[llvm] [CodeGen] [AMDGPU] Attempt DAGCombine for fmul with select to ldexp (PR #111109)
Matt Arsenault via llvm-commits
llvm-commits at lists.llvm.org
Fri Oct 4 00:52:08 PDT 2024
================
@@ -14476,6 +14477,57 @@ SDValue SITargetLowering::performFDivCombine(SDNode *N,
return SDValue();
}
+SDValue SITargetLowering::performFMulCombine(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);
+
+ // ldexp(x, zext(i1 y)) -> fmul x, (select y, 2.0, 1.0)
+ // ldexp(x, sext(i1 y)) -> fmul x, (select y, 0.5, 1.0)
+ //
+ // The above mentioned ldexp folding works fine for
+ // f16/f32, but as for f64 it creates f64 select which
+ // is costly to materealize as compared to f64 ldexp
+ // so here we undo the transform for f64 as follows :
+ //
+ // fmul x, (select y, 2.0, 1.0) -> ldexp(x, zext(i1 y))
+ // fmul x, (select y, 0.5, 1.0) -> ldexp(x, sext(i1 y))
+ // TODO : Need to handle vector of f64 type.
+ if (VT == MVT::f64) {
+ if (RHS.hasOneUse() && RHS.getOpcode() == ISD::SELECT) {
+ const ConstantFPSDNode *TrueNode =
+ dyn_cast<ConstantFPSDNode>(RHS.getOperand(1));
+ const ConstantFPSDNode *FalseNode =
+ dyn_cast<ConstantFPSDNode>(RHS.getOperand(2));
+
+ if (!TrueNode || !FalseNode)
+ return SDValue();
+
+ const double TrueVal = TrueNode->getValueAPF().convertToDouble();
+ const double FalseVal = FalseNode->getValueAPF().convertToDouble();
----------------
arsenm wrote:
Don't use convertToDouble. Use APFloat::isExactlyValue
https://github.com/llvm/llvm-project/pull/111109
More information about the llvm-commits
mailing list