Automerge: [AMDGPU] Ignore wavefront barrier latency during scheduling DAG mutation (#168500)

Do not add latency for wavefront and singlethread scope fences during
barrier latency DAG mutation.
These scopes do not typically introduce any latency and adjusting
schedules based on them significantly impacts latency hiding.

Author: perlfu

llvm/lib/Target/AMDGPU/AMDGPUBarrierLatency.cpp

@@ -27,8 +27,17 @@ using namespace llvm;
 namespace {
 
 class BarrierLatency : public ScheduleDAGMutation {
+private:
+  SmallSet<SyncScope::ID, 4> IgnoredScopes;
+
 public:
-  BarrierLatency() = default;
+  BarrierLatency(MachineFunction *MF) {
+    LLVMContext &Context = MF->getFunction().getContext();
+    IgnoredScopes.insert(SyncScope::SingleThread);
+    IgnoredScopes.insert(Context.getOrInsertSyncScopeID("wavefront"));
+    IgnoredScopes.insert(Context.getOrInsertSyncScopeID("wavefront-one-as"));
+    IgnoredScopes.insert(Context.getOrInsertSyncScopeID("singlethread-one-as"));
+  }
   void apply(ScheduleDAGInstrs *DAG) override;
 };
 
@@ -40,8 +49,11 @@ void BarrierLatency::apply(ScheduleDAGInstrs *DAG) {
       continue;
 
     // Update latency on barrier edges of ATOMIC_FENCE.
-    // We don't consider the scope of the fence or type of instruction
-    // involved in the barrier edge.
+    // Ignore scopes not expected to have any latency.
+    SyncScope::ID SSID = static_cast<SyncScope::ID>(MI->getOperand(1).getImm());
+    if (IgnoredScopes.contains(SSID))
+      continue;
+
     for (SDep &PredDep : SU.Preds) {
       if (!PredDep.isBarrier())
         continue;
@@ -68,6 +80,6 @@ void BarrierLatency::apply(ScheduleDAGInstrs *DAG) {
 } // end namespace
 
 std::unique_ptr<ScheduleDAGMutation>
-llvm::createAMDGPUBarrierLatencyDAGMutation() {
-  return std::make_unique<BarrierLatency>();
+llvm::createAMDGPUBarrierLatencyDAGMutation(MachineFunction *MF) {
+  return std::make_unique<BarrierLatency>(MF);
 }

llvm/lib/Target/AMDGPU/AMDGPUBarrierLatency.h

@@ -14,7 +14,10 @@
 
 namespace llvm {
 
-std::unique_ptr<ScheduleDAGMutation> createAMDGPUBarrierLatencyDAGMutation();
+class MachineFunction;
+
+std::unique_ptr<ScheduleDAGMutation>
+createAMDGPUBarrierLatencyDAGMutation(MachineFunction *MF);
 
 } // namespace llvm
 

llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

@@ -647,7 +647,7 @@ createGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) {
   DAG->addMutation(createIGroupLPDAGMutation(AMDGPU::SchedulingPhase::Initial));
   DAG->addMutation(createAMDGPUMacroFusionDAGMutation());
   DAG->addMutation(createAMDGPUExportClusteringDAGMutation());
-  DAG->addMutation(createAMDGPUBarrierLatencyDAGMutation());
+  DAG->addMutation(createAMDGPUBarrierLatencyDAGMutation(C->MF));
   return DAG;
 }
 
@@ -668,7 +668,7 @@ createGCNMaxMemoryClauseMachineScheduler(MachineSchedContext *C) {
   if (ST.shouldClusterStores())
     DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
   DAG->addMutation(createAMDGPUExportClusteringDAGMutation());
-  DAG->addMutation(createAMDGPUBarrierLatencyDAGMutation());
+  DAG->addMutation(createAMDGPUBarrierLatencyDAGMutation(C->MF));
   return DAG;
 }
 
@@ -1209,7 +1209,7 @@ GCNTargetMachine::createPostMachineScheduler(MachineSchedContext *C) const {
       EnableVOPD)
     DAG->addMutation(createVOPDPairingMutation());
   DAG->addMutation(createAMDGPUExportClusteringDAGMutation());
-  DAG->addMutation(createAMDGPUBarrierLatencyDAGMutation());
+  DAG->addMutation(createAMDGPUBarrierLatencyDAGMutation(C->MF));
   return DAG;
 }
 //===----------------------------------------------------------------------===//