1 /*
2 * Copyright (c) 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
ZSTD_updateDUBT(ZSTD_matchState_t * ms,const BYTE * ip,const BYTE * iend,U32 mls)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
ZSTD_insertDUBT1(const ZSTD_matchState_t * ms,U32 curr,const BYTE * inputEnd,U32 nbCompares,U32 btLow,const ZSTD_dictMode_e dictMode)64 ZSTD_insertDUBT1(const 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 for (; nbCompares && (matchIndex > windowLow); --nbCompares) {
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
ZSTD_DUBT_findBetterDictMatch(const ZSTD_matchState_t * ms,const BYTE * const ip,const BYTE * const iend,size_t * offsetPtr,size_t bestLength,U32 nbCompares,U32 const mls,const ZSTD_dictMode_e dictMode)153 ZSTD_DUBT_findBetterDictMatch (
154 const 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 for (; nbCompares && (dictMatchIndex > dictLowLimit); --nbCompares) {
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
ZSTD_DUBT_findBestMatch(ZSTD_matchState_t * ms,const BYTE * const ip,const BYTE * const iend,size_t * offsetPtr,U32 const mls,const ZSTD_dictMode_e dictMode)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 for (; nbCompares && (matchIndex > windowLow); --nbCompares) {
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 assert(nbCompares <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */
361 if (dictMode == ZSTD_dictMatchState && nbCompares) {
362 bestLength = ZSTD_DUBT_findBetterDictMatch(
363 ms, ip, iend,
364 offsetPtr, bestLength, nbCompares,
365 mls, dictMode);
366 }
367
368 assert(matchEndIdx > curr+8); /* ensure nextToUpdate is increased */
369 ms->nextToUpdate = matchEndIdx - 8; /* skip repetitive patterns */
370 if (bestLength >= MINMATCH) {
371 U32 const mIndex = curr - ((U32)*offsetPtr - ZSTD_REP_MOVE); (void)mIndex;
372 DEBUGLOG(8, "ZSTD_DUBT_findBestMatch(%u) : found match of length %u and offsetCode %u (pos %u)",
373 curr, (U32)bestLength, (U32)*offsetPtr, mIndex);
374 }
375 return bestLength;
376 }
377 }
378
379
380 /** ZSTD_BtFindBestMatch() : Tree updater, providing best match */
381 FORCE_INLINE_TEMPLATE size_t
ZSTD_BtFindBestMatch(ZSTD_matchState_t * ms,const BYTE * const ip,const BYTE * const iLimit,size_t * offsetPtr,const U32 mls,const ZSTD_dictMode_e dictMode)382 ZSTD_BtFindBestMatch( ZSTD_matchState_t* ms,
383 const BYTE* const ip, const BYTE* const iLimit,
384 size_t* offsetPtr,
385 const U32 mls /* template */,
386 const ZSTD_dictMode_e dictMode)
387 {
388 DEBUGLOG(7, "ZSTD_BtFindBestMatch");
389 if (ip < ms->window.base + ms->nextToUpdate) return 0; /* skipped area */
390 ZSTD_updateDUBT(ms, ip, iLimit, mls);
391 return ZSTD_DUBT_findBestMatch(ms, ip, iLimit, offsetPtr, mls, dictMode);
392 }
393
394 /***********************************
395 * Dedicated dict search
396 ***********************************/
397
ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t * ms,const BYTE * const ip)398 void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_matchState_t* ms, const BYTE* const ip)
399 {
400 const BYTE* const base = ms->window.base;
401 U32 const target = (U32)(ip - base);
402 U32* const hashTable = ms->hashTable;
403 U32* const chainTable = ms->chainTable;
404 U32 const chainSize = 1 << ms->cParams.chainLog;
405 U32 idx = ms->nextToUpdate;
406 U32 const minChain = chainSize < target - idx ? target - chainSize : idx;
407 U32 const bucketSize = 1 << ZSTD_LAZY_DDSS_BUCKET_LOG;
408 U32 const cacheSize = bucketSize - 1;
409 U32 const chainAttempts = (1 << ms->cParams.searchLog) - cacheSize;
410 U32 const chainLimit = chainAttempts > 255 ? 255 : chainAttempts;
411
412 /* We know the hashtable is oversized by a factor of `bucketSize`.
413 * We are going to temporarily pretend `bucketSize == 1`, keeping only a
414 * single entry. We will use the rest of the space to construct a temporary
415 * chaintable.
416 */
417 U32 const hashLog = ms->cParams.hashLog - ZSTD_LAZY_DDSS_BUCKET_LOG;
418 U32* const tmpHashTable = hashTable;
419 U32* const tmpChainTable = hashTable + ((size_t)1 << hashLog);
420 U32 const tmpChainSize = (U32)((1 << ZSTD_LAZY_DDSS_BUCKET_LOG) - 1) << hashLog;
421 U32 const tmpMinChain = tmpChainSize < target ? target - tmpChainSize : idx;
422 U32 hashIdx;
423
424 assert(ms->cParams.chainLog <= 24);
425 assert(ms->cParams.hashLog > ms->cParams.chainLog);
426 assert(idx != 0);
427 assert(tmpMinChain <= minChain);
428
429 /* fill conventional hash table and conventional chain table */
430 for ( ; idx < target; idx++) {
431 U32 const h = (U32)ZSTD_hashPtr(base + idx, hashLog, ms->cParams.minMatch);
432 if (idx >= tmpMinChain) {
433 tmpChainTable[idx - tmpMinChain] = hashTable[h];
434 }
435 tmpHashTable[h] = idx;
436 }
437
438 /* sort chains into ddss chain table */
439 {
440 U32 chainPos = 0;
441 for (hashIdx = 0; hashIdx < (1U << hashLog); hashIdx++) {
442 U32 count;
443 U32 countBeyondMinChain = 0;
444 U32 i = tmpHashTable[hashIdx];
445 for (count = 0; i >= tmpMinChain && count < cacheSize; count++) {
446 /* skip through the chain to the first position that won't be
447 * in the hash cache bucket */
448 if (i < minChain) {
449 countBeyondMinChain++;
450 }
451 i = tmpChainTable[i - tmpMinChain];
452 }
453 if (count == cacheSize) {
454 for (count = 0; count < chainLimit;) {
455 if (i < minChain) {
456 if (!i || ++countBeyondMinChain > cacheSize) {
457 /* only allow pulling `cacheSize` number of entries
458 * into the cache or chainTable beyond `minChain`,
459 * to replace the entries pulled out of the
460 * chainTable into the cache. This lets us reach
461 * back further without increasing the total number
462 * of entries in the chainTable, guaranteeing the
463 * DDSS chain table will fit into the space
464 * allocated for the regular one. */
465 break;
466 }
467 }
468 chainTable[chainPos++] = i;
469 count++;
470 if (i < tmpMinChain) {
471 break;
472 }
473 i = tmpChainTable[i - tmpMinChain];
474 }
475 } else {
476 count = 0;
477 }
478 if (count) {
479 tmpHashTable[hashIdx] = ((chainPos - count) << 8) + count;
480 } else {
481 tmpHashTable[hashIdx] = 0;
482 }
483 }
484 assert(chainPos <= chainSize); /* I believe this is guaranteed... */
485 }
486
487 /* move chain pointers into the last entry of each hash bucket */
488 for (hashIdx = (1 << hashLog); hashIdx; ) {
489 U32 const bucketIdx = --hashIdx << ZSTD_LAZY_DDSS_BUCKET_LOG;
490 U32 const chainPackedPointer = tmpHashTable[hashIdx];
491 U32 i;
492 for (i = 0; i < cacheSize; i++) {
493 hashTable[bucketIdx + i] = 0;
494 }
495 hashTable[bucketIdx + bucketSize - 1] = chainPackedPointer;
496 }
497
498 /* fill the buckets of the hash table */
499 for (idx = ms->nextToUpdate; idx < target; idx++) {
500 U32 const h = (U32)ZSTD_hashPtr(base + idx, hashLog, ms->cParams.minMatch)
501 << ZSTD_LAZY_DDSS_BUCKET_LOG;
502 U32 i;
503 /* Shift hash cache down 1. */
504 for (i = cacheSize - 1; i; i--)
505 hashTable[h + i] = hashTable[h + i - 1];
506 hashTable[h] = idx;
507 }
508
509 ms->nextToUpdate = target;
510 }
511
512 /* Returns the longest match length found in the dedicated dict search structure.
513 * If none are longer than the argument ml, then ml will be returned.
514 */
515 FORCE_INLINE_TEMPLATE
ZSTD_dedicatedDictSearch_lazy_search(size_t * offsetPtr,size_t ml,U32 nbAttempts,const ZSTD_matchState_t * const dms,const BYTE * const ip,const BYTE * const iLimit,const BYTE * const prefixStart,const U32 curr,const U32 dictLimit,const size_t ddsIdx)516 size_t ZSTD_dedicatedDictSearch_lazy_search(size_t* offsetPtr, size_t ml, U32 nbAttempts,
517 const ZSTD_matchState_t* const dms,
518 const BYTE* const ip, const BYTE* const iLimit,
519 const BYTE* const prefixStart, const U32 curr,
520 const U32 dictLimit, const size_t ddsIdx) {
521 const U32 ddsLowestIndex = dms->window.dictLimit;
522 const BYTE* const ddsBase = dms->window.base;
523 const BYTE* const ddsEnd = dms->window.nextSrc;
524 const U32 ddsSize = (U32)(ddsEnd - ddsBase);
525 const U32 ddsIndexDelta = dictLimit - ddsSize;
526 const U32 bucketSize = (1 << ZSTD_LAZY_DDSS_BUCKET_LOG);
527 const U32 bucketLimit = nbAttempts < bucketSize - 1 ? nbAttempts : bucketSize - 1;
528 U32 ddsAttempt;
529 U32 matchIndex;
530
531 for (ddsAttempt = 0; ddsAttempt < bucketSize - 1; ddsAttempt++) {
532 PREFETCH_L1(ddsBase + dms->hashTable[ddsIdx + ddsAttempt]);
533 }
534
535 {
536 U32 const chainPackedPointer = dms->hashTable[ddsIdx + bucketSize - 1];
537 U32 const chainIndex = chainPackedPointer >> 8;
538
539 PREFETCH_L1(&dms->chainTable[chainIndex]);
540 }
541
542 for (ddsAttempt = 0; ddsAttempt < bucketLimit; ddsAttempt++) {
543 size_t currentMl=0;
544 const BYTE* match;
545 matchIndex = dms->hashTable[ddsIdx + ddsAttempt];
546 match = ddsBase + matchIndex;
547
548 if (!matchIndex) {
549 return ml;
550 }
551
552 /* guaranteed by table construction */
553 (void)ddsLowestIndex;
554 assert(matchIndex >= ddsLowestIndex);
555 assert(match+4 <= ddsEnd);
556 if (MEM_read32(match) == MEM_read32(ip)) {
557 /* assumption : matchIndex <= dictLimit-4 (by table construction) */
558 currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, ddsEnd, prefixStart) + 4;
559 }
560
561 /* save best solution */
562 if (currentMl > ml) {
563 ml = currentMl;
564 *offsetPtr = curr - (matchIndex + ddsIndexDelta) + ZSTD_REP_MOVE;
565 if (ip+currentMl == iLimit) {
566 /* best possible, avoids read overflow on next attempt */
567 return ml;
568 }
569 }
570 }
571
572 {
573 U32 const chainPackedPointer = dms->hashTable[ddsIdx + bucketSize - 1];
574 U32 chainIndex = chainPackedPointer >> 8;
575 U32 const chainLength = chainPackedPointer & 0xFF;
576 U32 const chainAttempts = nbAttempts - ddsAttempt;
577 U32 const chainLimit = chainAttempts > chainLength ? chainLength : chainAttempts;
578 U32 chainAttempt;
579
580 for (chainAttempt = 0 ; chainAttempt < chainLimit; chainAttempt++) {
581 PREFETCH_L1(ddsBase + dms->chainTable[chainIndex + chainAttempt]);
582 }
583
584 for (chainAttempt = 0 ; chainAttempt < chainLimit; chainAttempt++, chainIndex++) {
585 size_t currentMl=0;
586 const BYTE* match;
587 matchIndex = dms->chainTable[chainIndex];
588 match = ddsBase + matchIndex;
589
590 /* guaranteed by table construction */
591 assert(matchIndex >= ddsLowestIndex);
592 assert(match+4 <= ddsEnd);
593 if (MEM_read32(match) == MEM_read32(ip)) {
594 /* assumption : matchIndex <= dictLimit-4 (by table construction) */
595 currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, ddsEnd, prefixStart) + 4;
596 }
597
598 /* save best solution */
599 if (currentMl > ml) {
600 ml = currentMl;
601 *offsetPtr = curr - (matchIndex + ddsIndexDelta) + ZSTD_REP_MOVE;
602 if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
603 }
604 }
605 }
606 return ml;
607 }
608
609
610 /* *********************************
611 * Hash Chain
612 ***********************************/
613 #define NEXT_IN_CHAIN(d, mask) chainTable[(d) & (mask)]
614
615 /* Update chains up to ip (excluded)
616 Assumption : always within prefix (i.e. not within extDict) */
ZSTD_insertAndFindFirstIndex_internal(ZSTD_matchState_t * ms,const ZSTD_compressionParameters * const cParams,const BYTE * ip,U32 const mls)617 FORCE_INLINE_TEMPLATE U32 ZSTD_insertAndFindFirstIndex_internal(
618 ZSTD_matchState_t* ms,
619 const ZSTD_compressionParameters* const cParams,
620 const BYTE* ip, U32 const mls)
621 {
622 U32* const hashTable = ms->hashTable;
623 const U32 hashLog = cParams->hashLog;
624 U32* const chainTable = ms->chainTable;
625 const U32 chainMask = (1 << cParams->chainLog) - 1;
626 const BYTE* const base = ms->window.base;
627 const U32 target = (U32)(ip - base);
628 U32 idx = ms->nextToUpdate;
629
630 while(idx < target) { /* catch up */
631 size_t const h = ZSTD_hashPtr(base+idx, hashLog, mls);
632 NEXT_IN_CHAIN(idx, chainMask) = hashTable[h];
633 hashTable[h] = idx;
634 idx++;
635 }
636
637 ms->nextToUpdate = target;
638 return hashTable[ZSTD_hashPtr(ip, hashLog, mls)];
639 }
640
ZSTD_insertAndFindFirstIndex(ZSTD_matchState_t * ms,const BYTE * ip)641 U32 ZSTD_insertAndFindFirstIndex(ZSTD_matchState_t* ms, const BYTE* ip) {
642 const ZSTD_compressionParameters* const cParams = &ms->cParams;
643 return ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, ms->cParams.minMatch);
644 }
645
646 /* inlining is important to hardwire a hot branch (template emulation) */
647 FORCE_INLINE_TEMPLATE
ZSTD_HcFindBestMatch(ZSTD_matchState_t * ms,const BYTE * const ip,const BYTE * const iLimit,size_t * offsetPtr,const U32 mls,const ZSTD_dictMode_e dictMode)648 size_t ZSTD_HcFindBestMatch(
649 ZSTD_matchState_t* ms,
650 const BYTE* const ip, const BYTE* const iLimit,
651 size_t* offsetPtr,
652 const U32 mls, const ZSTD_dictMode_e dictMode)
653 {
654 const ZSTD_compressionParameters* const cParams = &ms->cParams;
655 U32* const chainTable = ms->chainTable;
656 const U32 chainSize = (1 << cParams->chainLog);
657 const U32 chainMask = chainSize-1;
658 const BYTE* const base = ms->window.base;
659 const BYTE* const dictBase = ms->window.dictBase;
660 const U32 dictLimit = ms->window.dictLimit;
661 const BYTE* const prefixStart = base + dictLimit;
662 const BYTE* const dictEnd = dictBase + dictLimit;
663 const U32 curr = (U32)(ip-base);
664 const U32 maxDistance = 1U << cParams->windowLog;
665 const U32 lowestValid = ms->window.lowLimit;
666 const U32 withinMaxDistance = (curr - lowestValid > maxDistance) ? curr - maxDistance : lowestValid;
667 const U32 isDictionary = (ms->loadedDictEnd != 0);
668 const U32 lowLimit = isDictionary ? lowestValid : withinMaxDistance;
669 const U32 minChain = curr > chainSize ? curr - chainSize : 0;
670 U32 nbAttempts = 1U << cParams->searchLog;
671 size_t ml=4-1;
672
673 const ZSTD_matchState_t* const dms = ms->dictMatchState;
674 const U32 ddsHashLog = dictMode == ZSTD_dedicatedDictSearch
675 ? dms->cParams.hashLog - ZSTD_LAZY_DDSS_BUCKET_LOG : 0;
676 const size_t ddsIdx = dictMode == ZSTD_dedicatedDictSearch
677 ? ZSTD_hashPtr(ip, ddsHashLog, mls) << ZSTD_LAZY_DDSS_BUCKET_LOG : 0;
678
679 U32 matchIndex;
680
681 if (dictMode == ZSTD_dedicatedDictSearch) {
682 const U32* entry = &dms->hashTable[ddsIdx];
683 PREFETCH_L1(entry);
684 }
685
686 /* HC4 match finder */
687 matchIndex = ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, mls);
688
689 for ( ; (matchIndex>=lowLimit) & (nbAttempts>0) ; nbAttempts--) {
690 size_t currentMl=0;
691 if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) {
692 const BYTE* const match = base + matchIndex;
693 assert(matchIndex >= dictLimit); /* ensures this is true if dictMode != ZSTD_extDict */
694 if (match[ml] == ip[ml]) /* potentially better */
695 currentMl = ZSTD_count(ip, match, iLimit);
696 } else {
697 const BYTE* const match = dictBase + matchIndex;
698 assert(match+4 <= dictEnd);
699 if (MEM_read32(match) == MEM_read32(ip)) /* assumption : matchIndex <= dictLimit-4 (by table construction) */
700 currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dictEnd, prefixStart) + 4;
701 }
702
703 /* save best solution */
704 if (currentMl > ml) {
705 ml = currentMl;
706 *offsetPtr = curr - matchIndex + ZSTD_REP_MOVE;
707 if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
708 }
709
710 if (matchIndex <= minChain) break;
711 matchIndex = NEXT_IN_CHAIN(matchIndex, chainMask);
712 }
713
714 assert(nbAttempts <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */
715 if (dictMode == ZSTD_dedicatedDictSearch) {
716 ml = ZSTD_dedicatedDictSearch_lazy_search(offsetPtr, ml, nbAttempts, dms,
717 ip, iLimit, prefixStart, curr, dictLimit, ddsIdx);
718 } else if (dictMode == ZSTD_dictMatchState) {
719 const U32* const dmsChainTable = dms->chainTable;
720 const U32 dmsChainSize = (1 << dms->cParams.chainLog);
721 const U32 dmsChainMask = dmsChainSize - 1;
722 const U32 dmsLowestIndex = dms->window.dictLimit;
723 const BYTE* const dmsBase = dms->window.base;
724 const BYTE* const dmsEnd = dms->window.nextSrc;
725 const U32 dmsSize = (U32)(dmsEnd - dmsBase);
726 const U32 dmsIndexDelta = dictLimit - dmsSize;
727 const U32 dmsMinChain = dmsSize > dmsChainSize ? dmsSize - dmsChainSize : 0;
728
729 matchIndex = dms->hashTable[ZSTD_hashPtr(ip, dms->cParams.hashLog, mls)];
730
731 for ( ; (matchIndex>=dmsLowestIndex) & (nbAttempts>0) ; nbAttempts--) {
732 size_t currentMl=0;
733 const BYTE* const match = dmsBase + matchIndex;
734 assert(match+4 <= dmsEnd);
735 if (MEM_read32(match) == MEM_read32(ip)) /* assumption : matchIndex <= dictLimit-4 (by table construction) */
736 currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dmsEnd, prefixStart) + 4;
737
738 /* save best solution */
739 if (currentMl > ml) {
740 ml = currentMl;
741 *offsetPtr = curr - (matchIndex + dmsIndexDelta) + ZSTD_REP_MOVE;
742 if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
743 }
744
745 if (matchIndex <= dmsMinChain) break;
746
747 matchIndex = dmsChainTable[matchIndex & dmsChainMask];
748 }
749 }
750
751 return ml;
752 }
753
754 /* *********************************
755 * (SIMD) Row-based matchfinder
756 ***********************************/
757 /* Constants for row-based hash */
758 #define ZSTD_ROW_HASH_TAG_OFFSET 16 /* byte offset of hashes in the match state's tagTable from the beginning of a row */
759 #define ZSTD_ROW_HASH_TAG_BITS 8 /* nb bits to use for the tag */
760 #define ZSTD_ROW_HASH_TAG_MASK ((1u << ZSTD_ROW_HASH_TAG_BITS) - 1)
761 #define ZSTD_ROW_HASH_MAX_ENTRIES 64 /* absolute maximum number of entries per row, for all configurations */
762
763 #define ZSTD_ROW_HASH_CACHE_MASK (ZSTD_ROW_HASH_CACHE_SIZE - 1)
764
765 typedef U64 ZSTD_VecMask; /* Clarifies when we are interacting with a U64 representing a mask of matches */
766
767 /* ZSTD_VecMask_next():
768 * Starting from the LSB, returns the idx of the next non-zero bit.
769 * Basically counting the nb of trailing zeroes.
770 */
ZSTD_VecMask_next(ZSTD_VecMask val)771 static U32 ZSTD_VecMask_next(ZSTD_VecMask val) {
772 assert(val != 0);
773 # if defined(_MSC_VER) && defined(_WIN64)
774 if (val != 0) {
775 unsigned long r;
776 _BitScanForward64(&r, val);
777 return (U32)(r);
778 } else {
779 /* Should not reach this code path */
780 __assume(0);
781 }
782 # elif (defined(__GNUC__) && ((__GNUC__ > 3) || ((__GNUC__ == 3) && (__GNUC_MINOR__ >= 4))))
783 if (sizeof(size_t) == 4) {
784 U32 mostSignificantWord = (U32)(val >> 32);
785 U32 leastSignificantWord = (U32)val;
786 if (leastSignificantWord == 0) {
787 return 32 + (U32)__builtin_ctz(mostSignificantWord);
788 } else {
789 return (U32)__builtin_ctz(leastSignificantWord);
790 }
791 } else {
792 return (U32)__builtin_ctzll(val);
793 }
794 # else
795 /* Software ctz version: http://aggregate.org/MAGIC/#Trailing%20Zero%20Count
796 * and: https://stackoverflow.com/questions/2709430/count-number-of-bits-in-a-64-bit-long-big-integer
797 */
798 val = ~val & (val - 1ULL); /* Lowest set bit mask */
799 val = val - ((val >> 1) & 0x5555555555555555);
800 val = (val & 0x3333333333333333ULL) + ((val >> 2) & 0x3333333333333333ULL);
801 return (U32)((((val + (val >> 4)) & 0xF0F0F0F0F0F0F0FULL) * 0x101010101010101ULL) >> 56);
802 # endif
803 }
804
805 /* ZSTD_rotateRight_*():
806 * Rotates a bitfield to the right by "count" bits.
807 * https://en.wikipedia.org/w/index.php?title=Circular_shift&oldid=991635599#Implementing_circular_shifts
808 */
809 FORCE_INLINE_TEMPLATE
ZSTD_rotateRight_U64(U64 const value,U32 count)810 U64 ZSTD_rotateRight_U64(U64 const value, U32 count) {
811 assert(count < 64);
812 count &= 0x3F; /* for fickle pattern recognition */
813 return (value >> count) | (U64)(value << ((0U - count) & 0x3F));
814 }
815
816 FORCE_INLINE_TEMPLATE
ZSTD_rotateRight_U32(U32 const value,U32 count)817 U32 ZSTD_rotateRight_U32(U32 const value, U32 count) {
818 assert(count < 32);
819 count &= 0x1F; /* for fickle pattern recognition */
820 return (value >> count) | (U32)(value << ((0U - count) & 0x1F));
821 }
822
823 FORCE_INLINE_TEMPLATE
ZSTD_rotateRight_U16(U16 const value,U32 count)824 U16 ZSTD_rotateRight_U16(U16 const value, U32 count) {
825 assert(count < 16);
826 count &= 0x0F; /* for fickle pattern recognition */
827 return (value >> count) | (U16)(value << ((0U - count) & 0x0F));
828 }
829
830 /* ZSTD_row_nextIndex():
831 * Returns the next index to insert at within a tagTable row, and updates the "head"
832 * value to reflect the update. Essentially cycles backwards from [0, {entries per row})
833 */
ZSTD_row_nextIndex(BYTE * const tagRow,U32 const rowMask)834 FORCE_INLINE_TEMPLATE U32 ZSTD_row_nextIndex(BYTE* const tagRow, U32 const rowMask) {
835 U32 const next = (*tagRow - 1) & rowMask;
836 *tagRow = (BYTE)next;
837 return next;
838 }
839
840 /* ZSTD_isAligned():
841 * Checks that a pointer is aligned to "align" bytes which must be a power of 2.
842 */
ZSTD_isAligned(void const * ptr,size_t align)843 MEM_STATIC int ZSTD_isAligned(void const* ptr, size_t align) {
844 assert((align & (align - 1)) == 0);
845 return (((size_t)ptr) & (align - 1)) == 0;
846 }
847
848 /* ZSTD_row_prefetch():
849 * Performs prefetching for the hashTable and tagTable at a given row.
850 */
ZSTD_row_prefetch(U32 const * hashTable,U16 const * tagTable,U32 const relRow,U32 const rowLog)851 FORCE_INLINE_TEMPLATE void ZSTD_row_prefetch(U32 const* hashTable, U16 const* tagTable, U32 const relRow, U32 const rowLog) {
852 PREFETCH_L1(hashTable + relRow);
853 if (rowLog >= 5) {
854 PREFETCH_L1(hashTable + relRow + 16);
855 /* Note: prefetching more of the hash table does not appear to be beneficial for 128-entry rows */
856 }
857 PREFETCH_L1(tagTable + relRow);
858 if (rowLog == 6) {
859 PREFETCH_L1(tagTable + relRow + 32);
860 }
861 assert(rowLog == 4 || rowLog == 5 || rowLog == 6);
862 assert(ZSTD_isAligned(hashTable + relRow, 64)); /* prefetched hash row always 64-byte aligned */
863 assert(ZSTD_isAligned(tagTable + relRow, (size_t)1 << rowLog)); /* prefetched tagRow sits on correct multiple of bytes (32,64,128) */
864 }
865
866 /* ZSTD_row_fillHashCache():
867 * Fill up the hash cache starting at idx, prefetching up to ZSTD_ROW_HASH_CACHE_SIZE entries,
868 * but not beyond iLimit.
869 */
ZSTD_row_fillHashCache(ZSTD_matchState_t * ms,const BYTE * base,U32 const rowLog,U32 const mls,U32 idx,const BYTE * const iLimit)870 FORCE_INLINE_TEMPLATE void ZSTD_row_fillHashCache(ZSTD_matchState_t* ms, const BYTE* base,
871 U32 const rowLog, U32 const mls,
872 U32 idx, const BYTE* const iLimit)
873 {
874 U32 const* const hashTable = ms->hashTable;
875 U16 const* const tagTable = ms->tagTable;
876 U32 const hashLog = ms->rowHashLog;
877 U32 const maxElemsToPrefetch = (base + idx) > iLimit ? 0 : (U32)(iLimit - (base + idx) + 1);
878 U32 const lim = idx + MIN(ZSTD_ROW_HASH_CACHE_SIZE, maxElemsToPrefetch);
879
880 for (; idx < lim; ++idx) {
881 U32 const hash = (U32)ZSTD_hashPtr(base + idx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls);
882 U32 const row = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
883 ZSTD_row_prefetch(hashTable, tagTable, row, rowLog);
884 ms->hashCache[idx & ZSTD_ROW_HASH_CACHE_MASK] = hash;
885 }
886
887 DEBUGLOG(6, "ZSTD_row_fillHashCache(): [%u %u %u %u %u %u %u %u]", ms->hashCache[0], ms->hashCache[1],
888 ms->hashCache[2], ms->hashCache[3], ms->hashCache[4],
889 ms->hashCache[5], ms->hashCache[6], ms->hashCache[7]);
890 }
891
892 /* ZSTD_row_nextCachedHash():
893 * Returns the hash of base + idx, and replaces the hash in the hash cache with the byte at
894 * base + idx + ZSTD_ROW_HASH_CACHE_SIZE. Also prefetches the appropriate rows from hashTable and tagTable.
895 */
ZSTD_row_nextCachedHash(U32 * cache,U32 const * hashTable,U16 const * tagTable,BYTE const * base,U32 idx,U32 const hashLog,U32 const rowLog,U32 const mls)896 FORCE_INLINE_TEMPLATE U32 ZSTD_row_nextCachedHash(U32* cache, U32 const* hashTable,
897 U16 const* tagTable, BYTE const* base,
898 U32 idx, U32 const hashLog,
899 U32 const rowLog, U32 const mls)
900 {
901 U32 const newHash = (U32)ZSTD_hashPtr(base+idx+ZSTD_ROW_HASH_CACHE_SIZE, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls);
902 U32 const row = (newHash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
903 ZSTD_row_prefetch(hashTable, tagTable, row, rowLog);
904 { U32 const hash = cache[idx & ZSTD_ROW_HASH_CACHE_MASK];
905 cache[idx & ZSTD_ROW_HASH_CACHE_MASK] = newHash;
906 return hash;
907 }
908 }
909
910 /* ZSTD_row_update_internalImpl():
911 * Updates the hash table with positions starting from updateStartIdx until updateEndIdx.
912 */
ZSTD_row_update_internalImpl(ZSTD_matchState_t * ms,U32 updateStartIdx,U32 const updateEndIdx,U32 const mls,U32 const rowLog,U32 const rowMask,U32 const useCache)913 FORCE_INLINE_TEMPLATE void ZSTD_row_update_internalImpl(ZSTD_matchState_t* ms,
914 U32 updateStartIdx, U32 const updateEndIdx,
915 U32 const mls, U32 const rowLog,
916 U32 const rowMask, U32 const useCache)
917 {
918 U32* const hashTable = ms->hashTable;
919 U16* const tagTable = ms->tagTable;
920 U32 const hashLog = ms->rowHashLog;
921 const BYTE* const base = ms->window.base;
922
923 DEBUGLOG(6, "ZSTD_row_update_internalImpl(): updateStartIdx=%u, updateEndIdx=%u", updateStartIdx, updateEndIdx);
924 for (; updateStartIdx < updateEndIdx; ++updateStartIdx) {
925 U32 const hash = useCache ? ZSTD_row_nextCachedHash(ms->hashCache, hashTable, tagTable, base, updateStartIdx, hashLog, rowLog, mls)
926 : (U32)ZSTD_hashPtr(base + updateStartIdx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls);
927 U32 const relRow = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
928 U32* const row = hashTable + relRow;
929 BYTE* tagRow = (BYTE*)(tagTable + relRow); /* Though tagTable is laid out as a table of U16, each tag is only 1 byte.
930 Explicit cast allows us to get exact desired position within each row */
931 U32 const pos = ZSTD_row_nextIndex(tagRow, rowMask);
932
933 assert(hash == ZSTD_hashPtr(base + updateStartIdx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls));
934 ((BYTE*)tagRow)[pos + ZSTD_ROW_HASH_TAG_OFFSET] = hash & ZSTD_ROW_HASH_TAG_MASK;
935 row[pos] = updateStartIdx;
936 }
937 }
938
939 /* ZSTD_row_update_internal():
940 * Inserts the byte at ip into the appropriate position in the hash table, and updates ms->nextToUpdate.
941 * Skips sections of long matches as is necessary.
942 */
ZSTD_row_update_internal(ZSTD_matchState_t * ms,const BYTE * ip,U32 const mls,U32 const rowLog,U32 const rowMask,U32 const useCache)943 FORCE_INLINE_TEMPLATE void ZSTD_row_update_internal(ZSTD_matchState_t* ms, const BYTE* ip,
944 U32 const mls, U32 const rowLog,
945 U32 const rowMask, U32 const useCache)
946 {
947 U32 idx = ms->nextToUpdate;
948 const BYTE* const base = ms->window.base;
949 const U32 target = (U32)(ip - base);
950 const U32 kSkipThreshold = 384;
951 const U32 kMaxMatchStartPositionsToUpdate = 96;
952 const U32 kMaxMatchEndPositionsToUpdate = 32;
953
954 if (useCache) {
955 /* Only skip positions when using hash cache, i.e.
956 * if we are loading a dict, don't skip anything.
957 * If we decide to skip, then we only update a set number
958 * of positions at the beginning and end of the match.
959 */
960 if (UNLIKELY(target - idx > kSkipThreshold)) {
961 U32 const bound = idx + kMaxMatchStartPositionsToUpdate;
962 ZSTD_row_update_internalImpl(ms, idx, bound, mls, rowLog, rowMask, useCache);
963 idx = target - kMaxMatchEndPositionsToUpdate;
964 ZSTD_row_fillHashCache(ms, base, rowLog, mls, idx, ip+1);
965 }
966 }
967 assert(target >= idx);
968 ZSTD_row_update_internalImpl(ms, idx, target, mls, rowLog, rowMask, useCache);
969 ms->nextToUpdate = target;
970 }
971
972 /* ZSTD_row_update():
973 * External wrapper for ZSTD_row_update_internal(). Used for filling the hashtable during dictionary
974 * processing.
975 */
ZSTD_row_update(ZSTD_matchState_t * const ms,const BYTE * ip)976 void ZSTD_row_update(ZSTD_matchState_t* const ms, const BYTE* ip) {
977 const U32 rowLog = BOUNDED(4, ms->cParams.searchLog, 6);
978 const U32 rowMask = (1u << rowLog) - 1;
979 const U32 mls = MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */);
980
981 DEBUGLOG(5, "ZSTD_row_update(), rowLog=%u", rowLog);
982 ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 0 /* dont use cache */);
983 }
984
985 #if defined(ZSTD_ARCH_X86_SSE2)
986 FORCE_INLINE_TEMPLATE ZSTD_VecMask
ZSTD_row_getSSEMask(int nbChunks,const BYTE * const src,const BYTE tag,const U32 head)987 ZSTD_row_getSSEMask(int nbChunks, const BYTE* const src, const BYTE tag, const U32 head)
988 {
989 const __m128i comparisonMask = _mm_set1_epi8((char)tag);
990 int matches[4] = {0};
991 int i;
992 assert(nbChunks == 1 || nbChunks == 2 || nbChunks == 4);
993 for (i=0; i<nbChunks; i++) {
994 const __m128i chunk = _mm_loadu_si128((const __m128i*)(const void*)(src + 16*i));
995 const __m128i equalMask = _mm_cmpeq_epi8(chunk, comparisonMask);
996 matches[i] = _mm_movemask_epi8(equalMask);
997 }
998 if (nbChunks == 1) return ZSTD_rotateRight_U16((U16)matches[0], head);
999 if (nbChunks == 2) return ZSTD_rotateRight_U32((U32)matches[1] << 16 | (U32)matches[0], head);
1000 assert(nbChunks == 4);
1001 return ZSTD_rotateRight_U64((U64)matches[3] << 48 | (U64)matches[2] << 32 | (U64)matches[1] << 16 | (U64)matches[0], head);
1002 }
1003 #endif
1004
1005 /* Returns a ZSTD_VecMask (U32) that has the nth bit set to 1 if the newly-computed "tag" matches
1006 * the hash at the nth position in a row of the tagTable.
1007 * Each row is a circular buffer beginning at the value of "head". So we must rotate the "matches" bitfield
1008 * to match up with the actual layout of the entries within the hashTable */
1009 FORCE_INLINE_TEMPLATE ZSTD_VecMask
ZSTD_row_getMatchMask(const BYTE * const tagRow,const BYTE tag,const U32 head,const U32 rowEntries)1010 ZSTD_row_getMatchMask(const BYTE* const tagRow, const BYTE tag, const U32 head, const U32 rowEntries)
1011 {
1012 const BYTE* const src = tagRow + ZSTD_ROW_HASH_TAG_OFFSET;
1013 assert((rowEntries == 16) || (rowEntries == 32) || rowEntries == 64);
1014 assert(rowEntries <= ZSTD_ROW_HASH_MAX_ENTRIES);
1015
1016 #if defined(ZSTD_ARCH_X86_SSE2)
1017
1018 return ZSTD_row_getSSEMask(rowEntries / 16, src, tag, head);
1019
1020 #else /* SW or NEON-LE */
1021
1022 # if defined(ZSTD_ARCH_ARM_NEON)
1023 /* This NEON path only works for little endian - otherwise use SWAR below */
1024 if (MEM_isLittleEndian()) {
1025 if (rowEntries == 16) {
1026 const uint8x16_t chunk = vld1q_u8(src);
1027 const uint16x8_t equalMask = vreinterpretq_u16_u8(vceqq_u8(chunk, vdupq_n_u8(tag)));
1028 const uint16x8_t t0 = vshlq_n_u16(equalMask, 7);
1029 const uint32x4_t t1 = vreinterpretq_u32_u16(vsriq_n_u16(t0, t0, 14));
1030 const uint64x2_t t2 = vreinterpretq_u64_u32(vshrq_n_u32(t1, 14));
1031 const uint8x16_t t3 = vreinterpretq_u8_u64(vsraq_n_u64(t2, t2, 28));
1032 const U16 hi = (U16)vgetq_lane_u8(t3, 8);
1033 const U16 lo = (U16)vgetq_lane_u8(t3, 0);
1034 return ZSTD_rotateRight_U16((hi << 8) | lo, head);
1035 } else if (rowEntries == 32) {
1036 const uint16x8x2_t chunk = vld2q_u16((const U16*)(const void*)src);
1037 const uint8x16_t chunk0 = vreinterpretq_u8_u16(chunk.val[0]);
1038 const uint8x16_t chunk1 = vreinterpretq_u8_u16(chunk.val[1]);
1039 const uint8x16_t equalMask0 = vceqq_u8(chunk0, vdupq_n_u8(tag));
1040 const uint8x16_t equalMask1 = vceqq_u8(chunk1, vdupq_n_u8(tag));
1041 const int8x8_t pack0 = vqmovn_s16(vreinterpretq_s16_u8(equalMask0));
1042 const int8x8_t pack1 = vqmovn_s16(vreinterpretq_s16_u8(equalMask1));
1043 const uint8x8_t t0 = vreinterpret_u8_s8(pack0);
1044 const uint8x8_t t1 = vreinterpret_u8_s8(pack1);
1045 const uint8x8_t t2 = vsri_n_u8(t1, t0, 2);
1046 const uint8x8x2_t t3 = vuzp_u8(t2, t0);
1047 const uint8x8_t t4 = vsri_n_u8(t3.val[1], t3.val[0], 4);
1048 const U32 matches = vget_lane_u32(vreinterpret_u32_u8(t4), 0);
1049 return ZSTD_rotateRight_U32(matches, head);
1050 } else { /* rowEntries == 64 */
1051 const uint8x16x4_t chunk = vld4q_u8(src);
1052 const uint8x16_t dup = vdupq_n_u8(tag);
1053 const uint8x16_t cmp0 = vceqq_u8(chunk.val[0], dup);
1054 const uint8x16_t cmp1 = vceqq_u8(chunk.val[1], dup);
1055 const uint8x16_t cmp2 = vceqq_u8(chunk.val[2], dup);
1056 const uint8x16_t cmp3 = vceqq_u8(chunk.val[3], dup);
1057
1058 const uint8x16_t t0 = vsriq_n_u8(cmp1, cmp0, 1);
1059 const uint8x16_t t1 = vsriq_n_u8(cmp3, cmp2, 1);
1060 const uint8x16_t t2 = vsriq_n_u8(t1, t0, 2);
1061 const uint8x16_t t3 = vsriq_n_u8(t2, t2, 4);
1062 const uint8x8_t t4 = vshrn_n_u16(vreinterpretq_u16_u8(t3), 4);
1063 const U64 matches = vget_lane_u64(vreinterpret_u64_u8(t4), 0);
1064 return ZSTD_rotateRight_U64(matches, head);
1065 }
1066 }
1067 # endif /* ZSTD_ARCH_ARM_NEON */
1068 /* SWAR */
1069 { const size_t chunkSize = sizeof(size_t);
1070 const size_t shiftAmount = ((chunkSize * 8) - chunkSize);
1071 const size_t xFF = ~((size_t)0);
1072 const size_t x01 = xFF / 0xFF;
1073 const size_t x80 = x01 << 7;
1074 const size_t splatChar = tag * x01;
1075 ZSTD_VecMask matches = 0;
1076 int i = rowEntries - chunkSize;
1077 assert((sizeof(size_t) == 4) || (sizeof(size_t) == 8));
1078 if (MEM_isLittleEndian()) { /* runtime check so have two loops */
1079 const size_t extractMagic = (xFF / 0x7F) >> chunkSize;
1080 do {
1081 size_t chunk = MEM_readST(&src[i]);
1082 chunk ^= splatChar;
1083 chunk = (((chunk | x80) - x01) | chunk) & x80;
1084 matches <<= chunkSize;
1085 matches |= (chunk * extractMagic) >> shiftAmount;
1086 i -= chunkSize;
1087 } while (i >= 0);
1088 } else { /* big endian: reverse bits during extraction */
1089 const size_t msb = xFF ^ (xFF >> 1);
1090 const size_t extractMagic = (msb / 0x1FF) | msb;
1091 do {
1092 size_t chunk = MEM_readST(&src[i]);
1093 chunk ^= splatChar;
1094 chunk = (((chunk | x80) - x01) | chunk) & x80;
1095 matches <<= chunkSize;
1096 matches |= ((chunk >> 7) * extractMagic) >> shiftAmount;
1097 i -= chunkSize;
1098 } while (i >= 0);
1099 }
1100 matches = ~matches;
1101 if (rowEntries == 16) {
1102 return ZSTD_rotateRight_U16((U16)matches, head);
1103 } else if (rowEntries == 32) {
1104 return ZSTD_rotateRight_U32((U32)matches, head);
1105 } else {
1106 return ZSTD_rotateRight_U64((U64)matches, head);
1107 }
1108 }
1109 #endif
1110 }
1111
1112 /* The high-level approach of the SIMD row based match finder is as follows:
1113 * - Figure out where to insert the new entry:
1114 * - Generate a hash from a byte along with an additional 1-byte "short hash". The additional byte is our "tag"
1115 * - The hashTable is effectively split into groups or "rows" of 16 or 32 entries of U32, and the hash determines
1116 * which row to insert into.
1117 * - Determine the correct position within the row to insert the entry into. Each row of 16 or 32 can
1118 * be considered as a circular buffer with a "head" index that resides in the tagTable.
1119 * - Also insert the "tag" into the equivalent row and position in the tagTable.
1120 * - Note: The tagTable has 17 or 33 1-byte entries per row, due to 16 or 32 tags, and 1 "head" entry.
1121 * The 17 or 33 entry rows are spaced out to occur every 32 or 64 bytes, respectively,
1122 * for alignment/performance reasons, leaving some bytes unused.
1123 * - Use SIMD to efficiently compare the tags in the tagTable to the 1-byte "short hash" and
1124 * generate a bitfield that we can cycle through to check the collisions in the hash table.
1125 * - Pick the longest match.
1126 */
1127 FORCE_INLINE_TEMPLATE
ZSTD_RowFindBestMatch(ZSTD_matchState_t * ms,const BYTE * const ip,const BYTE * const iLimit,size_t * offsetPtr,const U32 mls,const ZSTD_dictMode_e dictMode,const U32 rowLog)1128 size_t ZSTD_RowFindBestMatch(
1129 ZSTD_matchState_t* ms,
1130 const BYTE* const ip, const BYTE* const iLimit,
1131 size_t* offsetPtr,
1132 const U32 mls, const ZSTD_dictMode_e dictMode,
1133 const U32 rowLog)
1134 {
1135 U32* const hashTable = ms->hashTable;
1136 U16* const tagTable = ms->tagTable;
1137 U32* const hashCache = ms->hashCache;
1138 const U32 hashLog = ms->rowHashLog;
1139 const ZSTD_compressionParameters* const cParams = &ms->cParams;
1140 const BYTE* const base = ms->window.base;
1141 const BYTE* const dictBase = ms->window.dictBase;
1142 const U32 dictLimit = ms->window.dictLimit;
1143 const BYTE* const prefixStart = base + dictLimit;
1144 const BYTE* const dictEnd = dictBase + dictLimit;
1145 const U32 curr = (U32)(ip-base);
1146 const U32 maxDistance = 1U << cParams->windowLog;
1147 const U32 lowestValid = ms->window.lowLimit;
1148 const U32 withinMaxDistance = (curr - lowestValid > maxDistance) ? curr - maxDistance : lowestValid;
1149 const U32 isDictionary = (ms->loadedDictEnd != 0);
1150 const U32 lowLimit = isDictionary ? lowestValid : withinMaxDistance;
1151 const U32 rowEntries = (1U << rowLog);
1152 const U32 rowMask = rowEntries - 1;
1153 const U32 cappedSearchLog = MIN(cParams->searchLog, rowLog); /* nb of searches is capped at nb entries per row */
1154 U32 nbAttempts = 1U << cappedSearchLog;
1155 size_t ml=4-1;
1156
1157 /* DMS/DDS variables that may be referenced laster */
1158 const ZSTD_matchState_t* const dms = ms->dictMatchState;
1159
1160 /* Initialize the following variables to satisfy static analyzer */
1161 size_t ddsIdx = 0;
1162 U32 ddsExtraAttempts = 0; /* cctx hash tables are limited in searches, but allow extra searches into DDS */
1163 U32 dmsTag = 0;
1164 U32* dmsRow = NULL;
1165 BYTE* dmsTagRow = NULL;
1166
1167 if (dictMode == ZSTD_dedicatedDictSearch) {
1168 const U32 ddsHashLog = dms->cParams.hashLog - ZSTD_LAZY_DDSS_BUCKET_LOG;
1169 { /* Prefetch DDS hashtable entry */
1170 ddsIdx = ZSTD_hashPtr(ip, ddsHashLog, mls) << ZSTD_LAZY_DDSS_BUCKET_LOG;
1171 PREFETCH_L1(&dms->hashTable[ddsIdx]);
1172 }
1173 ddsExtraAttempts = cParams->searchLog > rowLog ? 1U << (cParams->searchLog - rowLog) : 0;
1174 }
1175
1176 if (dictMode == ZSTD_dictMatchState) {
1177 /* Prefetch DMS rows */
1178 U32* const dmsHashTable = dms->hashTable;
1179 U16* const dmsTagTable = dms->tagTable;
1180 U32 const dmsHash = (U32)ZSTD_hashPtr(ip, dms->rowHashLog + ZSTD_ROW_HASH_TAG_BITS, mls);
1181 U32 const dmsRelRow = (dmsHash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
1182 dmsTag = dmsHash & ZSTD_ROW_HASH_TAG_MASK;
1183 dmsTagRow = (BYTE*)(dmsTagTable + dmsRelRow);
1184 dmsRow = dmsHashTable + dmsRelRow;
1185 ZSTD_row_prefetch(dmsHashTable, dmsTagTable, dmsRelRow, rowLog);
1186 }
1187
1188 /* Update the hashTable and tagTable up to (but not including) ip */
1189 ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 1 /* useCache */);
1190 { /* Get the hash for ip, compute the appropriate row */
1191 U32 const hash = ZSTD_row_nextCachedHash(hashCache, hashTable, tagTable, base, curr, hashLog, rowLog, mls);
1192 U32 const relRow = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
1193 U32 const tag = hash & ZSTD_ROW_HASH_TAG_MASK;
1194 U32* const row = hashTable + relRow;
1195 BYTE* tagRow = (BYTE*)(tagTable + relRow);
1196 U32 const head = *tagRow & rowMask;
1197 U32 matchBuffer[ZSTD_ROW_HASH_MAX_ENTRIES];
1198 size_t numMatches = 0;
1199 size_t currMatch = 0;
1200 ZSTD_VecMask matches = ZSTD_row_getMatchMask(tagRow, (BYTE)tag, head, rowEntries);
1201
1202 /* Cycle through the matches and prefetch */
1203 for (; (matches > 0) && (nbAttempts > 0); --nbAttempts, matches &= (matches - 1)) {
1204 U32 const matchPos = (head + ZSTD_VecMask_next(matches)) & rowMask;
1205 U32 const matchIndex = row[matchPos];
1206 assert(numMatches < rowEntries);
1207 if (matchIndex < lowLimit)
1208 break;
1209 if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) {
1210 PREFETCH_L1(base + matchIndex);
1211 } else {
1212 PREFETCH_L1(dictBase + matchIndex);
1213 }
1214 matchBuffer[numMatches++] = matchIndex;
1215 }
1216
1217 /* Speed opt: insert current byte into hashtable too. This allows us to avoid one iteration of the loop
1218 in ZSTD_row_update_internal() at the next search. */
1219 {
1220 U32 const pos = ZSTD_row_nextIndex(tagRow, rowMask);
1221 tagRow[pos + ZSTD_ROW_HASH_TAG_OFFSET] = (BYTE)tag;
1222 row[pos] = ms->nextToUpdate++;
1223 }
1224
1225 /* Return the longest match */
1226 for (; currMatch < numMatches; ++currMatch) {
1227 U32 const matchIndex = matchBuffer[currMatch];
1228 size_t currentMl=0;
1229 assert(matchIndex < curr);
1230 assert(matchIndex >= lowLimit);
1231
1232 if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) {
1233 const BYTE* const match = base + matchIndex;
1234 assert(matchIndex >= dictLimit); /* ensures this is true if dictMode != ZSTD_extDict */
1235 if (match[ml] == ip[ml]) /* potentially better */
1236 currentMl = ZSTD_count(ip, match, iLimit);
1237 } else {
1238 const BYTE* const match = dictBase + matchIndex;
1239 assert(match+4 <= dictEnd);
1240 if (MEM_read32(match) == MEM_read32(ip)) /* assumption : matchIndex <= dictLimit-4 (by table construction) */
1241 currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dictEnd, prefixStart) + 4;
1242 }
1243
1244 /* Save best solution */
1245 if (currentMl > ml) {
1246 ml = currentMl;
1247 *offsetPtr = curr - matchIndex + ZSTD_REP_MOVE;
1248 if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
1249 }
1250 }
1251 }
1252
1253 assert(nbAttempts <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */
1254 if (dictMode == ZSTD_dedicatedDictSearch) {
1255 ml = ZSTD_dedicatedDictSearch_lazy_search(offsetPtr, ml, nbAttempts + ddsExtraAttempts, dms,
1256 ip, iLimit, prefixStart, curr, dictLimit, ddsIdx);
1257 } else if (dictMode == ZSTD_dictMatchState) {
1258 /* TODO: Measure and potentially add prefetching to DMS */
1259 const U32 dmsLowestIndex = dms->window.dictLimit;
1260 const BYTE* const dmsBase = dms->window.base;
1261 const BYTE* const dmsEnd = dms->window.nextSrc;
1262 const U32 dmsSize = (U32)(dmsEnd - dmsBase);
1263 const U32 dmsIndexDelta = dictLimit - dmsSize;
1264
1265 { U32 const head = *dmsTagRow & rowMask;
1266 U32 matchBuffer[ZSTD_ROW_HASH_MAX_ENTRIES];
1267 size_t numMatches = 0;
1268 size_t currMatch = 0;
1269 ZSTD_VecMask matches = ZSTD_row_getMatchMask(dmsTagRow, (BYTE)dmsTag, head, rowEntries);
1270
1271 for (; (matches > 0) && (nbAttempts > 0); --nbAttempts, matches &= (matches - 1)) {
1272 U32 const matchPos = (head + ZSTD_VecMask_next(matches)) & rowMask;
1273 U32 const matchIndex = dmsRow[matchPos];
1274 if (matchIndex < dmsLowestIndex)
1275 break;
1276 PREFETCH_L1(dmsBase + matchIndex);
1277 matchBuffer[numMatches++] = matchIndex;
1278 }
1279
1280 /* Return the longest match */
1281 for (; currMatch < numMatches; ++currMatch) {
1282 U32 const matchIndex = matchBuffer[currMatch];
1283 size_t currentMl=0;
1284 assert(matchIndex >= dmsLowestIndex);
1285 assert(matchIndex < curr);
1286
1287 { const BYTE* const match = dmsBase + matchIndex;
1288 assert(match+4 <= dmsEnd);
1289 if (MEM_read32(match) == MEM_read32(ip))
1290 currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dmsEnd, prefixStart) + 4;
1291 }
1292
1293 if (currentMl > ml) {
1294 ml = currentMl;
1295 *offsetPtr = curr - (matchIndex + dmsIndexDelta) + ZSTD_REP_MOVE;
1296 if (ip+currentMl == iLimit) break;
1297 }
1298 }
1299 }
1300 }
1301 return ml;
1302 }
1303
1304
1305 typedef size_t (*searchMax_f)(
1306 ZSTD_matchState_t* ms,
1307 const BYTE* ip, const BYTE* iLimit, size_t* offsetPtr);
1308
1309 /**
1310 * This struct contains the functions necessary for lazy to search.
1311 * Currently, that is only searchMax. However, it is still valuable to have the
1312 * VTable because this makes it easier to add more functions to the VTable later.
1313 *
1314 * TODO: The start of the search function involves loading and calculating a
1315 * bunch of constants from the ZSTD_matchState_t. These computations could be
1316 * done in an initialization function, and saved somewhere in the match state.
1317 * Then we could pass a pointer to the saved state instead of the match state,
1318 * and avoid duplicate computations.
1319 *
1320 * TODO: Move the match re-winding into searchMax. This improves compression
1321 * ratio, and unlocks further simplifications with the next TODO.
1322 *
1323 * TODO: Try moving the repcode search into searchMax. After the re-winding
1324 * and repcode search are in searchMax, there is no more logic in the match
1325 * finder loop that requires knowledge about the dictMode. So we should be
1326 * able to avoid force inlining it, and we can join the extDict loop with
1327 * the single segment loop. It should go in searchMax instead of its own
1328 * function to avoid having multiple virtual function calls per search.
1329 */
1330 typedef struct {
1331 searchMax_f searchMax;
1332 } ZSTD_LazyVTable;
1333
1334 #define GEN_ZSTD_BT_VTABLE(dictMode, mls) \
1335 static size_t ZSTD_BtFindBestMatch_##dictMode##_##mls( \
1336 ZSTD_matchState_t* ms, \
1337 const BYTE* ip, const BYTE* const iLimit, \
1338 size_t* offsetPtr) \
1339 { \
1340 assert(MAX(4, MIN(6, ms->cParams.minMatch)) == mls); \
1341 return ZSTD_BtFindBestMatch(ms, ip, iLimit, offsetPtr, mls, ZSTD_##dictMode); \
1342 } \
1343 static const ZSTD_LazyVTable ZSTD_BtVTable_##dictMode##_##mls = { \
1344 ZSTD_BtFindBestMatch_##dictMode##_##mls \
1345 };
1346
1347 #define GEN_ZSTD_HC_VTABLE(dictMode, mls) \
1348 static size_t ZSTD_HcFindBestMatch_##dictMode##_##mls( \
1349 ZSTD_matchState_t* ms, \
1350 const BYTE* ip, const BYTE* const iLimit, \
1351 size_t* offsetPtr) \
1352 { \
1353 assert(MAX(4, MIN(6, ms->cParams.minMatch)) == mls); \
1354 return ZSTD_HcFindBestMatch(ms, ip, iLimit, offsetPtr, mls, ZSTD_##dictMode); \
1355 } \
1356 static const ZSTD_LazyVTable ZSTD_HcVTable_##dictMode##_##mls = { \
1357 ZSTD_HcFindBestMatch_##dictMode##_##mls \
1358 };
1359
1360 #define GEN_ZSTD_ROW_VTABLE(dictMode, mls, rowLog) \
1361 static size_t ZSTD_RowFindBestMatch_##dictMode##_##mls##_##rowLog( \
1362 ZSTD_matchState_t* ms, \
1363 const BYTE* ip, const BYTE* const iLimit, \
1364 size_t* offsetPtr) \
1365 { \
1366 assert(MAX(4, MIN(6, ms->cParams.minMatch)) == mls); \
1367 assert(MAX(4, MIN(6, ms->cParams.searchLog)) == rowLog); \
1368 return ZSTD_RowFindBestMatch(ms, ip, iLimit, offsetPtr, mls, ZSTD_##dictMode, rowLog); \
1369 } \
1370 static const ZSTD_LazyVTable ZSTD_RowVTable_##dictMode##_##mls##_##rowLog = { \
1371 ZSTD_RowFindBestMatch_##dictMode##_##mls##_##rowLog \
1372 };
1373
1374 #define ZSTD_FOR_EACH_ROWLOG(X, dictMode, mls) \
1375 X(dictMode, mls, 4) \
1376 X(dictMode, mls, 5) \
1377 X(dictMode, mls, 6)
1378
1379 #define ZSTD_FOR_EACH_MLS_ROWLOG(X, dictMode) \
1380 ZSTD_FOR_EACH_ROWLOG(X, dictMode, 4) \
1381 ZSTD_FOR_EACH_ROWLOG(X, dictMode, 5) \
1382 ZSTD_FOR_EACH_ROWLOG(X, dictMode, 6)
1383
1384 #define ZSTD_FOR_EACH_MLS(X, dictMode) \
1385 X(dictMode, 4) \
1386 X(dictMode, 5) \
1387 X(dictMode, 6)
1388
1389 #define ZSTD_FOR_EACH_DICT_MODE(X, ...) \
1390 X(__VA_ARGS__, noDict) \
1391 X(__VA_ARGS__, extDict) \
1392 X(__VA_ARGS__, dictMatchState) \
1393 X(__VA_ARGS__, dedicatedDictSearch)
1394
1395 /* Generate Row VTables for each combination of (dictMode, mls, rowLog) */
1396 ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS_ROWLOG, GEN_ZSTD_ROW_VTABLE)
1397 /* Generate Binary Tree VTables for each combination of (dictMode, mls) */
1398 ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS, GEN_ZSTD_BT_VTABLE)
1399 /* Generate Hash Chain VTables for each combination of (dictMode, mls) */
1400 ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS, GEN_ZSTD_HC_VTABLE)
1401
1402 #define GEN_ZSTD_BT_VTABLE_ARRAY(dictMode) \
1403 { \
1404 &ZSTD_BtVTable_##dictMode##_4, \
1405 &ZSTD_BtVTable_##dictMode##_5, \
1406 &ZSTD_BtVTable_##dictMode##_6 \
1407 }
1408
1409 #define GEN_ZSTD_HC_VTABLE_ARRAY(dictMode) \
1410 { \
1411 &ZSTD_HcVTable_##dictMode##_4, \
1412 &ZSTD_HcVTable_##dictMode##_5, \
1413 &ZSTD_HcVTable_##dictMode##_6 \
1414 }
1415
1416 #define GEN_ZSTD_ROW_VTABLE_ARRAY_(dictMode, mls) \
1417 { \
1418 &ZSTD_RowVTable_##dictMode##_##mls##_4, \
1419 &ZSTD_RowVTable_##dictMode##_##mls##_5, \
1420 &ZSTD_RowVTable_##dictMode##_##mls##_6 \
1421 }
1422
1423 #define GEN_ZSTD_ROW_VTABLE_ARRAY(dictMode) \
1424 { \
1425 GEN_ZSTD_ROW_VTABLE_ARRAY_(dictMode, 4), \
1426 GEN_ZSTD_ROW_VTABLE_ARRAY_(dictMode, 5), \
1427 GEN_ZSTD_ROW_VTABLE_ARRAY_(dictMode, 6) \
1428 }
1429
1430 #define GEN_ZSTD_VTABLE_ARRAY(X) \
1431 { \
1432 X(noDict), \
1433 X(extDict), \
1434 X(dictMatchState), \
1435 X(dedicatedDictSearch) \
1436 }
1437
1438 /* *******************************
1439 * Common parser - lazy strategy
1440 *********************************/
1441 typedef enum { search_hashChain=0, search_binaryTree=1, search_rowHash=2 } searchMethod_e;
1442
1443 /**
1444 * This table is indexed first by the four ZSTD_dictMode_e values, and then
1445 * by the two searchMethod_e values. NULLs are placed for configurations
1446 * that should never occur (extDict modes go to the other implementation
1447 * below and there is no DDSS for binary tree search yet).
1448 */
1449
1450 static ZSTD_LazyVTable const*
ZSTD_selectLazyVTable(ZSTD_matchState_t const * ms,searchMethod_e searchMethod,ZSTD_dictMode_e dictMode)1451 ZSTD_selectLazyVTable(ZSTD_matchState_t const* ms, searchMethod_e searchMethod, ZSTD_dictMode_e dictMode)
1452 {
1453 /* Fill the Hc/Bt VTable arrays with the right functions for the (dictMode, mls) combination. */
1454 ZSTD_LazyVTable const* const hcVTables[4][3] = GEN_ZSTD_VTABLE_ARRAY(GEN_ZSTD_HC_VTABLE_ARRAY);
1455 ZSTD_LazyVTable const* const btVTables[4][3] = GEN_ZSTD_VTABLE_ARRAY(GEN_ZSTD_BT_VTABLE_ARRAY);
1456 /* Fill the Row VTable array with the right functions for the (dictMode, mls, rowLog) combination. */
1457 ZSTD_LazyVTable const* const rowVTables[4][3][3] = GEN_ZSTD_VTABLE_ARRAY(GEN_ZSTD_ROW_VTABLE_ARRAY);
1458
1459 U32 const mls = MAX(4, MIN(6, ms->cParams.minMatch));
1460 U32 const rowLog = MAX(4, MIN(6, ms->cParams.searchLog));
1461 switch (searchMethod) {
1462 case search_hashChain:
1463 return hcVTables[dictMode][mls - 4];
1464 case search_binaryTree:
1465 return btVTables[dictMode][mls - 4];
1466 case search_rowHash:
1467 return rowVTables[dictMode][mls - 4][rowLog - 4];
1468 default:
1469 return NULL;
1470 }
1471 }
1472
1473 FORCE_INLINE_TEMPLATE size_t
ZSTD_compressBlock_lazy_generic(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],const void * src,size_t srcSize,const searchMethod_e searchMethod,const U32 depth,ZSTD_dictMode_e const dictMode)1474 ZSTD_compressBlock_lazy_generic(
1475 ZSTD_matchState_t* ms, seqStore_t* seqStore,
1476 U32 rep[ZSTD_REP_NUM],
1477 const void* src, size_t srcSize,
1478 const searchMethod_e searchMethod, const U32 depth,
1479 ZSTD_dictMode_e const dictMode)
1480 {
1481 const BYTE* const istart = (const BYTE*)src;
1482 const BYTE* ip = istart;
1483 const BYTE* anchor = istart;
1484 const BYTE* const iend = istart + srcSize;
1485 const BYTE* const ilimit = (searchMethod == search_rowHash) ? iend - 8 - ZSTD_ROW_HASH_CACHE_SIZE : iend - 8;
1486 const BYTE* const base = ms->window.base;
1487 const U32 prefixLowestIndex = ms->window.dictLimit;
1488 const BYTE* const prefixLowest = base + prefixLowestIndex;
1489
1490 searchMax_f const searchMax = ZSTD_selectLazyVTable(ms, searchMethod, dictMode)->searchMax;
1491 U32 offset_1 = rep[0], offset_2 = rep[1], savedOffset=0;
1492
1493 const int isDMS = dictMode == ZSTD_dictMatchState;
1494 const int isDDS = dictMode == ZSTD_dedicatedDictSearch;
1495 const int isDxS = isDMS || isDDS;
1496 const ZSTD_matchState_t* const dms = ms->dictMatchState;
1497 const U32 dictLowestIndex = isDxS ? dms->window.dictLimit : 0;
1498 const BYTE* const dictBase = isDxS ? dms->window.base : NULL;
1499 const BYTE* const dictLowest = isDxS ? dictBase + dictLowestIndex : NULL;
1500 const BYTE* const dictEnd = isDxS ? dms->window.nextSrc : NULL;
1501 const U32 dictIndexDelta = isDxS ?
1502 prefixLowestIndex - (U32)(dictEnd - dictBase) :
1503 0;
1504 const U32 dictAndPrefixLength = (U32)((ip - prefixLowest) + (dictEnd - dictLowest));
1505
1506 assert(searchMax != NULL);
1507
1508 DEBUGLOG(5, "ZSTD_compressBlock_lazy_generic (dictMode=%u) (searchFunc=%u)", (U32)dictMode, (U32)searchMethod);
1509 ip += (dictAndPrefixLength == 0);
1510 if (dictMode == ZSTD_noDict) {
1511 U32 const curr = (U32)(ip - base);
1512 U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, curr, ms->cParams.windowLog);
1513 U32 const maxRep = curr - windowLow;
1514 if (offset_2 > maxRep) savedOffset = offset_2, offset_2 = 0;
1515 if (offset_1 > maxRep) savedOffset = offset_1, offset_1 = 0;
1516 }
1517 if (isDxS) {
1518 /* dictMatchState repCode checks don't currently handle repCode == 0
1519 * disabling. */
1520 assert(offset_1 <= dictAndPrefixLength);
1521 assert(offset_2 <= dictAndPrefixLength);
1522 }
1523
1524 if (searchMethod == search_rowHash) {
1525 const U32 rowLog = MAX(4, MIN(6, ms->cParams.searchLog));
1526 ZSTD_row_fillHashCache(ms, base, rowLog,
1527 MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */),
1528 ms->nextToUpdate, ilimit);
1529 }
1530
1531 /* Match Loop */
1532 #if defined(__GNUC__) && defined(__x86_64__)
1533 /* I've measured random a 5% speed loss on levels 5 & 6 (greedy) when the
1534 * code alignment is perturbed. To fix the instability align the loop on 32-bytes.
1535 */
1536 __asm__(".p2align 5");
1537 #endif
1538 while (ip < ilimit) {
1539 size_t matchLength=0;
1540 size_t offset=0;
1541 const BYTE* start=ip+1;
1542
1543 /* check repCode */
1544 if (isDxS) {
1545 const U32 repIndex = (U32)(ip - base) + 1 - offset_1;
1546 const BYTE* repMatch = ((dictMode == ZSTD_dictMatchState || dictMode == ZSTD_dedicatedDictSearch)
1547 && repIndex < prefixLowestIndex) ?
1548 dictBase + (repIndex - dictIndexDelta) :
1549 base + repIndex;
1550 if (((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
1551 && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
1552 const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
1553 matchLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
1554 if (depth==0) goto _storeSequence;
1555 }
1556 }
1557 if ( dictMode == ZSTD_noDict
1558 && ((offset_1 > 0) & (MEM_read32(ip+1-offset_1) == MEM_read32(ip+1)))) {
1559 matchLength = ZSTD_count(ip+1+4, ip+1+4-offset_1, iend) + 4;
1560 if (depth==0) goto _storeSequence;
1561 }
1562
1563 /* first search (depth 0) */
1564 { size_t offsetFound = 999999999;
1565 size_t const ml2 = searchMax(ms, ip, iend, &offsetFound);
1566 if (ml2 > matchLength)
1567 matchLength = ml2, start = ip, offset=offsetFound;
1568 }
1569
1570 if (matchLength < 4) {
1571 ip += ((ip-anchor) >> kSearchStrength) + 1; /* jump faster over incompressible sections */
1572 continue;
1573 }
1574
1575 /* let's try to find a better solution */
1576 if (depth>=1)
1577 while (ip<ilimit) {
1578 ip ++;
1579 if ( (dictMode == ZSTD_noDict)
1580 && (offset) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
1581 size_t const mlRep = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4;
1582 int const gain2 = (int)(mlRep * 3);
1583 int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offset+1) + 1);
1584 if ((mlRep >= 4) && (gain2 > gain1))
1585 matchLength = mlRep, offset = 0, start = ip;
1586 }
1587 if (isDxS) {
1588 const U32 repIndex = (U32)(ip - base) - offset_1;
1589 const BYTE* repMatch = repIndex < prefixLowestIndex ?
1590 dictBase + (repIndex - dictIndexDelta) :
1591 base + repIndex;
1592 if (((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
1593 && (MEM_read32(repMatch) == MEM_read32(ip)) ) {
1594 const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
1595 size_t const mlRep = ZSTD_count_2segments(ip+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
1596 int const gain2 = (int)(mlRep * 3);
1597 int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offset+1) + 1);
1598 if ((mlRep >= 4) && (gain2 > gain1))
1599 matchLength = mlRep, offset = 0, start = ip;
1600 }
1601 }
1602 { size_t offset2=999999999;
1603 size_t const ml2 = searchMax(ms, ip, iend, &offset2);
1604 int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1)); /* raw approx */
1605 int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 4);
1606 if ((ml2 >= 4) && (gain2 > gain1)) {
1607 matchLength = ml2, offset = offset2, start = ip;
1608 continue; /* search a better one */
1609 } }
1610
1611 /* let's find an even better one */
1612 if ((depth==2) && (ip<ilimit)) {
1613 ip ++;
1614 if ( (dictMode == ZSTD_noDict)
1615 && (offset) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
1616 size_t const mlRep = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4;
1617 int const gain2 = (int)(mlRep * 4);
1618 int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 1);
1619 if ((mlRep >= 4) && (gain2 > gain1))
1620 matchLength = mlRep, offset = 0, start = ip;
1621 }
1622 if (isDxS) {
1623 const U32 repIndex = (U32)(ip - base) - offset_1;
1624 const BYTE* repMatch = repIndex < prefixLowestIndex ?
1625 dictBase + (repIndex - dictIndexDelta) :
1626 base + repIndex;
1627 if (((U32)((prefixLowestIndex-1) - repIndex) >= 3 /* intentional underflow */)
1628 && (MEM_read32(repMatch) == MEM_read32(ip)) ) {
1629 const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
1630 size_t const mlRep = ZSTD_count_2segments(ip+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
1631 int const gain2 = (int)(mlRep * 4);
1632 int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 1);
1633 if ((mlRep >= 4) && (gain2 > gain1))
1634 matchLength = mlRep, offset = 0, start = ip;
1635 }
1636 }
1637 { size_t offset2=999999999;
1638 size_t const ml2 = searchMax(ms, ip, iend, &offset2);
1639 int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1)); /* raw approx */
1640 int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 7);
1641 if ((ml2 >= 4) && (gain2 > gain1)) {
1642 matchLength = ml2, offset = offset2, start = ip;
1643 continue;
1644 } } }
1645 break; /* nothing found : store previous solution */
1646 }
1647
1648 /* NOTE:
1649 * start[-offset+ZSTD_REP_MOVE-1] is undefined behavior.
1650 * (-offset+ZSTD_REP_MOVE-1) is unsigned, and is added to start, which
1651 * overflows the pointer, which is undefined behavior.
1652 */
1653 /* catch up */
1654 if (offset) {
1655 if (dictMode == ZSTD_noDict) {
1656 while ( ((start > anchor) & (start - (offset-ZSTD_REP_MOVE) > prefixLowest))
1657 && (start[-1] == (start-(offset-ZSTD_REP_MOVE))[-1]) ) /* only search for offset within prefix */
1658 { start--; matchLength++; }
1659 }
1660 if (isDxS) {
1661 U32 const matchIndex = (U32)((size_t)(start-base) - (offset - ZSTD_REP_MOVE));
1662 const BYTE* match = (matchIndex < prefixLowestIndex) ? dictBase + matchIndex - dictIndexDelta : base + matchIndex;
1663 const BYTE* const mStart = (matchIndex < prefixLowestIndex) ? dictLowest : prefixLowest;
1664 while ((start>anchor) && (match>mStart) && (start[-1] == match[-1])) { start--; match--; matchLength++; } /* catch up */
1665 }
1666 offset_2 = offset_1; offset_1 = (U32)(offset - ZSTD_REP_MOVE);
1667 }
1668 /* store sequence */
1669 _storeSequence:
1670 { size_t const litLength = (size_t)(start - anchor);
1671 ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offset, matchLength-MINMATCH);
1672 anchor = ip = start + matchLength;
1673 }
1674
1675 /* check immediate repcode */
1676 if (isDxS) {
1677 while (ip <= ilimit) {
1678 U32 const current2 = (U32)(ip-base);
1679 U32 const repIndex = current2 - offset_2;
1680 const BYTE* repMatch = repIndex < prefixLowestIndex ?
1681 dictBase - dictIndexDelta + repIndex :
1682 base + repIndex;
1683 if ( ((U32)((prefixLowestIndex-1) - (U32)repIndex) >= 3 /* intentional overflow */)
1684 && (MEM_read32(repMatch) == MEM_read32(ip)) ) {
1685 const BYTE* const repEnd2 = repIndex < prefixLowestIndex ? dictEnd : iend;
1686 matchLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd2, prefixLowest) + 4;
1687 offset = offset_2; offset_2 = offset_1; offset_1 = (U32)offset; /* swap offset_2 <=> offset_1 */
1688 ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, matchLength-MINMATCH);
1689 ip += matchLength;
1690 anchor = ip;
1691 continue;
1692 }
1693 break;
1694 }
1695 }
1696
1697 if (dictMode == ZSTD_noDict) {
1698 while ( ((ip <= ilimit) & (offset_2>0))
1699 && (MEM_read32(ip) == MEM_read32(ip - offset_2)) ) {
1700 /* store sequence */
1701 matchLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4;
1702 offset = offset_2; offset_2 = offset_1; offset_1 = (U32)offset; /* swap repcodes */
1703 ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, matchLength-MINMATCH);
1704 ip += matchLength;
1705 anchor = ip;
1706 continue; /* faster when present ... (?) */
1707 } } }
1708
1709 /* Save reps for next block */
1710 rep[0] = offset_1 ? offset_1 : savedOffset;
1711 rep[1] = offset_2 ? offset_2 : savedOffset;
1712
1713 /* Return the last literals size */
1714 return (size_t)(iend - anchor);
1715 }
1716
1717
ZSTD_compressBlock_btlazy2(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)1718 size_t ZSTD_compressBlock_btlazy2(
1719 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1720 void const* src, size_t srcSize)
1721 {
1722 return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2, ZSTD_noDict);
1723 }
1724
ZSTD_compressBlock_lazy2(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)1725 size_t ZSTD_compressBlock_lazy2(
1726 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1727 void const* src, size_t srcSize)
1728 {
1729 return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_noDict);
1730 }
1731
ZSTD_compressBlock_lazy(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)1732 size_t ZSTD_compressBlock_lazy(
1733 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1734 void const* src, size_t srcSize)
1735 {
1736 return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_noDict);
1737 }
1738
ZSTD_compressBlock_greedy(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)1739 size_t ZSTD_compressBlock_greedy(
1740 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1741 void const* src, size_t srcSize)
1742 {
1743 return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_noDict);
1744 }
1745
ZSTD_compressBlock_btlazy2_dictMatchState(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)1746 size_t ZSTD_compressBlock_btlazy2_dictMatchState(
1747 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1748 void const* src, size_t srcSize)
1749 {
1750 return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2, ZSTD_dictMatchState);
1751 }
1752
ZSTD_compressBlock_lazy2_dictMatchState(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)1753 size_t ZSTD_compressBlock_lazy2_dictMatchState(
1754 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1755 void const* src, size_t srcSize)
1756 {
1757 return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_dictMatchState);
1758 }
1759
ZSTD_compressBlock_lazy_dictMatchState(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)1760 size_t ZSTD_compressBlock_lazy_dictMatchState(
1761 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1762 void const* src, size_t srcSize)
1763 {
1764 return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_dictMatchState);
1765 }
1766
ZSTD_compressBlock_greedy_dictMatchState(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)1767 size_t ZSTD_compressBlock_greedy_dictMatchState(
1768 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1769 void const* src, size_t srcSize)
1770 {
1771 return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_dictMatchState);
1772 }
1773
1774
ZSTD_compressBlock_lazy2_dedicatedDictSearch(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)1775 size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch(
1776 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1777 void const* src, size_t srcSize)
1778 {
1779 return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_dedicatedDictSearch);
1780 }
1781
ZSTD_compressBlock_lazy_dedicatedDictSearch(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)1782 size_t ZSTD_compressBlock_lazy_dedicatedDictSearch(
1783 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1784 void const* src, size_t srcSize)
1785 {
1786 return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_dedicatedDictSearch);
1787 }
1788
ZSTD_compressBlock_greedy_dedicatedDictSearch(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)1789 size_t ZSTD_compressBlock_greedy_dedicatedDictSearch(
1790 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1791 void const* src, size_t srcSize)
1792 {
1793 return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_dedicatedDictSearch);
1794 }
1795
1796 /* Row-based matchfinder */
ZSTD_compressBlock_lazy2_row(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)1797 size_t ZSTD_compressBlock_lazy2_row(
1798 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1799 void const* src, size_t srcSize)
1800 {
1801 return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_noDict);
1802 }
1803
ZSTD_compressBlock_lazy_row(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)1804 size_t ZSTD_compressBlock_lazy_row(
1805 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1806 void const* src, size_t srcSize)
1807 {
1808 return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_noDict);
1809 }
1810
ZSTD_compressBlock_greedy_row(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)1811 size_t ZSTD_compressBlock_greedy_row(
1812 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1813 void const* src, size_t srcSize)
1814 {
1815 return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_noDict);
1816 }
1817
ZSTD_compressBlock_lazy2_dictMatchState_row(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)1818 size_t ZSTD_compressBlock_lazy2_dictMatchState_row(
1819 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1820 void const* src, size_t srcSize)
1821 {
1822 return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_dictMatchState);
1823 }
1824
ZSTD_compressBlock_lazy_dictMatchState_row(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)1825 size_t ZSTD_compressBlock_lazy_dictMatchState_row(
1826 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1827 void const* src, size_t srcSize)
1828 {
1829 return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_dictMatchState);
1830 }
1831
ZSTD_compressBlock_greedy_dictMatchState_row(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)1832 size_t ZSTD_compressBlock_greedy_dictMatchState_row(
1833 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1834 void const* src, size_t srcSize)
1835 {
1836 return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_dictMatchState);
1837 }
1838
1839
ZSTD_compressBlock_lazy2_dedicatedDictSearch_row(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)1840 size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch_row(
1841 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1842 void const* src, size_t srcSize)
1843 {
1844 return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_dedicatedDictSearch);
1845 }
1846
ZSTD_compressBlock_lazy_dedicatedDictSearch_row(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)1847 size_t ZSTD_compressBlock_lazy_dedicatedDictSearch_row(
1848 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1849 void const* src, size_t srcSize)
1850 {
1851 return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_dedicatedDictSearch);
1852 }
1853
ZSTD_compressBlock_greedy_dedicatedDictSearch_row(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)1854 size_t ZSTD_compressBlock_greedy_dedicatedDictSearch_row(
1855 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
1856 void const* src, size_t srcSize)
1857 {
1858 return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_dedicatedDictSearch);
1859 }
1860
1861 FORCE_INLINE_TEMPLATE
ZSTD_compressBlock_lazy_extDict_generic(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],const void * src,size_t srcSize,const searchMethod_e searchMethod,const U32 depth)1862 size_t ZSTD_compressBlock_lazy_extDict_generic(
1863 ZSTD_matchState_t* ms, seqStore_t* seqStore,
1864 U32 rep[ZSTD_REP_NUM],
1865 const void* src, size_t srcSize,
1866 const searchMethod_e searchMethod, const U32 depth)
1867 {
1868 const BYTE* const istart = (const BYTE*)src;
1869 const BYTE* ip = istart;
1870 const BYTE* anchor = istart;
1871 const BYTE* const iend = istart + srcSize;
1872 const BYTE* const ilimit = searchMethod == search_rowHash ? iend - 8 - ZSTD_ROW_HASH_CACHE_SIZE : iend - 8;
1873 const BYTE* const base = ms->window.base;
1874 const U32 dictLimit = ms->window.dictLimit;
1875 const BYTE* const prefixStart = base + dictLimit;
1876 const BYTE* const dictBase = ms->window.dictBase;
1877 const BYTE* const dictEnd = dictBase + dictLimit;
1878 const BYTE* const dictStart = dictBase + ms->window.lowLimit;
1879 const U32 windowLog = ms->cParams.windowLog;
1880 const U32 rowLog = ms->cParams.searchLog < 5 ? 4 : 5;
1881
1882 searchMax_f const searchMax = ZSTD_selectLazyVTable(ms, searchMethod, ZSTD_extDict)->searchMax;
1883 U32 offset_1 = rep[0], offset_2 = rep[1];
1884
1885 DEBUGLOG(5, "ZSTD_compressBlock_lazy_extDict_generic (searchFunc=%u)", (U32)searchMethod);
1886
1887 /* init */
1888 ip += (ip == prefixStart);
1889 if (searchMethod == search_rowHash) {
1890 ZSTD_row_fillHashCache(ms, base, rowLog,
1891 MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */),
1892 ms->nextToUpdate, ilimit);
1893 }
1894
1895 /* Match Loop */
1896 #if defined(__GNUC__) && defined(__x86_64__)
1897 /* I've measured random a 5% speed loss on levels 5 & 6 (greedy) when the
1898 * code alignment is perturbed. To fix the instability align the loop on 32-bytes.
1899 */
1900 __asm__(".p2align 5");
1901 #endif
1902 while (ip < ilimit) {
1903 size_t matchLength=0;
1904 size_t offset=0;
1905 const BYTE* start=ip+1;
1906 U32 curr = (U32)(ip-base);
1907
1908 /* check repCode */
1909 { const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr+1, windowLog);
1910 const U32 repIndex = (U32)(curr+1 - offset_1);
1911 const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
1912 const BYTE* const repMatch = repBase + repIndex;
1913 if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow */
1914 & (offset_1 <= curr+1 - windowLow) ) /* note: we are searching at curr+1 */
1915 if (MEM_read32(ip+1) == MEM_read32(repMatch)) {
1916 /* repcode detected we should take it */
1917 const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
1918 matchLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repEnd, prefixStart) + 4;
1919 if (depth==0) goto _storeSequence;
1920 } }
1921
1922 /* first search (depth 0) */
1923 { size_t offsetFound = 999999999;
1924 size_t const ml2 = searchMax(ms, ip, iend, &offsetFound);
1925 if (ml2 > matchLength)
1926 matchLength = ml2, start = ip, offset=offsetFound;
1927 }
1928
1929 if (matchLength < 4) {
1930 ip += ((ip-anchor) >> kSearchStrength) + 1; /* jump faster over incompressible sections */
1931 continue;
1932 }
1933
1934 /* let's try to find a better solution */
1935 if (depth>=1)
1936 while (ip<ilimit) {
1937 ip ++;
1938 curr++;
1939 /* check repCode */
1940 if (offset) {
1941 const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr, windowLog);
1942 const U32 repIndex = (U32)(curr - offset_1);
1943 const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
1944 const BYTE* const repMatch = repBase + repIndex;
1945 if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow : do not test positions overlapping 2 memory segments */
1946 & (offset_1 <= curr - windowLow) ) /* equivalent to `curr > repIndex >= windowLow` */
1947 if (MEM_read32(ip) == MEM_read32(repMatch)) {
1948 /* repcode detected */
1949 const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
1950 size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
1951 int const gain2 = (int)(repLength * 3);
1952 int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offset+1) + 1);
1953 if ((repLength >= 4) && (gain2 > gain1))
1954 matchLength = repLength, offset = 0, start = ip;
1955 } }
1956
1957 /* search match, depth 1 */
1958 { size_t offset2=999999999;
1959 size_t const ml2 = searchMax(ms, ip, iend, &offset2);
1960 int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1)); /* raw approx */
1961 int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 4);
1962 if ((ml2 >= 4) && (gain2 > gain1)) {
1963 matchLength = ml2, offset = offset2, start = ip;
1964 continue; /* search a better one */
1965 } }
1966
1967 /* let's find an even better one */
1968 if ((depth==2) && (ip<ilimit)) {
1969 ip ++;
1970 curr++;
1971 /* check repCode */
1972 if (offset) {
1973 const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr, windowLog);
1974 const U32 repIndex = (U32)(curr - offset_1);
1975 const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
1976 const BYTE* const repMatch = repBase + repIndex;
1977 if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow : do not test positions overlapping 2 memory segments */
1978 & (offset_1 <= curr - windowLow) ) /* equivalent to `curr > repIndex >= windowLow` */
1979 if (MEM_read32(ip) == MEM_read32(repMatch)) {
1980 /* repcode detected */
1981 const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
1982 size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
1983 int const gain2 = (int)(repLength * 4);
1984 int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 1);
1985 if ((repLength >= 4) && (gain2 > gain1))
1986 matchLength = repLength, offset = 0, start = ip;
1987 } }
1988
1989 /* search match, depth 2 */
1990 { size_t offset2=999999999;
1991 size_t const ml2 = searchMax(ms, ip, iend, &offset2);
1992 int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)offset2+1)); /* raw approx */
1993 int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offset+1) + 7);
1994 if ((ml2 >= 4) && (gain2 > gain1)) {
1995 matchLength = ml2, offset = offset2, start = ip;
1996 continue;
1997 } } }
1998 break; /* nothing found : store previous solution */
1999 }
2000
2001 /* catch up */
2002 if (offset) {
2003 U32 const matchIndex = (U32)((size_t)(start-base) - (offset - ZSTD_REP_MOVE));
2004 const BYTE* match = (matchIndex < dictLimit) ? dictBase + matchIndex : base + matchIndex;
2005 const BYTE* const mStart = (matchIndex < dictLimit) ? dictStart : prefixStart;
2006 while ((start>anchor) && (match>mStart) && (start[-1] == match[-1])) { start--; match--; matchLength++; } /* catch up */
2007 offset_2 = offset_1; offset_1 = (U32)(offset - ZSTD_REP_MOVE);
2008 }
2009
2010 /* store sequence */
2011 _storeSequence:
2012 { size_t const litLength = (size_t)(start - anchor);
2013 ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offset, matchLength-MINMATCH);
2014 anchor = ip = start + matchLength;
2015 }
2016
2017 /* check immediate repcode */
2018 while (ip <= ilimit) {
2019 const U32 repCurrent = (U32)(ip-base);
2020 const U32 windowLow = ZSTD_getLowestMatchIndex(ms, repCurrent, windowLog);
2021 const U32 repIndex = repCurrent - offset_2;
2022 const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
2023 const BYTE* const repMatch = repBase + repIndex;
2024 if ( ((U32)((dictLimit-1) - repIndex) >= 3) /* intentional overflow : do not test positions overlapping 2 memory segments */
2025 & (offset_2 <= repCurrent - windowLow) ) /* equivalent to `curr > repIndex >= windowLow` */
2026 if (MEM_read32(ip) == MEM_read32(repMatch)) {
2027 /* repcode detected we should take it */
2028 const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
2029 matchLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
2030 offset = offset_2; offset_2 = offset_1; offset_1 = (U32)offset; /* swap offset history */
2031 ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, matchLength-MINMATCH);
2032 ip += matchLength;
2033 anchor = ip;
2034 continue; /* faster when present ... (?) */
2035 }
2036 break;
2037 } }
2038
2039 /* Save reps for next block */
2040 rep[0] = offset_1;
2041 rep[1] = offset_2;
2042
2043 /* Return the last literals size */
2044 return (size_t)(iend - anchor);
2045 }
2046
2047
ZSTD_compressBlock_greedy_extDict(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)2048 size_t ZSTD_compressBlock_greedy_extDict(
2049 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
2050 void const* src, size_t srcSize)
2051 {
2052 return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0);
2053 }
2054
ZSTD_compressBlock_lazy_extDict(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)2055 size_t ZSTD_compressBlock_lazy_extDict(
2056 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
2057 void const* src, size_t srcSize)
2058
2059 {
2060 return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1);
2061 }
2062
ZSTD_compressBlock_lazy2_extDict(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)2063 size_t ZSTD_compressBlock_lazy2_extDict(
2064 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
2065 void const* src, size_t srcSize)
2066
2067 {
2068 return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2);
2069 }
2070
ZSTD_compressBlock_btlazy2_extDict(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)2071 size_t ZSTD_compressBlock_btlazy2_extDict(
2072 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
2073 void const* src, size_t srcSize)
2074
2075 {
2076 return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2);
2077 }
2078
ZSTD_compressBlock_greedy_extDict_row(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)2079 size_t ZSTD_compressBlock_greedy_extDict_row(
2080 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
2081 void const* src, size_t srcSize)
2082 {
2083 return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0);
2084 }
2085
ZSTD_compressBlock_lazy_extDict_row(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)2086 size_t ZSTD_compressBlock_lazy_extDict_row(
2087 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
2088 void const* src, size_t srcSize)
2089
2090 {
2091 return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1);
2092 }
2093
ZSTD_compressBlock_lazy2_extDict_row(ZSTD_matchState_t * ms,seqStore_t * seqStore,U32 rep[ZSTD_REP_NUM],void const * src,size_t srcSize)2094 size_t ZSTD_compressBlock_lazy2_extDict_row(
2095 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
2096 void const* src, size_t srcSize)
2097
2098 {
2099 return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2);
2100 }
2101