//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "pre-RA-sched"
-#include "SDNodeDbgValue.h"
#include "ScheduleDAGSDNodes.h"
#include "InstrEmitter.h"
-#include "llvm/CodeGen/SelectionDAG.h"
-#include "llvm/CodeGen/MachineInstrBuilder.h"
-#include "llvm/CodeGen/MachineRegisterInfo.h"
-#include "llvm/MC/MCInstrItineraries.h"
-#include "llvm/Target/TargetMachine.h"
-#include "llvm/Target/TargetInstrInfo.h"
-#include "llvm/Target/TargetLowering.h"
-#include "llvm/Target/TargetRegisterInfo.h"
-#include "llvm/Target/TargetSubtargetInfo.h"
+#include "SDNodeDbgValue.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/SelectionDAG.h"
+#include "llvm/MC/MCInstrItineraries.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
+#include "llvm/Target/TargetInstrInfo.h"
+#include "llvm/Target/TargetLowering.h"
+#include "llvm/Target/TargetMachine.h"
+#include "llvm/Target/TargetRegisterInfo.h"
+#include "llvm/Target/TargetSubtargetInfo.h"
using namespace llvm;
STATISTIC(LoadsClustered, "Number of loads clustered together");
/// NewSUnit - Creates a new SUnit and return a ptr to it.
///
-SUnit *ScheduleDAGSDNodes::NewSUnit(SDNode *N) {
+SUnit *ScheduleDAGSDNodes::newSUnit(SDNode *N) {
#ifndef NDEBUG
const SUnit *Addr = 0;
if (!SUnits.empty())
}
SUnit *ScheduleDAGSDNodes::Clone(SUnit *Old) {
- SUnit *SU = NewSUnit(Old->getNode());
+ SUnit *SU = newSUnit(Old->getNode());
SU->OrigNode = Old->OrigNode;
SU->Latency = Old->Latency;
SU->isVRegCycle = Old->isVRegCycle;
}
}
-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);
-
+// Helper for AddGlue to clone node operands.
+static void CloneNodeWithValues(SDNode *N, SelectionDAG *DAG,
+ SmallVectorImpl<EVT> &VTs,
+ SDValue ExtraOper = SDValue()) {
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);
+ if (ExtraOper.getNode())
+ Ops.push_back(ExtraOper);
SDVTList VTList = DAG->getVTList(&VTs[0], VTs.size());
MachineSDNode::mmo_iterator Begin = 0, End = 0;
MN->setMemRefs(Begin, End);
}
+static bool 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 false;
+
+ // Don't add a glue operand to something that already uses glue.
+ if (GlueDestNode &&
+ N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Glue) {
+ return false;
+ }
+ // Don't add glue to something that already has a glue value.
+ if (N->getValueType(N->getNumValues() - 1) == MVT::Glue) return false;
+
+ for (unsigned I = 0, E = N->getNumValues(); I != E; ++I)
+ VTs.push_back(N->getValueType(I));
+
+ if (AddGlue)
+ VTs.push_back(MVT::Glue);
+
+ CloneNodeWithValues(N, DAG, VTs, Glue);
+
+ return true;
+}
+
+// Cleanup after unsuccessful AddGlue. Use the standard method of morphing the
+// node even though simply shrinking the value list is sufficient.
+static void RemoveUnusedGlue(SDNode *N, SelectionDAG *DAG) {
+ assert((N->getValueType(N->getNumValues() - 1) == MVT::Glue &&
+ !N->hasAnyUseOfValue(N->getNumValues() - 1)) &&
+ "expected an unused glue value");
+
+ SmallVector<EVT, 4> VTs;
+ for (unsigned I = 0, E = N->getNumValues()-1; I != E; ++I)
+ VTs.push_back(N->getValueType(I));
+
+ CloneNodeWithValues(N, DAG, VTs);
+}
+
/// 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
// 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);
+ SDValue InGlue = SDValue(0, 0);
+ if (AddGlue(Lead, InGlue, true, DAG))
+ 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);
+ // If AddGlue fails, we could leave an unsused glue value. This should not
+ // cause any
+ if (AddGlue(Load, InGlue, OutGlue, DAG)) {
+ if (OutGlue)
+ InGlue = SDValue(Load, Load->getNumValues() - 1);
- ++LoadsClustered;
+ ++LoadsClustered;
+ }
+ else if (!OutGlue && InGlue.getNode())
+ RemoveUnusedGlue(InGlue.getNode(), DAG);
}
}
// If this node has already been processed, stop now.
if (NI->getNodeId() != -1) continue;
- SUnit *NodeSUnit = NewSUnit(NI);
+ SUnit *NodeSUnit = newSUnit(NI);
// See if anything is glued to this node, if so, add them to glued
// nodes. Nodes can have at most one glue input and one glue output. Glue
InitNumRegDefsLeft(NodeSUnit);
// Assign the Latency field of NodeSUnit using target-provided information.
- ComputeLatency(NodeSUnit);
+ computeLatency(NodeSUnit);
}
// Find all call operands.
const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
// Check to see if the scheduler cares about latencies.
- bool UnitLatencies = ForceUnitLatencies();
+ bool UnitLatencies = forceUnitLatencies();
// Pass 2: add the preds, succs, etc.
for (unsigned su = 0, e = SUnits.size(); su != e; ++su) {
if(isChain && OpN->getOpcode() == ISD::TokenFactor)
OpLatency = 0;
- const SDep &dep = SDep(OpSU, isChain ? SDep::Order : SDep::Data,
- OpLatency, PhysReg);
+ SDep Dep = isChain ? SDep(OpSU, SDep::Barrier)
+ : SDep(OpSU, SDep::Data, PhysReg);
+ Dep.setLatency(OpLatency);
if (!isChain && !UnitLatencies) {
- ComputeOperandLatency(OpN, N, i, const_cast<SDep &>(dep));
- ST.adjustSchedDependency(OpSU, SU, const_cast<SDep &>(dep));
+ computeOperandLatency(OpN, N, i, Dep);
+ ST.adjustSchedDependency(OpSU, SU, Dep);
}
- if (!SU->addPred(dep) && !dep.isCtrl() && OpSU->NumRegDefsLeft > 1) {
+ if (!SU->addPred(Dep) && !Dep.isCtrl() && OpSU->NumRegDefsLeft > 1) {
// Multiple register uses are combined in the same SUnit. For example,
// we could have a set of glued nodes with all their defs consumed by
// another set of glued nodes. Register pressure tracking sees this as
for (;DefIdx < NodeNumDefs; ++DefIdx) {
if (!Node->hasAnyUseOfValue(DefIdx))
continue;
- ValueType = Node->getValueType(DefIdx);
+ ValueType = Node->getSimpleValueType(DefIdx);
++DefIdx;
return; // Found a normal regdef.
}
}
}
-void ScheduleDAGSDNodes::ComputeLatency(SUnit *SU) {
+void ScheduleDAGSDNodes::computeLatency(SUnit *SU) {
SDNode *N = SU->getNode();
// TokenFactor operands are considered zero latency, and some schedulers
}
// Check to see if the scheduler cares about latencies.
- if (ForceUnitLatencies()) {
+ if (forceUnitLatencies()) {
SU->Latency = 1;
return;
}
SU->Latency += TII->getInstrLatency(InstrItins, N);
}
-void ScheduleDAGSDNodes::ComputeOperandLatency(SDNode *Def, SDNode *Use,
+void ScheduleDAGSDNodes::computeOperandLatency(SDNode *Def, SDNode *Use,
unsigned OpIdx, SDep& dep) const{
// Check to see if the scheduler cares about latencies.
- if (ForceUnitLatencies())
+ if (forceUnitLatencies())
return;
if (dep.getKind() != SDep::Data)
}
void ScheduleDAGSDNodes::dumpNode(const SUnit *SU) const {
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
if (!SU->getNode()) {
dbgs() << "PHYS REG COPY\n";
return;
dbgs() << "\n";
GluedNodes.pop_back();
}
+#endif
}
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
void ScheduleDAGSDNodes::dumpSchedule() const {
for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
if (SUnit *SU = Sequence[i])
dbgs() << "**** NOOP ****\n";
}
}
+#endif
#ifndef NDEBUG
/// VerifyScheduledSequence - Verify that all SUnits were scheduled and that
}
SmallVector<SDNode *, 4> GluedNodes;
- for (SDNode *N = SU->getNode()->getGluedNode(); N;
- N = N->getGluedNode())
+ for (SDNode *N = SU->getNode()->getGluedNode(); N; N = N->getGluedNode())
GluedNodes.push_back(N);
while (!GluedNodes.empty()) {
SDNode *N = GluedNodes.back();
}
// Add trailing DbgValue's before the terminator. FIXME: May want to add
// some of them before one or more conditional branches?
+ SmallVector<MachineInstr*, 8> DbgMIs;
while (DI != DE) {
- MachineBasicBlock *InsertBB = Emitter.getBlock();
- MachineBasicBlock::iterator Pos= Emitter.getBlock()->getFirstTerminator();
- if (!(*DI)->isInvalidated()) {
- MachineInstr *DbgMI= Emitter.EmitDbgValue(*DI, VRBaseMap);
- if (DbgMI)
- InsertBB->insert(Pos, DbgMI);
- }
+ if (!(*DI)->isInvalidated())
+ if (MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, VRBaseMap))
+ DbgMIs.push_back(DbgMI);
++DI;
}
+
+ MachineBasicBlock *InsertBB = Emitter.getBlock();
+ MachineBasicBlock::iterator Pos = InsertBB->getFirstTerminator();
+ InsertBB->insert(Pos, DbgMIs.begin(), DbgMIs.end());
}
InsertPos = Emitter.getInsertPos();