[llvm] [AMDGPU] Model ordered XDL writes in expert scheduling (PR #218326)

[Jay Foad via llvm-commits]

================
@@ -670,8 +675,12 @@ class WaitcntBrackets {
 
   using CounterValueArray = std::array<unsigned, AMDGPU::NUM_INST_CNTS>;
 
+  std::optional<unsigned> getWaitForScore(AMDGPU::InstCounterType T,
+                                          unsigned Score) const;
   void determineWaitForScore(AMDGPU::InstCounterType T, unsigned Score,
                              AMDGPU::Waitcnt &Wait) const;
+  void determineWaitForVGPRScore(AMDGPU::InstCounterType T, VMEMID ID,
----------------
jayfoad wrote:

Is this saying that, for correctness, every VGPR wait _must_ go through determineWaitForVGPRScore instead of determineWaitForScore? Can you add an assert to enforce that?

https://github.com/llvm/llvm-project/pull/218326