[llvm] [DAG] Add ISD::FSQRT handling to computeKnownFPClass (PR #213620)

[Varad Rahul Kamthe via llvm-commits]

================
@@ -6191,6 +6191,25 @@ KnownFPClass SelectionDAG::computeKnownFPClass(SDValue Op,
     Known.fneg();
     break;
   }
+  case ISD::FSQRT: {
+    FPClassTest InterestedSrcs = InterestedClasses;
+    if (InterestedClasses & fcNan)
+      InterestedSrcs |= KnownFPClass::OrderedLessThanZeroMask;
+
+    KnownFPClass KnownSrc = computeKnownFPClass(Op.getOperand(0), DemandedElts,
+                                                InterestedSrcs, Depth + 1);
+
+    const fltSemantics &FltSem =
+        Op.getValueType().getScalarType().getFltSemantics();
+    DenormalMode Mode = getMachineFunction().getDenormalMode(FltSem);
+
+    Known = KnownFPClass::sqrt(KnownSrc, Mode);
+
+    if (Op->getFlags().hasNoNaNs())
----------------
varadk27 wrote:

Hi @arsenm!
Pushed one more update -- this generic change also improves an existing test: AMDGPU's `fsqrt.f32.ll` has a degenerate-value guard (checks `fcZero|fcPosInf`) that now correctly narrows to just `fcZero` for ninf-flagged sqrt calls, since `computeKnownFPClass` can now prove +inf is impossible there. 
Regenerated its CHECK lines to match. Full check-llvm is clean.

https://github.com/llvm/llvm-project/pull/213620