[Mlir-commits] [mlir] [Mlir] decompose generic by unfolding projected permutation crash fix (PR #122449)

[Andrzej Warzyński]

banach-space wrote:

> Personally, I often used this no_ACTION_CASE convention as I look at test case naming w.r.t. file name and/or intendent pass i.e., action taken.

That works really well for .e.g `@fold_<CASE>_<SUB-CASE>` (vs `@no_fold_<CASE>_<SUB-CASE>`), but "breaks" if there's no `ACTION`, e.g. here:
https://github.com/llvm/llvm-project/blob/ee655ca27aad466bcc54f6eba03f7e564940ad5a/mlir/test/Dialect/Vector/vector-transfer-collapse-inner-most-dims.mlir#L45-L50

We could re-name those test function names ...

> Negative cases don't need emphasis in invalid.mlir file

+1

> Thinking about it more... Perhaps fail_name could work? 

`fail` feels a bit strong though. While it is indeed `pattern application "failure"` (in most cases), some folks might think it's some more fundamental failure, no?

Tl;Dr I see two good/valid approaches here:

* `@no_ACTION_CASE`,
* `@negative_CASE`.

If we manage to reduce the number of options that people use to those two, that would already be a success 😅  I am happy to write a PR that will update MLIR recommendation for testing if you also feel that this would be beneficial.

Mostly thinking out loud, I agree that in general none of these is perfect.

https://github.com/llvm/llvm-project/pull/122449