rename CMOV_IS_FAST to HAVE_FAST_CMOV and simplify configure
[ffmpeg.git] / libavcodec / cabac.h
1 /*
2  * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder
3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  *
21  */
22
23 /**
24  * @file cabac.h
25  * Context Adaptive Binary Arithmetic Coder.
26  */
27
28
29 //#undef NDEBUG
30 #include <assert.h>
31 #ifdef ARCH_X86
32 #include "x86_cpu.h"
33 #endif
34
35 #define CABAC_BITS 16
36 #define CABAC_MASK ((1<<CABAC_BITS)-1)
37 #define BRANCHLESS_CABAC_DECODER 1
38 //#define ARCH_X86_DISABLED 1
39
40 typedef struct CABACContext{
41     int low;
42     int range;
43     int outstanding_count;
44 #ifdef STRICT_LIMITS
45     int symCount;
46 #endif
47     const uint8_t *bytestream_start;
48     const uint8_t *bytestream;
49     const uint8_t *bytestream_end;
50     PutBitContext pb;
51 }CABACContext;
52
53 extern uint8_t ff_h264_mlps_state[4*64];
54 extern uint8_t ff_h264_lps_range[4*2*64];  ///< rangeTabLPS
55 extern uint8_t ff_h264_mps_state[2*64];     ///< transIdxMPS
56 extern uint8_t ff_h264_lps_state[2*64];     ///< transIdxLPS
57 extern const uint8_t ff_h264_norm_shift[512];
58
59
60 void ff_init_cabac_encoder(CABACContext *c, uint8_t *buf, int buf_size);
61 void ff_init_cabac_decoder(CABACContext *c, const uint8_t *buf, int buf_size);
62 void ff_init_cabac_states(CABACContext *c);
63
64
65 static inline void put_cabac_bit(CABACContext *c, int b){
66     put_bits(&c->pb, 1, b);
67     for(;c->outstanding_count; c->outstanding_count--){
68         put_bits(&c->pb, 1, 1-b);
69     }
70 }
71
72 static inline void renorm_cabac_encoder(CABACContext *c){
73     while(c->range < 0x100){
74         //FIXME optimize
75         if(c->low<0x100){
76             put_cabac_bit(c, 0);
77         }else if(c->low<0x200){
78             c->outstanding_count++;
79             c->low -= 0x100;
80         }else{
81             put_cabac_bit(c, 1);
82             c->low -= 0x200;
83         }
84
85         c->range+= c->range;
86         c->low += c->low;
87     }
88 }
89
90 static void put_cabac(CABACContext *c, uint8_t * const state, int bit){
91     int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + *state];
92
93     if(bit == ((*state)&1)){
94         c->range -= RangeLPS;
95         *state= ff_h264_mps_state[*state];
96     }else{
97         c->low += c->range - RangeLPS;
98         c->range = RangeLPS;
99         *state= ff_h264_lps_state[*state];
100     }
101
102     renorm_cabac_encoder(c);
103
104 #ifdef STRICT_LIMITS
105     c->symCount++;
106 #endif
107 }
108
109 static void put_cabac_static(CABACContext *c, int RangeLPS, int bit){
110     assert(c->range > RangeLPS);
111
112     if(!bit){
113         c->range -= RangeLPS;
114     }else{
115         c->low += c->range - RangeLPS;
116         c->range = RangeLPS;
117     }
118
119     renorm_cabac_encoder(c);
120
121 #ifdef STRICT_LIMITS
122     c->symCount++;
123 #endif
124 }
125
126 /**
127  * @param bit 0 -> write zero bit, !=0 write one bit
128  */
129 static void put_cabac_bypass(CABACContext *c, int bit){
130     c->low += c->low;
131
132     if(bit){
133         c->low += c->range;
134     }
135 //FIXME optimize
136     if(c->low<0x200){
137         put_cabac_bit(c, 0);
138     }else if(c->low<0x400){
139         c->outstanding_count++;
140         c->low -= 0x200;
141     }else{
142         put_cabac_bit(c, 1);
143         c->low -= 0x400;
144     }
145
146 #ifdef STRICT_LIMITS
147     c->symCount++;
148 #endif
149 }
150
151 /**
152  *
153  * @return the number of bytes written
154  */
155 static int put_cabac_terminate(CABACContext *c, int bit){
156     c->range -= 2;
157
158     if(!bit){
159         renorm_cabac_encoder(c);
160     }else{
161         c->low += c->range;
162         c->range= 2;
163
164         renorm_cabac_encoder(c);
165
166         assert(c->low <= 0x1FF);
167         put_cabac_bit(c, c->low>>9);
168         put_bits(&c->pb, 2, ((c->low>>7)&3)|1);
169
170         flush_put_bits(&c->pb); //FIXME FIXME FIXME XXX wrong
171     }
172
173 #ifdef STRICT_LIMITS
174     c->symCount++;
175 #endif
176
177     return (put_bits_count(&c->pb)+7)>>3;
178 }
179
180 /**
181  * put (truncated) unary binarization.
182  */
183 static void put_cabac_u(CABACContext *c, uint8_t * state, int v, int max, int max_index, int truncated){
184     int i;
185
186     assert(v <= max);
187
188 #if 1
189     for(i=0; i<v; i++){
190         put_cabac(c, state, 1);
191         if(i < max_index) state++;
192     }
193     if(truncated==0 || v<max)
194         put_cabac(c, state, 0);
195 #else
196     if(v <= max_index){
197         for(i=0; i<v; i++){
198             put_cabac(c, state+i, 1);
199         }
200         if(truncated==0 || v<max)
201             put_cabac(c, state+i, 0);
202     }else{
203         for(i=0; i<=max_index; i++){
204             put_cabac(c, state+i, 1);
205         }
206         for(; i<v; i++){
207             put_cabac(c, state+max_index, 1);
208         }
209         if(truncated==0 || v<max)
210             put_cabac(c, state+max_index, 0);
211     }
212 #endif
213 }
214
215 /**
216  * put unary exp golomb k-th order binarization.
217  */
218 static void put_cabac_ueg(CABACContext *c, uint8_t * state, int v, int max, int is_signed, int k, int max_index){
219     int i;
220
221     if(v==0)
222         put_cabac(c, state, 0);
223     else{
224         const int sign= v < 0;
225
226         if(is_signed) v= FFABS(v);
227
228         if(v<max){
229             for(i=0; i<v; i++){
230                 put_cabac(c, state, 1);
231                 if(i < max_index) state++;
232             }
233
234             put_cabac(c, state, 0);
235         }else{
236             int m= 1<<k;
237
238             for(i=0; i<max; i++){
239                 put_cabac(c, state, 1);
240                 if(i < max_index) state++;
241             }
242
243             v -= max;
244             while(v >= m){ //FIXME optimize
245                 put_cabac_bypass(c, 1);
246                 v-= m;
247                 m+= m;
248             }
249             put_cabac_bypass(c, 0);
250             while(m>>=1){
251                 put_cabac_bypass(c, v&m);
252             }
253         }
254
255         if(is_signed)
256             put_cabac_bypass(c, sign);
257     }
258 }
259
260 static void refill(CABACContext *c){
261 #if CABAC_BITS == 16
262         c->low+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1);
263 #else
264         c->low+= c->bytestream[0]<<1;
265 #endif
266     c->low -= CABAC_MASK;
267     c->bytestream+= CABAC_BITS/8;
268 }
269
270 static void refill2(CABACContext *c){
271     int i, x;
272
273     x= c->low ^ (c->low-1);
274     i= 7 - ff_h264_norm_shift[x>>(CABAC_BITS-1)];
275
276     x= -CABAC_MASK;
277
278 #if CABAC_BITS == 16
279         x+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1);
280 #else
281         x+= c->bytestream[0]<<1;
282 #endif
283
284     c->low += x<<i;
285     c->bytestream+= CABAC_BITS/8;
286 }
287
288 static inline void renorm_cabac_decoder(CABACContext *c){
289     while(c->range < 0x100){
290         c->range+= c->range;
291         c->low+= c->low;
292         if(!(c->low & CABAC_MASK))
293             refill(c);
294     }
295 }
296
297 static inline void renorm_cabac_decoder_once(CABACContext *c){
298 #ifdef ARCH_X86_DISABLED
299     int temp;
300 #if 0
301     //P3:683    athlon:475
302     asm(
303         "lea -0x100(%0), %2         \n\t"
304         "shr $31, %2                \n\t"  //FIXME 31->63 for x86-64
305         "shl %%cl, %0               \n\t"
306         "shl %%cl, %1               \n\t"
307         : "+r"(c->range), "+r"(c->low), "+c"(temp)
308     );
309 #elif 0
310     //P3:680    athlon:474
311     asm(
312         "cmp $0x100, %0             \n\t"
313         "setb %%cl                  \n\t"  //FIXME 31->63 for x86-64
314         "shl %%cl, %0               \n\t"
315         "shl %%cl, %1               \n\t"
316         : "+r"(c->range), "+r"(c->low), "+c"(temp)
317     );
318 #elif 1
319     int temp2;
320     //P3:665    athlon:517
321     asm(
322         "lea -0x100(%0), %%eax      \n\t"
323         "cdq                        \n\t"
324         "mov %0, %%eax              \n\t"
325         "and %%edx, %0              \n\t"
326         "and %1, %%edx              \n\t"
327         "add %%eax, %0              \n\t"
328         "add %%edx, %1              \n\t"
329         : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
330     );
331 #elif 0
332     int temp2;
333     //P3:673    athlon:509
334     asm(
335         "cmp $0x100, %0             \n\t"
336         "sbb %%edx, %%edx           \n\t"
337         "mov %0, %%eax              \n\t"
338         "and %%edx, %0              \n\t"
339         "and %1, %%edx              \n\t"
340         "add %%eax, %0              \n\t"
341         "add %%edx, %1              \n\t"
342         : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
343     );
344 #else
345     int temp2;
346     //P3:677    athlon:511
347     asm(
348         "cmp $0x100, %0             \n\t"
349         "lea (%0, %0), %%eax        \n\t"
350         "lea (%1, %1), %%edx        \n\t"
351         "cmovb %%eax, %0            \n\t"
352         "cmovb %%edx, %1            \n\t"
353         : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
354     );
355 #endif
356 #else
357     //P3:675    athlon:476
358     int shift= (uint32_t)(c->range - 0x100)>>31;
359     c->range<<= shift;
360     c->low  <<= shift;
361 #endif
362     if(!(c->low & CABAC_MASK))
363         refill(c);
364 }
365
366 static int av_always_inline get_cabac_inline(CABACContext *c, uint8_t * const state){
367     //FIXME gcc generates duplicate load/stores for c->low and c->range
368 #define LOW          "0"
369 #define RANGE        "4"
370 #ifdef ARCH_X86_64
371 #define BYTESTART   "16"
372 #define BYTE        "24"
373 #define BYTEEND     "32"
374 #else
375 #define BYTESTART   "12"
376 #define BYTE        "16"
377 #define BYTEEND     "20"
378 #endif
379 #if defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__))
380     int bit;
381
382 #ifndef BRANCHLESS_CABAC_DECODER
383     asm volatile(
384         "movzbl (%1), %0                        \n\t"
385         "movl "RANGE    "(%2), %%ebx            \n\t"
386         "movl "RANGE    "(%2), %%edx            \n\t"
387         "andl $0xC0, %%ebx                      \n\t"
388         "movzbl "MANGLE(ff_h264_lps_range)"(%0, %%ebx, 2), %%esi\n\t"
389         "movl "LOW      "(%2), %%ebx            \n\t"
390 //eax:state ebx:low, edx:range, esi:RangeLPS
391         "subl %%esi, %%edx                      \n\t"
392         "movl %%edx, %%ecx                      \n\t"
393         "shll $17, %%ecx                        \n\t"
394         "cmpl %%ecx, %%ebx                      \n\t"
395         " ja 1f                                 \n\t"
396
397 #if 1
398         //athlon:4067 P3:4110
399         "lea -0x100(%%edx), %%ecx               \n\t"
400         "shr $31, %%ecx                         \n\t"
401         "shl %%cl, %%edx                        \n\t"
402         "shl %%cl, %%ebx                        \n\t"
403 #else
404         //athlon:4057 P3:4130
405         "cmp $0x100, %%edx                      \n\t" //FIXME avoidable
406         "setb %%cl                              \n\t"
407         "shl %%cl, %%edx                        \n\t"
408         "shl %%cl, %%ebx                        \n\t"
409 #endif
410         "movzbl "MANGLE(ff_h264_mps_state)"(%0), %%ecx   \n\t"
411         "movb %%cl, (%1)                        \n\t"
412 //eax:state ebx:low, edx:range, esi:RangeLPS
413         "test %%bx, %%bx                        \n\t"
414         " jnz 2f                                \n\t"
415         "mov  "BYTE     "(%2), %%"REG_S"        \n\t"
416         "subl $0xFFFF, %%ebx                    \n\t"
417         "movzwl (%%"REG_S"), %%ecx              \n\t"
418         "bswap %%ecx                            \n\t"
419         "shrl $15, %%ecx                        \n\t"
420         "add  $2, %%"REG_S"                     \n\t"
421         "addl %%ecx, %%ebx                      \n\t"
422         "mov  %%"REG_S", "BYTE    "(%2)         \n\t"
423         "jmp 2f                                 \n\t"
424         "1:                                     \n\t"
425 //eax:state ebx:low, edx:range, esi:RangeLPS
426         "subl %%ecx, %%ebx                      \n\t"
427         "movl %%esi, %%edx                      \n\t"
428         "movzbl " MANGLE(ff_h264_norm_shift) "(%%esi), %%ecx   \n\t"
429         "shll %%cl, %%ebx                       \n\t"
430         "shll %%cl, %%edx                       \n\t"
431         "movzbl "MANGLE(ff_h264_lps_state)"(%0), %%ecx   \n\t"
432         "movb %%cl, (%1)                        \n\t"
433         "add  $1, %0                            \n\t"
434         "test %%bx, %%bx                        \n\t"
435         " jnz 2f                                \n\t"
436
437         "mov  "BYTE     "(%2), %%"REG_c"        \n\t"
438         "movzwl (%%"REG_c"), %%esi              \n\t"
439         "bswap %%esi                            \n\t"
440         "shrl $15, %%esi                        \n\t"
441         "subl $0xFFFF, %%esi                    \n\t"
442         "add  $2, %%"REG_c"                     \n\t"
443         "mov  %%"REG_c", "BYTE    "(%2)         \n\t"
444
445         "leal -1(%%ebx), %%ecx                  \n\t"
446         "xorl %%ebx, %%ecx                      \n\t"
447         "shrl $15, %%ecx                        \n\t"
448         "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx   \n\t"
449         "neg %%ecx                              \n\t"
450         "add $7, %%ecx                          \n\t"
451
452         "shll %%cl , %%esi                      \n\t"
453         "addl %%esi, %%ebx                      \n\t"
454         "2:                                     \n\t"
455         "movl %%edx, "RANGE    "(%2)            \n\t"
456         "movl %%ebx, "LOW      "(%2)            \n\t"
457         :"=&a"(bit) //FIXME this is fragile gcc either runs out of registers or misscompiles it (for example if "+a"(bit) or "+m"(*state) is used
458         :"r"(state), "r"(c)
459         : "%"REG_c, "%ebx", "%edx", "%"REG_S, "memory"
460     );
461     bit&=1;
462 #else /* BRANCHLESS_CABAC_DECODER */
463
464
465 #if defined HAVE_FAST_CMOV
466 #define BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
467         "mov    "tmp"       , %%ecx                                     \n\t"\
468         "shl    $17         , "tmp"                                     \n\t"\
469         "cmp    "low"       , "tmp"                                     \n\t"\
470         "cmova  %%ecx       , "range"                                   \n\t"\
471         "sbb    %%ecx       , %%ecx                                     \n\t"\
472         "and    %%ecx       , "tmp"                                     \n\t"\
473         "sub    "tmp"       , "low"                                     \n\t"\
474         "xor    %%ecx       , "ret"                                     \n\t"
475 #else /* HAVE_FAST_CMOV */
476 #define BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
477         "mov    "tmp"       , %%ecx                                     \n\t"\
478         "shl    $17         , "tmp"                                     \n\t"\
479         "sub    "low"       , "tmp"                                     \n\t"\
480         "sar    $31         , "tmp"                                     \n\t" /*lps_mask*/\
481         "sub    %%ecx       , "range"                                   \n\t" /*RangeLPS - range*/\
482         "and    "tmp"       , "range"                                   \n\t" /*(RangeLPS - range)&lps_mask*/\
483         "add    %%ecx       , "range"                                   \n\t" /*new range*/\
484         "shl    $17         , %%ecx                                     \n\t"\
485         "and    "tmp"       , %%ecx                                     \n\t"\
486         "sub    %%ecx       , "low"                                     \n\t"\
487         "xor    "tmp"       , "ret"                                     \n\t"
488 #endif /* HAVE_FAST_CMOV */
489
490
491 #define BRANCHLESS_GET_CABAC(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
492         "movzbl "statep"    , "ret"                                     \n\t"\
493         "mov    "range"     , "tmp"                                     \n\t"\
494         "and    $0xC0       , "range"                                   \n\t"\
495         "movzbl "MANGLE(ff_h264_lps_range)"("ret", "range", 2), "range" \n\t"\
496         "sub    "range"     , "tmp"                                     \n\t"\
497         BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
498         "movzbl " MANGLE(ff_h264_norm_shift) "("range"), %%ecx          \n\t"\
499         "shl    %%cl        , "range"                                   \n\t"\
500         "movzbl "MANGLE(ff_h264_mlps_state)"+128("ret"), "tmp"          \n\t"\
501         "mov    "tmpbyte"   , "statep"                                  \n\t"\
502         "shl    %%cl        , "low"                                     \n\t"\
503         "test   "lowword"   , "lowword"                                 \n\t"\
504         " jnz   1f                                                      \n\t"\
505         "mov "BYTE"("cabac"), %%"REG_c"                                 \n\t"\
506         "movzwl (%%"REG_c")     , "tmp"                                 \n\t"\
507         "bswap  "tmp"                                                   \n\t"\
508         "shr    $15         , "tmp"                                     \n\t"\
509         "sub    $0xFFFF     , "tmp"                                     \n\t"\
510         "add    $2          , %%"REG_c"                                 \n\t"\
511         "mov    %%"REG_c"   , "BYTE    "("cabac")                       \n\t"\
512         "lea    -1("low")   , %%ecx                                     \n\t"\
513         "xor    "low"       , %%ecx                                     \n\t"\
514         "shr    $15         , %%ecx                                     \n\t"\
515         "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx            \n\t"\
516         "neg    %%ecx                                                   \n\t"\
517         "add    $7          , %%ecx                                     \n\t"\
518         "shl    %%cl        , "tmp"                                     \n\t"\
519         "add    "tmp"       , "low"                                     \n\t"\
520         "1:                                                             \n\t"
521
522     asm volatile(
523         "movl "RANGE    "(%2), %%esi            \n\t"
524         "movl "LOW      "(%2), %%ebx            \n\t"
525         BRANCHLESS_GET_CABAC("%0", "%2", "(%1)", "%%ebx", "%%bx", "%%esi", "%%edx", "%%dl")
526         "movl %%esi, "RANGE    "(%2)            \n\t"
527         "movl %%ebx, "LOW      "(%2)            \n\t"
528
529         :"=&a"(bit)
530         :"r"(state), "r"(c)
531         : "%"REG_c, "%ebx", "%edx", "%esi", "memory"
532     );
533     bit&=1;
534 #endif /* BRANCHLESS_CABAC_DECODER */
535 #else /* defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__)) */
536     int s = *state;
537     int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + s];
538     int bit, lps_mask attribute_unused;
539
540     c->range -= RangeLPS;
541 #ifndef BRANCHLESS_CABAC_DECODER
542     if(c->low < (c->range<<(CABAC_BITS+1))){
543         bit= s&1;
544         *state= ff_h264_mps_state[s];
545         renorm_cabac_decoder_once(c);
546     }else{
547         bit= ff_h264_norm_shift[RangeLPS];
548         c->low -= (c->range<<(CABAC_BITS+1));
549         *state= ff_h264_lps_state[s];
550         c->range = RangeLPS<<bit;
551         c->low <<= bit;
552         bit= (s&1)^1;
553
554         if(!(c->low & CABAC_MASK)){
555             refill2(c);
556         }
557     }
558 #else /* BRANCHLESS_CABAC_DECODER */
559     lps_mask= ((c->range<<(CABAC_BITS+1)) - c->low)>>31;
560
561     c->low -= (c->range<<(CABAC_BITS+1)) & lps_mask;
562     c->range += (RangeLPS - c->range) & lps_mask;
563
564     s^=lps_mask;
565     *state= (ff_h264_mlps_state+128)[s];
566     bit= s&1;
567
568     lps_mask= ff_h264_norm_shift[c->range];
569     c->range<<= lps_mask;
570     c->low  <<= lps_mask;
571     if(!(c->low & CABAC_MASK))
572         refill2(c);
573 #endif /* BRANCHLESS_CABAC_DECODER */
574 #endif /* defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__)) */
575     return bit;
576 }
577
578 static int __attribute((noinline)) get_cabac_noinline(CABACContext *c, uint8_t * const state){
579     return get_cabac_inline(c,state);
580 }
581
582 static int get_cabac(CABACContext *c, uint8_t * const state){
583     return get_cabac_inline(c,state);
584 }
585
586 static int get_cabac_bypass(CABACContext *c){
587 #if 0 //not faster
588     int bit;
589     asm volatile(
590         "movl "RANGE    "(%1), %%ebx            \n\t"
591         "movl "LOW      "(%1), %%eax            \n\t"
592         "shl $17, %%ebx                         \n\t"
593         "add %%eax, %%eax                       \n\t"
594         "sub %%ebx, %%eax                       \n\t"
595         "cdq                                    \n\t"
596         "and %%edx, %%ebx                       \n\t"
597         "add %%ebx, %%eax                       \n\t"
598         "test %%ax, %%ax                        \n\t"
599         " jnz 1f                                \n\t"
600         "movl "BYTE     "(%1), %%"REG_b"        \n\t"
601         "subl $0xFFFF, %%eax                    \n\t"
602         "movzwl (%%"REG_b"), %%ecx              \n\t"
603         "bswap %%ecx                            \n\t"
604         "shrl $15, %%ecx                        \n\t"
605         "addl $2, %%"REG_b"                     \n\t"
606         "addl %%ecx, %%eax                      \n\t"
607         "movl %%"REG_b", "BYTE     "(%1)        \n\t"
608         "1:                                     \n\t"
609         "movl %%eax, "LOW      "(%1)            \n\t"
610
611         :"=&d"(bit)
612         :"r"(c)
613         : "%eax", "%"REG_b, "%ecx", "memory"
614     );
615     return bit+1;
616 #else
617     int range;
618     c->low += c->low;
619
620     if(!(c->low & CABAC_MASK))
621         refill(c);
622
623     range= c->range<<(CABAC_BITS+1);
624     if(c->low < range){
625         return 0;
626     }else{
627         c->low -= range;
628         return 1;
629     }
630 #endif
631 }
632
633
634 static av_always_inline int get_cabac_bypass_sign(CABACContext *c, int val){
635 #if defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__))
636     asm volatile(
637         "movl "RANGE    "(%1), %%ebx            \n\t"
638         "movl "LOW      "(%1), %%eax            \n\t"
639         "shl $17, %%ebx                         \n\t"
640         "add %%eax, %%eax                       \n\t"
641         "sub %%ebx, %%eax                       \n\t"
642         "cdq                                    \n\t"
643         "and %%edx, %%ebx                       \n\t"
644         "add %%ebx, %%eax                       \n\t"
645         "xor %%edx, %%ecx                       \n\t"
646         "sub %%edx, %%ecx                       \n\t"
647         "test %%ax, %%ax                        \n\t"
648         " jnz 1f                                \n\t"
649         "mov  "BYTE     "(%1), %%"REG_b"        \n\t"
650         "subl $0xFFFF, %%eax                    \n\t"
651         "movzwl (%%"REG_b"), %%edx              \n\t"
652         "bswap %%edx                            \n\t"
653         "shrl $15, %%edx                        \n\t"
654         "add  $2, %%"REG_b"                     \n\t"
655         "addl %%edx, %%eax                      \n\t"
656         "mov  %%"REG_b", "BYTE     "(%1)        \n\t"
657         "1:                                     \n\t"
658         "movl %%eax, "LOW      "(%1)            \n\t"
659
660         :"+c"(val)
661         :"r"(c)
662         : "%eax", "%"REG_b, "%edx", "memory"
663     );
664     return val;
665 #else
666     int range, mask;
667     c->low += c->low;
668
669     if(!(c->low & CABAC_MASK))
670         refill(c);
671
672     range= c->range<<(CABAC_BITS+1);
673     c->low -= range;
674     mask= c->low >> 31;
675     range &= mask;
676     c->low += range;
677     return (val^mask)-mask;
678 #endif
679 }
680
681 //FIXME the x86 code from this file should be moved into i386/h264 or cabac something.c/h (note ill kill you if you move my code away from under my fingers before iam finished with it!)
682 //FIXME use some macros to avoid duplicatin get_cabac (cant be done yet as that would make optimization work hard)
683 #if defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__))
684 static int decode_significance_x86(CABACContext *c, int max_coeff, uint8_t *significant_coeff_ctx_base, int *index){
685     void *end= significant_coeff_ctx_base + max_coeff - 1;
686     int minusstart= -(int)significant_coeff_ctx_base;
687     int minusindex= 4-(int)index;
688     int coeff_count;
689     asm volatile(
690         "movl "RANGE    "(%3), %%esi            \n\t"
691         "movl "LOW      "(%3), %%ebx            \n\t"
692
693         "2:                                     \n\t"
694
695         BRANCHLESS_GET_CABAC("%%edx", "%3", "(%1)", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
696
697         "test $1, %%edx                         \n\t"
698         " jz 3f                                 \n\t"
699
700         BRANCHLESS_GET_CABAC("%%edx", "%3", "61(%1)", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
701
702         "mov  %2, %%"REG_a"                     \n\t"
703         "movl %4, %%ecx                         \n\t"
704         "add  %1, %%"REG_c"                     \n\t"
705         "movl %%ecx, (%%"REG_a")                \n\t"
706
707         "test $1, %%edx                         \n\t"
708         " jnz 4f                                \n\t"
709
710         "add  $4, %%"REG_a"                     \n\t"
711         "mov  %%"REG_a", %2                     \n\t"
712
713         "3:                                     \n\t"
714         "add  $1, %1                            \n\t"
715         "cmp  %5, %1                            \n\t"
716         " jb 2b                                 \n\t"
717         "mov  %2, %%"REG_a"                     \n\t"
718         "movl %4, %%ecx                         \n\t"
719         "add  %1, %%"REG_c"                     \n\t"
720         "movl %%ecx, (%%"REG_a")                \n\t"
721         "4:                                     \n\t"
722         "add  %6, %%eax                         \n\t"
723         "shr $2, %%eax                          \n\t"
724
725         "movl %%esi, "RANGE    "(%3)            \n\t"
726         "movl %%ebx, "LOW      "(%3)            \n\t"
727         :"=&a"(coeff_count), "+r"(significant_coeff_ctx_base), "+m"(index)\
728         :"r"(c), "m"(minusstart), "m"(end), "m"(minusindex)\
729         : "%"REG_c, "%ebx", "%edx", "%esi", "memory"\
730     );
731     return coeff_count;
732 }
733
734 static int decode_significance_8x8_x86(CABACContext *c, uint8_t *significant_coeff_ctx_base, int *index, uint8_t *sig_off){
735     int minusindex= 4-(int)index;
736     int coeff_count;
737     long last=0;
738     asm volatile(
739         "movl "RANGE    "(%3), %%esi            \n\t"
740         "movl "LOW      "(%3), %%ebx            \n\t"
741
742         "mov %1, %%"REG_D"                      \n\t"
743         "2:                                     \n\t"
744
745         "mov %6, %%"REG_a"                      \n\t"
746         "movzbl (%%"REG_a", %%"REG_D"), %%edi   \n\t"
747         "add %5, %%"REG_D"                      \n\t"
748
749         BRANCHLESS_GET_CABAC("%%edx", "%3", "(%%"REG_D")", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
750
751         "mov %1, %%edi                          \n\t"
752         "test $1, %%edx                         \n\t"
753         " jz 3f                                 \n\t"
754
755         "movzbl "MANGLE(last_coeff_flag_offset_8x8)"(%%edi), %%edi\n\t"
756         "add %5, %%"REG_D"                      \n\t"
757
758         BRANCHLESS_GET_CABAC("%%edx", "%3", "15(%%"REG_D")", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
759
760         "mov %2, %%"REG_a"                      \n\t"
761         "mov %1, %%edi                          \n\t"
762         "movl %%edi, (%%"REG_a")                \n\t"
763
764         "test $1, %%edx                         \n\t"
765         " jnz 4f                                \n\t"
766
767         "add $4, %%"REG_a"                      \n\t"
768         "mov %%"REG_a", %2                      \n\t"
769
770         "3:                                     \n\t"
771         "addl $1, %%edi                         \n\t"
772         "mov %%edi, %1                          \n\t"
773         "cmpl $63, %%edi                        \n\t"
774         " jb 2b                                 \n\t"
775         "mov %2, %%"REG_a"                      \n\t"
776         "movl %%edi, (%%"REG_a")                \n\t"
777         "4:                                     \n\t"
778         "addl %4, %%eax                         \n\t"
779         "shr $2, %%eax                          \n\t"
780
781         "movl %%esi, "RANGE    "(%3)            \n\t"
782         "movl %%ebx, "LOW      "(%3)            \n\t"
783         :"=&a"(coeff_count),"+m"(last), "+m"(index)\
784         :"r"(c), "m"(minusindex), "m"(significant_coeff_ctx_base), "m"(sig_off)\
785         : "%"REG_c, "%ebx", "%edx", "%esi", "%"REG_D, "memory"\
786     );
787     return coeff_count;
788 }
789 #endif /* defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__)) */
790
791 /**
792  *
793  * @return the number of bytes read or 0 if no end
794  */
795 static int get_cabac_terminate(CABACContext *c){
796     c->range -= 2;
797     if(c->low < c->range<<(CABAC_BITS+1)){
798         renorm_cabac_decoder_once(c);
799         return 0;
800     }else{
801         return c->bytestream - c->bytestream_start;
802     }
803 }
804
805 /**
806  * get (truncated) unnary binarization.
807  */
808 static int get_cabac_u(CABACContext *c, uint8_t * state, int max, int max_index, int truncated){
809     int i;
810
811     for(i=0; i<max; i++){
812         if(get_cabac(c, state)==0)
813             return i;
814
815         if(i< max_index) state++;
816     }
817
818     return truncated ? max : -1;
819 }
820
821 /**
822  * get unary exp golomb k-th order binarization.
823  */
824 static int get_cabac_ueg(CABACContext *c, uint8_t * state, int max, int is_signed, int k, int max_index){
825     int i, v;
826     int m= 1<<k;
827
828     if(get_cabac(c, state)==0)
829         return 0;
830
831     if(0 < max_index) state++;
832
833     for(i=1; i<max; i++){
834         if(get_cabac(c, state)==0){
835             if(is_signed && get_cabac_bypass(c)){
836                 return -i;
837             }else
838                 return i;
839         }
840
841         if(i < max_index) state++;
842     }
843
844     while(get_cabac_bypass(c)){
845         i+= m;
846         m+= m;
847     }
848
849     v=0;
850     while(m>>=1){
851         v+= v + get_cabac_bypass(c);
852     }
853     i += v;
854
855     if(is_signed && get_cabac_bypass(c)){
856         return -i;
857     }else
858         return i;
859 }