[llvm-branch-commits] [llvm] InstCombine: Handle rounding intrinsics in SimplifyDemandedFPClass (PR #174842)

[Jay Foad via llvm-branch-commits]

================
@@ -2519,6 +2519,73 @@ Value *InstCombinerImpl::SimplifyDemandedUseFPClass(Instruction *I,
 
       return getFPClassConstant(VTy, ValidResults, /*IsCanonicalizing=*/true);
     }
+    case Intrinsic::trunc:
+    case Intrinsic::floor:
+    case Intrinsic::ceil:
+    case Intrinsic::rint:
+    case Intrinsic::nearbyint:
+    case Intrinsic::round:
+    case Intrinsic::roundeven: {
+      FPClassTest DemandedSrcMask = DemandedMask;
+
+      // Zero results imply valid subnormal sources.
+      if (DemandedMask & fcNegZero)
+        DemandedSrcMask |= fcNegSubnormal;
+
+      if (DemandedMask & fcPosZero)
+        DemandedSrcMask |= fcPosSubnormal;
----------------
jayfoad wrote:

But you can also get zero results from normal inputs, e.g. trunc(0.5) or trunc(-0.5)

https://github.com/llvm/llvm-project/pull/174842