[llvm] [AMDGPU] Optionally Use GCNRPTrackers during scheduling (PR #93090)

[Jeffrey Byrnes via llvm-commits]

jrbyrnes wrote:

Added several tests based on the more significant lit changes.

Since this is a change which impacts scheduling heuristics, the changes are a mix of regressions and improvements. I've analyzed and included all the significant regressions, and added comments / fixmes to expplain the situation. The causes are: 1. not tracking PhysRegs, 2. the scheduler impoving RP in initial stage and not triggering high-RP stage, and 3. the scheduler being overly sensitive to minor changes in RP, and not being willing to incur a small penalty to enable scheduling of RP reducing instructions.

These will be addresed in a follow-up PR which enables the GCN RP Tracker features by default.

I've also included several tests where things go as expected -- more accurate RP Tracking leads to reductions in RP and, ultimately, improving occupancy

https://github.com/llvm/llvm-project/pull/93090