[PATCH] D141653: [X86] Improve instruction ordering of constant `srl/shl` with `and` to get better and-masks

[Noah Goldstein via Phabricator via llvm-commits]

goldstein.w.n added inline comments.


================
Comment at: llvm/test/CodeGen/X86/btc_bts_btr.ll:988
 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT:    andb $7, %cl
 ; X86-NEXT:    shlb $2, %cl
----------------
pengfei wrote:
> One more `andb`
> One more `andb`

Sorry, didn't see earlier. Think this breaks a few combine patterns that probably relied on semi-canonicalizarion of (and (shift x, y), z) instead of the other way around.

Fixing the cases caught here is doable, but is it possible to maybe move this to later in the process? Currently it's guarded behind AfterLegalize (for the exact reason of not breaking other patterns), but is there a higher level or different pass this might be better placed in?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D141653/new/

https://reviews.llvm.org/D141653