//
// 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 "sched"
-#include "llvm/CodeGen/MachineConstantPool.h"
-#include "llvm/CodeGen/MachineFunction.h"
+#define DEBUG_TYPE "pre-RA-sched"
#include "llvm/CodeGen/ScheduleDAG.h"
-#include "llvm/CodeGen/SSARegMap.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetInstrInfo.h"
-#include "llvm/Target/TargetInstrItineraries.h"
-#include "llvm/Target/TargetLowering.h"
+#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Support/Debug.h"
-#include <iostream>
using namespace llvm;
-
-/// 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.)
-static unsigned 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 an optional chain operand, then flag operands. Compute the
-/// number of actual operands that will go into the machine instr.
-static unsigned CountOperands(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;
+ScheduleDAG::ScheduleDAG(SelectionDAG *dag, MachineBasicBlock *bb,
+ const TargetMachine &tm)
+ : DAG(dag), BB(bb), TM(tm), MRI(BB->getParent()->getRegInfo()) {
+ TII = TM.getInstrInfo();
+ MF = BB->getParent();
+ TRI = TM.getRegisterInfo();
+ TLI = TM.getTargetLowering();
+ ConstPool = MF->getConstantPool();
}
-/// PrepareNodeInfo - Set up the basic minimum node info for scheduling.
-///
-void ScheduleDAG::PrepareNodeInfo() {
- // Allocate node information
- Info = new NodeInfo[NodeCount];
-
- unsigned i = 0;
- for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
- E = DAG.allnodes_end(); I != E; ++I, ++i) {
- // Fast reference to node schedule info
- NodeInfo* NI = &Info[i];
- // Set up map
- Map[I] = NI;
- // Set node
- NI->Node = I;
- // Set pending visit count
- NI->setPending(I->use_size());
+/// 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();
+ }
}
}
-/// IdentifyGroups - Put flagged nodes into groups.
-///
-void ScheduleDAG::IdentifyGroups() {
- for (unsigned i = 0, N = NodeCount; i < N; i++) {
- NodeInfo* NI = &Info[i];
- SDNode *Node = NI->Node;
-
- // For each operand (in reverse to only look at flags)
- for (unsigned N = Node->getNumOperands(); 0 < N--;) {
- // Get operand
- SDOperand Op = Node->getOperand(N);
- // No more flags to walk
- if (Op.getValueType() != MVT::Flag) break;
- // Add to node group
- NodeGroup::Add(getNI(Op.Val), NI);
- // Let everyone else know
- HasGroups = true;
- }
- }
+SUnit *ScheduleDAG::Clone(SUnit *Old) {
+ SUnit *SU = NewSUnit(Old->getNode());
+ SU->OrigNode = Old->OrigNode;
+ SU->Latency = Old->Latency;
+ SU->isTwoAddress = Old->isTwoAddress;
+ SU->isCommutable = Old->isCommutable;
+ SU->hasPhysRegDefs = Old->hasPhysRegDefs;
+ return SU;
}
-static unsigned CreateVirtualRegisters(MachineInstr *MI,
- unsigned NumResults,
- SSARegMap *RegMap,
- const TargetInstrDescriptor &II) {
- // Create the result registers for this node and add the result regs to
- // the machine instruction.
- const TargetOperandInfo *OpInfo = II.OpInfo;
- unsigned ResultReg = RegMap->createVirtualRegister(OpInfo[0].RegClass);
- MI->addRegOperand(ResultReg, MachineOperand::Def);
- for (unsigned i = 1; i != NumResults; ++i) {
- assert(OpInfo[i].RegClass && "Isn't a register operand!");
- MI->addRegOperand(RegMap->createVirtualRegister(OpInfo[i].RegClass),
- MachineOperand::Def);
+
+/// 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 ScheduleDAG::BuildSchedUnits() {
+ // For post-regalloc scheduling, build the SUnits from the MachineInstrs
+ // in the MachineBasicBlock.
+ if (!DAG) {
+ BuildSchedUnitsFromMBB();
+ return;
}
- return ResultReg;
-}
-/// EmitNode - Generate machine code for an node and needed dependencies.
-///
-void ScheduleDAG::EmitNode(NodeInfo *NI) {
- unsigned VRBase = 0; // First virtual register for node
- SDNode *Node = NI->Node;
+ // 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());
- // If machine instruction
- if (Node->isTargetOpcode()) {
- unsigned Opc = Node->getTargetOpcode();
- 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.numOperands == -1)&&
- "#operands for dag node doesn't match .td file!");
-#endif
-
- // Create the new machine instruction.
- MachineInstr *MI = new MachineInstr(Opc, NumMIOperands, true, true);
+ // 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;
- // Add result register values for things that are defined by this
- // instruction.
+ // If this node has already been processed, stop now.
+ if (NI->getNodeId() != -1) continue;
- // 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 (NumResults == 1) {
- 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 Reg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
- if (MRegisterInfo::isVirtualRegister(Reg)) {
- VRBase = Reg;
- MI->addRegOperand(Reg, MachineOperand::Def);
- break;
- }
+ SUnit *NodeSUnit = NewSUnit(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.
+
+ // 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);
+ }
+
+ // Scan down to find any flagged succs.
+ N = NI;
+ while (N->getValueType(N->getNumValues()-1) == MVT::Flag) {
+ 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.isOperandOf(*UI)) {
+ HasFlagUse = true;
+ assert(N->getNodeId() == -1 && "Node already inserted!");
+ N->setNodeId(NodeSUnit->NodeNum);
+ N = *UI;
+ break;
}
- }
+ if (!HasFlagUse) break;
}
- // Otherwise, create new virtual registers.
- if (NumResults && VRBase == 0)
- VRBase = CreateVirtualRegisters(MI, NumResults, RegMap, II);
+ // If there are flag operands involved, N is now the bottom-most node
+ // of the sequence of nodes that are flagged together.
+ // Update the SUnit.
+ NodeSUnit->setNode(N);
+ assert(N->getNodeId() == -1 && "Node already inserted!");
+ N->setNodeId(NodeSUnit->NodeNum);
+
+ ComputeLatency(NodeSUnit);
+ }
+
+ // Pass 2: add the preds, succs, etc.
+ for (unsigned su = 0, e = SUnits.size(); su != e; ++su) {
+ SUnit *SU = &SUnits[su];
+ SDNode *MainNode = SU->getNode();
- // Emit all of the actual operands of this instruction, adding them to the
- // instruction as appropriate.
- for (unsigned i = 0; i != NodeOperands; ++i) {
- if (Node->getOperand(i).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(Node->getOperand(i).getValueType() != MVT::Other &&
- Node->getOperand(i).getValueType() != MVT::Flag &&
- "Chain and flag operands should occur at end of operand list!");
-
- // Get/emit the operand.
- unsigned VReg = getVR(Node->getOperand(i));
- MI->addRegOperand(VReg, MachineOperand::Use);
-
- // Verify that it is right.
- assert(MRegisterInfo::isVirtualRegister(VReg) && "Not a vreg?");
- assert(II.OpInfo[i+NumResults].RegClass &&
- "Don't have operand info for this instruction!");
- assert(RegMap->getRegClass(VReg) == II.OpInfo[i+NumResults].RegClass &&
- "Register class of operand and regclass of use don't agree!");
- } else if (ConstantSDNode *C =
- dyn_cast<ConstantSDNode>(Node->getOperand(i))) {
- MI->addZeroExtImm64Operand(C->getValue());
- } else if (RegisterSDNode*R =
- dyn_cast<RegisterSDNode>(Node->getOperand(i))) {
- MI->addRegOperand(R->getReg(), MachineOperand::Use);
- } else if (GlobalAddressSDNode *TGA =
- dyn_cast<GlobalAddressSDNode>(Node->getOperand(i))) {
- MI->addGlobalAddressOperand(TGA->getGlobal(), false, TGA->getOffset());
- } else if (BasicBlockSDNode *BB =
- dyn_cast<BasicBlockSDNode>(Node->getOperand(i))) {
- MI->addMachineBasicBlockOperand(BB->getBasicBlock());
- } else if (FrameIndexSDNode *FI =
- dyn_cast<FrameIndexSDNode>(Node->getOperand(i))) {
- MI->addFrameIndexOperand(FI->getIndex());
- } else if (ConstantPoolSDNode *CP =
- dyn_cast<ConstantPoolSDNode>(Node->getOperand(i))) {
- unsigned Idx = ConstPool->getConstantPoolIndex(CP->get(),
- CP->getAlignment());
- MI->addConstantPoolIndexOperand(Idx);
- } else if (ExternalSymbolSDNode *ES =
- dyn_cast<ExternalSymbolSDNode>(Node->getOperand(i))) {
- MI->addExternalSymbolOperand(ES->getSymbol(), false);
- } else {
- assert(Node->getOperand(i).getValueType() != MVT::Other &&
- Node->getOperand(i).getValueType() != MVT::Flag &&
- "Chain and flag operands should occur at end of operand list!");
- unsigned VReg = getVR(Node->getOperand(i));
- MI->addRegOperand(VReg, MachineOperand::Use);
-
- // Verify that it is right.
- assert(MRegisterInfo::isVirtualRegister(VReg) && "Not a vreg?");
- assert(II.OpInfo[i+NumResults].RegClass &&
- "Don't have operand info for this instruction!");
- assert(RegMap->getRegClass(VReg) == II.OpInfo[i+NumResults].RegClass &&
- "Register class of operand and regclass of use don't agree!");
+ 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 (TID.isCommutable())
+ SU->isCommutable = true;
}
- // 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:
- Node->dump();
- assert(0 && "This target-independent node should have been selected!");
- case ISD::EntryToken: // fall thru
- case ISD::TokenFactor:
- break;
- case ISD::CopyToReg: {
- unsigned InReg = getVR(Node->getOperand(2));
- unsigned DestReg = cast<RegisterSDNode>(Node->getOperand(1))->getReg();
- if (InReg != DestReg) // Coallesced away the copy?
- MRI->copyRegToReg(*BB, BB->end(), DestReg, InReg,
- RegMap->getRegClass(InReg));
- break;
+ // 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;
+
+ for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
+ SDNode *OpN = N->getOperand(i).getNode();
+ if (isPassiveNode(OpN)) continue; // Not scheduled.
+ SUnit *OpSU = &SUnits[OpN->getNodeId()];
+ assert(OpSU && "Node has no SUnit!");
+ if (OpSU == SU) continue; // In the same group.
+
+ MVT 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, Cost);
+ SU->addPred(OpSU, isChain, false, PhysReg, Cost);
+ }
}
- case ISD::CopyFromReg: {
- unsigned SrcReg = cast<RegisterSDNode>(Node->getOperand(1))->getReg();
- if (MRegisterInfo::isVirtualRegister(SrcReg)) {
- VRBase = SrcReg; // Just use the input register directly!
- break;
+ }
+}
+
+void ScheduleDAG::BuildSchedUnitsFromMBB() {
+ SUnits.clear();
+ SUnits.reserve(BB->size());
+
+ std::vector<SUnit *> PendingLoads;
+ SUnit *Terminator = 0;
+ SUnit *Chain = 0;
+ SUnit *Defs[TargetRegisterInfo::FirstVirtualRegister] = {};
+ std::vector<SUnit *> Uses[TargetRegisterInfo::FirstVirtualRegister] = {};
+ int Cost = 1; // FIXME
+
+ for (MachineBasicBlock::iterator MII = BB->end(), MIE = BB->begin();
+ MII != MIE; --MII) {
+ MachineInstr *MI = prior(MII);
+ SUnit *SU = NewSUnit(MI);
+
+ for (unsigned j = 0, n = MI->getNumOperands(); j != n; ++j) {
+ const MachineOperand &MO = MI->getOperand(j);
+ if (!MO.isReg()) continue;
+ unsigned Reg = MO.getReg();
+ if (Reg == 0) continue;
+
+ assert(TRI->isPhysicalRegister(Reg) && "Virtual register encountered!");
+ std::vector<SUnit *> &UseList = Uses[Reg];
+ SUnit *&Def = Defs[Reg];
+ // Optionally add output and anti dependences
+ if (Def && Def != SU)
+ Def->addPred(SU, /*isCtrl=*/true, /*isSpecial=*/false,
+ /*PhyReg=*/Reg, Cost);
+ for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
+ SUnit *&Def = Defs[*Alias];
+ if (Def && Def != SU)
+ Def->addPred(SU, /*isCtrl=*/true, /*isSpecial=*/false,
+ /*PhyReg=*/*Alias, Cost);
}
- // 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;
- }
+ if (MO.isDef()) {
+ // Add any data dependencies.
+ for (unsigned i = 0, e = UseList.size(); i != e; ++i)
+ if (UseList[i] != SU)
+ UseList[i]->addPred(SU, /*isCtrl=*/false, /*isSpecial=*/false,
+ /*PhysReg=*/Reg, Cost);
+ for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
+ std::vector<SUnit *> &UseList = Uses[*Alias];
+ for (unsigned i = 0, e = UseList.size(); i != e; ++i)
+ if (UseList[i] != SU)
+ UseList[i]->addPred(SU, /*isCtrl=*/false, /*isSpecial=*/false,
+ /*PhysReg=*/*Alias, Cost);
}
- }
- // Figure out the register class to create for the destreg.
- const TargetRegisterClass *TRC = 0;
- if (VRBase) {
- TRC = RegMap->getRegClass(VRBase);
+ UseList.clear();
+ Def = SU;
} else {
-
- // 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(Node->getValueType(0)) &&
- (*I)->contains(SrcReg)) {
- TRC = *I;
- break;
- }
- assert(TRC && "Couldn't find register class for reg copy!");
-
- // Create the reg, emit the copy.
- VRBase = RegMap->createVirtualRegister(TRC);
- }
- MRI->copyRegToReg(*BB, BB->end(), VRBase, SrcReg, TRC);
- 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, TargetInstrInfo::INLINEASM, (NumOps-2)/2+1);
-
- // Add the asm string as an external symbol operand.
- const char *AsmStr =
- cast<ExternalSymbolSDNode>(Node->getOperand(1))->getSymbol();
- MI->addExternalSymbolOperand(AsmStr, false);
-
- // Add all of the operand registers to the instruction.
- for (unsigned i = 2; i != NumOps; i += 2) {
- unsigned Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg();
- unsigned Flags =cast<ConstantSDNode>(Node->getOperand(i+1))->getValue();
- MachineOperand::UseType UseTy;
- switch (Flags) {
- default: assert(0 && "Bad flags!");
- case 1: UseTy = MachineOperand::Use; break;
- case 2: UseTy = MachineOperand::Def; break;
- case 3: UseTy = MachineOperand::UseAndDef; break;
- }
- MI->addMachineRegOperand(Reg, UseTy);
+ UseList.push_back(SU);
}
- break;
}
+ bool False = false;
+ bool True = true;
+ if (!MI->isSafeToMove(TII, False)) {
+ if (Chain)
+ Chain->addPred(SU, /*isCtrl=*/false, /*isSpecial=*/false);
+ for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
+ PendingLoads[k]->addPred(SU, /*isCtrl=*/false, /*isSpecial=*/false);
+ PendingLoads.clear();
+ Chain = SU;
+ } else if (!MI->isSafeToMove(TII, True)) {
+ if (Chain)
+ Chain->addPred(SU, /*isCtrl=*/false, /*isSpecial=*/false);
+ PendingLoads.push_back(SU);
}
+ if (Terminator && SU->Succs.empty())
+ Terminator->addPred(SU, /*isCtrl=*/false, /*isSpecial=*/false);
+ if (MI->getDesc().isTerminator())
+ Terminator = SU;
}
-
- assert(NI->VRBase == 0 && "Node emitted out of order - early");
- NI->VRBase = VRBase;
}
-/// EmitAll - Emit all nodes in schedule sorted order.
-///
-void ScheduleDAG::EmitAll() {
- // For each node in the ordering
- for (unsigned i = 0, N = Ordering.size(); i < N; i++) {
- // Get the scheduling info
- NodeInfo *NI = Ordering[i];
- if (NI->isInGroup()) {
- NodeGroupIterator NGI(Ordering[i]);
- while (NodeInfo *NI = NGI.next()) EmitNode(NI);
- } else {
- EmitNode(NI);
+void ScheduleDAG::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;
+ for (SDNode *N = SU->getNode(); N; N = N->getFlaggedNode()) {
+ if (N->isMachineOpcode()) {
+ unsigned SchedClass = TII->get(N->getMachineOpcode()).getSchedClass();
+ const InstrStage *S = InstrItins.begin(SchedClass);
+ const InstrStage *E = InstrItins.end(SchedClass);
+ for (; S != E; ++S)
+ SU->Latency += S->Cycles;
}
}
}
-/// isFlagDefiner - Returns true if the node defines a flag result.
-static bool isFlagDefiner(SDNode *A) {
- unsigned N = A->getNumValues();
- return N && A->getValueType(N - 1) == MVT::Flag;
-}
+/// CalculateDepths - compute depths using algorithms for the longest
+/// paths in the DAG
+void ScheduleDAG::CalculateDepths() {
+ 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);
+ }
+ }
-/// isFlagUser - Returns true if the node uses a flag result.
-///
-static bool isFlagUser(SDNode *A) {
- unsigned N = A->getNumOperands();
- return N && A->getOperand(N - 1).getValueType() == MVT::Flag;
-}
+ // Process nodes in the topological order
+ while (!WorkList.empty()) {
+ SUnit *SU = WorkList.back();
+ WorkList.pop_back();
+ 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;
+ }
+ }
-/// printNI - Print node info.
-///
-void ScheduleDAG::printNI(std::ostream &O, NodeInfo *NI) const {
-#ifndef NDEBUG
- SDNode *Node = NI->Node;
- O << " "
- << std::hex << Node << std::dec
- << ", Lat=" << NI->Latency
- << ", Slot=" << NI->Slot
- << ", ARITY=(" << Node->getNumOperands() << ","
- << Node->getNumValues() << ")"
- << " " << Node->getOperationName(&DAG);
- if (isFlagDefiner(Node)) O << "<#";
- if (isFlagUser(Node)) O << ">#";
-#endif
-}
+ SU->Depth = SUDepth;
-/// printChanges - Hilight changes in order caused by scheduling.
-///
-void ScheduleDAG::printChanges(unsigned Index) const {
-#ifndef NDEBUG
- // Get the ordered node count
- unsigned N = Ordering.size();
- // Determine if any changes
- unsigned i = 0;
- for (; i < N; i++) {
- NodeInfo *NI = Ordering[i];
- if (NI->Preorder != i) break;
- }
-
- if (i < N) {
- std::cerr << Index << ". New Ordering\n";
-
- for (i = 0; i < N; i++) {
- NodeInfo *NI = Ordering[i];
- std::cerr << " " << NI->Preorder << ". ";
- printNI(std::cerr, NI);
- std::cerr << "\n";
- if (NI->isGroupDominator()) {
- NodeGroup *Group = NI->Group;
- for (NIIterator NII = Group->group_begin(), E = Group->group_end();
- NII != E; NII++) {
- std::cerr << " ";
- printNI(std::cerr, *NII);
- std::cerr << "\n";
- }
- }
+ // 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);
}
- } else {
- std::cerr << Index << ". No Changes\n";
}
-#endif
}
-/// print - Print ordering to specified output stream.
-///
-void ScheduleDAG::print(std::ostream &O) const {
-#ifndef NDEBUG
- using namespace std;
- O << "Ordering\n";
- for (unsigned i = 0, N = Ordering.size(); i < N; i++) {
- NodeInfo *NI = Ordering[i];
- printNI(O, NI);
- O << "\n";
- if (NI->isGroupDominator()) {
- NodeGroup *Group = NI->Group;
- for (NIIterator NII = Group->group_begin(), E = Group->group_end();
- NII != E; NII++) {
- O << " ";
- printNI(O, *NII);
- O << "\n";
+/// CalculateHeights - compute heights using algorithms for the longest
+/// paths in the DAG
+void ScheduleDAG::CalculateHeights() {
+ 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();
+ WorkList.pop_back();
+ 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);
+ }
}
-#endif
}
-void ScheduleDAG::dump(const char *tag) const {
- std::cerr << tag; dump();
+/// 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 ScheduleDAG::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;
}
-void ScheduleDAG::dump() const {
- print(std::cerr);
+/// 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;
}
-/// Run - perform scheduling.
-///
-MachineBasicBlock *ScheduleDAG::Run() {
- TII = TM.getInstrInfo();
- MRI = TM.getRegisterInfo();
- RegMap = BB->getParent()->getSSARegMap();
- ConstPool = BB->getParent()->getConstantPool();
+/// 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;
+ if (N && Node->getOperand(N - 1).getValueType() == MVT::Other)
+ --N; // Ignore chain if it exists.
+ return N;
+}
- // Number the nodes
- NodeCount = std::distance(DAG.allnodes_begin(), DAG.allnodes_end());
- // Set up minimum info for scheduling
- PrepareNodeInfo();
- // Construct node groups for flagged nodes
- IdentifyGroups();
- Schedule();
- return BB;
+/// dump - dump the schedule.
+void ScheduleDAG::dumpSchedule() const {
+ for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
+ if (SUnit *SU = Sequence[i])
+ SU->dump(this);
+ else
+ cerr << "**** NOOP ****\n";
+ }
}
-/// CountInternalUses - Returns the number of edges between the two nodes.
+/// Run - perform scheduling.
///
-static unsigned CountInternalUses(NodeInfo *D, NodeInfo *U) {
- unsigned N = 0;
- for (unsigned M = U->Node->getNumOperands(); 0 < M--;) {
- SDOperand Op = U->Node->getOperand(M);
- if (Op.Val == D->Node) N++;
- }
+void ScheduleDAG::Run() {
+ Schedule();
+
+ DOUT << "*** Final schedule ***\n";
+ DEBUG(dumpSchedule());
+ DOUT << "\n";
+}
- return N;
+/// SUnit - Scheduling unit. It's an wrapper around either a single SDNode or
+/// a group of nodes flagged together.
+void SUnit::dump(const ScheduleDAG *G) const {
+ cerr << "SU(" << NodeNum << "): ";
+ if (getNode())
+ getNode()->dump(G->DAG);
+ else
+ cerr << "CROSS RC COPY ";
+ cerr << "\n";
+ SmallVector<SDNode *, 4> FlaggedNodes;
+ for (SDNode *N = getNode()->getFlaggedNode(); N; N = N->getFlaggedNode())
+ FlaggedNodes.push_back(N);
+ while (!FlaggedNodes.empty()) {
+ cerr << " ";
+ FlaggedNodes.back()->dump(G->DAG);
+ cerr << "\n";
+ FlaggedNodes.pop_back();
+ }
}
-//===----------------------------------------------------------------------===//
-/// Add - Adds a definer and user pair to a node group.
-///
-void NodeGroup::Add(NodeInfo *D, NodeInfo *U) {
- // Get current groups
- NodeGroup *DGroup = D->Group;
- NodeGroup *UGroup = U->Group;
- // If both are members of groups
- if (DGroup && UGroup) {
- // There may have been another edge connecting
- if (DGroup == UGroup) return;
- // Add the pending users count
- DGroup->addPending(UGroup->getPending());
- // For each member of the users group
- NodeGroupIterator UNGI(U);
- while (NodeInfo *UNI = UNGI.next() ) {
- // Change the group
- UNI->Group = DGroup;
- // For each member of the definers group
- NodeGroupIterator DNGI(D);
- while (NodeInfo *DNI = DNGI.next() ) {
- // Remove internal edges
- DGroup->addPending(-CountInternalUses(DNI, UNI));
- }
- }
- // Merge the two lists
- DGroup->group_insert(DGroup->group_end(),
- UGroup->group_begin(), UGroup->group_end());
- } else if (DGroup) {
- // Make user member of definers group
- U->Group = DGroup;
- // Add users uses to definers group pending
- DGroup->addPending(U->Node->use_size());
- // For each member of the definers group
- NodeGroupIterator DNGI(D);
- while (NodeInfo *DNI = DNGI.next() ) {
- // Remove internal edges
- DGroup->addPending(-CountInternalUses(DNI, U));
+void SUnit::dumpAll(const ScheduleDAG *G) const {
+ dump(G);
+
+ cerr << " # preds left : " << NumPredsLeft << "\n";
+ cerr << " # succs left : " << NumSuccsLeft << "\n";
+ cerr << " Latency : " << Latency << "\n";
+ cerr << " Depth : " << Depth << "\n";
+ cerr << " Height : " << Height << "\n";
+
+ if (Preds.size() != 0) {
+ cerr << " Predecessors:\n";
+ for (SUnit::const_succ_iterator I = Preds.begin(), E = Preds.end();
+ I != E; ++I) {
+ if (I->isCtrl)
+ cerr << " ch #";
+ else
+ cerr << " val #";
+ cerr << I->Dep << " - SU(" << I->Dep->NodeNum << ")";
+ if (I->isSpecial)
+ cerr << " *";
+ cerr << "\n";
}
- DGroup->group_push_back(U);
- } else if (UGroup) {
- // Make definer member of users group
- D->Group = UGroup;
- // Add definers uses to users group pending
- UGroup->addPending(D->Node->use_size());
- // For each member of the users group
- NodeGroupIterator UNGI(U);
- while (NodeInfo *UNI = UNGI.next() ) {
- // Remove internal edges
- UGroup->addPending(-CountInternalUses(D, UNI));
+ }
+ if (Succs.size() != 0) {
+ cerr << " Successors:\n";
+ for (SUnit::const_succ_iterator I = Succs.begin(), E = Succs.end();
+ I != E; ++I) {
+ if (I->isCtrl)
+ cerr << " ch #";
+ else
+ cerr << " val #";
+ cerr << I->Dep << " - SU(" << I->Dep->NodeNum << ")";
+ if (I->isSpecial)
+ cerr << " *";
+ cerr << "\n";
}
- UGroup->group_insert(UGroup->group_begin(), D);
- } else {
- D->Group = U->Group = DGroup = new NodeGroup();
- DGroup->addPending(D->Node->use_size() + U->Node->use_size() -
- CountInternalUses(D, U));
- DGroup->group_push_back(D);
- DGroup->group_push_back(U);
}
+ cerr << "\n";
}