[llvm] [MachineScheduler][AMDGPU] Allow scheduling of single-MI regions (PR #128739)

[Lucas Ramirez via llvm-commits]

lucas-rami wrote:

> Can we not do this on targets that don't need it?

I am ok reintroducing the flag to only conditionally do this for targets that care (off-by-default).

> I thought the scheduler only ever built the DAG for one region at a time. What's the mechanism for having DAG for multiple regions live at the same time?

The AMDGPU backend doesn't actually schedule in the `ScheduleDAGInstrs::schedule` hook, it only collects the regions there. Scheduling only happens in the `ScheduleDAGInstrs::finalizeSchedule` hook, once all regions have been collected. In the end it still schedule regions one at a time, but there may be some instruction movement across regions between scheduling stages.

https://github.com/llvm/llvm-project/blob/main/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp#L793-L797

https://github.com/llvm/llvm-project/pull/128739