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

[Jeffrey Byrnes via llvm-commits]

jrbyrnes wrote:

I spent a while looking RP speculation in the face of dead lanes in superreg defs. In the current code (as propagated from the generic trackers), we basically just ignore these lanes when calcuting pressure impacts for speculative RP. However, this is inaccurate. For example, if we have an instruction which defines a vreg_128 and but only subregs 1 and 2 are used, then bumpDownwardPressure will speculate that this vgpr def increase VGPR pressure by 2. Since those dead regs are not needed after the instruction, this RP change is consistent with RAs view of conflicts for positions after the instruction. However, during the instruction (before the dead slot), we temporarily increase VGPR pressure by 4 to provide registers for those dead lanes. However, this may mislead the scheduler, particularly for bottom up scheduling in which defs reduce RP. Encoding this into RPTracker's CurPressure doesn't seem to make much sense, as this is used to model the pressure after the instruction (whether top-down or bottom-up direction). So, I think we should probably introduce a separate GCNRegPressure field in the trackers to model any temporary increases to RP, and the scheduler can use this to more accurately set heuristics. That said, this should probably be addressed in separate PR.

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