avcodec/x86/cabac: Disable get_cabac_bypass_x86() on broken llvm/clang
[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   (defined(__i386) && defined(__clang__) && (__clang_major__<2 || (__clang_major__==2 && __clang_minor__<10)))\
31    || (                  !defined(__clang__) && defined(__llvm__) && __GNUC__==4 && __GNUC_MINOR__==2 && __GNUC_PATCHLEVEL__<=1)
32 #       define BROKEN_COMPILER 1
33 #else
34 #       define BROKEN_COMPILER 0
35 #endif
36
37 #if HAVE_INLINE_ASM
38
39 #ifdef BROKEN_RELOCATIONS
40 #define TABLES_ARG , "r"(tables)
41
42 #if HAVE_FAST_CMOV
43 #define BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \
44         "cmp    "low"       , "tmp"                        \n\t"\
45         "cmova  %%ecx       , "range"                      \n\t"\
46         "sbb    %%rcx       , %%rcx                        \n\t"\
47         "and    %%ecx       , "tmp"                        \n\t"\
48         "xor    %%rcx       , "retq"                       \n\t"\
49         "sub    "tmp"       , "low"                        \n\t"
50 #else /* HAVE_FAST_CMOV */
51 #define BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \
52 /* P4 Prescott has crappy cmov,sbb,64bit shift so avoid them */ \
53         "sub    "low"       , "tmp"                        \n\t"\
54         "sar    $31         , "tmp"                        \n\t"\
55         "sub    %%ecx       , "range"                      \n\t"\
56         "and    "tmp"       , "range"                      \n\t"\
57         "add    %%ecx       , "range"                      \n\t"\
58         "shl    $17         , %%ecx                        \n\t"\
59         "and    "tmp"       , %%ecx                        \n\t"\
60         "sub    %%ecx       , "low"                        \n\t"\
61         "xor    "tmp"       , "ret"                        \n\t"\
62         "movslq "ret"       , "retq"                       \n\t"
63 #endif /* HAVE_FAST_CMOV */
64
65 #define BRANCHLESS_GET_CABAC(ret, retq, statep, low, lowword, range, rangeq, tmp, tmpbyte, byte, end, norm_off, lps_off, mlps_off, tables) \
66         "movzbl "statep"    , "ret"                                     \n\t"\
67         "mov    "range"     , "tmp"                                     \n\t"\
68         "and    $0xC0       , "range"                                   \n\t"\
69         "lea    ("ret", "range", 2), %%ecx                              \n\t"\
70         "movzbl "lps_off"("tables", %%rcx), "range"                     \n\t"\
71         "sub    "range"     , "tmp"                                     \n\t"\
72         "mov    "tmp"       , %%ecx                                     \n\t"\
73         "shl    $17         , "tmp"                                     \n\t"\
74         BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp)              \
75         "movzbl "norm_off"("tables", "rangeq"), %%ecx                   \n\t"\
76         "shl    %%cl        , "range"                                   \n\t"\
77         "movzbl "mlps_off"+128("tables", "retq"), "tmp"                 \n\t"\
78         "shl    %%cl        , "low"                                     \n\t"\
79         "mov    "tmpbyte"   , "statep"                                  \n\t"\
80         "test   "lowword"   , "lowword"                                 \n\t"\
81         "jnz    2f                                                      \n\t"\
82         "mov    "byte"      , %%"REG_c"                                 \n\t"\
83         "add"OPSIZE" $2     , "byte"                                    \n\t"\
84         "movzwl (%%"REG_c") , "tmp"                                     \n\t"\
85         "lea    -1("low")   , %%ecx                                     \n\t"\
86         "xor    "low"       , %%ecx                                     \n\t"\
87         "shr    $15         , %%ecx                                     \n\t"\
88         "bswap  "tmp"                                                   \n\t"\
89         "shr    $15         , "tmp"                                     \n\t"\
90         "movzbl "norm_off"("tables", %%rcx), %%ecx                      \n\t"\
91         "sub    $0xFFFF     , "tmp"                                     \n\t"\
92         "neg    %%ecx                                                   \n\t"\
93         "add    $7          , %%ecx                                     \n\t"\
94         "shl    %%cl        , "tmp"                                     \n\t"\
95         "add    "tmp"       , "low"                                     \n\t"\
96         "2:                                                             \n\t"
97
98 #else /* BROKEN_RELOCATIONS */
99 #define TABLES_ARG
100 #define RIP_ARG
101
102 #if HAVE_FAST_CMOV
103 #define BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)\
104         "mov    "tmp"       , %%ecx     \n\t"\
105         "shl    $17         , "tmp"     \n\t"\
106         "cmp    "low"       , "tmp"     \n\t"\
107         "cmova  %%ecx       , "range"   \n\t"\
108         "sbb    %%ecx       , %%ecx     \n\t"\
109         "and    %%ecx       , "tmp"     \n\t"\
110         "xor    %%ecx       , "ret"     \n\t"\
111         "sub    "tmp"       , "low"     \n\t"
112 #else /* HAVE_FAST_CMOV */
113 #define BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)\
114         "mov    "tmp"       , %%ecx     \n\t"\
115         "shl    $17         , "tmp"     \n\t"\
116         "sub    "low"       , "tmp"     \n\t"\
117         "sar    $31         , "tmp"     \n\t" /*lps_mask*/\
118         "sub    %%ecx       , "range"   \n\t" /*RangeLPS - range*/\
119         "and    "tmp"       , "range"   \n\t" /*(RangeLPS - range)&lps_mask*/\
120         "add    %%ecx       , "range"   \n\t" /*new range*/\
121         "shl    $17         , %%ecx     \n\t"\
122         "and    "tmp"       , %%ecx     \n\t"\
123         "sub    %%ecx       , "low"     \n\t"\
124         "xor    "tmp"       , "ret"     \n\t"
125 #endif /* HAVE_FAST_CMOV */
126
127 #define BRANCHLESS_GET_CABAC(ret, retq, statep, low, lowword, range, rangeq, tmp, tmpbyte, byte, end, norm_off, lps_off, mlps_off, tables) \
128         "movzbl "statep"    , "ret"                                     \n\t"\
129         "mov    "range"     , "tmp"                                     \n\t"\
130         "and    $0xC0       , "range"                                   \n\t"\
131         "movzbl "MANGLE(ff_h264_cabac_tables)"+"lps_off"("ret", "range", 2), "range" \n\t"\
132         "sub    "range"     , "tmp"                                     \n\t"\
133         BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)                    \
134         "movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"("range"), %%ecx    \n\t"\
135         "shl    %%cl        , "range"                                   \n\t"\
136         "movzbl "MANGLE(ff_h264_cabac_tables)"+"mlps_off"+128("ret"), "tmp"  \n\t"\
137         "shl    %%cl        , "low"                                     \n\t"\
138         "mov    "tmpbyte"   , "statep"                                  \n\t"\
139         "test   "lowword"   , "lowword"                                 \n\t"\
140         " jnz   2f                                                      \n\t"\
141         "mov    "byte"      , %%"REG_c"                                 \n\t"\
142         "add"OPSIZE" $2     , "byte"                                    \n\t"\
143         "movzwl (%%"REG_c")     , "tmp"                                 \n\t"\
144         "lea    -1("low")   , %%ecx                                     \n\t"\
145         "xor    "low"       , %%ecx                                     \n\t"\
146         "shr    $15         , %%ecx                                     \n\t"\
147         "bswap  "tmp"                                                   \n\t"\
148         "shr    $15         , "tmp"                                     \n\t"\
149         "movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"(%%ecx), %%ecx \n\t"\
150         "sub    $0xFFFF     , "tmp"                                     \n\t"\
151         "neg    %%ecx                                                   \n\t"\
152         "add    $7          , %%ecx                                     \n\t"\
153         "shl    %%cl        , "tmp"                                     \n\t"\
154         "add    "tmp"       , "low"                                     \n\t"\
155         "2:                                                             \n\t"
156
157 #endif /* BROKEN_RELOCATIONS */
158
159 #if HAVE_7REGS && !BROKEN_COMPILER
160 #define get_cabac_inline get_cabac_inline_x86
161 static av_always_inline int get_cabac_inline_x86(CABACContext *c,
162                                                  uint8_t *const state)
163 {
164     int bit, tmp;
165 #ifdef BROKEN_RELOCATIONS
166     void *tables;
167
168     __asm__ volatile(
169         "lea    "MANGLE(ff_h264_cabac_tables)", %0      \n\t"
170         : "=&r"(tables)
171     );
172 #endif
173
174     __asm__ volatile(
175         BRANCHLESS_GET_CABAC("%0", "%q0", "(%4)", "%1", "%w1",
176                              "%2", "%q2", "%3", "%b3",
177                              "%c6(%5)", "%c7(%5)",
178                              AV_STRINGIFY(H264_NORM_SHIFT_OFFSET),
179                              AV_STRINGIFY(H264_LPS_RANGE_OFFSET),
180                              AV_STRINGIFY(H264_MLPS_STATE_OFFSET),
181                              "%8")
182         : "=&r"(bit), "=&r"(c->low), "=&r"(c->range), "=&q"(tmp)
183         : "r"(state), "r"(c),
184           "i"(offsetof(CABACContext, bytestream)),
185           "i"(offsetof(CABACContext, bytestream_end))
186           TABLES_ARG
187           ,"1"(c->low), "2"(c->range)
188         : "%"REG_c, "memory"
189     );
190     return bit & 1;
191 }
192 #endif /* HAVE_7REGS */
193
194 #define get_cabac_bypass_sign get_cabac_bypass_sign_x86
195 static av_always_inline int get_cabac_bypass_sign_x86(CABACContext *c, int val)
196 {
197     x86_reg tmp;
198     __asm__ volatile(
199         "movl        %c6(%2), %k1       \n\t"
200         "movl        %c3(%2), %%eax     \n\t"
201         "shl             $17, %k1       \n\t"
202         "add           %%eax, %%eax     \n\t"
203         "sub             %k1, %%eax     \n\t"
204         "cltd                           \n\t"
205         "and           %%edx, %k1       \n\t"
206         "add             %k1, %%eax     \n\t"
207         "xor           %%edx, %%ecx     \n\t"
208         "sub           %%edx, %%ecx     \n\t"
209         "test           %%ax, %%ax      \n\t"
210         "jnz              1f            \n\t"
211         "mov         %c4(%2), %1        \n\t"
212         "subl        $0xFFFF, %%eax     \n\t"
213         "movzwl         (%1), %%edx     \n\t"
214         "bswap         %%edx            \n\t"
215         "shrl            $15, %%edx     \n\t"
216         "add              $2, %1        \n\t"
217         "addl          %%edx, %%eax     \n\t"
218         "mov              %1, %c4(%2)   \n\t"
219         "1:                             \n\t"
220         "movl          %%eax, %c3(%2)   \n\t"
221
222         : "+c"(val), "=&r"(tmp)
223         : "r"(c),
224           "i"(offsetof(CABACContext, low)),
225           "i"(offsetof(CABACContext, bytestream)),
226           "i"(offsetof(CABACContext, bytestream_end)),
227           "i"(offsetof(CABACContext, range))
228         : "%eax", "%edx", "memory"
229     );
230     return val;
231 }
232
233 #if !BROKEN_COMPILER
234 #define get_cabac_bypass get_cabac_bypass_x86
235 static av_always_inline int get_cabac_bypass_x86(CABACContext *c)
236 {
237     x86_reg tmp;
238     int res;
239     __asm__ volatile(
240         "movl        %c6(%2), %k1       \n\t"
241         "movl        %c3(%2), %%eax     \n\t"
242         "shl             $17, %k1       \n\t"
243         "add           %%eax, %%eax     \n\t"
244         "sub             %k1, %%eax     \n\t"
245         "cltd                           \n\t"
246         "and           %%edx, %k1       \n\t"
247         "add             %k1, %%eax     \n\t"
248         "inc           %%edx            \n\t"
249         "test           %%ax, %%ax      \n\t"
250         "jnz              1f            \n\t"
251         "mov         %c4(%2), %1        \n\t"
252         "subl        $0xFFFF, %%eax     \n\t"
253         "movzwl         (%1), %%ecx     \n\t"
254         "bswap         %%ecx            \n\t"
255         "shrl            $15, %%ecx     \n\t"
256         "addl          %%ecx, %%eax     \n\t"
257         "cmp         %c5(%2), %1        \n\t"
258         "jge              1f            \n\t"
259         "add"OPSIZE"      $2, %c4(%2)   \n\t"
260         "1:                             \n\t"
261         "movl          %%eax, %c3(%2)   \n\t"
262
263         : "=&d"(res), "=&r"(tmp)
264         : "r"(c),
265           "i"(offsetof(CABACContext, low)),
266           "i"(offsetof(CABACContext, bytestream)),
267           "i"(offsetof(CABACContext, bytestream_end)),
268           "i"(offsetof(CABACContext, range))
269         : "%eax", "%ecx", "memory"
270     );
271     return res;
272 }
273 #endif /* !BROKEN_COMPILER */
274
275 #endif /* HAVE_INLINE_ASM */
276 #endif /* AVCODEC_X86_CABAC_H */