[Mlir-commits] [mlir] [mlir] Re-land Loosen restrictions on folding dynamic reshapes (PR #142827)

[Artem Gindinson]

AGindinson wrote:

https://github.com/llvm/llvm-project/pull/144118 is the draft PR. However, I'm genuinely unsure which is the more expressive representation for scalars/pseudo-tensors:
- `tensor.collapse_shape %arg0 [] : tensor<?x?x?xf32> into tensor<f32>`, or
- tensor.collapse_shape %arg0 [[0, 1, 2]] : tensor<?x?x?xf32> into tensor<f32>

...because the essence is similar to `tensor<1x1xT> into tensor<1xT>`. Obviously, I don't want to assert the "new" representation is better before proposing adjustments to the validation routine.

https://github.com/llvm/llvm-project/pull/142827