[llvm-branch-commits] [llvm] DAG: Avoid forming shufflevector from a single extract_vector_elt (PR #122672)

[Matt Arsenault via llvm-branch-commits]

arsenm wrote:

> I don't think the heuristic here is quite what you want. I believe this heuristic disables both of the following cases:
> 
> * BuildVector w/one non-zero non-undef element
> * BuildVector w/one non-zero non-undef source, repeated 100 times (i.e. splat or select of two splats)
> 
> Disabling the former seems defensible, doing so for the second less so.

Depends if isExtractVecEltCheap is true or not for the non-zero index.

> * BuildVector w/one non-zero non-undef source, repeated 100 times (i.e. splat or select of two splats)

I don't follow, this is a 2 element vector, how can you have 100 variants?

>  If the target isn't optimally lowering the splat or select of splat case in the shuffle lowering, maybe we should just adjust the target lowering to do so?t

It's not a lowering issue, it's the effect on every other combine. We'd have to special case 1 element + 1 undef shuffles everywhere we handle extract_vector_elt now, which is just excessive complexity. #122671 is almost an alternative in one instance, but still shows expanding complexity of handling this edge case. 

https://github.com/llvm/llvm-project/pull/122672