[PATCH] D134605: [SLP] Add ScalarizationOverheadBuilder helper to track vector extractions

[Florian Hahn via Phabricator via llvm-commits]

fhahn added a comment.

Here's an example where this patch enables additional vectorization that harms performance: https://llvm.godbolt.org/z/KE1jv5sKs

I think there are the following issues that make this unprofitable:

1. vector version cannot use FMADD while scalar code can, so the vector code doesn't save that many math instructions (this is a known issue but with the patch it seems to trigger more frequently)
2. the vector code lengths the dependency chains in the loop, limiting out-of-order execution.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D134605/new/

https://reviews.llvm.org/D134605