2 * Copyright (c) Meta Platforms, Inc. and affiliates.
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.
11 #include "zstd_compress_internal.h" /* ZSTD_hashPtr, ZSTD_count, ZSTD_storeSeq */
12 #include "zstd_fast.h"
14 static void ZSTD_fillHashTableForCDict(ZSTD_matchState_t* ms,
15 const void* const end,
16 ZSTD_dictTableLoadMethod_e dtlm)
18 const ZSTD_compressionParameters* const cParams = &ms->cParams;
19 U32* const hashTable = ms->hashTable;
20 U32 const hBits = cParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS;
21 U32 const mls = cParams->minMatch;
22 const BYTE* const base = ms->window.base;
23 const BYTE* ip = base + ms->nextToUpdate;
24 const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE;
25 const U32 fastHashFillStep = 3;
27 /* Currently, we always use ZSTD_dtlm_full for filling CDict tables.
28 * Feel free to remove this assert if there's a good reason! */
29 assert(dtlm == ZSTD_dtlm_full);
31 /* Always insert every fastHashFillStep position into the hash table.
32 * Insert the other positions if their hash entry is empty.
34 for ( ; ip + fastHashFillStep < iend + 2; ip += fastHashFillStep) {
35 U32 const curr = (U32)(ip - base);
36 { size_t const hashAndTag = ZSTD_hashPtr(ip, hBits, mls);
37 ZSTD_writeTaggedIndex(hashTable, hashAndTag, curr); }
39 if (dtlm == ZSTD_dtlm_fast) continue;
40 /* Only load extra positions for ZSTD_dtlm_full */
42 for (p = 1; p < fastHashFillStep; ++p) {
43 size_t const hashAndTag = ZSTD_hashPtr(ip + p, hBits, mls);
44 if (hashTable[hashAndTag >> ZSTD_SHORT_CACHE_TAG_BITS] == 0) { /* not yet filled */
45 ZSTD_writeTaggedIndex(hashTable, hashAndTag, curr + p);
49 static void ZSTD_fillHashTableForCCtx(ZSTD_matchState_t* ms,
50 const void* const end,
51 ZSTD_dictTableLoadMethod_e dtlm)
53 const ZSTD_compressionParameters* const cParams = &ms->cParams;
54 U32* const hashTable = ms->hashTable;
55 U32 const hBits = cParams->hashLog;
56 U32 const mls = cParams->minMatch;
57 const BYTE* const base = ms->window.base;
58 const BYTE* ip = base + ms->nextToUpdate;
59 const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE;
60 const U32 fastHashFillStep = 3;
62 /* Currently, we always use ZSTD_dtlm_fast for filling CCtx tables.
63 * Feel free to remove this assert if there's a good reason! */
64 assert(dtlm == ZSTD_dtlm_fast);
66 /* Always insert every fastHashFillStep position into the hash table.
67 * Insert the other positions if their hash entry is empty.
69 for ( ; ip + fastHashFillStep < iend + 2; ip += fastHashFillStep) {
70 U32 const curr = (U32)(ip - base);
71 size_t const hash0 = ZSTD_hashPtr(ip, hBits, mls);
72 hashTable[hash0] = curr;
73 if (dtlm == ZSTD_dtlm_fast) continue;
74 /* Only load extra positions for ZSTD_dtlm_full */
76 for (p = 1; p < fastHashFillStep; ++p) {
77 size_t const hash = ZSTD_hashPtr(ip + p, hBits, mls);
78 if (hashTable[hash] == 0) { /* not yet filled */
79 hashTable[hash] = curr + p;
83 void ZSTD_fillHashTable(ZSTD_matchState_t* ms,
84 const void* const end,
85 ZSTD_dictTableLoadMethod_e dtlm,
86 ZSTD_tableFillPurpose_e tfp)
88 if (tfp == ZSTD_tfp_forCDict) {
89 ZSTD_fillHashTableForCDict(ms, end, dtlm);
91 ZSTD_fillHashTableForCCtx(ms, end, dtlm);
97 * If you squint hard enough (and ignore repcodes), the search operation at any
98 * given position is broken into 4 stages:
100 * 1. Hash (map position to hash value via input read)
101 * 2. Lookup (map hash val to index via hashtable read)
102 * 3. Load (map index to value at that position via input read)
105 * Each of these steps involves a memory read at an address which is computed
106 * from the previous step. This means these steps must be sequenced and their
107 * latencies are cumulative.
109 * Rather than do 1->2->3->4 sequentially for a single position before moving
110 * onto the next, this implementation interleaves these operations across the
111 * next few positions:
113 * R = Repcode Read & Compare
116 * M = Match Read & Compare
119 * ----+-------------------
128 * This is very much analogous to the pipelining of execution in a CPU. And just
129 * like a CPU, we have to dump the pipeline when we find a match (i.e., take a
132 * When this happens, we throw away our current state, and do the following prep
133 * to re-enter the loop:
136 * ----+-------------------
140 * This is also the work we do at the beginning to enter the loop initially.
142 FORCE_INLINE_TEMPLATE size_t
143 ZSTD_compressBlock_fast_noDict_generic(
144 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
145 void const* src, size_t srcSize,
146 U32 const mls, U32 const hasStep)
148 const ZSTD_compressionParameters* const cParams = &ms->cParams;
149 U32* const hashTable = ms->hashTable;
150 U32 const hlog = cParams->hashLog;
151 /* support stepSize of 0 */
152 size_t const stepSize = hasStep ? (cParams->targetLength + !(cParams->targetLength) + 1) : 2;
153 const BYTE* const base = ms->window.base;
154 const BYTE* const istart = (const BYTE*)src;
155 const U32 endIndex = (U32)((size_t)(istart - base) + srcSize);
156 const U32 prefixStartIndex = ZSTD_getLowestPrefixIndex(ms, endIndex, cParams->windowLog);
157 const BYTE* const prefixStart = base + prefixStartIndex;
158 const BYTE* const iend = istart + srcSize;
159 const BYTE* const ilimit = iend - HASH_READ_SIZE;
161 const BYTE* anchor = istart;
162 const BYTE* ip0 = istart;
168 U32 rep_offset1 = rep[0];
169 U32 rep_offset2 = rep[1];
170 U32 offsetSaved1 = 0, offsetSaved2 = 0;
172 size_t hash0; /* hash for ip0 */
173 size_t hash1; /* hash for ip1 */
174 U32 idx; /* match idx for ip0 */
175 U32 mval; /* src value at match idx */
181 /* ip0 and ip1 are always adjacent. The targetLength skipping and
182 * uncompressibility acceleration is applied to every other position,
183 * matching the behavior of #1562. step therefore represents the gap
184 * between pairs of positions, from ip0 to ip2 or ip1 to ip3. */
186 const BYTE* nextStep;
187 const size_t kStepIncr = (1 << (kSearchStrength - 1));
189 DEBUGLOG(5, "ZSTD_compressBlock_fast_generic");
190 ip0 += (ip0 == prefixStart);
191 { U32 const curr = (U32)(ip0 - base);
192 U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, curr, cParams->windowLog);
193 U32 const maxRep = curr - windowLow;
194 if (rep_offset2 > maxRep) offsetSaved2 = rep_offset2, rep_offset2 = 0;
195 if (rep_offset1 > maxRep) offsetSaved1 = rep_offset1, rep_offset1 = 0;
199 _start: /* Requires: ip0 */
202 nextStep = ip0 + kStepIncr;
204 /* calculate positions, ip0 - anchor == 0, so we skip step calc */
213 hash0 = ZSTD_hashPtr(ip0, hlog, mls);
214 hash1 = ZSTD_hashPtr(ip1, hlog, mls);
216 idx = hashTable[hash0];
219 /* load repcode match for ip[2]*/
220 const U32 rval = MEM_read32(ip2 - rep_offset1);
222 /* write back hash table entry */
223 current0 = (U32)(ip0 - base);
224 hashTable[hash0] = current0;
226 /* check repcode at ip[2] */
227 if ((MEM_read32(ip2) == rval) & (rep_offset1 > 0)) {
229 match0 = ip0 - rep_offset1;
230 mLength = ip0[-1] == match0[-1];
233 offcode = REPCODE1_TO_OFFBASE;
236 /* First write next hash table entry; we've already calculated it.
237 * This write is known to be safe because the ip1 is before the
239 hashTable[hash1] = (U32)(ip1 - base);
244 /* load match for ip[0] */
245 if (idx >= prefixStartIndex) {
246 mval = MEM_read32(base + idx);
248 mval = MEM_read32(ip0) ^ 1; /* guaranteed to not match. */
251 /* check match at ip[0] */
252 if (MEM_read32(ip0) == mval) {
255 /* First write next hash table entry; we've already calculated it.
256 * This write is known to be safe because the ip1 == ip0 + 1, so
257 * we know we will resume searching after ip1 */
258 hashTable[hash1] = (U32)(ip1 - base);
264 idx = hashTable[hash1];
268 hash1 = ZSTD_hashPtr(ip2, hlog, mls);
270 /* advance to next positions */
275 /* write back hash table entry */
276 current0 = (U32)(ip0 - base);
277 hashTable[hash0] = current0;
279 /* load match for ip[0] */
280 if (idx >= prefixStartIndex) {
281 mval = MEM_read32(base + idx);
283 mval = MEM_read32(ip0) ^ 1; /* guaranteed to not match. */
286 /* check match at ip[0] */
287 if (MEM_read32(ip0) == mval) {
290 /* first write next hash table entry; we've already calculated it */
292 /* We need to avoid writing an index into the hash table >= the
293 * position at which we will pick up our searching after we've
296 * The minimum possible match has length 4, so the earliest ip0
297 * can be after we take this match will be the current ip0 + 4.
298 * ip1 is ip0 + step - 1. If ip1 is >= ip0 + 4, we can't safely
299 * write this position.
301 hashTable[hash1] = (U32)(ip1 - base);
308 idx = hashTable[hash1];
312 hash1 = ZSTD_hashPtr(ip2, hlog, mls);
314 /* advance to next positions */
321 if (ip2 >= nextStep) {
323 PREFETCH_L1(ip1 + 64);
324 PREFETCH_L1(ip1 + 128);
325 nextStep += kStepIncr;
327 } while (ip3 < ilimit);
330 /* Note that there are probably still a couple positions we could search.
331 * However, it seems to be a meaningful performance hit to try to search
332 * them. So let's not. */
334 /* When the repcodes are outside of the prefix, we set them to zero before the loop.
335 * When the offsets are still zero, we need to restore them after the block to have a correct
336 * repcode history. If only one offset was invalid, it is easy. The tricky case is when both
337 * offsets were invalid. We need to figure out which offset to refill with.
338 * - If both offsets are zero they are in the same order.
339 * - If both offsets are non-zero, we won't restore the offsets from `offsetSaved[12]`.
340 * - If only one is zero, we need to decide which offset to restore.
341 * - If rep_offset1 is non-zero, then rep_offset2 must be offsetSaved1.
342 * - It is impossible for rep_offset2 to be non-zero.
344 * So if rep_offset1 started invalid (offsetSaved1 != 0) and became valid (rep_offset1 != 0), then
345 * set rep[0] = rep_offset1 and rep[1] = offsetSaved1.
347 offsetSaved2 = ((offsetSaved1 != 0) && (rep_offset1 != 0)) ? offsetSaved1 : offsetSaved2;
349 /* save reps for next block */
350 rep[0] = rep_offset1 ? rep_offset1 : offsetSaved1;
351 rep[1] = rep_offset2 ? rep_offset2 : offsetSaved2;
353 /* Return the last literals size */
354 return (size_t)(iend - anchor);
356 _offset: /* Requires: ip0, idx */
358 /* Compute the offset code. */
360 rep_offset2 = rep_offset1;
361 rep_offset1 = (U32)(ip0-match0);
362 offcode = OFFSET_TO_OFFBASE(rep_offset1);
365 /* Count the backwards match length. */
366 while (((ip0>anchor) & (match0>prefixStart)) && (ip0[-1] == match0[-1])) {
372 _match: /* Requires: ip0, match0, offcode */
374 /* Count the forward length. */
375 mLength += ZSTD_count(ip0 + mLength, match0 + mLength, iend);
377 ZSTD_storeSeq(seqStore, (size_t)(ip0 - anchor), anchor, iend, offcode, mLength);
382 /* Fill table and check for immediate repcode. */
385 assert(base+current0+2 > istart); /* check base overflow */
386 hashTable[ZSTD_hashPtr(base+current0+2, hlog, mls)] = current0+2; /* here because current+2 could be > iend-8 */
387 hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base);
389 if (rep_offset2 > 0) { /* rep_offset2==0 means rep_offset2 is invalidated */
390 while ( (ip0 <= ilimit) && (MEM_read32(ip0) == MEM_read32(ip0 - rep_offset2)) ) {
392 size_t const rLength = ZSTD_count(ip0+4, ip0+4-rep_offset2, iend) + 4;
393 { U32 const tmpOff = rep_offset2; rep_offset2 = rep_offset1; rep_offset1 = tmpOff; } /* swap rep_offset2 <=> rep_offset1 */
394 hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = (U32)(ip0-base);
396 ZSTD_storeSeq(seqStore, 0 /*litLen*/, anchor, iend, REPCODE1_TO_OFFBASE, rLength);
398 continue; /* faster when present (confirmed on gcc-8) ... (?) */
404 #define ZSTD_GEN_FAST_FN(dictMode, mls, step) \
405 static size_t ZSTD_compressBlock_fast_##dictMode##_##mls##_##step( \
406 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], \
407 void const* src, size_t srcSize) \
409 return ZSTD_compressBlock_fast_##dictMode##_generic(ms, seqStore, rep, src, srcSize, mls, step); \
412 ZSTD_GEN_FAST_FN(noDict, 4, 1)
413 ZSTD_GEN_FAST_FN(noDict, 5, 1)
414 ZSTD_GEN_FAST_FN(noDict, 6, 1)
415 ZSTD_GEN_FAST_FN(noDict, 7, 1)
417 ZSTD_GEN_FAST_FN(noDict, 4, 0)
418 ZSTD_GEN_FAST_FN(noDict, 5, 0)
419 ZSTD_GEN_FAST_FN(noDict, 6, 0)
420 ZSTD_GEN_FAST_FN(noDict, 7, 0)
422 size_t ZSTD_compressBlock_fast(
423 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
424 void const* src, size_t srcSize)
426 U32 const mls = ms->cParams.minMatch;
427 assert(ms->dictMatchState == NULL);
428 if (ms->cParams.targetLength > 1) {
431 default: /* includes case 3 */
433 return ZSTD_compressBlock_fast_noDict_4_1(ms, seqStore, rep, src, srcSize);
435 return ZSTD_compressBlock_fast_noDict_5_1(ms, seqStore, rep, src, srcSize);
437 return ZSTD_compressBlock_fast_noDict_6_1(ms, seqStore, rep, src, srcSize);
439 return ZSTD_compressBlock_fast_noDict_7_1(ms, seqStore, rep, src, srcSize);
444 default: /* includes case 3 */
446 return ZSTD_compressBlock_fast_noDict_4_0(ms, seqStore, rep, src, srcSize);
448 return ZSTD_compressBlock_fast_noDict_5_0(ms, seqStore, rep, src, srcSize);
450 return ZSTD_compressBlock_fast_noDict_6_0(ms, seqStore, rep, src, srcSize);
452 return ZSTD_compressBlock_fast_noDict_7_0(ms, seqStore, rep, src, srcSize);
458 FORCE_INLINE_TEMPLATE
459 size_t ZSTD_compressBlock_fast_dictMatchState_generic(
460 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
461 void const* src, size_t srcSize, U32 const mls, U32 const hasStep)
463 const ZSTD_compressionParameters* const cParams = &ms->cParams;
464 U32* const hashTable = ms->hashTable;
465 U32 const hlog = cParams->hashLog;
466 /* support stepSize of 0 */
467 U32 const stepSize = cParams->targetLength + !(cParams->targetLength);
468 const BYTE* const base = ms->window.base;
469 const BYTE* const istart = (const BYTE*)src;
470 const BYTE* ip0 = istart;
471 const BYTE* ip1 = ip0 + stepSize; /* we assert below that stepSize >= 1 */
472 const BYTE* anchor = istart;
473 const U32 prefixStartIndex = ms->window.dictLimit;
474 const BYTE* const prefixStart = base + prefixStartIndex;
475 const BYTE* const iend = istart + srcSize;
476 const BYTE* const ilimit = iend - HASH_READ_SIZE;
477 U32 offset_1=rep[0], offset_2=rep[1];
479 const ZSTD_matchState_t* const dms = ms->dictMatchState;
480 const ZSTD_compressionParameters* const dictCParams = &dms->cParams ;
481 const U32* const dictHashTable = dms->hashTable;
482 const U32 dictStartIndex = dms->window.dictLimit;
483 const BYTE* const dictBase = dms->window.base;
484 const BYTE* const dictStart = dictBase + dictStartIndex;
485 const BYTE* const dictEnd = dms->window.nextSrc;
486 const U32 dictIndexDelta = prefixStartIndex - (U32)(dictEnd - dictBase);
487 const U32 dictAndPrefixLength = (U32)(istart - prefixStart + dictEnd - dictStart);
488 const U32 dictHBits = dictCParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS;
490 /* if a dictionary is still attached, it necessarily means that
491 * it is within window size. So we just check it. */
492 const U32 maxDistance = 1U << cParams->windowLog;
493 const U32 endIndex = (U32)((size_t)(istart - base) + srcSize);
494 assert(endIndex - prefixStartIndex <= maxDistance);
495 (void)maxDistance; (void)endIndex; /* these variables are not used when assert() is disabled */
497 (void)hasStep; /* not currently specialized on whether it's accelerated */
499 /* ensure there will be no underflow
500 * when translating a dict index into a local index */
501 assert(prefixStartIndex >= (U32)(dictEnd - dictBase));
503 if (ms->prefetchCDictTables) {
504 size_t const hashTableBytes = (((size_t)1) << dictCParams->hashLog) * sizeof(U32);
505 PREFETCH_AREA(dictHashTable, hashTableBytes)
509 DEBUGLOG(5, "ZSTD_compressBlock_fast_dictMatchState_generic");
510 ip0 += (dictAndPrefixLength == 0);
511 /* dictMatchState repCode checks don't currently handle repCode == 0
513 assert(offset_1 <= dictAndPrefixLength);
514 assert(offset_2 <= dictAndPrefixLength);
516 /* Outer search loop */
517 assert(stepSize >= 1);
518 while (ip1 <= ilimit) { /* repcode check at (ip0 + 1) is safe because ip0 < ip1 */
520 size_t hash0 = ZSTD_hashPtr(ip0, hlog, mls);
522 size_t const dictHashAndTag0 = ZSTD_hashPtr(ip0, dictHBits, mls);
523 U32 dictMatchIndexAndTag = dictHashTable[dictHashAndTag0 >> ZSTD_SHORT_CACHE_TAG_BITS];
524 int dictTagsMatch = ZSTD_comparePackedTags(dictMatchIndexAndTag, dictHashAndTag0);
526 U32 matchIndex = hashTable[hash0];
527 U32 curr = (U32)(ip0 - base);
528 size_t step = stepSize;
529 const size_t kStepIncr = 1 << kSearchStrength;
530 const BYTE* nextStep = ip0 + kStepIncr;
532 /* Inner search loop */
534 const BYTE* match = base + matchIndex;
535 const U32 repIndex = curr + 1 - offset_1;
536 const BYTE* repMatch = (repIndex < prefixStartIndex) ?
537 dictBase + (repIndex - dictIndexDelta) :
539 const size_t hash1 = ZSTD_hashPtr(ip1, hlog, mls);
540 size_t const dictHashAndTag1 = ZSTD_hashPtr(ip1, dictHBits, mls);
541 hashTable[hash0] = curr; /* update hash table */
543 if (((U32) ((prefixStartIndex - 1) - repIndex) >=
544 3) /* intentional underflow : ensure repIndex isn't overlapping dict + prefix */
545 && (MEM_read32(repMatch) == MEM_read32(ip0 + 1))) {
546 const BYTE* const repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
547 mLength = ZSTD_count_2segments(ip0 + 1 + 4, repMatch + 4, iend, repMatchEnd, prefixStart) + 4;
549 ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength);
554 /* Found a possible dict match */
555 const U32 dictMatchIndex = dictMatchIndexAndTag >> ZSTD_SHORT_CACHE_TAG_BITS;
556 const BYTE* dictMatch = dictBase + dictMatchIndex;
557 if (dictMatchIndex > dictStartIndex &&
558 MEM_read32(dictMatch) == MEM_read32(ip0)) {
559 /* To replicate extDict parse behavior, we only use dict matches when the normal matchIndex is invalid */
560 if (matchIndex <= prefixStartIndex) {
561 U32 const offset = (U32) (curr - dictMatchIndex - dictIndexDelta);
562 mLength = ZSTD_count_2segments(ip0 + 4, dictMatch + 4, iend, dictEnd, prefixStart) + 4;
563 while (((ip0 > anchor) & (dictMatch > dictStart))
564 && (ip0[-1] == dictMatch[-1])) {
571 ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
577 if (matchIndex > prefixStartIndex && MEM_read32(match) == MEM_read32(ip0)) {
578 /* found a regular match */
579 U32 const offset = (U32) (ip0 - match);
580 mLength = ZSTD_count(ip0 + 4, match + 4, iend) + 4;
581 while (((ip0 > anchor) & (match > prefixStart))
582 && (ip0[-1] == match[-1])) {
589 ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
593 /* Prepare for next iteration */
594 dictMatchIndexAndTag = dictHashTable[dictHashAndTag1 >> ZSTD_SHORT_CACHE_TAG_BITS];
595 dictTagsMatch = ZSTD_comparePackedTags(dictMatchIndexAndTag, dictHashAndTag1);
596 matchIndex = hashTable[hash1];
598 if (ip1 >= nextStep) {
600 nextStep += kStepIncr;
604 if (ip1 > ilimit) goto _cleanup;
606 curr = (U32)(ip0 - base);
608 } /* end inner search loop */
617 assert(base+curr+2 > istart); /* check base overflow */
618 hashTable[ZSTD_hashPtr(base+curr+2, hlog, mls)] = curr+2; /* here because curr+2 could be > iend-8 */
619 hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base);
621 /* check immediate repcode */
622 while (ip0 <= ilimit) {
623 U32 const current2 = (U32)(ip0-base);
624 U32 const repIndex2 = current2 - offset_2;
625 const BYTE* repMatch2 = repIndex2 < prefixStartIndex ?
626 dictBase - dictIndexDelta + repIndex2 :
628 if ( ((U32)((prefixStartIndex-1) - (U32)repIndex2) >= 3 /* intentional overflow */)
629 && (MEM_read32(repMatch2) == MEM_read32(ip0))) {
630 const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
631 size_t const repLength2 = ZSTD_count_2segments(ip0+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
632 U32 tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; /* swap offset_2 <=> offset_1 */
633 ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
634 hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = current2;
643 /* Prepare for next iteration */
644 assert(ip0 == anchor);
645 ip1 = ip0 + stepSize;
649 /* save reps for next block */
653 /* Return the last literals size */
654 return (size_t)(iend - anchor);
658 ZSTD_GEN_FAST_FN(dictMatchState, 4, 0)
659 ZSTD_GEN_FAST_FN(dictMatchState, 5, 0)
660 ZSTD_GEN_FAST_FN(dictMatchState, 6, 0)
661 ZSTD_GEN_FAST_FN(dictMatchState, 7, 0)
663 size_t ZSTD_compressBlock_fast_dictMatchState(
664 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
665 void const* src, size_t srcSize)
667 U32 const mls = ms->cParams.minMatch;
668 assert(ms->dictMatchState != NULL);
671 default: /* includes case 3 */
673 return ZSTD_compressBlock_fast_dictMatchState_4_0(ms, seqStore, rep, src, srcSize);
675 return ZSTD_compressBlock_fast_dictMatchState_5_0(ms, seqStore, rep, src, srcSize);
677 return ZSTD_compressBlock_fast_dictMatchState_6_0(ms, seqStore, rep, src, srcSize);
679 return ZSTD_compressBlock_fast_dictMatchState_7_0(ms, seqStore, rep, src, srcSize);
684 static size_t ZSTD_compressBlock_fast_extDict_generic(
685 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
686 void const* src, size_t srcSize, U32 const mls, U32 const hasStep)
688 const ZSTD_compressionParameters* const cParams = &ms->cParams;
689 U32* const hashTable = ms->hashTable;
690 U32 const hlog = cParams->hashLog;
691 /* support stepSize of 0 */
692 size_t const stepSize = cParams->targetLength + !(cParams->targetLength) + 1;
693 const BYTE* const base = ms->window.base;
694 const BYTE* const dictBase = ms->window.dictBase;
695 const BYTE* const istart = (const BYTE*)src;
696 const BYTE* anchor = istart;
697 const U32 endIndex = (U32)((size_t)(istart - base) + srcSize);
698 const U32 lowLimit = ZSTD_getLowestMatchIndex(ms, endIndex, cParams->windowLog);
699 const U32 dictStartIndex = lowLimit;
700 const BYTE* const dictStart = dictBase + dictStartIndex;
701 const U32 dictLimit = ms->window.dictLimit;
702 const U32 prefixStartIndex = dictLimit < lowLimit ? lowLimit : dictLimit;
703 const BYTE* const prefixStart = base + prefixStartIndex;
704 const BYTE* const dictEnd = dictBase + prefixStartIndex;
705 const BYTE* const iend = istart + srcSize;
706 const BYTE* const ilimit = iend - 8;
707 U32 offset_1=rep[0], offset_2=rep[1];
708 U32 offsetSaved1 = 0, offsetSaved2 = 0;
710 const BYTE* ip0 = istart;
717 size_t hash0; /* hash for ip0 */
718 size_t hash1; /* hash for ip1 */
719 U32 idx; /* match idx for ip0 */
720 const BYTE* idxBase; /* base pointer for idx */
725 const BYTE* matchEnd = 0; /* initialize to avoid warning, assert != 0 later */
728 const BYTE* nextStep;
729 const size_t kStepIncr = (1 << (kSearchStrength - 1));
731 (void)hasStep; /* not currently specialized on whether it's accelerated */
733 DEBUGLOG(5, "ZSTD_compressBlock_fast_extDict_generic (offset_1=%u)", offset_1);
735 /* switch to "regular" variant if extDict is invalidated due to maxDistance */
736 if (prefixStartIndex == dictStartIndex)
737 return ZSTD_compressBlock_fast(ms, seqStore, rep, src, srcSize);
739 { U32 const curr = (U32)(ip0 - base);
740 U32 const maxRep = curr - dictStartIndex;
741 if (offset_2 >= maxRep) offsetSaved2 = offset_2, offset_2 = 0;
742 if (offset_1 >= maxRep) offsetSaved1 = offset_1, offset_1 = 0;
746 _start: /* Requires: ip0 */
749 nextStep = ip0 + kStepIncr;
751 /* calculate positions, ip0 - anchor == 0, so we skip step calc */
760 hash0 = ZSTD_hashPtr(ip0, hlog, mls);
761 hash1 = ZSTD_hashPtr(ip1, hlog, mls);
763 idx = hashTable[hash0];
764 idxBase = idx < prefixStartIndex ? dictBase : base;
767 { /* load repcode match for ip[2] */
768 U32 const current2 = (U32)(ip2 - base);
769 U32 const repIndex = current2 - offset_1;
770 const BYTE* const repBase = repIndex < prefixStartIndex ? dictBase : base;
772 if ( ((U32)(prefixStartIndex - repIndex) >= 4) /* intentional underflow */
774 rval = MEM_read32(repBase + repIndex);
776 rval = MEM_read32(ip2) ^ 1; /* guaranteed to not match. */
779 /* write back hash table entry */
780 current0 = (U32)(ip0 - base);
781 hashTable[hash0] = current0;
783 /* check repcode at ip[2] */
784 if (MEM_read32(ip2) == rval) {
786 match0 = repBase + repIndex;
787 matchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
788 assert((match0 != prefixStart) & (match0 != dictStart));
789 mLength = ip0[-1] == match0[-1];
792 offcode = REPCODE1_TO_OFFBASE;
797 { /* load match for ip[0] */
798 U32 const mval = idx >= dictStartIndex ?
799 MEM_read32(idxBase + idx) :
800 MEM_read32(ip0) ^ 1; /* guaranteed not to match */
802 /* check match at ip[0] */
803 if (MEM_read32(ip0) == mval) {
809 idx = hashTable[hash1];
810 idxBase = idx < prefixStartIndex ? dictBase : base;
814 hash1 = ZSTD_hashPtr(ip2, hlog, mls);
816 /* advance to next positions */
821 /* write back hash table entry */
822 current0 = (U32)(ip0 - base);
823 hashTable[hash0] = current0;
825 { /* load match for ip[0] */
826 U32 const mval = idx >= dictStartIndex ?
827 MEM_read32(idxBase + idx) :
828 MEM_read32(ip0) ^ 1; /* guaranteed not to match */
830 /* check match at ip[0] */
831 if (MEM_read32(ip0) == mval) {
837 idx = hashTable[hash1];
838 idxBase = idx < prefixStartIndex ? dictBase : base;
842 hash1 = ZSTD_hashPtr(ip2, hlog, mls);
844 /* advance to next positions */
851 if (ip2 >= nextStep) {
853 PREFETCH_L1(ip1 + 64);
854 PREFETCH_L1(ip1 + 128);
855 nextStep += kStepIncr;
857 } while (ip3 < ilimit);
860 /* Note that there are probably still a couple positions we could search.
861 * However, it seems to be a meaningful performance hit to try to search
862 * them. So let's not. */
864 /* If offset_1 started invalid (offsetSaved1 != 0) and became valid (offset_1 != 0),
865 * rotate saved offsets. See comment in ZSTD_compressBlock_fast_noDict for more context. */
866 offsetSaved2 = ((offsetSaved1 != 0) && (offset_1 != 0)) ? offsetSaved1 : offsetSaved2;
868 /* save reps for next block */
869 rep[0] = offset_1 ? offset_1 : offsetSaved1;
870 rep[1] = offset_2 ? offset_2 : offsetSaved2;
872 /* Return the last literals size */
873 return (size_t)(iend - anchor);
875 _offset: /* Requires: ip0, idx, idxBase */
877 /* Compute the offset code. */
878 { U32 const offset = current0 - idx;
879 const BYTE* const lowMatchPtr = idx < prefixStartIndex ? dictStart : prefixStart;
880 matchEnd = idx < prefixStartIndex ? dictEnd : iend;
881 match0 = idxBase + idx;
884 offcode = OFFSET_TO_OFFBASE(offset);
887 /* Count the backwards match length. */
888 while (((ip0>anchor) & (match0>lowMatchPtr)) && (ip0[-1] == match0[-1])) {
894 _match: /* Requires: ip0, match0, offcode, matchEnd */
896 /* Count the forward length. */
897 assert(matchEnd != 0);
898 mLength += ZSTD_count_2segments(ip0 + mLength, match0 + mLength, iend, matchEnd, prefixStart);
900 ZSTD_storeSeq(seqStore, (size_t)(ip0 - anchor), anchor, iend, offcode, mLength);
905 /* write next hash table entry */
907 hashTable[hash1] = (U32)(ip1 - base);
910 /* Fill table and check for immediate repcode. */
913 assert(base+current0+2 > istart); /* check base overflow */
914 hashTable[ZSTD_hashPtr(base+current0+2, hlog, mls)] = current0+2; /* here because current+2 could be > iend-8 */
915 hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base);
917 while (ip0 <= ilimit) {
918 U32 const repIndex2 = (U32)(ip0-base) - offset_2;
919 const BYTE* const repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2;
920 if ( (((U32)((prefixStartIndex-1) - repIndex2) >= 3) & (offset_2 > 0)) /* intentional underflow */
921 && (MEM_read32(repMatch2) == MEM_read32(ip0)) ) {
922 const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
923 size_t const repLength2 = ZSTD_count_2segments(ip0+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
924 { U32 const tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; } /* swap offset_2 <=> offset_1 */
925 ZSTD_storeSeq(seqStore, 0 /*litlen*/, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
926 hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = (U32)(ip0-base);
937 ZSTD_GEN_FAST_FN(extDict, 4, 0)
938 ZSTD_GEN_FAST_FN(extDict, 5, 0)
939 ZSTD_GEN_FAST_FN(extDict, 6, 0)
940 ZSTD_GEN_FAST_FN(extDict, 7, 0)
942 size_t ZSTD_compressBlock_fast_extDict(
943 ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
944 void const* src, size_t srcSize)
946 U32 const mls = ms->cParams.minMatch;
947 assert(ms->dictMatchState == NULL);
950 default: /* includes case 3 */
952 return ZSTD_compressBlock_fast_extDict_4_0(ms, seqStore, rep, src, srcSize);
954 return ZSTD_compressBlock_fast_extDict_5_0(ms, seqStore, rep, src, srcSize);
956 return ZSTD_compressBlock_fast_extDict_6_0(ms, seqStore, rep, src, srcSize);
958 return ZSTD_compressBlock_fast_extDict_7_0(ms, seqStore, rep, src, srcSize);