389f155745ecfbad69b198b77b3ad1f72a42f0b6
[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         "cmp    "end"       , %%"REG_c"                                 \n\t"\
77         "jge    1f                                                      \n\t"\
78         "add"OPSIZE" $2     , "byte"                                    \n\t"\
79         "1:                                                             \n\t"\
80         "movzwl (%%"REG_c") , "tmp"                                     \n\t"\
81         "lea    -1("low")   , %%ecx                                     \n\t"\
82         "xor    "low"       , %%ecx                                     \n\t"\
83         "shr    $15         , %%ecx                                     \n\t"\
84         "bswap  "tmp"                                                   \n\t"\
85         "shr    $15         , "tmp"                                     \n\t"\
86         "movzbl "norm_off"("tables", %%rcx), %%ecx                      \n\t"\
87         "sub    $0xFFFF     , "tmp"                                     \n\t"\
88         "neg    %%ecx                                                   \n\t"\
89         "add    $7          , %%ecx                                     \n\t"\
90         "shl    %%cl        , "tmp"                                     \n\t"\
91         "add    "tmp"       , "low"                                     \n\t"\
92         "2:                                                             \n\t"
93
94 #else /* BROKEN_RELOCATIONS */
95 #define TABLES_ARG
96
97 #if HAVE_FAST_CMOV
98 #define BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)\
99         "mov    "tmp"       , %%ecx     \n\t"\
100         "shl    $17         , "tmp"     \n\t"\
101         "cmp    "low"       , "tmp"     \n\t"\
102         "cmova  %%ecx       , "range"   \n\t"\
103         "sbb    %%ecx       , %%ecx     \n\t"\
104         "and    %%ecx       , "tmp"     \n\t"\
105         "xor    %%ecx       , "ret"     \n\t"\
106         "sub    "tmp"       , "low"     \n\t"
107 #else /* HAVE_FAST_CMOV */
108 #define BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)\
109         "mov    "tmp"       , %%ecx     \n\t"\
110         "shl    $17         , "tmp"     \n\t"\
111         "sub    "low"       , "tmp"     \n\t"\
112         "sar    $31         , "tmp"     \n\t" /*lps_mask*/\
113         "sub    %%ecx       , "range"   \n\t" /*RangeLPS - range*/\
114         "and    "tmp"       , "range"   \n\t" /*(RangeLPS - range)&lps_mask*/\
115         "add    %%ecx       , "range"   \n\t" /*new range*/\
116         "shl    $17         , %%ecx     \n\t"\
117         "and    "tmp"       , %%ecx     \n\t"\
118         "sub    %%ecx       , "low"     \n\t"\
119         "xor    "tmp"       , "ret"     \n\t"
120 #endif /* HAVE_FAST_CMOV */
121
122 #define BRANCHLESS_GET_CABAC(ret, retq, statep, low, lowword, range, rangeq, tmp, tmpbyte, byte, end, norm_off, lps_off, mlps_off, tables) \
123         "movzbl "statep"    , "ret"                                     \n\t"\
124         "mov    "range"     , "tmp"                                     \n\t"\
125         "and    $0xC0       , "range"                                   \n\t"\
126         "movzbl "MANGLE(ff_h264_cabac_tables)"+"lps_off"("ret", "range", 2), "range" \n\t"\
127         "sub    "range"     , "tmp"                                     \n\t"\
128         BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)                    \
129         "movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"("range"), %%ecx    \n\t"\
130         "shl    %%cl        , "range"                                   \n\t"\
131         "movzbl "MANGLE(ff_h264_cabac_tables)"+"mlps_off"+128("ret"), "tmp"  \n\t"\
132         "shl    %%cl        , "low"                                     \n\t"\
133         "mov    "tmpbyte"   , "statep"                                  \n\t"\
134         "test   "lowword"   , "lowword"                                 \n\t"\
135         " jnz   2f                                                      \n\t"\
136         "mov    "byte"      , %%"REG_c"                                 \n\t"\
137         "cmp    "end"       , %%"REG_c"                                 \n\t"\
138         "jge    1f                                                      \n\t"\
139         "add"OPSIZE" $2     , "byte"                                    \n\t"\
140         "1:                                                             \n\t"\
141         "movzwl (%%"REG_c")     , "tmp"                                 \n\t"\
142         "lea    -1("low")   , %%ecx                                     \n\t"\
143         "xor    "low"       , %%ecx                                     \n\t"\
144         "shr    $15         , %%ecx                                     \n\t"\
145         "bswap  "tmp"                                                   \n\t"\
146         "shr    $15         , "tmp"                                     \n\t"\
147         "movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"(%%ecx), %%ecx \n\t"\
148         "sub    $0xFFFF     , "tmp"                                     \n\t"\
149         "neg    %%ecx                                                   \n\t"\
150         "add    $7          , %%ecx                                     \n\t"\
151         "shl    %%cl        , "tmp"                                     \n\t"\
152         "add    "tmp"       , "low"                                     \n\t"\
153         "2:                                                             \n\t"
154
155 #endif /* BROKEN_RELOCATIONS */
156
157
158 #if HAVE_7REGS
159 #define get_cabac_inline get_cabac_inline_x86
160 static av_always_inline int get_cabac_inline_x86(CABACContext *c,
161                                                  uint8_t *const state)
162 {
163     int bit, tmp;
164 #ifdef BROKEN_RELOCATIONS
165     void *tables;
166
167     __asm__ volatile(
168         "lea    "MANGLE(ff_h264_cabac_tables)", %0      \n\t"
169         : "=&r"(tables)
170     );
171 #endif
172
173     __asm__ volatile(
174         BRANCHLESS_GET_CABAC("%0", "%q0", "(%4)", "%1", "%w1",
175                              "%2", "%q2", "%3", "%b3",
176                              "%a6(%5)", "%a7(%5)", "%a8", "%a9", "%a10", "%11")
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           "i"(H264_NORM_SHIFT_OFFSET),
182           "i"(H264_LPS_RANGE_OFFSET),
183           "i"(H264_MLPS_STATE_OFFSET) TABLES_ARG
184         : "%"REG_c, "memory"
185     );
186     return bit & 1;
187 }
188 #endif /* HAVE_7REGS */
189
190 #define get_cabac_bypass_sign get_cabac_bypass_sign_x86
191 static av_always_inline int get_cabac_bypass_sign_x86(CABACContext *c, int val)
192 {
193     x86_reg tmp;
194     __asm__ volatile(
195         "movl        %a6(%2), %k1       \n\t"
196         "movl        %a3(%2), %%eax     \n\t"
197         "shl             $17, %k1       \n\t"
198         "add           %%eax, %%eax     \n\t"
199         "sub             %k1, %%eax     \n\t"
200         "cltd                           \n\t"
201         "and           %%edx, %k1       \n\t"
202         "add             %k1, %%eax     \n\t"
203         "xor           %%edx, %%ecx     \n\t"
204         "sub           %%edx, %%ecx     \n\t"
205         "test           %%ax, %%ax      \n\t"
206         "jnz              1f            \n\t"
207         "mov         %a4(%2), %1        \n\t"
208         "subl        $0xFFFF, %%eax     \n\t"
209         "movzwl         (%1), %%edx     \n\t"
210         "bswap         %%edx            \n\t"
211         "shrl            $15, %%edx     \n\t"
212         "addl          %%edx, %%eax     \n\t"
213         "cmp         %a5(%2), %1        \n\t"
214         "jge              1f            \n\t"
215         "add"OPSIZE"      $2, %a4(%2)   \n\t"
216         "1:                             \n\t"
217         "movl          %%eax, %a3(%2)   \n\t"
218
219         : "+c"(val), "=&r"(tmp)
220         : "r"(c),
221           "i"(offsetof(CABACContext, low)),
222           "i"(offsetof(CABACContext, bytestream)),
223           "i"(offsetof(CABACContext, bytestream_end)),
224           "i"(offsetof(CABACContext, range))
225         : "%eax", "%edx", "memory"
226     );
227     return val;
228 }
229
230 #endif /* HAVE_INLINE_ASM */
231 #endif /* AVCODEC_X86_CABAC_H */