llvm-project

7274ae97 - [AMDGPU][Scheduler] Simplify scheduling revert logic (#177203)

[AMDGPU][Scheduler] Simplify scheduling revert logic (#177203) When scheduling must be reverted for a region, the current implementation re-orders non-debug instructions and debug instructions separately; the former in a first pass and the latter in a second pass handled by a generic machine scheduler helper whose state is tied to the current region being scheduled, in turns limiting the revert logic to only work on the active scheduling region. This makes the revert logic work in a single pass for all MIs, and removes the restriction that it works exclusively on the active scheduling region. The latter enables future use cases such as reverting scheduling of multiple regions at once.