[Mlir-commits] [mlir] [mlir][vector] Support complete folding in single pass for vector.insert/vector.extract (PR #142124)

[Diego Caballero]

dcaballe wrote:

In my view, this issue and its fix state that canonicalization is not equivalent to folding when it comes to applying only folding patterns, which some of us (or at least me) were not aware of. I believe the solution should go in the direction of introducing a new `-fold` pass, similar to the `-canonicalize` one, and ensuring that we test both approaches as needed, MLIR-wide.

That said, I don't think we should block this fix or request a major shift in the canonicalization/folding testing in MLIR within this PR. Adding a test to `canonicalize.mlir` or using the constant fold approach should be enough to land this fix. However, we should file a bug to revisit canonicalization/folding testing across MLIR and reference this PR. This is something that may need some broader discussion.

WDYAT?

https://github.com/llvm/llvm-project/pull/142124