Merge pull request #416 from lz4/newopt

Improve Optimal parser

6 files changed, 415 insertions, 391 deletions

diff --git a/lib/README.md b/lib/README.md
index 7082fe3..fc5d4e9 100644
--- a/lib/README.md
+++ b/lib/README.md
@@ -15,7 +15,7 @@ They generate and decode data using [LZ4 block format].
 
 For more compression ratio at the cost of compression speed,
 the High Compression variant called **lz4hc** is available.
-Add files **`lz4hc.c`** and **`lz4hc.h`**.
+Add files **`lz4hc.c`**, **`lz4hc.h`** and **`lz4opt.h`**.
 The variant still depends on regular `lib/lz4.*` source files.
 
 
diff --git a/lib/lz4.c b/lib/lz4.c
index 6157285..cc76eba 100644
--- a/lib/lz4.c
+++ b/lib/lz4.c
@@ -323,15 +323,24 @@ static const int LZ4_minLength = (MFLIMIT+1);
 /*-************************************
 *  Error detection
 **************************************/
+#if defined(LZ4_DEBUG) && (LZ4_DEBUG>=1)
+#  include <assert.h>
+#else
+#  ifndef assert
+#    define assert(condition) ((void)0)
+#  endif
+#endif
+
 #define LZ4_STATIC_ASSERT(c)   { enum { LZ4_static_assert = 1/(int)(!!(c)) }; }   /* use only *after* variable declarations */
 
 #if defined(LZ4_DEBUG) && (LZ4_DEBUG>=2)
 #  include <stdio.h>
-#  define DEBUGLOG(l, ...) {                          \
-                if (l<=LZ4_DEBUG) {                   \
-                    fprintf(stderr, __FILE__ ": ");   \
-                    fprintf(stderr, __VA_ARGS__);     \
-                    fprintf(stderr, " \n");           \
+static int g_debuglog_enable = 1;
+#  define DEBUGLOG(l, ...) {                                  \
+                if ((g_debuglog_enable) && (l<=LZ4_DEBUG)) {  \
+                    fprintf(stderr, __FILE__ ": ");           \
+                    fprintf(stderr, __VA_ARGS__);             \
+                    fprintf(stderr, " \n");                   \
             }   }
 #else
 #  define DEBUGLOG(l, ...)      {}    /* disabled */
@@ -990,6 +999,7 @@ LZ4_stream_t* LZ4_createStream(void)
 
 void LZ4_resetStream (LZ4_stream_t* LZ4_stream)
 {
+    DEBUGLOG(4, "LZ4_resetStream");
     MEM_INIT(LZ4_stream, 0, sizeof(LZ4_stream_t));
 }
 
diff --git a/lib/lz4frame.c b/lib/lz4frame.c
index 3adbdd9..ebd1089 100644
--- a/lib/lz4frame.c
+++ b/lib/lz4frame.c
@@ -322,7 +322,7 @@ size_t LZ4F_compressFrame_usingCDict(void* dstBuffer, size_t dstCapacity,
                                const LZ4F_preferences_t* preferencesPtr)
 {
     LZ4F_cctx_t cctxI;
-    LZ4_stream_t lz4ctx;
+    LZ4_stream_t lz4ctx;   /* pretty large on stack */
     LZ4F_preferences_t prefs;
     LZ4F_compressOptions_t options;
     BYTE* const dstStart = (BYTE*) dstBuffer;
@@ -504,15 +504,15 @@ size_t LZ4F_compressBegin_usingCDict(LZ4F_cctx* cctxPtr,
     cctxPtr->prefs = *preferencesPtr;
 
     /* Ctx Management */
-    {   U32 const tableID = (cctxPtr->prefs.compressionLevel < LZ4HC_CLEVEL_MIN) ? 1 : 2;  /* 0:nothing ; 1:LZ4 table ; 2:HC tables */
-        if (cctxPtr->lz4CtxLevel < tableID) {
+    {   U32 const ctxTypeID = (cctxPtr->prefs.compressionLevel < LZ4HC_CLEVEL_MIN) ? 1 : 2;  /* 0:nothing ; 1:LZ4 table ; 2:HC tables */
+        if (cctxPtr->lz4CtxLevel < ctxTypeID) {
             FREEMEM(cctxPtr->lz4CtxPtr);
             if (cctxPtr->prefs.compressionLevel < LZ4HC_CLEVEL_MIN)
                 cctxPtr->lz4CtxPtr = (void*)LZ4_createStream();
             else
                 cctxPtr->lz4CtxPtr = (void*)LZ4_createStreamHC();
             if (cctxPtr->lz4CtxPtr == NULL) return err0r(LZ4F_ERROR_allocation_failed);
-            cctxPtr->lz4CtxLevel = tableID;
+            cctxPtr->lz4CtxLevel = ctxTypeID;
     }   }
 
     /* Buffer Management */
diff --git a/lib/lz4hc.c b/lib/lz4hc.c
index d7f8d23..60690a0 100644
--- a/lib/lz4hc.c
+++ b/lib/lz4hc.c
@@ -49,6 +49,7 @@
 
 
 /*===    Dependency    ===*/
+#define LZ4_HC_STATIC_LINKING_ONLY
 #include "lz4hc.h"
 
 
@@ -116,54 +117,71 @@ LZ4_FORCE_INLINE void LZ4HC_Insert (LZ4HC_CCtx_internal* hc4, const BYTE* ip)
     hc4->nextToUpdate = target;
 }
 
-
-LZ4_FORCE_INLINE int LZ4HC_InsertAndFindBestMatch (LZ4HC_CCtx_internal* const hc4,   /* Index table will be updated */
-                                               const BYTE* const ip, const BYTE* const iLimit,
-                                               const BYTE** matchpos,
-                                               const int maxNbAttempts)
+/** LZ4HC_countBack() :
+ * @return : negative value, nb of common bytes before ip/match */
+LZ4_FORCE_INLINE
+int LZ4HC_countBack(const BYTE* const ip, const BYTE* const match,
+                    const BYTE* const iMin, const BYTE* const mMin)
 {
-    U16* const chainTable = hc4->chainTable;
-    U32* const HashTable = hc4->hashTable;
-    const BYTE* const base = hc4->base;
-    const BYTE* const dictBase = hc4->dictBase;
-    const U32 dictLimit = hc4->dictLimit;
-    const U32 lowLimit = (hc4->lowLimit + 64 KB > (U32)(ip-base)) ? hc4->lowLimit : (U32)(ip - base) - (64 KB - 1);
-    U32 matchIndex;
-    int nbAttempts = maxNbAttempts;
-    size_t ml = 0;
+    int back=0;
+    while ( (ip+back > iMin)
+         && (match+back > mMin)
+         && (ip[back-1] == match[back-1]))
+            back--;
+    return back;
+}
 
-    /* HC4 match finder */
-    LZ4HC_Insert(hc4, ip);
-    matchIndex = HashTable[LZ4HC_hashPtr(ip)];
+/* LZ4HC_countPattern() :
+ * pattern32 must be a sample of repetitive pattern of length 1, 2 or 4 (but not 3!) */
+static unsigned LZ4HC_countPattern(const BYTE* ip, const BYTE* const iEnd, U32 const pattern32)
+{
+    const BYTE* const iStart = ip;
+    reg_t const pattern = (sizeof(pattern)==8) ? (reg_t)pattern32 + (((reg_t)pattern32) << 32) : pattern32;
+
+    while (likely(ip < iEnd-(sizeof(pattern)-1))) {
+        reg_t const diff = LZ4_read_ARCH(ip) ^ pattern;
+        if (!diff) { ip+=sizeof(pattern); continue; }
+        ip += LZ4_NbCommonBytes(diff);
+        return (unsigned)(ip - iStart);
+    }
 
-    while ((matchIndex>=lowLimit) && (nbAttempts)) {
-        nbAttempts--;
-        if (matchIndex >= dictLimit) {
-            const BYTE* const match = base + matchIndex;
-            if ( (*(match+ml) == *(ip+ml))   /* can be longer */
-               && (LZ4_read32(match) == LZ4_read32(ip)) )
-            {
-                size_t const mlt = LZ4_count(ip+MINMATCH, match+MINMATCH, iLimit) + MINMATCH;
-                if (mlt > ml) { ml = mlt; *matchpos = match; }
-            }
-        } else {
-            const BYTE* const match = dictBase + matchIndex;
-            if (LZ4_read32(match) == LZ4_read32(ip)) {
-                size_t mlt;
-                const BYTE* vLimit = ip + (dictLimit - matchIndex);
-                if (vLimit > iLimit) vLimit = iLimit;
-                mlt = LZ4_count(ip+MINMATCH, match+MINMATCH, vLimit) + MINMATCH;
-                if ((ip+mlt == vLimit) && (vLimit < iLimit))
-                    mlt += LZ4_count(ip+mlt, base+dictLimit, iLimit);
-                if (mlt > ml) { ml = mlt; *matchpos = base + matchIndex; }   /* virtual matchpos */
-            }
+    if (LZ4_isLittleEndian()) {
+        reg_t patternByte = pattern;
+        while ((ip<iEnd) && (*ip == (BYTE)patternByte)) {
+            ip++; patternByte >>= 8;
+        }
+    } else {  /* big endian */
+        U32 bitOffset = (sizeof(pattern)*8) - 8;
+        while (ip < iEnd) {
+            BYTE const byte = (BYTE)(pattern >> bitOffset);
+            if (*ip != byte) break;
+            ip ++; bitOffset -= 8;
         }
-        matchIndex -= DELTANEXTU16(chainTable, matchIndex);
     }
 
-    return (int)ml;
+    return (unsigned)(ip - iStart);
+}
+
+/* LZ4HC_reverseCountPattern() :
+ * pattern must be a sample of repetitive pattern of length 1, 2 or 4 (but not 3!)
+ * read using natural platform endianess */
+static unsigned LZ4HC_reverseCountPattern(const BYTE* ip, const BYTE* const iLow, U32 pattern)
+{
+    const BYTE* const iStart = ip;
+
+    while (likely(ip >= iLow+4)) {
+        if (LZ4_read32(ip-4) != pattern) break;
+        ip -= 4;
+    }
+    {   const BYTE* bytePtr = (const BYTE*)(&pattern) + 3; /* works for any endianess */
+        while (likely(ip>iLow)) {
+            if (ip[-1] != *bytePtr) break;
+            ip--; bytePtr--;
+    }   }
+    return (unsigned)(iStart - ip);
 }
 
+typedef enum { rep_untested, rep_not, rep_confirmed } repeat_state_e;
 
 LZ4_FORCE_INLINE int LZ4HC_InsertAndGetWiderMatch (
     LZ4HC_CCtx_internal* hc4,
@@ -180,60 +198,117 @@ LZ4_FORCE_INLINE int LZ4HC_InsertAndGetWiderMatch (
     const BYTE* const base = hc4->base;
     const U32 dictLimit = hc4->dictLimit;
     const BYTE* const lowPrefixPtr = base + dictLimit;
-    const U32 lowLimit = (hc4->lowLimit + 64 KB > (U32)(ip-base)) ? hc4->lowLimit : (U32)(ip - base) - (64 KB - 1);
+    const U32 lowLimit = (hc4->lowLimit + 64 KB > (U32)(ip-base)) ? hc4->lowLimit : (U32)(ip - base) - MAX_DISTANCE;
     const BYTE* const dictBase = hc4->dictBase;
     int const delta = (int)(ip-iLowLimit);
     int nbAttempts = maxNbAttempts;
+    U32 const pattern = LZ4_read32(ip);
     U32 matchIndex;
+    repeat_state_e repeat = rep_untested;
+    size_t srcPatternLength = 0;
 
-
+    DEBUGLOG(7, "LZ4HC_InsertAndGetWiderMatch");
     /* First Match */
     LZ4HC_Insert(hc4, ip);
     matchIndex = HashTable[LZ4HC_hashPtr(ip)];
+    DEBUGLOG(7, "First match at index %u / %u (lowLimit)",
+                matchIndex, lowLimit);
 
     while ((matchIndex>=lowLimit) && (nbAttempts)) {
+        DEBUGLOG(7, "remaining attempts : %i", nbAttempts);
         nbAttempts--;
         if (matchIndex >= dictLimit) {
             const BYTE* const matchPtr = base + matchIndex;
             if (*(iLowLimit + longest) == *(matchPtr - delta + longest)) {
-                if (LZ4_read32(matchPtr) == LZ4_read32(ip)) {
+                if (LZ4_read32(matchPtr) == pattern) {
                     int mlt = MINMATCH + LZ4_count(ip+MINMATCH, matchPtr+MINMATCH, iHighLimit);
+    #if 0
+                    /* more generic but unfortunately slower on clang */
+                    int const back = LZ4HC_countBack(ip, matchPtr, iLowLimit, lowPrefixPtr);
+    #else
                     int back = 0;
-
                     while ( (ip+back > iLowLimit)
                          && (matchPtr+back > lowPrefixPtr)
                          && (ip[back-1] == matchPtr[back-1])) {
                             back--;
                     }
-
+    #endif
                     mlt -= back;
 
                     if (mlt > longest) {
                         longest = mlt;
                         *matchpos = matchPtr+back;
                         *startpos = ip+back;
-            }   }   }
-        } else {
+                }   }
+            }
+        } else {   /* matchIndex < dictLimit */
             const BYTE* const matchPtr = dictBase + matchIndex;
-            if (LZ4_read32(matchPtr) == LZ4_read32(ip)) {
+            if (LZ4_read32(matchPtr) == pattern) {
                 int mlt;
-                int back=0;
+                int back = 0;
                 const BYTE* vLimit = ip + (dictLimit - matchIndex);
                 if (vLimit > iHighLimit) vLimit = iHighLimit;
                 mlt = LZ4_count(ip+MINMATCH, matchPtr+MINMATCH, vLimit) + MINMATCH;
                 if ((ip+mlt == vLimit) && (vLimit < iHighLimit))
                     mlt += LZ4_count(ip+mlt, base+dictLimit, iHighLimit);
-                while ((ip+back > iLowLimit) && (matchIndex+back > lowLimit) && (ip[back-1] == matchPtr[back-1])) back--;
+                while ( (ip+back > iLowLimit)
+                     && (matchIndex+back > lowLimit)
+                     && (ip[back-1] == matchPtr[back-1]))
+                        back--;
                 mlt -= back;
-                if (mlt > longest) { longest = mlt; *matchpos = base + matchIndex + back; *startpos = ip+back; }
-            }
-        }
-        matchIndex -= DELTANEXTU16(chainTable, matchIndex);
-    }
+                if (mlt > longest) {
+                    longest = mlt;
+                    *matchpos = base + matchIndex + back;
+                    *startpos = ip + back;
+        }   }   }
+
+        {   U32 const nextOffset = DELTANEXTU16(chainTable, matchIndex);
+            matchIndex -= nextOffset;
+            if (nextOffset==1) {
+                /* may be a repeated pattern */
+                if (repeat == rep_untested) {
+                    if ( ((pattern & 0xFFFF) == (pattern >> 16))
+                      &  ((pattern & 0xFF)   == (pattern >> 24)) ) {
+                        repeat = rep_confirmed;
+                        srcPatternLength = LZ4HC_countPattern(ip+4, iHighLimit, pattern) + 4;
+                    } else {
+                        repeat = rep_not;
+                }   }
+                if ( (repeat == rep_confirmed)
+                  && (matchIndex >= dictLimit) ) {   /* same segment only */
+                    const BYTE* const matchPtr = base + matchIndex;
+                    if (LZ4_read32(matchPtr) == pattern) {  /* good candidate */
+                        size_t const forwardPatternLength = LZ4HC_countPattern(matchPtr+sizeof(pattern), iHighLimit, pattern) + sizeof(pattern);
+                        const BYTE* const maxLowPtr = (lowPrefixPtr + MAX_DISTANCE >= ip) ? lowPrefixPtr : ip - MAX_DISTANCE;
+                        size_t const backLength = LZ4HC_reverseCountPattern(matchPtr, maxLowPtr, pattern);
+                        size_t const currentSegmentLength = backLength + forwardPatternLength;
+
+                        if ( (currentSegmentLength >= srcPatternLength)   /* current pattern segment large enough to contain full srcPatternLength */
+                          && (forwardPatternLength <= srcPatternLength) ) { /* haven't reached this position yet */
+                            matchIndex += (U32)forwardPatternLength - (U32)srcPatternLength;  /* best position, full pattern, might be followed by more match */
+                        } else {
+                            matchIndex -= (U32)backLength;   /* let's go to farthest segment position, will find a match of length currentSegmentLength + maybe some back */
+                        }
+        }   }   }   }
+    }  /* while ((matchIndex>=lowLimit) && (nbAttempts)) */
 
     return longest;
 }
 
+LZ4_FORCE_INLINE
+int LZ4HC_InsertAndFindBestMatch(LZ4HC_CCtx_internal* const hc4,   /* Index table will be updated */
+                                 const BYTE* const ip, const BYTE* const iLimit,
+                                 const BYTE** matchpos,
+                                 const int maxNbAttempts)
+{
+    const BYTE* uselessPtr = ip;
+    /* note : LZ4HC_InsertAndGetWiderMatch() is able to modify the starting position of a match (*startpos),
+     * but this won't be the case here, as we define iLowLimit==ip,
+     * so LZ4HC_InsertAndGetWiderMatch() won't be allowed to search past ip */
+    return LZ4HC_InsertAndGetWiderMatch(hc4, ip, ip, iLimit, MINMATCH-1, matchpos, &uselessPtr, maxNbAttempts);
+}
+
+
 
 typedef enum {
     noLimit = 0,
@@ -241,10 +316,6 @@ typedef enum {
     limitedDestSize = 2,
 } limitedOutput_directive;
 
-#ifndef LZ4HC_DEBUG
-#  define LZ4HC_DEBUG 0
-#endif
-
 /* LZ4HC_encodeSequence() :
  * @return : 0 if ok,
  *           1 if buffer issue detected */
@@ -260,9 +331,21 @@ LZ4_FORCE_INLINE int LZ4HC_encodeSequence (
     size_t length;
     BYTE* const token = (*op)++;
 
-#if LZ4HC_DEBUG
-    printf("literal : %u  --  match : %u  --  offset : %u\n",
-           (U32)(*ip - *anchor), (U32)matchLength, (U32)(*ip-match));
+#if defined(LZ4_DEBUG) && (LZ4_DEBUG >= 2)
+    static const BYTE* start = NULL;
+    static U32 totalCost = 0;
+    U32 const pos = (start==NULL) ? 0 : (U32)(*anchor - start);
+    U32 const ll = (U32)(*ip - *anchor);
+    U32 const llAdd = (ll>=15) ? ((ll-15) / 255) + 1 : 0;
+    U32 const mlAdd = (matchLength>=19) ? ((matchLength-19) / 255) + 1 : 0;
+    U32 const cost = 1 + llAdd + ll + 2 + mlAdd;
+    if (start==NULL) start = *anchor;  /* only works for single segment */
+    //g_debuglog_enable = (pos >= 2228) & (pos <= 2262);
+    DEBUGLOG(2, "pos:%7u -- literals:%3u, match:%4i, offset:%5u, cost:%3u + %u",
+                pos,
+                (U32)(*ip - *anchor), matchLength, (U32)(*ip-match),
+                cost, totalCost);
+    totalCost += cost;
 #endif
 
     /* Encode Literal length */
@@ -285,6 +368,7 @@ LZ4_FORCE_INLINE int LZ4HC_encodeSequence (
     LZ4_writeLE16(*op, (U16)(*ip-match)); *op += 2;
 
     /* Encode MatchLength */
+    assert(matchLength >= MINMATCH);
     length = (size_t)(matchLength - MINMATCH);
     if ((limit) && (*op + (length >> 8) + (1 + LASTLITERALS) > oend)) return 1;   /* Check output limit */
     if (length >= ML_MASK) {
@@ -344,16 +428,13 @@ static int LZ4HC_compress_hashChain (
     if (limit == limitedDestSize && maxOutputSize < 1) return 0;         /* Impossible to store anything */
     if ((U32)inputSize > (U32)LZ4_MAX_INPUT_SIZE) return 0;              /* Unsupported input size, too large (or negative) */
 
-    ctx->end += inputSize;
     if (limit == limitedDestSize) oend -= LASTLITERALS;                  /* Hack for support limitations LZ4 decompressor */
     if (inputSize < LZ4_minLength) goto _last_literals;                  /* Input too small, no compression (all literals) */
 
-    ip++;
-
     /* Main Loop */
     while (ip < mflimit) {
-        ml = LZ4HC_InsertAndFindBestMatch (ctx, ip, matchlimit, (&ref), maxNbAttempts);
-        if (!ml) { ip++; continue; }
+        ml = LZ4HC_InsertAndFindBestMatch (ctx, ip, matchlimit, &ref, maxNbAttempts);
+        if (ml<MINMATCH) { ip++; continue; }
 
         /* saved, in case we would skip too much */
         start0 = ip;
@@ -527,14 +608,6 @@ _dest_overflow:
     return 0;
 }
 
-static int LZ4HC_getSearchNum(int compressionLevel)
-{
-    switch (compressionLevel) {
-        default: return 0; /* unused */
-        case 11: return 128;
-        case 12: return 1<<10;
-    }
-}
 
 static int LZ4HC_compress_generic (
     LZ4HC_CCtx_internal* const ctx,
@@ -546,21 +619,19 @@ static int LZ4HC_compress_generic (
     limitedOutput_directive limit
     )
 {
-    if (cLevel < 1) cLevel = LZ4HC_CLEVEL_DEFAULT;   /* note : convention is different from lz4frame, maybe to reconsider */
+    ctx->end += *srcSizePtr;
+    if (cLevel < 1) cLevel = LZ4HC_CLEVEL_DEFAULT;   /* note : convention is different from lz4frame, maybe something to review */
     if (cLevel > 9) {
         if (limit == limitedDestSize) cLevel = 10;
         switch (cLevel) {
             case 10:
-                return LZ4HC_compress_hashChain(ctx, src, dst, srcSizePtr, dstCapacity, 1 << 12, limit);
+                return LZ4HC_compress_hashChain(ctx, src, dst, srcSizePtr, dstCapacity, 1<<12, limit);
             case 11:
-                ctx->searchNum = LZ4HC_getSearchNum(cLevel);
-                return LZ4HC_compress_optimal(ctx, src, dst, *srcSizePtr, dstCapacity, limit, 128, 0);
+                return LZ4HC_compress_optimal(ctx, src, dst, *srcSizePtr, dstCapacity, limit, 512, 128, 0);
             default:
-                cLevel = 12;
                 /* fall-through */
             case 12:
-                ctx->searchNum = LZ4HC_getSearchNum(cLevel);
-                return LZ4HC_compress_optimal(ctx, src, dst, *srcSizePtr, dstCapacity, limit, LZ4_OPT_NUM, 1);
+                return LZ4HC_compress_optimal(ctx, src, dst, *srcSizePtr, dstCapacity, limit, 1<<13, LZ4_OPT_NUM, 1);
         }
     }
     return LZ4HC_compress_hashChain(ctx, src, dst, srcSizePtr, dstCapacity, 1 << (cLevel-1), limit);  /* levels 1-9 */
@@ -596,8 +667,7 @@ int LZ4_compress_HC(const char* src, char* dst, int srcSize, int dstCapacity, in
 }
 
 /* LZ4_compress_HC_destSize() :
- * currently, only compatible with Hash Chain implementation,
- * hence limit compression level to LZ4HC_CLEVEL_OPT_MIN-1*/
+ * only compatible with Hash Chain match finder */
 int LZ4_compress_HC_destSize(void* LZ4HC_Data, const char* source, char* dest, int* sourceSizePtr, int targetDestSize, int cLevel)
 {
     LZ4HC_CCtx_internal* const ctx = &((LZ4_streamHC_t*)LZ4HC_Data)->internal_donotuse;
@@ -624,18 +694,13 @@ void LZ4_resetStreamHC (LZ4_streamHC_t* LZ4_streamHCPtr, int compressionLevel)
 {
     LZ4_STATIC_ASSERT(sizeof(LZ4HC_CCtx_internal) <= sizeof(size_t) * LZ4_STREAMHCSIZE_SIZET);   /* if compilation fails here, LZ4_STREAMHCSIZE must be increased */
     LZ4_streamHCPtr->internal_donotuse.base = NULL;
-    if (compressionLevel > LZ4HC_CLEVEL_MAX) compressionLevel = LZ4HC_CLEVEL_MAX;  /* cap compression level */
-    LZ4_streamHCPtr->internal_donotuse.compressionLevel = compressionLevel;
-    LZ4_streamHCPtr->internal_donotuse.searchNum = LZ4HC_getSearchNum(compressionLevel);
+    LZ4_setCompressionLevel(LZ4_streamHCPtr, compressionLevel);
 }
 
 void LZ4_setCompressionLevel(LZ4_streamHC_t* LZ4_streamHCPtr, int compressionLevel)
 {
-    int const currentCLevel = LZ4_streamHCPtr->internal_donotuse.compressionLevel;
-    int const minCLevel = currentCLevel < LZ4HC_CLEVEL_OPT_MIN ? 1 : LZ4HC_CLEVEL_OPT_MIN;
-    int const maxCLevel = currentCLevel < LZ4HC_CLEVEL_OPT_MIN ? LZ4HC_CLEVEL_OPT_MIN-1 : LZ4HC_CLEVEL_MAX;
-    compressionLevel = MIN(compressionLevel, minCLevel);
-    compressionLevel = MAX(compressionLevel, maxCLevel);
+    if (compressionLevel < 1) compressionLevel = 1;
+    if (compressionLevel > LZ4HC_CLEVEL_MAX) compressionLevel = LZ4HC_CLEVEL_MAX;
     LZ4_streamHCPtr->internal_donotuse.compressionLevel = compressionLevel;
 }
 
@@ -648,10 +713,7 @@ int LZ4_loadDictHC (LZ4_streamHC_t* LZ4_streamHCPtr, const char* dictionary, int
     }
     LZ4HC_init (ctxPtr, (const BYTE*)dictionary);
     ctxPtr->end = (const BYTE*)dictionary + dictSize;
-    if (ctxPtr->compressionLevel >= LZ4HC_CLEVEL_OPT_MIN)
-        LZ4HC_updateBinTree(ctxPtr, ctxPtr->end - MFLIMIT, ctxPtr->end - LASTLITERALS);
-    else
-        if (dictSize >= 4) LZ4HC_Insert (ctxPtr, ctxPtr->end-3);
+    if (dictSize >= 4) LZ4HC_Insert (ctxPtr, ctxPtr->end-3);
     return dictSize;
 }
 
@@ -660,10 +722,7 @@ int LZ4_loadDictHC (LZ4_streamHC_t* LZ4_streamHCPtr, const char* dictionary, int
 
 static void LZ4HC_setExternalDict(LZ4HC_CCtx_internal* ctxPtr, const BYTE* newBlock)
 {
-    if (ctxPtr->compressionLevel >= LZ4HC_CLEVEL_OPT_MIN)
-        LZ4HC_updateBinTree(ctxPtr, ctxPtr->end - MFLIMIT, ctxPtr->end - LASTLITERALS);
-    else
-        if (ctxPtr->end >= ctxPtr->base + 4) LZ4HC_Insert (ctxPtr, ctxPtr->end-3);   /* Referencing remaining dictionary content */
+    if (ctxPtr->end >= ctxPtr->base + 4) LZ4HC_Insert (ctxPtr, ctxPtr->end-3);   /* Referencing remaining dictionary content */
 
     /* Only one memory segment for extDict, so any previous extDict is lost at this stage */
     ctxPtr->lowLimit  = ctxPtr->dictLimit;
@@ -717,8 +776,6 @@ int LZ4_compress_HC_continue (LZ4_streamHC_t* LZ4_streamHCPtr, const char* src,
 
 int LZ4_compress_HC_continue_destSize (LZ4_streamHC_t* LZ4_streamHCPtr, const char* src, char* dst, int* srcSizePtr, int targetDestSize)
 {
-    LZ4HC_CCtx_internal* const ctxPtr = &LZ4_streamHCPtr->internal_donotuse;
-    if (ctxPtr->compressionLevel >= LZ4HC_CLEVEL_OPT_MIN) LZ4HC_init(ctxPtr, (const BYTE*)src);   /* not compatible with btopt implementation */
     return LZ4_compressHC_continue_generic(LZ4_streamHCPtr, src, dst, srcSizePtr, targetDestSize, limitedDestSize);
 }
 
diff --git a/lib/lz4hc.h b/lib/lz4hc.h
index 66d5636..d791062 100644
--- a/lib/lz4hc.h
+++ b/lib/lz4hc.h
@@ -39,7 +39,7 @@ extern "C" {
 #endif
 
 /* --- Dependency --- */
-/* note : lz4hc is not an independent module, it requires lz4.h/lz4.c for proper compilation */
+/* note : lz4hc requires lz4.h/lz4.c for compilation */
 #include "lz4.h"   /* stddef, LZ4LIB_API, LZ4_DEPRECATED */
 
 
@@ -54,12 +54,12 @@ extern "C" {
  *  Block Compression
  **************************************/
 /*! LZ4_compress_HC() :
- * Compress data from `src` into `dst`, using the more powerful but slower "HC" algorithm.
+ *  Compress data from `src` into `dst`, using the more powerful but slower "HC" algorithm.
  * `dst` must be already allocated.
- * Compression is guaranteed to succeed if `dstCapacity >= LZ4_compressBound(srcSize)` (see "lz4.h")
- * Max supported `srcSize` value is LZ4_MAX_INPUT_SIZE (see "lz4.h")
- * `compressionLevel` : Recommended values are between 4 and 9, although any value between 1 and LZ4HC_CLEVEL_MAX will work.
- *                      Values >LZ4HC_CLEVEL_MAX behave the same as LZ4HC_CLEVEL_MAX.
+ *  Compression is guaranteed to succeed if `dstCapacity >= LZ4_compressBound(srcSize)` (see "lz4.h")
+ *  Max supported `srcSize` value is LZ4_MAX_INPUT_SIZE (see "lz4.h")
+ * `compressionLevel` : any value between 1 and LZ4HC_CLEVEL_MAX will work.
+ *                      Values > LZ4HC_CLEVEL_MAX behave the same as LZ4HC_CLEVEL_MAX.
  * @return : the number of bytes written into 'dst'
  *           or 0 if compression fails.
  */
@@ -72,12 +72,12 @@ LZ4LIB_API int LZ4_compress_HC (const char* src, char* dst, int srcSize, int dst
 
 
 /*! LZ4_compress_HC_extStateHC() :
- * Same as LZ4_compress_HC(), but using an externally allocated memory segment for `state`.
+ *  Same as LZ4_compress_HC(), but using an externally allocated memory segment for `state`.
  * `state` size is provided by LZ4_sizeofStateHC().
- * Memory segment must be aligned on 8-bytes boundaries (which a normal malloc() will do properly).
+ *  Memory segment must be aligned on 8-bytes boundaries (which a normal malloc() should do properly).
  */
-LZ4LIB_API int LZ4_compress_HC_extStateHC(void* state, const char* src, char* dst, int srcSize, int maxDstSize, int compressionLevel);
 LZ4LIB_API int LZ4_sizeofStateHC(void);
+LZ4LIB_API int LZ4_compress_HC_extStateHC(void* state, const char* src, char* dst, int srcSize, int maxDstSize, int compressionLevel);
 
 
 /*-************************************
@@ -87,10 +87,10 @@ LZ4LIB_API int LZ4_sizeofStateHC(void);
  typedef union LZ4_streamHC_u LZ4_streamHC_t;   /* incomplete type (defined later) */
 
 /*! LZ4_createStreamHC() and LZ4_freeStreamHC() :
- * These functions create and release memory for LZ4 HC streaming state.
- * Newly created states are automatically initialized.
- * Existing states can be re-used several times, using LZ4_resetStreamHC().
- * These methods are API and ABI stable, they can be used in combination with a DLL.
+ *  These functions create and release memory for LZ4 HC streaming state.
+ *  Newly created states are automatically initialized.
+ *  Existing states can be re-used several times, using LZ4_resetStreamHC().
+ *  These methods are API and ABI stable, they can be used in combination with a DLL.
  */
 LZ4LIB_API LZ4_streamHC_t* LZ4_createStreamHC(void);
 LZ4LIB_API int             LZ4_freeStreamHC (LZ4_streamHC_t* streamHCPtr);
@@ -123,13 +123,13 @@ LZ4LIB_API int LZ4_saveDictHC (LZ4_streamHC_t* streamHCPtr, char* safeBuffer, in
 */
 
 
- /*-*************************************
+/*-**************************************************************
  * PRIVATE DEFINITIONS :
  * Do not use these definitions.
  * They are exposed to allow static allocation of `LZ4_streamHC_t`.
  * Using these definitions makes the code vulnerable to potential API break when upgrading LZ4
- **************************************/
-#define LZ4HC_DICTIONARY_LOGSIZE 17   /* because of btopt, hc would only need 16 */
+ ****************************************************************/
+#define LZ4HC_DICTIONARY_LOGSIZE 16
 #define LZ4HC_MAXD (1<<LZ4HC_DICTIONARY_LOGSIZE)
 #define LZ4HC_MAXD_MASK (LZ4HC_MAXD - 1)
 
@@ -152,8 +152,7 @@ typedef struct
     uint32_t   dictLimit;       /* below that point, need extDict */
     uint32_t   lowLimit;        /* below that point, no more dict */
     uint32_t   nextToUpdate;    /* index from which to continue dictionary update */
-    uint32_t   searchNum;       /* only for optimal parser */
-    uint32_t   compressionLevel;
+    int        compressionLevel;
 } LZ4HC_CCtx_internal;
 
 #else
@@ -169,13 +168,12 @@ typedef struct
     unsigned int   dictLimit;        /* below that point, need extDict */
     unsigned int   lowLimit;         /* below that point, no more dict */
     unsigned int   nextToUpdate;     /* index from which to continue dictionary update */
-    unsigned int   searchNum;        /* only for optimal parser */
     int            compressionLevel;
 } LZ4HC_CCtx_internal;
 
 #endif
 
-#define LZ4_STREAMHCSIZE       (4*LZ4HC_HASHTABLESIZE + 2*LZ4HC_MAXD + 56) /* 393268 */
+#define LZ4_STREAMHCSIZE       (4*LZ4HC_HASHTABLESIZE + 2*LZ4HC_MAXD + 56) /* 262200 */
 #define LZ4_STREAMHCSIZE_SIZET (LZ4_STREAMHCSIZE / sizeof(size_t))
 union LZ4_streamHC_u {
     size_t table[LZ4_STREAMHCSIZE_SIZET];
@@ -197,7 +195,6 @@ union LZ4_streamHC_u {
 /* see lz4.h LZ4_DISABLE_DEPRECATE_WARNINGS to turn off deprecation warnings */
 
 /* deprecated compression functions */
-/* these functions will trigger warning messages in future releases */
 LZ4LIB_API LZ4_DEPRECATED("use LZ4_compress_HC() instead") int LZ4_compressHC               (const char* source, char* dest, int inputSize);
 LZ4LIB_API LZ4_DEPRECATED("use LZ4_compress_HC() instead") int LZ4_compressHC_limitedOutput (const char* source, char* dest, int inputSize, int maxOutputSize);
 LZ4LIB_API LZ4_DEPRECATED("use LZ4_compress_HC() instead") int LZ4_compressHC2 (const char* source, char* dest, int inputSize, int compressionLevel);
@@ -225,14 +222,15 @@ LZ4LIB_API LZ4_DEPRECATED("use LZ4_resetStreamHC() instead")  int   LZ4_resetStr
 
 #endif /* LZ4_HC_H_19834876238432 */
 
-/*-************************************************
+
+/*-**************************************************
  * !!!!!     STATIC LINKING ONLY     !!!!!
  * Following definitions are considered experimental.
  * They should not be linked from DLL,
  * as there is no guarantee of API stability yet.
  * Prototypes will be promoted to "stable" status
  * after successfull usage in real-life scenarios.
- *************************************************/
+ ***************************************************/
 #ifdef LZ4_HC_STATIC_LINKING_ONLY   /* protection macro */
 #ifndef LZ4_HC_SLO_098092834
 #define LZ4_HC_SLO_098092834
@@ -258,17 +256,13 @@ int LZ4_compress_HC_destSize(void* LZ4HC_Data,
  * @return : the number of bytes written into 'dst'
  *           or 0 if compression fails.
  * `srcSizePtr` : value will be updated to indicate how much bytes were read from `src`.
- *  Important : due to limitations, this prototype only works well up to cLevel < LZ4HC_CLEVEL_OPT_MIN
- *              beyond that level, compression performance will be much reduced due to internal incompatibilities
  */
 int LZ4_compress_HC_continue_destSize(LZ4_streamHC_t* LZ4_streamHCPtr,
                             const char* src, char* dst,
                             int* srcSizePtr, int targetDstSize);
 
 /*! LZ4_setCompressionLevel() : v1.8.0 (experimental)
- *  It's possible to change compression level after LZ4_resetStreamHC(), between 2 invocations of LZ4_compress_HC_continue*(),
- *  but that requires to stay in the same mode (aka 1-10 or 11-12).
- *  This function ensures this condition.
+ *  It's possible to change compression level between 2 invocations of LZ4_compress_HC_continue*()
  */
 void LZ4_setCompressionLevel(LZ4_streamHC_t* LZ4_streamHCPtr, int compressionLevel);
 
diff --git a/lib/lz4opt.h b/lib/lz4opt.h
index 584dc97..9917851 100644
--- a/lib/lz4opt.h
+++ b/lib/lz4opt.h
@@ -35,12 +35,6 @@
 
 #define LZ4_OPT_NUM   (1<<12)
 
-
-typedef struct {
-    int off;
-    int len;
-} LZ4HC_match_t;
-
 typedef struct {
     int price;
     int off;
@@ -50,317 +44,286 @@ typedef struct {
 
 
 /* price in bytes */
-LZ4_FORCE_INLINE size_t LZ4HC_literalsPrice(size_t litlen)
+LZ4_FORCE_INLINE int LZ4HC_literalsPrice(int const litlen)
 {
-    size_t price = litlen;
-    if (litlen >= (size_t)RUN_MASK)
+    int price = litlen;
+    if (litlen >= (int)RUN_MASK)
         price += 1 + (litlen-RUN_MASK)/255;
     return price;
 }
 
 
 /* requires mlen >= MINMATCH */
-LZ4_FORCE_INLINE size_t LZ4HC_sequencePrice(size_t litlen, size_t mlen)
+LZ4_FORCE_INLINE int LZ4HC_sequencePrice(int litlen, int mlen)
 {
-    size_t price = 2 + 1; /* 16-bit offset + token */
+    int price = 1 + 2 ; /* token + 16-bit offset */
 
     price += LZ4HC_literalsPrice(litlen);
 
-    if (mlen >= (size_t)(ML_MASK+MINMATCH))
-        price+= 1 + (mlen-(ML_MASK+MINMATCH))/255;
+    if (mlen >= (int)(ML_MASK+MINMATCH))
+        price += 1 + (mlen-(ML_MASK+MINMATCH))/255;
 
     return price;
 }
 
 
 /*-*************************************
-*  Binary Tree search
+*  Match finder
 ***************************************/
-LZ4_FORCE_INLINE int LZ4HC_BinTree_InsertAndGetAllMatches (
-    LZ4HC_CCtx_internal* ctx,
-    const BYTE* const ip,
-    const BYTE* const iHighLimit,
-    size_t best_mlen,
-    LZ4HC_match_t* matches,
-    int* matchNum)
-{
-    U16* const chainTable = ctx->chainTable;
-    U32* const HashTable = ctx->hashTable;
-    const BYTE* const base = ctx->base;
-    const U32 dictLimit = ctx->dictLimit;
-    const U32 current = (U32)(ip - base);
-    const U32 lowLimit = (ctx->lowLimit + MAX_DISTANCE > current) ? ctx->lowLimit : current - (MAX_DISTANCE - 1);
-    const BYTE* const dictBase = ctx->dictBase;
-    const BYTE* match;
-    int nbAttempts = ctx->searchNum;
-    int mnum = 0;
-    U16 *ptr0, *ptr1, delta0, delta1;
-    U32 matchIndex;
-    size_t matchLength = 0;
-    U32* HashPos;
-
-    if (ip + MINMATCH > iHighLimit) return 1;
-
-    /* HC4 match finder */
-    HashPos = &HashTable[LZ4HC_hashPtr(ip)];
-    matchIndex = *HashPos;
-    *HashPos = current;
-
-    ptr0 = &DELTANEXTMAXD(current*2+1);
-    ptr1 = &DELTANEXTMAXD(current*2);
-    delta0 = delta1 = (U16)(current - matchIndex);
-
-    while ((matchIndex < current) && (matchIndex>=lowLimit) && (nbAttempts)) {
-        nbAttempts--;
-        if (matchIndex >= dictLimit) {
-            match = base + matchIndex;
-            matchLength = LZ4_count(ip, match, iHighLimit);
-        } else {
-            const BYTE* vLimit = ip + (dictLimit - matchIndex);
-            match = dictBase + matchIndex;
-            if (vLimit > iHighLimit) vLimit = iHighLimit;
-            matchLength = LZ4_count(ip, match, vLimit);
-            if ((ip+matchLength == vLimit) && (vLimit < iHighLimit))
-                matchLength += LZ4_count(ip+matchLength, base+dictLimit, iHighLimit);
-            if (matchIndex+matchLength >= dictLimit)
-                match = base + matchIndex;   /* to prepare for next usage of match[matchLength] */
-        }
-
-        if (matchLength > best_mlen) {
-            best_mlen = matchLength;
-            if (matches) {
-                if (matchIndex >= dictLimit)
-                    matches[mnum].off = (int)(ip - match);
-                else
-                    matches[mnum].off = (int)(ip - (base + matchIndex)); /* virtual matchpos */
-                matches[mnum].len = (int)matchLength;
-                mnum++;
-            }
-            if (best_mlen > LZ4_OPT_NUM) break;
-        }
-
-        if (ip+matchLength >= iHighLimit)   /* equal : no way to know if inf or sup */
-            break;   /* drop , to guarantee consistency ; miss a bit of compression, but other solutions can corrupt the tree */
-
-        DEBUGLOG(6, "ip   :%016llX", (U64)ip);
-        DEBUGLOG(6, "match:%016llX", (U64)match);
-        if (*(ip+matchLength) < *(match+matchLength)) {
-            *ptr0 = delta0;
-            ptr0 = &DELTANEXTMAXD(matchIndex*2);
-            if (*ptr0 == (U16)-1) break;
-            delta0 = *ptr0;
-            delta1 += delta0;
-            matchIndex -= delta0;
-        } else {
-            *ptr1 = delta1;
-            ptr1 = &DELTANEXTMAXD(matchIndex*2+1);
-            if (*ptr1 == (U16)-1) break;
-            delta1 = *ptr1;
-            delta0 += delta1;
-            matchIndex -= delta1;
-        }
-    }
-
-    *ptr0 = (U16)-1;
-    *ptr1 = (U16)-1;
-    if (matchNum) *matchNum = mnum;
-  /*  if (best_mlen > 8) return best_mlen-8; */
-    if (!matchNum) return 1;
-    return 1;
-}
-
-
-LZ4_FORCE_INLINE void LZ4HC_updateBinTree(LZ4HC_CCtx_internal* ctx, const BYTE* const ip, const BYTE* const iHighLimit)
-{
-    const BYTE* const base = ctx->base;
-    const U32 target = (U32)(ip - base);
-    U32 idx = ctx->nextToUpdate;
-    while(idx < target)
-        idx += LZ4HC_BinTree_InsertAndGetAllMatches(ctx, base+idx, iHighLimit, 8, NULL, NULL);
-}
-
+typedef struct {
+    int off;
+    int len;
+} LZ4HC_match_t;
 
-/** Tree updater, providing best match */
-LZ4_FORCE_INLINE int LZ4HC_BinTree_GetAllMatches (
-                        LZ4HC_CCtx_internal* ctx,
-                        const BYTE* const ip, const BYTE* const iHighLimit,
-                        size_t best_mlen, LZ4HC_match_t* matches, const int fullUpdate)
+LZ4_FORCE_INLINE
+LZ4HC_match_t LZ4HC_FindLongerMatch(LZ4HC_CCtx_internal* const ctx,
+                        const BYTE* ip, const BYTE* const iHighLimit,
+                        int minLen, int nbSearches)
 {
-    int mnum = 0;
-    if (ip < ctx->base + ctx->nextToUpdate) return 0;   /* skipped area */
-    if (fullUpdate) LZ4HC_updateBinTree(ctx, ip, iHighLimit);
-    best_mlen = LZ4HC_BinTree_InsertAndGetAllMatches(ctx, ip, iHighLimit, best_mlen, matches, &mnum);
-    ctx->nextToUpdate = (U32)(ip - ctx->base + best_mlen);
-    return mnum;
-}
-
-
-#define SET_PRICE(pos, ml, offset, ll, cost)           \
-{                                                      \
-    while (last_pos < pos)  { opt[last_pos+1].price = 1<<30; last_pos++; } \
-    opt[pos].mlen = (int)ml;                           \
-    opt[pos].off = (int)offset;                        \
-    opt[pos].litlen = (int)ll;                         \
-    opt[pos].price = (int)cost;                        \
+    LZ4HC_match_t match = { 0 , 0 };
+    const BYTE* matchPtr = NULL;
+    /* note : LZ4HC_InsertAndGetWiderMatch() is able to modify the starting position of a match (*startpos),
+     * but this won't be the case here, as we define iLowLimit==ip,
+     * so LZ4HC_InsertAndGetWiderMatch() won't be allowed to search past ip */
+    int const matchLength = LZ4HC_InsertAndGetWiderMatch(ctx, ip, ip, iHighLimit, minLen, &matchPtr, &ip, nbSearches);
+    if (matchLength <= minLen) return match;
+    match.len = matchLength;
+    match.off = (int)(ip-matchPtr);
+    return match;
 }
 
 
 static int LZ4HC_compress_optimal (
     LZ4HC_CCtx_internal* ctx,
     const char* const source,
-    char* dest,
+    char* dst,
     int inputSize,
-    int maxOutputSize,
+    int dstCapacity,
     limitedOutput_directive limit,
+    int const nbSearches,
     size_t sufficient_len,
-    const int fullUpdate
+    int const fullUpdate
     )
 {
-    LZ4HC_optimal_t opt[LZ4_OPT_NUM + 1];   /* this uses a bit too much stack memory to my taste ... */
-    LZ4HC_match_t matches[LZ4_OPT_NUM + 1];
+#define TRAILING_LITERALS 3
+    LZ4HC_optimal_t opt[LZ4_OPT_NUM + TRAILING_LITERALS];   /* this uses a bit too much stack memory to my taste ... */
 
     const BYTE* ip = (const BYTE*) source;
     const BYTE* anchor = ip;
     const BYTE* const iend = ip + inputSize;
     const BYTE* const mflimit = iend - MFLIMIT;
-    const BYTE* const matchlimit = (iend - LASTLITERALS);
-    BYTE* op = (BYTE*) dest;
-    BYTE* const oend = op + maxOutputSize;
+    const BYTE* const matchlimit = iend - LASTLITERALS;
+    BYTE* op = (BYTE*) dst;
+    BYTE* const oend = op + dstCapacity;
 
     /* init */
     DEBUGLOG(5, "LZ4HC_compress_optimal");
     if (sufficient_len >= LZ4_OPT_NUM) sufficient_len = LZ4_OPT_NUM-1;
-    ctx->end += inputSize;
-    ip++;
 
     /* Main Loop */
+    assert(ip - anchor < LZ4_MAX_INPUT_SIZE);
     while (ip < mflimit) {
-        size_t const llen = ip - anchor;
-        size_t last_pos = 0;
-        size_t match_num, cur, best_mlen, best_off;
-        memset(opt, 0, sizeof(LZ4HC_optimal_t));  /* memset only the first one */
+        int const llen = (int)(ip - anchor);
+        int best_mlen, best_off;
+        int cur, last_match_pos = 0;
 
-        match_num = LZ4HC_BinTree_GetAllMatches(ctx, ip, matchlimit, MINMATCH-1, matches, fullUpdate);
-        if (!match_num) { ip++; continue; }
+        LZ4HC_match_t const firstMatch = LZ4HC_FindLongerMatch(ctx, ip, matchlimit, MINMATCH-1, nbSearches);
+        if (firstMatch.len==0) { ip++; continue; }
 
-        if ((size_t)matches[match_num-1].len > sufficient_len) {
+        if ((size_t)firstMatch.len > sufficient_len) {
             /* good enough solution : immediate encoding */
-            best_mlen = matches[match_num-1].len;
-            best_off = matches[match_num-1].off;
-            cur = 0;
-            last_pos = 1;
-            goto encode;
+            int const firstML = firstMatch.len;
+            const BYTE* const matchPos = ip - firstMatch.off;
+            if ( LZ4HC_encodeSequence(&ip, &op, &anchor, firstML, matchPos, limit, oend) )   /* updates ip, op and anchor */
+                return 0;  /* error */
+            continue;
         }
 
-        /* set prices using matches at position = 0 */
-        {   size_t matchNb;
-            for (matchNb = 0; matchNb < match_num; matchNb++) {
-                size_t mlen = (matchNb>0) ? (size_t)matches[matchNb-1].len+1 : MINMATCH;
-                best_mlen = matches[matchNb].len;   /* necessarily < sufficient_len < LZ4_OPT_NUM */
-                for ( ; mlen <= best_mlen ; mlen++) {
-                    size_t const cost = LZ4HC_sequencePrice(llen, mlen) - LZ4HC_literalsPrice(llen);
-                    SET_PRICE(mlen, mlen, matches[matchNb].off, 0, cost);   /* updates last_pos and opt[pos] */
-        }   }   }
-
-        if (last_pos < MINMATCH) { ip++; continue; }  /* note : on clang at least, this test improves performance */
+        /* set prices for first positions (literals) */
+        {   int rPos;
+            for (rPos = 0 ; rPos < MINMATCH ; rPos++) {
+                int const cost = LZ4HC_literalsPrice(llen + rPos);
+                opt[rPos].mlen = 1;
+                opt[rPos].off = 0;
+                opt[rPos].litlen = llen + rPos;
+                opt[rPos].price = cost;
+                DEBUGLOG(7, "rPos:%3i => price:%3i (litlen=%i) -- initial setup",
+                            rPos, cost, opt[rPos].litlen);
+        }   }
+        /* set prices using initial match */
+        {   int mlen = MINMATCH;
+            int const matchML = firstMatch.len;   /* necessarily < sufficient_len < LZ4_OPT_NUM */
+            int const offset = firstMatch.off;
+            assert(matchML < LZ4_OPT_NUM);
+            for ( ; mlen <= matchML ; mlen++) {
+                int const cost = LZ4HC_sequencePrice(llen, mlen);
+                opt[mlen].mlen = mlen;
+                opt[mlen].off = offset;
+                opt[mlen].litlen = llen;
+                opt[mlen].price = cost;
+                DEBUGLOG(7, "rPos:%3i => price:%3i (matchlen=%i) -- initial setup",
+                            mlen, cost, mlen);
+        }   }
+        last_match_pos = firstMatch.len;
+        {   int addLit;
+            for (addLit = 1; addLit <= TRAILING_LITERALS; addLit ++) {
+                opt[last_match_pos+addLit].mlen = 1; /* literal */
+                opt[last_match_pos+addLit].off = 0;
+                opt[last_match_pos+addLit].litlen = addLit;
+                opt[last_match_pos+addLit].price = opt[last_match_pos].price + LZ4HC_literalsPrice(addLit);
+                DEBUGLOG(7, "rPos:%3i => price:%3i (litlen=%i) -- initial setup",
+                            last_match_pos+addLit, opt[last_match_pos+addLit].price, addLit);
+        }   }
 
         /* check further positions */
-        opt[0].mlen = opt[1].mlen = 1;
-        for (cur = 1; cur <= last_pos; cur++) {
+        for (cur = 1; cur < last_match_pos; cur++) {
             const BYTE* const curPtr = ip + cur;
-
-            /* establish baseline price if cur is literal */
-            {   size_t price, litlen;
-                if (opt[cur-1].mlen == 1) {
-                    /* no match at previous position */
-                    litlen = opt[cur-1].litlen + 1;
-                    if (cur > litlen) {
-                        price = opt[cur - litlen].price + LZ4HC_literalsPrice(litlen);
-                    } else {
-                        price = LZ4HC_literalsPrice(llen + litlen) - LZ4HC_literalsPrice(llen);
-                    }
-                } else {
-                    litlen = 1;
-                    price = opt[cur - 1].price + LZ4HC_literalsPrice(1);
-                }
-
-                if (price < (size_t)opt[cur].price)
-                    SET_PRICE(cur, 1 /*mlen*/, 0 /*off*/, litlen, price);   /* note : increases last_pos */
+            LZ4HC_match_t newMatch;
+
+            if (curPtr >= mflimit) break;
+            DEBUGLOG(7, "rPos:%u[%u] vs [%u]%u",
+                    cur, opt[cur].price, opt[cur+1].price, cur+1);
+            if (fullUpdate) {
+                /* not useful to search here if next position has same (or lower) cost */
+                if ( (opt[cur+1].price <= opt[cur].price)
+                  /* in some cases, next position has same cost, but cost rises sharply after, so a small match would still be beneficial */
+                  && (opt[cur+MINMATCH].price < opt[cur].price + 3/*min seq price*/) )
+                    continue;
+            } else {
+                /* not useful to search here if next position has same (or lower) cost */
+                if (opt[cur+1].price <= opt[cur].price) continue;
             }
 
-            if (cur == last_pos || curPtr >= mflimit) break;
+            DEBUGLOG(7, "search at rPos:%u", cur);
+            if (fullUpdate)
+                newMatch = LZ4HC_FindLongerMatch(ctx, curPtr, matchlimit, MINMATCH-1, nbSearches);
+            else
+                /* only test matches of minimum length; slightly faster, but misses a few bytes */
+                newMatch = LZ4HC_FindLongerMatch(ctx, curPtr, matchlimit, last_match_pos - cur, nbSearches);
+            if (!newMatch.len) continue;
 
-            match_num = LZ4HC_BinTree_GetAllMatches(ctx, curPtr, matchlimit, MINMATCH-1, matches, fullUpdate);
-            if ((match_num > 0) && (size_t)matches[match_num-1].len > sufficient_len) {
+            if ( ((size_t)newMatch.len > sufficient_len)
+              || (newMatch.len + cur >= LZ4_OPT_NUM) ) {
                 /* immediate encoding */
-                best_mlen = matches[match_num-1].len;
-                best_off = matches[match_num-1].off;
-                last_pos = cur + 1;
+                best_mlen = newMatch.len;
+                best_off = newMatch.off;
+                last_match_pos = cur + 1;
                 goto encode;
             }
 
-            /* set prices using matches at position = cur */
-            {   size_t matchNb;
-                for (matchNb = 0; matchNb < match_num; matchNb++) {
-                    size_t ml = (matchNb>0) ? (size_t)matches[matchNb-1].len+1 : MINMATCH;
-                    best_mlen = (cur + matches[matchNb].len < LZ4_OPT_NUM) ?
-                                (size_t)matches[matchNb].len : LZ4_OPT_NUM - cur;
-
-                    for ( ; ml <= best_mlen ; ml++) {
-                        size_t ll, price;
-                        if (opt[cur].mlen == 1) {
-                            ll = opt[cur].litlen;
-                            if (cur > ll)
-                                price = opt[cur - ll].price + LZ4HC_sequencePrice(ll, ml);
-                            else
-                                price = LZ4HC_sequencePrice(llen + ll, ml) - LZ4HC_literalsPrice(llen);
-                        } else {
-                            ll = 0;
-                            price = opt[cur].price + LZ4HC_sequencePrice(0, ml);
-                        }
-
-                        if (cur + ml > last_pos || price < (size_t)opt[cur + ml].price) {
-                            SET_PRICE(cur + ml, ml, matches[matchNb].off, ll, price);
-            }   }   }   }
-        } /* for (cur = 1; cur <= last_pos; cur++) */
-
-        best_mlen = opt[last_pos].mlen;
-        best_off = opt[last_pos].off;
-        cur = last_pos - best_mlen;
-
-encode: /* cur, last_pos, best_mlen, best_off must be set */
-        opt[0].mlen = 1;
-        while (1) {  /* from end to beginning */
-            size_t const ml = opt[cur].mlen;
-            int const offset = opt[cur].off;
-            opt[cur].mlen = (int)best_mlen;
-            opt[cur].off = (int)best_off;
-            best_mlen = ml;
-            best_off = offset;
-            if (ml > cur) break;   /* can this happen ? */
-            cur -= ml;
-        }
+            /* before match : set price with literals at beginning */
+            {   int const baseLitlen = opt[cur].litlen;
+                int litlen;
+                for (litlen = 1; litlen < MINMATCH; litlen++) {
+                    int const price = opt[cur].price - LZ4HC_literalsPrice(baseLitlen) + LZ4HC_literalsPrice(baseLitlen+litlen);
+                    int const pos = cur + litlen;
+                    if (price < opt[pos].price) {
+                        opt[pos].mlen = 1; /* literal */
+                        opt[pos].off = 0;
+                        opt[pos].litlen = baseLitlen+litlen;
+                        opt[pos].price = price;
+                        DEBUGLOG(7, "rPos:%3i => price:%3i (litlen=%i)",
+                                    pos, price, opt[pos].litlen);
+            }   }   }
+
+            /* set prices using match at position = cur */
+            {   int const matchML = newMatch.len;
+                int ml = MINMATCH;
+
+                assert(cur + newMatch.len < LZ4_OPT_NUM);
+                for ( ; ml <= matchML ; ml++) {
+                    int const pos = cur + ml;
+                    int const offset = newMatch.off;
+                    int price;
+                    int ll;
+                    DEBUGLOG(7, "testing price rPos %i (last_match_pos=%i)",
+                                pos, last_match_pos);
+                    if (opt[cur].mlen == 1) {
+                        ll = opt[cur].litlen;
+                        price = ((cur > ll) ? opt[cur - ll].price : 0)
+                              + LZ4HC_sequencePrice(ll, ml);
+                    } else {
+                        ll = 0;
+                        price = opt[cur].price + LZ4HC_sequencePrice(0, ml);
+                    }
 
-        /* encode all recorded sequences */
-        cur = 0;
-        while (cur < last_pos) {
-            int const ml = opt[cur].mlen;
-            int const offset = opt[cur].off;
-            if (ml == 1) { ip++; cur++; continue; }
-            cur += ml;
-            if ( LZ4HC_encodeSequence(&ip, &op, &anchor, ml, ip - offset, limit, oend) ) return 0;
-        }
+                    if (pos > last_match_pos+TRAILING_LITERALS || price <= opt[pos].price) {
+                        DEBUGLOG(7, "rPos:%3i => price:%3i (matchlen=%i)",
+                                    pos, price, ml);
+                        assert(pos < LZ4_OPT_NUM);
+                        if ( (ml == matchML)  /* last pos of last match */
+                          && (last_match_pos < pos) )
+                            last_match_pos = pos;
+                        opt[pos].mlen = ml;
+                        opt[pos].off = offset;
+                        opt[pos].litlen = ll;
+                        opt[pos].price = price;
+            }   }   }
+            /* complete following positions with literals */
+            {   int addLit;
+                for (addLit = 1; addLit <= TRAILING_LITERALS; addLit ++) {
+                    opt[last_match_pos+addLit].mlen = 1; /* literal */
+                    opt[last_match_pos+addLit].off = 0;
+                    opt[last_match_pos+addLit].litlen = addLit;
+                    opt[last_match_pos+addLit].price = opt[last_match_pos].price + LZ4HC_literalsPrice(addLit);
+                    DEBUGLOG(7, "rPos:%3i => price:%3i (litlen=%i)", last_match_pos+addLit, opt[last_match_pos+addLit].price, addLit);
+            }   }
+        }  /* for (cur = 1; cur <= last_match_pos; cur++) */
+
+        best_mlen = opt[last_match_pos].mlen;
+        best_off = opt[last_match_pos].off;
+        cur = last_match_pos - best_mlen;
+
+encode: /* cur, last_match_pos, best_mlen, best_off must be set */
+        assert(cur < LZ4_OPT_NUM);
+        assert(last_match_pos >= 1);  /* == 1 when only one candidate */
+        DEBUGLOG(6, "reverse traversal, looking for shortest path")
+        DEBUGLOG(6, "last_match_pos = %i", last_match_pos);
+        {   int candidate_pos = cur;
+            int selected_matchLength = best_mlen;
+            int selected_offset = best_off;
+            while (1) {  /* from end to beginning */
+                int const next_matchLength = opt[candidate_pos].mlen;  /* can be 1, means literal */
+                int const next_offset = opt[candidate_pos].off;
+                DEBUGLOG(6, "pos %i: sequence length %i", candidate_pos, selected_matchLength);
+                opt[candidate_pos].mlen = selected_matchLength;
+                opt[candidate_pos].off = selected_offset;
+                selected_matchLength = next_matchLength;
+                selected_offset = next_offset;
+                if (next_matchLength > candidate_pos) break; /* last match elected, first match to encode */
+                assert(next_matchLength > 0);  /* can be 1, means literal */
+                candidate_pos -= next_matchLength;
+        }   }
+
+        /* encode all recorded sequences in order */
+        {   int rPos = 0;  /* relative position (to ip) */
+            while (rPos < last_match_pos) {
+                int const ml = opt[rPos].mlen;
+                int const offset = opt[rPos].off;
+                if (ml == 1) { ip++; rPos++; continue; }  /* literal; note: can end up with several literals, in which case, skip them */
+                rPos += ml;
+                assert(ml >= MINMATCH);
+                assert((offset >= 1) && (offset <= MAX_DISTANCE));
+                if ( LZ4HC_encodeSequence(&ip, &op, &anchor, ml, ip - offset, limit, oend) )   /* updates ip, op and anchor */
+                    return 0;  /* error */
+        }   }
     }  /* while (ip < mflimit) */
 
     /* Encode Last Literals */
     {   int lastRun = (int)(iend - anchor);
-        if ((limit) && (((char*)op - dest) + lastRun + 1 + ((lastRun+255-RUN_MASK)/255) > (U32)maxOutputSize)) return 0;  /* Check output limit */
-        if (lastRun>=(int)RUN_MASK) { *op++=(RUN_MASK<<ML_BITS); lastRun-=RUN_MASK; for(; lastRun > 254 ; lastRun-=255) *op++ = 255; *op++ = (BYTE) lastRun; }
-        else *op++ = (BYTE)(lastRun<<ML_BITS);
+        if ( (limit)
+          && (((char*)op - dst) + lastRun + 1 + ((lastRun+255-RUN_MASK)/255) > (U32)dstCapacity))
+            return 0;  /* Check output limit */
+        if (lastRun >= (int)RUN_MASK) {
+            *op++=(RUN_MASK<<ML_BITS);
+            lastRun-=RUN_MASK;
+            for (; lastRun > 254 ; lastRun-=255) *op++ = 255;
+            *op++ = (BYTE) lastRun;
+        } else *op++ = (BYTE)(lastRun<<ML_BITS);
         memcpy(op, anchor, iend - anchor);
         op += iend-anchor;
     }
 
     /* End */
-    return (int) ((char*)op-dest);
+    return (int) ((char*)op-dst);
 }