[llvm] [RFC][VPlan] Simplify reverse patterns in EVL-based vectorization (PR #199510)

[Luke Lau via llvm-commits]

lukel97 wrote:

> > > This patch is not opposed to separating permutation elimination into a standalone transformation; fully eliminating these combinations requires iterating through the simplification process. However, I am not yet certain if a fixed-point iteration of simplifyRecipes would severely impact compile-time.
> > 
> > 
> > Tagging #199234 in case you missed it. In that PR it looks like it's enough to just do a single pass sinking reverses as far as possible, before the `reverse(reverse(x)) -> x` fold.
> 
> That makes sense. But if that's the case, wouldn't implementing it directly in simplifyRecipes be sufficient?

I think so, the exact location isn't too important as long as it happens before the `reverse(reverse(x)) -> x` fold.

https://github.com/llvm/llvm-project/pull/199510