xref: /dflybsd-src/contrib/zstd/lib/common/fse_decompress.c (revision a28cd43d19e8b720a6c852a4bbc5ae147a26165a)
1*a28cd43dSSascha Wildner /* ******************************************************************
2*a28cd43dSSascha Wildner  * FSE : Finite State Entropy decoder
3*a28cd43dSSascha Wildner  * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
4*a28cd43dSSascha Wildner  *
5*a28cd43dSSascha Wildner  *  You can contact the author at :
6*a28cd43dSSascha Wildner  *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
7*a28cd43dSSascha Wildner  *  - Public forum : https://groups.google.com/forum/#!forum/lz4c
8*a28cd43dSSascha Wildner  *
9*a28cd43dSSascha Wildner  * This source code is licensed under both the BSD-style license (found in the
10*a28cd43dSSascha Wildner  * LICENSE file in the root directory of this source tree) and the GPLv2 (found
11*a28cd43dSSascha Wildner  * in the COPYING file in the root directory of this source tree).
12*a28cd43dSSascha Wildner  * You may select, at your option, one of the above-listed licenses.
13*a28cd43dSSascha Wildner ****************************************************************** */
14*a28cd43dSSascha Wildner 
15*a28cd43dSSascha Wildner 
16*a28cd43dSSascha Wildner /* **************************************************************
17*a28cd43dSSascha Wildner *  Includes
18*a28cd43dSSascha Wildner ****************************************************************/
19*a28cd43dSSascha Wildner #include "debug.h"      /* assert */
20*a28cd43dSSascha Wildner #include "bitstream.h"
21*a28cd43dSSascha Wildner #include "compiler.h"
22*a28cd43dSSascha Wildner #define FSE_STATIC_LINKING_ONLY
23*a28cd43dSSascha Wildner #include "fse.h"
24*a28cd43dSSascha Wildner #include "error_private.h"
25*a28cd43dSSascha Wildner #define ZSTD_DEPS_NEED_MALLOC
26*a28cd43dSSascha Wildner #include "zstd_deps.h"
27*a28cd43dSSascha Wildner 
28*a28cd43dSSascha Wildner 
29*a28cd43dSSascha Wildner /* **************************************************************
30*a28cd43dSSascha Wildner *  Error Management
31*a28cd43dSSascha Wildner ****************************************************************/
32*a28cd43dSSascha Wildner #define FSE_isError ERR_isError
33*a28cd43dSSascha Wildner #define FSE_STATIC_ASSERT(c) DEBUG_STATIC_ASSERT(c)   /* use only *after* variable declarations */
34*a28cd43dSSascha Wildner 
35*a28cd43dSSascha Wildner 
36*a28cd43dSSascha Wildner /* **************************************************************
37*a28cd43dSSascha Wildner *  Templates
38*a28cd43dSSascha Wildner ****************************************************************/
39*a28cd43dSSascha Wildner /*
40*a28cd43dSSascha Wildner   designed to be included
41*a28cd43dSSascha Wildner   for type-specific functions (template emulation in C)
42*a28cd43dSSascha Wildner   Objective is to write these functions only once, for improved maintenance
43*a28cd43dSSascha Wildner */
44*a28cd43dSSascha Wildner 
45*a28cd43dSSascha Wildner /* safety checks */
46*a28cd43dSSascha Wildner #ifndef FSE_FUNCTION_EXTENSION
47*a28cd43dSSascha Wildner #  error "FSE_FUNCTION_EXTENSION must be defined"
48*a28cd43dSSascha Wildner #endif
49*a28cd43dSSascha Wildner #ifndef FSE_FUNCTION_TYPE
50*a28cd43dSSascha Wildner #  error "FSE_FUNCTION_TYPE must be defined"
51*a28cd43dSSascha Wildner #endif
52*a28cd43dSSascha Wildner 
53*a28cd43dSSascha Wildner /* Function names */
54*a28cd43dSSascha Wildner #define FSE_CAT(X,Y) X##Y
55*a28cd43dSSascha Wildner #define FSE_FUNCTION_NAME(X,Y) FSE_CAT(X,Y)
56*a28cd43dSSascha Wildner #define FSE_TYPE_NAME(X,Y) FSE_CAT(X,Y)
57*a28cd43dSSascha Wildner 
58*a28cd43dSSascha Wildner 
59*a28cd43dSSascha Wildner /* Function templates */
FSE_createDTable(unsigned tableLog)60*a28cd43dSSascha Wildner FSE_DTable* FSE_createDTable (unsigned tableLog)
61*a28cd43dSSascha Wildner {
62*a28cd43dSSascha Wildner     if (tableLog > FSE_TABLELOG_ABSOLUTE_MAX) tableLog = FSE_TABLELOG_ABSOLUTE_MAX;
63*a28cd43dSSascha Wildner     return (FSE_DTable*)ZSTD_malloc( FSE_DTABLE_SIZE_U32(tableLog) * sizeof (U32) );
64*a28cd43dSSascha Wildner }
65*a28cd43dSSascha Wildner 
FSE_freeDTable(FSE_DTable * dt)66*a28cd43dSSascha Wildner void FSE_freeDTable (FSE_DTable* dt)
67*a28cd43dSSascha Wildner {
68*a28cd43dSSascha Wildner     ZSTD_free(dt);
69*a28cd43dSSascha Wildner }
70*a28cd43dSSascha Wildner 
FSE_buildDTable_internal(FSE_DTable * dt,const short * normalizedCounter,unsigned maxSymbolValue,unsigned tableLog,void * workSpace,size_t wkspSize)71*a28cd43dSSascha Wildner static size_t FSE_buildDTable_internal(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize)
72*a28cd43dSSascha Wildner {
73*a28cd43dSSascha Wildner     void* const tdPtr = dt+1;   /* because *dt is unsigned, 32-bits aligned on 32-bits */
74*a28cd43dSSascha Wildner     FSE_DECODE_TYPE* const tableDecode = (FSE_DECODE_TYPE*) (tdPtr);
75*a28cd43dSSascha Wildner     U16* symbolNext = (U16*)workSpace;
76*a28cd43dSSascha Wildner     BYTE* spread = (BYTE*)(symbolNext + maxSymbolValue + 1);
77*a28cd43dSSascha Wildner 
78*a28cd43dSSascha Wildner     U32 const maxSV1 = maxSymbolValue + 1;
79*a28cd43dSSascha Wildner     U32 const tableSize = 1 << tableLog;
80*a28cd43dSSascha Wildner     U32 highThreshold = tableSize-1;
81*a28cd43dSSascha Wildner 
82*a28cd43dSSascha Wildner     /* Sanity Checks */
83*a28cd43dSSascha Wildner     if (FSE_BUILD_DTABLE_WKSP_SIZE(tableLog, maxSymbolValue) > wkspSize) return ERROR(maxSymbolValue_tooLarge);
84*a28cd43dSSascha Wildner     if (maxSymbolValue > FSE_MAX_SYMBOL_VALUE) return ERROR(maxSymbolValue_tooLarge);
85*a28cd43dSSascha Wildner     if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);
86*a28cd43dSSascha Wildner 
87*a28cd43dSSascha Wildner     /* Init, lay down lowprob symbols */
88*a28cd43dSSascha Wildner     {   FSE_DTableHeader DTableH;
89*a28cd43dSSascha Wildner         DTableH.tableLog = (U16)tableLog;
90*a28cd43dSSascha Wildner         DTableH.fastMode = 1;
91*a28cd43dSSascha Wildner         {   S16 const largeLimit= (S16)(1 << (tableLog-1));
92*a28cd43dSSascha Wildner             U32 s;
93*a28cd43dSSascha Wildner             for (s=0; s<maxSV1; s++) {
94*a28cd43dSSascha Wildner                 if (normalizedCounter[s]==-1) {
95*a28cd43dSSascha Wildner                     tableDecode[highThreshold--].symbol = (FSE_FUNCTION_TYPE)s;
96*a28cd43dSSascha Wildner                     symbolNext[s] = 1;
97*a28cd43dSSascha Wildner                 } else {
98*a28cd43dSSascha Wildner                     if (normalizedCounter[s] >= largeLimit) DTableH.fastMode=0;
99*a28cd43dSSascha Wildner                     symbolNext[s] = normalizedCounter[s];
100*a28cd43dSSascha Wildner         }   }   }
101*a28cd43dSSascha Wildner         ZSTD_memcpy(dt, &DTableH, sizeof(DTableH));
102*a28cd43dSSascha Wildner     }
103*a28cd43dSSascha Wildner 
104*a28cd43dSSascha Wildner     /* Spread symbols */
105*a28cd43dSSascha Wildner     if (highThreshold == tableSize - 1) {
106*a28cd43dSSascha Wildner         size_t const tableMask = tableSize-1;
107*a28cd43dSSascha Wildner         size_t const step = FSE_TABLESTEP(tableSize);
108*a28cd43dSSascha Wildner         /* First lay down the symbols in order.
109*a28cd43dSSascha Wildner          * We use a uint64_t to lay down 8 bytes at a time. This reduces branch
110*a28cd43dSSascha Wildner          * misses since small blocks generally have small table logs, so nearly
111*a28cd43dSSascha Wildner          * all symbols have counts <= 8. We ensure we have 8 bytes at the end of
112*a28cd43dSSascha Wildner          * our buffer to handle the over-write.
113*a28cd43dSSascha Wildner          */
114*a28cd43dSSascha Wildner         {
115*a28cd43dSSascha Wildner             U64 const add = 0x0101010101010101ull;
116*a28cd43dSSascha Wildner             size_t pos = 0;
117*a28cd43dSSascha Wildner             U64 sv = 0;
118*a28cd43dSSascha Wildner             U32 s;
119*a28cd43dSSascha Wildner             for (s=0; s<maxSV1; ++s, sv += add) {
120*a28cd43dSSascha Wildner                 int i;
121*a28cd43dSSascha Wildner                 int const n = normalizedCounter[s];
122*a28cd43dSSascha Wildner                 MEM_write64(spread + pos, sv);
123*a28cd43dSSascha Wildner                 for (i = 8; i < n; i += 8) {
124*a28cd43dSSascha Wildner                     MEM_write64(spread + pos + i, sv);
125*a28cd43dSSascha Wildner                 }
126*a28cd43dSSascha Wildner                 pos += n;
127*a28cd43dSSascha Wildner             }
128*a28cd43dSSascha Wildner         }
129*a28cd43dSSascha Wildner         /* Now we spread those positions across the table.
130*a28cd43dSSascha Wildner          * The benefit of doing it in two stages is that we avoid the the
131*a28cd43dSSascha Wildner          * variable size inner loop, which caused lots of branch misses.
132*a28cd43dSSascha Wildner          * Now we can run through all the positions without any branch misses.
133*a28cd43dSSascha Wildner          * We unroll the loop twice, since that is what emperically worked best.
134*a28cd43dSSascha Wildner          */
135*a28cd43dSSascha Wildner         {
136*a28cd43dSSascha Wildner             size_t position = 0;
137*a28cd43dSSascha Wildner             size_t s;
138*a28cd43dSSascha Wildner             size_t const unroll = 2;
139*a28cd43dSSascha Wildner             assert(tableSize % unroll == 0); /* FSE_MIN_TABLELOG is 5 */
140*a28cd43dSSascha Wildner             for (s = 0; s < (size_t)tableSize; s += unroll) {
141*a28cd43dSSascha Wildner                 size_t u;
142*a28cd43dSSascha Wildner                 for (u = 0; u < unroll; ++u) {
143*a28cd43dSSascha Wildner                     size_t const uPosition = (position + (u * step)) & tableMask;
144*a28cd43dSSascha Wildner                     tableDecode[uPosition].symbol = spread[s + u];
145*a28cd43dSSascha Wildner                 }
146*a28cd43dSSascha Wildner                 position = (position + (unroll * step)) & tableMask;
147*a28cd43dSSascha Wildner             }
148*a28cd43dSSascha Wildner             assert(position == 0);
149*a28cd43dSSascha Wildner         }
150*a28cd43dSSascha Wildner     } else {
151*a28cd43dSSascha Wildner         U32 const tableMask = tableSize-1;
152*a28cd43dSSascha Wildner         U32 const step = FSE_TABLESTEP(tableSize);
153*a28cd43dSSascha Wildner         U32 s, position = 0;
154*a28cd43dSSascha Wildner         for (s=0; s<maxSV1; s++) {
155*a28cd43dSSascha Wildner             int i;
156*a28cd43dSSascha Wildner             for (i=0; i<normalizedCounter[s]; i++) {
157*a28cd43dSSascha Wildner                 tableDecode[position].symbol = (FSE_FUNCTION_TYPE)s;
158*a28cd43dSSascha Wildner                 position = (position + step) & tableMask;
159*a28cd43dSSascha Wildner                 while (position > highThreshold) position = (position + step) & tableMask;   /* lowprob area */
160*a28cd43dSSascha Wildner         }   }
161*a28cd43dSSascha Wildner         if (position!=0) return ERROR(GENERIC);   /* position must reach all cells once, otherwise normalizedCounter is incorrect */
162*a28cd43dSSascha Wildner     }
163*a28cd43dSSascha Wildner 
164*a28cd43dSSascha Wildner     /* Build Decoding table */
165*a28cd43dSSascha Wildner     {   U32 u;
166*a28cd43dSSascha Wildner         for (u=0; u<tableSize; u++) {
167*a28cd43dSSascha Wildner             FSE_FUNCTION_TYPE const symbol = (FSE_FUNCTION_TYPE)(tableDecode[u].symbol);
168*a28cd43dSSascha Wildner             U32 const nextState = symbolNext[symbol]++;
169*a28cd43dSSascha Wildner             tableDecode[u].nbBits = (BYTE) (tableLog - BIT_highbit32(nextState) );
170*a28cd43dSSascha Wildner             tableDecode[u].newState = (U16) ( (nextState << tableDecode[u].nbBits) - tableSize);
171*a28cd43dSSascha Wildner     }   }
172*a28cd43dSSascha Wildner 
173*a28cd43dSSascha Wildner     return 0;
174*a28cd43dSSascha Wildner }
175*a28cd43dSSascha Wildner 
FSE_buildDTable_wksp(FSE_DTable * dt,const short * normalizedCounter,unsigned maxSymbolValue,unsigned tableLog,void * workSpace,size_t wkspSize)176*a28cd43dSSascha Wildner size_t FSE_buildDTable_wksp(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize)
177*a28cd43dSSascha Wildner {
178*a28cd43dSSascha Wildner     return FSE_buildDTable_internal(dt, normalizedCounter, maxSymbolValue, tableLog, workSpace, wkspSize);
179*a28cd43dSSascha Wildner }
180*a28cd43dSSascha Wildner 
181*a28cd43dSSascha Wildner 
182*a28cd43dSSascha Wildner #ifndef FSE_COMMONDEFS_ONLY
183*a28cd43dSSascha Wildner 
184*a28cd43dSSascha Wildner /*-*******************************************************
185*a28cd43dSSascha Wildner *  Decompression (Byte symbols)
186*a28cd43dSSascha Wildner *********************************************************/
FSE_buildDTable_rle(FSE_DTable * dt,BYTE symbolValue)187*a28cd43dSSascha Wildner size_t FSE_buildDTable_rle (FSE_DTable* dt, BYTE symbolValue)
188*a28cd43dSSascha Wildner {
189*a28cd43dSSascha Wildner     void* ptr = dt;
190*a28cd43dSSascha Wildner     FSE_DTableHeader* const DTableH = (FSE_DTableHeader*)ptr;
191*a28cd43dSSascha Wildner     void* dPtr = dt + 1;
192*a28cd43dSSascha Wildner     FSE_decode_t* const cell = (FSE_decode_t*)dPtr;
193*a28cd43dSSascha Wildner 
194*a28cd43dSSascha Wildner     DTableH->tableLog = 0;
195*a28cd43dSSascha Wildner     DTableH->fastMode = 0;
196*a28cd43dSSascha Wildner 
197*a28cd43dSSascha Wildner     cell->newState = 0;
198*a28cd43dSSascha Wildner     cell->symbol = symbolValue;
199*a28cd43dSSascha Wildner     cell->nbBits = 0;
200*a28cd43dSSascha Wildner 
201*a28cd43dSSascha Wildner     return 0;
202*a28cd43dSSascha Wildner }
203*a28cd43dSSascha Wildner 
204*a28cd43dSSascha Wildner 
FSE_buildDTable_raw(FSE_DTable * dt,unsigned nbBits)205*a28cd43dSSascha Wildner size_t FSE_buildDTable_raw (FSE_DTable* dt, unsigned nbBits)
206*a28cd43dSSascha Wildner {
207*a28cd43dSSascha Wildner     void* ptr = dt;
208*a28cd43dSSascha Wildner     FSE_DTableHeader* const DTableH = (FSE_DTableHeader*)ptr;
209*a28cd43dSSascha Wildner     void* dPtr = dt + 1;
210*a28cd43dSSascha Wildner     FSE_decode_t* const dinfo = (FSE_decode_t*)dPtr;
211*a28cd43dSSascha Wildner     const unsigned tableSize = 1 << nbBits;
212*a28cd43dSSascha Wildner     const unsigned tableMask = tableSize - 1;
213*a28cd43dSSascha Wildner     const unsigned maxSV1 = tableMask+1;
214*a28cd43dSSascha Wildner     unsigned s;
215*a28cd43dSSascha Wildner 
216*a28cd43dSSascha Wildner     /* Sanity checks */
217*a28cd43dSSascha Wildner     if (nbBits < 1) return ERROR(GENERIC);         /* min size */
218*a28cd43dSSascha Wildner 
219*a28cd43dSSascha Wildner     /* Build Decoding Table */
220*a28cd43dSSascha Wildner     DTableH->tableLog = (U16)nbBits;
221*a28cd43dSSascha Wildner     DTableH->fastMode = 1;
222*a28cd43dSSascha Wildner     for (s=0; s<maxSV1; s++) {
223*a28cd43dSSascha Wildner         dinfo[s].newState = 0;
224*a28cd43dSSascha Wildner         dinfo[s].symbol = (BYTE)s;
225*a28cd43dSSascha Wildner         dinfo[s].nbBits = (BYTE)nbBits;
226*a28cd43dSSascha Wildner     }
227*a28cd43dSSascha Wildner 
228*a28cd43dSSascha Wildner     return 0;
229*a28cd43dSSascha Wildner }
230*a28cd43dSSascha Wildner 
FSE_decompress_usingDTable_generic(void * dst,size_t maxDstSize,const void * cSrc,size_t cSrcSize,const FSE_DTable * dt,const unsigned fast)231*a28cd43dSSascha Wildner FORCE_INLINE_TEMPLATE size_t FSE_decompress_usingDTable_generic(
232*a28cd43dSSascha Wildner           void* dst, size_t maxDstSize,
233*a28cd43dSSascha Wildner     const void* cSrc, size_t cSrcSize,
234*a28cd43dSSascha Wildner     const FSE_DTable* dt, const unsigned fast)
235*a28cd43dSSascha Wildner {
236*a28cd43dSSascha Wildner     BYTE* const ostart = (BYTE*) dst;
237*a28cd43dSSascha Wildner     BYTE* op = ostart;
238*a28cd43dSSascha Wildner     BYTE* const omax = op + maxDstSize;
239*a28cd43dSSascha Wildner     BYTE* const olimit = omax-3;
240*a28cd43dSSascha Wildner 
241*a28cd43dSSascha Wildner     BIT_DStream_t bitD;
242*a28cd43dSSascha Wildner     FSE_DState_t state1;
243*a28cd43dSSascha Wildner     FSE_DState_t state2;
244*a28cd43dSSascha Wildner 
245*a28cd43dSSascha Wildner     /* Init */
246*a28cd43dSSascha Wildner     CHECK_F(BIT_initDStream(&bitD, cSrc, cSrcSize));
247*a28cd43dSSascha Wildner 
248*a28cd43dSSascha Wildner     FSE_initDState(&state1, &bitD, dt);
249*a28cd43dSSascha Wildner     FSE_initDState(&state2, &bitD, dt);
250*a28cd43dSSascha Wildner 
251*a28cd43dSSascha Wildner #define FSE_GETSYMBOL(statePtr) fast ? FSE_decodeSymbolFast(statePtr, &bitD) : FSE_decodeSymbol(statePtr, &bitD)
252*a28cd43dSSascha Wildner 
253*a28cd43dSSascha Wildner     /* 4 symbols per loop */
254*a28cd43dSSascha Wildner     for ( ; (BIT_reloadDStream(&bitD)==BIT_DStream_unfinished) & (op<olimit) ; op+=4) {
255*a28cd43dSSascha Wildner         op[0] = FSE_GETSYMBOL(&state1);
256*a28cd43dSSascha Wildner 
257*a28cd43dSSascha Wildner         if (FSE_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
258*a28cd43dSSascha Wildner             BIT_reloadDStream(&bitD);
259*a28cd43dSSascha Wildner 
260*a28cd43dSSascha Wildner         op[1] = FSE_GETSYMBOL(&state2);
261*a28cd43dSSascha Wildner 
262*a28cd43dSSascha Wildner         if (FSE_MAX_TABLELOG*4+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
263*a28cd43dSSascha Wildner             { if (BIT_reloadDStream(&bitD) > BIT_DStream_unfinished) { op+=2; break; } }
264*a28cd43dSSascha Wildner 
265*a28cd43dSSascha Wildner         op[2] = FSE_GETSYMBOL(&state1);
266*a28cd43dSSascha Wildner 
267*a28cd43dSSascha Wildner         if (FSE_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
268*a28cd43dSSascha Wildner             BIT_reloadDStream(&bitD);
269*a28cd43dSSascha Wildner 
270*a28cd43dSSascha Wildner         op[3] = FSE_GETSYMBOL(&state2);
271*a28cd43dSSascha Wildner     }
272*a28cd43dSSascha Wildner 
273*a28cd43dSSascha Wildner     /* tail */
274*a28cd43dSSascha Wildner     /* note : BIT_reloadDStream(&bitD) >= FSE_DStream_partiallyFilled; Ends at exactly BIT_DStream_completed */
275*a28cd43dSSascha Wildner     while (1) {
276*a28cd43dSSascha Wildner         if (op>(omax-2)) return ERROR(dstSize_tooSmall);
277*a28cd43dSSascha Wildner         *op++ = FSE_GETSYMBOL(&state1);
278*a28cd43dSSascha Wildner         if (BIT_reloadDStream(&bitD)==BIT_DStream_overflow) {
279*a28cd43dSSascha Wildner             *op++ = FSE_GETSYMBOL(&state2);
280*a28cd43dSSascha Wildner             break;
281*a28cd43dSSascha Wildner         }
282*a28cd43dSSascha Wildner 
283*a28cd43dSSascha Wildner         if (op>(omax-2)) return ERROR(dstSize_tooSmall);
284*a28cd43dSSascha Wildner         *op++ = FSE_GETSYMBOL(&state2);
285*a28cd43dSSascha Wildner         if (BIT_reloadDStream(&bitD)==BIT_DStream_overflow) {
286*a28cd43dSSascha Wildner             *op++ = FSE_GETSYMBOL(&state1);
287*a28cd43dSSascha Wildner             break;
288*a28cd43dSSascha Wildner     }   }
289*a28cd43dSSascha Wildner 
290*a28cd43dSSascha Wildner     return op-ostart;
291*a28cd43dSSascha Wildner }
292*a28cd43dSSascha Wildner 
293*a28cd43dSSascha Wildner 
FSE_decompress_usingDTable(void * dst,size_t originalSize,const void * cSrc,size_t cSrcSize,const FSE_DTable * dt)294*a28cd43dSSascha Wildner size_t FSE_decompress_usingDTable(void* dst, size_t originalSize,
295*a28cd43dSSascha Wildner                             const void* cSrc, size_t cSrcSize,
296*a28cd43dSSascha Wildner                             const FSE_DTable* dt)
297*a28cd43dSSascha Wildner {
298*a28cd43dSSascha Wildner     const void* ptr = dt;
299*a28cd43dSSascha Wildner     const FSE_DTableHeader* DTableH = (const FSE_DTableHeader*)ptr;
300*a28cd43dSSascha Wildner     const U32 fastMode = DTableH->fastMode;
301*a28cd43dSSascha Wildner 
302*a28cd43dSSascha Wildner     /* select fast mode (static) */
303*a28cd43dSSascha Wildner     if (fastMode) return FSE_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 1);
304*a28cd43dSSascha Wildner     return FSE_decompress_usingDTable_generic(dst, originalSize, cSrc, cSrcSize, dt, 0);
305*a28cd43dSSascha Wildner }
306*a28cd43dSSascha Wildner 
307*a28cd43dSSascha Wildner 
FSE_decompress_wksp(void * dst,size_t dstCapacity,const void * cSrc,size_t cSrcSize,unsigned maxLog,void * workSpace,size_t wkspSize)308*a28cd43dSSascha Wildner size_t FSE_decompress_wksp(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize)
309*a28cd43dSSascha Wildner {
310*a28cd43dSSascha Wildner     return FSE_decompress_wksp_bmi2(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize, /* bmi2 */ 0);
311*a28cd43dSSascha Wildner }
312*a28cd43dSSascha Wildner 
FSE_decompress_wksp_body(void * dst,size_t dstCapacity,const void * cSrc,size_t cSrcSize,unsigned maxLog,void * workSpace,size_t wkspSize,int bmi2)313*a28cd43dSSascha Wildner FORCE_INLINE_TEMPLATE size_t FSE_decompress_wksp_body(
314*a28cd43dSSascha Wildner         void* dst, size_t dstCapacity,
315*a28cd43dSSascha Wildner         const void* cSrc, size_t cSrcSize,
316*a28cd43dSSascha Wildner         unsigned maxLog, void* workSpace, size_t wkspSize,
317*a28cd43dSSascha Wildner         int bmi2)
318*a28cd43dSSascha Wildner {
319*a28cd43dSSascha Wildner     const BYTE* const istart = (const BYTE*)cSrc;
320*a28cd43dSSascha Wildner     const BYTE* ip = istart;
321*a28cd43dSSascha Wildner     short counting[FSE_MAX_SYMBOL_VALUE+1];
322*a28cd43dSSascha Wildner     unsigned tableLog;
323*a28cd43dSSascha Wildner     unsigned maxSymbolValue = FSE_MAX_SYMBOL_VALUE;
324*a28cd43dSSascha Wildner     FSE_DTable* const dtable = (FSE_DTable*)workSpace;
325*a28cd43dSSascha Wildner 
326*a28cd43dSSascha Wildner     /* normal FSE decoding mode */
327*a28cd43dSSascha Wildner     size_t const NCountLength = FSE_readNCount_bmi2(counting, &maxSymbolValue, &tableLog, istart, cSrcSize, bmi2);
328*a28cd43dSSascha Wildner     if (FSE_isError(NCountLength)) return NCountLength;
329*a28cd43dSSascha Wildner     if (tableLog > maxLog) return ERROR(tableLog_tooLarge);
330*a28cd43dSSascha Wildner     assert(NCountLength <= cSrcSize);
331*a28cd43dSSascha Wildner     ip += NCountLength;
332*a28cd43dSSascha Wildner     cSrcSize -= NCountLength;
333*a28cd43dSSascha Wildner 
334*a28cd43dSSascha Wildner     if (FSE_DECOMPRESS_WKSP_SIZE(tableLog, maxSymbolValue) > wkspSize) return ERROR(tableLog_tooLarge);
335*a28cd43dSSascha Wildner     workSpace = dtable + FSE_DTABLE_SIZE_U32(tableLog);
336*a28cd43dSSascha Wildner     wkspSize -= FSE_DTABLE_SIZE(tableLog);
337*a28cd43dSSascha Wildner 
338*a28cd43dSSascha Wildner     CHECK_F( FSE_buildDTable_internal(dtable, counting, maxSymbolValue, tableLog, workSpace, wkspSize) );
339*a28cd43dSSascha Wildner 
340*a28cd43dSSascha Wildner     {
341*a28cd43dSSascha Wildner         const void* ptr = dtable;
342*a28cd43dSSascha Wildner         const FSE_DTableHeader* DTableH = (const FSE_DTableHeader*)ptr;
343*a28cd43dSSascha Wildner         const U32 fastMode = DTableH->fastMode;
344*a28cd43dSSascha Wildner 
345*a28cd43dSSascha Wildner         /* select fast mode (static) */
346*a28cd43dSSascha Wildner         if (fastMode) return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, dtable, 1);
347*a28cd43dSSascha Wildner         return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, dtable, 0);
348*a28cd43dSSascha Wildner     }
349*a28cd43dSSascha Wildner }
350*a28cd43dSSascha Wildner 
351*a28cd43dSSascha Wildner /* Avoids the FORCE_INLINE of the _body() function. */
FSE_decompress_wksp_body_default(void * dst,size_t dstCapacity,const void * cSrc,size_t cSrcSize,unsigned maxLog,void * workSpace,size_t wkspSize)352*a28cd43dSSascha Wildner static size_t FSE_decompress_wksp_body_default(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize)
353*a28cd43dSSascha Wildner {
354*a28cd43dSSascha Wildner     return FSE_decompress_wksp_body(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize, 0);
355*a28cd43dSSascha Wildner }
356*a28cd43dSSascha Wildner 
357*a28cd43dSSascha Wildner #if DYNAMIC_BMI2
FSE_decompress_wksp_body_bmi2(void * dst,size_t dstCapacity,const void * cSrc,size_t cSrcSize,unsigned maxLog,void * workSpace,size_t wkspSize)358*a28cd43dSSascha Wildner TARGET_ATTRIBUTE("bmi2") static size_t FSE_decompress_wksp_body_bmi2(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize)
359*a28cd43dSSascha Wildner {
360*a28cd43dSSascha Wildner     return FSE_decompress_wksp_body(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize, 1);
361*a28cd43dSSascha Wildner }
362*a28cd43dSSascha Wildner #endif
363*a28cd43dSSascha Wildner 
FSE_decompress_wksp_bmi2(void * dst,size_t dstCapacity,const void * cSrc,size_t cSrcSize,unsigned maxLog,void * workSpace,size_t wkspSize,int bmi2)364*a28cd43dSSascha Wildner size_t FSE_decompress_wksp_bmi2(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize, int bmi2)
365*a28cd43dSSascha Wildner {
366*a28cd43dSSascha Wildner #if DYNAMIC_BMI2
367*a28cd43dSSascha Wildner     if (bmi2) {
368*a28cd43dSSascha Wildner         return FSE_decompress_wksp_body_bmi2(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize);
369*a28cd43dSSascha Wildner     }
370*a28cd43dSSascha Wildner #endif
371*a28cd43dSSascha Wildner     (void)bmi2;
372*a28cd43dSSascha Wildner     return FSE_decompress_wksp_body_default(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize);
373*a28cd43dSSascha Wildner }
374*a28cd43dSSascha Wildner 
375*a28cd43dSSascha Wildner 
376*a28cd43dSSascha Wildner typedef FSE_DTable DTable_max_t[FSE_DTABLE_SIZE_U32(FSE_MAX_TABLELOG)];
377*a28cd43dSSascha Wildner 
378*a28cd43dSSascha Wildner #ifndef ZSTD_NO_UNUSED_FUNCTIONS
FSE_buildDTable(FSE_DTable * dt,const short * normalizedCounter,unsigned maxSymbolValue,unsigned tableLog)379*a28cd43dSSascha Wildner size_t FSE_buildDTable(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog) {
380*a28cd43dSSascha Wildner     U32 wksp[FSE_BUILD_DTABLE_WKSP_SIZE_U32(FSE_TABLELOG_ABSOLUTE_MAX, FSE_MAX_SYMBOL_VALUE)];
381*a28cd43dSSascha Wildner     return FSE_buildDTable_wksp(dt, normalizedCounter, maxSymbolValue, tableLog, wksp, sizeof(wksp));
382*a28cd43dSSascha Wildner }
383*a28cd43dSSascha Wildner 
FSE_decompress(void * dst,size_t dstCapacity,const void * cSrc,size_t cSrcSize)384*a28cd43dSSascha Wildner size_t FSE_decompress(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize)
385*a28cd43dSSascha Wildner {
386*a28cd43dSSascha Wildner     /* Static analyzer seems unable to understand this table will be properly initialized later */
387*a28cd43dSSascha Wildner     U32 wksp[FSE_DECOMPRESS_WKSP_SIZE_U32(FSE_MAX_TABLELOG, FSE_MAX_SYMBOL_VALUE)];
388*a28cd43dSSascha Wildner     return FSE_decompress_wksp(dst, dstCapacity, cSrc, cSrcSize, FSE_MAX_TABLELOG, wksp, sizeof(wksp));
389*a28cd43dSSascha Wildner }
390*a28cd43dSSascha Wildner #endif
391*a28cd43dSSascha Wildner 
392*a28cd43dSSascha Wildner 
393*a28cd43dSSascha Wildner #endif   /* FSE_COMMONDEFS_ONLY */
394