lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp

   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/Target/TargetSubtarget.h"
  22 #include "llvm/Support/Debug.h"
  23 #include "llvm/Support/raw_ostream.h"
  24 using namespace llvm;
  25
  26 ScheduleDAGSDNodes::ScheduleDAGSDNodes(MachineFunction &mf)
  27   : ScheduleDAG(mf) {
  28 }
  29
  30 /// Run - perform scheduling.
  31 ///
  32 void ScheduleDAGSDNodes::Run(SelectionDAG *dag, MachineBasicBlock *bb,
  33                              MachineBasicBlock::iterator insertPos) {
  34   DAG = dag;
  35   ScheduleDAG::Run(bb, insertPos);
  36 }
  37
  38 SUnit *ScheduleDAGSDNodes::Clone(SUnit *Old) {
  39   SUnit *SU = NewSUnit(Old->getNode());
  40   SU->OrigNode = Old->OrigNode;
  41   SU->Latency = Old->Latency;
  42   SU->isTwoAddress = Old->isTwoAddress;
  43   SU->isCommutable = Old->isCommutable;
  44   SU->hasPhysRegDefs = Old->hasPhysRegDefs;
  45   SU->hasPhysRegClobbers = Old->hasPhysRegClobbers;
  46   Old->isCloned = true;
  47   return SU;
  48 }
  49
  50 /// CheckForPhysRegDependency - Check if the dependency between def and use of
  51 /// a specified operand is a physical register dependency. If so, returns the
  52 /// register and the cost of copying the register.
  53 static void CheckForPhysRegDependency(SDNode *Def, SDNode *User, unsigned Op,
  54                                       const TargetRegisterInfo *TRI,
  55                                       const TargetInstrInfo *TII,
  56                                       unsigned &PhysReg, int &Cost) {
  57   if (Op != 2 || User->getOpcode() != ISD::CopyToReg)
  58     return;
  59
  60   unsigned Reg = cast<RegisterSDNode>(User->getOperand(1))->getReg();
  61   if (TargetRegisterInfo::isVirtualRegister(Reg))
  62     return;
  63
  64   unsigned ResNo = User->getOperand(2).getResNo();
  65   if (Def->isMachineOpcode()) {
  66     const TargetInstrDesc &II = TII->get(Def->getMachineOpcode());
  67     if (ResNo >= II.getNumDefs() &&
  68         II.ImplicitDefs[ResNo - II.getNumDefs()] == Reg) {
  69       PhysReg = Reg;
  70       const TargetRegisterClass *RC =
  71         TRI->getPhysicalRegisterRegClass(Reg, Def->getValueType(ResNo));
  72       Cost = RC->getCopyCost();
  73     }
  74   }
  75 }
  76
  77 void ScheduleDAGSDNodes::BuildSchedUnits() {
  78   // During scheduling, the NodeId field of SDNode is used to map SDNodes
  79   // to their associated SUnits by holding SUnits table indices. A value
  80   // of -1 means the SDNode does not yet have an associated SUnit.
  81   unsigned NumNodes = 0;
  82   for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(),
  83        E = DAG->allnodes_end(); NI != E; ++NI) {
  84     NI->setNodeId(-1);
  85     ++NumNodes;
  86   }
  87
  88   // Reserve entries in the vector for each of the SUnits we are creating.  This
  89   // ensure that reallocation of the vector won't happen, so SUnit*'s won't get
  90   // invalidated.
  91   // FIXME: Multiply by 2 because we may clone nodes during scheduling.
  92   // This is a temporary workaround.
  93   SUnits.reserve(NumNodes * 2);
  94
  95   // Check to see if the scheduler cares about latencies.
  96   bool UnitLatencies = ForceUnitLatencies();
  97
  98   for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(),
  99        E = DAG->allnodes_end(); NI != E; ++NI) {
 100     if (isPassiveNode(NI))  // Leaf node, e.g. a TargetImmediate.
 101       continue;
 102
 103     // If this node has already been processed, stop now.
 104     if (NI->getNodeId() != -1) continue;
 105
 106     SUnit *NodeSUnit = NewSUnit(NI);
 107
 108     // See if anything is flagged to this node, if so, add them to flagged
 109     // nodes.  Nodes can have at most one flag input and one flag output.  Flags
 110     // are required to be the last operand and result of a node.
 111
 112     // Scan up to find flagged preds.
 113     SDNode *N = NI;
 114     while (N->getNumOperands() &&
 115            N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag) {
 116       N = N->getOperand(N->getNumOperands()-1).getNode();
 117       assert(N->getNodeId() == -1 && "Node already inserted!");
 118       N->setNodeId(NodeSUnit->NodeNum);
 119     }
 120
 121     // Scan down to find any flagged succs.
 122     N = NI;
 123     while (N->getValueType(N->getNumValues()-1) == MVT::Flag) {
 124       SDValue FlagVal(N, N->getNumValues()-1);
 125
 126       // There are either zero or one users of the Flag result.
 127       bool HasFlagUse = false;
 128       for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end();
 129            UI != E; ++UI)
 130         if (FlagVal.isOperandOf(*UI)) {
 131           HasFlagUse = true;
 132           assert(N->getNodeId() == -1 && "Node already inserted!");
 133           N->setNodeId(NodeSUnit->NodeNum);
 134           N = *UI;
 135           break;
 136         }
 137       if (!HasFlagUse) break;
 138     }
 139
 140     // If there are flag operands involved, N is now the bottom-most node
 141     // of the sequence of nodes that are flagged together.
 142     // Update the SUnit.
 143     NodeSUnit->setNode(N);
 144     assert(N->getNodeId() == -1 && "Node already inserted!");
 145     N->setNodeId(NodeSUnit->NodeNum);
 146
 147     // Assign the Latency field of NodeSUnit using target-provided information.
 148     if (UnitLatencies)
 149       NodeSUnit->Latency = 1;
 150     else
 151       ComputeLatency(NodeSUnit);
 152   }
 153 }
 154
 155 void ScheduleDAGSDNodes::AddSchedEdges() {
 156   const TargetSubtarget &ST = TM.getSubtarget<TargetSubtarget>();
 157
 158   // Check to see if the scheduler cares about latencies.
 159   bool UnitLatencies = ForceUnitLatencies();
 160
 161   // Pass 2: add the preds, succs, etc.
 162   for (unsigned su = 0, e = SUnits.size(); su != e; ++su) {
 163     SUnit *SU = &SUnits[su];
 164     SDNode *MainNode = SU->getNode();
 165
 166     if (MainNode->isMachineOpcode()) {
 167       unsigned Opc = MainNode->getMachineOpcode();
 168       const TargetInstrDesc &TID = TII->get(Opc);
 169       for (unsigned i = 0; i != TID.getNumOperands(); ++i) {
 170         if (TID.getOperandConstraint(i, TOI::TIED_TO) != -1) {
 171           SU->isTwoAddress = true;
 172           break;
 173         }
 174       }
 175       if (TID.isCommutable())
 176         SU->isCommutable = true;
 177     }
 178
 179     // Find all predecessors and successors of the group.
 180     for (SDNode *N = SU->getNode(); N; N = N->getFlaggedNode()) {
 181       if (N->isMachineOpcode() &&
 182           TII->get(N->getMachineOpcode()).getImplicitDefs()) {
 183         SU->hasPhysRegClobbers = true;
 184         unsigned NumUsed = CountResults(N);
 185         while (NumUsed != 0 && !N->hasAnyUseOfValue(NumUsed - 1))
 186           --NumUsed;    // Skip over unused values at the end.
 187         if (NumUsed > TII->get(N->getMachineOpcode()).getNumDefs())
 188           SU->hasPhysRegDefs = true;
 189       }
 190
 191       for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
 192         SDNode *OpN = N->getOperand(i).getNode();
 193         if (isPassiveNode(OpN)) continue;   // Not scheduled.
 194         SUnit *OpSU = &SUnits[OpN->getNodeId()];
 195         assert(OpSU && "Node has no SUnit!");
 196         if (OpSU == SU) continue;           // In the same group.
 197
 198         EVT OpVT = N->getOperand(i).getValueType();
 199         assert(OpVT != MVT::Flag && "Flagged nodes should be in same sunit!");
 200         bool isChain = OpVT == MVT::Other;
 201
 202         unsigned PhysReg = 0;
 203         int Cost = 1;
 204         // Determine if this is a physical register dependency.
 205         CheckForPhysRegDependency(OpN, N, i, TRI, TII, PhysReg, Cost);
 206         assert((PhysReg == 0 || !isChain) &&
 207                "Chain dependence via physreg data?");
 208         // FIXME: See ScheduleDAGSDNodes::EmitCopyFromReg. For now, scheduler
 209         // emits a copy from the physical register to a virtual register unless
 210         // it requires a cross class copy (cost < 0). That means we are only
 211         // treating "expensive to copy" register dependency as physical register
 212         // dependency. This may change in the future though.
 213         if (Cost >= 0)
 214           PhysReg = 0;
 215
 216         const SDep& dep = SDep(OpSU, isChain ? SDep::Order : SDep::Data,
 217                                OpSU->Latency, PhysReg);
 218         if (!isChain && !UnitLatencies) {
 219           ComputeOperandLatency(OpSU, SU, (SDep &)dep);
 220           ST.adjustSchedDependency(OpSU, SU, (SDep &)dep);
 221         }
 222
 223         SU->addPred(dep);
 224       }
 225     }
 226   }
 227 }
 228
 229 /// BuildSchedGraph - Build the SUnit graph from the selection dag that we
 230 /// are input.  This SUnit graph is similar to the SelectionDAG, but
 231 /// excludes nodes that aren't interesting to scheduling, and represents
 232 /// flagged together nodes with a single SUnit.
 233 void ScheduleDAGSDNodes::BuildSchedGraph() {
 234   // Populate the SUnits array.
 235   BuildSchedUnits();
 236   // Compute all the scheduling dependencies between nodes.
 237   AddSchedEdges();
 238 }
 239
 240 void ScheduleDAGSDNodes::ComputeLatency(SUnit *SU) {
 241   const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
 242
 243   // Compute the latency for the node.  We use the sum of the latencies for
 244   // all nodes flagged together into this SUnit.
 245   SU->Latency = 0;
 246   for (SDNode *N = SU->getNode(); N; N = N->getFlaggedNode())
 247     if (N->isMachineOpcode()) {
 248       SU->Latency += InstrItins.
 249         getStageLatency(TII->get(N->getMachineOpcode()).getSchedClass());
 250     }
 251 }
 252
 253 /// CountResults - The results of target nodes have register or immediate
 254 /// operands first, then an optional chain, and optional flag operands (which do
 255 /// not go into the resulting MachineInstr).
 256 unsigned ScheduleDAGSDNodes::CountResults(SDNode *Node) {
 257   unsigned N = Node->getNumValues();
 258   while (N && Node->getValueType(N - 1) == MVT::Flag)
 259     --N;
 260   if (N && Node->getValueType(N - 1) == MVT::Other)
 261     --N;    // Skip over chain result.
 262   return N;
 263 }
 264
 265 /// CountOperands - The inputs to target nodes have any actual inputs first,
 266 /// followed by special operands that describe memory references, then an
 267 /// optional chain operand, then an optional flag operand.  Compute the number
 268 /// of actual operands that will go into the resulting MachineInstr.
 269 unsigned ScheduleDAGSDNodes::CountOperands(SDNode *Node) {
 270   unsigned N = ComputeMemOperandsEnd(Node);
 271   while (N && isa<MemOperandSDNode>(Node->getOperand(N - 1).getNode()))
 272     --N; // Ignore MEMOPERAND nodes
 273   return N;
 274 }
 275
 276 /// ComputeMemOperandsEnd - Find the index one past the last MemOperandSDNode
 277 /// operand
 278 unsigned ScheduleDAGSDNodes::ComputeMemOperandsEnd(SDNode *Node) {
 279   unsigned N = Node->getNumOperands();
 280   while (N && Node->getOperand(N - 1).getValueType() == MVT::Flag)
 281     --N;
 282   if (N && Node->getOperand(N - 1).getValueType() == MVT::Other)
 283     --N; // Ignore chain if it exists.
 284   return N;
 285 }
 286
 287
 288 void ScheduleDAGSDNodes::dumpNode(const SUnit *SU) const {
 289   if (!SU->getNode()) {
 290     errs() << "PHYS REG COPY\n";
 291     return;
 292   }
 293
 294   SU->getNode()->dump(DAG);
 295   errs() << "\n";
 296   SmallVector<SDNode *, 4> FlaggedNodes;
 297   for (SDNode *N = SU->getNode()->getFlaggedNode(); N; N = N->getFlaggedNode())
 298     FlaggedNodes.push_back(N);
 299   while (!FlaggedNodes.empty()) {
 300     errs() << "    ";
 301     FlaggedNodes.back()->dump(DAG);
 302     errs() << "\n";
 303     FlaggedNodes.pop_back();
 304   }
 305 }