[llvm] [LV][VPlan] Add initial support for CSA vectorization (PR #121222)

[via llvm-commits]

ayalz wrote:

> @ayalz thanks for the review!
> 
> > Continuing with this argument, better to also sink the loading of a[i] to after the loop, instead of loading vectorized va with mask inside the loop?
> 
> I agree in the example you bring up in your example.

This is "The motivating example" rather than one I brought up.

> But consider a case where a[i] is needed unconditionally in the loop. For example if `cond[i]` is replaced with `a[i]`. Then doing it after the loop is doing duplicate loads. Additionally, consider that `t = f(a, b, c)`, then we need to keep track of `a, b, c` live out of the loop, which may be tricky.

If `a, b, c` are loop invariant they should be available just the same right after the loop?
There is indeed potential for reuse inside the vectorized loop, but even then maintaining a vector `va` live across iterations and live out of the loop may incur a cost (register pressure?).
Would be good to reason about concrete examples along with their associated costs.
> 
> I think there is some room to do such an optimization, but I prefer to leave this as future work. WDYT?
> 
> > The reduction becomes a "FindLast" reduction once this function is sunk. Sound reasonable?

I think such patterns are essentially extensions of "FindLast" reduction and should be developed as such, rather than being considered distinct unrelated patterns.
> 
> According to [#106560 (comment)](https://github.com/llvm/llvm-project/pull/106560#issuecomment-2419166743), there is no plan to support `FindLast` at the moment. Above I describe a scenario where it is not beneficial to do this sinking, so we would not be able to rely on `FindLast` in all instances.

Note that by replacing `t = a[i]` with `t = i` the above "CSA" example becomes a `FindLast` one, hopefully suggesting such a supportive plan?
> 
> > Finally, note that LV already supports some sort of "CSA", introduced by @annamthomas some years ago in https://reviews.llvm.org/D52656; see test variant_val_store_to_inv_address_conditional in llvm/test/Transforms/LoopVectorize/X86/invariant-store-vectorization.ll - a conditional store to the same address could be converted into a conditional scalar assignment coupled with sinking the conditional store to after the loop.
> 
> This sounds like a good idea for future improvement of these loops.

 -- one which would also seem to benefit from sinking `f(i)=c[i]` of `%t4` to after the loop, based on `FindLast` reduction.

https://github.com/llvm/llvm-project/pull/121222