yading@10: /* yading@10: * Copyright (c) 2003 Michael Niedermayer yading@10: * yading@10: * This file is part of Libav. yading@10: * yading@10: * Libav is free software; you can redistribute it and/or yading@10: * modify it under the terms of the GNU Lesser General Public yading@10: * License as published by the Free Software Foundation; either yading@10: * version 2.1 of the License, or (at your option) any later version. yading@10: * yading@10: * Libav is distributed in the hope that it will be useful, yading@10: * but WITHOUT ANY WARRANTY; without even the implied warranty of yading@10: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU yading@10: * Lesser General Public License for more details. yading@10: * yading@10: * You should have received a copy of the GNU Lesser General Public yading@10: * License along with Libav; if not, write to the Free Software yading@10: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA yading@10: */ yading@10: yading@10: #ifndef AVCODEC_X86_CABAC_H yading@10: #define AVCODEC_X86_CABAC_H yading@10: yading@10: #include "libavcodec/cabac.h" yading@10: #include "libavutil/attributes.h" yading@10: #include "libavutil/x86/asm.h" yading@10: #include "libavutil/internal.h" yading@10: #include "config.h" yading@10: yading@10: #if HAVE_INLINE_ASM yading@10: yading@10: #ifdef BROKEN_RELOCATIONS yading@10: #define TABLES_ARG , "r"(tables) yading@10: yading@10: #if HAVE_FAST_CMOV yading@10: #define BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \ yading@10: "cmp "low" , "tmp" \n\t"\ yading@10: "cmova %%ecx , "range" \n\t"\ yading@10: "sbb %%rcx , %%rcx \n\t"\ yading@10: "and %%ecx , "tmp" \n\t"\ yading@10: "xor %%rcx , "retq" \n\t"\ yading@10: "sub "tmp" , "low" \n\t" yading@10: #else /* HAVE_FAST_CMOV */ yading@10: #define BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \ yading@10: /* P4 Prescott has crappy cmov,sbb,64bit shift so avoid them */ \ yading@10: "sub "low" , "tmp" \n\t"\ yading@10: "sar $31 , "tmp" \n\t"\ yading@10: "sub %%ecx , "range" \n\t"\ yading@10: "and "tmp" , "range" \n\t"\ yading@10: "add %%ecx , "range" \n\t"\ yading@10: "shl $17 , %%ecx \n\t"\ yading@10: "and "tmp" , %%ecx \n\t"\ yading@10: "sub %%ecx , "low" \n\t"\ yading@10: "xor "tmp" , "ret" \n\t"\ yading@10: "movslq "ret" , "retq" \n\t" yading@10: #endif /* HAVE_FAST_CMOV */ yading@10: yading@10: #define BRANCHLESS_GET_CABAC(ret, retq, statep, low, lowword, range, rangeq, tmp, tmpbyte, byte, end, norm_off, lps_off, mlps_off, tables) \ yading@10: "movzbl "statep" , "ret" \n\t"\ yading@10: "mov "range" , "tmp" \n\t"\ yading@10: "and $0xC0 , "range" \n\t"\ yading@10: "lea ("ret", "range", 2), %%ecx \n\t"\ yading@10: "movzbl "lps_off"("tables", %%rcx), "range" \n\t"\ yading@10: "sub "range" , "tmp" \n\t"\ yading@10: "mov "tmp" , %%ecx \n\t"\ yading@10: "shl $17 , "tmp" \n\t"\ yading@10: BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \ yading@10: "movzbl "norm_off"("tables", "rangeq"), %%ecx \n\t"\ yading@10: "shl %%cl , "range" \n\t"\ yading@10: "movzbl "mlps_off"+128("tables", "retq"), "tmp" \n\t"\ yading@10: "shl %%cl , "low" \n\t"\ yading@10: "mov "tmpbyte" , "statep" \n\t"\ yading@10: "test "lowword" , "lowword" \n\t"\ yading@10: "jnz 2f \n\t"\ yading@10: "mov "byte" , %%"REG_c" \n\t"\ yading@10: "add"OPSIZE" $2 , "byte" \n\t"\ yading@10: "movzwl (%%"REG_c") , "tmp" \n\t"\ yading@10: "lea -1("low") , %%ecx \n\t"\ yading@10: "xor "low" , %%ecx \n\t"\ yading@10: "shr $15 , %%ecx \n\t"\ yading@10: "bswap "tmp" \n\t"\ yading@10: "shr $15 , "tmp" \n\t"\ yading@10: "movzbl "norm_off"("tables", %%rcx), %%ecx \n\t"\ yading@10: "sub $0xFFFF , "tmp" \n\t"\ yading@10: "neg %%ecx \n\t"\ yading@10: "add $7 , %%ecx \n\t"\ yading@10: "shl %%cl , "tmp" \n\t"\ yading@10: "add "tmp" , "low" \n\t"\ yading@10: "2: \n\t" yading@10: yading@10: #else /* BROKEN_RELOCATIONS */ yading@10: #define TABLES_ARG yading@10: #define RIP_ARG yading@10: yading@10: #if HAVE_FAST_CMOV yading@10: #define BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)\ yading@10: "mov "tmp" , %%ecx \n\t"\ yading@10: "shl $17 , "tmp" \n\t"\ yading@10: "cmp "low" , "tmp" \n\t"\ yading@10: "cmova %%ecx , "range" \n\t"\ yading@10: "sbb %%ecx , %%ecx \n\t"\ yading@10: "and %%ecx , "tmp" \n\t"\ yading@10: "xor %%ecx , "ret" \n\t"\ yading@10: "sub "tmp" , "low" \n\t" yading@10: #else /* HAVE_FAST_CMOV */ yading@10: #define BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)\ yading@10: "mov "tmp" , %%ecx \n\t"\ yading@10: "shl $17 , "tmp" \n\t"\ yading@10: "sub "low" , "tmp" \n\t"\ yading@10: "sar $31 , "tmp" \n\t" /*lps_mask*/\ yading@10: "sub %%ecx , "range" \n\t" /*RangeLPS - range*/\ yading@10: "and "tmp" , "range" \n\t" /*(RangeLPS - range)&lps_mask*/\ yading@10: "add %%ecx , "range" \n\t" /*new range*/\ yading@10: "shl $17 , %%ecx \n\t"\ yading@10: "and "tmp" , %%ecx \n\t"\ yading@10: "sub %%ecx , "low" \n\t"\ yading@10: "xor "tmp" , "ret" \n\t" yading@10: #endif /* HAVE_FAST_CMOV */ yading@10: yading@10: #define BRANCHLESS_GET_CABAC(ret, retq, statep, low, lowword, range, rangeq, tmp, tmpbyte, byte, end, norm_off, lps_off, mlps_off, tables) \ yading@10: "movzbl "statep" , "ret" \n\t"\ yading@10: "mov "range" , "tmp" \n\t"\ yading@10: "and $0xC0 , "range" \n\t"\ yading@10: "movzbl "MANGLE(ff_h264_cabac_tables)"+"lps_off"("ret", "range", 2), "range" \n\t"\ yading@10: "sub "range" , "tmp" \n\t"\ yading@10: BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp) \ yading@10: "movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"("range"), %%ecx \n\t"\ yading@10: "shl %%cl , "range" \n\t"\ yading@10: "movzbl "MANGLE(ff_h264_cabac_tables)"+"mlps_off"+128("ret"), "tmp" \n\t"\ yading@10: "shl %%cl , "low" \n\t"\ yading@10: "mov "tmpbyte" , "statep" \n\t"\ yading@10: "test "lowword" , "lowword" \n\t"\ yading@10: " jnz 2f \n\t"\ yading@10: "mov "byte" , %%"REG_c" \n\t"\ yading@10: "add"OPSIZE" $2 , "byte" \n\t"\ yading@10: "movzwl (%%"REG_c") , "tmp" \n\t"\ yading@10: "lea -1("low") , %%ecx \n\t"\ yading@10: "xor "low" , %%ecx \n\t"\ yading@10: "shr $15 , %%ecx \n\t"\ yading@10: "bswap "tmp" \n\t"\ yading@10: "shr $15 , "tmp" \n\t"\ yading@10: "movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"(%%ecx), %%ecx \n\t"\ yading@10: "sub $0xFFFF , "tmp" \n\t"\ yading@10: "neg %%ecx \n\t"\ yading@10: "add $7 , %%ecx \n\t"\ yading@10: "shl %%cl , "tmp" \n\t"\ yading@10: "add "tmp" , "low" \n\t"\ yading@10: "2: \n\t" yading@10: yading@10: #endif /* BROKEN_RELOCATIONS */ yading@10: yading@10: yading@10: #if HAVE_7REGS && !(defined(__i386) && defined(__clang__) && (__clang_major__<2 || (__clang_major__==2 && __clang_minor__<10)))\ yading@10: && !( !defined(__clang__) && defined(__llvm__) && __GNUC__==4 && __GNUC_MINOR__==2 && __GNUC_PATCHLEVEL__<=1) yading@10: #define get_cabac_inline get_cabac_inline_x86 yading@10: static av_always_inline int get_cabac_inline_x86(CABACContext *c, yading@10: uint8_t *const state) yading@10: { yading@10: int bit, tmp; yading@10: #ifdef BROKEN_RELOCATIONS yading@10: void *tables; yading@10: yading@10: __asm__ volatile( yading@10: "lea "MANGLE(ff_h264_cabac_tables)", %0 \n\t" yading@10: : "=&r"(tables) yading@10: ); yading@10: #endif yading@10: yading@10: __asm__ volatile( yading@10: BRANCHLESS_GET_CABAC("%0", "%q0", "(%4)", "%1", "%w1", yading@10: "%2", "%q2", "%3", "%b3", yading@10: "%c6(%5)", "%c7(%5)", yading@10: AV_STRINGIFY(H264_NORM_SHIFT_OFFSET), yading@10: AV_STRINGIFY(H264_LPS_RANGE_OFFSET), yading@10: AV_STRINGIFY(H264_MLPS_STATE_OFFSET), yading@10: "%8") yading@10: : "=&r"(bit), "=&r"(c->low), "=&r"(c->range), "=&q"(tmp) yading@10: : "r"(state), "r"(c), yading@10: "i"(offsetof(CABACContext, bytestream)), yading@10: "i"(offsetof(CABACContext, bytestream_end)) yading@10: TABLES_ARG yading@10: ,"1"(c->low), "2"(c->range) yading@10: : "%"REG_c, "memory" yading@10: ); yading@10: return bit & 1; yading@10: } yading@10: #endif /* HAVE_7REGS */ yading@10: yading@10: #define get_cabac_bypass_sign get_cabac_bypass_sign_x86 yading@10: static av_always_inline int get_cabac_bypass_sign_x86(CABACContext *c, int val) yading@10: { yading@10: x86_reg tmp; yading@10: __asm__ volatile( yading@10: "movl %c6(%2), %k1 \n\t" yading@10: "movl %c3(%2), %%eax \n\t" yading@10: "shl $17, %k1 \n\t" yading@10: "add %%eax, %%eax \n\t" yading@10: "sub %k1, %%eax \n\t" yading@10: "cltd \n\t" yading@10: "and %%edx, %k1 \n\t" yading@10: "add %k1, %%eax \n\t" yading@10: "xor %%edx, %%ecx \n\t" yading@10: "sub %%edx, %%ecx \n\t" yading@10: "test %%ax, %%ax \n\t" yading@10: "jnz 1f \n\t" yading@10: "mov %c4(%2), %1 \n\t" yading@10: "subl $0xFFFF, %%eax \n\t" yading@10: "movzwl (%1), %%edx \n\t" yading@10: "bswap %%edx \n\t" yading@10: "shrl $15, %%edx \n\t" yading@10: "add $2, %1 \n\t" yading@10: "addl %%edx, %%eax \n\t" yading@10: "mov %1, %c4(%2) \n\t" yading@10: "1: \n\t" yading@10: "movl %%eax, %c3(%2) \n\t" yading@10: yading@10: : "+c"(val), "=&r"(tmp) yading@10: : "r"(c), yading@10: "i"(offsetof(CABACContext, low)), yading@10: "i"(offsetof(CABACContext, bytestream)), yading@10: "i"(offsetof(CABACContext, bytestream_end)), yading@10: "i"(offsetof(CABACContext, range)) yading@10: : "%eax", "%edx", "memory" yading@10: ); yading@10: return val; yading@10: } yading@10: yading@10: #endif /* HAVE_INLINE_ASM */ yading@10: #endif /* AVCODEC_X86_CABAC_H */