindBestMatch_selectMLS(ms, ip, iLimit, ZSTD_dedicatedDictSearch, offsetPtr, 5);
+    case 6 : return ZSTD_RowFindBestMatch_selectMLS(ms, ip, iLimit, ZSTD_dedicatedDictSearch, offsetPtr, 6);
+    }
+}
+
+FORCE_INLINE_TEMPLATE size_t ZSTD_RowFindBestMatch_extDict_selectRowLog (
+                        ZSTD_matchState_t* ms,
+                        const BYTE* ip, const BYTE* const iLimit,
+                        size_t* offsetPtr)
+{
+    const U32 cappedSearchLog = MIN(ms->cParams.searchLog, 6);
+    switch(cappedSearchLog)
+    {
+    default :
+    case 4 : return ZSTD_RowFindBestMatch_selectMLS(ms, ip, iLimit, ZSTD_extDict, offsetPtr, 4);
+    case 5 : return ZSTD_RowFindBestMatch_selectMLS(ms, ip, iLimit, ZSTD_extDict, offsetPtr, 5);
+    case 6 : return ZSTD_RowFindBestMatch_selectMLS(ms, ip, iLimit, ZSTD_extDict, offsetPtr, 6);
+    }
+}
+
 
 /* *******************************
 *  Common parser - lazy strategy
 *********************************/
-typedef enum { search_hashChain, search_binaryTree } searchMethod_e;
+typedef enum { search_hashChain=0, search_binaryTree=1, search_rowHash=2 } searchMethod_e;
 
 FORCE_INLINE_TEMPLATE size_t
 ZSTD_compressBlock_lazy_generic(
@@ -863,10 +1470,11 @@ ZSTD_compressBlock_lazy_generic(
     const BYTE* ip = istart;
     const BYTE* anchor = istart;
     const BYTE* const iend = istart + srcSize;
-    const BYTE* const ilimit = iend - 8;
+    const BYTE* const ilimit = searchMethod == search_rowHash ? iend - 8 - ZSTD_ROW_HASH_CACHE_SIZE : iend - 8;
     const BYTE* const base = ms->window.base;
     const U32 prefixLowestIndex = ms->window.dictLimit;
     const BYTE* const prefixLowest = base + prefixLowestIndex;
+    const U32 rowLog = ms->cParams.searchLog < 5 ? 4 : 5;
 
     typedef size_t (*searchMax_f)(
                         ZSTD_matchState_t* ms,
@@ -878,26 +1486,30 @@ ZSTD_compressBlock_lazy_generic(
      * that should never occur (extDict modes go to the other implementation
      * below and there is no DDSS for binary tree search yet).
      */
-    const searchMax_f searchFuncs[4][2] = {
+    const searchMax_f searchFuncs[4][3] = {
         {
             ZSTD_HcFindBestMatch_selectMLS,
-            ZSTD_BtFindBestMatch_selectMLS
+            ZSTD_BtFindBestMatch_selectMLS,
+            ZSTD_RowFindBestMatch_selectRowLog
         },
         {
+            NULL,
             NULL,
             NULL
         },
         {
             ZSTD_HcFindBestMatch_dictMatchState_selectMLS,
-            ZSTD_BtFindBestMatch_dictMatchState_selectMLS
+            ZSTD_BtFindBestMatch_dictMatchState_selectMLS,
+            ZSTD_RowFindBestMatch_dictMatchState_selectRowLog
         },
         {
             ZSTD_HcFindBestMatch_dedicatedDictSearch_selectMLS,
-            NULL
+            NULL,
+            ZSTD_RowFindBestMatch_dedicatedDictSearch_selectRowLog
         }
     };
 
-    searchMax_f const searchMax = searchFuncs[dictMode][searchMethod == search_binaryTree];
+    searchMax_f const searchMax = searchFuncs[dictMode][(int)searchMethod];
     U32 offset_1 = rep[0], offset_2 = rep[1], savedOffset=0;
 
     const int isDMS = dictMode == ZSTD_dictMatchState;
@@ -915,9 +1527,7 @@ ZSTD_compressBlock_lazy_generic(
 
     assert(searchMax != NULL);
 
-    DEBUGLOG(5, "ZSTD_compressBlock_lazy_generic (dictMode=%u)", (U32)dictMode);
-
-    /* init */
+    DEBUGLOG(5, "ZSTD_compressBlock_lazy_generic (dictMode=%u) (searchFunc=%u)", (U32)dictMode, (U32)searchMethod);
     ip += (dictAndPrefixLength == 0);
     if (dictMode == ZSTD_noDict) {
         U32 const curr = (U32)(ip - base);
@@ -933,6 +1543,12 @@ ZSTD_compressBlock_lazy_generic(
         assert(offset_2 <= dictAndPrefixLength);
     }
 
+    if (searchMethod == search_rowHash) {
+        ZSTD_row_fillHashCache(ms, base, rowLog,
+                            MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */),
+                            ms->nextToUpdate, ilimit);
+    }
+
     /* Match Loop */
 #if defined(__GNUC__) && defined(__x86_64__)
     /* I've measured random a 5% speed loss on levels 5 & 6 (greedy) when the
@@ -1198,6 +1814,70 @@ size_t ZSTD_compressBlock_greedy_dedicatedDictSearch(
     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_dedicatedDictSearch);
 }
 
+/* Row-based matchfinder */
+size_t ZSTD_compressBlock_lazy2_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_noDict);
+}
+
+size_t ZSTD_compressBlock_lazy_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_noDict);
+}
+
+size_t ZSTD_compressBlock_greedy_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_noDict);
+}
+
+size_t ZSTD_compressBlock_lazy2_dictMatchState_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_dictMatchState);
+}
+
+size_t ZSTD_compressBlock_lazy_dictMatchState_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_dictMatchState);
+}
+
+size_t ZSTD_compressBlock_greedy_dictMatchState_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_dictMatchState);
+}
+
+
+size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_dedicatedDictSearch);
+}
+
+size_t ZSTD_compressBlock_lazy_dedicatedDictSearch_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_dedicatedDictSearch);
+}
+
+size_t ZSTD_compressBlock_greedy_dedicatedDictSearch_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_dedicatedDictSearch);
+}
 
 FORCE_INLINE_TEMPLATE
 size_t ZSTD_compressBlock_lazy_extDict_generic(
@@ -1210,7 +1890,7 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
     const BYTE* ip = istart;
     const BYTE* anchor = istart;
     const BYTE* const iend = istart + srcSize;
-    const BYTE* const ilimit = iend - 8;
+    const BYTE* const ilimit = searchMethod == search_rowHash ? iend - 8 - ZSTD_ROW_HASH_CACHE_SIZE : iend - 8;
     const BYTE* const base = ms->window.base;
     const U32 dictLimit = ms->window.dictLimit;
     const BYTE* const prefixStart = base + dictLimit;
@@ -1218,18 +1898,28 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
     const BYTE* const dictEnd  = dictBase + dictLimit;
     const BYTE* const dictStart  = dictBase + ms->window.lowLimit;
     const U32 windowLog = ms->cParams.windowLog;
+    const U32 rowLog = ms->cParams.searchLog < 5 ? 4 : 5;
 
     typedef size_t (*searchMax_f)(
                         ZSTD_matchState_t* ms,
                         const BYTE* ip, const BYTE* iLimit, size_t* offsetPtr);
-    searchMax_f searchMax = searchMethod==search_binaryTree ? ZSTD_BtFindBestMatch_extDict_selectMLS : ZSTD_HcFindBestMatch_extDict_selectMLS;
-
+    const searchMax_f searchFuncs[3] = {
+        ZSTD_HcFindBestMatch_extDict_selectMLS,
+        ZSTD_BtFindBestMatch_extDict_selectMLS,
+        ZSTD_RowFindBestMatch_extDict_selectRowLog
+    };
+    searchMax_f searchMax = searchFuncs[(int)searchMethod];
     U32 offset_1 = rep[0], offset_2 = rep[1];
 
-    DEBUGLOG(5, "ZSTD_compressBlock_lazy_extDict_generic");
+    DEBUGLOG(5, "ZSTD_compressBlock_lazy_extDict_generic (searchFunc=%u)", (U32)searchMethod);
 
     /* init */
     ip += (ip == prefixStart);
+    if (searchMethod == search_rowHash) {
+        ZSTD_row_fillHashCache(ms, base, rowLog,
+                               MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */),
+                               ms->nextToUpdate, ilimit);
+    }
 
     /* Match Loop */
 #if defined(__GNUC__) && defined(__x86_64__)
@@ -1249,7 +1939,8 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
             const U32 repIndex = (U32)(curr+1 - offset_1);
             const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
             const BYTE* const repMatch = repBase + repIndex;
-            if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > windowLow))   /* intentional overflow */
+            if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow */
+               & (offset_1 <= curr+1 - windowLow) ) /* note: we are searching at curr+1 */
             if (MEM_read32(ip+1) == MEM_read32(repMatch)) {
                 /* repcode detected we should take it */
                 const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
@@ -1280,7 +1971,8 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
                 const U32 repIndex = (U32)(curr - offset_1);
                 const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
                 const BYTE* const repMatch = repBase + repIndex;
-                if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > windowLow))  /* intentional overflow */
+                if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow : do not test positions overlapping 2 memory segments  */
+                   & (offset_1 <= curr - windowLow) ) /* equivalent to `curr > repIndex >= windowLow` */
                 if (MEM_read32(ip) == MEM_read32(repMatch)) {
                     /* repcode detected */
                     const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
@@ -1311,7 +2003,8 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
                     const U32 repIndex = (U32)(curr - offset_1);
                     const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
                     const BYTE* const repMatch = repBase + repIndex;
-                    if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > windowLow))  /* intentional overflow */
+                    if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow : do not test positions overlapping 2 memory segments  */
+                       & (offset_1 <= curr - windowLow) ) /* equivalent to `curr > repIndex >= windowLow` */
                     if (MEM_read32(ip) == MEM_read32(repMatch)) {
                         /* repcode detected */
                         const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
@@ -1357,7 +2050,8 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
             const U32 repIndex = repCurrent - offset_2;
             const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
             const BYTE* const repMatch = repBase + repIndex;
-            if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > windowLow))  /* intentional overflow */
+            if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow : do not test positions overlapping 2 memory segments  */
+               & (offset_2 <= repCurrent - windowLow) ) /* equivalent to `curr > repIndex >= windowLow` */
             if (MEM_read32(ip) == MEM_read32(repMatch)) {
                 /* repcode detected we should take it */
                 const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
@@ -1410,3 +2104,26 @@ size_t ZSTD_compressBlock_btlazy2_extDict(
 {
     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2);
 }
+
+size_t ZSTD_compressBlock_greedy_extDict_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+{
+    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0);
+}
+
+size_t ZSTD_compressBlock_lazy_extDict_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+
+{
+    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1);
+}
+
+size_t ZSTD_compressBlock_lazy2_extDict_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize)
+
+{
+    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2);
+}
diff --git a/zstd_lazy.h b/zstd_lazy.h
index d0214d5..150f7b3 100644
--- a/zstd_lazy.h
+++ b/zstd_lazy.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -26,6 +26,7 @@ extern "C" {
 #define ZSTD_LAZY_DDSS_BUCKET_LOG 2
 
 U32 ZSTD_insertAndFindFirstIndex(ZSTD_matchState_t* ms, const BYTE* ip);
+void ZSTD_row_update(ZSTD_matchState_t* const ms, const BYTE* ip);
 
 void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t* ms, const BYTE* const ip);
 
@@ -43,6 +44,15 @@ size_t ZSTD_compressBlock_lazy(
 size_t ZSTD_compressBlock_greedy(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy2_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_greedy_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
 
 size_t ZSTD_compressBlock_btlazy2_dictMatchState(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
@@ -56,6 +66,15 @@ size_t ZSTD_compressBlock_lazy_dictMatchState(
 size_t ZSTD_compressBlock_greedy_dictMatchState(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy2_dictMatchState_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy_dictMatchState_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_greedy_dictMatchState_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
 
 size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
@@ -66,6 +85,15 @@ size_t ZSTD_compressBlock_lazy_dedicatedDictSearch(
 size_t ZSTD_compressBlock_greedy_dedicatedDictSearch(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy_dedicatedDictSearch_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_greedy_dedicatedDictSearch_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
 
 size_t ZSTD_compressBlock_greedy_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
@@ -76,9 +104,19 @@ size_t ZSTD_compressBlock_lazy_extDict(
 size_t ZSTD_compressBlock_lazy2_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_greedy_extDict_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy_extDict_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
+size_t ZSTD_compressBlock_lazy2_extDict_row(
+        ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+        void const* src, size_t srcSize);
 size_t ZSTD_compressBlock_btlazy2_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize);
+        
 
 #if defined (__cplusplus)
 }
diff --git a/zstd_ldm.c b/zstd_ldm.c
index 3776dd1..9714438 100644
--- a/zstd_ldm.c
+++ b/zstd_ldm.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -11,13 +11,126 @@
 #include "zstd_ldm.h"
 
 #include "debug.h"
+#include "xxhash.h"
 #include "zstd_fast.h"          /* ZSTD_fillHashTable() */
 #include "zstd_double_fast.h"   /* ZSTD_fillDoubleHashTable() */
+#include "zstd_ldm_geartab.h"
 
 #define LDM_BUCKET_SIZE_LOG 3
 #define LDM_MIN_MATCH_LENGTH 64
 #define LDM_HASH_RLOG 7
-#define LDM_HASH_CHAR_OFFSET 10
+
+typedef struct {
+    U64 rolling;
+    U64 stopMask;
+} ldmRollingHashState_t;
+
+/** ZSTD_ldm_gear_init():
+ *
+ * Initializes the rolling hash state such that it will honor the
+ * settings in params. */
+static void ZSTD_ldm_gear_init(ldmRollingHashState_t* state, ldmParams_t const* params)
+{
+    unsigned maxBitsInMask = MIN(params->minMatchLength, 64);
+    unsigned hashRateLog = params->hashRateLog;
+
+    state->rolling = ~(U32)0;
+
+    /* The choice of the splitting criterion is subject to two conditions:
+     *   1. it has to trigger on average every 2^(hashRateLog) bytes;
+     *   2. ideally, it has to depend on a window of minMatchLength bytes.
+     *
+     * In the gear hash algorithm, bit n depends on the last n bytes;
+     * so in order to obtain a good quality splitting criterion it is
+     * preferable to use bits with high weight.
+     *
+     * To match condition 1 we use a mask with hashRateLog bits set
+     * and, because of the previous remark, we make sure these bits
+     * have the highest possible weight while still respecting
+     * condition 2.
+     */
+    if (hashRateLog > 0 && hashRateLog <= maxBitsInMask) {
+        state->stopMask = (((U64)1 << hashRateLog) - 1) << (maxBitsInMask - hashRateLog);
+    } else {
+        /* In this degenerate case we simply honor the hash rate. */
+        state->stopMask = ((U64)1 << hashRateLog) - 1;
+    }
+}
+
+/** ZSTD_ldm_gear_reset()
+ * Feeds [data, data + minMatchLength) into the hash without registering any
+ * splits. This effectively resets the hash state. This is used when skipping
+ * over data, either at the beginning of a block, or skipping sections.
+ */
+static void ZSTD_ldm_gear_reset(ldmRollingHashState_t* state,
+                                BYTE const* data, size_t minMatchLength)
+{
+    U64 hash = state->rolling;
+    size_t n = 0;
+
+#define GEAR_ITER_ONCE() do {                                  \
+        hash = (hash << 1) + ZSTD_ldm_gearTab[data[n] & 0xff]; \
+        n += 1;                                                \
+    } while (0)
+    while (n + 3 < minMatchLength) {
+        GEAR_ITER_ONCE();
+        GEAR_ITER_ONCE();
+        GEAR_ITER_ONCE();
+        GEAR_ITER_ONCE();
+    }
+    while (n < minMatchLength) {
+        GEAR_ITER_ONCE();
+    }
+#undef GEAR_ITER_ONCE
+}
+
+/** ZSTD_ldm_gear_feed():
+ *
+ * Registers in the splits array all the split points found in the first
+ * size bytes following the data pointer. This function terminates when
+ * either all the data has been processed or LDM_BATCH_SIZE splits are
+ * present in the splits array.
+ *
+ * Precondition: The splits array must not be full.
+ * Returns: The number of bytes processed. */
+static size_t ZSTD_ldm_gear_feed(ldmRollingHashState_t* state,
+                                 BYTE const* data, size_t size,
+                                 size_t* splits, unsigned* numSplits)
+{
+    size_t n;
+    U64 hash, mask;
+
+    hash = state->rolling;
+    mask = state->stopMask;
+    n = 0;
+
+#define GEAR_ITER_ONCE() do { \
+        hash = (hash << 1) + ZSTD_ldm_gearTab[data[n] & 0xff]; \
+        n += 1; \
+        if (UNLIKELY((hash & mask) == 0)) { \
+            splits[*numSplits] = n; \
+            *numSplits += 1; \
+            if (*numSplits == LDM_BATCH_SIZE) \
+                goto done; \
+        } \
+    } while (0)
+
+    while (n + 3 < size) {
+        GEAR_ITER_ONCE();
+        GEAR_ITER_ONCE();
+        GEAR_ITER_ONCE();
+        GEAR_ITER_ONCE();
+    }
+    while (n < size) {
+        GEAR_ITER_ONCE();
+    }
+
+#undef GEAR_ITER_ONCE
+
+done:
+    state->rolling = hash;
+    return n;
+}
 
 void ZSTD_ldm_adjustParameters(ldmParams_t* params,
                                ZSTD_compressionParameters const* cParams)
@@ -54,41 +167,6 @@ size_t ZSTD_ldm_getMaxNbSeq(ldmParams_t params, size_t maxChunkSize)
     return params.enableLdm ? (maxChunkSize / params.minMatchLength) : 0;
 }
 
-/** ZSTD_ldm_getSmallHash() :
- *  numBits should be <= 32
- *  If numBits==0, returns 0.
- *  @return : the most significant numBits of value. */
-static U32 ZSTD_ldm_getSmallHash(U64 value, U32 numBits)
-{
-    assert(numBits <= 32);
-    return numBits == 0 ? 0 : (U32)(value >> (64 - numBits));
-}
-
-/** ZSTD_ldm_getChecksum() :
- *  numBitsToDiscard should be <= 32
- *  @return : the next most significant 32 bits after numBitsToDiscard */
-static U32 ZSTD_ldm_getChecksum(U64 hash, U32 numBitsToDiscard)
-{
-    assert(numBitsToDiscard <= 32);
-    return (hash >> (64 - 32 - numBitsToDiscard)) & 0xFFFFFFFF;
-}
-
-/** ZSTD_ldm_getTag() ;
- *  Given the hash, returns the most significant numTagBits bits
- *  after (32 + hbits) bits.
- *
- *  If there are not enough bits remaining, return the last
- *  numTagBits bits. */
-static U32 ZSTD_ldm_getTag(U64 hash, U32 hbits, U32 numTagBits)
-{
-    assert(numTagBits < 32 && hbits <= 32);
-    if (32 - hbits < numTagBits) {
-        return hash & (((U32)1 << numTagBits) - 1);
-    } else {
-        return (hash >> (32 - hbits - numTagBits)) & (((U32)1 << numTagBits) - 1);
-    }
-}
-
 /** ZSTD_ldm_getBucket() :
  *  Returns a pointer to the start of the bucket associated with hash. */
 static ldmEntry_t* ZSTD_ldm_getBucket(
@@ -103,38 +181,12 @@ static void ZSTD_ldm_insertEntry(ldmState_t* ldmState,
                                  size_t const hash, const ldmEntry_t entry,
                                  ldmParams_t const ldmParams)
 {
-    BYTE* const bucketOffsets = ldmState->bucketOffsets;
-    *(ZSTD_ldm_getBucket(ldmState, hash, ldmParams) + bucketOffsets[hash]) = entry;
-    bucketOffsets[hash]++;
-    bucketOffsets[hash] &= ((U32)1 << ldmParams.bucketSizeLog) - 1;
-}
+    BYTE* const pOffset = ldmState->bucketOffsets + hash;
+    unsigned const offset = *pOffset;
+
+    *(ZSTD_ldm_getBucket(ldmState, hash, ldmParams) + offset) = entry;
+    *pOffset = (BYTE)((offset + 1) & ((1u << ldmParams.bucketSizeLog) - 1));
 
-/** ZSTD_ldm_makeEntryAndInsertByTag() :
- *
- *  Gets the small hash, checksum, and tag from the rollingHash.
- *
- *  If the tag matches (1 << ldmParams.hashRateLog)-1, then
- *  creates an ldmEntry from the offset, and inserts it into the hash table.
- *
- *  hBits is the length of the small hash, which is the most significant hBits
- *  of rollingHash. The checksum is the next 32 most significant bits, followed
- *  by ldmParams.hashRateLog bits that make up the tag. */
-static void ZSTD_ldm_makeEntryAndInsertByTag(ldmState_t* ldmState,
-                                             U64 const rollingHash,
-                                             U32 const hBits,
-                                             U32 const offset,
-                                             ldmParams_t const ldmParams)
-{
-    U32 const tag = ZSTD_ldm_getTag(rollingHash, hBits, ldmParams.hashRateLog);
-    U32 const tagMask = ((U32)1 << ldmParams.hashRateLog) - 1;
-    if (tag == tagMask) {
-        U32 const hash = ZSTD_ldm_getSmallHash(rollingHash, hBits);
-        U32 const checksum = ZSTD_ldm_getChecksum(rollingHash, hBits);
-        ldmEntry_t entry;
-        entry.offset = offset;
-        entry.checksum = checksum;
-        ZSTD_ldm_insertEntry(ldmState, hash, entry, ldmParams);
-    }
 }
 
 /** ZSTD_ldm_countBackwardsMatch() :
@@ -212,43 +264,42 @@ static size_t ZSTD_ldm_fillFastTables(ZSTD_matchState_t* ms,
     return 0;
 }
 
-/** ZSTD_ldm_fillLdmHashTable() :
- *
- *  Fills hashTable from (lastHashed + 1) to iend (non-inclusive).
- *  lastHash is the rolling hash that corresponds to lastHashed.
- *
- *  Returns the rolling hash corresponding to position iend-1. */
-static U64 ZSTD_ldm_fillLdmHashTable(ldmState_t* state,
-                                     U64 lastHash, const BYTE* lastHashed,
-                                     const BYTE* iend, const BYTE* base,
-                                     U32 hBits, ldmParams_t const ldmParams)
-{
-    U64 rollingHash = lastHash;
-    const BYTE* cur = lastHashed + 1;
-
-    while (cur < iend) {
-        rollingHash = ZSTD_rollingHash_rotate(rollingHash, cur[-1],
-                                              cur[ldmParams.minMatchLength-1],
-                                              state->hashPower);
-        ZSTD_ldm_makeEntryAndInsertByTag(state,
-                                         rollingHash, hBits,
-                                         (U32)(cur - base), ldmParams);
-        ++cur;
-    }
-    return rollingHash;
-}
-
 void ZSTD_ldm_fillHashTable(
-            ldmState_t* state, const BYTE* ip,
+            ldmState_t* ldmState, const BYTE* ip,
             const BYTE* iend, ldmParams_t const* params)
 {
+    U32 const minMatchLength = params->minMatchLength;
+    U32 const hBits = params->hashLog - params->bucketSizeLog;
+    BYTE const* const base = ldmState->window.base;
+    BYTE const* const istart = ip;
+    ldmRollingHashState_t hashState;
+    size_t* const splits = ldmState->splitIndices;
+    unsigned numSplits;
+
     DEBUGLOG(5, "ZSTD_ldm_fillHashTable");
-    if ((size_t)(iend - ip) >= params->minMatchLength) {
-        U64 startingHash = ZSTD_rollingHash_compute(ip, params->minMatchLength);
-        ZSTD_ldm_fillLdmHashTable(
-            state, startingHash, ip, iend - params->minMatchLength, state->window.base,
-            params->hashLog - params->bucketSizeLog,
-            *params);
+
+    ZSTD_ldm_gear_init(&hashState, params);
+    while (ip < iend) {
+        size_t hashed;
+        unsigned n;
+
+        numSplits = 0;
+        hashed = ZSTD_ldm_gear_feed(&hashState, ip, iend - ip, splits, &numSplits);
+
+        for (n = 0; n < numSplits; n++) {
+            if (ip + splits[n] >= istart + minMatchLength) {
+                BYTE const* const split = ip + splits[n] - minMatchLength;
+                U64 const xxhash = XXH64(split, minMatchLength, 0);
+                U32 const hash = (U32)(xxhash & (((U32)1 << hBits) - 1));
+                ldmEntry_t entry;
+
+                entry.offset = (U32)(split - base);
+                entry.checksum = (U32)(xxhash >> 32);
+                ZSTD_ldm_insertEntry(ldmState, hash, entry, *params);
+            }
+        }
+
+        ip += hashed;
     }
 }
 
@@ -274,11 +325,8 @@ static size_t ZSTD_ldm_generateSequences_internal(
     /* LDM parameters */
     int const extDict = ZSTD_window_hasExtDict(ldmState->window);
     U32 const minMatchLength = params->minMatchLength;
-    U64 const hashPower = ldmState->hashPower;
+    U32 const entsPerBucket = 1U << params->bucketSizeLog;
     U32 const hBits = params->hashLog - params->bucketSizeLog;
-    U32 const ldmBucketSize = 1U << params->bucketSizeLog;
-    U32 const hashRateLog = params->hashRateLog;
-    U32 const ldmTagMask = (1U << params->hashRateLog) - 1;
     /* Prefix and extDict parameters */
     U32 const dictLimit = ldmState->window.dictLimit;
     U32 const lowestIndex = extDict ? ldmState->window.lowLimit : dictLimit;
@@ -290,45 +338,69 @@ static size_t ZSTD_ldm_generateSequences_internal(
     /* Input bounds */
     BYTE const* const istart = (BYTE const*)src;
     BYTE const* const iend = istart + srcSize;
-    BYTE const* const ilimit = iend - MAX(minMatchLength, HASH_READ_SIZE);
+    BYTE const* const ilimit = iend - HASH_READ_SIZE;
     /* Input positions */
     BYTE const* anchor = istart;
     BYTE const* ip = istart;
-    /* Rolling hash */
-    BYTE const* lastHashed = NULL;
-    U64 rollingHash = 0;
-
-    while (ip <= ilimit) {
-        size_t mLength;
-        U32 const curr = (U32)(ip - base);
-        size_t forwardMatchLength = 0, backwardMatchLength = 0;
-        ldmEntry_t* bestEntry = NULL;
-        if (ip != istart) {
-            rollingHash = ZSTD_rollingHash_rotate(rollingHash, lastHashed[0],
-                                                  lastHashed[minMatchLength],
-                                                  hashPower);
-        } else {
-            rollingHash = ZSTD_rollingHash_compute(ip, minMatchLength);
+    /* Rolling hash state */
+    ldmRollingHashState_t hashState;
+    /* Arrays for staged-processing */
+    size_t* const splits = ldmState->splitIndices;
+    ldmMatchCandidate_t* const candidates = ldmState->matchCandidates;
+    unsigned numSplits;
+
+    if (srcSize < minMatchLength)
+        return iend - anchor;
+
+    /* Initialize the rolling hash state with the first minMatchLength bytes */
+    ZSTD_ldm_gear_init(&hashState, params);
+    ZSTD_ldm_gear_reset(&hashState, ip, minMatchLength);
+    ip += minMatchLength;
+
+    while (ip < ilimit) {
+        size_t hashed;
+        unsigned n;
+
+        numSplits = 0;
+        hashed = ZSTD_ldm_gear_feed(&hashState, ip, ilimit - ip,
+                                    splits, &numSplits);
+
+        for (n = 0; n < numSplits; n++) {
+            BYTE const* const split = ip + splits[n] - minMatchLength;
+            U64 const xxhash = XXH64(split, minMatchLength, 0);
+            U32 const hash = (U32)(xxhash & (((U32)1 << hBits) - 1));
+
+            candidates[n].split = split;
+            candidates[n].hash = hash;
+            candidates[n].checksum = (U32)(xxhash >> 32);
+            candidates[n].bucket = ZSTD_ldm_getBucket(ldmState, hash, *params);
+            PREFETCH_L1(candidates[n].bucket);
         }
-        lastHashed = ip;
 
-        /* Do not insert and do not look for a match */
-        if (ZSTD_ldm_getTag(rollingHash, hBits, hashRateLog) != ldmTagMask) {
-           ip++;
-           continue;
-        }
+        for (n = 0; n < numSplits; n++) {
+            size_t forwardMatchLength = 0, backwardMatchLength = 0,
+                   bestMatchLength = 0, mLength;
+            U32 offset;
+            BYTE const* const split = candidates[n].split;
+            U32 const checksum = candidates[n].checksum;
+            U32 const hash = candidates[n].hash;
+            ldmEntry_t* const bucket = candidates[n].bucket;
+            ldmEntry_t const* cur;
+            ldmEntry_t const* bestEntry = NULL;
+            ldmEntry_t newEntry;
+
+            newEntry.offset = (U32)(split - base);
+            newEntry.checksum = checksum;
+
+            /* If a split point would generate a sequence overlapping with
+             * the previous one, we merely register it in the hash table and
+             * move on */
+            if (split < anchor) {
+                ZSTD_ldm_insertEntry(ldmState, hash, newEntry, *params);
+                continue;
+            }
 
-        /* Get the best entry and compute the match lengths */
-        {
-            ldmEntry_t* const bucket =
-                ZSTD_ldm_getBucket(ldmState,
-                                   ZSTD_ldm_getSmallHash(rollingHash, hBits),
-                                   *params);
-            ldmEntry_t* cur;
-            size_t bestMatchLength = 0;
-            U32 const checksum = ZSTD_ldm_getChecksum(rollingHash, hBits);
-
-            for (cur = bucket; cur < bucket + ldmBucketSize; ++cur) {
+            for (cur = bucket; cur < bucket + entsPerBucket; cur++) {
                 size_t curForwardMatchLength, curBackwardMatchLength,
                        curTotalMatchLength;
                 if (cur->checksum != checksum || cur->offset <= lowestIndex) {
@@ -342,31 +414,23 @@ static size_t ZSTD_ldm_generateSequences_internal(
                         cur->offset < dictLimit ? dictEnd : iend;
                     BYTE const* const lowMatchPtr =
                         cur->offset < dictLimit ? dictStart : lowPrefixPtr;
-
-                    curForwardMatchLength = ZSTD_count_2segments(
-                                                ip, pMatch, iend,
-                                                matchEnd, lowPrefixPtr);
+                    curForwardMatchLength =
+                        ZSTD_count_2segments(split, pMatch, iend, matchEnd, lowPrefixPtr);
                     if (curForwardMatchLength < minMatchLength) {
                         continue;
                     }
-                    curBackwardMatchLength =
-                        ZSTD_ldm_countBackwardsMatch_2segments(ip, anchor,
-                                                               pMatch, lowMatchPtr,
-                                                               dictStart, dictEnd);
-                    curTotalMatchLength = curForwardMatchLength +
-                                          curBackwardMatchLength;
+                    curBackwardMatchLength = ZSTD_ldm_countBackwardsMatch_2segments(
+                            split, anchor, pMatch, lowMatchPtr, dictStart, dictEnd);
                 } else { /* !extDict */
                     BYTE const* const pMatch = base + cur->offset;
-                    curForwardMatchLength = ZSTD_count(ip, pMatch, iend);
+                    curForwardMatchLength = ZSTD_count(split, pMatch, iend);
                     if (curForwardMatchLength < minMatchLength) {
                         continue;
                     }
                     curBackwardMatchLength =
-                        ZSTD_ldm_countBackwardsMatch(ip, anchor, pMatch,
-                                                     lowPrefixPtr);
-                    curTotalMatchLength = curForwardMatchLength +
-                                          curBackwardMatchLength;
+                        ZSTD_ldm_countBackwardsMatch(split, anchor, pMatch, lowPrefixPtr);
                 }
+                curTotalMatchLength = curForwardMatchLength + curBackwardMatchLength;
 
                 if (curTotalMatchLength > bestMatchLength) {
                     bestMatchLength = curTotalMatchLength;
@@ -375,57 +439,54 @@ static size_t ZSTD_ldm_generateSequences_internal(
                     bestEntry = cur;
                 }
             }
-        }
-
-        /* No match found -- continue searching */
-        if (bestEntry == NULL) {
-            ZSTD_ldm_makeEntryAndInsertByTag(ldmState, rollingHash,
-                                             hBits, curr,
-                                             *params);
-            ip++;
-            continue;
-        }
 
-        /* Match found */
-        mLength = forwardMatchLength + backwardMatchLength;
-        ip -= backwardMatchLength;
+            /* No match found -- insert an entry into the hash table
+             * and process the next candidate match */
+            if (bestEntry == NULL) {
+                ZSTD_ldm_insertEntry(ldmState, hash, newEntry, *params);
+                continue;
+            }
 
-        {
-            /* Store the sequence:
-             * ip = curr - backwardMatchLength
-             * The match is at (bestEntry->offset - backwardMatchLength)
-             */
-            U32 const matchIndex = bestEntry->offset;
-            U32 const offset = curr - matchIndex;
-            rawSeq* const seq = rawSeqStore->seq + rawSeqStore->size;
-
-            /* Out of sequence storage */
-            if (rawSeqStore->size == rawSeqStore->capacity)
-                return ERROR(dstSize_tooSmall);
-            seq->litLength = (U32)(ip - anchor);
-            seq->matchLength = (U32)mLength;
-            seq->offset = offset;
-            rawSeqStore->size++;
-        }
+            /* Match found */
+            offset = (U32)(split - base) - bestEntry->offset;
+            mLength = forwardMatchLength + backwardMatchLength;
+            {
+                rawSeq* const seq = rawSeqStore->seq + rawSeqStore->size;
+
+                /* Out of sequence storage */
+                if (rawSeqStore->size == rawSeqStore->capacity)
+                    return ERROR(dstSize_tooSmall);
+                seq->litLength = (U32)(split - backwardMatchLength - anchor);
+                seq->matchLength = (U32)mLength;
+                seq->offset = offset;
+                rawSeqStore->size++;
+            }
 
-        /* Insert the current entry into the hash table */
-        ZSTD_ldm_makeEntryAndInsertByTag(ldmState, rollingHash, hBits,
-                                         (U32)(lastHashed - base),
-                                         *params);
+            /* Insert the current entry into the hash table --- it must be
+             * done after the previous block to avoid clobbering bestEntry */
+            ZSTD_ldm_insertEntry(ldmState, hash, newEntry, *params);
 
-        assert(ip + backwardMatchLength == lastHashed);
+            anchor = split + forwardMatchLength;
 
-        /* Fill the hash table from lastHashed+1 to ip+mLength*/
-        /* Heuristic: don't need to fill the entire table at end of block */
-        if (ip + mLength <= ilimit) {
-            rollingHash = ZSTD_ldm_fillLdmHashTable(
-                              ldmState, rollingHash, lastHashed,
-                              ip + mLength, base, hBits, *params);
-            lastHashed = ip + mLength - 1;
+            /* If we find a match that ends after the data that we've hashed
+             * then we have a repeating, overlapping, pattern. E.g. all zeros.
+             * If one repetition of the pattern matches our `stopMask` then all
+             * repetitions will. We don't need to insert them all into out table,
+             * only the first one. So skip over overlapping matches.
+             * This is a major speed boost (20x) for compressing a single byte
+             * repeated, when that byte ends up in the table.
+             */
+            if (anchor > ip + hashed) {
+                ZSTD_ldm_gear_reset(&hashState, anchor - minMatchLength, minMatchLength);
+                /* Continue the outter loop at anchor (ip + hashed == anchor). */
+                ip = anchor - hashed;
+                break;
+            }
         }
-        ip += mLength;
-        anchor = ip;
+
+        ip += hashed;
     }
+
     return iend - anchor;
 }
 
@@ -474,7 +535,7 @@ size_t ZSTD_ldm_generateSequences(
 
         assert(chunkStart < iend);
         /* 1. Perform overflow correction if necessary. */
-        if (ZSTD_window_needOverflowCorrection(ldmState->window, chunkEnd)) {
+        if (ZSTD_window_needOverflowCorrection(ldmState->window, 0, maxDist, ldmState->loadedDictEnd, chunkStart, chunkEnd)) {
             U32 const ldmHSize = 1U << params->hashLog;
             U32 const correction = ZSTD_window_correctOverflow(
                 &ldmState->window, /* cycleLog */ 0, maxDist, chunkStart);
@@ -596,12 +657,13 @@ 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,
     void const* src, size_t srcSize)
 {
     const ZSTD_compressionParameters* const cParams = &ms->cParams;
     unsigned const minMatch = cParams->minMatch;
     ZSTD_blockCompressor const blockCompressor =
-        ZSTD_selectBlockCompressor(cParams->strategy, ZSTD_matchState_dictMode(ms));
+        ZSTD_selectBlockCompressor(cParams->strategy, useRowMatchFinder, ZSTD_matchState_dictMode(ms));
     /* Input bounds */
     BYTE const* const istart = (BYTE const*)src;
     BYTE const* const iend = istart + srcSize;
@@ -620,7 +682,7 @@ size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore,
 
     assert(rawSeqStore->pos <= rawSeqStore->size);
     assert(rawSeqStore->size <= rawSeqStore->capacity);
-    /* Loop through each sequence and apply the block compressor to the lits */
+    /* Loop through each sequence and apply the block compressor to the literals */
     while (rawSeqStore->pos < rawSeqStore->size && ip < iend) {
         /* maybeSplitSequence updates rawSeqStore->pos */
         rawSeq const sequence = maybeSplitSequence(rawSeqStore,
diff --git a/zstd_ldm.h b/zstd_ldm.h
index e95cfd1..700a98b 100644
--- a/zstd_ldm.h
+++ b/zstd_ldm.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -66,6 +66,7 @@ size_t ZSTD_ldm_generateSequences(
  */
 size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore,
             ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
+            ZSTD_useRowMatchFinderMode_e useRowMatchFinder,
             void const* src, size_t srcSize);
 
 /**
@@ -73,7 +74,7 @@ size_t ZSTD_ldm_blockCompress(rawSeqStore_t* rawSeqStore,
  *
  * Skip past `srcSize` bytes worth of sequences in `rawSeqStore`.
  * Avoids emitting matches less than `minMatch` bytes.
- * Must be called for data with is not passed to ZSTD_ldm_blockCompress().
+ * Must be called for data that is not passed to ZSTD_ldm_blockCompress().
  */
 void ZSTD_ldm_skipSequences(rawSeqStore_t* rawSeqStore, size_t srcSize,
     U32 const minMatch);
diff --git a/zstd_ldm_geartab.h b/zstd_ldm_geartab.h
new file mode 100644
index 0000000..2b15c02
--- /dev/null
+++ b/zstd_ldm_geartab.h
@@ -0,0 +1,106 @@
+/*
+ * 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_LDM_GEARTAB_H
+#define ZSTD_LDM_GEARTAB_H
+
+#include "compiler.h" /* UNUSED_ATTR */
+#include "mem.h"      /* U64 */
+
+static UNUSED_ATTR const U64 ZSTD_ldm_gearTab[256] = {
+    0xf5b8f72c5f77775c, 0x84935f266b7ac412, 0xb647ada9ca730ccc,
+    0xb065bb4b114fb1de, 0x34584e7e8c3a9fd0, 0x4e97e17c6ae26b05,
+    0x3a03d743bc99a604, 0xcecd042422c4044f, 0x76de76c58524259e,
+    0x9c8528f65badeaca, 0x86563706e2097529, 0x2902475fa375d889,
+    0xafb32a9739a5ebe6, 0xce2714da3883e639, 0x21eaf821722e69e,
+    0x37b628620b628,    0x49a8d455d88caf5,  0x8556d711e6958140,
+    0x4f7ae74fc605c1f,  0x829f0c3468bd3a20, 0x4ffdc885c625179e,
+    0x8473de048a3daf1b, 0x51008822b05646b2, 0x69d75d12b2d1cc5f,
+    0x8c9d4a19159154bc, 0xc3cc10f4abbd4003, 0xd06ddc1cecb97391,
+    0xbe48e6e7ed80302e, 0x3481db31cee03547, 0xacc3f67cdaa1d210,
+    0x65cb771d8c7f96cc, 0x8eb27177055723dd, 0xc789950d44cd94be,
+    0x934feadc3700b12b, 0x5e485f11edbdf182, 0x1e2e2a46fd64767a,
+    0x2969ca71d82efa7c, 0x9d46e9935ebbba2e, 0xe056b67e05e6822b,
+    0x94d73f55739d03a0, 0xcd7010bdb69b5a03, 0x455ef9fcd79b82f4,
+    0x869cb54a8749c161, 0x38d1a4fa6185d225, 0xb475166f94bbe9bb,
+    0xa4143548720959f1, 0x7aed4780ba6b26ba, 0xd0ce264439e02312,
+    0x84366d746078d508, 0xa8ce973c72ed17be, 0x21c323a29a430b01,
+    0x9962d617e3af80ee, 0xab0ce91d9c8cf75b, 0x530e8ee6d19a4dbc,
+    0x2ef68c0cf53f5d72, 0xc03a681640a85506, 0x496e4e9f9c310967,
+    0x78580472b59b14a0, 0x273824c23b388577, 0x66bf923ad45cb553,
+    0x47ae1a5a2492ba86, 0x35e304569e229659, 0x4765182a46870b6f,
+    0x6cbab625e9099412, 0xddac9a2e598522c1, 0x7172086e666624f2,
+    0xdf5003ca503b7837, 0x88c0c1db78563d09, 0x58d51865acfc289d,
+    0x177671aec65224f1, 0xfb79d8a241e967d7, 0x2be1e101cad9a49a,
+    0x6625682f6e29186b, 0x399553457ac06e50, 0x35dffb4c23abb74,
+    0x429db2591f54aade, 0xc52802a8037d1009, 0x6acb27381f0b25f3,
+    0xf45e2551ee4f823b, 0x8b0ea2d99580c2f7, 0x3bed519cbcb4e1e1,
+    0xff452823dbb010a,  0x9d42ed614f3dd267, 0x5b9313c06257c57b,
+    0xa114b8008b5e1442, 0xc1fe311c11c13d4b, 0x66e8763ea34c5568,
+    0x8b982af1c262f05d, 0xee8876faaa75fbb7, 0x8a62a4d0d172bb2a,
+    0xc13d94a3b7449a97, 0x6dbbba9dc15d037c, 0xc786101f1d92e0f1,
+    0xd78681a907a0b79b, 0xf61aaf2962c9abb9, 0x2cfd16fcd3cb7ad9,
+    0x868c5b6744624d21, 0x25e650899c74ddd7, 0xba042af4a7c37463,
+    0x4eb1a539465a3eca, 0xbe09dbf03b05d5ca, 0x774e5a362b5472ba,
+    0x47a1221229d183cd, 0x504b0ca18ef5a2df, 0xdffbdfbde2456eb9,
+    0x46cd2b2fbee34634, 0xf2aef8fe819d98c3, 0x357f5276d4599d61,
+    0x24a5483879c453e3, 0x88026889192b4b9,  0x28da96671782dbec,
+    0x4ef37c40588e9aaa, 0x8837b90651bc9fb3, 0xc164f741d3f0e5d6,
+    0xbc135a0a704b70ba, 0x69cd868f7622ada,  0xbc37ba89e0b9c0ab,
+    0x47c14a01323552f6, 0x4f00794bacee98bb, 0x7107de7d637a69d5,
+    0x88af793bb6f2255e, 0xf3c6466b8799b598, 0xc288c616aa7f3b59,
+    0x81ca63cf42fca3fd, 0x88d85ace36a2674b, 0xd056bd3792389e7,
+    0xe55c396c4e9dd32d, 0xbefb504571e6c0a6, 0x96ab32115e91e8cc,
+    0xbf8acb18de8f38d1, 0x66dae58801672606, 0x833b6017872317fb,
+    0xb87c16f2d1c92864, 0xdb766a74e58b669c, 0x89659f85c61417be,
+    0xc8daad856011ea0c, 0x76a4b565b6fe7eae, 0xa469d085f6237312,
+    0xaaf0365683a3e96c, 0x4dbb746f8424f7b8, 0x638755af4e4acc1,
+    0x3d7807f5bde64486, 0x17be6d8f5bbb7639, 0x903f0cd44dc35dc,
+    0x67b672eafdf1196c, 0xa676ff93ed4c82f1, 0x521d1004c5053d9d,
+    0x37ba9ad09ccc9202, 0x84e54d297aacfb51, 0xa0b4b776a143445,
+    0x820d471e20b348e,  0x1874383cb83d46dc, 0x97edeec7a1efe11c,
+    0xb330e50b1bdc42aa, 0x1dd91955ce70e032, 0xa514cdb88f2939d5,
+    0x2791233fd90db9d3, 0x7b670a4cc50f7a9b, 0x77c07d2a05c6dfa5,
+    0xe3778b6646d0a6fa, 0xb39c8eda47b56749, 0x933ed448addbef28,
+    0xaf846af6ab7d0bf4, 0xe5af208eb666e49,  0x5e6622f73534cd6a,
+    0x297daeca42ef5b6e, 0x862daef3d35539a6, 0xe68722498f8e1ea9,
+    0x981c53093dc0d572, 0xfa09b0bfbf86fbf5, 0x30b1e96166219f15,
+    0x70e7d466bdc4fb83, 0x5a66736e35f2a8e9, 0xcddb59d2b7c1baef,
+    0xd6c7d247d26d8996, 0xea4e39eac8de1ba3, 0x539c8bb19fa3aff2,
+    0x9f90e4c5fd508d8,  0xa34e5956fbaf3385, 0x2e2f8e151d3ef375,
+    0x173691e9b83faec1, 0xb85a8d56bf016379, 0x8382381267408ae3,
+    0xb90f901bbdc0096d, 0x7c6ad32933bcec65, 0x76bb5e2f2c8ad595,
+    0x390f851a6cf46d28, 0xc3e6064da1c2da72, 0xc52a0c101cfa5389,
+    0xd78eaf84a3fbc530, 0x3781b9e2288b997e, 0x73c2f6dea83d05c4,
+    0x4228e364c5b5ed7,  0x9d7a3edf0da43911, 0x8edcfeda24686756,
+    0x5e7667a7b7a9b3a1, 0x4c4f389fa143791d, 0xb08bc1023da7cddc,
+    0x7ab4be3ae529b1cc, 0x754e6132dbe74ff9, 0x71635442a839df45,
+    0x2f6fb1643fbe52de, 0x961e0a42cf7a8177, 0xf3b45d83d89ef2ea,
+    0xee3de4cf4a6e3e9b, 0xcd6848542c3295e7, 0xe4cee1664c78662f,
+    0x9947548b474c68c4, 0x25d73777a5ed8b0b, 0xc915b1d636b7fc,
+    0x21c2ba75d9b0d2da, 0x5f6b5dcf608a64a1, 0xdcf333255ff9570c,
+    0x633b922418ced4ee, 0xc136dde0b004b34a, 0x58cc83b05d4b2f5a,
+    0x5eb424dda28e42d2, 0x62df47369739cd98, 0xb4e0b42485e4ce17,
+    0x16e1f0c1f9a8d1e7, 0x8ec3916707560ebf, 0x62ba6e2df2cc9db3,
+    0xcbf9f4ff77d83a16, 0x78d9d7d07d2bbcc4, 0xef554ce1e02c41f4,
+    0x8d7581127eccf94d, 0xa9b53336cb3c8a05, 0x38c42c0bf45c4f91,
+    0x640893cdf4488863, 0x80ec34bc575ea568, 0x39f324f5b48eaa40,
+    0xe9d9ed1f8eff527f, 0x9224fc058cc5a214, 0xbaba00b04cfe7741,
+    0x309a9f120fcf52af, 0xa558f3ec65626212, 0x424bec8b7adabe2f,
+    0x41622513a6aea433, 0xb88da2d5324ca798, 0xd287733b245528a4,
+    0x9a44697e6d68aec3, 0x7b1093be2f49bb28, 0x50bbec632e3d8aad,
+    0x6cd90723e1ea8283, 0x897b9e7431b02bf3, 0x219efdcb338a7047,
+    0x3b0311f0a27c0656, 0xdb17bf91c0db96e7, 0x8cd4fd6b4e85a5b2,
+    0xfab071054ba6409d, 0x40d6fe831fa9dfd9, 0xaf358debad7d791e,
+    0xeb8d0e25a65e3e58, 0xbbcbd3df14e08580, 0xcf751f27ecdab2b,
+    0x2b4da14f2613d8f4
+};
+
+#endif /* ZSTD_LDM_GEARTAB_H */
diff --git a/zstd_legacy.h b/zstd_legacy.h
index ecd9682..6ba3637 100644
--- a/zstd_legacy.h
+++ b/zstd_legacy.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_opt.c b/zstd_opt.c
index e55c459..301f985 100644
--- a/zstd_opt.c
+++ b/zstd_opt.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Przemyslaw Skibinski, Yann Collet, Facebook, Inc.
+ * Copyright (c) Przemyslaw Skibinski, Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -126,7 +126,7 @@ 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];
@@ -253,11 +253,10 @@ static U32 ZSTD_litLengthPrice(U32 const litLength, const optState_t* const optP
  * Provides the cost of the match part (offset + matchLength) of a sequence
  * Must be combined with ZSTD_fullLiteralsCost() to get the full cost of a sequence.
  * optLevel: when <2, favors small offset for decompression speed (improved cache efficiency) */
-FORCE_INLINE_TEMPLATE U32
-ZSTD_getMatchPrice(U32 const offset,
-                   U32 const matchLength,
-             const optState_t* const optPtr,
-                   int const optLevel)
+static U32 ZSTD_getMatchPrice(U32 const offset,
+                              U32 const matchLength,
+                        const optState_t* const optPtr,
+                              int const optLevel)
 {
     U32 price;
     U32 const offCode = ZSTD_highbit32(offset+1);
@@ -364,11 +363,13 @@ 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 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 +392,10 @@ 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're 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;
@@ -404,6 +408,7 @@ 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 */
 
@@ -479,11 +484,9 @@ static U32 ZSTD_insertBt1(
     }
 }
 
-FORCE_INLINE_TEMPLATE
-void ZSTD_updateTree_internal(
-                ZSTD_matchState_t* ms,
-                const BYTE* const ip, const BYTE* const iend,
-                const U32 mls, const ZSTD_dictMode_e dictMode)
+static void ZSTD_updateTree_internal(ZSTD_matchState_t* ms,
+                               const BYTE* const ip, const BYTE* const iend,
+                               const U32 mls, const ZSTD_dictMode_e dictMode)
 {
     const BYTE* const base = ms->window.base;
     U32 const target = (U32)(ip - base);
@@ -492,7 +495,7 @@ 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;
     }
@@ -505,8 +508,7 @@ void ZSTD_updateTree(ZSTD_matchState_t* ms, const BYTE* ip, const BYTE* iend) {
     ZSTD_updateTree_internal(ms, ip, iend, ms->cParams.minMatch, ZSTD_noDict);
 }
 
-FORCE_INLINE_TEMPLATE
-U32 ZSTD_insertBtAndGetAllMatches (
+static U32 ZSTD_insertBtAndGetAllMatches (
                     ZSTD_match_t* matches,   /* store result (found matches) in this table (presumed large enough) */
                     ZSTD_matchState_t* ms,
                     U32* nextToUpdate3,
@@ -739,7 +741,7 @@ U32 ZSTD_insertBtAndGetAllMatches (
 }
 
 
-FORCE_INLINE_TEMPLATE U32 ZSTD_BtGetAllMatches (
+static U32 ZSTD_BtGetAllMatches (
                         ZSTD_match_t* matches,   /* store result (match found, increasing size) in this table */
                         ZSTD_matchState_t* ms,
                         U32* nextToUpdate3,
@@ -893,7 +895,7 @@ static void ZSTD_optLdm_processMatchCandidate(ZSTD_optLdm_t* optLdm, ZSTD_match_
              */
             U32 posOvershoot = currPosInBlock - optLdm->endPosInBlock;
             ZSTD_optLdm_skipRawSeqStoreBytes(&optLdm->seqStore, posOvershoot);
-        } 
+        }
         ZSTD_opt_getNextMatchAndUpdateSeqStore(optLdm, currPosInBlock, remainingBytes);
     }
     ZSTD_optLdm_maybeAddMatch(matches, nbMatches, optLdm, currPosInBlock);
@@ -926,7 +928,7 @@ listStats(const U32* table, int lastEltID)
 
 #endif
 
-FORCE_INLINE_TEMPLATE size_t
+static size_t
 ZSTD_compressBlock_opt_generic(ZSTD_matchState_t* ms,
                                seqStore_t* seqStore,
                                U32 rep[ZSTD_REP_NUM],
diff --git a/zstd_opt.h b/zstd_opt.h
index 9aba8a9..627255f 100644
--- a/zstd_opt.h
+++ b/zstd_opt.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_trace.h b/zstd_trace.h
new file mode 100644
index 0000000..f9121f7
--- /dev/null
+++ b/zstd_trace.h
@@ -0,0 +1,163 @@
+/*
+ * 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_TRACE_H
+#define ZSTD_TRACE_H
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+#include <stddef.h>
+
+/* weak symbol support
+ * For now, enable conservatively:
+ * - Only GNUC
+ * - Only ELF
+ * - Only x86-64 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(_AIX)
+#  define ZSTD_HAVE_WEAK_SYMBOLS 1
+#else
+#  define ZSTD_HAVE_WEAK_SYMBOLS 0
+#endif
+#if ZSTD_HAVE_WEAK_SYMBOLS
+#  define ZSTD_WEAK_ATTR __attribute__((__weak__))
+#else
+#  define ZSTD_WEAK_ATTR
+#endif
+
+/* Only enable tracing when weak symbols are available. */
+#ifndef ZSTD_TRACE
+#  define ZSTD_TRACE ZSTD_HAVE_WEAK_SYMBOLS
+#endif
+
+#if ZSTD_TRACE
+
+struct ZSTD_CCtx_s;
+struct ZSTD_DCtx_s;
+struct ZSTD_CCtx_params_s;
+
+typedef struct {
+    /**
+     * ZSTD_VERSION_NUMBER
+     *
+     * This is guaranteed to be the first member of ZSTD_trace.
+     * Otherwise, this struct is not stable between versions. If
+     * the version number does not match your expectation, you
+     * should not interpret the rest of the struct.
+     */
+    unsigned version;
+    /**
+     * Non-zero if streaming (de)compression is used.
+     */
+    unsigned streaming;
+    /**
+     * The dictionary ID.
+     */
+    unsigned dictionaryID;
+    /**
+     * Is the dictionary cold?
+     * Only set on decompression.
+     */
+    unsigned dictionaryIsCold;
+    /**
+     * The dictionary size or zero if no dictionary.
+     */
+    size_t dictionarySize;
+    /**
+     * The uncompressed size of the data.
+     */
+    size_t uncompressedSize;
+    /**
+     * The compressed size of the data.
+     */
+    size_t compressedSize;
+    /**
+     * The fully resolved CCtx parameters (NULL on decompression).
+     */
+    struct ZSTD_CCtx_params_s const* params;
+    /**
+     * The ZSTD_CCtx pointer (NULL on decompression).
+     */
+    struct ZSTD_CCtx_s const* cctx;
+    /**
+     * The ZSTD_DCtx pointer (NULL on compression).
+     */
+    struct ZSTD_DCtx_s const* dctx;
+} ZSTD_Trace;
+
+/**
+ * A tracing context. It must be 0 when tracing is disabled.
+ * Otherwise, any non-zero value returned by a tracing begin()
+ * function is presented to any subsequent calls to end().
+ *
+ * Any non-zero value is treated as tracing is enabled and not
+ * interpreted by the library.
+ *
+ * Two possible uses are:
+ * * A timestamp for when the begin() function was called.
+ * * A unique key identifying the (de)compression, like the
+ *   address of the [dc]ctx pointer if you need to track
+ *   more information than just a timestamp.
+ */
+typedef unsigned long long ZSTD_TraceCtx;
+
+/**
+ * Trace the beginning of a compression call.
+ * @param cctx The dctx pointer for the compression.
+ *             It can be used as a key to map begin() to end().
+ * @returns Non-zero if tracing is enabled. The return value is
+ *          passed to ZSTD_trace_compress_end().
+ */
+ZSTD_WEAK_ATTR ZSTD_TraceCtx ZSTD_trace_compress_begin(
+    struct ZSTD_CCtx_s const* cctx);
+
+/**
+ * Trace the end of a compression call.
+ * @param ctx The return value of ZSTD_trace_compress_begin().
+ * @param trace The zstd tracing info.
+ */
+ZSTD_WEAK_ATTR void ZSTD_trace_compress_end(
+    ZSTD_TraceCtx ctx,
+    ZSTD_Trace const* trace);
+
+/**
+ * Trace the beginning of a decompression call.
+ * @param dctx The dctx pointer for the decompression.
+ *             It can be used as a key to map begin() to end().
+ * @returns Non-zero if tracing is enabled. The return value is
+ *          passed to ZSTD_trace_compress_end().
+ */
+ZSTD_WEAK_ATTR ZSTD_TraceCtx ZSTD_trace_decompress_begin(
+    struct ZSTD_DCtx_s const* dctx);
+
+/**
+ * Trace the end of a decompression call.
+ * @param ctx The return value of ZSTD_trace_decompress_begin().
+ * @param trace The zstd tracing info.
+ */
+ZSTD_WEAK_ATTR void ZSTD_trace_decompress_end(
+    ZSTD_TraceCtx ctx,
+    ZSTD_Trace const* trace);
+
+#endif /* ZSTD_TRACE */
+
+#if defined (__cplusplus)
+}
+#endif
+
+#endif /* ZSTD_TRACE_H */
diff --git a/zstd_v01.c b/zstd_v01.c
index 3eb34eb..d6e1920 100644
--- a/zstd_v01.c
+++ b/zstd_v01.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_v01.h b/zstd_v01.h
index 7910351..f777eb6 100644
--- a/zstd_v01.h
+++ b/zstd_v01.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_v02.c b/zstd_v02.c
index 3bb580f..2091ee9 100644
--- a/zstd_v02.c
+++ b/zstd_v02.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_v02.h b/zstd_v02.h
index 5f8f6cd..1b37195 100644
--- a/zstd_v02.h
+++ b/zstd_v02.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_v03.c b/zstd_v03.c
index 7dedf03..272d6da 100644
--- a/zstd_v03.c
+++ b/zstd_v03.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_v03.h b/zstd_v03.h
index 5fc7273..7a00d43 100644
--- a/zstd_v03.h
+++ b/zstd_v03.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_v04.c b/zstd_v04.c
index 7abb698..ad07e33 100644
--- a/zstd_v04.c
+++ b/zstd_v04.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_v04.h b/zstd_v04.h
index 15fce0d..66b97ab 100644
--- a/zstd_v04.h
+++ b/zstd_v04.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_v05.c b/zstd_v05.c
index b2053be..70a01ee 100644
--- a/zstd_v05.c
+++ b/zstd_v05.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -2833,7 +2833,7 @@ static size_t ZSTDv05_decodeFrameHeader_Part2(ZSTDv05_DCtx* zc, const void* src,
 
 static size_t ZSTDv05_getcBlockSize(const void* src, size_t srcSize, blockProperties_t* bpPtr)
 {
-    const BYTE* const in = (const BYTE* const)src;
+    const BYTE* const in = (const BYTE*)src;
     BYTE headerFlags;
     U32 cSize;
 
@@ -3002,7 +3002,7 @@ static size_t ZSTDv05_decodeSeqHeaders(int* nbSeq, const BYTE** dumpsPtr, size_t
                          FSEv05_DTable* DTableLL, FSEv05_DTable* DTableML, FSEv05_DTable* DTableOffb,
                          const void* src, size_t srcSize, U32 flagStaticTable)
 {
-    const BYTE* const istart = (const BYTE* const)src;
+    const BYTE* const istart = (const BYTE*)src;
     const BYTE* ip = istart;
     const BYTE* const iend = istart + srcSize;
     U32 LLtype, Offtype, MLtype;
@@ -3310,7 +3310,7 @@ static size_t ZSTDv05_decompressSequences(
 {
     const BYTE* ip = (const BYTE*)seqStart;
     const BYTE* const iend = ip + seqSize;
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     BYTE* op = ostart;
     BYTE* const oend = ostart + maxDstSize;
     size_t errorCode, dumpsLength=0;
@@ -3423,7 +3423,7 @@ static size_t ZSTDv05_decompress_continueDCtx(ZSTDv05_DCtx* dctx,
 {
     const BYTE* ip = (const BYTE*)src;
     const BYTE* iend = ip + srcSize;
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     BYTE* op = ostart;
     BYTE* const oend = ostart + maxDstSize;
     size_t remainingSize = srcSize;
diff --git a/zstd_v05.h b/zstd_v05.h
index 6c46e8f..d555175 100644
--- a/zstd_v05.h
+++ b/zstd_v05.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_v06.c b/zstd_v06.c
index cf17623..3b684b6 100644
--- a/zstd_v06.c
+++ b/zstd_v06.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -3029,7 +3029,7 @@ typedef struct
 *   Provides the size of compressed block from block header `src` */
 static size_t ZSTDv06_getcBlockSize(const void* src, size_t srcSize, blockProperties_t* bpPtr)
 {
-    const BYTE* const in = (const BYTE* const)src;
+    const BYTE* const in = (const BYTE*)src;
     U32 cSize;
 
     if (srcSize < ZSTDv06_blockHeaderSize) return ERROR(srcSize_wrong);
@@ -3223,7 +3223,7 @@ static size_t ZSTDv06_decodeSeqHeaders(int* nbSeqPtr,
                              FSEv06_DTable* DTableLL, FSEv06_DTable* DTableML, FSEv06_DTable* DTableOffb, U32 flagRepeatTable,
                              const void* src, size_t srcSize)
 {
-    const BYTE* const istart = (const BYTE* const)src;
+    const BYTE* const istart = (const BYTE*)src;
     const BYTE* const iend = istart + srcSize;
     const BYTE* ip = istart;
 
@@ -3445,7 +3445,7 @@ static size_t ZSTDv06_decompressSequences(
 {
     const BYTE* ip = (const BYTE*)seqStart;
     const BYTE* const iend = ip + seqSize;
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     BYTE* const oend = ostart + maxDstSize;
     BYTE* op = ostart;
     const BYTE* litPtr = dctx->litPtr;
@@ -3561,7 +3561,7 @@ static size_t ZSTDv06_decompressFrame(ZSTDv06_DCtx* dctx,
 {
     const BYTE* ip = (const BYTE*)src;
     const BYTE* const iend = ip + srcSize;
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     BYTE* op = ostart;
     BYTE* const oend = ostart + dstCapacity;
     size_t remainingSize = srcSize;
diff --git a/zstd_v06.h b/zstd_v06.h
index 2fd99e6..9e32b76 100644
--- a/zstd_v06.h
+++ b/zstd_v06.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstd_v07.c b/zstd_v07.c
index f486eb9..c87b033 100644
--- a/zstd_v07.c
+++ b/zstd_v07.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -3258,7 +3258,7 @@ typedef struct
 *   Provides the size of compressed block from block header `src` */
 static size_t ZSTDv07_getcBlockSize(const void* src, size_t srcSize, blockProperties_t* bpPtr)
 {
-    const BYTE* const in = (const BYTE* const)src;
+    const BYTE* const in = (const BYTE*)src;
     U32 cSize;
 
     if (srcSize < ZSTDv07_blockHeaderSize) return ERROR(srcSize_wrong);
@@ -3453,7 +3453,7 @@ static size_t ZSTDv07_decodeSeqHeaders(int* nbSeqPtr,
                              FSEv07_DTable* DTableLL, FSEv07_DTable* DTableML, FSEv07_DTable* DTableOffb, U32 flagRepeatTable,
                              const void* src, size_t srcSize)
 {
-    const BYTE* const istart = (const BYTE* const)src;
+    const BYTE* const istart = (const BYTE*)src;
     const BYTE* const iend = istart + srcSize;
     const BYTE* ip = istart;
 
@@ -3672,7 +3672,7 @@ static size_t ZSTDv07_decompressSequences(
 {
     const BYTE* ip = (const BYTE*)seqStart;
     const BYTE* const iend = ip + seqSize;
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     BYTE* const oend = ostart + maxDstSize;
     BYTE* op = ostart;
     const BYTE* litPtr = dctx->litPtr;
@@ -3799,7 +3799,7 @@ static size_t ZSTDv07_decompressFrame(ZSTDv07_DCtx* dctx,
 {
     const BYTE* ip = (const BYTE*)src;
     const BYTE* const iend = ip + srcSize;
-    BYTE* const ostart = (BYTE* const)dst;
+    BYTE* const ostart = (BYTE*)dst;
     BYTE* const oend = ostart + dstCapacity;
     BYTE* op = ostart;
     size_t remainingSize = srcSize;
diff --git a/zstd_v07.h b/zstd_v07.h
index 9da50c4..bc35cfa 100644
--- a/zstd_v07.h
+++ b/zstd_v07.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
diff --git a/zstdmt_compress.c b/zstdmt_compress.c
index 084e48b..8e5dd15 100644
--- a/zstdmt_compress.c
+++ b/zstdmt_compress.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -472,8 +472,6 @@ ZSTDMT_serialState_reset(serialState_t* serialState,
         ZSTD_ldm_adjustParameters(&params.ldmParams, &params.cParams);
         assert(params.ldmParams.hashLog >= params.ldmParams.bucketSizeLog);
         assert(params.ldmParams.hashRateLog < 32);
-        serialState->ldmState.hashPower =
-                ZSTD_rollingHash_primePower(params.ldmParams.minMatchLength);
     } else {
         ZSTD_memset(&params.ldmParams, 0, sizeof(params.ldmParams));
     }
@@ -486,10 +484,10 @@ ZSTDMT_serialState_reset(serialState_t* serialState,
         size_t const hashSize = ((size_t)1 << hashLog) * sizeof(ldmEntry_t);
         unsigned const bucketLog =
             params.ldmParams.hashLog - params.ldmParams.bucketSizeLog;
-        size_t const bucketSize = (size_t)1 << bucketLog;
         unsigned const prevBucketLog =
             serialState->params.ldmParams.hashLog -
             serialState->params.ldmParams.bucketSizeLog;
+        size_t const numBuckets = (size_t)1 << bucketLog;
         /* Size the seq pool tables */
         ZSTDMT_setNbSeq(seqPool, ZSTD_ldm_getMaxNbSeq(params.ldmParams, jobSize));
         /* Reset the window */
@@ -501,20 +499,20 @@ ZSTDMT_serialState_reset(serialState_t* serialState,
         }
         if (serialState->ldmState.bucketOffsets == NULL || prevBucketLog < bucketLog) {
             ZSTD_customFree(serialState->ldmState.bucketOffsets, cMem);
-            serialState->ldmState.bucketOffsets = (BYTE*)ZSTD_customMalloc(bucketSize, cMem);
+            serialState->ldmState.bucketOffsets = (BYTE*)ZSTD_customMalloc(numBuckets, cMem);
         }
         if (!serialState->ldmState.hashTable || !serialState->ldmState.bucketOffsets)
             return 1;
         /* Zero the tables */
         ZSTD_memset(serialState->ldmState.hashTable, 0, hashSize);
-        ZSTD_memset(serialState->ldmState.bucketOffsets, 0, bucketSize);
+        ZSTD_memset(serialState->ldmState.bucketOffsets, 0, numBuckets);
 
         /* Update window state and fill hash table with dict */
         serialState->ldmState.loadedDictEnd = 0;
         if (dictSize > 0) {
             if (dictContentType == ZSTD_dct_rawContent) {
                 BYTE const* const dictEnd = (const BYTE*)dict + dictSize;
-                ZSTD_window_update(&serialState->ldmState.window, dict, dictSize);
+                ZSTD_window_update(&serialState->ldmState.window, dict, dictSize, /* forceNonContiguous */ 0);
                 ZSTD_ldm_fillHashTable(&serialState->ldmState, (const BYTE*)dict, dictEnd, &params.ldmParams);
                 serialState->ldmState.loadedDictEnd = params.forceWindow ? 0 : (U32)(dictEnd - serialState->ldmState.window.base);
             } else {
@@ -571,7 +569,7 @@ static void ZSTDMT_serialState_update(serialState_t* serialState,
             assert(seqStore.seq != NULL && seqStore.pos == 0 &&
                    seqStore.size == 0 && seqStore.capacity > 0);
             assert(src.size <= serialState->params.jobSize);
-            ZSTD_window_update(&serialState->ldmState.window, src.start, src.size);
+            ZSTD_window_update(&serialState->ldmState.window, src.start, src.size, /* forceNonContiguous */ 0);
             error = ZSTD_ldm_generateSequences(
                 &serialState->ldmState, &seqStore,
                 &serialState->params.ldmParams, src.start, src.size);
@@ -683,6 +681,8 @@ 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;
+    /* Correct nbWorkers to 0. */
+    jobParams.nbWorkers = 0;
 
 
     /* init */
@@ -695,6 +695,10 @@ static void ZSTDMT_compressionJob(void* jobDescription)
         {   size_t const forceWindowError = ZSTD_CCtxParams_setParameter(&jobParams, ZSTD_c_forceMaxWindow, !job->firstJob);
             if (ZSTD_isError(forceWindowError)) JOB_ERROR(forceWindowError);
         }
+        if (!job->firstJob) {
+            size_t const err = ZSTD_CCtxParams_setParameter(&jobParams, ZSTD_c_deterministicRefPrefix, 0);
+            if (ZSTD_isError(err)) JOB_ERROR(err);
+        }
         {   size_t const initError = ZSTD_compressBegin_advanced_internal(cctx,
                                         job->prefix.start, job->prefix.size, ZSTD_dct_rawContent, /* load dictionary in "content-only" mode (no header analysis) */
                                         ZSTD_dtlm_fast,
@@ -750,6 +754,13 @@ static void ZSTDMT_compressionJob(void* jobDescription)
             if (ZSTD_isError(cSize)) JOB_ERROR(cSize);
             lastCBlockSize = cSize;
     }   }
+    if (!job->firstJob) {
+        /* Double check that we don't have an ext-dict, because then our
+         * repcode invalidation doesn't work.
+         */
+        assert(!ZSTD_window_hasExtDict(cctx->blockState.matchState.window));
+    }
+    ZSTD_CCtx_trace(cctx, 0);
 
 _endJob:
     ZSTDMT_serialState_ensureFinished(job->serial, job->jobID, job->cSize);
@@ -1239,9 +1250,8 @@ size_t ZSTDMT_initCStream_internal(
 
     if (params.rsyncable) {
         /* Aim for the targetsectionSize as the average job size. */
-        U32 const jobSizeMB = (U32)(mtctx->targetSectionSize >> 20);
-        U32 const rsyncBits = ZSTD_highbit32(jobSizeMB) + 20;
-        assert(jobSizeMB >= 1);
+        U32 const jobSizeKB = (U32)(mtctx->targetSectionSize >> 10);
+        U32 const rsyncBits = (assert(jobSizeKB >= 1), ZSTD_highbit32(jobSizeKB) + 10);
         DEBUGLOG(4, "rsyncLog = %u", rsyncBits);
         mtctx->rsync.hash = 0;
         mtctx->rsync.hitMask = (1ULL << rsyncBits) - 1;
diff --git a/zstdmt_compress.h b/zstdmt_compress.h
index e306964..5b9565f 100644
--- a/zstdmt_compress.h
+++ b/zstdmt_compress.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Yann Collet, Facebook, Inc.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -32,11 +32,11 @@
 
 
 /* ===   Constants   === */
-#ifndef ZSTDMT_NBWORKERS_MAX
-#  define ZSTDMT_NBWORKERS_MAX 200
+#ifndef ZSTDMT_NBWORKERS_MAX /* a different value can be selected at compile time */
+#  define ZSTDMT_NBWORKERS_MAX ((sizeof(void*)==4) /*32-bit*/ ? 64 : 256)
 #endif
-#ifndef ZSTDMT_JOBSIZE_MIN
-#  define ZSTDMT_JOBSIZE_MIN (1 MB)
+#ifndef ZSTDMT_JOBSIZE_MIN   /* a different value can be selected at compile time */
+#  define ZSTDMT_JOBSIZE_MIN (512 KB)
 #endif
 #define ZSTDMT_JOBLOG_MAX   (MEM_32bits() ? 29 : 30)
 #define ZSTDMT_JOBSIZE_MAX  (MEM_32bits() ? (512 MB) : (1024 MB))