[PATCH] D134723: [AMDGPU] Set memory bound occupancy based on addressable VGPRs

[Carl Ritson via Phabricator via llvm-commits]

critson added a comment.

So this was based on dealing with the temporary 128 VGPR limit on GFX11, which has now been removed.
On that basis I will probably abandon this revision; however, I do still have concerns about this magic number 4.

Because of this number the scheduler would pick a schedule which used more than 128 VGPRs and spill, when in fact a schedule without spilling was possible.
Essentially when memory bound heuristic has been triggered the scheduler behaves in a manner which artificially reduces occupancy to this value of 4.
With 256 VGPRs addressable then 4 does work as a minimum without spilling for all hardware variants.
Although logically the value is going to be 6 for GFX11 with extra VGPRs in Wave32.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D134723/new/

https://reviews.llvm.org/D134723