extzstd 0.3.2 → 0.4

data/contrib/zstd/lib/compress/zstd_lazy.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -10,14 +10,23 @@
 
 #include "zstd_compress_internal.h"
 #include "zstd_lazy.h"
+#include "../common/bits.h" /* ZSTD_countTrailingZeros64 */
+
+#if !defined(ZSTD_EXCLUDE_GREEDY_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_LAZY_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_LAZY2_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR)
+
+#define kLazySkippingStep 8
 
 
 /*-*************************************
 *  Binary Tree search
 ***************************************/
 
-static void
-ZSTD_updateDUBT(ZSTD_matchState_t* ms,
+static
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+void ZSTD_updateDUBT(ZSTD_matchState_t* ms,
                 const BYTE* ip, const BYTE* iend,
                 U32 mls)
 {
@@ -60,8 +69,9 @@ ZSTD_updateDUBT(ZSTD_matchState_t* ms,
  *  sort one already inserted but unsorted position
  *  assumption : curr >= btlow == (curr - btmask)
  *  doesn't fail */
-static void
-ZSTD_insertDUBT1(ZSTD_matchState_t* ms,
+static
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+void ZSTD_insertDUBT1(const ZSTD_matchState_t* ms,
                  U32 curr, const BYTE* inputEnd,
                  U32 nbCompares, U32 btLow,
                  const ZSTD_dictMode_e dictMode)
@@ -93,7 +103,7 @@ ZSTD_insertDUBT1(ZSTD_matchState_t* ms,
     assert(curr >= btLow);
     assert(ip < iend);   /* condition for ZSTD_count */
 
-    while (nbCompares-- && (matchIndex > windowLow)) {
+    for (; nbCompares && (matchIndex > windowLow); --nbCompares) {
         U32* const nextPtr = bt + 2*(matchIndex & btMask);
         size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
         assert(matchIndex < curr);
@@ -149,9 +159,10 @@ ZSTD_insertDUBT1(ZSTD_matchState_t* ms,
 }
 
 
-static size_t
-ZSTD_DUBT_findBetterDictMatch (
-        ZSTD_matchState_t* ms,
+static
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+size_t ZSTD_DUBT_findBetterDictMatch (
+        const ZSTD_matchState_t* ms,
         const BYTE* const ip, const BYTE* const iend,
         size_t* offsetPtr,
         size_t bestLength,
@@ -185,7 +196,7 @@ ZSTD_DUBT_findBetterDictMatch (
     (void)dictMode;
     assert(dictMode == ZSTD_dictMatchState);
 
-    while (nbCompares-- && (dictMatchIndex > dictLowLimit)) {
+    for (; nbCompares && (dictMatchIndex > dictLowLimit); --nbCompares) {
         U32* const nextPtr = dictBt + 2*(dictMatchIndex & btMask);
         size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
         const BYTE* match = dictBase + dictMatchIndex;
@@ -197,8 +208,8 @@ ZSTD_DUBT_findBetterDictMatch (
             U32 matchIndex = dictMatchIndex + dictIndexDelta;
             if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(curr-matchIndex+1) - ZSTD_highbit32((U32)offsetPtr[0]+1)) ) {
                 DEBUGLOG(9, "ZSTD_DUBT_findBetterDictMatch(%u) : found better match length %u -> %u and offsetCode %u -> %u (dictMatchIndex %u, matchIndex %u)",
-                    curr, (U32)bestLength, (U32)matchLength, (U32)*offsetPtr, ZSTD_REP_MOVE + curr - matchIndex, dictMatchIndex, matchIndex);
-                bestLength = matchLength, *offsetPtr = ZSTD_REP_MOVE + curr - matchIndex;
+                    curr, (U32)bestLength, (U32)matchLength, (U32)*offsetPtr, OFFSET_TO_OFFBASE(curr - matchIndex), dictMatchIndex, matchIndex);
+                bestLength = matchLength, *offsetPtr = OFFSET_TO_OFFBASE(curr - matchIndex);
             }
             if (ip+matchLength == iend) {   /* reached end of input : ip[matchLength] is not valid, no way to know if it's larger or smaller than match */
                 break;   /* drop, to guarantee consistency (miss a little bit of compression) */
@@ -218,7 +229,7 @@ ZSTD_DUBT_findBetterDictMatch (
     }
 
     if (bestLength >= MINMATCH) {
-        U32 const mIndex = curr - ((U32)*offsetPtr - ZSTD_REP_MOVE); (void)mIndex;
+        U32 const mIndex = curr - (U32)OFFBASE_TO_OFFSET(*offsetPtr); (void)mIndex;
         DEBUGLOG(8, "ZSTD_DUBT_findBetterDictMatch(%u) : found match of length %u and offsetCode %u (pos %u)",
                     curr, (U32)bestLength, (U32)*offsetPtr, mIndex);
     }
@@ -227,10 +238,11 @@ ZSTD_DUBT_findBetterDictMatch (
 }
 
 
-static size_t
-ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms,
+static
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+size_t ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms,
                         const BYTE* const ip, const BYTE* const iend,
-                        size_t* offsetPtr,
+                        size_t* offBasePtr,
                         U32 const mls,
                         const ZSTD_dictMode_e dictMode)
 {
@@ -309,7 +321,7 @@ ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms,
         matchIndex  = hashTable[h];
         hashTable[h] = curr;   /* Update Hash Table */
 
-        while (nbCompares-- && (matchIndex > windowLow)) {
+        for (; nbCompares && (matchIndex > windowLow); --nbCompares) {
             U32* const nextPtr = bt + 2*(matchIndex & btMask);
             size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
             const BYTE* match;
@@ -327,8 +339,8 @@ ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms,
             if (matchLength > bestLength) {
                 if (matchLength > matchEndIdx - matchIndex)
                     matchEndIdx = matchIndex + (U32)matchLength;
-                if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(curr-matchIndex+1) - ZSTD_highbit32((U32)offsetPtr[0]+1)) )
-                    bestLength = matchLength, *offsetPtr = ZSTD_REP_MOVE + curr - matchIndex;
+                if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(curr - matchIndex + 1) - ZSTD_highbit32((U32)*offBasePtr)) )
+                    bestLength = matchLength, *offBasePtr = OFFSET_TO_OFFBASE(curr - matchIndex);
                 if (ip+matchLength == iend) {   /* equal : no way to know if inf or sup */
                     if (dictMode == ZSTD_dictMatchState) {
                         nbCompares = 0; /* in addition to avoiding checking any
@@ -357,19 +369,20 @@ ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms,
 
         *smallerPtr = *largerPtr = 0;
 
+        assert(nbCompares <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */
         if (dictMode == ZSTD_dictMatchState && nbCompares) {
             bestLength = ZSTD_DUBT_findBetterDictMatch(
                     ms, ip, iend,
-                    offsetPtr, bestLength, nbCompares,
+                    offBasePtr, bestLength, nbCompares,
                     mls, dictMode);
         }
 
         assert(matchEndIdx > curr+8); /* ensure nextToUpdate is increased */
         ms->nextToUpdate = matchEndIdx - 8;   /* skip repetitive patterns */
         if (bestLength >= MINMATCH) {
-            U32 const mIndex = curr - ((U32)*offsetPtr - ZSTD_REP_MOVE); (void)mIndex;
+            U32 const mIndex = curr - (U32)OFFBASE_TO_OFFSET(*offBasePtr); (void)mIndex;
             DEBUGLOG(8, "ZSTD_DUBT_findBestMatch(%u) : found match of length %u and offsetCode %u (pos %u)",
-                        curr, (U32)bestLength, (U32)*offsetPtr, mIndex);
+                        curr, (U32)bestLength, (U32)*offBasePtr, mIndex);
         }
         return bestLength;
     }
@@ -377,104 +390,23 @@ ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms,
 
 
 /** ZSTD_BtFindBestMatch() : Tree updater, providing best match */
-FORCE_INLINE_TEMPLATE size_t
-ZSTD_BtFindBestMatch( ZSTD_matchState_t* ms,
+FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+size_t ZSTD_BtFindBestMatch( ZSTD_matchState_t* ms,
                 const BYTE* const ip, const BYTE* const iLimit,
-                      size_t* offsetPtr,
+                      size_t* offBasePtr,
                 const U32 mls /* template */,
                 const ZSTD_dictMode_e dictMode)
 {
     DEBUGLOG(7, "ZSTD_BtFindBestMatch");
     if (ip < ms->window.base + ms->nextToUpdate) return 0;   /* skipped area */
     ZSTD_updateDUBT(ms, ip, iLimit, mls);
-    return ZSTD_DUBT_findBestMatch(ms, ip, iLimit, offsetPtr, mls, dictMode);
-}
-
-
-static size_t
-ZSTD_BtFindBestMatch_selectMLS (  ZSTD_matchState_t* ms,
-                            const BYTE* ip, const BYTE* const iLimit,
-                                  size_t* offsetPtr)
-{
-    switch(ms->cParams.minMatch)
-    {
-    default : /* includes case 3 */
-    case 4 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 4, ZSTD_noDict);
-    case 5 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 5, ZSTD_noDict);
-    case 7 :
-    case 6 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 6, ZSTD_noDict);
-    }
+    return ZSTD_DUBT_findBestMatch(ms, ip, iLimit, offBasePtr, mls, dictMode);
 }
 
-
-static size_t ZSTD_BtFindBestMatch_dictMatchState_selectMLS (
-                        ZSTD_matchState_t* ms,
-                        const BYTE* ip, const BYTE* const iLimit,
-                        size_t* offsetPtr)
-{
-    switch(ms->cParams.minMatch)
-    {
-    default : /* includes case 3 */
-    case 4 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 4, ZSTD_dictMatchState);
-    case 5 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 5, ZSTD_dictMatchState);
-    case 7 :
-    case 6 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 6, ZSTD_dictMatchState);
-    }
-}
-
-
-static size_t ZSTD_BtFindBestMatch_extDict_selectMLS (
-                        ZSTD_matchState_t* ms,
-                        const BYTE* ip, const BYTE* const iLimit,
-                        size_t* offsetPtr)
-{
-    switch(ms->cParams.minMatch)
-    {
-    default : /* includes case 3 */
-    case 4 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 4, ZSTD_extDict);
-    case 5 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 5, ZSTD_extDict);
-    case 7 :
-    case 6 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 6, ZSTD_extDict);
-    }
-}
-
-
-
-/* *********************************
-*  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 +416,7 @@ 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;
@@ -498,13 +430,12 @@ void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t* ms, const B
     U32 const hashLog = ms->cParams.hashLog - ZSTD_LAZY_DDSS_BUCKET_LOG;
     U32* const tmpHashTable = hashTable;
     U32* const tmpChainTable = hashTable + ((size_t)1 << hashLog);
-    U32 const tmpChainSize = ((1 << ZSTD_LAZY_DDSS_BUCKET_LOG) - 1) << hashLog;
+    U32 const tmpChainSize = (U32)((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 +466,7 @@ 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
@@ -591,10 +522,149 @@ 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 = OFFSET_TO_OFFBASE(curr - (matchIndex + ddsIndexDelta));
+            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 = OFFSET_TO_OFFBASE(curr - (matchIndex + ddsIndexDelta));
+                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
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+U32 ZSTD_insertAndFindFirstIndex_internal(
+                        ZSTD_matchState_t* ms,
+                        const ZSTD_compressionParameters* const cParams,
+                        const BYTE* ip, U32 const mls, U32 const lazySkipping)
+{
+    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++;
+        /* Stop inserting every position when in the lazy skipping mode. */
+        if (lazySkipping)
+            break;
+    }
+
+    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, /* lazySkipping*/ 0);
+}
 
 /* inlining is important to hardwire a hot branch (template emulation) */
 FORCE_INLINE_TEMPLATE
-size_t ZSTD_HcFindBestMatch_generic (
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+size_t ZSTD_HcFindBestMatch(
                         ZSTD_matchState_t* ms,
                         const BYTE* const ip, const BYTE* const iLimit,
                         size_t* offsetPtr,
@@ -633,14 +703,15 @@ size_t ZSTD_HcFindBestMatch_generic (
     }
 
     /* HC4 match finder */
-    matchIndex = ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, mls);
+    matchIndex = ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, mls, ms->lazySkipping);
 
     for ( ; (matchIndex>=lowLimit) & (nbAttempts>0) ; nbAttempts--) {
         size_t currentMl=0;
         if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) {
             const BYTE* const match = base + matchIndex;
             assert(matchIndex >= dictLimit);   /* ensures this is true if dictMode != ZSTD_extDict */
-            if (match[ml] == ip[ml])   /* potentially better */
+            /* read 4B starting from (match + ml + 1 - sizeof(U32)) */
+            if (MEM_read32(match + ml - 3) == MEM_read32(ip + ml - 3))   /* potentially better */
                 currentMl = ZSTD_count(ip, match, iLimit);
         } else {
             const BYTE* const match = dictBase + matchIndex;
@@ -652,7 +723,7 @@ size_t ZSTD_HcFindBestMatch_generic (
         /* save best solution */
         if (currentMl > ml) {
             ml = currentMl;
-            *offsetPtr = curr - matchIndex + ZSTD_REP_MOVE;
+            *offsetPtr = OFFSET_TO_OFFBASE(curr - matchIndex);
             if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
         }
 
@@ -660,91 +731,10 @@ size_t ZSTD_HcFindBestMatch_generic (
         matchIndex = NEXT_IN_CHAIN(matchIndex, chainMask);
     }
 
+    assert(nbAttempts <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */
     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);
@@ -768,7 +758,8 @@ size_t ZSTD_HcFindBestMatch_generic (
             /* save best solution */
             if (currentMl > ml) {
                 ml = currentMl;
-                *offsetPtr = curr - (matchIndex + dmsIndexDelta) + ZSTD_REP_MOVE;
+                assert(curr > matchIndex + dmsIndexDelta);
+                *offsetPtr = OFFSET_TO_OFFBASE(curr - (matchIndex + dmsIndexDelta));
                 if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
             }
 
@@ -781,78 +772,748 @@ size_t ZSTD_HcFindBestMatch_generic (
     return ml;
 }
 
+/* *********************************
+* (SIMD) Row-based matchfinder
+***********************************/
+/* Constants for row-based hash */
+#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 */
 
-FORCE_INLINE_TEMPLATE size_t ZSTD_HcFindBestMatch_selectMLS (
-                        ZSTD_matchState_t* ms,
-                        const BYTE* ip, const BYTE* const iLimit,
-                        size_t* offsetPtr)
+#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.
+ */
+MEM_STATIC U32 ZSTD_VecMask_next(ZSTD_VecMask val) {
+    return ZSTD_countTrailingZeros64(val);
+}
+
+/* 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 [1, {entries per row})
+ */
+FORCE_INLINE_TEMPLATE U32 ZSTD_row_nextIndex(BYTE* const tagRow, U32 const rowMask) {
+    U32 next = (*tagRow-1) & rowMask;
+    next += (next == 0) ? rowMask : 0; /* skip first position */
+    *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, BYTE 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.
+ */
+FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+void ZSTD_row_fillHashCache(ZSTD_matchState_t* ms, const BYTE* base,
+                                   U32 const rowLog, U32 const mls,
+                                   U32 idx, const BYTE* const iLimit)
 {
-    switch(ms->cParams.minMatch)
-    {
-    default : /* includes case 3 */
-    case 4 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 4, ZSTD_noDict);
-    case 5 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 5, ZSTD_noDict);
-    case 7 :
-    case 6 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 6, ZSTD_noDict);
+    U32 const* const hashTable = ms->hashTable;
+    BYTE 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_hashPtrSalted(base + idx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls, ms->hashSalt);
+        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
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+U32 ZSTD_row_nextCachedHash(U32* cache, U32 const* hashTable,
+                                                  BYTE const* tagTable, BYTE const* base,
+                                                  U32 idx, U32 const hashLog,
+                                                  U32 const rowLog, U32 const mls,
+                                                  U64 const hashSalt)
+{
+    U32 const newHash = (U32)ZSTD_hashPtrSalted(base+idx+ZSTD_ROW_HASH_CACHE_SIZE, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls, hashSalt);
+    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;
+    }
+}
 
-static size_t ZSTD_HcFindBestMatch_dictMatchState_selectMLS (
-                        ZSTD_matchState_t* ms,
-                        const BYTE* ip, const BYTE* const iLimit,
-                        size_t* offsetPtr)
+/* ZSTD_row_update_internalImpl():
+ * Updates the hash table with positions starting from updateStartIdx until updateEndIdx.
+ */
+FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+void ZSTD_row_update_internalImpl(ZSTD_matchState_t* ms,
+                                  U32 updateStartIdx, U32 const updateEndIdx,
+                                  U32 const mls, U32 const rowLog,
+                                  U32 const rowMask, U32 const useCache)
 {
-    switch(ms->cParams.minMatch)
-    {
-    default : /* includes case 3 */
-    case 4 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 4, ZSTD_dictMatchState);
-    case 5 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 5, ZSTD_dictMatchState);
-    case 7 :
-    case 6 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 6, ZSTD_dictMatchState);
+    U32* const hashTable = ms->hashTable;
+    BYTE* const tagTable = ms->tagTable;
+    U32 const hashLog = ms->rowHashLog;
+    const BYTE* const base = ms->window.base;
+
+    DEBUGLOG(6, "ZSTD_row_update_internalImpl(): updateStartIdx=%u, updateEndIdx=%u", updateStartIdx, updateEndIdx);
+    for (; updateStartIdx < updateEndIdx; ++updateStartIdx) {
+        U32 const hash = useCache ? ZSTD_row_nextCachedHash(ms->hashCache, hashTable, tagTable, base, updateStartIdx, hashLog, rowLog, mls, ms->hashSalt)
+                                  : (U32)ZSTD_hashPtrSalted(base + updateStartIdx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls, ms->hashSalt);
+        U32 const relRow = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
+        U32* const row = hashTable + relRow;
+        BYTE* tagRow = tagTable + relRow;
+        U32 const pos = ZSTD_row_nextIndex(tagRow, rowMask);
+
+        assert(hash == ZSTD_hashPtrSalted(base + updateStartIdx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls, ms->hashSalt));
+        tagRow[pos] = hash & ZSTD_ROW_HASH_TAG_MASK;
+        row[pos] = updateStartIdx;
+    }
+}
+
+/* ZSTD_row_update_internal():
+ * Inserts the byte at ip into the appropriate position in the hash table, and updates ms->nextToUpdate.
+ * Skips sections of long matches as is necessary.
+ */
+FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+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 idx = ms->nextToUpdate;
+    const BYTE* const base = ms->window.base;
+    const U32 target = (U32)(ip - base);
+    const U32 kSkipThreshold = 384;
+    const U32 kMaxMatchStartPositionsToUpdate = 96;
+    const U32 kMaxMatchEndPositionsToUpdate = 32;
+
+    if (useCache) {
+        /* Only skip positions when using hash cache, i.e.
+         * if we are loading a dict, don't skip anything.
+         * If we decide to skip, then we only update a set number
+         * of positions at the beginning and end of the match.
+         */
+        if (UNLIKELY(target - idx > kSkipThreshold)) {
+            U32 const bound = idx + kMaxMatchStartPositionsToUpdate;
+            ZSTD_row_update_internalImpl(ms, idx, bound, mls, rowLog, rowMask, useCache);
+            idx = target - kMaxMatchEndPositionsToUpdate;
+            ZSTD_row_fillHashCache(ms, base, rowLog, mls, idx, ip+1);
+        }
     }
+    assert(target >= idx);
+    ZSTD_row_update_internalImpl(ms, idx, target, mls, rowLog, rowMask, useCache);
+    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 = BOUNDED(4, ms->cParams.searchLog, 6);
+    const U32 rowMask = (1u << rowLog) - 1;
+    const U32 mls = MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */);
 
-static size_t ZSTD_HcFindBestMatch_dedicatedDictSearch_selectMLS (
-                        ZSTD_matchState_t* ms,
-                        const BYTE* ip, const BYTE* const iLimit,
-                        size_t* offsetPtr)
+    DEBUGLOG(5, "ZSTD_row_update(), rowLog=%u", rowLog);
+    ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 0 /* don't use cache */);
+}
+
+/* Returns the mask width of bits group of which will be set to 1. Given not all
+ * architectures have easy movemask instruction, this helps to iterate over
+ * groups of bits easier and faster.
+ */
+FORCE_INLINE_TEMPLATE U32
+ZSTD_row_matchMaskGroupWidth(const U32 rowEntries)
 {
-    switch(ms->cParams.minMatch)
-    {
-    default : /* includes case 3 */
-    case 4 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 4, ZSTD_dedicatedDictSearch);
-    case 5 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 5, ZSTD_dedicatedDictSearch);
-    case 7 :
-    case 6 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 6, ZSTD_dedicatedDictSearch);
+    assert((rowEntries == 16) || (rowEntries == 32) || rowEntries == 64);
+    assert(rowEntries <= ZSTD_ROW_HASH_MAX_ENTRIES);
+    (void)rowEntries;
+#if defined(ZSTD_ARCH_ARM_NEON)
+    /* NEON path only works for little endian */
+    if (!MEM_isLittleEndian()) {
+        return 1;
     }
+    if (rowEntries == 16) {
+        return 4;
+    }
+    if (rowEntries == 32) {
+        return 2;
+    }
+    if (rowEntries == 64) {
+        return 1;
+    }
+#endif
+    return 1;
 }
 
+#if defined(ZSTD_ARCH_X86_SSE2)
+FORCE_INLINE_TEMPLATE ZSTD_VecMask
+ZSTD_row_getSSEMask(int nbChunks, const BYTE* const src, const BYTE tag, const U32 head)
+{
+    const __m128i comparisonMask = _mm_set1_epi8((char)tag);
+    int matches[4] = {0};
+    int i;
+    assert(nbChunks == 1 || nbChunks == 2 || nbChunks == 4);
+    for (i=0; i<nbChunks; i++) {
+        const __m128i chunk = _mm_loadu_si128((const __m128i*)(const void*)(src + 16*i));
+        const __m128i equalMask = _mm_cmpeq_epi8(chunk, comparisonMask);
+        matches[i] = _mm_movemask_epi8(equalMask);
+    }
+    if (nbChunks == 1) return ZSTD_rotateRight_U16((U16)matches[0], head);
+    if (nbChunks == 2) return ZSTD_rotateRight_U32((U32)matches[1] << 16 | (U32)matches[0], head);
+    assert(nbChunks == 4);
+    return ZSTD_rotateRight_U64((U64)matches[3] << 48 | (U64)matches[2] << 32 | (U64)matches[1] << 16 | (U64)matches[0], head);
+}
+#endif
+
+#if defined(ZSTD_ARCH_ARM_NEON)
+FORCE_INLINE_TEMPLATE ZSTD_VecMask
+ZSTD_row_getNEONMask(const U32 rowEntries, const BYTE* const src, const BYTE tag, const U32 headGrouped)
+{
+    assert((rowEntries == 16) || (rowEntries == 32) || rowEntries == 64);
+    if (rowEntries == 16) {
+        /* vshrn_n_u16 shifts by 4 every u16 and narrows to 8 lower bits.
+         * After that groups of 4 bits represent the equalMask. We lower
+         * all bits except the highest in these groups by doing AND with
+         * 0x88 = 0b10001000.
+         */
+        const uint8x16_t chunk = vld1q_u8(src);
+        const uint16x8_t equalMask = vreinterpretq_u16_u8(vceqq_u8(chunk, vdupq_n_u8(tag)));
+        const uint8x8_t res = vshrn_n_u16(equalMask, 4);
+        const U64 matches = vget_lane_u64(vreinterpret_u64_u8(res), 0);
+        return ZSTD_rotateRight_U64(matches, headGrouped) & 0x8888888888888888ull;
+    } else if (rowEntries == 32) {
+        /* Same idea as with rowEntries == 16 but doing AND with
+         * 0x55 = 0b01010101.
+         */
+        const uint16x8x2_t chunk = vld2q_u16((const uint16_t*)(const void*)src);
+        const uint8x16_t chunk0 = vreinterpretq_u8_u16(chunk.val[0]);
+        const uint8x16_t chunk1 = vreinterpretq_u8_u16(chunk.val[1]);
+        const uint8x16_t dup = vdupq_n_u8(tag);
+        const uint8x8_t t0 = vshrn_n_u16(vreinterpretq_u16_u8(vceqq_u8(chunk0, dup)), 6);
+        const uint8x8_t t1 = vshrn_n_u16(vreinterpretq_u16_u8(vceqq_u8(chunk1, dup)), 6);
+        const uint8x8_t res = vsli_n_u8(t0, t1, 4);
+        const U64 matches = vget_lane_u64(vreinterpret_u64_u8(res), 0) ;
+        return ZSTD_rotateRight_U64(matches, headGrouped) & 0x5555555555555555ull;
+    } 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, headGrouped);
+    }
+}
+#endif
+
+/* Returns a ZSTD_VecMask (U64) that has the nth group (determined by
+ * ZSTD_row_matchMaskGroupWidth) of bits set to 1 if the newly-computed "tag"
+ * matches the hash at the nth position in a row of the tagTable.
+ * Each row is a circular buffer beginning at the value of "headGrouped". So we
+ * must rotate the "matches" bitfield to match up with the actual layout of the
+ * entries within the hashTable */
+FORCE_INLINE_TEMPLATE ZSTD_VecMask
+ZSTD_row_getMatchMask(const BYTE* const tagRow, const BYTE tag, const U32 headGrouped, const U32 rowEntries)
+{
+    const BYTE* const src = tagRow;
+    assert((rowEntries == 16) || (rowEntries == 32) || rowEntries == 64);
+    assert(rowEntries <= ZSTD_ROW_HASH_MAX_ENTRIES);
+    assert(ZSTD_row_matchMaskGroupWidth(rowEntries) * rowEntries <= sizeof(ZSTD_VecMask) * 8);
+
+#if defined(ZSTD_ARCH_X86_SSE2)
+
+    return ZSTD_row_getSSEMask(rowEntries / 16, src, tag, headGrouped);
+
+#else /* SW or NEON-LE */
+
+# if defined(ZSTD_ARCH_ARM_NEON)
+  /* This NEON path only works for little endian - otherwise use SWAR below */
+    if (MEM_isLittleEndian()) {
+        return ZSTD_row_getNEONMask(rowEntries, src, tag, headGrouped);
+    }
+# endif /* ZSTD_ARCH_ARM_NEON */
+    /* SWAR */
+    {   const int chunkSize = sizeof(size_t);
+        const size_t shiftAmount = ((chunkSize * 8) - chunkSize);
+        const size_t xFF = ~((size_t)0);
+        const size_t x01 = xFF / 0xFF;
+        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, headGrouped);
+        } else if (rowEntries == 32) {
+            return ZSTD_rotateRight_U32((U32)matches, headGrouped);
+        } else {
+            return ZSTD_rotateRight_U64((U64)matches, headGrouped);
+        }
+    }
+#endif
+}
 
-FORCE_INLINE_TEMPLATE size_t ZSTD_HcFindBestMatch_extDict_selectMLS (
+/* 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 for current input posistion and split it into a one byte of tag and `rowHashLog` bits of index.
+ *           - The hash is salted by a value that changes on every contex reset, so when the same table is used
+ *             we will avoid collisions that would otherwise slow us down by intorducing phantom matches.
+ *      - The hashTable is effectively split into groups or "rows" of 15 or 31 entries of U32, and the index determines
+ *        which row to insert into.
+ *      - Determine the correct position within the row to insert the entry into. Each row of 15 or 31 can
+ *        be considered as a circular buffer with a "head" index that resides in the tagTable (overall 16 or 32 bytes
+ *        per row).
+ * - Use SIMD to efficiently compare the tags in the tagTable to the 1-byte tag calculated for the position and
+ *   generate a bitfield that we can cycle through to check the collisions in the hash table.
+ * - Pick the longest match.
+ * - Insert the tag into the equivalent row and position in the tagTable.
+ */
+FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+size_t ZSTD_RowFindBestMatch(
                         ZSTD_matchState_t* ms,
-                        const BYTE* ip, const BYTE* const iLimit,
-                        size_t* offsetPtr)
+                        const BYTE* const ip, const BYTE* const iLimit,
+                        size_t* offsetPtr,
+                        const U32 mls, const ZSTD_dictMode_e dictMode,
+                        const U32 rowLog)
 {
-    switch(ms->cParams.minMatch)
-    {
-    default : /* includes case 3 */
-    case 4 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 4, ZSTD_extDict);
-    case 5 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 5, ZSTD_extDict);
-    case 7 :
-    case 6 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 6, ZSTD_extDict);
+    U32* const hashTable = ms->hashTable;
+    BYTE* 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 */
+    const U32 groupWidth = ZSTD_row_matchMaskGroupWidth(rowEntries);
+    const U64 hashSalt = ms->hashSalt;
+    U32 nbAttempts = 1U << cappedSearchLog;
+    size_t ml=4-1;
+    U32 hash;
+
+    /* 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;
+        BYTE* 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 */
+    if (!ms->lazySkipping) {
+        ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 1 /* useCache */);
+        hash = ZSTD_row_nextCachedHash(hashCache, hashTable, tagTable, base, curr, hashLog, rowLog, mls, hashSalt);
+    } else {
+        /* Stop inserting every position when in the lazy skipping mode.
+         * The hash cache is also not kept up to date in this mode.
+         */
+        hash = (U32)ZSTD_hashPtrSalted(ip, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls, hashSalt);
+        ms->nextToUpdate = curr;
+    }
+    ms->hashSaltEntropy += hash; /* collect salt entropy */
+
+    {   /* Get the hash for ip, compute the appropriate row */
+        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 headGrouped = (*tagRow & rowMask) * groupWidth;
+        U32 matchBuffer[ZSTD_ROW_HASH_MAX_ENTRIES];
+        size_t numMatches = 0;
+        size_t currMatch = 0;
+        ZSTD_VecMask matches = ZSTD_row_getMatchMask(tagRow, (BYTE)tag, headGrouped, rowEntries);
+
+        /* Cycle through the matches and prefetch */
+        for (; (matches > 0) && (nbAttempts > 0); matches &= (matches - 1)) {
+            U32 const matchPos = ((headGrouped + ZSTD_VecMask_next(matches)) / groupWidth) & rowMask;
+            U32 const matchIndex = row[matchPos];
+            if(matchPos == 0) continue;
+            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;
+            --nbAttempts;
+        }
+
+        /* 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] = (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 */
+                /* read 4B starting from (match + ml + 1 - sizeof(U32)) */
+                if (MEM_read32(match + ml - 3) == MEM_read32(ip + ml - 3))   /* potentially better */
+                    currentMl = ZSTD_count(ip, match, iLimit);
+            } else {
+                const BYTE* const match = dictBase + matchIndex;
+                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 = OFFSET_TO_OFFBASE(curr - matchIndex);
+                if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
+            }
+        }
+    }
+
+    assert(nbAttempts <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */
+    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 headGrouped = (*dmsTagRow & rowMask) * groupWidth;
+            U32 matchBuffer[ZSTD_ROW_HASH_MAX_ENTRIES];
+            size_t numMatches = 0;
+            size_t currMatch = 0;
+            ZSTD_VecMask matches = ZSTD_row_getMatchMask(dmsTagRow, (BYTE)dmsTag, headGrouped, rowEntries);
+
+            for (; (matches > 0) && (nbAttempts > 0); matches &= (matches - 1)) {
+                U32 const matchPos = ((headGrouped + ZSTD_VecMask_next(matches)) / groupWidth) & rowMask;
+                U32 const matchIndex = dmsRow[matchPos];
+                if(matchPos == 0) continue;
+                if (matchIndex < dmsLowestIndex)
+                    break;
+                PREFETCH_L1(dmsBase + matchIndex);
+                matchBuffer[numMatches++] = matchIndex;
+                --nbAttempts;
+            }
+
+            /* 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;
+                    assert(curr > matchIndex + dmsIndexDelta);
+                    *offsetPtr = OFFSET_TO_OFFBASE(curr - (matchIndex + dmsIndexDelta));
+                    if (ip+currentMl == iLimit) break;
+                }
+            }
+        }
     }
+    return ml;
 }
 
 
+/**
+ * Generate search functions templated on (dictMode, mls, rowLog).
+ * These functions are outlined for code size & compilation time.
+ * ZSTD_searchMax() dispatches to the correct implementation function.
+ *
+ * TODO: The start of the search function involves loading and calculating a
+ * bunch of constants from the ZSTD_matchState_t. These computations could be
+ * done in an initialization function, and saved somewhere in the match state.
+ * Then we could pass a pointer to the saved state instead of the match state,
+ * and avoid duplicate computations.
+ *
+ * TODO: Move the match re-winding into searchMax. This improves compression
+ * ratio, and unlocks further simplifications with the next TODO.
+ *
+ * TODO: Try moving the repcode search into searchMax. After the re-winding
+ * and repcode search are in searchMax, there is no more logic in the match
+ * finder loop that requires knowledge about the dictMode. So we should be
+ * able to avoid force inlining it, and we can join the extDict loop with
+ * the single segment loop. It should go in searchMax instead of its own
+ * function to avoid having multiple virtual function calls per search.
+ */
+
+#define ZSTD_BT_SEARCH_FN(dictMode, mls) ZSTD_BtFindBestMatch_##dictMode##_##mls
+#define ZSTD_HC_SEARCH_FN(dictMode, mls) ZSTD_HcFindBestMatch_##dictMode##_##mls
+#define ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog) ZSTD_RowFindBestMatch_##dictMode##_##mls##_##rowLog
+
+#define ZSTD_SEARCH_FN_ATTRS FORCE_NOINLINE
+
+#define GEN_ZSTD_BT_SEARCH_FN(dictMode, mls)                                           \
+    ZSTD_SEARCH_FN_ATTRS size_t ZSTD_BT_SEARCH_FN(dictMode, mls)(                      \
+            ZSTD_matchState_t* ms,                                                     \
+            const BYTE* ip, const BYTE* const iLimit,                                  \
+            size_t* offBasePtr)                                                        \
+    {                                                                                  \
+        assert(MAX(4, MIN(6, ms->cParams.minMatch)) == mls);                           \
+        return ZSTD_BtFindBestMatch(ms, ip, iLimit, offBasePtr, mls, ZSTD_##dictMode); \
+    }                                                                                  \
+
+#define GEN_ZSTD_HC_SEARCH_FN(dictMode, mls)                                          \
+    ZSTD_SEARCH_FN_ATTRS size_t ZSTD_HC_SEARCH_FN(dictMode, mls)(                     \
+            ZSTD_matchState_t* ms,                                                    \
+            const BYTE* ip, const BYTE* const iLimit,                                 \
+            size_t* offsetPtr)                                                        \
+    {                                                                                 \
+        assert(MAX(4, MIN(6, ms->cParams.minMatch)) == mls);                          \
+        return ZSTD_HcFindBestMatch(ms, ip, iLimit, offsetPtr, mls, ZSTD_##dictMode); \
+    }                                                                                 \
+
+#define GEN_ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog)                                          \
+    ZSTD_SEARCH_FN_ATTRS size_t ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog)(                     \
+            ZSTD_matchState_t* ms,                                                             \
+            const BYTE* ip, const BYTE* const iLimit,                                          \
+            size_t* offsetPtr)                                                                 \
+    {                                                                                          \
+        assert(MAX(4, MIN(6, ms->cParams.minMatch)) == mls);                                   \
+        assert(MAX(4, MIN(6, ms->cParams.searchLog)) == rowLog);                               \
+        return ZSTD_RowFindBestMatch(ms, ip, iLimit, offsetPtr, mls, ZSTD_##dictMode, rowLog); \
+    }                                                                                          \
+
+#define ZSTD_FOR_EACH_ROWLOG(X, dictMode, mls) \
+    X(dictMode, mls, 4)                        \
+    X(dictMode, mls, 5)                        \
+    X(dictMode, mls, 6)
+
+#define ZSTD_FOR_EACH_MLS_ROWLOG(X, dictMode) \
+    ZSTD_FOR_EACH_ROWLOG(X, dictMode, 4)      \
+    ZSTD_FOR_EACH_ROWLOG(X, dictMode, 5)      \
+    ZSTD_FOR_EACH_ROWLOG(X, dictMode, 6)
+
+#define ZSTD_FOR_EACH_MLS(X, dictMode) \
+    X(dictMode, 4)                     \
+    X(dictMode, 5)                     \
+    X(dictMode, 6)
+
+#define ZSTD_FOR_EACH_DICT_MODE(X, ...) \
+    X(__VA_ARGS__, noDict)              \
+    X(__VA_ARGS__, extDict)             \
+    X(__VA_ARGS__, dictMatchState)      \
+    X(__VA_ARGS__, dedicatedDictSearch)
+
+/* Generate row search fns for each combination of (dictMode, mls, rowLog) */
+ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS_ROWLOG, GEN_ZSTD_ROW_SEARCH_FN)
+/* Generate binary Tree search fns for each combination of (dictMode, mls) */
+ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS, GEN_ZSTD_BT_SEARCH_FN)
+/* Generate hash chain search fns for each combination of (dictMode, mls) */
+ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS, GEN_ZSTD_HC_SEARCH_FN)
+
+typedef enum { search_hashChain=0, search_binaryTree=1, search_rowHash=2 } searchMethod_e;
+
+#define GEN_ZSTD_CALL_BT_SEARCH_FN(dictMode, mls)                         \
+    case mls:                                                             \
+        return ZSTD_BT_SEARCH_FN(dictMode, mls)(ms, ip, iend, offsetPtr);
+#define GEN_ZSTD_CALL_HC_SEARCH_FN(dictMode, mls)                         \
+    case mls:                                                             \
+        return ZSTD_HC_SEARCH_FN(dictMode, mls)(ms, ip, iend, offsetPtr);
+#define GEN_ZSTD_CALL_ROW_SEARCH_FN(dictMode, mls, rowLog)                         \
+    case rowLog:                                                                   \
+        return ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog)(ms, ip, iend, offsetPtr);
+
+#define ZSTD_SWITCH_MLS(X, dictMode)   \
+    switch (mls) {                     \
+        ZSTD_FOR_EACH_MLS(X, dictMode) \
+    }
+
+#define ZSTD_SWITCH_ROWLOG(dictMode, mls)                                    \
+    case mls:                                                                \
+        switch (rowLog) {                                                    \
+            ZSTD_FOR_EACH_ROWLOG(GEN_ZSTD_CALL_ROW_SEARCH_FN, dictMode, mls) \
+        }                                                                    \
+        ZSTD_UNREACHABLE;                                                    \
+        break;
+
+#define ZSTD_SWITCH_SEARCH_METHOD(dictMode)                       \
+    switch (searchMethod) {                                       \
+        case search_hashChain:                                    \
+            ZSTD_SWITCH_MLS(GEN_ZSTD_CALL_HC_SEARCH_FN, dictMode) \
+            break;                                                \
+        case search_binaryTree:                                   \
+            ZSTD_SWITCH_MLS(GEN_ZSTD_CALL_BT_SEARCH_FN, dictMode) \
+            break;                                                \
+        case search_rowHash:                                      \
+            ZSTD_SWITCH_MLS(ZSTD_SWITCH_ROWLOG, dictMode)         \
+            break;                                                \
+    }                                                             \
+    ZSTD_UNREACHABLE;
+
+/**
+ * Searches for the longest match at @p ip.
+ * Dispatches to the correct implementation function based on the
+ * (searchMethod, dictMode, mls, rowLog). We use switch statements
+ * here instead of using an indirect function call through a function
+ * pointer because after Spectre and Meltdown mitigations, indirect
+ * function calls can be very costly, especially in the kernel.
+ *
+ * NOTE: dictMode and searchMethod should be templated, so those switch
+ * statements should be optimized out. Only the mls & rowLog switches
+ * should be left.
+ *
+ * @param ms The match state.
+ * @param ip The position to search at.
+ * @param iend The end of the input data.
+ * @param[out] offsetPtr Stores the match offset into this pointer.
+ * @param mls The minimum search length, in the range [4, 6].
+ * @param rowLog The row log (if applicable), in the range [4, 6].
+ * @param searchMethod The search method to use (templated).
+ * @param dictMode The dictMode (templated).
+ *
+ * @returns The length of the longest match found, or < mls if no match is found.
+ * If a match is found its offset is stored in @p offsetPtr.
+ */
+FORCE_INLINE_TEMPLATE size_t ZSTD_searchMax(
+    ZSTD_matchState_t* ms,
+    const BYTE* ip,
+    const BYTE* iend,
+    size_t* offsetPtr,
+    U32 const mls,
+    U32 const rowLog,
+    searchMethod_e const searchMethod,
+    ZSTD_dictMode_e const dictMode)
+{
+    if (dictMode == ZSTD_noDict) {
+        ZSTD_SWITCH_SEARCH_METHOD(noDict)
+    } else if (dictMode == ZSTD_extDict) {
+        ZSTD_SWITCH_SEARCH_METHOD(extDict)
+    } else if (dictMode == ZSTD_dictMatchState) {
+        ZSTD_SWITCH_SEARCH_METHOD(dictMatchState)
+    } else if (dictMode == ZSTD_dedicatedDictSearch) {
+        ZSTD_SWITCH_SEARCH_METHOD(dedicatedDictSearch)
+    }
+    ZSTD_UNREACHABLE;
+    return 0;
+}
+
 /* *******************************
 *  Common parser - lazy strategy
 *********************************/
-typedef enum { search_hashChain, search_binaryTree } searchMethod_e;
 
-FORCE_INLINE_TEMPLATE size_t
-ZSTD_compressBlock_lazy_generic(
+FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
+size_t ZSTD_compressBlock_lazy_generic(
                         ZSTD_matchState_t* ms, seqStore_t* seqStore,
                         U32 rep[ZSTD_REP_NUM],
                         const void* src, size_t srcSize,
@@ -863,42 +1524,15 @@ 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 mls = BOUNDED(4, ms->cParams.minMatch, 6);
+    const U32 rowLog = BOUNDED(4, ms->cParams.searchLog, 6);
 
-    typedef size_t (*searchMax_f)(
-                        ZSTD_matchState_t* ms,
-                        const BYTE* ip, const BYTE* iLimit, size_t* offsetPtr);
-
-    /**
-     * This table is indexed first by the four ZSTD_dictMode_e values, and then
-     * by the two searchMethod_e values. NULLs are placed for configurations
-     * that should never occur (extDict modes go to the other implementation
-     * below and there is no DDSS for binary tree search yet).
-     */
-    const searchMax_f searchFuncs[4][2] = {
-        {
-            ZSTD_HcFindBestMatch_selectMLS,
-            ZSTD_BtFindBestMatch_selectMLS
-        },
-        {
-            NULL,
-            NULL
-        },
-        {
-            ZSTD_HcFindBestMatch_dictMatchState_selectMLS,
-            ZSTD_BtFindBestMatch_dictMatchState_selectMLS
-        },
-        {
-            ZSTD_HcFindBestMatch_dedicatedDictSearch_selectMLS,
-            NULL
-        }
-    };
-
-    searchMax_f const searchMax = searchFuncs[dictMode][searchMethod == search_binaryTree];
-    U32 offset_1 = rep[0], offset_2 = rep[1], savedOffset=0;
+    U32 offset_1 = rep[0], offset_2 = rep[1];
+    U32 offsetSaved1 = 0, offsetSaved2 = 0;
 
     const int isDMS = dictMode == ZSTD_dictMatchState;
     const int isDDS = dictMode == ZSTD_dedicatedDictSearch;
@@ -913,18 +1547,14 @@ ZSTD_compressBlock_lazy_generic(
                                      0;
     const U32 dictAndPrefixLength = (U32)((ip - prefixLowest) + (dictEnd - dictLowest));
 
-    assert(searchMax != NULL);
-
-    DEBUGLOG(5, "ZSTD_compressBlock_lazy_generic (dictMode=%u)", (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);
         U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, curr, ms->cParams.windowLog);
         U32 const maxRep = curr - windowLow;
-        if (offset_2 > maxRep) savedOffset = offset_2, offset_2 = 0;
-        if (offset_1 > maxRep) savedOffset = offset_1, offset_1 = 0;
+        if (offset_2 > maxRep) offsetSaved2 = offset_2, offset_2 = 0;
+        if (offset_1 > maxRep) offsetSaved1 = offset_1, offset_1 = 0;
     }
     if (isDxS) {
         /* dictMatchState repCode checks don't currently handle repCode == 0
@@ -933,6 +1563,13 @@ ZSTD_compressBlock_lazy_generic(
         assert(offset_2 <= dictAndPrefixLength);
     }
 
+    /* Reset the lazy skipping state */
+    ms->lazySkipping = 0;
+
+    if (searchMethod == search_rowHash) {
+        ZSTD_row_fillHashCache(ms, base, rowLog, mls, 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
@@ -942,8 +1579,9 @@ ZSTD_compressBlock_lazy_generic(
 #endif
     while (ip < ilimit) {
         size_t matchLength=0;
-        size_t offset=0;
+        size_t offBase = REPCODE1_TO_OFFBASE;
         const BYTE* start=ip+1;
+        DEBUGLOG(7, "search baseline (depth 0)");
 
         /* check repCode */
         if (isDxS) {
@@ -966,28 +1604,38 @@ ZSTD_compressBlock_lazy_generic(
         }
 
         /* first search (depth 0) */
-        {   size_t offsetFound = 999999999;
-            size_t const ml2 = searchMax(ms, ip, iend, &offsetFound);
+        {   size_t offbaseFound = 999999999;
+            size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &offbaseFound, mls, rowLog, searchMethod, dictMode);
             if (ml2 > matchLength)
-                matchLength = ml2, start = ip, offset=offsetFound;
+                matchLength = ml2, start = ip, offBase = offbaseFound;
         }
 
         if (matchLength < 4) {
-            ip += ((ip-anchor) >> kSearchStrength) + 1;   /* jump faster over incompressible sections */
+            size_t const step = ((size_t)(ip-anchor) >> kSearchStrength) + 1;   /* jump faster over incompressible sections */;
+            ip += step;
+            /* Enter the lazy skipping mode once we are skipping more than 8 bytes at a time.
+             * In this mode we stop inserting every position into our tables, and only insert
+             * positions that we search, which is one in step positions.
+             * The exact cutoff is flexible, I've just chosen a number that is reasonably high,
+             * so we minimize the compression ratio loss in "normal" scenarios. This mode gets
+             * triggered once we've gone 2KB without finding any matches.
+             */
+            ms->lazySkipping = step > kLazySkippingStep;
             continue;
         }
 
         /* let's try to find a better solution */
         if (depth>=1)
         while (ip<ilimit) {
+            DEBUGLOG(7, "search depth 1");
             ip ++;
             if ( (dictMode == ZSTD_noDict)
-              && (offset) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
+              && (offBase) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
                 size_t const mlRep = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4;
                 int const gain2 = (int)(mlRep * 3);
-                int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offset+1) + 1);
+                int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offBase) + 1);
                 if ((mlRep >= 4) && (gain2 > gain1))
-                    matchLength = mlRep, offset = 0, start = ip;
+                    matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
             }
             if (isDxS) {
                 const U32 repIndex = (U32)(ip - base) - offset_1;
@@ -999,30 +1647,31 @@ ZSTD_compressBlock_lazy_generic(
                     const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
                     size_t const mlRep = ZSTD_count_2segments(ip+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
                     int const gain2 = (int)(mlRep * 3);
-                    int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offset+1) + 1);
+                    int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offBase) + 1);
                     if ((mlRep >= 4) && (gain2 > gain1))
-                        matchLength = mlRep, offset = 0, start = ip;
+                        matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
                 }
             }
-            {   size_t offset2=999999999;
-                size_t const ml2 = searchMax(ms, ip, iend, &offset2);
-                int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1));   /* raw approx */
-                int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 4);
+            {   size_t ofbCandidate=999999999;
+                size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, dictMode);
+                int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */
+                int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 4);
                 if ((ml2 >= 4) && (gain2 > gain1)) {
-                    matchLength = ml2, offset = offset2, start = ip;
+                    matchLength = ml2, offBase = ofbCandidate, start = ip;
                     continue;   /* search a better one */
             }   }
 
             /* let's find an even better one */
             if ((depth==2) && (ip<ilimit)) {
+                DEBUGLOG(7, "search depth 2");
                 ip ++;
                 if ( (dictMode == ZSTD_noDict)
-                  && (offset) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
+                  && (offBase) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
                     size_t const mlRep = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4;
                     int const gain2 = (int)(mlRep * 4);
-                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 1);
+                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 1);
                     if ((mlRep >= 4) && (gain2 > gain1))
-                        matchLength = mlRep, offset = 0, start = ip;
+                        matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
                 }
                 if (isDxS) {
                     const U32 repIndex = (U32)(ip - base) - offset_1;
@@ -1034,48 +1683,54 @@ ZSTD_compressBlock_lazy_generic(
                         const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
                         size_t const mlRep = ZSTD_count_2segments(ip+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
                         int const gain2 = (int)(mlRep * 4);
-                        int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 1);
+                        int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 1);
                         if ((mlRep >= 4) && (gain2 > gain1))
-                            matchLength = mlRep, offset = 0, start = ip;
+                            matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
                     }
                 }
-                {   size_t offset2=999999999;
-                    size_t const ml2 = searchMax(ms, ip, iend, &offset2);
-                    int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1));   /* raw approx */
-                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 7);
+                {   size_t ofbCandidate=999999999;
+                    size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, dictMode);
+                    int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */
+                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 7);
                     if ((ml2 >= 4) && (gain2 > gain1)) {
-                        matchLength = ml2, offset = offset2, start = ip;
+                        matchLength = ml2, offBase = ofbCandidate, start = ip;
                         continue;
             }   }   }
             break;  /* nothing found : store previous solution */
         }
 
         /* NOTE:
-         * start[-offset+ZSTD_REP_MOVE-1] is undefined behavior.
-         * (-offset+ZSTD_REP_MOVE-1) is unsigned, and is added to start, which
-         * overflows the pointer, which is undefined behavior.
+         * Pay attention that `start[-value]` can lead to strange undefined behavior
+         * notably if `value` is unsigned, resulting in a large positive `-value`.
          */
         /* catch up */
-        if (offset) {
+        if (OFFBASE_IS_OFFSET(offBase)) {
             if (dictMode == ZSTD_noDict) {
-                while ( ((start > anchor) & (start - (offset-ZSTD_REP_MOVE) > prefixLowest))
-                     && (start[-1] == (start-(offset-ZSTD_REP_MOVE))[-1]) )  /* only search for offset within prefix */
+                while ( ((start > anchor) & (start - OFFBASE_TO_OFFSET(offBase) > prefixLowest))
+                     && (start[-1] == (start-OFFBASE_TO_OFFSET(offBase))[-1]) )  /* only search for offset within prefix */
                     { start--; matchLength++; }
             }
             if (isDxS) {
-                U32 const matchIndex = (U32)((start-base) - (offset - ZSTD_REP_MOVE));
+                U32 const matchIndex = (U32)((size_t)(start-base) - OFFBASE_TO_OFFSET(offBase));
                 const BYTE* match = (matchIndex < prefixLowestIndex) ? dictBase + matchIndex - dictIndexDelta : base + matchIndex;
                 const BYTE* const mStart = (matchIndex < prefixLowestIndex) ? dictLowest : prefixLowest;
                 while ((start>anchor) && (match>mStart) && (start[-1] == match[-1])) { start--; match--; matchLength++; }  /* catch up */
             }
-            offset_2 = offset_1; offset_1 = (U32)(offset - ZSTD_REP_MOVE);
+            offset_2 = offset_1; offset_1 = (U32)OFFBASE_TO_OFFSET(offBase);
         }
         /* store sequence */
 _storeSequence:
-        {   size_t const litLength = start - anchor;
-            ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offset, matchLength-MINMATCH);
+        {   size_t const litLength = (size_t)(start - anchor);
+            ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offBase, matchLength);
             anchor = ip = start + matchLength;
         }
+        if (ms->lazySkipping) {
+            /* We've found a match, disable lazy skipping mode, and refill the hash cache. */
+            if (searchMethod == search_rowHash) {
+                ZSTD_row_fillHashCache(ms, base, rowLog, mls, ms->nextToUpdate, ilimit);
+            }
+            ms->lazySkipping = 0;
+        }
 
         /* check immediate repcode */
         if (isDxS) {
@@ -1089,8 +1744,8 @@ _storeSequence:
                    && (MEM_read32(repMatch) == MEM_read32(ip)) ) {
                     const BYTE* const repEnd2 = repIndex < prefixLowestIndex ? dictEnd : iend;
                     matchLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd2, prefixLowest) + 4;
-                    offset = offset_2; offset_2 = offset_1; offset_1 = (U32)offset;   /* swap offset_2 <=> offset_1 */
-                    ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, matchLength-MINMATCH);
+                    offBase = offset_2; offset_2 = offset_1; offset_1 = (U32)offBase;   /* swap offset_2 <=> offset_1 */
+                    ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, matchLength);
                     ip += matchLength;
                     anchor = ip;
                     continue;
@@ -1104,62 +1759,77 @@ _storeSequence:
                  && (MEM_read32(ip) == MEM_read32(ip - offset_2)) ) {
                 /* store sequence */
                 matchLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4;
-                offset = offset_2; offset_2 = offset_1; offset_1 = (U32)offset; /* swap repcodes */
-                ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, matchLength-MINMATCH);
+                offBase = offset_2; offset_2 = offset_1; offset_1 = (U32)offBase; /* swap repcodes */
+                ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, matchLength);
                 ip += matchLength;
                 anchor = ip;
                 continue;   /* faster when present ... (?) */
     }   }   }
 
-    /* Save reps for next block */
-    rep[0] = offset_1 ? offset_1 : savedOffset;
-    rep[1] = offset_2 ? offset_2 : savedOffset;
+    /* If offset_1 started invalid (offsetSaved1 != 0) and became valid (offset_1 != 0),
+     * rotate saved offsets. See comment in ZSTD_compressBlock_fast_noDict for more context. */
+    offsetSaved2 = ((offsetSaved1 != 0) && (offset_1 != 0)) ? offsetSaved1 : offsetSaved2;
+
+    /* save reps for next block */
+    rep[0] = offset_1 ? offset_1 : offsetSaved1;
+    rep[1] = offset_2 ? offset_2 : offsetSaved2;
 
     /* Return the last literals size */
     return (size_t)(iend - anchor);
 }
+#endif /* build exclusions */
 
 
-size_t ZSTD_compressBlock_btlazy2(
+#ifndef ZSTD_EXCLUDE_GREEDY_BLOCK_COMPRESSOR
+size_t ZSTD_compressBlock_greedy(
         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_binaryTree, 2, ZSTD_noDict);
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_noDict);
 }
 
-size_t ZSTD_compressBlock_lazy2(
+size_t ZSTD_compressBlock_greedy_dictMatchState(
         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_hashChain, 2, ZSTD_noDict);
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_dictMatchState);
 }
 
-size_t ZSTD_compressBlock_lazy(
+size_t ZSTD_compressBlock_greedy_dedicatedDictSearch(
         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_hashChain, 1, ZSTD_noDict);
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_dedicatedDictSearch);
 }
 
-size_t ZSTD_compressBlock_greedy(
+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_hashChain, 0, ZSTD_noDict);
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_noDict);
 }
 
-size_t ZSTD_compressBlock_btlazy2_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_binaryTree, 2, ZSTD_dictMatchState);
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_dictMatchState);
 }
 
-size_t ZSTD_compressBlock_lazy2_dictMatchState(
+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_hashChain, 2, ZSTD_dictMatchState);
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_dedicatedDictSearch);
+}
+#endif
+
+#ifndef ZSTD_EXCLUDE_LAZY_BLOCK_COMPRESSOR
+size_t ZSTD_compressBlock_lazy(
+        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_hashChain, 1, ZSTD_noDict);
 }
 
 size_t ZSTD_compressBlock_lazy_dictMatchState(
@@ -1169,13 +1839,49 @@ size_t ZSTD_compressBlock_lazy_dictMatchState(
     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_dictMatchState);
 }
 
-size_t ZSTD_compressBlock_greedy_dictMatchState(
+size_t ZSTD_compressBlock_lazy_dedicatedDictSearch(
         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_hashChain, 0, ZSTD_dictMatchState);
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_dedicatedDictSearch);
+}
+
+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_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_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);
+}
+#endif
+
+#ifndef ZSTD_EXCLUDE_LAZY2_BLOCK_COMPRESSOR
+size_t ZSTD_compressBlock_lazy2(
+        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_hashChain, 2, ZSTD_noDict);
 }
 
+size_t ZSTD_compressBlock_lazy2_dictMatchState(
+        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_hashChain, 2, ZSTD_dictMatchState);
+}
 
 size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
@@ -1184,22 +1890,50 @@ size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch(
     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_dedicatedDictSearch);
 }
 
-size_t ZSTD_compressBlock_lazy_dedicatedDictSearch(
+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_hashChain, 1, ZSTD_dedicatedDictSearch);
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_noDict);
 }
 
-size_t ZSTD_compressBlock_greedy_dedicatedDictSearch(
+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_hashChain, 0, ZSTD_dedicatedDictSearch);
+    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, 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);
 }
+#endif
 
+#ifndef ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR
+size_t ZSTD_compressBlock_btlazy2(
+        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_binaryTree, 2, ZSTD_noDict);
+}
+
+size_t ZSTD_compressBlock_btlazy2_dictMatchState(
+        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_binaryTree, 2, ZSTD_dictMatchState);
+}
+#endif
 
+#if !defined(ZSTD_EXCLUDE_GREEDY_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_LAZY_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_LAZY2_BLOCK_COMPRESSOR) \
+ || !defined(ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR)
 FORCE_INLINE_TEMPLATE
+ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
 size_t ZSTD_compressBlock_lazy_extDict_generic(
                         ZSTD_matchState_t* ms, seqStore_t* seqStore,
                         U32 rep[ZSTD_REP_NUM],
@@ -1210,7 +1944,7 @@ 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 +1952,21 @@ 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;
-
-    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 U32 mls = BOUNDED(4, ms->cParams.minMatch, 6);
+    const U32 rowLog = BOUNDED(4, ms->cParams.searchLog, 6);
 
     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);
+
+    /* Reset the lazy skipping state */
+    ms->lazySkipping = 0;
 
     /* init */
     ip += (ip == prefixStart);
+    if (searchMethod == search_rowHash) {
+        ZSTD_row_fillHashCache(ms, base, rowLog, mls, ms->nextToUpdate, ilimit);
+    }
 
     /* Match Loop */
 #if defined(__GNUC__) && defined(__x86_64__)
@@ -1240,7 +1977,7 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
 #endif
     while (ip < ilimit) {
         size_t matchLength=0;
-        size_t offset=0;
+        size_t offBase = REPCODE1_TO_OFFBASE;
         const BYTE* start=ip+1;
         U32 curr = (U32)(ip-base);
 
@@ -1249,7 +1986,8 @@ 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;
@@ -1258,14 +1996,23 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
         }   }
 
         /* first search (depth 0) */
-        {   size_t offsetFound = 999999999;
-            size_t const ml2 = searchMax(ms, ip, iend, &offsetFound);
+        {   size_t ofbCandidate = 999999999;
+            size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, ZSTD_extDict);
             if (ml2 > matchLength)
-                matchLength = ml2, start = ip, offset=offsetFound;
+                matchLength = ml2, start = ip, offBase = ofbCandidate;
         }
 
-         if (matchLength < 4) {
-            ip += ((ip-anchor) >> kSearchStrength) + 1;   /* jump faster over incompressible sections */
+        if (matchLength < 4) {
+            size_t const step = ((size_t)(ip-anchor) >> kSearchStrength);
+            ip += step + 1;   /* jump faster over incompressible sections */
+            /* Enter the lazy skipping mode once we are skipping more than 8 bytes at a time.
+             * In this mode we stop inserting every position into our tables, and only insert
+             * positions that we search, which is one in step positions.
+             * The exact cutoff is flexible, I've just chosen a number that is reasonably high,
+             * so we minimize the compression ratio loss in "normal" scenarios. This mode gets
+             * triggered once we've gone 2KB without finding any matches.
+             */
+            ms->lazySkipping = step > kLazySkippingStep;
             continue;
         }
 
@@ -1275,29 +2022,30 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
             ip ++;
             curr++;
             /* check repCode */
-            if (offset) {
+            if (offBase) {
                 const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr, windowLog);
                 const U32 repIndex = (U32)(curr - offset_1);
                 const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
                 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;
                     size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
                     int const gain2 = (int)(repLength * 3);
-                    int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offset+1) + 1);
+                    int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offBase) + 1);
                     if ((repLength >= 4) && (gain2 > gain1))
-                        matchLength = repLength, offset = 0, start = ip;
+                        matchLength = repLength, offBase = REPCODE1_TO_OFFBASE, start = ip;
             }   }
 
             /* search match, depth 1 */
-            {   size_t offset2=999999999;
-                size_t const ml2 = searchMax(ms, ip, iend, &offset2);
-                int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1));   /* raw approx */
-                int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 4);
+            {   size_t ofbCandidate = 999999999;
+                size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, ZSTD_extDict);
+                int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */
+                int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 4);
                 if ((ml2 >= 4) && (gain2 > gain1)) {
-                    matchLength = ml2, offset = offset2, start = ip;
+                    matchLength = ml2, offBase = ofbCandidate, start = ip;
                     continue;   /* search a better one */
             }   }
 
@@ -1306,49 +2054,57 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
                 ip ++;
                 curr++;
                 /* check repCode */
-                if (offset) {
+                if (offBase) {
                     const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr, windowLog);
                     const U32 repIndex = (U32)(curr - offset_1);
                     const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
                     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;
                         size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
                         int const gain2 = (int)(repLength * 4);
-                        int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 1);
+                        int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 1);
                         if ((repLength >= 4) && (gain2 > gain1))
-                            matchLength = repLength, offset = 0, start = ip;
+                            matchLength = repLength, offBase = REPCODE1_TO_OFFBASE, start = ip;
                 }   }
 
                 /* search match, depth 2 */
-                {   size_t offset2=999999999;
-                    size_t const ml2 = searchMax(ms, ip, iend, &offset2);
-                    int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1));   /* raw approx */
-                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 7);
+                {   size_t ofbCandidate = 999999999;
+                    size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, ZSTD_extDict);
+                    int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */
+                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 7);
                     if ((ml2 >= 4) && (gain2 > gain1)) {
-                        matchLength = ml2, offset = offset2, start = ip;
+                        matchLength = ml2, offBase = ofbCandidate, start = ip;
                         continue;
             }   }   }
             break;  /* nothing found : store previous solution */
         }
 
         /* catch up */
-        if (offset) {
-            U32 const matchIndex = (U32)((start-base) - (offset - ZSTD_REP_MOVE));
+        if (OFFBASE_IS_OFFSET(offBase)) {
+            U32 const matchIndex = (U32)((size_t)(start-base) - OFFBASE_TO_OFFSET(offBase));
             const BYTE* match = (matchIndex < dictLimit) ? dictBase + matchIndex : base + matchIndex;
             const BYTE* const mStart = (matchIndex < dictLimit) ? dictStart : prefixStart;
             while ((start>anchor) && (match>mStart) && (start[-1] == match[-1])) { start--; match--; matchLength++; }  /* catch up */
-            offset_2 = offset_1; offset_1 = (U32)(offset - ZSTD_REP_MOVE);
+            offset_2 = offset_1; offset_1 = (U32)OFFBASE_TO_OFFSET(offBase);
         }
 
         /* store sequence */
 _storeSequence:
-        {   size_t const litLength = start - anchor;
-            ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offset, matchLength-MINMATCH);
+        {   size_t const litLength = (size_t)(start - anchor);
+            ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offBase, matchLength);
             anchor = ip = start + matchLength;
         }
+        if (ms->lazySkipping) {
+            /* We've found a match, disable lazy skipping mode, and refill the hash cache. */
+            if (searchMethod == search_rowHash) {
+                ZSTD_row_fillHashCache(ms, base, rowLog, mls, ms->nextToUpdate, ilimit);
+            }
+            ms->lazySkipping = 0;
+        }
 
         /* check immediate repcode */
         while (ip <= ilimit) {
@@ -1357,13 +2113,14 @@ _storeSequence:
             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;
                 matchLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
-                offset = offset_2; offset_2 = offset_1; offset_1 = (U32)offset;   /* swap offset history */
-                ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, matchLength-MINMATCH);
+                offBase = offset_2; offset_2 = offset_1; offset_1 = (U32)offBase;   /* swap offset history */
+                ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, matchLength);
                 ip += matchLength;
                 anchor = ip;
                 continue;   /* faster when present ... (?) */
@@ -1378,8 +2135,9 @@ _storeSequence:
     /* Return the last literals size */
     return (size_t)(iend - anchor);
 }
+#endif /* build exclusions */
 
-
+#ifndef ZSTD_EXCLUDE_GREEDY_BLOCK_COMPRESSOR
 size_t ZSTD_compressBlock_greedy_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
@@ -1387,6 +2145,15 @@ size_t ZSTD_compressBlock_greedy_extDict(
     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0);
 }
 
+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);
+}
+#endif
+
+#ifndef ZSTD_EXCLUDE_LAZY_BLOCK_COMPRESSOR
 size_t ZSTD_compressBlock_lazy_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
@@ -1395,6 +2162,16 @@ size_t ZSTD_compressBlock_lazy_extDict(
     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1);
 }
 
+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);
+}
+#endif
+
+#ifndef ZSTD_EXCLUDE_LAZY2_BLOCK_COMPRESSOR
 size_t ZSTD_compressBlock_lazy2_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
@@ -1403,6 +2180,15 @@ size_t ZSTD_compressBlock_lazy2_extDict(
     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2);
 }
 
+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);
+}
+#endif
+
+#ifndef ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR
 size_t ZSTD_compressBlock_btlazy2_extDict(
         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
         void const* src, size_t srcSize)
@@ -1410,3 +2196,4 @@ size_t ZSTD_compressBlock_btlazy2_extDict(
 {
     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2);
 }
+#endif