//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "pre-RA-sched"
-#include "llvm/CodeGen/ScheduleDAGSDNodes.h"
+#include "SDDbgValue.h"
+#include "ScheduleDAGSDNodes.h"
+#include "InstrEmitter.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetRegisterInfo.h"
+#include "llvm/Target/TargetSubtarget.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/Statistic.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
using namespace llvm;
-ScheduleDAGSDNodes::ScheduleDAGSDNodes(SelectionDAG *dag, MachineBasicBlock *bb,
- const TargetMachine &tm)
- : ScheduleDAG(dag, bb, tm) {
+STATISTIC(LoadsClustered, "Number of loads clustered together");
+
+ScheduleDAGSDNodes::ScheduleDAGSDNodes(MachineFunction &mf)
+ : ScheduleDAG(mf) {
+}
+
+/// Run - perform scheduling.
+///
+void ScheduleDAGSDNodes::Run(SelectionDAG *dag, MachineBasicBlock *bb,
+ MachineBasicBlock::iterator insertPos) {
+ DAG = dag;
+ ScheduleDAG::Run(bb, insertPos);
}
SUnit *ScheduleDAGSDNodes::Clone(SUnit *Old) {
SU->isTwoAddress = Old->isTwoAddress;
SU->isCommutable = Old->isCommutable;
SU->hasPhysRegDefs = Old->hasPhysRegDefs;
+ SU->hasPhysRegClobbers = Old->hasPhysRegClobbers;
+ Old->isCloned = true;
return SU;
}
/// CheckForPhysRegDependency - Check if the dependency between def and use of
/// a specified operand is a physical register dependency. If so, returns the
-/// register.
+/// register and the cost of copying the register.
static void CheckForPhysRegDependency(SDNode *Def, SDNode *User, unsigned Op,
const TargetRegisterInfo *TRI,
const TargetInstrInfo *TII,
- unsigned &PhysReg) {
+ unsigned &PhysReg, int &Cost) {
if (Op != 2 || User->getOpcode() != ISD::CopyToReg)
return;
if (Def->isMachineOpcode()) {
const TargetInstrDesc &II = TII->get(Def->getMachineOpcode());
if (ResNo >= II.getNumDefs() &&
- II.ImplicitDefs[ResNo - II.getNumDefs()] == Reg)
+ II.ImplicitDefs[ResNo - II.getNumDefs()] == Reg) {
PhysReg = Reg;
+ const TargetRegisterClass *RC =
+ TRI->getPhysicalRegisterRegClass(Reg, Def->getValueType(ResNo));
+ Cost = RC->getCopyCost();
+ }
+ }
+}
+
+static void AddFlags(SDNode *N, SDValue Flag, bool AddFlag,
+ SelectionDAG *DAG) {
+ SmallVector<EVT, 4> VTs;
+ for (unsigned i = 0, e = N->getNumValues(); i != e; ++i)
+ VTs.push_back(N->getValueType(i));
+ if (AddFlag)
+ VTs.push_back(MVT::Flag);
+ SmallVector<SDValue, 4> Ops;
+ for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
+ Ops.push_back(N->getOperand(i));
+ if (Flag.getNode())
+ Ops.push_back(Flag);
+ SDVTList VTList = DAG->getVTList(&VTs[0], VTs.size());
+ DAG->MorphNodeTo(N, N->getOpcode(), VTList, &Ops[0], Ops.size());
+}
+
+/// ClusterNeighboringLoads - Force nearby loads together by "flagging" them.
+/// This function finds loads of the same base and different offsets. If the
+/// offsets are not far apart (target specific), it add MVT::Flag 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() {
+ SmallPtrSet<SDNode*, 16> Visited;
+ SmallVector<int64_t, 4> Offsets;
+ DenseMap<long long, SDNode*> O2SMap; // Map from offset to SDNode.
+ 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())
+ continue;
+
+ SDNode *Chain = 0;
+ unsigned NumOps = Node->getNumOperands();
+ if (Node->getOperand(NumOps-1).getValueType() == MVT::Other)
+ Chain = Node->getOperand(NumOps-1).getNode();
+ if (!Chain)
+ continue;
+
+ // Look for other loads of the same chain. Find loads that are loading from
+ // the same base pointer and different offsets.
+ Visited.clear();
+ Offsets.clear();
+ O2SMap.clear();
+ bool Cluster = false;
+ SDNode *Base = Node;
+ int64_t BaseOffset;
+ 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;
+ BaseOffset = Offset2;
+ } else {
+ BaseOffset = Offset1;
+ }
+ Cluster = true;
+ }
+
+ if (!Cluster)
+ continue;
+
+ // 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)
+ continue;
+
+ // Cluster loads by adding MVT::Flag outputs and inputs. This also
+ // ensure they are scheduled in order of increasing addresses.
+ SDNode *Lead = Loads[0];
+ AddFlags(Lead, SDValue(0,0), true, DAG);
+ SDValue InFlag = SDValue(Lead, Lead->getNumValues()-1);
+ for (unsigned i = 1, e = Loads.size(); i != e; ++i) {
+ bool OutFlag = i < e-1;
+ SDNode *Load = Loads[i];
+ AddFlags(Load, InFlag, OutFlag, DAG);
+ if (OutFlag)
+ InFlag = SDValue(Load, Load->getNumValues()-1);
+ ++LoadsClustered;
+ }
}
}
// Check to see if the scheduler cares about latencies.
bool UnitLatencies = ForceUnitLatencies();
- for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(),
- E = DAG->allnodes_end(); NI != E; ++NI) {
+ // Add all nodes in depth first order.
+ SmallVector<SDNode*, 64> Worklist;
+ SmallPtrSet<SDNode*, 64> Visited;
+ Worklist.push_back(DAG->getRoot().getNode());
+ Visited.insert(DAG->getRoot().getNode());
+
+ while (!Worklist.empty()) {
+ SDNode *NI = Worklist.pop_back_val();
+
+ // Add all operands to the worklist unless they've already been added.
+ for (unsigned i = 0, e = NI->getNumOperands(); i != e; ++i)
+ if (Visited.insert(NI->getOperand(i).getNode()))
+ Worklist.push_back(NI->getOperand(i).getNode());
+
if (isPassiveNode(NI)) // Leaf node, e.g. a TargetImmediate.
continue;
// See if anything is flagged to this node, if so, add them to flagged
// nodes. Nodes can have at most one flag input and one flag output. Flags
- // are required the be the last operand and result of a node.
+ // are required to be the last operand and result of a node.
// Scan up to find flagged preds.
SDNode *N = NI;
- if (N->getNumOperands() &&
- N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag) {
- do {
- N = N->getOperand(N->getNumOperands()-1).getNode();
- assert(N->getNodeId() == -1 && "Node already inserted!");
- N->setNodeId(NodeSUnit->NodeNum);
- } while (N->getNumOperands() &&
- N->getOperand(N->getNumOperands()-1).getValueType()== MVT::Flag);
+ while (N->getNumOperands() &&
+ N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag) {
+ N = N->getOperand(N->getNumOperands()-1).getNode();
+ assert(N->getNodeId() == -1 && "Node already inserted!");
+ N->setNodeId(NodeSUnit->NodeNum);
}
// Scan down to find any flagged succs.
}
void ScheduleDAGSDNodes::AddSchedEdges() {
+ const TargetSubtarget &ST = TM.getSubtarget<TargetSubtarget>();
+
+ // Check to see if the scheduler cares about latencies.
+ bool UnitLatencies = ForceUnitLatencies();
+
// Pass 2: add the preds, succs, etc.
for (unsigned su = 0, e = SUnits.size(); su != e; ++su) {
SUnit *SU = &SUnits[su];
// Find all predecessors and successors of the group.
for (SDNode *N = SU->getNode(); N; N = N->getFlaggedNode()) {
if (N->isMachineOpcode() &&
- TII->get(N->getMachineOpcode()).getImplicitDefs() &&
- CountResults(N) > TII->get(N->getMachineOpcode()).getNumDefs())
- SU->hasPhysRegDefs = true;
+ TII->get(N->getMachineOpcode()).getImplicitDefs()) {
+ SU->hasPhysRegClobbers = true;
+ unsigned NumUsed = InstrEmitter::CountResults(N);
+ while (NumUsed != 0 && !N->hasAnyUseOfValue(NumUsed - 1))
+ --NumUsed; // Skip over unused values at the end.
+ if (NumUsed > TII->get(N->getMachineOpcode()).getNumDefs())
+ SU->hasPhysRegDefs = true;
+ }
for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
SDNode *OpN = N->getOperand(i).getNode();
assert(OpSU && "Node has no SUnit!");
if (OpSU == SU) continue; // In the same group.
- MVT OpVT = N->getOperand(i).getValueType();
+ EVT OpVT = N->getOperand(i).getValueType();
assert(OpVT != MVT::Flag && "Flagged nodes should be in same sunit!");
bool isChain = OpVT == MVT::Other;
unsigned PhysReg = 0;
+ int Cost = 1;
// Determine if this is a physical register dependency.
- CheckForPhysRegDependency(OpN, N, i, TRI, TII, PhysReg);
+ CheckForPhysRegDependency(OpN, N, i, TRI, TII, PhysReg, Cost);
assert((PhysReg == 0 || !isChain) &&
"Chain dependence via physreg data?");
- SU->addPred(SDep(OpSU, isChain ? SDep::Order : SDep::Data,
- OpSU->Latency, PhysReg));
+ // FIXME: See ScheduleDAGSDNodes::EmitCopyFromReg. For now, scheduler
+ // emits a copy from the physical register to a virtual register unless
+ // it requires a cross class copy (cost < 0). That means we are only
+ // treating "expensive to copy" register dependency as physical register
+ // dependency. This may change in the future though.
+ if (Cost >= 0)
+ PhysReg = 0;
+
+ const SDep& dep = SDep(OpSU, isChain ? SDep::Order : SDep::Data,
+ OpSU->Latency, PhysReg);
+ if (!isChain && !UnitLatencies) {
+ ComputeOperandLatency(OpSU, SU, (SDep &)dep);
+ ST.adjustSchedDependency(OpSU, SU, (SDep &)dep);
+ }
+
+ SU->addPred(dep);
}
}
}
/// are input. This SUnit graph is similar to the SelectionDAG, but
/// excludes nodes that aren't interesting to scheduling, and represents
/// flagged together nodes with a single SUnit.
-void ScheduleDAGSDNodes::BuildSchedGraph() {
+void ScheduleDAGSDNodes::BuildSchedGraph(AliasAnalysis *AA) {
+ // Cluster loads from "near" addresses into combined SUnits.
+ ClusterNeighboringLoads();
// Populate the SUnits array.
BuildSchedUnits();
// Compute all the scheduling dependencies between nodes.
// Compute the latency for the node. We use the sum of the latencies for
// all nodes flagged together into this SUnit.
SU->Latency = 0;
- bool SawMachineOpcode = false;
for (SDNode *N = SU->getNode(); N; N = N->getFlaggedNode())
if (N->isMachineOpcode()) {
- SawMachineOpcode = true;
- SU->Latency +=
- InstrItins.getLatency(TII->get(N->getMachineOpcode()).getSchedClass());
+ SU->Latency += InstrItins.
+ getStageLatency(TII->get(N->getMachineOpcode()).getSchedClass());
}
}
-/// CountResults - The results of target nodes have register or immediate
-/// operands first, then an optional chain, and optional flag operands (which do
-/// not go into the resulting MachineInstr).
-unsigned ScheduleDAGSDNodes::CountResults(SDNode *Node) {
- unsigned N = Node->getNumValues();
- while (N && Node->getValueType(N - 1) == MVT::Flag)
- --N;
- if (N && Node->getValueType(N - 1) == MVT::Other)
- --N; // Skip over chain result.
- return N;
-}
-
-/// CountOperands - The inputs to target nodes have any actual inputs first,
-/// followed by special operands that describe memory references, then an
-/// optional chain operand, then an optional flag operand. Compute the number
-/// of actual operands that will go into the resulting MachineInstr.
-unsigned ScheduleDAGSDNodes::CountOperands(SDNode *Node) {
- unsigned N = ComputeMemOperandsEnd(Node);
- while (N && isa<MemOperandSDNode>(Node->getOperand(N - 1).getNode()))
- --N; // Ignore MEMOPERAND nodes
- return N;
-}
-
-/// ComputeMemOperandsEnd - Find the index one past the last MemOperandSDNode
-/// operand
-unsigned ScheduleDAGSDNodes::ComputeMemOperandsEnd(SDNode *Node) {
- unsigned N = Node->getNumOperands();
- while (N && Node->getOperand(N - 1).getValueType() == MVT::Flag)
- --N;
- if (N && Node->getOperand(N - 1).getValueType() == MVT::Other)
- --N; // Ignore chain if it exists.
- return N;
-}
-
-
void ScheduleDAGSDNodes::dumpNode(const SUnit *SU) const {
- if (SU->getNode())
- SU->getNode()->dump(DAG);
- else
- cerr << "CROSS RC COPY ";
- cerr << "\n";
+ if (!SU->getNode()) {
+ dbgs() << "PHYS REG COPY\n";
+ return;
+ }
+
+ SU->getNode()->dump(DAG);
+ dbgs() << "\n";
SmallVector<SDNode *, 4> FlaggedNodes;
for (SDNode *N = SU->getNode()->getFlaggedNode(); N; N = N->getFlaggedNode())
FlaggedNodes.push_back(N);
while (!FlaggedNodes.empty()) {
- cerr << " ";
+ dbgs() << " ";
FlaggedNodes.back()->dump(DAG);
- cerr << "\n";
+ dbgs() << "\n";
FlaggedNodes.pop_back();
}
}
+
+/// EmitSchedule - Emit the machine code in scheduled order.
+MachineBasicBlock *ScheduleDAGSDNodes::
+EmitSchedule(DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) {
+ InstrEmitter Emitter(BB, InsertPos);
+ DenseMap<SDValue, unsigned> VRBaseMap;
+ DenseMap<SUnit*, unsigned> CopyVRBaseMap;
+
+ // For now, any constant debug info nodes go at the beginning.
+ for (SDDbgInfo::ConstDbgIterator I = DAG->DbgConstBegin(),
+ E = DAG->DbgConstEnd(); I!=E; I++) {
+ Emitter.EmitDbgValue(*I, EM);
+ delete *I;
+ }
+
+ for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
+ SUnit *SU = Sequence[i];
+ if (!SU) {
+ // Null SUnit* is a noop.
+ EmitNoop();
+ continue;
+ }
+
+ // For pre-regalloc scheduling, create instructions corresponding to the
+ // SDNode and any flagged SDNodes and append them to the block.
+ if (!SU->getNode()) {
+ // Emit a copy.
+ EmitPhysRegCopy(SU, CopyVRBaseMap);
+ continue;
+ }
+
+ SmallVector<SDNode *, 4> FlaggedNodes;
+ for (SDNode *N = SU->getNode()->getFlaggedNode(); N;
+ N = N->getFlaggedNode())
+ FlaggedNodes.push_back(N);
+ while (!FlaggedNodes.empty()) {
+ Emitter.EmitNode(FlaggedNodes.back(), SU->OrigNode != SU, SU->isCloned,
+ VRBaseMap, EM);
+ if (FlaggedNodes.back()->getHasDebugValue())
+ if (SDDbgValue *sd = DAG->GetDbgInfo(FlaggedNodes.back())) {
+ Emitter.EmitDbgValue(FlaggedNodes.back(), VRBaseMap, sd);
+ delete sd;
+ }
+ FlaggedNodes.pop_back();
+ }
+ Emitter.EmitNode(SU->getNode(), SU->OrigNode != SU, SU->isCloned,
+ VRBaseMap, EM);
+ if (SU->getNode()->getHasDebugValue())
+ if (SDDbgValue *sd = DAG->GetDbgInfo(SU->getNode())) {
+ Emitter.EmitDbgValue(SU->getNode(), VRBaseMap, sd);
+ delete sd;
+ }
+ }
+
+ BB = Emitter.getBlock();
+ InsertPos = Emitter.getInsertPos();
+ return BB;
+}