split h264.c to move parser in its own file
[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 #ifndef CABAC_H
29 #define CABAC_H
30
31 #include "bitstream.h"
32
33 //#undef NDEBUG
34 #include <assert.h>
35 #ifdef ARCH_X86
36 #include "x86_cpu.h"
37 #endif
38
39 #define CABAC_BITS 16
40 #define CABAC_MASK ((1<<CABAC_BITS)-1)
41 #define BRANCHLESS_CABAC_DECODER 1
42 //#define ARCH_X86_DISABLED 1
43
44 typedef struct CABACContext{
45     int low;
46     int range;
47     int outstanding_count;
48 #ifdef STRICT_LIMITS
49     int symCount;
50 #endif
51     const uint8_t *bytestream_start;
52     const uint8_t *bytestream;
53     const uint8_t *bytestream_end;
54     PutBitContext pb;
55 }CABACContext;
56
57 extern uint8_t ff_h264_mlps_state[4*64];
58 extern uint8_t ff_h264_lps_range[4*2*64];  ///< rangeTabLPS
59 extern uint8_t ff_h264_mps_state[2*64];     ///< transIdxMPS
60 extern uint8_t ff_h264_lps_state[2*64];     ///< transIdxLPS
61 extern const uint8_t ff_h264_norm_shift[512];
62
63
64 void ff_init_cabac_encoder(CABACContext *c, uint8_t *buf, int buf_size);
65 void ff_init_cabac_decoder(CABACContext *c, const uint8_t *buf, int buf_size);
66 void ff_init_cabac_states(CABACContext *c);
67
68
69 static inline void put_cabac_bit(CABACContext *c, int b){
70     put_bits(&c->pb, 1, b);
71     for(;c->outstanding_count; c->outstanding_count--){
72         put_bits(&c->pb, 1, 1-b);
73     }
74 }
75
76 static inline void renorm_cabac_encoder(CABACContext *c){
77     while(c->range < 0x100){
78         //FIXME optimize
79         if(c->low<0x100){
80             put_cabac_bit(c, 0);
81         }else if(c->low<0x200){
82             c->outstanding_count++;
83             c->low -= 0x100;
84         }else{
85             put_cabac_bit(c, 1);
86             c->low -= 0x200;
87         }
88
89         c->range+= c->range;
90         c->low += c->low;
91     }
92 }
93
94 static void put_cabac(CABACContext *c, uint8_t * const state, int bit){
95     int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + *state];
96
97     if(bit == ((*state)&1)){
98         c->range -= RangeLPS;
99         *state= ff_h264_mps_state[*state];
100     }else{
101         c->low += c->range - RangeLPS;
102         c->range = RangeLPS;
103         *state= ff_h264_lps_state[*state];
104     }
105
106     renorm_cabac_encoder(c);
107
108 #ifdef STRICT_LIMITS
109     c->symCount++;
110 #endif
111 }
112
113 static void put_cabac_static(CABACContext *c, int RangeLPS, int bit){
114     assert(c->range > RangeLPS);
115
116     if(!bit){
117         c->range -= RangeLPS;
118     }else{
119         c->low += c->range - RangeLPS;
120         c->range = RangeLPS;
121     }
122
123     renorm_cabac_encoder(c);
124
125 #ifdef STRICT_LIMITS
126     c->symCount++;
127 #endif
128 }
129
130 /**
131  * @param bit 0 -> write zero bit, !=0 write one bit
132  */
133 static void put_cabac_bypass(CABACContext *c, int bit){
134     c->low += c->low;
135
136     if(bit){
137         c->low += c->range;
138     }
139 //FIXME optimize
140     if(c->low<0x200){
141         put_cabac_bit(c, 0);
142     }else if(c->low<0x400){
143         c->outstanding_count++;
144         c->low -= 0x200;
145     }else{
146         put_cabac_bit(c, 1);
147         c->low -= 0x400;
148     }
149
150 #ifdef STRICT_LIMITS
151     c->symCount++;
152 #endif
153 }
154
155 /**
156  *
157  * @return the number of bytes written
158  */
159 static int put_cabac_terminate(CABACContext *c, int bit){
160     c->range -= 2;
161
162     if(!bit){
163         renorm_cabac_encoder(c);
164     }else{
165         c->low += c->range;
166         c->range= 2;
167
168         renorm_cabac_encoder(c);
169
170         assert(c->low <= 0x1FF);
171         put_cabac_bit(c, c->low>>9);
172         put_bits(&c->pb, 2, ((c->low>>7)&3)|1);
173
174         flush_put_bits(&c->pb); //FIXME FIXME FIXME XXX wrong
175     }
176
177 #ifdef STRICT_LIMITS
178     c->symCount++;
179 #endif
180
181     return (put_bits_count(&c->pb)+7)>>3;
182 }
183
184 /**
185  * put (truncated) unary binarization.
186  */
187 static void put_cabac_u(CABACContext *c, uint8_t * state, int v, int max, int max_index, int truncated){
188     int i;
189
190     assert(v <= max);
191
192 #if 1
193     for(i=0; i<v; i++){
194         put_cabac(c, state, 1);
195         if(i < max_index) state++;
196     }
197     if(truncated==0 || v<max)
198         put_cabac(c, state, 0);
199 #else
200     if(v <= max_index){
201         for(i=0; i<v; i++){
202             put_cabac(c, state+i, 1);
203         }
204         if(truncated==0 || v<max)
205             put_cabac(c, state+i, 0);
206     }else{
207         for(i=0; i<=max_index; i++){
208             put_cabac(c, state+i, 1);
209         }
210         for(; i<v; i++){
211             put_cabac(c, state+max_index, 1);
212         }
213         if(truncated==0 || v<max)
214             put_cabac(c, state+max_index, 0);
215     }
216 #endif
217 }
218
219 /**
220  * put unary exp golomb k-th order binarization.
221  */
222 static void put_cabac_ueg(CABACContext *c, uint8_t * state, int v, int max, int is_signed, int k, int max_index){
223     int i;
224
225     if(v==0)
226         put_cabac(c, state, 0);
227     else{
228         const int sign= v < 0;
229
230         if(is_signed) v= FFABS(v);
231
232         if(v<max){
233             for(i=0; i<v; i++){
234                 put_cabac(c, state, 1);
235                 if(i < max_index) state++;
236             }
237
238             put_cabac(c, state, 0);
239         }else{
240             int m= 1<<k;
241
242             for(i=0; i<max; i++){
243                 put_cabac(c, state, 1);
244                 if(i < max_index) state++;
245             }
246
247             v -= max;
248             while(v >= m){ //FIXME optimize
249                 put_cabac_bypass(c, 1);
250                 v-= m;
251                 m+= m;
252             }
253             put_cabac_bypass(c, 0);
254             while(m>>=1){
255                 put_cabac_bypass(c, v&m);
256             }
257         }
258
259         if(is_signed)
260             put_cabac_bypass(c, sign);
261     }
262 }
263
264 static void refill(CABACContext *c){
265 #if CABAC_BITS == 16
266         c->low+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1);
267 #else
268         c->low+= c->bytestream[0]<<1;
269 #endif
270     c->low -= CABAC_MASK;
271     c->bytestream+= CABAC_BITS/8;
272 }
273
274 static void refill2(CABACContext *c){
275     int i, x;
276
277     x= c->low ^ (c->low-1);
278     i= 7 - ff_h264_norm_shift[x>>(CABAC_BITS-1)];
279
280     x= -CABAC_MASK;
281
282 #if CABAC_BITS == 16
283         x+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1);
284 #else
285         x+= c->bytestream[0]<<1;
286 #endif
287
288     c->low += x<<i;
289     c->bytestream+= CABAC_BITS/8;
290 }
291
292 static inline void renorm_cabac_decoder(CABACContext *c){
293     while(c->range < 0x100){
294         c->range+= c->range;
295         c->low+= c->low;
296         if(!(c->low & CABAC_MASK))
297             refill(c);
298     }
299 }
300
301 static inline void renorm_cabac_decoder_once(CABACContext *c){
302 #ifdef ARCH_X86_DISABLED
303     int temp;
304 #if 0
305     //P3:683    athlon:475
306     asm(
307         "lea -0x100(%0), %2         \n\t"
308         "shr $31, %2                \n\t"  //FIXME 31->63 for x86-64
309         "shl %%cl, %0               \n\t"
310         "shl %%cl, %1               \n\t"
311         : "+r"(c->range), "+r"(c->low), "+c"(temp)
312     );
313 #elif 0
314     //P3:680    athlon:474
315     asm(
316         "cmp $0x100, %0             \n\t"
317         "setb %%cl                  \n\t"  //FIXME 31->63 for x86-64
318         "shl %%cl, %0               \n\t"
319         "shl %%cl, %1               \n\t"
320         : "+r"(c->range), "+r"(c->low), "+c"(temp)
321     );
322 #elif 1
323     int temp2;
324     //P3:665    athlon:517
325     asm(
326         "lea -0x100(%0), %%eax      \n\t"
327         "cdq                        \n\t"
328         "mov %0, %%eax              \n\t"
329         "and %%edx, %0              \n\t"
330         "and %1, %%edx              \n\t"
331         "add %%eax, %0              \n\t"
332         "add %%edx, %1              \n\t"
333         : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
334     );
335 #elif 0
336     int temp2;
337     //P3:673    athlon:509
338     asm(
339         "cmp $0x100, %0             \n\t"
340         "sbb %%edx, %%edx           \n\t"
341         "mov %0, %%eax              \n\t"
342         "and %%edx, %0              \n\t"
343         "and %1, %%edx              \n\t"
344         "add %%eax, %0              \n\t"
345         "add %%edx, %1              \n\t"
346         : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
347     );
348 #else
349     int temp2;
350     //P3:677    athlon:511
351     asm(
352         "cmp $0x100, %0             \n\t"
353         "lea (%0, %0), %%eax        \n\t"
354         "lea (%1, %1), %%edx        \n\t"
355         "cmovb %%eax, %0            \n\t"
356         "cmovb %%edx, %1            \n\t"
357         : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
358     );
359 #endif
360 #else
361     //P3:675    athlon:476
362     int shift= (uint32_t)(c->range - 0x100)>>31;
363     c->range<<= shift;
364     c->low  <<= shift;
365 #endif
366     if(!(c->low & CABAC_MASK))
367         refill(c);
368 }
369
370 static av_always_inline int get_cabac_inline(CABACContext *c, uint8_t * const state){
371     //FIXME gcc generates duplicate load/stores for c->low and c->range
372 #define LOW          "0"
373 #define RANGE        "4"
374 #ifdef ARCH_X86_64
375 #define BYTESTART   "16"
376 #define BYTE        "24"
377 #define BYTEEND     "32"
378 #else
379 #define BYTESTART   "12"
380 #define BYTE        "16"
381 #define BYTEEND     "20"
382 #endif
383 #if defined(ARCH_X86) && defined(CONFIG_7REGS) && defined(HAVE_EBX_AVAILABLE) && !defined(BROKEN_RELOCATIONS)
384     int bit;
385
386 #ifndef BRANCHLESS_CABAC_DECODER
387     asm volatile(
388         "movzbl (%1), %0                        \n\t"
389         "movl "RANGE    "(%2), %%ebx            \n\t"
390         "movl "RANGE    "(%2), %%edx            \n\t"
391         "andl $0xC0, %%ebx                      \n\t"
392         "movzbl "MANGLE(ff_h264_lps_range)"(%0, %%ebx, 2), %%esi\n\t"
393         "movl "LOW      "(%2), %%ebx            \n\t"
394 //eax:state ebx:low, edx:range, esi:RangeLPS
395         "subl %%esi, %%edx                      \n\t"
396         "movl %%edx, %%ecx                      \n\t"
397         "shll $17, %%ecx                        \n\t"
398         "cmpl %%ecx, %%ebx                      \n\t"
399         " ja 1f                                 \n\t"
400
401 #if 1
402         //athlon:4067 P3:4110
403         "lea -0x100(%%edx), %%ecx               \n\t"
404         "shr $31, %%ecx                         \n\t"
405         "shl %%cl, %%edx                        \n\t"
406         "shl %%cl, %%ebx                        \n\t"
407 #else
408         //athlon:4057 P3:4130
409         "cmp $0x100, %%edx                      \n\t" //FIXME avoidable
410         "setb %%cl                              \n\t"
411         "shl %%cl, %%edx                        \n\t"
412         "shl %%cl, %%ebx                        \n\t"
413 #endif
414         "movzbl "MANGLE(ff_h264_mps_state)"(%0), %%ecx   \n\t"
415         "movb %%cl, (%1)                        \n\t"
416 //eax:state ebx:low, edx:range, esi:RangeLPS
417         "test %%bx, %%bx                        \n\t"
418         " jnz 2f                                \n\t"
419         "mov  "BYTE     "(%2), %%"REG_S"        \n\t"
420         "subl $0xFFFF, %%ebx                    \n\t"
421         "movzwl (%%"REG_S"), %%ecx              \n\t"
422         "bswap %%ecx                            \n\t"
423         "shrl $15, %%ecx                        \n\t"
424         "add  $2, %%"REG_S"                     \n\t"
425         "addl %%ecx, %%ebx                      \n\t"
426         "mov  %%"REG_S", "BYTE    "(%2)         \n\t"
427         "jmp 2f                                 \n\t"
428         "1:                                     \n\t"
429 //eax:state ebx:low, edx:range, esi:RangeLPS
430         "subl %%ecx, %%ebx                      \n\t"
431         "movl %%esi, %%edx                      \n\t"
432         "movzbl " MANGLE(ff_h264_norm_shift) "(%%esi), %%ecx   \n\t"
433         "shll %%cl, %%ebx                       \n\t"
434         "shll %%cl, %%edx                       \n\t"
435         "movzbl "MANGLE(ff_h264_lps_state)"(%0), %%ecx   \n\t"
436         "movb %%cl, (%1)                        \n\t"
437         "add  $1, %0                            \n\t"
438         "test %%bx, %%bx                        \n\t"
439         " jnz 2f                                \n\t"
440
441         "mov  "BYTE     "(%2), %%"REG_c"        \n\t"
442         "movzwl (%%"REG_c"), %%esi              \n\t"
443         "bswap %%esi                            \n\t"
444         "shrl $15, %%esi                        \n\t"
445         "subl $0xFFFF, %%esi                    \n\t"
446         "add  $2, %%"REG_c"                     \n\t"
447         "mov  %%"REG_c", "BYTE    "(%2)         \n\t"
448
449         "leal -1(%%ebx), %%ecx                  \n\t"
450         "xorl %%ebx, %%ecx                      \n\t"
451         "shrl $15, %%ecx                        \n\t"
452         "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx   \n\t"
453         "neg %%ecx                              \n\t"
454         "add $7, %%ecx                          \n\t"
455
456         "shll %%cl , %%esi                      \n\t"
457         "addl %%esi, %%ebx                      \n\t"
458         "2:                                     \n\t"
459         "movl %%edx, "RANGE    "(%2)            \n\t"
460         "movl %%ebx, "LOW      "(%2)            \n\t"
461         :"=&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
462         :"r"(state), "r"(c)
463         : "%"REG_c, "%ebx", "%edx", "%"REG_S, "memory"
464     );
465     bit&=1;
466 #else /* BRANCHLESS_CABAC_DECODER */
467
468
469 #if defined HAVE_FAST_CMOV
470 #define BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
471         "mov    "tmp"       , %%ecx                                     \n\t"\
472         "shl    $17         , "tmp"                                     \n\t"\
473         "cmp    "low"       , "tmp"                                     \n\t"\
474         "cmova  %%ecx       , "range"                                   \n\t"\
475         "sbb    %%ecx       , %%ecx                                     \n\t"\
476         "and    %%ecx       , "tmp"                                     \n\t"\
477         "sub    "tmp"       , "low"                                     \n\t"\
478         "xor    %%ecx       , "ret"                                     \n\t"
479 #else /* HAVE_FAST_CMOV */
480 #define BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
481         "mov    "tmp"       , %%ecx                                     \n\t"\
482         "shl    $17         , "tmp"                                     \n\t"\
483         "sub    "low"       , "tmp"                                     \n\t"\
484         "sar    $31         , "tmp"                                     \n\t" /*lps_mask*/\
485         "sub    %%ecx       , "range"                                   \n\t" /*RangeLPS - range*/\
486         "and    "tmp"       , "range"                                   \n\t" /*(RangeLPS - range)&lps_mask*/\
487         "add    %%ecx       , "range"                                   \n\t" /*new range*/\
488         "shl    $17         , %%ecx                                     \n\t"\
489         "and    "tmp"       , %%ecx                                     \n\t"\
490         "sub    %%ecx       , "low"                                     \n\t"\
491         "xor    "tmp"       , "ret"                                     \n\t"
492 #endif /* HAVE_FAST_CMOV */
493
494
495 #define BRANCHLESS_GET_CABAC(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
496         "movzbl "statep"    , "ret"                                     \n\t"\
497         "mov    "range"     , "tmp"                                     \n\t"\
498         "and    $0xC0       , "range"                                   \n\t"\
499         "movzbl "MANGLE(ff_h264_lps_range)"("ret", "range", 2), "range" \n\t"\
500         "sub    "range"     , "tmp"                                     \n\t"\
501         BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
502         "movzbl " MANGLE(ff_h264_norm_shift) "("range"), %%ecx          \n\t"\
503         "shl    %%cl        , "range"                                   \n\t"\
504         "movzbl "MANGLE(ff_h264_mlps_state)"+128("ret"), "tmp"          \n\t"\
505         "mov    "tmpbyte"   , "statep"                                  \n\t"\
506         "shl    %%cl        , "low"                                     \n\t"\
507         "test   "lowword"   , "lowword"                                 \n\t"\
508         " jnz   1f                                                      \n\t"\
509         "mov "BYTE"("cabac"), %%"REG_c"                                 \n\t"\
510         "movzwl (%%"REG_c")     , "tmp"                                 \n\t"\
511         "bswap  "tmp"                                                   \n\t"\
512         "shr    $15         , "tmp"                                     \n\t"\
513         "sub    $0xFFFF     , "tmp"                                     \n\t"\
514         "add    $2          , %%"REG_c"                                 \n\t"\
515         "mov    %%"REG_c"   , "BYTE    "("cabac")                       \n\t"\
516         "lea    -1("low")   , %%ecx                                     \n\t"\
517         "xor    "low"       , %%ecx                                     \n\t"\
518         "shr    $15         , %%ecx                                     \n\t"\
519         "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx            \n\t"\
520         "neg    %%ecx                                                   \n\t"\
521         "add    $7          , %%ecx                                     \n\t"\
522         "shl    %%cl        , "tmp"                                     \n\t"\
523         "add    "tmp"       , "low"                                     \n\t"\
524         "1:                                                             \n\t"
525
526     asm volatile(
527         "movl "RANGE    "(%2), %%esi            \n\t"
528         "movl "LOW      "(%2), %%ebx            \n\t"
529         BRANCHLESS_GET_CABAC("%0", "%2", "(%1)", "%%ebx", "%%bx", "%%esi", "%%edx", "%%dl")
530         "movl %%esi, "RANGE    "(%2)            \n\t"
531         "movl %%ebx, "LOW      "(%2)            \n\t"
532
533         :"=&a"(bit)
534         :"r"(state), "r"(c)
535         : "%"REG_c, "%ebx", "%edx", "%esi", "memory"
536     );
537     bit&=1;
538 #endif /* BRANCHLESS_CABAC_DECODER */
539 #else /* defined(ARCH_X86) && defined(CONFIG_7REGS) && defined(HAVE_EBX_AVAILABLE) && !defined(BROKEN_RELOCATIONS) */
540     int s = *state;
541     int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + s];
542     int bit, lps_mask attribute_unused;
543
544     c->range -= RangeLPS;
545 #ifndef BRANCHLESS_CABAC_DECODER
546     if(c->low < (c->range<<(CABAC_BITS+1))){
547         bit= s&1;
548         *state= ff_h264_mps_state[s];
549         renorm_cabac_decoder_once(c);
550     }else{
551         bit= ff_h264_norm_shift[RangeLPS];
552         c->low -= (c->range<<(CABAC_BITS+1));
553         *state= ff_h264_lps_state[s];
554         c->range = RangeLPS<<bit;
555         c->low <<= bit;
556         bit= (s&1)^1;
557
558         if(!(c->low & CABAC_MASK)){
559             refill2(c);
560         }
561     }
562 #else /* BRANCHLESS_CABAC_DECODER */
563     lps_mask= ((c->range<<(CABAC_BITS+1)) - c->low)>>31;
564
565     c->low -= (c->range<<(CABAC_BITS+1)) & lps_mask;
566     c->range += (RangeLPS - c->range) & lps_mask;
567
568     s^=lps_mask;
569     *state= (ff_h264_mlps_state+128)[s];
570     bit= s&1;
571
572     lps_mask= ff_h264_norm_shift[c->range];
573     c->range<<= lps_mask;
574     c->low  <<= lps_mask;
575     if(!(c->low & CABAC_MASK))
576         refill2(c);
577 #endif /* BRANCHLESS_CABAC_DECODER */
578 #endif /* defined(ARCH_X86) && defined(CONFIG_7REGS) && defined(HAVE_EBX_AVAILABLE) && !defined(BROKEN_RELOCATIONS) */
579     return bit;
580 }
581
582 static int av_noinline get_cabac_noinline(CABACContext *c, uint8_t * const state){
583     return get_cabac_inline(c,state);
584 }
585
586 static int get_cabac(CABACContext *c, uint8_t * const state){
587     return get_cabac_inline(c,state);
588 }
589
590 static int get_cabac_bypass(CABACContext *c){
591 #if 0 //not faster
592     int bit;
593     asm volatile(
594         "movl "RANGE    "(%1), %%ebx            \n\t"
595         "movl "LOW      "(%1), %%eax            \n\t"
596         "shl $17, %%ebx                         \n\t"
597         "add %%eax, %%eax                       \n\t"
598         "sub %%ebx, %%eax                       \n\t"
599         "cdq                                    \n\t"
600         "and %%edx, %%ebx                       \n\t"
601         "add %%ebx, %%eax                       \n\t"
602         "test %%ax, %%ax                        \n\t"
603         " jnz 1f                                \n\t"
604         "movl "BYTE     "(%1), %%"REG_b"        \n\t"
605         "subl $0xFFFF, %%eax                    \n\t"
606         "movzwl (%%"REG_b"), %%ecx              \n\t"
607         "bswap %%ecx                            \n\t"
608         "shrl $15, %%ecx                        \n\t"
609         "addl $2, %%"REG_b"                     \n\t"
610         "addl %%ecx, %%eax                      \n\t"
611         "movl %%"REG_b", "BYTE     "(%1)        \n\t"
612         "1:                                     \n\t"
613         "movl %%eax, "LOW      "(%1)            \n\t"
614
615         :"=&d"(bit)
616         :"r"(c)
617         : "%eax", "%"REG_b, "%ecx", "memory"
618     );
619     return bit+1;
620 #else
621     int range;
622     c->low += c->low;
623
624     if(!(c->low & CABAC_MASK))
625         refill(c);
626
627     range= c->range<<(CABAC_BITS+1);
628     if(c->low < range){
629         return 0;
630     }else{
631         c->low -= range;
632         return 1;
633     }
634 #endif
635 }
636
637
638 static av_always_inline int get_cabac_bypass_sign(CABACContext *c, int val){
639 #if defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__))
640     asm volatile(
641         "movl "RANGE    "(%1), %%ebx            \n\t"
642         "movl "LOW      "(%1), %%eax            \n\t"
643         "shl $17, %%ebx                         \n\t"
644         "add %%eax, %%eax                       \n\t"
645         "sub %%ebx, %%eax                       \n\t"
646         "cdq                                    \n\t"
647         "and %%edx, %%ebx                       \n\t"
648         "add %%ebx, %%eax                       \n\t"
649         "xor %%edx, %%ecx                       \n\t"
650         "sub %%edx, %%ecx                       \n\t"
651         "test %%ax, %%ax                        \n\t"
652         " jnz 1f                                \n\t"
653         "mov  "BYTE     "(%1), %%"REG_b"        \n\t"
654         "subl $0xFFFF, %%eax                    \n\t"
655         "movzwl (%%"REG_b"), %%edx              \n\t"
656         "bswap %%edx                            \n\t"
657         "shrl $15, %%edx                        \n\t"
658         "add  $2, %%"REG_b"                     \n\t"
659         "addl %%edx, %%eax                      \n\t"
660         "mov  %%"REG_b", "BYTE     "(%1)        \n\t"
661         "1:                                     \n\t"
662         "movl %%eax, "LOW      "(%1)            \n\t"
663
664         :"+c"(val)
665         :"r"(c)
666         : "%eax", "%"REG_b, "%edx", "memory"
667     );
668     return val;
669 #else
670     int range, mask;
671     c->low += c->low;
672
673     if(!(c->low & CABAC_MASK))
674         refill(c);
675
676     range= c->range<<(CABAC_BITS+1);
677     c->low -= range;
678     mask= c->low >> 31;
679     range &= mask;
680     c->low += range;
681     return (val^mask)-mask;
682 #endif
683 }
684
685 //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!)
686 //FIXME use some macros to avoid duplicatin get_cabac (cant be done yet as that would make optimization work hard)
687 #if defined(ARCH_X86) && defined(CONFIG_7REGS) && defined(HAVE_EBX_AVAILABLE) && !defined(BROKEN_RELOCATIONS)
688 static int decode_significance_x86(CABACContext *c, int max_coeff, uint8_t *significant_coeff_ctx_base, int *index){
689     void *end= significant_coeff_ctx_base + max_coeff - 1;
690     int minusstart= -(int)significant_coeff_ctx_base;
691     int minusindex= 4-(int)index;
692     int coeff_count;
693     asm volatile(
694         "movl "RANGE    "(%3), %%esi            \n\t"
695         "movl "LOW      "(%3), %%ebx            \n\t"
696
697         "2:                                     \n\t"
698
699         BRANCHLESS_GET_CABAC("%%edx", "%3", "(%1)", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
700
701         "test $1, %%edx                         \n\t"
702         " jz 3f                                 \n\t"
703
704         BRANCHLESS_GET_CABAC("%%edx", "%3", "61(%1)", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
705
706         "mov  %2, %%"REG_a"                     \n\t"
707         "movl %4, %%ecx                         \n\t"
708         "add  %1, %%"REG_c"                     \n\t"
709         "movl %%ecx, (%%"REG_a")                \n\t"
710
711         "test $1, %%edx                         \n\t"
712         " jnz 4f                                \n\t"
713
714         "add  $4, %%"REG_a"                     \n\t"
715         "mov  %%"REG_a", %2                     \n\t"
716
717         "3:                                     \n\t"
718         "add  $1, %1                            \n\t"
719         "cmp  %5, %1                            \n\t"
720         " jb 2b                                 \n\t"
721         "mov  %2, %%"REG_a"                     \n\t"
722         "movl %4, %%ecx                         \n\t"
723         "add  %1, %%"REG_c"                     \n\t"
724         "movl %%ecx, (%%"REG_a")                \n\t"
725         "4:                                     \n\t"
726         "add  %6, %%eax                         \n\t"
727         "shr $2, %%eax                          \n\t"
728
729         "movl %%esi, "RANGE    "(%3)            \n\t"
730         "movl %%ebx, "LOW      "(%3)            \n\t"
731         :"=&a"(coeff_count), "+r"(significant_coeff_ctx_base), "+m"(index)\
732         :"r"(c), "m"(minusstart), "m"(end), "m"(minusindex)\
733         : "%"REG_c, "%ebx", "%edx", "%esi", "memory"\
734     );
735     return coeff_count;
736 }
737
738 static int decode_significance_8x8_x86(CABACContext *c, uint8_t *significant_coeff_ctx_base, int *index, uint8_t *sig_off){
739     int minusindex= 4-(int)index;
740     int coeff_count;
741     long last=0;
742     asm volatile(
743         "movl "RANGE    "(%3), %%esi            \n\t"
744         "movl "LOW      "(%3), %%ebx            \n\t"
745
746         "mov %1, %%"REG_D"                      \n\t"
747         "2:                                     \n\t"
748
749         "mov %6, %%"REG_a"                      \n\t"
750         "movzbl (%%"REG_a", %%"REG_D"), %%edi   \n\t"
751         "add %5, %%"REG_D"                      \n\t"
752
753         BRANCHLESS_GET_CABAC("%%edx", "%3", "(%%"REG_D")", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
754
755         "mov %1, %%edi                          \n\t"
756         "test $1, %%edx                         \n\t"
757         " jz 3f                                 \n\t"
758
759         "movzbl "MANGLE(last_coeff_flag_offset_8x8)"(%%edi), %%edi\n\t"
760         "add %5, %%"REG_D"                      \n\t"
761
762         BRANCHLESS_GET_CABAC("%%edx", "%3", "15(%%"REG_D")", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
763
764         "mov %2, %%"REG_a"                      \n\t"
765         "mov %1, %%edi                          \n\t"
766         "movl %%edi, (%%"REG_a")                \n\t"
767
768         "test $1, %%edx                         \n\t"
769         " jnz 4f                                \n\t"
770
771         "add $4, %%"REG_a"                      \n\t"
772         "mov %%"REG_a", %2                      \n\t"
773
774         "3:                                     \n\t"
775         "addl $1, %%edi                         \n\t"
776         "mov %%edi, %1                          \n\t"
777         "cmpl $63, %%edi                        \n\t"
778         " jb 2b                                 \n\t"
779         "mov %2, %%"REG_a"                      \n\t"
780         "movl %%edi, (%%"REG_a")                \n\t"
781         "4:                                     \n\t"
782         "addl %4, %%eax                         \n\t"
783         "shr $2, %%eax                          \n\t"
784
785         "movl %%esi, "RANGE    "(%3)            \n\t"
786         "movl %%ebx, "LOW      "(%3)            \n\t"
787         :"=&a"(coeff_count),"+m"(last), "+m"(index)\
788         :"r"(c), "m"(minusindex), "m"(significant_coeff_ctx_base), "m"(sig_off)\
789         : "%"REG_c, "%ebx", "%edx", "%esi", "%"REG_D, "memory"\
790     );
791     return coeff_count;
792 }
793 #endif /* defined(ARCH_X86) && && defined(CONFIG_7REGS) && defined(HAVE_EBX_AVAILABLE) && !defined(BROKEN_RELOCATIONS) */
794
795 /**
796  *
797  * @return the number of bytes read or 0 if no end
798  */
799 static int get_cabac_terminate(CABACContext *c){
800     c->range -= 2;
801     if(c->low < c->range<<(CABAC_BITS+1)){
802         renorm_cabac_decoder_once(c);
803         return 0;
804     }else{
805         return c->bytestream - c->bytestream_start;
806     }
807 }
808
809 /**
810  * get (truncated) unnary binarization.
811  */
812 static int get_cabac_u(CABACContext *c, uint8_t * state, int max, int max_index, int truncated){
813     int i;
814
815     for(i=0; i<max; i++){
816         if(get_cabac(c, state)==0)
817             return i;
818
819         if(i< max_index) state++;
820     }
821
822     return truncated ? max : -1;
823 }
824
825 /**
826  * get unary exp golomb k-th order binarization.
827  */
828 static int get_cabac_ueg(CABACContext *c, uint8_t * state, int max, int is_signed, int k, int max_index){
829     int i, v;
830     int m= 1<<k;
831
832     if(get_cabac(c, state)==0)
833         return 0;
834
835     if(0 < max_index) state++;
836
837     for(i=1; i<max; i++){
838         if(get_cabac(c, state)==0){
839             if(is_signed && get_cabac_bypass(c)){
840                 return -i;
841             }else
842                 return i;
843         }
844
845         if(i < max_index) state++;
846     }
847
848     while(get_cabac_bypass(c)){
849         i+= m;
850         m+= m;
851     }
852
853     v=0;
854     while(m>>=1){
855         v+= v + get_cabac_bypass(c);
856     }
857     i += v;
858
859     if(is_signed && get_cabac_bypass(c)){
860         return -i;
861     }else
862         return i;
863 }
864
865 #endif /* CABAC_H */