[DAGCombine] Fix an ICE in combineMinNumMaxNum(...)

[cameron.mcinally]

Summary

65420c8041f4 introduced an ICE in combineMinNumMaxNum(...) when combineMinNumMaxNumImpl(...) returns an SDValue(). Make sure to check that a value is returned before trying to perform an FNEG on it.

Note that this ICE occurs in the release/16.x branch as well, so it will need to be backported. @tstellar

[cameron.mcinally]

Update the diff to include context.

[arsenm]

LGTM with test nits

llvm/test/CodeGen/X86/2023-02-22-combineMinNumMaxNum.ll

6

Don't need local_unnamed_addr

8

I suspect you can simplify this too

14

Use named values in test

22

Do we really need the attributes (IIRC, yes for this case which really should be fixed)

[cameron.mcinally]

Updated patch for Matt's reviews...

[cameron.mcinally]

llvm/test/CodeGen/X86/2023-02-22-combineMinNumMaxNum.ll

22

We do. Those are the requirements from isLegalToCombineMinNumMaxNum(...).

[arsenm]

llvm/test/CodeGen/X86/2023-02-22-combineMinNumMaxNum.ll

19

I'm surprised dso_local is allowed on intrinsic declarations

[cameron.mcinally]

llvm/test/CodeGen/X86/2023-02-22-combineMinNumMaxNum.ll

19

Ah, missed that. Will correct before submitting.

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

if (LHS == NegTrue) {
     // the select.
     SDValue NegRHS = TLI.getCheaperOrNeutralNegatedExpression(
         RHS, DAG, LegalOperations, ForCodeSize);
     if (NegRHS) {
       HandleSDNode NegRHSHandle(NegRHS);
       if (NegRHS == False) {
         SDValue Combined = combineMinNumMaxNumImpl(DL, VT, LHS, RHS, NegTrue,
                                                    False, CC, TLI, DAG);
+        if (Combined)
           return DAG.getNode(ISD::FNEG, DL, VT, Combined);
       }
     }
   }
 
   return SDValue();
 }
 
 /// If a (v)select has a condition value that is a sign-bit test, try to smear

llvm/test/CodeGen/X86/2023-02-22-combineMinNumMaxNum.ll

+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mcpu=skylake
+
+; Checking for a DAGCombine ICE.
+
+define float @test_combinemaxnum(float %sub) #0 {
+L.entry:
+  %maxnum1 = call float @llvm.maxnum.f32(float 0.000000e+00, float 0.000000e+00)
+  br label %L.LB21_850
+
+L.LB21_850:
+  %neg1 = fneg fast float %maxnum1
+  %neg2 = fneg fast float %sub
+  %mask = fcmp fast ule float %maxnum1, %neg2
+  %maxnum2 = select i1 %mask, float %neg1, float %sub
+  ret float %maxnum2
+}
+
+declare float @llvm.maxnum.f32(float, float) 
+
+attributes #0 = { "no-nans-fp-math"="true" "no-signed-zeros-fp-math"="true" }