[all-commits] [llvm/llvm-project] 775451: [AArch64] Cost-model vector splat LD1Rs to avoid u...

[sjoerdmeijer via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 775451b66a4c726f1c3925f74be8cefcb308e4f8
      https://github.com/llvm/llvm-project/commit/775451b66a4c726f1c3925f74be8cefcb308e4f8
  Author: Sjoerd Meijer <smeijer at nvidia.com>
  Date:   2023-03-13 (Mon, 13 Mar 2023)

  Changed paths:
    M llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
    M llvm/test/Analysis/CostModel/AArch64/shuffle-load.ll
    M llvm/test/Transforms/SLPVectorizer/AArch64/slp-fma-loss.ll

  Log Message:
  -----------
  [AArch64] Cost-model vector splat LD1Rs to avoid unprofitable SLP vectorisation

This slightly increases the costs of InsertElement instructions that are part
of a vector splat sequence, i.e. a load, InsertElement and a shuffle (load +
dup). The resulting LD1R is a high latency instruction, and this slight
increase in costs avoids SLP vectorisation for a couple of cases where this
isn't profitable.

Fixes: https://github.com/llvm/llvm-project/issues/61047

Differential Revision: https://reviews.llvm.org/D145578