jcs.org
"Das U-Boot" Source Tree
1/* ******************************************************************
2 * Common functions 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* Dependencies
17***************************************/
18#include "mem.h"
19#include "error_private.h" /* ERR_*, ERROR */
20#define FSE_STATIC_LINKING_ONLY /* FSE_MIN_TABLELOG */
21#include "fse.h"
22#define HUF_STATIC_LINKING_ONLY /* HUF_TABLELOG_ABSOLUTEMAX */
23#include "huf.h"
24
25/*=== Version ===*/
26unsigned FSE_versionNumber(void) { return FSE_VERSION_NUMBER; }
27
28/*=== Error Management ===*/
29unsigned FSE_isError(size_t code) { return ERR_isError(code); }
30const char* FSE_getErrorName(size_t code) { return ERR_getErrorName(code); }
31
32unsigned HUF_isError(size_t code) { return ERR_isError(code); }
33const char* HUF_getErrorName(size_t code) { return ERR_getErrorName(code); }
34
35/*-**************************************************************
36* FSE NCount encoding-decoding
37****************************************************************/
38static U32 FSE_ctz(U32 val)
39{
40 assert(val != 0);
41 {
42# if (__GNUC__ >= 3) /* GCC Intrinsic */
43 return __builtin_ctz(val);
44# else /* Software version */
45 U32 count = 0;
46 while ((val & 1) == 0) {
47 val >>= 1;
48 ++count;
49 }
50 return count;
51# endif
52 }
53}
54
55FORCE_INLINE_TEMPLATE
56size_t FSE_readNCount_body(short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
57 const void* headerBuffer, size_t hbSize)
58{
59 const BYTE* const istart = (const BYTE*) headerBuffer;
60 const BYTE* const iend = istart + hbSize;
61 const BYTE* ip = istart;
62 int nbBits;
63 int remaining;
64 int threshold;
65 U32 bitStream;
66 int bitCount;
67 unsigned charnum = 0;
68 unsigned const maxSV1 = *maxSVPtr + 1;
69 int previous0 = 0;
70
71 if (hbSize < 8) {
72 /* This function only works when hbSize >= 8 */
73 char buffer[8] = {0};
74 ZSTD_memcpy(buffer, headerBuffer, hbSize);
75 { size_t const countSize = FSE_readNCount(normalizedCounter, maxSVPtr, tableLogPtr,
76 buffer, sizeof(buffer));
77 if (FSE_isError(countSize)) return countSize;
78 if (countSize > hbSize) return ERROR(corruption_detected);
79 return countSize;
80 } }
81 assert(hbSize >= 8);
82
83 /* init */
84 ZSTD_memset(normalizedCounter, 0, (*maxSVPtr+1) * sizeof(normalizedCounter[0])); /* all symbols not present in NCount have a frequency of 0 */
85 bitStream = MEM_readLE32(ip);
86 nbBits = (bitStream & 0xF) + FSE_MIN_TABLELOG; /* extract tableLog */
87 if (nbBits > FSE_TABLELOG_ABSOLUTE_MAX) return ERROR(tableLog_tooLarge);
88 bitStream >>= 4;
89 bitCount = 4;
90 *tableLogPtr = nbBits;
91 remaining = (1<<nbBits)+1;
92 threshold = 1<<nbBits;
93 nbBits++;
94
95 for (;;) {
96 if (previous0) {
97 /* Count the number of repeats. Each time the
98 * 2-bit repeat code is 0b11 there is another
99 * repeat.
100 * Avoid UB by setting the high bit to 1.
101 */
102 int repeats = FSE_ctz(~bitStream | 0x80000000) >> 1;
103 while (repeats >= 12) {
104 charnum += 3 * 12;
105 if (LIKELY(ip <= iend-7)) {
106 ip += 3;
107 } else {
108 bitCount -= (int)(8 * (iend - 7 - ip));
109 bitCount &= 31;
110 ip = iend - 4;
111 }
112 bitStream = MEM_readLE32(ip) >> bitCount;
113 repeats = FSE_ctz(~bitStream | 0x80000000) >> 1;
114 }
115 charnum += 3 * repeats;
116 bitStream >>= 2 * repeats;
117 bitCount += 2 * repeats;
118
119 /* Add the final repeat which isn't 0b11. */
120 assert((bitStream & 3) < 3);
121 charnum += bitStream & 3;
122 bitCount += 2;
123
124 /* This is an error, but break and return an error
125 * at the end, because returning out of a loop makes
126 * it harder for the compiler to optimize.
127 */
128 if (charnum >= maxSV1) break;
129
130 /* We don't need to set the normalized count to 0
131 * because we already memset the whole buffer to 0.
132 */
133
134 if (LIKELY(ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
135 assert((bitCount >> 3) <= 3); /* For first condition to work */
136 ip += bitCount>>3;
137 bitCount &= 7;
138 } else {
139 bitCount -= (int)(8 * (iend - 4 - ip));
140 bitCount &= 31;
141 ip = iend - 4;
142 }
143 bitStream = MEM_readLE32(ip) >> bitCount;
144 }
145 {
146 int const max = (2*threshold-1) - remaining;
147 int count;
148
149 if ((bitStream & (threshold-1)) < (U32)max) {
150 count = bitStream & (threshold-1);
151 bitCount += nbBits-1;
152 } else {
153 count = bitStream & (2*threshold-1);
154 if (count >= threshold) count -= max;
155 bitCount += nbBits;
156 }
157
158 count--; /* extra accuracy */
159 /* When it matters (small blocks), this is a
160 * predictable branch, because we don't use -1.
161 */
162 if (count >= 0) {
163 remaining -= count;
164 } else {
165 assert(count == -1);
166 remaining += count;
167 }
168 normalizedCounter[charnum++] = (short)count;
169 previous0 = !count;
170
171 assert(threshold > 1);
172 if (remaining < threshold) {
173 /* This branch can be folded into the
174 * threshold update condition because we
175 * know that threshold > 1.
176 */
177 if (remaining <= 1) break;
178 nbBits = BIT_highbit32(remaining) + 1;
179 threshold = 1 << (nbBits - 1);
180 }
181 if (charnum >= maxSV1) break;
182
183 if (LIKELY(ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
184 ip += bitCount>>3;
185 bitCount &= 7;
186 } else {
187 bitCount -= (int)(8 * (iend - 4 - ip));
188 bitCount &= 31;
189 ip = iend - 4;
190 }
191 bitStream = MEM_readLE32(ip) >> bitCount;
192 } }
193 if (remaining != 1) return ERROR(corruption_detected);
194 /* Only possible when there are too many zeros. */
195 if (charnum > maxSV1) return ERROR(maxSymbolValue_tooSmall);
196 if (bitCount > 32) return ERROR(corruption_detected);
197 *maxSVPtr = charnum-1;
198
199 ip += (bitCount+7)>>3;
200 return ip-istart;
201}
202
203/* Avoids the FORCE_INLINE of the _body() function. */
204static size_t FSE_readNCount_body_default(
205 short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
206 const void* headerBuffer, size_t hbSize)
207{
208 return FSE_readNCount_body(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
209}
210
211#if DYNAMIC_BMI2
212BMI2_TARGET_ATTRIBUTE static size_t FSE_readNCount_body_bmi2(
213 short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
214 const void* headerBuffer, size_t hbSize)
215{
216 return FSE_readNCount_body(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
217}
218#endif
219
220size_t FSE_readNCount_bmi2(
221 short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
222 const void* headerBuffer, size_t hbSize, int bmi2)
223{
224#if DYNAMIC_BMI2
225 if (bmi2) {
226 return FSE_readNCount_body_bmi2(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
227 }
228#endif
229 (void)bmi2;
230 return FSE_readNCount_body_default(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
231}
232
233size_t FSE_readNCount(
234 short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
235 const void* headerBuffer, size_t hbSize)
236{
237 return FSE_readNCount_bmi2(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize, /* bmi2 */ 0);
238}
239
240/*! HUF_readStats() :
241 Read compact Huffman tree, saved by HUF_writeCTable().
242 `huffWeight` is destination buffer.
243 `rankStats` is assumed to be a table of at least HUF_TABLELOG_MAX U32.
244 @return : size read from `src` , or an error Code .
245 Note : Needed by HUF_readCTable() and HUF_readDTableX?() .
246*/
247size_t HUF_readStats(BYTE* huffWeight, size_t hwSize, U32* rankStats,
248 U32* nbSymbolsPtr, U32* tableLogPtr,
249 const void* src, size_t srcSize)
250{
251 U32 wksp[HUF_READ_STATS_WORKSPACE_SIZE_U32];
252 return HUF_readStats_wksp(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, wksp, sizeof(wksp), /* bmi2 */ 0);
253}
254
255FORCE_INLINE_TEMPLATE size_t
256HUF_readStats_body(BYTE* huffWeight, size_t hwSize, U32* rankStats,
257 U32* nbSymbolsPtr, U32* tableLogPtr,
258 const void* src, size_t srcSize,
259 void* workSpace, size_t wkspSize,
260 int bmi2)
261{
262 U32 weightTotal;
263 const BYTE* ip = (const BYTE*) src;
264 size_t iSize;
265 size_t oSize;
266
267 if (!srcSize) return ERROR(srcSize_wrong);
268 iSize = ip[0];
269 /* ZSTD_memset(huffWeight, 0, hwSize); *//* is not necessary, even though some analyzer complain ... */
270
271 if (iSize >= 128) { /* special header */
272 oSize = iSize - 127;
273 iSize = ((oSize+1)/2);
274 if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
275 if (oSize >= hwSize) return ERROR(corruption_detected);
276 ip += 1;
277 { U32 n;
278 for (n=0; n<oSize; n+=2) {
279 huffWeight[n] = ip[n/2] >> 4;
280 huffWeight[n+1] = ip[n/2] & 15;
281 } } }
282 else { /* header compressed with FSE (normal case) */
283 if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
284 /* max (hwSize-1) values decoded, as last one is implied */
285 oSize = FSE_decompress_wksp_bmi2(huffWeight, hwSize-1, ip+1, iSize, 6, workSpace, wkspSize, bmi2);
286 if (FSE_isError(oSize)) return oSize;
287 }
288
289 /* collect weight stats */
290 ZSTD_memset(rankStats, 0, (HUF_TABLELOG_MAX + 1) * sizeof(U32));
291 weightTotal = 0;
292 { U32 n; for (n=0; n<oSize; n++) {
293 if (huffWeight[n] > HUF_TABLELOG_MAX) return ERROR(corruption_detected);
294 rankStats[huffWeight[n]]++;
295 weightTotal += (1 << huffWeight[n]) >> 1;
296 } }
297 if (weightTotal == 0) return ERROR(corruption_detected);
298
299 /* get last non-null symbol weight (implied, total must be 2^n) */
300 { U32 const tableLog = BIT_highbit32(weightTotal) + 1;
301 if (tableLog > HUF_TABLELOG_MAX) return ERROR(corruption_detected);
302 *tableLogPtr = tableLog;
303 /* determine last weight */
304 { U32 const total = 1 << tableLog;
305 U32 const rest = total - weightTotal;
306 U32 const verif = 1 << BIT_highbit32(rest);
307 U32 const lastWeight = BIT_highbit32(rest) + 1;
308 if (verif != rest) return ERROR(corruption_detected); /* last value must be a clean power of 2 */
309 huffWeight[oSize] = (BYTE)lastWeight;
310 rankStats[lastWeight]++;
311 } }
312
313 /* check tree construction validity */
314 if ((rankStats[1] < 2) || (rankStats[1] & 1)) return ERROR(corruption_detected); /* by construction : at least 2 elts of rank 1, must be even */
315
316 /* results */
317 *nbSymbolsPtr = (U32)(oSize+1);
318 return iSize+1;
319}
320
321/* Avoids the FORCE_INLINE of the _body() function. */
322static size_t HUF_readStats_body_default(BYTE* huffWeight, size_t hwSize, U32* rankStats,
323 U32* nbSymbolsPtr, U32* tableLogPtr,
324 const void* src, size_t srcSize,
325 void* workSpace, size_t wkspSize)
326{
327 return HUF_readStats_body(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize, 0);
328}
329
330#if DYNAMIC_BMI2
331static BMI2_TARGET_ATTRIBUTE size_t HUF_readStats_body_bmi2(BYTE* huffWeight, size_t hwSize, U32* rankStats,
332 U32* nbSymbolsPtr, U32* tableLogPtr,
333 const void* src, size_t srcSize,
334 void* workSpace, size_t wkspSize)
335{
336 return HUF_readStats_body(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize, 1);
337}
338#endif
339
340size_t HUF_readStats_wksp(BYTE* huffWeight, size_t hwSize, U32* rankStats,
341 U32* nbSymbolsPtr, U32* tableLogPtr,
342 const void* src, size_t srcSize,
343 void* workSpace, size_t wkspSize,
344 int bmi2)
345{
346#if DYNAMIC_BMI2
347 if (bmi2) {
348 return HUF_readStats_body_bmi2(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize);
349 }
350#endif
351 (void)bmi2;
352 return HUF_readStats_body_default(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize);
353}