Merge pull request #72 from Cyan4973/dev

Dev

Makefile

@@ -32,7 +32,7 @@
 # ################################################################
 
 # Version number
-export VERSION := 0.3.4
+export VERSION := 0.3.6
 
 PRGDIR  = programs
 ZSTDDIR = lib

NEWS

@@ -1,3 +1,9 @@
+v0.3.6
+small blocks params
+
+v0.3.5
+minor generic compression improvements
+
 v0.3.4
 Faster fast cLevels
 

lib/zstd.h

@@ -48,7 +48,7 @@ extern "C" {
 ***************************************/
 #define ZSTD_VERSION_MAJOR    0    /* for breaking interface changes  */
 #define ZSTD_VERSION_MINOR    3    /* for new (non-breaking) interface capabilities */
-#define ZSTD_VERSION_RELEASE  4    /* for tweaks, bug-fixes, or development */
+#define ZSTD_VERSION_RELEASE  6    /* for tweaks, bug-fixes, or development */
 #define ZSTD_VERSION_NUMBER  (ZSTD_VERSION_MAJOR *100*100 + ZSTD_VERSION_MINOR *100 + ZSTD_VERSION_RELEASE)
 unsigned ZSTD_versionNumber (void);
 

lib/zstd_internal.h

@@ -226,6 +226,14 @@ static const U32 g_searchStrength = 8;
 */
 MEM_STATIC void ZSTD_storeSeq(seqStore_t* seqStorePtr, size_t litLength, const BYTE* literals, size_t offsetCode, size_t matchCode)
 {
+#if 0
+    static const BYTE* g_start = NULL;
+    if (g_start==NULL) g_start = literals;
+    if (literals - g_start == 8695)
+    printf("pos %6u : %3u literals & match %3u bytes at distance %6u \n",
+           (U32)(literals - g_start), (U32)litLength, (U32)matchCode+4, (U32)offsetCode);
+#endif
+
     /* copy Literals */
     ZSTD_wildcopy(seqStorePtr->lit, literals, litLength);
     seqStorePtr->lit += litLength;

lib/zstdhc.c

@@ -111,7 +111,7 @@ size_t ZSTD_HC_freeCCtx(ZSTD_HC_CCtx* cctx)
 /** ZSTD_HC_validateParams
     correct params value to remain within authorized range
     optimize for srcSize if srcSize > 0 */
-void ZSTD_HC_validateParams(ZSTD_HC_parameters* params, size_t srcSize)
+void ZSTD_HC_validateParams(ZSTD_HC_parameters* params, U64 srcSizeHint)
 {
     const U32 btPlus = (params->strategy == ZSTD_HC_btlazy2);
 
@@ -120,9 +120,9 @@ void ZSTD_HC_validateParams(ZSTD_HC_parameters* params, size_t srcSize)
     if (params->windowLog   < ZSTD_HC_WINDOWLOG_MIN) params->windowLog = ZSTD_HC_WINDOWLOG_MIN;
 
     /* correct params, to use less memory */
-    if (srcSize > 0)
+    if ((srcSizeHint > 0) && (srcSizeHint < (1<<ZSTD_HC_WINDOWLOG_MAX)))
     {
-        U32 srcLog = ZSTD_highbit((U32)srcSize-1) + 1;
+        U32 srcLog = ZSTD_highbit((U32)srcSizeHint-1) + 1;
         if (params->windowLog > srcLog) params->windowLog = srcLog;
     }
 
@@ -139,9 +139,10 @@ void ZSTD_HC_validateParams(ZSTD_HC_parameters* params, size_t srcSize)
 
 
 static size_t ZSTD_HC_resetCCtx_advanced (ZSTD_HC_CCtx* zc,
-                                          ZSTD_HC_parameters params)
+                                          ZSTD_HC_parameters params,
+                                          U64 srcSizeHint)
 {
-    ZSTD_HC_validateParams(&params, 0);
+    ZSTD_HC_validateParams(&params, srcSizeHint);
 
     /* reserve table memory */
     {
@@ -161,7 +162,7 @@ static size_t ZSTD_HC_resetCCtx_advanced (ZSTD_HC_CCtx* zc,
         zc->seqStore.buffer = (void*) (zc->contentTable + ((size_t)1 << contentLog));
     }
 
-    zc->nextToUpdate = 0;
+    zc->nextToUpdate = 1;
     zc->end = NULL;
     zc->base = NULL;
     zc->dictBase = NULL;
@@ -316,7 +317,7 @@ size_t ZSTD_HC_compressBlock_fast(ZSTD_HC_CCtx* ctx,
 ***************************************/
 /** ZSTD_HC_insertBt1 : add one ptr to tree
     @ip : assumed <= iend-8 */
-static void ZSTD_HC_insertBt1(ZSTD_HC_CCtx* zc, const BYTE* const ip, const U32 mls, const BYTE* const iend, U32 nbCompares)
+static U32 ZSTD_HC_insertBt1(ZSTD_HC_CCtx* zc, const BYTE* const ip, const U32 mls, const BYTE* const iend, U32 nbCompares)
 {
     U32* const hashTable = zc->hashTable;
     const U32 hashLog = zc->params.hashLog;
@@ -327,7 +328,8 @@ static void ZSTD_HC_insertBt1(ZSTD_HC_CCtx* zc, const BYTE* const ip, const U32
     U32 matchIndex  = hashTable[h];
     size_t commonLengthSmaller=0, commonLengthLarger=0;
     const BYTE* const base = zc->base;
-    const U32 current = (U32)(ip-base);
+    const BYTE* match = base + matchIndex;
+    U32 current = (U32)(ip-base);
     const U32 btLow = btMask >= current ? 0 : current - btMask;
     U32* smallerPtr = bt + 2*(current&btMask);
     U32* largerPtr  = bt + 2*(current&btMask) + 1;
@@ -335,14 +337,21 @@ static void ZSTD_HC_insertBt1(ZSTD_HC_CCtx* zc, const BYTE* const ip, const U32
     const U32 windowSize = 1 << zc->params.windowLog;
     const U32 windowLow = windowSize >= current ? 0 : current - windowSize;
 
-    hashTable[h] = (U32)(ip-base);   /* Update Hash Table */
+    if ((current-matchIndex == 1)   /* RLE */
+        && MEM_read64(match) == MEM_read64(ip))
+    {
+        size_t rleLength = ZSTD_count(ip+sizeof(size_t), match+sizeof(size_t), iend) + sizeof(size_t);
+        return (U32)(rleLength - mls);
+    }
+
+    hashTable[h] = (U32)(ip - base);   /* Update Hash Table */
 
     while (nbCompares-- && (matchIndex > windowLow))
     {
         U32* nextPtr = bt + 2*(matchIndex & btMask);
-        const BYTE* match = base + matchIndex;
         size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
 
+        match = base + matchIndex;
         matchLength += ZSTD_count(ip+matchLength, match+matchLength, iend);
 
         if (ip+matchLength == iend)   /* equal : no way to know if inf or sup */
@@ -369,6 +378,7 @@ static void ZSTD_HC_insertBt1(ZSTD_HC_CCtx* zc, const BYTE* const ip, const U32
     }
 
     *smallerPtr = *largerPtr = 0;
+    return 1;
 }
 
 
@@ -412,7 +422,7 @@ size_t ZSTD_HC_insertBtAndFindBestMatch (
             if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit(current-matchIndex+1) - ZSTD_highbit((U32)offsetPtr[0]+1)) )
                 bestLength = matchLength, *offsetPtr = current - matchIndex;
             if (ip+matchLength == iend)   /* equal : no way to know if inf or sup */
-                break;   /* drop, next to null, to guarantee consistency (is there a way to do better ?) */
+                break;   /* just drop, to guarantee consistency (miss a little bit of compression) */
         }
 
         if (match[matchLength] < ip[matchLength])
@@ -420,41 +430,41 @@ size_t ZSTD_HC_insertBtAndFindBestMatch (
             /* match is smaller than current */
             *smallerPtr = matchIndex;             /* update smaller idx */
             commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
-            if (matchIndex <= btLow) { smallerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
             smallerPtr = nextPtr+1;               /* new "smaller" => larger of match */
-            matchIndex = nextPtr[1];              /* new matchIndex larger than previous (closer to current) */
+            if (matchIndex <= btLow) smallerPtr=&dummy32;  /* beyond tree size, stop the search */
+            matchIndex = (matchIndex <= btLow) ? windowLow : nextPtr[1];
         }
         else
         {
             /* match is larger than current */
             *largerPtr = matchIndex;
             commonLengthLarger = matchLength;
-            if (matchIndex <= btLow) { largerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
             largerPtr = nextPtr;
-            matchIndex = nextPtr[0];
+            if (matchIndex <= btLow) largerPtr=&dummy32; /* beyond tree size, stop the search */
+            matchIndex = (matchIndex <= btLow) ? windowLow : nextPtr[0];
         }
     }
 
     *smallerPtr = *largerPtr = 0;
 
     zc->nextToUpdate = current+1;   /* current has been inserted */
-    if (bestLength < MINMATCH) return 0;
     return bestLength;
 }
 
 
-static void ZSTD_HC_updateTree(ZSTD_HC_CCtx* zc, const BYTE* const ip, const BYTE* const iend, const U32 nbCompares, const U32 mls)
+static const BYTE* ZSTD_HC_updateTree(ZSTD_HC_CCtx* zc, const BYTE* const ip, const BYTE* const iend, const U32 nbCompares, const U32 mls)
 {
     const BYTE* const base = zc->base;
     const U32 target = (U32)(ip - base);
     U32 idx = zc->nextToUpdate;
     //size_t dummy;
 
-    for( ; idx < target ; idx++)
-        ZSTD_HC_insertBt1(zc, base+idx, mls, iend, nbCompares);
+    for( ; idx < target ; )
+        idx += ZSTD_HC_insertBt1(zc, base+idx, mls, iend, nbCompares);
         //ZSTD_HC_insertBtAndFindBestMatch(zc, base+idx, iend, &dummy, nbCompares, mls);
 
-    zc->nextToUpdate = target;
+    zc->nextToUpdate = idx;
+    return base + idx;
 }
 
 
@@ -466,7 +476,13 @@ size_t ZSTD_HC_BtFindBestMatch (
                         size_t* offsetPtr,
                         const U32 maxNbAttempts, const U32 mls)
 {
-    ZSTD_HC_updateTree(zc, ip, iLimit, maxNbAttempts, mls);
+    const BYTE* nextToUpdate = ZSTD_HC_updateTree(zc, ip, iLimit, maxNbAttempts, mls);
+    if (nextToUpdate > ip)
+    {
+        /* RLE data */
+        *offsetPtr = 1;
+        return ZSTD_count(ip, ip-1, iLimit);
+    }
     return ZSTD_HC_insertBtAndFindBestMatch(zc, ip, iLimit, offsetPtr, maxNbAttempts, mls);
 }
 
@@ -546,22 +562,21 @@ size_t ZSTD_HC_HcFindBestMatch (
         if (matchIndex >= dictLimit)
         {
             match = base + matchIndex;
-            if ( (match[ml] == ip[ml])
-              && (MEM_read32(match) == MEM_read32(ip)) )   /* ensures minimum match of 4 */
+            if (match[ml] == ip[ml])   /* potentially better */
             {
-                const size_t mlt = ZSTD_count(ip+MINMATCH, match+MINMATCH, iLimit) + MINMATCH;
+                const size_t mlt = ZSTD_count(ip, match, iLimit);
                 if (mlt > ml)
                 //if (((int)(4*mlt) - (int)ZSTD_highbit((U32)(ip-match)+1)) > ((int)(4*ml) - (int)ZSTD_highbit((U32)((*offsetPtr)+1))))
                 {
                     ml = mlt; *offsetPtr = ip-match;
-                    if (ip+ml >= iLimit) break;
+                    if (ip+mlt >= iLimit) break;
                 }
             }
         }
         else
         {
             match = dictBase + matchIndex;
-            if (MEM_read32(match) == MEM_read32(ip))
+            if (MEM_read32(match) == MEM_read32(ip))   /* beware of end of dict */
             {
                 size_t mlt;
                 const BYTE* vLimit = ip + (dictLimit - matchIndex);
@@ -647,7 +662,11 @@ size_t ZSTD_HC_compressBlock_lazy_generic(ZSTD_HC_CCtx* ctx,
 
         offset_2 = offset_1;
         matchLength = searchMax(ctx, ip, iend, &offset, maxSearches, mls);
-        if (!matchLength) { ip++; continue; }
+        if (matchLength < MINMATCH)
+        {
+            ip += ((ip-anchor) >> g_searchStrength) + 1;   /* jump faster over incompressible sections */
+            continue;
+        }
 
         /* let's try to find a better solution */
         start = ip;
@@ -660,7 +679,7 @@ size_t ZSTD_HC_compressBlock_lazy_generic(ZSTD_HC_CCtx* ctx,
                 size_t ml2 = ZSTD_count(ip+MINMATCH, ip+MINMATCH-offset_1, iend) + MINMATCH;
                 int gain2 = (int)(ml2 * 3);
                 int gain1 = (int)(matchLength*3 - ZSTD_highbit((U32)offset+1) + 1);
-                if (gain2 > gain1)
+                if ((ml2 >= MINMATCH) && (gain2 > gain1))
                     matchLength = ml2, offset = 0, start = ip;
             }
             {
@@ -668,7 +687,7 @@ size_t ZSTD_HC_compressBlock_lazy_generic(ZSTD_HC_CCtx* ctx,
                 size_t ml2 = searchMax(ctx, ip, iend, &offset2, maxSearches, mls);
                 int gain2 = (int)(ml2*(3+deep) - ZSTD_highbit((U32)offset2+1));   /* raw approx */
                 int gain1 = (int)(matchLength*(3+deep) - ZSTD_highbit((U32)offset+1) + (3+deep));
-                if (gain2 > gain1)
+                if ((ml2 >= MINMATCH) && (gain2 > gain1))
                 {
                     matchLength = ml2, offset = offset2, start = ip;
                     continue;   /* search a better one */
@@ -684,7 +703,7 @@ size_t ZSTD_HC_compressBlock_lazy_generic(ZSTD_HC_CCtx* ctx,
                     size_t ml2 = ZSTD_count(ip+MINMATCH, ip+MINMATCH-offset_1, iend) + MINMATCH;
                     int gain2 = (int)(ml2 * 4);
                     int gain1 = (int)(matchLength*4 - ZSTD_highbit((U32)offset+1) + 1);
-                    if (gain2 > gain1)
+                    if ((ml2 >= MINMATCH) && (gain2 > gain1))
                         matchLength = ml2, offset = 0, start = ip;
                 }
                 {
@@ -692,7 +711,7 @@ size_t ZSTD_HC_compressBlock_lazy_generic(ZSTD_HC_CCtx* ctx,
                     size_t ml2 = searchMax(ctx, ip, iend, &offset2, maxSearches, mls);
                     int gain2 = (int)(ml2*4 - ZSTD_highbit((U32)offset2+1));   /* raw approx */
                     int gain1 = (int)(matchLength*4 - ZSTD_highbit((U32)offset+1) + 7);
-                    if (gain2 > gain1)
+                    if ((ml2 >= MINMATCH) && (gain2 > gain1))
                     {
                         matchLength = ml2, offset = offset2, start = ip;
                         continue;
@@ -702,13 +721,19 @@ size_t ZSTD_HC_compressBlock_lazy_generic(ZSTD_HC_CCtx* ctx,
             break;  /* nothing found : store previous solution */
         }
 
+        /* catch up */
+        if (offset)
+        {
+            while ((start>anchor) && (start>ctx->base+offset) && (start[-1] == start[-1-offset]))
+                { start--; matchLength++; }
+        }
+
         /* store sequence */
         {
             size_t litLength = start - anchor;
             if (offset) offset_1 = offset;
             ZSTD_storeSeq(seqStorePtr, litLength, anchor, offset, matchLength-MINMATCH);
-            ip = start + matchLength;
-            anchor = ip;
+            anchor = ip = start + matchLength;
         }
 
     }
@@ -793,8 +818,12 @@ size_t ZSTD_HC_compressBlock_greedy(ZSTD_HC_CCtx* ctx, void* dst, size_t maxDstS
         {
             size_t offset=999999;
             size_t matchLength = ZSTD_HC_HcFindBestMatch_selectMLS(ctx, ip, iend, &offset, maxSearches, mls);
-            if (!matchLength) { ip++; continue; }
-            while ((ip>anchor) && (ip-offset>ctx->base) && (ip[-1] == ip[-1-offset])) { ip--; }  /* catch up */
+            if (matchLength < MINMATCH)
+            {
+                ip += ((ip-anchor) >> g_searchStrength) + 1;   /* jump faster over incompressible sections */
+                continue;
+            }
+            while ((ip>anchor) && (ip-offset>ctx->base) && (ip[-1] == ip[-1-offset])) { ip--; matchLength++; }  /* catch up */
             /* store sequence */
             {
                 size_t litLength = ip-anchor;
@@ -902,7 +931,7 @@ size_t ZSTD_HC_compressContinue (ZSTD_HC_CCtx* ctxPtr,
     if (ip != ctxPtr->end)
     {
         if (ctxPtr->end != NULL)
-            ZSTD_HC_resetCCtx_advanced(ctxPtr, ctxPtr->params);
+            ZSTD_HC_resetCCtx_advanced(ctxPtr, ctxPtr->params, srcSize);
         ctxPtr->base = ip;
     }
 
@@ -913,22 +942,24 @@ size_t ZSTD_HC_compressContinue (ZSTD_HC_CCtx* ctxPtr,
 
 size_t ZSTD_HC_compressBegin_advanced(ZSTD_HC_CCtx* ctx,
                                       void* dst, size_t maxDstSize,
-                                      const ZSTD_HC_parameters params)
+                                      const ZSTD_HC_parameters params,
+                                      U64 srcSizeHint)
 {
     size_t errorCode;
     if (maxDstSize < 4) return ERROR(dstSize_tooSmall);
-    errorCode = ZSTD_HC_resetCCtx_advanced(ctx, params);
+    errorCode = ZSTD_HC_resetCCtx_advanced(ctx, params, srcSizeHint);
     if (ZSTD_isError(errorCode)) return errorCode;
     MEM_writeLE32(dst, ZSTD_magicNumber); /* Write Header */
     return 4;
 }
 
 
-size_t ZSTD_HC_compressBegin(ZSTD_HC_CCtx* ctx, void* dst, size_t maxDstSize, int compressionLevel)
+size_t ZSTD_HC_compressBegin(ZSTD_HC_CCtx* ctx, void* dst, size_t maxDstSize, int compressionLevel, U64 srcSizeHint)
 {
+    int tableID = ((srcSizeHint-1) > 128 KB);   /* intentional underflow for 0 */
     if (compressionLevel<=0) compressionLevel = 1;
     if (compressionLevel > ZSTD_HC_MAX_CLEVEL) compressionLevel = ZSTD_HC_MAX_CLEVEL;
-    return ZSTD_HC_compressBegin_advanced(ctx, dst, maxDstSize, ZSTD_HC_defaultParameters[compressionLevel]);
+    return ZSTD_HC_compressBegin_advanced(ctx, dst, maxDstSize, ZSTD_HC_defaultParameters[tableID][compressionLevel], srcSizeHint);
 }
 
 
@@ -966,7 +997,7 @@ size_t ZSTD_HC_compress_advanced (ZSTD_HC_CCtx* ctx,
     }
 
     /* Header */
-    oSize = ZSTD_HC_compressBegin_advanced(ctx, dst, maxDstSize, params);
+    oSize = ZSTD_HC_compressBegin_advanced(ctx, dst, maxDstSize, params, srcSize);
     if(ZSTD_isError(oSize)) return oSize;
     op += oSize;
     maxDstSize -= oSize;
@@ -988,9 +1019,10 @@ size_t ZSTD_HC_compress_advanced (ZSTD_HC_CCtx* ctx,
 
 size_t ZSTD_HC_compressCCtx (ZSTD_HC_CCtx* ctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize, int compressionLevel)
 {
+    const int tableID = (srcSize > 128 KB);
     if (compressionLevel<=1) return ZSTD_compress(dst, maxDstSize, src, srcSize);   /* fast mode */
     if (compressionLevel > ZSTD_HC_MAX_CLEVEL) compressionLevel = ZSTD_HC_MAX_CLEVEL;
-    return ZSTD_HC_compress_advanced(ctx, dst, maxDstSize, src, srcSize, ZSTD_HC_defaultParameters[compressionLevel]);
+    return ZSTD_HC_compress_advanced(ctx, dst, maxDstSize, src, srcSize, ZSTD_HC_defaultParameters[tableID][compressionLevel]);
 }
 
 size_t ZSTD_HC_compress(void* dst, size_t maxDstSize, const void* src, size_t srcSize, int compressionLevel)

lib/zstdhc_static.h

@@ -83,14 +83,14 @@ size_t ZSTD_HC_compress_advanced (ZSTD_HC_CCtx* ctx,
 
 /** ZSTD_HC_validateParams
     correct params value to remain within authorized range
-    optimize for srcSize if srcSize > 0 */
-void ZSTD_HC_validateParams(ZSTD_HC_parameters* params, size_t srcSize);
+    srcSizeHint value is optional, select 0 if not known */
+void ZSTD_HC_validateParams(ZSTD_HC_parameters* params, U64 srcSizeHint);
 
 
 /* *************************************
 *  Streaming functions
 ***************************************/
-size_t ZSTD_HC_compressBegin(ZSTD_HC_CCtx* ctx, void* dst, size_t maxDstSize, int compressionLevel);
+size_t ZSTD_HC_compressBegin(ZSTD_HC_CCtx* ctx, void* dst, size_t maxDstSize, int compressionLevel, U64 srcSizeHint);
 size_t ZSTD_HC_compressContinue(ZSTD_HC_CCtx* ctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize);
 size_t ZSTD_HC_compressEnd(ZSTD_HC_CCtx* ctx, void* dst, size_t maxDstSize);
 
@@ -99,33 +99,58 @@ size_t ZSTD_HC_compressEnd(ZSTD_HC_CCtx* ctx, void* dst, size_t maxDstSize);
 *  Pre-defined compression levels
 ***************************************/
 #define ZSTD_HC_MAX_CLEVEL 20
-static const ZSTD_HC_parameters ZSTD_HC_defaultParameters[ZSTD_HC_MAX_CLEVEL+1] = {
+static const ZSTD_HC_parameters ZSTD_HC_defaultParameters[2][ZSTD_HC_MAX_CLEVEL+1] = {
+{   /* for <= 128 KB */
+    /* W,  C,  H,  S,  L, strat */
+    { 17, 12, 12,  1,  4, ZSTD_HC_fast    },  /* level  0 - never used */
+    { 17, 12, 13,  1,  6, ZSTD_HC_fast    },  /* level  1 */
+    { 17, 15, 16,  1,  5, ZSTD_HC_fast    },  /* level  2 */
+    { 17, 16, 17,  1,  5, ZSTD_HC_fast    },  /* level  3 */
+    { 17, 13, 15,  2,  4, ZSTD_HC_greedy  },  /* level  4 */
+    { 17, 15, 17,  3,  4, ZSTD_HC_greedy  },  /* level  5 */
+    { 17, 14, 17,  3,  4, ZSTD_HC_lazy    },  /* level  6 */
+    { 17, 16, 17,  4,  4, ZSTD_HC_lazy    },  /* level  7 */
+    { 17, 16, 17,  4,  4, ZSTD_HC_lazy2   },  /* level  8 */
+    { 17, 17, 16,  5,  4, ZSTD_HC_lazy2   },  /* level  9 */
+    { 17, 17, 16,  6,  4, ZSTD_HC_lazy2   },  /* level 10 */
+    { 17, 17, 16,  7,  4, ZSTD_HC_lazy2   },  /* level 11 */
+    { 17, 17, 16,  8,  4, ZSTD_HC_lazy2   },  /* level 12 */
+    { 17, 18, 16,  4,  4, ZSTD_HC_btlazy2 },  /* level 13 */
+    { 17, 18, 16,  5,  4, ZSTD_HC_btlazy2 },  /* level 14 */
+    { 17, 18, 16,  6,  4, ZSTD_HC_btlazy2 },  /* level 15 */
+    { 17, 18, 16,  7,  4, ZSTD_HC_btlazy2 },  /* level 16 */
+    { 17, 18, 16,  8,  4, ZSTD_HC_btlazy2 },  /* level 17 */
+    { 17, 18, 16,  9,  4, ZSTD_HC_btlazy2 },  /* level 18 */
+    { 17, 18, 16, 10,  4, ZSTD_HC_btlazy2 },  /* level 19 */
+    { 17, 18, 18, 12,  4, ZSTD_HC_btlazy2 },  /* level 20 */
+},
+{   /* for > 128 KB */
     /* W,  C,  H,  S,  L, strat */
     { 18, 12, 12,  1,  4, ZSTD_HC_fast    },  /* level  0 - never used */
     { 18, 14, 14,  1,  7, ZSTD_HC_fast    },  /* level  1 - in fact redirected towards zstd fast */
     { 19, 15, 16,  1,  6, ZSTD_HC_fast    },  /* level  2 */
     { 20, 18, 20,  1,  6, ZSTD_HC_fast    },  /* level  3 */
     { 21, 19, 21,  1,  6, ZSTD_HC_fast    },  /* level  4 */
-    { 19, 14, 19,  2,  5, ZSTD_HC_greedy  },  /* level  5 */
+    { 20, 13, 18,  5,  5, ZSTD_HC_greedy  },  /* level  5 */
     { 20, 17, 19,  3,  5, ZSTD_HC_greedy  },  /* level  6 */
     { 21, 17, 20,  3,  5, ZSTD_HC_lazy    },  /* level  7 */
     { 21, 19, 20,  3,  5, ZSTD_HC_lazy    },  /* level  8 */
-    { 21, 19, 20,  4,  5, ZSTD_HC_lazy    },  /* level  9 */
-    { 21, 19, 20,  5,  5, ZSTD_HC_lazy    },  /* level 10 */
-    { 21, 20, 20,  5,  5, ZSTD_HC_lazy    },  /* level 11 */
+    { 21, 20, 20,  3,  5, ZSTD_HC_lazy2   },  /* level  9 */
+    { 21, 19, 20,  4,  5, ZSTD_HC_lazy2   },  /* level 10 */
+    { 22, 20, 22,  4,  5, ZSTD_HC_lazy2   },  /* level 11 */
     { 22, 20, 22,  5,  5, ZSTD_HC_lazy2   },  /* level 12 */
     { 22, 21, 22,  5,  5, ZSTD_HC_lazy2   },  /* level 13 */
     { 22, 22, 23,  5,  5, ZSTD_HC_lazy2   },  /* level 14 */
-    { 22, 21, 22,  6,  5, ZSTD_HC_lazy2   },  /* level 15 */
-    { 22, 21, 22,  4,  5, ZSTD_HC_btlazy2 },  /* level 16 */
-    { 23, 23, 23,  4,  5, ZSTD_HC_btlazy2 },  /* level 17 */
+    { 23, 23, 23,  5,  5, ZSTD_HC_lazy2   },  /* level 15 */
+    { 23, 21, 22,  5,  5, ZSTD_HC_btlazy2 },  /* level 16 */
+    { 23, 24, 23,  4,  5, ZSTD_HC_btlazy2 },  /* level 17 */
     { 25, 24, 23,  5,  5, ZSTD_HC_btlazy2 },  /* level 18 */
     { 25, 26, 23,  5,  5, ZSTD_HC_btlazy2 },  /* level 19 */
     { 26, 27, 24,  6,  5, ZSTD_HC_btlazy2 },  /* level 20 */
+}
 };
 
 
-
 #if defined (__cplusplus)
 }
 #endif

programs/Makefile

@@ -30,7 +30,7 @@
 # fullbench32: Same as fullbench, but forced to compile in 32-bits mode
 # ##########################################################################
 
-VERSION?= 0.3.4
+VERSION?= 0.3.6
 
 DESTDIR?=
 PREFIX ?= /usr/local

programs/bench.c

@@ -380,7 +380,7 @@ static int BMK_benchMem(void* srcBuffer, size_t srcSize, const char* fileName, i
 }
 
 
-static U64 BMK_GetFileSize(char* infilename)
+static U64 BMK_GetFileSize(const char* infilename)
 {
     int r;
 #if defined(_MSC_VER)

programs/fileio.c

@@ -36,7 +36,7 @@
 /**LEGACY_SUPPORT :
 *  decompressor can decode older formats (starting from Zstd 0.1+) */
 #  define ZSTD_LEGACY_SUPPORT 1
-#endif // ZSTD_LEGACY_SUPPORT
+#endif
 
 
 /* *************************************
@@ -58,11 +58,13 @@
 /* *************************************
 *  Includes
 ***************************************/
-#include <stdio.h>     /* fprintf, fopen, fread, _fileno, stdin, stdout */
-#include <stdlib.h>    /* malloc, free */
-#include <string.h>    /* strcmp, strlen */
-#include <time.h>      /* clock */
-#include <errno.h>     /* errno */
+#include <stdio.h>      /* fprintf, fopen, fread, _fileno, stdin, stdout */
+#include <stdlib.h>     /* malloc, free */
+#include <string.h>     /* strcmp, strlen */
+#include <time.h>       /* clock */
+#include <errno.h>      /* errno */
+#include <sys/types.h>  /* stat64 */
+#include <sys/stat.h>   /* stat64 */
 #include "mem.h"
 #include "fileio.h"
 #include "zstd_static.h"
@@ -91,6 +93,10 @@
 #  define IS_CONSOLE(stdStream) isatty(fileno(stdStream))
 #endif
 
+#if !defined(S_ISREG)
+#  define S_ISREG(x) (((x) & S_IFMT) == S_IFREG)
+#endif
+
 
 /* *************************************
 *  Constants
@@ -215,19 +221,35 @@ static void FIO_getFileHandles(FILE** pfinput, FILE** pfoutput, const char* inpu
     if ( *pfoutput==0) EXM_THROW(13, "Pb opening dst : %s", output_filename);
 }
 
+
+static U64 FIO_getFileSize(const char* infilename)
+{
+    int r;
+#if defined(_MSC_VER)
+    struct _stat64 statbuf;
+    r = _stat64(infilename, &statbuf);
+#else
+    struct stat statbuf;
+    r = stat(infilename, &statbuf);
+#endif
+    if (r || !S_ISREG(statbuf.st_mode)) return 0;
+    return (U64)statbuf.st_size;
+}
+
+
 typedef void* (*FIO_createC) (void);
 static void* local_ZSTD_createCCtx(void) { return (void*) ZSTD_createCCtx(); }
 static void* local_ZSTD_HC_createCCtx(void) { return (void*) ZSTD_HC_createCCtx(); }
 
-typedef size_t (*FIO_initC) (void* ctx, void* dst, size_t maxDstSize, int cLevel);
-static size_t local_ZSTD_compressBegin (void* ctx, void* dst, size_t maxDstSize, int cLevel)
+typedef size_t (*FIO_initC) (void* ctx, void* dst, size_t maxDstSize, int cLevel, U64 srcSizeHint);
+static size_t local_ZSTD_compressBegin (void* ctx, void* dst, size_t maxDstSize, int cLevel, U64 srcSizeHint)
 {
-    (void)cLevel;
+    (void)cLevel; (void)srcSizeHint;
     return ZSTD_compressBegin((ZSTD_CCtx*)ctx, dst, maxDstSize);
 }
-static size_t local_ZSTD_HC_compressBegin (void* ctx, void* dst, size_t maxDstSize, int cLevel)
+static size_t local_ZSTD_HC_compressBegin (void* ctx, void* dst, size_t maxDstSize, int cLevel, U64 srcSizeHint)
 {
-    return ZSTD_HC_compressBegin((ZSTD_HC_CCtx*)ctx, dst, maxDstSize, cLevel);
+    return ZSTD_HC_compressBegin((ZSTD_HC_CCtx*)ctx, dst, maxDstSize, cLevel, srcSizeHint);
 }
 
 typedef size_t (*FIO_continueC) (void* ctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize);
@@ -294,6 +316,7 @@ unsigned long long FIO_compressFilename(const char* output_filename, const char*
         freeC = local_ZSTD_HC_freeCCtx;
     }
     FIO_getFileHandles(&finput, &foutput, input_filename, output_filename);
+    filesize = FIO_getFileSize(input_filename);
 
     /* Allocate Memory */
     ctx = createC();
@@ -304,12 +327,13 @@ unsigned long long FIO_compressFilename(const char* output_filename, const char*
     inEnd = inBuff + inBuffSize;
 
     /* Write Frame Header */
-    cSize = initC(ctx, outBuff, outBuffSize, cLevel);
+    cSize = initC(ctx, outBuff, outBuffSize, cLevel, filesize);
     if (ZSTD_isError(cSize)) EXM_THROW(22, "Compression error : cannot create frame header");
 
     sizeCheck = fwrite(outBuff, 1, cSize, foutput);
     if (sizeCheck!=cSize) EXM_THROW(23, "Write error : cannot write header into %s", output_filename);
     compressedfilesize += cSize;
+    filesize = 0;
 
     /* Main compression loop */
     while (1)

programs/fuzzer.c

@@ -297,7 +297,7 @@ int fuzzerTests(U32 seed, U32 nbTests, unsigned startTest, double compressibilit
     {
         size_t sampleSize, sampleStart;
         size_t cSize, dSize, dSupSize;
-        U32 sampleSizeLog, buffNb;
+        U32 sampleSizeLog, buffNb, cLevelMod;
         U64 crcOrig, crcDest;
         int cLevel;
 
@@ -328,13 +328,10 @@ int fuzzerTests(U32 seed, U32 nbTests, unsigned startTest, double compressibilit
         sampleStart = FUZ_rand(&lseed) % (srcBufferSize - sampleSize);
         crcOrig = XXH64(srcBuffer + sampleStart, sampleSize, 0);
 
-        /* compression test */
-        /* covered by HC cLevel 1
-        cSize = ZSTD_compressCCtx(ctx, cBuffer, cBufferSize, srcBuffer + sampleStart, sampleSize);
-        CHECK(ZSTD_isError(cSize), "ZSTD_compress failed"); */
-
         /* HC compression test */
-        cLevel = (FUZ_rand(&lseed) & 3) +1;
+#define MAX(a,b) ((a)>(b)?(a):(b))
+        cLevelMod = MAX(1, 38 - (int)(MAX(9, sampleSizeLog) * 2));   /* use high compression levels with small samples, for speed */
+        cLevel = (FUZ_rand(&lseed) % cLevelMod) +1;
         cSize = ZSTD_HC_compressCCtx(hcctx, cBuffer, cBufferSize, srcBuffer + sampleStart, sampleSize, cLevel);
         CHECK(ZSTD_isError(cSize), "ZSTD_HC_compressCCtx failed");
 
@@ -358,7 +355,7 @@ int fuzzerTests(U32 seed, U32 nbTests, unsigned startTest, double compressibilit
         dSize = ZSTD_decompress(dstBuffer, sampleSize + dSupSize, cBuffer, cSize);
         CHECK(dSize != sampleSize, "ZSTD_decompress failed (%s) (srcSize : %u ; cSize : %u)", ZSTD_getErrorName(dSize), (U32)sampleSize, (U32)cSize);
         crcDest = XXH64(dstBuffer, sampleSize, 0);
-        CHECK(crcOrig != crcDest, "dstBuffer corrupted (pos %u / %u)", (U32)findDiff(srcBuffer+sampleStart, dstBuffer, sampleSize), (U32)sampleSize);
+        CHECK(crcOrig != crcDest, "decompression result corrupted (pos %u / %u)", (U32)findDiff(srcBuffer+sampleStart, dstBuffer, sampleSize), (U32)sampleSize);
 
         /* truncated src decompression test */
         {

programs/paramgrill.c

@@ -122,7 +122,7 @@ static U32 g_rand = 1;
 static U32 g_singleRun = 0;
 static U32 g_target = 0;
 static U32 g_noSeed = 0;
-static const ZSTD_HC_parameters* g_seedParams = ZSTD_HC_defaultParameters;
+static const ZSTD_HC_parameters* g_seedParams = ZSTD_HC_defaultParameters[0];
 static ZSTD_HC_parameters g_params = { 0, 0, 0, 0, 0, ZSTD_HC_greedy };
 
 void BMK_SetNbIterations(int nbLoops)
@@ -289,7 +289,7 @@ static size_t BMK_benchParam(BMK_result_t* resultPtr,
     U64 crcOrig;
 
     /* Memory allocation & restrictions */
-    snprintf(name, 30, "W%02uC%02uH%02uS%02uL%1ust%1u", Wlog, Clog, Hlog, Slog, Slength, strat);
+    snprintf(name, 30, "Sw%02uc%02uh%02us%02ul%1ut%1u", Wlog, Clog, Hlog, Slog, Slength, strat);
     if (!compressedBuffer || !resultBuffer || !blockTable)
     {
         DISPLAY("\nError: not enough memory!\n");
@@ -579,24 +579,30 @@ static int BMK_seed(winnerInfo_t* winners, const ZSTD_HC_parameters params,
     return better;
 }
 
-#define MAX(a,b)   ( (a) > (b) ? (a) : (b) )
 
-static BYTE g_alreadyTested[ZSTD_HC_WINDOWLOG_MAX+1-ZSTD_HC_WINDOWLOG_MIN]
-                           [ZSTD_HC_CONTENTLOG_MAX+1-ZSTD_HC_CONTENTLOG_MIN]
-                           [ZSTD_HC_HASHLOG_MAX+1-ZSTD_HC_HASHLOG_MIN]
-                           [ZSTD_HC_SEARCHLOG_MAX+1-ZSTD_HC_SEARCHLOG_MIN]
-                           [ZSTD_HC_SEARCHLENGTH_MAX+1-ZSTD_HC_SEARCHLENGTH_MIN]
-                           [ZSTD_HC_btlazy2+1 /* strategy */ ] = {};   /* init to zero */
+/* nullified useless params, to ensure count stats */
+static ZSTD_HC_parameters* sanitizeParams(ZSTD_HC_parameters params)
+{
+    g_params = params;
+    if (params.strategy == ZSTD_HC_fast)
+    {
+        g_params.contentLog = 0;
+        g_params.searchLog = 0;
+    }
+    return &g_params;
+}
+
+#define PARAMTABLELOG   25
+#define PARAMTABLESIZE (1<<PARAMTABLELOG)
+#define PARAMTABLEMASK (PARAMTABLESIZE-1)
+static BYTE g_alreadyTested[PARAMTABLESIZE] = {0};   /* init to zero */
 
 #define NB_TESTS_PLAYED(p) \
-    g_alreadyTested[p.windowLog-ZSTD_HC_WINDOWLOG_MIN] \
-                   [p.contentLog-ZSTD_HC_CONTENTLOG_MIN]   \
-                   [p.hashLog-ZSTD_HC_HASHLOG_MIN]     \
-                   [p.searchLog-ZSTD_HC_SEARCHLOG_MIN] \
-                   [p.searchLength-ZSTD_HC_SEARCHLENGTH_MIN] \
-                   [(U32)p.strategy]
+    g_alreadyTested[(XXH64(sanitizeParams(p), sizeof(p), 0) >> 3) & PARAMTABLEMASK]
 
 
+#define MAX(a,b)   ( (a) > (b) ? (a) : (b) )
+
 static void playAround(FILE* f, winnerInfo_t* winners,
                        ZSTD_HC_parameters params,
                        const void* srcBuffer, size_t srcSize,
@@ -697,12 +703,13 @@ static void BMK_benchMem(void* srcBuffer, size_t srcSize)
     int i;
     const char* rfName = "grillResults.txt";
     FILE* f;
-    const U32 srcLog = BMK_highbit((U32)( (g_blockSize ? g_blockSize : srcSize) -1))+1;
+    const size_t blockSize = g_blockSize ? g_blockSize : srcSize;
+    const U32 srcLog = BMK_highbit((U32)(blockSize-1))+1;
 
     if (g_singleRun)
     {
         BMK_result_t testResult;
-        ZSTD_HC_validateParams(&g_params, g_blockSize ? g_blockSize : srcSize);
+        ZSTD_HC_validateParams(&g_params, blockSize);
         BMK_benchParam(&testResult, srcBuffer, srcSize, ctx, g_params);
         DISPLAY("\n");
         return;
@@ -719,12 +726,13 @@ static void BMK_benchMem(void* srcBuffer, size_t srcSize)
     {
         /* baseline config for level 1 */
         BMK_result_t testResult;
-        params.windowLog = MIN(srcLog, 18);
+        params.windowLog = 18;
         params.hashLog = 14;
         params.contentLog = 1;
         params.searchLog = 1;
         params.searchLength = 7;
         params.strategy = ZSTD_HC_fast;
+        ZSTD_HC_validateParams(&params, blockSize);
         BMK_benchParam(&testResult, srcBuffer, srcSize, ctx, params);
         g_cSpeedTarget[1] = (testResult.cSpeed * 15) >> 4;
     }
@@ -735,7 +743,9 @@ static void BMK_benchMem(void* srcBuffer, size_t srcSize)
 
     /* populate initial solution */
     {
+        const int tableID = (blockSize > 128 KB);
         const int maxSeeds = g_noSeed ? 1 : ZSTD_HC_MAX_CLEVEL;
+        g_seedParams = ZSTD_HC_defaultParameters[tableID];
         for (i=1; i<=maxSeeds; i++)
         {
             const U32 btPlus = (params.strategy == ZSTD_HC_btlazy2);
@@ -1031,6 +1041,7 @@ int main(int argc, char** argv)
                         if (*argument=='K') g_blockSize<<=10, argument++;  /* allows using KB notation */
                         if (*argument=='M') g_blockSize<<=20, argument++;
                         if (*argument=='B') argument++;
+                        DISPLAY("using %u KB block size \n", g_blockSize>>10);
                     }
                     break;
 

programs/zstdcli.c

@@ -70,7 +70,7 @@
 **************************************/
 #define COMPRESSOR_NAME "zstd command line interface"
 #ifndef ZSTD_VERSION
-#  define ZSTD_VERSION "v0.3.4"
+#  define ZSTD_VERSION "v0.3.6"
 #endif
 #define AUTHOR "Yann Collet"
 #define WELCOME_MESSAGE "*** %s %i-bits %s, by %s (%s) ***\n", COMPRESSOR_NAME, (int)(sizeof(void*)*8), ZSTD_VERSION, AUTHOR, __DATE__