xref: /openbmc/linux/lib/zstd/compress/zstd_fast.c (revision e0c1b49f5b674cca7b10549c53b3791d0bbc90a8)
1*e0c1b49fSNick Terrell /*
2*e0c1b49fSNick Terrell  * Copyright (c) Yann Collet, Facebook, Inc.
3*e0c1b49fSNick Terrell  * All rights reserved.
4*e0c1b49fSNick Terrell  *
5*e0c1b49fSNick Terrell  * This source code is licensed under both the BSD-style license (found in the
6*e0c1b49fSNick Terrell  * LICENSE file in the root directory of this source tree) and the GPLv2 (found
7*e0c1b49fSNick Terrell  * in the COPYING file in the root directory of this source tree).
8*e0c1b49fSNick Terrell  * You may select, at your option, one of the above-listed licenses.
9*e0c1b49fSNick Terrell  */
10*e0c1b49fSNick Terrell 
11*e0c1b49fSNick Terrell #include "zstd_compress_internal.h"  /* ZSTD_hashPtr, ZSTD_count, ZSTD_storeSeq */
12*e0c1b49fSNick Terrell #include "zstd_fast.h"
13*e0c1b49fSNick Terrell 
14*e0c1b49fSNick Terrell 
15*e0c1b49fSNick Terrell void ZSTD_fillHashTable(ZSTD_matchState_t* ms,
16*e0c1b49fSNick Terrell                         const void* const end,
17*e0c1b49fSNick Terrell                         ZSTD_dictTableLoadMethod_e dtlm)
18*e0c1b49fSNick Terrell {
19*e0c1b49fSNick Terrell     const ZSTD_compressionParameters* const cParams = &ms->cParams;
20*e0c1b49fSNick Terrell     U32* const hashTable = ms->hashTable;
21*e0c1b49fSNick Terrell     U32  const hBits = cParams->hashLog;
22*e0c1b49fSNick Terrell     U32  const mls = cParams->minMatch;
23*e0c1b49fSNick Terrell     const BYTE* const base = ms->window.base;
24*e0c1b49fSNick Terrell     const BYTE* ip = base + ms->nextToUpdate;
25*e0c1b49fSNick Terrell     const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE;
26*e0c1b49fSNick Terrell     const U32 fastHashFillStep = 3;
27*e0c1b49fSNick Terrell 
28*e0c1b49fSNick Terrell     /* Always insert every fastHashFillStep position into the hash table.
29*e0c1b49fSNick Terrell      * Insert the other positions if their hash entry is empty.
30*e0c1b49fSNick Terrell      */
31*e0c1b49fSNick Terrell     for ( ; ip + fastHashFillStep < iend + 2; ip += fastHashFillStep) {
32*e0c1b49fSNick Terrell         U32 const curr = (U32)(ip - base);
33*e0c1b49fSNick Terrell         size_t const hash0 = ZSTD_hashPtr(ip, hBits, mls);
34*e0c1b49fSNick Terrell         hashTable[hash0] = curr;
35*e0c1b49fSNick Terrell         if (dtlm == ZSTD_dtlm_fast) continue;
36*e0c1b49fSNick Terrell         /* Only load extra positions for ZSTD_dtlm_full */
37*e0c1b49fSNick Terrell         {   U32 p;
38*e0c1b49fSNick Terrell             for (p = 1; p < fastHashFillStep; ++p) {
39*e0c1b49fSNick Terrell                 size_t const hash = ZSTD_hashPtr(ip + p, hBits, mls);
40*e0c1b49fSNick Terrell                 if (hashTable[hash] == 0) {  /* not yet filled */
41*e0c1b49fSNick Terrell                     hashTable[hash] = curr + p;
42*e0c1b49fSNick Terrell     }   }   }   }
43*e0c1b49fSNick Terrell }
44*e0c1b49fSNick Terrell 
45*e0c1b49fSNick Terrell 
46*e0c1b49fSNick Terrell FORCE_INLINE_TEMPLATE size_t
47*e0c1b49fSNick Terrell ZSTD_compressBlock_fast_generic(
48*e0c1b49fSNick Terrell         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
49*e0c1b49fSNick Terrell         void const* src, size_t srcSize,
50*e0c1b49fSNick Terrell         U32 const mls)
51*e0c1b49fSNick Terrell {
52*e0c1b49fSNick Terrell     const ZSTD_compressionParameters* const cParams = &ms->cParams;
53*e0c1b49fSNick Terrell     U32* const hashTable = ms->hashTable;
54*e0c1b49fSNick Terrell     U32 const hlog = cParams->hashLog;
55*e0c1b49fSNick Terrell     /* support stepSize of 0 */
56*e0c1b49fSNick Terrell     size_t const stepSize = cParams->targetLength + !(cParams->targetLength) + 1;
57*e0c1b49fSNick Terrell     const BYTE* const base = ms->window.base;
58*e0c1b49fSNick Terrell     const BYTE* const istart = (const BYTE*)src;
59*e0c1b49fSNick Terrell     /* We check ip0 (ip + 0) and ip1 (ip + 1) each loop */
60*e0c1b49fSNick Terrell     const BYTE* ip0 = istart;
61*e0c1b49fSNick Terrell     const BYTE* ip1;
62*e0c1b49fSNick Terrell     const BYTE* anchor = istart;
63*e0c1b49fSNick Terrell     const U32   endIndex = (U32)((size_t)(istart - base) + srcSize);
64*e0c1b49fSNick Terrell     const U32   prefixStartIndex = ZSTD_getLowestPrefixIndex(ms, endIndex, cParams->windowLog);
65*e0c1b49fSNick Terrell     const BYTE* const prefixStart = base + prefixStartIndex;
66*e0c1b49fSNick Terrell     const BYTE* const iend = istart + srcSize;
67*e0c1b49fSNick Terrell     const BYTE* const ilimit = iend - HASH_READ_SIZE;
68*e0c1b49fSNick Terrell     U32 offset_1=rep[0], offset_2=rep[1];
69*e0c1b49fSNick Terrell     U32 offsetSaved = 0;
70*e0c1b49fSNick Terrell 
71*e0c1b49fSNick Terrell     /* init */
72*e0c1b49fSNick Terrell     DEBUGLOG(5, "ZSTD_compressBlock_fast_generic");
73*e0c1b49fSNick Terrell     ip0 += (ip0 == prefixStart);
74*e0c1b49fSNick Terrell     ip1 = ip0 + 1;
75*e0c1b49fSNick Terrell     {   U32 const curr = (U32)(ip0 - base);
76*e0c1b49fSNick Terrell         U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, curr, cParams->windowLog);
77*e0c1b49fSNick Terrell         U32 const maxRep = curr - windowLow;
78*e0c1b49fSNick Terrell         if (offset_2 > maxRep) offsetSaved = offset_2, offset_2 = 0;
79*e0c1b49fSNick Terrell         if (offset_1 > maxRep) offsetSaved = offset_1, offset_1 = 0;
80*e0c1b49fSNick Terrell     }
81*e0c1b49fSNick Terrell 
82*e0c1b49fSNick Terrell     /* Main Search Loop */
83*e0c1b49fSNick Terrell #ifdef __INTEL_COMPILER
84*e0c1b49fSNick Terrell     /* From intel 'The vector pragma indicates that the loop should be
85*e0c1b49fSNick Terrell      * vectorized if it is legal to do so'. Can be used together with
86*e0c1b49fSNick Terrell      * #pragma ivdep (but have opted to exclude that because intel
87*e0c1b49fSNick Terrell      * warns against using it).*/
88*e0c1b49fSNick Terrell     #pragma vector always
89*e0c1b49fSNick Terrell #endif
90*e0c1b49fSNick Terrell     while (ip1 < ilimit) {   /* < instead of <=, because check at ip0+2 */
91*e0c1b49fSNick Terrell         size_t mLength;
92*e0c1b49fSNick Terrell         BYTE const* ip2 = ip0 + 2;
93*e0c1b49fSNick Terrell         size_t const h0 = ZSTD_hashPtr(ip0, hlog, mls);
94*e0c1b49fSNick Terrell         U32 const val0 = MEM_read32(ip0);
95*e0c1b49fSNick Terrell         size_t const h1 = ZSTD_hashPtr(ip1, hlog, mls);
96*e0c1b49fSNick Terrell         U32 const val1 = MEM_read32(ip1);
97*e0c1b49fSNick Terrell         U32 const current0 = (U32)(ip0-base);
98*e0c1b49fSNick Terrell         U32 const current1 = (U32)(ip1-base);
99*e0c1b49fSNick Terrell         U32 const matchIndex0 = hashTable[h0];
100*e0c1b49fSNick Terrell         U32 const matchIndex1 = hashTable[h1];
101*e0c1b49fSNick Terrell         BYTE const* repMatch = ip2 - offset_1;
102*e0c1b49fSNick Terrell         const BYTE* match0 = base + matchIndex0;
103*e0c1b49fSNick Terrell         const BYTE* match1 = base + matchIndex1;
104*e0c1b49fSNick Terrell         U32 offcode;
105*e0c1b49fSNick Terrell 
106*e0c1b49fSNick Terrell #if defined(__aarch64__)
107*e0c1b49fSNick Terrell         PREFETCH_L1(ip0+256);
108*e0c1b49fSNick Terrell #endif
109*e0c1b49fSNick Terrell 
110*e0c1b49fSNick Terrell         hashTable[h0] = current0;   /* update hash table */
111*e0c1b49fSNick Terrell         hashTable[h1] = current1;   /* update hash table */
112*e0c1b49fSNick Terrell 
113*e0c1b49fSNick Terrell         assert(ip0 + 1 == ip1);
114*e0c1b49fSNick Terrell 
115*e0c1b49fSNick Terrell         if ((offset_1 > 0) & (MEM_read32(repMatch) == MEM_read32(ip2))) {
116*e0c1b49fSNick Terrell             mLength = (ip2[-1] == repMatch[-1]) ? 1 : 0;
117*e0c1b49fSNick Terrell             ip0 = ip2 - mLength;
118*e0c1b49fSNick Terrell             match0 = repMatch - mLength;
119*e0c1b49fSNick Terrell             mLength += 4;
120*e0c1b49fSNick Terrell             offcode = 0;
121*e0c1b49fSNick Terrell             goto _match;
122*e0c1b49fSNick Terrell         }
123*e0c1b49fSNick Terrell         if ((matchIndex0 > prefixStartIndex) && MEM_read32(match0) == val0) {
124*e0c1b49fSNick Terrell             /* found a regular match */
125*e0c1b49fSNick Terrell             goto _offset;
126*e0c1b49fSNick Terrell         }
127*e0c1b49fSNick Terrell         if ((matchIndex1 > prefixStartIndex) && MEM_read32(match1) == val1) {
128*e0c1b49fSNick Terrell             /* found a regular match after one literal */
129*e0c1b49fSNick Terrell             ip0 = ip1;
130*e0c1b49fSNick Terrell             match0 = match1;
131*e0c1b49fSNick Terrell             goto _offset;
132*e0c1b49fSNick Terrell         }
133*e0c1b49fSNick Terrell         {   size_t const step = ((size_t)(ip0-anchor) >> (kSearchStrength - 1)) + stepSize;
134*e0c1b49fSNick Terrell             assert(step >= 2);
135*e0c1b49fSNick Terrell             ip0 += step;
136*e0c1b49fSNick Terrell             ip1 += step;
137*e0c1b49fSNick Terrell             continue;
138*e0c1b49fSNick Terrell         }
139*e0c1b49fSNick Terrell _offset: /* Requires: ip0, match0 */
140*e0c1b49fSNick Terrell         /* Compute the offset code */
141*e0c1b49fSNick Terrell         offset_2 = offset_1;
142*e0c1b49fSNick Terrell         offset_1 = (U32)(ip0-match0);
143*e0c1b49fSNick Terrell         offcode = offset_1 + ZSTD_REP_MOVE;
144*e0c1b49fSNick Terrell         mLength = 4;
145*e0c1b49fSNick Terrell         /* Count the backwards match length */
146*e0c1b49fSNick Terrell         while (((ip0>anchor) & (match0>prefixStart))
147*e0c1b49fSNick Terrell              && (ip0[-1] == match0[-1])) { ip0--; match0--; mLength++; } /* catch up */
148*e0c1b49fSNick Terrell 
149*e0c1b49fSNick Terrell _match: /* Requires: ip0, match0, offcode */
150*e0c1b49fSNick Terrell         /* Count the forward length */
151*e0c1b49fSNick Terrell         mLength += ZSTD_count(ip0+mLength, match0+mLength, iend);
152*e0c1b49fSNick Terrell         ZSTD_storeSeq(seqStore, (size_t)(ip0-anchor), anchor, iend, offcode, mLength-MINMATCH);
153*e0c1b49fSNick Terrell         /* match found */
154*e0c1b49fSNick Terrell         ip0 += mLength;
155*e0c1b49fSNick Terrell         anchor = ip0;
156*e0c1b49fSNick Terrell 
157*e0c1b49fSNick Terrell         if (ip0 <= ilimit) {
158*e0c1b49fSNick Terrell             /* Fill Table */
159*e0c1b49fSNick Terrell             assert(base+current0+2 > istart);  /* check base overflow */
160*e0c1b49fSNick Terrell             hashTable[ZSTD_hashPtr(base+current0+2, hlog, mls)] = current0+2;  /* here because current+2 could be > iend-8 */
161*e0c1b49fSNick Terrell             hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base);
162*e0c1b49fSNick Terrell 
163*e0c1b49fSNick Terrell             if (offset_2 > 0) { /* offset_2==0 means offset_2 is invalidated */
164*e0c1b49fSNick Terrell                 while ( (ip0 <= ilimit) && (MEM_read32(ip0) == MEM_read32(ip0 - offset_2)) ) {
165*e0c1b49fSNick Terrell                     /* store sequence */
166*e0c1b49fSNick Terrell                     size_t const rLength = ZSTD_count(ip0+4, ip0+4-offset_2, iend) + 4;
167*e0c1b49fSNick Terrell                     { U32 const tmpOff = offset_2; offset_2 = offset_1; offset_1 = tmpOff; } /* swap offset_2 <=> offset_1 */
168*e0c1b49fSNick Terrell                     hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = (U32)(ip0-base);
169*e0c1b49fSNick Terrell                     ip0 += rLength;
170*e0c1b49fSNick Terrell                     ZSTD_storeSeq(seqStore, 0 /*litLen*/, anchor, iend, 0 /*offCode*/, rLength-MINMATCH);
171*e0c1b49fSNick Terrell                     anchor = ip0;
172*e0c1b49fSNick Terrell                     continue;   /* faster when present (confirmed on gcc-8) ... (?) */
173*e0c1b49fSNick Terrell         }   }   }
174*e0c1b49fSNick Terrell         ip1 = ip0 + 1;
175*e0c1b49fSNick Terrell     }
176*e0c1b49fSNick Terrell 
177*e0c1b49fSNick Terrell     /* save reps for next block */
178*e0c1b49fSNick Terrell     rep[0] = offset_1 ? offset_1 : offsetSaved;
179*e0c1b49fSNick Terrell     rep[1] = offset_2 ? offset_2 : offsetSaved;
180*e0c1b49fSNick Terrell 
181*e0c1b49fSNick Terrell     /* Return the last literals size */
182*e0c1b49fSNick Terrell     return (size_t)(iend - anchor);
183*e0c1b49fSNick Terrell }
184*e0c1b49fSNick Terrell 
185*e0c1b49fSNick Terrell 
186*e0c1b49fSNick Terrell size_t ZSTD_compressBlock_fast(
187*e0c1b49fSNick Terrell         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
188*e0c1b49fSNick Terrell         void const* src, size_t srcSize)
189*e0c1b49fSNick Terrell {
190*e0c1b49fSNick Terrell     U32 const mls = ms->cParams.minMatch;
191*e0c1b49fSNick Terrell     assert(ms->dictMatchState == NULL);
192*e0c1b49fSNick Terrell     switch(mls)
193*e0c1b49fSNick Terrell     {
194*e0c1b49fSNick Terrell     default: /* includes case 3 */
195*e0c1b49fSNick Terrell     case 4 :
196*e0c1b49fSNick Terrell         return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, 4);
197*e0c1b49fSNick Terrell     case 5 :
198*e0c1b49fSNick Terrell         return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, 5);
199*e0c1b49fSNick Terrell     case 6 :
200*e0c1b49fSNick Terrell         return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, 6);
201*e0c1b49fSNick Terrell     case 7 :
202*e0c1b49fSNick Terrell         return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, 7);
203*e0c1b49fSNick Terrell     }
204*e0c1b49fSNick Terrell }
205*e0c1b49fSNick Terrell 
206*e0c1b49fSNick Terrell FORCE_INLINE_TEMPLATE
207*e0c1b49fSNick Terrell size_t ZSTD_compressBlock_fast_dictMatchState_generic(
208*e0c1b49fSNick Terrell         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
209*e0c1b49fSNick Terrell         void const* src, size_t srcSize, U32 const mls)
210*e0c1b49fSNick Terrell {
211*e0c1b49fSNick Terrell     const ZSTD_compressionParameters* const cParams = &ms->cParams;
212*e0c1b49fSNick Terrell     U32* const hashTable = ms->hashTable;
213*e0c1b49fSNick Terrell     U32 const hlog = cParams->hashLog;
214*e0c1b49fSNick Terrell     /* support stepSize of 0 */
215*e0c1b49fSNick Terrell     U32 const stepSize = cParams->targetLength + !(cParams->targetLength);
216*e0c1b49fSNick Terrell     const BYTE* const base = ms->window.base;
217*e0c1b49fSNick Terrell     const BYTE* const istart = (const BYTE*)src;
218*e0c1b49fSNick Terrell     const BYTE* ip = istart;
219*e0c1b49fSNick Terrell     const BYTE* anchor = istart;
220*e0c1b49fSNick Terrell     const U32   prefixStartIndex = ms->window.dictLimit;
221*e0c1b49fSNick Terrell     const BYTE* const prefixStart = base + prefixStartIndex;
222*e0c1b49fSNick Terrell     const BYTE* const iend = istart + srcSize;
223*e0c1b49fSNick Terrell     const BYTE* const ilimit = iend - HASH_READ_SIZE;
224*e0c1b49fSNick Terrell     U32 offset_1=rep[0], offset_2=rep[1];
225*e0c1b49fSNick Terrell     U32 offsetSaved = 0;
226*e0c1b49fSNick Terrell 
227*e0c1b49fSNick Terrell     const ZSTD_matchState_t* const dms = ms->dictMatchState;
228*e0c1b49fSNick Terrell     const ZSTD_compressionParameters* const dictCParams = &dms->cParams ;
229*e0c1b49fSNick Terrell     const U32* const dictHashTable = dms->hashTable;
230*e0c1b49fSNick Terrell     const U32 dictStartIndex       = dms->window.dictLimit;
231*e0c1b49fSNick Terrell     const BYTE* const dictBase     = dms->window.base;
232*e0c1b49fSNick Terrell     const BYTE* const dictStart    = dictBase + dictStartIndex;
233*e0c1b49fSNick Terrell     const BYTE* const dictEnd      = dms->window.nextSrc;
234*e0c1b49fSNick Terrell     const U32 dictIndexDelta       = prefixStartIndex - (U32)(dictEnd - dictBase);
235*e0c1b49fSNick Terrell     const U32 dictAndPrefixLength  = (U32)(ip - prefixStart + dictEnd - dictStart);
236*e0c1b49fSNick Terrell     const U32 dictHLog             = dictCParams->hashLog;
237*e0c1b49fSNick Terrell 
238*e0c1b49fSNick Terrell     /* if a dictionary is still attached, it necessarily means that
239*e0c1b49fSNick Terrell      * it is within window size. So we just check it. */
240*e0c1b49fSNick Terrell     const U32 maxDistance = 1U << cParams->windowLog;
241*e0c1b49fSNick Terrell     const U32 endIndex = (U32)((size_t)(ip - base) + srcSize);
242*e0c1b49fSNick Terrell     assert(endIndex - prefixStartIndex <= maxDistance);
243*e0c1b49fSNick Terrell     (void)maxDistance; (void)endIndex;   /* these variables are not used when assert() is disabled */
244*e0c1b49fSNick Terrell 
245*e0c1b49fSNick Terrell     /* ensure there will be no underflow
246*e0c1b49fSNick Terrell      * when translating a dict index into a local index */
247*e0c1b49fSNick Terrell     assert(prefixStartIndex >= (U32)(dictEnd - dictBase));
248*e0c1b49fSNick Terrell 
249*e0c1b49fSNick Terrell     /* init */
250*e0c1b49fSNick Terrell     DEBUGLOG(5, "ZSTD_compressBlock_fast_dictMatchState_generic");
251*e0c1b49fSNick Terrell     ip += (dictAndPrefixLength == 0);
252*e0c1b49fSNick Terrell     /* dictMatchState repCode checks don't currently handle repCode == 0
253*e0c1b49fSNick Terrell      * disabling. */
254*e0c1b49fSNick Terrell     assert(offset_1 <= dictAndPrefixLength);
255*e0c1b49fSNick Terrell     assert(offset_2 <= dictAndPrefixLength);
256*e0c1b49fSNick Terrell 
257*e0c1b49fSNick Terrell     /* Main Search Loop */
258*e0c1b49fSNick Terrell     while (ip < ilimit) {   /* < instead of <=, because repcode check at (ip+1) */
259*e0c1b49fSNick Terrell         size_t mLength;
260*e0c1b49fSNick Terrell         size_t const h = ZSTD_hashPtr(ip, hlog, mls);
261*e0c1b49fSNick Terrell         U32 const curr = (U32)(ip-base);
262*e0c1b49fSNick Terrell         U32 const matchIndex = hashTable[h];
263*e0c1b49fSNick Terrell         const BYTE* match = base + matchIndex;
264*e0c1b49fSNick Terrell         const U32 repIndex = curr + 1 - offset_1;
265*e0c1b49fSNick Terrell         const BYTE* repMatch = (repIndex < prefixStartIndex) ?
266*e0c1b49fSNick Terrell                                dictBase + (repIndex - dictIndexDelta) :
267*e0c1b49fSNick Terrell                                base + repIndex;
268*e0c1b49fSNick Terrell         hashTable[h] = curr;   /* update hash table */
269*e0c1b49fSNick Terrell 
270*e0c1b49fSNick Terrell         if ( ((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow : ensure repIndex isn't overlapping dict + prefix */
271*e0c1b49fSNick Terrell           && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
272*e0c1b49fSNick Terrell             const BYTE* const repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
273*e0c1b49fSNick Terrell             mLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixStart) + 4;
274*e0c1b49fSNick Terrell             ip++;
275*e0c1b49fSNick Terrell             ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, 0, mLength-MINMATCH);
276*e0c1b49fSNick Terrell         } else if ( (matchIndex <= prefixStartIndex) ) {
277*e0c1b49fSNick Terrell             size_t const dictHash = ZSTD_hashPtr(ip, dictHLog, mls);
278*e0c1b49fSNick Terrell             U32 const dictMatchIndex = dictHashTable[dictHash];
279*e0c1b49fSNick Terrell             const BYTE* dictMatch = dictBase + dictMatchIndex;
280*e0c1b49fSNick Terrell             if (dictMatchIndex <= dictStartIndex ||
281*e0c1b49fSNick Terrell                 MEM_read32(dictMatch) != MEM_read32(ip)) {
282*e0c1b49fSNick Terrell                 assert(stepSize >= 1);
283*e0c1b49fSNick Terrell                 ip += ((ip-anchor) >> kSearchStrength) + stepSize;
284*e0c1b49fSNick Terrell                 continue;
285*e0c1b49fSNick Terrell             } else {
286*e0c1b49fSNick Terrell                 /* found a dict match */
287*e0c1b49fSNick Terrell                 U32 const offset = (U32)(curr-dictMatchIndex-dictIndexDelta);
288*e0c1b49fSNick Terrell                 mLength = ZSTD_count_2segments(ip+4, dictMatch+4, iend, dictEnd, prefixStart) + 4;
289*e0c1b49fSNick Terrell                 while (((ip>anchor) & (dictMatch>dictStart))
290*e0c1b49fSNick Terrell                      && (ip[-1] == dictMatch[-1])) {
291*e0c1b49fSNick Terrell                     ip--; dictMatch--; mLength++;
292*e0c1b49fSNick Terrell                 } /* catch up */
293*e0c1b49fSNick Terrell                 offset_2 = offset_1;
294*e0c1b49fSNick Terrell                 offset_1 = offset;
295*e0c1b49fSNick Terrell                 ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH);
296*e0c1b49fSNick Terrell             }
297*e0c1b49fSNick Terrell         } else if (MEM_read32(match) != MEM_read32(ip)) {
298*e0c1b49fSNick Terrell             /* it's not a match, and we're not going to check the dictionary */
299*e0c1b49fSNick Terrell             assert(stepSize >= 1);
300*e0c1b49fSNick Terrell             ip += ((ip-anchor) >> kSearchStrength) + stepSize;
301*e0c1b49fSNick Terrell             continue;
302*e0c1b49fSNick Terrell         } else {
303*e0c1b49fSNick Terrell             /* found a regular match */
304*e0c1b49fSNick Terrell             U32 const offset = (U32)(ip-match);
305*e0c1b49fSNick Terrell             mLength = ZSTD_count(ip+4, match+4, iend) + 4;
306*e0c1b49fSNick Terrell             while (((ip>anchor) & (match>prefixStart))
307*e0c1b49fSNick Terrell                  && (ip[-1] == match[-1])) { ip--; match--; mLength++; } /* catch up */
308*e0c1b49fSNick Terrell             offset_2 = offset_1;
309*e0c1b49fSNick Terrell             offset_1 = offset;
310*e0c1b49fSNick Terrell             ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH);
311*e0c1b49fSNick Terrell         }
312*e0c1b49fSNick Terrell 
313*e0c1b49fSNick Terrell         /* match found */
314*e0c1b49fSNick Terrell         ip += mLength;
315*e0c1b49fSNick Terrell         anchor = ip;
316*e0c1b49fSNick Terrell 
317*e0c1b49fSNick Terrell         if (ip <= ilimit) {
318*e0c1b49fSNick Terrell             /* Fill Table */
319*e0c1b49fSNick Terrell             assert(base+curr+2 > istart);  /* check base overflow */
320*e0c1b49fSNick Terrell             hashTable[ZSTD_hashPtr(base+curr+2, hlog, mls)] = curr+2;  /* here because curr+2 could be > iend-8 */
321*e0c1b49fSNick Terrell             hashTable[ZSTD_hashPtr(ip-2, hlog, mls)] = (U32)(ip-2-base);
322*e0c1b49fSNick Terrell 
323*e0c1b49fSNick Terrell             /* check immediate repcode */
324*e0c1b49fSNick Terrell             while (ip <= ilimit) {
325*e0c1b49fSNick Terrell                 U32 const current2 = (U32)(ip-base);
326*e0c1b49fSNick Terrell                 U32 const repIndex2 = current2 - offset_2;
327*e0c1b49fSNick Terrell                 const BYTE* repMatch2 = repIndex2 < prefixStartIndex ?
328*e0c1b49fSNick Terrell                         dictBase - dictIndexDelta + repIndex2 :
329*e0c1b49fSNick Terrell                         base + repIndex2;
330*e0c1b49fSNick Terrell                 if ( ((U32)((prefixStartIndex-1) - (U32)repIndex2) >= 3 /* intentional overflow */)
331*e0c1b49fSNick Terrell                    && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
332*e0c1b49fSNick Terrell                     const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
333*e0c1b49fSNick Terrell                     size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
334*e0c1b49fSNick Terrell                     U32 tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset;   /* swap offset_2 <=> offset_1 */
335*e0c1b49fSNick Terrell                     ZSTD_storeSeq(seqStore, 0, anchor, iend, 0, repLength2-MINMATCH);
336*e0c1b49fSNick Terrell                     hashTable[ZSTD_hashPtr(ip, hlog, mls)] = current2;
337*e0c1b49fSNick Terrell                     ip += repLength2;
338*e0c1b49fSNick Terrell                     anchor = ip;
339*e0c1b49fSNick Terrell                     continue;
340*e0c1b49fSNick Terrell                 }
341*e0c1b49fSNick Terrell                 break;
342*e0c1b49fSNick Terrell             }
343*e0c1b49fSNick Terrell         }
344*e0c1b49fSNick Terrell     }
345*e0c1b49fSNick Terrell 
346*e0c1b49fSNick Terrell     /* save reps for next block */
347*e0c1b49fSNick Terrell     rep[0] = offset_1 ? offset_1 : offsetSaved;
348*e0c1b49fSNick Terrell     rep[1] = offset_2 ? offset_2 : offsetSaved;
349*e0c1b49fSNick Terrell 
350*e0c1b49fSNick Terrell     /* Return the last literals size */
351*e0c1b49fSNick Terrell     return (size_t)(iend - anchor);
352*e0c1b49fSNick Terrell }
353*e0c1b49fSNick Terrell 
354*e0c1b49fSNick Terrell size_t ZSTD_compressBlock_fast_dictMatchState(
355*e0c1b49fSNick Terrell         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
356*e0c1b49fSNick Terrell         void const* src, size_t srcSize)
357*e0c1b49fSNick Terrell {
358*e0c1b49fSNick Terrell     U32 const mls = ms->cParams.minMatch;
359*e0c1b49fSNick Terrell     assert(ms->dictMatchState != NULL);
360*e0c1b49fSNick Terrell     switch(mls)
361*e0c1b49fSNick Terrell     {
362*e0c1b49fSNick Terrell     default: /* includes case 3 */
363*e0c1b49fSNick Terrell     case 4 :
364*e0c1b49fSNick Terrell         return ZSTD_compressBlock_fast_dictMatchState_generic(ms, seqStore, rep, src, srcSize, 4);
365*e0c1b49fSNick Terrell     case 5 :
366*e0c1b49fSNick Terrell         return ZSTD_compressBlock_fast_dictMatchState_generic(ms, seqStore, rep, src, srcSize, 5);
367*e0c1b49fSNick Terrell     case 6 :
368*e0c1b49fSNick Terrell         return ZSTD_compressBlock_fast_dictMatchState_generic(ms, seqStore, rep, src, srcSize, 6);
369*e0c1b49fSNick Terrell     case 7 :
370*e0c1b49fSNick Terrell         return ZSTD_compressBlock_fast_dictMatchState_generic(ms, seqStore, rep, src, srcSize, 7);
371*e0c1b49fSNick Terrell     }
372*e0c1b49fSNick Terrell }
373*e0c1b49fSNick Terrell 
374*e0c1b49fSNick Terrell 
375*e0c1b49fSNick Terrell static size_t ZSTD_compressBlock_fast_extDict_generic(
376*e0c1b49fSNick Terrell         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
377*e0c1b49fSNick Terrell         void const* src, size_t srcSize, U32 const mls)
378*e0c1b49fSNick Terrell {
379*e0c1b49fSNick Terrell     const ZSTD_compressionParameters* const cParams = &ms->cParams;
380*e0c1b49fSNick Terrell     U32* const hashTable = ms->hashTable;
381*e0c1b49fSNick Terrell     U32 const hlog = cParams->hashLog;
382*e0c1b49fSNick Terrell     /* support stepSize of 0 */
383*e0c1b49fSNick Terrell     U32 const stepSize = cParams->targetLength + !(cParams->targetLength);
384*e0c1b49fSNick Terrell     const BYTE* const base = ms->window.base;
385*e0c1b49fSNick Terrell     const BYTE* const dictBase = ms->window.dictBase;
386*e0c1b49fSNick Terrell     const BYTE* const istart = (const BYTE*)src;
387*e0c1b49fSNick Terrell     const BYTE* ip = istart;
388*e0c1b49fSNick Terrell     const BYTE* anchor = istart;
389*e0c1b49fSNick Terrell     const U32   endIndex = (U32)((size_t)(istart - base) + srcSize);
390*e0c1b49fSNick Terrell     const U32   lowLimit = ZSTD_getLowestMatchIndex(ms, endIndex, cParams->windowLog);
391*e0c1b49fSNick Terrell     const U32   dictStartIndex = lowLimit;
392*e0c1b49fSNick Terrell     const BYTE* const dictStart = dictBase + dictStartIndex;
393*e0c1b49fSNick Terrell     const U32   dictLimit = ms->window.dictLimit;
394*e0c1b49fSNick Terrell     const U32   prefixStartIndex = dictLimit < lowLimit ? lowLimit : dictLimit;
395*e0c1b49fSNick Terrell     const BYTE* const prefixStart = base + prefixStartIndex;
396*e0c1b49fSNick Terrell     const BYTE* const dictEnd = dictBase + prefixStartIndex;
397*e0c1b49fSNick Terrell     const BYTE* const iend = istart + srcSize;
398*e0c1b49fSNick Terrell     const BYTE* const ilimit = iend - 8;
399*e0c1b49fSNick Terrell     U32 offset_1=rep[0], offset_2=rep[1];
400*e0c1b49fSNick Terrell 
401*e0c1b49fSNick Terrell     DEBUGLOG(5, "ZSTD_compressBlock_fast_extDict_generic (offset_1=%u)", offset_1);
402*e0c1b49fSNick Terrell 
403*e0c1b49fSNick Terrell     /* switch to "regular" variant if extDict is invalidated due to maxDistance */
404*e0c1b49fSNick Terrell     if (prefixStartIndex == dictStartIndex)
405*e0c1b49fSNick Terrell         return ZSTD_compressBlock_fast_generic(ms, seqStore, rep, src, srcSize, mls);
406*e0c1b49fSNick Terrell 
407*e0c1b49fSNick Terrell     /* Search Loop */
408*e0c1b49fSNick Terrell     while (ip < ilimit) {  /* < instead of <=, because (ip+1) */
409*e0c1b49fSNick Terrell         const size_t h = ZSTD_hashPtr(ip, hlog, mls);
410*e0c1b49fSNick Terrell         const U32    matchIndex = hashTable[h];
411*e0c1b49fSNick Terrell         const BYTE* const matchBase = matchIndex < prefixStartIndex ? dictBase : base;
412*e0c1b49fSNick Terrell         const BYTE*  match = matchBase + matchIndex;
413*e0c1b49fSNick Terrell         const U32    curr = (U32)(ip-base);
414*e0c1b49fSNick Terrell         const U32    repIndex = curr + 1 - offset_1;
415*e0c1b49fSNick Terrell         const BYTE* const repBase = repIndex < prefixStartIndex ? dictBase : base;
416*e0c1b49fSNick Terrell         const BYTE* const repMatch = repBase + repIndex;
417*e0c1b49fSNick Terrell         hashTable[h] = curr;   /* update hash table */
418*e0c1b49fSNick Terrell         DEBUGLOG(7, "offset_1 = %u , curr = %u", offset_1, curr);
419*e0c1b49fSNick Terrell         assert(offset_1 <= curr +1);   /* check repIndex */
420*e0c1b49fSNick Terrell 
421*e0c1b49fSNick Terrell         if ( (((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow */ & (repIndex > dictStartIndex))
422*e0c1b49fSNick Terrell            && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
423*e0c1b49fSNick Terrell             const BYTE* const repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
424*e0c1b49fSNick Terrell             size_t const rLength = ZSTD_count_2segments(ip+1 +4, repMatch +4, iend, repMatchEnd, prefixStart) + 4;
425*e0c1b49fSNick Terrell             ip++;
426*e0c1b49fSNick Terrell             ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, 0, rLength-MINMATCH);
427*e0c1b49fSNick Terrell             ip += rLength;
428*e0c1b49fSNick Terrell             anchor = ip;
429*e0c1b49fSNick Terrell         } else {
430*e0c1b49fSNick Terrell             if ( (matchIndex < dictStartIndex) ||
431*e0c1b49fSNick Terrell                  (MEM_read32(match) != MEM_read32(ip)) ) {
432*e0c1b49fSNick Terrell                 assert(stepSize >= 1);
433*e0c1b49fSNick Terrell                 ip += ((ip-anchor) >> kSearchStrength) + stepSize;
434*e0c1b49fSNick Terrell                 continue;
435*e0c1b49fSNick Terrell             }
436*e0c1b49fSNick Terrell             {   const BYTE* const matchEnd = matchIndex < prefixStartIndex ? dictEnd : iend;
437*e0c1b49fSNick Terrell                 const BYTE* const lowMatchPtr = matchIndex < prefixStartIndex ? dictStart : prefixStart;
438*e0c1b49fSNick Terrell                 U32 const offset = curr - matchIndex;
439*e0c1b49fSNick Terrell                 size_t mLength = ZSTD_count_2segments(ip+4, match+4, iend, matchEnd, prefixStart) + 4;
440*e0c1b49fSNick Terrell                 while (((ip>anchor) & (match>lowMatchPtr)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; }   /* catch up */
441*e0c1b49fSNick Terrell                 offset_2 = offset_1; offset_1 = offset;  /* update offset history */
442*e0c1b49fSNick Terrell                 ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, offset + ZSTD_REP_MOVE, mLength-MINMATCH);
443*e0c1b49fSNick Terrell                 ip += mLength;
444*e0c1b49fSNick Terrell                 anchor = ip;
445*e0c1b49fSNick Terrell         }   }
446*e0c1b49fSNick Terrell 
447*e0c1b49fSNick Terrell         if (ip <= ilimit) {
448*e0c1b49fSNick Terrell             /* Fill Table */
449*e0c1b49fSNick Terrell             hashTable[ZSTD_hashPtr(base+curr+2, hlog, mls)] = curr+2;
450*e0c1b49fSNick Terrell             hashTable[ZSTD_hashPtr(ip-2, hlog, mls)] = (U32)(ip-2-base);
451*e0c1b49fSNick Terrell             /* check immediate repcode */
452*e0c1b49fSNick Terrell             while (ip <= ilimit) {
453*e0c1b49fSNick Terrell                 U32 const current2 = (U32)(ip-base);
454*e0c1b49fSNick Terrell                 U32 const repIndex2 = current2 - offset_2;
455*e0c1b49fSNick Terrell                 const BYTE* const repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2;
456*e0c1b49fSNick Terrell                 if ( (((U32)((prefixStartIndex-1) - repIndex2) >= 3) & (repIndex2 > dictStartIndex))  /* intentional overflow */
457*e0c1b49fSNick Terrell                    && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
458*e0c1b49fSNick Terrell                     const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
459*e0c1b49fSNick Terrell                     size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
460*e0c1b49fSNick Terrell                     { U32 const tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; }  /* swap offset_2 <=> offset_1 */
461*e0c1b49fSNick Terrell                     ZSTD_storeSeq(seqStore, 0 /*litlen*/, anchor, iend, 0 /*offcode*/, repLength2-MINMATCH);
462*e0c1b49fSNick Terrell                     hashTable[ZSTD_hashPtr(ip, hlog, mls)] = current2;
463*e0c1b49fSNick Terrell                     ip += repLength2;
464*e0c1b49fSNick Terrell                     anchor = ip;
465*e0c1b49fSNick Terrell                     continue;
466*e0c1b49fSNick Terrell                 }
467*e0c1b49fSNick Terrell                 break;
468*e0c1b49fSNick Terrell     }   }   }
469*e0c1b49fSNick Terrell 
470*e0c1b49fSNick Terrell     /* save reps for next block */
471*e0c1b49fSNick Terrell     rep[0] = offset_1;
472*e0c1b49fSNick Terrell     rep[1] = offset_2;
473*e0c1b49fSNick Terrell 
474*e0c1b49fSNick Terrell     /* Return the last literals size */
475*e0c1b49fSNick Terrell     return (size_t)(iend - anchor);
476*e0c1b49fSNick Terrell }
477*e0c1b49fSNick Terrell 
478*e0c1b49fSNick Terrell 
479*e0c1b49fSNick Terrell size_t ZSTD_compressBlock_fast_extDict(
480*e0c1b49fSNick Terrell         ZSTD_matchState_t* ms, seqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
481*e0c1b49fSNick Terrell         void const* src, size_t srcSize)
482*e0c1b49fSNick Terrell {
483*e0c1b49fSNick Terrell     U32 const mls = ms->cParams.minMatch;
484*e0c1b49fSNick Terrell     switch(mls)
485*e0c1b49fSNick Terrell     {
486*e0c1b49fSNick Terrell     default: /* includes case 3 */
487*e0c1b49fSNick Terrell     case 4 :
488*e0c1b49fSNick Terrell         return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, 4);
489*e0c1b49fSNick Terrell     case 5 :
490*e0c1b49fSNick Terrell         return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, 5);
491*e0c1b49fSNick Terrell     case 6 :
492*e0c1b49fSNick Terrell         return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, 6);
493*e0c1b49fSNick Terrell     case 7 :
494*e0c1b49fSNick Terrell         return ZSTD_compressBlock_fast_extDict_generic(ms, seqStore, rep, src, srcSize, 7);
495*e0c1b49fSNick Terrell     }
496*e0c1b49fSNick Terrell }
497