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

emit_x64_floating_point: Add AVX512F conversion operations to EmitFPU32ToSingle() and EmitFPU32ToDouble()

Browse source 
AVX-512F provides convenient instructions for these kinds of conversions
directly
This commit is contained in:
Lioncash 2018-07-09 17:46:31 -04:00 committed by MerryMage
parent 3a41465eaf
commit fbd7623fe5
1 changed files with 22 additions and 12 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

22
src/backend_x64/emit_x64_floating_point.cpp
View file

@ -1031,14 +1031,19 @@ void EmitX64::EmitFPS32ToSingle(EmitContext& ctx, IR::Inst* inst) {
void EmitX64::EmitFPU32ToSingle(EmitContext& ctx, IR::Inst* inst) { void EmitX64::EmitFPU32ToSingle(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst); auto args = ctx.reg_alloc.GetArgumentInfo(inst);
Xbyak::Reg64 from = ctx.reg_alloc.UseGpr(args[0]);
Xbyak::Xmm to = ctx.reg_alloc.ScratchXmm(); const Xbyak::Reg64 from = ctx.reg_alloc.UseGpr(args[0]);
bool round_to_nearest = args[1].GetImmediateU1(); const Xbyak::Xmm to = ctx.reg_alloc.ScratchXmm();
const bool round_to_nearest = args[1].GetImmediateU1();
ASSERT_MSG(!round_to_nearest, "round_to_nearest unimplemented"); ASSERT_MSG(!round_to_nearest, "round_to_nearest unimplemented");
if (code.DoesCpuSupport(Xbyak::util::Cpu::tAVX512F)) {
code.vcvtusi2ss(to, to, from.cvt32());
} else {
// We are using a 64-bit GPR register to ensure we don't end up treating the input as signed // We are using a 64-bit GPR register to ensure we don't end up treating the input as signed
code.mov(from.cvt32(), from.cvt32()); // TODO: Verify if this is necessary code.mov(from.cvt32(), from.cvt32()); // TODO: Verify if this is necessary
code.cvtsi2ss(to, from); code.cvtsi2ss(to, from);
}
ctx.reg_alloc.DefineValue(inst, to); ctx.reg_alloc.DefineValue(inst, to);
} }
@ -1070,14 +1075,19 @@ void EmitX64::EmitFPS64ToDouble(EmitContext& ctx, IR::Inst* inst) {
void EmitX64::EmitFPU32ToDouble(EmitContext& ctx, IR::Inst* inst) { void EmitX64::EmitFPU32ToDouble(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst); auto args = ctx.reg_alloc.GetArgumentInfo(inst);
Xbyak::Reg64 from = ctx.reg_alloc.UseGpr(args[0]);
Xbyak::Xmm to = ctx.reg_alloc.ScratchXmm(); const Xbyak::Reg64 from = ctx.reg_alloc.UseGpr(args[0]);
bool round_to_nearest = args[1].GetImmediateU1(); const Xbyak::Xmm to = ctx.reg_alloc.ScratchXmm();
const bool round_to_nearest = args[1].GetImmediateU1();
ASSERT_MSG(!round_to_nearest, "round_to_nearest unimplemented"); ASSERT_MSG(!round_to_nearest, "round_to_nearest unimplemented");
if (code.DoesCpuSupport(Xbyak::util::Cpu::tAVX512F)) {
code.vcvtusi2sd(to, to, from.cvt32());
} else {
// We are using a 64-bit GPR register to ensure we don't end up treating the input as signed // We are using a 64-bit GPR register to ensure we don't end up treating the input as signed
code.mov(from.cvt32(), from.cvt32()); // TODO: Verify if this is necessary code.mov(from.cvt32(), from.cvt32()); // TODO: Verify if this is necessary
code.cvtsi2sd(to, from); code.cvtsi2sd(to, from);
}
ctx.reg_alloc.DefineValue(inst, to); ctx.reg_alloc.DefineValue(inst, to);
} }