[all-commits] [llvm/llvm-project] db231e: [InstCombine] fold fake vector extract to shift+trunc

[RotateRight via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: db231ebdb07f38f8b5d92311a4dc89e63c5a9d05
      https://github.com/llvm/llvm-project/commit/db231ebdb07f38f8b5d92311a4dc89e63c5a9d05
  Author: Sanjay Patel <spatel at rotateright.com>
  Date:   2021-10-06 (Wed, 06 Oct 2021)

  Changed paths:
    M llvm/lib/Transforms/InstCombine/InstCombineInternal.h
    M llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
    M llvm/test/Transforms/InstCombine/X86/x86-avx512-inseltpoison.ll
    M llvm/test/Transforms/InstCombine/X86/x86-avx512.ll
    M llvm/test/Transforms/InstCombine/extractelement.ll

  Log Message:
  -----------
  [InstCombine] fold fake vector extract to shift+trunc

We already handle more complicated cases like:
extelt (bitcast (inselt poison, X, 0)) --> trunc (lshr X)

But we missed this simpler pattern:
https://alive2.llvm.org/ce/z/D55h64 / https://alive2.llvm.org/ce/z/GKzzRq

This is part of solving:
https://llvm.org/PR52057

I made the transform depend on legal/desirable int type to avoid creating
a shift of an illegal type (for example i128). I'm not sure if that
restriction is actually necessary, but we can change that as a follow-up
if the backend can deal with integer ops on too-wide illegal types.

The pile of AVX512 test changes are all neutral AFAICT - the x86 backend
seems to know how to turn that into the expected "kmov" instructions.

Differential Revision: https://reviews.llvm.org/D111082