[llvm] [AMDGPU] Examine instructions in pending queues during scheduling (PR #147653)

[Jeffrey Byrnes via llvm-commits]

https://github.com/jrbyrnes approved this pull request.

LGTM - please wait a little while to see if Jay still has an outstanding concern.

It feels like theres a deeper flaw in the interaction between the schedmodel and the scheduler. Probably the pending queue should be replaced with a stall / resource availability heuristic. 

That said, it seems like this PR bridges the gap in that the scheduler makes sensible decisions without the deeper rework. All the alternate ways I could think of to do this broke the intent of various other mechanisms.

https://github.com/llvm/llvm-project/pull/147653