[Mlir-commits] [mlir] [mlir] Drop outermost dims in slice rank reduction inference (PR #95020)

[llvmlistbot at llvm.org]

================
@@ -282,23 +282,6 @@ func.func @insert_slice_of_insert_slice(%t: tensor<f32>, %r0: tensor<1x1xf32>, %
 
 // -----
 
-// CHECK-LABEL: func @insert_slice_of_insert_slice(
----------------
Max191 wrote:

This test is failing with the new changes, but I think the test itself is ambiguous and doesn't make much sense. This is testing that an out of bounds offset into a single element tensor gets composed specifically with the out of bounds offset into the unit dimension of another tensor when the two inserts are combined.

(1) Based on the conventional way of inferring rank reduced dimensions, the offset should be composed with the inner offset into the `14` dim, so the test is incorrect.
(2) I would have simply updated the test, but the test doesn't make much sense to me anyway. It is composing 2 out of bounds inserts with each other to make an even more out of bounds insert. The behavior of both of these inserts is undefined, so composing them into a new, different undefined insert seems like a strange test to me.

If you see a reason to keep the test, then I can add it back and fix it, but otherwise I figured it made more sense to just remove it.

https://github.com/llvm/llvm-project/pull/95020