Use cltd instead of cdq asm mnemonic, ICC and gcc support both, but
[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  * @file cabac.h
24  * Context Adaptive Binary Arithmetic Coder.
25  */
26
27 #ifndef FFMPEG_CABAC_H
28 #define FFMPEG_CABAC_H
29
30 #include "bitstream.h"
31
32 //#undef NDEBUG
33 #include <assert.h>
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 #ifdef TEST
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 #endif /* TEST */
264
265 static void refill(CABACContext *c){
266 #if CABAC_BITS == 16
267         c->low+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1);
268 #else
269         c->low+= c->bytestream[0]<<1;
270 #endif
271     c->low -= CABAC_MASK;
272     c->bytestream+= CABAC_BITS/8;
273 }
274
275 #if ! ( defined(ARCH_X86) && defined(HAVE_7REGS) && defined(HAVE_EBX_AVAILABLE) && !defined(BROKEN_RELOCATIONS) )
276 static void refill2(CABACContext *c){
277     int i, x;
278
279     x= c->low ^ (c->low-1);
280     i= 7 - ff_h264_norm_shift[x>>(CABAC_BITS-1)];
281
282     x= -CABAC_MASK;
283
284 #if CABAC_BITS == 16
285         x+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1);
286 #else
287         x+= c->bytestream[0]<<1;
288 #endif
289
290     c->low += x<<i;
291     c->bytestream+= CABAC_BITS/8;
292 }
293 #endif
294
295 static inline void renorm_cabac_decoder(CABACContext *c){
296     while(c->range < 0x100){
297         c->range+= c->range;
298         c->low+= c->low;
299         if(!(c->low & CABAC_MASK))
300             refill(c);
301     }
302 }
303
304 static inline void renorm_cabac_decoder_once(CABACContext *c){
305 #ifdef ARCH_X86_DISABLED
306     int temp;
307 #if 0
308     //P3:683    athlon:475
309     asm(
310         "lea -0x100(%0), %2         \n\t"
311         "shr $31, %2                \n\t"  //FIXME 31->63 for x86-64
312         "shl %%cl, %0               \n\t"
313         "shl %%cl, %1               \n\t"
314         : "+r"(c->range), "+r"(c->low), "+c"(temp)
315     );
316 #elif 0
317     //P3:680    athlon:474
318     asm(
319         "cmp $0x100, %0             \n\t"
320         "setb %%cl                  \n\t"  //FIXME 31->63 for x86-64
321         "shl %%cl, %0               \n\t"
322         "shl %%cl, %1               \n\t"
323         : "+r"(c->range), "+r"(c->low), "+c"(temp)
324     );
325 #elif 1
326     int temp2;
327     //P3:665    athlon:517
328     asm(
329         "lea -0x100(%0), %%eax      \n\t"
330         "cltd                       \n\t"
331         "mov %0, %%eax              \n\t"
332         "and %%edx, %0              \n\t"
333         "and %1, %%edx              \n\t"
334         "add %%eax, %0              \n\t"
335         "add %%edx, %1              \n\t"
336         : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
337     );
338 #elif 0
339     int temp2;
340     //P3:673    athlon:509
341     asm(
342         "cmp $0x100, %0             \n\t"
343         "sbb %%edx, %%edx           \n\t"
344         "mov %0, %%eax              \n\t"
345         "and %%edx, %0              \n\t"
346         "and %1, %%edx              \n\t"
347         "add %%eax, %0              \n\t"
348         "add %%edx, %1              \n\t"
349         : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
350     );
351 #else
352     int temp2;
353     //P3:677    athlon:511
354     asm(
355         "cmp $0x100, %0             \n\t"
356         "lea (%0, %0), %%eax        \n\t"
357         "lea (%1, %1), %%edx        \n\t"
358         "cmovb %%eax, %0            \n\t"
359         "cmovb %%edx, %1            \n\t"
360         : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
361     );
362 #endif
363 #else
364     //P3:675    athlon:476
365     int shift= (uint32_t)(c->range - 0x100)>>31;
366     c->range<<= shift;
367     c->low  <<= shift;
368 #endif
369     if(!(c->low & CABAC_MASK))
370         refill(c);
371 }
372
373 static av_always_inline int get_cabac_inline(CABACContext *c, uint8_t * const state){
374     //FIXME gcc generates duplicate load/stores for c->low and c->range
375 #define LOW          "0"
376 #define RANGE        "4"
377 #ifdef ARCH_X86_64
378 #define BYTESTART   "16"
379 #define BYTE        "24"
380 #define BYTEEND     "32"
381 #else
382 #define BYTESTART   "12"
383 #define BYTE        "16"
384 #define BYTEEND     "20"
385 #endif
386 #if defined(ARCH_X86) && defined(HAVE_7REGS) && defined(HAVE_EBX_AVAILABLE) && !defined(BROKEN_RELOCATIONS)
387     int bit;
388
389 #ifndef BRANCHLESS_CABAC_DECODER
390     asm volatile(
391         "movzbl (%1), %0                        \n\t"
392         "movl "RANGE    "(%2), %%ebx            \n\t"
393         "movl "RANGE    "(%2), %%edx            \n\t"
394         "andl $0xC0, %%ebx                      \n\t"
395         "movzbl "MANGLE(ff_h264_lps_range)"(%0, %%ebx, 2), %%esi\n\t"
396         "movl "LOW      "(%2), %%ebx            \n\t"
397 //eax:state ebx:low, edx:range, esi:RangeLPS
398         "subl %%esi, %%edx                      \n\t"
399         "movl %%edx, %%ecx                      \n\t"
400         "shll $17, %%ecx                        \n\t"
401         "cmpl %%ecx, %%ebx                      \n\t"
402         " ja 1f                                 \n\t"
403
404 #if 1
405         //athlon:4067 P3:4110
406         "lea -0x100(%%edx), %%ecx               \n\t"
407         "shr $31, %%ecx                         \n\t"
408         "shl %%cl, %%edx                        \n\t"
409         "shl %%cl, %%ebx                        \n\t"
410 #else
411         //athlon:4057 P3:4130
412         "cmp $0x100, %%edx                      \n\t" //FIXME avoidable
413         "setb %%cl                              \n\t"
414         "shl %%cl, %%edx                        \n\t"
415         "shl %%cl, %%ebx                        \n\t"
416 #endif
417         "movzbl "MANGLE(ff_h264_mps_state)"(%0), %%ecx   \n\t"
418         "movb %%cl, (%1)                        \n\t"
419 //eax:state ebx:low, edx:range, esi:RangeLPS
420         "test %%bx, %%bx                        \n\t"
421         " jnz 2f                                \n\t"
422         "mov  "BYTE     "(%2), %%"REG_S"        \n\t"
423         "subl $0xFFFF, %%ebx                    \n\t"
424         "movzwl (%%"REG_S"), %%ecx              \n\t"
425         "bswap %%ecx                            \n\t"
426         "shrl $15, %%ecx                        \n\t"
427         "add  $2, %%"REG_S"                     \n\t"
428         "addl %%ecx, %%ebx                      \n\t"
429         "mov  %%"REG_S", "BYTE    "(%2)         \n\t"
430         "jmp 2f                                 \n\t"
431         "1:                                     \n\t"
432 //eax:state ebx:low, edx:range, esi:RangeLPS
433         "subl %%ecx, %%ebx                      \n\t"
434         "movl %%esi, %%edx                      \n\t"
435         "movzbl " MANGLE(ff_h264_norm_shift) "(%%esi), %%ecx   \n\t"
436         "shll %%cl, %%ebx                       \n\t"
437         "shll %%cl, %%edx                       \n\t"
438         "movzbl "MANGLE(ff_h264_lps_state)"(%0), %%ecx   \n\t"
439         "movb %%cl, (%1)                        \n\t"
440         "add  $1, %0                            \n\t"
441         "test %%bx, %%bx                        \n\t"
442         " jnz 2f                                \n\t"
443
444         "mov  "BYTE     "(%2), %%"REG_c"        \n\t"
445         "movzwl (%%"REG_c"), %%esi              \n\t"
446         "bswap %%esi                            \n\t"
447         "shrl $15, %%esi                        \n\t"
448         "subl $0xFFFF, %%esi                    \n\t"
449         "add  $2, %%"REG_c"                     \n\t"
450         "mov  %%"REG_c", "BYTE    "(%2)         \n\t"
451
452         "leal -1(%%ebx), %%ecx                  \n\t"
453         "xorl %%ebx, %%ecx                      \n\t"
454         "shrl $15, %%ecx                        \n\t"
455         "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx   \n\t"
456         "neg %%ecx                              \n\t"
457         "add $7, %%ecx                          \n\t"
458
459         "shll %%cl , %%esi                      \n\t"
460         "addl %%esi, %%ebx                      \n\t"
461         "2:                                     \n\t"
462         "movl %%edx, "RANGE    "(%2)            \n\t"
463         "movl %%ebx, "LOW      "(%2)            \n\t"
464         :"=&a"(bit) //FIXME this is fragile gcc either runs out of registers or miscompiles it (for example if "+a"(bit) or "+m"(*state) is used
465         :"r"(state), "r"(c)
466         : "%"REG_c, "%ebx", "%edx", "%"REG_S, "memory"
467     );
468     bit&=1;
469 #else /* BRANCHLESS_CABAC_DECODER */
470
471
472 #if defined HAVE_FAST_CMOV
473 #define BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
474         "mov    "tmp"       , %%ecx                                     \n\t"\
475         "shl    $17         , "tmp"                                     \n\t"\
476         "cmp    "low"       , "tmp"                                     \n\t"\
477         "cmova  %%ecx       , "range"                                   \n\t"\
478         "sbb    %%ecx       , %%ecx                                     \n\t"\
479         "and    %%ecx       , "tmp"                                     \n\t"\
480         "sub    "tmp"       , "low"                                     \n\t"\
481         "xor    %%ecx       , "ret"                                     \n\t"
482 #else /* HAVE_FAST_CMOV */
483 #define BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
484         "mov    "tmp"       , %%ecx                                     \n\t"\
485         "shl    $17         , "tmp"                                     \n\t"\
486         "sub    "low"       , "tmp"                                     \n\t"\
487         "sar    $31         , "tmp"                                     \n\t" /*lps_mask*/\
488         "sub    %%ecx       , "range"                                   \n\t" /*RangeLPS - range*/\
489         "and    "tmp"       , "range"                                   \n\t" /*(RangeLPS - range)&lps_mask*/\
490         "add    %%ecx       , "range"                                   \n\t" /*new range*/\
491         "shl    $17         , %%ecx                                     \n\t"\
492         "and    "tmp"       , %%ecx                                     \n\t"\
493         "sub    %%ecx       , "low"                                     \n\t"\
494         "xor    "tmp"       , "ret"                                     \n\t"
495 #endif /* HAVE_FAST_CMOV */
496
497
498 #define BRANCHLESS_GET_CABAC(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
499         "movzbl "statep"    , "ret"                                     \n\t"\
500         "mov    "range"     , "tmp"                                     \n\t"\
501         "and    $0xC0       , "range"                                   \n\t"\
502         "movzbl "MANGLE(ff_h264_lps_range)"("ret", "range", 2), "range" \n\t"\
503         "sub    "range"     , "tmp"                                     \n\t"\
504         BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
505         "movzbl " MANGLE(ff_h264_norm_shift) "("range"), %%ecx          \n\t"\
506         "shl    %%cl        , "range"                                   \n\t"\
507         "movzbl "MANGLE(ff_h264_mlps_state)"+128("ret"), "tmp"          \n\t"\
508         "mov    "tmpbyte"   , "statep"                                  \n\t"\
509         "shl    %%cl        , "low"                                     \n\t"\
510         "test   "lowword"   , "lowword"                                 \n\t"\
511         " jnz   1f                                                      \n\t"\
512         "mov "BYTE"("cabac"), %%"REG_c"                                 \n\t"\
513         "movzwl (%%"REG_c")     , "tmp"                                 \n\t"\
514         "bswap  "tmp"                                                   \n\t"\
515         "shr    $15         , "tmp"                                     \n\t"\
516         "sub    $0xFFFF     , "tmp"                                     \n\t"\
517         "add    $2          , %%"REG_c"                                 \n\t"\
518         "mov    %%"REG_c"   , "BYTE    "("cabac")                       \n\t"\
519         "lea    -1("low")   , %%ecx                                     \n\t"\
520         "xor    "low"       , %%ecx                                     \n\t"\
521         "shr    $15         , %%ecx                                     \n\t"\
522         "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx            \n\t"\
523         "neg    %%ecx                                                   \n\t"\
524         "add    $7          , %%ecx                                     \n\t"\
525         "shl    %%cl        , "tmp"                                     \n\t"\
526         "add    "tmp"       , "low"                                     \n\t"\
527         "1:                                                             \n\t"
528
529     asm volatile(
530         "movl "RANGE    "(%2), %%esi            \n\t"
531         "movl "LOW      "(%2), %%ebx            \n\t"
532         BRANCHLESS_GET_CABAC("%0", "%2", "(%1)", "%%ebx", "%%bx", "%%esi", "%%edx", "%%dl")
533         "movl %%esi, "RANGE    "(%2)            \n\t"
534         "movl %%ebx, "LOW      "(%2)            \n\t"
535
536         :"=&a"(bit)
537         :"r"(state), "r"(c)
538         : "%"REG_c, "%ebx", "%edx", "%esi", "memory"
539     );
540     bit&=1;
541 #endif /* BRANCHLESS_CABAC_DECODER */
542 #else /* defined(ARCH_X86) && defined(HAVE_7REGS) && defined(HAVE_EBX_AVAILABLE) && !defined(BROKEN_RELOCATIONS) */
543     int s = *state;
544     int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + s];
545     int bit, lps_mask av_unused;
546
547     c->range -= RangeLPS;
548 #ifndef BRANCHLESS_CABAC_DECODER
549     if(c->low < (c->range<<(CABAC_BITS+1))){
550         bit= s&1;
551         *state= ff_h264_mps_state[s];
552         renorm_cabac_decoder_once(c);
553     }else{
554         bit= ff_h264_norm_shift[RangeLPS];
555         c->low -= (c->range<<(CABAC_BITS+1));
556         *state= ff_h264_lps_state[s];
557         c->range = RangeLPS<<bit;
558         c->low <<= bit;
559         bit= (s&1)^1;
560
561         if(!(c->low & CABAC_MASK)){
562             refill2(c);
563         }
564     }
565 #else /* BRANCHLESS_CABAC_DECODER */
566     lps_mask= ((c->range<<(CABAC_BITS+1)) - c->low)>>31;
567
568     c->low -= (c->range<<(CABAC_BITS+1)) & lps_mask;
569     c->range += (RangeLPS - c->range) & lps_mask;
570
571     s^=lps_mask;
572     *state= (ff_h264_mlps_state+128)[s];
573     bit= s&1;
574
575     lps_mask= ff_h264_norm_shift[c->range];
576     c->range<<= lps_mask;
577     c->low  <<= lps_mask;
578     if(!(c->low & CABAC_MASK))
579         refill2(c);
580 #endif /* BRANCHLESS_CABAC_DECODER */
581 #endif /* defined(ARCH_X86) && defined(HAVE_7REGS) && defined(HAVE_EBX_AVAILABLE) && !defined(BROKEN_RELOCATIONS) */
582     return bit;
583 }
584
585 static int av_noinline get_cabac_noinline(CABACContext *c, uint8_t * const state){
586     return get_cabac_inline(c,state);
587 }
588
589 static int get_cabac(CABACContext *c, uint8_t * const state){
590     return get_cabac_inline(c,state);
591 }
592
593 static int get_cabac_bypass(CABACContext *c){
594 #if 0 //not faster
595     int bit;
596     asm volatile(
597         "movl "RANGE    "(%1), %%ebx            \n\t"
598         "movl "LOW      "(%1), %%eax            \n\t"
599         "shl $17, %%ebx                         \n\t"
600         "add %%eax, %%eax                       \n\t"
601         "sub %%ebx, %%eax                       \n\t"
602         "cltd                                   \n\t"
603         "and %%edx, %%ebx                       \n\t"
604         "add %%ebx, %%eax                       \n\t"
605         "test %%ax, %%ax                        \n\t"
606         " jnz 1f                                \n\t"
607         "movl "BYTE     "(%1), %%"REG_b"        \n\t"
608         "subl $0xFFFF, %%eax                    \n\t"
609         "movzwl (%%"REG_b"), %%ecx              \n\t"
610         "bswap %%ecx                            \n\t"
611         "shrl $15, %%ecx                        \n\t"
612         "addl $2, %%"REG_b"                     \n\t"
613         "addl %%ecx, %%eax                      \n\t"
614         "movl %%"REG_b", "BYTE     "(%1)        \n\t"
615         "1:                                     \n\t"
616         "movl %%eax, "LOW      "(%1)            \n\t"
617
618         :"=&d"(bit)
619         :"r"(c)
620         : "%eax", "%"REG_b, "%ecx", "memory"
621     );
622     return bit+1;
623 #else
624     int range;
625     c->low += c->low;
626
627     if(!(c->low & CABAC_MASK))
628         refill(c);
629
630     range= c->range<<(CABAC_BITS+1);
631     if(c->low < range){
632         return 0;
633     }else{
634         c->low -= range;
635         return 1;
636     }
637 #endif
638 }
639
640
641 static av_always_inline int get_cabac_bypass_sign(CABACContext *c, int val){
642 #if defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__))
643     asm volatile(
644         "movl "RANGE    "(%1), %%ebx            \n\t"
645         "movl "LOW      "(%1), %%eax            \n\t"
646         "shl $17, %%ebx                         \n\t"
647         "add %%eax, %%eax                       \n\t"
648         "sub %%ebx, %%eax                       \n\t"
649         "cltd                                   \n\t"
650         "and %%edx, %%ebx                       \n\t"
651         "add %%ebx, %%eax                       \n\t"
652         "xor %%edx, %%ecx                       \n\t"
653         "sub %%edx, %%ecx                       \n\t"
654         "test %%ax, %%ax                        \n\t"
655         " jnz 1f                                \n\t"
656         "mov  "BYTE     "(%1), %%"REG_b"        \n\t"
657         "subl $0xFFFF, %%eax                    \n\t"
658         "movzwl (%%"REG_b"), %%edx              \n\t"
659         "bswap %%edx                            \n\t"
660         "shrl $15, %%edx                        \n\t"
661         "add  $2, %%"REG_b"                     \n\t"
662         "addl %%edx, %%eax                      \n\t"
663         "mov  %%"REG_b", "BYTE     "(%1)        \n\t"
664         "1:                                     \n\t"
665         "movl %%eax, "LOW      "(%1)            \n\t"
666
667         :"+c"(val)
668         :"r"(c)
669         : "%eax", "%"REG_b, "%edx", "memory"
670     );
671     return val;
672 #else
673     int range, mask;
674     c->low += c->low;
675
676     if(!(c->low & CABAC_MASK))
677         refill(c);
678
679     range= c->range<<(CABAC_BITS+1);
680     c->low -= range;
681     mask= c->low >> 31;
682     range &= mask;
683     c->low += range;
684     return (val^mask)-mask;
685 #endif
686 }
687
688 //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!)
689 //FIXME use some macros to avoid duplicatin get_cabac (cannot be done yet as that would make optimization work hard)
690 #if defined(ARCH_X86) && defined(HAVE_7REGS) && defined(HAVE_EBX_AVAILABLE) && !defined(BROKEN_RELOCATIONS)
691 static int decode_significance_x86(CABACContext *c, int max_coeff, uint8_t *significant_coeff_ctx_base, int *index){
692     void *end= significant_coeff_ctx_base + max_coeff - 1;
693     int minusstart= -(int)significant_coeff_ctx_base;
694     int minusindex= 4-(int)index;
695     int coeff_count;
696     asm volatile(
697         "movl "RANGE    "(%3), %%esi            \n\t"
698         "movl "LOW      "(%3), %%ebx            \n\t"
699
700         "2:                                     \n\t"
701
702         BRANCHLESS_GET_CABAC("%%edx", "%3", "(%1)", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
703
704         "test $1, %%edx                         \n\t"
705         " jz 3f                                 \n\t"
706
707         BRANCHLESS_GET_CABAC("%%edx", "%3", "61(%1)", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
708
709         "mov  %2, %%"REG_a"                     \n\t"
710         "movl %4, %%ecx                         \n\t"
711         "add  %1, %%"REG_c"                     \n\t"
712         "movl %%ecx, (%%"REG_a")                \n\t"
713
714         "test $1, %%edx                         \n\t"
715         " jnz 4f                                \n\t"
716
717         "add  $4, %%"REG_a"                     \n\t"
718         "mov  %%"REG_a", %2                     \n\t"
719
720         "3:                                     \n\t"
721         "add  $1, %1                            \n\t"
722         "cmp  %5, %1                            \n\t"
723         " jb 2b                                 \n\t"
724         "mov  %2, %%"REG_a"                     \n\t"
725         "movl %4, %%ecx                         \n\t"
726         "add  %1, %%"REG_c"                     \n\t"
727         "movl %%ecx, (%%"REG_a")                \n\t"
728         "4:                                     \n\t"
729         "add  %6, %%eax                         \n\t"
730         "shr $2, %%eax                          \n\t"
731
732         "movl %%esi, "RANGE    "(%3)            \n\t"
733         "movl %%ebx, "LOW      "(%3)            \n\t"
734         :"=&a"(coeff_count), "+r"(significant_coeff_ctx_base), "+m"(index)\
735         :"r"(c), "m"(minusstart), "m"(end), "m"(minusindex)\
736         : "%"REG_c, "%ebx", "%edx", "%esi", "memory"\
737     );
738     return coeff_count;
739 }
740
741 static int decode_significance_8x8_x86(CABACContext *c, uint8_t *significant_coeff_ctx_base, int *index, const uint8_t *sig_off){
742     int minusindex= 4-(int)index;
743     int coeff_count;
744     long last=0;
745     asm volatile(
746         "movl "RANGE    "(%3), %%esi            \n\t"
747         "movl "LOW      "(%3), %%ebx            \n\t"
748
749         "mov %1, %%"REG_D"                      \n\t"
750         "2:                                     \n\t"
751
752         "mov %6, %%"REG_a"                      \n\t"
753         "movzbl (%%"REG_a", %%"REG_D"), %%edi   \n\t"
754         "add %5, %%"REG_D"                      \n\t"
755
756         BRANCHLESS_GET_CABAC("%%edx", "%3", "(%%"REG_D")", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
757
758         "mov %1, %%edi                          \n\t"
759         "test $1, %%edx                         \n\t"
760         " jz 3f                                 \n\t"
761
762         "movzbl "MANGLE(last_coeff_flag_offset_8x8)"(%%edi), %%edi\n\t"
763         "add %5, %%"REG_D"                      \n\t"
764
765         BRANCHLESS_GET_CABAC("%%edx", "%3", "15(%%"REG_D")", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
766
767         "mov %2, %%"REG_a"                      \n\t"
768         "mov %1, %%edi                          \n\t"
769         "movl %%edi, (%%"REG_a")                \n\t"
770
771         "test $1, %%edx                         \n\t"
772         " jnz 4f                                \n\t"
773
774         "add $4, %%"REG_a"                      \n\t"
775         "mov %%"REG_a", %2                      \n\t"
776
777         "3:                                     \n\t"
778         "addl $1, %%edi                         \n\t"
779         "mov %%edi, %1                          \n\t"
780         "cmpl $63, %%edi                        \n\t"
781         " jb 2b                                 \n\t"
782         "mov %2, %%"REG_a"                      \n\t"
783         "movl %%edi, (%%"REG_a")                \n\t"
784         "4:                                     \n\t"
785         "addl %4, %%eax                         \n\t"
786         "shr $2, %%eax                          \n\t"
787
788         "movl %%esi, "RANGE    "(%3)            \n\t"
789         "movl %%ebx, "LOW      "(%3)            \n\t"
790         :"=&a"(coeff_count),"+m"(last), "+m"(index)\
791         :"r"(c), "m"(minusindex), "m"(significant_coeff_ctx_base), "m"(sig_off)\
792         : "%"REG_c, "%ebx", "%edx", "%esi", "%"REG_D, "memory"\
793     );
794     return coeff_count;
795 }
796 #endif /* defined(ARCH_X86) && && defined(HAVE_7REGS) && defined(HAVE_EBX_AVAILABLE) && !defined(BROKEN_RELOCATIONS) */
797
798 /**
799  *
800  * @return the number of bytes read or 0 if no end
801  */
802 static int get_cabac_terminate(CABACContext *c){
803     c->range -= 2;
804     if(c->low < c->range<<(CABAC_BITS+1)){
805         renorm_cabac_decoder_once(c);
806         return 0;
807     }else{
808         return c->bytestream - c->bytestream_start;
809     }
810 }
811
812 #if 0
813 /**
814  * Get (truncated) unary binarization.
815  */
816 static int get_cabac_u(CABACContext *c, uint8_t * state, int max, int max_index, int truncated){
817     int i;
818
819     for(i=0; i<max; i++){
820         if(get_cabac(c, state)==0)
821             return i;
822
823         if(i< max_index) state++;
824     }
825
826     return truncated ? max : -1;
827 }
828
829 /**
830  * get unary exp golomb k-th order binarization.
831  */
832 static int get_cabac_ueg(CABACContext *c, uint8_t * state, int max, int is_signed, int k, int max_index){
833     int i, v;
834     int m= 1<<k;
835
836     if(get_cabac(c, state)==0)
837         return 0;
838
839     if(0 < max_index) state++;
840
841     for(i=1; i<max; i++){
842         if(get_cabac(c, state)==0){
843             if(is_signed && get_cabac_bypass(c)){
844                 return -i;
845             }else
846                 return i;
847         }
848
849         if(i < max_index) state++;
850     }
851
852     while(get_cabac_bypass(c)){
853         i+= m;
854         m+= m;
855     }
856
857     v=0;
858     while(m>>=1){
859         v+= v + get_cabac_bypass(c);
860     }
861     i += v;
862
863     if(is_signed && get_cabac_bypass(c)){
864         return -i;
865     }else
866         return i;
867 }
868 #endif /* 0 */
869
870 #endif /* FFMPEG_CABAC_H */