From 018c33838eeb41944af46287e7d8be82c5c427d8 Mon Sep 17 00:00:00 2001 From: Mans Rullgard Date: Mon, 20 Jun 2011 02:31:53 +0100 Subject: [PATCH] x86: cabac: remove hardcoded ebx in inline asm Signed-off-by: Mans Rullgard --- libavcodec/cabac.h | 53 ++++++++++++++++++----------------- libavcodec/x86/h264_i386.h | 70 ++++++++++++++++++++++++---------------------- 2 files changed, 63 insertions(+), 60 deletions(-) diff --git a/libavcodec/cabac.h b/libavcodec/cabac.h index 67a332eba..b868f77f7 100644 --- a/libavcodec/cabac.h +++ b/libavcodec/cabac.h @@ -270,7 +270,7 @@ static void refill(CABACContext *c){ c->bytestream+= CABAC_BITS/8; } -#if ! ( ARCH_X86 && HAVE_7REGS && HAVE_EBX_AVAILABLE && !defined(BROKEN_RELOCATIONS) ) +#if ! ( ARCH_X86 && HAVE_7REGS && !defined(BROKEN_RELOCATIONS) ) static void refill2(CABACContext *c){ int i, x; @@ -309,8 +309,8 @@ static inline void renorm_cabac_decoder_once(CABACContext *c){ static av_always_inline int get_cabac_inline(CABACContext *c, uint8_t * const state){ //FIXME gcc generates duplicate load/stores for c->low and c->range -#if ARCH_X86 && HAVE_7REGS && HAVE_EBX_AVAILABLE && !defined(BROKEN_RELOCATIONS) - int bit; +#if ARCH_X86 && HAVE_7REGS && !defined(BROKEN_RELOCATIONS) + int bit, low; #if HAVE_FAST_CMOV #define BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\ @@ -370,20 +370,20 @@ static av_always_inline int get_cabac_inline(CABACContext *c, uint8_t * const st "1: \n\t" __asm__ volatile( - "movl %a3(%2), %%esi \n\t" - "movl %a4(%2), %%ebx \n\t" - BRANCHLESS_GET_CABAC("%0", "%2", "(%1)", "%%ebx", "%%bx", "%%esi", "%%edx", "%%dl", "%a5") - "movl %%esi, %a3(%2) \n\t" - "movl %%ebx, %a4(%2) \n\t" + "movl %a4(%3), %%esi \n\t" + "movl %a5(%3), %1 \n\t" + BRANCHLESS_GET_CABAC("%0", "%3", "(%2)", "%1", "%w1", "%%esi", "%%edx", "%%dl", "%a6") + "movl %%esi, %a4(%3) \n\t" + "movl %1, %a5(%3) \n\t" - :"=&a"(bit) + :"=&a"(bit), "=&r"(low) :"r"(state), "r"(c), "i"(offsetof(CABACContext, range)), "i"(offsetof(CABACContext, low)), "i"(offsetof(CABACContext, bytestream)) - : "%"REG_c, "%ebx", "%edx", "%esi", "memory" + : "%"REG_c, "%edx", "%esi", "memory" ); bit&=1; -#else /* ARCH_X86 && HAVE_7REGS && HAVE_EBX_AVAILABLE && !defined(BROKEN_RELOCATIONS) */ +#else /* ARCH_X86 && HAVE_7REGS && !defined(BROKEN_RELOCATIONS) */ int s = *state; int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + s]; int bit, lps_mask; @@ -403,7 +403,7 @@ static av_always_inline int get_cabac_inline(CABACContext *c, uint8_t * const st c->low <<= lps_mask; if(!(c->low & CABAC_MASK)) refill2(c); -#endif /* ARCH_X86 && HAVE_7REGS && HAVE_EBX_AVAILABLE && !defined(BROKEN_RELOCATIONS) */ +#endif /* ARCH_X86 && HAVE_7REGS && !defined(BROKEN_RELOCATIONS) */ return bit; } @@ -433,36 +433,37 @@ static int av_unused get_cabac_bypass(CABACContext *c){ static av_always_inline int get_cabac_bypass_sign(CABACContext *c, int val){ -#if ARCH_X86 && HAVE_EBX_AVAILABLE +#if ARCH_X86 + x86_reg tmp; __asm__ volatile( - "movl %a2(%1), %%ebx \n\t" - "movl %a3(%1), %%eax \n\t" - "shl $17, %%ebx \n\t" + "movl %a3(%2), %k1 \n\t" + "movl %a4(%2), %%eax \n\t" + "shl $17, %k1 \n\t" "add %%eax, %%eax \n\t" - "sub %%ebx, %%eax \n\t" + "sub %k1, %%eax \n\t" "cltd \n\t" - "and %%edx, %%ebx \n\t" - "add %%ebx, %%eax \n\t" + "and %%edx, %k1 \n\t" + "add %k1, %%eax \n\t" "xor %%edx, %%ecx \n\t" "sub %%edx, %%ecx \n\t" "test %%ax, %%ax \n\t" " jnz 1f \n\t" - "mov %a4(%1), %%"REG_b" \n\t" + "mov %a5(%2), %1 \n\t" "subl $0xFFFF, %%eax \n\t" - "movzwl (%%"REG_b"), %%edx \n\t" + "movzwl (%1), %%edx \n\t" "bswap %%edx \n\t" "shrl $15, %%edx \n\t" - "add $2, %%"REG_b" \n\t" + "add $2, %1 \n\t" "addl %%edx, %%eax \n\t" - "mov %%"REG_b", %a4(%1) \n\t" + "mov %1, %a5(%2) \n\t" "1: \n\t" - "movl %%eax, %a3(%1) \n\t" + "movl %%eax, %a4(%2) \n\t" - :"+c"(val) + :"+c"(val), "=&r"(tmp) :"r"(c), "i"(offsetof(CABACContext, range)), "i"(offsetof(CABACContext, low)), "i"(offsetof(CABACContext, bytestream)) - : "%eax", "%"REG_b, "%edx", "memory" + : "%eax", "%edx", "memory" ); return val; #else diff --git a/libavcodec/x86/h264_i386.h b/libavcodec/x86/h264_i386.h index b6c225a2f..ba3a5ddfa 100644 --- a/libavcodec/x86/h264_i386.h +++ b/libavcodec/x86/h264_i386.h @@ -35,7 +35,7 @@ //FIXME use some macros to avoid duplicating get_cabac (cannot be done yet //as that would make optimization work hard) -#if ARCH_X86 && HAVE_7REGS && HAVE_EBX_AVAILABLE && !defined(BROKEN_RELOCATIONS) +#if ARCH_X86 && HAVE_7REGS && !defined(BROKEN_RELOCATIONS) static int decode_significance_x86(CABACContext *c, int max_coeff, uint8_t *significant_coeff_ctx_base, int *index, x86_reg last_off){ @@ -43,25 +43,26 @@ static int decode_significance_x86(CABACContext *c, int max_coeff, int minusstart= -(int)significant_coeff_ctx_base; int minusindex= 4-(int)index; int coeff_count; + int low; __asm__ volatile( - "movl %a8(%3), %%esi \n\t" - "movl %a9(%3), %%ebx \n\t" + "movl %a9(%4), %%esi \n\t" + "movl %a10(%4), %3 \n\t" "2: \n\t" - BRANCHLESS_GET_CABAC("%%edx", "%3", "(%1)", "%%ebx", - "%%bx", "%%esi", "%%eax", "%%al", "%a10") + BRANCHLESS_GET_CABAC("%%edx", "%4", "(%1)", "%3", + "%w3", "%%esi", "%%eax", "%%al", "%a11") "test $1, %%edx \n\t" " jz 3f \n\t" - "add %7, %1 \n\t" + "add %8, %1 \n\t" - BRANCHLESS_GET_CABAC("%%edx", "%3", "(%1)", "%%ebx", - "%%bx", "%%esi", "%%eax", "%%al", "%a10") + BRANCHLESS_GET_CABAC("%%edx", "%4", "(%1)", "%3", + "%w3", "%%esi", "%%eax", "%%al", "%a11") - "sub %7, %1 \n\t" + "sub %8, %1 \n\t" "mov %2, %%"REG_a" \n\t" - "movl %4, %%ecx \n\t" + "movl %5, %%ecx \n\t" "add %1, %%"REG_c" \n\t" "movl %%ecx, (%%"REG_a") \n\t" @@ -73,23 +74,24 @@ static int decode_significance_x86(CABACContext *c, int max_coeff, "3: \n\t" "add $1, %1 \n\t" - "cmp %5, %1 \n\t" + "cmp %6, %1 \n\t" " jb 2b \n\t" "mov %2, %%"REG_a" \n\t" - "movl %4, %%ecx \n\t" + "movl %5, %%ecx \n\t" "add %1, %%"REG_c" \n\t" "movl %%ecx, (%%"REG_a") \n\t" "4: \n\t" - "add %6, %%eax \n\t" + "add %7, %%eax \n\t" "shr $2, %%eax \n\t" - "movl %%esi, %a8(%3) \n\t" - "movl %%ebx, %a9(%3) \n\t" - :"=&a"(coeff_count), "+r"(significant_coeff_ctx_base), "+m"(index) + "movl %%esi, %a9(%4) \n\t" + "movl %3, %a10(%4) \n\t" + :"=&a"(coeff_count), "+r"(significant_coeff_ctx_base), "+m"(index), + "=&r"(low) :"r"(c), "m"(minusstart), "m"(end), "m"(minusindex), "m"(last_off), "i"(offsetof(CABACContext, range)), "i"(offsetof(CABACContext, low)), "i"(offsetof(CABACContext, bytestream)) - : "%"REG_c, "%ebx", "%edx", "%esi", "memory" + : "%"REG_c, "%edx", "%esi", "memory" ); return coeff_count; } @@ -99,31 +101,32 @@ static int decode_significance_8x8_x86(CABACContext *c, int *index, x86_reg last_off, const uint8_t *sig_off){ int minusindex= 4-(int)index; int coeff_count; + int low; x86_reg last=0; __asm__ volatile( - "movl %a8(%3), %%esi \n\t" - "movl %a9(%3), %%ebx \n\t" + "movl %a9(%4), %%esi \n\t" + "movl %a10(%4), %3 \n\t" "mov %1, %%"REG_D" \n\t" "2: \n\t" - "mov %6, %%"REG_a" \n\t" + "mov %7, %%"REG_a" \n\t" "movzbl (%%"REG_a", %%"REG_D"), %%edi \n\t" - "add %5, %%"REG_D" \n\t" + "add %6, %%"REG_D" \n\t" - BRANCHLESS_GET_CABAC("%%edx", "%3", "(%%"REG_D")", "%%ebx", - "%%bx", "%%esi", "%%eax", "%%al", "%a10") + BRANCHLESS_GET_CABAC("%%edx", "%4", "(%%"REG_D")", "%3", + "%w3", "%%esi", "%%eax", "%%al", "%a11") "mov %1, %%edi \n\t" "test $1, %%edx \n\t" " jz 3f \n\t" "movzbl "MANGLE(last_coeff_flag_offset_8x8)"(%%edi), %%edi\n\t" - "add %5, %%"REG_D" \n\t" - "add %7, %%"REG_D" \n\t" + "add %6, %%"REG_D" \n\t" + "add %8, %%"REG_D" \n\t" - BRANCHLESS_GET_CABAC("%%edx", "%3", "(%%"REG_D")", "%%ebx", - "%%bx", "%%esi", "%%eax", "%%al", "%a10") + BRANCHLESS_GET_CABAC("%%edx", "%4", "(%%"REG_D")", "%3", + "%w3", "%%esi", "%%eax", "%%al", "%a11") "mov %2, %%"REG_a" \n\t" "mov %1, %%edi \n\t" @@ -143,20 +146,19 @@ static int decode_significance_8x8_x86(CABACContext *c, "mov %2, %%"REG_a" \n\t" "movl %%edi, (%%"REG_a") \n\t" "4: \n\t" - "addl %4, %%eax \n\t" + "addl %5, %%eax \n\t" "shr $2, %%eax \n\t" - "movl %%esi, %a8(%3) \n\t" - "movl %%ebx, %a9(%3) \n\t" - :"=&a"(coeff_count),"+m"(last), "+m"(index) + "movl %%esi, %a9(%4) \n\t" + "movl %3, %a10(%4) \n\t" + :"=&a"(coeff_count),"+m"(last), "+m"(index), "=&r"(low) :"r"(c), "m"(minusindex), "m"(significant_coeff_ctx_base), "m"(sig_off), "m"(last_off), "i"(offsetof(CABACContext, range)), "i"(offsetof(CABACContext, low)), "i"(offsetof(CABACContext, bytestream)) - : "%"REG_c, "%ebx", "%edx", "%esi", "%"REG_D, "memory" + : "%"REG_c, "%edx", "%esi", "%"REG_D, "memory" ); return coeff_count; } -#endif /* ARCH_X86 && HAVE_7REGS && HAVE_EBX_AVAILABLE */ - /* !defined(BROKEN_RELOCATIONS) */ +#endif /* ARCH_X86 && HAVE_7REGS && !defined(BROKEN_RELOCATIONS) */ #endif /* AVCODEC_X86_H264_I386_H */ -- 2.11.0