[llvm-branch-commits] [llvm] InstCombine: Handle fmul in SimplifyDemandedFPClass (PR #173872)

[Matt Arsenault via llvm-branch-commits]

================
@@ -2280,6 +2290,127 @@ Value *InstCombinerImpl::SimplifyDemandedUseFPClass(Value *V,
     Known = KnownLHS | KnownRHS;
     break;
   }
+  case Instruction::FMul: {
+    KnownFPClass KnownLHS, KnownRHS;
+
+    Value *X = I->getOperand(0);
+    Value *Y = I->getOperand(1);
+
+    FPClassTest SrcDemandedMask =
+        DemandedMask & (fcNan | fcZero | fcSubnormal | fcNormal);
+
+    if (DemandedMask & fcInf) {
+      // mul x, inf = inf
+      // mul large_x, large_y = inf
+      SrcDemandedMask |= fcSubnormal | fcNormal | fcInf;
+    }
+
+    if (DemandedMask & fcNan) {
+      // mul +/-inf, 0 => nan
+      SrcDemandedMask |= fcZero | fcInf;
+
+      // TODO: Mode check
+      // mul +/-inf, sub => nan if daz
+      SrcDemandedMask |= fcSubnormal;
+    }
+
+    if (X == Y) {
+      if (SimplifyDemandedFPClass(I, 0, SrcDemandedMask, KnownLHS, Depth + 1))
+        return I;
----------------
arsenm wrote:

I think this is kind of broken. It works out in these tests since the only select simplifications performed can support multiple uses. Really this needs to be broken up into single and multiple use versions like SimpilfyDemandedBits is 

https://github.com/llvm/llvm-project/pull/173872