[llvm] [LAA][NFC] Add a test for #191867 (PR #213722)

[Andrei Elovikov via llvm-commits]

eas wrote:

> Not sure if this test makes sense: there's a stride1 variant, followed by a stride2 variant, followed again by a stride1 variant; why would the second stride1 variant be influenced by the stride2 variant? Was it happening before your fix?

It's basically a test for https://github.com/llvm/llvm-project/pull/191867/changes/1d5711517d7ed48dbefebebe335102b6d9bcf281 (for the change in the loop's upper bound). It basically guards against

* assume `%stride1` dep is safe (both) and `%stride2` is unsafe for any VF
* Process one safe dep, report as safe,
* process unsafe dep, limit `MaxVFWithoutSLForwardIssuesPowerOf2` to `1`
* Process another safe dep, if the loop upper bound is still `MaxVFWithoutSLForwardIssuesPowerOf2`, then we'd incorrectly report that safe dep as unsafe as well.

I'm adding "safe" deps on both sides so that the test would actually "work" for whatever the iteration order is (because my first attempt at writing it I've actually chosen the wrong one).



https://github.com/llvm/llvm-project/pull/213722