[all-commits] [llvm/llvm-project] 69a3ac: [Test] We can benefit from pipelining of ymm load/...

[Max Kazantsev via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 69a3acffdf1b3f5fc040aaeafc1c77588a607d1a
      https://github.com/llvm/llvm-project/commit/69a3acffdf1b3f5fc040aaeafc1c77588a607d1a
  Author: Max Kazantsev <mkazantsev at azul.com>
  Date:   2021-07-15 (Thu, 15 Jul 2021)

  Changed paths:
    A llvm/test/CodeGen/X86/ymm-ordering.ll

  Log Message:
  -----------
  [Test] We can benefit from pipelining of ymm load/stores

This patch demonstrates a scenario when we need to load/store a single
64-byte value, which is done by 2 ymm loads and stores in AVX. The current
codegen choses the following sequence:

  load ymm0
  load ymm1
  store ymm1
  store ymm0

If we instead stored ymm0 before ymm1, we could execute 2nd load and 1st store
in parallel.