[Mlir-commits] [mlir] [mlir][arith] Move canonicalization patterns to fold (PR #184381)

[Razvan Lupusoru]

================
@@ -1845,6 +1845,14 @@ bool arith::IndexCastOp::areCastCompatible(TypeRange inputs,
 }
 
 OpFoldResult arith::IndexCastOp::fold(FoldAdaptor adaptor) {
+  // index_cast(index_cast(x, exact)) -> x, when result type matches x's type.
+  // The inner exact guarantees the iN -> index conversion is lossless,
+  // so the roundtrip through index preserves the value.
+  if (auto innerCast = getIn().getDefiningOp<IndexCastOp>()) {
+    if (innerCast.getIn().getType() == getType() && innerCast.getIsExact())
----------------
razvanlupusoru wrote:

I posted a question about this scenario in your previous PR:
https://github.com/llvm/llvm-project/pull/183395/changes#r2880422529
I hope we reach a reasonable resolution :) Thank you!

https://github.com/llvm/llvm-project/pull/184381