[PATCH] D109726: [X86] Improve `matchBinaryShuffle()`'s `BLEND` lowering with per-element all-zero/all-ones knowledge

[Simon Pilgrim via Phabricator via llvm-commits]

RKSimon added a comment.

In D109726#2999993 <https://reviews.llvm.org/D109726#2999993>, @lebedev.ri wrote:

> In D109726#2999987 <https://reviews.llvm.org/D109726#2999987>, @RKSimon wrote:
>
>> Why couldn't this be put in more generic ISD::OR combines somewhere? Possibly TargetLowering::SimplifyDemandedVectorElts?
>
> What do you mean by 'it'? `blend`->`or` lowering?

The insert tests look like we're just failing to recognize that we're OR'ing all-ones elements, which should allow the SimplifyDemanded call to realise that we're going to completely overwrite those elements, so we don't need to mask them. Its most of the logic you've written but it feels like it should be somewhere more generic than inside x86 shuffle combining.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D109726/new/

https://reviews.llvm.org/D109726