[PATCH] D112085: [DAGCombiner] fold bit-hack form of usubsat

[Sanjay Patel via Phabricator via llvm-commits]

spatel created this revision.
spatel added reviewers: arsenm, craig.topper, lebedev.ri, pengfei, RKSimon.
Herald added subscribers: ecnelises, kerbowa, hiraditya, nhaehnle, jvesely, mcrosier.
spatel requested review of this revision.
Herald added a subscriber: wdng.
Herald added a project: LLVM.

(i8 X ^ 128) & (i8 X s>> 7) --> usubsat X, 128

I haven't found a generalization of this identity:
https://alive2.llvm.org/ce/z/_sriEQ

Note: I was actually looking at the first form of the pattern in that link, but that's part of a long chain of potential missed transforms in codegen and IR....that I hope ends here!

The predicates for when this is profitable are a bit tricky. This version of the patch allows multi-use but excludes `custom` lowering (as opposed to `legal` only).

On x86, for example, we have custom lowering for some vector types and that uses umax and sub. So if we want to enable that fold, we need add use checks to avoid regressions. Even here, we see that legacy SSE code may need an extra instruction with extra uses, so I can add a constraint to make this more conservative if that seems better.


https://reviews.llvm.org/D112085

Files:
  llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
  llvm/test/CodeGen/AMDGPU/usubsat.ll
  llvm/test/CodeGen/X86/psubus.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D112085.380722.patch
Type: text/x-patch
Size: 6038 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20211019/d45b7f6b/attachment.bin>