[PATCH] D89566: [LV] Epilogue Vectorization with Optimal Control Flow

Mon Nov 9 17:06:44 PST 2020

bmahjour added a comment.

In D89566#2383507 <https://reviews.llvm.org/D89566#2383507>, @SjoerdMeijer wrote:

> My observations here are:
>
> - Performance of both patches for SPEC are the same. The advantage of D88819 <https://reviews.llvm.org/D88819> on ThunderX2 might be a coincidence (knock on effect, e.g. (loop) alignment) or some micro-architecture reason.
> - This patch D89566 <https://reviews.llvm.org/D89566> can handle more loops, but it is not reflected in the SPEC numbers (cold/not executed code?).
> - The beauty of D88819 <https://reviews.llvm.org/D88819> is that the changes are very minimal, but looks like it's worth spending extra lines of code here in D89566 <https://reviews.llvm.org/D89566> to helps in vectorising more.
>
> To me, that shows there's more potential with this patch.
> I don't think I can be the arbiter in this, so it's best if other reviewers comment too.

As you pointed out, the reason D89566 <https://reviews.llvm.org/D89566> causes no additional gain despite transforming more loops, could be because the loops are cold or the main vectorized loop dominates the execution profile. These issues can probably be addressed with an improved cost-model or with PGO data.  It's important to be able to have the infrastructure to transform these loops for when the cost-model gets an upgrade.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D89566/new/

https://reviews.llvm.org/D89566