[llvm] [VPlan] Don't convert widen recipes to VP intrinsics in EVL transform (PR #126177)

[Luke Lau via llvm-commits]

lukel97 wrote:

> But overall hardware that have vl-dependent RVV implementation will benefit of having proper EVL set.

@npanchen As Alexey said the RISCVVLOptimizer pass does exactly that, see #108640. It's more general than the EVL on VP intrinsics as it's able to reason about demanded elements, and with this patch it looks like it's actually able to reduce VL in a few more places due to there being less constraints.

> passes such InstCombine need to be updated.

Ideally yes, but I think this is a large amount of work that's unlikely to be finished any time soon. The goal here is to enable EVL tail folding within a reasonable time frame, and this addresses the codegen issues due to the lack of VP support in InstCombine etc.

https://github.com/llvm/llvm-project/pull/126177