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

emit_x64_vector: Use VPSRAQ in EmitVectorArithmeticShiftRight64() if AVX-512VL is available

Browse source
This commit is contained in:
Lioncash 2018-05-02 16:04:02 -04:00 committed by MerryMage
parent b38dd191bd
commit d4ee878cbd
1 changed files with 15 additions and 11 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

26
src/backend_x64/emit_x64_vector.cpp
View file

@ -385,20 +385,24 @@ void EmitX64::EmitVectorArithmeticShiftRight32(EmitContext& ctx, IR::Inst* inst)
void EmitX64::EmitVectorArithmeticShiftRight64(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
Xbyak::Xmm tmp1 = ctx.reg_alloc.ScratchXmm();
Xbyak::Xmm tmp2 = ctx.reg_alloc.ScratchXmm();
const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
const u8 shift_amount = std::min(args[1].GetImmediateU8(), u8(63));
const u64 sign_bit = 0x80000000'00000000u >> shift_amount;
if (code.DoesCpuSupport(Xbyak::util::Cpu::tAVX512VL)) {
code.vpsraq(result, result, shift_amount);
} else {
const Xbyak::Xmm tmp1 = ctx.reg_alloc.ScratchXmm();
const Xbyak::Xmm tmp2 = ctx.reg_alloc.ScratchXmm();
code.pxor(tmp2, tmp2);
code.psrlq(result, shift_amount);
code.movdqa(tmp1, code.MConst(xword, sign_bit, sign_bit));
code.pand(tmp1, result);
code.psubq(tmp2, tmp1);
code.por(result, tmp2);
const u64 sign_bit = 0x80000000'00000000u >> shift_amount;
code.pxor(tmp2, tmp2);
code.psrlq(result, shift_amount);
code.movdqa(tmp1, code.MConst(xword, sign_bit, sign_bit));
code.pand(tmp1, result);
code.psubq(tmp2, tmp1);
code.por(result, tmp2);
}
ctx.reg_alloc.DefineValue(inst, result);
}