Merge commit 'e6d8acf6a8fba4743eb56eabe72a741d1bbee3cb'
[ffmpeg.git] / libavcodec / x86 / cabac.h
1 /*
2  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
3  *
4  * This file is part of Libav.
5  *
6  * Libav is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU Lesser General Public
8  * License as published by the Free Software Foundation; either
9  * version 2.1 of the License, or (at your option) any later version.
10  *
11  * Libav is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14  * Lesser General Public License for more details.
15  *
16  * You should have received a copy of the GNU Lesser General Public
17  * License along with Libav; if not, write to the Free Software
18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19  */
20
21 #ifndef AVCODEC_X86_CABAC_H
22 #define AVCODEC_X86_CABAC_H
23
24 #include "libavcodec/cabac.h"
25 #include "libavutil/attributes.h"
26 #include "libavutil/x86/asm.h"
27 #include "libavutil/internal.h"
28 #include "config.h"
29
30 #if HAVE_INLINE_ASM
31
32 #ifdef BROKEN_RELOCATIONS
33 #define TABLES_ARG , "r"(tables)
34
35 #if HAVE_FAST_CMOV
36 #define BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \
37         "cmp    "low"       , "tmp"                        \n\t"\
38         "cmova  %%ecx       , "range"                      \n\t"\
39         "sbb    %%rcx       , %%rcx                        \n\t"\
40         "and    %%ecx       , "tmp"                        \n\t"\
41         "xor    %%rcx       , "retq"                       \n\t"\
42         "sub    "tmp"       , "low"                        \n\t"
43 #else /* HAVE_FAST_CMOV */
44 #define BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \
45 /* P4 Prescott has crappy cmov,sbb,64bit shift so avoid them */ \
46         "sub    "low"       , "tmp"                        \n\t"\
47         "sar    $31         , "tmp"                        \n\t"\
48         "sub    %%ecx       , "range"                      \n\t"\
49         "and    "tmp"       , "range"                      \n\t"\
50         "add    %%ecx       , "range"                      \n\t"\
51         "shl    $17         , %%ecx                        \n\t"\
52         "and    "tmp"       , %%ecx                        \n\t"\
53         "sub    %%ecx       , "low"                        \n\t"\
54         "xor    "tmp"       , "ret"                        \n\t"\
55         "movslq "ret"       , "retq"                       \n\t"
56 #endif /* HAVE_FAST_CMOV */
57
58 #define BRANCHLESS_GET_CABAC(ret, retq, statep, low, lowword, range, rangeq, tmp, tmpbyte, byte, end, norm_off, lps_off, mlps_off, tables) \
59         "movzbl "statep"    , "ret"                                     \n\t"\
60         "mov    "range"     , "tmp"                                     \n\t"\
61         "and    $0xC0       , "range"                                   \n\t"\
62         "lea    ("ret", "range", 2), %%ecx                              \n\t"\
63         "movzbl "lps_off"("tables", %%rcx), "range"                     \n\t"\
64         "sub    "range"     , "tmp"                                     \n\t"\
65         "mov    "tmp"       , %%ecx                                     \n\t"\
66         "shl    $17         , "tmp"                                     \n\t"\
67         BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp)              \
68         "movzbl "norm_off"("tables", "rangeq"), %%ecx                   \n\t"\
69         "shl    %%cl        , "range"                                   \n\t"\
70         "movzbl "mlps_off"+128("tables", "retq"), "tmp"                 \n\t"\
71         "shl    %%cl        , "low"                                     \n\t"\
72         "mov    "tmpbyte"   , "statep"                                  \n\t"\
73         "test   "lowword"   , "lowword"                                 \n\t"\
74         "jnz    2f                                                      \n\t"\
75         "mov    "byte"      , %%"REG_c"                                 \n\t"\
76         "add"OPSIZE" $2     , "byte"                                    \n\t"\
77         "movzwl (%%"REG_c") , "tmp"                                     \n\t"\
78         "lea    -1("low")   , %%ecx                                     \n\t"\
79         "xor    "low"       , %%ecx                                     \n\t"\
80         "shr    $15         , %%ecx                                     \n\t"\
81         "bswap  "tmp"                                                   \n\t"\
82         "shr    $15         , "tmp"                                     \n\t"\
83         "movzbl "norm_off"("tables", %%rcx), %%ecx                      \n\t"\
84         "sub    $0xFFFF     , "tmp"                                     \n\t"\
85         "neg    %%ecx                                                   \n\t"\
86         "add    $7          , %%ecx                                     \n\t"\
87         "shl    %%cl        , "tmp"                                     \n\t"\
88         "add    "tmp"       , "low"                                     \n\t"\
89         "2:                                                             \n\t"
90
91 #else /* BROKEN_RELOCATIONS */
92 #define TABLES_ARG
93 #define RIP_ARG
94
95 #if HAVE_FAST_CMOV
96 #define BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)\
97         "mov    "tmp"       , %%ecx     \n\t"\
98         "shl    $17         , "tmp"     \n\t"\
99         "cmp    "low"       , "tmp"     \n\t"\
100         "cmova  %%ecx       , "range"   \n\t"\
101         "sbb    %%ecx       , %%ecx     \n\t"\
102         "and    %%ecx       , "tmp"     \n\t"\
103         "xor    %%ecx       , "ret"     \n\t"\
104         "sub    "tmp"       , "low"     \n\t"
105 #else /* HAVE_FAST_CMOV */
106 #define BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)\
107         "mov    "tmp"       , %%ecx     \n\t"\
108         "shl    $17         , "tmp"     \n\t"\
109         "sub    "low"       , "tmp"     \n\t"\
110         "sar    $31         , "tmp"     \n\t" /*lps_mask*/\
111         "sub    %%ecx       , "range"   \n\t" /*RangeLPS - range*/\
112         "and    "tmp"       , "range"   \n\t" /*(RangeLPS - range)&lps_mask*/\
113         "add    %%ecx       , "range"   \n\t" /*new range*/\
114         "shl    $17         , %%ecx     \n\t"\
115         "and    "tmp"       , %%ecx     \n\t"\
116         "sub    %%ecx       , "low"     \n\t"\
117         "xor    "tmp"       , "ret"     \n\t"
118 #endif /* HAVE_FAST_CMOV */
119
120 #define BRANCHLESS_GET_CABAC(ret, retq, statep, low, lowword, range, rangeq, tmp, tmpbyte, byte, end, norm_off, lps_off, mlps_off, tables) \
121         "movzbl "statep"    , "ret"                                     \n\t"\
122         "mov    "range"     , "tmp"                                     \n\t"\
123         "and    $0xC0       , "range"                                   \n\t"\
124         "movzbl "MANGLE(ff_h264_cabac_tables)"+"lps_off"("ret", "range", 2), "range" \n\t"\
125         "sub    "range"     , "tmp"                                     \n\t"\
126         BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)                    \
127         "movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"("range"), %%ecx    \n\t"\
128         "shl    %%cl        , "range"                                   \n\t"\
129         "movzbl "MANGLE(ff_h264_cabac_tables)"+"mlps_off"+128("ret"), "tmp"  \n\t"\
130         "shl    %%cl        , "low"                                     \n\t"\
131         "mov    "tmpbyte"   , "statep"                                  \n\t"\
132         "test   "lowword"   , "lowword"                                 \n\t"\
133         " jnz   2f                                                      \n\t"\
134         "mov    "byte"      , %%"REG_c"                                 \n\t"\
135         "add"OPSIZE" $2     , "byte"                                    \n\t"\
136         "movzwl (%%"REG_c")     , "tmp"                                 \n\t"\
137         "lea    -1("low")   , %%ecx                                     \n\t"\
138         "xor    "low"       , %%ecx                                     \n\t"\
139         "shr    $15         , %%ecx                                     \n\t"\
140         "bswap  "tmp"                                                   \n\t"\
141         "shr    $15         , "tmp"                                     \n\t"\
142         "movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"(%%ecx), %%ecx \n\t"\
143         "sub    $0xFFFF     , "tmp"                                     \n\t"\
144         "neg    %%ecx                                                   \n\t"\
145         "add    $7          , %%ecx                                     \n\t"\
146         "shl    %%cl        , "tmp"                                     \n\t"\
147         "add    "tmp"       , "low"                                     \n\t"\
148         "2:                                                             \n\t"
149
150 #endif /* BROKEN_RELOCATIONS */
151
152
153 #if HAVE_7REGS && !(defined(__i386) && defined(__clang__) && (__clang_major__<2 || (__clang_major__==2 && __clang_minor__<10)))\
154                && !(                  !defined(__clang__) && defined(__llvm__) && __GNUC__==4 && __GNUC_MINOR__==2 && __GNUC_PATCHLEVEL__<=1)
155 #define get_cabac_inline get_cabac_inline_x86
156 static av_always_inline int get_cabac_inline_x86(CABACContext *c,
157                                                  uint8_t *const state)
158 {
159     int bit, tmp;
160 #ifdef BROKEN_RELOCATIONS
161     void *tables;
162
163     __asm__ volatile(
164         "lea    "MANGLE(ff_h264_cabac_tables)", %0      \n\t"
165         : "=&r"(tables)
166     );
167 #endif
168
169     __asm__ volatile(
170         BRANCHLESS_GET_CABAC("%0", "%q0", "(%4)", "%1", "%w1",
171                              "%2", "%q2", "%3", "%b3",
172                              "%c6(%5)", "%c7(%5)",
173                              AV_STRINGIFY(H264_NORM_SHIFT_OFFSET),
174                              AV_STRINGIFY(H264_LPS_RANGE_OFFSET),
175                              AV_STRINGIFY(H264_MLPS_STATE_OFFSET),
176                              "%8")
177         : "=&r"(bit), "=&r"(c->low), "=&r"(c->range), "=&q"(tmp)
178         : "r"(state), "r"(c),
179           "i"(offsetof(CABACContext, bytestream)),
180           "i"(offsetof(CABACContext, bytestream_end))
181           TABLES_ARG
182           ,"1"(c->low), "2"(c->range)
183         : "%"REG_c, "memory"
184     );
185     return bit & 1;
186 }
187 #endif /* HAVE_7REGS */
188
189 #define get_cabac_bypass_sign get_cabac_bypass_sign_x86
190 static av_always_inline int get_cabac_bypass_sign_x86(CABACContext *c, int val)
191 {
192     x86_reg tmp;
193     __asm__ volatile(
194         "movl        %c6(%2), %k1       \n\t"
195         "movl        %c3(%2), %%eax     \n\t"
196         "shl             $17, %k1       \n\t"
197         "add           %%eax, %%eax     \n\t"
198         "sub             %k1, %%eax     \n\t"
199         "cltd                           \n\t"
200         "and           %%edx, %k1       \n\t"
201         "add             %k1, %%eax     \n\t"
202         "xor           %%edx, %%ecx     \n\t"
203         "sub           %%edx, %%ecx     \n\t"
204         "test           %%ax, %%ax      \n\t"
205         "jnz              1f            \n\t"
206         "mov         %c4(%2), %1        \n\t"
207         "subl        $0xFFFF, %%eax     \n\t"
208         "movzwl         (%1), %%edx     \n\t"
209         "bswap         %%edx            \n\t"
210         "shrl            $15, %%edx     \n\t"
211         "add              $2, %1        \n\t"
212         "addl          %%edx, %%eax     \n\t"
213         "mov              %1, %c4(%2)   \n\t"
214         "1:                             \n\t"
215         "movl          %%eax, %c3(%2)   \n\t"
216
217         : "+c"(val), "=&r"(tmp)
218         : "r"(c),
219           "i"(offsetof(CABACContext, low)),
220           "i"(offsetof(CABACContext, bytestream)),
221           "i"(offsetof(CABACContext, bytestream_end)),
222           "i"(offsetof(CABACContext, range))
223         : "%eax", "%edx", "memory"
224     );
225     return val;
226 }
227
228 #define get_cabac_bypass get_cabac_bypass_x86
229 static av_always_inline int get_cabac_bypass_x86(CABACContext *c)
230 {
231     x86_reg tmp;
232     int res;
233     __asm__ volatile(
234         "movl        %c6(%2), %k1       \n\t"
235         "movl        %c3(%2), %%eax     \n\t"
236         "shl             $17, %k1       \n\t"
237         "add           %%eax, %%eax     \n\t"
238         "sub             %k1, %%eax     \n\t"
239         "cltd                           \n\t"
240         "and           %%edx, %k1       \n\t"
241         "add             %k1, %%eax     \n\t"
242         "inc           %%edx            \n\t"
243         "test           %%ax, %%ax      \n\t"
244         "jnz              1f            \n\t"
245         "mov         %c4(%2), %1        \n\t"
246         "subl        $0xFFFF, %%eax     \n\t"
247         "movzwl         (%1), %%ecx     \n\t"
248         "bswap         %%ecx            \n\t"
249         "shrl            $15, %%ecx     \n\t"
250         "addl          %%ecx, %%eax     \n\t"
251         "cmp         %c5(%2), %1        \n\t"
252         "jge              1f            \n\t"
253         "add"OPSIZE"      $2, %c4(%2)   \n\t"
254         "1:                             \n\t"
255         "movl          %%eax, %c3(%2)   \n\t"
256
257         : "=&d"(res), "=&r"(tmp)
258         : "r"(c),
259           "i"(offsetof(CABACContext, low)),
260           "i"(offsetof(CABACContext, bytestream)),
261           "i"(offsetof(CABACContext, bytestream_end)),
262           "i"(offsetof(CABACContext, range))
263         : "%eax", "%ecx", "memory"
264     );
265     return res;
266 }
267
268 #endif /* HAVE_INLINE_ASM */
269 #endif /* AVCODEC_X86_CABAC_H */