[llvm] [AMDGPU] Use V_FMAC_F64 in "if (cond) a -= c" (PR #168710)
Matt Arsenault via llvm-commits
llvm-commits at lists.llvm.org
Fri Jan 30 14:53:45 PST 2026
================
@@ -16284,18 +16284,75 @@ SITargetLowering::performAddCarrySubCarryCombine(SDNode *N,
return SDValue();
}
-SDValue SITargetLowering::performFAddCombine(SDNode *N,
- DAGCombinerInfo &DCI) const {
- if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
+// Create fma((cond ? -1.0, 0.0), c, a)
+static SDValue createConditionalSubToFMA(SelectionDAG &DAG, SDLoc SL, EVT VT,
+ SDValue Cond, SDValue C, SDValue A) {
+ // FIXME: This should work but currently generates incorrect code.
+ if (isa<ConstantFPSDNode>(C) || isa<ConstantSDNode>(C))
return SDValue();
+ const SDValue MinusOne = DAG.getConstantFP(-1.0, SL, VT);
+ const SDValue Zero = DAG.getConstantFP(0.0, SL, VT);
+ SDValue Correction = DAG.getNode(ISD::SELECT, SL, VT, Cond, MinusOne, Zero);
+ return DAG.getNode(ISD::FMA, SL, VT, Correction, C, A);
+}
+SDValue SITargetLowering::performFAddCombine(SDNode *N,
+ DAGCombinerInfo &DCI) const {
SelectionDAG &DAG = DCI.DAG;
EVT VT = N->getValueType(0);
SDLoc SL(N);
SDValue LHS = N->getOperand(0);
SDValue RHS = N->getOperand(1);
+ // Match conditional subtraction patterns for FMA optimization:
+ // result = a - (cond ? c : 0.0) (see performFSubCombine)
+ // result = a + (cond ? -c : 0.0)
+ // result = a + (-(cond ? c : 0.0))
+ // and convert to:
+ // result = fma((cond ? -1.0, 0.0), c, a)
+ //
+ // This saves one v_cndmask per pattern on a pre-gfx12.
+ // The optimization doesn't make sense on gfx12 due to dual-issued v_cndmask.
+ if (Subtarget->shouldUseConditionalSubToFMAF64() &&
+ VT == MVT::f64) { // Only optimize if FMA is available
----------------
arsenm wrote:
FMA is always available
https://github.com/llvm/llvm-project/pull/168710
More information about the llvm-commits
mailing list