diff options
Diffstat (limited to 'src/codegen_sse.h')
| -rw-r--r-- | src/codegen_sse.h | 42 |
1 files changed, 12 insertions, 30 deletions
diff --git a/src/codegen_sse.h b/src/codegen_sse.h index ed81d1e..fb4fbfc 100644 --- a/src/codegen_sse.h +++ b/src/codegen_sse.h @@ -314,24 +314,6 @@ static FFTS_INLINE void SHUFPS(uint8_t **p, uint8_t reg2, uint8_t reg1, const in *(*p)++ = (select & 0xFF); } -static FFTS_INLINE void SUBPS(uint8_t **p, uint8_t reg2, uint8_t reg1) -{ - uint8_t r1 = (reg1 & 7); - uint8_t r2 = (reg2 & 7); - - /* REX prefix */ - if ((reg1 & 8) || (reg2 & 8)) { - *(*p)++ = 0x40 | ((reg1 & 8) >> 3) | ((reg2 & 8) >> 1); - } - - /* esacape opcode */ - *(*p)++ = 0x0F; - - /* opcode */ - *(*p)++ = 0x5C; - *(*p)++ = 0xC0 | r1 | (r2 << 3); -} - static FFTS_INLINE void ffts_insert_nops(uint8_t **p, uint32_t count) { if (count >= 9) { @@ -629,7 +611,7 @@ static FFTS_INLINE insns_t* generate_size8_base_case(insns_t **fp, int sign) SHUFPS(fp, XMM6, XMM6, 0xB1); MULPS(fp, XMM6, XMM8); SHUFPS(fp, XMM7, XMM7, 0xB1); - SUBPS(fp, XMM11, XMM6); + x64_sse_subps_reg_reg(*fp, X64_XMM11, X64_XMM6); MULPS(fp, XMM8, XMM7); /* movaps xmm10, xmm11 */ @@ -649,7 +631,7 @@ static FFTS_INLINE insns_t* generate_size8_base_case(insns_t **fp, int sign) *(*fp)++ = 0x20; x64_sse_addps_reg_reg(*fp, X64_XMM10, X64_XMM9); - SUBPS(fp, XMM11, XMM9); + x64_sse_subps_reg_reg(*fp, X64_XMM11, X64_XMM9); /* movaps xmm5, [rcx] */ /* output + 0 * output_stride */ @@ -694,7 +676,7 @@ static FFTS_INLINE insns_t* generate_size8_base_case(insns_t **fp, int sign) *(*fp)++ = 0x70; *(*fp)++ = 0x30; - SUBPS(fp, XMM2, XMM10); + x64_sse_subps_reg_reg(*fp, X64_XMM2, X64_XMM10); MULPS(fp, XMM6, XMM12); x64_sse_addps_reg_reg(*fp, X64_XMM5, X64_XMM10); @@ -716,7 +698,7 @@ static FFTS_INLINE insns_t* generate_size8_base_case(insns_t **fp, int sign) SHUFPS(fp, XMM13, XMM13, 0xB1); MULPS(fp, XMM12, XMM14); MULPS(fp, XMM14, XMM13); - SUBPS(fp, XMM6, XMM12); + x64_sse_subps_reg_reg(*fp, X64_XMM6, X64_XMM12); x64_sse_addps_reg_reg(*fp, X64_XMM15, X64_XMM14); /* movaps xmm7, [rcx + r10] */ @@ -756,10 +738,10 @@ static FFTS_INLINE insns_t* generate_size8_base_case(insns_t **fp, int sign) x64_alu_reg_imm_size(*fp, X86_ADD, X64_RAX, 0x60, 8); MULPS(fp, XMM13, XMM7); - SUBPS(fp, XMM6, XMM15); + x64_sse_subps_reg_reg(*fp, X64_XMM6, X64_XMM15); x64_sse_addps_reg_reg(*fp, X64_XMM12, X64_XMM15); MULPS(fp, XMM10, XMM8); - SUBPS(fp, XMM0, XMM12); + x64_sse_subps_reg_reg(*fp, X64_XMM0, X64_XMM12); x64_sse_addps_reg_reg(*fp, X64_XMM5, X64_XMM12); SHUFPS(fp, XMM7, XMM7, 0xB1); x64_sse_xorps_reg_reg(*fp, X64_XMM6, X64_XMM3); @@ -774,7 +756,7 @@ static FFTS_INLINE insns_t* generate_size8_base_case(insns_t **fp, int sign) MULPS(fp, XMM7, XMM9); MULPS(fp, XMM9, XMM8); - SUBPS(fp, XMM13, XMM7); + x64_sse_subps_reg_reg(*fp, X64_XMM13, X64_XMM7); x64_sse_addps_reg_reg(*fp, X64_XMM10, X64_XMM9); /* movaps xmm4, [rcx + rbx] */ @@ -793,9 +775,9 @@ static FFTS_INLINE insns_t* generate_size8_base_case(insns_t **fp, int sign) SHUFPS(fp, XMM6, XMM6, 0xB1); x64_sse_addps_reg_reg(*fp, X64_XMM1, X64_XMM11); - SUBPS(fp, XMM4, XMM11); + x64_sse_subps_reg_reg(*fp, X64_XMM4, X64_XMM11); x64_sse_addps_reg_reg(*fp, X64_XMM12, X64_XMM6); - SUBPS(fp, XMM2, XMM6); + x64_sse_subps_reg_reg(*fp, X64_XMM2, X64_XMM6); /* movaps xmm11, xmm13 */ *(*fp)++ = 0x45; @@ -814,11 +796,11 @@ static FFTS_INLINE insns_t* generate_size8_base_case(insns_t **fp, int sign) *(*fp)++ = 0x28; *(*fp)++ = 0xF1; - SUBPS(fp, XMM13, XMM10); + x64_sse_subps_reg_reg(*fp, X64_XMM13, X64_XMM10); x64_sse_addps_reg_reg(*fp, X64_XMM11, X64_XMM10); x64_sse_xorps_reg_reg(*fp, X64_XMM13, X64_XMM3); x64_sse_addps_reg_reg(*fp, X64_XMM4, X64_XMM11); - SUBPS(fp, XMM14, XMM11); + x64_sse_subps_reg_reg(*fp, X64_XMM14, X64_XMM11); SHUFPS(fp, XMM13, XMM13, 0xB1); /* movaps [rcx], xmm5 */ @@ -841,7 +823,7 @@ static FFTS_INLINE insns_t* generate_size8_base_case(insns_t **fp, int sign) *(*fp)++ = 0x14; *(*fp)++ = 0x59; - SUBPS(fp, XMM1, XMM13); + x64_sse_subps_reg_reg(*fp, X64_XMM1, X64_XMM13); x64_sse_addps_reg_reg(*fp, X64_XMM6, X64_XMM13); /* movaps [rcx + rsi], xmm1 */ |
