[llvm] [AMDGPU] Add OMOD folding for TRANS bfloat16 instructions (PR #208595)
Shilei Tian via llvm-commits
llvm-commits at lists.llvm.org
Thu Jul 9 19:29:46 PDT 2026
================
@@ -2363,6 +2358,117 @@ SIFoldOperandsImpl::isOMod(const MachineInstr &MI) const {
return std::pair(nullptr, SIOutMods::NONE);
}
+ case AMDGPU::V_FMA_MIXLO_BF16:
+ case AMDGPU::V_FMA_MIX_BF16_t16: {
+ if (MFI->getMode().FP32Denormals.Output != DenormalMode::PreserveSign ||
+ MI.mayRaiseFPException())
+ return std::pair(nullptr, SIOutMods::NONE);
+
+ const MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
+ const MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
+ const MachineOperand *Src2 = TII->getNamedOperand(MI, AMDGPU::OpName::src2);
+
+ if (!Src2 || !Src2->isImm() || Src2->getImm() != 0)
+ return std::pair(nullptr, SIOutMods::NONE);
+
+ // op_sel bits are in src*_modifiers but don't affect folding
+ const MachineOperand *Src0Mods =
+ TII->getNamedOperand(MI, AMDGPU::OpName::src0_modifiers);
+ const MachineOperand *Src1Mods =
+ TII->getNamedOperand(MI, AMDGPU::OpName::src1_modifiers);
+ if ((Src0Mods &&
+ (Src0Mods->getImm() & (SISrcMods::NEG | SISrcMods::ABS))) ||
+ (Src1Mods &&
+ (Src1Mods->getImm() & (SISrcMods::NEG | SISrcMods::ABS))) ||
+ TII->hasModifiersSet(MI, AMDGPU::OpName::omod) ||
+ TII->hasModifiersSet(MI, AMDGPU::OpName::clamp))
+ return std::pair(nullptr, SIOutMods::NONE);
+
+ if (!Src0->isReg() || !Src1->isReg())
+ return std::pair(nullptr, SIOutMods::NONE);
+
+ const MachineOperand *RegOp = nullptr;
+ int64_t ImmValue = 0;
+
+ const MachineInstr *Src1Def = MRI->getVRegDef(Src1->getReg());
+ if (Src1Def && Src1Def->isMoveImmediate() &&
+ Src1Def->getOperand(1).isImm()) {
+ ImmValue = Src1Def->getOperand(1).getImm();
+ RegOp = Src0;
+ } else {
+ const MachineInstr *Src0Def = MRI->getVRegDef(Src0->getReg());
+ if (Src0Def && Src0Def->isMoveImmediate() &&
+ Src0Def->getOperand(1).isImm()) {
+ ImmValue = Src0Def->getOperand(1).getImm();
+ RegOp = Src1;
+ } else {
+ return std::pair(nullptr, SIOutMods::NONE);
+ }
+ }
+
+ int OMod = getOModValue(AMDGPU::V_MUL_F32_e64, ImmValue);
----------------
shiltian wrote:
What does `V_MUL_F32_e64` do here?
https://github.com/llvm/llvm-project/pull/208595
More information about the llvm-commits
mailing list