notaz.gp2x.de / pcsx_rearmed.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
git subrepo pull (merge) --force deps/libchdr
[pcsx_rearmed.git] / deps / libchdr / deps / zstd-1.5.5 / lib / decompress / huf_decompress.c
CommitLineData
648db22b 1/* ******************************************************************
2 * huff0 huffman decoder,
3 * part of Finite State Entropy library
4 * Copyright (c) Meta Platforms, Inc. and affiliates.
5 *
6 * You can contact the author at :
7 * - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
8 *
9 * This source code is licensed under both the BSD-style license (found in the
10 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
11 * in the COPYING file in the root directory of this source tree).
12 * You may select, at your option, one of the above-listed licenses.
13****************************************************************** */
14
15/* **************************************************************
16* Dependencies
17****************************************************************/
18#include "../common/zstd_deps.h" /* ZSTD_memcpy, ZSTD_memset */
19#include "../common/compiler.h"
20#include "../common/bitstream.h" /* BIT_* */
21#include "../common/fse.h" /* to compress headers */
22#include "../common/huf.h"
23#include "../common/error_private.h"
24#include "../common/zstd_internal.h"
25#include "../common/bits.h" /* ZSTD_highbit32, ZSTD_countTrailingZeros64 */
26
27/* **************************************************************
28* Constants
29****************************************************************/
30
31#define HUF_DECODER_FAST_TABLELOG 11
32
33/* **************************************************************
34* Macros
35****************************************************************/
36
37/* These two optional macros force the use one way or another of the two
38 * Huffman decompression implementations. You can't force in both directions
39 * at the same time.
40 */
41#if defined(HUF_FORCE_DECOMPRESS_X1) && \
42 defined(HUF_FORCE_DECOMPRESS_X2)
43#error "Cannot force the use of the X1 and X2 decoders at the same time!"
44#endif
45
46/* When DYNAMIC_BMI2 is enabled, fast decoders are only called when bmi2 is
47 * supported at runtime, so we can add the BMI2 target attribute.
48 * When it is disabled, we will still get BMI2 if it is enabled statically.
49 */
50#if DYNAMIC_BMI2
51# define HUF_FAST_BMI2_ATTRS BMI2_TARGET_ATTRIBUTE
52#else
53# define HUF_FAST_BMI2_ATTRS
54#endif
55
56#ifdef __cplusplus
57# define HUF_EXTERN_C extern "C"
58#else
59# define HUF_EXTERN_C
60#endif
61#define HUF_ASM_DECL HUF_EXTERN_C
62
63#if DYNAMIC_BMI2
64# define HUF_NEED_BMI2_FUNCTION 1
65#else
66# define HUF_NEED_BMI2_FUNCTION 0
67#endif
68
69/* **************************************************************
70* Error Management
71****************************************************************/
72#define HUF_isError ERR_isError
73
74
75/* **************************************************************
76* Byte alignment for workSpace management
77****************************************************************/
78#define HUF_ALIGN(x, a) HUF_ALIGN_MASK((x), (a) - 1)
79#define HUF_ALIGN_MASK(x, mask) (((x) + (mask)) & ~(mask))
80
81
82/* **************************************************************
83* BMI2 Variant Wrappers
84****************************************************************/
85typedef size_t (*HUF_DecompressUsingDTableFn)(void *dst, size_t dstSize,
86 const void *cSrc,
87 size_t cSrcSize,
88 const HUF_DTable *DTable);
89
90#if DYNAMIC_BMI2
91
92#define HUF_DGEN(fn) \
93 \
94 static size_t fn##_default( \
95 void* dst, size_t dstSize, \
96 const void* cSrc, size_t cSrcSize, \
97 const HUF_DTable* DTable) \
98 { \
99 return fn##_body(dst, dstSize, cSrc, cSrcSize, DTable); \
100 } \
101 \
102 static BMI2_TARGET_ATTRIBUTE size_t fn##_bmi2( \
103 void* dst, size_t dstSize, \
104 const void* cSrc, size_t cSrcSize, \
105 const HUF_DTable* DTable) \
106 { \
107 return fn##_body(dst, dstSize, cSrc, cSrcSize, DTable); \
108 } \
109 \
110 static size_t fn(void* dst, size_t dstSize, void const* cSrc, \
111 size_t cSrcSize, HUF_DTable const* DTable, int flags) \
112 { \
113 if (flags & HUF_flags_bmi2) { \
114 return fn##_bmi2(dst, dstSize, cSrc, cSrcSize, DTable); \
115 } \
116 return fn##_default(dst, dstSize, cSrc, cSrcSize, DTable); \
117 }
118
119#else
120
121#define HUF_DGEN(fn) \
122 static size_t fn(void* dst, size_t dstSize, void const* cSrc, \
123 size_t cSrcSize, HUF_DTable const* DTable, int flags) \
124 { \
125 (void)flags; \
126 return fn##_body(dst, dstSize, cSrc, cSrcSize, DTable); \
127 }
128
129#endif
130
131
132/*-***************************/
133/* generic DTableDesc */
134/*-***************************/
135typedef struct { BYTE maxTableLog; BYTE tableType; BYTE tableLog; BYTE reserved; } DTableDesc;
136
137static DTableDesc HUF_getDTableDesc(const HUF_DTable* table)
138{
139 DTableDesc dtd;
140 ZSTD_memcpy(&dtd, table, sizeof(dtd));
141 return dtd;
142}
143
144static size_t HUF_initFastDStream(BYTE const* ip) {
145 BYTE const lastByte = ip[7];
146 size_t const bitsConsumed = lastByte ? 8 - ZSTD_highbit32(lastByte) : 0;
147 size_t const value = MEM_readLEST(ip) | 1;
148 assert(bitsConsumed <= 8);
149 assert(sizeof(size_t) == 8);
150 return value << bitsConsumed;
151}
152
153
154/**
155 * The input/output arguments to the Huffman fast decoding loop:
156 *
157 * ip [in/out] - The input pointers, must be updated to reflect what is consumed.
158 * op [in/out] - The output pointers, must be updated to reflect what is written.
159 * bits [in/out] - The bitstream containers, must be updated to reflect the current state.
160 * dt [in] - The decoding table.
161 * ilimit [in] - The input limit, stop when any input pointer is below ilimit.
162 * oend [in] - The end of the output stream. op[3] must not cross oend.
163 * iend [in] - The end of each input stream. ip[i] may cross iend[i],
164 * as long as it is above ilimit, but that indicates corruption.
165 */
166typedef struct {
167 BYTE const* ip[4];
168 BYTE* op[4];
169 U64 bits[4];
170 void const* dt;
171 BYTE const* ilimit;
172 BYTE* oend;
173 BYTE const* iend[4];
174} HUF_DecompressFastArgs;
175
176typedef void (*HUF_DecompressFastLoopFn)(HUF_DecompressFastArgs*);
177
178/**
179 * Initializes args for the fast decoding loop.
180 * @returns 1 on success
181 * 0 if the fallback implementation should be used.
182 * Or an error code on failure.
183 */
184static size_t HUF_DecompressFastArgs_init(HUF_DecompressFastArgs* args, void* dst, size_t dstSize, void const* src, size_t srcSize, const HUF_DTable* DTable)
185{
186 void const* dt = DTable + 1;
187 U32 const dtLog = HUF_getDTableDesc(DTable).tableLog;
188
189 const BYTE* const ilimit = (const BYTE*)src + 6 + 8;
190
191 BYTE* const oend = (BYTE*)dst + dstSize;
192
193 /* The fast decoding loop assumes 64-bit little-endian.
194 * This condition is false on x32.
195 */
196 if (!MEM_isLittleEndian() || MEM_32bits())
197 return 0;
198
199 /* strict minimum : jump table + 1 byte per stream */
200 if (srcSize < 10)
201 return ERROR(corruption_detected);
202
203 /* Must have at least 8 bytes per stream because we don't handle initializing smaller bit containers.
204 * If table log is not correct at this point, fallback to the old decoder.
205 * On small inputs we don't have enough data to trigger the fast loop, so use the old decoder.
206 */
207 if (dtLog != HUF_DECODER_FAST_TABLELOG)
208 return 0;
209
210 /* Read the jump table. */
211 {
212 const BYTE* const istart = (const BYTE*)src;
213 size_t const length1 = MEM_readLE16(istart);
214 size_t const length2 = MEM_readLE16(istart+2);
215 size_t const length3 = MEM_readLE16(istart+4);
216 size_t const length4 = srcSize - (length1 + length2 + length3 + 6);
217 args->iend[0] = istart + 6; /* jumpTable */
218 args->iend[1] = args->iend[0] + length1;
219 args->iend[2] = args->iend[1] + length2;
220 args->iend[3] = args->iend[2] + length3;
221
222 /* HUF_initFastDStream() requires this, and this small of an input
223 * won't benefit from the ASM loop anyways.
224 * length1 must be >= 16 so that ip[0] >= ilimit before the loop
225 * starts.
226 */
227 if (length1 < 16 || length2 < 8 || length3 < 8 || length4 < 8)
228 return 0;
229 if (length4 > srcSize) return ERROR(corruption_detected); /* overflow */
230 }
231 /* ip[] contains the position that is currently loaded into bits[]. */
232 args->ip[0] = args->iend[1] - sizeof(U64);
233 args->ip[1] = args->iend[2] - sizeof(U64);
234 args->ip[2] = args->iend[3] - sizeof(U64);
235 args->ip[3] = (BYTE const*)src + srcSize - sizeof(U64);
236
237 /* op[] contains the output pointers. */
238 args->op[0] = (BYTE*)dst;
239 args->op[1] = args->op[0] + (dstSize+3)/4;
240 args->op[2] = args->op[1] + (dstSize+3)/4;
241 args->op[3] = args->op[2] + (dstSize+3)/4;
242
243 /* No point to call the ASM loop for tiny outputs. */
244 if (args->op[3] >= oend)
245 return 0;
246
247 /* bits[] is the bit container.
248 * It is read from the MSB down to the LSB.
249 * It is shifted left as it is read, and zeros are
250 * shifted in. After the lowest valid bit a 1 is
251 * set, so that CountTrailingZeros(bits[]) can be used
252 * to count how many bits we've consumed.
253 */
254 args->bits[0] = HUF_initFastDStream(args->ip[0]);
255 args->bits[1] = HUF_initFastDStream(args->ip[1]);
256 args->bits[2] = HUF_initFastDStream(args->ip[2]);
257 args->bits[3] = HUF_initFastDStream(args->ip[3]);
258
259 /* If ip[] >= ilimit, it is guaranteed to be safe to
260 * reload bits[]. It may be beyond its section, but is
261 * guaranteed to be valid (>= istart).
262 */
263 args->ilimit = ilimit;
264
265 args->oend = oend;
266 args->dt = dt;
267
268 return 1;
269}
270
271static size_t HUF_initRemainingDStream(BIT_DStream_t* bit, HUF_DecompressFastArgs const* args, int stream, BYTE* segmentEnd)
272{
273 /* Validate that we haven't overwritten. */
274 if (args->op[stream] > segmentEnd)
275 return ERROR(corruption_detected);
276 /* Validate that we haven't read beyond iend[].
277 * Note that ip[] may be < iend[] because the MSB is
278 * the next bit to read, and we may have consumed 100%
279 * of the stream, so down to iend[i] - 8 is valid.
280 */
281 if (args->ip[stream] < args->iend[stream] - 8)
282 return ERROR(corruption_detected);
283
284 /* Construct the BIT_DStream_t. */
285 assert(sizeof(size_t) == 8);
286 bit->bitContainer = MEM_readLEST(args->ip[stream]);
287 bit->bitsConsumed = ZSTD_countTrailingZeros64(args->bits[stream]);
288 bit->start = (const char*)args->iend[0];
289 bit->limitPtr = bit->start + sizeof(size_t);
290 bit->ptr = (const char*)args->ip[stream];
291
292 return 0;
293}
294
295
296#ifndef HUF_FORCE_DECOMPRESS_X2
297
298/*-***************************/
299/* single-symbol decoding */
300/*-***************************/
301typedef struct { BYTE nbBits; BYTE byte; } HUF_DEltX1; /* single-symbol decoding */
302
303/**
304 * Packs 4 HUF_DEltX1 structs into a U64. This is used to lay down 4 entries at
305 * a time.
306 */
307static U64 HUF_DEltX1_set4(BYTE symbol, BYTE nbBits) {
308 U64 D4;
309 if (MEM_isLittleEndian()) {
310 D4 = (U64)((symbol << 8) + nbBits);
311 } else {
312 D4 = (U64)(symbol + (nbBits << 8));
313 }
314 assert(D4 < (1U << 16));
315 D4 *= 0x0001000100010001ULL;
316 return D4;
317}
318
319/**
320 * Increase the tableLog to targetTableLog and rescales the stats.
321 * If tableLog > targetTableLog this is a no-op.
322 * @returns New tableLog
323 */
324static U32 HUF_rescaleStats(BYTE* huffWeight, U32* rankVal, U32 nbSymbols, U32 tableLog, U32 targetTableLog)
325{
326 if (tableLog > targetTableLog)
327 return tableLog;
328 if (tableLog < targetTableLog) {
329 U32 const scale = targetTableLog - tableLog;
330 U32 s;
331 /* Increase the weight for all non-zero probability symbols by scale. */
332 for (s = 0; s < nbSymbols; ++s) {
333 huffWeight[s] += (BYTE)((huffWeight[s] == 0) ? 0 : scale);
334 }
335 /* Update rankVal to reflect the new weights.
336 * All weights except 0 get moved to weight + scale.
337 * Weights [1, scale] are empty.
338 */
339 for (s = targetTableLog; s > scale; --s) {
340 rankVal[s] = rankVal[s - scale];
341 }
342 for (s = scale; s > 0; --s) {
343 rankVal[s] = 0;
344 }
345 }
346 return targetTableLog;
347}
348
349typedef struct {
350 U32 rankVal[HUF_TABLELOG_ABSOLUTEMAX + 1];
351 U32 rankStart[HUF_TABLELOG_ABSOLUTEMAX + 1];
352 U32 statsWksp[HUF_READ_STATS_WORKSPACE_SIZE_U32];
353 BYTE symbols[HUF_SYMBOLVALUE_MAX + 1];
354 BYTE huffWeight[HUF_SYMBOLVALUE_MAX + 1];
355} HUF_ReadDTableX1_Workspace;
356
357size_t HUF_readDTableX1_wksp(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize, int flags)
358{
359 U32 tableLog = 0;
360 U32 nbSymbols = 0;
361 size_t iSize;
362 void* const dtPtr = DTable + 1;
363 HUF_DEltX1* const dt = (HUF_DEltX1*)dtPtr;
364 HUF_ReadDTableX1_Workspace* wksp = (HUF_ReadDTableX1_Workspace*)workSpace;
365
366 DEBUG_STATIC_ASSERT(HUF_DECOMPRESS_WORKSPACE_SIZE >= sizeof(*wksp));
367 if (sizeof(*wksp) > wkspSize) return ERROR(tableLog_tooLarge);
368
369 DEBUG_STATIC_ASSERT(sizeof(DTableDesc) == sizeof(HUF_DTable));
370 /* ZSTD_memset(huffWeight, 0, sizeof(huffWeight)); */ /* is not necessary, even though some analyzer complain ... */
371
372 iSize = HUF_readStats_wksp(wksp->huffWeight, HUF_SYMBOLVALUE_MAX + 1, wksp->rankVal, &nbSymbols, &tableLog, src, srcSize, wksp->statsWksp, sizeof(wksp->statsWksp), flags);
373 if (HUF_isError(iSize)) return iSize;
374
375
376 /* Table header */
377 { DTableDesc dtd = HUF_getDTableDesc(DTable);
378 U32 const maxTableLog = dtd.maxTableLog + 1;
379 U32 const targetTableLog = MIN(maxTableLog, HUF_DECODER_FAST_TABLELOG);
380 tableLog = HUF_rescaleStats(wksp->huffWeight, wksp->rankVal, nbSymbols, tableLog, targetTableLog);
381 if (tableLog > (U32)(dtd.maxTableLog+1)) return ERROR(tableLog_tooLarge); /* DTable too small, Huffman tree cannot fit in */
382 dtd.tableType = 0;
383 dtd.tableLog = (BYTE)tableLog;
384 ZSTD_memcpy(DTable, &dtd, sizeof(dtd));
385 }
386
387 /* Compute symbols and rankStart given rankVal:
388 *
389 * rankVal already contains the number of values of each weight.
390 *
391 * symbols contains the symbols ordered by weight. First are the rankVal[0]
392 * weight 0 symbols, followed by the rankVal[1] weight 1 symbols, and so on.
393 * symbols[0] is filled (but unused) to avoid a branch.
394 *
395 * rankStart contains the offset where each rank belongs in the DTable.
396 * rankStart[0] is not filled because there are no entries in the table for
397 * weight 0.
398 */
399 { int n;
400 U32 nextRankStart = 0;
401 int const unroll = 4;
402 int const nLimit = (int)nbSymbols - unroll + 1;
403 for (n=0; n<(int)tableLog+1; n++) {
404 U32 const curr = nextRankStart;
405 nextRankStart += wksp->rankVal[n];
406 wksp->rankStart[n] = curr;
407 }
408 for (n=0; n < nLimit; n += unroll) {
409 int u;
410 for (u=0; u < unroll; ++u) {
411 size_t const w = wksp->huffWeight[n+u];
412 wksp->symbols[wksp->rankStart[w]++] = (BYTE)(n+u);
413 }
414 }
415 for (; n < (int)nbSymbols; ++n) {
416 size_t const w = wksp->huffWeight[n];
417 wksp->symbols[wksp->rankStart[w]++] = (BYTE)n;
418 }
419 }
420
421 /* fill DTable
422 * We fill all entries of each weight in order.
423 * That way length is a constant for each iteration of the outer loop.
424 * We can switch based on the length to a different inner loop which is
425 * optimized for that particular case.
426 */
427 { U32 w;
428 int symbol = wksp->rankVal[0];
429 int rankStart = 0;
430 for (w=1; w<tableLog+1; ++w) {
431 int const symbolCount = wksp->rankVal[w];
432 int const length = (1 << w) >> 1;
433 int uStart = rankStart;
434 BYTE const nbBits = (BYTE)(tableLog + 1 - w);
435 int s;
436 int u;
437 switch (length) {
438 case 1:
439 for (s=0; s<symbolCount; ++s) {
440 HUF_DEltX1 D;
441 D.byte = wksp->symbols[symbol + s];
442 D.nbBits = nbBits;
443 dt[uStart] = D;
444 uStart += 1;
445 }
446 break;
447 case 2:
448 for (s=0; s<symbolCount; ++s) {
449 HUF_DEltX1 D;
450 D.byte = wksp->symbols[symbol + s];
451 D.nbBits = nbBits;
452 dt[uStart+0] = D;
453 dt[uStart+1] = D;
454 uStart += 2;
455 }
456 break;
457 case 4:
458 for (s=0; s<symbolCount; ++s) {
459 U64 const D4 = HUF_DEltX1_set4(wksp->symbols[symbol + s], nbBits);
460 MEM_write64(dt + uStart, D4);
461 uStart += 4;
462 }
463 break;
464 case 8:
465 for (s=0; s<symbolCount; ++s) {
466 U64 const D4 = HUF_DEltX1_set4(wksp->symbols[symbol + s], nbBits);
467 MEM_write64(dt + uStart, D4);
468 MEM_write64(dt + uStart + 4, D4);
469 uStart += 8;
470 }
471 break;
472 default:
473 for (s=0; s<symbolCount; ++s) {
474 U64 const D4 = HUF_DEltX1_set4(wksp->symbols[symbol + s], nbBits);
475 for (u=0; u < length; u += 16) {
476 MEM_write64(dt + uStart + u + 0, D4);
477 MEM_write64(dt + uStart + u + 4, D4);
478 MEM_write64(dt + uStart + u + 8, D4);
479 MEM_write64(dt + uStart + u + 12, D4);
480 }
481 assert(u == length);
482 uStart += length;
483 }
484 break;
485 }
486 symbol += symbolCount;
487 rankStart += symbolCount * length;
488 }
489 }
490 return iSize;
491}
492
493FORCE_INLINE_TEMPLATE BYTE
494HUF_decodeSymbolX1(BIT_DStream_t* Dstream, const HUF_DEltX1* dt, const U32 dtLog)
495{
496 size_t const val = BIT_lookBitsFast(Dstream, dtLog); /* note : dtLog >= 1 */
497 BYTE const c = dt[val].byte;
498 BIT_skipBits(Dstream, dt[val].nbBits);
499 return c;
500}
501
502#define HUF_DECODE_SYMBOLX1_0(ptr, DStreamPtr) \
503 *ptr++ = HUF_decodeSymbolX1(DStreamPtr, dt, dtLog)
504
505#define HUF_DECODE_SYMBOLX1_1(ptr, DStreamPtr) \
506 if (MEM_64bits() || (HUF_TABLELOG_MAX<=12)) \
507 HUF_DECODE_SYMBOLX1_0(ptr, DStreamPtr)
508
509#define HUF_DECODE_SYMBOLX1_2(ptr, DStreamPtr) \
510 if (MEM_64bits()) \
511 HUF_DECODE_SYMBOLX1_0(ptr, DStreamPtr)
512
513HINT_INLINE size_t
514HUF_decodeStreamX1(BYTE* p, BIT_DStream_t* const bitDPtr, BYTE* const pEnd, const HUF_DEltX1* const dt, const U32 dtLog)
515{
516 BYTE* const pStart = p;
517
518 /* up to 4 symbols at a time */
519 if ((pEnd - p) > 3) {
520 while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) & (p < pEnd-3)) {
521 HUF_DECODE_SYMBOLX1_2(p, bitDPtr);
522 HUF_DECODE_SYMBOLX1_1(p, bitDPtr);
523 HUF_DECODE_SYMBOLX1_2(p, bitDPtr);
524 HUF_DECODE_SYMBOLX1_0(p, bitDPtr);
525 }
526 } else {
527 BIT_reloadDStream(bitDPtr);
528 }
529
530 /* [0-3] symbols remaining */
531 if (MEM_32bits())
532 while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) & (p < pEnd))
533 HUF_DECODE_SYMBOLX1_0(p, bitDPtr);
534
535 /* no more data to retrieve from bitstream, no need to reload */
536 while (p < pEnd)
537 HUF_DECODE_SYMBOLX1_0(p, bitDPtr);
538
539 return (size_t)(pEnd-pStart);
540}
541
542FORCE_INLINE_TEMPLATE size_t
543HUF_decompress1X1_usingDTable_internal_body(
544 void* dst, size_t dstSize,
545 const void* cSrc, size_t cSrcSize,
546 const HUF_DTable* DTable)
547{
548 BYTE* op = (BYTE*)dst;
549 BYTE* const oend = op + dstSize;
550 const void* dtPtr = DTable + 1;
551 const HUF_DEltX1* const dt = (const HUF_DEltX1*)dtPtr;
552 BIT_DStream_t bitD;
553 DTableDesc const dtd = HUF_getDTableDesc(DTable);
554 U32 const dtLog = dtd.tableLog;
555
556 CHECK_F( BIT_initDStream(&bitD, cSrc, cSrcSize) );
557
558 HUF_decodeStreamX1(op, &bitD, oend, dt, dtLog);
559
560 if (!BIT_endOfDStream(&bitD)) return ERROR(corruption_detected);
561
562 return dstSize;
563}
564
565/* HUF_decompress4X1_usingDTable_internal_body():
566 * Conditions :
567 * @dstSize >= 6
568 */
569FORCE_INLINE_TEMPLATE size_t
570HUF_decompress4X1_usingDTable_internal_body(
571 void* dst, size_t dstSize,
572 const void* cSrc, size_t cSrcSize,
573 const HUF_DTable* DTable)
574{
575 /* Check */
576 if (cSrcSize < 10) return ERROR(corruption_detected); /* strict minimum : jump table + 1 byte per stream */
577
578 { const BYTE* const istart = (const BYTE*) cSrc;
579 BYTE* const ostart = (BYTE*) dst;
580 BYTE* const oend = ostart + dstSize;
581 BYTE* const olimit = oend - 3;
582 const void* const dtPtr = DTable + 1;
583 const HUF_DEltX1* const dt = (const HUF_DEltX1*)dtPtr;
584
585 /* Init */
586 BIT_DStream_t bitD1;
587 BIT_DStream_t bitD2;
588 BIT_DStream_t bitD3;
589 BIT_DStream_t bitD4;
590 size_t const length1 = MEM_readLE16(istart);
591 size_t const length2 = MEM_readLE16(istart+2);
592 size_t const length3 = MEM_readLE16(istart+4);
593 size_t const length4 = cSrcSize - (length1 + length2 + length3 + 6);
594 const BYTE* const istart1 = istart + 6; /* jumpTable */
595 const BYTE* const istart2 = istart1 + length1;
596 const BYTE* const istart3 = istart2 + length2;
597 const BYTE* const istart4 = istart3 + length3;
598 const size_t segmentSize = (dstSize+3) / 4;
599 BYTE* const opStart2 = ostart + segmentSize;
600 BYTE* const opStart3 = opStart2 + segmentSize;
601 BYTE* const opStart4 = opStart3 + segmentSize;
602 BYTE* op1 = ostart;
603 BYTE* op2 = opStart2;
604 BYTE* op3 = opStart3;
605 BYTE* op4 = opStart4;
606 DTableDesc const dtd = HUF_getDTableDesc(DTable);
607 U32 const dtLog = dtd.tableLog;
608 U32 endSignal = 1;
609
610 if (length4 > cSrcSize) return ERROR(corruption_detected); /* overflow */
611 if (opStart4 > oend) return ERROR(corruption_detected); /* overflow */
612 if (dstSize < 6) return ERROR(corruption_detected); /* stream 4-split doesn't work */
613 CHECK_F( BIT_initDStream(&bitD1, istart1, length1) );
614 CHECK_F( BIT_initDStream(&bitD2, istart2, length2) );
615 CHECK_F( BIT_initDStream(&bitD3, istart3, length3) );
616 CHECK_F( BIT_initDStream(&bitD4, istart4, length4) );
617
618 /* up to 16 symbols per loop (4 symbols per stream) in 64-bit mode */
619 if ((size_t)(oend - op4) >= sizeof(size_t)) {
620 for ( ; (endSignal) & (op4 < olimit) ; ) {
621 HUF_DECODE_SYMBOLX1_2(op1, &bitD1);
622 HUF_DECODE_SYMBOLX1_2(op2, &bitD2);
623 HUF_DECODE_SYMBOLX1_2(op3, &bitD3);
624 HUF_DECODE_SYMBOLX1_2(op4, &bitD4);
625 HUF_DECODE_SYMBOLX1_1(op1, &bitD1);
626 HUF_DECODE_SYMBOLX1_1(op2, &bitD2);
627 HUF_DECODE_SYMBOLX1_1(op3, &bitD3);
628 HUF_DECODE_SYMBOLX1_1(op4, &bitD4);
629 HUF_DECODE_SYMBOLX1_2(op1, &bitD1);
630 HUF_DECODE_SYMBOLX1_2(op2, &bitD2);
631 HUF_DECODE_SYMBOLX1_2(op3, &bitD3);
632 HUF_DECODE_SYMBOLX1_2(op4, &bitD4);
633 HUF_DECODE_SYMBOLX1_0(op1, &bitD1);
634 HUF_DECODE_SYMBOLX1_0(op2, &bitD2);
635 HUF_DECODE_SYMBOLX1_0(op3, &bitD3);
636 HUF_DECODE_SYMBOLX1_0(op4, &bitD4);
637 endSignal &= BIT_reloadDStreamFast(&bitD1) == BIT_DStream_unfinished;
638 endSignal &= BIT_reloadDStreamFast(&bitD2) == BIT_DStream_unfinished;
639 endSignal &= BIT_reloadDStreamFast(&bitD3) == BIT_DStream_unfinished;
640 endSignal &= BIT_reloadDStreamFast(&bitD4) == BIT_DStream_unfinished;
641 }
642 }
643
644 /* check corruption */
645 /* note : should not be necessary : op# advance in lock step, and we control op4.
646 * but curiously, binary generated by gcc 7.2 & 7.3 with -mbmi2 runs faster when >=1 test is present */
647 if (op1 > opStart2) return ERROR(corruption_detected);
648 if (op2 > opStart3) return ERROR(corruption_detected);
649 if (op3 > opStart4) return ERROR(corruption_detected);
650 /* note : op4 supposed already verified within main loop */
651
652 /* finish bitStreams one by one */
653 HUF_decodeStreamX1(op1, &bitD1, opStart2, dt, dtLog);
654 HUF_decodeStreamX1(op2, &bitD2, opStart3, dt, dtLog);
655 HUF_decodeStreamX1(op3, &bitD3, opStart4, dt, dtLog);
656 HUF_decodeStreamX1(op4, &bitD4, oend, dt, dtLog);
657
658 /* check */
659 { U32 const endCheck = BIT_endOfDStream(&bitD1) & BIT_endOfDStream(&bitD2) & BIT_endOfDStream(&bitD3) & BIT_endOfDStream(&bitD4);
660 if (!endCheck) return ERROR(corruption_detected); }
661
662 /* decoded size */
663 return dstSize;
664 }
665}
666
667#if HUF_NEED_BMI2_FUNCTION
668static BMI2_TARGET_ATTRIBUTE
669size_t HUF_decompress4X1_usingDTable_internal_bmi2(void* dst, size_t dstSize, void const* cSrc,
670 size_t cSrcSize, HUF_DTable const* DTable) {
671 return HUF_decompress4X1_usingDTable_internal_body(dst, dstSize, cSrc, cSrcSize, DTable);
672}
673#endif
674
675static
676size_t HUF_decompress4X1_usingDTable_internal_default(void* dst, size_t dstSize, void const* cSrc,
677 size_t cSrcSize, HUF_DTable const* DTable) {
678 return HUF_decompress4X1_usingDTable_internal_body(dst, dstSize, cSrc, cSrcSize, DTable);
679}
680
681#if ZSTD_ENABLE_ASM_X86_64_BMI2
682
683HUF_ASM_DECL void HUF_decompress4X1_usingDTable_internal_fast_asm_loop(HUF_DecompressFastArgs* args) ZSTDLIB_HIDDEN;
684
685#endif
686
687static HUF_FAST_BMI2_ATTRS
688void HUF_decompress4X1_usingDTable_internal_fast_c_loop(HUF_DecompressFastArgs* args)
689{
690 U64 bits[4];
691 BYTE const* ip[4];
692 BYTE* op[4];
693 U16 const* const dtable = (U16 const*)args->dt;
694 BYTE* const oend = args->oend;
695 BYTE const* const ilimit = args->ilimit;
696
697 /* Copy the arguments to local variables */
698 ZSTD_memcpy(&bits, &args->bits, sizeof(bits));
699 ZSTD_memcpy((void*)(&ip), &args->ip, sizeof(ip));
700 ZSTD_memcpy(&op, &args->op, sizeof(op));
701
702 assert(MEM_isLittleEndian());
703 assert(!MEM_32bits());
704
705 for (;;) {
706 BYTE* olimit;
707 int stream;
708 int symbol;
709
710 /* Assert loop preconditions */
711#ifndef NDEBUG
712 for (stream = 0; stream < 4; ++stream) {
713 assert(op[stream] <= (stream == 3 ? oend : op[stream + 1]));
714 assert(ip[stream] >= ilimit);
715 }
716#endif
717 /* Compute olimit */
718 {
719 /* Each iteration produces 5 output symbols per stream */
720 size_t const oiters = (size_t)(oend - op[3]) / 5;
721 /* Each iteration consumes up to 11 bits * 5 = 55 bits < 7 bytes
722 * per stream.
723 */
724 size_t const iiters = (size_t)(ip[0] - ilimit) / 7;
725 /* We can safely run iters iterations before running bounds checks */
726 size_t const iters = MIN(oiters, iiters);
727 size_t const symbols = iters * 5;
728
729 /* We can simply check that op[3] < olimit, instead of checking all
730 * of our bounds, since we can't hit the other bounds until we've run
731 * iters iterations, which only happens when op[3] == olimit.
732 */
733 olimit = op[3] + symbols;
734
735 /* Exit fast decoding loop once we get close to the end. */
736 if (op[3] + 20 > olimit)
737 break;
738
739 /* Exit the decoding loop if any input pointer has crossed the
740 * previous one. This indicates corruption, and a precondition
741 * to our loop is that ip[i] >= ip[0].
742 */
743 for (stream = 1; stream < 4; ++stream) {
744 if (ip[stream] < ip[stream - 1])
745 goto _out;
746 }
747 }
748
749#ifndef NDEBUG
750 for (stream = 1; stream < 4; ++stream) {
751 assert(ip[stream] >= ip[stream - 1]);
752 }
753#endif
754
755 do {
756 /* Decode 5 symbols in each of the 4 streams */
757 for (symbol = 0; symbol < 5; ++symbol) {
758 for (stream = 0; stream < 4; ++stream) {
759 int const index = (int)(bits[stream] >> 53);
760 int const entry = (int)dtable[index];
761 bits[stream] <<= (entry & 63);
762 op[stream][symbol] = (BYTE)((entry >> 8) & 0xFF);
763 }
764 }
765 /* Reload the bitstreams */
766 for (stream = 0; stream < 4; ++stream) {
767 int const ctz = ZSTD_countTrailingZeros64(bits[stream]);
768 int const nbBits = ctz & 7;
769 int const nbBytes = ctz >> 3;
770 op[stream] += 5;
771 ip[stream] -= nbBytes;
772 bits[stream] = MEM_read64(ip[stream]) | 1;
773 bits[stream] <<= nbBits;
774 }
775 } while (op[3] < olimit);
776 }
777
778_out:
779
780 /* Save the final values of each of the state variables back to args. */
781 ZSTD_memcpy(&args->bits, &bits, sizeof(bits));
782 ZSTD_memcpy((void*)(&args->ip), &ip, sizeof(ip));
783 ZSTD_memcpy(&args->op, &op, sizeof(op));
784}
785
786/**
787 * @returns @p dstSize on success (>= 6)
788 * 0 if the fallback implementation should be used
789 * An error if an error occurred
790 */
791static HUF_FAST_BMI2_ATTRS
792size_t
793HUF_decompress4X1_usingDTable_internal_fast(
794 void* dst, size_t dstSize,
795 const void* cSrc, size_t cSrcSize,
796 const HUF_DTable* DTable,
797 HUF_DecompressFastLoopFn loopFn)
798{
799 void const* dt = DTable + 1;
800 const BYTE* const iend = (const BYTE*)cSrc + 6;
801 BYTE* const oend = (BYTE*)dst + dstSize;
802 HUF_DecompressFastArgs args;
803 { size_t const ret = HUF_DecompressFastArgs_init(&args, dst, dstSize, cSrc, cSrcSize, DTable);
804 FORWARD_IF_ERROR(ret, "Failed to init fast loop args");
805 if (ret == 0)
806 return 0;
807 }
808
809 assert(args.ip[0] >= args.ilimit);
810 loopFn(&args);
811
812 /* Our loop guarantees that ip[] >= ilimit and that we haven't
813 * overwritten any op[].
814 */
815 assert(args.ip[0] >= iend);
816 assert(args.ip[1] >= iend);
817 assert(args.ip[2] >= iend);
818 assert(args.ip[3] >= iend);
819 assert(args.op[3] <= oend);
820 (void)iend;
821
822 /* finish bit streams one by one. */
823 { size_t const segmentSize = (dstSize+3) / 4;
824 BYTE* segmentEnd = (BYTE*)dst;
825 int i;
826 for (i = 0; i < 4; ++i) {
827 BIT_DStream_t bit;
828 if (segmentSize <= (size_t)(oend - segmentEnd))
829 segmentEnd += segmentSize;
830 else
831 segmentEnd = oend;
832 FORWARD_IF_ERROR(HUF_initRemainingDStream(&bit, &args, i, segmentEnd), "corruption");
833 /* Decompress and validate that we've produced exactly the expected length. */
834 args.op[i] += HUF_decodeStreamX1(args.op[i], &bit, segmentEnd, (HUF_DEltX1 const*)dt, HUF_DECODER_FAST_TABLELOG);
835 if (args.op[i] != segmentEnd) return ERROR(corruption_detected);
836 }
837 }
838
839 /* decoded size */
840 assert(dstSize != 0);
841 return dstSize;
842}
843
844HUF_DGEN(HUF_decompress1X1_usingDTable_internal)
845
846static size_t HUF_decompress4X1_usingDTable_internal(void* dst, size_t dstSize, void const* cSrc,
847 size_t cSrcSize, HUF_DTable const* DTable, int flags)
848{
849 HUF_DecompressUsingDTableFn fallbackFn = HUF_decompress4X1_usingDTable_internal_default;
850 HUF_DecompressFastLoopFn loopFn = HUF_decompress4X1_usingDTable_internal_fast_c_loop;
851
852#if DYNAMIC_BMI2
853 if (flags & HUF_flags_bmi2) {
854 fallbackFn = HUF_decompress4X1_usingDTable_internal_bmi2;
855# if ZSTD_ENABLE_ASM_X86_64_BMI2
856 if (!(flags & HUF_flags_disableAsm)) {
857 loopFn = HUF_decompress4X1_usingDTable_internal_fast_asm_loop;
858 }
859# endif
860 } else {
861 return fallbackFn(dst, dstSize, cSrc, cSrcSize, DTable);
862 }
863#endif
864
865#if ZSTD_ENABLE_ASM_X86_64_BMI2 && defined(__BMI2__)
866 if (!(flags & HUF_flags_disableAsm)) {
867 loopFn = HUF_decompress4X1_usingDTable_internal_fast_asm_loop;
868 }
869#endif
870
871 if (!(flags & HUF_flags_disableFast)) {
872 size_t const ret = HUF_decompress4X1_usingDTable_internal_fast(dst, dstSize, cSrc, cSrcSize, DTable, loopFn);
873 if (ret != 0)
874 return ret;
875 }
876 return fallbackFn(dst, dstSize, cSrc, cSrcSize, DTable);
877}
878
879static size_t HUF_decompress4X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
880 const void* cSrc, size_t cSrcSize,
881 void* workSpace, size_t wkspSize, int flags)
882{
883 const BYTE* ip = (const BYTE*) cSrc;
884
885 size_t const hSize = HUF_readDTableX1_wksp(dctx, cSrc, cSrcSize, workSpace, wkspSize, flags);
886 if (HUF_isError(hSize)) return hSize;
887 if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
888 ip += hSize; cSrcSize -= hSize;
889
890 return HUF_decompress4X1_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx, flags);
891}
892
893#endif /* HUF_FORCE_DECOMPRESS_X2 */
894
895
896#ifndef HUF_FORCE_DECOMPRESS_X1
897
898/* *************************/
899/* double-symbols decoding */
900/* *************************/
901
902typedef struct { U16 sequence; BYTE nbBits; BYTE length; } HUF_DEltX2; /* double-symbols decoding */
903typedef struct { BYTE symbol; } sortedSymbol_t;
904typedef U32 rankValCol_t[HUF_TABLELOG_MAX + 1];
905typedef rankValCol_t rankVal_t[HUF_TABLELOG_MAX];
906
907/**
908 * Constructs a HUF_DEltX2 in a U32.
909 */
910static U32 HUF_buildDEltX2U32(U32 symbol, U32 nbBits, U32 baseSeq, int level)
911{
912 U32 seq;
913 DEBUG_STATIC_ASSERT(offsetof(HUF_DEltX2, sequence) == 0);
914 DEBUG_STATIC_ASSERT(offsetof(HUF_DEltX2, nbBits) == 2);
915 DEBUG_STATIC_ASSERT(offsetof(HUF_DEltX2, length) == 3);
916 DEBUG_STATIC_ASSERT(sizeof(HUF_DEltX2) == sizeof(U32));
917 if (MEM_isLittleEndian()) {
918 seq = level == 1 ? symbol : (baseSeq + (symbol << 8));
919 return seq + (nbBits << 16) + ((U32)level << 24);
920 } else {
921 seq = level == 1 ? (symbol << 8) : ((baseSeq << 8) + symbol);
922 return (seq << 16) + (nbBits << 8) + (U32)level;
923 }
924}
925
926/**
927 * Constructs a HUF_DEltX2.
928 */
929static HUF_DEltX2 HUF_buildDEltX2(U32 symbol, U32 nbBits, U32 baseSeq, int level)
930{
931 HUF_DEltX2 DElt;
932 U32 const val = HUF_buildDEltX2U32(symbol, nbBits, baseSeq, level);
933 DEBUG_STATIC_ASSERT(sizeof(DElt) == sizeof(val));
934 ZSTD_memcpy(&DElt, &val, sizeof(val));
935 return DElt;
936}
937
938/**
939 * Constructs 2 HUF_DEltX2s and packs them into a U64.
940 */
941static U64 HUF_buildDEltX2U64(U32 symbol, U32 nbBits, U16 baseSeq, int level)
942{
943 U32 DElt = HUF_buildDEltX2U32(symbol, nbBits, baseSeq, level);
944 return (U64)DElt + ((U64)DElt << 32);
945}
946
947/**
948 * Fills the DTable rank with all the symbols from [begin, end) that are each
949 * nbBits long.
950 *
951 * @param DTableRank The start of the rank in the DTable.
952 * @param begin The first symbol to fill (inclusive).
953 * @param end The last symbol to fill (exclusive).
954 * @param nbBits Each symbol is nbBits long.
955 * @param tableLog The table log.
956 * @param baseSeq If level == 1 { 0 } else { the first level symbol }
957 * @param level The level in the table. Must be 1 or 2.
958 */
959static void HUF_fillDTableX2ForWeight(
960 HUF_DEltX2* DTableRank,
961 sortedSymbol_t const* begin, sortedSymbol_t const* end,
962 U32 nbBits, U32 tableLog,
963 U16 baseSeq, int const level)
964{
965 U32 const length = 1U << ((tableLog - nbBits) & 0x1F /* quiet static-analyzer */);
966 const sortedSymbol_t* ptr;
967 assert(level >= 1 && level <= 2);
968 switch (length) {
969 case 1:
970 for (ptr = begin; ptr != end; ++ptr) {
971 HUF_DEltX2 const DElt = HUF_buildDEltX2(ptr->symbol, nbBits, baseSeq, level);
972 *DTableRank++ = DElt;
973 }
974 break;
975 case 2:
976 for (ptr = begin; ptr != end; ++ptr) {
977 HUF_DEltX2 const DElt = HUF_buildDEltX2(ptr->symbol, nbBits, baseSeq, level);
978 DTableRank[0] = DElt;
979 DTableRank[1] = DElt;
980 DTableRank += 2;
981 }
982 break;
983 case 4:
984 for (ptr = begin; ptr != end; ++ptr) {
985 U64 const DEltX2 = HUF_buildDEltX2U64(ptr->symbol, nbBits, baseSeq, level);
986 ZSTD_memcpy(DTableRank + 0, &DEltX2, sizeof(DEltX2));
987 ZSTD_memcpy(DTableRank + 2, &DEltX2, sizeof(DEltX2));
988 DTableRank += 4;
989 }
990 break;
991 case 8:
992 for (ptr = begin; ptr != end; ++ptr) {
993 U64 const DEltX2 = HUF_buildDEltX2U64(ptr->symbol, nbBits, baseSeq, level);
994 ZSTD_memcpy(DTableRank + 0, &DEltX2, sizeof(DEltX2));
995 ZSTD_memcpy(DTableRank + 2, &DEltX2, sizeof(DEltX2));
996 ZSTD_memcpy(DTableRank + 4, &DEltX2, sizeof(DEltX2));
997 ZSTD_memcpy(DTableRank + 6, &DEltX2, sizeof(DEltX2));
998 DTableRank += 8;
999 }
1000 break;
1001 default:
1002 for (ptr = begin; ptr != end; ++ptr) {
1003 U64 const DEltX2 = HUF_buildDEltX2U64(ptr->symbol, nbBits, baseSeq, level);
1004 HUF_DEltX2* const DTableRankEnd = DTableRank + length;
1005 for (; DTableRank != DTableRankEnd; DTableRank += 8) {
1006 ZSTD_memcpy(DTableRank + 0, &DEltX2, sizeof(DEltX2));
1007 ZSTD_memcpy(DTableRank + 2, &DEltX2, sizeof(DEltX2));
1008 ZSTD_memcpy(DTableRank + 4, &DEltX2, sizeof(DEltX2));
1009 ZSTD_memcpy(DTableRank + 6, &DEltX2, sizeof(DEltX2));
1010 }
1011 }
1012 break;
1013 }
1014}
1015
1016/* HUF_fillDTableX2Level2() :
1017 * `rankValOrigin` must be a table of at least (HUF_TABLELOG_MAX + 1) U32 */
1018static void HUF_fillDTableX2Level2(HUF_DEltX2* DTable, U32 targetLog, const U32 consumedBits,
1019 const U32* rankVal, const int minWeight, const int maxWeight1,
1020 const sortedSymbol_t* sortedSymbols, U32 const* rankStart,
1021 U32 nbBitsBaseline, U16 baseSeq)
1022{
1023 /* Fill skipped values (all positions up to rankVal[minWeight]).
1024 * These are positions only get a single symbol because the combined weight
1025 * is too large.
1026 */
1027 if (minWeight>1) {
1028 U32 const length = 1U << ((targetLog - consumedBits) & 0x1F /* quiet static-analyzer */);
1029 U64 const DEltX2 = HUF_buildDEltX2U64(baseSeq, consumedBits, /* baseSeq */ 0, /* level */ 1);
1030 int const skipSize = rankVal[minWeight];
1031 assert(length > 1);
1032 assert((U32)skipSize < length);
1033 switch (length) {
1034 case 2:
1035 assert(skipSize == 1);
1036 ZSTD_memcpy(DTable, &DEltX2, sizeof(DEltX2));
1037 break;
1038 case 4:
1039 assert(skipSize <= 4);
1040 ZSTD_memcpy(DTable + 0, &DEltX2, sizeof(DEltX2));
1041 ZSTD_memcpy(DTable + 2, &DEltX2, sizeof(DEltX2));
1042 break;
1043 default:
1044 {
1045 int i;
1046 for (i = 0; i < skipSize; i += 8) {
1047 ZSTD_memcpy(DTable + i + 0, &DEltX2, sizeof(DEltX2));
1048 ZSTD_memcpy(DTable + i + 2, &DEltX2, sizeof(DEltX2));
1049 ZSTD_memcpy(DTable + i + 4, &DEltX2, sizeof(DEltX2));
1050 ZSTD_memcpy(DTable + i + 6, &DEltX2, sizeof(DEltX2));
1051 }
1052 }
1053 }
1054 }
1055
1056 /* Fill each of the second level symbols by weight. */
1057 {
1058 int w;
1059 for (w = minWeight; w < maxWeight1; ++w) {
1060 int const begin = rankStart[w];
1061 int const end = rankStart[w+1];
1062 U32 const nbBits = nbBitsBaseline - w;
1063 U32 const totalBits = nbBits + consumedBits;
1064 HUF_fillDTableX2ForWeight(
1065 DTable + rankVal[w],
1066 sortedSymbols + begin, sortedSymbols + end,
1067 totalBits, targetLog,
1068 baseSeq, /* level */ 2);
1069 }
1070 }
1071}
1072
1073static void HUF_fillDTableX2(HUF_DEltX2* DTable, const U32 targetLog,
1074 const sortedSymbol_t* sortedList,
1075 const U32* rankStart, rankValCol_t* rankValOrigin, const U32 maxWeight,
1076 const U32 nbBitsBaseline)
1077{
1078 U32* const rankVal = rankValOrigin[0];
1079 const int scaleLog = nbBitsBaseline - targetLog; /* note : targetLog >= srcLog, hence scaleLog <= 1 */
1080 const U32 minBits = nbBitsBaseline - maxWeight;
1081 int w;
1082 int const wEnd = (int)maxWeight + 1;
1083
1084 /* Fill DTable in order of weight. */
1085 for (w = 1; w < wEnd; ++w) {
1086 int const begin = (int)rankStart[w];
1087 int const end = (int)rankStart[w+1];
1088 U32 const nbBits = nbBitsBaseline - w;
1089
1090 if (targetLog-nbBits >= minBits) {
1091 /* Enough room for a second symbol. */
1092 int start = rankVal[w];
1093 U32 const length = 1U << ((targetLog - nbBits) & 0x1F /* quiet static-analyzer */);
1094 int minWeight = nbBits + scaleLog;
1095 int s;
1096 if (minWeight < 1) minWeight = 1;
1097 /* Fill the DTable for every symbol of weight w.
1098 * These symbols get at least 1 second symbol.
1099 */
1100 for (s = begin; s != end; ++s) {
1101 HUF_fillDTableX2Level2(
1102 DTable + start, targetLog, nbBits,
1103 rankValOrigin[nbBits], minWeight, wEnd,
1104 sortedList, rankStart,
1105 nbBitsBaseline, sortedList[s].symbol);
1106 start += length;
1107 }
1108 } else {
1109 /* Only a single symbol. */
1110 HUF_fillDTableX2ForWeight(
1111 DTable + rankVal[w],
1112 sortedList + begin, sortedList + end,
1113 nbBits, targetLog,
1114 /* baseSeq */ 0, /* level */ 1);
1115 }
1116 }
1117}
1118
1119typedef struct {
1120 rankValCol_t rankVal[HUF_TABLELOG_MAX];
1121 U32 rankStats[HUF_TABLELOG_MAX + 1];
1122 U32 rankStart0[HUF_TABLELOG_MAX + 3];
1123 sortedSymbol_t sortedSymbol[HUF_SYMBOLVALUE_MAX + 1];
1124 BYTE weightList[HUF_SYMBOLVALUE_MAX + 1];
1125 U32 calleeWksp[HUF_READ_STATS_WORKSPACE_SIZE_U32];
1126} HUF_ReadDTableX2_Workspace;
1127
1128size_t HUF_readDTableX2_wksp(HUF_DTable* DTable,
1129 const void* src, size_t srcSize,
1130 void* workSpace, size_t wkspSize, int flags)
1131{
1132 U32 tableLog, maxW, nbSymbols;
1133 DTableDesc dtd = HUF_getDTableDesc(DTable);
1134 U32 maxTableLog = dtd.maxTableLog;
1135 size_t iSize;
1136 void* dtPtr = DTable+1; /* force compiler to avoid strict-aliasing */
1137 HUF_DEltX2* const dt = (HUF_DEltX2*)dtPtr;
1138 U32 *rankStart;
1139
1140 HUF_ReadDTableX2_Workspace* const wksp = (HUF_ReadDTableX2_Workspace*)workSpace;
1141
1142 if (sizeof(*wksp) > wkspSize) return ERROR(GENERIC);
1143
1144 rankStart = wksp->rankStart0 + 1;
1145 ZSTD_memset(wksp->rankStats, 0, sizeof(wksp->rankStats));
1146 ZSTD_memset(wksp->rankStart0, 0, sizeof(wksp->rankStart0));
1147
1148 DEBUG_STATIC_ASSERT(sizeof(HUF_DEltX2) == sizeof(HUF_DTable)); /* if compiler fails here, assertion is wrong */
1149 if (maxTableLog > HUF_TABLELOG_MAX) return ERROR(tableLog_tooLarge);
1150 /* ZSTD_memset(weightList, 0, sizeof(weightList)); */ /* is not necessary, even though some analyzer complain ... */
1151
1152 iSize = HUF_readStats_wksp(wksp->weightList, HUF_SYMBOLVALUE_MAX + 1, wksp->rankStats, &nbSymbols, &tableLog, src, srcSize, wksp->calleeWksp, sizeof(wksp->calleeWksp), flags);
1153 if (HUF_isError(iSize)) return iSize;
1154
1155 /* check result */
1156 if (tableLog > maxTableLog) return ERROR(tableLog_tooLarge); /* DTable can't fit code depth */
1157 if (tableLog <= HUF_DECODER_FAST_TABLELOG && maxTableLog > HUF_DECODER_FAST_TABLELOG) maxTableLog = HUF_DECODER_FAST_TABLELOG;
1158
1159 /* find maxWeight */
1160 for (maxW = tableLog; wksp->rankStats[maxW]==0; maxW--) {} /* necessarily finds a solution before 0 */
1161
1162 /* Get start index of each weight */
1163 { U32 w, nextRankStart = 0;
1164 for (w=1; w<maxW+1; w++) {
1165 U32 curr = nextRankStart;
1166 nextRankStart += wksp->rankStats[w];
1167 rankStart[w] = curr;
1168 }
1169 rankStart[0] = nextRankStart; /* put all 0w symbols at the end of sorted list*/
1170 rankStart[maxW+1] = nextRankStart;
1171 }
1172
1173 /* sort symbols by weight */
1174 { U32 s;
1175 for (s=0; s<nbSymbols; s++) {
1176 U32 const w = wksp->weightList[s];
1177 U32 const r = rankStart[w]++;
1178 wksp->sortedSymbol[r].symbol = (BYTE)s;
1179 }
1180 rankStart[0] = 0; /* forget 0w symbols; this is beginning of weight(1) */
1181 }
1182
1183 /* Build rankVal */
1184 { U32* const rankVal0 = wksp->rankVal[0];
1185 { int const rescale = (maxTableLog-tableLog) - 1; /* tableLog <= maxTableLog */
1186 U32 nextRankVal = 0;
1187 U32 w;
1188 for (w=1; w<maxW+1; w++) {
1189 U32 curr = nextRankVal;
1190 nextRankVal += wksp->rankStats[w] << (w+rescale);
1191 rankVal0[w] = curr;
1192 } }
1193 { U32 const minBits = tableLog+1 - maxW;
1194 U32 consumed;
1195 for (consumed = minBits; consumed < maxTableLog - minBits + 1; consumed++) {
1196 U32* const rankValPtr = wksp->rankVal[consumed];
1197 U32 w;
1198 for (w = 1; w < maxW+1; w++) {
1199 rankValPtr[w] = rankVal0[w] >> consumed;
1200 } } } }
1201
1202 HUF_fillDTableX2(dt, maxTableLog,
1203 wksp->sortedSymbol,
1204 wksp->rankStart0, wksp->rankVal, maxW,
1205 tableLog+1);
1206
1207 dtd.tableLog = (BYTE)maxTableLog;
1208 dtd.tableType = 1;
1209 ZSTD_memcpy(DTable, &dtd, sizeof(dtd));
1210 return iSize;
1211}
1212
1213
1214FORCE_INLINE_TEMPLATE U32
1215HUF_decodeSymbolX2(void* op, BIT_DStream_t* DStream, const HUF_DEltX2* dt, const U32 dtLog)
1216{
1217 size_t const val = BIT_lookBitsFast(DStream, dtLog); /* note : dtLog >= 1 */
1218 ZSTD_memcpy(op, &dt[val].sequence, 2);
1219 BIT_skipBits(DStream, dt[val].nbBits);
1220 return dt[val].length;
1221}
1222
1223FORCE_INLINE_TEMPLATE U32
1224HUF_decodeLastSymbolX2(void* op, BIT_DStream_t* DStream, const HUF_DEltX2* dt, const U32 dtLog)
1225{
1226 size_t const val = BIT_lookBitsFast(DStream, dtLog); /* note : dtLog >= 1 */
1227 ZSTD_memcpy(op, &dt[val].sequence, 1);
1228 if (dt[val].length==1) {
1229 BIT_skipBits(DStream, dt[val].nbBits);
1230 } else {
1231 if (DStream->bitsConsumed < (sizeof(DStream->bitContainer)*8)) {
1232 BIT_skipBits(DStream, dt[val].nbBits);
1233 if (DStream->bitsConsumed > (sizeof(DStream->bitContainer)*8))
1234 /* ugly hack; works only because it's the last symbol. Note : can't easily extract nbBits from just this symbol */
1235 DStream->bitsConsumed = (sizeof(DStream->bitContainer)*8);
1236 }
1237 }
1238 return 1;
1239}
1240
1241#define HUF_DECODE_SYMBOLX2_0(ptr, DStreamPtr) \
1242 ptr += HUF_decodeSymbolX2(ptr, DStreamPtr, dt, dtLog)
1243
1244#define HUF_DECODE_SYMBOLX2_1(ptr, DStreamPtr) \
1245 if (MEM_64bits() || (HUF_TABLELOG_MAX<=12)) \
1246 ptr += HUF_decodeSymbolX2(ptr, DStreamPtr, dt, dtLog)
1247
1248#define HUF_DECODE_SYMBOLX2_2(ptr, DStreamPtr) \
1249 if (MEM_64bits()) \
1250 ptr += HUF_decodeSymbolX2(ptr, DStreamPtr, dt, dtLog)
1251
1252HINT_INLINE size_t
1253HUF_decodeStreamX2(BYTE* p, BIT_DStream_t* bitDPtr, BYTE* const pEnd,
1254 const HUF_DEltX2* const dt, const U32 dtLog)
1255{
1256 BYTE* const pStart = p;
1257
1258 /* up to 8 symbols at a time */
1259 if ((size_t)(pEnd - p) >= sizeof(bitDPtr->bitContainer)) {
1260 if (dtLog <= 11 && MEM_64bits()) {
1261 /* up to 10 symbols at a time */
1262 while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) & (p < pEnd-9)) {
1263 HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
1264 HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
1265 HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
1266 HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
1267 HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
1268 }
1269 } else {
1270 /* up to 8 symbols at a time */
1271 while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) & (p < pEnd-(sizeof(bitDPtr->bitContainer)-1))) {
1272 HUF_DECODE_SYMBOLX2_2(p, bitDPtr);
1273 HUF_DECODE_SYMBOLX2_1(p, bitDPtr);
1274 HUF_DECODE_SYMBOLX2_2(p, bitDPtr);
1275 HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
1276 }
1277 }
1278 } else {
1279 BIT_reloadDStream(bitDPtr);
1280 }
1281
1282 /* closer to end : up to 2 symbols at a time */
1283 if ((size_t)(pEnd - p) >= 2) {
1284 while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) & (p <= pEnd-2))
1285 HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
1286
1287 while (p <= pEnd-2)
1288 HUF_DECODE_SYMBOLX2_0(p, bitDPtr); /* no need to reload : reached the end of DStream */
1289 }
1290
1291 if (p < pEnd)
1292 p += HUF_decodeLastSymbolX2(p, bitDPtr, dt, dtLog);
1293
1294 return p-pStart;
1295}
1296
1297FORCE_INLINE_TEMPLATE size_t
1298HUF_decompress1X2_usingDTable_internal_body(
1299 void* dst, size_t dstSize,
1300 const void* cSrc, size_t cSrcSize,
1301 const HUF_DTable* DTable)
1302{
1303 BIT_DStream_t bitD;
1304
1305 /* Init */
1306 CHECK_F( BIT_initDStream(&bitD, cSrc, cSrcSize) );
1307
1308 /* decode */
1309 { BYTE* const ostart = (BYTE*) dst;
1310 BYTE* const oend = ostart + dstSize;
1311 const void* const dtPtr = DTable+1; /* force compiler to not use strict-aliasing */
1312 const HUF_DEltX2* const dt = (const HUF_DEltX2*)dtPtr;
1313 DTableDesc const dtd = HUF_getDTableDesc(DTable);
1314 HUF_decodeStreamX2(ostart, &bitD, oend, dt, dtd.tableLog);
1315 }
1316
1317 /* check */
1318 if (!BIT_endOfDStream(&bitD)) return ERROR(corruption_detected);
1319
1320 /* decoded size */
1321 return dstSize;
1322}
1323
1324/* HUF_decompress4X2_usingDTable_internal_body():
1325 * Conditions:
1326 * @dstSize >= 6
1327 */
1328FORCE_INLINE_TEMPLATE size_t
1329HUF_decompress4X2_usingDTable_internal_body(
1330 void* dst, size_t dstSize,
1331 const void* cSrc, size_t cSrcSize,
1332 const HUF_DTable* DTable)
1333{
1334 if (cSrcSize < 10) return ERROR(corruption_detected); /* strict minimum : jump table + 1 byte per stream */
1335
1336 { const BYTE* const istart = (const BYTE*) cSrc;
1337 BYTE* const ostart = (BYTE*) dst;
1338 BYTE* const oend = ostart + dstSize;
1339 BYTE* const olimit = oend - (sizeof(size_t)-1);
1340 const void* const dtPtr = DTable+1;
1341 const HUF_DEltX2* const dt = (const HUF_DEltX2*)dtPtr;
1342
1343 /* Init */
1344 BIT_DStream_t bitD1;
1345 BIT_DStream_t bitD2;
1346 BIT_DStream_t bitD3;
1347 BIT_DStream_t bitD4;
1348 size_t const length1 = MEM_readLE16(istart);
1349 size_t const length2 = MEM_readLE16(istart+2);
1350 size_t const length3 = MEM_readLE16(istart+4);
1351 size_t const length4 = cSrcSize - (length1 + length2 + length3 + 6);
1352 const BYTE* const istart1 = istart + 6; /* jumpTable */
1353 const BYTE* const istart2 = istart1 + length1;
1354 const BYTE* const istart3 = istart2 + length2;
1355 const BYTE* const istart4 = istart3 + length3;
1356 size_t const segmentSize = (dstSize+3) / 4;
1357 BYTE* const opStart2 = ostart + segmentSize;
1358 BYTE* const opStart3 = opStart2 + segmentSize;
1359 BYTE* const opStart4 = opStart3 + segmentSize;
1360 BYTE* op1 = ostart;
1361 BYTE* op2 = opStart2;
1362 BYTE* op3 = opStart3;
1363 BYTE* op4 = opStart4;
1364 U32 endSignal = 1;
1365 DTableDesc const dtd = HUF_getDTableDesc(DTable);
1366 U32 const dtLog = dtd.tableLog;
1367
1368 if (length4 > cSrcSize) return ERROR(corruption_detected); /* overflow */
1369 if (opStart4 > oend) return ERROR(corruption_detected); /* overflow */
1370 if (dstSize < 6) return ERROR(corruption_detected); /* stream 4-split doesn't work */
1371 CHECK_F( BIT_initDStream(&bitD1, istart1, length1) );
1372 CHECK_F( BIT_initDStream(&bitD2, istart2, length2) );
1373 CHECK_F( BIT_initDStream(&bitD3, istart3, length3) );
1374 CHECK_F( BIT_initDStream(&bitD4, istart4, length4) );
1375
1376 /* 16-32 symbols per loop (4-8 symbols per stream) */
1377 if ((size_t)(oend - op4) >= sizeof(size_t)) {
1378 for ( ; (endSignal) & (op4 < olimit); ) {
1379#if defined(__clang__) && (defined(__x86_64__) || defined(__i386__))
1380 HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
1381 HUF_DECODE_SYMBOLX2_1(op1, &bitD1);
1382 HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
1383 HUF_DECODE_SYMBOLX2_0(op1, &bitD1);
1384 HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
1385 HUF_DECODE_SYMBOLX2_1(op2, &bitD2);
1386 HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
1387 HUF_DECODE_SYMBOLX2_0(op2, &bitD2);
1388 endSignal &= BIT_reloadDStreamFast(&bitD1) == BIT_DStream_unfinished;
1389 endSignal &= BIT_reloadDStreamFast(&bitD2) == BIT_DStream_unfinished;
1390 HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
1391 HUF_DECODE_SYMBOLX2_1(op3, &bitD3);
1392 HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
1393 HUF_DECODE_SYMBOLX2_0(op3, &bitD3);
1394 HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
1395 HUF_DECODE_SYMBOLX2_1(op4, &bitD4);
1396 HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
1397 HUF_DECODE_SYMBOLX2_0(op4, &bitD4);
1398 endSignal &= BIT_reloadDStreamFast(&bitD3) == BIT_DStream_unfinished;
1399 endSignal &= BIT_reloadDStreamFast(&bitD4) == BIT_DStream_unfinished;
1400#else
1401 HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
1402 HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
1403 HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
1404 HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
1405 HUF_DECODE_SYMBOLX2_1(op1, &bitD1);
1406 HUF_DECODE_SYMBOLX2_1(op2, &bitD2);
1407 HUF_DECODE_SYMBOLX2_1(op3, &bitD3);
1408 HUF_DECODE_SYMBOLX2_1(op4, &bitD4);
1409 HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
1410 HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
1411 HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
1412 HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
1413 HUF_DECODE_SYMBOLX2_0(op1, &bitD1);
1414 HUF_DECODE_SYMBOLX2_0(op2, &bitD2);
1415 HUF_DECODE_SYMBOLX2_0(op3, &bitD3);
1416 HUF_DECODE_SYMBOLX2_0(op4, &bitD4);
1417 endSignal = (U32)LIKELY((U32)
1418 (BIT_reloadDStreamFast(&bitD1) == BIT_DStream_unfinished)
1419 & (BIT_reloadDStreamFast(&bitD2) == BIT_DStream_unfinished)
1420 & (BIT_reloadDStreamFast(&bitD3) == BIT_DStream_unfinished)
1421 & (BIT_reloadDStreamFast(&bitD4) == BIT_DStream_unfinished));
1422#endif
1423 }
1424 }
1425
1426 /* check corruption */
1427 if (op1 > opStart2) return ERROR(corruption_detected);
1428 if (op2 > opStart3) return ERROR(corruption_detected);
1429 if (op3 > opStart4) return ERROR(corruption_detected);
1430 /* note : op4 already verified within main loop */
1431
1432 /* finish bitStreams one by one */
1433 HUF_decodeStreamX2(op1, &bitD1, opStart2, dt, dtLog);
1434 HUF_decodeStreamX2(op2, &bitD2, opStart3, dt, dtLog);
1435 HUF_decodeStreamX2(op3, &bitD3, opStart4, dt, dtLog);
1436 HUF_decodeStreamX2(op4, &bitD4, oend, dt, dtLog);
1437
1438 /* check */
1439 { U32 const endCheck = BIT_endOfDStream(&bitD1) & BIT_endOfDStream(&bitD2) & BIT_endOfDStream(&bitD3) & BIT_endOfDStream(&bitD4);
1440 if (!endCheck) return ERROR(corruption_detected); }
1441
1442 /* decoded size */
1443 return dstSize;
1444 }
1445}
1446
1447#if HUF_NEED_BMI2_FUNCTION
1448static BMI2_TARGET_ATTRIBUTE
1449size_t HUF_decompress4X2_usingDTable_internal_bmi2(void* dst, size_t dstSize, void const* cSrc,
1450 size_t cSrcSize, HUF_DTable const* DTable) {
1451 return HUF_decompress4X2_usingDTable_internal_body(dst, dstSize, cSrc, cSrcSize, DTable);
1452}
1453#endif
1454
1455static
1456size_t HUF_decompress4X2_usingDTable_internal_default(void* dst, size_t dstSize, void const* cSrc,
1457 size_t cSrcSize, HUF_DTable const* DTable) {
1458 return HUF_decompress4X2_usingDTable_internal_body(dst, dstSize, cSrc, cSrcSize, DTable);
1459}
1460
1461#if ZSTD_ENABLE_ASM_X86_64_BMI2
1462
1463HUF_ASM_DECL void HUF_decompress4X2_usingDTable_internal_fast_asm_loop(HUF_DecompressFastArgs* args) ZSTDLIB_HIDDEN;
1464
1465#endif
1466
1467static HUF_FAST_BMI2_ATTRS
1468void HUF_decompress4X2_usingDTable_internal_fast_c_loop(HUF_DecompressFastArgs* args)
1469{
1470 U64 bits[4];
1471 BYTE const* ip[4];
1472 BYTE* op[4];
1473 BYTE* oend[4];
1474 HUF_DEltX2 const* const dtable = (HUF_DEltX2 const*)args->dt;
1475 BYTE const* const ilimit = args->ilimit;
1476
1477 /* Copy the arguments to local registers. */
1478 ZSTD_memcpy(&bits, &args->bits, sizeof(bits));
1479 ZSTD_memcpy((void*)(&ip), &args->ip, sizeof(ip));
1480 ZSTD_memcpy(&op, &args->op, sizeof(op));
1481
1482 oend[0] = op[1];
1483 oend[1] = op[2];
1484 oend[2] = op[3];
1485 oend[3] = args->oend;
1486
1487 assert(MEM_isLittleEndian());
1488 assert(!MEM_32bits());
1489
1490 for (;;) {
1491 BYTE* olimit;
1492 int stream;
1493 int symbol;
1494
1495 /* Assert loop preconditions */
1496#ifndef NDEBUG
1497 for (stream = 0; stream < 4; ++stream) {
1498 assert(op[stream] <= oend[stream]);
1499 assert(ip[stream] >= ilimit);
1500 }
1501#endif
1502 /* Compute olimit */
1503 {
1504 /* Each loop does 5 table lookups for each of the 4 streams.
1505 * Each table lookup consumes up to 11 bits of input, and produces
1506 * up to 2 bytes of output.
1507 */
1508 /* We can consume up to 7 bytes of input per iteration per stream.
1509 * We also know that each input pointer is >= ip[0]. So we can run
1510 * iters loops before running out of input.
1511 */
1512 size_t iters = (size_t)(ip[0] - ilimit) / 7;
1513 /* Each iteration can produce up to 10 bytes of output per stream.
1514 * Each output stream my advance at different rates. So take the
1515 * minimum number of safe iterations among all the output streams.
1516 */
1517 for (stream = 0; stream < 4; ++stream) {
1518 size_t const oiters = (size_t)(oend[stream] - op[stream]) / 10;
1519 iters = MIN(iters, oiters);
1520 }
1521
1522 /* Each iteration produces at least 5 output symbols. So until
1523 * op[3] crosses olimit, we know we haven't executed iters
1524 * iterations yet. This saves us maintaining an iters counter,
1525 * at the expense of computing the remaining # of iterations
1526 * more frequently.
1527 */
1528 olimit = op[3] + (iters * 5);
1529
1530 /* Exit the fast decoding loop if we are too close to the end. */
1531 if (op[3] + 10 > olimit)
1532 break;
1533
1534 /* Exit the decoding loop if any input pointer has crossed the
1535 * previous one. This indicates corruption, and a precondition
1536 * to our loop is that ip[i] >= ip[0].
1537 */
1538 for (stream = 1; stream < 4; ++stream) {
1539 if (ip[stream] < ip[stream - 1])
1540 goto _out;
1541 }
1542 }
1543
1544#ifndef NDEBUG
1545 for (stream = 1; stream < 4; ++stream) {
1546 assert(ip[stream] >= ip[stream - 1]);
1547 }
1548#endif
1549
1550 do {
1551 /* Do 5 table lookups for each of the first 3 streams */
1552 for (symbol = 0; symbol < 5; ++symbol) {
1553 for (stream = 0; stream < 3; ++stream) {
1554 int const index = (int)(bits[stream] >> 53);
1555 HUF_DEltX2 const entry = dtable[index];
1556 MEM_write16(op[stream], entry.sequence);
1557 bits[stream] <<= (entry.nbBits);
1558 op[stream] += (entry.length);
1559 }
1560 }
1561 /* Do 1 table lookup from the final stream */
1562 {
1563 int const index = (int)(bits[3] >> 53);
1564 HUF_DEltX2 const entry = dtable[index];
1565 MEM_write16(op[3], entry.sequence);
1566 bits[3] <<= (entry.nbBits);
1567 op[3] += (entry.length);
1568 }
1569 /* Do 4 table lookups from the final stream & reload bitstreams */
1570 for (stream = 0; stream < 4; ++stream) {
1571 /* Do a table lookup from the final stream.
1572 * This is interleaved with the reloading to reduce register
1573 * pressure. This shouldn't be necessary, but compilers can
1574 * struggle with codegen with high register pressure.
1575 */
1576 {
1577 int const index = (int)(bits[3] >> 53);
1578 HUF_DEltX2 const entry = dtable[index];
1579 MEM_write16(op[3], entry.sequence);
1580 bits[3] <<= (entry.nbBits);
1581 op[3] += (entry.length);
1582 }
1583 /* Reload the bistreams. The final bitstream must be reloaded
1584 * after the 5th symbol was decoded.
1585 */
1586 {
1587 int const ctz = ZSTD_countTrailingZeros64(bits[stream]);
1588 int const nbBits = ctz & 7;
1589 int const nbBytes = ctz >> 3;
1590 ip[stream] -= nbBytes;
1591 bits[stream] = MEM_read64(ip[stream]) | 1;
1592 bits[stream] <<= nbBits;
1593 }
1594 }
1595 } while (op[3] < olimit);
1596 }
1597
1598_out:
1599
1600 /* Save the final values of each of the state variables back to args. */
1601 ZSTD_memcpy(&args->bits, &bits, sizeof(bits));
1602 ZSTD_memcpy((void*)(&args->ip), &ip, sizeof(ip));
1603 ZSTD_memcpy(&args->op, &op, sizeof(op));
1604}
1605
1606
1607static HUF_FAST_BMI2_ATTRS size_t
1608HUF_decompress4X2_usingDTable_internal_fast(
1609 void* dst, size_t dstSize,
1610 const void* cSrc, size_t cSrcSize,
1611 const HUF_DTable* DTable,
1612 HUF_DecompressFastLoopFn loopFn) {
1613 void const* dt = DTable + 1;
1614 const BYTE* const iend = (const BYTE*)cSrc + 6;
1615 BYTE* const oend = (BYTE*)dst + dstSize;
1616 HUF_DecompressFastArgs args;
1617 {
1618 size_t const ret = HUF_DecompressFastArgs_init(&args, dst, dstSize, cSrc, cSrcSize, DTable);
1619 FORWARD_IF_ERROR(ret, "Failed to init asm args");
1620 if (ret == 0)
1621 return 0;
1622 }
1623
1624 assert(args.ip[0] >= args.ilimit);
1625 loopFn(&args);
1626
1627 /* note : op4 already verified within main loop */
1628 assert(args.ip[0] >= iend);
1629 assert(args.ip[1] >= iend);
1630 assert(args.ip[2] >= iend);
1631 assert(args.ip[3] >= iend);
1632 assert(args.op[3] <= oend);
1633 (void)iend;
1634
1635 /* finish bitStreams one by one */
1636 {
1637 size_t const segmentSize = (dstSize+3) / 4;
1638 BYTE* segmentEnd = (BYTE*)dst;
1639 int i;
1640 for (i = 0; i < 4; ++i) {
1641 BIT_DStream_t bit;
1642 if (segmentSize <= (size_t)(oend - segmentEnd))
1643 segmentEnd += segmentSize;
1644 else
1645 segmentEnd = oend;
1646 FORWARD_IF_ERROR(HUF_initRemainingDStream(&bit, &args, i, segmentEnd), "corruption");
1647 args.op[i] += HUF_decodeStreamX2(args.op[i], &bit, segmentEnd, (HUF_DEltX2 const*)dt, HUF_DECODER_FAST_TABLELOG);
1648 if (args.op[i] != segmentEnd)
1649 return ERROR(corruption_detected);
1650 }
1651 }
1652
1653 /* decoded size */
1654 return dstSize;
1655}
1656
1657static size_t HUF_decompress4X2_usingDTable_internal(void* dst, size_t dstSize, void const* cSrc,
1658 size_t cSrcSize, HUF_DTable const* DTable, int flags)
1659{
1660 HUF_DecompressUsingDTableFn fallbackFn = HUF_decompress4X2_usingDTable_internal_default;
1661 HUF_DecompressFastLoopFn loopFn = HUF_decompress4X2_usingDTable_internal_fast_c_loop;
1662
1663#if DYNAMIC_BMI2
1664 if (flags & HUF_flags_bmi2) {
1665 fallbackFn = HUF_decompress4X2_usingDTable_internal_bmi2;
1666# if ZSTD_ENABLE_ASM_X86_64_BMI2
1667 if (!(flags & HUF_flags_disableAsm)) {
1668 loopFn = HUF_decompress4X2_usingDTable_internal_fast_asm_loop;
1669 }
1670# endif
1671 } else {
1672 return fallbackFn(dst, dstSize, cSrc, cSrcSize, DTable);
1673 }
1674#endif
1675
1676#if ZSTD_ENABLE_ASM_X86_64_BMI2 && defined(__BMI2__)
1677 if (!(flags & HUF_flags_disableAsm)) {
1678 loopFn = HUF_decompress4X2_usingDTable_internal_fast_asm_loop;
1679 }
1680#endif
1681
1682 if (!(flags & HUF_flags_disableFast)) {
1683 size_t const ret = HUF_decompress4X2_usingDTable_internal_fast(dst, dstSize, cSrc, cSrcSize, DTable, loopFn);
1684 if (ret != 0)
1685 return ret;
1686 }
1687 return fallbackFn(dst, dstSize, cSrc, cSrcSize, DTable);
1688}
1689
1690HUF_DGEN(HUF_decompress1X2_usingDTable_internal)
1691
1692size_t HUF_decompress1X2_DCtx_wksp(HUF_DTable* DCtx, void* dst, size_t dstSize,
1693 const void* cSrc, size_t cSrcSize,
1694 void* workSpace, size_t wkspSize, int flags)
1695{
1696 const BYTE* ip = (const BYTE*) cSrc;
1697
1698 size_t const hSize = HUF_readDTableX2_wksp(DCtx, cSrc, cSrcSize,
1699 workSpace, wkspSize, flags);
1700 if (HUF_isError(hSize)) return hSize;
1701 if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
1702 ip += hSize; cSrcSize -= hSize;
1703
1704 return HUF_decompress1X2_usingDTable_internal(dst, dstSize, ip, cSrcSize, DCtx, flags);
1705}
1706
1707static size_t HUF_decompress4X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
1708 const void* cSrc, size_t cSrcSize,
1709 void* workSpace, size_t wkspSize, int flags)
1710{
1711 const BYTE* ip = (const BYTE*) cSrc;
1712
1713 size_t hSize = HUF_readDTableX2_wksp(dctx, cSrc, cSrcSize,
1714 workSpace, wkspSize, flags);
1715 if (HUF_isError(hSize)) return hSize;
1716 if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
1717 ip += hSize; cSrcSize -= hSize;
1718
1719 return HUF_decompress4X2_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx, flags);
1720}
1721
1722#endif /* HUF_FORCE_DECOMPRESS_X1 */
1723
1724
1725/* ***********************************/
1726/* Universal decompression selectors */
1727/* ***********************************/
1728
1729
1730#if !defined(HUF_FORCE_DECOMPRESS_X1) && !defined(HUF_FORCE_DECOMPRESS_X2)
1731typedef struct { U32 tableTime; U32 decode256Time; } algo_time_t;
1732static const algo_time_t algoTime[16 /* Quantization */][2 /* single, double */] =
1733{
1734 /* single, double, quad */
1735 {{0,0}, {1,1}}, /* Q==0 : impossible */
1736 {{0,0}, {1,1}}, /* Q==1 : impossible */
1737 {{ 150,216}, { 381,119}}, /* Q == 2 : 12-18% */
1738 {{ 170,205}, { 514,112}}, /* Q == 3 : 18-25% */
1739 {{ 177,199}, { 539,110}}, /* Q == 4 : 25-32% */
1740 {{ 197,194}, { 644,107}}, /* Q == 5 : 32-38% */
1741 {{ 221,192}, { 735,107}}, /* Q == 6 : 38-44% */
1742 {{ 256,189}, { 881,106}}, /* Q == 7 : 44-50% */
1743 {{ 359,188}, {1167,109}}, /* Q == 8 : 50-56% */
1744 {{ 582,187}, {1570,114}}, /* Q == 9 : 56-62% */
1745 {{ 688,187}, {1712,122}}, /* Q ==10 : 62-69% */
1746 {{ 825,186}, {1965,136}}, /* Q ==11 : 69-75% */
1747 {{ 976,185}, {2131,150}}, /* Q ==12 : 75-81% */
1748 {{1180,186}, {2070,175}}, /* Q ==13 : 81-87% */
1749 {{1377,185}, {1731,202}}, /* Q ==14 : 87-93% */
1750 {{1412,185}, {1695,202}}, /* Q ==15 : 93-99% */
1751};
1752#endif
1753
1754/** HUF_selectDecoder() :
1755 * Tells which decoder is likely to decode faster,
1756 * based on a set of pre-computed metrics.
1757 * @return : 0==HUF_decompress4X1, 1==HUF_decompress4X2 .
1758 * Assumption : 0 < dstSize <= 128 KB */
1759U32 HUF_selectDecoder (size_t dstSize, size_t cSrcSize)
1760{
1761 assert(dstSize > 0);
1762 assert(dstSize <= 128*1024);
1763#if defined(HUF_FORCE_DECOMPRESS_X1)
1764 (void)dstSize;
1765 (void)cSrcSize;
1766 return 0;
1767#elif defined(HUF_FORCE_DECOMPRESS_X2)
1768 (void)dstSize;
1769 (void)cSrcSize;
1770 return 1;
1771#else
1772 /* decoder timing evaluation */
1773 { U32 const Q = (cSrcSize >= dstSize) ? 15 : (U32)(cSrcSize * 16 / dstSize); /* Q < 16 */
1774 U32 const D256 = (U32)(dstSize >> 8);
1775 U32 const DTime0 = algoTime[Q][0].tableTime + (algoTime[Q][0].decode256Time * D256);
1776 U32 DTime1 = algoTime[Q][1].tableTime + (algoTime[Q][1].decode256Time * D256);
1777 DTime1 += DTime1 >> 5; /* small advantage to algorithm using less memory, to reduce cache eviction */
1778 return DTime1 < DTime0;
1779 }
1780#endif
1781}
1782
1783size_t HUF_decompress1X_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
1784 const void* cSrc, size_t cSrcSize,
1785 void* workSpace, size_t wkspSize, int flags)
1786{
1787 /* validation checks */
1788 if (dstSize == 0) return ERROR(dstSize_tooSmall);
1789 if (cSrcSize > dstSize) return ERROR(corruption_detected); /* invalid */
1790 if (cSrcSize == dstSize) { ZSTD_memcpy(dst, cSrc, dstSize); return dstSize; } /* not compressed */
1791 if (cSrcSize == 1) { ZSTD_memset(dst, *(const BYTE*)cSrc, dstSize); return dstSize; } /* RLE */
1792
1793 { U32 const algoNb = HUF_selectDecoder(dstSize, cSrcSize);
1794#if defined(HUF_FORCE_DECOMPRESS_X1)
1795 (void)algoNb;
1796 assert(algoNb == 0);
1797 return HUF_decompress1X1_DCtx_wksp(dctx, dst, dstSize, cSrc,
1798 cSrcSize, workSpace, wkspSize, flags);
1799#elif defined(HUF_FORCE_DECOMPRESS_X2)
1800 (void)algoNb;
1801 assert(algoNb == 1);
1802 return HUF_decompress1X2_DCtx_wksp(dctx, dst, dstSize, cSrc,
1803 cSrcSize, workSpace, wkspSize, flags);
1804#else
1805 return algoNb ? HUF_decompress1X2_DCtx_wksp(dctx, dst, dstSize, cSrc,
1806 cSrcSize, workSpace, wkspSize, flags):
1807 HUF_decompress1X1_DCtx_wksp(dctx, dst, dstSize, cSrc,
1808 cSrcSize, workSpace, wkspSize, flags);
1809#endif
1810 }
1811}
1812
1813
1814size_t HUF_decompress1X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int flags)
1815{
1816 DTableDesc const dtd = HUF_getDTableDesc(DTable);
1817#if defined(HUF_FORCE_DECOMPRESS_X1)
1818 (void)dtd;
1819 assert(dtd.tableType == 0);
1820 return HUF_decompress1X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
1821#elif defined(HUF_FORCE_DECOMPRESS_X2)
1822 (void)dtd;
1823 assert(dtd.tableType == 1);
1824 return HUF_decompress1X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
1825#else
1826 return dtd.tableType ? HUF_decompress1X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags) :
1827 HUF_decompress1X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
1828#endif
1829}
1830
1831#ifndef HUF_FORCE_DECOMPRESS_X2
1832size_t HUF_decompress1X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags)
1833{
1834 const BYTE* ip = (const BYTE*) cSrc;
1835
1836 size_t const hSize = HUF_readDTableX1_wksp(dctx, cSrc, cSrcSize, workSpace, wkspSize, flags);
1837 if (HUF_isError(hSize)) return hSize;
1838 if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
1839 ip += hSize; cSrcSize -= hSize;
1840
1841 return HUF_decompress1X1_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx, flags);
1842}
1843#endif
1844
1845size_t HUF_decompress4X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int flags)
1846{
1847 DTableDesc const dtd = HUF_getDTableDesc(DTable);
1848#if defined(HUF_FORCE_DECOMPRESS_X1)
1849 (void)dtd;
1850 assert(dtd.tableType == 0);
1851 return HUF_decompress4X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
1852#elif defined(HUF_FORCE_DECOMPRESS_X2)
1853 (void)dtd;
1854 assert(dtd.tableType == 1);
1855 return HUF_decompress4X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
1856#else
1857 return dtd.tableType ? HUF_decompress4X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags) :
1858 HUF_decompress4X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
1859#endif
1860}
1861
1862size_t HUF_decompress4X_hufOnly_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags)
1863{
1864 /* validation checks */
1865 if (dstSize == 0) return ERROR(dstSize_tooSmall);
1866 if (cSrcSize == 0) return ERROR(corruption_detected);
1867
1868 { U32 const algoNb = HUF_selectDecoder(dstSize, cSrcSize);
1869#if defined(HUF_FORCE_DECOMPRESS_X1)
1870 (void)algoNb;
1871 assert(algoNb == 0);
1872 return HUF_decompress4X1_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, flags);
1873#elif defined(HUF_FORCE_DECOMPRESS_X2)
1874 (void)algoNb;
1875 assert(algoNb == 1);
1876 return HUF_decompress4X2_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, flags);
1877#else
1878 return algoNb ? HUF_decompress4X2_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, flags) :
1879 HUF_decompress4X1_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, flags);
1880#endif
1881 }
1882}
ARM optimized PCSX fork