[Mlir-commits] [mlir] [mlir][memref] Add foldUseDominateCast function to castOp (PR #168337)

[Mehdi Amini]

================
@@ -793,8 +795,32 @@ bool CastOp::areCastCompatible(TypeRange inputs, TypeRange outputs) {
   return false;
 }
 
+static OpFoldResult foldUseDominateCast(CastOp castOp) {
+  auto funcOp = castOp->getParentOfType<FunctionOpInterface>();
+  if (!funcOp)
+    return {};
+  auto castOps = castOp->getBlock()->getOps<CastOp>();
+  CastOp dominateCastOp = castOp;
+  SmallVector<CastOp> ops(castOps);
+  mlir::DominanceInfo dominanceInfo(castOp);
+  for (auto it : castOps) {
+    if (it.getSource() == dominateCastOp.getSource() &&
+        it.getDest().getType() == dominateCastOp.getDest().getType() &&
+        dominanceInfo.dominates(it.getOperation(),
+                                dominateCastOp.getOperation())) {
+      dominateCastOp = it;
+    }
+  }
----------------
joker-eph wrote:

Why do you need the dominanceInfo when you're just processing operations in a single block? Seems heavy for something that is naturally ordered.

https://github.com/llvm/llvm-project/pull/168337