[llvm-branch-commits] [llvm] AMDGPU: Handle amdgcn_rcp in computeKnownFPClass (PR #172490)

[Yingwei Zheng via llvm-branch-commits]

================
@@ -5553,6 +5553,39 @@ void computeKnownFPClass(const Value *V, const APInt &DemandedElts,
 
       // TODO: Copy inf handling from instructions
       break;
+    case Intrinsic::amdgcn_rcp: {
+      KnownFPClass KnownSrc;
+      computeKnownFPClass(II->getArgOperand(0), DemandedElts, InterestedClasses,
+                          KnownSrc, Q, Depth + 1);
+
+      Known.propagateNaN(KnownSrc);
+
+      Type *EltTy = II->getType()->getScalarType();
+
+      // f32 denormal always flushed.
+      if (EltTy->isFloatTy())
+        Known.knownNot(fcSubnormal);
+
+      if (KnownSrc.isKnownNever(fcNegative))
+        Known.knownNot(fcNegative);
+      if (KnownSrc.isKnownNever(fcPositive))
+        Known.knownNot(fcPositive);
+
+      if (KnownSrc.isKnownNeverNegInfinity())
+        Known.knownNot(fcNegZero);
+      if (KnownSrc.isKnownNeverPosInfinity())
+        Known.knownNot(fcPosZero);
----------------
dtcxzyw wrote:

It is incorrect. `rcp(FLT_MAX)` is a denormal value: https://godbolt.org/z/odoWseG4h

https://github.com/llvm/llvm-project/pull/172490