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

[Jeffrey Byrnes via llvm-commits]

================
@@ -414,6 +508,64 @@ void GCNUpwardRPTracker::recede(const MachineInstr &MI) {
   assert(CurPressure == getRegPressure(*MRI, LiveRegs));
 }
 
+void GCNUpwardRPTracker::bumpUpwardPressure(const MachineInstr *MI,
----------------
jrbyrnes wrote:

> It's not really clear what this method does to the RP tracker object - modifies CurPressure and MaxPressure? What is the state of the object after the method call? 

This method modifies CurPressure and MaxPressure w.r.t the uses/defs in MI. It is meant to be used for RP Speculation, so MI is not required to be the next instruction in the LIS. Due to this requirement, we can't use recede which uses the implied program order in LIS 
to calculate live lane masks.

> From your code I only see one call against a temp object which may indicate that the state of a temp becomes undefined after a call. If so why don't make a non-method function returning max pressure?

Since we are calculting the impact on RP for a SchedCandidate, we don't want to commit the new pressure state to the tracker (hence why we use the temp object). However, calculating RP w.r.t some instruction feels like the responsibility of the tracker, and keeping it as a method is consistent with the generic implementation.

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