mirrors/dynarmic
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

emit_x64_vector_floating_point: Prefer blendvp{s,d} to vblendvp{s,d} where possible

Browse source 
It's a cheaper instruction.
This commit is contained in:
MerryMage 2018-07-31 21:03:48 +01:00
parent 476c0f15da
commit dfb660cd16
1 changed files with 10 additions and 10 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

20
src/backend_x64/emit_x64_vector_floating_point.cpp
View file

@ -153,10 +153,10 @@ Xbyak::Address GetSmallestNormalVector(BlockOfCode& code) {
template<size_t fsize> template<size_t fsize>
void ForceToDefaultNaN(BlockOfCode& code, EmitContext& ctx, Xbyak::Xmm result) { void ForceToDefaultNaN(BlockOfCode& code, EmitContext& ctx, Xbyak::Xmm result) {
if (ctx.FPSCR_DN()) { if (ctx.FPSCR_DN()) {
const Xbyak::Xmm nan_mask = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm nan_mask = xmm0;
if (code.DoesCpuSupport(Xbyak::util::Cpu::tAVX)) { if (code.DoesCpuSupport(Xbyak::util::Cpu::tAVX)) {
FCODE(vcmpunordp)(nan_mask, result, result); FCODE(vcmpunordp)(nan_mask, result, result);
FCODE(vblendvp)(result, result, GetNaNVector<fsize>(code), nan_mask); FCODE(blendvp)(result, GetNaNVector<fsize>(code));
} else { } else {
code.movaps(nan_mask, result); code.movaps(nan_mask, result);
FCODE(cmpordp)(nan_mask, nan_mask); FCODE(cmpordp)(nan_mask, nan_mask);
@ -572,7 +572,7 @@ static void EmitFPVectorMax(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst)
const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]); const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
const Xbyak::Xmm xmm_b = ctx.FPSCR_FTZ() ? ctx.reg_alloc.UseScratchXmm(args[1]) : ctx.reg_alloc.UseXmm(args[1]); const Xbyak::Xmm xmm_b = ctx.FPSCR_FTZ() ? ctx.reg_alloc.UseScratchXmm(args[1]) : ctx.reg_alloc.UseXmm(args[1]);
const Xbyak::Xmm mask = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm mask = xmm0;
const Xbyak::Xmm anded = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm anded = ctx.reg_alloc.ScratchXmm();
const Xbyak::Xmm nan_mask = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm nan_mask = ctx.reg_alloc.ScratchXmm();
@ -583,7 +583,7 @@ static void EmitFPVectorMax(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst)
FCODE(vcmpunordp)(nan_mask, result, xmm_b); FCODE(vcmpunordp)(nan_mask, result, xmm_b);
FCODE(vandp)(anded, result, xmm_b); FCODE(vandp)(anded, result, xmm_b);
FCODE(vmaxp)(result, result, xmm_b); FCODE(vmaxp)(result, result, xmm_b);
FCODE(vblendvp)(result, result, anded, mask); FCODE(blendvp)(result, anded);
FCODE(vblendvp)(result, result, GetNaNVector<fsize>(code), nan_mask); FCODE(vblendvp)(result, result, GetNaNVector<fsize>(code), nan_mask);
} else { } else {
code.movaps(mask, result); code.movaps(mask, result);
@ -610,7 +610,7 @@ static void EmitFPVectorMax(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst)
} }
EmitThreeOpVectorOperation<fsize, DefaultIndexer>(code, ctx, inst, [&](const Xbyak::Xmm& result, Xbyak::Xmm xmm_b){ EmitThreeOpVectorOperation<fsize, DefaultIndexer>(code, ctx, inst, [&](const Xbyak::Xmm& result, Xbyak::Xmm xmm_b){
const Xbyak::Xmm mask = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm mask = xmm0;
const Xbyak::Xmm anded = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm anded = ctx.reg_alloc.ScratchXmm();
if (ctx.FPSCR_FTZ()) { if (ctx.FPSCR_FTZ()) {
@ -628,7 +628,7 @@ static void EmitFPVectorMax(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst)
FCODE(vcmpeqp)(mask, result, xmm_b); FCODE(vcmpeqp)(mask, result, xmm_b);
FCODE(vandp)(anded, result, xmm_b); FCODE(vandp)(anded, result, xmm_b);
FCODE(vmaxp)(result, result, xmm_b); FCODE(vmaxp)(result, result, xmm_b);
FCODE(vblendvp)(result, result, anded, mask); FCODE(blendvp)(result, anded);
} else { } else {
code.movaps(mask, result); code.movaps(mask, result);
code.movaps(anded, result); code.movaps(anded, result);
@ -659,7 +659,7 @@ static void EmitFPVectorMin(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst)
const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]); const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
const Xbyak::Xmm xmm_b = ctx.FPSCR_FTZ() ? ctx.reg_alloc.UseScratchXmm(args[1]) : ctx.reg_alloc.UseXmm(args[1]); const Xbyak::Xmm xmm_b = ctx.FPSCR_FTZ() ? ctx.reg_alloc.UseScratchXmm(args[1]) : ctx.reg_alloc.UseXmm(args[1]);
const Xbyak::Xmm mask = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm mask = xmm0;
const Xbyak::Xmm ored = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm ored = ctx.reg_alloc.ScratchXmm();
const Xbyak::Xmm nan_mask = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm nan_mask = ctx.reg_alloc.ScratchXmm();
@ -670,7 +670,7 @@ static void EmitFPVectorMin(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst)
FCODE(vcmpunordp)(nan_mask, result, xmm_b); FCODE(vcmpunordp)(nan_mask, result, xmm_b);
FCODE(vorp)(ored, result, xmm_b); FCODE(vorp)(ored, result, xmm_b);
FCODE(vminp)(result, result, xmm_b); FCODE(vminp)(result, result, xmm_b);
FCODE(vblendvp)(result, result, ored, mask); FCODE(blendvp)(result, ored);
FCODE(vblendvp)(result, result, GetNaNVector<fsize>(code), nan_mask); FCODE(vblendvp)(result, result, GetNaNVector<fsize>(code), nan_mask);
} else { } else {
code.movaps(mask, result); code.movaps(mask, result);
@ -697,7 +697,7 @@ static void EmitFPVectorMin(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst)
} }
EmitThreeOpVectorOperation<fsize, DefaultIndexer>(code, ctx, inst, [&](const Xbyak::Xmm& result, Xbyak::Xmm xmm_b){ EmitThreeOpVectorOperation<fsize, DefaultIndexer>(code, ctx, inst, [&](const Xbyak::Xmm& result, Xbyak::Xmm xmm_b){
const Xbyak::Xmm mask = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm mask = xmm0;
const Xbyak::Xmm ored = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm ored = ctx.reg_alloc.ScratchXmm();
if (ctx.FPSCR_FTZ()) { if (ctx.FPSCR_FTZ()) {
@ -715,7 +715,7 @@ static void EmitFPVectorMin(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst)
FCODE(vcmpeqp)(mask, result, xmm_b); FCODE(vcmpeqp)(mask, result, xmm_b);
FCODE(vorp)(ored, result, xmm_b); FCODE(vorp)(ored, result, xmm_b);
FCODE(vminp)(result, result, xmm_b); FCODE(vminp)(result, result, xmm_b);
FCODE(vblendvp)(result, result, ored, mask); FCODE(blendvp)(result, ored);
} else { } else {
code.movaps(mask, result); code.movaps(mask, result);
code.movaps(ored, result); code.movaps(ored, result);