[PATCH] D136918: [AMDGPU] Scheduler: fix RP calculation for a MBB with one successor

Thu Nov 3 05:59:43 PDT 2022

vpykhtin added a comment.

In D136918#3902766 <https://reviews.llvm.org/D136918#3902766>, @arsenm wrote:

> In D136918#3902686 <https://reviews.llvm.org/D136918#3902686>, @arsenm wrote:
>
>> In D136918#3902682 <https://reviews.llvm.org/D136918#3902682>, @vpykhtin wrote:
>>
>>>> (Liveness is really "due to" uses, not defs.)
>>>
>>> This sounds like if we need to track RP upward from uses to defs.
>>
>> Yes, this is how things are supposed to work. There's a lot of infrastructure that needs to be inverted
>
> In particular RegScavenging is used in the forward direction and relies on accurate kill flags. stepBackwards doesn't, but most everything needs to move to use that instead

We have GCNUpwardRPTracker which can be used at in AMDGPU code. It was written with the idea of making passes over tentative schedules, which aren't yet implemented on instructions yet.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D136918/new/

https://reviews.llvm.org/D136918