[PATCH] D72325: [AMDGPU] Fix cluster size threshold calculation

[Nicolai Hähnle via Phabricator via llvm-commits]

nhaehnle added a comment.

Don't we *want* clusters that large, and even larger?

Consider some code that loads an array-of-structures (AoS). We really want to cluster that as aggressively as possible, to increase the chance of lowest-level cache hits on successive instructions? I would say the method is *very* inexact :)

The comment talks about not wanting to drive register pressure up too much. That's a legitimate concern, but this approach here seems to be quite wrong to me. The scheduler ought to track register pressure properly, and that's where the knowledge about whether to break clusters based on register pressure should be.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D72325/new/

https://reviews.llvm.org/D72325