[PATCH] D57156: [x86] narrow a shuffle that doesn't use or set any high elements

[Sanjay Patel via Phabricator via llvm-commits]

spatel added a comment.

In D57156#1370844 <https://reviews.llvm.org/D57156#1370844>, @RKSimon wrote:

> LGTM - thanks!
>
> In D57156#1370481 <https://reviews.llvm.org/D57156#1370481>, @spatel wrote:
>
> > Note: Tests with AVX512 and 512-bit vectors are not showing the same improvements as other targets/tests because we don't know how to collapse an (extract (extract V, 0), 0) into 1 extract node yet. I'm imagining that patch is similar to the insert-of-insert that in D56604 <https://reviews.llvm.org/D56604>. Once we have that fold, we'll get several more narrowing of binops like we see here. But like this patch, that one also appears to expose a shuffle lowering hole, so it might need a preliminary patch to avoid a regression.
>
>
> Please can you raise a bug about this?


Yes - I still need to minimize that one. If I find an obvious fix, I'll post a patch.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D57156/new/

https://reviews.llvm.org/D57156