1 //===---- ScheduleDAG.cpp - Implement the ScheduleDAG 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 "llvm/CodeGen/ScheduleDAG.h" 17 #include "llvm/CodeGen/ScheduleHazardRecognizer.h" 18 #include "llvm/CodeGen/SelectionDAGNodes.h" 19 #include "llvm/Target/TargetMachine.h" 20 #include "llvm/Target/TargetInstrInfo.h" 21 #include "llvm/Target/TargetRegisterInfo.h" 22 #include "llvm/Support/CommandLine.h" 23 #include "llvm/Support/Debug.h" 24 #include "llvm/Support/raw_ostream.h" 25 #include <climits> 26 using namespace llvm; 27 28 #ifndef NDEBUG 29 static cl::opt<bool> StressSchedOpt( 30 "stress-sched", cl::Hidden, cl::init(false), 31 cl::desc("Stress test instruction scheduling")); 32 #endif 33 34 void SchedulingPriorityQueue::anchor() { } 35 36 ScheduleDAG::ScheduleDAG(MachineFunction &mf) 37 : TM(mf.getTarget()), 38 TII(TM.getInstrInfo()), 39 TRI(TM.getRegisterInfo()), 40 MF(mf), MRI(mf.getRegInfo()), 41 EntrySU(), ExitSU() { 42 #ifndef NDEBUG 43 StressSched = StressSchedOpt; 44 #endif 45 } 46 47 ScheduleDAG::~ScheduleDAG() {} 48 49 /// Clear the DAG state (e.g. between scheduling regions). 50 void ScheduleDAG::clearDAG() { 51 SUnits.clear(); 52 EntrySU = SUnit(); 53 ExitSU = SUnit(); 54 } 55 56 /// getInstrDesc helper to handle SDNodes. 57 const MCInstrDesc *ScheduleDAG::getNodeDesc(const SDNode *Node) const { 58 if (!Node || !Node->isMachineOpcode()) return NULL; 59 return &TII->get(Node->getMachineOpcode()); 60 } 61 62 /// addPred - This adds the specified edge as a pred of the current node if 63 /// not already. It also adds the current node as a successor of the 64 /// specified node. 65 bool SUnit::addPred(const SDep &D) { 66 // If this node already has this depenence, don't add a redundant one. 67 for (SmallVector<SDep, 4>::const_iterator I = Preds.begin(), E = Preds.end(); 68 I != E; ++I) 69 if (*I == D) 70 return false; 71 // Now add a corresponding succ to N. 72 SDep P = D; 73 P.setSUnit(this); 74 SUnit *N = D.getSUnit(); 75 // Update the bookkeeping. 76 if (D.getKind() == SDep::Data) { 77 assert(NumPreds < UINT_MAX && "NumPreds will overflow!"); 78 assert(N->NumSuccs < UINT_MAX && "NumSuccs will overflow!"); 79 ++NumPreds; 80 ++N->NumSuccs; 81 } 82 if (!N->isScheduled) { 83 assert(NumPredsLeft < UINT_MAX && "NumPredsLeft will overflow!"); 84 ++NumPredsLeft; 85 } 86 if (!isScheduled) { 87 assert(N->NumSuccsLeft < UINT_MAX && "NumSuccsLeft will overflow!"); 88 ++N->NumSuccsLeft; 89 } 90 Preds.push_back(D); 91 N->Succs.push_back(P); 92 if (P.getLatency() != 0) { 93 this->setDepthDirty(); 94 N->setHeightDirty(); 95 } 96 return true; 97 } 98 99 /// removePred - This removes the specified edge as a pred of the current 100 /// node if it exists. It also removes the current node as a successor of 101 /// the specified node. 102 void SUnit::removePred(const SDep &D) { 103 // Find the matching predecessor. 104 for (SmallVector<SDep, 4>::iterator I = Preds.begin(), E = Preds.end(); 105 I != E; ++I) 106 if (*I == D) { 107 bool FoundSucc = false; 108 // Find the corresponding successor in N. 109 SDep P = D; 110 P.setSUnit(this); 111 SUnit *N = D.getSUnit(); 112 for (SmallVector<SDep, 4>::iterator II = N->Succs.begin(), 113 EE = N->Succs.end(); II != EE; ++II) 114 if (*II == P) { 115 FoundSucc = true; 116 N->Succs.erase(II); 117 break; 118 } 119 assert(FoundSucc && "Mismatching preds / succs lists!"); 120 (void)FoundSucc; 121 Preds.erase(I); 122 // Update the bookkeeping. 123 if (P.getKind() == SDep::Data) { 124 assert(NumPreds > 0 && "NumPreds will underflow!"); 125 assert(N->NumSuccs > 0 && "NumSuccs will underflow!"); 126 --NumPreds; 127 --N->NumSuccs; 128 } 129 if (!N->isScheduled) { 130 assert(NumPredsLeft > 0 && "NumPredsLeft will underflow!"); 131 --NumPredsLeft; 132 } 133 if (!isScheduled) { 134 assert(N->NumSuccsLeft > 0 && "NumSuccsLeft will underflow!"); 135 --N->NumSuccsLeft; 136 } 137 if (P.getLatency() != 0) { 138 this->setDepthDirty(); 139 N->setHeightDirty(); 140 } 141 return; 142 } 143 } 144 145 void SUnit::setDepthDirty() { 146 if (!isDepthCurrent) return; 147 SmallVector<SUnit*, 8> WorkList; 148 WorkList.push_back(this); 149 do { 150 SUnit *SU = WorkList.pop_back_val(); 151 SU->isDepthCurrent = false; 152 for (SUnit::const_succ_iterator I = SU->Succs.begin(), 153 E = SU->Succs.end(); I != E; ++I) { 154 SUnit *SuccSU = I->getSUnit(); 155 if (SuccSU->isDepthCurrent) 156 WorkList.push_back(SuccSU); 157 } 158 } while (!WorkList.empty()); 159 } 160 161 void SUnit::setHeightDirty() { 162 if (!isHeightCurrent) return; 163 SmallVector<SUnit*, 8> WorkList; 164 WorkList.push_back(this); 165 do { 166 SUnit *SU = WorkList.pop_back_val(); 167 SU->isHeightCurrent = false; 168 for (SUnit::const_pred_iterator I = SU->Preds.begin(), 169 E = SU->Preds.end(); I != E; ++I) { 170 SUnit *PredSU = I->getSUnit(); 171 if (PredSU->isHeightCurrent) 172 WorkList.push_back(PredSU); 173 } 174 } while (!WorkList.empty()); 175 } 176 177 /// setDepthToAtLeast - Update this node's successors to reflect the 178 /// fact that this node's depth just increased. 179 /// 180 void SUnit::setDepthToAtLeast(unsigned NewDepth) { 181 if (NewDepth <= getDepth()) 182 return; 183 setDepthDirty(); 184 Depth = NewDepth; 185 isDepthCurrent = true; 186 } 187 188 /// setHeightToAtLeast - Update this node's predecessors to reflect the 189 /// fact that this node's height just increased. 190 /// 191 void SUnit::setHeightToAtLeast(unsigned NewHeight) { 192 if (NewHeight <= getHeight()) 193 return; 194 setHeightDirty(); 195 Height = NewHeight; 196 isHeightCurrent = true; 197 } 198 199 /// ComputeDepth - Calculate the maximal path from the node to the exit. 200 /// 201 void SUnit::ComputeDepth() { 202 SmallVector<SUnit*, 8> WorkList; 203 WorkList.push_back(this); 204 do { 205 SUnit *Cur = WorkList.back(); 206 207 bool Done = true; 208 unsigned MaxPredDepth = 0; 209 for (SUnit::const_pred_iterator I = Cur->Preds.begin(), 210 E = Cur->Preds.end(); I != E; ++I) { 211 SUnit *PredSU = I->getSUnit(); 212 if (PredSU->isDepthCurrent) 213 MaxPredDepth = std::max(MaxPredDepth, 214 PredSU->Depth + I->getLatency()); 215 else { 216 Done = false; 217 WorkList.push_back(PredSU); 218 } 219 } 220 221 if (Done) { 222 WorkList.pop_back(); 223 if (MaxPredDepth != Cur->Depth) { 224 Cur->setDepthDirty(); 225 Cur->Depth = MaxPredDepth; 226 } 227 Cur->isDepthCurrent = true; 228 } 229 } while (!WorkList.empty()); 230 } 231 232 /// ComputeHeight - Calculate the maximal path from the node to the entry. 233 /// 234 void SUnit::ComputeHeight() { 235 SmallVector<SUnit*, 8> WorkList; 236 WorkList.push_back(this); 237 do { 238 SUnit *Cur = WorkList.back(); 239 240 bool Done = true; 241 unsigned MaxSuccHeight = 0; 242 for (SUnit::const_succ_iterator I = Cur->Succs.begin(), 243 E = Cur->Succs.end(); I != E; ++I) { 244 SUnit *SuccSU = I->getSUnit(); 245 if (SuccSU->isHeightCurrent) 246 MaxSuccHeight = std::max(MaxSuccHeight, 247 SuccSU->Height + I->getLatency()); 248 else { 249 Done = false; 250 WorkList.push_back(SuccSU); 251 } 252 } 253 254 if (Done) { 255 WorkList.pop_back(); 256 if (MaxSuccHeight != Cur->Height) { 257 Cur->setHeightDirty(); 258 Cur->Height = MaxSuccHeight; 259 } 260 Cur->isHeightCurrent = true; 261 } 262 } while (!WorkList.empty()); 263 } 264 265 /// SUnit - Scheduling unit. It's an wrapper around either a single SDNode or 266 /// a group of nodes flagged together. 267 void SUnit::dump(const ScheduleDAG *G) const { 268 dbgs() << "SU(" << NodeNum << "): "; 269 G->dumpNode(this); 270 } 271 272 void SUnit::dumpAll(const ScheduleDAG *G) const { 273 dump(G); 274 275 dbgs() << " # preds left : " << NumPredsLeft << "\n"; 276 dbgs() << " # succs left : " << NumSuccsLeft << "\n"; 277 dbgs() << " # rdefs left : " << NumRegDefsLeft << "\n"; 278 dbgs() << " Latency : " << Latency << "\n"; 279 dbgs() << " Depth : " << Depth << "\n"; 280 dbgs() << " Height : " << Height << "\n"; 281 282 if (Preds.size() != 0) { 283 dbgs() << " Predecessors:\n"; 284 for (SUnit::const_succ_iterator I = Preds.begin(), E = Preds.end(); 285 I != E; ++I) { 286 dbgs() << " "; 287 switch (I->getKind()) { 288 case SDep::Data: dbgs() << "val "; break; 289 case SDep::Anti: dbgs() << "anti"; break; 290 case SDep::Output: dbgs() << "out "; break; 291 case SDep::Order: dbgs() << "ch "; break; 292 } 293 dbgs() << "SU(" << I->getSUnit()->NodeNum << ")"; 294 if (I->isArtificial()) 295 dbgs() << " *"; 296 dbgs() << ": Latency=" << I->getLatency(); 297 if (I->isAssignedRegDep()) 298 dbgs() << " Reg=" << PrintReg(I->getReg(), G->TRI); 299 dbgs() << "\n"; 300 } 301 } 302 if (Succs.size() != 0) { 303 dbgs() << " Successors:\n"; 304 for (SUnit::const_succ_iterator I = Succs.begin(), E = Succs.end(); 305 I != E; ++I) { 306 dbgs() << " "; 307 switch (I->getKind()) { 308 case SDep::Data: dbgs() << "val "; break; 309 case SDep::Anti: dbgs() << "anti"; break; 310 case SDep::Output: dbgs() << "out "; break; 311 case SDep::Order: dbgs() << "ch "; break; 312 } 313 dbgs() << "SU(" << I->getSUnit()->NodeNum << ")"; 314 if (I->isArtificial()) 315 dbgs() << " *"; 316 dbgs() << ": Latency=" << I->getLatency(); 317 dbgs() << "\n"; 318 } 319 } 320 dbgs() << "\n"; 321 } 322 323 #ifndef NDEBUG 324 /// VerifyScheduledDAG - Verify that all SUnits were scheduled and that 325 /// their state is consistent. Return the number of scheduled nodes. 326 /// 327 unsigned ScheduleDAG::VerifyScheduledDAG(bool isBottomUp) { 328 bool AnyNotSched = false; 329 unsigned DeadNodes = 0; 330 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) { 331 if (!SUnits[i].isScheduled) { 332 if (SUnits[i].NumPreds == 0 && SUnits[i].NumSuccs == 0) { 333 ++DeadNodes; 334 continue; 335 } 336 if (!AnyNotSched) 337 dbgs() << "*** Scheduling failed! ***\n"; 338 SUnits[i].dump(this); 339 dbgs() << "has not been scheduled!\n"; 340 AnyNotSched = true; 341 } 342 if (SUnits[i].isScheduled && 343 (isBottomUp ? SUnits[i].getHeight() : SUnits[i].getDepth()) > 344 unsigned(INT_MAX)) { 345 if (!AnyNotSched) 346 dbgs() << "*** Scheduling failed! ***\n"; 347 SUnits[i].dump(this); 348 dbgs() << "has an unexpected " 349 << (isBottomUp ? "Height" : "Depth") << " value!\n"; 350 AnyNotSched = true; 351 } 352 if (isBottomUp) { 353 if (SUnits[i].NumSuccsLeft != 0) { 354 if (!AnyNotSched) 355 dbgs() << "*** Scheduling failed! ***\n"; 356 SUnits[i].dump(this); 357 dbgs() << "has successors left!\n"; 358 AnyNotSched = true; 359 } 360 } else { 361 if (SUnits[i].NumPredsLeft != 0) { 362 if (!AnyNotSched) 363 dbgs() << "*** Scheduling failed! ***\n"; 364 SUnits[i].dump(this); 365 dbgs() << "has predecessors left!\n"; 366 AnyNotSched = true; 367 } 368 } 369 } 370 assert(!AnyNotSched); 371 return SUnits.size() - DeadNodes; 372 } 373 #endif 374 375 /// InitDAGTopologicalSorting - create the initial topological 376 /// ordering from the DAG to be scheduled. 377 /// 378 /// The idea of the algorithm is taken from 379 /// "Online algorithms for managing the topological order of 380 /// a directed acyclic graph" by David J. Pearce and Paul H.J. Kelly 381 /// This is the MNR algorithm, which was first introduced by 382 /// A. Marchetti-Spaccamela, U. Nanni and H. Rohnert in 383 /// "Maintaining a topological order under edge insertions". 384 /// 385 /// Short description of the algorithm: 386 /// 387 /// Topological ordering, ord, of a DAG maps each node to a topological 388 /// index so that for all edges X->Y it is the case that ord(X) < ord(Y). 389 /// 390 /// This means that if there is a path from the node X to the node Z, 391 /// then ord(X) < ord(Z). 392 /// 393 /// This property can be used to check for reachability of nodes: 394 /// if Z is reachable from X, then an insertion of the edge Z->X would 395 /// create a cycle. 396 /// 397 /// The algorithm first computes a topological ordering for the DAG by 398 /// initializing the Index2Node and Node2Index arrays and then tries to keep 399 /// the ordering up-to-date after edge insertions by reordering the DAG. 400 /// 401 /// On insertion of the edge X->Y, the algorithm first marks by calling DFS 402 /// the nodes reachable from Y, and then shifts them using Shift to lie 403 /// immediately after X in Index2Node. 404 void ScheduleDAGTopologicalSort::InitDAGTopologicalSorting() { 405 unsigned DAGSize = SUnits.size(); 406 std::vector<SUnit*> WorkList; 407 WorkList.reserve(DAGSize); 408 409 Index2Node.resize(DAGSize); 410 Node2Index.resize(DAGSize); 411 412 // Initialize the data structures. 413 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 414 SUnit *SU = &SUnits[i]; 415 int NodeNum = SU->NodeNum; 416 unsigned Degree = SU->Succs.size(); 417 // Temporarily use the Node2Index array as scratch space for degree counts. 418 Node2Index[NodeNum] = Degree; 419 420 // Is it a node without dependencies? 421 if (Degree == 0) { 422 assert(SU->Succs.empty() && "SUnit should have no successors"); 423 // Collect leaf nodes. 424 WorkList.push_back(SU); 425 } 426 } 427 428 int Id = DAGSize; 429 while (!WorkList.empty()) { 430 SUnit *SU = WorkList.back(); 431 WorkList.pop_back(); 432 Allocate(SU->NodeNum, --Id); 433 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); 434 I != E; ++I) { 435 SUnit *SU = I->getSUnit(); 436 if (!--Node2Index[SU->NodeNum]) 437 // If all dependencies of the node are processed already, 438 // then the node can be computed now. 439 WorkList.push_back(SU); 440 } 441 } 442 443 Visited.resize(DAGSize); 444 445 #ifndef NDEBUG 446 // Check correctness of the ordering 447 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 448 SUnit *SU = &SUnits[i]; 449 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); 450 I != E; ++I) { 451 assert(Node2Index[SU->NodeNum] > Node2Index[I->getSUnit()->NodeNum] && 452 "Wrong topological sorting"); 453 } 454 } 455 #endif 456 } 457 458 /// AddPred - Updates the topological ordering to accommodate an edge 459 /// to be added from SUnit X to SUnit Y. 460 void ScheduleDAGTopologicalSort::AddPred(SUnit *Y, SUnit *X) { 461 int UpperBound, LowerBound; 462 LowerBound = Node2Index[Y->NodeNum]; 463 UpperBound = Node2Index[X->NodeNum]; 464 bool HasLoop = false; 465 // Is Ord(X) < Ord(Y) ? 466 if (LowerBound < UpperBound) { 467 // Update the topological order. 468 Visited.reset(); 469 DFS(Y, UpperBound, HasLoop); 470 assert(!HasLoop && "Inserted edge creates a loop!"); 471 // Recompute topological indexes. 472 Shift(Visited, LowerBound, UpperBound); 473 } 474 } 475 476 /// RemovePred - Updates the topological ordering to accommodate an 477 /// an edge to be removed from the specified node N from the predecessors 478 /// of the current node M. 479 void ScheduleDAGTopologicalSort::RemovePred(SUnit *M, SUnit *N) { 480 // InitDAGTopologicalSorting(); 481 } 482 483 /// DFS - Make a DFS traversal to mark all nodes reachable from SU and mark 484 /// all nodes affected by the edge insertion. These nodes will later get new 485 /// topological indexes by means of the Shift method. 486 void ScheduleDAGTopologicalSort::DFS(const SUnit *SU, int UpperBound, 487 bool &HasLoop) { 488 std::vector<const SUnit*> WorkList; 489 WorkList.reserve(SUnits.size()); 490 491 WorkList.push_back(SU); 492 do { 493 SU = WorkList.back(); 494 WorkList.pop_back(); 495 Visited.set(SU->NodeNum); 496 for (int I = SU->Succs.size()-1; I >= 0; --I) { 497 int s = SU->Succs[I].getSUnit()->NodeNum; 498 if (Node2Index[s] == UpperBound) { 499 HasLoop = true; 500 return; 501 } 502 // Visit successors if not already and in affected region. 503 if (!Visited.test(s) && Node2Index[s] < UpperBound) { 504 WorkList.push_back(SU->Succs[I].getSUnit()); 505 } 506 } 507 } while (!WorkList.empty()); 508 } 509 510 /// Shift - Renumber the nodes so that the topological ordering is 511 /// preserved. 512 void ScheduleDAGTopologicalSort::Shift(BitVector& Visited, int LowerBound, 513 int UpperBound) { 514 std::vector<int> L; 515 int shift = 0; 516 int i; 517 518 for (i = LowerBound; i <= UpperBound; ++i) { 519 // w is node at topological index i. 520 int w = Index2Node[i]; 521 if (Visited.test(w)) { 522 // Unmark. 523 Visited.reset(w); 524 L.push_back(w); 525 shift = shift + 1; 526 } else { 527 Allocate(w, i - shift); 528 } 529 } 530 531 for (unsigned j = 0; j < L.size(); ++j) { 532 Allocate(L[j], i - shift); 533 i = i + 1; 534 } 535 } 536 537 538 /// WillCreateCycle - Returns true if adding an edge from SU to TargetSU will 539 /// create a cycle. 540 bool ScheduleDAGTopologicalSort::WillCreateCycle(SUnit *SU, SUnit *TargetSU) { 541 if (IsReachable(TargetSU, SU)) 542 return true; 543 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); 544 I != E; ++I) 545 if (I->isAssignedRegDep() && 546 IsReachable(TargetSU, I->getSUnit())) 547 return true; 548 return false; 549 } 550 551 /// IsReachable - Checks if SU is reachable from TargetSU. 552 bool ScheduleDAGTopologicalSort::IsReachable(const SUnit *SU, 553 const SUnit *TargetSU) { 554 // If insertion of the edge SU->TargetSU would create a cycle 555 // then there is a path from TargetSU to SU. 556 int UpperBound, LowerBound; 557 LowerBound = Node2Index[TargetSU->NodeNum]; 558 UpperBound = Node2Index[SU->NodeNum]; 559 bool HasLoop = false; 560 // Is Ord(TargetSU) < Ord(SU) ? 561 if (LowerBound < UpperBound) { 562 Visited.reset(); 563 // There may be a path from TargetSU to SU. Check for it. 564 DFS(TargetSU, UpperBound, HasLoop); 565 } 566 return HasLoop; 567 } 568 569 /// Allocate - assign the topological index to the node n. 570 void ScheduleDAGTopologicalSort::Allocate(int n, int index) { 571 Node2Index[n] = index; 572 Index2Node[index] = n; 573 } 574 575 ScheduleDAGTopologicalSort:: 576 ScheduleDAGTopologicalSort(std::vector<SUnit> &sunits) : SUnits(sunits) {} 577 578 ScheduleHazardRecognizer::~ScheduleHazardRecognizer() {} 579