ScheduleDAG.cpp revision 54e4c36a7349e94a84773afb56eccd4ca65b49e9
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/Target/TargetMachine.h" 18#include "llvm/Target/TargetInstrInfo.h" 19#include "llvm/Target/TargetRegisterInfo.h" 20#include "llvm/Support/Debug.h" 21#include <climits> 22using namespace llvm; 23 24ScheduleDAG::ScheduleDAG(SelectionDAG *dag, MachineBasicBlock *bb, 25 const TargetMachine &tm) 26 : DAG(dag), BB(bb), TM(tm), MRI(BB->getParent()->getRegInfo()) { 27 TII = TM.getInstrInfo(); 28 MF = BB->getParent(); 29 TRI = TM.getRegisterInfo(); 30 TLI = TM.getTargetLowering(); 31 ConstPool = MF->getConstantPool(); 32} 33 34ScheduleDAG::~ScheduleDAG() {} 35 36/// CalculateDepths - compute depths using algorithms for the longest 37/// paths in the DAG 38void ScheduleDAG::CalculateDepths() { 39 unsigned DAGSize = SUnits.size(); 40 std::vector<SUnit*> WorkList; 41 WorkList.reserve(DAGSize); 42 43 // Initialize the data structures 44 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 45 SUnit *SU = &SUnits[i]; 46 unsigned Degree = SU->Preds.size(); 47 // Temporarily use the Depth field as scratch space for the degree count. 48 SU->Depth = Degree; 49 50 // Is it a node without dependencies? 51 if (Degree == 0) { 52 assert(SU->Preds.empty() && "SUnit should have no predecessors"); 53 // Collect leaf nodes 54 WorkList.push_back(SU); 55 } 56 } 57 58 // Process nodes in the topological order 59 while (!WorkList.empty()) { 60 SUnit *SU = WorkList.back(); 61 WorkList.pop_back(); 62 unsigned SUDepth = 0; 63 64 // Use dynamic programming: 65 // When current node is being processed, all of its dependencies 66 // are already processed. 67 // So, just iterate over all predecessors and take the longest path 68 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); 69 I != E; ++I) { 70 unsigned PredDepth = I->getSUnit()->Depth; 71 if (PredDepth+1 > SUDepth) { 72 SUDepth = PredDepth + 1; 73 } 74 } 75 76 SU->Depth = SUDepth; 77 78 // Update degrees of all nodes depending on current SUnit 79 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end(); 80 I != E; ++I) { 81 SUnit *SU = I->getSUnit(); 82 if (!--SU->Depth) 83 // If all dependencies of the node are processed already, 84 // then the longest path for the node can be computed now 85 WorkList.push_back(SU); 86 } 87 } 88} 89 90/// CalculateHeights - compute heights using algorithms for the longest 91/// paths in the DAG 92void ScheduleDAG::CalculateHeights() { 93 unsigned DAGSize = SUnits.size(); 94 std::vector<SUnit*> WorkList; 95 WorkList.reserve(DAGSize); 96 97 // Initialize the data structures 98 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 99 SUnit *SU = &SUnits[i]; 100 unsigned Degree = SU->Succs.size(); 101 // Temporarily use the Height field as scratch space for the degree count. 102 SU->Height = Degree; 103 104 // Is it a node without dependencies? 105 if (Degree == 0) { 106 assert(SU->Succs.empty() && "Something wrong"); 107 assert(WorkList.empty() && "Should be empty"); 108 // Collect leaf nodes 109 WorkList.push_back(SU); 110 } 111 } 112 113 // Process nodes in the topological order 114 while (!WorkList.empty()) { 115 SUnit *SU = WorkList.back(); 116 WorkList.pop_back(); 117 unsigned SUHeight = 0; 118 119 // Use dynamic programming: 120 // When current node is being processed, all of its dependencies 121 // are already processed. 122 // So, just iterate over all successors and take the longest path 123 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end(); 124 I != E; ++I) { 125 unsigned SuccHeight = I->getSUnit()->Height; 126 if (SuccHeight+1 > SUHeight) { 127 SUHeight = SuccHeight + 1; 128 } 129 } 130 131 SU->Height = SUHeight; 132 133 // Update degrees of all nodes depending on current SUnit 134 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); 135 I != E; ++I) { 136 SUnit *SU = I->getSUnit(); 137 if (!--SU->Height) 138 // If all dependencies of the node are processed already, 139 // then the longest path for the node can be computed now 140 WorkList.push_back(SU); 141 } 142 } 143} 144 145/// dump - dump the schedule. 146void ScheduleDAG::dumpSchedule() const { 147 for (unsigned i = 0, e = Sequence.size(); i != e; i++) { 148 if (SUnit *SU = Sequence[i]) 149 SU->dump(this); 150 else 151 cerr << "**** NOOP ****\n"; 152 } 153} 154 155 156/// Run - perform scheduling. 157/// 158void ScheduleDAG::Run() { 159 Schedule(); 160 161 DOUT << "*** Final schedule ***\n"; 162 DEBUG(dumpSchedule()); 163 DOUT << "\n"; 164} 165 166/// SUnit - Scheduling unit. It's an wrapper around either a single SDNode or 167/// a group of nodes flagged together. 168void SUnit::dump(const ScheduleDAG *G) const { 169 cerr << "SU(" << NodeNum << "): "; 170 G->dumpNode(this); 171} 172 173void SUnit::dumpAll(const ScheduleDAG *G) const { 174 dump(G); 175 176 cerr << " # preds left : " << NumPredsLeft << "\n"; 177 cerr << " # succs left : " << NumSuccsLeft << "\n"; 178 cerr << " Latency : " << Latency << "\n"; 179 cerr << " Depth : " << Depth << "\n"; 180 cerr << " Height : " << Height << "\n"; 181 182 if (Preds.size() != 0) { 183 cerr << " Predecessors:\n"; 184 for (SUnit::const_succ_iterator I = Preds.begin(), E = Preds.end(); 185 I != E; ++I) { 186 cerr << " "; 187 switch (I->getKind()) { 188 case SDep::Data: cerr << "val "; break; 189 case SDep::Anti: cerr << "anti"; break; 190 case SDep::Output: cerr << "out "; break; 191 case SDep::Order: cerr << "ch "; break; 192 } 193 cerr << "#"; 194 cerr << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")"; 195 if (I->isArtificial()) 196 cerr << " *"; 197 cerr << "\n"; 198 } 199 } 200 if (Succs.size() != 0) { 201 cerr << " Successors:\n"; 202 for (SUnit::const_succ_iterator I = Succs.begin(), E = Succs.end(); 203 I != E; ++I) { 204 cerr << " "; 205 switch (I->getKind()) { 206 case SDep::Data: cerr << "val "; break; 207 case SDep::Anti: cerr << "anti"; break; 208 case SDep::Output: cerr << "out "; break; 209 case SDep::Order: cerr << "ch "; break; 210 } 211 cerr << "#"; 212 cerr << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")"; 213 if (I->isArtificial()) 214 cerr << " *"; 215 cerr << "\n"; 216 } 217 } 218 cerr << "\n"; 219} 220 221#ifndef NDEBUG 222/// VerifySchedule - Verify that all SUnits were scheduled and that 223/// their state is consistent. 224/// 225void ScheduleDAG::VerifySchedule(bool isBottomUp) { 226 bool AnyNotSched = false; 227 unsigned DeadNodes = 0; 228 unsigned Noops = 0; 229 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) { 230 if (!SUnits[i].isScheduled) { 231 if (SUnits[i].NumPreds == 0 && SUnits[i].NumSuccs == 0) { 232 ++DeadNodes; 233 continue; 234 } 235 if (!AnyNotSched) 236 cerr << "*** Scheduling failed! ***\n"; 237 SUnits[i].dump(this); 238 cerr << "has not been scheduled!\n"; 239 AnyNotSched = true; 240 } 241 if (SUnits[i].isScheduled && SUnits[i].Cycle > (unsigned)INT_MAX) { 242 if (!AnyNotSched) 243 cerr << "*** Scheduling failed! ***\n"; 244 SUnits[i].dump(this); 245 cerr << "has an unexpected Cycle value!\n"; 246 AnyNotSched = true; 247 } 248 if (isBottomUp) { 249 if (SUnits[i].NumSuccsLeft != 0) { 250 if (!AnyNotSched) 251 cerr << "*** Scheduling failed! ***\n"; 252 SUnits[i].dump(this); 253 cerr << "has successors left!\n"; 254 AnyNotSched = true; 255 } 256 } else { 257 if (SUnits[i].NumPredsLeft != 0) { 258 if (!AnyNotSched) 259 cerr << "*** Scheduling failed! ***\n"; 260 SUnits[i].dump(this); 261 cerr << "has predecessors left!\n"; 262 AnyNotSched = true; 263 } 264 } 265 } 266 for (unsigned i = 0, e = Sequence.size(); i != e; ++i) 267 if (!Sequence[i]) 268 ++Noops; 269 assert(!AnyNotSched); 270 assert(Sequence.size() + DeadNodes - Noops == SUnits.size() && 271 "The number of nodes scheduled doesn't match the expected number!"); 272} 273#endif 274 275/// InitDAGTopologicalSorting - create the initial topological 276/// ordering from the DAG to be scheduled. 277/// 278/// The idea of the algorithm is taken from 279/// "Online algorithms for managing the topological order of 280/// a directed acyclic graph" by David J. Pearce and Paul H.J. Kelly 281/// This is the MNR algorithm, which was first introduced by 282/// A. Marchetti-Spaccamela, U. Nanni and H. Rohnert in 283/// "Maintaining a topological order under edge insertions". 284/// 285/// Short description of the algorithm: 286/// 287/// Topological ordering, ord, of a DAG maps each node to a topological 288/// index so that for all edges X->Y it is the case that ord(X) < ord(Y). 289/// 290/// This means that if there is a path from the node X to the node Z, 291/// then ord(X) < ord(Z). 292/// 293/// This property can be used to check for reachability of nodes: 294/// if Z is reachable from X, then an insertion of the edge Z->X would 295/// create a cycle. 296/// 297/// The algorithm first computes a topological ordering for the DAG by 298/// initializing the Index2Node and Node2Index arrays and then tries to keep 299/// the ordering up-to-date after edge insertions by reordering the DAG. 300/// 301/// On insertion of the edge X->Y, the algorithm first marks by calling DFS 302/// the nodes reachable from Y, and then shifts them using Shift to lie 303/// immediately after X in Index2Node. 304void ScheduleDAGTopologicalSort::InitDAGTopologicalSorting() { 305 unsigned DAGSize = SUnits.size(); 306 std::vector<SUnit*> WorkList; 307 WorkList.reserve(DAGSize); 308 309 Index2Node.resize(DAGSize); 310 Node2Index.resize(DAGSize); 311 312 // Initialize the data structures. 313 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 314 SUnit *SU = &SUnits[i]; 315 int NodeNum = SU->NodeNum; 316 unsigned Degree = SU->Succs.size(); 317 // Temporarily use the Node2Index array as scratch space for degree counts. 318 Node2Index[NodeNum] = Degree; 319 320 // Is it a node without dependencies? 321 if (Degree == 0) { 322 assert(SU->Succs.empty() && "SUnit should have no successors"); 323 // Collect leaf nodes. 324 WorkList.push_back(SU); 325 } 326 } 327 328 int Id = DAGSize; 329 while (!WorkList.empty()) { 330 SUnit *SU = WorkList.back(); 331 WorkList.pop_back(); 332 Allocate(SU->NodeNum, --Id); 333 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); 334 I != E; ++I) { 335 SUnit *SU = I->getSUnit(); 336 if (!--Node2Index[SU->NodeNum]) 337 // If all dependencies of the node are processed already, 338 // then the node can be computed now. 339 WorkList.push_back(SU); 340 } 341 } 342 343 Visited.resize(DAGSize); 344 345#ifndef NDEBUG 346 // Check correctness of the ordering 347 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 348 SUnit *SU = &SUnits[i]; 349 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); 350 I != E; ++I) { 351 assert(Node2Index[SU->NodeNum] > Node2Index[I->getSUnit()->NodeNum] && 352 "Wrong topological sorting"); 353 } 354 } 355#endif 356} 357 358/// AddPred - Updates the topological ordering to accomodate an edge 359/// to be added from SUnit X to SUnit Y. 360void ScheduleDAGTopologicalSort::AddPred(SUnit *Y, SUnit *X) { 361 int UpperBound, LowerBound; 362 LowerBound = Node2Index[Y->NodeNum]; 363 UpperBound = Node2Index[X->NodeNum]; 364 bool HasLoop = false; 365 // Is Ord(X) < Ord(Y) ? 366 if (LowerBound < UpperBound) { 367 // Update the topological order. 368 Visited.reset(); 369 DFS(Y, UpperBound, HasLoop); 370 assert(!HasLoop && "Inserted edge creates a loop!"); 371 // Recompute topological indexes. 372 Shift(Visited, LowerBound, UpperBound); 373 } 374} 375 376/// RemovePred - Updates the topological ordering to accomodate an 377/// an edge to be removed from the specified node N from the predecessors 378/// of the current node M. 379void ScheduleDAGTopologicalSort::RemovePred(SUnit *M, SUnit *N) { 380 // InitDAGTopologicalSorting(); 381} 382 383/// DFS - Make a DFS traversal to mark all nodes reachable from SU and mark 384/// all nodes affected by the edge insertion. These nodes will later get new 385/// topological indexes by means of the Shift method. 386void ScheduleDAGTopologicalSort::DFS(const SUnit *SU, int UpperBound, 387 bool& HasLoop) { 388 std::vector<const SUnit*> WorkList; 389 WorkList.reserve(SUnits.size()); 390 391 WorkList.push_back(SU); 392 while (!WorkList.empty()) { 393 SU = WorkList.back(); 394 WorkList.pop_back(); 395 Visited.set(SU->NodeNum); 396 for (int I = SU->Succs.size()-1; I >= 0; --I) { 397 int s = SU->Succs[I].getSUnit()->NodeNum; 398 if (Node2Index[s] == UpperBound) { 399 HasLoop = true; 400 return; 401 } 402 // Visit successors if not already and in affected region. 403 if (!Visited.test(s) && Node2Index[s] < UpperBound) { 404 WorkList.push_back(SU->Succs[I].getSUnit()); 405 } 406 } 407 } 408} 409 410/// Shift - Renumber the nodes so that the topological ordering is 411/// preserved. 412void ScheduleDAGTopologicalSort::Shift(BitVector& Visited, int LowerBound, 413 int UpperBound) { 414 std::vector<int> L; 415 int shift = 0; 416 int i; 417 418 for (i = LowerBound; i <= UpperBound; ++i) { 419 // w is node at topological index i. 420 int w = Index2Node[i]; 421 if (Visited.test(w)) { 422 // Unmark. 423 Visited.reset(w); 424 L.push_back(w); 425 shift = shift + 1; 426 } else { 427 Allocate(w, i - shift); 428 } 429 } 430 431 for (unsigned j = 0; j < L.size(); ++j) { 432 Allocate(L[j], i - shift); 433 i = i + 1; 434 } 435} 436 437 438/// WillCreateCycle - Returns true if adding an edge from SU to TargetSU will 439/// create a cycle. 440bool ScheduleDAGTopologicalSort::WillCreateCycle(SUnit *SU, SUnit *TargetSU) { 441 if (IsReachable(TargetSU, SU)) 442 return true; 443 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); 444 I != E; ++I) 445 if (I->isAssignedRegDep() && 446 IsReachable(TargetSU, I->getSUnit())) 447 return true; 448 return false; 449} 450 451/// IsReachable - Checks if SU is reachable from TargetSU. 452bool ScheduleDAGTopologicalSort::IsReachable(const SUnit *SU, 453 const SUnit *TargetSU) { 454 // If insertion of the edge SU->TargetSU would create a cycle 455 // then there is a path from TargetSU to SU. 456 int UpperBound, LowerBound; 457 LowerBound = Node2Index[TargetSU->NodeNum]; 458 UpperBound = Node2Index[SU->NodeNum]; 459 bool HasLoop = false; 460 // Is Ord(TargetSU) < Ord(SU) ? 461 if (LowerBound < UpperBound) { 462 Visited.reset(); 463 // There may be a path from TargetSU to SU. Check for it. 464 DFS(TargetSU, UpperBound, HasLoop); 465 } 466 return HasLoop; 467} 468 469/// Allocate - assign the topological index to the node n. 470void ScheduleDAGTopologicalSort::Allocate(int n, int index) { 471 Node2Index[n] = index; 472 Index2Node[index] = n; 473} 474 475ScheduleDAGTopologicalSort::ScheduleDAGTopologicalSort( 476 std::vector<SUnit> &sunits) 477 : SUnits(sunits) {} 478