[Mlir-commits] [mlir] [mlir] Handle arith.const expr in dispatchIndexOpFoldResult func (PR #122432)

[Matthias Springer]

================
@@ -20,6 +20,26 @@ func.func @bubble_parallel_reshapes(%arg0: tensor<?x?x?x?xf32>, %s0: index, %s1:
 
 // -----
 
+func.func @bubble_parallel_reshapes2(%arg0: tensor<?x2x2x6xf32>, %s0: index, %s1: index) -> tensor<?x4x2x3xf32> {
+  %c2 = arith.constant 2 : index
+  %c3 = arith.constant 3 : index
+  %collapse = tensor.collapse_shape %arg0 [[0], [1, 2], [3]] : tensor<?x2x2x6xf32> into tensor<?x4x6xf32>
+  %expand = tensor.expand_shape %collapse [[0], [1], [2, 3]]
+              output_shape [%s0, %s1, %c2, %c3] : tensor<?x4x6xf32> into tensor<?x4x2x3xf32>
----------------
matthias-springer wrote:

For consistency with other tensor dialect ops, I would recommend to make the verifier stricter. See discussion [here](https://discourse.llvm.org/t/tensor-ops-with-dynamic-sizes-which-behaviour-is-more-correct/82612).

https://github.com/llvm/llvm-project/pull/122432