[Mlir-commits] [mlir] [mlir][arith] Fix arith maxnumf/minnumf folder (PR #114595)

[Clément Fournier]

================
@@ -1014,13 +1014,14 @@ OpFoldResult arith::MaxNumFOp::fold(FoldAdaptor adaptor) {
   if (getLhs() == getRhs())
     return getRhs();
 
-  // maxnumf(x, -inf) -> x
-  if (matchPattern(adaptor.getRhs(), m_NegInfFloat()))
+  // maxnumf(x, NaN) -> x
+  if (matchPattern(adaptor.getRhs(), m_NaNFloat()))
     return getLhs();
 
-  return constFoldBinaryOp<FloatAttr>(
-      adaptor.getOperands(),
-      [](const APFloat &a, const APFloat &b) { return llvm::maximum(a, b); });
----------------
oowekyala wrote:

This matters if the rhs folds to a number, and the LHS folds to NaN. Because this used `maximum` and not `maximumnum`, it would have folded the operation to `NaN` and not to the LHS.

https://github.com/llvm/llvm-project/pull/114595