[PATCH] D26297: [AVX-512] Add support for lowering shuffles to VALIGND/VALIGNQ

[Simon Pilgrim via llvm-commits]

RKSimon added a comment.

Please can you add the new tests to trunk with the existing lowering?

> For 128-bit vectors, we favor VALIGND/Q over PALIGNR so that we naturely match the vector element size in case the shuffle is part of a masked operation

Would we be better off putting this logic into combineSelect() and keep to PALIGNR? Detect mask/maskz selects taking a bitcasted target shuffles/logicals/whatever and attempt to narrow/widen them back? Not certain if this would actually work in general, its just a thought.


https://reviews.llvm.org/D26297