[PATCH] D41436: [X86][AVX512] Enable variable shuffle combining by default on AVX512 targets

[Simon Pilgrim via Phabricator via llvm-commits]

RKSimon added inline comments.


================
Comment at: lib/Target/X86/X86ISelLowering.cpp:28592
+  int VariableShuffleDepth =
+      Subtarget.hasFastVariableShuffle() || Subtarget.hasAVX512() ? 2 : 3;
   bool AllowVariableMask = (Depth >= VariableShuffleDepth) || HasVariableMask;
----------------
craig.topper wrote:
> Should we just do the OR in the implementation of hasFastVariableShuffle()? That way if we use this method in other places they'll have consistent behavior.
I'm happy to do that if the AVX512 gurus agree, or if they want to go all the way and make FeatureAVX512 inherit FeatureFastVariableShuffle - but AFAICT we don't tend to include the fast/slow 'characteristic' features like that with 'hardware' features.


Repository:
  rL LLVM

https://reviews.llvm.org/D41436