[llvm] AMDGPU: Match fract pattern with swapped edge case check (PR #189081)

[Matt Arsenault via llvm-commits]

================
@@ -1608,18 +1608,44 @@ bool AMDGPUCodeGenPrepareImpl::visitSelectInst(SelectInst &I) {
   IRBuilder<> Builder(&I);
   Builder.setFastMathFlags(FPOp->getFastMathFlags());
 
-  auto *IITrue = dyn_cast<IntrinsicInst>(TrueVal);
-  auto *IIFalse = dyn_cast<IntrinsicInst>(FalseVal);
-
   Value *Fract = nullptr;
-  if (Pred == FCmpInst::FCMP_UNO && TrueVal == CmpVal && IIFalse &&
-      CmpVal == matchFractPat(*IIFalse)) {
+  if (IsNanPred == FCmpInst::FCMP_UNO && TrueVal == CmpVal &&
+      CmpVal == matchFractPat(*FalseVal)) {
     // isnan(x) ? x : fract(x)
     Fract = applyFractPat(Builder, CmpVal);
-  } else if (Pred == FCmpInst::FCMP_ORD && FalseVal == CmpVal && IITrue &&
-             CmpVal == matchFractPat(*IITrue)) {
-    // !isnan(x) ? fract(x) : x
-    Fract = applyFractPat(Builder, CmpVal);
+  } else if (IsNanPred == FCmpInst::FCMP_ORD && FalseVal == CmpVal) {
+    if (CmpVal == matchFractPat(*TrueVal)) {
+      // !isnan(x) ? fract(x) : x
+      Fract = applyFractPat(Builder, CmpVal);
+    } else {
+      // Match an intermediate clamp infinity to 0 pattern. i.e.
+      // !isnan(x) ? (!isinf(x) ? fract(x) : 0.0) : x
----------------
arsenm wrote:

Instcombine does swap these around, and this is the canonical output. I did try all the combinations and they folded to this. This particular ordering of which check comes first came from the libclc and device libs implementations, which did the opposite 

https://github.com/llvm/llvm-project/pull/189081