1 /* ******************************************************************
2 * Huffman encoder, part of New Generation Entropy library
3 * Copyright (c) Yann Collet, Facebook, Inc.
4 *
5 * You can contact the author at :
6 * - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
7 * - Public forum : https://groups.google.com/forum/#!forum/lz4c
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 * Compiler specifics
17 ****************************************************************/
18
19
20 /* **************************************************************
21 * Includes
22 ****************************************************************/
23 #include "../common/zstd_deps.h" /* ZSTD_memcpy, ZSTD_memset */
24 #include "../common/compiler.h"
25 #include "../common/bitstream.h"
26 #include "hist.h"
27 #define FSE_STATIC_LINKING_ONLY /* FSE_optimalTableLog_internal */
28 #include "../common/fse.h" /* header compression */
29 #define HUF_STATIC_LINKING_ONLY
30 #include "../common/huf.h"
31 #include "../common/error_private.h"
32
33
34 /* **************************************************************
35 * Error Management
36 ****************************************************************/
37 #define HUF_isError ERR_isError
38 #define HUF_STATIC_ASSERT(c) DEBUG_STATIC_ASSERT(c) /* use only *after* variable declarations */
39
40
41 /* **************************************************************
42 * Utils
43 ****************************************************************/
HUF_optimalTableLog(unsigned maxTableLog,size_t srcSize,unsigned maxSymbolValue)44 unsigned HUF_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue)
45 {
46 return FSE_optimalTableLog_internal(maxTableLog, srcSize, maxSymbolValue, 1);
47 }
48
49
50 /* *******************************************************
51 * HUF : Huffman block compression
52 *********************************************************/
53 /* HUF_compressWeights() :
54 * Same as FSE_compress(), but dedicated to huff0's weights compression.
55 * The use case needs much less stack memory.
56 * Note : all elements within weightTable are supposed to be <= HUF_TABLELOG_MAX.
57 */
58 #define MAX_FSE_TABLELOG_FOR_HUFF_HEADER 6
59
60 typedef struct {
61 FSE_CTable CTable[FSE_CTABLE_SIZE_U32(MAX_FSE_TABLELOG_FOR_HUFF_HEADER, HUF_TABLELOG_MAX)];
62 U32 scratchBuffer[FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(HUF_TABLELOG_MAX, MAX_FSE_TABLELOG_FOR_HUFF_HEADER)];
63 unsigned count[HUF_TABLELOG_MAX+1];
64 S16 norm[HUF_TABLELOG_MAX+1];
65 } HUF_CompressWeightsWksp;
66
HUF_compressWeights(void * dst,size_t dstSize,const void * weightTable,size_t wtSize,void * workspace,size_t workspaceSize)67 static size_t HUF_compressWeights(void* dst, size_t dstSize, const void* weightTable, size_t wtSize, void* workspace, size_t workspaceSize)
68 {
69 BYTE* const ostart = (BYTE*) dst;
70 BYTE* op = ostart;
71 BYTE* const oend = ostart + dstSize;
72
73 unsigned maxSymbolValue = HUF_TABLELOG_MAX;
74 U32 tableLog = MAX_FSE_TABLELOG_FOR_HUFF_HEADER;
75 HUF_CompressWeightsWksp* wksp = (HUF_CompressWeightsWksp*)workspace;
76
77 if (workspaceSize < sizeof(HUF_CompressWeightsWksp)) return ERROR(GENERIC);
78
79 /* init conditions */
80 if (wtSize <= 1) return 0; /* Not compressible */
81
82 /* Scan input and build symbol stats */
83 { unsigned const maxCount = HIST_count_simple(wksp->count, &maxSymbolValue, weightTable, wtSize); /* never fails */
84 if (maxCount == wtSize) return 1; /* only a single symbol in src : rle */
85 if (maxCount == 1) return 0; /* each symbol present maximum once => not compressible */
86 }
87
88 tableLog = FSE_optimalTableLog(tableLog, wtSize, maxSymbolValue);
89 CHECK_F( FSE_normalizeCount(wksp->norm, tableLog, wksp->count, wtSize, maxSymbolValue, /* useLowProbCount */ 0) );
90
91 /* Write table description header */
92 { CHECK_V_F(hSize, FSE_writeNCount(op, (size_t)(oend-op), wksp->norm, maxSymbolValue, tableLog) );
93 op += hSize;
94 }
95
96 /* Compress */
97 CHECK_F( FSE_buildCTable_wksp(wksp->CTable, wksp->norm, maxSymbolValue, tableLog, wksp->scratchBuffer, sizeof(wksp->scratchBuffer)) );
98 { CHECK_V_F(cSize, FSE_compress_usingCTable(op, (size_t)(oend - op), weightTable, wtSize, wksp->CTable) );
99 if (cSize == 0) return 0; /* not enough space for compressed data */
100 op += cSize;
101 }
102
103 return (size_t)(op-ostart);
104 }
105
106
107 typedef struct {
108 HUF_CompressWeightsWksp wksp;
109 BYTE bitsToWeight[HUF_TABLELOG_MAX + 1]; /* precomputed conversion table */
110 BYTE huffWeight[HUF_SYMBOLVALUE_MAX];
111 } HUF_WriteCTableWksp;
112
HUF_writeCTable_wksp(void * dst,size_t maxDstSize,const HUF_CElt * CTable,unsigned maxSymbolValue,unsigned huffLog,void * workspace,size_t workspaceSize)113 size_t HUF_writeCTable_wksp(void* dst, size_t maxDstSize,
114 const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog,
115 void* workspace, size_t workspaceSize)
116 {
117 BYTE* op = (BYTE*)dst;
118 U32 n;
119 HUF_WriteCTableWksp* wksp = (HUF_WriteCTableWksp*)workspace;
120
121 /* check conditions */
122 if (workspaceSize < sizeof(HUF_WriteCTableWksp)) return ERROR(GENERIC);
123 if (maxSymbolValue > HUF_SYMBOLVALUE_MAX) return ERROR(maxSymbolValue_tooLarge);
124
125 /* convert to weight */
126 wksp->bitsToWeight[0] = 0;
127 for (n=1; n<huffLog+1; n++)
128 wksp->bitsToWeight[n] = (BYTE)(huffLog + 1 - n);
129 for (n=0; n<maxSymbolValue; n++)
130 wksp->huffWeight[n] = wksp->bitsToWeight[CTable[n].nbBits];
131
132 /* attempt weights compression by FSE */
133 { CHECK_V_F(hSize, HUF_compressWeights(op+1, maxDstSize-1, wksp->huffWeight, maxSymbolValue, &wksp->wksp, sizeof(wksp->wksp)) );
134 if ((hSize>1) & (hSize < maxSymbolValue/2)) { /* FSE compressed */
135 op[0] = (BYTE)hSize;
136 return hSize+1;
137 } }
138
139 /* write raw values as 4-bits (max : 15) */
140 if (maxSymbolValue > (256-128)) return ERROR(GENERIC); /* should not happen : likely means source cannot be compressed */
141 if (((maxSymbolValue+1)/2) + 1 > maxDstSize) return ERROR(dstSize_tooSmall); /* not enough space within dst buffer */
142 op[0] = (BYTE)(128 /*special case*/ + (maxSymbolValue-1));
143 wksp->huffWeight[maxSymbolValue] = 0; /* to be sure it doesn't cause msan issue in final combination */
144 for (n=0; n<maxSymbolValue; n+=2)
145 op[(n/2)+1] = (BYTE)((wksp->huffWeight[n] << 4) + wksp->huffWeight[n+1]);
146 return ((maxSymbolValue+1)/2) + 1;
147 }
148
149 /*! HUF_writeCTable() :
150 `CTable` : Huffman tree to save, using huf representation.
151 @return : size of saved CTable */
HUF_writeCTable(void * dst,size_t maxDstSize,const HUF_CElt * CTable,unsigned maxSymbolValue,unsigned huffLog)152 size_t HUF_writeCTable (void* dst, size_t maxDstSize,
153 const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog)
154 {
155 HUF_WriteCTableWksp wksp;
156 return HUF_writeCTable_wksp(dst, maxDstSize, CTable, maxSymbolValue, huffLog, &wksp, sizeof(wksp));
157 }
158
159
HUF_readCTable(HUF_CElt * CTable,unsigned * maxSymbolValuePtr,const void * src,size_t srcSize,unsigned * hasZeroWeights)160 size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void* src, size_t srcSize, unsigned* hasZeroWeights)
161 {
162 BYTE huffWeight[HUF_SYMBOLVALUE_MAX + 1]; /* init not required, even though some static analyzer may complain */
163 U32 rankVal[HUF_TABLELOG_ABSOLUTEMAX + 1]; /* large enough for values from 0 to 16 */
164 U32 tableLog = 0;
165 U32 nbSymbols = 0;
166
167 /* get symbol weights */
168 CHECK_V_F(readSize, HUF_readStats(huffWeight, HUF_SYMBOLVALUE_MAX+1, rankVal, &nbSymbols, &tableLog, src, srcSize));
169 *hasZeroWeights = (rankVal[0] > 0);
170
171 /* check result */
172 if (tableLog > HUF_TABLELOG_MAX) return ERROR(tableLog_tooLarge);
173 if (nbSymbols > *maxSymbolValuePtr+1) return ERROR(maxSymbolValue_tooSmall);
174
175 /* Prepare base value per rank */
176 { U32 n, nextRankStart = 0;
177 for (n=1; n<=tableLog; n++) {
178 U32 curr = nextRankStart;
179 nextRankStart += (rankVal[n] << (n-1));
180 rankVal[n] = curr;
181 } }
182
183 /* fill nbBits */
184 { U32 n; for (n=0; n<nbSymbols; n++) {
185 const U32 w = huffWeight[n];
186 CTable[n].nbBits = (BYTE)(tableLog + 1 - w) & -(w != 0);
187 } }
188
189 /* fill val */
190 { U16 nbPerRank[HUF_TABLELOG_MAX+2] = {0}; /* support w=0=>n=tableLog+1 */
191 U16 valPerRank[HUF_TABLELOG_MAX+2] = {0};
192 { U32 n; for (n=0; n<nbSymbols; n++) nbPerRank[CTable[n].nbBits]++; }
193 /* determine stating value per rank */
194 valPerRank[tableLog+1] = 0; /* for w==0 */
195 { U16 min = 0;
196 U32 n; for (n=tableLog; n>0; n--) { /* start at n=tablelog <-> w=1 */
197 valPerRank[n] = min; /* get starting value within each rank */
198 min += nbPerRank[n];
199 min >>= 1;
200 } }
201 /* assign value within rank, symbol order */
202 { U32 n; for (n=0; n<nbSymbols; n++) CTable[n].val = valPerRank[CTable[n].nbBits]++; }
203 }
204
205 *maxSymbolValuePtr = nbSymbols - 1;
206 return readSize;
207 }
208
HUF_getNbBits(const void * symbolTable,U32 symbolValue)209 U32 HUF_getNbBits(const void* symbolTable, U32 symbolValue)
210 {
211 const HUF_CElt* table = (const HUF_CElt*)symbolTable;
212 assert(symbolValue <= HUF_SYMBOLVALUE_MAX);
213 return table[symbolValue].nbBits;
214 }
215
216
217 typedef struct nodeElt_s {
218 U32 count;
219 U16 parent;
220 BYTE byte;
221 BYTE nbBits;
222 } nodeElt;
223
224 /*
225 * HUF_setMaxHeight():
226 * Enforces maxNbBits on the Huffman tree described in huffNode.
227 *
228 * It sets all nodes with nbBits > maxNbBits to be maxNbBits. Then it adjusts
229 * the tree to so that it is a valid canonical Huffman tree.
230 *
231 * @pre The sum of the ranks of each symbol == 2^largestBits,
232 * where largestBits == huffNode[lastNonNull].nbBits.
233 * @post The sum of the ranks of each symbol == 2^largestBits,
234 * where largestBits is the return value <= maxNbBits.
235 *
236 * @param huffNode The Huffman tree modified in place to enforce maxNbBits.
237 * @param lastNonNull The symbol with the lowest count in the Huffman tree.
238 * @param maxNbBits The maximum allowed number of bits, which the Huffman tree
239 * may not respect. After this function the Huffman tree will
240 * respect maxNbBits.
241 * @return The maximum number of bits of the Huffman tree after adjustment,
242 * necessarily no more than maxNbBits.
243 */
HUF_setMaxHeight(nodeElt * huffNode,U32 lastNonNull,U32 maxNbBits)244 static U32 HUF_setMaxHeight(nodeElt* huffNode, U32 lastNonNull, U32 maxNbBits)
245 {
246 const U32 largestBits = huffNode[lastNonNull].nbBits;
247 /* early exit : no elt > maxNbBits, so the tree is already valid. */
248 if (largestBits <= maxNbBits) return largestBits;
249
250 /* there are several too large elements (at least >= 2) */
251 { int totalCost = 0;
252 const U32 baseCost = 1 << (largestBits - maxNbBits);
253 int n = (int)lastNonNull;
254
255 /* Adjust any ranks > maxNbBits to maxNbBits.
256 * Compute totalCost, which is how far the sum of the ranks is
257 * we are over 2^largestBits after adjust the offending ranks.
258 */
259 while (huffNode[n].nbBits > maxNbBits) {
260 totalCost += baseCost - (1 << (largestBits - huffNode[n].nbBits));
261 huffNode[n].nbBits = (BYTE)maxNbBits;
262 n--;
263 }
264 /* n stops at huffNode[n].nbBits <= maxNbBits */
265 assert(huffNode[n].nbBits <= maxNbBits);
266 /* n end at index of smallest symbol using < maxNbBits */
267 while (huffNode[n].nbBits == maxNbBits) --n;
268
269 /* renorm totalCost from 2^largestBits to 2^maxNbBits
270 * note : totalCost is necessarily a multiple of baseCost */
271 assert((totalCost & (baseCost - 1)) == 0);
272 totalCost >>= (largestBits - maxNbBits);
273 assert(totalCost > 0);
274
275 /* repay normalized cost */
276 { U32 const noSymbol = 0xF0F0F0F0;
277 U32 rankLast[HUF_TABLELOG_MAX+2];
278
279 /* Get pos of last (smallest = lowest cum. count) symbol per rank */
280 ZSTD_memset(rankLast, 0xF0, sizeof(rankLast));
281 { U32 currentNbBits = maxNbBits;
282 int pos;
283 for (pos=n ; pos >= 0; pos--) {
284 if (huffNode[pos].nbBits >= currentNbBits) continue;
285 currentNbBits = huffNode[pos].nbBits; /* < maxNbBits */
286 rankLast[maxNbBits-currentNbBits] = (U32)pos;
287 } }
288
289 while (totalCost > 0) {
290 /* Try to reduce the next power of 2 above totalCost because we
291 * gain back half the rank.
292 */
293 U32 nBitsToDecrease = BIT_highbit32((U32)totalCost) + 1;
294 for ( ; nBitsToDecrease > 1; nBitsToDecrease--) {
295 U32 const highPos = rankLast[nBitsToDecrease];
296 U32 const lowPos = rankLast[nBitsToDecrease-1];
297 if (highPos == noSymbol) continue;
298 /* Decrease highPos if no symbols of lowPos or if it is
299 * not cheaper to remove 2 lowPos than highPos.
300 */
301 if (lowPos == noSymbol) break;
302 { U32 const highTotal = huffNode[highPos].count;
303 U32 const lowTotal = 2 * huffNode[lowPos].count;
304 if (highTotal <= lowTotal) break;
305 } }
306 /* only triggered when no more rank 1 symbol left => find closest one (note : there is necessarily at least one !) */
307 assert(rankLast[nBitsToDecrease] != noSymbol || nBitsToDecrease == 1);
308 /* HUF_MAX_TABLELOG test just to please gcc 5+; but it should not be necessary */
309 while ((nBitsToDecrease<=HUF_TABLELOG_MAX) && (rankLast[nBitsToDecrease] == noSymbol))
310 nBitsToDecrease++;
311 assert(rankLast[nBitsToDecrease] != noSymbol);
312 /* Increase the number of bits to gain back half the rank cost. */
313 totalCost -= 1 << (nBitsToDecrease-1);
314 huffNode[rankLast[nBitsToDecrease]].nbBits++;
315
316 /* Fix up the new rank.
317 * If the new rank was empty, this symbol is now its smallest.
318 * Otherwise, this symbol will be the largest in the new rank so no adjustment.
319 */
320 if (rankLast[nBitsToDecrease-1] == noSymbol)
321 rankLast[nBitsToDecrease-1] = rankLast[nBitsToDecrease];
322 /* Fix up the old rank.
323 * If the symbol was at position 0, meaning it was the highest weight symbol in the tree,
324 * it must be the only symbol in its rank, so the old rank now has no symbols.
325 * Otherwise, since the Huffman nodes are sorted by count, the previous position is now
326 * the smallest node in the rank. If the previous position belongs to a different rank,
327 * then the rank is now empty.
328 */
329 if (rankLast[nBitsToDecrease] == 0) /* special case, reached largest symbol */
330 rankLast[nBitsToDecrease] = noSymbol;
331 else {
332 rankLast[nBitsToDecrease]--;
333 if (huffNode[rankLast[nBitsToDecrease]].nbBits != maxNbBits-nBitsToDecrease)
334 rankLast[nBitsToDecrease] = noSymbol; /* this rank is now empty */
335 }
336 } /* while (totalCost > 0) */
337
338 /* If we've removed too much weight, then we have to add it back.
339 * To avoid overshooting again, we only adjust the smallest rank.
340 * We take the largest nodes from the lowest rank 0 and move them
341 * to rank 1. There's guaranteed to be enough rank 0 symbols because
342 * TODO.
343 */
344 while (totalCost < 0) { /* Sometimes, cost correction overshoot */
345 /* special case : no rank 1 symbol (using maxNbBits-1);
346 * let's create one from largest rank 0 (using maxNbBits).
347 */
348 if (rankLast[1] == noSymbol) {
349 while (huffNode[n].nbBits == maxNbBits) n--;
350 huffNode[n+1].nbBits--;
351 assert(n >= 0);
352 rankLast[1] = (U32)(n+1);
353 totalCost++;
354 continue;
355 }
356 huffNode[ rankLast[1] + 1 ].nbBits--;
357 rankLast[1]++;
358 totalCost ++;
359 }
360 } /* repay normalized cost */
361 } /* there are several too large elements (at least >= 2) */
362
363 return maxNbBits;
364 }
365
366 typedef struct {
367 U32 base;
368 U32 curr;
369 } rankPos;
370
371 typedef nodeElt huffNodeTable[HUF_CTABLE_WORKSPACE_SIZE_U32];
372
373 #define RANK_POSITION_TABLE_SIZE 32
374
375 typedef struct {
376 huffNodeTable huffNodeTbl;
377 rankPos rankPosition[RANK_POSITION_TABLE_SIZE];
378 } HUF_buildCTable_wksp_tables;
379
380 /*
381 * HUF_sort():
382 * Sorts the symbols [0, maxSymbolValue] by count[symbol] in decreasing order.
383 *
384 * @param[out] huffNode Sorted symbols by decreasing count. Only members `.count` and `.byte` are filled.
385 * Must have (maxSymbolValue + 1) entries.
386 * @param[in] count Histogram of the symbols.
387 * @param[in] maxSymbolValue Maximum symbol value.
388 * @param rankPosition This is a scratch workspace. Must have RANK_POSITION_TABLE_SIZE entries.
389 */
HUF_sort(nodeElt * huffNode,const unsigned * count,U32 maxSymbolValue,rankPos * rankPosition)390 static void HUF_sort(nodeElt* huffNode, const unsigned* count, U32 maxSymbolValue, rankPos* rankPosition)
391 {
392 int n;
393 int const maxSymbolValue1 = (int)maxSymbolValue + 1;
394
395 /* Compute base and set curr to base.
396 * For symbol s let lowerRank = BIT_highbit32(count[n]+1) and rank = lowerRank + 1.
397 * Then 2^lowerRank <= count[n]+1 <= 2^rank.
398 * We attribute each symbol to lowerRank's base value, because we want to know where
399 * each rank begins in the output, so for rank R we want to count ranks R+1 and above.
400 */
401 ZSTD_memset(rankPosition, 0, sizeof(*rankPosition) * RANK_POSITION_TABLE_SIZE);
402 for (n = 0; n < maxSymbolValue1; ++n) {
403 U32 lowerRank = BIT_highbit32(count[n] + 1);
404 rankPosition[lowerRank].base++;
405 }
406 assert(rankPosition[RANK_POSITION_TABLE_SIZE - 1].base == 0);
407 for (n = RANK_POSITION_TABLE_SIZE - 1; n > 0; --n) {
408 rankPosition[n-1].base += rankPosition[n].base;
409 rankPosition[n-1].curr = rankPosition[n-1].base;
410 }
411 /* Sort */
412 for (n = 0; n < maxSymbolValue1; ++n) {
413 U32 const c = count[n];
414 U32 const r = BIT_highbit32(c+1) + 1;
415 U32 pos = rankPosition[r].curr++;
416 /* Insert into the correct position in the rank.
417 * We have at most 256 symbols, so this insertion should be fine.
418 */
419 while ((pos > rankPosition[r].base) && (c > huffNode[pos-1].count)) {
420 huffNode[pos] = huffNode[pos-1];
421 pos--;
422 }
423 huffNode[pos].count = c;
424 huffNode[pos].byte = (BYTE)n;
425 }
426 }
427
428
429 /* HUF_buildCTable_wksp() :
430 * Same as HUF_buildCTable(), but using externally allocated scratch buffer.
431 * `workSpace` must be aligned on 4-bytes boundaries, and be at least as large as sizeof(HUF_buildCTable_wksp_tables).
432 */
433 #define STARTNODE (HUF_SYMBOLVALUE_MAX+1)
434
435 /* HUF_buildTree():
436 * Takes the huffNode array sorted by HUF_sort() and builds an unlimited-depth Huffman tree.
437 *
438 * @param huffNode The array sorted by HUF_sort(). Builds the Huffman tree in this array.
439 * @param maxSymbolValue The maximum symbol value.
440 * @return The smallest node in the Huffman tree (by count).
441 */
HUF_buildTree(nodeElt * huffNode,U32 maxSymbolValue)442 static int HUF_buildTree(nodeElt* huffNode, U32 maxSymbolValue)
443 {
444 nodeElt* const huffNode0 = huffNode - 1;
445 int nonNullRank;
446 int lowS, lowN;
447 int nodeNb = STARTNODE;
448 int n, nodeRoot;
449 /* init for parents */
450 nonNullRank = (int)maxSymbolValue;
451 while(huffNode[nonNullRank].count == 0) nonNullRank--;
452 lowS = nonNullRank; nodeRoot = nodeNb + lowS - 1; lowN = nodeNb;
453 huffNode[nodeNb].count = huffNode[lowS].count + huffNode[lowS-1].count;
454 huffNode[lowS].parent = huffNode[lowS-1].parent = (U16)nodeNb;
455 nodeNb++; lowS-=2;
456 for (n=nodeNb; n<=nodeRoot; n++) huffNode[n].count = (U32)(1U<<30);
457 huffNode0[0].count = (U32)(1U<<31); /* fake entry, strong barrier */
458
459 /* create parents */
460 while (nodeNb <= nodeRoot) {
461 int const n1 = (huffNode[lowS].count < huffNode[lowN].count) ? lowS-- : lowN++;
462 int const n2 = (huffNode[lowS].count < huffNode[lowN].count) ? lowS-- : lowN++;
463 huffNode[nodeNb].count = huffNode[n1].count + huffNode[n2].count;
464 huffNode[n1].parent = huffNode[n2].parent = (U16)nodeNb;
465 nodeNb++;
466 }
467
468 /* distribute weights (unlimited tree height) */
469 huffNode[nodeRoot].nbBits = 0;
470 for (n=nodeRoot-1; n>=STARTNODE; n--)
471 huffNode[n].nbBits = huffNode[ huffNode[n].parent ].nbBits + 1;
472 for (n=0; n<=nonNullRank; n++)
473 huffNode[n].nbBits = huffNode[ huffNode[n].parent ].nbBits + 1;
474
475 return nonNullRank;
476 }
477
478 /*
479 * HUF_buildCTableFromTree():
480 * Build the CTable given the Huffman tree in huffNode.
481 *
482 * @param[out] CTable The output Huffman CTable.
483 * @param huffNode The Huffman tree.
484 * @param nonNullRank The last and smallest node in the Huffman tree.
485 * @param maxSymbolValue The maximum symbol value.
486 * @param maxNbBits The exact maximum number of bits used in the Huffman tree.
487 */
HUF_buildCTableFromTree(HUF_CElt * CTable,nodeElt const * huffNode,int nonNullRank,U32 maxSymbolValue,U32 maxNbBits)488 static void HUF_buildCTableFromTree(HUF_CElt* CTable, nodeElt const* huffNode, int nonNullRank, U32 maxSymbolValue, U32 maxNbBits)
489 {
490 /* fill result into ctable (val, nbBits) */
491 int n;
492 U16 nbPerRank[HUF_TABLELOG_MAX+1] = {0};
493 U16 valPerRank[HUF_TABLELOG_MAX+1] = {0};
494 int const alphabetSize = (int)(maxSymbolValue + 1);
495 for (n=0; n<=nonNullRank; n++)
496 nbPerRank[huffNode[n].nbBits]++;
497 /* determine starting value per rank */
498 { U16 min = 0;
499 for (n=(int)maxNbBits; n>0; n--) {
500 valPerRank[n] = min; /* get starting value within each rank */
501 min += nbPerRank[n];
502 min >>= 1;
503 } }
504 for (n=0; n<alphabetSize; n++)
505 CTable[huffNode[n].byte].nbBits = huffNode[n].nbBits; /* push nbBits per symbol, symbol order */
506 for (n=0; n<alphabetSize; n++)
507 CTable[n].val = valPerRank[CTable[n].nbBits]++; /* assign value within rank, symbol order */
508 }
509
HUF_buildCTable_wksp(HUF_CElt * tree,const unsigned * count,U32 maxSymbolValue,U32 maxNbBits,void * workSpace,size_t wkspSize)510 size_t HUF_buildCTable_wksp (HUF_CElt* tree, const unsigned* count, U32 maxSymbolValue, U32 maxNbBits, void* workSpace, size_t wkspSize)
511 {
512 HUF_buildCTable_wksp_tables* const wksp_tables = (HUF_buildCTable_wksp_tables*)workSpace;
513 nodeElt* const huffNode0 = wksp_tables->huffNodeTbl;
514 nodeElt* const huffNode = huffNode0+1;
515 int nonNullRank;
516
517 /* safety checks */
518 if (((size_t)workSpace & 3) != 0) return ERROR(GENERIC); /* must be aligned on 4-bytes boundaries */
519 if (wkspSize < sizeof(HUF_buildCTable_wksp_tables))
520 return ERROR(workSpace_tooSmall);
521 if (maxNbBits == 0) maxNbBits = HUF_TABLELOG_DEFAULT;
522 if (maxSymbolValue > HUF_SYMBOLVALUE_MAX)
523 return ERROR(maxSymbolValue_tooLarge);
524 ZSTD_memset(huffNode0, 0, sizeof(huffNodeTable));
525
526 /* sort, decreasing order */
527 HUF_sort(huffNode, count, maxSymbolValue, wksp_tables->rankPosition);
528
529 /* build tree */
530 nonNullRank = HUF_buildTree(huffNode, maxSymbolValue);
531
532 /* enforce maxTableLog */
533 maxNbBits = HUF_setMaxHeight(huffNode, (U32)nonNullRank, maxNbBits);
534 if (maxNbBits > HUF_TABLELOG_MAX) return ERROR(GENERIC); /* check fit into table */
535
536 HUF_buildCTableFromTree(tree, huffNode, nonNullRank, maxSymbolValue, maxNbBits);
537
538 return maxNbBits;
539 }
540
HUF_estimateCompressedSize(const HUF_CElt * CTable,const unsigned * count,unsigned maxSymbolValue)541 size_t HUF_estimateCompressedSize(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue)
542 {
543 size_t nbBits = 0;
544 int s;
545 for (s = 0; s <= (int)maxSymbolValue; ++s) {
546 nbBits += CTable[s].nbBits * count[s];
547 }
548 return nbBits >> 3;
549 }
550
HUF_validateCTable(const HUF_CElt * CTable,const unsigned * count,unsigned maxSymbolValue)551 int HUF_validateCTable(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue) {
552 int bad = 0;
553 int s;
554 for (s = 0; s <= (int)maxSymbolValue; ++s) {
555 bad |= (count[s] != 0) & (CTable[s].nbBits == 0);
556 }
557 return !bad;
558 }
559
HUF_compressBound(size_t size)560 size_t HUF_compressBound(size_t size) { return HUF_COMPRESSBOUND(size); }
561
562 FORCE_INLINE_TEMPLATE void
HUF_encodeSymbol(BIT_CStream_t * bitCPtr,U32 symbol,const HUF_CElt * CTable)563 HUF_encodeSymbol(BIT_CStream_t* bitCPtr, U32 symbol, const HUF_CElt* CTable)
564 {
565 BIT_addBitsFast(bitCPtr, CTable[symbol].val, CTable[symbol].nbBits);
566 }
567
568 #define HUF_FLUSHBITS(s) BIT_flushBits(s)
569
570 #define HUF_FLUSHBITS_1(stream) \
571 if (sizeof((stream)->bitContainer)*8 < HUF_TABLELOG_MAX*2+7) HUF_FLUSHBITS(stream)
572
573 #define HUF_FLUSHBITS_2(stream) \
574 if (sizeof((stream)->bitContainer)*8 < HUF_TABLELOG_MAX*4+7) HUF_FLUSHBITS(stream)
575
576 FORCE_INLINE_TEMPLATE size_t
HUF_compress1X_usingCTable_internal_body(void * dst,size_t dstSize,const void * src,size_t srcSize,const HUF_CElt * CTable)577 HUF_compress1X_usingCTable_internal_body(void* dst, size_t dstSize,
578 const void* src, size_t srcSize,
579 const HUF_CElt* CTable)
580 {
581 const BYTE* ip = (const BYTE*) src;
582 BYTE* const ostart = (BYTE*)dst;
583 BYTE* const oend = ostart + dstSize;
584 BYTE* op = ostart;
585 size_t n;
586 BIT_CStream_t bitC;
587
588 /* init */
589 if (dstSize < 8) return 0; /* not enough space to compress */
590 { size_t const initErr = BIT_initCStream(&bitC, op, (size_t)(oend-op));
591 if (HUF_isError(initErr)) return 0; }
592
593 n = srcSize & ~3; /* join to mod 4 */
594 switch (srcSize & 3)
595 {
596 case 3:
597 HUF_encodeSymbol(&bitC, ip[n+ 2], CTable);
598 HUF_FLUSHBITS_2(&bitC);
599 ZSTD_FALLTHROUGH;
600 case 2:
601 HUF_encodeSymbol(&bitC, ip[n+ 1], CTable);
602 HUF_FLUSHBITS_1(&bitC);
603 ZSTD_FALLTHROUGH;
604 case 1:
605 HUF_encodeSymbol(&bitC, ip[n+ 0], CTable);
606 HUF_FLUSHBITS(&bitC);
607 ZSTD_FALLTHROUGH;
608 case 0: ZSTD_FALLTHROUGH;
609 default: break;
610 }
611
612 for (; n>0; n-=4) { /* note : n&3==0 at this stage */
613 HUF_encodeSymbol(&bitC, ip[n- 1], CTable);
614 HUF_FLUSHBITS_1(&bitC);
615 HUF_encodeSymbol(&bitC, ip[n- 2], CTable);
616 HUF_FLUSHBITS_2(&bitC);
617 HUF_encodeSymbol(&bitC, ip[n- 3], CTable);
618 HUF_FLUSHBITS_1(&bitC);
619 HUF_encodeSymbol(&bitC, ip[n- 4], CTable);
620 HUF_FLUSHBITS(&bitC);
621 }
622
623 return BIT_closeCStream(&bitC);
624 }
625
626 #if DYNAMIC_BMI2
627
628 static TARGET_ATTRIBUTE("bmi2") size_t
HUF_compress1X_usingCTable_internal_bmi2(void * dst,size_t dstSize,const void * src,size_t srcSize,const HUF_CElt * CTable)629 HUF_compress1X_usingCTable_internal_bmi2(void* dst, size_t dstSize,
630 const void* src, size_t srcSize,
631 const HUF_CElt* CTable)
632 {
633 return HUF_compress1X_usingCTable_internal_body(dst, dstSize, src, srcSize, CTable);
634 }
635
636 static size_t
HUF_compress1X_usingCTable_internal_default(void * dst,size_t dstSize,const void * src,size_t srcSize,const HUF_CElt * CTable)637 HUF_compress1X_usingCTable_internal_default(void* dst, size_t dstSize,
638 const void* src, size_t srcSize,
639 const HUF_CElt* CTable)
640 {
641 return HUF_compress1X_usingCTable_internal_body(dst, dstSize, src, srcSize, CTable);
642 }
643
644 static size_t
HUF_compress1X_usingCTable_internal(void * dst,size_t dstSize,const void * src,size_t srcSize,const HUF_CElt * CTable,const int bmi2)645 HUF_compress1X_usingCTable_internal(void* dst, size_t dstSize,
646 const void* src, size_t srcSize,
647 const HUF_CElt* CTable, const int bmi2)
648 {
649 if (bmi2) {
650 return HUF_compress1X_usingCTable_internal_bmi2(dst, dstSize, src, srcSize, CTable);
651 }
652 return HUF_compress1X_usingCTable_internal_default(dst, dstSize, src, srcSize, CTable);
653 }
654
655 #else
656
657 static size_t
HUF_compress1X_usingCTable_internal(void * dst,size_t dstSize,const void * src,size_t srcSize,const HUF_CElt * CTable,const int bmi2)658 HUF_compress1X_usingCTable_internal(void* dst, size_t dstSize,
659 const void* src, size_t srcSize,
660 const HUF_CElt* CTable, const int bmi2)
661 {
662 (void)bmi2;
663 return HUF_compress1X_usingCTable_internal_body(dst, dstSize, src, srcSize, CTable);
664 }
665
666 #endif
667
HUF_compress1X_usingCTable(void * dst,size_t dstSize,const void * src,size_t srcSize,const HUF_CElt * CTable)668 size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable)
669 {
670 return HUF_compress1X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, /* bmi2 */ 0);
671 }
672
673
674 static size_t
HUF_compress4X_usingCTable_internal(void * dst,size_t dstSize,const void * src,size_t srcSize,const HUF_CElt * CTable,int bmi2)675 HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize,
676 const void* src, size_t srcSize,
677 const HUF_CElt* CTable, int bmi2)
678 {
679 size_t const segmentSize = (srcSize+3)/4; /* first 3 segments */
680 const BYTE* ip = (const BYTE*) src;
681 const BYTE* const iend = ip + srcSize;
682 BYTE* const ostart = (BYTE*) dst;
683 BYTE* const oend = ostart + dstSize;
684 BYTE* op = ostart;
685
686 if (dstSize < 6 + 1 + 1 + 1 + 8) return 0; /* minimum space to compress successfully */
687 if (srcSize < 12) return 0; /* no saving possible : too small input */
688 op += 6; /* jumpTable */
689
690 assert(op <= oend);
691 { CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, bmi2) );
692 if (cSize==0) return 0;
693 assert(cSize <= 65535);
694 MEM_writeLE16(ostart, (U16)cSize);
695 op += cSize;
696 }
697
698 ip += segmentSize;
699 assert(op <= oend);
700 { CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, bmi2) );
701 if (cSize==0) return 0;
702 assert(cSize <= 65535);
703 MEM_writeLE16(ostart+2, (U16)cSize);
704 op += cSize;
705 }
706
707 ip += segmentSize;
708 assert(op <= oend);
709 { CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, bmi2) );
710 if (cSize==0) return 0;
711 assert(cSize <= 65535);
712 MEM_writeLE16(ostart+4, (U16)cSize);
713 op += cSize;
714 }
715
716 ip += segmentSize;
717 assert(op <= oend);
718 assert(ip <= iend);
719 { CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, (size_t)(iend-ip), CTable, bmi2) );
720 if (cSize==0) return 0;
721 op += cSize;
722 }
723
724 return (size_t)(op-ostart);
725 }
726
HUF_compress4X_usingCTable(void * dst,size_t dstSize,const void * src,size_t srcSize,const HUF_CElt * CTable)727 size_t HUF_compress4X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable)
728 {
729 return HUF_compress4X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, /* bmi2 */ 0);
730 }
731
732 typedef enum { HUF_singleStream, HUF_fourStreams } HUF_nbStreams_e;
733
HUF_compressCTable_internal(BYTE * const ostart,BYTE * op,BYTE * const oend,const void * src,size_t srcSize,HUF_nbStreams_e nbStreams,const HUF_CElt * CTable,const int bmi2)734 static size_t HUF_compressCTable_internal(
735 BYTE* const ostart, BYTE* op, BYTE* const oend,
736 const void* src, size_t srcSize,
737 HUF_nbStreams_e nbStreams, const HUF_CElt* CTable, const int bmi2)
738 {
739 size_t const cSize = (nbStreams==HUF_singleStream) ?
740 HUF_compress1X_usingCTable_internal(op, (size_t)(oend - op), src, srcSize, CTable, bmi2) :
741 HUF_compress4X_usingCTable_internal(op, (size_t)(oend - op), src, srcSize, CTable, bmi2);
742 if (HUF_isError(cSize)) { return cSize; }
743 if (cSize==0) { return 0; } /* uncompressible */
744 op += cSize;
745 /* check compressibility */
746 assert(op >= ostart);
747 if ((size_t)(op-ostart) >= srcSize-1) { return 0; }
748 return (size_t)(op-ostart);
749 }
750
751 typedef struct {
752 unsigned count[HUF_SYMBOLVALUE_MAX + 1];
753 HUF_CElt CTable[HUF_SYMBOLVALUE_MAX + 1];
754 union {
755 HUF_buildCTable_wksp_tables buildCTable_wksp;
756 HUF_WriteCTableWksp writeCTable_wksp;
757 } wksps;
758 } HUF_compress_tables_t;
759
760 /* HUF_compress_internal() :
761 * `workSpace_align4` must be aligned on 4-bytes boundaries,
762 * and occupies the same space as a table of HUF_WORKSPACE_SIZE_U32 unsigned */
763 static size_t
HUF_compress_internal(void * dst,size_t dstSize,const void * src,size_t srcSize,unsigned maxSymbolValue,unsigned huffLog,HUF_nbStreams_e nbStreams,void * workSpace_align4,size_t wkspSize,HUF_CElt * oldHufTable,HUF_repeat * repeat,int preferRepeat,const int bmi2)764 HUF_compress_internal (void* dst, size_t dstSize,
765 const void* src, size_t srcSize,
766 unsigned maxSymbolValue, unsigned huffLog,
767 HUF_nbStreams_e nbStreams,
768 void* workSpace_align4, size_t wkspSize,
769 HUF_CElt* oldHufTable, HUF_repeat* repeat, int preferRepeat,
770 const int bmi2)
771 {
772 HUF_compress_tables_t* const table = (HUF_compress_tables_t*)workSpace_align4;
773 BYTE* const ostart = (BYTE*)dst;
774 BYTE* const oend = ostart + dstSize;
775 BYTE* op = ostart;
776
777 HUF_STATIC_ASSERT(sizeof(*table) <= HUF_WORKSPACE_SIZE);
778 assert(((size_t)workSpace_align4 & 3) == 0); /* must be aligned on 4-bytes boundaries */
779
780 /* checks & inits */
781 if (wkspSize < HUF_WORKSPACE_SIZE) return ERROR(workSpace_tooSmall);
782 if (!srcSize) return 0; /* Uncompressed */
783 if (!dstSize) return 0; /* cannot fit anything within dst budget */
784 if (srcSize > HUF_BLOCKSIZE_MAX) return ERROR(srcSize_wrong); /* current block size limit */
785 if (huffLog > HUF_TABLELOG_MAX) return ERROR(tableLog_tooLarge);
786 if (maxSymbolValue > HUF_SYMBOLVALUE_MAX) return ERROR(maxSymbolValue_tooLarge);
787 if (!maxSymbolValue) maxSymbolValue = HUF_SYMBOLVALUE_MAX;
788 if (!huffLog) huffLog = HUF_TABLELOG_DEFAULT;
789
790 /* Heuristic : If old table is valid, use it for small inputs */
791 if (preferRepeat && repeat && *repeat == HUF_repeat_valid) {
792 return HUF_compressCTable_internal(ostart, op, oend,
793 src, srcSize,
794 nbStreams, oldHufTable, bmi2);
795 }
796
797 /* Scan input and build symbol stats */
798 { CHECK_V_F(largest, HIST_count_wksp (table->count, &maxSymbolValue, (const BYTE*)src, srcSize, workSpace_align4, wkspSize) );
799 if (largest == srcSize) { *ostart = ((const BYTE*)src)[0]; return 1; } /* single symbol, rle */
800 if (largest <= (srcSize >> 7)+4) return 0; /* heuristic : probably not compressible enough */
801 }
802
803 /* Check validity of previous table */
804 if ( repeat
805 && *repeat == HUF_repeat_check
806 && !HUF_validateCTable(oldHufTable, table->count, maxSymbolValue)) {
807 *repeat = HUF_repeat_none;
808 }
809 /* Heuristic : use existing table for small inputs */
810 if (preferRepeat && repeat && *repeat != HUF_repeat_none) {
811 return HUF_compressCTable_internal(ostart, op, oend,
812 src, srcSize,
813 nbStreams, oldHufTable, bmi2);
814 }
815
816 /* Build Huffman Tree */
817 huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue);
818 { size_t const maxBits = HUF_buildCTable_wksp(table->CTable, table->count,
819 maxSymbolValue, huffLog,
820 &table->wksps.buildCTable_wksp, sizeof(table->wksps.buildCTable_wksp));
821 CHECK_F(maxBits);
822 huffLog = (U32)maxBits;
823 /* Zero unused symbols in CTable, so we can check it for validity */
824 ZSTD_memset(table->CTable + (maxSymbolValue + 1), 0,
825 sizeof(table->CTable) - ((maxSymbolValue + 1) * sizeof(HUF_CElt)));
826 }
827
828 /* Write table description header */
829 { CHECK_V_F(hSize, HUF_writeCTable_wksp(op, dstSize, table->CTable, maxSymbolValue, huffLog,
830 &table->wksps.writeCTable_wksp, sizeof(table->wksps.writeCTable_wksp)) );
831 /* Check if using previous huffman table is beneficial */
832 if (repeat && *repeat != HUF_repeat_none) {
833 size_t const oldSize = HUF_estimateCompressedSize(oldHufTable, table->count, maxSymbolValue);
834 size_t const newSize = HUF_estimateCompressedSize(table->CTable, table->count, maxSymbolValue);
835 if (oldSize <= hSize + newSize || hSize + 12 >= srcSize) {
836 return HUF_compressCTable_internal(ostart, op, oend,
837 src, srcSize,
838 nbStreams, oldHufTable, bmi2);
839 } }
840
841 /* Use the new huffman table */
842 if (hSize + 12ul >= srcSize) { return 0; }
843 op += hSize;
844 if (repeat) { *repeat = HUF_repeat_none; }
845 if (oldHufTable)
846 ZSTD_memcpy(oldHufTable, table->CTable, sizeof(table->CTable)); /* Save new table */
847 }
848 return HUF_compressCTable_internal(ostart, op, oend,
849 src, srcSize,
850 nbStreams, table->CTable, bmi2);
851 }
852
853
HUF_compress1X_wksp(void * dst,size_t dstSize,const void * src,size_t srcSize,unsigned maxSymbolValue,unsigned huffLog,void * workSpace,size_t wkspSize)854 size_t HUF_compress1X_wksp (void* dst, size_t dstSize,
855 const void* src, size_t srcSize,
856 unsigned maxSymbolValue, unsigned huffLog,
857 void* workSpace, size_t wkspSize)
858 {
859 return HUF_compress_internal(dst, dstSize, src, srcSize,
860 maxSymbolValue, huffLog, HUF_singleStream,
861 workSpace, wkspSize,
862 NULL, NULL, 0, 0 /*bmi2*/);
863 }
864
HUF_compress1X_repeat(void * dst,size_t dstSize,const void * src,size_t srcSize,unsigned maxSymbolValue,unsigned huffLog,void * workSpace,size_t wkspSize,HUF_CElt * hufTable,HUF_repeat * repeat,int preferRepeat,int bmi2)865 size_t HUF_compress1X_repeat (void* dst, size_t dstSize,
866 const void* src, size_t srcSize,
867 unsigned maxSymbolValue, unsigned huffLog,
868 void* workSpace, size_t wkspSize,
869 HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2)
870 {
871 return HUF_compress_internal(dst, dstSize, src, srcSize,
872 maxSymbolValue, huffLog, HUF_singleStream,
873 workSpace, wkspSize, hufTable,
874 repeat, preferRepeat, bmi2);
875 }
876
877 /* HUF_compress4X_repeat():
878 * compress input using 4 streams.
879 * provide workspace to generate compression tables */
HUF_compress4X_wksp(void * dst,size_t dstSize,const void * src,size_t srcSize,unsigned maxSymbolValue,unsigned huffLog,void * workSpace,size_t wkspSize)880 size_t HUF_compress4X_wksp (void* dst, size_t dstSize,
881 const void* src, size_t srcSize,
882 unsigned maxSymbolValue, unsigned huffLog,
883 void* workSpace, size_t wkspSize)
884 {
885 return HUF_compress_internal(dst, dstSize, src, srcSize,
886 maxSymbolValue, huffLog, HUF_fourStreams,
887 workSpace, wkspSize,
888 NULL, NULL, 0, 0 /*bmi2*/);
889 }
890
891 /* HUF_compress4X_repeat():
892 * compress input using 4 streams.
893 * re-use an existing huffman compression table */
HUF_compress4X_repeat(void * dst,size_t dstSize,const void * src,size_t srcSize,unsigned maxSymbolValue,unsigned huffLog,void * workSpace,size_t wkspSize,HUF_CElt * hufTable,HUF_repeat * repeat,int preferRepeat,int bmi2)894 size_t HUF_compress4X_repeat (void* dst, size_t dstSize,
895 const void* src, size_t srcSize,
896 unsigned maxSymbolValue, unsigned huffLog,
897 void* workSpace, size_t wkspSize,
898 HUF_CElt* hufTable, HUF_repeat* repeat, int preferRepeat, int bmi2)
899 {
900 return HUF_compress_internal(dst, dstSize, src, srcSize,
901 maxSymbolValue, huffLog, HUF_fourStreams,
902 workSpace, wkspSize,
903 hufTable, repeat, preferRepeat, bmi2);
904 }
905
906