AMDGPU: Constant fold rcp node
When doing arcp optimization with a constant denominator,
this was leaving behind rcps with constant inputs.
llvm-svn: 297248
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index edaab00..41a4f76 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -1295,8 +1295,8 @@
SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FltVT, fq);
// float fr = mad(fqneg, fb, fa);
- unsigned OpCode = Subtarget->hasFP32Denormals() ?
- (unsigned)AMDGPUISD::FMAD_FTZ :
+ unsigned OpCode = Subtarget->hasFP32Denormals() ?
+ (unsigned)AMDGPUISD::FMAD_FTZ :
(unsigned)ISD::FMAD;
SDValue fr = DAG.getNode(OpCode, DL, FltVT, fqneg, fb, fa);
@@ -3351,6 +3351,16 @@
return performStoreCombine(N, DCI);
case AMDGPUISD::CLAMP:
return performClampCombine(N, DCI);
+ case AMDGPUISD::RCP: {
+ if (const auto *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0))) {
+ // XXX - Should this flush denormals?
+ const APFloat &Val = CFP->getValueAPF();
+ APFloat One(Val.getSemantics(), "1.0");
+ return DAG.getConstantFP(One / Val, SDLoc(N), N->getValueType(0));
+ }
+
+ break;
+ }
}
return SDValue();
}