[llvm] [LV] Fix incorrect cost kind in VPReplicateRecipe::computeCost (PR #153216)

[David Sherwood via llvm-commits]

david-arm wrote:

> > > Thanks. Would be great to have a test case for this. I can check if I can surface anything
> > 
> > 
> > I don't think this is even possible at the moment, because this will currently only apply if you build with `-Oz -fvectorize`, but you can't even write a test case for this because we hit the error at the end of `LoopVectorizationCostModel::computeMaxVF`:
> > ```
> >   reportVectorizationFailure(
> >       "Cannot optimize for size and vectorize at the same time.",
> > ```
> > 
> > 
> >     
> >       
> >     
> > 
> >       
> >     
> > 
> >     
> >   
> > I suppose that really means this patch is NFC and should be harmless. I just wanted to change the CostKind for completeness.
> 
> Hm, I think we should definitely hit VPReplicateRecipe::computeCost with different CostKinds, e.g. for this test https://github.com/llvm/llvm-project/blob/main/llvm/test/Transforms/LoopVectorize/AArch64/optsize_minsize.ll#L221
> 
> For Os/Oz, the requirement is that there is no scalar tail + no runtime checks IIRC, and the message is for that case.

Nope, we never create replicate recipes at all due to this from the debug output:

```
LV: Not considering vector loop of width 2 because it would cause replicated blocks to be generated, which isn't allowed when optimizing for size.
LV: Not considering vector loop of width 4 because it would cause replicated blocks to be generated, which isn't allowed when optimizing for size.
LV: Not considering vector loop of width 8 because it would cause replicated blocks to be generated, which isn't allowed when optimizing for size.
LV: Not considering vector loop of width 16 because it would cause replicated blocks to be generated, which isn't allowed when optimizing for size.
```

https://github.com/llvm/llvm-project/pull/153216