[PATCH] D123409: [AArch64] Use PerfectShuffle costs in AArch64TTIImpl::getShuffleCost

Fri May 13 08:14:58 PDT 2022

fhahn added a comment.

In D123409#3496279 <https://reviews.llvm.org/D123409#3496279>, @dmgreen wrote:

> Did you ever manage to come up with a reproducer? I'm hope this new cost model is generally more accurate, but you know cost modelling.. The codegen might be off or their might be any number of second order effects going wrong. Let me know.

The issue was that after this patch some code got vectorized when it wasn't profitable, but it looked like a general SLP issue. Previously it just didn't get vectorized because some ridiculously high costs where used for some shuffles. It looks like D115750 <https://reviews.llvm.org/D115750> fixed the underlying issue and the code is back to not getting vectorized and original performance is restored :)

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D123409/new/

https://reviews.llvm.org/D123409