[PATCH] D116191: [X86] isVectorShiftByScalarCheap - vXi8 select(shift(x,splat0),shift(x,splat1)) is better than shift(x,select(splat0,splat1))

[Phoebe Wang via Phabricator via llvm-commits]

pengfei accepted this revision.
pengfei added a comment.
This revision is now accepted and ready to land.

LGTM.



================
Comment at: llvm/lib/Target/X86/X86ISelLowering.cpp:32997-33001
   // XOP has v16i8/v8i16/v4i32/v2i64 variable vector shifts.
   // Splitting for v32i8/v16i16 on XOP+AVX2 targets is still preferred.
   if (Subtarget.hasXOP() &&
       (Bits == 8 || Bits == 16 || Bits == 32 || Bits == 64))
     return false;
----------------
RKSimon wrote:
> pengfei wrote:
> > I'm confused by the summary and the code here. Do we prefer to shifts on both has XOP and not? Why do we always return false for XOP?
> XOP has awesome per-element shift ops for all 128-bit vector types - so there's no need to prefer vector shifts by scalar amounts over per-element amounts.
Got it. Thanks!


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D116191/new/

https://reviews.llvm.org/D116191