[Mlir-commits] [mlir] [mlir] Speed up resolving ExtractValueOp source (PR #176478)

[Matthias Springer]

https://github.com/matthias-springer approved this pull request.

Overall, this looks good to me, but let's wait for the discussion with @gysit to settle.

One more question before you merge: can you confirm that this pattern produces the exact same IR as the existing folder (at the end of the canonicalization)? If the IR differs depending on the order/selection of pattern/folder applications, we could have flaky test cases.

> This requires patterns to be applied in top-down order, which is the default for the canonicaliser pass.

Note: This is the case only when the initial worklist is built. Afterwards, operations are processed LIFO.

> I am separately working on simplifying `ExtractValueOp::fold` to do less traversal, but that requires some care to ensure existing cases are not pessimised.

That would be great. Not only to make sure that your new pattern applies all the time. We'd also like to avoid "duplicate" patterns/foldings.


https://github.com/llvm/llvm-project/pull/176478