AMDGPU: Fold undef fcanonicalize to qNaN

We could choose a free 0 for this, but this
matches the behavior for fmul undef, 1.0. Also,
the NaN use is more useful for folding use operations
although if it's not eliminated it is more expensive
in terms of code size.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@338376 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AMDGPU/SIISelLowering.cpp b/lib/Target/AMDGPU/SIISelLowering.cpp
index 97c38e44e408a89d47a4c2e7e3463c187b2ca522..acde638a00651c11dc58b9812c185ae3445f3218 100644 (file)
--- a/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -6845,8 +6845,16 @@ SDValue SITargetLowering::performFCanonicalizeCombine(
   SDNode *N,
   DAGCombinerInfo &DCI) const {
   SelectionDAG &DAG = DCI.DAG;
-  ConstantFPSDNode *CFP = isConstOrConstSplatFP(N->getOperand(0));
+  SDValue N0 = N->getOperand(0);
+
+  // fcanonicalize undef -> qnan
+  if (N0.isUndef()) {
+    EVT VT = N->getValueType(0);
+    APFloat QNaN = APFloat::getQNaN(SelectionDAG::EVTToAPFloatSemantics(VT));
+    return DAG.getConstantFP(QNaN, SDLoc(N), VT);
+  }
 
+  ConstantFPSDNode *CFP = isConstOrConstSplatFP(N0);
   if (!CFP) {
     SDValue N0 = N->getOperand(0);
     EVT VT = N0.getValueType().getScalarType();
@@ -6899,7 +6907,7 @@ SDValue SITargetLowering::performFCanonicalizeCombine(
       return DAG.getConstantFP(CanonicalQNaN, SDLoc(N), VT);
   }
 
-  return N->getOperand(0);
+  return N0;
 }
 
 static unsigned minMaxOpcToMin3Max3Opc(unsigned Opc) {

diff --git a/test/CodeGen/AMDGPU/fcanonicalize.ll b/test/CodeGen/AMDGPU/fcanonicalize.ll
index 7c4698b8cfaf9f51e21a55039659366e3a87a2b3..6b2d58db804ee7c7e94a4205aeedbdcba2e83f2c 100644 (file)
--- a/test/CodeGen/AMDGPU/fcanonicalize.ll
+++ b/test/CodeGen/AMDGPU/fcanonicalize.ll
@@ -62,6 +62,15 @@ define amdgpu_kernel void @v_test_canonicalize_fneg_var_f32(float addrspace(1)*
   ret void
 }
 
+; GCN-LABEL: {{^}}test_fold_canonicalize_undef_f32:
+; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7fc00000{{$}}
+; GCN: buffer_store_dword [[REG]]
+define amdgpu_kernel void @test_fold_canonicalize_undef_f32(float addrspace(1)* %out) #1 {
+  %canonicalized = call float @llvm.canonicalize.f32(float undef)
+  store float %canonicalized, float addrspace(1)* %out
+  ret void
+}
+
 ; GCN-LABEL: {{^}}test_fold_canonicalize_p0_f32:
 ; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0{{$}}
 ; GCN: buffer_store_dword [[REG]]