[Mlir-commits] [mlir] [MLIR][Tensor] Remove tensor.dim canonicalization patterns registered on tensor.expand_shape/tensor.collapse_shape (PR #134219)

[Nicolas Vasilache]

https://github.com/nicolasvasilache commented:

Spoke about this PR with @Groverkss today and I believe this PR may trying to cure a symptom by going into a local minimum.

For future reference if needed:
1. collapse/expand semantics had evolved and have internal assumptions that were not mirrored in the canonicalization pattern
2. there is machinery missing to properly support collapse/expand shape and other things more generally (e.g. divisibility information)
3. the canonicalization pattern was dropping information (due to 2.)

This should probably be revisited in the future.

https://github.com/llvm/llvm-project/pull/134219