[PATCH] D84779: [AMDGPU] Add amdgpu specific loop threshold metadata

[Tim Corringham via Phabricator via llvm-commits]

timcorringham added a comment.

To answer Matt's question, there isn't currently metadata to set the threshold - there is a function attribute that I added that allows the default to be set which is similar, but that of course uses the same value for all loops within a function. However, that doesn't give the per-loop control we have since discovered we need for graphics applications.

For Johannes concerns, I could undertake to remove the function attribute in a future change if that helps? I would have to modify our front-end to use this new metadata rather than the function attribute first, but then the function attribute could be removed from LLVM (although that is a single line of code). As the new metadata is amdgpu specific it has no impact on the generic loop metadata, and is only used by AMDGPUTTIImpl::getUnrollingPreferences().


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D84779/new/

https://reviews.llvm.org/D84779