[llvm-branch-commits] [llvm] ValueTracking: Handle amdgcn.rsq intrinsic in computeKnownFPClass (PR #171837)

[Yingwei Zheng via llvm-branch-commits]

================
@@ -5553,6 +5553,48 @@ void computeKnownFPClass(const Value *V, const APInt &DemandedElts,
 
       // TODO: Copy inf handling from instructions
       break;
+    case Intrinsic::amdgcn_rsq: {
+      KnownFPClass KnownSrc;
+      // The only negative value that can be returned is -inf for -0 inputs.
+      Known.knownNot(fcNegZero | fcNegSubnormal | fcNegNormal);
+
+      computeKnownFPClass(II->getArgOperand(0), DemandedElts, InterestedClasses,
+                          KnownSrc, Q, Depth + 1);
+
+      // Negative -> nan
+      if (KnownSrc.isKnownNeverNaN() && KnownSrc.cannotBeOrderedLessThanZero())
+        Known.knownNot(fcNan);
+      else if (KnownSrc.isKnownNever(fcSNan))
+        Known.knownNot(fcSNan);
+
+      // -inf -> -0
+      if (KnownSrc.isKnownNeverNegInfinity())
+        Known.knownNot(fcNegZero);
----------------
dtcxzyw wrote:

```suggestion
```
`rsq(-inf) = nan`
<img width="1366" height="754" alt="image" src="https://github.com/user-attachments/assets/4c25b96a-ebf1-4b33-98cb-6ec66b0e7a60" />


https://github.com/llvm/llvm-project/pull/171837