| 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/macros.h" |
| 27 | #include "libavutil/x86/asm.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 | || (defined(__INTEL_COMPILER) && defined(_MSC_VER)) |
| 33 | # define BROKEN_COMPILER 1 |
| 34 | #else |
| 35 | # define BROKEN_COMPILER 0 |
| 36 | #endif |
| 37 | |
| 38 | #if HAVE_INLINE_ASM |
| 39 | |
| 40 | #ifndef UNCHECKED_BITSTREAM_READER |
| 41 | #define UNCHECKED_BITSTREAM_READER !CONFIG_SAFE_BITSTREAM_READER |
| 42 | #endif |
| 43 | |
| 44 | #if UNCHECKED_BITSTREAM_READER |
| 45 | #define END_CHECK(end) "" |
| 46 | #else |
| 47 | #define END_CHECK(end) \ |
| 48 | "cmp "end" , %%"REG_c" \n\t"\ |
| 49 | "jge 1f \n\t" |
| 50 | #endif |
| 51 | |
| 52 | #ifdef BROKEN_RELOCATIONS |
| 53 | #define TABLES_ARG , "r"(tables) |
| 54 | |
| 55 | #if HAVE_FAST_CMOV |
| 56 | #define BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \ |
| 57 | "cmp "low" , "tmp" \n\t"\ |
| 58 | "cmova %%ecx , "range" \n\t"\ |
| 59 | "sbb %%rcx , %%rcx \n\t"\ |
| 60 | "and %%ecx , "tmp" \n\t"\ |
| 61 | "xor %%rcx , "retq" \n\t"\ |
| 62 | "sub "tmp" , "low" \n\t" |
| 63 | #else /* HAVE_FAST_CMOV */ |
| 64 | #define BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \ |
| 65 | /* P4 Prescott has crappy cmov,sbb,64bit shift so avoid them */ \ |
| 66 | "sub "low" , "tmp" \n\t"\ |
| 67 | "sar $31 , "tmp" \n\t"\ |
| 68 | "sub %%ecx , "range" \n\t"\ |
| 69 | "and "tmp" , "range" \n\t"\ |
| 70 | "add %%ecx , "range" \n\t"\ |
| 71 | "shl $17 , %%ecx \n\t"\ |
| 72 | "and "tmp" , %%ecx \n\t"\ |
| 73 | "sub %%ecx , "low" \n\t"\ |
| 74 | "xor "tmp" , "ret" \n\t"\ |
| 75 | "movslq "ret" , "retq" \n\t" |
| 76 | #endif /* HAVE_FAST_CMOV */ |
| 77 | |
| 78 | #define BRANCHLESS_GET_CABAC(ret, retq, statep, low, lowword, range, rangeq, tmp, tmpbyte, byte, end, norm_off, lps_off, mlps_off, tables) \ |
| 79 | "movzbl "statep" , "ret" \n\t"\ |
| 80 | "mov "range" , "tmp" \n\t"\ |
| 81 | "and $0xC0 , "range" \n\t"\ |
| 82 | "lea ("ret", "range", 2), %%ecx \n\t"\ |
| 83 | "movzbl "lps_off"("tables", %%rcx), "range" \n\t"\ |
| 84 | "sub "range" , "tmp" \n\t"\ |
| 85 | "mov "tmp" , %%ecx \n\t"\ |
| 86 | "shl $17 , "tmp" \n\t"\ |
| 87 | BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \ |
| 88 | "movzbl "norm_off"("tables", "rangeq"), %%ecx \n\t"\ |
| 89 | "shl %%cl , "range" \n\t"\ |
| 90 | "movzbl "mlps_off"+128("tables", "retq"), "tmp" \n\t"\ |
| 91 | "shl %%cl , "low" \n\t"\ |
| 92 | "mov "tmpbyte" , "statep" \n\t"\ |
| 93 | "test "lowword" , "lowword" \n\t"\ |
| 94 | "jnz 2f \n\t"\ |
| 95 | "mov "byte" , %%"REG_c" \n\t"\ |
| 96 | END_CHECK(end)\ |
| 97 | "add"OPSIZE" $2 , "byte" \n\t"\ |
| 98 | "1: \n\t"\ |
| 99 | "movzwl (%%"REG_c") , "tmp" \n\t"\ |
| 100 | "lea -1("low") , %%ecx \n\t"\ |
| 101 | "xor "low" , %%ecx \n\t"\ |
| 102 | "shr $15 , %%ecx \n\t"\ |
| 103 | "bswap "tmp" \n\t"\ |
| 104 | "shr $15 , "tmp" \n\t"\ |
| 105 | "movzbl "norm_off"("tables", %%rcx), %%ecx \n\t"\ |
| 106 | "sub $0xFFFF , "tmp" \n\t"\ |
| 107 | "neg %%ecx \n\t"\ |
| 108 | "add $7 , %%ecx \n\t"\ |
| 109 | "shl %%cl , "tmp" \n\t"\ |
| 110 | "add "tmp" , "low" \n\t"\ |
| 111 | "2: \n\t" |
| 112 | |
| 113 | #else /* BROKEN_RELOCATIONS */ |
| 114 | #define TABLES_ARG NAMED_CONSTRAINTS_ARRAY_ADD(ff_h264_cabac_tables) |
| 115 | #define RIP_ARG |
| 116 | |
| 117 | #if HAVE_FAST_CMOV |
| 118 | #define BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)\ |
| 119 | "mov "tmp" , %%ecx \n\t"\ |
| 120 | "shl $17 , "tmp" \n\t"\ |
| 121 | "cmp "low" , "tmp" \n\t"\ |
| 122 | "cmova %%ecx , "range" \n\t"\ |
| 123 | "sbb %%ecx , %%ecx \n\t"\ |
| 124 | "and %%ecx , "tmp" \n\t"\ |
| 125 | "xor %%ecx , "ret" \n\t"\ |
| 126 | "sub "tmp" , "low" \n\t" |
| 127 | #else /* HAVE_FAST_CMOV */ |
| 128 | #define BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)\ |
| 129 | "mov "tmp" , %%ecx \n\t"\ |
| 130 | "shl $17 , "tmp" \n\t"\ |
| 131 | "sub "low" , "tmp" \n\t"\ |
| 132 | "sar $31 , "tmp" \n\t" /*lps_mask*/\ |
| 133 | "sub %%ecx , "range" \n\t" /*RangeLPS - range*/\ |
| 134 | "and "tmp" , "range" \n\t" /*(RangeLPS - range)&lps_mask*/\ |
| 135 | "add %%ecx , "range" \n\t" /*new range*/\ |
| 136 | "shl $17 , %%ecx \n\t"\ |
| 137 | "and "tmp" , %%ecx \n\t"\ |
| 138 | "sub %%ecx , "low" \n\t"\ |
| 139 | "xor "tmp" , "ret" \n\t" |
| 140 | #endif /* HAVE_FAST_CMOV */ |
| 141 | |
| 142 | #define BRANCHLESS_GET_CABAC(ret, retq, statep, low, lowword, range, rangeq, tmp, tmpbyte, byte, end, norm_off, lps_off, mlps_off, tables) \ |
| 143 | "movzbl "statep" , "ret" \n\t"\ |
| 144 | "mov "range" , "tmp" \n\t"\ |
| 145 | "and $0xC0 , "range" \n\t"\ |
| 146 | "movzbl "MANGLE(ff_h264_cabac_tables)"+"lps_off"("ret", "range", 2), "range" \n\t"\ |
| 147 | "sub "range" , "tmp" \n\t"\ |
| 148 | BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp) \ |
| 149 | "movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"("range"), %%ecx \n\t"\ |
| 150 | "shl %%cl , "range" \n\t"\ |
| 151 | "movzbl "MANGLE(ff_h264_cabac_tables)"+"mlps_off"+128("ret"), "tmp" \n\t"\ |
| 152 | "shl %%cl , "low" \n\t"\ |
| 153 | "mov "tmpbyte" , "statep" \n\t"\ |
| 154 | "test "lowword" , "lowword" \n\t"\ |
| 155 | " jnz 2f \n\t"\ |
| 156 | "mov "byte" , %%"REG_c" \n\t"\ |
| 157 | END_CHECK(end)\ |
| 158 | "add"OPSIZE" $2 , "byte" \n\t"\ |
| 159 | "1: \n\t"\ |
| 160 | "movzwl (%%"REG_c") , "tmp" \n\t"\ |
| 161 | "lea -1("low") , %%ecx \n\t"\ |
| 162 | "xor "low" , %%ecx \n\t"\ |
| 163 | "shr $15 , %%ecx \n\t"\ |
| 164 | "bswap "tmp" \n\t"\ |
| 165 | "shr $15 , "tmp" \n\t"\ |
| 166 | "movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"(%%ecx), %%ecx \n\t"\ |
| 167 | "sub $0xFFFF , "tmp" \n\t"\ |
| 168 | "neg %%ecx \n\t"\ |
| 169 | "add $7 , %%ecx \n\t"\ |
| 170 | "shl %%cl , "tmp" \n\t"\ |
| 171 | "add "tmp" , "low" \n\t"\ |
| 172 | "2: \n\t" |
| 173 | |
| 174 | #endif /* BROKEN_RELOCATIONS */ |
| 175 | |
| 176 | #if HAVE_7REGS && !BROKEN_COMPILER |
| 177 | #define get_cabac_inline get_cabac_inline_x86 |
| 178 | static av_always_inline int get_cabac_inline_x86(CABACContext *c, |
| 179 | uint8_t *const state) |
| 180 | { |
| 181 | int bit, tmp; |
| 182 | #ifdef BROKEN_RELOCATIONS |
| 183 | void *tables; |
| 184 | |
| 185 | __asm__ volatile( |
| 186 | "lea "MANGLE(ff_h264_cabac_tables)", %0 \n\t" |
| 187 | : "=&r"(tables) |
| 188 | : NAMED_CONSTRAINTS_ARRAY(ff_h264_cabac_tables) |
| 189 | ); |
| 190 | #endif |
| 191 | |
| 192 | __asm__ volatile( |
| 193 | BRANCHLESS_GET_CABAC("%0", "%q0", "(%4)", "%1", "%w1", |
| 194 | "%2", "%q2", "%3", "%b3", |
| 195 | "%c6(%5)", "%c7(%5)", |
| 196 | AV_STRINGIFY(H264_NORM_SHIFT_OFFSET), |
| 197 | AV_STRINGIFY(H264_LPS_RANGE_OFFSET), |
| 198 | AV_STRINGIFY(H264_MLPS_STATE_OFFSET), |
| 199 | "%8") |
| 200 | : "=&r"(bit), "=&r"(c->low), "=&r"(c->range), "=&q"(tmp) |
| 201 | : "r"(state), "r"(c), |
| 202 | "i"(offsetof(CABACContext, bytestream)), |
| 203 | "i"(offsetof(CABACContext, bytestream_end)) |
| 204 | TABLES_ARG |
| 205 | ,"1"(c->low), "2"(c->range) |
| 206 | : "%"REG_c, "memory" |
| 207 | ); |
| 208 | return bit & 1; |
| 209 | } |
| 210 | #endif /* HAVE_7REGS && !BROKEN_COMPILER */ |
| 211 | |
| 212 | #if !BROKEN_COMPILER |
| 213 | #define get_cabac_bypass_sign get_cabac_bypass_sign_x86 |
| 214 | static av_always_inline int get_cabac_bypass_sign_x86(CABACContext *c, int val) |
| 215 | { |
| 216 | x86_reg tmp; |
| 217 | __asm__ volatile( |
| 218 | "movl %c6(%2), %k1 \n\t" |
| 219 | "movl %c3(%2), %%eax \n\t" |
| 220 | "shl $17, %k1 \n\t" |
| 221 | "add %%eax, %%eax \n\t" |
| 222 | "sub %k1, %%eax \n\t" |
| 223 | "cdq \n\t" |
| 224 | "and %%edx, %k1 \n\t" |
| 225 | "add %k1, %%eax \n\t" |
| 226 | "xor %%edx, %%ecx \n\t" |
| 227 | "sub %%edx, %%ecx \n\t" |
| 228 | "test %%ax, %%ax \n\t" |
| 229 | "jnz 1f \n\t" |
| 230 | "mov %c4(%2), %1 \n\t" |
| 231 | "subl $0xFFFF, %%eax \n\t" |
| 232 | "movzwl (%1), %%edx \n\t" |
| 233 | "bswap %%edx \n\t" |
| 234 | "shrl $15, %%edx \n\t" |
| 235 | #if UNCHECKED_BITSTREAM_READER |
| 236 | "add $2, %1 \n\t" |
| 237 | "addl %%edx, %%eax \n\t" |
| 238 | "mov %1, %c4(%2) \n\t" |
| 239 | #else |
| 240 | "addl %%edx, %%eax \n\t" |
| 241 | "cmp %c5(%2), %1 \n\t" |
| 242 | "jge 1f \n\t" |
| 243 | "add"OPSIZE" $2, %c4(%2) \n\t" |
| 244 | #endif |
| 245 | "1: \n\t" |
| 246 | "movl %%eax, %c3(%2) \n\t" |
| 247 | |
| 248 | : "+c"(val), "=&r"(tmp) |
| 249 | : "r"(c), |
| 250 | "i"(offsetof(CABACContext, low)), |
| 251 | "i"(offsetof(CABACContext, bytestream)), |
| 252 | "i"(offsetof(CABACContext, bytestream_end)), |
| 253 | "i"(offsetof(CABACContext, range)) |
| 254 | : "%eax", "%edx", "memory" |
| 255 | ); |
| 256 | return val; |
| 257 | } |
| 258 | |
| 259 | #define get_cabac_bypass get_cabac_bypass_x86 |
| 260 | static av_always_inline int get_cabac_bypass_x86(CABACContext *c) |
| 261 | { |
| 262 | x86_reg tmp; |
| 263 | int res; |
| 264 | __asm__ volatile( |
| 265 | "movl %c6(%2), %k1 \n\t" |
| 266 | "movl %c3(%2), %%eax \n\t" |
| 267 | "shl $17, %k1 \n\t" |
| 268 | "add %%eax, %%eax \n\t" |
| 269 | "sub %k1, %%eax \n\t" |
| 270 | "cdq \n\t" |
| 271 | "and %%edx, %k1 \n\t" |
| 272 | "add %k1, %%eax \n\t" |
| 273 | "inc %%edx \n\t" |
| 274 | "test %%ax, %%ax \n\t" |
| 275 | "jnz 1f \n\t" |
| 276 | "mov %c4(%2), %1 \n\t" |
| 277 | "subl $0xFFFF, %%eax \n\t" |
| 278 | "movzwl (%1), %%ecx \n\t" |
| 279 | "bswap %%ecx \n\t" |
| 280 | "shrl $15, %%ecx \n\t" |
| 281 | "addl %%ecx, %%eax \n\t" |
| 282 | "cmp %c5(%2), %1 \n\t" |
| 283 | "jge 1f \n\t" |
| 284 | "add"OPSIZE" $2, %c4(%2) \n\t" |
| 285 | "1: \n\t" |
| 286 | "movl %%eax, %c3(%2) \n\t" |
| 287 | |
| 288 | : "=&d"(res), "=&r"(tmp) |
| 289 | : "r"(c), |
| 290 | "i"(offsetof(CABACContext, low)), |
| 291 | "i"(offsetof(CABACContext, bytestream)), |
| 292 | "i"(offsetof(CABACContext, bytestream_end)), |
| 293 | "i"(offsetof(CABACContext, range)) |
| 294 | : "%eax", "%ecx", "memory" |
| 295 | ); |
| 296 | return res; |
| 297 | } |
| 298 | #endif /* !BROKEN_COMPILER */ |
| 299 | |
| 300 | #endif /* HAVE_INLINE_ASM */ |
| 301 | #endif /* AVCODEC_X86_CABAC_H */ |