ScheduleDAGSDNodes.cpp revision b872c7c9784c3b9a6c965379ed3ed2115c55f5b3
1//===--- ScheduleDAGSDNodes.cpp - Implement the ScheduleDAGSDNodes class --===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This implements the ScheduleDAG class, which is a base class used by 11// scheduling implementation classes. 12// 13//===----------------------------------------------------------------------===// 14 15#define DEBUG_TYPE "pre-RA-sched" 16#include "ScheduleDAGSDNodes.h" 17#include "llvm/CodeGen/SelectionDAG.h" 18#include "llvm/Target/TargetMachine.h" 19#include "llvm/Target/TargetInstrInfo.h" 20#include "llvm/Target/TargetRegisterInfo.h" 21#include "llvm/Support/Debug.h" 22#include "llvm/Support/raw_ostream.h" 23using namespace llvm; 24 25ScheduleDAGSDNodes::ScheduleDAGSDNodes(MachineFunction &mf) 26 : ScheduleDAG(mf) { 27} 28 29/// Run - perform scheduling. 30/// 31void ScheduleDAGSDNodes::Run(SelectionDAG *dag, MachineBasicBlock *bb, 32 MachineBasicBlock::iterator insertPos) { 33 DAG = dag; 34 ScheduleDAG::Run(bb, insertPos); 35} 36 37SUnit *ScheduleDAGSDNodes::Clone(SUnit *Old) { 38 SUnit *SU = NewSUnit(Old->getNode()); 39 SU->OrigNode = Old->OrigNode; 40 SU->Latency = Old->Latency; 41 SU->isTwoAddress = Old->isTwoAddress; 42 SU->isCommutable = Old->isCommutable; 43 SU->hasPhysRegDefs = Old->hasPhysRegDefs; 44 SU->hasPhysRegClobbers = Old->hasPhysRegClobbers; 45 Old->isCloned = true; 46 return SU; 47} 48 49/// CheckForPhysRegDependency - Check if the dependency between def and use of 50/// a specified operand is a physical register dependency. If so, returns the 51/// register and the cost of copying the register. 52static void CheckForPhysRegDependency(SDNode *Def, SDNode *User, unsigned Op, 53 const TargetRegisterInfo *TRI, 54 const TargetInstrInfo *TII, 55 unsigned &PhysReg, int &Cost) { 56 if (Op != 2 || User->getOpcode() != ISD::CopyToReg) 57 return; 58 59 unsigned Reg = cast<RegisterSDNode>(User->getOperand(1))->getReg(); 60 if (TargetRegisterInfo::isVirtualRegister(Reg)) 61 return; 62 63 unsigned ResNo = User->getOperand(2).getResNo(); 64 if (Def->isMachineOpcode()) { 65 const TargetInstrDesc &II = TII->get(Def->getMachineOpcode()); 66 if (ResNo >= II.getNumDefs() && 67 II.ImplicitDefs[ResNo - II.getNumDefs()] == Reg) { 68 PhysReg = Reg; 69 const TargetRegisterClass *RC = 70 TRI->getPhysicalRegisterRegClass(Reg, Def->getValueType(ResNo)); 71 Cost = RC->getCopyCost(); 72 } 73 } 74} 75 76void ScheduleDAGSDNodes::BuildSchedUnits() { 77 // During scheduling, the NodeId field of SDNode is used to map SDNodes 78 // to their associated SUnits by holding SUnits table indices. A value 79 // of -1 means the SDNode does not yet have an associated SUnit. 80 unsigned NumNodes = 0; 81 for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(), 82 E = DAG->allnodes_end(); NI != E; ++NI) { 83 NI->setNodeId(-1); 84 ++NumNodes; 85 } 86 87 // Reserve entries in the vector for each of the SUnits we are creating. This 88 // ensure that reallocation of the vector won't happen, so SUnit*'s won't get 89 // invalidated. 90 // FIXME: Multiply by 2 because we may clone nodes during scheduling. 91 // This is a temporary workaround. 92 SUnits.reserve(NumNodes * 2); 93 94 // Check to see if the scheduler cares about latencies. 95 bool UnitLatencies = ForceUnitLatencies(); 96 97 for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(), 98 E = DAG->allnodes_end(); NI != E; ++NI) { 99 if (isPassiveNode(NI)) // Leaf node, e.g. a TargetImmediate. 100 continue; 101 102 // If this node has already been processed, stop now. 103 if (NI->getNodeId() != -1) continue; 104 105 SUnit *NodeSUnit = NewSUnit(NI); 106 107 // See if anything is flagged to this node, if so, add them to flagged 108 // nodes. Nodes can have at most one flag input and one flag output. Flags 109 // are required to be the last operand and result of a node. 110 111 // Scan up to find flagged preds. 112 SDNode *N = NI; 113 while (N->getNumOperands() && 114 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag) { 115 N = N->getOperand(N->getNumOperands()-1).getNode(); 116 assert(N->getNodeId() == -1 && "Node already inserted!"); 117 N->setNodeId(NodeSUnit->NodeNum); 118 } 119 120 // Scan down to find any flagged succs. 121 N = NI; 122 while (N->getValueType(N->getNumValues()-1) == MVT::Flag) { 123 SDValue FlagVal(N, N->getNumValues()-1); 124 125 // There are either zero or one users of the Flag result. 126 bool HasFlagUse = false; 127 for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end(); 128 UI != E; ++UI) 129 if (FlagVal.isOperandOf(*UI)) { 130 HasFlagUse = true; 131 assert(N->getNodeId() == -1 && "Node already inserted!"); 132 N->setNodeId(NodeSUnit->NodeNum); 133 N = *UI; 134 break; 135 } 136 if (!HasFlagUse) break; 137 } 138 139 // If there are flag operands involved, N is now the bottom-most node 140 // of the sequence of nodes that are flagged together. 141 // Update the SUnit. 142 NodeSUnit->setNode(N); 143 assert(N->getNodeId() == -1 && "Node already inserted!"); 144 N->setNodeId(NodeSUnit->NodeNum); 145 146 // Assign the Latency field of NodeSUnit using target-provided information. 147 if (UnitLatencies) 148 NodeSUnit->Latency = 1; 149 else 150 ComputeLatency(NodeSUnit); 151 } 152} 153 154void ScheduleDAGSDNodes::AddSchedEdges() { 155 // Pass 2: add the preds, succs, etc. 156 for (unsigned su = 0, e = SUnits.size(); su != e; ++su) { 157 SUnit *SU = &SUnits[su]; 158 SDNode *MainNode = SU->getNode(); 159 160 if (MainNode->isMachineOpcode()) { 161 unsigned Opc = MainNode->getMachineOpcode(); 162 const TargetInstrDesc &TID = TII->get(Opc); 163 for (unsigned i = 0; i != TID.getNumOperands(); ++i) { 164 if (TID.getOperandConstraint(i, TOI::TIED_TO) != -1) { 165 SU->isTwoAddress = true; 166 break; 167 } 168 } 169 if (TID.isCommutable()) 170 SU->isCommutable = true; 171 } 172 173 // Find all predecessors and successors of the group. 174 for (SDNode *N = SU->getNode(); N; N = N->getFlaggedNode()) { 175 if (N->isMachineOpcode() && 176 TII->get(N->getMachineOpcode()).getImplicitDefs()) { 177 SU->hasPhysRegClobbers = true; 178 if (CountResults(N) > TII->get(N->getMachineOpcode()).getNumDefs()) 179 SU->hasPhysRegDefs = true; 180 } 181 182 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 183 SDNode *OpN = N->getOperand(i).getNode(); 184 if (isPassiveNode(OpN)) continue; // Not scheduled. 185 SUnit *OpSU = &SUnits[OpN->getNodeId()]; 186 assert(OpSU && "Node has no SUnit!"); 187 if (OpSU == SU) continue; // In the same group. 188 189 MVT OpVT = N->getOperand(i).getValueType(); 190 assert(OpVT != MVT::Flag && "Flagged nodes should be in same sunit!"); 191 bool isChain = OpVT == MVT::Other; 192 193 unsigned PhysReg = 0; 194 int Cost = 1; 195 // Determine if this is a physical register dependency. 196 CheckForPhysRegDependency(OpN, N, i, TRI, TII, PhysReg, Cost); 197 assert((PhysReg == 0 || !isChain) && 198 "Chain dependence via physreg data?"); 199 // FIXME: See ScheduleDAGSDNodes::EmitCopyFromReg. For now, scheduler 200 // emits a copy from the physical register to a virtual register unless 201 // it requires a cross class copy (cost < 0). That means we are only 202 // treating "expensive to copy" register dependency as physical register 203 // dependency. This may change in the future though. 204 if (Cost >= 0) 205 PhysReg = 0; 206 SU->addPred(SDep(OpSU, isChain ? SDep::Order : SDep::Data, 207 OpSU->Latency, PhysReg)); 208 } 209 } 210 } 211} 212 213/// BuildSchedGraph - Build the SUnit graph from the selection dag that we 214/// are input. This SUnit graph is similar to the SelectionDAG, but 215/// excludes nodes that aren't interesting to scheduling, and represents 216/// flagged together nodes with a single SUnit. 217void ScheduleDAGSDNodes::BuildSchedGraph() { 218 // Populate the SUnits array. 219 BuildSchedUnits(); 220 // Compute all the scheduling dependencies between nodes. 221 AddSchedEdges(); 222} 223 224void ScheduleDAGSDNodes::ComputeLatency(SUnit *SU) { 225 const InstrItineraryData &InstrItins = TM.getInstrItineraryData(); 226 227 // Compute the latency for the node. We use the sum of the latencies for 228 // all nodes flagged together into this SUnit. 229 SU->Latency = 0; 230 bool SawMachineOpcode = false; 231 for (SDNode *N = SU->getNode(); N; N = N->getFlaggedNode()) 232 if (N->isMachineOpcode()) { 233 SawMachineOpcode = true; 234 SU->Latency += 235 InstrItins.getLatency(TII->get(N->getMachineOpcode()).getSchedClass()); 236 } 237} 238 239/// CountResults - The results of target nodes have register or immediate 240/// operands first, then an optional chain, and optional flag operands (which do 241/// not go into the resulting MachineInstr). 242unsigned ScheduleDAGSDNodes::CountResults(SDNode *Node) { 243 unsigned N = Node->getNumValues(); 244 while (N && Node->getValueType(N - 1) == MVT::Flag) 245 --N; 246 if (N && Node->getValueType(N - 1) == MVT::Other) 247 --N; // Skip over chain result. 248 return N; 249} 250 251/// CountOperands - The inputs to target nodes have any actual inputs first, 252/// followed by special operands that describe memory references, then an 253/// optional chain operand, then an optional flag operand. Compute the number 254/// of actual operands that will go into the resulting MachineInstr. 255unsigned ScheduleDAGSDNodes::CountOperands(SDNode *Node) { 256 unsigned N = ComputeMemOperandsEnd(Node); 257 while (N && isa<MemOperandSDNode>(Node->getOperand(N - 1).getNode())) 258 --N; // Ignore MEMOPERAND nodes 259 return N; 260} 261 262/// ComputeMemOperandsEnd - Find the index one past the last MemOperandSDNode 263/// operand 264unsigned ScheduleDAGSDNodes::ComputeMemOperandsEnd(SDNode *Node) { 265 unsigned N = Node->getNumOperands(); 266 while (N && Node->getOperand(N - 1).getValueType() == MVT::Flag) 267 --N; 268 if (N && Node->getOperand(N - 1).getValueType() == MVT::Other) 269 --N; // Ignore chain if it exists. 270 return N; 271} 272 273 274void ScheduleDAGSDNodes::dumpNode(const SUnit *SU) const { 275 if (!SU->getNode()) { 276 cerr << "PHYS REG COPY\n"; 277 return; 278 } 279 280 SU->getNode()->dump(DAG); 281 cerr << "\n"; 282 SmallVector<SDNode *, 4> FlaggedNodes; 283 for (SDNode *N = SU->getNode()->getFlaggedNode(); N; N = N->getFlaggedNode()) 284 FlaggedNodes.push_back(N); 285 while (!FlaggedNodes.empty()) { 286 cerr << " "; 287 FlaggedNodes.back()->dump(DAG); 288 cerr << "\n"; 289 FlaggedNodes.pop_back(); 290 } 291} 292