[Mlir-commits] [mlir] [MLIR] Enable pattern only for scf.forall op (PR #110230)

[llvmlistbot at llvm.org]

================
@@ -131,11 +132,25 @@ struct IterArgsToInitArgs : public OpRewritePattern<tensor::DimOp> {
     auto blockArg = dyn_cast<BlockArgument>(dimOp.getSource());
     if (!blockArg)
       return failure();
-    auto loopLikeOp =
-        dyn_cast<LoopLikeOpInterface>(blockArg.getParentBlock()->getParentOp());
-    if (!loopLikeOp)
+    // TODO: Enable this for loopLikeInterface. Restricting for scf.for
----------------
MaheshRavishankar wrote:

I think that pattern is wrong for the reason why this is being restricted to forall only. The init arg shape can vary from iteration to iteration. There it is added as a canonicalization which is even worse.

https://github.com/llvm/llvm-project/pull/110230