[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;
+    auto Reg = Op.getReg();
+    if (!Reg.isVirtual())
+      continue;
+
+    for (auto &Def : DAG->MRI.def_instructions(Reg)) {
+      // We don't have the proper modelling to accurately measure all carried
+      // latency. Just try to measure carried latency for long latency loads to
+      // avoid long stalls.
+      if (!Def.mayLoad())
+        continue;
+
+      unsigned Latency = getHWUICyclesForMI(&Def);
+
+      // Load is carried across block
+      if (Def.getParent() != MI->getParent()) {
+        bool FoundUseInDefBlock = false;
+        for (auto &Use : DAG->MRI.use_nodbg_instructions(Reg)) {
----------------
arsenm wrote:

No auto 

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