[PATCH] D113827: [X86] Widen 128/256-bit VPTERNLOG patterns to 512-bit on non-VLX targets

[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/test/CodeGen/X86/combine-bitselect.ll:35-40
+; AVX512F-NEXT:    # kill: def $xmm1 killed $xmm1 def $zmm1
+; AVX512F-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
+; AVX512F-NEXT:    vmovdqa {{.*#+}} xmm2 = [18446744069414584319,18446744060824649725]
+; AVX512F-NEXT:    vpternlogq $216, %zmm2, %zmm1, %zmm0
+; AVX512F-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0
+; AVX512F-NEXT:    vzeroupper
----------------
RKSimon wrote:
> pengfei wrote:
> > I wonder if the left side code wins sometime, e.g. optsize?
> We've replaced 2 x 128-bit (folded) loads with 1 x 128-bit (unfolded) load.
> 
> The issue I was most concerned with is increased register pressure - this is what the broadcast code helps with, but even that is of low concern.
Agreed. Thanks Simon.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D113827/new/

https://reviews.llvm.org/D113827