[llvm] [AMDGPU] Add block carried latency to CoExecSched (PR #187413)

[Matt Arsenault via llvm-commits]

================
@@ -216,16 +236,77 @@ void CandidateHeuristics::initialize(ScheduleDAGMI *SchedDAG,
   HWUInfo[(int)InstructionFlavor::MultiCycleVALU].setProducesCoexecWindow(true);
   HWUInfo[(int)InstructionFlavor::TRANS].setProducesCoexecWindow(true);
 
-  collectHWUIPressure();
+  collectRegionSummary();
+}
+
+unsigned CandidateHeuristics::getCarriedLatency(SUnit *SU) {
+  MachineInstr *MI = SU->getInstr();
+  unsigned CarriedLatency = 0;
+  for (auto &Op : MI->operands()) {
+    if (!Op.isReg())
+      continue;
+    if (!Op.isUse())
+      continue;
----------------
arsenm wrote:

all_uses?

https://github.com/llvm/llvm-project/pull/187413