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

[Lucas Ramirez via llvm-commits]

lucas-rami wrote:

> Testcase where this makes a difference?

I spent some time trying to come up with one but didn't manage to. This change is motivated by ongoing pre-RA rematerialization work on the AMDGPU backend (see #125885) which requires that we are able to track regions with a single MI. The current implementation follows a different logic to identify rematerializable instructions, and I can't manage to make this fail/succeed depending on whether single-MI regions are filtered out/in.

> Am I looking at the right diff here? The only code change shown here is the removal of I == std::prev(RegionEnd).

A few unit tests across various targets were also updated manually. 

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