+static void AddGlue(SDNode *N, SDValue Glue, bool AddGlue, SelectionDAG *DAG) {
+ SmallVector<EVT, 4> VTs;
+ SDNode *GlueDestNode = Glue.getNode();
+
+ // Don't add glue from a node to itself.
+ if (GlueDestNode == N) return;
+
+ // Don't add glue to something which already has glue.
+ if (N->getValueType(N->getNumValues() - 1) == MVT::Glue) return;
+
+ for (unsigned I = 0, E = N->getNumValues(); I != E; ++I)
+ VTs.push_back(N->getValueType(I));
+
+ if (AddGlue)
+ VTs.push_back(MVT::Glue);
+
+ SmallVector<SDValue, 4> Ops;
+ for (unsigned I = 0, E = N->getNumOperands(); I != E; ++I)
+ Ops.push_back(N->getOperand(I));
+
+ if (GlueDestNode)
+ Ops.push_back(Glue);
+
+ SDVTList VTList = DAG->getVTList(&VTs[0], VTs.size());
+ MachineSDNode::mmo_iterator Begin = 0, End = 0;
+ MachineSDNode *MN = dyn_cast<MachineSDNode>(N);
+
+ // Store memory references.
+ if (MN) {
+ Begin = MN->memoperands_begin();
+ End = MN->memoperands_end();
+ }
+
+ DAG->MorphNodeTo(N, N->getOpcode(), VTList, &Ops[0], Ops.size());
+
+ // Reset the memory references
+ if (MN)
+ MN->setMemRefs(Begin, End);
+}
+
+/// ClusterNeighboringLoads - Force nearby loads together by "gluing" them.
+/// This function finds loads of the same base and different offsets. If the
+/// offsets are not far apart (target specific), it add MVT::Glue inputs and
+/// outputs to ensure they are scheduled together and in order. This
+/// optimization may benefit some targets by improving cache locality.
+void ScheduleDAGSDNodes::ClusterNeighboringLoads(SDNode *Node) {
+ SDNode *Chain = 0;
+ unsigned NumOps = Node->getNumOperands();
+ if (Node->getOperand(NumOps-1).getValueType() == MVT::Other)
+ Chain = Node->getOperand(NumOps-1).getNode();
+ if (!Chain)
+ return;
+
+ // Look for other loads of the same chain. Find loads that are loading from
+ // the same base pointer and different offsets.
+ SmallPtrSet<SDNode*, 16> Visited;
+ SmallVector<int64_t, 4> Offsets;
+ DenseMap<long long, SDNode*> O2SMap; // Map from offset to SDNode.
+ bool Cluster = false;
+ SDNode *Base = Node;
+ for (SDNode::use_iterator I = Chain->use_begin(), E = Chain->use_end();
+ I != E; ++I) {
+ SDNode *User = *I;
+ if (User == Node || !Visited.insert(User))
+ continue;
+ int64_t Offset1, Offset2;
+ if (!TII->areLoadsFromSameBasePtr(Base, User, Offset1, Offset2) ||
+ Offset1 == Offset2)
+ // FIXME: Should be ok if they addresses are identical. But earlier
+ // optimizations really should have eliminated one of the loads.
+ continue;
+ if (O2SMap.insert(std::make_pair(Offset1, Base)).second)
+ Offsets.push_back(Offset1);
+ O2SMap.insert(std::make_pair(Offset2, User));
+ Offsets.push_back(Offset2);
+ if (Offset2 < Offset1)
+ Base = User;
+ Cluster = true;
+ }
+
+ if (!Cluster)
+ return;
+
+ // Sort them in increasing order.
+ std::sort(Offsets.begin(), Offsets.end());
+
+ // Check if the loads are close enough.
+ SmallVector<SDNode*, 4> Loads;
+ unsigned NumLoads = 0;
+ int64_t BaseOff = Offsets[0];
+ SDNode *BaseLoad = O2SMap[BaseOff];
+ Loads.push_back(BaseLoad);
+ for (unsigned i = 1, e = Offsets.size(); i != e; ++i) {
+ int64_t Offset = Offsets[i];
+ SDNode *Load = O2SMap[Offset];
+ if (!TII->shouldScheduleLoadsNear(BaseLoad, Load, BaseOff, Offset,NumLoads))
+ break; // Stop right here. Ignore loads that are further away.
+ Loads.push_back(Load);
+ ++NumLoads;
+ }
+
+ if (NumLoads == 0)
+ return;
+
+ // Cluster loads by adding MVT::Glue outputs and inputs. This also
+ // ensure they are scheduled in order of increasing addresses.
+ SDNode *Lead = Loads[0];
+ AddGlue(Lead, SDValue(0, 0), true, DAG);
+
+ SDValue InGlue = SDValue(Lead, Lead->getNumValues() - 1);
+ for (unsigned I = 1, E = Loads.size(); I != E; ++I) {
+ bool OutGlue = I < E - 1;
+ SDNode *Load = Loads[I];
+
+ AddGlue(Load, InGlue, OutGlue, DAG);
+
+ if (OutGlue)
+ InGlue = SDValue(Load, Load->getNumValues() - 1);
+
+ ++LoadsClustered;
+ }
+}
+
+/// ClusterNodes - Cluster certain nodes which should be scheduled together.
+///
+void ScheduleDAGSDNodes::ClusterNodes() {
+ for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(),
+ E = DAG->allnodes_end(); NI != E; ++NI) {
+ SDNode *Node = &*NI;
+ if (!Node || !Node->isMachineOpcode())
+ continue;
+
+ unsigned Opc = Node->getMachineOpcode();
+ const TargetInstrDesc &TID = TII->get(Opc);
+ if (TID.mayLoad())
+ // Cluster loads from "near" addresses into combined SUnits.
+ ClusterNeighboringLoads(Node);
+ }
+}
+