Update Zstandard to Version 1.5.4

Signed-off-by: Tino Reichardt <milky-7zip@mcmilk.de>

C/zstd/zstd_fast.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) Yann Collet, Facebook, Inc.
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  * All rights reserved.
  *
  * This source code is licensed under both the BSD-style license (found in the
@@ -11,8 +11,42 @@
 #include "zstd_compress_internal.h"  /* ZSTD_hashPtr, ZSTD_count, ZSTD_storeSeq */
 #include "zstd_fast.h"
 
+static void ZSTD_fillHashTableForCDict(ZSTD_matchState_t* ms,
+                        const void* const end,
+                        ZSTD_dictTableLoadMethod_e dtlm)
+{
+    const ZSTD_compressionParameters* const cParams = &ms->cParams;
+    U32* const hashTable = ms->hashTable;
+    U32  const hBits = cParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS;
+    U32  const mls = cParams->minMatch;
+    const BYTE* const base = ms->window.base;
+    const BYTE* ip = base + ms->nextToUpdate;
+    const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE;
+    const U32 fastHashFillStep = 3;
 
-void ZSTD_fillHashTable(ZSTD_matchState_t* ms,
+    /* Currently, we always use ZSTD_dtlm_full for filling CDict tables.
+     * Feel free to remove this assert if there's a good reason! */
+    assert(dtlm == ZSTD_dtlm_full);
+
+    /* Always insert every fastHashFillStep position into the hash table.
+     * Insert the other positions if their hash entry is empty.
+     */
+    for ( ; ip + fastHashFillStep < iend + 2; ip += fastHashFillStep) {
+        U32 const curr = (U32)(ip - base);
+        {   size_t const hashAndTag = ZSTD_hashPtr(ip, hBits, mls);
+            ZSTD_writeTaggedIndex(hashTable, hashAndTag, curr);   }
+
+        if (dtlm == ZSTD_dtlm_fast) continue;
+        /* Only load extra positions for ZSTD_dtlm_full */
+        {   U32 p;
+            for (p = 1; p < fastHashFillStep; ++p) {
+                size_t const hashAndTag = ZSTD_hashPtr(ip + p, hBits, mls);
+                if (hashTable[hashAndTag >> ZSTD_SHORT_CACHE_TAG_BITS] == 0) {  /* not yet filled */
+                    ZSTD_writeTaggedIndex(hashTable, hashAndTag, curr + p);
+                }   }   }   }
+}
+
+static void ZSTD_fillHashTableForCCtx(ZSTD_matchState_t* ms,
                         const void* const end,
                         ZSTD_dictTableLoadMethod_e dtlm)
 {
@@ -25,6 +59,10 @@ void ZSTD_fillHashTable(ZSTD_matchState_t* ms,
     const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE;
     const U32 fastHashFillStep = 3;
 
+    /* Currently, we always use ZSTD_dtlm_fast for filling CCtx tables.
+     * Feel free to remove this assert if there's a good reason! */
+    assert(dtlm == ZSTD_dtlm_fast);
+
     /* Always insert every fastHashFillStep position into the hash table.
      * Insert the other positions if their hash entry is empty.
      */
@@ -42,6 +80,18 @@ void ZSTD_fillHashTable(ZSTD_matchState_t* ms,
     }   }   }   }
 }
 
+void ZSTD_fillHashTable(ZSTD_matchState_t* ms,
+                        const void* const end,
+                        ZSTD_dictTableLoadMethod_e dtlm,
+                        ZSTD_tableFillPurpose_e tfp)
+{
+    if (tfp == ZSTD_tfp_forCDict) {
+        ZSTD_fillHashTableForCDict(ms, end, dtlm);
+    } else {
+        ZSTD_fillHashTableForCCtx(ms, end, dtlm);
+    }
+}
+
 
 /**
  * If you squint hard enough (and ignore repcodes), the search operation at any
@@ -117,7 +167,7 @@ ZSTD_compressBlock_fast_noDict_generic(
 
     U32 rep_offset1 = rep[0];
     U32 rep_offset2 = rep[1];
-    U32 offsetSaved = 0;
+    U32 offsetSaved1 = 0, offsetSaved2 = 0;
 
     size_t hash0; /* hash for ip0 */
     size_t hash1; /* hash for ip1 */
@@ -141,8 +191,8 @@ ZSTD_compressBlock_fast_noDict_generic(
     {   U32 const curr = (U32)(ip0 - base);
         U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, curr, cParams->windowLog);
         U32 const maxRep = curr - windowLow;
-        if (rep_offset2 > maxRep) offsetSaved = rep_offset2, rep_offset2 = 0;
-        if (rep_offset1 > maxRep) offsetSaved = rep_offset1, rep_offset1 = 0;
+        if (rep_offset2 > maxRep) offsetSaved2 = rep_offset2, rep_offset2 = 0;
+        if (rep_offset1 > maxRep) offsetSaved1 = rep_offset1, rep_offset1 = 0;
     }
 
     /* start each op */
@@ -180,8 +230,14 @@ _start: /* Requires: ip0 */
             mLength = ip0[-1] == match0[-1];
             ip0 -= mLength;
             match0 -= mLength;
-            offcode = STORE_REPCODE_1;
+            offcode = REPCODE1_TO_OFFBASE;
             mLength += 4;
+
+            /* First write next hash table entry; we've already calculated it.
+             * This write is known to be safe because the ip1 is before the
+             * repcode (ip2). */
+            hashTable[hash1] = (U32)(ip1 - base);
+
             goto _match;
         }
 
@@ -195,6 +251,12 @@ _start: /* Requires: ip0 */
         /* check match at ip[0] */
         if (MEM_read32(ip0) == mval) {
             /* found a match! */
+
+            /* First write next hash table entry; we've already calculated it.
+             * This write is known to be safe because the ip1 == ip0 + 1, so
+             * we know we will resume searching after ip1 */
+            hashTable[hash1] = (U32)(ip1 - base);
+
             goto _offset;
         }
 
@@ -224,6 +286,21 @@ _start: /* Requires: ip0 */
         /* check match at ip[0] */
         if (MEM_read32(ip0) == mval) {
             /* found a match! */
+
+            /* first write next hash table entry; we've already calculated it */
+            if (step <= 4) {
+                /* We need to avoid writing an index into the hash table >= the
+                 * position at which we will pick up our searching after we've
+                 * taken this match.
+                 *
+                 * The minimum possible match has length 4, so the earliest ip0
+                 * can be after we take this match will be the current ip0 + 4.
+                 * ip1 is ip0 + step - 1. If ip1 is >= ip0 + 4, we can't safely
+                 * write this position.
+                 */
+                hashTable[hash1] = (U32)(ip1 - base);
+            }
+
             goto _offset;
         }
 
@@ -254,9 +331,24 @@ _cleanup:
      * However, it seems to be a meaningful performance hit to try to search
      * them. So let's not. */
 
+    /* When the repcodes are outside of the prefix, we set them to zero before the loop.
+     * When the offsets are still zero, we need to restore them after the block to have a correct
+     * repcode history. If only one offset was invalid, it is easy. The tricky case is when both
+     * offsets were invalid. We need to figure out which offset to refill with.
+     *     - If both offsets are zero they are in the same order.
+     *     - If both offsets are non-zero, we won't restore the offsets from `offsetSaved[12]`.
+     *     - If only one is zero, we need to decide which offset to restore.
+     *         - If rep_offset1 is non-zero, then rep_offset2 must be offsetSaved1.
+     *         - It is impossible for rep_offset2 to be non-zero.
+     *
+     * So if rep_offset1 started invalid (offsetSaved1 != 0) and became valid (rep_offset1 != 0), then
+     * set rep[0] = rep_offset1 and rep[1] = offsetSaved1.
+     */
+    offsetSaved2 = ((offsetSaved1 != 0) && (rep_offset1 != 0)) ? offsetSaved1 : offsetSaved2;
+
     /* save reps for next block */
-    rep[0] = rep_offset1 ? rep_offset1 : offsetSaved;
-    rep[1] = rep_offset2 ? rep_offset2 : offsetSaved;
+    rep[0] = rep_offset1 ? rep_offset1 : offsetSaved1;
+    rep[1] = rep_offset2 ? rep_offset2 : offsetSaved2;
 
     /* Return the last literals size */
     return (size_t)(iend - anchor);
@@ -267,7 +359,7 @@ _offset: /* Requires: ip0, idx */
     match0 = base + idx;
     rep_offset2 = rep_offset1;
     rep_offset1 = (U32)(ip0-match0);
-    offcode = STORE_OFFSET(rep_offset1);
+    offcode = OFFSET_TO_OFFBASE(rep_offset1);
     mLength = 4;
 
     /* Count the backwards match length. */
@@ -287,11 +379,6 @@ _match: /* Requires: ip0, match0, offcode */
     ip0 += mLength;
     anchor = ip0;
 
-    /* write next hash table entry */
-    if (ip1 < ip0) {
-        hashTable[hash1] = (U32)(ip1 - base);
-    }
-
     /* Fill table and check for immediate repcode. */
     if (ip0 <= ilimit) {
         /* Fill Table */
@@ -306,7 +393,7 @@ _match: /* Requires: ip0, match0, offcode */
                 { U32 const tmpOff = rep_offset2; rep_offset2 = rep_offset1; rep_offset1 = tmpOff; } /* swap rep_offset2 <=> rep_offset1 */
                 hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = (U32)(ip0-base);
                 ip0 += rLength;
-                ZSTD_storeSeq(seqStore, 0 /*litLen*/, anchor, iend, STORE_REPCODE_1, rLength);
+                ZSTD_storeSeq(seqStore, 0 /*litLen*/, anchor, iend, REPCODE1_TO_OFFBASE, rLength);
                 anchor = ip0;
                 continue;   /* faster when present (confirmed on gcc-8) ... (?) */
     }   }   }
@@ -380,14 +467,14 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic(
     U32 const stepSize = cParams->targetLength + !(cParams->targetLength);
     const BYTE* const base = ms->window.base;
     const BYTE* const istart = (const BYTE*)src;
-    const BYTE* ip = istart;
+    const BYTE* ip0 = istart;
+    const BYTE* ip1 = ip0 + stepSize; /* we assert below that stepSize >= 1 */
     const BYTE* anchor = istart;
     const U32   prefixStartIndex = ms->window.dictLimit;
     const BYTE* const prefixStart = base + prefixStartIndex;
     const BYTE* const iend = istart + srcSize;
     const BYTE* const ilimit = iend - HASH_READ_SIZE;
     U32 offset_1=rep[0], offset_2=rep[1];
-    U32 offsetSaved = 0;
 
     const ZSTD_matchState_t* const dms = ms->dictMatchState;
     const ZSTD_compressionParameters* const dictCParams = &dms->cParams ;
@@ -397,13 +484,13 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic(
     const BYTE* const dictStart    = dictBase + dictStartIndex;
     const BYTE* const dictEnd      = dms->window.nextSrc;
     const U32 dictIndexDelta       = prefixStartIndex - (U32)(dictEnd - dictBase);
-    const U32 dictAndPrefixLength  = (U32)(ip - prefixStart + dictEnd - dictStart);
-    const U32 dictHLog             = dictCParams->hashLog;
+    const U32 dictAndPrefixLength  = (U32)(istart - prefixStart + dictEnd - dictStart);
+    const U32 dictHBits            = dictCParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS;
 
     /* if a dictionary is still attached, it necessarily means that
      * it is within window size. So we just check it. */
     const U32 maxDistance = 1U << cParams->windowLog;
-    const U32 endIndex = (U32)((size_t)(ip - base) + srcSize);
+    const U32 endIndex = (U32)((size_t)(istart - base) + srcSize);
     assert(endIndex - prefixStartIndex <= maxDistance);
     (void)maxDistance; (void)endIndex;   /* these variables are not used when assert() is disabled */
 
@@ -413,106 +500,155 @@ size_t ZSTD_compressBlock_fast_dictMatchState_generic(
      * when translating a dict index into a local index */
     assert(prefixStartIndex >= (U32)(dictEnd - dictBase));
 
+    if (ms->prefetchCDictTables) {
+        size_t const hashTableBytes = (((size_t)1) << dictCParams->hashLog) * sizeof(U32);
+        PREFETCH_AREA(dictHashTable, hashTableBytes)
+    }
+
     /* init */
     DEBUGLOG(5, "ZSTD_compressBlock_fast_dictMatchState_generic");
-    ip += (dictAndPrefixLength == 0);
+    ip0 += (dictAndPrefixLength == 0);
     /* dictMatchState repCode checks don't currently handle repCode == 0
      * disabling. */
     assert(offset_1 <= dictAndPrefixLength);
     assert(offset_2 <= dictAndPrefixLength);
 
-    /* Main Search Loop */
-    while (ip < ilimit) {   /* < instead of <=, because repcode check at (ip+1) */
+    /* Outer search loop */
+    assert(stepSize >= 1);
+    while (ip1 <= ilimit) {   /* repcode check at (ip0 + 1) is safe because ip0 < ip1 */
         size_t mLength;
-        size_t const h = ZSTD_hashPtr(ip, hlog, mls);
-        U32 const curr = (U32)(ip-base);
-        U32 const matchIndex = hashTable[h];
-        const BYTE* match = base + matchIndex;
-        const U32 repIndex = curr + 1 - offset_1;
-        const BYTE* repMatch = (repIndex < prefixStartIndex) ?
-                               dictBase + (repIndex - dictIndexDelta) :
-                               base + repIndex;
-        hashTable[h] = curr;   /* update hash table */
+        size_t hash0 = ZSTD_hashPtr(ip0, hlog, mls);
 
-        if ( ((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow : ensure repIndex isn't overlapping dict + prefix */
-          && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
-            const BYTE* const repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
-            mLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixStart) + 4;
-            ip++;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_REPCODE_1, mLength);
-        } else if ( (matchIndex <= prefixStartIndex) ) {
-            size_t const dictHash = ZSTD_hashPtr(ip, dictHLog, mls);
-            U32 const dictMatchIndex = dictHashTable[dictHash];
-            const BYTE* dictMatch = dictBase + dictMatchIndex;
-            if (dictMatchIndex <= dictStartIndex ||
-                MEM_read32(dictMatch) != MEM_read32(ip)) {
-                assert(stepSize >= 1);
-                ip += ((ip-anchor) >> kSearchStrength) + stepSize;
-                continue;
-            } else {
-                /* found a dict match */
-                U32 const offset = (U32)(curr-dictMatchIndex-dictIndexDelta);
-                mLength = ZSTD_count_2segments(ip+4, dictMatch+4, iend, dictEnd, prefixStart) + 4;
-                while (((ip>anchor) & (dictMatch>dictStart))
-                     && (ip[-1] == dictMatch[-1])) {
-                    ip--; dictMatch--; mLength++;
+        size_t const dictHashAndTag0 = ZSTD_hashPtr(ip0, dictHBits, mls);
+        U32 dictMatchIndexAndTag = dictHashTable[dictHashAndTag0 >> ZSTD_SHORT_CACHE_TAG_BITS];
+        int dictTagsMatch = ZSTD_comparePackedTags(dictMatchIndexAndTag, dictHashAndTag0);
+
+        U32 matchIndex = hashTable[hash0];
+        U32 curr = (U32)(ip0 - base);
+        size_t step = stepSize;
+        const size_t kStepIncr = 1 << kSearchStrength;
+        const BYTE* nextStep = ip0 + kStepIncr;
+
+        /* Inner search loop */
+        while (1) {
+            const BYTE* match = base + matchIndex;
+            const U32 repIndex = curr + 1 - offset_1;
+            const BYTE* repMatch = (repIndex < prefixStartIndex) ?
+                                   dictBase + (repIndex - dictIndexDelta) :
+                                   base + repIndex;
+            const size_t hash1 = ZSTD_hashPtr(ip1, hlog, mls);
+            size_t const dictHashAndTag1 = ZSTD_hashPtr(ip1, dictHBits, mls);
+            hashTable[hash0] = curr;   /* update hash table */
+
+            if (((U32) ((prefixStartIndex - 1) - repIndex) >=
+                 3) /* intentional underflow : ensure repIndex isn't overlapping dict + prefix */
+                && (MEM_read32(repMatch) == MEM_read32(ip0 + 1))) {
+                const BYTE* const repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
+                mLength = ZSTD_count_2segments(ip0 + 1 + 4, repMatch + 4, iend, repMatchEnd, prefixStart) + 4;
+                ip0++;
+                ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength);
+                break;
+            }
+
+            if (dictTagsMatch) {
+                /* Found a possible dict match */
+                const U32 dictMatchIndex = dictMatchIndexAndTag >> ZSTD_SHORT_CACHE_TAG_BITS;
+                const BYTE* dictMatch = dictBase + dictMatchIndex;
+                if (dictMatchIndex > dictStartIndex &&
+                    MEM_read32(dictMatch) == MEM_read32(ip0)) {
+                    /* To replicate extDict parse behavior, we only use dict matches when the normal matchIndex is invalid */
+                    if (matchIndex <= prefixStartIndex) {
+                        U32 const offset = (U32) (curr - dictMatchIndex - dictIndexDelta);
+                        mLength = ZSTD_count_2segments(ip0 + 4, dictMatch + 4, iend, dictEnd, prefixStart) + 4;
+                        while (((ip0 > anchor) & (dictMatch > dictStart))
+                            && (ip0[-1] == dictMatch[-1])) {
+                            ip0--;
+                            dictMatch--;
+                            mLength++;
+                        } /* catch up */
+                        offset_2 = offset_1;
+                        offset_1 = offset;
+                        ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
+                        break;
+                    }
+                }
+            }
+
+            if (matchIndex > prefixStartIndex && MEM_read32(match) == MEM_read32(ip0)) {
+                /* found a regular match */
+                U32 const offset = (U32) (ip0 - match);
+                mLength = ZSTD_count(ip0 + 4, match + 4, iend) + 4;
+                while (((ip0 > anchor) & (match > prefixStart))
+                       && (ip0[-1] == match[-1])) {
+                    ip0--;
+                    match--;
+                    mLength++;
                 } /* catch up */
                 offset_2 = offset_1;
                 offset_1 = offset;
-                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength);
+                ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
+                break;
             }
-        } else if (MEM_read32(match) != MEM_read32(ip)) {
-            /* it's not a match, and we're not going to check the dictionary */
-            assert(stepSize >= 1);
-            ip += ((ip-anchor) >> kSearchStrength) + stepSize;
-            continue;
-        } else {
-            /* found a regular match */
-            U32 const offset = (U32)(ip-match);
-            mLength = ZSTD_count(ip+4, match+4, iend) + 4;
-            while (((ip>anchor) & (match>prefixStart))
-                 && (ip[-1] == match[-1])) { ip--; match--; mLength++; } /* catch up */
-            offset_2 = offset_1;
-            offset_1 = offset;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength);
-        }
+
+            /* Prepare for next iteration */
+            dictMatchIndexAndTag = dictHashTable[dictHashAndTag1 >> ZSTD_SHORT_CACHE_TAG_BITS];
+            dictTagsMatch = ZSTD_comparePackedTags(dictMatchIndexAndTag, dictHashAndTag1);
+            matchIndex = hashTable[hash1];
+
+            if (ip1 >= nextStep) {
+                step++;
+                nextStep += kStepIncr;
+            }
+            ip0 = ip1;
+            ip1 = ip1 + step;
+            if (ip1 > ilimit) goto _cleanup;
+
+            curr = (U32)(ip0 - base);
+            hash0 = hash1;
+        }   /* end inner search loop */
 
         /* match found */
-        ip += mLength;
-        anchor = ip;
+        assert(mLength);
+        ip0 += mLength;
+        anchor = ip0;
 
-        if (ip <= ilimit) {
+        if (ip0 <= ilimit) {
             /* Fill Table */
             assert(base+curr+2 > istart);  /* check base overflow */
             hashTable[ZSTD_hashPtr(base+curr+2, hlog, mls)] = curr+2;  /* here because curr+2 could be > iend-8 */
-            hashTable[ZSTD_hashPtr(ip-2, hlog, mls)] = (U32)(ip-2-base);
+            hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base);
 
             /* check immediate repcode */
-            while (ip <= ilimit) {
-                U32 const current2 = (U32)(ip-base);
+            while (ip0 <= ilimit) {
+                U32 const current2 = (U32)(ip0-base);
                 U32 const repIndex2 = current2 - offset_2;
                 const BYTE* repMatch2 = repIndex2 < prefixStartIndex ?
                         dictBase - dictIndexDelta + repIndex2 :
                         base + repIndex2;
                 if ( ((U32)((prefixStartIndex-1) - (U32)repIndex2) >= 3 /* intentional overflow */)
-                   && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
+                   && (MEM_read32(repMatch2) == MEM_read32(ip0))) {
                     const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
-                    size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
+                    size_t const repLength2 = ZSTD_count_2segments(ip0+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
                     U32 tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset;   /* swap offset_2 <=> offset_1 */
-                    ZSTD_storeSeq(seqStore, 0, anchor, iend, STORE_REPCODE_1, repLength2);
-                    hashTable[ZSTD_hashPtr(ip, hlog, mls)] = current2;
-                    ip += repLength2;
-                    anchor = ip;
+                    ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
+                    hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = current2;
+                    ip0 += repLength2;
+                    anchor = ip0;
                     continue;
                 }
                 break;
             }
         }
+
+        /* Prepare for next iteration */
+        assert(ip0 == anchor);
+        ip1 = ip0 + stepSize;
     }
 
+_cleanup:
     /* save reps for next block */
-    rep[0] = offset_1 ? offset_1 : offsetSaved;
-    rep[1] = offset_2 ? offset_2 : offsetSaved;
+    rep[0] = offset_1;
+    rep[1] = offset_2;
 
     /* Return the last literals size */
     return (size_t)(iend - anchor);
@@ -553,11 +689,10 @@ static size_t ZSTD_compressBlock_fast_extDict_generic(
     U32* const hashTable = ms->hashTable;
     U32 const hlog = cParams->hashLog;
     /* support stepSize of 0 */
-    U32 const stepSize = cParams->targetLength + !(cParams->targetLength);
+    size_t const stepSize = cParams->targetLength + !(cParams->targetLength) + 1;
     const BYTE* const base = ms->window.base;
     const BYTE* const dictBase = ms->window.dictBase;
     const BYTE* const istart = (const BYTE*)src;
-    const BYTE* ip = istart;
     const BYTE* anchor = istart;
     const U32   endIndex = (U32)((size_t)(istart - base) + srcSize);
     const U32   lowLimit = ZSTD_getLowestMatchIndex(ms, endIndex, cParams->windowLog);
@@ -570,6 +705,28 @@ static size_t ZSTD_compressBlock_fast_extDict_generic(
     const BYTE* const iend = istart + srcSize;
     const BYTE* const ilimit = iend - 8;
     U32 offset_1=rep[0], offset_2=rep[1];
+    U32 offsetSaved1 = 0, offsetSaved2 = 0;
+
+    const BYTE* ip0 = istart;
+    const BYTE* ip1;
+    const BYTE* ip2;
+    const BYTE* ip3;
+    U32 current0;
+
+
+    size_t hash0; /* hash for ip0 */
+    size_t hash1; /* hash for ip1 */
+    U32 idx; /* match idx for ip0 */
+    const BYTE* idxBase; /* base pointer for idx */
+
+    U32 offcode;
+    const BYTE* match0;
+    size_t mLength;
+    const BYTE* matchEnd = 0; /* initialize to avoid warning, assert != 0 later */
+
+    size_t step;
+    const BYTE* nextStep;
+    const size_t kStepIncr = (1 << (kSearchStrength - 1));
 
     (void)hasStep; /* not currently specialized on whether it's accelerated */
 
@@ -579,75 +736,202 @@ static size_t ZSTD_compressBlock_fast_extDict_generic(
     if (prefixStartIndex == dictStartIndex)
         return ZSTD_compressBlock_fast(ms, seqStore, rep, src, srcSize);
 
-    /* Search Loop */
-    while (ip < ilimit) {  /* < instead of <=, because (ip+1) */
-        const size_t h = ZSTD_hashPtr(ip, hlog, mls);
-        const U32    matchIndex = hashTable[h];
-        const BYTE* const matchBase = matchIndex < prefixStartIndex ? dictBase : base;
-        const BYTE*  match = matchBase + matchIndex;
-        const U32    curr = (U32)(ip-base);
-        const U32    repIndex = curr + 1 - offset_1;
-        const BYTE* const repBase = repIndex < prefixStartIndex ? dictBase : base;
-        const BYTE* const repMatch = repBase + repIndex;
-        hashTable[h] = curr;   /* update hash table */
-        DEBUGLOG(7, "offset_1 = %u , curr = %u", offset_1, curr);
+    {   U32 const curr = (U32)(ip0 - base);
+        U32 const maxRep = curr - dictStartIndex;
+        if (offset_2 >= maxRep) offsetSaved2 = offset_2, offset_2 = 0;
+        if (offset_1 >= maxRep) offsetSaved1 = offset_1, offset_1 = 0;
+    }
 
-        if ( ( ((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow */
-             & (offset_1 <= curr+1 - dictStartIndex) ) /* note: we are searching at curr+1 */
-           && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
-            const BYTE* const repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
-            size_t const rLength = ZSTD_count_2segments(ip+1 +4, repMatch +4, iend, repMatchEnd, prefixStart) + 4;
-            ip++;
-            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_REPCODE_1, rLength);
-            ip += rLength;
-            anchor = ip;
-        } else {
-            if ( (matchIndex < dictStartIndex) ||
-                 (MEM_read32(match) != MEM_read32(ip)) ) {
-                assert(stepSize >= 1);
-                ip += ((ip-anchor) >> kSearchStrength) + stepSize;
-                continue;
+    /* start each op */
+_start: /* Requires: ip0 */
+
+    step = stepSize;
+    nextStep = ip0 + kStepIncr;
+
+    /* calculate positions, ip0 - anchor == 0, so we skip step calc */
+    ip1 = ip0 + 1;
+    ip2 = ip0 + step;
+    ip3 = ip2 + 1;
+
+    if (ip3 >= ilimit) {
+        goto _cleanup;
+    }
+
+    hash0 = ZSTD_hashPtr(ip0, hlog, mls);
+    hash1 = ZSTD_hashPtr(ip1, hlog, mls);
+
+    idx = hashTable[hash0];
+    idxBase = idx < prefixStartIndex ? dictBase : base;
+
+    do {
+        {   /* load repcode match for ip[2] */
+            U32 const current2 = (U32)(ip2 - base);
+            U32 const repIndex = current2 - offset_1;
+            const BYTE* const repBase = repIndex < prefixStartIndex ? dictBase : base;
+            U32 rval;
+            if ( ((U32)(prefixStartIndex - repIndex) >= 4) /* intentional underflow */
+                 & (offset_1 > 0) ) {
+                rval = MEM_read32(repBase + repIndex);
+            } else {
+                rval = MEM_read32(ip2) ^ 1; /* guaranteed to not match. */
             }
-            {   const BYTE* const matchEnd = matchIndex < prefixStartIndex ? dictEnd : iend;
-                const BYTE* const lowMatchPtr = matchIndex < prefixStartIndex ? dictStart : prefixStart;
-                U32 const offset = curr - matchIndex;
-                size_t mLength = ZSTD_count_2segments(ip+4, match+4, iend, matchEnd, prefixStart) + 4;
-                while (((ip>anchor) & (match>lowMatchPtr)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; }   /* catch up */
-                offset_2 = offset_1; offset_1 = offset;  /* update offset history */
-                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, STORE_OFFSET(offset), mLength);
-                ip += mLength;
-                anchor = ip;
+
+            /* write back hash table entry */
+            current0 = (U32)(ip0 - base);
+            hashTable[hash0] = current0;
+
+            /* check repcode at ip[2] */
+            if (MEM_read32(ip2) == rval) {
+                ip0 = ip2;
+                match0 = repBase + repIndex;
+                matchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
+                assert((match0 != prefixStart) & (match0 != dictStart));
+                mLength = ip0[-1] == match0[-1];
+                ip0 -= mLength;
+                match0 -= mLength;
+                offcode = REPCODE1_TO_OFFBASE;
+                mLength += 4;
+                goto _match;
         }   }
 
-        if (ip <= ilimit) {
-            /* Fill Table */
-            hashTable[ZSTD_hashPtr(base+curr+2, hlog, mls)] = curr+2;
-            hashTable[ZSTD_hashPtr(ip-2, hlog, mls)] = (U32)(ip-2-base);
-            /* check immediate repcode */
-            while (ip <= ilimit) {
-                U32 const current2 = (U32)(ip-base);
-                U32 const repIndex2 = current2 - offset_2;
-                const BYTE* const repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2;
-                if ( (((U32)((prefixStartIndex-1) - repIndex2) >= 3) & (offset_2 <= curr - dictStartIndex))  /* intentional overflow */
-                   && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
-                    const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
-                    size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
-                    { U32 const tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; }  /* swap offset_2 <=> offset_1 */
-                    ZSTD_storeSeq(seqStore, 0 /*litlen*/, anchor, iend, STORE_REPCODE_1, repLength2);
-                    hashTable[ZSTD_hashPtr(ip, hlog, mls)] = current2;
-                    ip += repLength2;
-                    anchor = ip;
-                    continue;
-                }
-                break;
-    }   }   }
+        {   /* load match for ip[0] */
+            U32 const mval = idx >= dictStartIndex ?
+                    MEM_read32(idxBase + idx) :
+                    MEM_read32(ip0) ^ 1; /* guaranteed not to match */
+
+            /* check match at ip[0] */
+            if (MEM_read32(ip0) == mval) {
+                /* found a match! */
+                goto _offset;
+        }   }
+
+        /* lookup ip[1] */
+        idx = hashTable[hash1];
+        idxBase = idx < prefixStartIndex ? dictBase : base;
+
+        /* hash ip[2] */
+        hash0 = hash1;
+        hash1 = ZSTD_hashPtr(ip2, hlog, mls);
+
+        /* advance to next positions */
+        ip0 = ip1;
+        ip1 = ip2;
+        ip2 = ip3;
+
+        /* write back hash table entry */
+        current0 = (U32)(ip0 - base);
+        hashTable[hash0] = current0;
+
+        {   /* load match for ip[0] */
+            U32 const mval = idx >= dictStartIndex ?
+                    MEM_read32(idxBase + idx) :
+                    MEM_read32(ip0) ^ 1; /* guaranteed not to match */
+
+            /* check match at ip[0] */
+            if (MEM_read32(ip0) == mval) {
+                /* found a match! */
+                goto _offset;
+        }   }
+
+        /* lookup ip[1] */
+        idx = hashTable[hash1];
+        idxBase = idx < prefixStartIndex ? dictBase : base;
+
+        /* hash ip[2] */
+        hash0 = hash1;
+        hash1 = ZSTD_hashPtr(ip2, hlog, mls);
+
+        /* advance to next positions */
+        ip0 = ip1;
+        ip1 = ip2;
+        ip2 = ip0 + step;
+        ip3 = ip1 + step;
+
+        /* calculate step */
+        if (ip2 >= nextStep) {
+            step++;
+            PREFETCH_L1(ip1 + 64);
+            PREFETCH_L1(ip1 + 128);
+            nextStep += kStepIncr;
+        }
+    } while (ip3 < ilimit);
+
+_cleanup:
+    /* Note that there are probably still a couple positions we could search.
+     * However, it seems to be a meaningful performance hit to try to search
+     * them. So let's not. */
+
+    /* If offset_1 started invalid (offsetSaved1 != 0) and became valid (offset_1 != 0),
+     * rotate saved offsets. See comment in ZSTD_compressBlock_fast_noDict for more context. */
+    offsetSaved2 = ((offsetSaved1 != 0) && (offset_1 != 0)) ? offsetSaved1 : offsetSaved2;
 
     /* save reps for next block */
-    rep[0] = offset_1;
-    rep[1] = offset_2;
+    rep[0] = offset_1 ? offset_1 : offsetSaved1;
+    rep[1] = offset_2 ? offset_2 : offsetSaved2;
 
     /* Return the last literals size */
     return (size_t)(iend - anchor);
+
+_offset: /* Requires: ip0, idx, idxBase */
+
+    /* Compute the offset code. */
+    {   U32 const offset = current0 - idx;
+        const BYTE* const lowMatchPtr = idx < prefixStartIndex ? dictStart : prefixStart;
+        matchEnd = idx < prefixStartIndex ? dictEnd : iend;
+        match0 = idxBase + idx;
+        offset_2 = offset_1;
+        offset_1 = offset;
+        offcode = OFFSET_TO_OFFBASE(offset);
+        mLength = 4;
+
+        /* Count the backwards match length. */
+        while (((ip0>anchor) & (match0>lowMatchPtr)) && (ip0[-1] == match0[-1])) {
+            ip0--;
+            match0--;
+            mLength++;
+    }   }
+
+_match: /* Requires: ip0, match0, offcode, matchEnd */
+
+    /* Count the forward length. */
+    assert(matchEnd != 0);
+    mLength += ZSTD_count_2segments(ip0 + mLength, match0 + mLength, iend, matchEnd, prefixStart);
+
+    ZSTD_storeSeq(seqStore, (size_t)(ip0 - anchor), anchor, iend, offcode, mLength);
+
+    ip0 += mLength;
+    anchor = ip0;
+
+    /* write next hash table entry */
+    if (ip1 < ip0) {
+        hashTable[hash1] = (U32)(ip1 - base);
+    }
+
+    /* Fill table and check for immediate repcode. */
+    if (ip0 <= ilimit) {
+        /* Fill Table */
+        assert(base+current0+2 > istart);  /* check base overflow */
+        hashTable[ZSTD_hashPtr(base+current0+2, hlog, mls)] = current0+2;  /* here because current+2 could be > iend-8 */
+        hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base);
+
+        while (ip0 <= ilimit) {
+            U32 const repIndex2 = (U32)(ip0-base) - offset_2;
+            const BYTE* const repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2;
+            if ( (((U32)((prefixStartIndex-1) - repIndex2) >= 3) & (offset_2 > 0))  /* intentional underflow */
+                 && (MEM_read32(repMatch2) == MEM_read32(ip0)) ) {
+                const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
+                size_t const repLength2 = ZSTD_count_2segments(ip0+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
+                { U32 const tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; }  /* swap offset_2 <=> offset_1 */
+                ZSTD_storeSeq(seqStore, 0 /*litlen*/, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
+                hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = (U32)(ip0-base);
+                ip0 += repLength2;
+                anchor = ip0;
+                continue;
+            }
+            break;
+    }   }
+
+    goto _start;
 }
 
 ZSTD_GEN_FAST_FN(extDict, 4, 0)
@@ -660,6 +944,7 @@ size_t ZSTD_compressBlock_fast_extDict(
         void const* src, size_t srcSize)
 {
     U32 const mls = ms->cParams.minMatch;
+    assert(ms->dictMatchState == NULL);
     switch(mls)
     {
     default: /* includes case 3 */