[llvm] [LAA] Avoid unprofitable unit-stride versioning (PR #225331)

[Boyao Wang via llvm-commits]

BoyaoWang430 wrote:

> > [dtcxzyw/llvm-opt-benchmark-nightly#1477](https://github.com/dtcxzyw/llvm-opt-benchmark-nightly/pull/1477) seems to show a significant decrease in loop loads eliminated. Presumably this is coming from the LoopLoadElimination pass? It would be good to have some tests similar to test/Transforms/LoopLoadElim/unknown-stride-known-dep.ll that expose the cases this PR is changing.
> 
> I added a few representative tests for the LLE differences. Most of the delta comes from unit-stride clones that are unreachable after versioning.

I checked the affected LLE cases. The original loops can run more than once, but their strides can never be 1. The 20 reductions in JPEG/PNG come from unreachable unit-stride clones.

https://github.com/llvm/llvm-project/pull/225331