suyu/dynarmic
26
1
Fork 1
Code Issues Pull requests Projects Releases Packages Wiki Activity

emit_x64_vector_floating_point: Reduce codesize of EmitTwoOpVectorOperation

Browse source
This commit is contained in:
MerryMage 2018-07-31 20:25:03 +01:00
parent bb93353f94
commit 7731dcdca9
1 changed files with 21 additions and 13 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

34
src/backend_x64/emit_x64_vector_floating_point.cpp
View file

@ -241,17 +241,21 @@ void EmitTwoOpVectorOperation(BlockOfCode& code, EmitContext& ctx, IR::Inst* ins
if (!ctx.AccurateNaN() || ctx.FPSCR_DN()) { if (!ctx.AccurateNaN() || ctx.FPSCR_DN()) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst); auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const Xbyak::Xmm xmm_a = ctx.reg_alloc.UseScratchXmm(args[0]);
Xbyak::Xmm result;
if constexpr (std::is_member_function_pointer_v<Function>) { if constexpr (std::is_member_function_pointer_v<Function>) {
(code.*fn)(xmm_a); result = ctx.reg_alloc.UseScratchXmm(args[0]);
(code.*fn)(result);
} else { } else {
fn(xmm_a); const Xbyak::Xmm xmm_a = ctx.reg_alloc.UseXmm(args[0]);
result = ctx.reg_alloc.ScratchXmm();
fn(result, xmm_a);
} }
ForceToDefaultNaN<fsize>(code, ctx, xmm_a); ForceToDefaultNaN<fsize>(code, ctx, result);
ctx.reg_alloc.DefineValue(inst, xmm_a); ctx.reg_alloc.DefineValue(inst, result);
return; return;
} }
@ -261,15 +265,19 @@ void EmitTwoOpVectorOperation(BlockOfCode& code, EmitContext& ctx, IR::Inst* ins
const Xbyak::Xmm xmm_a = ctx.reg_alloc.UseXmm(args[0]); const Xbyak::Xmm xmm_a = ctx.reg_alloc.UseXmm(args[0]);
const Xbyak::Xmm nan_mask = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm nan_mask = ctx.reg_alloc.ScratchXmm();
code.movaps(nan_mask, xmm_a);
code.movaps(result, xmm_a);
FCODE(cmpunordp)(nan_mask, nan_mask);
if constexpr (std::is_member_function_pointer_v<Function>) { if constexpr (std::is_member_function_pointer_v<Function>) {
code.movaps(result, xmm_a);
(code.*fn)(result); (code.*fn)(result);
} else { } else {
fn(result); fn(result, xmm_a);
}
if (code.DoesCpuSupport(Xbyak::util::Cpu::tAVX)) {
FCODE(vcmpunordp)(nan_mask, result, result);
} else {
code.movaps(nan_mask, result);
FCODE(cmpunordp)(nan_mask, nan_mask);
} }
FCODE(cmpunordp)(nan_mask, result);
HandleNaNs<fsize, 1>(code, ctx, {result, xmm_a}, nan_mask, nan_handler); HandleNaNs<fsize, 1>(code, ctx, {result, xmm_a}, nan_mask, nan_handler);
@ -907,11 +915,11 @@ void EmitFPVectorRoundInt(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
return 0; return 0;
}(); }();
EmitTwoOpVectorOperation<fsize, DefaultIndexer>(code, ctx, inst, [&](const Xbyak::Xmm& result){ EmitTwoOpVectorOperation<fsize, DefaultIndexer>(code, ctx, inst, [&](const Xbyak::Xmm& result, const Xbyak::Xmm& xmm_a){
if constexpr (fsize == 32) { if constexpr (fsize == 32) {
code.roundps(result, result, round_imm); code.roundps(result, xmm_a, round_imm);
} else { } else {
code.roundpd(result, result, round_imm); code.roundpd(result, xmm_a, round_imm);
} }
}); });