llvm.org GIT mirror llvm / release_24 lib / CodeGen / SelectionDAG / ScheduleDAG.cpp
release_24

Tree @release_24 (Download .tar.gz)

ScheduleDAG.cpp @release_24raw · history · blame

  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
//===---- ScheduleDAG.cpp - Implement the ScheduleDAG class ---------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This implements the ScheduleDAG class, which is a base class used by
// scheduling implementation classes.
//
//===----------------------------------------------------------------------===//

#define DEBUG_TYPE "pre-RA-sched"
#include "llvm/CodeGen/ScheduleDAG.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Support/Debug.h"
using namespace llvm;

ScheduleDAG::ScheduleDAG(SelectionDAG &dag, MachineBasicBlock *bb,
                         const TargetMachine &tm)
  : DAG(dag), BB(bb), TM(tm), MRI(BB->getParent()->getRegInfo()) {
  TII = TM.getInstrInfo();
  MF  = &DAG.getMachineFunction();
  TRI = TM.getRegisterInfo();
  TLI = &DAG.getTargetLoweringInfo();
  ConstPool = BB->getParent()->getConstantPool();
}

/// CheckForPhysRegDependency - Check if the dependency between def and use of
/// a specified operand is a physical register dependency. If so, returns the
/// register and the cost of copying the register.
static void CheckForPhysRegDependency(SDNode *Def, SDNode *User, unsigned Op,
                                      const TargetRegisterInfo *TRI, 
                                      const TargetInstrInfo *TII,
                                      unsigned &PhysReg, int &Cost) {
  if (Op != 2 || User->getOpcode() != ISD::CopyToReg)
    return;

  unsigned Reg = cast<RegisterSDNode>(User->getOperand(1))->getReg();
  if (TargetRegisterInfo::isVirtualRegister(Reg))
    return;

  unsigned ResNo = User->getOperand(2).getResNo();
  if (Def->isMachineOpcode()) {
    const TargetInstrDesc &II = TII->get(Def->getMachineOpcode());
    if (ResNo >= II.getNumDefs() &&
        II.ImplicitDefs[ResNo - II.getNumDefs()] == Reg) {
      PhysReg = Reg;
      const TargetRegisterClass *RC =
        TRI->getPhysicalRegisterRegClass(Reg, Def->getValueType(ResNo));
      Cost = RC->getCopyCost();
    }
  }
}

SUnit *ScheduleDAG::Clone(SUnit *Old) {
  SUnit *SU = NewSUnit(Old->Node);
  SU->OrigNode = Old->OrigNode;
  SU->FlaggedNodes = Old->FlaggedNodes;
  SU->Latency = Old->Latency;
  SU->isTwoAddress = Old->isTwoAddress;
  SU->isCommutable = Old->isCommutable;
  SU->hasPhysRegDefs = Old->hasPhysRegDefs;
  return SU;
}


/// BuildSchedUnits - Build SUnits from the selection dag that we are input.
/// This SUnit graph is similar to the SelectionDAG, but represents flagged
/// together nodes with a single SUnit.
void ScheduleDAG::BuildSchedUnits() {
  // 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());
  
  // During scheduling, the NodeId field of SDNode is used to map SDNodes
  // to their associated SUnits by holding SUnits table indices. A value
  // of -1 means the SDNode does not yet have an associated SUnit.
  for (SelectionDAG::allnodes_iterator NI = DAG.allnodes_begin(),
       E = DAG.allnodes_end(); NI != E; ++NI)
    NI->setNodeId(-1);

  for (SelectionDAG::allnodes_iterator NI = DAG.allnodes_begin(),
       E = DAG.allnodes_end(); NI != E; ++NI) {
    if (isPassiveNode(NI))  // Leaf node, e.g. a TargetImmediate.
      continue;
    
    // If this node has already been processed, stop now.
    if (NI->getNodeId() != -1) continue;
    
    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, adding flagged preds to FlaggedNodes.
    SDNode *N = NI;
    if (N->getNumOperands() &&
        N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag) {
      do {
        N = N->getOperand(N->getNumOperands()-1).getNode();
        NodeSUnit->FlaggedNodes.push_back(N);
        assert(N->getNodeId() == -1 && "Node already inserted!");
        N->setNodeId(NodeSUnit->NodeNum);
      } while (N->getNumOperands() &&
               N->getOperand(N->getNumOperands()-1).getValueType()== MVT::Flag);
      std::reverse(NodeSUnit->FlaggedNodes.begin(),
                   NodeSUnit->FlaggedNodes.end());
    }
    
    // Scan down, adding this node and any flagged succs to FlaggedNodes if they
    // have a user of the flag operand.
    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;
          NodeSUnit->FlaggedNodes.push_back(N);
          assert(N->getNodeId() == -1 && "Node already inserted!");
          N->setNodeId(NodeSUnit->NodeNum);
          N = *UI;
          break;
        }
      if (!HasFlagUse) break;
    }
    
    // Now all flagged nodes are in FlaggedNodes and N is the bottom-most node.
    // Update the SUnit
    NodeSUnit->Node = 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->Node;
    
    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;
    }
    
    // Find all predecessors and successors of the group.
    // Temporarily add N to make code simpler.
    SU->FlaggedNodes.push_back(MainNode);
    
    for (unsigned n = 0, e = SU->FlaggedNodes.size(); n != e; ++n) {
      SDNode *N = SU->FlaggedNodes[n];
      if (N->isMachineOpcode() &&
          TII->get(N->getMachineOpcode()).getImplicitDefs() &&
          CountResults(N) > TII->get(N->getMachineOpcode()).getNumDefs())
        SU->hasPhysRegDefs = true;
      
      for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
        SDNode *OpN = N->getOperand(i).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);
      }
    }
    
    // Remove MainNode from FlaggedNodes again.
    SU->FlaggedNodes.pop_back();
  }
}

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;
  if (SU->Node->isMachineOpcode()) {
    unsigned SchedClass = TII->get(SU->Node->getMachineOpcode()).getSchedClass();
    const InstrStage *S = InstrItins.begin(SchedClass);
    const InstrStage *E = InstrItins.end(SchedClass);
    for (; S != E; ++S)
      SU->Latency += S->Cycles;
  }
  for (unsigned i = 0, e = SU->FlaggedNodes.size(); i != e; ++i) {
    SDNode *FNode = SU->FlaggedNodes[i];
    if (FNode->isMachineOpcode()) {
      unsigned SchedClass = TII->get(FNode->getMachineOpcode()).getSchedClass();
      const InstrStage *S = InstrItins.begin(SchedClass);
      const InstrStage *E = InstrItins.end(SchedClass);
      for (; S != E; ++S)
        SU->Latency += S->Cycles;
    }
  }
}

/// CalculateDepths - compute depths using algorithms for the longest
/// paths in the DAG
void ScheduleDAG::CalculateDepths() {
  unsigned DAGSize = SUnits.size();
  std::vector<SUnit*> WorkList;
  WorkList.reserve(DAGSize);

  // Initialize the data structures
  for (unsigned i = 0, e = DAGSize; i != e; ++i) {
    SUnit *SU = &SUnits[i];
    unsigned Degree = SU->Preds.size();
    // Temporarily use the Depth field as scratch space for the degree count.
    SU->Depth = Degree;

    // Is it a node without dependencies?
    if (Degree == 0) {
        assert(SU->Preds.empty() && "SUnit should have no predecessors");
        // Collect leaf nodes
        WorkList.push_back(SU);
    }
  }

  // Process nodes in the topological order
  while (!WorkList.empty()) {
    SUnit *SU = WorkList.back();
    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;
      }
    }

    SU->Depth = SUDepth;

    // Update degrees of all nodes depending on current SUnit
    for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
         I != E; ++I) {
      SUnit *SU = I->Dep;
      if (!--SU->Depth)
        // If all dependencies of the node are processed already,
        // then the longest path for the node can be computed now
        WorkList.push_back(SU);
    }
  }
}

/// CalculateHeights - compute heights using algorithms for the longest
/// paths in the DAG
void ScheduleDAG::CalculateHeights() {
  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);
    }
  }
}

/// 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;
}

/// 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<MemOperandSDNode>(Node->getOperand(N - 1).getNode()))
    --N; // Ignore MEMOPERAND nodes
  return N;
}

/// ComputeMemOperandsEnd - Find the index one past the last MemOperandSDNode
/// operand
unsigned ScheduleDAG::ComputeMemOperandsEnd(SDNode *Node) {
  unsigned N = Node->getNumOperands();
  while (N && Node->getOperand(N - 1).getValueType() == MVT::Flag)
    --N;
  if (N && Node->getOperand(N - 1).getValueType() == MVT::Other)
    --N; // Ignore chain if it exists.
  return N;
}


/// 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(&DAG);
    else
      cerr << "**** NOOP ****\n";
  }
}


/// Run - perform scheduling.
///
void ScheduleDAG::Run() {
  Schedule();
  
  DOUT << "*** Final schedule ***\n";
  DEBUG(dumpSchedule());
  DOUT << "\n";
}

/// SUnit - Scheduling unit. It's an wrapper around either a single SDNode or
/// a group of nodes flagged together.
void SUnit::dump(const SelectionDAG *G) const {
  cerr << "SU(" << NodeNum << "): ";
  if (Node)
    Node->dump(G);
  else
    cerr << "CROSS RC COPY ";
  cerr << "\n";
  if (FlaggedNodes.size() != 0) {
    for (unsigned i = 0, e = FlaggedNodes.size(); i != e; i++) {
      cerr << "    ";
      FlaggedNodes[i]->dump(G);
      cerr << "\n";
    }
  }
}

void SUnit::dumpAll(const SelectionDAG *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";
    }
  }
  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";
    }
  }
  cerr << "\n";
}