luau-lang/luau
1
0
Fork 0
mirror of https://github.com/luau-lang/luau.git synced 2025-04-03 18:30:54 +01:00
Code Issues Projects Releases Packages Wiki Activity

CodeGen: Rewrite DOT_VEC lowering for A64 using faddp

Browse source 
This results in about the same performance as a naive version on M2,
but uses fewer registers and is what clang generates for a similar
source.
This commit is contained in:
Arseny Kapoulkine 2024-11-08 10:59:44 -08:00
parent 6ebac70495
commit 81b691bf91
1 changed files with 8 additions and 10 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

18
CodeGen/src/IrLoweringA64.cpp
View file

@ -732,17 +732,15 @@ void IrLoweringA64::lowerInst(IrInst& inst, uint32_t index, const IrBlock& next)
{
inst.regA64 = regs.allocReg(KindA64::d, index);
RegisterA64 temp1 = regs.allocTemp(KindA64::q);
RegisterA64 temp2 = regs.allocTemp(KindA64::q);
RegisterA64 temp3 = regs.allocTemp(KindA64::q);
RegisterA64 temp = regs.allocTemp(KindA64::q);
RegisterA64 temps = castReg(KindA64::s, temp);
RegisterA64 regs = castReg(KindA64::s, inst.regA64);
build.fmul(temp1, regOp(inst.a), regOp(inst.b));
build.dup_4s(temp2, temp1, 1);
build.dup_4s(temp3, temp1, 2);
build.fadd(castReg(KindA64::s, temp1), castReg(KindA64::s, temp1), castReg(KindA64::s, temp2));
build.fadd(castReg(KindA64::s, temp1), castReg(KindA64::s, temp1), castReg(KindA64::s, temp3));
build.fcvt(inst.regA64, castReg(KindA64::s, temp1));
build.fmul(temp, regOp(inst.a), regOp(inst.b));
build.faddp(regs, temps); // x+y
build.dup_4s(temp, temp, 2);
build.fadd(regs, regs, temps); // +z
build.fcvt(inst.regA64, regs);
break;
}
case IrCmd::NOT_ANY: