[PATCH] D50041: AMDGPU: Fold undef fcanonicalize to qNaN

[Matt Arsenault via Phabricator via llvm-commits]

arsenm created this revision.
arsenm added a reviewer: rampitec.
Herald added subscribers: t-tye, tpr, dstuttard, yaxunl, nhaehnle, wdng, kzhuravl.

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.


https://reviews.llvm.org/D50041

Files:
  lib/Target/AMDGPU/SIISelLowering.cpp
  test/CodeGen/AMDGPU/fcanonicalize.ll


Index: test/CodeGen/AMDGPU/fcanonicalize.ll
===================================================================
--- test/CodeGen/AMDGPU/fcanonicalize.ll
+++ test/CodeGen/AMDGPU/fcanonicalize.ll
@@ -62,6 +62,15 @@
   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]]
Index: lib/Target/AMDGPU/SIISelLowering.cpp
===================================================================
--- lib/Target/AMDGPU/SIISelLowering.cpp
+++ lib/Target/AMDGPU/SIISelLowering.cpp
@@ -6876,10 +6876,16 @@
   SDNode *N,
   DAGCombinerInfo &DCI) const {
   SelectionDAG &DAG = DCI.DAG;
-  ConstantFPSDNode *CFP = isConstOrConstSplatFP(N->getOperand(0));
+  SDValue N0 = N->getOperand(0);
+
+  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);
     return isCanonicalized(DAG, N0) ? N0 : SDValue();
   }
 
@@ -6915,7 +6921,7 @@
       return DAG.getConstantFP(CanonicalQNaN, SDLoc(N), VT);
   }
 
-  return N->getOperand(0);
+  return N0;
 }
 
 static unsigned minMaxOpcToMin3Max3Opc(unsigned Opc) {


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D50041.158184.patch
Type: text/x-patch
Size: 1692 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180731/4b22c8e1/attachment.bin>