[Mlir-commits] [mlir] [Linalg] Update Conv Decomposition patterns to work with generic convolution ops as well (PR #174196)

[Andrzej Warzyński]

banach-space wrote:

High-level question: **should Linalg transformations be taking this path?**

This PR updates convolution "decomposition" (and also "vectorization" patterns) to match both named conv ops and _equivalent_ `linalg.generic` forms. However, there are two ways to enable transformations to work on generics:

1. **Generalize/specialize within other transformations** (e.g. inside decomposition) - this is the approach taken in this PR.
2. **Rely on an explicit pre-processing pipeline step** (i.e. run two transformations independently, e.g. generalization followed by decomposition).

Given that (2) provides a cleaner separation of concerns and keeps transformations orthogonal, **why choose (1) here?** In particular, what cannot be achieved by relying on an explicit generalization step?

Also, tests in this PR appear to duplicate inputs to cover both named and generic forms inside decomposition/vectorization tests. We already test specialization and decomposition extensively on their own, so this looks like test duplication rather than additional coverage.

https://github.com/llvm/llvm-project/pull/174196