[llvm] [LV][EVL] Generate negative strided load/store for reversed load/store (PR #123608)

[Mel Chen via llvm-commits]

Mel-Chen wrote:

> > Currently, due to the existing dependency on VPWidenIntOrFpInductionRecipe, the VPlan for tail folding with EVL will be bailed out. I will work on eliminating this dependency to resolve the issue this week.
> 
> Is this related to #118638?

Yes, they are currently related, but ultimately, the strided memory pointer should not depend on VPWidenIntOrFpInductionRecipe, as strided accesses only require a scalar address, unlike gather/scatter, which require a vectorized address.

However, supporting VPWidenIntOrFpInductionRecipe for tail folding with EVL remains crucial. This would enable the vectorization of non-unit stride strided accesses under folding by EVL, which I believe would be a significant improvement.

I will start reviewing #118638 next week since this Friday is a holiday at my location:)


https://github.com/llvm/llvm-project/pull/123608