[PATCH] D147507: [X86] Disable masked UNPCKLPD/UNPCKHPD -> SHUFPS transformation

[Simon Pilgrim via Phabricator via llvm-commits]

RKSimon added a comment.

In D147507#4243369 <https://reviews.llvm.org/D147507#4243369>, @goldstein.w.n wrote:

> In D147507#4242658 <https://reviews.llvm.org/D147507#4242658>, @RKSimon wrote:
>
>> @goldstein.w.n You might want to investigate if its worth using VSHUFPD instead?
>
> @pengfei and @RKSimon what about using `{VP}UNPCK{L|H}QDQ{...}`? I tested on ICL and didn't see any domain penalty. Wasn't able to find the hardware to test hsw/skl/.... and not sure if it falls under no-shuffle `hasNoDomainDelayShuffle` or something else but it is the ideal replacement both from perf and codesize perspective.

Why not both? We can try with VSHUFPD to see if it has better scheduling, else try integer unpack if we don't have a domain penalty


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D147507/new/

https://reviews.llvm.org/D147507