xref: /freebsd/sys/contrib/zstd/lib/compress/zstd_lazy.c (revision cfd6422a5217410fbd66f7a7a8a64d9d85e61229)
1 /*
2  * Copyright (c) 2016-2020, Yann Collet, Facebook, Inc.
3  * All rights reserved.
4  *
5  * This source code is licensed under both the BSD-style license (found in the
6  * LICENSE file in the root directory of this source tree) and the GPLv2 (found
7  * in the COPYING file in the root directory of this source tree).
8  * You may select, at your option, one of the above-listed licenses.
9  */
10 
11 #include "zstd_compress_internal.h"
12 #include "zstd_lazy.h"
13 
14 
15 /*-*************************************
16 *  Binary Tree search
17 ***************************************/
18 
19 static void
20 ZSTD_updateDUBT(ZSTD_matchState_t* ms,
21                 const BYTE* ip, const BYTE* iend,
22                 U32 mls)
23 {
24     const ZSTD_compressionParameters* const cParams = &ms->cParams;
25     U32* const hashTable = ms->hashTable;
26     U32  const hashLog = cParams->hashLog;
27 
28     U32* const bt = ms->chainTable;
29     U32  const btLog  = cParams->chainLog - 1;
30     U32  const btMask = (1 << btLog) - 1;
31 
32     const BYTE* const base = ms->window.base;
33     U32 const target = (U32)(ip - base);
34     U32 idx = ms->nextToUpdate;
35 
36     if (idx != target)
37         DEBUGLOG(7, "ZSTD_updateDUBT, from %u to %u (dictLimit:%u)",
38                     idx, target, ms->window.dictLimit);
39     assert(ip + 8 <= iend);   /* condition for ZSTD_hashPtr */
40     (void)iend;
41 
42     assert(idx >= ms->window.dictLimit);   /* condition for valid base+idx */
43     for ( ; idx < target ; idx++) {
44         size_t const h  = ZSTD_hashPtr(base + idx, hashLog, mls);   /* assumption : ip + 8 <= iend */
45         U32    const matchIndex = hashTable[h];
46 
47         U32*   const nextCandidatePtr = bt + 2*(idx&btMask);
48         U32*   const sortMarkPtr  = nextCandidatePtr + 1;
49 
50         DEBUGLOG(8, "ZSTD_updateDUBT: insert %u", idx);
51         hashTable[h] = idx;   /* Update Hash Table */
52         *nextCandidatePtr = matchIndex;   /* update BT like a chain */
53         *sortMarkPtr = ZSTD_DUBT_UNSORTED_MARK;
54     }
55     ms->nextToUpdate = target;
56 }
57 
58 
59 /** ZSTD_insertDUBT1() :
60  *  sort one already inserted but unsorted position
61  *  assumption : curr >= btlow == (curr - btmask)
62  *  doesn't fail */
63 static void
64 ZSTD_insertDUBT1(ZSTD_matchState_t* ms,
65                  U32 curr, const BYTE* inputEnd,
66                  U32 nbCompares, U32 btLow,
67                  const ZSTD_dictMode_e dictMode)
68 {
69     const ZSTD_compressionParameters* const cParams = &ms->cParams;
70     U32* const bt = ms->chainTable;
71     U32  const btLog  = cParams->chainLog - 1;
72     U32  const btMask = (1 << btLog) - 1;
73     size_t commonLengthSmaller=0, commonLengthLarger=0;
74     const BYTE* const base = ms->window.base;
75     const BYTE* const dictBase = ms->window.dictBase;
76     const U32 dictLimit = ms->window.dictLimit;
77     const BYTE* const ip = (curr>=dictLimit) ? base + curr : dictBase + curr;
78     const BYTE* const iend = (curr>=dictLimit) ? inputEnd : dictBase + dictLimit;
79     const BYTE* const dictEnd = dictBase + dictLimit;
80     const BYTE* const prefixStart = base + dictLimit;
81     const BYTE* match;
82     U32* smallerPtr = bt + 2*(curr&btMask);
83     U32* largerPtr  = smallerPtr + 1;
84     U32 matchIndex = *smallerPtr;   /* this candidate is unsorted : next sorted candidate is reached through *smallerPtr, while *largerPtr contains previous unsorted candidate (which is already saved and can be overwritten) */
85     U32 dummy32;   /* to be nullified at the end */
86     U32 const windowValid = ms->window.lowLimit;
87     U32 const maxDistance = 1U << cParams->windowLog;
88     U32 const windowLow = (curr - windowValid > maxDistance) ? curr - maxDistance : windowValid;
89 
90 
91     DEBUGLOG(8, "ZSTD_insertDUBT1(%u) (dictLimit=%u, lowLimit=%u)",
92                 curr, dictLimit, windowLow);
93     assert(curr >= btLow);
94     assert(ip < iend);   /* condition for ZSTD_count */
95 
96     while (nbCompares-- && (matchIndex > windowLow)) {
97         U32* const nextPtr = bt + 2*(matchIndex & btMask);
98         size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
99         assert(matchIndex < curr);
100         /* note : all candidates are now supposed sorted,
101          * but it's still possible to have nextPtr[1] == ZSTD_DUBT_UNSORTED_MARK
102          * when a real index has the same value as ZSTD_DUBT_UNSORTED_MARK */
103 
104         if ( (dictMode != ZSTD_extDict)
105           || (matchIndex+matchLength >= dictLimit)  /* both in current segment*/
106           || (curr < dictLimit) /* both in extDict */) {
107             const BYTE* const mBase = ( (dictMode != ZSTD_extDict)
108                                      || (matchIndex+matchLength >= dictLimit)) ?
109                                         base : dictBase;
110             assert( (matchIndex+matchLength >= dictLimit)   /* might be wrong if extDict is incorrectly set to 0 */
111                  || (curr < dictLimit) );
112             match = mBase + matchIndex;
113             matchLength += ZSTD_count(ip+matchLength, match+matchLength, iend);
114         } else {
115             match = dictBase + matchIndex;
116             matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iend, dictEnd, prefixStart);
117             if (matchIndex+matchLength >= dictLimit)
118                 match = base + matchIndex;   /* preparation for next read of match[matchLength] */
119         }
120 
121         DEBUGLOG(8, "ZSTD_insertDUBT1: comparing %u with %u : found %u common bytes ",
122                     curr, matchIndex, (U32)matchLength);
123 
124         if (ip+matchLength == iend) {   /* equal : no way to know if inf or sup */
125             break;   /* drop , to guarantee consistency ; miss a bit of compression, but other solutions can corrupt tree */
126         }
127 
128         if (match[matchLength] < ip[matchLength]) {  /* necessarily within buffer */
129             /* match is smaller than current */
130             *smallerPtr = matchIndex;             /* update smaller idx */
131             commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
132             if (matchIndex <= btLow) { smallerPtr=&dummy32; break; }   /* beyond tree size, stop searching */
133             DEBUGLOG(8, "ZSTD_insertDUBT1: %u (>btLow=%u) is smaller : next => %u",
134                         matchIndex, btLow, nextPtr[1]);
135             smallerPtr = nextPtr+1;               /* new "candidate" => larger than match, which was smaller than target */
136             matchIndex = nextPtr[1];              /* new matchIndex, larger than previous and closer to current */
137         } else {
138             /* match is larger than current */
139             *largerPtr = matchIndex;
140             commonLengthLarger = matchLength;
141             if (matchIndex <= btLow) { largerPtr=&dummy32; break; }   /* beyond tree size, stop searching */
142             DEBUGLOG(8, "ZSTD_insertDUBT1: %u (>btLow=%u) is larger => %u",
143                         matchIndex, btLow, nextPtr[0]);
144             largerPtr = nextPtr;
145             matchIndex = nextPtr[0];
146     }   }
147 
148     *smallerPtr = *largerPtr = 0;
149 }
150 
151 
152 static size_t
153 ZSTD_DUBT_findBetterDictMatch (
154         ZSTD_matchState_t* ms,
155         const BYTE* const ip, const BYTE* const iend,
156         size_t* offsetPtr,
157         size_t bestLength,
158         U32 nbCompares,
159         U32 const mls,
160         const ZSTD_dictMode_e dictMode)
161 {
162     const ZSTD_matchState_t * const dms = ms->dictMatchState;
163     const ZSTD_compressionParameters* const dmsCParams = &dms->cParams;
164     const U32 * const dictHashTable = dms->hashTable;
165     U32         const hashLog = dmsCParams->hashLog;
166     size_t      const h  = ZSTD_hashPtr(ip, hashLog, mls);
167     U32               dictMatchIndex = dictHashTable[h];
168 
169     const BYTE* const base = ms->window.base;
170     const BYTE* const prefixStart = base + ms->window.dictLimit;
171     U32         const curr = (U32)(ip-base);
172     const BYTE* const dictBase = dms->window.base;
173     const BYTE* const dictEnd = dms->window.nextSrc;
174     U32         const dictHighLimit = (U32)(dms->window.nextSrc - dms->window.base);
175     U32         const dictLowLimit = dms->window.lowLimit;
176     U32         const dictIndexDelta = ms->window.lowLimit - dictHighLimit;
177 
178     U32*        const dictBt = dms->chainTable;
179     U32         const btLog  = dmsCParams->chainLog - 1;
180     U32         const btMask = (1 << btLog) - 1;
181     U32         const btLow = (btMask >= dictHighLimit - dictLowLimit) ? dictLowLimit : dictHighLimit - btMask;
182 
183     size_t commonLengthSmaller=0, commonLengthLarger=0;
184 
185     (void)dictMode;
186     assert(dictMode == ZSTD_dictMatchState);
187 
188     while (nbCompares-- && (dictMatchIndex > dictLowLimit)) {
189         U32* const nextPtr = dictBt + 2*(dictMatchIndex & btMask);
190         size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
191         const BYTE* match = dictBase + dictMatchIndex;
192         matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iend, dictEnd, prefixStart);
193         if (dictMatchIndex+matchLength >= dictHighLimit)
194             match = base + dictMatchIndex + dictIndexDelta;   /* to prepare for next usage of match[matchLength] */
195 
196         if (matchLength > bestLength) {
197             U32 matchIndex = dictMatchIndex + dictIndexDelta;
198             if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(curr-matchIndex+1) - ZSTD_highbit32((U32)offsetPtr[0]+1)) ) {
199                 DEBUGLOG(9, "ZSTD_DUBT_findBetterDictMatch(%u) : found better match length %u -> %u and offsetCode %u -> %u (dictMatchIndex %u, matchIndex %u)",
200                     curr, (U32)bestLength, (U32)matchLength, (U32)*offsetPtr, ZSTD_REP_MOVE + curr - matchIndex, dictMatchIndex, matchIndex);
201                 bestLength = matchLength, *offsetPtr = ZSTD_REP_MOVE + curr - matchIndex;
202             }
203             if (ip+matchLength == iend) {   /* reached end of input : ip[matchLength] is not valid, no way to know if it's larger or smaller than match */
204                 break;   /* drop, to guarantee consistency (miss a little bit of compression) */
205             }
206         }
207 
208         if (match[matchLength] < ip[matchLength]) {
209             if (dictMatchIndex <= btLow) { break; }   /* beyond tree size, stop the search */
210             commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
211             dictMatchIndex = nextPtr[1];              /* new matchIndex larger than previous (closer to current) */
212         } else {
213             /* match is larger than current */
214             if (dictMatchIndex <= btLow) { break; }   /* beyond tree size, stop the search */
215             commonLengthLarger = matchLength;
216             dictMatchIndex = nextPtr[0];
217         }
218     }
219 
220     if (bestLength >= MINMATCH) {
221         U32 const mIndex = curr - ((U32)*offsetPtr - ZSTD_REP_MOVE); (void)mIndex;
222         DEBUGLOG(8, "ZSTD_DUBT_findBetterDictMatch(%u) : found match of length %u and offsetCode %u (pos %u)",
223                     curr, (U32)bestLength, (U32)*offsetPtr, mIndex);
224     }
225     return bestLength;
226 
227 }
228 
229 
230 static size_t
231 ZSTD_DUBT_findBestMatch(ZSTD_matchState_t* ms,
232                         const BYTE* const ip, const BYTE* const iend,
233                         size_t* offsetPtr,
234                         U32 const mls,
235                         const ZSTD_dictMode_e dictMode)
236 {
237     const ZSTD_compressionParameters* const cParams = &ms->cParams;
238     U32*   const hashTable = ms->hashTable;
239     U32    const hashLog = cParams->hashLog;
240     size_t const h  = ZSTD_hashPtr(ip, hashLog, mls);
241     U32          matchIndex  = hashTable[h];
242 
243     const BYTE* const base = ms->window.base;
244     U32    const curr = (U32)(ip-base);
245     U32    const windowLow = ZSTD_getLowestMatchIndex(ms, curr, cParams->windowLog);
246 
247     U32*   const bt = ms->chainTable;
248     U32    const btLog  = cParams->chainLog - 1;
249     U32    const btMask = (1 << btLog) - 1;
250     U32    const btLow = (btMask >= curr) ? 0 : curr - btMask;
251     U32    const unsortLimit = MAX(btLow, windowLow);
252 
253     U32*         nextCandidate = bt + 2*(matchIndex&btMask);
254     U32*         unsortedMark = bt + 2*(matchIndex&btMask) + 1;
255     U32          nbCompares = 1U << cParams->searchLog;
256     U32          nbCandidates = nbCompares;
257     U32          previousCandidate = 0;
258 
259     DEBUGLOG(7, "ZSTD_DUBT_findBestMatch (%u) ", curr);
260     assert(ip <= iend-8);   /* required for h calculation */
261     assert(dictMode != ZSTD_dedicatedDictSearch);
262 
263     /* reach end of unsorted candidates list */
264     while ( (matchIndex > unsortLimit)
265          && (*unsortedMark == ZSTD_DUBT_UNSORTED_MARK)
266          && (nbCandidates > 1) ) {
267         DEBUGLOG(8, "ZSTD_DUBT_findBestMatch: candidate %u is unsorted",
268                     matchIndex);
269         *unsortedMark = previousCandidate;  /* the unsortedMark becomes a reversed chain, to move up back to original position */
270         previousCandidate = matchIndex;
271         matchIndex = *nextCandidate;
272         nextCandidate = bt + 2*(matchIndex&btMask);
273         unsortedMark = bt + 2*(matchIndex&btMask) + 1;
274         nbCandidates --;
275     }
276 
277     /* nullify last candidate if it's still unsorted
278      * simplification, detrimental to compression ratio, beneficial for speed */
279     if ( (matchIndex > unsortLimit)
280       && (*unsortedMark==ZSTD_DUBT_UNSORTED_MARK) ) {
281         DEBUGLOG(7, "ZSTD_DUBT_findBestMatch: nullify last unsorted candidate %u",
282                     matchIndex);
283         *nextCandidate = *unsortedMark = 0;
284     }
285 
286     /* batch sort stacked candidates */
287     matchIndex = previousCandidate;
288     while (matchIndex) {  /* will end on matchIndex == 0 */
289         U32* const nextCandidateIdxPtr = bt + 2*(matchIndex&btMask) + 1;
290         U32 const nextCandidateIdx = *nextCandidateIdxPtr;
291         ZSTD_insertDUBT1(ms, matchIndex, iend,
292                          nbCandidates, unsortLimit, dictMode);
293         matchIndex = nextCandidateIdx;
294         nbCandidates++;
295     }
296 
297     /* find longest match */
298     {   size_t commonLengthSmaller = 0, commonLengthLarger = 0;
299         const BYTE* const dictBase = ms->window.dictBase;
300         const U32 dictLimit = ms->window.dictLimit;
301         const BYTE* const dictEnd = dictBase + dictLimit;
302         const BYTE* const prefixStart = base + dictLimit;
303         U32* smallerPtr = bt + 2*(curr&btMask);
304         U32* largerPtr  = bt + 2*(curr&btMask) + 1;
305         U32 matchEndIdx = curr + 8 + 1;
306         U32 dummy32;   /* to be nullified at the end */
307         size_t bestLength = 0;
308 
309         matchIndex  = hashTable[h];
310         hashTable[h] = curr;   /* Update Hash Table */
311 
312         while (nbCompares-- && (matchIndex > windowLow)) {
313             U32* const nextPtr = bt + 2*(matchIndex & btMask);
314             size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
315             const BYTE* match;
316 
317             if ((dictMode != ZSTD_extDict) || (matchIndex+matchLength >= dictLimit)) {
318                 match = base + matchIndex;
319                 matchLength += ZSTD_count(ip+matchLength, match+matchLength, iend);
320             } else {
321                 match = dictBase + matchIndex;
322                 matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iend, dictEnd, prefixStart);
323                 if (matchIndex+matchLength >= dictLimit)
324                     match = base + matchIndex;   /* to prepare for next usage of match[matchLength] */
325             }
326 
327             if (matchLength > bestLength) {
328                 if (matchLength > matchEndIdx - matchIndex)
329                     matchEndIdx = matchIndex + (U32)matchLength;
330                 if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(curr-matchIndex+1) - ZSTD_highbit32((U32)offsetPtr[0]+1)) )
331                     bestLength = matchLength, *offsetPtr = ZSTD_REP_MOVE + curr - matchIndex;
332                 if (ip+matchLength == iend) {   /* equal : no way to know if inf or sup */
333                     if (dictMode == ZSTD_dictMatchState) {
334                         nbCompares = 0; /* in addition to avoiding checking any
335                                          * further in this loop, make sure we
336                                          * skip checking in the dictionary. */
337                     }
338                     break;   /* drop, to guarantee consistency (miss a little bit of compression) */
339                 }
340             }
341 
342             if (match[matchLength] < ip[matchLength]) {
343                 /* match is smaller than current */
344                 *smallerPtr = matchIndex;             /* update smaller idx */
345                 commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
346                 if (matchIndex <= btLow) { smallerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
347                 smallerPtr = nextPtr+1;               /* new "smaller" => larger of match */
348                 matchIndex = nextPtr[1];              /* new matchIndex larger than previous (closer to current) */
349             } else {
350                 /* match is larger than current */
351                 *largerPtr = matchIndex;
352                 commonLengthLarger = matchLength;
353                 if (matchIndex <= btLow) { largerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
354                 largerPtr = nextPtr;
355                 matchIndex = nextPtr[0];
356         }   }
357 
358         *smallerPtr = *largerPtr = 0;
359 
360         if (dictMode == ZSTD_dictMatchState && nbCompares) {
361             bestLength = ZSTD_DUBT_findBetterDictMatch(
362                     ms, ip, iend,
363                     offsetPtr, bestLength, nbCompares,
364                     mls, dictMode);
365         }
366 
367         assert(matchEndIdx > curr+8); /* ensure nextToUpdate is increased */
368         ms->nextToUpdate = matchEndIdx - 8;   /* skip repetitive patterns */
369         if (bestLength >= MINMATCH) {
370             U32 const mIndex = curr - ((U32)*offsetPtr - ZSTD_REP_MOVE); (void)mIndex;
371             DEBUGLOG(8, "ZSTD_DUBT_findBestMatch(%u) : found match of length %u and offsetCode %u (pos %u)",
372                         curr, (U32)bestLength, (U32)*offsetPtr, mIndex);
373         }
374         return bestLength;
375     }
376 }
377 
378 
379 /** ZSTD_BtFindBestMatch() : Tree updater, providing best match */
380 FORCE_INLINE_TEMPLATE size_t
381 ZSTD_BtFindBestMatch( ZSTD_matchState_t* ms,
382                 const BYTE* const ip, const BYTE* const iLimit,
383                       size_t* offsetPtr,
384                 const U32 mls /* template */,
385                 const ZSTD_dictMode_e dictMode)
386 {
387     DEBUGLOG(7, "ZSTD_BtFindBestMatch");
388     if (ip < ms->window.base + ms->nextToUpdate) return 0;   /* skipped area */
389     ZSTD_updateDUBT(ms, ip, iLimit, mls);
390     return ZSTD_DUBT_findBestMatch(ms, ip, iLimit, offsetPtr, mls, dictMode);
391 }
392 
393 
394 static size_t
395 ZSTD_BtFindBestMatch_selectMLS (  ZSTD_matchState_t* ms,
396                             const BYTE* ip, const BYTE* const iLimit,
397                                   size_t* offsetPtr)
398 {
399     switch(ms->cParams.minMatch)
400     {
401     default : /* includes case 3 */
402     case 4 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 4, ZSTD_noDict);
403     case 5 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 5, ZSTD_noDict);
404     case 7 :
405     case 6 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 6, ZSTD_noDict);
406     }
407 }
408 
409 
410 static size_t ZSTD_BtFindBestMatch_dictMatchState_selectMLS (
411                         ZSTD_matchState_t* ms,
412                         const BYTE* ip, const BYTE* const iLimit,
413                         size_t* offsetPtr)
414 {
415     switch(ms->cParams.minMatch)
416     {
417     default : /* includes case 3 */
418     case 4 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 4, ZSTD_dictMatchState);
419     case 5 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 5, ZSTD_dictMatchState);
420     case 7 :
421     case 6 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 6, ZSTD_dictMatchState);
422     }
423 }
424 
425 
426 static size_t ZSTD_BtFindBestMatch_extDict_selectMLS (
427                         ZSTD_matchState_t* ms,
428                         const BYTE* ip, const BYTE* const iLimit,
429                         size_t* offsetPtr)
430 {
431     switch(ms->cParams.minMatch)
432     {
433     default : /* includes case 3 */
434     case 4 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 4, ZSTD_extDict);
435     case 5 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 5, ZSTD_extDict);
436     case 7 :
437     case 6 : return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, 6, ZSTD_extDict);
438     }
439 }
440 
441 
442 
443 /* *********************************
444 *  Hash Chain
445 ***********************************/
446 #define NEXT_IN_CHAIN(d, mask)   chainTable[(d) & (mask)]
447 
448 /* Update chains up to ip (excluded)
449    Assumption : always within prefix (i.e. not within extDict) */
450 FORCE_INLINE_TEMPLATE U32 ZSTD_insertAndFindFirstIndex_internal(
451                         ZSTD_matchState_t* ms,
452                         const ZSTD_compressionParameters* const cParams,
453                         const BYTE* ip, U32 const mls)
454 {
455     U32* const hashTable  = ms->hashTable;
456     const U32 hashLog = cParams->hashLog;
457     U32* const chainTable = ms->chainTable;
458     const U32 chainMask = (1 << cParams->chainLog) - 1;
459     const BYTE* const base = ms->window.base;
460     const U32 target = (U32)(ip - base);
461     U32 idx = ms->nextToUpdate;
462 
463     while(idx < target) { /* catch up */
464         size_t const h = ZSTD_hashPtr(base+idx, hashLog, mls);
465         NEXT_IN_CHAIN(idx, chainMask) = hashTable[h];
466         hashTable[h] = idx;
467         idx++;
468     }
469 
470     ms->nextToUpdate = target;
471     return hashTable[ZSTD_hashPtr(ip, hashLog, mls)];
472 }
473 
474 U32 ZSTD_insertAndFindFirstIndex(ZSTD_matchState_t* ms, const BYTE* ip) {
475     const ZSTD_compressionParameters* const cParams = &ms->cParams;
476     return ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, ms->cParams.minMatch);
477 }
478 
479 void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t* ms, const BYTE* const ip)
480 {
481     const BYTE* const base = ms->window.base;
482     U32 const target = (U32)(ip - base);
483     U32* const hashTable = ms->hashTable;
484     U32* const chainTable = ms->chainTable;
485     U32 const chainSize = 1 << ms->cParams.chainLog;
486     U32 idx = ms->nextToUpdate;
487     U32 const minChain = chainSize < target ? target - chainSize : idx;
488     U32 const bucketSize = 1 << ZSTD_LAZY_DDSS_BUCKET_LOG;
489     U32 const cacheSize = bucketSize - 1;
490     U32 const chainAttempts = (1 << ms->cParams.searchLog) - cacheSize;
491     U32 const chainLimit = chainAttempts > 255 ? 255 : chainAttempts;
492 
493     /* We know the hashtable is oversized by a factor of `bucketSize`.
494      * We are going to temporarily pretend `bucketSize == 1`, keeping only a
495      * single entry. We will use the rest of the space to construct a temporary
496      * chaintable.
497      */
498     U32 const hashLog = ms->cParams.hashLog - ZSTD_LAZY_DDSS_BUCKET_LOG;
499     U32* const tmpHashTable = hashTable;
500     U32* const tmpChainTable = hashTable + ((size_t)1 << hashLog);
501     U32 const tmpChainSize = ((1 << ZSTD_LAZY_DDSS_BUCKET_LOG) - 1) << hashLog;
502     U32 const tmpMinChain = tmpChainSize < target ? target - tmpChainSize : idx;
503 
504     U32 hashIdx;
505 
506     assert(ms->cParams.chainLog <= 24);
507     assert(ms->cParams.hashLog >= ms->cParams.chainLog);
508     assert(idx != 0);
509     assert(tmpMinChain <= minChain);
510 
511     /* fill conventional hash table and conventional chain table */
512     for ( ; idx < target; idx++) {
513         U32 const h = (U32)ZSTD_hashPtr(base + idx, hashLog, ms->cParams.minMatch);
514         if (idx >= tmpMinChain) {
515             tmpChainTable[idx - tmpMinChain] = hashTable[h];
516         }
517         tmpHashTable[h] = idx;
518     }
519 
520     /* sort chains into ddss chain table */
521     {
522         U32 chainPos = 0;
523         for (hashIdx = 0; hashIdx < (1U << hashLog); hashIdx++) {
524             U32 count;
525             U32 countBeyondMinChain = 0;
526             U32 i = tmpHashTable[hashIdx];
527             for (count = 0; i >= tmpMinChain && count < cacheSize; count++) {
528                 /* skip through the chain to the first position that won't be
529                  * in the hash cache bucket */
530                 if (i < minChain) {
531                     countBeyondMinChain++;
532                 }
533                 i = tmpChainTable[i - tmpMinChain];
534             }
535             if (count == cacheSize) {
536                 for (count = 0; count < chainLimit;) {
537                     if (i < minChain) {
538                         if (!i || countBeyondMinChain++ > cacheSize) {
539                             /* only allow pulling `cacheSize` number of entries
540                              * into the cache or chainTable beyond `minChain`,
541                              * to replace the entries pulled out of the
542                              * chainTable into the cache. This lets us reach
543                              * back further without increasing the total number
544                              * of entries in the chainTable, guaranteeing the
545                              * DDSS chain table will fit into the space
546                              * allocated for the regular one. */
547                             break;
548                         }
549                     }
550                     chainTable[chainPos++] = i;
551                     count++;
552                     if (i < tmpMinChain) {
553                         break;
554                     }
555                     i = tmpChainTable[i - tmpMinChain];
556                 }
557             } else {
558                 count = 0;
559             }
560             if (count) {
561                 tmpHashTable[hashIdx] = ((chainPos - count) << 8) + count;
562             } else {
563                 tmpHashTable[hashIdx] = 0;
564             }
565         }
566         assert(chainPos <= chainSize); /* I believe this is guaranteed... */
567     }
568 
569     /* move chain pointers into the last entry of each hash bucket */
570     for (hashIdx = (1 << hashLog); hashIdx; ) {
571         U32 const bucketIdx = --hashIdx << ZSTD_LAZY_DDSS_BUCKET_LOG;
572         U32 const chainPackedPointer = tmpHashTable[hashIdx];
573         U32 i;
574         for (i = 0; i < cacheSize; i++) {
575             hashTable[bucketIdx + i] = 0;
576         }
577         hashTable[bucketIdx + bucketSize - 1] = chainPackedPointer;
578     }
579 
580     /* fill the buckets of the hash table */
581     for (idx = ms->nextToUpdate; idx < target; idx++) {
582         U32 const h = (U32)ZSTD_hashPtr(base + idx, hashLog, ms->cParams.minMatch)
583                    << ZSTD_LAZY_DDSS_BUCKET_LOG;
584         U32 i;
585         /* Shift hash cache down 1. */
586         for (i = cacheSize - 1; i; i--)
587             hashTable[h + i] = hashTable[h + i - 1];
588         hashTable[h] = idx;
589     }
590 
591     ms->nextToUpdate = target;
592 }
593 
594 
595 /* inlining is important to hardwire a hot branch (template emulation) */
596 FORCE_INLINE_TEMPLATE
597 size_t ZSTD_HcFindBestMatch_generic (
598                         ZSTD_matchState_t* ms,
599                         const BYTE* const ip, const BYTE* const iLimit,
600                         size_t* offsetPtr,
601                         const U32 mls, const ZSTD_dictMode_e dictMode)
602 {
603     const ZSTD_compressionParameters* const cParams = &ms->cParams;
604     U32* const chainTable = ms->chainTable;
605     const U32 chainSize = (1 << cParams->chainLog);
606     const U32 chainMask = chainSize-1;
607     const BYTE* const base = ms->window.base;
608     const BYTE* const dictBase = ms->window.dictBase;
609     const U32 dictLimit = ms->window.dictLimit;
610     const BYTE* const prefixStart = base + dictLimit;
611     const BYTE* const dictEnd = dictBase + dictLimit;
612     const U32 curr = (U32)(ip-base);
613     const U32 maxDistance = 1U << cParams->windowLog;
614     const U32 lowestValid = ms->window.lowLimit;
615     const U32 withinMaxDistance = (curr - lowestValid > maxDistance) ? curr - maxDistance : lowestValid;
616     const U32 isDictionary = (ms->loadedDictEnd != 0);
617     const U32 lowLimit = isDictionary ? lowestValid : withinMaxDistance;
618     const U32 minChain = curr > chainSize ? curr - chainSize : 0;
619     U32 nbAttempts = 1U << cParams->searchLog;
620     size_t ml=4-1;
621 
622     const ZSTD_matchState_t* const dms = ms->dictMatchState;
623     const U32 ddsHashLog = dictMode == ZSTD_dedicatedDictSearch
624                          ? dms->cParams.hashLog - ZSTD_LAZY_DDSS_BUCKET_LOG : 0;
625     const size_t ddsIdx = dictMode == ZSTD_dedicatedDictSearch
626                         ? ZSTD_hashPtr(ip, ddsHashLog, mls) << ZSTD_LAZY_DDSS_BUCKET_LOG : 0;
627 
628     U32 matchIndex;
629 
630     if (dictMode == ZSTD_dedicatedDictSearch) {
631         const U32* entry = &dms->hashTable[ddsIdx];
632         PREFETCH_L1(entry);
633     }
634 
635     /* HC4 match finder */
636     matchIndex = ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, mls);
637 
638     for ( ; (matchIndex>=lowLimit) & (nbAttempts>0) ; nbAttempts--) {
639         size_t currentMl=0;
640         if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) {
641             const BYTE* const match = base + matchIndex;
642             assert(matchIndex >= dictLimit);   /* ensures this is true if dictMode != ZSTD_extDict */
643             if (match[ml] == ip[ml])   /* potentially better */
644                 currentMl = ZSTD_count(ip, match, iLimit);
645         } else {
646             const BYTE* const match = dictBase + matchIndex;
647             assert(match+4 <= dictEnd);
648             if (MEM_read32(match) == MEM_read32(ip))   /* assumption : matchIndex <= dictLimit-4 (by table construction) */
649                 currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dictEnd, prefixStart) + 4;
650         }
651 
652         /* save best solution */
653         if (currentMl > ml) {
654             ml = currentMl;
655             *offsetPtr = curr - matchIndex + ZSTD_REP_MOVE;
656             if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
657         }
658 
659         if (matchIndex <= minChain) break;
660         matchIndex = NEXT_IN_CHAIN(matchIndex, chainMask);
661     }
662 
663     if (dictMode == ZSTD_dedicatedDictSearch) {
664         const U32 ddsLowestIndex  = dms->window.dictLimit;
665         const BYTE* const ddsBase = dms->window.base;
666         const BYTE* const ddsEnd  = dms->window.nextSrc;
667         const U32 ddsSize         = (U32)(ddsEnd - ddsBase);
668         const U32 ddsIndexDelta   = dictLimit - ddsSize;
669         const U32 bucketSize      = (1 << ZSTD_LAZY_DDSS_BUCKET_LOG);
670         const U32 bucketLimit     = nbAttempts < bucketSize - 1 ? nbAttempts : bucketSize - 1;
671         U32 ddsAttempt;
672 
673         for (ddsAttempt = 0; ddsAttempt < bucketSize - 1; ddsAttempt++) {
674             PREFETCH_L1(ddsBase + dms->hashTable[ddsIdx + ddsAttempt]);
675         }
676 
677         {
678             U32 const chainPackedPointer = dms->hashTable[ddsIdx + bucketSize - 1];
679             U32 const chainIndex = chainPackedPointer >> 8;
680 
681             PREFETCH_L1(&dms->chainTable[chainIndex]);
682         }
683 
684         for (ddsAttempt = 0; ddsAttempt < bucketLimit; ddsAttempt++) {
685             size_t currentMl=0;
686             const BYTE* match;
687             matchIndex = dms->hashTable[ddsIdx + ddsAttempt];
688             match = ddsBase + matchIndex;
689 
690             if (!matchIndex) {
691                 return ml;
692             }
693 
694             /* guaranteed by table construction */
695             (void)ddsLowestIndex;
696             assert(matchIndex >= ddsLowestIndex);
697             assert(match+4 <= ddsEnd);
698             if (MEM_read32(match) == MEM_read32(ip)) {
699                 /* assumption : matchIndex <= dictLimit-4 (by table construction) */
700                 currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, ddsEnd, prefixStart) + 4;
701             }
702 
703             /* save best solution */
704             if (currentMl > ml) {
705                 ml = currentMl;
706                 *offsetPtr = curr - (matchIndex + ddsIndexDelta) + ZSTD_REP_MOVE;
707                 if (ip+currentMl == iLimit) {
708                     /* best possible, avoids read overflow on next attempt */
709                     return ml;
710                 }
711             }
712         }
713 
714         {
715             U32 const chainPackedPointer = dms->hashTable[ddsIdx + bucketSize - 1];
716             U32 chainIndex = chainPackedPointer >> 8;
717             U32 const chainLength = chainPackedPointer & 0xFF;
718             U32 const chainAttempts = nbAttempts - ddsAttempt;
719             U32 const chainLimit = chainAttempts > chainLength ? chainLength : chainAttempts;
720             U32 chainAttempt;
721 
722             for (chainAttempt = 0 ; chainAttempt < chainLimit; chainAttempt++) {
723                 PREFETCH_L1(ddsBase + dms->chainTable[chainIndex + chainAttempt]);
724             }
725 
726             for (chainAttempt = 0 ; chainAttempt < chainLimit; chainAttempt++, chainIndex++) {
727                 size_t currentMl=0;
728                 const BYTE* match;
729                 matchIndex = dms->chainTable[chainIndex];
730                 match = ddsBase + matchIndex;
731 
732                 /* guaranteed by table construction */
733                 assert(matchIndex >= ddsLowestIndex);
734                 assert(match+4 <= ddsEnd);
735                 if (MEM_read32(match) == MEM_read32(ip)) {
736                     /* assumption : matchIndex <= dictLimit-4 (by table construction) */
737                     currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, ddsEnd, prefixStart) + 4;
738                 }
739 
740                 /* save best solution */
741                 if (currentMl > ml) {
742                     ml = currentMl;
743                     *offsetPtr = curr - (matchIndex + ddsIndexDelta) + ZSTD_REP_MOVE;
744                     if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
745                 }
746             }
747         }
748     } else if (dictMode == ZSTD_dictMatchState) {
749         const U32* const dmsChainTable = dms->chainTable;
750         const U32 dmsChainSize         = (1 << dms->cParams.chainLog);
751         const U32 dmsChainMask         = dmsChainSize - 1;
752         const U32 dmsLowestIndex       = dms->window.dictLimit;
753         const BYTE* const dmsBase      = dms->window.base;
754         const BYTE* const dmsEnd       = dms->window.nextSrc;
755         const U32 dmsSize              = (U32)(dmsEnd - dmsBase);
756         const U32 dmsIndexDelta        = dictLimit - dmsSize;
757         const U32 dmsMinChain = dmsSize > dmsChainSize ? dmsSize - dmsChainSize : 0;
758 
759         matchIndex = dms->hashTable[ZSTD_hashPtr(ip, dms->cParams.hashLog, mls)];
760 
761         for ( ; (matchIndex>=dmsLowestIndex) & (nbAttempts>0) ; nbAttempts--) {
762             size_t currentMl=0;
763             const BYTE* const match = dmsBase + matchIndex;
764             assert(match+4 <= dmsEnd);
765             if (MEM_read32(match) == MEM_read32(ip))   /* assumption : matchIndex <= dictLimit-4 (by table construction) */
766                 currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dmsEnd, prefixStart) + 4;
767 
768             /* save best solution */
769             if (currentMl > ml) {
770                 ml = currentMl;
771                 *offsetPtr = curr - (matchIndex + dmsIndexDelta) + ZSTD_REP_MOVE;
772                 if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
773             }
774 
775             if (matchIndex <= dmsMinChain) break;
776 
777             matchIndex = dmsChainTable[matchIndex & dmsChainMask];
778         }
779     }
780 
781     return ml;
782 }
783 
784 
785 FORCE_INLINE_TEMPLATE size_t ZSTD_HcFindBestMatch_selectMLS (
786                         ZSTD_matchState_t* ms,
787                         const BYTE* ip, const BYTE* const iLimit,
788                         size_t* offsetPtr)
789 {
790     switch(ms->cParams.minMatch)
791     {
792     default : /* includes case 3 */
793     case 4 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 4, ZSTD_noDict);
794     case 5 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 5, ZSTD_noDict);
795     case 7 :
796     case 6 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 6, ZSTD_noDict);
797     }
798 }
799 
800 
801 static size_t ZSTD_HcFindBestMatch_dictMatchState_selectMLS (
802                         ZSTD_matchState_t* ms,
803                         const BYTE* ip, const BYTE* const iLimit,
804                         size_t* offsetPtr)
805 {
806     switch(ms->cParams.minMatch)
807     {
808     default : /* includes case 3 */
809     case 4 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 4, ZSTD_dictMatchState);
810     case 5 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 5, ZSTD_dictMatchState);
811     case 7 :
812     case 6 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 6, ZSTD_dictMatchState);
813     }
814 }
815 
816 
817 static size_t ZSTD_HcFindBestMatch_dedicatedDictSearch_selectMLS (
818                         ZSTD_matchState_t* ms,
819                         const BYTE* ip, const BYTE* const iLimit,
820                         size_t* offsetPtr)
821 {
822     switch(ms->cParams.minMatch)
823     {
824     default : /* includes case 3 */
825     case 4 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 4, ZSTD_dedicatedDictSearch);
826     case 5 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 5, ZSTD_dedicatedDictSearch);
827     case 7 :
828     case 6 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 6, ZSTD_dedicatedDictSearch);
829     }
830 }
831 
832 
833 FORCE_INLINE_TEMPLATE size_t ZSTD_HcFindBestMatch_extDict_selectMLS (
834                         ZSTD_matchState_t* ms,
835                         const BYTE* ip, const BYTE* const iLimit,
836                         size_t* offsetPtr)
837 {
838     switch(ms->cParams.minMatch)
839     {
840     default : /* includes case 3 */
841     case 4 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 4, ZSTD_extDict);
842     case 5 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 5, ZSTD_extDict);
843     case 7 :
844     case 6 : return ZSTD_HcFindBestMatch_generic(ms, ip, iLimit, offsetPtr, 6, ZSTD_extDict);
845     }
846 }
847 
848 
849 /* *******************************
850 *  Common parser - lazy strategy
851 *********************************/
852 typedef enum { search_hashChain, search_binaryTree } searchMethod_e;
853 
854 FORCE_INLINE_TEMPLATE size_t
855 ZSTD_compressBlock_lazy_generic(
856                         ZSTD_matchState_t* ms, seqStore_t* seqStore,
857                         U32 rep[ZSTD_REP_NUM],
858                         const void* src, size_t srcSize,
859                         const searchMethod_e searchMethod, const U32 depth,
860                         ZSTD_dictMode_e const dictMode)
861 {
862     const BYTE* const istart = (const BYTE*)src;
863     const BYTE* ip = istart;
864     const BYTE* anchor = istart;
865     const BYTE* const iend = istart + srcSize;
866     const BYTE* const ilimit = iend - 8;
867     const BYTE* const base = ms->window.base;
868     const U32 prefixLowestIndex = ms->window.dictLimit;
869     const BYTE* const prefixLowest = base + prefixLowestIndex;
870 
871     typedef size_t (*searchMax_f)(
872                         ZSTD_matchState_t* ms,
873                         const BYTE* ip, const BYTE* iLimit, size_t* offsetPtr);
874 
875     /**
876      * This table is indexed first by the four ZSTD_dictMode_e values, and then
877      * by the two searchMethod_e values. NULLs are placed for configurations
878      * that should never occur (extDict modes go to the other implementation
879      * below and there is no DDSS for binary tree search yet).
880      */
881     const searchMax_f searchFuncs[4][2] = {
882         {
883             ZSTD_HcFindBestMatch_selectMLS,
884             ZSTD_BtFindBestMatch_selectMLS
885         },
886         {
887             NULL,
888             NULL
889         },
890         {
891             ZSTD_HcFindBestMatch_dictMatchState_selectMLS,
892             ZSTD_BtFindBestMatch_dictMatchState_selectMLS
893         },
894         {
895             ZSTD_HcFindBestMatch_dedicatedDictSearch_selectMLS,
896             NULL
897         }
898     };
899 
900     searchMax_f const searchMax = searchFuncs[dictMode][searchMethod == search_binaryTree];
901     U32 offset_1 = rep[0], offset_2 = rep[1], savedOffset=0;
902 
903     const int isDMS = dictMode == ZSTD_dictMatchState;
904     const int isDDS = dictMode == ZSTD_dedicatedDictSearch;
905     const int isDxS = isDMS || isDDS;
906     const ZSTD_matchState_t* const dms = ms->dictMatchState;
907     const U32 dictLowestIndex      = isDxS ? dms->window.dictLimit : 0;
908     const BYTE* const dictBase     = isDxS ? dms->window.base : NULL;
909     const BYTE* const dictLowest   = isDxS ? dictBase + dictLowestIndex : NULL;
910     const BYTE* const dictEnd      = isDxS ? dms->window.nextSrc : NULL;
911     const U32 dictIndexDelta       = isDxS ?
912                                      prefixLowestIndex - (U32)(dictEnd - dictBase) :
913                                      0;
914     const U32 dictAndPrefixLength = (U32)((ip - prefixLowest) + (dictEnd - dictLowest));
915 
916     assert(searchMax != NULL);
917 
918     DEBUGLOG(5, "ZSTD_compressBlock_lazy_generic (dictMode=%u)", (U32)dictMode);
919 
920     /* init */
921     ip += (dictAndPrefixLength == 0);
922     if (dictMode == ZSTD_noDict) {
923         U32 const curr = (U32)(ip - base);
924         U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, curr, ms->cParams.windowLog);
925         U32 const maxRep = curr - windowLow;
926         if (offset_2 > maxRep) savedOffset = offset_2, offset_2 = 0;
927         if (offset_1 > maxRep) savedOffset = offset_1, offset_1 = 0;
928     }
929     if (isDxS) {
930         /* dictMatchState repCode checks don't currently handle repCode == 0
931          * disabling. */
932         assert(offset_1 <= dictAndPrefixLength);
933         assert(offset_2 <= dictAndPrefixLength);
934     }
935 
936     /* Match Loop */
937 #if defined(__GNUC__) && defined(__x86_64__)
938     /* I've measured random a 5% speed loss on levels 5 & 6 (greedy) when the
939      * code alignment is perturbed. To fix the instability align the loop on 32-bytes.
940      */
941     __asm__(".p2align 5");
942 #endif
943     while (ip < ilimit) {
944         size_t matchLength=0;
945         size_t offset=0;
946         const BYTE* start=ip+1;
947 
948         /* check repCode */
949         if (isDxS) {
950             const U32 repIndex = (U32)(ip - base) + 1 - offset_1;
951             const BYTE* repMatch = ((dictMode == ZSTD_dictMatchState || dictMode == ZSTD_dedicatedDictSearch)
952                                 && repIndex < prefixLowestIndex) ?
953                                    dictBase + (repIndex - dictIndexDelta) :
954                                    base + repIndex;
955             if (((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
956                 && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
957                 const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
958                 matchLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
959                 if (depth==0) goto _storeSequence;
960             }
961         }
962         if ( dictMode == ZSTD_noDict
963           && ((offset_1 > 0) & (MEM_read32(ip+1-offset_1) == MEM_read32(ip+1)))) {
964             matchLength = ZSTD_count(ip+1+4, ip+1+4-offset_1, iend) + 4;
965             if (depth==0) goto _storeSequence;
966         }
967 
968         /* first search (depth 0) */
969         {   size_t offsetFound = 999999999;
970             size_t const ml2 = searchMax(ms, ip, iend, &offsetFound);
971             if (ml2 > matchLength)
972                 matchLength = ml2, start = ip, offset=offsetFound;
973         }
974 
975         if (matchLength < 4) {
976             ip += ((ip-anchor) >> kSearchStrength) + 1;   /* jump faster over incompressible sections */
977             continue;
978         }
979 
980         /* let's try to find a better solution */
981         if (depth>=1)
982         while (ip<ilimit) {
983             ip ++;
984             if ( (dictMode == ZSTD_noDict)
985               && (offset) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
986                 size_t const mlRep = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4;
987                 int const gain2 = (int)(mlRep * 3);
988                 int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offset+1) + 1);
989                 if ((mlRep >= 4) && (gain2 > gain1))
990                     matchLength = mlRep, offset = 0, start = ip;
991             }
992             if (isDxS) {
993                 const U32 repIndex = (U32)(ip - base) - offset_1;
994                 const BYTE* repMatch = repIndex < prefixLowestIndex ?
995                                dictBase + (repIndex - dictIndexDelta) :
996                                base + repIndex;
997                 if (((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
998                     && (MEM_read32(repMatch) == MEM_read32(ip)) ) {
999                     const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
1000                     size_t const mlRep = ZSTD_count_2segments(ip+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
1001                     int const gain2 = (int)(mlRep * 3);
1002                     int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offset+1) + 1);
1003                     if ((mlRep >= 4) && (gain2 > gain1))
1004                         matchLength = mlRep, offset = 0, start = ip;
1005                 }
1006             }
1007             {   size_t offset2=999999999;
1008                 size_t const ml2 = searchMax(ms, ip, iend, &offset2);
1009                 int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1));   /* raw approx */
1010                 int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 4);
1011                 if ((ml2 >= 4) && (gain2 > gain1)) {
1012                     matchLength = ml2, offset = offset2, start = ip;
1013                     continue;   /* search a better one */
1014             }   }
1015 
1016             /* let's find an even better one */
1017             if ((depth==2) && (ip<ilimit)) {
1018                 ip ++;
1019                 if ( (dictMode == ZSTD_noDict)
1020                   && (offset) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
1021                     size_t const mlRep = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4;
1022                     int const gain2 = (int)(mlRep * 4);
1023                     int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 1);
1024                     if ((mlRep >= 4) && (gain2 > gain1))
1025                         matchLength = mlRep, offset = 0, start = ip;
1026                 }
1027                 if (isDxS) {
1028                     const U32 repIndex = (U32)(ip - base) - offset_1;
1029                     const BYTE* repMatch = repIndex < prefixLowestIndex ?
1030                                    dictBase + (repIndex - dictIndexDelta) :
1031                                    base + repIndex;
1032                     if (((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
1033                         && (MEM_read32(repMatch) == MEM_read32(ip)) ) {
1034                         const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
1035                         size_t const mlRep = ZSTD_count_2segments(ip+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
1036                         int const gain2 = (int)(mlRep * 4);
1037                         int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 1);
1038                         if ((mlRep >= 4) && (gain2 > gain1))
1039                             matchLength = mlRep, offset = 0, start = ip;
1040                     }
1041                 }
1042                 {   size_t offset2=999999999;
1043                     size_t const ml2 = searchMax(ms, ip, iend, &offset2);
1044                     int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1));   /* raw approx */
1045                     int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 7);
1046                     if ((ml2 >= 4) && (gain2 > gain1)) {
1047                         matchLength = ml2, offset = offset2, start = ip;
1048                         continue;
1049             }   }   }
1050             break;  /* nothing found : store previous solution */
1051         }
1052 
1053         /* NOTE:
1054          * start[-offset+ZSTD_REP_MOVE-1] is undefined behavior.
1055          * (-offset+ZSTD_REP_MOVE-1) is unsigned, and is added to start, which
1056          * overflows the pointer, which is undefined behavior.
1057          */
1058         /* catch up */
1059         if (offset) {
1060             if (dictMode == ZSTD_noDict) {
1061                 while ( ((start > anchor) & (start - (offset-ZSTD_REP_MOVE) > prefixLowest))
1062                      && (start[-1] == (start-(offset-ZSTD_REP_MOVE))[-1]) )  /* only search for offset within prefix */
1063                     { start--; matchLength++; }
1064             }
1065             if (isDxS) {
1066                 U32 const matchIndex = (U32)((start-base) - (offset - ZSTD_REP_MOVE));
1067                 const BYTE* match = (matchIndex < prefixLowestIndex) ? dictBase + matchIndex - dictIndexDelta : base + matchIndex;
1068                 const BYTE* const mStart = (matchIndex < prefixLowestIndex) ? dictLowest : prefixLowest;
1069                 while ((start>anchor) && (match>mStart) && (start[-1] == match[-1])) { start--; match--; matchLength++; }  /* catch up */
1070             }
1071             offset_2 = offset_1; offset_1 = (U32)(offset - ZSTD_REP_MOVE);
1072         }
1073         /* store sequence */
1074 _storeSequence:
1075         {   size_t const litLength = start - anchor;
1076             ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offset, matchLength-MINMATCH);
1077             anchor = ip = start + matchLength;
1078         }
1079 
1080         /* check immediate repcode */
1081         if (isDxS) {
1082             while (ip <= ilimit) {
1083                 U32 const current2 = (U32)(ip-base);
1084                 U32 const repIndex = current2 - offset_2;
1085                 const BYTE* repMatch = repIndex < prefixLowestIndex ?
1086                         dictBase - dictIndexDelta + repIndex :
1087                         base + repIndex;
1088                 if ( ((U32)((prefixLowestIndex-1) - (U32)repIndex) >= 3 /* intentional overflow */)
1089                    && (MEM_read32(repMatch) == MEM_read32(ip)) ) {
1090                     const BYTE* const repEnd2 = repIndex < prefixLowestIndex ? dictEnd : iend;
1091                     matchLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd2, prefixLowest) + 4;
1092                     offset = offset_2; offset_2 = offset_1; offset_1 = (U32)offset;   /* swap offset_2 <=> offset_1 */
1093                     ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, matchLength-MINMATCH);
1094                     ip += matchLength;
1095                     anchor = ip;
1096                     continue;
1097                 }
1098                 break;
1099             }
1100         }
1101 
1102         if (dictMode == ZSTD_noDict) {
1103             while ( ((ip <= ilimit) & (offset_2>0))
1104                  && (MEM_read32(ip) == MEM_read32(ip - offset_2)) ) {
1105                 /* store sequence */
1106                 matchLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4;
1107                 offset = offset_2; offset_2 = offset_1; offset_1 = (U32)offset; /* swap repcodes */
1108                 ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, matchLength-MINMATCH);
1109                 ip += matchLength;
1110                 anchor = ip;
1111                 continue;   /* faster when present ... (?) */
1112     }   }   }
1113 
1114     /* Save reps for next block */
1115     rep[0] = offset_1 ? offset_1 : savedOffset;
1116     rep[1] = offset_2 ? offset_2 : savedOffset;
1117 
1118     /* Return the last literals size */
1119     return (size_t)(iend - anchor);
1120 }
1121 
1122 
1123 size_t ZSTD_compressBlock_btlazy2(
1124         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1125         void const* src, size_t srcSize)
1126 {
1127     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2, ZSTD_noDict);
1128 }
1129 
1130 size_t ZSTD_compressBlock_lazy2(
1131         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1132         void const* src, size_t srcSize)
1133 {
1134     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_noDict);
1135 }
1136 
1137 size_t ZSTD_compressBlock_lazy(
1138         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1139         void const* src, size_t srcSize)
1140 {
1141     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_noDict);
1142 }
1143 
1144 size_t ZSTD_compressBlock_greedy(
1145         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1146         void const* src, size_t srcSize)
1147 {
1148     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_noDict);
1149 }
1150 
1151 size_t ZSTD_compressBlock_btlazy2_dictMatchState(
1152         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1153         void const* src, size_t srcSize)
1154 {
1155     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2, ZSTD_dictMatchState);
1156 }
1157 
1158 size_t ZSTD_compressBlock_lazy2_dictMatchState(
1159         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1160         void const* src, size_t srcSize)
1161 {
1162     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_dictMatchState);
1163 }
1164 
1165 size_t ZSTD_compressBlock_lazy_dictMatchState(
1166         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1167         void const* src, size_t srcSize)
1168 {
1169     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_dictMatchState);
1170 }
1171 
1172 size_t ZSTD_compressBlock_greedy_dictMatchState(
1173         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1174         void const* src, size_t srcSize)
1175 {
1176     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_dictMatchState);
1177 }
1178 
1179 
1180 size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch(
1181         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1182         void const* src, size_t srcSize)
1183 {
1184     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_dedicatedDictSearch);
1185 }
1186 
1187 size_t ZSTD_compressBlock_lazy_dedicatedDictSearch(
1188         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1189         void const* src, size_t srcSize)
1190 {
1191     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_dedicatedDictSearch);
1192 }
1193 
1194 size_t ZSTD_compressBlock_greedy_dedicatedDictSearch(
1195         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1196         void const* src, size_t srcSize)
1197 {
1198     return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_dedicatedDictSearch);
1199 }
1200 
1201 
1202 FORCE_INLINE_TEMPLATE
1203 size_t ZSTD_compressBlock_lazy_extDict_generic(
1204                         ZSTD_matchState_t* ms, seqStore_t* seqStore,
1205                         U32 rep[ZSTD_REP_NUM],
1206                         const void* src, size_t srcSize,
1207                         const searchMethod_e searchMethod, const U32 depth)
1208 {
1209     const BYTE* const istart = (const BYTE*)src;
1210     const BYTE* ip = istart;
1211     const BYTE* anchor = istart;
1212     const BYTE* const iend = istart + srcSize;
1213     const BYTE* const ilimit = iend - 8;
1214     const BYTE* const base = ms->window.base;
1215     const U32 dictLimit = ms->window.dictLimit;
1216     const BYTE* const prefixStart = base + dictLimit;
1217     const BYTE* const dictBase = ms->window.dictBase;
1218     const BYTE* const dictEnd  = dictBase + dictLimit;
1219     const BYTE* const dictStart  = dictBase + ms->window.lowLimit;
1220     const U32 windowLog = ms->cParams.windowLog;
1221 
1222     typedef size_t (*searchMax_f)(
1223                         ZSTD_matchState_t* ms,
1224                         const BYTE* ip, const BYTE* iLimit, size_t* offsetPtr);
1225     searchMax_f searchMax = searchMethod==search_binaryTree ? ZSTD_BtFindBestMatch_extDict_selectMLS : ZSTD_HcFindBestMatch_extDict_selectMLS;
1226 
1227     U32 offset_1 = rep[0], offset_2 = rep[1];
1228 
1229     DEBUGLOG(5, "ZSTD_compressBlock_lazy_extDict_generic");
1230 
1231     /* init */
1232     ip += (ip == prefixStart);
1233 
1234     /* Match Loop */
1235 #if defined(__GNUC__) && defined(__x86_64__)
1236     /* I've measured random a 5% speed loss on levels 5 & 6 (greedy) when the
1237      * code alignment is perturbed. To fix the instability align the loop on 32-bytes.
1238      */
1239     __asm__(".p2align 5");
1240 #endif
1241     while (ip < ilimit) {
1242         size_t matchLength=0;
1243         size_t offset=0;
1244         const BYTE* start=ip+1;
1245         U32 curr = (U32)(ip-base);
1246 
1247         /* check repCode */
1248         {   const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr+1, windowLog);
1249             const U32 repIndex = (U32)(curr+1 - offset_1);
1250             const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
1251             const BYTE* const repMatch = repBase + repIndex;
1252             if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > windowLow))   /* intentional overflow */
1253             if (MEM_read32(ip+1) == MEM_read32(repMatch)) {
1254                 /* repcode detected we should take it */
1255                 const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
1256                 matchLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repEnd, prefixStart) + 4;
1257                 if (depth==0) goto _storeSequence;
1258         }   }
1259 
1260         /* first search (depth 0) */
1261         {   size_t offsetFound = 999999999;
1262             size_t const ml2 = searchMax(ms, ip, iend, &offsetFound);
1263             if (ml2 > matchLength)
1264                 matchLength = ml2, start = ip, offset=offsetFound;
1265         }
1266 
1267          if (matchLength < 4) {
1268             ip += ((ip-anchor) >> kSearchStrength) + 1;   /* jump faster over incompressible sections */
1269             continue;
1270         }
1271 
1272         /* let's try to find a better solution */
1273         if (depth>=1)
1274         while (ip<ilimit) {
1275             ip ++;
1276             curr++;
1277             /* check repCode */
1278             if (offset) {
1279                 const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr, windowLog);
1280                 const U32 repIndex = (U32)(curr - offset_1);
1281                 const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
1282                 const BYTE* const repMatch = repBase + repIndex;
1283                 if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > windowLow))  /* intentional overflow */
1284                 if (MEM_read32(ip) == MEM_read32(repMatch)) {
1285                     /* repcode detected */
1286                     const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
1287                     size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
1288                     int const gain2 = (int)(repLength * 3);
1289                     int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offset+1) + 1);
1290                     if ((repLength >= 4) && (gain2 > gain1))
1291                         matchLength = repLength, offset = 0, start = ip;
1292             }   }
1293 
1294             /* search match, depth 1 */
1295             {   size_t offset2=999999999;
1296                 size_t const ml2 = searchMax(ms, ip, iend, &offset2);
1297                 int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1));   /* raw approx */
1298                 int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 4);
1299                 if ((ml2 >= 4) && (gain2 > gain1)) {
1300                     matchLength = ml2, offset = offset2, start = ip;
1301                     continue;   /* search a better one */
1302             }   }
1303 
1304             /* let's find an even better one */
1305             if ((depth==2) && (ip<ilimit)) {
1306                 ip ++;
1307                 curr++;
1308                 /* check repCode */
1309                 if (offset) {
1310                     const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr, windowLog);
1311                     const U32 repIndex = (U32)(curr - offset_1);
1312                     const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
1313                     const BYTE* const repMatch = repBase + repIndex;
1314                     if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > windowLow))  /* intentional overflow */
1315                     if (MEM_read32(ip) == MEM_read32(repMatch)) {
1316                         /* repcode detected */
1317                         const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
1318                         size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
1319                         int const gain2 = (int)(repLength * 4);
1320                         int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 1);
1321                         if ((repLength >= 4) && (gain2 > gain1))
1322                             matchLength = repLength, offset = 0, start = ip;
1323                 }   }
1324 
1325                 /* search match, depth 2 */
1326                 {   size_t offset2=999999999;
1327                     size_t const ml2 = searchMax(ms, ip, iend, &offset2);
1328                     int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1));   /* raw approx */
1329                     int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 7);
1330                     if ((ml2 >= 4) && (gain2 > gain1)) {
1331                         matchLength = ml2, offset = offset2, start = ip;
1332                         continue;
1333             }   }   }
1334             break;  /* nothing found : store previous solution */
1335         }
1336 
1337         /* catch up */
1338         if (offset) {
1339             U32 const matchIndex = (U32)((start-base) - (offset - ZSTD_REP_MOVE));
1340             const BYTE* match = (matchIndex < dictLimit) ? dictBase + matchIndex : base + matchIndex;
1341             const BYTE* const mStart = (matchIndex < dictLimit) ? dictStart : prefixStart;
1342             while ((start>anchor) && (match>mStart) && (start[-1] == match[-1])) { start--; match--; matchLength++; }  /* catch up */
1343             offset_2 = offset_1; offset_1 = (U32)(offset - ZSTD_REP_MOVE);
1344         }
1345 
1346         /* store sequence */
1347 _storeSequence:
1348         {   size_t const litLength = start - anchor;
1349             ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offset, matchLength-MINMATCH);
1350             anchor = ip = start + matchLength;
1351         }
1352 
1353         /* check immediate repcode */
1354         while (ip <= ilimit) {
1355             const U32 repCurrent = (U32)(ip-base);
1356             const U32 windowLow = ZSTD_getLowestMatchIndex(ms, repCurrent, windowLog);
1357             const U32 repIndex = repCurrent - offset_2;
1358             const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
1359             const BYTE* const repMatch = repBase + repIndex;
1360             if (((U32)((dictLimit-1) - repIndex) >= 3) & (repIndex > windowLow))  /* intentional overflow */
1361             if (MEM_read32(ip) == MEM_read32(repMatch)) {
1362                 /* repcode detected we should take it */
1363                 const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
1364                 matchLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
1365                 offset = offset_2; offset_2 = offset_1; offset_1 = (U32)offset;   /* swap offset history */
1366                 ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, matchLength-MINMATCH);
1367                 ip += matchLength;
1368                 anchor = ip;
1369                 continue;   /* faster when present ... (?) */
1370             }
1371             break;
1372     }   }
1373 
1374     /* Save reps for next block */
1375     rep[0] = offset_1;
1376     rep[1] = offset_2;
1377 
1378     /* Return the last literals size */
1379     return (size_t)(iend - anchor);
1380 }
1381 
1382 
1383 size_t ZSTD_compressBlock_greedy_extDict(
1384         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1385         void const* src, size_t srcSize)
1386 {
1387     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0);
1388 }
1389 
1390 size_t ZSTD_compressBlock_lazy_extDict(
1391         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1392         void const* src, size_t srcSize)
1393 
1394 {
1395     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1);
1396 }
1397 
1398 size_t ZSTD_compressBlock_lazy2_extDict(
1399         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1400         void const* src, size_t srcSize)
1401 
1402 {
1403     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2);
1404 }
1405 
1406 size_t ZSTD_compressBlock_btlazy2_extDict(
1407         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1408         void const* src, size_t srcSize)
1409 
1410 {
1411     return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2);
1412 }
1413