//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "pre-RA-sched"
-#include "llvm/CodeGen/ScheduleDAGSDNodes.h"
+#include "ScheduleDAGSDNodes.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/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) {
+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;
}
}
}
-/// BuildSchedUnits - Build SUnits from the selection dag that we are input.
-/// This SUnit graph is similar to the SelectionDAG, but represents flagged
-/// together nodes with a single SUnit.
void ScheduleDAGSDNodes::BuildSchedUnits() {
- // Reserve entries in the vector for each of the SUnits we are creating. This
- // ensure that reallocation of the vector won't happen, so SUnit*'s won't get
- // invalidated.
- SUnits.reserve(DAG->allnodes_size());
-
// During scheduling, the NodeId field of SDNode is used to map SDNodes
// to their associated SUnits by holding SUnits table indices. A value
// of -1 means the SDNode does not yet have an associated SUnit.
+ unsigned NumNodes = 0;
for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(),
- E = DAG->allnodes_end(); NI != E; ++NI)
+ E = DAG->allnodes_end(); NI != E; ++NI) {
NI->setNodeId(-1);
+ ++NumNodes;
+ }
+
+ // Reserve entries in the vector for each of the SUnits we are creating. This
+ // ensure that reallocation of the vector won't happen, so SUnit*'s won't get
+ // invalidated.
+ // FIXME: Multiply by 2 because we may clone nodes during scheduling.
+ // This is a temporary workaround.
+ SUnits.reserve(NumNodes * 2);
+
+ // 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) {
// 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.
assert(N->getNodeId() == -1 && "Node already inserted!");
N->setNodeId(NodeSUnit->NodeNum);
- ComputeLatency(NodeSUnit);
+ // Assign the Latency field of NodeSUnit using target-provided information.
+ if (UnitLatencies)
+ NodeSUnit->Latency = 1;
+ else
+ ComputeLatency(NodeSUnit);
}
-
+}
+
+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 = 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;
int Cost = 1;
// Determine if this is a physical register dependency.
CheckForPhysRegDependency(OpN, N, i, TRI, TII, PhysReg, Cost);
- SU->addPred(OpSU, isChain, false, PhysReg, Cost);
+ assert((PhysReg == 0 || !isChain) &&
+ "Chain dependence via physreg data?");
+ // 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);
}
}
}
}
+/// BuildSchedGraph - Build the SUnit graph from the selection dag that we
+/// 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() {
+ // Populate the SUnits array.
+ BuildSchedUnits();
+ // Compute all the scheduling dependencies between nodes.
+ AddSchedEdges();
+}
+
void ScheduleDAGSDNodes::ComputeLatency(SUnit *SU) {
const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
// Compute the latency for the node. We use the sum of the latencies for
// all nodes flagged together into this SUnit.
- if (InstrItins.isEmpty()) {
- // No latency information.
- SU->Latency = 1;
- return;
- }
-
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());
}
-
- // Ensure that CopyToReg and similar nodes have a non-zero latency.
- if (!SawMachineOpcode)
- SU->Latency = 1;
}
/// CountResults - The results of target nodes have register or immediate
void ScheduleDAGSDNodes::dumpNode(const SUnit *SU) const {
- if (SU->getNode())
- SU->getNode()->dump(DAG);
- else
- cerr << "CROSS RC COPY ";
- cerr << "\n";
+ if (!SU->getNode()) {
+ errs() << "PHYS REG COPY\n";
+ return;
+ }
+
+ SU->getNode()->dump(DAG);
+ errs() << "\n";
SmallVector<SDNode *, 4> FlaggedNodes;
for (SDNode *N = SU->getNode()->getFlaggedNode(); N; N = N->getFlaggedNode())
FlaggedNodes.push_back(N);
while (!FlaggedNodes.empty()) {
- cerr << " ";
+ errs() << " ";
FlaggedNodes.back()->dump(DAG);
- cerr << "\n";
+ errs() << "\n";
FlaggedNodes.pop_back();
}
}