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

emit_x64_vector_floating_point: Avoid checking inputs for NaNs for three-ops where able

Browse source
This commit is contained in:
MerryMage 2021-03-28 21:50:49 +01:00
parent e06933f123
commit c4cff773b9
1 changed files with 25 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

30
src/backend/x64/emit_x64_vector_floating_point.cpp
View file

@ -337,8 +337,12 @@ void EmitTwoOpVectorOperation(BlockOfCode& code, EmitContext& ctx, IR::Inst* ins
ctx.reg_alloc.DefineValue(inst, result); ctx.reg_alloc.DefineValue(inst, result);
} }
enum CheckInputNaN {
Yes, No,
};
template<size_t fsize, template<typename> class Indexer, typename Function> template<size_t fsize, template<typename> class Indexer, typename Function>
void EmitThreeOpVectorOperation(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst, Function fn, typename NaNHandler<fsize, Indexer, 3>::function_type nan_handler = NaNHandler<fsize, Indexer, 3>::GetDefault()) { void EmitThreeOpVectorOperation(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst, Function fn, CheckInputNaN check_input_nan = CheckInputNaN::No, typename NaNHandler<fsize, Indexer, 3>::function_type nan_handler = NaNHandler<fsize, Indexer, 3>::GetDefault()) {
static_assert(fsize == 32 || fsize == 64, "fsize must be either 32 or 64"); static_assert(fsize == 32 || fsize == 64, "fsize must be either 32 or 64");
auto args = ctx.reg_alloc.GetArgumentInfo(inst); auto args = ctx.reg_alloc.GetArgumentInfo(inst);
@ -371,15 +375,31 @@ void EmitThreeOpVectorOperation(BlockOfCode& code, EmitContext& ctx, IR::Inst* i
const Xbyak::Xmm xmm_b = ctx.reg_alloc.UseXmm(args[1]); const Xbyak::Xmm xmm_b = ctx.reg_alloc.UseXmm(args[1]);
const Xbyak::Xmm nan_mask = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm nan_mask = ctx.reg_alloc.ScratchXmm();
code.movaps(nan_mask, xmm_b);
code.movaps(result, xmm_a); code.movaps(result, xmm_a);
FCODE(cmpunordp)(nan_mask, xmm_a);
if (check_input_nan == CheckInputNaN::Yes) {
if (code.HasAVX()) {
FCODE(vcmpunordp)(nan_mask, xmm_a, xmm_b);
} else {
code.movaps(nan_mask, xmm_b);
FCODE(cmpunordp)(nan_mask, xmm_a);
}
}
if constexpr (std::is_member_function_pointer_v<Function>) { if constexpr (std::is_member_function_pointer_v<Function>) {
(code.*fn)(result, xmm_b); (code.*fn)(result, xmm_b);
} else { } else {
fn(result, xmm_b); fn(result, xmm_b);
} }
FCODE(cmpunordp)(nan_mask, result);
if (check_input_nan == CheckInputNaN::Yes) {
FCODE(cmpunordp)(nan_mask, result);
} else if (code.HasAVX()) {
FCODE(vcmpunordp)(nan_mask, result, result);
} else {
code.movaps(nan_mask, result);
FCODE(cmpunordp)(nan_mask, nan_mask);
}
HandleNaNs<fsize, 2>(code, ctx, fpcr_controlled, {result, xmm_a, xmm_b}, nan_mask, nan_handler); HandleNaNs<fsize, 2>(code, ctx, fpcr_controlled, {result, xmm_a, xmm_b}, nan_mask, nan_handler);
@ -951,7 +971,7 @@ static void EmitFPVectorMinMax(BlockOfCode& code, EmitContext& ctx, IR::Inst* in
code.andnps(mask, eq); code.andnps(mask, eq);
code.orps(result, mask); code.orps(result, mask);
} }
}); }, CheckInputNaN::Yes);
} }
void EmitX64::EmitFPVectorMax32(EmitContext& ctx, IR::Inst* inst) { void EmitX64::EmitFPVectorMax32(EmitContext& ctx, IR::Inst* inst) {