git.sesse.net Git - ffmpeg/blob - libavcodec/x86/cabac.h

   1 /*
   2  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
   3  *
   4  * This file is part of FFmpeg.
   5  *
   6  * FFmpeg 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  * FFmpeg 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 FFmpeg; 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 #ifndef UNCHECKED_BITSTREAM_READER
  40 #define UNCHECKED_BITSTREAM_READER !CONFIG_SAFE_BITSTREAM_READER
  41 #endif
  42
  43 #if UNCHECKED_BITSTREAM_READER
  44 #define END_CHECK(end) ""
  45 #else
  46 #define END_CHECK(end) \
  47         "cmp    "end"       , %%"REG_c"                                 \n\t"\
  48         "jge    1f                                                      \n\t"
  49 #endif
  50
  51 #ifdef BROKEN_RELOCATIONS
  52 #define TABLES_ARG , "r"(tables)
  53
  54 #if HAVE_FAST_CMOV
  55 #define BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \
  56         "cmp    "low"       , "tmp"                        \n\t"\
  57         "cmova  %%ecx       , "range"                      \n\t"\
  58         "sbb    %%rcx       , %%rcx                        \n\t"\
  59         "and    %%ecx       , "tmp"                        \n\t"\
  60         "xor    %%rcx       , "retq"                       \n\t"\
  61         "sub    "tmp"       , "low"                        \n\t"
  62 #else /* HAVE_FAST_CMOV */
  63 #define BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \
  64 /* P4 Prescott has crappy cmov,sbb,64bit shift so avoid them */ \
  65         "sub    "low"       , "tmp"                        \n\t"\
  66         "sar    $31         , "tmp"                        \n\t"\
  67         "sub    %%ecx       , "range"                      \n\t"\
  68         "and    "tmp"       , "range"                      \n\t"\
  69         "add    %%ecx       , "range"                      \n\t"\
  70         "shl    $17         , %%ecx                        \n\t"\
  71         "and    "tmp"       , %%ecx                        \n\t"\
  72         "sub    %%ecx       , "low"                        \n\t"\
  73         "xor    "tmp"       , "ret"                        \n\t"\
  74         "movslq "ret"       , "retq"                       \n\t"
  75 #endif /* HAVE_FAST_CMOV */
  76
  77 #define BRANCHLESS_GET_CABAC(ret, retq, statep, low, lowword, range, rangeq, tmp, tmpbyte, byte, end, norm_off, lps_off, mlps_off, tables) \
  78         "movzbl "statep"    , "ret"                                     \n\t"\
  79         "mov    "range"     , "tmp"                                     \n\t"\
  80         "and    $0xC0       , "range"                                   \n\t"\
  81         "lea    ("ret", "range", 2), %%ecx                              \n\t"\
  82         "movzbl "lps_off"("tables", %%rcx), "range"                     \n\t"\
  83         "sub    "range"     , "tmp"                                     \n\t"\
  84         "mov    "tmp"       , %%ecx                                     \n\t"\
  85         "shl    $17         , "tmp"                                     \n\t"\
  86         BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp)              \
  87         "movzbl "norm_off"("tables", "rangeq"), %%ecx                   \n\t"\
  88         "shl    %%cl        , "range"                                   \n\t"\
  89         "movzbl "mlps_off"+128("tables", "retq"), "tmp"                 \n\t"\
  90         "shl    %%cl        , "low"                                     \n\t"\
  91         "mov    "tmpbyte"   , "statep"                                  \n\t"\
  92         "test   "lowword"   , "lowword"                                 \n\t"\
  93         "jnz    2f                                                      \n\t"\
  94         "mov    "byte"      , %%"REG_c"                                 \n\t"\
  95         END_CHECK(end)\
  96         "add"OPSIZE" $2     , "byte"                                    \n\t"\
  97         "1:                                                             \n\t"\
  98         "movzwl (%%"REG_c") , "tmp"                                     \n\t"\
  99         "lea    -1("low")   , %%ecx                                     \n\t"\
 100         "xor    "low"       , %%ecx                                     \n\t"\
 101         "shr    $15         , %%ecx                                     \n\t"\
 102         "bswap  "tmp"                                                   \n\t"\
 103         "shr    $15         , "tmp"                                     \n\t"\
 104         "movzbl "norm_off"("tables", %%rcx), %%ecx                      \n\t"\
 105         "sub    $0xFFFF     , "tmp"                                     \n\t"\
 106         "neg    %%ecx                                                   \n\t"\
 107         "add    $7          , %%ecx                                     \n\t"\
 108         "shl    %%cl        , "tmp"                                     \n\t"\
 109         "add    "tmp"       , "low"                                     \n\t"\
 110         "2:                                                             \n\t"
 111
 112 #else /* BROKEN_RELOCATIONS */
 113 #define TABLES_ARG
 114 #define RIP_ARG
 115
 116 #if HAVE_FAST_CMOV
 117 #define BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)\
 118         "mov    "tmp"       , %%ecx     \n\t"\
 119         "shl    $17         , "tmp"     \n\t"\
 120         "cmp    "low"       , "tmp"     \n\t"\
 121         "cmova  %%ecx       , "range"   \n\t"\
 122         "sbb    %%ecx       , %%ecx     \n\t"\
 123         "and    %%ecx       , "tmp"     \n\t"\
 124         "xor    %%ecx       , "ret"     \n\t"\
 125         "sub    "tmp"       , "low"     \n\t"
 126 #else /* HAVE_FAST_CMOV */
 127 #define BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)\
 128         "mov    "tmp"       , %%ecx     \n\t"\
 129         "shl    $17         , "tmp"     \n\t"\
 130         "sub    "low"       , "tmp"     \n\t"\
 131         "sar    $31         , "tmp"     \n\t" /*lps_mask*/\
 132         "sub    %%ecx       , "range"   \n\t" /*RangeLPS - range*/\
 133         "and    "tmp"       , "range"   \n\t" /*(RangeLPS - range)&lps_mask*/\
 134         "add    %%ecx       , "range"   \n\t" /*new range*/\
 135         "shl    $17         , %%ecx     \n\t"\
 136         "and    "tmp"       , %%ecx     \n\t"\
 137         "sub    %%ecx       , "low"     \n\t"\
 138         "xor    "tmp"       , "ret"     \n\t"
 139 #endif /* HAVE_FAST_CMOV */
 140
 141 #define BRANCHLESS_GET_CABAC(ret, retq, statep, low, lowword, range, rangeq, tmp, tmpbyte, byte, end, norm_off, lps_off, mlps_off, tables) \
 142         "movzbl "statep"    , "ret"                                     \n\t"\
 143         "mov    "range"     , "tmp"                                     \n\t"\
 144         "and    $0xC0       , "range"                                   \n\t"\
 145         "movzbl "MANGLE(ff_h264_cabac_tables)"+"lps_off"("ret", "range", 2), "range" \n\t"\
 146         "sub    "range"     , "tmp"                                     \n\t"\
 147         BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)                    \
 148         "movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"("range"), %%ecx    \n\t"\
 149         "shl    %%cl        , "range"                                   \n\t"\
 150         "movzbl "MANGLE(ff_h264_cabac_tables)"+"mlps_off"+128("ret"), "tmp"  \n\t"\
 151         "shl    %%cl        , "low"                                     \n\t"\
 152         "mov    "tmpbyte"   , "statep"                                  \n\t"\
 153         "test   "lowword"   , "lowword"                                 \n\t"\
 154         " jnz   2f                                                      \n\t"\
 155         "mov    "byte"      , %%"REG_c"                                 \n\t"\
 156         END_CHECK(end)\
 157         "add"OPSIZE" $2     , "byte"                                    \n\t"\
 158         "1:                                                             \n\t"\
 159         "movzwl (%%"REG_c")     , "tmp"                                 \n\t"\
 160         "lea    -1("low")   , %%ecx                                     \n\t"\
 161         "xor    "low"       , %%ecx                                     \n\t"\
 162         "shr    $15         , %%ecx                                     \n\t"\
 163         "bswap  "tmp"                                                   \n\t"\
 164         "shr    $15         , "tmp"                                     \n\t"\
 165         "movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"(%%ecx), %%ecx \n\t"\
 166         "sub    $0xFFFF     , "tmp"                                     \n\t"\
 167         "neg    %%ecx                                                   \n\t"\
 168         "add    $7          , %%ecx                                     \n\t"\
 169         "shl    %%cl        , "tmp"                                     \n\t"\
 170         "add    "tmp"       , "low"                                     \n\t"\
 171         "2:                                                             \n\t"
 172
 173 #endif /* BROKEN_RELOCATIONS */
 174
 175 #if HAVE_7REGS && !BROKEN_COMPILER
 176 #define get_cabac_inline get_cabac_inline_x86
 177 static av_always_inline int get_cabac_inline_x86(CABACContext *c,
 178                                                  uint8_t *const state)
 179 {
 180     int bit, tmp;
 181 #ifdef BROKEN_RELOCATIONS
 182     void *tables;
 183
 184     __asm__ volatile(
 185         "lea    "MANGLE(ff_h264_cabac_tables)", %0      \n\t"
 186         : "=&r"(tables)
 187     );
 188 #endif
 189
 190     __asm__ volatile(
 191         BRANCHLESS_GET_CABAC("%0", "%q0", "(%4)", "%1", "%w1",
 192                              "%2", "%q2", "%3", "%b3",
 193                              "%c6(%5)", "%c7(%5)",
 194                              AV_STRINGIFY(H264_NORM_SHIFT_OFFSET),
 195                              AV_STRINGIFY(H264_LPS_RANGE_OFFSET),
 196                              AV_STRINGIFY(H264_MLPS_STATE_OFFSET),
 197                              "%8")
 198         : "=&r"(bit), "=&r"(c->low), "=&r"(c->range), "=&q"(tmp)
 199         : "r"(state), "r"(c),
 200           "i"(offsetof(CABACContext, bytestream)),
 201           "i"(offsetof(CABACContext, bytestream_end))
 202           TABLES_ARG
 203           ,"1"(c->low), "2"(c->range)
 204         : "%"REG_c, "memory"
 205     );
 206     return bit & 1;
 207 }
 208 #endif /* HAVE_7REGS */
 209
 210 #if !BROKEN_COMPILER
 211 #define get_cabac_bypass_sign get_cabac_bypass_sign_x86
 212 static av_always_inline int get_cabac_bypass_sign_x86(CABACContext *c, int val)
 213 {
 214     x86_reg tmp;
 215     __asm__ volatile(
 216         "movl        %c6(%2), %k1       \n\t"
 217         "movl        %c3(%2), %%eax     \n\t"
 218         "shl             $17, %k1       \n\t"
 219         "add           %%eax, %%eax     \n\t"
 220         "sub             %k1, %%eax     \n\t"
 221         "cltd                           \n\t"
 222         "and           %%edx, %k1       \n\t"
 223         "add             %k1, %%eax     \n\t"
 224         "xor           %%edx, %%ecx     \n\t"
 225         "sub           %%edx, %%ecx     \n\t"
 226         "test           %%ax, %%ax      \n\t"
 227         "jnz              1f            \n\t"
 228         "mov         %c4(%2), %1        \n\t"
 229         "subl        $0xFFFF, %%eax     \n\t"
 230         "movzwl         (%1), %%edx     \n\t"
 231         "bswap         %%edx            \n\t"
 232         "shrl            $15, %%edx     \n\t"
 233 #if UNCHECKED_BITSTREAM_READER
 234         "add              $2, %1        \n\t"
 235         "addl          %%edx, %%eax     \n\t"
 236         "mov              %1, %c4(%2)   \n\t"
 237 #else
 238         "addl          %%edx, %%eax     \n\t"
 239         "cmp         %c5(%2), %1        \n\t"
 240         "jge              1f            \n\t"
 241         "add"OPSIZE"      $2, %c4(%2)   \n\t"
 242 #endif
 243         "1:                             \n\t"
 244         "movl          %%eax, %c3(%2)   \n\t"
 245
 246         : "+c"(val), "=&r"(tmp)
 247         : "r"(c),
 248           "i"(offsetof(CABACContext, low)),
 249           "i"(offsetof(CABACContext, bytestream)),
 250           "i"(offsetof(CABACContext, bytestream_end)),
 251           "i"(offsetof(CABACContext, range))
 252         : "%eax", "%edx", "memory"
 253     );
 254     return val;
 255 }
 256
 257 #define get_cabac_bypass get_cabac_bypass_x86
 258 static av_always_inline int get_cabac_bypass_x86(CABACContext *c)
 259 {
 260     x86_reg tmp;
 261     int res;
 262     __asm__ volatile(
 263         "movl        %c6(%2), %k1       \n\t"
 264         "movl        %c3(%2), %%eax     \n\t"
 265         "shl             $17, %k1       \n\t"
 266         "add           %%eax, %%eax     \n\t"
 267         "sub             %k1, %%eax     \n\t"
 268         "cltd                           \n\t"
 269         "and           %%edx, %k1       \n\t"
 270         "add             %k1, %%eax     \n\t"
 271         "inc           %%edx            \n\t"
 272         "test           %%ax, %%ax      \n\t"
 273         "jnz              1f            \n\t"
 274         "mov         %c4(%2), %1        \n\t"
 275         "subl        $0xFFFF, %%eax     \n\t"
 276         "movzwl         (%1), %%ecx     \n\t"
 277         "bswap         %%ecx            \n\t"
 278         "shrl            $15, %%ecx     \n\t"
 279         "addl          %%ecx, %%eax     \n\t"
 280         "cmp         %c5(%2), %1        \n\t"
 281         "jge              1f            \n\t"
 282         "add"OPSIZE"      $2, %c4(%2)   \n\t"
 283         "1:                             \n\t"
 284         "movl          %%eax, %c3(%2)   \n\t"
 285
 286         : "=&d"(res), "=&r"(tmp)
 287         : "r"(c),
 288           "i"(offsetof(CABACContext, low)),
 289           "i"(offsetof(CABACContext, bytestream)),
 290           "i"(offsetof(CABACContext, bytestream_end)),
 291           "i"(offsetof(CABACContext, range))
 292         : "%eax", "%ecx", "memory"
 293     );
 294     return res;
 295 }
 296 #endif /* !BROKEN_COMPILER */
 297
 298 #endif /* HAVE_INLINE_ASM */
 299 #endif /* AVCODEC_X86_CABAC_H */