//
// The LLVM Compiler Infrastructure
//
-// This file was developed by James M. Laskey and is distributed under the
-// University of Illinois Open Source License. See LICENSE.TXT for details.
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
-// This implements a simple two pass scheduler. The first pass attempts to push
-// backward any lengthy instructions and critical paths. The second pass packs
-// instructions into semi-optimal time slots.
+// This implements the ScheduleDAG class, which is a base class used by
+// scheduling implementation classes.
//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "pre-RA-sched"
-#include "llvm/Type.h"
#include "llvm/CodeGen/ScheduleDAG.h"
-#include "llvm/CodeGen/MachineConstantPool.h"
-#include "llvm/CodeGen/MachineFunction.h"
-#include "llvm/CodeGen/SSARegMap.h"
-#include "llvm/Target/TargetData.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetInstrInfo.h"
-#include "llvm/Target/TargetLowering.h"
+#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Support/Debug.h"
-#include "llvm/Support/MathExtras.h"
using namespace llvm;
+ScheduleDAG::ScheduleDAG(SelectionDAG &dag, MachineBasicBlock *bb,
+ const TargetMachine &tm)
+ : DAG(dag), BB(bb), TM(tm), MRI(BB->getParent()->getRegInfo()) {
+ TII = TM.getInstrInfo();
+ MF = &DAG.getMachineFunction();
+ TRI = TM.getRegisterInfo();
+ TLI = &DAG.getTargetLoweringInfo();
+ ConstPool = BB->getParent()->getConstantPool();
+}
+
+/// CheckForPhysRegDependency - Check if the dependency between def and use of
+/// a specified operand is a physical register dependency. If so, returns the
+/// 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, int &Cost) {
+ if (Op != 2 || User->getOpcode() != ISD::CopyToReg)
+ return;
+
+ unsigned Reg = cast<RegisterSDNode>(User->getOperand(1))->getReg();
+ if (TargetRegisterInfo::isVirtualRegister(Reg))
+ return;
+
+ unsigned ResNo = User->getOperand(2).getResNo();
+ if (Def->isMachineOpcode()) {
+ const TargetInstrDesc &II = TII->get(Def->getMachineOpcode());
+ if (ResNo >= II.getNumDefs() &&
+ II.ImplicitDefs[ResNo - II.getNumDefs()] == Reg) {
+ PhysReg = Reg;
+ const TargetRegisterClass *RC =
+ TRI->getPhysicalRegisterRegClass(Reg, Def->getValueType(ResNo));
+ Cost = RC->getCopyCost();
+ }
+ }
+}
+
+SUnit *ScheduleDAG::Clone(SUnit *Old) {
+ SUnit *SU = NewSUnit(Old->Node);
+ SU->OrigNode = Old->OrigNode;
+ SU->FlaggedNodes = Old->FlaggedNodes;
+ SU->Latency = Old->Latency;
+ SU->isTwoAddress = Old->isTwoAddress;
+ SU->isCommutable = Old->isCommutable;
+ SU->hasPhysRegDefs = Old->hasPhysRegDefs;
+ 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.
// 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(std::distance(DAG.allnodes_begin(), DAG.allnodes_end()));
-
- const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
+ 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.
+ for (SelectionDAG::allnodes_iterator NI = DAG.allnodes_begin(),
+ E = DAG.allnodes_end(); NI != E; ++NI)
+ NI->setNodeId(-1);
+
for (SelectionDAG::allnodes_iterator NI = DAG.allnodes_begin(),
E = DAG.allnodes_end(); NI != E; ++NI) {
if (isPassiveNode(NI)) // Leaf node, e.g. a TargetImmediate.
continue;
// If this node has already been processed, stop now.
- if (SUnitMap[NI]) continue;
+ if (NI->getNodeId() != -1) continue;
SUnit *NodeSUnit = NewSUnit(NI);
if (N->getNumOperands() &&
N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag) {
do {
- N = N->getOperand(N->getNumOperands()-1).Val;
+ N = N->getOperand(N->getNumOperands()-1).getNode();
NodeSUnit->FlaggedNodes.push_back(N);
- SUnitMap[N] = NodeSUnit;
+ assert(N->getNodeId() == -1 && "Node already inserted!");
+ N->setNodeId(NodeSUnit->NodeNum);
} while (N->getNumOperands() &&
N->getOperand(N->getNumOperands()-1).getValueType()== MVT::Flag);
std::reverse(NodeSUnit->FlaggedNodes.begin(),
// have a user of the flag operand.
N = NI;
while (N->getValueType(N->getNumValues()-1) == MVT::Flag) {
- SDOperand FlagVal(N, N->getNumValues()-1);
+ SDValue FlagVal(N, N->getNumValues()-1);
// There are either zero or one users of the Flag result.
bool HasFlagUse = false;
for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end();
UI != E; ++UI)
- if (FlagVal.isOperand(*UI)) {
+ if (FlagVal.isOperandOf(*UI)) {
HasFlagUse = true;
NodeSUnit->FlaggedNodes.push_back(N);
- SUnitMap[N] = NodeSUnit;
+ assert(N->getNodeId() == -1 && "Node already inserted!");
+ N->setNodeId(NodeSUnit->NodeNum);
N = *UI;
break;
}
// Now all flagged nodes are in FlaggedNodes and N is the bottom-most node.
// Update the SUnit
NodeSUnit->Node = N;
- SUnitMap[N] = NodeSUnit;
-
- // 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.
- NodeSUnit->Latency = 1;
- } else {
- NodeSUnit->Latency = 0;
- if (N->isTargetOpcode()) {
- unsigned SchedClass = TII->getSchedClass(N->getTargetOpcode());
- InstrStage *S = InstrItins.begin(SchedClass);
- InstrStage *E = InstrItins.end(SchedClass);
- for (; S != E; ++S)
- NodeSUnit->Latency += S->Cycles;
- }
- for (unsigned i = 0, e = NodeSUnit->FlaggedNodes.size(); i != e; ++i) {
- SDNode *FNode = NodeSUnit->FlaggedNodes[i];
- if (FNode->isTargetOpcode()) {
- unsigned SchedClass = TII->getSchedClass(FNode->getTargetOpcode());
- InstrStage *S = InstrItins.begin(SchedClass);
- InstrStage *E = InstrItins.end(SchedClass);
- for (; S != E; ++S)
- NodeSUnit->Latency += S->Cycles;
- }
- }
- }
+ assert(N->getNodeId() == -1 && "Node already inserted!");
+ N->setNodeId(NodeSUnit->NodeNum);
+
+ ComputeLatency(NodeSUnit);
}
// Pass 2: add the preds, succs, etc.
SUnit *SU = &SUnits[su];
SDNode *MainNode = SU->Node;
- if (MainNode->isTargetOpcode()) {
- unsigned Opc = MainNode->getTargetOpcode();
- for (unsigned i = 0, ee = TII->getNumOperands(Opc); i != ee; ++i) {
- if (TII->getOperandConstraint(Opc, i, TOI::TIED_TO) != -1) {
+ if (MainNode->isMachineOpcode()) {
+ unsigned Opc = MainNode->getMachineOpcode();
+ const TargetInstrDesc &TID = TII->get(Opc);
+ for (unsigned i = 0; i != TID.getNumOperands(); ++i) {
+ if (TID.getOperandConstraint(i, TOI::TIED_TO) != -1) {
SU->isTwoAddress = true;
break;
}
}
- if (TII->isCommutableInstr(Opc))
+ if (TID.isCommutable())
SU->isCommutable = true;
}
for (unsigned n = 0, e = SU->FlaggedNodes.size(); n != e; ++n) {
SDNode *N = SU->FlaggedNodes[n];
+ if (N->isMachineOpcode() &&
+ TII->get(N->getMachineOpcode()).getImplicitDefs() &&
+ CountResults(N) > TII->get(N->getMachineOpcode()).getNumDefs())
+ SU->hasPhysRegDefs = true;
for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
- SDNode *OpN = N->getOperand(i).Val;
+ SDNode *OpN = N->getOperand(i).getNode();
if (isPassiveNode(OpN)) continue; // Not scheduled.
- SUnit *OpSU = SUnitMap[OpN];
+ SUnit *OpSU = &SUnits[OpN->getNodeId()];
assert(OpSU && "Node has no SUnit!");
if (OpSU == SU) continue; // In the same group.
- MVT::ValueType OpVT = N->getOperand(i).getValueType();
+ MVT OpVT = N->getOperand(i).getValueType();
assert(OpVT != MVT::Flag && "Flagged nodes should be in same sunit!");
bool isChain = OpVT == MVT::Other;
-
- if (SU->addPred(OpSU, isChain)) {
- if (!isChain) {
- SU->NumPreds++;
- SU->NumPredsLeft++;
- } else {
- SU->NumChainPredsLeft++;
- }
- }
- if (OpSU->addSucc(SU, isChain)) {
- if (!isChain) {
- OpSU->NumSuccs++;
- OpSU->NumSuccsLeft++;
- } else {
- OpSU->NumChainSuccsLeft++;
- }
- }
+
+ unsigned PhysReg = 0;
+ 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);
}
}
// Remove MainNode from FlaggedNodes again.
SU->FlaggedNodes.pop_back();
}
+}
+
+void ScheduleDAG::ComputeLatency(SUnit *SU) {
+ const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
- return;
+ // 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;
+ if (SU->Node->isMachineOpcode()) {
+ unsigned SchedClass = TII->get(SU->Node->getMachineOpcode()).getSchedClass();
+ const InstrStage *S = InstrItins.begin(SchedClass);
+ const InstrStage *E = InstrItins.end(SchedClass);
+ for (; S != E; ++S)
+ SU->Latency += S->Cycles;
+ }
+ for (unsigned i = 0, e = SU->FlaggedNodes.size(); i != e; ++i) {
+ SDNode *FNode = SU->FlaggedNodes[i];
+ if (FNode->isMachineOpcode()) {
+ unsigned SchedClass = TII->get(FNode->getMachineOpcode()).getSchedClass();
+ const InstrStage *S = InstrItins.begin(SchedClass);
+ const InstrStage *E = InstrItins.end(SchedClass);
+ for (; S != E; ++S)
+ SU->Latency += S->Cycles;
+ }
+ }
}
+/// CalculateDepths - compute depths using algorithms for the longest
+/// paths in the DAG
void ScheduleDAG::CalculateDepths() {
- std::vector<std::pair<SUnit*, unsigned> > WorkList;
- for (unsigned i = 0, e = SUnits.size(); i != e; ++i)
- if (SUnits[i].Preds.size() == 0/* && &SUnits[i] != Entry*/)
- WorkList.push_back(std::make_pair(&SUnits[i], 0U));
+ unsigned DAGSize = SUnits.size();
+ std::vector<SUnit*> WorkList;
+ WorkList.reserve(DAGSize);
+
+ // Initialize the data structures
+ for (unsigned i = 0, e = DAGSize; i != e; ++i) {
+ SUnit *SU = &SUnits[i];
+ unsigned Degree = SU->Preds.size();
+ // Temporarily use the Depth field as scratch space for the degree count.
+ SU->Depth = Degree;
+
+ // Is it a node without dependencies?
+ if (Degree == 0) {
+ assert(SU->Preds.empty() && "SUnit should have no predecessors");
+ // Collect leaf nodes
+ WorkList.push_back(SU);
+ }
+ }
+ // Process nodes in the topological order
while (!WorkList.empty()) {
- SUnit *SU = WorkList.back().first;
- unsigned Depth = WorkList.back().second;
+ SUnit *SU = WorkList.back();
WorkList.pop_back();
- if (SU->Depth == 0 || Depth > SU->Depth) {
- SU->Depth = Depth;
- for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
- I != E; ++I)
- WorkList.push_back(std::make_pair(I->first, Depth+1));
+ unsigned SUDepth = 0;
+
+ // Use dynamic programming:
+ // When current node is being processed, all of its dependencies
+ // are already processed.
+ // So, just iterate over all predecessors and take the longest path
+ for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
+ I != E; ++I) {
+ unsigned PredDepth = I->Dep->Depth;
+ if (PredDepth+1 > SUDepth) {
+ SUDepth = PredDepth + 1;
+ }
+ }
+
+ SU->Depth = SUDepth;
+
+ // Update degrees of all nodes depending on current SUnit
+ for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
+ I != E; ++I) {
+ SUnit *SU = I->Dep;
+ if (!--SU->Depth)
+ // If all dependencies of the node are processed already,
+ // then the longest path for the node can be computed now
+ WorkList.push_back(SU);
}
}
}
+/// CalculateHeights - compute heights using algorithms for the longest
+/// paths in the DAG
void ScheduleDAG::CalculateHeights() {
- std::vector<std::pair<SUnit*, unsigned> > WorkList;
- SUnit *Root = SUnitMap[DAG.getRoot().Val];
- WorkList.push_back(std::make_pair(Root, 0U));
+ unsigned DAGSize = SUnits.size();
+ std::vector<SUnit*> WorkList;
+ WorkList.reserve(DAGSize);
+
+ // Initialize the data structures
+ for (unsigned i = 0, e = DAGSize; i != e; ++i) {
+ SUnit *SU = &SUnits[i];
+ unsigned Degree = SU->Succs.size();
+ // Temporarily use the Height field as scratch space for the degree count.
+ SU->Height = Degree;
+
+ // Is it a node without dependencies?
+ if (Degree == 0) {
+ assert(SU->Succs.empty() && "Something wrong");
+ assert(WorkList.empty() && "Should be empty");
+ // Collect leaf nodes
+ WorkList.push_back(SU);
+ }
+ }
+ // Process nodes in the topological order
while (!WorkList.empty()) {
- SUnit *SU = WorkList.back().first;
- unsigned Height = WorkList.back().second;
+ SUnit *SU = WorkList.back();
WorkList.pop_back();
- if (SU->Height == 0 || Height > SU->Height) {
- SU->Height = Height;
- for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
- I != E; ++I)
- WorkList.push_back(std::make_pair(I->first, Height+1));
+ unsigned SUHeight = 0;
+
+ // Use dynamic programming:
+ // When current node is being processed, all of its dependencies
+ // are already processed.
+ // So, just iterate over all successors and take the longest path
+ for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
+ I != E; ++I) {
+ unsigned SuccHeight = I->Dep->Height;
+ if (SuccHeight+1 > SUHeight) {
+ SUHeight = SuccHeight + 1;
+ }
+ }
+
+ SU->Height = SUHeight;
+
+ // Update degrees of all nodes depending on current SUnit
+ for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
+ I != E; ++I) {
+ SUnit *SU = I->Dep;
+ if (!--SU->Height)
+ // If all dependencies of the node are processed already,
+ // then the longest path for the node can be computed now
+ WorkList.push_back(SU);
}
}
}
/// 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 machine instrs.)
+/// not go into the resulting MachineInstr).
unsigned ScheduleDAG::CountResults(SDNode *Node) {
unsigned N = Node->getNumValues();
while (N && Node->getValueType(N - 1) == MVT::Flag)
return N;
}
-/// CountOperands The inputs to target nodes have any actual inputs first,
-/// followed by an optional chain operand, then flag operands. Compute the
-/// number of actual operands that will go into the machine instr.
+/// 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 ScheduleDAG::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 ScheduleDAG::ComputeMemOperandsEnd(SDNode *Node) {
unsigned N = Node->getNumOperands();
while (N && Node->getOperand(N - 1).getValueType() == MVT::Flag)
--N;
return N;
}
-static const TargetRegisterClass *getInstrOperandRegClass(
- const MRegisterInfo *MRI,
- const TargetInstrInfo *TII,
- const TargetInstrDescriptor *II,
- unsigned Op) {
- if (Op >= II->numOperands) {
- assert((II->Flags & M_VARIABLE_OPS)&& "Invalid operand # of instruction");
- return NULL;
- }
- const TargetOperandInfo &toi = II->OpInfo[Op];
- return (toi.Flags & M_LOOK_UP_PTR_REG_CLASS)
- ? TII->getPointerRegClass() : MRI->getRegClass(toi.RegClass);
-}
-
-static void CreateVirtualRegisters(SDNode *Node,
- unsigned NumResults,
- const MRegisterInfo *MRI,
- MachineInstr *MI,
- SSARegMap *RegMap,
- const TargetInstrInfo *TII,
- const TargetInstrDescriptor &II,
- DenseMap<SDOperand, unsigned> &VRBaseMap) {
- for (unsigned i = 0; i < NumResults; ++i) {
- // If the specific node value is only used by a CopyToReg and the dest reg
- // is a vreg, use the CopyToReg'd destination register instead of creating
- // a new vreg.
- unsigned VRBase = 0;
- for (SDNode::use_iterator UI = Node->use_begin(), E = Node->use_end();
- UI != E; ++UI) {
- SDNode *Use = *UI;
- if (Use->getOpcode() == ISD::CopyToReg &&
- Use->getOperand(2).Val == Node &&
- Use->getOperand(2).ResNo == i) {
- unsigned Reg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
- if (MRegisterInfo::isVirtualRegister(Reg)) {
- VRBase = Reg;
- MI->addRegOperand(Reg, true);
- break;
- }
- }
- }
-
- if (VRBase == 0) {
- // Create the result registers for this node and add the result regs to
- // the machine instruction.
- const TargetRegisterClass *RC = getInstrOperandRegClass(MRI, TII, &II, i);
- assert(RC && "Isn't a register operand!");
- VRBase = RegMap->createVirtualRegister(RC);
- MI->addRegOperand(VRBase, true);
- }
-
- bool isNew = VRBaseMap.insert(std::make_pair(SDOperand(Node,i), VRBase));
- assert(isNew && "Node emitted out of order - early");
- }
-}
-
-/// getVR - Return the virtual register corresponding to the specified result
-/// of the specified node.
-static unsigned getVR(SDOperand Op, DenseMap<SDOperand, unsigned> &VRBaseMap) {
- DenseMap<SDOperand, unsigned>::iterator I = VRBaseMap.find(Op);
- assert(I != VRBaseMap.end() && "Node emitted out of order - late");
- return I->second;
-}
-
-
-/// AddOperand - Add the specified operand to the specified machine instr. II
-/// specifies the instruction information for the node, and IIOpNum is the
-/// operand number (in the II) that we are adding. IIOpNum and II are used for
-/// assertions only.
-void ScheduleDAG::AddOperand(MachineInstr *MI, SDOperand Op,
- unsigned IIOpNum,
- const TargetInstrDescriptor *II,
- DenseMap<SDOperand, unsigned> &VRBaseMap) {
- if (Op.isTargetOpcode()) {
- // Note that this case is redundant with the final else block, but we
- // include it because it is the most common and it makes the logic
- // simpler here.
- assert(Op.getValueType() != MVT::Other &&
- Op.getValueType() != MVT::Flag &&
- "Chain and flag operands should occur at end of operand list!");
-
- // Get/emit the operand.
- unsigned VReg = getVR(Op, VRBaseMap);
- const TargetInstrDescriptor *TID = MI->getInstrDescriptor();
- bool isOptDef = (IIOpNum < TID->numOperands)
- ? (TID->OpInfo[IIOpNum].Flags & M_OPTIONAL_DEF_OPERAND) : false;
- MI->addRegOperand(VReg, isOptDef);
-
- // Verify that it is right.
- assert(MRegisterInfo::isVirtualRegister(VReg) && "Not a vreg?");
- if (II) {
- const TargetRegisterClass *RC =
- getInstrOperandRegClass(MRI, TII, II, IIOpNum);
- assert(RC && "Don't have operand info for this instruction!");
- const TargetRegisterClass *VRC = RegMap->getRegClass(VReg);
- if (VRC != RC) {
- cerr << "Register class of operand and regclass of use don't agree!\n";
-#ifndef NDEBUG
- cerr << "Operand = " << IIOpNum << "\n";
- cerr << "Op->Val = "; Op.Val->dump(&DAG); cerr << "\n";
- cerr << "MI = "; MI->print(cerr);
- cerr << "VReg = " << VReg << "\n";
- cerr << "VReg RegClass size = " << VRC->getSize()
- << ", align = " << VRC->getAlignment() << "\n";
- cerr << "Expected RegClass size = " << RC->getSize()
- << ", align = " << RC->getAlignment() << "\n";
-#endif
- cerr << "Fatal error, aborting.\n";
- abort();
- }
- }
- } else if (ConstantSDNode *C =
- dyn_cast<ConstantSDNode>(Op)) {
- MI->addImmOperand(C->getValue());
- } else if (RegisterSDNode *R =
- dyn_cast<RegisterSDNode>(Op)) {
- MI->addRegOperand(R->getReg(), false);
- } else if (GlobalAddressSDNode *TGA =
- dyn_cast<GlobalAddressSDNode>(Op)) {
- MI->addGlobalAddressOperand(TGA->getGlobal(), TGA->getOffset());
- } else if (BasicBlockSDNode *BB =
- dyn_cast<BasicBlockSDNode>(Op)) {
- MI->addMachineBasicBlockOperand(BB->getBasicBlock());
- } else if (FrameIndexSDNode *FI =
- dyn_cast<FrameIndexSDNode>(Op)) {
- MI->addFrameIndexOperand(FI->getIndex());
- } else if (JumpTableSDNode *JT =
- dyn_cast<JumpTableSDNode>(Op)) {
- MI->addJumpTableIndexOperand(JT->getIndex());
- } else if (ConstantPoolSDNode *CP =
- dyn_cast<ConstantPoolSDNode>(Op)) {
- int Offset = CP->getOffset();
- unsigned Align = CP->getAlignment();
- const Type *Type = CP->getType();
- // MachineConstantPool wants an explicit alignment.
- if (Align == 0) {
- Align = TM.getTargetData()->getPreferredTypeAlignmentShift(Type);
- if (Align == 0) {
- // Alignment of vector types. FIXME!
- Align = TM.getTargetData()->getTypeSize(Type);
- Align = Log2_64(Align);
- }
- }
-
- unsigned Idx;
- if (CP->isMachineConstantPoolEntry())
- Idx = ConstPool->getConstantPoolIndex(CP->getMachineCPVal(), Align);
- else
- Idx = ConstPool->getConstantPoolIndex(CP->getConstVal(), Align);
- MI->addConstantPoolIndexOperand(Idx, Offset);
- } else if (ExternalSymbolSDNode *ES =
- dyn_cast<ExternalSymbolSDNode>(Op)) {
- MI->addExternalSymbolOperand(ES->getSymbol());
- } else {
- assert(Op.getValueType() != MVT::Other &&
- Op.getValueType() != MVT::Flag &&
- "Chain and flag operands should occur at end of operand list!");
- unsigned VReg = getVR(Op, VRBaseMap);
- MI->addRegOperand(VReg, false);
-
- // Verify that it is right.
- assert(MRegisterInfo::isVirtualRegister(VReg) && "Not a vreg?");
- if (II) {
- const TargetRegisterClass *RC =
- getInstrOperandRegClass(MRI, TII, II, IIOpNum);
- assert(RC && "Don't have operand info for this instruction!");
- assert(RegMap->getRegClass(VReg) == RC &&
- "Register class of operand and regclass of use don't agree!");
- }
- }
-
-}
-
-// Returns the Register Class of a physical register
-static const TargetRegisterClass *getPhysicalRegisterRegClass(
- const MRegisterInfo *MRI,
- MVT::ValueType VT,
- unsigned reg) {
- assert(MRegisterInfo::isPhysicalRegister(reg) &&
- "reg must be a physical register");
- // Pick the register class of the right type that contains this physreg.
- for (MRegisterInfo::regclass_iterator I = MRI->regclass_begin(),
- E = MRI->regclass_end(); I != E; ++I)
- if ((*I)->hasType(VT) && (*I)->contains(reg))
- return *I;
- assert(false && "Couldn't find the register class");
- return 0;
-}
-
-// Returns the Register Class of a subregister
-static const TargetRegisterClass *getSubRegisterRegClass(
- const TargetRegisterClass *TRC,
- unsigned SubIdx) {
- // Pick the register class of the subregister
- MRegisterInfo::regclass_iterator I = TRC->subregclasses_begin() + SubIdx-1;
- assert(I < TRC->subregclasses_end() &&
- "Invalid subregister index for register class");
- return *I;
-}
-
-static const TargetRegisterClass *getSuperregRegisterClass(
- const TargetRegisterClass *TRC,
- unsigned SubIdx,
- MVT::ValueType VT) {
- // Pick the register class of the superegister for this type
- for (MRegisterInfo::regclass_iterator I = TRC->superregclasses_begin(),
- E = TRC->superregclasses_end(); I != E; ++I)
- if ((*I)->hasType(VT) && getSubRegisterRegClass(*I, SubIdx) == TRC)
- return *I;
- assert(false && "Couldn't find the register class");
- return 0;
-}
-
-/// EmitSubregNode - Generate machine code for subreg nodes.
-///
-void ScheduleDAG::EmitSubregNode(SDNode *Node,
- DenseMap<SDOperand, unsigned> &VRBaseMap) {
- unsigned VRBase = 0;
- unsigned Opc = Node->getTargetOpcode();
- if (Opc == TargetInstrInfo::EXTRACT_SUBREG) {
- // If the node is only used by a CopyToReg and the dest reg is a vreg, use
- // the CopyToReg'd destination register instead of creating a new vreg.
- for (SDNode::use_iterator UI = Node->use_begin(), E = Node->use_end();
- UI != E; ++UI) {
- SDNode *Use = *UI;
- if (Use->getOpcode() == ISD::CopyToReg &&
- Use->getOperand(2).Val == Node) {
- unsigned DestReg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
- if (MRegisterInfo::isVirtualRegister(DestReg)) {
- VRBase = DestReg;
- break;
- }
- }
- }
-
- unsigned SubIdx = cast<ConstantSDNode>(Node->getOperand(1))->getValue();
-
- // TODO: If the node is a use of a CopyFromReg from a physical register
- // fold the extract into the copy now
-
- // TODO: Add tracking info to SSARegMap of which vregs are subregs
- // to allow coalescing in the allocator
-
- // Create the extract_subreg machine instruction.
- MachineInstr *MI =
- new MachineInstr(BB, TII->get(TargetInstrInfo::EXTRACT_SUBREG));
-
- // Figure out the register class to create for the destreg.
- unsigned VReg = getVR(Node->getOperand(0), VRBaseMap);
- const TargetRegisterClass *TRC = RegMap->getRegClass(VReg);
- const TargetRegisterClass *SRC = getSubRegisterRegClass(TRC, SubIdx);
-
- if (VRBase) {
- // Grab the destination register
- const TargetRegisterClass *DRC = 0;
- DRC = RegMap->getRegClass(VRBase);
- assert(SRC == DRC &&
- "Source subregister and destination must have the same class");
- } else {
- // Create the reg
- VRBase = RegMap->createVirtualRegister(SRC);
- }
-
- // Add def, source, and subreg index
- MI->addRegOperand(VRBase, true);
- AddOperand(MI, Node->getOperand(0), 0, 0, VRBaseMap);
- MI->addImmOperand(SubIdx);
-
- } else if (Opc == TargetInstrInfo::INSERT_SUBREG) {
- assert((Node->getNumOperands() == 2 || Node->getNumOperands() == 3) &&
- "Malformed insert_subreg node");
- bool isUndefInput = (Node->getNumOperands() == 2);
- unsigned SubReg = 0;
- unsigned SubIdx = 0;
-
- if (isUndefInput) {
- SubReg = getVR(Node->getOperand(0), VRBaseMap);
- SubIdx = cast<ConstantSDNode>(Node->getOperand(1))->getValue();
- } else {
- SubReg = getVR(Node->getOperand(1), VRBaseMap);
- SubIdx = cast<ConstantSDNode>(Node->getOperand(2))->getValue();
- }
-
- // TODO: Add tracking info to SSARegMap of which vregs are subregs
- // to allow coalescing in the allocator
-
- // If the node is only used by a CopyToReg and the dest reg is a vreg, use
- // the CopyToReg'd destination register instead of creating a new vreg.
- // If the CopyToReg'd destination register is physical, then fold the
- // insert into the copy
- for (SDNode::use_iterator UI = Node->use_begin(), E = Node->use_end();
- UI != E; ++UI) {
- SDNode *Use = *UI;
- if (Use->getOpcode() == ISD::CopyToReg &&
- Use->getOperand(2).Val == Node) {
- unsigned DestReg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
- if (MRegisterInfo::isVirtualRegister(DestReg)) {
- VRBase = DestReg;
- break;
- }
- }
- }
-
- // Create the insert_subreg machine instruction.
- MachineInstr *MI =
- new MachineInstr(BB, TII->get(TargetInstrInfo::INSERT_SUBREG));
-
- // Figure out the register class to create for the destreg.
- const TargetRegisterClass *TRC = 0;
- if (VRBase) {
- TRC = RegMap->getRegClass(VRBase);
- } else {
- TRC = getSuperregRegisterClass(RegMap->getRegClass(SubReg),
- SubIdx,
- Node->getValueType(0));
- assert(TRC && "Couldn't determine register class for insert_subreg");
- VRBase = RegMap->createVirtualRegister(TRC); // Create the reg
- }
-
- MI->addRegOperand(VRBase, true);
- AddOperand(MI, Node->getOperand(0), 0, 0, VRBaseMap);
- if (!isUndefInput)
- AddOperand(MI, Node->getOperand(1), 0, 0, VRBaseMap);
- MI->addImmOperand(SubIdx);
- } else
- assert(0 && "Node is not a subreg insert or extract");
-
- bool isNew = VRBaseMap.insert(std::make_pair(SDOperand(Node,0), VRBase));
- assert(isNew && "Node emitted out of order - early");
-}
-
-/// EmitNode - Generate machine code for an node and needed dependencies.
-///
-void ScheduleDAG::EmitNode(SDNode *Node,
- DenseMap<SDOperand, unsigned> &VRBaseMap) {
- // If machine instruction
- if (Node->isTargetOpcode()) {
- unsigned Opc = Node->getTargetOpcode();
-
- // Handle subreg insert/extract specially
- if (Opc == TargetInstrInfo::EXTRACT_SUBREG ||
- Opc == TargetInstrInfo::INSERT_SUBREG) {
- EmitSubregNode(Node, VRBaseMap);
- return;
- }
-
- const TargetInstrDescriptor &II = TII->get(Opc);
-
- unsigned NumResults = CountResults(Node);
- unsigned NodeOperands = CountOperands(Node);
- unsigned NumMIOperands = NodeOperands + NumResults;
-#ifndef NDEBUG
- assert((unsigned(II.numOperands) == NumMIOperands ||
- (II.Flags & M_VARIABLE_OPS)) &&
- "#operands for dag node doesn't match .td file!");
-#endif
-
- // Create the new machine instruction.
- MachineInstr *MI = new MachineInstr(II);
-
- // Add result register values for things that are defined by this
- // instruction.
- if (NumResults)
- CreateVirtualRegisters(Node, NumResults, MRI, MI, RegMap,
- TII, II, VRBaseMap);
-
- // Emit all of the actual operands of this instruction, adding them to the
- // instruction as appropriate.
- for (unsigned i = 0; i != NodeOperands; ++i)
- AddOperand(MI, Node->getOperand(i), i+NumResults, &II, VRBaseMap);
-
- // Commute node if it has been determined to be profitable.
- if (CommuteSet.count(Node)) {
- MachineInstr *NewMI = TII->commuteInstruction(MI);
- if (NewMI == 0)
- DOUT << "Sched: COMMUTING FAILED!\n";
- else {
- DOUT << "Sched: COMMUTED TO: " << *NewMI;
- if (MI != NewMI) {
- delete MI;
- MI = NewMI;
- }
- }
- }
-
- // Now that we have emitted all operands, emit this instruction itself.
- if ((II.Flags & M_USES_CUSTOM_DAG_SCHED_INSERTION) == 0) {
- BB->insert(BB->end(), MI);
- } else {
- // Insert this instruction into the end of the basic block, potentially
- // taking some custom action.
- BB = DAG.getTargetLoweringInfo().InsertAtEndOfBasicBlock(MI, BB);
- }
- } else {
- switch (Node->getOpcode()) {
- default:
-#ifndef NDEBUG
- Node->dump(&DAG);
-#endif
- assert(0 && "This target-independent node should have been selected!");
- case ISD::EntryToken: // fall thru
- case ISD::TokenFactor:
- case ISD::LABEL:
- break;
- case ISD::CopyToReg: {
- unsigned InReg;
- if (RegisterSDNode *R = dyn_cast<RegisterSDNode>(Node->getOperand(2)))
- InReg = R->getReg();
- else
- InReg = getVR(Node->getOperand(2), VRBaseMap);
- unsigned DestReg = cast<RegisterSDNode>(Node->getOperand(1))->getReg();
- if (InReg != DestReg) {// Coalesced away the copy?
- const TargetRegisterClass *TRC = 0;
- // Get the target register class
- if (MRegisterInfo::isVirtualRegister(InReg))
- TRC = RegMap->getRegClass(InReg);
- else
- TRC = getPhysicalRegisterRegClass(MRI,
- Node->getOperand(2).getValueType(),
- InReg);
- MRI->copyRegToReg(*BB, BB->end(), DestReg, InReg, TRC);
- }
- break;
- }
- case ISD::CopyFromReg: {
- unsigned VRBase = 0;
- unsigned SrcReg = cast<RegisterSDNode>(Node->getOperand(1))->getReg();
- if (MRegisterInfo::isVirtualRegister(SrcReg)) {
- // Just use the input register directly!
- bool isNew = VRBaseMap.insert(std::make_pair(SDOperand(Node,0),SrcReg));
- assert(isNew && "Node emitted out of order - early");
- break;
- }
-
- // If the node is only used by a CopyToReg and the dest reg is a vreg, use
- // the CopyToReg'd destination register instead of creating a new vreg.
- for (SDNode::use_iterator UI = Node->use_begin(), E = Node->use_end();
- UI != E; ++UI) {
- SDNode *Use = *UI;
- if (Use->getOpcode() == ISD::CopyToReg &&
- Use->getOperand(2).Val == Node) {
- unsigned DestReg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
- if (MRegisterInfo::isVirtualRegister(DestReg)) {
- VRBase = DestReg;
- break;
- }
- }
- }
-
- // Figure out the register class to create for the destreg.
- const TargetRegisterClass *TRC = 0;
- if (VRBase) {
- TRC = RegMap->getRegClass(VRBase);
- } else {
- TRC = getPhysicalRegisterRegClass(MRI, Node->getValueType(0), SrcReg);
-
- // Create the reg, emit the copy.
- VRBase = RegMap->createVirtualRegister(TRC);
- }
- MRI->copyRegToReg(*BB, BB->end(), VRBase, SrcReg, TRC);
-
- bool isNew = VRBaseMap.insert(std::make_pair(SDOperand(Node,0), VRBase));
- assert(isNew && "Node emitted out of order - early");
- break;
- }
- case ISD::INLINEASM: {
- unsigned NumOps = Node->getNumOperands();
- if (Node->getOperand(NumOps-1).getValueType() == MVT::Flag)
- --NumOps; // Ignore the flag operand.
-
- // Create the inline asm machine instruction.
- MachineInstr *MI =
- new MachineInstr(BB, TII->get(TargetInstrInfo::INLINEASM));
-
- // Add the asm string as an external symbol operand.
- const char *AsmStr =
- cast<ExternalSymbolSDNode>(Node->getOperand(1))->getSymbol();
- MI->addExternalSymbolOperand(AsmStr);
-
- // Add all of the operand registers to the instruction.
- for (unsigned i = 2; i != NumOps;) {
- unsigned Flags = cast<ConstantSDNode>(Node->getOperand(i))->getValue();
- unsigned NumVals = Flags >> 3;
-
- MI->addImmOperand(Flags);
- ++i; // Skip the ID value.
-
- switch (Flags & 7) {
- default: assert(0 && "Bad flags!");
- case 1: // Use of register.
- for (; NumVals; --NumVals, ++i) {
- unsigned Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg();
- MI->addRegOperand(Reg, false);
- }
- break;
- case 2: // Def of register.
- for (; NumVals; --NumVals, ++i) {
- unsigned Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg();
- MI->addRegOperand(Reg, true);
- }
- break;
- case 3: { // Immediate.
- assert(NumVals == 1 && "Unknown immediate value!");
- if (ConstantSDNode *CS=dyn_cast<ConstantSDNode>(Node->getOperand(i))){
- MI->addImmOperand(CS->getValue());
- } else {
- GlobalAddressSDNode *GA =
- cast<GlobalAddressSDNode>(Node->getOperand(i));
- MI->addGlobalAddressOperand(GA->getGlobal(), GA->getOffset());
- }
- ++i;
- break;
- }
- case 4: // Addressing mode.
- // The addressing mode has been selected, just add all of the
- // operands to the machine instruction.
- for (; NumVals; --NumVals, ++i)
- AddOperand(MI, Node->getOperand(i), 0, 0, VRBaseMap);
- break;
- }
- }
- break;
- }
- }
- }
-}
-
-void ScheduleDAG::EmitNoop() {
- TII->insertNoop(*BB, BB->end());
-}
-
-/// EmitSchedule - Emit the machine code in scheduled order.
-void ScheduleDAG::EmitSchedule() {
- // If this is the first basic block in the function, and if it has live ins
- // that need to be copied into vregs, emit the copies into the top of the
- // block before emitting the code for the block.
- MachineFunction &MF = DAG.getMachineFunction();
- if (&MF.front() == BB && MF.livein_begin() != MF.livein_end()) {
- for (MachineFunction::livein_iterator LI = MF.livein_begin(),
- E = MF.livein_end(); LI != E; ++LI)
- if (LI->second)
- MRI->copyRegToReg(*MF.begin(), MF.begin()->end(), LI->second,
- LI->first, RegMap->getRegClass(LI->second));
- }
-
-
- // Finally, emit the code for all of the scheduled instructions.
- DenseMap<SDOperand, unsigned> VRBaseMap;
- for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
- if (SUnit *SU = Sequence[i]) {
- for (unsigned j = 0, ee = SU->FlaggedNodes.size(); j != ee; j++)
- EmitNode(SU->FlaggedNodes[j], VRBaseMap);
- EmitNode(SU->Node, VRBaseMap);
- } else {
- // Null SUnit* is a noop.
- EmitNoop();
- }
- }
-}
/// dump - dump the schedule.
void ScheduleDAG::dumpSchedule() const {
/// Run - perform scheduling.
///
-MachineBasicBlock *ScheduleDAG::Run() {
- TII = TM.getInstrInfo();
- MRI = TM.getRegisterInfo();
- RegMap = BB->getParent()->getSSARegMap();
- ConstPool = BB->getParent()->getConstantPool();
-
+void ScheduleDAG::Run() {
Schedule();
- return BB;
+
+ DOUT << "*** Final schedule ***\n";
+ DEBUG(dumpSchedule());
+ DOUT << "\n";
}
/// SUnit - Scheduling unit. It's an wrapper around either a single SDNode or
/// a group of nodes flagged together.
void SUnit::dump(const SelectionDAG *G) const {
cerr << "SU(" << NodeNum << "): ";
- Node->dump(G);
+ if (Node)
+ Node->dump(G);
+ else
+ cerr << "CROSS RC COPY ";
cerr << "\n";
if (FlaggedNodes.size() != 0) {
for (unsigned i = 0, e = FlaggedNodes.size(); i != e; i++) {
cerr << " # preds left : " << NumPredsLeft << "\n";
cerr << " # succs left : " << NumSuccsLeft << "\n";
- cerr << " # chain preds left : " << NumChainPredsLeft << "\n";
- cerr << " # chain succs left : " << NumChainSuccsLeft << "\n";
cerr << " Latency : " << Latency << "\n";
cerr << " Depth : " << Depth << "\n";
cerr << " Height : " << Height << "\n";
cerr << " Predecessors:\n";
for (SUnit::const_succ_iterator I = Preds.begin(), E = Preds.end();
I != E; ++I) {
- if (I->second)
+ if (I->isCtrl)
cerr << " ch #";
else
cerr << " val #";
- cerr << I->first << " - SU(" << I->first->NodeNum << ")\n";
+ cerr << I->Dep << " - SU(" << I->Dep->NodeNum << ")";
+ if (I->isSpecial)
+ cerr << " *";
+ cerr << "\n";
}
}
if (Succs.size() != 0) {
cerr << " Successors:\n";
for (SUnit::const_succ_iterator I = Succs.begin(), E = Succs.end();
I != E; ++I) {
- if (I->second)
+ if (I->isCtrl)
cerr << " ch #";
else
cerr << " val #";
- cerr << I->first << " - SU(" << I->first->NodeNum << ")\n";
+ cerr << I->Dep << " - SU(" << I->Dep->NodeNum << ")";
+ if (I->isSpecial)
+ cerr << " *";
+ cerr << "\n";
}
}
cerr << "\n";