extzstd 0.3.1 → 0.3.3

data/contrib/zstd/lib/compress/zstd_compress_internal.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -23,12 +23,12 @@
 #ifdef ZSTD_MULTITHREAD
 #  include "zstdmt_compress.h"
 #endif
+#include "../common/bits.h" /* ZSTD_highbit32, ZSTD_NbCommonBytes */
 
 #if defined (__cplusplus)
 extern "C" {
 #endif
 
-
 /*-*************************************
 *  Constants
 ***************************************/
@@ -64,7 +64,7 @@ typedef struct {
 } ZSTD_localDict;
 
 typedef struct {
-    U32 CTable[HUF_CTABLE_SIZE_U32(255)];
+    HUF_CElt CTable[HUF_CTABLE_SIZE_ST(255)];
     HUF_repeat repeatMode;
 } ZSTD_hufCTables_t;
 
@@ -82,11 +82,82 @@ 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 {
-    U32 off;
-    U32 len;
+    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(
+                    const 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 sumtype code for the match, using ZSTD_storeSeq() format */
+    U32 len;            /* Raw length of match */
 } ZSTD_match_t;
 
+typedef struct {
+    U32 offset;         /* Offset of sequence */
+    U32 litLength;      /* Length of literals prior to match */
+    U32 matchLength;    /* Raw length of match */
+} rawSeq;
+
+typedef struct {
+  rawSeq* seq;          /* The start of the sequences */
+  size_t pos;           /* The index in seq where reading stopped. pos <= size. */
+  size_t posInSequence; /* The position within the sequence at seq[pos] where reading
+                           stopped. posInSequence <= seq[pos].litLength + seq[pos].matchLength */
+  size_t size;          /* The number of sequences. <= capacity. */
+  size_t capacity;      /* The capacity starting from `seq` pointer */
+} rawSeqStore_t;
+
+typedef struct {
+    U32 idx;            /* Index in array of ZSTD_Sequence */
+    U32 posInSequence;  /* Position within sequence at idx */
+    size_t posInSrc;    /* Number of bytes given by sequences provided so far */
+} ZSTD_sequencePosition;
+
+UNUSED_ATTR static const rawSeqStore_t kNullRawSeqStore = {NULL, 0, 0, 0, 0};
+
 typedef struct {
     int price;
     U32 off;
@@ -116,7 +187,7 @@ typedef struct {
     U32  offCodeSumBasePrice;    /* to compare to log2(offreq)  */
     ZSTD_OptPrice_e priceType;   /* prices can be determined dynamically, or follow a pre-defined cost structure */
     const ZSTD_entropyCTables_t* symbolCosts;  /* pre-calculated dictionary statistics */
-    ZSTD_literalCompressionMode_e literalCompressionMode;
+    ZSTD_paramSwitch_e literalCompressionMode;
 } optState_t;
 
 typedef struct {
@@ -125,14 +196,23 @@ 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.
@@ -144,12 +224,38 @@ 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.*/
+    BYTE* 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 */
+    U64 hashSalt;                            /* For row-based matchFinder: salts the hash for re-use of tag table */
+    U32 hashSaltEntropy;                     /* For row-based matchFinder: collects entropy for salt generation */
+
     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.
+                               */
     optState_t opt;         /* optimal parser state */
     const ZSTD_matchState_t* dictMatchState;
     ZSTD_compressionParameters cParams;
+    const rawSeqStore_t* ldmSeqStore;
+
+    /* Controls prefetching in some dictMatchState matchfinders.
+     * This behavior is controlled from the cctx ms.
+     * This parameter has no effect in the cdict ms. */
+    int prefetchCDictTables;
+
+    /* When == 0, lazy match finders insert every position.
+     * When != 0, lazy match finders only insert positions they search.
+     * This allows them to skip much faster over incompressible data,
+     * at a small cost to compression ratio.
+     */
+    int lazySkipping;
 };
 
 typedef struct {
@@ -163,17 +269,26 @@ 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 {
-    U32 enableLdm;          /* 1 if enable long distance matching */
+    ZSTD_paramSwitch_e enableLdm; /* ZSTD_ps_enable to enable LDM. ZSTD_ps_auto by default */
     U32 hashLog;            /* Log size of hashTable */
     U32 bucketSizeLog;      /* Log bucket size for collision resolution, at most 8 */
     U32 minMatchLength;     /* Minimum match length */
@@ -181,19 +296,6 @@ typedef struct {
     U32 windowLog;          /* Window log for the LDM */
 } ldmParams_t;
 
-typedef struct {
-    U32 offset;
-    U32 litLength;
-    U32 matchLength;
-} rawSeq;
-
-typedef struct {
-  rawSeq* seq;     /* The start of the sequences */
-  size_t pos;      /* The position where reading stopped. <= size. */
-  size_t size;     /* The number of sequences. <= capacity. */
-  size_t capacity; /* The capacity starting from `seq` pointer */
-} rawSeqStore_t;
-
 typedef struct {
     int collectSequences;
     ZSTD_Sequence* seqStart;
@@ -217,7 +319,7 @@ struct ZSTD_CCtx_params_s {
                                 * There is no guarantee that hint is close to actual source size */
 
     ZSTD_dictAttachPref_e attachDictPref;
-    ZSTD_literalCompressionMode_e literalCompressionMode;
+    ZSTD_paramSwitch_e literalCompressionMode;
 
     /* Multithreading: used to pass parameters to mtctx */
     int nbWorkers;
@@ -228,17 +330,94 @@ struct ZSTD_CCtx_params_s {
     /* Long distance matching parameters */
     ldmParams_t ldmParams;
 
+    /* Dedicated dict search algorithm trigger */
+    int enableDedicatedDictSearch;
+
+    /* Input/output buffer modes */
+    ZSTD_bufferMode_e inBufferMode;
+    ZSTD_bufferMode_e outBufferMode;
+
+    /* Sequence compression API */
+    ZSTD_sequenceFormat_e blockDelimiters;
+    int validateSequences;
+
+    /* Block splitting */
+    ZSTD_paramSwitch_e useBlockSplitter;
+
+    /* Param for deciding whether to use row-based matchfinder */
+    ZSTD_paramSwitch_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;
+
+    /* Controls prefetching in some dictMatchState matchfinders */
+    ZSTD_paramSwitch_e prefetchCDictTables;
+
+    /* Controls whether zstd will fall back to an internal matchfinder
+     * if the external matchfinder returns an error code. */
+    int enableMatchFinderFallback;
+
+    /* Indicates whether an external matchfinder has been referenced.
+     * Users can't set this externally.
+     * It is set internally in ZSTD_registerSequenceProducer(). */
+    int useSequenceProducer;
+
+    /* Adjust the max block size*/
+    size_t maxBlockSize;
+
+    /* Controls repcode search in external sequence parsing */
+    ZSTD_paramSwitch_e searchForExternalRepcodes;
 };  /* typedef'd to ZSTD_CCtx_params within "zstd.h" */
 
+#define COMPRESS_SEQUENCES_WORKSPACE_SIZE (sizeof(unsigned) * (MaxSeq + 2))
+#define ENTROPY_WORKSPACE_SIZE (HUF_WORKSPACE_SIZE + COMPRESS_SEQUENCES_WORKSPACE_SIZE)
+
+/**
+ * Indicates whether this compression proceeds directly from user-provided
+ * source buffer to user-provided destination buffer (ZSTDb_not_buffered), or
+ * whether the context needs to buffer the input/output (ZSTDb_buffered).
+ */
+typedef enum {
+    ZSTDb_not_buffered,
+    ZSTDb_buffered
+} ZSTD_buffered_policy_e;
+
+/**
+ * Struct that contains all elements of block splitter that should be allocated
+ * in a wksp.
+ */
+#define ZSTD_MAX_NB_BLOCK_SPLITS 196
+typedef struct {
+    seqStore_t fullSeqStoreChunk;
+    seqStore_t firstHalfSeqStore;
+    seqStore_t secondHalfSeqStore;
+    seqStore_t currSeqStore;
+    seqStore_t nextSeqStore;
+
+    U32 partitions[ZSTD_MAX_NB_BLOCK_SPLITS];
+    ZSTD_entropyCTablesMetadata_t entropyMetadata;
+} ZSTD_blockSplitCtx;
+
+/* Context for block-level external matchfinder API */
+typedef struct {
+  void* mState;
+  ZSTD_sequenceProducer_F* mFinder;
+  ZSTD_Sequence* seqBuffer;
+  size_t seqBufferCapacity;
+} ZSTD_externalMatchCtx;
+
 struct ZSTD_CCtx_s {
     ZSTD_compressionStage_e stage;
     int cParamsChanged;                  /* == 1 if cParams(except wlog) or compression level are changed in requestedParams. Triggers transmission of new params to ZSTDMT (if available) then reset to 0. */
     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;
@@ -247,6 +426,7 @@ struct ZSTD_CCtx_s {
     unsigned long long producedCSize;
     XXH64_state_t xxhState;
     ZSTD_customMem customMem;
+    ZSTD_threadPool* pool;
     size_t staticSize;
     SeqCollector seqCollector;
     int isFirstBlock;
@@ -258,7 +438,10 @@ struct ZSTD_CCtx_s {
     size_t maxNbLdmSequences;
     rawSeqStore_t externSeqStore; /* Mutable reference to external sequences */
     ZSTD_blockState_t blockState;
-    U32* entropyWorkspace;  /* entropy workspace of HUF_WORKSPACE_SIZE bytes */
+    U32* entropyWorkspace;  /* entropy workspace of ENTROPY_WORKSPACE_SIZE bytes */
+
+    /* Whether we are streaming or not */
+    ZSTD_buffered_policy_e bufferedPolicy;
 
     /* streaming */
     char*  inBuff;
@@ -273,6 +456,11 @@ struct ZSTD_CCtx_s {
     ZSTD_cStreamStage streamStage;
     U32    frameEnded;
 
+    /* Stable in/out buffer verification */
+    ZSTD_inBuffer expectedInBuffer;
+    size_t stableIn_notConsumed; /* nb bytes within stable input buffer that are said to be consumed but are not */
+    size_t expectedOutBufferSize;
+
     /* Dictionary */
     ZSTD_localDict localDict;
     const ZSTD_CDict* cdict;
@@ -282,17 +470,53 @@ struct ZSTD_CCtx_s {
 #ifdef ZSTD_MULTITHREAD
     ZSTDMT_CCtx* mtctx;
 #endif
-};
 
-typedef enum { ZSTD_dtlm_fast, ZSTD_dtlm_full } ZSTD_dictTableLoadMethod_e;
+    /* Tracing */
+#if ZSTD_TRACE
+    ZSTD_TraceCtx traceCtx;
+#endif
 
-typedef enum { ZSTD_noDict = 0, ZSTD_extDict = 1, ZSTD_dictMatchState = 2 } ZSTD_dictMode_e;
+    /* Workspace for block splitter */
+    ZSTD_blockSplitCtx blockSplitCtx;
 
+    /* Workspace for external matchfinder */
+    ZSTD_externalMatchCtx externalMatchCtx;
+};
+
+typedef enum { ZSTD_dtlm_fast, ZSTD_dtlm_full } ZSTD_dictTableLoadMethod_e;
+typedef enum { ZSTD_tfp_forCCtx, ZSTD_tfp_forCDict } ZSTD_tableFillPurpose_e;
+
+typedef enum {
+    ZSTD_noDict = 0,
+    ZSTD_extDict = 1,
+    ZSTD_dictMatchState = 2,
+    ZSTD_dedicatedDictSearch = 3
+} ZSTD_dictMode_e;
+
+typedef enum {
+    ZSTD_cpm_noAttachDict = 0,  /* Compression with ZSTD_noDict or ZSTD_extDict.
+                                 * In this mode we use both the srcSize and the dictSize
+                                 * when selecting and adjusting parameters.
+                                 */
+    ZSTD_cpm_attachDict = 1,    /* Compression with ZSTD_dictMatchState or ZSTD_dedicatedDictSearch.
+                                 * In this mode we only take the srcSize into account when selecting
+                                 * and adjusting parameters.
+                                 */
+    ZSTD_cpm_createCDict = 2,   /* Creating a CDict.
+                                 * In this mode we take both the source size and the dictionary size
+                                 * into account when selecting and adjusting the parameters.
+                                 */
+    ZSTD_cpm_unknown = 3        /* ZSTD_getCParams, ZSTD_getParams, ZSTD_adjustParams.
+                                 * We don't know what these parameters are for. We default to the legacy
+                                 * behavior of taking both the source size and the dict size into account
+                                 * when selecting and adjusting parameters.
+                                 */
+} ZSTD_cParamMode_e;
 
 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_paramSwitch_e rowMatchfinderMode, ZSTD_dictMode_e dictMode);
 
 
 MEM_STATIC U32 ZSTD_LLcode(U32 litLength)
@@ -326,31 +550,6 @@ MEM_STATIC U32 ZSTD_MLcode(U32 mlBase)
     return (mlBase > 127) ? ZSTD_highbit32(mlBase) + ML_deltaCode : ML_Code[mlBase];
 }
 
-typedef struct repcodes_s {
-    U32 rep[3];
-} repcodes_t;
-
-MEM_STATIC repcodes_t ZSTD_updateRep(U32 const rep[3], U32 const offset, U32 const ll0)
-{
-    repcodes_t newReps;
-    if (offset >= ZSTD_REP_NUM) {  /* full offset */
-        newReps.rep[2] = rep[1];
-        newReps.rep[1] = rep[0];
-        newReps.rep[0] = offset - ZSTD_REP_MOVE;
-    } else {   /* repcode */
-        U32 const repCode = offset + ll0;
-        if (repCode > 0) {  /* note : if repCode==0, no change */
-            U32 const currentOffset = (repCode==ZSTD_REP_NUM) ? (rep[0] - 1) : rep[repCode];
-            newReps.rep[2] = (repCode >= 2) ? rep[1] : rep[2];
-            newReps.rep[1] = rep[0];
-            newReps.rep[0] = currentOffset;
-        } else {   /* repCode == 0 */
-            memcpy(&newReps, rep, sizeof(newReps));
-        }
-    }
-    return newReps;
-}
-
 /* ZSTD_cParam_withinBounds:
  * @return 1 if value is within cParam bounds,
  * 0 otherwise */
@@ -366,17 +565,20 @@ MEM_STATIC int ZSTD_cParam_withinBounds(ZSTD_cParameter cParam, int value)
 /* ZSTD_noCompressBlock() :
  * Writes uncompressed block to dst buffer from given src.
  * Returns the size of the block */
-MEM_STATIC size_t ZSTD_noCompressBlock (void* dst, size_t dstCapacity, const void* src, size_t srcSize, U32 lastBlock)
+MEM_STATIC size_t
+ZSTD_noCompressBlock(void* dst, size_t dstCapacity, const void* src, size_t srcSize, U32 lastBlock)
 {
     U32 const cBlockHeader24 = lastBlock + (((U32)bt_raw)<<1) + (U32)(srcSize << 3);
+    DEBUGLOG(5, "ZSTD_noCompressBlock (srcSize=%zu, dstCapacity=%zu)", srcSize, dstCapacity);
     RETURN_ERROR_IF(srcSize + ZSTD_blockHeaderSize > dstCapacity,
                     dstSize_tooSmall, "dst buf too small for uncompressed block");
     MEM_writeLE24(dst, cBlockHeader24);
-    memcpy((BYTE*)dst + ZSTD_blockHeaderSize, src, srcSize);
+    ZSTD_memcpy((BYTE*)dst + ZSTD_blockHeaderSize, src, srcSize);
     return ZSTD_blockHeaderSize + srcSize;
 }
 
-MEM_STATIC size_t ZSTD_rleCompressBlock (void* dst, size_t dstCapacity, BYTE src, size_t srcSize, U32 lastBlock)
+MEM_STATIC size_t
+ZSTD_rleCompressBlock(void* dst, size_t dstCapacity, BYTE src, size_t srcSize, U32 lastBlock)
 {
     BYTE* const op = (BYTE*)dst;
     U32 const cBlockHeader = lastBlock + (((U32)bt_rle)<<1) + (U32)(srcSize << 3);
@@ -395,21 +597,21 @@ MEM_STATIC size_t ZSTD_minGain(size_t srcSize, ZSTD_strategy strat)
 {
     U32 const minlog = (strat>=ZSTD_btultra) ? (U32)(strat) - 1 : 6;
     ZSTD_STATIC_ASSERT(ZSTD_btultra == 8);
-    assert(ZSTD_cParam_withinBounds(ZSTD_c_strategy, strat));
+    assert(ZSTD_cParam_withinBounds(ZSTD_c_strategy, (int)strat));
     return (srcSize >> minlog) + 2;
 }
 
-MEM_STATIC int ZSTD_disableLiteralsCompression(const ZSTD_CCtx_params* cctxParams)
+MEM_STATIC int ZSTD_literalsCompressionIsDisabled(const ZSTD_CCtx_params* cctxParams)
 {
     switch (cctxParams->literalCompressionMode) {
-    case ZSTD_lcm_huffman:
+    case ZSTD_ps_enable:
         return 0;
-    case ZSTD_lcm_uncompressed:
+    case ZSTD_ps_disable:
         return 1;
     default:
         assert(0 /* impossible: pre-validated */);
-        /* fall-through */
-    case ZSTD_lcm_auto:
+        ZSTD_FALLTHROUGH;
+    case ZSTD_ps_auto:
         return (cctxParams->cParams.strategy == ZSTD_fast) && (cctxParams->cParams.targetLength > 0);
     }
 }
@@ -419,7 +621,9 @@ MEM_STATIC int ZSTD_disableLiteralsCompression(const ZSTD_CCtx_params* cctxParam
  *  Only called when the sequence ends past ilimit_w, so it only needs to be optimized for single
  *  large copies.
  */
-static void ZSTD_safecopyLiterals(BYTE* op, BYTE const* ip, BYTE const* const iend, BYTE const* ilimit_w) {
+static void
+ZSTD_safecopyLiterals(BYTE* op, BYTE const* ip, BYTE const* const iend, BYTE const* ilimit_w)
+{
     assert(iend > ilimit_w);
     if (ip <= ilimit_w) {
         ZSTD_wildcopy(op, ip, ilimit_w - ip, ZSTD_no_overlap);
@@ -429,14 +633,28 @@ static void ZSTD_safecopyLiterals(BYTE* op, BYTE const* ip, BYTE const* const ie
     while (ip < iend) *op++ = *ip++;
 }
 
+
+#define REPCODE1_TO_OFFBASE REPCODE_TO_OFFBASE(1)
+#define REPCODE2_TO_OFFBASE REPCODE_TO_OFFBASE(2)
+#define REPCODE3_TO_OFFBASE REPCODE_TO_OFFBASE(3)
+#define REPCODE_TO_OFFBASE(r) (assert((r)>=1), assert((r)<=ZSTD_REP_NUM), (r)) /* accepts IDs 1,2,3 */
+#define OFFSET_TO_OFFBASE(o)  (assert((o)>0), o + ZSTD_REP_NUM)
+#define OFFBASE_IS_OFFSET(o)  ((o) > ZSTD_REP_NUM)
+#define OFFBASE_IS_REPCODE(o) ( 1 <= (o) && (o) <= ZSTD_REP_NUM)
+#define OFFBASE_TO_OFFSET(o)  (assert(OFFBASE_IS_OFFSET(o)), (o) - ZSTD_REP_NUM)
+#define OFFBASE_TO_REPCODE(o) (assert(OFFBASE_IS_REPCODE(o)), (o))  /* returns ID 1,2,3 */
+
 /*! ZSTD_storeSeq() :
- *  Store a sequence (litlen, litPtr, offCode and mlBase) into seqStore_t.
- *  `offCode` : distance to match + ZSTD_REP_MOVE (values <= ZSTD_REP_MOVE are repCodes).
- *  `mlBase` : matchLength - MINMATCH
- *  Allowed to overread literals up to litLimit.
+ *  Store a sequence (litlen, litPtr, offBase and matchLength) into seqStore_t.
+ *  @offBase : Users should employ macros REPCODE_TO_OFFBASE() and OFFSET_TO_OFFBASE().
+ *  @matchLength : must be >= MINMATCH
+ *  Allowed to over-read literals up to litLimit.
 */
-HINT_INLINE UNUSED_ATTR
-void ZSTD_storeSeq(seqStore_t* seqStorePtr, size_t litLength, const BYTE* literals, const BYTE* litLimit, U32 offCode, size_t mlBase)
+HINT_INLINE UNUSED_ATTR void
+ZSTD_storeSeq(seqStore_t* seqStorePtr,
+              size_t litLength, const BYTE* literals, const BYTE* litLimit,
+              U32 offBase,
+              size_t matchLength)
 {
     BYTE const* const litLimit_w = litLimit - WILDCOPY_OVERLENGTH;
     BYTE const* const litEnd = literals + litLength;
@@ -444,8 +662,8 @@ void ZSTD_storeSeq(seqStore_t* seqStorePtr, size_t litLength, const BYTE* litera
     static const BYTE* g_start = NULL;
     if (g_start==NULL) g_start = (const BYTE*)literals;  /* note : index only works for compression within a single segment */
     {   U32 const pos = (U32)((const BYTE*)literals - g_start);
-        DEBUGLOG(6, "Cpos%7u :%3u literals, match%4u bytes at offCode%7u",
-               pos, (U32)litLength, (U32)mlBase+MINMATCH, (U32)offCode);
+        DEBUGLOG(6, "Cpos%7u :%3u literals, match%4u bytes at offBase%7u",
+               pos, (U32)litLength, (U32)matchLength, (U32)offBase);
     }
 #endif
     assert((size_t)(seqStorePtr->sequences - seqStorePtr->sequencesStart) < seqStorePtr->maxNbSeq);
@@ -455,9 +673,9 @@ void ZSTD_storeSeq(seqStore_t* seqStorePtr, size_t litLength, const BYTE* litera
     assert(literals + litLength <= litLimit);
     if (litEnd <= litLimit_w) {
         /* Common case we can use wildcopy.
-	 * First copy 16 bytes, because literals are likely short.
-	 */
-        assert(WILDCOPY_OVERLENGTH >= 16);
+         * First copy 16 bytes, because literals are likely short.
+         */
+        ZSTD_STATIC_ASSERT(WILDCOPY_OVERLENGTH >= 16);
         ZSTD_copy16(seqStorePtr->lit, literals);
         if (litLength > 16) {
             ZSTD_wildcopy(seqStorePtr->lit+16, literals+16, (ptrdiff_t)litLength-16, ZSTD_no_overlap);
@@ -469,95 +687,70 @@ 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;
 
     /* match offset */
-    seqStorePtr->sequences[0].offset = offCode + 1;
+    seqStorePtr->sequences[0].offBase = offBase;
 
     /* match Length */
-    if (mlBase>0xFFFF) {
-        assert(seqStorePtr->longLengthID == 0); /* there can only be a single long length */
-        seqStorePtr->longLengthID = 2;
-        seqStorePtr->longLengthPos = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
+    assert(matchLength >= MINMATCH);
+    {   size_t const mlBase = matchLength - MINMATCH;
+        if (mlBase>0xFFFF) {
+            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].mlBase = (U16)mlBase;
     }
-    seqStorePtr->sequences[0].matchLength = (U16)mlBase;
 
     seqStorePtr->sequences++;
 }
 
-
-/*-*************************************
-*  Match length counter
-***************************************/
-static unsigned ZSTD_NbCommonBytes (size_t val)
+/* ZSTD_updateRep() :
+ * updates in-place @rep (array of repeat offsets)
+ * @offBase : sum-type, using numeric representation of ZSTD_storeSeq()
+ */
+MEM_STATIC void
+ZSTD_updateRep(U32 rep[ZSTD_REP_NUM], U32 const offBase, U32 const ll0)
 {
-    if (MEM_isLittleEndian()) {
-        if (MEM_64bits()) {
-#       if defined(_MSC_VER) && defined(_WIN64)
-            unsigned long r = 0;
-            return _BitScanForward64( &r, (U64)val ) ? (unsigned)(r >> 3) : 0;
-#       elif defined(__GNUC__) && (__GNUC__ >= 4)
-            return (__builtin_ctzll((U64)val) >> 3);
-#       else
-            static const int DeBruijnBytePos[64] = { 0, 0, 0, 0, 0, 1, 1, 2,
-                                                     0, 3, 1, 3, 1, 4, 2, 7,
-                                                     0, 2, 3, 6, 1, 5, 3, 5,
-                                                     1, 3, 4, 4, 2, 5, 6, 7,
-                                                     7, 0, 1, 2, 3, 3, 4, 6,
-                                                     2, 6, 5, 5, 3, 4, 5, 6,
-                                                     7, 1, 2, 4, 6, 4, 4, 5,
-                                                     7, 2, 6, 5, 7, 6, 7, 7 };
-            return DeBruijnBytePos[((U64)((val & -(long long)val) * 0x0218A392CDABBD3FULL)) >> 58];
-#       endif
-        } else { /* 32 bits */
-#       if defined(_MSC_VER)
-            unsigned long r=0;
-            return _BitScanForward( &r, (U32)val ) ? (unsigned)(r >> 3) : 0;
-#       elif defined(__GNUC__) && (__GNUC__ >= 3)
-            return (__builtin_ctz((U32)val) >> 3);
-#       else
-            static const int DeBruijnBytePos[32] = { 0, 0, 3, 0, 3, 1, 3, 0,
-                                                     3, 2, 2, 1, 3, 2, 0, 1,
-                                                     3, 3, 1, 2, 2, 2, 2, 0,
-                                                     3, 1, 2, 0, 1, 0, 1, 1 };
-            return DeBruijnBytePos[((U32)((val & -(S32)val) * 0x077CB531U)) >> 27];
-#       endif
+    if (OFFBASE_IS_OFFSET(offBase)) {  /* full offset */
+        rep[2] = rep[1];
+        rep[1] = rep[0];
+        rep[0] = OFFBASE_TO_OFFSET(offBase);
+    } else {   /* repcode */
+        U32 const repCode = OFFBASE_TO_REPCODE(offBase) - 1 + ll0;
+        if (repCode > 0) {  /* note : if repCode==0, no change */
+            U32 const currentOffset = (repCode==ZSTD_REP_NUM) ? (rep[0] - 1) : rep[repCode];
+            rep[2] = (repCode >= 2) ? rep[1] : rep[2];
+            rep[1] = rep[0];
+            rep[0] = currentOffset;
+        } else {   /* repCode == 0 */
+            /* nothing to do */
         }
-    } else {  /* Big Endian CPU */
-        if (MEM_64bits()) {
-#       if defined(_MSC_VER) && defined(_WIN64)
-            unsigned long r = 0;
-            return _BitScanReverse64( &r, val ) ? (unsigned)(r >> 3) : 0;
-#       elif defined(__GNUC__) && (__GNUC__ >= 4)
-            return (__builtin_clzll(val) >> 3);
-#       else
-            unsigned r;
-            const unsigned n32 = sizeof(size_t)*4;   /* calculate this way due to compiler complaining in 32-bits mode */
-            if (!(val>>n32)) { r=4; } else { r=0; val>>=n32; }
-            if (!(val>>16)) { r+=2; val>>=8; } else { val>>=24; }
-            r += (!val);
-            return r;
-#       endif
-        } else { /* 32 bits */
-#       if defined(_MSC_VER)
-            unsigned long r = 0;
-            return _BitScanReverse( &r, (unsigned long)val ) ? (unsigned)(r >> 3) : 0;
-#       elif defined(__GNUC__) && (__GNUC__ >= 3)
-            return (__builtin_clz((U32)val) >> 3);
-#       else
-            unsigned r;
-            if (!(val>>16)) { r=2; val>>=8; } else { r=0; val>>=24; }
-            r += (!val);
-            return r;
-#       endif
-    }   }
+    }
+}
+
+typedef struct repcodes_s {
+    U32 rep[3];
+} repcodes_t;
+
+MEM_STATIC repcodes_t
+ZSTD_newRep(U32 const rep[ZSTD_REP_NUM], U32 const offBase, U32 const ll0)
+{
+    repcodes_t newReps;
+    ZSTD_memcpy(&newReps, rep, sizeof(newReps));
+    ZSTD_updateRep(newReps.rep, offBase, ll0);
+    return newReps;
 }
 
 
+/*-*************************************
+*  Match length counter
+***************************************/
 MEM_STATIC size_t ZSTD_count(const BYTE* pIn, const BYTE* pMatch, const BYTE* const pInLimit)
 {
     const BYTE* const pStart = pIn;
@@ -603,31 +796,43 @@ ZSTD_count_2segments(const BYTE* ip, const BYTE* match,
  *  Hashes
  ***************************************/
 static const U32 prime3bytes = 506832829U;
-static U32    ZSTD_hash3(U32 u, U32 h) { return ((u << (32-24)) * prime3bytes)  >> (32-h) ; }
-MEM_STATIC size_t ZSTD_hash3Ptr(const void* ptr, U32 h) { return ZSTD_hash3(MEM_readLE32(ptr), h); } /* only in zstd_opt.h */
+static U32    ZSTD_hash3(U32 u, U32 h, U32 s) { assert(h <= 32); return (((u << (32-24)) * prime3bytes) ^ s)  >> (32-h) ; }
+MEM_STATIC size_t ZSTD_hash3Ptr(const void* ptr, U32 h) { return ZSTD_hash3(MEM_readLE32(ptr), h, 0); } /* only in zstd_opt.h */
+MEM_STATIC size_t ZSTD_hash3PtrS(const void* ptr, U32 h, U32 s) { return ZSTD_hash3(MEM_readLE32(ptr), h, s); }
 
 static const U32 prime4bytes = 2654435761U;
-static U32    ZSTD_hash4(U32 u, U32 h) { return (u * prime4bytes) >> (32-h) ; }
-static size_t ZSTD_hash4Ptr(const void* ptr, U32 h) { return ZSTD_hash4(MEM_read32(ptr), h); }
+static U32    ZSTD_hash4(U32 u, U32 h, U32 s) { assert(h <= 32); return ((u * prime4bytes) ^ s) >> (32-h) ; }
+static size_t ZSTD_hash4Ptr(const void* ptr, U32 h) { return ZSTD_hash4(MEM_readLE32(ptr), h, 0); }
+static size_t ZSTD_hash4PtrS(const void* ptr, U32 h, U32 s) { return ZSTD_hash4(MEM_readLE32(ptr), h, s); }
 
 static const U64 prime5bytes = 889523592379ULL;
-static size_t ZSTD_hash5(U64 u, U32 h) { return (size_t)(((u  << (64-40)) * prime5bytes) >> (64-h)) ; }
-static size_t ZSTD_hash5Ptr(const void* p, U32 h) { return ZSTD_hash5(MEM_readLE64(p), h); }
+static size_t ZSTD_hash5(U64 u, U32 h, U64 s) { assert(h <= 64); return (size_t)((((u  << (64-40)) * prime5bytes) ^ s) >> (64-h)) ; }
+static size_t ZSTD_hash5Ptr(const void* p, U32 h) { return ZSTD_hash5(MEM_readLE64(p), h, 0); }
+static size_t ZSTD_hash5PtrS(const void* p, U32 h, U64 s) { return ZSTD_hash5(MEM_readLE64(p), h, s); }
 
 static const U64 prime6bytes = 227718039650203ULL;
-static size_t ZSTD_hash6(U64 u, U32 h) { return (size_t)(((u  << (64-48)) * prime6bytes) >> (64-h)) ; }
-static size_t ZSTD_hash6Ptr(const void* p, U32 h) { return ZSTD_hash6(MEM_readLE64(p), h); }
+static size_t ZSTD_hash6(U64 u, U32 h, U64 s) { assert(h <= 64); return (size_t)((((u  << (64-48)) * prime6bytes) ^ s) >> (64-h)) ; }
+static size_t ZSTD_hash6Ptr(const void* p, U32 h) { return ZSTD_hash6(MEM_readLE64(p), h, 0); }
+static size_t ZSTD_hash6PtrS(const void* p, U32 h, U64 s) { return ZSTD_hash6(MEM_readLE64(p), h, s); }
 
 static const U64 prime7bytes = 58295818150454627ULL;
-static size_t ZSTD_hash7(U64 u, U32 h) { return (size_t)(((u  << (64-56)) * prime7bytes) >> (64-h)) ; }
-static size_t ZSTD_hash7Ptr(const void* p, U32 h) { return ZSTD_hash7(MEM_readLE64(p), h); }
+static size_t ZSTD_hash7(U64 u, U32 h, U64 s) { assert(h <= 64); return (size_t)((((u  << (64-56)) * prime7bytes) ^ s) >> (64-h)) ; }
+static size_t ZSTD_hash7Ptr(const void* p, U32 h) { return ZSTD_hash7(MEM_readLE64(p), h, 0); }
+static size_t ZSTD_hash7PtrS(const void* p, U32 h, U64 s) { return ZSTD_hash7(MEM_readLE64(p), h, s); }
 
 static const U64 prime8bytes = 0xCF1BBCDCB7A56463ULL;
-static size_t ZSTD_hash8(U64 u, U32 h) { return (size_t)(((u) * prime8bytes) >> (64-h)) ; }
-static size_t ZSTD_hash8Ptr(const void* p, U32 h) { return ZSTD_hash8(MEM_readLE64(p), h); }
+static size_t ZSTD_hash8(U64 u, U32 h, U64 s) { assert(h <= 64); return (size_t)((((u) * prime8bytes)  ^ s) >> (64-h)) ; }
+static size_t ZSTD_hash8Ptr(const void* p, U32 h) { return ZSTD_hash8(MEM_readLE64(p), h, 0); }
+static size_t ZSTD_hash8PtrS(const void* p, U32 h, U64 s) { return ZSTD_hash8(MEM_readLE64(p), h, s); }
+
 
-MEM_STATIC size_t ZSTD_hashPtr(const void* p, U32 hBits, U32 mls)
+MEM_STATIC FORCE_INLINE_ATTR
+size_t ZSTD_hashPtr(const void* p, U32 hBits, U32 mls)
 {
+    /* Although some of these hashes do support hBits up to 64, some do not.
+     * To be on the safe side, always avoid hBits > 32. */
+    assert(hBits <= 32);
+
     switch(mls)
     {
     default:
@@ -639,6 +844,24 @@ MEM_STATIC size_t ZSTD_hashPtr(const void* p, U32 hBits, U32 mls)
     }
 }
 
+MEM_STATIC FORCE_INLINE_ATTR
+size_t ZSTD_hashPtrSalted(const void* p, U32 hBits, U32 mls, const U64 hashSalt) {
+    /* Although some of these hashes do support hBits up to 64, some do not.
+     * To be on the safe side, always avoid hBits > 32. */
+    assert(hBits <= 32);
+
+    switch(mls)
+    {
+        default:
+        case 4: return ZSTD_hash4PtrS(p, hBits, (U32)hashSalt);
+        case 5: return ZSTD_hash5PtrS(p, hBits, hashSalt);
+        case 6: return ZSTD_hash6PtrS(p, hBits, hashSalt);
+        case 7: return ZSTD_hash7PtrS(p, hBits, hashSalt);
+        case 8: return ZSTD_hash8PtrS(p, hBits, hashSalt);
+    }
+}
+
+
 /** ZSTD_ipow() :
  * Return base^exponent.
  */
@@ -723,6 +946,13 @@ 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.
@@ -742,20 +972,76 @@ MEM_STATIC ZSTD_dictMode_e ZSTD_matchState_dictMode(const ZSTD_matchState_t *ms)
     return ZSTD_window_hasExtDict(ms->window) ?
         ZSTD_extDict :
         ms->dictMatchState != NULL ?
-            ZSTD_dictMatchState :
+            (ms->dictMatchState->dedicatedDictSearch ? ZSTD_dedicatedDictSearch : ZSTD_dictMatchState) :
             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)
+                                        + ZSTD_WINDOW_START_INDEX;
+
+    /* Adjust the min index to backoff the overflow correction frequency,
+     * so we don't waste too much CPU in overflow correction. If this
+     * 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 current = (U32)((BYTE const*)srcEnd - window.base);
-    return current > ZSTD_CURRENT_MAX;
+    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;
 }
 
 /**
@@ -766,7 +1052,6 @@ 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)
@@ -790,32 +1075,52 @@ 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 current = (U32)((BYTE const*)src - window->base);
-    U32 const currentCycle0 = current & cycleMask;
-    /* Exclude zero so that newCurrent - maxDist >= 1. */
-    U32 const currentCycle1 = currentCycle0 == 0 ? (1U << cycleLog) : currentCycle0;
-    U32 const newCurrent = currentCycle1 + maxDist;
-    U32 const correction = current - newCurrent;
-    assert((maxDist & cycleMask) == 0);
-    assert(current > newCurrent);
-    /* Loose bound, should be around 1<<29 (see above) */
-    assert(correction > 1<<28);
+    U32 const cycleSize = 1u << cycleLog;
+    U32 const cycleMask = cycleSize - 1;
+    U32 const curr = (U32)((BYTE const*)src - window->base);
+    U32 const currentCycle = curr & cycleMask;
+    /* Ensure newCurrent - maxDist >= ZSTD_WINDOW_START_INDEX. */
+    U32 const currentCycleCorrection = currentCycle < ZSTD_WINDOW_START_INDEX
+                                     ? MAX(cycleSize, ZSTD_WINDOW_START_INDEX)
+                                     : 0;
+    U32 const newCurrent = currentCycle
+                         + currentCycleCorrection
+                         + MAX(maxDist, cycleSize);
+    U32 const correction = curr - newCurrent;
+    /* maxDist must be a power of two so that:
+     *   (newCurrent & cycleMask) == (curr & cycleMask)
+     * This is required to not corrupt the chains / binary tree.
+     */
+    assert((maxDist & (maxDist - 1)) == 0);
+    assert((curr & cycleMask) == (newCurrent & cycleMask));
+    assert(curr > newCurrent);
+    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;
-    if (window->lowLimit <= correction) window->lowLimit = 1;
-    else window->lowLimit -= correction;
-    if (window->dictLimit <= correction) window->dictLimit = 1;
-    else window->dictLimit -= correction;
+    if (window->lowLimit < correction + ZSTD_WINDOW_START_INDEX) {
+        window->lowLimit = ZSTD_WINDOW_START_INDEX;
+    } else {
+        window->lowLimit -= correction;
+    }
+    if (window->dictLimit < correction + ZSTD_WINDOW_START_INDEX) {
+        window->dictLimit = ZSTD_WINDOW_START_INDEX;
+    } else {
+        window->dictLimit -= correction;
+    }
 
     /* Ensure we can still reference the full window. */
     assert(newCurrent >= maxDist);
-    assert(newCurrent - maxDist >= 1);
+    assert(newCurrent - maxDist >= ZSTD_WINDOW_START_INDEX);
     /* Ensure that lowLimit and dictLimit didn't underflow. */
     assert(window->lowLimit <= newCurrent);
     assert(window->dictLimit <= newCurrent);
 
+    ++window->nbOverflowCorrections;
+
     DEBUGLOG(4, "Correction of 0x%x bytes to lowLimit=0x%x", correction,
              window->lowLimit);
     return correction;
@@ -904,10 +1209,15 @@ ZSTD_checkDictValidity(const ZSTD_window_t* window,
                     (unsigned)blockEndIdx, (unsigned)maxDist, (unsigned)loadedDictEnd);
         assert(blockEndIdx >= loadedDictEnd);
 
-        if (blockEndIdx > loadedDictEnd + maxDist) {
+        if (blockEndIdx > loadedDictEnd + maxDist || loadedDictEnd != window->dictLimit) {
             /* On reaching window size, dictionaries are invalidated.
              * For simplification, if window size is reached anywhere within next block,
              * the dictionary is invalidated for the full block.
+             *
+             * We also have to invalidate the dictionary if ZSTD_window_update() has detected
+             * non-contiguous segments, which means that loadedDictEnd != window->dictLimit.
+             * loadedDictEnd may be 0, if forceWindow is true, but in that case we never use
+             * dictMatchState, so setting it to NULL is not a problem.
              */
             DEBUGLOG(6, "invalidating dictionary for current block (distance > windowSize)");
             *loadedDictEndPtr = 0;
@@ -919,12 +1229,14 @@ ZSTD_checkDictValidity(const ZSTD_window_t* window,
 }
 
 MEM_STATIC void ZSTD_window_init(ZSTD_window_t* window) {
-    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 */
+    ZSTD_memset(window, 0, sizeof(*window));
+    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;
 }
 
 /**
@@ -935,7 +1247,8 @@ 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;
@@ -945,7 +1258,7 @@ 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);
@@ -973,25 +1286,32 @@ MEM_STATIC U32 ZSTD_window_update(ZSTD_window_t* window,
 /**
  * Returns the lowest allowed match index. It may either be in the ext-dict or the prefix.
  */
-MEM_STATIC U32 ZSTD_getLowestMatchIndex(const ZSTD_matchState_t* ms, U32 current, unsigned windowLog)
+MEM_STATIC U32 ZSTD_getLowestMatchIndex(const ZSTD_matchState_t* ms, U32 curr, unsigned windowLog)
 {
-    U32    const maxDistance = 1U << windowLog;
-    U32    const lowestValid = ms->window.lowLimit;
-    U32    const withinWindow = (current - lowestValid > maxDistance) ? current - maxDistance : lowestValid;
-    U32    const isDictionary = (ms->loadedDictEnd != 0);
-    U32    const matchLowest = isDictionary ? lowestValid : withinWindow;
+    U32 const maxDistance = 1U << windowLog;
+    U32 const lowestValid = ms->window.lowLimit;
+    U32 const withinWindow = (curr - lowestValid > maxDistance) ? curr - maxDistance : lowestValid;
+    U32 const isDictionary = (ms->loadedDictEnd != 0);
+    /* When using a dictionary the entire dictionary is valid if a single byte of the dictionary
+     * is within the window. We invalidate the dictionary (and set loadedDictEnd to 0) when it isn't
+     * valid for the entire block. So this check is sufficient to find the lowest valid match index.
+     */
+    U32 const matchLowest = isDictionary ? lowestValid : withinWindow;
     return matchLowest;
 }
 
 /**
  * Returns the lowest allowed match index in the prefix.
  */
-MEM_STATIC U32 ZSTD_getLowestPrefixIndex(const ZSTD_matchState_t* ms, U32 current, unsigned windowLog)
+MEM_STATIC U32 ZSTD_getLowestPrefixIndex(const ZSTD_matchState_t* ms, U32 curr, unsigned windowLog)
 {
     U32    const maxDistance = 1U << windowLog;
     U32    const lowestValid = ms->window.dictLimit;
-    U32    const withinWindow = (current - lowestValid > maxDistance) ? current - maxDistance : lowestValid;
+    U32    const withinWindow = (curr - lowestValid > maxDistance) ? curr - maxDistance : lowestValid;
     U32    const isDictionary = (ms->loadedDictEnd != 0);
+    /* When computing the lowest prefix index we need to take the dictionary into account to handle
+     * the edge case where the dictionary and the source are contiguous in memory.
+     */
     U32    const matchLowest = isDictionary ? lowestValid : withinWindow;
     return matchLowest;
 }
@@ -1029,6 +1349,42 @@ MEM_STATIC void ZSTD_debugTable(const U32* table, U32 max)
 
 #endif
 
+/* Short Cache */
+
+/* Normally, zstd matchfinders follow this flow:
+ *     1. Compute hash at ip
+ *     2. Load index from hashTable[hash]
+ *     3. Check if *ip == *(base + index)
+ * In dictionary compression, loading *(base + index) is often an L2 or even L3 miss.
+ *
+ * Short cache is an optimization which allows us to avoid step 3 most of the time
+ * when the data doesn't actually match. With short cache, the flow becomes:
+ *     1. Compute (hash, currentTag) at ip. currentTag is an 8-bit independent hash at ip.
+ *     2. Load (index, matchTag) from hashTable[hash]. See ZSTD_writeTaggedIndex to understand how this works.
+ *     3. Only if currentTag == matchTag, check *ip == *(base + index). Otherwise, continue.
+ *
+ * Currently, short cache is only implemented in CDict hashtables. Thus, its use is limited to
+ * dictMatchState matchfinders.
+ */
+#define ZSTD_SHORT_CACHE_TAG_BITS 8
+#define ZSTD_SHORT_CACHE_TAG_MASK ((1u << ZSTD_SHORT_CACHE_TAG_BITS) - 1)
+
+/* Helper function for ZSTD_fillHashTable and ZSTD_fillDoubleHashTable.
+ * Unpacks hashAndTag into (hash, tag), then packs (index, tag) into hashTable[hash]. */
+MEM_STATIC void ZSTD_writeTaggedIndex(U32* const hashTable, size_t hashAndTag, U32 index) {
+    size_t const hash = hashAndTag >> ZSTD_SHORT_CACHE_TAG_BITS;
+    U32 const tag = (U32)(hashAndTag & ZSTD_SHORT_CACHE_TAG_MASK);
+    assert(index >> (32 - ZSTD_SHORT_CACHE_TAG_BITS) == 0);
+    hashTable[hash] = (index << ZSTD_SHORT_CACHE_TAG_BITS) | tag;
+}
+
+/* Helper function for short cache matchfinders.
+ * Unpacks tag1 and tag2 from lower bits of packedTag1 and packedTag2, then checks if the tags match. */
+MEM_STATIC int ZSTD_comparePackedTags(size_t packedTag1, size_t packedTag2) {
+    U32 const tag1 = packedTag1 & ZSTD_SHORT_CACHE_TAG_MASK;
+    U32 const tag2 = packedTag2 & ZSTD_SHORT_CACHE_TAG_MASK;
+    return tag1 == tag2;
+}
 
 #if defined (__cplusplus)
 }
@@ -1045,7 +1401,6 @@ MEM_STATIC void ZSTD_debugTable(const U32* table, U32 max)
  * assumptions : magic number supposed already checked
  *               and dictSize >= 8 */
 size_t ZSTD_loadCEntropy(ZSTD_compressedBlockState_t* bs, void* workspace,
-                         short* offcodeNCount, unsigned* offcodeMaxValue,
                          const void* const dict, size_t dictSize);
 
 void ZSTD_reset_compressedBlockState(ZSTD_compressedBlockState_t* bs);
@@ -1061,7 +1416,7 @@ void ZSTD_reset_compressedBlockState(ZSTD_compressedBlockState_t* bs);
  * Note: srcSizeHint == 0 means 0!
  */
 ZSTD_compressionParameters ZSTD_getCParamsFromCCtxParams(
-        const ZSTD_CCtx_params* CCtxParams, U64 srcSizeHint, size_t dictSize);
+        const ZSTD_CCtx_params* CCtxParams, U64 srcSizeHint, size_t dictSize, ZSTD_cParamMode_e mode);
 
 /*! ZSTD_initCStream_internal() :
  *  Private use only. Init streaming operation.
@@ -1122,4 +1477,56 @@ 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);
+
+/* Returns 0 on success, and a ZSTD_error otherwise. This function scans through an array of
+ * ZSTD_Sequence, storing the sequences it finds, until it reaches a block delimiter.
+ * Note that the block delimiter must include the last literals of the block.
+ */
+size_t
+ZSTD_copySequencesToSeqStoreExplicitBlockDelim(ZSTD_CCtx* cctx,
+                                              ZSTD_sequencePosition* seqPos,
+                                        const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
+                                        const void* src, size_t blockSize, ZSTD_paramSwitch_e externalRepSearch);
+
+/* Returns the number of bytes to move the current read position back by.
+ * Only non-zero if we ended up splitting a sequence.
+ * Otherwise, it may return a ZSTD error if something went wrong.
+ *
+ * This function will attempt to scan through blockSize bytes
+ * represented by the sequences in @inSeqs,
+ * storing any (partial) sequences.
+ *
+ * Occasionally, we may want to change the actual number of bytes we consumed from inSeqs to
+ * avoid splitting a match, or to avoid splitting a match such that it would produce a match
+ * smaller than MINMATCH. In this case, we return the number of bytes that we didn't read from this block.
+ */
+size_t
+ZSTD_copySequencesToSeqStoreNoBlockDelim(ZSTD_CCtx* cctx, ZSTD_sequencePosition* seqPos,
+                                   const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
+                                   const void* src, size_t blockSize, ZSTD_paramSwitch_e externalRepSearch);
+
+
+/* ===============================================================
+ * Deprecated definitions that are still used internally to avoid
+ * deprecation warnings. These functions are exactly equivalent to
+ * their public variants, but avoid the deprecation warnings.
+ * =============================================================== */
+
+size_t ZSTD_compressBegin_usingCDict_deprecated(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict);
+
+size_t ZSTD_compressContinue_public(ZSTD_CCtx* cctx,
+                                    void* dst, size_t dstCapacity,
+                              const void* src, size_t srcSize);
+
+size_t ZSTD_compressEnd_public(ZSTD_CCtx* cctx,
+                               void* dst, size_t dstCapacity,
+                         const void* src, size_t srcSize);
+
+size_t ZSTD_compressBlock_deprecated(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);
+
+
 #endif /* ZSTD_COMPRESS_H */