Experimental FA scheduling.

WIP

Changes
- Custom XDL resource tracking. Assume first 8 cycles of 8-pass mfma stalls the whole pipeline. Prefer MFMA if XDL is free, see `tryXDL` in `tryCand` for heuristic priorities.
- Fix exp latency.
- Force topdown scheduling in marked regions.
- Add `GCNPostSchedStrategy` and `tryCand` overrides for both the default schedulers.
- PostRA prefer MFMA with more successors.

Author: kerbowa

llvm/lib/Target/AMDGPU/AMDGPUIGroupLP.cpp

@@ -2669,6 +2669,12 @@ void IGroupLPDAGMutation::initSchedGroupBarrierPipelineStage(
 bool IGroupLPDAGMutation::initIGLPOpt(SUnit &SU) {
   IGLPStrategyID StrategyID =
       (IGLPStrategyID)SU.getInstr()->getOperand(0).getImm();
+  if (StrategyID == 10) {
+    for (auto &SU : DAG->SUnits)
+      SU.hasReservedResource = false;
+
+    return false;
+  }
   auto S = createIGLPStrategy(StrategyID, DAG, TII);
   if (!S->shouldApplyStrategy(DAG, Phase))
     return false;

llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

@@ -1118,7 +1118,7 @@ GCNTargetMachine::createMachineScheduler(MachineSchedContext *C) const {
 ScheduleDAGInstrs *
 GCNTargetMachine::createPostMachineScheduler(MachineSchedContext *C) const {
   ScheduleDAGMI *DAG =
-      new GCNPostScheduleDAGMILive(C, std::make_unique<PostGenericScheduler>(C),
+      new GCNPostScheduleDAGMILive(C, std::make_unique<GCNPostSchedStrategy>(C),
                                    /*RemoveKillFlags=*/true);
   const GCNSubtarget &ST = C->MF->getSubtarget<GCNSubtarget>();
   DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));