[llvm] [LAA] Hoist setting condition for RT-checks (PR #128045)

[Ramkumar Ramachandra via llvm-commits]

artagnon wrote:

> From what I can tell, we can have multiple `DepType::Unknown` dependences that, by themselves, do not trigger the retry with runtime memory checks, and therefore block vectorisation. But once a single `FoundNonConstantDistanceDependence` is found, the analysis seems to switch to the **"LAA: Retrying with memory checks"** path and allows all these dependences to be handled via runtime checks.

Yes, this is correct.

> Is this the intended behaviour?

Using a single `FoundNonConstantDistanceDependence` is not ideal, to begin with: it is not exactly the most principled condition to retry with runtime checks. But yes, that is the intended side-effect of the patch, that was originally unintended.

> If so, is there a rationale for treating `DepType::Unknown` differently depending on the presence of a non-constant distance dependence?

As far as I can tell, there is no rationale for this, although I'm unable to dissect the original intent of the code from the history. It should be noted that `FoundNonConstantDistanceDependence` is a misnomer:

```cpp
    // TODO: Relax requirement that there is a common unscaled stride to retry
    // with non-constant distance dependencies.
```

It is actually an ad-hoc condition for retrying with runtime checks, which is set when either there is a non-constant dependence distance, or if there is no common unscaled stride (what is the semantic meaning of the latter?).

Thanks for your write-up! I couldn't have written it more clearly myself, and I will adapt your write-up for the commit message.

https://github.com/llvm/llvm-project/pull/128045