| Line | Branch | Exec | Source |
|---|---|---|---|
| 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 <stddef.h> | ||
| 25 | |||
| 26 | #include "libavcodec/cabac.h" | ||
| 27 | #include "libavutil/attributes.h" | ||
| 28 | #include "libavutil/macros.h" | ||
| 29 | #include "libavutil/x86/asm.h" | ||
| 30 | #include "config.h" | ||
| 31 | |||
| 32 | #if (defined(__i386) && defined(__clang__) && (__clang_major__<2 || (__clang_major__==2 && __clang_minor__<10)))\ | ||
| 33 | || ( !defined(__clang__) && defined(__llvm__) && __GNUC__==4 && __GNUC_MINOR__==2 && __GNUC_PATCHLEVEL__<=1)\ | ||
| 34 | || (defined(__INTEL_COMPILER) && defined(_MSC_VER)) | ||
| 35 | # define BROKEN_COMPILER 1 | ||
| 36 | #else | ||
| 37 | # define BROKEN_COMPILER 0 | ||
| 38 | #endif | ||
| 39 | |||
| 40 | #if HAVE_INLINE_ASM | ||
| 41 | |||
| 42 | #ifndef UNCHECKED_BITSTREAM_READER | ||
| 43 | #define UNCHECKED_BITSTREAM_READER !CONFIG_SAFE_BITSTREAM_READER | ||
| 44 | #endif | ||
| 45 | |||
| 46 | #if UNCHECKED_BITSTREAM_READER | ||
| 47 | #define END_CHECK(end) "" | ||
| 48 | #else | ||
| 49 | #define END_CHECK(end) \ | ||
| 50 | "cmp "end" , %%"FF_REG_c" \n\t"\ | ||
| 51 | "jge 1f \n\t" | ||
| 52 | #endif | ||
| 53 | |||
| 54 | #ifdef BROKEN_RELOCATIONS | ||
| 55 | #define TABLES_ARG , "r"(tables) | ||
| 56 | |||
| 57 | #if HAVE_FAST_CMOV | ||
| 58 | #define BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \ | ||
| 59 | "cmp "low" , "tmp" \n\t"\ | ||
| 60 | "cmova %%ecx , "range" \n\t"\ | ||
| 61 | "sbb %%rcx , %%rcx \n\t"\ | ||
| 62 | "and %%ecx , "tmp" \n\t"\ | ||
| 63 | "xor %%rcx , "retq" \n\t"\ | ||
| 64 | "sub "tmp" , "low" \n\t" | ||
| 65 | #else /* HAVE_FAST_CMOV */ | ||
| 66 | #define BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \ | ||
| 67 | /* P4 Prescott has crappy cmov,sbb,64-bit shift so avoid them */ \ | ||
| 68 | "sub "low" , "tmp" \n\t"\ | ||
| 69 | "sar $31 , "tmp" \n\t"\ | ||
| 70 | "sub %%ecx , "range" \n\t"\ | ||
| 71 | "and "tmp" , "range" \n\t"\ | ||
| 72 | "add %%ecx , "range" \n\t"\ | ||
| 73 | "shl $17 , %%ecx \n\t"\ | ||
| 74 | "and "tmp" , %%ecx \n\t"\ | ||
| 75 | "sub %%ecx , "low" \n\t"\ | ||
| 76 | "xor "tmp" , "ret" \n\t"\ | ||
| 77 | "movslq "ret" , "retq" \n\t" | ||
| 78 | #endif /* HAVE_FAST_CMOV */ | ||
| 79 | |||
| 80 | #ifdef __LZCNT__ | ||
| 81 | /* renormalisation shift = lzcnt(range) - 23, refill shift = tzcnt(low) - 16 */ | ||
| 82 | #define BRANCHLESS_GET_CABAC_RENORM(range, rangeq, norm_off, tables) \ | ||
| 83 | "lzcnt "range" , %%ecx \n\t"\ | ||
| 84 | "sub $23 , %%ecx \n\t" | ||
| 85 | #define BRANCHLESS_GET_CABAC_REFILL(low, tmp, norm_off, tables) \ | ||
| 86 | "movzwl (%%"FF_REG_c") , "tmp" \n\t"\ | ||
| 87 | "tzcnt "low" , %%ecx \n\t"\ | ||
| 88 | "bswap "tmp" \n\t"\ | ||
| 89 | "shr $15 , "tmp" \n\t"\ | ||
| 90 | "sub $16 , %%ecx \n\t"\ | ||
| 91 | "sub $0xFFFF , "tmp" \n\t"\ | ||
| 92 | "shl %%cl , "tmp" \n\t"\ | ||
| 93 | "add "tmp" , "low" \n\t" | ||
| 94 | #else /* __LZCNT__ */ | ||
| 95 | #define BRANCHLESS_GET_CABAC_RENORM(range, rangeq, norm_off, tables) \ | ||
| 96 | "movzbl "norm_off"("tables", "rangeq"), %%ecx \n\t" | ||
| 97 | #define BRANCHLESS_GET_CABAC_REFILL(low, tmp, norm_off, tables) \ | ||
| 98 | "movzwl (%%"FF_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 | #endif /* __LZCNT__ */ | ||
| 111 | |||
| 112 | #define BRANCHLESS_GET_CABAC(ret, retq, statep, low, lowword, range, rangeq, tmp, tmpbyte, byte, end, norm_off, lps_off, mlps_off, tables) \ | ||
| 113 | "movzbl "statep" , "ret" \n\t"\ | ||
| 114 | "mov "range" , "tmp" \n\t"\ | ||
| 115 | "and $0xC0 , "range" \n\t"\ | ||
| 116 | "lea ("ret", "range", 2), %%ecx \n\t"\ | ||
| 117 | "movzbl "lps_off"("tables", %%rcx), "range" \n\t"\ | ||
| 118 | "sub "range" , "tmp" \n\t"\ | ||
| 119 | "mov "tmp" , %%ecx \n\t"\ | ||
| 120 | "shl $17 , "tmp" \n\t"\ | ||
| 121 | BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \ | ||
| 122 | BRANCHLESS_GET_CABAC_RENORM(range, rangeq, norm_off, tables) \ | ||
| 123 | "shl %%cl , "range" \n\t"\ | ||
| 124 | "movzbl "mlps_off"+128("tables", "retq"), "tmp" \n\t"\ | ||
| 125 | "shl %%cl , "low" \n\t"\ | ||
| 126 | "mov "tmpbyte" , "statep" \n\t"\ | ||
| 127 | "test "lowword" , "lowword" \n\t"\ | ||
| 128 | "jnz 2f \n\t"\ | ||
| 129 | "mov "byte" , %%"FF_REG_c" \n\t"\ | ||
| 130 | END_CHECK(end)\ | ||
| 131 | "add"FF_OPSIZE" $2 , "byte" \n\t"\ | ||
| 132 | "1: \n\t"\ | ||
| 133 | BRANCHLESS_GET_CABAC_REFILL(low, tmp, norm_off, tables) \ | ||
| 134 | "2: \n\t" | ||
| 135 | |||
| 136 | #else /* BROKEN_RELOCATIONS */ | ||
| 137 | #define TABLES_ARG NAMED_CONSTRAINTS_ARRAY_ADD(ff_h264_cabac_tables) | ||
| 138 | #define RIP_ARG | ||
| 139 | |||
| 140 | #if HAVE_FAST_CMOV | ||
| 141 | #define BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)\ | ||
| 142 | "mov "tmp" , %%ecx \n\t"\ | ||
| 143 | "shl $17 , "tmp" \n\t"\ | ||
| 144 | "cmp "low" , "tmp" \n\t"\ | ||
| 145 | "cmova %%ecx , "range" \n\t"\ | ||
| 146 | "sbb %%ecx , %%ecx \n\t"\ | ||
| 147 | "and %%ecx , "tmp" \n\t"\ | ||
| 148 | "xor %%ecx , "ret" \n\t"\ | ||
| 149 | "sub "tmp" , "low" \n\t" | ||
| 150 | #else /* HAVE_FAST_CMOV */ | ||
| 151 | #define BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)\ | ||
| 152 | "mov "tmp" , %%ecx \n\t"\ | ||
| 153 | "shl $17 , "tmp" \n\t"\ | ||
| 154 | "sub "low" , "tmp" \n\t"\ | ||
| 155 | "sar $31 , "tmp" \n\t" /*lps_mask*/\ | ||
| 156 | "sub %%ecx , "range" \n\t" /*RangeLPS - range*/\ | ||
| 157 | "and "tmp" , "range" \n\t" /*(RangeLPS - range)&lps_mask*/\ | ||
| 158 | "add %%ecx , "range" \n\t" /*new range*/\ | ||
| 159 | "shl $17 , %%ecx \n\t"\ | ||
| 160 | "and "tmp" , %%ecx \n\t"\ | ||
| 161 | "sub %%ecx , "low" \n\t"\ | ||
| 162 | "xor "tmp" , "ret" \n\t" | ||
| 163 | #endif /* HAVE_FAST_CMOV */ | ||
| 164 | |||
| 165 | #ifdef __LZCNT__ | ||
| 166 | #define BRANCHLESS_GET_CABAC_RENORM(range, rangeq, norm_off, tables) \ | ||
| 167 | "lzcnt "range" , %%ecx \n\t"\ | ||
| 168 | "sub $23 , %%ecx \n\t" | ||
| 169 | #define BRANCHLESS_GET_CABAC_REFILL(low, tmp, norm_off, tables) \ | ||
| 170 | "movzwl (%%"FF_REG_c") , "tmp" \n\t"\ | ||
| 171 | "tzcnt "low" , %%ecx \n\t"\ | ||
| 172 | "bswap "tmp" \n\t"\ | ||
| 173 | "shr $15 , "tmp" \n\t"\ | ||
| 174 | "sub $16 , %%ecx \n\t"\ | ||
| 175 | "sub $0xFFFF , "tmp" \n\t"\ | ||
| 176 | "shl %%cl , "tmp" \n\t"\ | ||
| 177 | "add "tmp" , "low" \n\t" | ||
| 178 | #else /* __LZCNT__ */ | ||
| 179 | #define BRANCHLESS_GET_CABAC_RENORM(range, rangeq, norm_off, tables) \ | ||
| 180 | "movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"("range"), %%ecx \n\t" | ||
| 181 | #define BRANCHLESS_GET_CABAC_REFILL(low, tmp, norm_off, tables) \ | ||
| 182 | "movzwl (%%"FF_REG_c") , "tmp" \n\t"\ | ||
| 183 | "lea -1("low") , %%ecx \n\t"\ | ||
| 184 | "xor "low" , %%ecx \n\t"\ | ||
| 185 | "shr $15 , %%ecx \n\t"\ | ||
| 186 | "bswap "tmp" \n\t"\ | ||
| 187 | "shr $15 , "tmp" \n\t"\ | ||
| 188 | "movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"(%%ecx), %%ecx \n\t"\ | ||
| 189 | "sub $0xFFFF , "tmp" \n\t"\ | ||
| 190 | "neg %%ecx \n\t"\ | ||
| 191 | "add $7 , %%ecx \n\t"\ | ||
| 192 | "shl %%cl , "tmp" \n\t"\ | ||
| 193 | "add "tmp" , "low" \n\t" | ||
| 194 | #endif /* __LZCNT__ */ | ||
| 195 | |||
| 196 | #define BRANCHLESS_GET_CABAC(ret, retq, statep, low, lowword, range, rangeq, tmp, tmpbyte, byte, end, norm_off, lps_off, mlps_off, tables) \ | ||
| 197 | "movzbl "statep" , "ret" \n\t"\ | ||
| 198 | "mov "range" , "tmp" \n\t"\ | ||
| 199 | "and $0xC0 , "range" \n\t"\ | ||
| 200 | "movzbl "MANGLE(ff_h264_cabac_tables)"+"lps_off"("ret", "range", 2), "range" \n\t"\ | ||
| 201 | "sub "range" , "tmp" \n\t"\ | ||
| 202 | BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp) \ | ||
| 203 | BRANCHLESS_GET_CABAC_RENORM(range, rangeq, norm_off, tables) \ | ||
| 204 | "shl %%cl , "range" \n\t"\ | ||
| 205 | "movzbl "MANGLE(ff_h264_cabac_tables)"+"mlps_off"+128("ret"), "tmp" \n\t"\ | ||
| 206 | "shl %%cl , "low" \n\t"\ | ||
| 207 | "mov "tmpbyte" , "statep" \n\t"\ | ||
| 208 | "test "lowword" , "lowword" \n\t"\ | ||
| 209 | " jnz 2f \n\t"\ | ||
| 210 | "mov "byte" , %%"FF_REG_c" \n\t"\ | ||
| 211 | END_CHECK(end)\ | ||
| 212 | "add"FF_OPSIZE" $2 , "byte" \n\t"\ | ||
| 213 | "1: \n\t"\ | ||
| 214 | BRANCHLESS_GET_CABAC_REFILL(low, tmp, norm_off, tables) \ | ||
| 215 | "2: \n\t" | ||
| 216 | |||
| 217 | #endif /* BROKEN_RELOCATIONS */ | ||
| 218 | |||
| 219 | #if HAVE_X86_7REGS && !BROKEN_COMPILER | ||
| 220 | #define get_cabac_inline get_cabac_inline_x86 | ||
| 221 | static | ||
| 222 | #if ARCH_X86_32 | ||
| 223 | av_noinline | ||
| 224 | #else | ||
| 225 | av_always_inline | ||
| 226 | #endif | ||
| 227 | 1188752578 | int get_cabac_inline_x86(CABACContext *c, uint8_t *const state) | |
| 228 | { | ||
| 229 | int bit, tmp; | ||
| 230 | #ifdef BROKEN_RELOCATIONS | ||
| 231 | void *tables; | ||
| 232 | |||
| 233 | 1188752578 | __asm__ volatile( | |
| 234 | "lea "MANGLE(ff_h264_cabac_tables)", %0 \n\t" | ||
| 235 | : "=&r"(tables) | ||
| 236 | : NAMED_CONSTRAINTS_ARRAY(ff_h264_cabac_tables) | ||
| 237 | ); | ||
| 238 | #endif | ||
| 239 | |||
| 240 | 1188752578 | __asm__ volatile( | |
| 241 | BRANCHLESS_GET_CABAC("%0", "%q0", "(%4)", "%1", "%w1", | ||
| 242 | "%2", "%q2", "%3", "%b3", | ||
| 243 | "%c6(%5)", "%c7(%5)", | ||
| 244 | AV_STRINGIFY(H264_NORM_SHIFT_OFFSET), | ||
| 245 | AV_STRINGIFY(H264_LPS_RANGE_OFFSET), | ||
| 246 | AV_STRINGIFY(H264_MLPS_STATE_OFFSET), | ||
| 247 | "%8") | ||
| 248 | : "=&r"(bit), "=&r"(c->low), "=&r"(c->range), "=&q"(tmp) | ||
| 249 | : "r"(state), "r"(c), | ||
| 250 | "i"(offsetof(CABACContext, bytestream)), | ||
| 251 | "i"(offsetof(CABACContext, bytestream_end)) | ||
| 252 | TABLES_ARG | ||
| 253 | 1188752578 | ,"1"(c->low), "2"(c->range) | |
| 254 | : "%"FF_REG_c, "memory" | ||
| 255 | ); | ||
| 256 | 1188752578 | return bit & 1; | |
| 257 | } | ||
| 258 | #endif /* HAVE_X86_7REGS && !BROKEN_COMPILER */ | ||
| 259 | |||
| 260 | #if !BROKEN_COMPILER | ||
| 261 | #define get_cabac_bypass_sign get_cabac_bypass_sign_x86 | ||
| 262 | 182011838 | static av_always_inline int get_cabac_bypass_sign_x86(CABACContext *c, int val) | |
| 263 | { | ||
| 264 | x86_reg tmp; | ||
| 265 | 182011838 | __asm__ volatile( | |
| 266 | "movl %c6(%2), %k1 \n\t" | ||
| 267 | "movl %c3(%2), %%eax \n\t" | ||
| 268 | "shl $17, %k1 \n\t" | ||
| 269 | "add %%eax, %%eax \n\t" | ||
| 270 | "sub %k1, %%eax \n\t" | ||
| 271 | "cdq \n\t" | ||
| 272 | "and %%edx, %k1 \n\t" | ||
| 273 | "add %k1, %%eax \n\t" | ||
| 274 | "xor %%edx, %%ecx \n\t" | ||
| 275 | "sub %%edx, %%ecx \n\t" | ||
| 276 | "test %%ax, %%ax \n\t" | ||
| 277 | "jnz 1f \n\t" | ||
| 278 | "mov %c4(%2), %1 \n\t" | ||
| 279 | "subl $0xFFFF, %%eax \n\t" | ||
| 280 | "movzwl (%1), %%edx \n\t" | ||
| 281 | "bswap %%edx \n\t" | ||
| 282 | "shrl $15, %%edx \n\t" | ||
| 283 | #if UNCHECKED_BITSTREAM_READER | ||
| 284 | "add $2, %1 \n\t" | ||
| 285 | "addl %%edx, %%eax \n\t" | ||
| 286 | "mov %1, %c4(%2) \n\t" | ||
| 287 | #else | ||
| 288 | "addl %%edx, %%eax \n\t" | ||
| 289 | "cmp %c5(%2), %1 \n\t" | ||
| 290 | "jge 1f \n\t" | ||
| 291 | "add"FF_OPSIZE" $2, %c4(%2) \n\t" | ||
| 292 | #endif | ||
| 293 | "1: \n\t" | ||
| 294 | "movl %%eax, %c3(%2) \n\t" | ||
| 295 | |||
| 296 | : "+c"(val), "=&r"(tmp) | ||
| 297 | : "r"(c), | ||
| 298 | "i"(offsetof(CABACContext, low)), | ||
| 299 | "i"(offsetof(CABACContext, bytestream)), | ||
| 300 | "i"(offsetof(CABACContext, bytestream_end)), | ||
| 301 | "i"(offsetof(CABACContext, range)) | ||
| 302 | : "%eax", "%edx", "memory" | ||
| 303 | ); | ||
| 304 | 182011838 | return val; | |
| 305 | } | ||
| 306 | |||
| 307 | #define get_cabac_bypass get_cabac_bypass_x86 | ||
| 308 | 164919889 | static av_always_inline int get_cabac_bypass_x86(CABACContext *c) | |
| 309 | { | ||
| 310 | x86_reg tmp; | ||
| 311 | int res; | ||
| 312 | 164919889 | __asm__ volatile( | |
| 313 | "movl %c6(%2), %k1 \n\t" | ||
| 314 | "movl %c3(%2), %%eax \n\t" | ||
| 315 | "shl $17, %k1 \n\t" | ||
| 316 | "add %%eax, %%eax \n\t" | ||
| 317 | "sub %k1, %%eax \n\t" | ||
| 318 | "cdq \n\t" | ||
| 319 | "and %%edx, %k1 \n\t" | ||
| 320 | "add %k1, %%eax \n\t" | ||
| 321 | "inc %%edx \n\t" | ||
| 322 | "test %%ax, %%ax \n\t" | ||
| 323 | "jnz 1f \n\t" | ||
| 324 | "mov %c4(%2), %1 \n\t" | ||
| 325 | "subl $0xFFFF, %%eax \n\t" | ||
| 326 | "movzwl (%1), %%ecx \n\t" | ||
| 327 | "bswap %%ecx \n\t" | ||
| 328 | "shrl $15, %%ecx \n\t" | ||
| 329 | "addl %%ecx, %%eax \n\t" | ||
| 330 | "cmp %c5(%2), %1 \n\t" | ||
| 331 | "jge 1f \n\t" | ||
| 332 | "add"FF_OPSIZE" $2, %c4(%2) \n\t" | ||
| 333 | "1: \n\t" | ||
| 334 | "movl %%eax, %c3(%2) \n\t" | ||
| 335 | |||
| 336 | : "=&d"(res), "=&r"(tmp) | ||
| 337 | : "r"(c), | ||
| 338 | "i"(offsetof(CABACContext, low)), | ||
| 339 | "i"(offsetof(CABACContext, bytestream)), | ||
| 340 | "i"(offsetof(CABACContext, bytestream_end)), | ||
| 341 | "i"(offsetof(CABACContext, range)) | ||
| 342 | : "%eax", "%ecx", "memory" | ||
| 343 | ); | ||
| 344 | 164919889 | return res; | |
| 345 | } | ||
| 346 | #endif /* !BROKEN_COMPILER */ | ||
| 347 | |||
| 348 | #endif /* HAVE_INLINE_ASM */ | ||
| 349 | #endif /* AVCODEC_X86_CABAC_H */ | ||
| 350 |