ScheduleDAG.cpp revision 7a2bdde0a0eebcd2125055e0eacaca040f0b766c
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/Debug.h" 23#include "llvm/Support/raw_ostream.h" 24#include <climits> 25using namespace llvm; 26 27ScheduleDAG::ScheduleDAG(MachineFunction &mf) 28 : TM(mf.getTarget()), 29 TII(TM.getInstrInfo()), 30 TRI(TM.getRegisterInfo()), 31 MF(mf), MRI(mf.getRegInfo()), 32 EntrySU(), ExitSU() { 33} 34 35ScheduleDAG::~ScheduleDAG() {} 36 37/// getInstrDesc helper to handle SDNodes. 38const TargetInstrDesc *ScheduleDAG::getNodeDesc(const SDNode *Node) const { 39 if (!Node || !Node->isMachineOpcode()) return NULL; 40 return &TII->get(Node->getMachineOpcode()); 41} 42 43/// dump - dump the schedule. 44void ScheduleDAG::dumpSchedule() const { 45 for (unsigned i = 0, e = Sequence.size(); i != e; i++) { 46 if (SUnit *SU = Sequence[i]) 47 SU->dump(this); 48 else 49 dbgs() << "**** NOOP ****\n"; 50 } 51} 52 53 54/// Run - perform scheduling. 55/// 56void ScheduleDAG::Run(MachineBasicBlock *bb, 57 MachineBasicBlock::iterator insertPos) { 58 BB = bb; 59 InsertPos = insertPos; 60 61 SUnits.clear(); 62 Sequence.clear(); 63 EntrySU = SUnit(); 64 ExitSU = SUnit(); 65 66 Schedule(); 67 68 DEBUG({ 69 dbgs() << "*** Final schedule ***\n"; 70 dumpSchedule(); 71 dbgs() << '\n'; 72 }); 73} 74 75/// addPred - This adds the specified edge as a pred of the current node if 76/// not already. It also adds the current node as a successor of the 77/// specified node. 78bool SUnit::addPred(const SDep &D) { 79 // If this node already has this depenence, don't add a redundant one. 80 for (SmallVector<SDep, 4>::const_iterator I = Preds.begin(), E = Preds.end(); 81 I != E; ++I) 82 if (*I == D) 83 return false; 84 // Now add a corresponding succ to N. 85 SDep P = D; 86 P.setSUnit(this); 87 SUnit *N = D.getSUnit(); 88 // Update the bookkeeping. 89 if (D.getKind() == SDep::Data) { 90 assert(NumPreds < UINT_MAX && "NumPreds will overflow!"); 91 assert(N->NumSuccs < UINT_MAX && "NumSuccs will overflow!"); 92 ++NumPreds; 93 ++N->NumSuccs; 94 } 95 if (!N->isScheduled) { 96 assert(NumPredsLeft < UINT_MAX && "NumPredsLeft will overflow!"); 97 ++NumPredsLeft; 98 } 99 if (!isScheduled) { 100 assert(N->NumSuccsLeft < UINT_MAX && "NumSuccsLeft will overflow!"); 101 ++N->NumSuccsLeft; 102 } 103 Preds.push_back(D); 104 N->Succs.push_back(P); 105 if (P.getLatency() != 0) { 106 this->setDepthDirty(); 107 N->setHeightDirty(); 108 } 109 return true; 110} 111 112/// removePred - This removes the specified edge as a pred of the current 113/// node if it exists. It also removes the current node as a successor of 114/// the specified node. 115void SUnit::removePred(const SDep &D) { 116 // Find the matching predecessor. 117 for (SmallVector<SDep, 4>::iterator I = Preds.begin(), E = Preds.end(); 118 I != E; ++I) 119 if (*I == D) { 120 bool FoundSucc = false; 121 // Find the corresponding successor in N. 122 SDep P = D; 123 P.setSUnit(this); 124 SUnit *N = D.getSUnit(); 125 for (SmallVector<SDep, 4>::iterator II = N->Succs.begin(), 126 EE = N->Succs.end(); II != EE; ++II) 127 if (*II == P) { 128 FoundSucc = true; 129 N->Succs.erase(II); 130 break; 131 } 132 assert(FoundSucc && "Mismatching preds / succs lists!"); 133 Preds.erase(I); 134 // Update the bookkeeping. 135 if (P.getKind() == SDep::Data) { 136 assert(NumPreds > 0 && "NumPreds will underflow!"); 137 assert(N->NumSuccs > 0 && "NumSuccs will underflow!"); 138 --NumPreds; 139 --N->NumSuccs; 140 } 141 if (!N->isScheduled) { 142 assert(NumPredsLeft > 0 && "NumPredsLeft will underflow!"); 143 --NumPredsLeft; 144 } 145 if (!isScheduled) { 146 assert(N->NumSuccsLeft > 0 && "NumSuccsLeft will underflow!"); 147 --N->NumSuccsLeft; 148 } 149 if (P.getLatency() != 0) { 150 this->setDepthDirty(); 151 N->setHeightDirty(); 152 } 153 return; 154 } 155} 156 157void SUnit::setDepthDirty() { 158 if (!isDepthCurrent) return; 159 SmallVector<SUnit*, 8> WorkList; 160 WorkList.push_back(this); 161 do { 162 SUnit *SU = WorkList.pop_back_val(); 163 SU->isDepthCurrent = false; 164 for (SUnit::const_succ_iterator I = SU->Succs.begin(), 165 E = SU->Succs.end(); I != E; ++I) { 166 SUnit *SuccSU = I->getSUnit(); 167 if (SuccSU->isDepthCurrent) 168 WorkList.push_back(SuccSU); 169 } 170 } while (!WorkList.empty()); 171} 172 173void SUnit::setHeightDirty() { 174 if (!isHeightCurrent) return; 175 SmallVector<SUnit*, 8> WorkList; 176 WorkList.push_back(this); 177 do { 178 SUnit *SU = WorkList.pop_back_val(); 179 SU->isHeightCurrent = false; 180 for (SUnit::const_pred_iterator I = SU->Preds.begin(), 181 E = SU->Preds.end(); I != E; ++I) { 182 SUnit *PredSU = I->getSUnit(); 183 if (PredSU->isHeightCurrent) 184 WorkList.push_back(PredSU); 185 } 186 } while (!WorkList.empty()); 187} 188 189/// setDepthToAtLeast - Update this node's successors to reflect the 190/// fact that this node's depth just increased. 191/// 192void SUnit::setDepthToAtLeast(unsigned NewDepth) { 193 if (NewDepth <= getDepth()) 194 return; 195 setDepthDirty(); 196 Depth = NewDepth; 197 isDepthCurrent = true; 198} 199 200/// setHeightToAtLeast - Update this node's predecessors to reflect the 201/// fact that this node's height just increased. 202/// 203void SUnit::setHeightToAtLeast(unsigned NewHeight) { 204 if (NewHeight <= getHeight()) 205 return; 206 setHeightDirty(); 207 Height = NewHeight; 208 isHeightCurrent = true; 209} 210 211/// ComputeDepth - Calculate the maximal path from the node to the exit. 212/// 213void SUnit::ComputeDepth() { 214 SmallVector<SUnit*, 8> WorkList; 215 WorkList.push_back(this); 216 do { 217 SUnit *Cur = WorkList.back(); 218 219 bool Done = true; 220 unsigned MaxPredDepth = 0; 221 for (SUnit::const_pred_iterator I = Cur->Preds.begin(), 222 E = Cur->Preds.end(); I != E; ++I) { 223 SUnit *PredSU = I->getSUnit(); 224 if (PredSU->isDepthCurrent) 225 MaxPredDepth = std::max(MaxPredDepth, 226 PredSU->Depth + I->getLatency()); 227 else { 228 Done = false; 229 WorkList.push_back(PredSU); 230 } 231 } 232 233 if (Done) { 234 WorkList.pop_back(); 235 if (MaxPredDepth != Cur->Depth) { 236 Cur->setDepthDirty(); 237 Cur->Depth = MaxPredDepth; 238 } 239 Cur->isDepthCurrent = true; 240 } 241 } while (!WorkList.empty()); 242} 243 244/// ComputeHeight - Calculate the maximal path from the node to the entry. 245/// 246void SUnit::ComputeHeight() { 247 SmallVector<SUnit*, 8> WorkList; 248 WorkList.push_back(this); 249 do { 250 SUnit *Cur = WorkList.back(); 251 252 bool Done = true; 253 unsigned MaxSuccHeight = 0; 254 for (SUnit::const_succ_iterator I = Cur->Succs.begin(), 255 E = Cur->Succs.end(); I != E; ++I) { 256 SUnit *SuccSU = I->getSUnit(); 257 if (SuccSU->isHeightCurrent) 258 MaxSuccHeight = std::max(MaxSuccHeight, 259 SuccSU->Height + I->getLatency()); 260 else { 261 Done = false; 262 WorkList.push_back(SuccSU); 263 } 264 } 265 266 if (Done) { 267 WorkList.pop_back(); 268 if (MaxSuccHeight != Cur->Height) { 269 Cur->setHeightDirty(); 270 Cur->Height = MaxSuccHeight; 271 } 272 Cur->isHeightCurrent = true; 273 } 274 } while (!WorkList.empty()); 275} 276 277/// SUnit - Scheduling unit. It's an wrapper around either a single SDNode or 278/// a group of nodes flagged together. 279void SUnit::dump(const ScheduleDAG *G) const { 280 dbgs() << "SU(" << NodeNum << "): "; 281 G->dumpNode(this); 282} 283 284void SUnit::dumpAll(const ScheduleDAG *G) const { 285 dump(G); 286 287 dbgs() << " # preds left : " << NumPredsLeft << "\n"; 288 dbgs() << " # succs left : " << NumSuccsLeft << "\n"; 289 dbgs() << " # rdefs left : " << NumRegDefsLeft << "\n"; 290 dbgs() << " Latency : " << Latency << "\n"; 291 dbgs() << " Depth : " << Depth << "\n"; 292 dbgs() << " Height : " << Height << "\n"; 293 294 if (Preds.size() != 0) { 295 dbgs() << " Predecessors:\n"; 296 for (SUnit::const_succ_iterator I = Preds.begin(), E = Preds.end(); 297 I != E; ++I) { 298 dbgs() << " "; 299 switch (I->getKind()) { 300 case SDep::Data: dbgs() << "val "; break; 301 case SDep::Anti: dbgs() << "anti"; break; 302 case SDep::Output: dbgs() << "out "; break; 303 case SDep::Order: dbgs() << "ch "; break; 304 } 305 dbgs() << "#"; 306 dbgs() << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")"; 307 if (I->isArtificial()) 308 dbgs() << " *"; 309 dbgs() << ": Latency=" << I->getLatency(); 310 dbgs() << "\n"; 311 } 312 } 313 if (Succs.size() != 0) { 314 dbgs() << " Successors:\n"; 315 for (SUnit::const_succ_iterator I = Succs.begin(), E = Succs.end(); 316 I != E; ++I) { 317 dbgs() << " "; 318 switch (I->getKind()) { 319 case SDep::Data: dbgs() << "val "; break; 320 case SDep::Anti: dbgs() << "anti"; break; 321 case SDep::Output: dbgs() << "out "; break; 322 case SDep::Order: dbgs() << "ch "; break; 323 } 324 dbgs() << "#"; 325 dbgs() << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")"; 326 if (I->isArtificial()) 327 dbgs() << " *"; 328 dbgs() << ": Latency=" << I->getLatency(); 329 dbgs() << "\n"; 330 } 331 } 332 dbgs() << "\n"; 333} 334 335#ifndef NDEBUG 336/// VerifySchedule - Verify that all SUnits were scheduled and that 337/// their state is consistent. 338/// 339void ScheduleDAG::VerifySchedule(bool isBottomUp) { 340 bool AnyNotSched = false; 341 unsigned DeadNodes = 0; 342 unsigned Noops = 0; 343 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) { 344 if (!SUnits[i].isScheduled) { 345 if (SUnits[i].NumPreds == 0 && SUnits[i].NumSuccs == 0) { 346 ++DeadNodes; 347 continue; 348 } 349 if (!AnyNotSched) 350 dbgs() << "*** Scheduling failed! ***\n"; 351 SUnits[i].dump(this); 352 dbgs() << "has not been scheduled!\n"; 353 AnyNotSched = true; 354 } 355 if (SUnits[i].isScheduled && 356 (isBottomUp ? SUnits[i].getHeight() : SUnits[i].getDepth()) > 357 unsigned(INT_MAX)) { 358 if (!AnyNotSched) 359 dbgs() << "*** Scheduling failed! ***\n"; 360 SUnits[i].dump(this); 361 dbgs() << "has an unexpected " 362 << (isBottomUp ? "Height" : "Depth") << " value!\n"; 363 AnyNotSched = true; 364 } 365 if (isBottomUp) { 366 if (SUnits[i].NumSuccsLeft != 0) { 367 if (!AnyNotSched) 368 dbgs() << "*** Scheduling failed! ***\n"; 369 SUnits[i].dump(this); 370 dbgs() << "has successors left!\n"; 371 AnyNotSched = true; 372 } 373 } else { 374 if (SUnits[i].NumPredsLeft != 0) { 375 if (!AnyNotSched) 376 dbgs() << "*** Scheduling failed! ***\n"; 377 SUnits[i].dump(this); 378 dbgs() << "has predecessors left!\n"; 379 AnyNotSched = true; 380 } 381 } 382 } 383 for (unsigned i = 0, e = Sequence.size(); i != e; ++i) 384 if (!Sequence[i]) 385 ++Noops; 386 assert(!AnyNotSched); 387 assert(Sequence.size() + DeadNodes - Noops == SUnits.size() && 388 "The number of nodes scheduled doesn't match the expected number!"); 389} 390#endif 391 392/// InitDAGTopologicalSorting - create the initial topological 393/// ordering from the DAG to be scheduled. 394/// 395/// The idea of the algorithm is taken from 396/// "Online algorithms for managing the topological order of 397/// a directed acyclic graph" by David J. Pearce and Paul H.J. Kelly 398/// This is the MNR algorithm, which was first introduced by 399/// A. Marchetti-Spaccamela, U. Nanni and H. Rohnert in 400/// "Maintaining a topological order under edge insertions". 401/// 402/// Short description of the algorithm: 403/// 404/// Topological ordering, ord, of a DAG maps each node to a topological 405/// index so that for all edges X->Y it is the case that ord(X) < ord(Y). 406/// 407/// This means that if there is a path from the node X to the node Z, 408/// then ord(X) < ord(Z). 409/// 410/// This property can be used to check for reachability of nodes: 411/// if Z is reachable from X, then an insertion of the edge Z->X would 412/// create a cycle. 413/// 414/// The algorithm first computes a topological ordering for the DAG by 415/// initializing the Index2Node and Node2Index arrays and then tries to keep 416/// the ordering up-to-date after edge insertions by reordering the DAG. 417/// 418/// On insertion of the edge X->Y, the algorithm first marks by calling DFS 419/// the nodes reachable from Y, and then shifts them using Shift to lie 420/// immediately after X in Index2Node. 421void ScheduleDAGTopologicalSort::InitDAGTopologicalSorting() { 422 unsigned DAGSize = SUnits.size(); 423 std::vector<SUnit*> WorkList; 424 WorkList.reserve(DAGSize); 425 426 Index2Node.resize(DAGSize); 427 Node2Index.resize(DAGSize); 428 429 // Initialize the data structures. 430 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 431 SUnit *SU = &SUnits[i]; 432 int NodeNum = SU->NodeNum; 433 unsigned Degree = SU->Succs.size(); 434 // Temporarily use the Node2Index array as scratch space for degree counts. 435 Node2Index[NodeNum] = Degree; 436 437 // Is it a node without dependencies? 438 if (Degree == 0) { 439 assert(SU->Succs.empty() && "SUnit should have no successors"); 440 // Collect leaf nodes. 441 WorkList.push_back(SU); 442 } 443 } 444 445 int Id = DAGSize; 446 while (!WorkList.empty()) { 447 SUnit *SU = WorkList.back(); 448 WorkList.pop_back(); 449 Allocate(SU->NodeNum, --Id); 450 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); 451 I != E; ++I) { 452 SUnit *SU = I->getSUnit(); 453 if (!--Node2Index[SU->NodeNum]) 454 // If all dependencies of the node are processed already, 455 // then the node can be computed now. 456 WorkList.push_back(SU); 457 } 458 } 459 460 Visited.resize(DAGSize); 461 462#ifndef NDEBUG 463 // Check correctness of the ordering 464 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 465 SUnit *SU = &SUnits[i]; 466 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); 467 I != E; ++I) { 468 assert(Node2Index[SU->NodeNum] > Node2Index[I->getSUnit()->NodeNum] && 469 "Wrong topological sorting"); 470 } 471 } 472#endif 473} 474 475/// AddPred - Updates the topological ordering to accommodate an edge 476/// to be added from SUnit X to SUnit Y. 477void ScheduleDAGTopologicalSort::AddPred(SUnit *Y, SUnit *X) { 478 int UpperBound, LowerBound; 479 LowerBound = Node2Index[Y->NodeNum]; 480 UpperBound = Node2Index[X->NodeNum]; 481 bool HasLoop = false; 482 // Is Ord(X) < Ord(Y) ? 483 if (LowerBound < UpperBound) { 484 // Update the topological order. 485 Visited.reset(); 486 DFS(Y, UpperBound, HasLoop); 487 assert(!HasLoop && "Inserted edge creates a loop!"); 488 // Recompute topological indexes. 489 Shift(Visited, LowerBound, UpperBound); 490 } 491} 492 493/// RemovePred - Updates the topological ordering to accommodate an 494/// an edge to be removed from the specified node N from the predecessors 495/// of the current node M. 496void ScheduleDAGTopologicalSort::RemovePred(SUnit *M, SUnit *N) { 497 // InitDAGTopologicalSorting(); 498} 499 500/// DFS - Make a DFS traversal to mark all nodes reachable from SU and mark 501/// all nodes affected by the edge insertion. These nodes will later get new 502/// topological indexes by means of the Shift method. 503void ScheduleDAGTopologicalSort::DFS(const SUnit *SU, int UpperBound, 504 bool &HasLoop) { 505 std::vector<const SUnit*> WorkList; 506 WorkList.reserve(SUnits.size()); 507 508 WorkList.push_back(SU); 509 do { 510 SU = WorkList.back(); 511 WorkList.pop_back(); 512 Visited.set(SU->NodeNum); 513 for (int I = SU->Succs.size()-1; I >= 0; --I) { 514 int s = SU->Succs[I].getSUnit()->NodeNum; 515 if (Node2Index[s] == UpperBound) { 516 HasLoop = true; 517 return; 518 } 519 // Visit successors if not already and in affected region. 520 if (!Visited.test(s) && Node2Index[s] < UpperBound) { 521 WorkList.push_back(SU->Succs[I].getSUnit()); 522 } 523 } 524 } while (!WorkList.empty()); 525} 526 527/// Shift - Renumber the nodes so that the topological ordering is 528/// preserved. 529void ScheduleDAGTopologicalSort::Shift(BitVector& Visited, int LowerBound, 530 int UpperBound) { 531 std::vector<int> L; 532 int shift = 0; 533 int i; 534 535 for (i = LowerBound; i <= UpperBound; ++i) { 536 // w is node at topological index i. 537 int w = Index2Node[i]; 538 if (Visited.test(w)) { 539 // Unmark. 540 Visited.reset(w); 541 L.push_back(w); 542 shift = shift + 1; 543 } else { 544 Allocate(w, i - shift); 545 } 546 } 547 548 for (unsigned j = 0; j < L.size(); ++j) { 549 Allocate(L[j], i - shift); 550 i = i + 1; 551 } 552} 553 554 555/// WillCreateCycle - Returns true if adding an edge from SU to TargetSU will 556/// create a cycle. 557bool ScheduleDAGTopologicalSort::WillCreateCycle(SUnit *SU, SUnit *TargetSU) { 558 if (IsReachable(TargetSU, SU)) 559 return true; 560 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); 561 I != E; ++I) 562 if (I->isAssignedRegDep() && 563 IsReachable(TargetSU, I->getSUnit())) 564 return true; 565 return false; 566} 567 568/// IsReachable - Checks if SU is reachable from TargetSU. 569bool ScheduleDAGTopologicalSort::IsReachable(const SUnit *SU, 570 const SUnit *TargetSU) { 571 // If insertion of the edge SU->TargetSU would create a cycle 572 // then there is a path from TargetSU to SU. 573 int UpperBound, LowerBound; 574 LowerBound = Node2Index[TargetSU->NodeNum]; 575 UpperBound = Node2Index[SU->NodeNum]; 576 bool HasLoop = false; 577 // Is Ord(TargetSU) < Ord(SU) ? 578 if (LowerBound < UpperBound) { 579 Visited.reset(); 580 // There may be a path from TargetSU to SU. Check for it. 581 DFS(TargetSU, UpperBound, HasLoop); 582 } 583 return HasLoop; 584} 585 586/// Allocate - assign the topological index to the node n. 587void ScheduleDAGTopologicalSort::Allocate(int n, int index) { 588 Node2Index[n] = index; 589 Index2Node[index] = n; 590} 591 592ScheduleDAGTopologicalSort:: 593ScheduleDAGTopologicalSort(std::vector<SUnit> &sunits) : SUnits(sunits) {} 594 595ScheduleHazardRecognizer::~ScheduleHazardRecognizer() {} 596