[PATCH] D42088: [x86] shrink 'and' immediate values by setting the high bits (PR35907)

[Sanjay Patel via Phabricator via llvm-commits]

spatel updated this revision to Diff 130411.
spatel added a comment.

Patch updated:

1. Use DAG.MaskedValueIsZero() to make the code cleaner.
2. Move the check for shrinkability ahead of the more expensive call to compute known bits.
3. Check for a -1 'and' mask. In this case, replace and return. Don't create an unnecessary constant and don't select out-of-turn. Allow the normal selection sequence to continue.

Thinking about this some more -- and as we mentioned in https://reviews.llvm.org/D42090 -- I don't think it's possible to guarantee that we never see a dead 'and' op here in selection. We don't run DAGCombine to fix-point, so nothing is certain.

So even though we could try to fix the cases seen in the regression tests upstream from here, it won't prevent needing to handle the case of a -1 mask. Thus, I've added code for that situation to avoid unexpected selection behavior.

We could still move this over to PreprocessISelDAG() if that seems better.


https://reviews.llvm.org/D42088

Files:
  lib/Target/X86/X86ISelDAGToDAG.cpp
  test/CodeGen/X86/and-encoding.ll
  test/CodeGen/X86/divide-by-constant.ll
  test/CodeGen/X86/known-bits.ll
  test/CodeGen/X86/popcnt.ll
  test/CodeGen/X86/pr33844.ll
  test/CodeGen/X86/shift-pair.ll
  test/CodeGen/X86/urem-i8-constant.ll
  test/CodeGen/X86/vector-sext.ll
  test/CodeGen/X86/win64_frame.ll
  test/CodeGen/X86/x86-64-baseptr.ll
  test/CodeGen/X86/zext-fold.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D42088.130411.patch
Type: text/x-patch
Size: 31927 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180118/8f5f75a2/attachment.bin>