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

emit_x64_vector: Remove fallback in EmitVectorSignExtend64()

Browse source 
This is fairly trivial to do manually.
This commit is contained in:
Lioncash 2018-08-29 14:45:52 -04:00 committed by MerryMage
parent 39593fcd26
commit fda19095ea
1 changed files with 22 additions and 18 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

40
src/backend/x64/emit_x64_vector.cpp
View file

@ -2471,31 +2471,35 @@ void EmitX64::EmitVectorSignExtend32(EmitContext& ctx, IR::Inst* inst) {
if (code.DoesCpuSupport(Xbyak::util::Cpu::tSSE41)) {
code.pmovsxdq(a, a);
} else {
const Xbyak::Xmm xmm_tmp = ctx.reg_alloc.ScratchXmm();
const Xbyak::Reg64 gpr_tmp1 = ctx.reg_alloc.ScratchGpr();
const Xbyak::Reg64 gpr_tmp2 = ctx.reg_alloc.ScratchGpr();
const Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
code.movd(gpr_tmp1.cvt32(), a);
code.movsxd(gpr_tmp1, gpr_tmp1.cvt32());
code.pshufd(a, a, 0b11100101);
code.movd(gpr_tmp2.cvt32(), a);
code.movsxd(gpr_tmp2, gpr_tmp2.cvt32());
code.movq(a, gpr_tmp1);
code.movq(xmm_tmp, gpr_tmp2);
code.punpcklqdq(a, xmm_tmp);
code.movaps(tmp, a);
code.psrad(tmp, 31);
code.punpckldq(a, tmp);
}
ctx.reg_alloc.DefineValue(inst, a);
}
void EmitX64::EmitVectorSignExtend64(EmitContext& ctx, IR::Inst* inst) {
EmitOneArgumentFallback(code, ctx, inst, [](VectorArray<u64>& result, const VectorArray<u64>& a) {
result[1] = (a[0] >> 63) ? ~u64(0) : 0;
result[0] = a[0];
});
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const Xbyak::Xmm data = ctx.reg_alloc.UseScratchXmm(args[0]);
const Xbyak::Reg64 gpr_tmp = ctx.reg_alloc.ScratchGpr();
code.movq(gpr_tmp, data);
code.sar(gpr_tmp, 63);
if (code.DoesCpuSupport(Xbyak::util::Cpu::tSSE41)) {
code.pinsrq(data, gpr_tmp, 1);
} else {
const Xbyak::Xmm xmm_tmp = ctx.reg_alloc.ScratchXmm();
code.movq(xmm_tmp, gpr_tmp);
code.punpcklqdq(data, xmm_tmp);
}
ctx.reg_alloc.DefineValue(inst, data);
}
static void EmitVectorSignedAbsoluteDifference(size_t esize, EmitContext& ctx, IR::Inst* inst, BlockOfCode& code) {