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
26/*===   Version   ===*/
27unsigned FSE_versionNumber(void) { return FSE_VERSION_NUMBER; }
28
29
30/*===   Error Management   ===*/
31unsigned FSE_isError(size_t code) { return ERR_isError(code); }
32const char* FSE_getErrorName(size_t code) { return ERR_getErrorName(code); }
33
34unsigned HUF_isError(size_t code) { return ERR_isError(code); }
35const char* HUF_getErrorName(size_t code) { return ERR_getErrorName(code); }
36
37
38/*-**************************************************************
39*  FSE NCount encoding-decoding
40****************************************************************/
41static U32 FSE_ctz(U32 val)
42{
43    assert(val != 0);
44    {
45#   if defined(_MSC_VER)   /* Visual */
46        if (val != 0) {
47            unsigned long r;
48            _BitScanForward(&r, val);
49            return (unsigned)r;
50        } else {
51            /* Should not reach this code path */
52            __assume(0);
53        }
54#   elif defined(__GNUC__) && (__GNUC__ >= 3)   /* GCC Intrinsic */
55        return __builtin_ctz(val);
56#   elif defined(__ICCARM__)    /* IAR Intrinsic */
57        return __CTZ(val);
58#   else   /* Software version */
59        U32 count = 0;
60        while ((val & 1) == 0) {
61            val >>= 1;
62            ++count;
63        }
64        return count;
65#   endif
66    }
67}
68
69FORCE_INLINE_TEMPLATE
70size_t FSE_readNCount_body(short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
71                           const void* headerBuffer, size_t hbSize)
72{
73    const BYTE* const istart = (const BYTE*) headerBuffer;
74    const BYTE* const iend = istart + hbSize;
75    const BYTE* ip = istart;
76    int nbBits;
77    int remaining;
78    int threshold;
79    U32 bitStream;
80    int bitCount;
81    unsigned charnum = 0;
82    unsigned const maxSV1 = *maxSVPtr + 1;
83    int previous0 = 0;
84
85    if (hbSize < 8) {
86        /* This function only works when hbSize >= 8 */
87        char buffer[8] = {0};
88        ZSTD_memcpy(buffer, headerBuffer, hbSize);
89        {   size_t const countSize = FSE_readNCount(normalizedCounter, maxSVPtr, tableLogPtr,
90                                                    buffer, sizeof(buffer));
91            if (FSE_isError(countSize)) return countSize;
92            if (countSize > hbSize) return ERROR(corruption_detected);
93            return countSize;
94    }   }
95    assert(hbSize >= 8);
96
97    /* init */
98    ZSTD_memset(normalizedCounter, 0, (*maxSVPtr+1) * sizeof(normalizedCounter[0]));   /* all symbols not present in NCount have a frequency of 0 */
99    bitStream = MEM_readLE32(ip);
100    nbBits = (bitStream & 0xF) + FSE_MIN_TABLELOG;   /* extract tableLog */
101    if (nbBits > FSE_TABLELOG_ABSOLUTE_MAX) return ERROR(tableLog_tooLarge);
102    bitStream >>= 4;
103    bitCount = 4;
104    *tableLogPtr = nbBits;
105    remaining = (1<<nbBits)+1;
106    threshold = 1<<nbBits;
107    nbBits++;
108
109    for (;;) {
110        if (previous0) {
111            /* Count the number of repeats. Each time the
112             * 2-bit repeat code is 0b11 there is another
113             * repeat.
114             * Avoid UB by setting the high bit to 1.
115             */
116            int repeats = FSE_ctz(~bitStream | 0x80000000) >> 1;
117            while (repeats >= 12) {
118                charnum += 3 * 12;
119                if (LIKELY(ip <= iend-7)) {
120                    ip += 3;
121                } else {
122                    bitCount -= (int)(8 * (iend - 7 - ip));
123                    bitCount &= 31;
124                    ip = iend - 4;
125                }
126                bitStream = MEM_readLE32(ip) >> bitCount;
127                repeats = FSE_ctz(~bitStream | 0x80000000) >> 1;
128            }
129            charnum += 3 * repeats;
130            bitStream >>= 2 * repeats;
131            bitCount += 2 * repeats;
132
133            /* Add the final repeat which isn't 0b11. */
134            assert((bitStream & 3) < 3);
135            charnum += bitStream & 3;
136            bitCount += 2;
137
138            /* This is an error, but break and return an error
139             * at the end, because returning out of a loop makes
140             * it harder for the compiler to optimize.
141             */
142            if (charnum >= maxSV1) break;
143
144            /* We don't need to set the normalized count to 0
145             * because we already memset the whole buffer to 0.
146             */
147
148            if (LIKELY(ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
149                assert((bitCount >> 3) <= 3); /* For first condition to work */
150                ip += bitCount>>3;
151                bitCount &= 7;
152            } else {
153                bitCount -= (int)(8 * (iend - 4 - ip));
154                bitCount &= 31;
155                ip = iend - 4;
156            }
157            bitStream = MEM_readLE32(ip) >> bitCount;
158        }
159        {
160            int const max = (2*threshold-1) - remaining;
161            int count;
162
163            if ((bitStream & (threshold-1)) < (U32)max) {
164                count = bitStream & (threshold-1);
165                bitCount += nbBits-1;
166            } else {
167                count = bitStream & (2*threshold-1);
168                if (count >= threshold) count -= max;
169                bitCount += nbBits;
170            }
171
172            count--;   /* extra accuracy */
173            /* When it matters (small blocks), this is a
174             * predictable branch, because we don't use -1.
175             */
176            if (count >= 0) {
177                remaining -= count;
178            } else {
179                assert(count == -1);
180                remaining += count;
181            }
182            normalizedCounter[charnum++] = (short)count;
183            previous0 = !count;
184
185            assert(threshold > 1);
186            if (remaining < threshold) {
187                /* This branch can be folded into the
188                 * threshold update condition because we
189                 * know that threshold > 1.
190                 */
191                if (remaining <= 1) break;
192                nbBits = BIT_highbit32(remaining) + 1;
193                threshold = 1 << (nbBits - 1);
194            }
195            if (charnum >= maxSV1) break;
196
197            if (LIKELY(ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
198                ip += bitCount>>3;
199                bitCount &= 7;
200            } else {
201                bitCount -= (int)(8 * (iend - 4 - ip));
202                bitCount &= 31;
203                ip = iend - 4;
204            }
205            bitStream = MEM_readLE32(ip) >> bitCount;
206    }   }
207    if (remaining != 1) return ERROR(corruption_detected);
208    /* Only possible when there are too many zeros. */
209    if (charnum > maxSV1) return ERROR(maxSymbolValue_tooSmall);
210    if (bitCount > 32) return ERROR(corruption_detected);
211    *maxSVPtr = charnum-1;
212
213    ip += (bitCount+7)>>3;
214    return ip-istart;
215}
216
217/* Avoids the FORCE_INLINE of the _body() function. */
218static size_t FSE_readNCount_body_default(
219        short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
220        const void* headerBuffer, size_t hbSize)
221{
222    return FSE_readNCount_body(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
223}
224
225#if DYNAMIC_BMI2
226BMI2_TARGET_ATTRIBUTE static size_t FSE_readNCount_body_bmi2(
227        short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
228        const void* headerBuffer, size_t hbSize)
229{
230    return FSE_readNCount_body(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
231}
232#endif
233
234size_t FSE_readNCount_bmi2(
235        short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
236        const void* headerBuffer, size_t hbSize, int bmi2)
237{
238#if DYNAMIC_BMI2
239    if (bmi2) {
240        return FSE_readNCount_body_bmi2(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
241    }
242#endif
243    (void)bmi2;
244    return FSE_readNCount_body_default(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
245}
246
247size_t FSE_readNCount(
248        short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
249        const void* headerBuffer, size_t hbSize)
250{
251    return FSE_readNCount_bmi2(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize, /* bmi2 */ 0);
252}
253
254
255/*! HUF_readStats() :
256    Read compact Huffman tree, saved by HUF_writeCTable().
257    `huffWeight` is destination buffer.
258    `rankStats` is assumed to be a table of at least HUF_TABLELOG_MAX U32.
259    @return : size read from `src` , or an error Code .
260    Note : Needed by HUF_readCTable() and HUF_readDTableX?() .
261*/
262size_t HUF_readStats(BYTE* huffWeight, size_t hwSize, U32* rankStats,
263                     U32* nbSymbolsPtr, U32* tableLogPtr,
264                     const void* src, size_t srcSize)
265{
266    U32 wksp[HUF_READ_STATS_WORKSPACE_SIZE_U32];
267    return HUF_readStats_wksp(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, wksp, sizeof(wksp), /* bmi2 */ 0);
268}
269
270FORCE_INLINE_TEMPLATE size_t
271HUF_readStats_body(BYTE* huffWeight, size_t hwSize, U32* rankStats,
272                   U32* nbSymbolsPtr, U32* tableLogPtr,
273                   const void* src, size_t srcSize,
274                   void* workSpace, size_t wkspSize,
275                   int bmi2)
276{
277    U32 weightTotal;
278    const BYTE* ip = (const BYTE*) src;
279    size_t iSize;
280    size_t oSize;
281
282    if (!srcSize) return ERROR(srcSize_wrong);
283    iSize = ip[0];
284    /* ZSTD_memset(huffWeight, 0, hwSize);   *//* is not necessary, even though some analyzer complain ... */
285
286    if (iSize >= 128) {  /* special header */
287        oSize = iSize - 127;
288        iSize = ((oSize+1)/2);
289        if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
290        if (oSize >= hwSize) return ERROR(corruption_detected);
291        ip += 1;
292        {   U32 n;
293            for (n=0; n<oSize; n+=2) {
294                huffWeight[n]   = ip[n/2] >> 4;
295                huffWeight[n+1] = ip[n/2] & 15;
296    }   }   }
297    else  {   /* header compressed with FSE (normal case) */
298        if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
299        /* max (hwSize-1) values decoded, as last one is implied */
300        oSize = FSE_decompress_wksp_bmi2(huffWeight, hwSize-1, ip+1, iSize, 6, workSpace, wkspSize, bmi2);
301        if (FSE_isError(oSize)) return oSize;
302    }
303
304    /* collect weight stats */
305    ZSTD_memset(rankStats, 0, (HUF_TABLELOG_MAX + 1) * sizeof(U32));
306    weightTotal = 0;
307    {   U32 n; for (n=0; n<oSize; n++) {
308            if (huffWeight[n] > HUF_TABLELOG_MAX) return ERROR(corruption_detected);
309            rankStats[huffWeight[n]]++;
310            weightTotal += (1 << huffWeight[n]) >> 1;
311    }   }
312    if (weightTotal == 0) return ERROR(corruption_detected);
313
314    /* get last non-null symbol weight (implied, total must be 2^n) */
315    {   U32 const tableLog = BIT_highbit32(weightTotal) + 1;
316        if (tableLog > HUF_TABLELOG_MAX) return ERROR(corruption_detected);
317        *tableLogPtr = tableLog;
318        /* determine last weight */
319        {   U32 const total = 1 << tableLog;
320            U32 const rest = total - weightTotal;
321            U32 const verif = 1 << BIT_highbit32(rest);
322            U32 const lastWeight = BIT_highbit32(rest) + 1;
323            if (verif != rest) return ERROR(corruption_detected);    /* last value must be a clean power of 2 */
324            huffWeight[oSize] = (BYTE)lastWeight;
325            rankStats[lastWeight]++;
326    }   }
327
328    /* check tree construction validity */
329    if ((rankStats[1] < 2) || (rankStats[1] & 1)) return ERROR(corruption_detected);   /* by construction : at least 2 elts of rank 1, must be even */
330
331    /* results */
332    *nbSymbolsPtr = (U32)(oSize+1);
333    return iSize+1;
334}
335
336/* Avoids the FORCE_INLINE of the _body() function. */
337static size_t HUF_readStats_body_default(BYTE* huffWeight, size_t hwSize, U32* rankStats,
338                     U32* nbSymbolsPtr, U32* tableLogPtr,
339                     const void* src, size_t srcSize,
340                     void* workSpace, size_t wkspSize)
341{
342    return HUF_readStats_body(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize, 0);
343}
344
345#if DYNAMIC_BMI2
346static BMI2_TARGET_ATTRIBUTE size_t HUF_readStats_body_bmi2(BYTE* huffWeight, size_t hwSize, U32* rankStats,
347                     U32* nbSymbolsPtr, U32* tableLogPtr,
348                     const void* src, size_t srcSize,
349                     void* workSpace, size_t wkspSize)
350{
351    return HUF_readStats_body(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize, 1);
352}
353#endif
354
355size_t HUF_readStats_wksp(BYTE* huffWeight, size_t hwSize, U32* rankStats,
356                     U32* nbSymbolsPtr, U32* tableLogPtr,
357                     const void* src, size_t srcSize,
358                     void* workSpace, size_t wkspSize,
359                     int bmi2)
360{
361#if DYNAMIC_BMI2
362    if (bmi2) {
363        return HUF_readStats_body_bmi2(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize);
364    }
365#endif
366    (void)bmi2;
367    return HUF_readStats_body_default(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize);
368}
369