X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FSelectionDAG%2FScheduleDAG.cpp;h=5a2b4ede27807ac5156e4a0f924519c972fcda43;hb=02b6d25a2702d8857b82d333f290550e3c6ec4dc;hp=13465957b986ab7c5f866e73f466984a76b90133;hpb=713a98dee8ab07a3066d1707a07648d27dd0c19c;p=oota-llvm.git diff --git a/lib/CodeGen/SelectionDAG/ScheduleDAG.cpp b/lib/CodeGen/SelectionDAG/ScheduleDAG.cpp index 13465957b98..5a2b4ede278 100644 --- a/lib/CodeGen/SelectionDAG/ScheduleDAG.cpp +++ b/lib/CodeGen/SelectionDAG/ScheduleDAG.cpp @@ -2,8 +2,8 @@ // // 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. // //===----------------------------------------------------------------------===// // @@ -14,19 +14,73 @@ //===----------------------------------------------------------------------===// #define DEBUG_TYPE "pre-RA-sched" +#include "llvm/Constants.h" #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/CodeGen/MachineRegisterInfo.h" #include "llvm/Target/TargetData.h" #include "llvm/Target/TargetMachine.h" #include "llvm/Target/TargetInstrInfo.h" #include "llvm/Target/TargetLowering.h" +#include "llvm/ADT/Statistic.h" #include "llvm/Support/Debug.h" #include "llvm/Support/MathExtras.h" using namespace llvm; +STATISTIC(NumCommutes, "Number of instructions commuted"); + +ScheduleDAG::ScheduleDAG(SelectionDAG &dag, MachineBasicBlock *bb, + const TargetMachine &tm) + : DAG(dag), BB(bb), TM(tm), RegInfo(BB->getParent()->getRegInfo()) { + TII = TM.getInstrInfo(); + MF = &DAG.getMachineFunction(); + TRI = TM.getRegisterInfo(); + 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 *Use, unsigned Op, + const TargetRegisterInfo *TRI, + const TargetInstrInfo *TII, + unsigned &PhysReg, int &Cost) { + if (Op != 2 || Use->getOpcode() != ISD::CopyToReg) + return; + + unsigned Reg = cast(Use->getOperand(1))->getReg(); + if (TargetRegisterInfo::isVirtualRegister(Reg)) + return; + + unsigned ResNo = Use->getOperand(2).ResNo; + if (Def->isTargetOpcode()) { + const TargetInstrDesc &II = TII->get(Def->getTargetOpcode()); + if (ResNo >= II.getNumDefs() && + II.ImplicitDefs[ResNo - II.getNumDefs()] == Reg) { + PhysReg = Reg; + const TargetRegisterClass *RC = + TRI->getPhysicalRegisterRegClass(Def->getValueType(ResNo), Reg); + Cost = RC->getCopyCost(); + } + } +} + +SUnit *ScheduleDAG::Clone(SUnit *Old) { + SUnit *SU = NewSUnit(Old->Node); + for (unsigned i = 0, e = SU->FlaggedNodes.size(); i != e; ++i) + SU->FlaggedNodes.push_back(SU->FlaggedNodes[i]); + SU->InstanceNo = SUnitMap[Old->Node].size(); + SU->Latency = Old->Latency; + SU->isTwoAddress = Old->isTwoAddress; + SU->isCommutable = Old->isCommutable; + SU->hasPhysRegDefs = Old->hasPhysRegDefs; + SUnitMap[Old->Node].push_back(SU); + 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. @@ -36,15 +90,13 @@ void ScheduleDAG::BuildSchedUnits() { // invalidated. SUnits.reserve(std::distance(DAG.allnodes_begin(), DAG.allnodes_end())); - const InstrItineraryData &InstrItins = TM.getInstrItineraryData(); - 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 (SUnitMap[NI].size()) continue; SUnit *NodeSUnit = NewSUnit(NI); @@ -59,7 +111,7 @@ void ScheduleDAG::BuildSchedUnits() { do { N = N->getOperand(N->getNumOperands()-1).Val; NodeSUnit->FlaggedNodes.push_back(N); - SUnitMap[N] = NodeSUnit; + SUnitMap[N].push_back(NodeSUnit); } while (N->getNumOperands() && N->getOperand(N->getNumOperands()-1).getValueType()== MVT::Flag); std::reverse(NodeSUnit->FlaggedNodes.begin(), @@ -76,10 +128,10 @@ void ScheduleDAG::BuildSchedUnits() { 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; + SUnitMap[N].push_back(NodeSUnit); N = *UI; break; } @@ -89,33 +141,9 @@ void ScheduleDAG::BuildSchedUnits() { // 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; - } - } - } + SUnitMap[N].push_back(NodeSUnit); + + ComputeLatency(NodeSUnit); } // Pass 2: add the preds, succs, etc. @@ -125,13 +153,14 @@ void ScheduleDAG::BuildSchedUnits() { 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) { + 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; } @@ -141,34 +170,27 @@ void ScheduleDAG::BuildSchedUnits() { for (unsigned n = 0, e = SU->FlaggedNodes.size(); n != e; ++n) { SDNode *N = SU->FlaggedNodes[n]; + if (N->isTargetOpcode() && + TII->get(N->getTargetOpcode()).getImplicitDefs() && + CountResults(N) > TII->get(N->getTargetOpcode()).getNumDefs()) + SU->hasPhysRegDefs = true; for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { SDNode *OpN = N->getOperand(i).Val; if (isPassiveNode(OpN)) continue; // Not scheduled. - SUnit *OpSU = SUnitMap[OpN]; + SUnit *OpSU = SUnitMap[OpN].front(); assert(OpSU && "Node has no SUnit!"); if (OpSU == SU) continue; // In the same group. MVT::ValueType 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); } } @@ -179,46 +201,149 @@ void ScheduleDAG::BuildSchedUnits() { return; } +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; + } else { + SU->Latency = 0; + if (SU->Node->isTargetOpcode()) { + unsigned SchedClass = + TII->get(SU->Node->getTargetOpcode()).getSchedClass(); + InstrStage *S = InstrItins.begin(SchedClass); + 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->isTargetOpcode()) { + unsigned SchedClass =TII->get(FNode->getTargetOpcode()).getSchedClass(); + InstrStage *S = InstrItins.begin(SchedClass); + 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 > WorkList; - for (unsigned i = 0, e = SUnits.size(); i != e; ++i) - if (SUnits[i].Preds.size() == 0) - WorkList.push_back(std::make_pair(&SUnits[i], 0U)); + unsigned DAGSize = SUnits.size(); + std::vector InDegree(DAGSize); + std::vector WorkList; + WorkList.reserve(DAGSize); + + // Initialize the data structures + for (unsigned i = 0, e = DAGSize; i != e; ++i) { + SUnit *SU = &SUnits[i]; + int NodeNum = SU->NodeNum; + unsigned Degree = SU->Preds.size(); + InDegree[NodeNum] = Degree; + SU->Depth = 0; + // 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->Dep, Depth+1)); + unsigned &SUDepth = SU->Depth; + + // 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; + } + } + + // Update InDegrees 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 (!--InDegree[SU->NodeNum]) + // 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 > WorkList; - SUnit *Root = SUnitMap[DAG.getRoot().Val]; - WorkList.push_back(std::make_pair(Root, 0U)); + unsigned DAGSize = SUnits.size(); + std::vector InDegree(DAGSize); + std::vector WorkList; + WorkList.reserve(DAGSize); + + // Initialize the data structures + for (unsigned i = 0, e = DAGSize; i != e; ++i) { + SUnit *SU = &SUnits[i]; + int NodeNum = SU->NodeNum; + unsigned Degree = SU->Succs.size(); + InDegree[NodeNum] = Degree; + SU->Height = 0; + + // 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->Dep, Height+1)); + unsigned &SUHeight = SU->Height; + + // 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; + } + } + + // Update InDegrees 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 (!--InDegree[SU->NodeNum]) + // 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) @@ -228,10 +353,20 @@ unsigned ScheduleDAG::CountResults(SDNode *Node) { 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 flag operands. 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(Node->getOperand(N - 1).Val)) + --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; @@ -241,40 +376,27 @@ unsigned ScheduleDAG::CountOperands(SDNode *Node) { } static const TargetRegisterClass *getInstrOperandRegClass( - const MRegisterInfo *MRI, + const TargetRegisterInfo *TRI, const TargetInstrInfo *TII, - const TargetInstrDescriptor *II, + const TargetInstrDesc &II, unsigned Op) { - if (Op >= II->numOperands) { - assert((II->Flags & M_VARIABLE_OPS)&& "Invalid operand # of instruction"); + if (Op >= II.getNumOperands()) { + assert(II.isVariadic() && "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); + if (II.OpInfo[Op].isLookupPtrRegClass()) + return TII->getPointerRegClass(); + return TRI->getRegClass(II.OpInfo[Op].RegClass); } -// 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; -} - -void ScheduleDAG::EmitCopyFromReg(SDNode *Node, unsigned ResNo, unsigned SrcReg, +void ScheduleDAG::EmitCopyFromReg(SDNode *Node, unsigned ResNo, + unsigned InstanceNo, unsigned SrcReg, DenseMap &VRBaseMap) { unsigned VRBase = 0; - if (MRegisterInfo::isVirtualRegister(SrcReg)) { + if (TargetRegisterInfo::isVirtualRegister(SrcReg)) { // Just use the input register directly! + if (InstanceNo > 0) + VRBaseMap.erase(SDOperand(Node, ResNo)); bool isNew = VRBaseMap.insert(std::make_pair(SDOperand(Node,ResNo),SrcReg)); assert(isNew && "Node emitted out of order - early"); return; @@ -282,41 +404,67 @@ void ScheduleDAG::EmitCopyFromReg(SDNode *Node, unsigned ResNo, unsigned SrcReg, // 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. + bool MatchReg = true; for (SDNode::use_iterator UI = Node->use_begin(), E = Node->use_end(); UI != E; ++UI) { SDNode *Use = *UI; + bool Match = true; if (Use->getOpcode() == ISD::CopyToReg && Use->getOperand(2).Val == Node && Use->getOperand(2).ResNo == ResNo) { unsigned DestReg = cast(Use->getOperand(1))->getReg(); - if (MRegisterInfo::isVirtualRegister(DestReg)) { + if (TargetRegisterInfo::isVirtualRegister(DestReg)) { VRBase = DestReg; - break; + Match = false; + } else if (DestReg != SrcReg) + Match = false; + } else { + for (unsigned i = 0, e = Use->getNumOperands(); i != e; ++i) { + SDOperand Op = Use->getOperand(i); + if (Op.Val != Node || Op.ResNo != ResNo) + continue; + MVT::ValueType VT = Node->getValueType(Op.ResNo); + if (VT != MVT::Other && VT != MVT::Flag) + Match = false; } } + MatchReg &= Match; + if (VRBase) + break; } + const TargetRegisterClass *SrcRC = 0, *DstRC = 0; + SrcRC = TRI->getPhysicalRegisterRegClass(Node->getValueType(ResNo), SrcReg); + // Figure out the register class to create for the destreg. - const TargetRegisterClass *TRC = 0; if (VRBase) { - TRC = RegMap->getRegClass(VRBase); + DstRC = RegInfo.getRegClass(VRBase); + } else { + DstRC = DAG.getTargetLoweringInfo() + .getRegClassFor(Node->getValueType(ResNo)); + } + + // If all uses are reading from the src physical register and copying the + // register is either impossible or very expensive, then don't create a copy. + if (MatchReg && SrcRC->getCopyCost() < 0) { + VRBase = SrcReg; } else { - TRC = getPhysicalRegisterRegClass(MRI, Node->getValueType(ResNo), SrcReg); - // Create the reg, emit the copy. - VRBase = RegMap->createVirtualRegister(TRC); + VRBase = RegInfo.createVirtualRegister(DstRC); + TII->copyRegToReg(*BB, BB->end(), VRBase, SrcReg, DstRC, SrcRC); } - MRI->copyRegToReg(*BB, BB->end(), VRBase, SrcReg, TRC); + if (InstanceNo > 0) + VRBaseMap.erase(SDOperand(Node, ResNo)); bool isNew = VRBaseMap.insert(std::make_pair(SDOperand(Node,ResNo), VRBase)); assert(isNew && "Node emitted out of order - early"); } void ScheduleDAG::CreateVirtualRegisters(SDNode *Node, MachineInstr *MI, - const TargetInstrDescriptor &II, + const TargetInstrDesc &II, DenseMap &VRBaseMap) { - for (unsigned i = 0; i < II.numDefs; ++i) { + for (unsigned i = 0; i < II.getNumDefs(); ++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. @@ -328,9 +476,9 @@ void ScheduleDAG::CreateVirtualRegisters(SDNode *Node, Use->getOperand(2).Val == Node && Use->getOperand(2).ResNo == i) { unsigned Reg = cast(Use->getOperand(1))->getReg(); - if (MRegisterInfo::isVirtualRegister(Reg)) { + if (TargetRegisterInfo::isVirtualRegister(Reg)) { VRBase = Reg; - MI->addRegOperand(Reg, true); + MI->addOperand(MachineOperand::CreateReg(Reg, true)); break; } } @@ -339,10 +487,10 @@ void ScheduleDAG::CreateVirtualRegisters(SDNode *Node, // Create the result registers for this node and add the result regs to // the machine instruction. if (VRBase == 0) { - const TargetRegisterClass *RC = getInstrOperandRegClass(MRI, TII, &II, i); + const TargetRegisterClass *RC = getInstrOperandRegClass(TRI, TII, II, i); assert(RC && "Isn't a register operand!"); - VRBase = RegMap->createVirtualRegister(RC); - MI->addRegOperand(VRBase, true); + VRBase = RegInfo.createVirtualRegister(RC); + MI->addOperand(MachineOperand::CreateReg(VRBase, true)); } bool isNew = VRBaseMap.insert(std::make_pair(SDOperand(Node,i), VRBase)); @@ -365,7 +513,7 @@ static unsigned getVR(SDOperand Op, DenseMap &VRBaseMap) { /// assertions only. void ScheduleDAG::AddOperand(MachineInstr *MI, SDOperand Op, unsigned IIOpNum, - const TargetInstrDescriptor *II, + const TargetInstrDesc *II, DenseMap &VRBaseMap) { if (Op.isTargetOpcode()) { // Note that this case is redundant with the final else block, but we @@ -377,18 +525,18 @@ void ScheduleDAG::AddOperand(MachineInstr *MI, SDOperand Op, // 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); + const TargetInstrDesc &TID = MI->getDesc(); + bool isOptDef = (IIOpNum < TID.getNumOperands()) + ? (TID.OpInfo[IIOpNum].isOptionalDef()) : false; + MI->addOperand(MachineOperand::CreateReg(VReg, isOptDef)); // Verify that it is right. - assert(MRegisterInfo::isVirtualRegister(VReg) && "Not a vreg?"); + assert(TargetRegisterInfo::isVirtualRegister(VReg) && "Not a vreg?"); if (II) { const TargetRegisterClass *RC = - getInstrOperandRegClass(MRI, TII, II, IIOpNum); + getInstrOperandRegClass(TRI, TII, *II, IIOpNum); assert(RC && "Don't have operand info for this instruction!"); - const TargetRegisterClass *VRC = RegMap->getRegClass(VReg); + const TargetRegisterClass *VRC = RegInfo.getRegClass(VReg); if (VRC != RC) { cerr << "Register class of operand and regclass of use don't agree!\n"; #ifndef NDEBUG @@ -405,26 +553,23 @@ void ScheduleDAG::AddOperand(MachineInstr *MI, SDOperand Op, abort(); } } - } else if (ConstantSDNode *C = - dyn_cast(Op)) { - MI->addImmOperand(C->getValue()); - } else if (RegisterSDNode *R = - dyn_cast(Op)) { - MI->addRegOperand(R->getReg(), false); - } else if (GlobalAddressSDNode *TGA = - dyn_cast(Op)) { - MI->addGlobalAddressOperand(TGA->getGlobal(), TGA->getOffset()); - } else if (BasicBlockSDNode *BB = - dyn_cast(Op)) { - MI->addMachineBasicBlockOperand(BB->getBasicBlock()); - } else if (FrameIndexSDNode *FI = - dyn_cast(Op)) { - MI->addFrameIndexOperand(FI->getIndex()); - } else if (JumpTableSDNode *JT = - dyn_cast(Op)) { - MI->addJumpTableIndexOperand(JT->getIndex()); - } else if (ConstantPoolSDNode *CP = - dyn_cast(Op)) { + } else if (ConstantSDNode *C = dyn_cast(Op)) { + MI->addOperand(MachineOperand::CreateImm(C->getValue())); + } else if (ConstantFPSDNode *F = dyn_cast(Op)) { + const Type *FType = MVT::getTypeForValueType(Op.getValueType()); + ConstantFP *CFP = ConstantFP::get(FType, F->getValueAPF()); + MI->addOperand(MachineOperand::CreateFPImm(CFP)); + } else if (RegisterSDNode *R = dyn_cast(Op)) { + MI->addOperand(MachineOperand::CreateReg(R->getReg(), false)); + } else if (GlobalAddressSDNode *TGA = dyn_cast(Op)) { + MI->addOperand(MachineOperand::CreateGA(TGA->getGlobal(),TGA->getOffset())); + } else if (BasicBlockSDNode *BB = dyn_cast(Op)) { + MI->addOperand(MachineOperand::CreateMBB(BB->getBasicBlock())); + } else if (FrameIndexSDNode *FI = dyn_cast(Op)) { + MI->addOperand(MachineOperand::CreateFI(FI->getIndex())); + } else if (JumpTableSDNode *JT = dyn_cast(Op)) { + MI->addOperand(MachineOperand::CreateJTI(JT->getIndex())); + } else if (ConstantPoolSDNode *CP = dyn_cast(Op)) { int Offset = CP->getOffset(); unsigned Align = CP->getAlignment(); const Type *Type = CP->getType(); @@ -433,7 +578,7 @@ void ScheduleDAG::AddOperand(MachineInstr *MI, SDOperand Op, Align = TM.getTargetData()->getPreferredTypeAlignmentShift(Type); if (Align == 0) { // Alignment of vector types. FIXME! - Align = TM.getTargetData()->getTypeSize(Type); + Align = TM.getTargetData()->getABITypeSize(Type); Align = Log2_64(Align); } } @@ -443,36 +588,40 @@ void ScheduleDAG::AddOperand(MachineInstr *MI, SDOperand Op, Idx = ConstPool->getConstantPoolIndex(CP->getMachineCPVal(), Align); else Idx = ConstPool->getConstantPoolIndex(CP->getConstVal(), Align); - MI->addConstantPoolIndexOperand(Idx, Offset); - } else if (ExternalSymbolSDNode *ES = - dyn_cast(Op)) { - MI->addExternalSymbolOperand(ES->getSymbol()); + MI->addOperand(MachineOperand::CreateCPI(Idx, Offset)); + } else if (ExternalSymbolSDNode *ES = dyn_cast(Op)) { + MI->addOperand(MachineOperand::CreateES(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); + MI->addOperand(MachineOperand::CreateReg(VReg, false)); - // Verify that it is right. - assert(MRegisterInfo::isVirtualRegister(VReg) && "Not a vreg?"); + // Verify that it is right. Note that the reg class of the physreg and the + // vreg don't necessarily need to match, but the target copy insertion has + // to be able to handle it. This handles things like copies from ST(0) to + // an FP vreg on x86. + assert(TargetRegisterInfo::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!"); + assert(getInstrOperandRegClass(TRI, TII, *II, IIOpNum) && + "Don't have operand info for this instruction!"); } } } +void ScheduleDAG::AddMemOperand(MachineInstr *MI, const MemOperand &MO) { + MI->addMemOperand(MO); +} + // 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; + TargetRegisterInfo::regclass_iterator I = + TRC->subregclasses_begin() + SubIdx-1; assert(I < TRC->subregclasses_end() && "Invalid subregister index for register class"); return *I; @@ -483,7 +632,7 @@ static const TargetRegisterClass *getSuperregRegisterClass( unsigned SubIdx, MVT::ValueType VT) { // Pick the register class of the superegister for this type - for (MRegisterInfo::regclass_iterator I = TRC->superregclasses_begin(), + for (TargetRegisterInfo::regclass_iterator I = TRC->superregclasses_begin(), E = TRC->superregclasses_end(); I != E; ++I) if ((*I)->hasType(VT) && getSubRegisterRegClass(*I, SubIdx) == TRC) return *I; @@ -506,7 +655,7 @@ void ScheduleDAG::EmitSubregNode(SDNode *Node, if (Use->getOpcode() == ISD::CopyToReg && Use->getOperand(2).Val == Node) { unsigned DestReg = cast(Use->getOperand(1))->getReg(); - if (MRegisterInfo::isVirtualRegister(DestReg)) { + if (TargetRegisterInfo::isVirtualRegister(DestReg)) { VRBase = DestReg; break; } @@ -518,33 +667,30 @@ void ScheduleDAG::EmitSubregNode(SDNode *Node, // 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 *TRC = RegInfo.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 && + const TargetRegisterClass *DRC = RegInfo.getRegClass(VRBase); + assert(SRC && DRC && SRC == DRC && "Source subregister and destination must have the same class"); } else { // Create the reg - VRBase = RegMap->createVirtualRegister(SRC); + assert(SRC && "Couldn't find source register class"); + VRBase = RegInfo.createVirtualRegister(SRC); } // Add def, source, and subreg index - MI->addRegOperand(VRBase, true); + MI->addOperand(MachineOperand::CreateReg(VRBase, true)); AddOperand(MI, Node->getOperand(0), 0, 0, VRBaseMap); - MI->addImmOperand(SubIdx); + MI->addOperand(MachineOperand::CreateImm(SubIdx)); } else if (Opc == TargetInstrInfo::INSERT_SUBREG) { assert((Node->getNumOperands() == 2 || Node->getNumOperands() == 3) && @@ -561,7 +707,7 @@ void ScheduleDAG::EmitSubregNode(SDNode *Node, SubIdx = cast(Node->getOperand(2))->getValue(); } - // TODO: Add tracking info to SSARegMap of which vregs are subregs + // TODO: Add tracking info to MachineRegisterInfo 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 @@ -574,7 +720,7 @@ void ScheduleDAG::EmitSubregNode(SDNode *Node, if (Use->getOpcode() == ISD::CopyToReg && Use->getOperand(2).Val == Node) { unsigned DestReg = cast(Use->getOperand(1))->getReg(); - if (MRegisterInfo::isVirtualRegister(DestReg)) { + if (TargetRegisterInfo::isVirtualRegister(DestReg)) { VRBase = DestReg; break; } @@ -588,20 +734,19 @@ void ScheduleDAG::EmitSubregNode(SDNode *Node, // Figure out the register class to create for the destreg. const TargetRegisterClass *TRC = 0; if (VRBase) { - TRC = RegMap->getRegClass(VRBase); + TRC = RegInfo.getRegClass(VRBase); } else { - TRC = getSuperregRegisterClass(RegMap->getRegClass(SubReg), - SubIdx, + TRC = getSuperregRegisterClass(RegInfo.getRegClass(SubReg), SubIdx, Node->getValueType(0)); assert(TRC && "Couldn't determine register class for insert_subreg"); - VRBase = RegMap->createVirtualRegister(TRC); // Create the reg + VRBase = RegInfo.createVirtualRegister(TRC); // Create the reg } - MI->addRegOperand(VRBase, true); + MI->addOperand(MachineOperand::CreateReg(VRBase, true)); AddOperand(MI, Node->getOperand(0), 0, 0, VRBaseMap); if (!isUndefInput) AddOperand(MI, Node->getOperand(1), 0, 0, VRBaseMap); - MI->addImmOperand(SubIdx); + MI->addOperand(MachineOperand::CreateImm(SubIdx)); } else assert(0 && "Node is not a subreg insert or extract"); @@ -611,7 +756,7 @@ void ScheduleDAG::EmitSubregNode(SDNode *Node, /// EmitNode - Generate machine code for an node and needed dependencies. /// -void ScheduleDAG::EmitNode(SDNode *Node, +void ScheduleDAG::EmitNode(SDNode *Node, unsigned InstanceNo, DenseMap &VRBaseMap) { // If machine instruction if (Node->isTargetOpcode()) { @@ -624,15 +769,17 @@ void ScheduleDAG::EmitNode(SDNode *Node, return; } - const TargetInstrDescriptor &II = TII->get(Opc); + const TargetInstrDesc &II = TII->get(Opc); unsigned NumResults = CountResults(Node); unsigned NodeOperands = CountOperands(Node); + unsigned MemOperandsEnd = ComputeMemOperandsEnd(Node); unsigned NumMIOperands = NodeOperands + NumResults; - bool HasPhysRegOuts = (NumResults > II.numDefs) && II.ImplicitDefs; + bool HasPhysRegOuts = (NumResults > II.getNumDefs()) && + II.getImplicitDefs() != 0; #ifndef NDEBUG - assert((unsigned(II.numOperands) == NumMIOperands || - HasPhysRegOuts || (II.Flags & M_VARIABLE_OPS)) && + assert((II.getNumOperands() == NumMIOperands || + HasPhysRegOuts || II.isVariadic()) && "#operands for dag node doesn't match .td file!"); #endif @@ -647,7 +794,11 @@ void ScheduleDAG::EmitNode(SDNode *Node, // 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+II.numDefs, &II, VRBaseMap); + AddOperand(MI, Node->getOperand(i), i+II.getNumDefs(), &II, VRBaseMap); + + // Emit all of the memory operands of this instruction + for (unsigned i = NodeOperands; i != MemOperandsEnd; ++i) + AddMemOperand(MI, cast(Node->getOperand(i))->MO); // Commute node if it has been determined to be profitable. if (CommuteSet.count(Node)) { @@ -660,24 +811,23 @@ void ScheduleDAG::EmitNode(SDNode *Node, delete MI; MI = NewMI; } + ++NumCommutes; } } - // 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); - } + if (II.usesCustomDAGSchedInsertionHook()) + // Insert this instruction into the basic block using a target + // specific inserter which may returns a new basic block. + BB = DAG.getTargetLoweringInfo().EmitInstrWithCustomInserter(MI, BB); + else + BB->push_back(MI); // Additional results must be an physical register def. if (HasPhysRegOuts) { - for (unsigned i = II.numDefs; i < NumResults; ++i) { - unsigned Reg = II.ImplicitDefs[i - II.numDefs]; + for (unsigned i = II.getNumDefs(); i < NumResults; ++i) { + unsigned Reg = II.getImplicitDefs()[i - II.getNumDefs()]; if (Node->hasAnyUseOfValue(i)) - EmitCopyFromReg(Node, i, Reg, VRBaseMap); + EmitCopyFromReg(Node, i, InstanceNo, Reg, VRBaseMap); } } } else { @@ -690,6 +840,8 @@ void ScheduleDAG::EmitNode(SDNode *Node, case ISD::EntryToken: // fall thru case ISD::TokenFactor: case ISD::LABEL: + case ISD::DECLARE: + case ISD::SRCVALUE: break; case ISD::CopyToReg: { unsigned InReg; @@ -701,19 +853,19 @@ void ScheduleDAG::EmitNode(SDNode *Node, if (InReg != DestReg) {// Coalesced away the copy? const TargetRegisterClass *TRC = 0; // Get the target register class - if (MRegisterInfo::isVirtualRegister(InReg)) - TRC = RegMap->getRegClass(InReg); + if (TargetRegisterInfo::isVirtualRegister(InReg)) + TRC = RegInfo.getRegClass(InReg); else - TRC = getPhysicalRegisterRegClass(MRI, - Node->getOperand(2).getValueType(), + TRC = + TRI->getPhysicalRegisterRegClass(Node->getOperand(2).getValueType(), InReg); - MRI->copyRegToReg(*BB, BB->end(), DestReg, InReg, TRC); + TII->copyRegToReg(*BB, BB->end(), DestReg, InReg, TRC, TRC); } break; } case ISD::CopyFromReg: { unsigned SrcReg = cast(Node->getOperand(1))->getReg(); - EmitCopyFromReg(Node, 0, SrcReg, VRBaseMap); + EmitCopyFromReg(Node, 0, InstanceNo, SrcReg, VRBaseMap); break; } case ISD::INLINEASM: { @@ -728,14 +880,14 @@ void ScheduleDAG::EmitNode(SDNode *Node, // Add the asm string as an external symbol operand. const char *AsmStr = cast(Node->getOperand(1))->getSymbol(); - MI->addExternalSymbolOperand(AsmStr); + MI->addOperand(MachineOperand::CreateES(AsmStr)); // Add all of the operand registers to the instruction. for (unsigned i = 2; i != NumOps;) { unsigned Flags = cast(Node->getOperand(i))->getValue(); unsigned NumVals = Flags >> 3; - MI->addImmOperand(Flags); + MI->addOperand(MachineOperand::CreateImm(Flags)); ++i; // Skip the ID value. switch (Flags & 7) { @@ -743,24 +895,27 @@ void ScheduleDAG::EmitNode(SDNode *Node, case 1: // Use of register. for (; NumVals; --NumVals, ++i) { unsigned Reg = cast(Node->getOperand(i))->getReg(); - MI->addRegOperand(Reg, false); + MI->addOperand(MachineOperand::CreateReg(Reg, false)); } break; case 2: // Def of register. for (; NumVals; --NumVals, ++i) { unsigned Reg = cast(Node->getOperand(i))->getReg(); - MI->addRegOperand(Reg, true); + MI->addOperand(MachineOperand::CreateReg(Reg, true)); } break; case 3: { // Immediate. for (; NumVals; --NumVals, ++i) { if (ConstantSDNode *CS = dyn_cast(Node->getOperand(i))) { - MI->addImmOperand(CS->getValue()); + MI->addOperand(MachineOperand::CreateImm(CS->getValue())); + } else if (GlobalAddressSDNode *GA = + dyn_cast(Node->getOperand(i))) { + MI->addOperand(MachineOperand::CreateGA(GA->getGlobal(), + GA->getOffset())); } else { - GlobalAddressSDNode *GA = - cast(Node->getOperand(i)); - MI->addGlobalAddressOperand(GA->getGlobal(), GA->getOffset()); + BasicBlockSDNode *BB =cast(Node->getOperand(i)); + MI->addOperand(MachineOperand::CreateMBB(BB->getBasicBlock())); } } break; @@ -783,28 +938,67 @@ void ScheduleDAG::EmitNoop() { TII->insertNoop(*BB, BB->end()); } +void ScheduleDAG::EmitCrossRCCopy(SUnit *SU, + DenseMap &VRBaseMap) { + for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); + I != E; ++I) { + if (I->isCtrl) continue; // ignore chain preds + if (!I->Dep->Node) { + // Copy to physical register. + DenseMap::iterator VRI = VRBaseMap.find(I->Dep); + assert(VRI != VRBaseMap.end() && "Node emitted out of order - late"); + // Find the destination physical register. + unsigned Reg = 0; + for (SUnit::const_succ_iterator II = SU->Succs.begin(), + EE = SU->Succs.end(); II != EE; ++II) { + if (I->Reg) { + Reg = I->Reg; + break; + } + } + assert(I->Reg && "Unknown physical register!"); + TII->copyRegToReg(*BB, BB->end(), Reg, VRI->second, + SU->CopyDstRC, SU->CopySrcRC); + } else { + // Copy from physical register. + assert(I->Reg && "Unknown physical register!"); + unsigned VRBase = RegInfo.createVirtualRegister(SU->CopyDstRC); + bool isNew = VRBaseMap.insert(std::make_pair(SU, VRBase)); + assert(isNew && "Node emitted out of order - early"); + TII->copyRegToReg(*BB, BB->end(), VRBase, I->Reg, + SU->CopyDstRC, SU->CopySrcRC); + } + break; + } +} + /// 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)); + if (&MF->front() == BB) { + for (MachineRegisterInfo::livein_iterator LI = RegInfo.livein_begin(), + E = RegInfo.livein_end(); LI != E; ++LI) + if (LI->second) { + const TargetRegisterClass *RC = RegInfo.getRegClass(LI->second); + TII->copyRegToReg(*MF->begin(), MF->begin()->end(), LI->second, + LI->first, RC, RC); + } } // Finally, emit the code for all of the scheduled instructions. DenseMap VRBaseMap; + DenseMap CopyVRBaseMap; 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); + for (unsigned j = 0, ee = SU->FlaggedNodes.size(); j != ee; ++j) + EmitNode(SU->FlaggedNodes[j], SU->InstanceNo, VRBaseMap); + if (SU->Node) + EmitNode(SU->Node, SU->InstanceNo, VRBaseMap); + else + EmitCrossRCCopy(SU, CopyVRBaseMap); } else { // Null SUnit* is a noop. EmitNoop(); @@ -826,11 +1020,6 @@ void ScheduleDAG::dumpSchedule() const { /// Run - perform scheduling. /// MachineBasicBlock *ScheduleDAG::Run() { - TII = TM.getInstrInfo(); - MRI = TM.getRegisterInfo(); - RegMap = BB->getParent()->getSSARegMap(); - ConstPool = BB->getParent()->getConstantPool(); - Schedule(); return BB; } @@ -839,7 +1028,10 @@ MachineBasicBlock *ScheduleDAG::Run() { /// 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++) { @@ -855,8 +1047,6 @@ void SUnit::dumpAll(const SelectionDAG *G) const { 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"; @@ -869,7 +1059,10 @@ void SUnit::dumpAll(const SelectionDAG *G) const { cerr << " ch #"; else cerr << " val #"; - cerr << I->Dep << " - SU(" << I->Dep->NodeNum << ")\n"; + cerr << I->Dep << " - SU(" << I->Dep->NodeNum << ")"; + if (I->isSpecial) + cerr << " *"; + cerr << "\n"; } } if (Succs.size() != 0) { @@ -880,7 +1073,10 @@ void SUnit::dumpAll(const SelectionDAG *G) const { cerr << " ch #"; else cerr << " val #"; - cerr << I->Dep << " - SU(" << I->Dep->NodeNum << ")\n"; + cerr << I->Dep << " - SU(" << I->Dep->NodeNum << ")"; + if (I->isSpecial) + cerr << " *"; + cerr << "\n"; } } cerr << "\n";