Fix compilation with PIC enabled, BRANCHLESS_GET_CABAC is defined under
[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_64
32 #define ARCH_X86
33 #endif
34 #ifdef ARCH_X86
35 #include "x86_cpu.h"
36 #endif
37
38 #define CABAC_BITS 16
39 #define CABAC_MASK ((1<<CABAC_BITS)-1)
40 #define BRANCHLESS_CABAC_DECODER 1
41 //#define ARCH_X86_DISABLED 1
42
43 typedef struct CABACContext{
44     int low;
45     int range;
46     int outstanding_count;
47 #ifdef STRICT_LIMITS
48     int symCount;
49 #endif
50     const uint8_t *bytestream_start;
51     const uint8_t *bytestream;
52     const uint8_t *bytestream_end;
53     PutBitContext pb;
54 }CABACContext;
55
56 extern uint8_t ff_h264_mlps_state[4*64];
57 extern uint8_t ff_h264_lps_range[4*2*64];  ///< rangeTabLPS
58 extern uint8_t ff_h264_mps_state[2*64];     ///< transIdxMPS
59 extern uint8_t ff_h264_lps_state[2*64];     ///< transIdxLPS
60 extern const uint8_t ff_h264_norm_shift[512];
61
62
63 void ff_init_cabac_encoder(CABACContext *c, uint8_t *buf, int buf_size);
64 void ff_init_cabac_decoder(CABACContext *c, const uint8_t *buf, int buf_size);
65 void ff_init_cabac_states(CABACContext *c);
66
67
68 static inline void put_cabac_bit(CABACContext *c, int b){
69     put_bits(&c->pb, 1, b);
70     for(;c->outstanding_count; c->outstanding_count--){
71         put_bits(&c->pb, 1, 1-b);
72     }
73 }
74
75 static inline void renorm_cabac_encoder(CABACContext *c){
76     while(c->range < 0x100){
77         //FIXME optimize
78         if(c->low<0x100){
79             put_cabac_bit(c, 0);
80         }else if(c->low<0x200){
81             c->outstanding_count++;
82             c->low -= 0x100;
83         }else{
84             put_cabac_bit(c, 1);
85             c->low -= 0x200;
86         }
87
88         c->range+= c->range;
89         c->low += c->low;
90     }
91 }
92
93 static void put_cabac(CABACContext *c, uint8_t * const state, int bit){
94     int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + *state];
95
96     if(bit == ((*state)&1)){
97         c->range -= RangeLPS;
98         *state= ff_h264_mps_state[*state];
99     }else{
100         c->low += c->range - RangeLPS;
101         c->range = RangeLPS;
102         *state= ff_h264_lps_state[*state];
103     }
104
105     renorm_cabac_encoder(c);
106
107 #ifdef STRICT_LIMITS
108     c->symCount++;
109 #endif
110 }
111
112 static void put_cabac_static(CABACContext *c, int RangeLPS, int bit){
113     assert(c->range > RangeLPS);
114
115     if(!bit){
116         c->range -= RangeLPS;
117     }else{
118         c->low += c->range - RangeLPS;
119         c->range = RangeLPS;
120     }
121
122     renorm_cabac_encoder(c);
123
124 #ifdef STRICT_LIMITS
125     c->symCount++;
126 #endif
127 }
128
129 /**
130  * @param bit 0 -> write zero bit, !=0 write one bit
131  */
132 static void put_cabac_bypass(CABACContext *c, int bit){
133     c->low += c->low;
134
135     if(bit){
136         c->low += c->range;
137     }
138 //FIXME optimize
139     if(c->low<0x200){
140         put_cabac_bit(c, 0);
141     }else if(c->low<0x400){
142         c->outstanding_count++;
143         c->low -= 0x200;
144     }else{
145         put_cabac_bit(c, 1);
146         c->low -= 0x400;
147     }
148
149 #ifdef STRICT_LIMITS
150     c->symCount++;
151 #endif
152 }
153
154 /**
155  *
156  * @return the number of bytes written
157  */
158 static int put_cabac_terminate(CABACContext *c, int bit){
159     c->range -= 2;
160
161     if(!bit){
162         renorm_cabac_encoder(c);
163     }else{
164         c->low += c->range;
165         c->range= 2;
166
167         renorm_cabac_encoder(c);
168
169         assert(c->low <= 0x1FF);
170         put_cabac_bit(c, c->low>>9);
171         put_bits(&c->pb, 2, ((c->low>>7)&3)|1);
172
173         flush_put_bits(&c->pb); //FIXME FIXME FIXME XXX wrong
174     }
175
176 #ifdef STRICT_LIMITS
177     c->symCount++;
178 #endif
179
180     return (put_bits_count(&c->pb)+7)>>3;
181 }
182
183 /**
184  * put (truncated) unary binarization.
185  */
186 static void put_cabac_u(CABACContext *c, uint8_t * state, int v, int max, int max_index, int truncated){
187     int i;
188
189     assert(v <= max);
190
191 #if 1
192     for(i=0; i<v; i++){
193         put_cabac(c, state, 1);
194         if(i < max_index) state++;
195     }
196     if(truncated==0 || v<max)
197         put_cabac(c, state, 0);
198 #else
199     if(v <= max_index){
200         for(i=0; i<v; i++){
201             put_cabac(c, state+i, 1);
202         }
203         if(truncated==0 || v<max)
204             put_cabac(c, state+i, 0);
205     }else{
206         for(i=0; i<=max_index; i++){
207             put_cabac(c, state+i, 1);
208         }
209         for(; i<v; i++){
210             put_cabac(c, state+max_index, 1);
211         }
212         if(truncated==0 || v<max)
213             put_cabac(c, state+max_index, 0);
214     }
215 #endif
216 }
217
218 /**
219  * put unary exp golomb k-th order binarization.
220  */
221 static void put_cabac_ueg(CABACContext *c, uint8_t * state, int v, int max, int is_signed, int k, int max_index){
222     int i;
223
224     if(v==0)
225         put_cabac(c, state, 0);
226     else{
227         const int sign= v < 0;
228
229         if(is_signed) v= FFABS(v);
230
231         if(v<max){
232             for(i=0; i<v; i++){
233                 put_cabac(c, state, 1);
234                 if(i < max_index) state++;
235             }
236
237             put_cabac(c, state, 0);
238         }else{
239             int m= 1<<k;
240
241             for(i=0; i<max; i++){
242                 put_cabac(c, state, 1);
243                 if(i < max_index) state++;
244             }
245
246             v -= max;
247             while(v >= m){ //FIXME optimize
248                 put_cabac_bypass(c, 1);
249                 v-= m;
250                 m+= m;
251             }
252             put_cabac_bypass(c, 0);
253             while(m>>=1){
254                 put_cabac_bypass(c, v&m);
255             }
256         }
257
258         if(is_signed)
259             put_cabac_bypass(c, sign);
260     }
261 }
262
263 static void refill(CABACContext *c){
264 #if CABAC_BITS == 16
265         c->low+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1);
266 #else
267         c->low+= c->bytestream[0]<<1;
268 #endif
269     c->low -= CABAC_MASK;
270     c->bytestream+= CABAC_BITS/8;
271 }
272
273 static void refill2(CABACContext *c){
274     int i, x;
275
276     x= c->low ^ (c->low-1);
277     i= 7 - ff_h264_norm_shift[x>>(CABAC_BITS-1)];
278
279     x= -CABAC_MASK;
280
281 #if CABAC_BITS == 16
282         x+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1);
283 #else
284         x+= c->bytestream[0]<<1;
285 #endif
286
287     c->low += x<<i;
288     c->bytestream+= CABAC_BITS/8;
289 }
290
291 static inline void renorm_cabac_decoder(CABACContext *c){
292     while(c->range < 0x100){
293         c->range+= c->range;
294         c->low+= c->low;
295         if(!(c->low & CABAC_MASK))
296             refill(c);
297     }
298 }
299
300 static inline void renorm_cabac_decoder_once(CABACContext *c){
301 #ifdef ARCH_X86_DISABLED
302     int temp;
303 #if 0
304     //P3:683    athlon:475
305     asm(
306         "lea -0x100(%0), %2         \n\t"
307         "shr $31, %2                \n\t"  //FIXME 31->63 for x86-64
308         "shl %%cl, %0               \n\t"
309         "shl %%cl, %1               \n\t"
310         : "+r"(c->range), "+r"(c->low), "+c"(temp)
311     );
312 #elif 0
313     //P3:680    athlon:474
314     asm(
315         "cmp $0x100, %0             \n\t"
316         "setb %%cl                  \n\t"  //FIXME 31->63 for x86-64
317         "shl %%cl, %0               \n\t"
318         "shl %%cl, %1               \n\t"
319         : "+r"(c->range), "+r"(c->low), "+c"(temp)
320     );
321 #elif 1
322     int temp2;
323     //P3:665    athlon:517
324     asm(
325         "lea -0x100(%0), %%eax      \n\t"
326         "cdq                        \n\t"
327         "mov %0, %%eax              \n\t"
328         "and %%edx, %0              \n\t"
329         "and %1, %%edx              \n\t"
330         "add %%eax, %0              \n\t"
331         "add %%edx, %1              \n\t"
332         : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
333     );
334 #elif 0
335     int temp2;
336     //P3:673    athlon:509
337     asm(
338         "cmp $0x100, %0             \n\t"
339         "sbb %%edx, %%edx           \n\t"
340         "mov %0, %%eax              \n\t"
341         "and %%edx, %0              \n\t"
342         "and %1, %%edx              \n\t"
343         "add %%eax, %0              \n\t"
344         "add %%edx, %1              \n\t"
345         : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
346     );
347 #else
348     int temp2;
349     //P3:677    athlon:511
350     asm(
351         "cmp $0x100, %0             \n\t"
352         "lea (%0, %0), %%eax        \n\t"
353         "lea (%1, %1), %%edx        \n\t"
354         "cmovb %%eax, %0            \n\t"
355         "cmovb %%edx, %1            \n\t"
356         : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
357     );
358 #endif
359 #else
360     //P3:675    athlon:476
361     int shift= (uint32_t)(c->range - 0x100)>>31;
362     c->range<<= shift;
363     c->low  <<= shift;
364 #endif
365     if(!(c->low & CABAC_MASK))
366         refill(c);
367 }
368
369 static int always_inline get_cabac_inline(CABACContext *c, uint8_t * const state){
370     //FIXME gcc generates duplicate load/stores for c->low and c->range
371 #define LOW          "0"
372 #define RANGE        "4"
373 #ifdef ARCH_X86_64
374 #define BYTESTART   "16"
375 #define BYTE        "24"
376 #define BYTEEND     "32"
377 #else
378 #define BYTESTART   "12"
379 #define BYTE        "16"
380 #define BYTEEND     "20"
381 #endif
382 #if defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__))
383     int bit;
384
385 #ifndef BRANCHLESS_CABAC_DECODER
386     asm volatile(
387         "movzbl (%1), %0                        \n\t"
388         "movl "RANGE    "(%2), %%ebx            \n\t"
389         "movl "RANGE    "(%2), %%edx            \n\t"
390         "andl $0xC0, %%ebx                      \n\t"
391         "movzbl "MANGLE(ff_h264_lps_range)"(%0, %%ebx, 2), %%esi\n\t"
392         "movl "LOW      "(%2), %%ebx            \n\t"
393 //eax:state ebx:low, edx:range, esi:RangeLPS
394         "subl %%esi, %%edx                      \n\t"
395         "movl %%edx, %%ecx                      \n\t"
396         "shll $17, %%ecx                        \n\t"
397         "cmpl %%ecx, %%ebx                      \n\t"
398         " ja 1f                                 \n\t"
399
400 #if 1
401         //athlon:4067 P3:4110
402         "lea -0x100(%%edx), %%ecx               \n\t"
403         "shr $31, %%ecx                         \n\t"
404         "shl %%cl, %%edx                        \n\t"
405         "shl %%cl, %%ebx                        \n\t"
406 #else
407         //athlon:4057 P3:4130
408         "cmp $0x100, %%edx                      \n\t" //FIXME avoidable
409         "setb %%cl                              \n\t"
410         "shl %%cl, %%edx                        \n\t"
411         "shl %%cl, %%ebx                        \n\t"
412 #endif
413         "movzbl "MANGLE(ff_h264_mps_state)"(%0), %%ecx   \n\t"
414         "movb %%cl, (%1)                        \n\t"
415 //eax:state ebx:low, edx:range, esi:RangeLPS
416         "test %%bx, %%bx                        \n\t"
417         " jnz 2f                                \n\t"
418         "mov  "BYTE     "(%2), %%"REG_S"        \n\t"
419         "subl $0xFFFF, %%ebx                    \n\t"
420         "movzwl (%%"REG_S"), %%ecx              \n\t"
421         "bswap %%ecx                            \n\t"
422         "shrl $15, %%ecx                        \n\t"
423         "add  $2, %%"REG_S"                     \n\t"
424         "addl %%ecx, %%ebx                      \n\t"
425         "mov  %%"REG_S", "BYTE    "(%2)         \n\t"
426         "jmp 2f                                 \n\t"
427         "1:                                     \n\t"
428 //eax:state ebx:low, edx:range, esi:RangeLPS
429         "subl %%ecx, %%ebx                      \n\t"
430         "movl %%esi, %%edx                      \n\t"
431         "movzbl " MANGLE(ff_h264_norm_shift) "(%%esi), %%ecx   \n\t"
432         "shll %%cl, %%ebx                       \n\t"
433         "shll %%cl, %%edx                       \n\t"
434         "movzbl "MANGLE(ff_h264_lps_state)"(%0), %%ecx   \n\t"
435         "movb %%cl, (%1)                        \n\t"
436         "add  $1, %0                            \n\t"
437         "test %%bx, %%bx                        \n\t"
438         " jnz 2f                                \n\t"
439
440         "mov  "BYTE     "(%2), %%"REG_c"        \n\t"
441         "movzwl (%%"REG_c"), %%esi              \n\t"
442         "bswap %%esi                            \n\t"
443         "shrl $15, %%esi                        \n\t"
444         "subl $0xFFFF, %%esi                    \n\t"
445         "add  $2, %%"REG_c"                     \n\t"
446         "mov  %%"REG_c", "BYTE    "(%2)         \n\t"
447
448         "leal -1(%%ebx), %%ecx                  \n\t"
449         "xorl %%ebx, %%ecx                      \n\t"
450         "shrl $15, %%ecx                        \n\t"
451         "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx   \n\t"
452         "neg %%ecx                              \n\t"
453         "add $7, %%ecx                          \n\t"
454
455         "shll %%cl , %%esi                      \n\t"
456         "addl %%esi, %%ebx                      \n\t"
457         "2:                                     \n\t"
458         "movl %%edx, "RANGE    "(%2)            \n\t"
459         "movl %%ebx, "LOW      "(%2)            \n\t"
460         :"=&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
461         :"r"(state), "r"(c)
462         : "%"REG_c, "%ebx", "%edx", "%"REG_S, "memory"
463     );
464     bit&=1;
465 #else /* BRANCHLESS_CABAC_DECODER */
466
467
468 #if defined CMOV_IS_FAST
469 #define BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
470         "mov    "tmp"       , %%ecx                                     \n\t"\
471         "shl    $17         , "tmp"                                     \n\t"\
472         "cmp    "low"       , "tmp"                                     \n\t"\
473         "cmova  %%ecx       , "range"                                   \n\t"\
474         "sbb    %%ecx       , %%ecx                                     \n\t"\
475         "and    %%ecx       , "tmp"                                     \n\t"\
476         "sub    "tmp"       , "low"                                     \n\t"\
477         "xor    %%ecx       , "ret"                                     \n\t"
478 #else /* CMOV_IS_FAST */
479 #define BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
480         "mov    "tmp"       , %%ecx                                     \n\t"\
481         "shl    $17         , "tmp"                                     \n\t"\
482         "sub    "low"       , "tmp"                                     \n\t"\
483         "sar    $31         , "tmp"                                     \n\t" /*lps_mask*/\
484         "sub    %%ecx       , "range"                                   \n\t" /*RangeLPS - range*/\
485         "and    "tmp"       , "range"                                   \n\t" /*(RangeLPS - range)&lps_mask*/\
486         "add    %%ecx       , "range"                                   \n\t" /*new range*/\
487         "shl    $17         , %%ecx                                     \n\t"\
488         "and    "tmp"       , %%ecx                                     \n\t"\
489         "sub    %%ecx       , "low"                                     \n\t"\
490         "xor    "tmp"       , "ret"                                     \n\t"
491 #endif /* CMOV_IS_FAST */
492
493
494 #define BRANCHLESS_GET_CABAC(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
495         "movzbl "statep"    , "ret"                                     \n\t"\
496         "mov    "range"     , "tmp"                                     \n\t"\
497         "and    $0xC0       , "range"                                   \n\t"\
498         "movzbl "MANGLE(ff_h264_lps_range)"("ret", "range", 2), "range" \n\t"\
499         "sub    "range"     , "tmp"                                     \n\t"\
500         BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
501         "movzbl " MANGLE(ff_h264_norm_shift) "("range"), %%ecx          \n\t"\
502         "shl    %%cl        , "range"                                   \n\t"\
503         "movzbl "MANGLE(ff_h264_mlps_state)"+128("ret"), "tmp"          \n\t"\
504         "mov    "tmpbyte"   , "statep"                                  \n\t"\
505         "shl    %%cl        , "low"                                     \n\t"\
506         "test   "lowword"   , "lowword"                                 \n\t"\
507         " jnz   1f                                                      \n\t"\
508         "mov "BYTE"("cabac"), %%"REG_c"                                 \n\t"\
509         "movzwl (%%"REG_c")     , "tmp"                                 \n\t"\
510         "bswap  "tmp"                                                   \n\t"\
511         "shr    $15         , "tmp"                                     \n\t"\
512         "sub    $0xFFFF     , "tmp"                                     \n\t"\
513         "add    $2          , %%"REG_c"                                 \n\t"\
514         "mov    %%"REG_c"   , "BYTE    "("cabac")                       \n\t"\
515         "lea    -1("low")   , %%ecx                                     \n\t"\
516         "xor    "low"       , %%ecx                                     \n\t"\
517         "shr    $15         , %%ecx                                     \n\t"\
518         "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx            \n\t"\
519         "neg    %%ecx                                                   \n\t"\
520         "add    $7          , %%ecx                                     \n\t"\
521         "shl    %%cl        , "tmp"                                     \n\t"\
522         "add    "tmp"       , "low"                                     \n\t"\
523         "1:                                                             \n\t"
524
525     asm volatile(
526         "movl "RANGE    "(%2), %%esi            \n\t"
527         "movl "LOW      "(%2), %%ebx            \n\t"
528         BRANCHLESS_GET_CABAC("%0", "%2", "(%1)", "%%ebx", "%%bx", "%%esi", "%%edx", "%%dl")
529         "movl %%esi, "RANGE    "(%2)            \n\t"
530         "movl %%ebx, "LOW      "(%2)            \n\t"
531
532         :"=&a"(bit)
533         :"r"(state), "r"(c)
534         : "%"REG_c, "%ebx", "%edx", "%esi", "memory"
535     );
536     bit&=1;
537 #endif /* BRANCHLESS_CABAC_DECODER */
538 #else /* defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__)) */
539     int s = *state;
540     int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + s];
541     int bit, lps_mask attribute_unused;
542
543     c->range -= RangeLPS;
544 #ifndef BRANCHLESS_CABAC_DECODER
545     if(c->low < (c->range<<17)){
546         bit= s&1;
547         *state= ff_h264_mps_state[s];
548         renorm_cabac_decoder_once(c);
549     }else{
550         bit= ff_h264_norm_shift[RangeLPS];
551         c->low -= (c->range<<17);
552         *state= ff_h264_lps_state[s];
553         c->range = RangeLPS<<bit;
554         c->low <<= bit;
555         bit= (s&1)^1;
556
557         if(!(c->low & 0xFFFF)){
558             refill2(c);
559         }
560     }
561 #else /* BRANCHLESS_CABAC_DECODER */
562     lps_mask= ((c->range<<17) - c->low)>>31;
563
564     c->low -= (c->range<<17) & lps_mask;
565     c->range += (RangeLPS - c->range) & lps_mask;
566
567     s^=lps_mask;
568     *state= (ff_h264_mlps_state+128)[s];
569     bit= s&1;
570
571     lps_mask= ff_h264_norm_shift[c->range];
572     c->range<<= lps_mask;
573     c->low  <<= lps_mask;
574     if(!(c->low & CABAC_MASK))
575         refill2(c);
576 #endif /* BRANCHLESS_CABAC_DECODER */
577 #endif /* defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__)) */
578     return bit;
579 }
580
581 static int __attribute((noinline)) get_cabac_noinline(CABACContext *c, uint8_t * const state){
582     return get_cabac_inline(c,state);
583 }
584
585 static int get_cabac(CABACContext *c, uint8_t * const state){
586     return get_cabac_inline(c,state);
587 }
588
589 static int get_cabac_bypass(CABACContext *c){
590 #if 0 //not faster
591     int bit;
592     asm volatile(
593         "movl "RANGE    "(%1), %%ebx            \n\t"
594         "movl "LOW      "(%1), %%eax            \n\t"
595         "shl $17, %%ebx                         \n\t"
596         "add %%eax, %%eax                       \n\t"
597         "sub %%ebx, %%eax                       \n\t"
598         "cdq                                    \n\t"
599         "and %%edx, %%ebx                       \n\t"
600         "add %%ebx, %%eax                       \n\t"
601         "test %%ax, %%ax                        \n\t"
602         " jnz 1f                                \n\t"
603         "movl "BYTE     "(%1), %%"REG_b"        \n\t"
604         "subl $0xFFFF, %%eax                    \n\t"
605         "movzwl (%%"REG_b"), %%ecx              \n\t"
606         "bswap %%ecx                            \n\t"
607         "shrl $15, %%ecx                        \n\t"
608         "addl $2, %%"REG_b"                     \n\t"
609         "addl %%ecx, %%eax                      \n\t"
610         "movl %%"REG_b", "BYTE     "(%1)        \n\t"
611         "1:                                     \n\t"
612         "movl %%eax, "LOW      "(%1)            \n\t"
613
614         :"=&d"(bit)
615         :"r"(c)
616         : "%eax", "%"REG_b, "%ecx", "memory"
617     );
618     return bit+1;
619 #else
620     int range;
621     c->low += c->low;
622
623     if(!(c->low & CABAC_MASK))
624         refill(c);
625
626     range= c->range<<17;
627     if(c->low < range){
628         return 0;
629     }else{
630         c->low -= range;
631         return 1;
632     }
633 #endif
634 }
635
636
637 static always_inline int get_cabac_bypass_sign(CABACContext *c, int val){
638 #ifdef ARCH_X86
639     asm volatile(
640         "movl "RANGE    "(%1), %%ebx            \n\t"
641         "movl "LOW      "(%1), %%eax            \n\t"
642         "shl $17, %%ebx                         \n\t"
643         "add %%eax, %%eax                       \n\t"
644         "sub %%ebx, %%eax                       \n\t"
645         "cdq                                    \n\t"
646         "and %%edx, %%ebx                       \n\t"
647         "add %%ebx, %%eax                       \n\t"
648         "xor %%edx, %%ecx                       \n\t"
649         "sub %%edx, %%ecx                       \n\t"
650         "test %%ax, %%ax                        \n\t"
651         " jnz 1f                                \n\t"
652         "mov  "BYTE     "(%1), %%"REG_b"        \n\t"
653         "subl $0xFFFF, %%eax                    \n\t"
654         "movzwl (%%"REG_b"), %%edx              \n\t"
655         "bswap %%edx                            \n\t"
656         "shrl $15, %%edx                        \n\t"
657         "add  $2, %%"REG_b"                     \n\t"
658         "addl %%edx, %%eax                      \n\t"
659         "mov  %%"REG_b", "BYTE     "(%1)        \n\t"
660         "1:                                     \n\t"
661         "movl %%eax, "LOW      "(%1)            \n\t"
662
663         :"+c"(val)
664         :"r"(c)
665         : "%eax", "%"REG_b, "%edx", "memory"
666     );
667     return val;
668 #else
669     int range, mask;
670     c->low += c->low;
671
672     if(!(c->low & CABAC_MASK))
673         refill(c);
674
675     range= c->range<<17;
676     c->low -= range;
677     mask= c->low >> 31;
678     range &= mask;
679     c->low += range;
680     return (val^mask)-mask;
681 #endif
682 }
683
684 //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!)
685 //FIXME use some macros to avoid duplicatin get_cabac (cant be done yet as that would make optimization work hard)
686 #if defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__))
687 static int decode_significance_x86(CABACContext *c, int max_coeff, uint8_t *significant_coeff_ctx_base, int *index){
688     void *end= significant_coeff_ctx_base + max_coeff - 1;
689     int minusstart= -(int)significant_coeff_ctx_base;
690     int minusindex= 4-(int)index;
691     int coeff_count;
692     asm volatile(
693         "movl "RANGE    "(%3), %%esi            \n\t"
694         "movl "LOW      "(%3), %%ebx            \n\t"
695
696         "2:                                     \n\t"
697
698         BRANCHLESS_GET_CABAC("%%edx", "%3", "(%1)", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
699
700         "test $1, %%edx                         \n\t"
701         " jz 3f                                 \n\t"
702
703         BRANCHLESS_GET_CABAC("%%edx", "%3", "61(%1)", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
704
705         "mov  %2, %%"REG_a"                     \n\t"
706         "movl %4, %%ecx                         \n\t"
707         "add  %1, %%"REG_c"                     \n\t"
708         "movl %%ecx, (%%"REG_a")                \n\t"
709
710         "test $1, %%edx                         \n\t"
711         " jnz 4f                                \n\t"
712
713         "add  $4, %%"REG_a"                     \n\t"
714         "mov  %%"REG_a", %2                     \n\t"
715
716         "3:                                     \n\t"
717         "add  $1, %1                            \n\t"
718         "cmp  %5, %1                            \n\t"
719         " jb 2b                                 \n\t"
720         "mov  %2, %%"REG_a"                     \n\t"
721         "movl %4, %%ecx                         \n\t"
722         "add  %1, %%"REG_c"                     \n\t"
723         "movl %%ecx, (%%"REG_a")                \n\t"
724         "4:                                     \n\t"
725         "add  %6, %%eax                         \n\t"
726         "shr $2, %%eax                          \n\t"
727
728         "movl %%esi, "RANGE    "(%3)            \n\t"
729         "movl %%ebx, "LOW      "(%3)            \n\t"
730         :"=&a"(coeff_count), "+r"(significant_coeff_ctx_base), "+m"(index)\
731         :"r"(c), "m"(minusstart), "m"(end), "m"(minusindex)\
732         : "%"REG_c, "%ebx", "%edx", "%esi", "memory"\
733     );
734     return coeff_count;
735 }
736
737 static int decode_significance_8x8_x86(CABACContext *c, uint8_t *significant_coeff_ctx_base, int *index, uint8_t *sig_off){
738     int minusindex= 4-(int)index;
739     int coeff_count;
740     long last=0;
741     asm volatile(
742         "movl "RANGE    "(%3), %%esi            \n\t"
743         "movl "LOW      "(%3), %%ebx            \n\t"
744
745         "mov %1, %%"REG_D"                      \n\t"
746         "2:                                     \n\t"
747
748         "mov %6, %%"REG_a"                      \n\t"
749         "movzbl (%%"REG_a", %%"REG_D"), %%edi   \n\t"
750         "add %5, %%"REG_D"                      \n\t"
751
752         BRANCHLESS_GET_CABAC("%%edx", "%3", "(%%"REG_D")", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
753
754         "mov %1, %%edi                          \n\t"
755         "test $1, %%edx                         \n\t"
756         " jz 3f                                 \n\t"
757
758         "movzbl "MANGLE(last_coeff_flag_offset_8x8)"(%%edi), %%edi\n\t"
759         "add %5, %%"REG_D"                      \n\t"
760
761         BRANCHLESS_GET_CABAC("%%edx", "%3", "15(%%"REG_D")", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
762
763         "mov %2, %%"REG_a"                      \n\t"
764         "mov %1, %%edi                          \n\t"
765         "movl %%edi, (%%"REG_a")                \n\t"
766
767         "test $1, %%edx                         \n\t"
768         " jnz 4f                                \n\t"
769
770         "add $4, %%"REG_a"                      \n\t"
771         "mov %%"REG_a", %2                      \n\t"
772
773         "3:                                     \n\t"
774         "addl $1, %%edi                         \n\t"
775         "mov %%edi, %1                          \n\t"
776         "cmpl $63, %%edi                        \n\t"
777         " jb 2b                                 \n\t"
778         "mov %2, %%"REG_a"                      \n\t"
779         "movl %%edi, (%%"REG_a")                \n\t"
780         "4:                                     \n\t"
781         "addl %4, %%eax                         \n\t"
782         "shr $2, %%eax                          \n\t"
783
784         "movl %%esi, "RANGE    "(%3)            \n\t"
785         "movl %%ebx, "LOW      "(%3)            \n\t"
786         :"=&a"(coeff_count),"+m"(last), "+m"(index)\
787         :"r"(c), "m"(minusindex), "m"(significant_coeff_ctx_base), "m"(sig_off)\
788         : "%"REG_c, "%ebx", "%edx", "%esi", "%"REG_D, "memory"\
789     );
790     return coeff_count;
791 }
792 #endif /* defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__)) */
793
794 /**
795  *
796  * @return the number of bytes read or 0 if no end
797  */
798 static int get_cabac_terminate(CABACContext *c){
799     c->range -= 2;
800     if(c->low < c->range<<17){
801         renorm_cabac_decoder_once(c);
802         return 0;
803     }else{
804         return c->bytestream - c->bytestream_start;
805     }
806 }
807
808 /**
809  * get (truncated) unnary binarization.
810  */
811 static int get_cabac_u(CABACContext *c, uint8_t * state, int max, int max_index, int truncated){
812     int i;
813
814     for(i=0; i<max; i++){
815         if(get_cabac(c, state)==0)
816             return i;
817
818         if(i< max_index) state++;
819     }
820
821     return truncated ? max : -1;
822 }
823
824 /**
825  * get unary exp golomb k-th order binarization.
826  */
827 static int get_cabac_ueg(CABACContext *c, uint8_t * state, int max, int is_signed, int k, int max_index){
828     int i, v;
829     int m= 1<<k;
830
831     if(get_cabac(c, state)==0)
832         return 0;
833
834     if(0 < max_index) state++;
835
836     for(i=1; i<max; i++){
837         if(get_cabac(c, state)==0){
838             if(is_signed && get_cabac_bypass(c)){
839                 return -i;
840             }else
841                 return i;
842         }
843
844         if(i < max_index) state++;
845     }
846
847     while(get_cabac_bypass(c)){
848         i+= m;
849         m+= m;
850     }
851
852     v=0;
853     while(m>>=1){
854         v+= v + get_cabac_bypass(c);
855     }
856     i += v;
857
858     if(is_signed && get_cabac_bypass(c)){
859         return -i;
860     }else
861         return i;
862 }