ScheduleDAG.cpp revision d04a8d4b33ff316ca4cf961e06c9e312eff8e64f
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/Support/CommandLine.h" 20#include "llvm/Support/Debug.h" 21#include "llvm/Support/raw_ostream.h" 22#include "llvm/Target/TargetInstrInfo.h" 23#include "llvm/Target/TargetMachine.h" 24#include "llvm/Target/TargetRegisterInfo.h" 25#include <climits> 26using namespace llvm; 27 28#ifndef NDEBUG 29static cl::opt<bool> StressSchedOpt( 30 "stress-sched", cl::Hidden, cl::init(false), 31 cl::desc("Stress test instruction scheduling")); 32#endif 33 34void SchedulingPriorityQueue::anchor() { } 35 36ScheduleDAG::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 47ScheduleDAG::~ScheduleDAG() {} 48 49/// Clear the DAG state (e.g. between scheduling regions). 50void ScheduleDAG::clearDAG() { 51 SUnits.clear(); 52 EntrySU = SUnit(); 53 ExitSU = SUnit(); 54} 55 56/// getInstrDesc helper to handle SDNodes. 57const 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. 65bool SUnit::addPred(const SDep &D, bool Required) { 66 // If this node already has this depenence, don't add a redundant one. 67 for (SmallVector<SDep, 4>::iterator I = Preds.begin(), E = Preds.end(); 68 I != E; ++I) { 69 // Zero-latency weak edges may be added purely for heuristic ordering. Don't 70 // add them if another kind of edge already exists. 71 if (!Required && I->getSUnit() == D.getSUnit()) 72 return false; 73 if (I->overlaps(D)) { 74 // Extend the latency if needed. Equivalent to removePred(I) + addPred(D). 75 if (I->getLatency() < D.getLatency()) { 76 SUnit *PredSU = I->getSUnit(); 77 // Find the corresponding successor in N. 78 SDep ForwardD = *I; 79 ForwardD.setSUnit(this); 80 for (SmallVector<SDep, 4>::iterator II = PredSU->Succs.begin(), 81 EE = PredSU->Succs.end(); II != EE; ++II) { 82 if (*II == ForwardD) { 83 II->setLatency(D.getLatency()); 84 break; 85 } 86 } 87 I->setLatency(D.getLatency()); 88 } 89 return false; 90 } 91 } 92 // Now add a corresponding succ to N. 93 SDep P = D; 94 P.setSUnit(this); 95 SUnit *N = D.getSUnit(); 96 // Update the bookkeeping. 97 if (D.getKind() == SDep::Data) { 98 assert(NumPreds < UINT_MAX && "NumPreds will overflow!"); 99 assert(N->NumSuccs < UINT_MAX && "NumSuccs will overflow!"); 100 ++NumPreds; 101 ++N->NumSuccs; 102 } 103 if (!N->isScheduled) { 104 if (D.isWeak()) { 105 ++WeakPredsLeft; 106 } 107 else { 108 assert(NumPredsLeft < UINT_MAX && "NumPredsLeft will overflow!"); 109 ++NumPredsLeft; 110 } 111 } 112 if (!isScheduled) { 113 if (D.isWeak()) { 114 ++N->WeakSuccsLeft; 115 } 116 else { 117 assert(N->NumSuccsLeft < UINT_MAX && "NumSuccsLeft will overflow!"); 118 ++N->NumSuccsLeft; 119 } 120 } 121 Preds.push_back(D); 122 N->Succs.push_back(P); 123 if (P.getLatency() != 0) { 124 this->setDepthDirty(); 125 N->setHeightDirty(); 126 } 127 return true; 128} 129 130/// removePred - This removes the specified edge as a pred of the current 131/// node if it exists. It also removes the current node as a successor of 132/// the specified node. 133void SUnit::removePred(const SDep &D) { 134 // Find the matching predecessor. 135 for (SmallVector<SDep, 4>::iterator I = Preds.begin(), E = Preds.end(); 136 I != E; ++I) 137 if (*I == D) { 138 bool FoundSucc = false; 139 // Find the corresponding successor in N. 140 SDep P = D; 141 P.setSUnit(this); 142 SUnit *N = D.getSUnit(); 143 for (SmallVector<SDep, 4>::iterator II = N->Succs.begin(), 144 EE = N->Succs.end(); II != EE; ++II) 145 if (*II == P) { 146 FoundSucc = true; 147 N->Succs.erase(II); 148 break; 149 } 150 assert(FoundSucc && "Mismatching preds / succs lists!"); 151 (void)FoundSucc; 152 Preds.erase(I); 153 // Update the bookkeeping. 154 if (P.getKind() == SDep::Data) { 155 assert(NumPreds > 0 && "NumPreds will underflow!"); 156 assert(N->NumSuccs > 0 && "NumSuccs will underflow!"); 157 --NumPreds; 158 --N->NumSuccs; 159 } 160 if (!N->isScheduled) { 161 if (D.isWeak()) 162 --WeakPredsLeft; 163 else { 164 assert(NumPredsLeft > 0 && "NumPredsLeft will underflow!"); 165 --NumPredsLeft; 166 } 167 } 168 if (!isScheduled) { 169 if (D.isWeak()) 170 --N->WeakSuccsLeft; 171 else { 172 assert(N->NumSuccsLeft > 0 && "NumSuccsLeft will underflow!"); 173 --N->NumSuccsLeft; 174 } 175 } 176 if (P.getLatency() != 0) { 177 this->setDepthDirty(); 178 N->setHeightDirty(); 179 } 180 return; 181 } 182} 183 184void SUnit::setDepthDirty() { 185 if (!isDepthCurrent) return; 186 SmallVector<SUnit*, 8> WorkList; 187 WorkList.push_back(this); 188 do { 189 SUnit *SU = WorkList.pop_back_val(); 190 SU->isDepthCurrent = false; 191 for (SUnit::const_succ_iterator I = SU->Succs.begin(), 192 E = SU->Succs.end(); I != E; ++I) { 193 SUnit *SuccSU = I->getSUnit(); 194 if (SuccSU->isDepthCurrent) 195 WorkList.push_back(SuccSU); 196 } 197 } while (!WorkList.empty()); 198} 199 200void SUnit::setHeightDirty() { 201 if (!isHeightCurrent) return; 202 SmallVector<SUnit*, 8> WorkList; 203 WorkList.push_back(this); 204 do { 205 SUnit *SU = WorkList.pop_back_val(); 206 SU->isHeightCurrent = false; 207 for (SUnit::const_pred_iterator I = SU->Preds.begin(), 208 E = SU->Preds.end(); I != E; ++I) { 209 SUnit *PredSU = I->getSUnit(); 210 if (PredSU->isHeightCurrent) 211 WorkList.push_back(PredSU); 212 } 213 } while (!WorkList.empty()); 214} 215 216/// setDepthToAtLeast - Update this node's successors to reflect the 217/// fact that this node's depth just increased. 218/// 219void SUnit::setDepthToAtLeast(unsigned NewDepth) { 220 if (NewDepth <= getDepth()) 221 return; 222 setDepthDirty(); 223 Depth = NewDepth; 224 isDepthCurrent = true; 225} 226 227/// setHeightToAtLeast - Update this node's predecessors to reflect the 228/// fact that this node's height just increased. 229/// 230void SUnit::setHeightToAtLeast(unsigned NewHeight) { 231 if (NewHeight <= getHeight()) 232 return; 233 setHeightDirty(); 234 Height = NewHeight; 235 isHeightCurrent = true; 236} 237 238/// ComputeDepth - Calculate the maximal path from the node to the exit. 239/// 240void SUnit::ComputeDepth() { 241 SmallVector<SUnit*, 8> WorkList; 242 WorkList.push_back(this); 243 do { 244 SUnit *Cur = WorkList.back(); 245 246 bool Done = true; 247 unsigned MaxPredDepth = 0; 248 for (SUnit::const_pred_iterator I = Cur->Preds.begin(), 249 E = Cur->Preds.end(); I != E; ++I) { 250 SUnit *PredSU = I->getSUnit(); 251 if (PredSU->isDepthCurrent) 252 MaxPredDepth = std::max(MaxPredDepth, 253 PredSU->Depth + I->getLatency()); 254 else { 255 Done = false; 256 WorkList.push_back(PredSU); 257 } 258 } 259 260 if (Done) { 261 WorkList.pop_back(); 262 if (MaxPredDepth != Cur->Depth) { 263 Cur->setDepthDirty(); 264 Cur->Depth = MaxPredDepth; 265 } 266 Cur->isDepthCurrent = true; 267 } 268 } while (!WorkList.empty()); 269} 270 271/// ComputeHeight - Calculate the maximal path from the node to the entry. 272/// 273void SUnit::ComputeHeight() { 274 SmallVector<SUnit*, 8> WorkList; 275 WorkList.push_back(this); 276 do { 277 SUnit *Cur = WorkList.back(); 278 279 bool Done = true; 280 unsigned MaxSuccHeight = 0; 281 for (SUnit::const_succ_iterator I = Cur->Succs.begin(), 282 E = Cur->Succs.end(); I != E; ++I) { 283 SUnit *SuccSU = I->getSUnit(); 284 if (SuccSU->isHeightCurrent) 285 MaxSuccHeight = std::max(MaxSuccHeight, 286 SuccSU->Height + I->getLatency()); 287 else { 288 Done = false; 289 WorkList.push_back(SuccSU); 290 } 291 } 292 293 if (Done) { 294 WorkList.pop_back(); 295 if (MaxSuccHeight != Cur->Height) { 296 Cur->setHeightDirty(); 297 Cur->Height = MaxSuccHeight; 298 } 299 Cur->isHeightCurrent = true; 300 } 301 } while (!WorkList.empty()); 302} 303 304#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 305/// SUnit - Scheduling unit. It's an wrapper around either a single SDNode or 306/// a group of nodes flagged together. 307void SUnit::dump(const ScheduleDAG *G) const { 308 dbgs() << "SU(" << NodeNum << "): "; 309 G->dumpNode(this); 310} 311 312void SUnit::dumpAll(const ScheduleDAG *G) const { 313 dump(G); 314 315 dbgs() << " # preds left : " << NumPredsLeft << "\n"; 316 dbgs() << " # succs left : " << NumSuccsLeft << "\n"; 317 if (WeakPredsLeft) 318 dbgs() << " # weak preds left : " << WeakPredsLeft << "\n"; 319 if (WeakSuccsLeft) 320 dbgs() << " # weak succs left : " << WeakSuccsLeft << "\n"; 321 dbgs() << " # rdefs left : " << NumRegDefsLeft << "\n"; 322 dbgs() << " Latency : " << Latency << "\n"; 323 dbgs() << " Depth : " << Depth << "\n"; 324 dbgs() << " Height : " << Height << "\n"; 325 326 if (Preds.size() != 0) { 327 dbgs() << " Predecessors:\n"; 328 for (SUnit::const_succ_iterator I = Preds.begin(), E = Preds.end(); 329 I != E; ++I) { 330 dbgs() << " "; 331 switch (I->getKind()) { 332 case SDep::Data: dbgs() << "val "; break; 333 case SDep::Anti: dbgs() << "anti"; break; 334 case SDep::Output: dbgs() << "out "; break; 335 case SDep::Order: dbgs() << "ch "; break; 336 } 337 dbgs() << "SU(" << I->getSUnit()->NodeNum << ")"; 338 if (I->isArtificial()) 339 dbgs() << " *"; 340 dbgs() << ": Latency=" << I->getLatency(); 341 if (I->isAssignedRegDep()) 342 dbgs() << " Reg=" << PrintReg(I->getReg(), G->TRI); 343 dbgs() << "\n"; 344 } 345 } 346 if (Succs.size() != 0) { 347 dbgs() << " Successors:\n"; 348 for (SUnit::const_succ_iterator I = Succs.begin(), E = Succs.end(); 349 I != E; ++I) { 350 dbgs() << " "; 351 switch (I->getKind()) { 352 case SDep::Data: dbgs() << "val "; break; 353 case SDep::Anti: dbgs() << "anti"; break; 354 case SDep::Output: dbgs() << "out "; break; 355 case SDep::Order: dbgs() << "ch "; break; 356 } 357 dbgs() << "SU(" << I->getSUnit()->NodeNum << ")"; 358 if (I->isArtificial()) 359 dbgs() << " *"; 360 dbgs() << ": Latency=" << I->getLatency(); 361 dbgs() << "\n"; 362 } 363 } 364 dbgs() << "\n"; 365} 366#endif 367 368#ifndef NDEBUG 369/// VerifyScheduledDAG - Verify that all SUnits were scheduled and that 370/// their state is consistent. Return the number of scheduled nodes. 371/// 372unsigned ScheduleDAG::VerifyScheduledDAG(bool isBottomUp) { 373 bool AnyNotSched = false; 374 unsigned DeadNodes = 0; 375 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) { 376 if (!SUnits[i].isScheduled) { 377 if (SUnits[i].NumPreds == 0 && SUnits[i].NumSuccs == 0) { 378 ++DeadNodes; 379 continue; 380 } 381 if (!AnyNotSched) 382 dbgs() << "*** Scheduling failed! ***\n"; 383 SUnits[i].dump(this); 384 dbgs() << "has not been scheduled!\n"; 385 AnyNotSched = true; 386 } 387 if (SUnits[i].isScheduled && 388 (isBottomUp ? SUnits[i].getHeight() : SUnits[i].getDepth()) > 389 unsigned(INT_MAX)) { 390 if (!AnyNotSched) 391 dbgs() << "*** Scheduling failed! ***\n"; 392 SUnits[i].dump(this); 393 dbgs() << "has an unexpected " 394 << (isBottomUp ? "Height" : "Depth") << " value!\n"; 395 AnyNotSched = true; 396 } 397 if (isBottomUp) { 398 if (SUnits[i].NumSuccsLeft != 0) { 399 if (!AnyNotSched) 400 dbgs() << "*** Scheduling failed! ***\n"; 401 SUnits[i].dump(this); 402 dbgs() << "has successors left!\n"; 403 AnyNotSched = true; 404 } 405 } else { 406 if (SUnits[i].NumPredsLeft != 0) { 407 if (!AnyNotSched) 408 dbgs() << "*** Scheduling failed! ***\n"; 409 SUnits[i].dump(this); 410 dbgs() << "has predecessors left!\n"; 411 AnyNotSched = true; 412 } 413 } 414 } 415 assert(!AnyNotSched); 416 return SUnits.size() - DeadNodes; 417} 418#endif 419 420/// InitDAGTopologicalSorting - create the initial topological 421/// ordering from the DAG to be scheduled. 422/// 423/// The idea of the algorithm is taken from 424/// "Online algorithms for managing the topological order of 425/// a directed acyclic graph" by David J. Pearce and Paul H.J. Kelly 426/// This is the MNR algorithm, which was first introduced by 427/// A. Marchetti-Spaccamela, U. Nanni and H. Rohnert in 428/// "Maintaining a topological order under edge insertions". 429/// 430/// Short description of the algorithm: 431/// 432/// Topological ordering, ord, of a DAG maps each node to a topological 433/// index so that for all edges X->Y it is the case that ord(X) < ord(Y). 434/// 435/// This means that if there is a path from the node X to the node Z, 436/// then ord(X) < ord(Z). 437/// 438/// This property can be used to check for reachability of nodes: 439/// if Z is reachable from X, then an insertion of the edge Z->X would 440/// create a cycle. 441/// 442/// The algorithm first computes a topological ordering for the DAG by 443/// initializing the Index2Node and Node2Index arrays and then tries to keep 444/// the ordering up-to-date after edge insertions by reordering the DAG. 445/// 446/// On insertion of the edge X->Y, the algorithm first marks by calling DFS 447/// the nodes reachable from Y, and then shifts them using Shift to lie 448/// immediately after X in Index2Node. 449void ScheduleDAGTopologicalSort::InitDAGTopologicalSorting() { 450 unsigned DAGSize = SUnits.size(); 451 std::vector<SUnit*> WorkList; 452 WorkList.reserve(DAGSize); 453 454 Index2Node.resize(DAGSize); 455 Node2Index.resize(DAGSize); 456 457 // Initialize the data structures. 458 if (ExitSU) 459 WorkList.push_back(ExitSU); 460 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 461 SUnit *SU = &SUnits[i]; 462 int NodeNum = SU->NodeNum; 463 unsigned Degree = SU->Succs.size(); 464 // Temporarily use the Node2Index array as scratch space for degree counts. 465 Node2Index[NodeNum] = Degree; 466 467 // Is it a node without dependencies? 468 if (Degree == 0) { 469 assert(SU->Succs.empty() && "SUnit should have no successors"); 470 // Collect leaf nodes. 471 WorkList.push_back(SU); 472 } 473 } 474 475 int Id = DAGSize; 476 while (!WorkList.empty()) { 477 SUnit *SU = WorkList.back(); 478 WorkList.pop_back(); 479 if (SU->NodeNum < DAGSize) 480 Allocate(SU->NodeNum, --Id); 481 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); 482 I != E; ++I) { 483 SUnit *SU = I->getSUnit(); 484 if (SU->NodeNum < DAGSize && !--Node2Index[SU->NodeNum]) 485 // If all dependencies of the node are processed already, 486 // then the node can be computed now. 487 WorkList.push_back(SU); 488 } 489 } 490 491 Visited.resize(DAGSize); 492 493#ifndef NDEBUG 494 // Check correctness of the ordering 495 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 496 SUnit *SU = &SUnits[i]; 497 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); 498 I != E; ++I) { 499 assert(Node2Index[SU->NodeNum] > Node2Index[I->getSUnit()->NodeNum] && 500 "Wrong topological sorting"); 501 } 502 } 503#endif 504} 505 506/// AddPred - Updates the topological ordering to accommodate an edge 507/// to be added from SUnit X to SUnit Y. 508void ScheduleDAGTopologicalSort::AddPred(SUnit *Y, SUnit *X) { 509 int UpperBound, LowerBound; 510 LowerBound = Node2Index[Y->NodeNum]; 511 UpperBound = Node2Index[X->NodeNum]; 512 bool HasLoop = false; 513 // Is Ord(X) < Ord(Y) ? 514 if (LowerBound < UpperBound) { 515 // Update the topological order. 516 Visited.reset(); 517 DFS(Y, UpperBound, HasLoop); 518 assert(!HasLoop && "Inserted edge creates a loop!"); 519 // Recompute topological indexes. 520 Shift(Visited, LowerBound, UpperBound); 521 } 522} 523 524/// RemovePred - Updates the topological ordering to accommodate an 525/// an edge to be removed from the specified node N from the predecessors 526/// of the current node M. 527void ScheduleDAGTopologicalSort::RemovePred(SUnit *M, SUnit *N) { 528 // InitDAGTopologicalSorting(); 529} 530 531/// DFS - Make a DFS traversal to mark all nodes reachable from SU and mark 532/// all nodes affected by the edge insertion. These nodes will later get new 533/// topological indexes by means of the Shift method. 534void ScheduleDAGTopologicalSort::DFS(const SUnit *SU, int UpperBound, 535 bool &HasLoop) { 536 std::vector<const SUnit*> WorkList; 537 WorkList.reserve(SUnits.size()); 538 539 WorkList.push_back(SU); 540 do { 541 SU = WorkList.back(); 542 WorkList.pop_back(); 543 Visited.set(SU->NodeNum); 544 for (int I = SU->Succs.size()-1; I >= 0; --I) { 545 unsigned s = SU->Succs[I].getSUnit()->NodeNum; 546 // Edges to non-SUnits are allowed but ignored (e.g. ExitSU). 547 if (s >= Node2Index.size()) 548 continue; 549 if (Node2Index[s] == UpperBound) { 550 HasLoop = true; 551 return; 552 } 553 // Visit successors if not already and in affected region. 554 if (!Visited.test(s) && Node2Index[s] < UpperBound) { 555 WorkList.push_back(SU->Succs[I].getSUnit()); 556 } 557 } 558 } while (!WorkList.empty()); 559} 560 561/// Shift - Renumber the nodes so that the topological ordering is 562/// preserved. 563void ScheduleDAGTopologicalSort::Shift(BitVector& Visited, int LowerBound, 564 int UpperBound) { 565 std::vector<int> L; 566 int shift = 0; 567 int i; 568 569 for (i = LowerBound; i <= UpperBound; ++i) { 570 // w is node at topological index i. 571 int w = Index2Node[i]; 572 if (Visited.test(w)) { 573 // Unmark. 574 Visited.reset(w); 575 L.push_back(w); 576 shift = shift + 1; 577 } else { 578 Allocate(w, i - shift); 579 } 580 } 581 582 for (unsigned j = 0; j < L.size(); ++j) { 583 Allocate(L[j], i - shift); 584 i = i + 1; 585 } 586} 587 588 589/// WillCreateCycle - Returns true if adding an edge to TargetSU from SU will 590/// create a cycle. If so, it is not safe to call AddPred(TargetSU, SU). 591bool ScheduleDAGTopologicalSort::WillCreateCycle(SUnit *TargetSU, SUnit *SU) { 592 // Is SU reachable from TargetSU via successor edges? 593 if (IsReachable(SU, TargetSU)) 594 return true; 595 for (SUnit::pred_iterator 596 I = TargetSU->Preds.begin(), E = TargetSU->Preds.end(); I != E; ++I) 597 if (I->isAssignedRegDep() && 598 IsReachable(SU, I->getSUnit())) 599 return true; 600 return false; 601} 602 603/// IsReachable - Checks if SU is reachable from TargetSU. 604bool ScheduleDAGTopologicalSort::IsReachable(const SUnit *SU, 605 const SUnit *TargetSU) { 606 // If insertion of the edge SU->TargetSU would create a cycle 607 // then there is a path from TargetSU to SU. 608 int UpperBound, LowerBound; 609 LowerBound = Node2Index[TargetSU->NodeNum]; 610 UpperBound = Node2Index[SU->NodeNum]; 611 bool HasLoop = false; 612 // Is Ord(TargetSU) < Ord(SU) ? 613 if (LowerBound < UpperBound) { 614 Visited.reset(); 615 // There may be a path from TargetSU to SU. Check for it. 616 DFS(TargetSU, UpperBound, HasLoop); 617 } 618 return HasLoop; 619} 620 621/// Allocate - assign the topological index to the node n. 622void ScheduleDAGTopologicalSort::Allocate(int n, int index) { 623 Node2Index[n] = index; 624 Index2Node[index] = n; 625} 626 627ScheduleDAGTopologicalSort:: 628ScheduleDAGTopologicalSort(std::vector<SUnit> &sunits, SUnit *exitsu) 629 : SUnits(sunits), ExitSU(exitsu) {} 630 631ScheduleHazardRecognizer::~ScheduleHazardRecognizer() {} 632