1 //===- MachinePipeliner.cpp - Machine Software Pipeliner Pass -------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // An implementation of the Swing Modulo Scheduling (SMS) software pipeliner. 10 // 11 // This SMS implementation is a target-independent back-end pass. When enabled, 12 // the pass runs just prior to the register allocation pass, while the machine 13 // IR is in SSA form. If software pipelining is successful, then the original 14 // loop is replaced by the optimized loop. The optimized loop contains one or 15 // more prolog blocks, the pipelined kernel, and one or more epilog blocks. If 16 // the instructions cannot be scheduled in a given MII, we increase the MII by 17 // one and try again. 18 // 19 // The SMS implementation is an extension of the ScheduleDAGInstrs class. We 20 // represent loop carried dependences in the DAG as order edges to the Phi 21 // nodes. We also perform several passes over the DAG to eliminate unnecessary 22 // edges that inhibit the ability to pipeline. The implementation uses the 23 // DFAPacketizer class to compute the minimum initiation interval and the check 24 // where an instruction may be inserted in the pipelined schedule. 25 // 26 // In order for the SMS pass to work, several target specific hooks need to be 27 // implemented to get information about the loop structure and to rewrite 28 // instructions. 29 // 30 //===----------------------------------------------------------------------===// 31 32 #include "llvm/ADT/ArrayRef.h" 33 #include "llvm/ADT/BitVector.h" 34 #include "llvm/ADT/DenseMap.h" 35 #include "llvm/ADT/MapVector.h" 36 #include "llvm/ADT/PriorityQueue.h" 37 #include "llvm/ADT/SetVector.h" 38 #include "llvm/ADT/SmallPtrSet.h" 39 #include "llvm/ADT/SmallSet.h" 40 #include "llvm/ADT/SmallVector.h" 41 #include "llvm/ADT/Statistic.h" 42 #include "llvm/ADT/iterator_range.h" 43 #include "llvm/Analysis/AliasAnalysis.h" 44 #include "llvm/Analysis/MemoryLocation.h" 45 #include "llvm/Analysis/ValueTracking.h" 46 #include "llvm/CodeGen/DFAPacketizer.h" 47 #include "llvm/CodeGen/LiveIntervals.h" 48 #include "llvm/CodeGen/MachineBasicBlock.h" 49 #include "llvm/CodeGen/MachineDominators.h" 50 #include "llvm/CodeGen/MachineFunction.h" 51 #include "llvm/CodeGen/MachineFunctionPass.h" 52 #include "llvm/CodeGen/MachineInstr.h" 53 #include "llvm/CodeGen/MachineInstrBuilder.h" 54 #include "llvm/CodeGen/MachineLoopInfo.h" 55 #include "llvm/CodeGen/MachineMemOperand.h" 56 #include "llvm/CodeGen/MachineOperand.h" 57 #include "llvm/CodeGen/MachinePipeliner.h" 58 #include "llvm/CodeGen/MachineRegisterInfo.h" 59 #include "llvm/CodeGen/ModuloSchedule.h" 60 #include "llvm/CodeGen/RegisterPressure.h" 61 #include "llvm/CodeGen/ScheduleDAG.h" 62 #include "llvm/CodeGen/ScheduleDAGMutation.h" 63 #include "llvm/CodeGen/TargetOpcodes.h" 64 #include "llvm/CodeGen/TargetRegisterInfo.h" 65 #include "llvm/CodeGen/TargetSubtargetInfo.h" 66 #include "llvm/Config/llvm-config.h" 67 #include "llvm/IR/Attributes.h" 68 #include "llvm/IR/DebugLoc.h" 69 #include "llvm/IR/Function.h" 70 #include "llvm/MC/LaneBitmask.h" 71 #include "llvm/MC/MCInstrDesc.h" 72 #include "llvm/MC/MCInstrItineraries.h" 73 #include "llvm/MC/MCRegisterInfo.h" 74 #include "llvm/Pass.h" 75 #include "llvm/Support/CommandLine.h" 76 #include "llvm/Support/Compiler.h" 77 #include "llvm/Support/Debug.h" 78 #include "llvm/Support/MathExtras.h" 79 #include "llvm/Support/raw_ostream.h" 80 #include <algorithm> 81 #include <cassert> 82 #include <climits> 83 #include <cstdint> 84 #include <deque> 85 #include <functional> 86 #include <iterator> 87 #include <map> 88 #include <memory> 89 #include <tuple> 90 #include <utility> 91 #include <vector> 92 93 using namespace llvm; 94 95 #define DEBUG_TYPE "pipeliner" 96 97 STATISTIC(NumTrytoPipeline, "Number of loops that we attempt to pipeline"); 98 STATISTIC(NumPipelined, "Number of loops software pipelined"); 99 STATISTIC(NumNodeOrderIssues, "Number of node order issues found"); 100 STATISTIC(NumFailBranch, "Pipeliner abort due to unknown branch"); 101 STATISTIC(NumFailLoop, "Pipeliner abort due to unsupported loop"); 102 STATISTIC(NumFailPreheader, "Pipeliner abort due to missing preheader"); 103 STATISTIC(NumFailLargeMaxMII, "Pipeliner abort due to MaxMII too large"); 104 STATISTIC(NumFailZeroMII, "Pipeliner abort due to zero MII"); 105 STATISTIC(NumFailNoSchedule, "Pipeliner abort due to no schedule found"); 106 STATISTIC(NumFailZeroStage, "Pipeliner abort due to zero stage"); 107 STATISTIC(NumFailLargeMaxStage, "Pipeliner abort due to too many stages"); 108 109 /// A command line option to turn software pipelining on or off. 110 static cl::opt<bool> EnableSWP("enable-pipeliner", cl::Hidden, cl::init(true), 111 cl::ZeroOrMore, 112 cl::desc("Enable Software Pipelining")); 113 114 /// A command line option to enable SWP at -Os. 115 static cl::opt<bool> EnableSWPOptSize("enable-pipeliner-opt-size", 116 cl::desc("Enable SWP at Os."), cl::Hidden, 117 cl::init(false)); 118 119 /// A command line argument to limit minimum initial interval for pipelining. 120 static cl::opt<int> SwpMaxMii("pipeliner-max-mii", 121 cl::desc("Size limit for the MII."), 122 cl::Hidden, cl::init(27)); 123 124 /// A command line argument to limit the number of stages in the pipeline. 125 static cl::opt<int> 126 SwpMaxStages("pipeliner-max-stages", 127 cl::desc("Maximum stages allowed in the generated scheduled."), 128 cl::Hidden, cl::init(3)); 129 130 /// A command line option to disable the pruning of chain dependences due to 131 /// an unrelated Phi. 132 static cl::opt<bool> 133 SwpPruneDeps("pipeliner-prune-deps", 134 cl::desc("Prune dependences between unrelated Phi nodes."), 135 cl::Hidden, cl::init(true)); 136 137 /// A command line option to disable the pruning of loop carried order 138 /// dependences. 139 static cl::opt<bool> 140 SwpPruneLoopCarried("pipeliner-prune-loop-carried", 141 cl::desc("Prune loop carried order dependences."), 142 cl::Hidden, cl::init(true)); 143 144 #ifndef NDEBUG 145 static cl::opt<int> SwpLoopLimit("pipeliner-max", cl::Hidden, cl::init(-1)); 146 #endif 147 148 static cl::opt<bool> SwpIgnoreRecMII("pipeliner-ignore-recmii", 149 cl::ReallyHidden, cl::init(false), 150 cl::ZeroOrMore, cl::desc("Ignore RecMII")); 151 152 static cl::opt<bool> SwpShowResMask("pipeliner-show-mask", cl::Hidden, 153 cl::init(false)); 154 static cl::opt<bool> SwpDebugResource("pipeliner-dbg-res", cl::Hidden, 155 cl::init(false)); 156 157 static cl::opt<bool> EmitTestAnnotations( 158 "pipeliner-annotate-for-testing", cl::Hidden, cl::init(false), 159 cl::desc("Instead of emitting the pipelined code, annotate instructions " 160 "with the generated schedule for feeding into the " 161 "-modulo-schedule-test pass")); 162 163 static cl::opt<bool> ExperimentalCodeGen( 164 "pipeliner-experimental-cg", cl::Hidden, cl::init(false), 165 cl::desc( 166 "Use the experimental peeling code generator for software pipelining")); 167 168 namespace llvm { 169 170 // A command line option to enable the CopyToPhi DAG mutation. 171 cl::opt<bool> 172 SwpEnableCopyToPhi("pipeliner-enable-copytophi", cl::ReallyHidden, 173 cl::init(true), cl::ZeroOrMore, 174 cl::desc("Enable CopyToPhi DAG Mutation")); 175 176 } // end namespace llvm 177 178 unsigned SwingSchedulerDAG::Circuits::MaxPaths = 5; 179 char MachinePipeliner::ID = 0; 180 #ifndef NDEBUG 181 int MachinePipeliner::NumTries = 0; 182 #endif 183 char &llvm::MachinePipelinerID = MachinePipeliner::ID; 184 185 INITIALIZE_PASS_BEGIN(MachinePipeliner, DEBUG_TYPE, 186 "Modulo Software Pipelining", false, false) 187 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 188 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) 189 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) 190 INITIALIZE_PASS_DEPENDENCY(LiveIntervals) 191 INITIALIZE_PASS_END(MachinePipeliner, DEBUG_TYPE, 192 "Modulo Software Pipelining", false, false) 193 194 /// The "main" function for implementing Swing Modulo Scheduling. 195 bool MachinePipeliner::runOnMachineFunction(MachineFunction &mf) { 196 if (skipFunction(mf.getFunction())) 197 return false; 198 199 if (!EnableSWP) 200 return false; 201 202 if (mf.getFunction().getAttributes().hasAttribute( 203 AttributeList::FunctionIndex, Attribute::OptimizeForSize) && 204 !EnableSWPOptSize.getPosition()) 205 return false; 206 207 if (!mf.getSubtarget().enableMachinePipeliner()) 208 return false; 209 210 // Cannot pipeline loops without instruction itineraries if we are using 211 // DFA for the pipeliner. 212 if (mf.getSubtarget().useDFAforSMS() && 213 (!mf.getSubtarget().getInstrItineraryData() || 214 mf.getSubtarget().getInstrItineraryData()->isEmpty())) 215 return false; 216 217 MF = &mf; 218 MLI = &getAnalysis<MachineLoopInfo>(); 219 MDT = &getAnalysis<MachineDominatorTree>(); 220 TII = MF->getSubtarget().getInstrInfo(); 221 RegClassInfo.runOnMachineFunction(*MF); 222 223 for (auto &L : *MLI) 224 scheduleLoop(*L); 225 226 return false; 227 } 228 229 /// Attempt to perform the SMS algorithm on the specified loop. This function is 230 /// the main entry point for the algorithm. The function identifies candidate 231 /// loops, calculates the minimum initiation interval, and attempts to schedule 232 /// the loop. 233 bool MachinePipeliner::scheduleLoop(MachineLoop &L) { 234 bool Changed = false; 235 for (auto &InnerLoop : L) 236 Changed |= scheduleLoop(*InnerLoop); 237 238 #ifndef NDEBUG 239 // Stop trying after reaching the limit (if any). 240 int Limit = SwpLoopLimit; 241 if (Limit >= 0) { 242 if (NumTries >= SwpLoopLimit) 243 return Changed; 244 NumTries++; 245 } 246 #endif 247 248 setPragmaPipelineOptions(L); 249 if (!canPipelineLoop(L)) { 250 LLVM_DEBUG(dbgs() << "\n!!! Can not pipeline loop.\n"); 251 return Changed; 252 } 253 254 ++NumTrytoPipeline; 255 256 Changed = swingModuloScheduler(L); 257 258 return Changed; 259 } 260 261 void MachinePipeliner::setPragmaPipelineOptions(MachineLoop &L) { 262 MachineBasicBlock *LBLK = L.getTopBlock(); 263 264 if (LBLK == nullptr) 265 return; 266 267 const BasicBlock *BBLK = LBLK->getBasicBlock(); 268 if (BBLK == nullptr) 269 return; 270 271 const Instruction *TI = BBLK->getTerminator(); 272 if (TI == nullptr) 273 return; 274 275 MDNode *LoopID = TI->getMetadata(LLVMContext::MD_loop); 276 if (LoopID == nullptr) 277 return; 278 279 assert(LoopID->getNumOperands() > 0 && "requires atleast one operand"); 280 assert(LoopID->getOperand(0) == LoopID && "invalid loop"); 281 282 for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) { 283 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i)); 284 285 if (MD == nullptr) 286 continue; 287 288 MDString *S = dyn_cast<MDString>(MD->getOperand(0)); 289 290 if (S == nullptr) 291 continue; 292 293 if (S->getString() == "llvm.loop.pipeline.initiationinterval") { 294 assert(MD->getNumOperands() == 2 && 295 "Pipeline initiation interval hint metadata should have two operands."); 296 II_setByPragma = 297 mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue(); 298 assert(II_setByPragma >= 1 && "Pipeline initiation interval must be positive."); 299 } else if (S->getString() == "llvm.loop.pipeline.disable") { 300 disabledByPragma = true; 301 } 302 } 303 } 304 305 /// Return true if the loop can be software pipelined. The algorithm is 306 /// restricted to loops with a single basic block. Make sure that the 307 /// branch in the loop can be analyzed. 308 bool MachinePipeliner::canPipelineLoop(MachineLoop &L) { 309 if (L.getNumBlocks() != 1) 310 return false; 311 312 if (disabledByPragma) 313 return false; 314 315 // Check if the branch can't be understood because we can't do pipelining 316 // if that's the case. 317 LI.TBB = nullptr; 318 LI.FBB = nullptr; 319 LI.BrCond.clear(); 320 if (TII->analyzeBranch(*L.getHeader(), LI.TBB, LI.FBB, LI.BrCond)) { 321 LLVM_DEBUG( 322 dbgs() << "Unable to analyzeBranch, can NOT pipeline current Loop\n"); 323 NumFailBranch++; 324 return false; 325 } 326 327 LI.LoopInductionVar = nullptr; 328 LI.LoopCompare = nullptr; 329 if (!TII->analyzeLoopForPipelining(L.getTopBlock())) { 330 LLVM_DEBUG( 331 dbgs() << "Unable to analyzeLoop, can NOT pipeline current Loop\n"); 332 NumFailLoop++; 333 return false; 334 } 335 336 if (!L.getLoopPreheader()) { 337 LLVM_DEBUG( 338 dbgs() << "Preheader not found, can NOT pipeline current Loop\n"); 339 NumFailPreheader++; 340 return false; 341 } 342 343 // Remove any subregisters from inputs to phi nodes. 344 preprocessPhiNodes(*L.getHeader()); 345 return true; 346 } 347 348 void MachinePipeliner::preprocessPhiNodes(MachineBasicBlock &B) { 349 MachineRegisterInfo &MRI = MF->getRegInfo(); 350 SlotIndexes &Slots = *getAnalysis<LiveIntervals>().getSlotIndexes(); 351 352 for (MachineInstr &PI : make_range(B.begin(), B.getFirstNonPHI())) { 353 MachineOperand &DefOp = PI.getOperand(0); 354 assert(DefOp.getSubReg() == 0); 355 auto *RC = MRI.getRegClass(DefOp.getReg()); 356 357 for (unsigned i = 1, n = PI.getNumOperands(); i != n; i += 2) { 358 MachineOperand &RegOp = PI.getOperand(i); 359 if (RegOp.getSubReg() == 0) 360 continue; 361 362 // If the operand uses a subregister, replace it with a new register 363 // without subregisters, and generate a copy to the new register. 364 Register NewReg = MRI.createVirtualRegister(RC); 365 MachineBasicBlock &PredB = *PI.getOperand(i+1).getMBB(); 366 MachineBasicBlock::iterator At = PredB.getFirstTerminator(); 367 const DebugLoc &DL = PredB.findDebugLoc(At); 368 auto Copy = BuildMI(PredB, At, DL, TII->get(TargetOpcode::COPY), NewReg) 369 .addReg(RegOp.getReg(), getRegState(RegOp), 370 RegOp.getSubReg()); 371 Slots.insertMachineInstrInMaps(*Copy); 372 RegOp.setReg(NewReg); 373 RegOp.setSubReg(0); 374 } 375 } 376 } 377 378 /// The SMS algorithm consists of the following main steps: 379 /// 1. Computation and analysis of the dependence graph. 380 /// 2. Ordering of the nodes (instructions). 381 /// 3. Attempt to Schedule the loop. 382 bool MachinePipeliner::swingModuloScheduler(MachineLoop &L) { 383 assert(L.getBlocks().size() == 1 && "SMS works on single blocks only."); 384 385 SwingSchedulerDAG SMS(*this, L, getAnalysis<LiveIntervals>(), RegClassInfo, 386 II_setByPragma); 387 388 MachineBasicBlock *MBB = L.getHeader(); 389 // The kernel should not include any terminator instructions. These 390 // will be added back later. 391 SMS.startBlock(MBB); 392 393 // Compute the number of 'real' instructions in the basic block by 394 // ignoring terminators. 395 unsigned size = MBB->size(); 396 for (MachineBasicBlock::iterator I = MBB->getFirstTerminator(), 397 E = MBB->instr_end(); 398 I != E; ++I, --size) 399 ; 400 401 SMS.enterRegion(MBB, MBB->begin(), MBB->getFirstTerminator(), size); 402 SMS.schedule(); 403 SMS.exitRegion(); 404 405 SMS.finishBlock(); 406 return SMS.hasNewSchedule(); 407 } 408 409 void SwingSchedulerDAG::setMII(unsigned ResMII, unsigned RecMII) { 410 if (II_setByPragma > 0) 411 MII = II_setByPragma; 412 else 413 MII = std::max(ResMII, RecMII); 414 } 415 416 void SwingSchedulerDAG::setMAX_II() { 417 if (II_setByPragma > 0) 418 MAX_II = II_setByPragma; 419 else 420 MAX_II = MII + 10; 421 } 422 423 /// We override the schedule function in ScheduleDAGInstrs to implement the 424 /// scheduling part of the Swing Modulo Scheduling algorithm. 425 void SwingSchedulerDAG::schedule() { 426 AliasAnalysis *AA = &Pass.getAnalysis<AAResultsWrapperPass>().getAAResults(); 427 buildSchedGraph(AA); 428 addLoopCarriedDependences(AA); 429 updatePhiDependences(); 430 Topo.InitDAGTopologicalSorting(); 431 changeDependences(); 432 postprocessDAG(); 433 LLVM_DEBUG(dump()); 434 435 NodeSetType NodeSets; 436 findCircuits(NodeSets); 437 NodeSetType Circuits = NodeSets; 438 439 // Calculate the MII. 440 unsigned ResMII = calculateResMII(); 441 unsigned RecMII = calculateRecMII(NodeSets); 442 443 fuseRecs(NodeSets); 444 445 // This flag is used for testing and can cause correctness problems. 446 if (SwpIgnoreRecMII) 447 RecMII = 0; 448 449 setMII(ResMII, RecMII); 450 setMAX_II(); 451 452 LLVM_DEBUG(dbgs() << "MII = " << MII << " MAX_II = " << MAX_II 453 << " (rec=" << RecMII << ", res=" << ResMII << ")\n"); 454 455 // Can't schedule a loop without a valid MII. 456 if (MII == 0) { 457 LLVM_DEBUG( 458 dbgs() 459 << "0 is not a valid Minimal Initiation Interval, can NOT schedule\n"); 460 NumFailZeroMII++; 461 return; 462 } 463 464 // Don't pipeline large loops. 465 if (SwpMaxMii != -1 && (int)MII > SwpMaxMii) { 466 LLVM_DEBUG(dbgs() << "MII > " << SwpMaxMii 467 << ", we don't pipleline large loops\n"); 468 NumFailLargeMaxMII++; 469 return; 470 } 471 472 computeNodeFunctions(NodeSets); 473 474 registerPressureFilter(NodeSets); 475 476 colocateNodeSets(NodeSets); 477 478 checkNodeSets(NodeSets); 479 480 LLVM_DEBUG({ 481 for (auto &I : NodeSets) { 482 dbgs() << " Rec NodeSet "; 483 I.dump(); 484 } 485 }); 486 487 llvm::stable_sort(NodeSets, std::greater<NodeSet>()); 488 489 groupRemainingNodes(NodeSets); 490 491 removeDuplicateNodes(NodeSets); 492 493 LLVM_DEBUG({ 494 for (auto &I : NodeSets) { 495 dbgs() << " NodeSet "; 496 I.dump(); 497 } 498 }); 499 500 computeNodeOrder(NodeSets); 501 502 // check for node order issues 503 checkValidNodeOrder(Circuits); 504 505 SMSchedule Schedule(Pass.MF); 506 Scheduled = schedulePipeline(Schedule); 507 508 if (!Scheduled){ 509 LLVM_DEBUG(dbgs() << "No schedule found, return\n"); 510 NumFailNoSchedule++; 511 return; 512 } 513 514 unsigned numStages = Schedule.getMaxStageCount(); 515 // No need to generate pipeline if there are no overlapped iterations. 516 if (numStages == 0) { 517 LLVM_DEBUG( 518 dbgs() << "No overlapped iterations, no need to generate pipeline\n"); 519 NumFailZeroStage++; 520 return; 521 } 522 // Check that the maximum stage count is less than user-defined limit. 523 if (SwpMaxStages > -1 && (int)numStages > SwpMaxStages) { 524 LLVM_DEBUG(dbgs() << "numStages:" << numStages << ">" << SwpMaxStages 525 << " : too many stages, abort\n"); 526 NumFailLargeMaxStage++; 527 return; 528 } 529 530 // Generate the schedule as a ModuloSchedule. 531 DenseMap<MachineInstr *, int> Cycles, Stages; 532 std::vector<MachineInstr *> OrderedInsts; 533 for (int Cycle = Schedule.getFirstCycle(); Cycle <= Schedule.getFinalCycle(); 534 ++Cycle) { 535 for (SUnit *SU : Schedule.getInstructions(Cycle)) { 536 OrderedInsts.push_back(SU->getInstr()); 537 Cycles[SU->getInstr()] = Cycle; 538 Stages[SU->getInstr()] = Schedule.stageScheduled(SU); 539 } 540 } 541 DenseMap<MachineInstr *, std::pair<unsigned, int64_t>> NewInstrChanges; 542 for (auto &KV : NewMIs) { 543 Cycles[KV.first] = Cycles[KV.second]; 544 Stages[KV.first] = Stages[KV.second]; 545 NewInstrChanges[KV.first] = InstrChanges[getSUnit(KV.first)]; 546 } 547 548 ModuloSchedule MS(MF, &Loop, std::move(OrderedInsts), std::move(Cycles), 549 std::move(Stages)); 550 if (EmitTestAnnotations) { 551 assert(NewInstrChanges.empty() && 552 "Cannot serialize a schedule with InstrChanges!"); 553 ModuloScheduleTestAnnotater MSTI(MF, MS); 554 MSTI.annotate(); 555 return; 556 } 557 // The experimental code generator can't work if there are InstChanges. 558 if (ExperimentalCodeGen && NewInstrChanges.empty()) { 559 PeelingModuloScheduleExpander MSE(MF, MS, &LIS); 560 MSE.expand(); 561 } else { 562 ModuloScheduleExpander MSE(MF, MS, LIS, std::move(NewInstrChanges)); 563 MSE.expand(); 564 MSE.cleanup(); 565 } 566 ++NumPipelined; 567 } 568 569 /// Clean up after the software pipeliner runs. 570 void SwingSchedulerDAG::finishBlock() { 571 for (auto &KV : NewMIs) 572 MF.DeleteMachineInstr(KV.second); 573 NewMIs.clear(); 574 575 // Call the superclass. 576 ScheduleDAGInstrs::finishBlock(); 577 } 578 579 /// Return the register values for the operands of a Phi instruction. 580 /// This function assume the instruction is a Phi. 581 static void getPhiRegs(MachineInstr &Phi, MachineBasicBlock *Loop, 582 unsigned &InitVal, unsigned &LoopVal) { 583 assert(Phi.isPHI() && "Expecting a Phi."); 584 585 InitVal = 0; 586 LoopVal = 0; 587 for (unsigned i = 1, e = Phi.getNumOperands(); i != e; i += 2) 588 if (Phi.getOperand(i + 1).getMBB() != Loop) 589 InitVal = Phi.getOperand(i).getReg(); 590 else 591 LoopVal = Phi.getOperand(i).getReg(); 592 593 assert(InitVal != 0 && LoopVal != 0 && "Unexpected Phi structure."); 594 } 595 596 /// Return the Phi register value that comes the loop block. 597 static unsigned getLoopPhiReg(MachineInstr &Phi, MachineBasicBlock *LoopBB) { 598 for (unsigned i = 1, e = Phi.getNumOperands(); i != e; i += 2) 599 if (Phi.getOperand(i + 1).getMBB() == LoopBB) 600 return Phi.getOperand(i).getReg(); 601 return 0; 602 } 603 604 /// Return true if SUb can be reached from SUa following the chain edges. 605 static bool isSuccOrder(SUnit *SUa, SUnit *SUb) { 606 SmallPtrSet<SUnit *, 8> Visited; 607 SmallVector<SUnit *, 8> Worklist; 608 Worklist.push_back(SUa); 609 while (!Worklist.empty()) { 610 const SUnit *SU = Worklist.pop_back_val(); 611 for (auto &SI : SU->Succs) { 612 SUnit *SuccSU = SI.getSUnit(); 613 if (SI.getKind() == SDep::Order) { 614 if (Visited.count(SuccSU)) 615 continue; 616 if (SuccSU == SUb) 617 return true; 618 Worklist.push_back(SuccSU); 619 Visited.insert(SuccSU); 620 } 621 } 622 } 623 return false; 624 } 625 626 /// Return true if the instruction causes a chain between memory 627 /// references before and after it. 628 static bool isDependenceBarrier(MachineInstr &MI, AliasAnalysis *AA) { 629 return MI.isCall() || MI.mayRaiseFPException() || 630 MI.hasUnmodeledSideEffects() || 631 (MI.hasOrderedMemoryRef() && 632 (!MI.mayLoad() || !MI.isDereferenceableInvariantLoad(AA))); 633 } 634 635 /// Return the underlying objects for the memory references of an instruction. 636 /// This function calls the code in ValueTracking, but first checks that the 637 /// instruction has a memory operand. 638 static void getUnderlyingObjects(const MachineInstr *MI, 639 SmallVectorImpl<const Value *> &Objs, 640 const DataLayout &DL) { 641 if (!MI->hasOneMemOperand()) 642 return; 643 MachineMemOperand *MM = *MI->memoperands_begin(); 644 if (!MM->getValue()) 645 return; 646 GetUnderlyingObjects(MM->getValue(), Objs, DL); 647 for (const Value *V : Objs) { 648 if (!isIdentifiedObject(V)) { 649 Objs.clear(); 650 return; 651 } 652 Objs.push_back(V); 653 } 654 } 655 656 /// Add a chain edge between a load and store if the store can be an 657 /// alias of the load on a subsequent iteration, i.e., a loop carried 658 /// dependence. This code is very similar to the code in ScheduleDAGInstrs 659 /// but that code doesn't create loop carried dependences. 660 void SwingSchedulerDAG::addLoopCarriedDependences(AliasAnalysis *AA) { 661 MapVector<const Value *, SmallVector<SUnit *, 4>> PendingLoads; 662 Value *UnknownValue = 663 UndefValue::get(Type::getVoidTy(MF.getFunction().getContext())); 664 for (auto &SU : SUnits) { 665 MachineInstr &MI = *SU.getInstr(); 666 if (isDependenceBarrier(MI, AA)) 667 PendingLoads.clear(); 668 else if (MI.mayLoad()) { 669 SmallVector<const Value *, 4> Objs; 670 getUnderlyingObjects(&MI, Objs, MF.getDataLayout()); 671 if (Objs.empty()) 672 Objs.push_back(UnknownValue); 673 for (auto V : Objs) { 674 SmallVector<SUnit *, 4> &SUs = PendingLoads[V]; 675 SUs.push_back(&SU); 676 } 677 } else if (MI.mayStore()) { 678 SmallVector<const Value *, 4> Objs; 679 getUnderlyingObjects(&MI, Objs, MF.getDataLayout()); 680 if (Objs.empty()) 681 Objs.push_back(UnknownValue); 682 for (auto V : Objs) { 683 MapVector<const Value *, SmallVector<SUnit *, 4>>::iterator I = 684 PendingLoads.find(V); 685 if (I == PendingLoads.end()) 686 continue; 687 for (auto Load : I->second) { 688 if (isSuccOrder(Load, &SU)) 689 continue; 690 MachineInstr &LdMI = *Load->getInstr(); 691 // First, perform the cheaper check that compares the base register. 692 // If they are the same and the load offset is less than the store 693 // offset, then mark the dependence as loop carried potentially. 694 const MachineOperand *BaseOp1, *BaseOp2; 695 int64_t Offset1, Offset2; 696 if (TII->getMemOperandWithOffset(LdMI, BaseOp1, Offset1, TRI) && 697 TII->getMemOperandWithOffset(MI, BaseOp2, Offset2, TRI)) { 698 if (BaseOp1->isIdenticalTo(*BaseOp2) && 699 (int)Offset1 < (int)Offset2) { 700 assert(TII->areMemAccessesTriviallyDisjoint(LdMI, MI) && 701 "What happened to the chain edge?"); 702 SDep Dep(Load, SDep::Barrier); 703 Dep.setLatency(1); 704 SU.addPred(Dep); 705 continue; 706 } 707 } 708 // Second, the more expensive check that uses alias analysis on the 709 // base registers. If they alias, and the load offset is less than 710 // the store offset, the mark the dependence as loop carried. 711 if (!AA) { 712 SDep Dep(Load, SDep::Barrier); 713 Dep.setLatency(1); 714 SU.addPred(Dep); 715 continue; 716 } 717 MachineMemOperand *MMO1 = *LdMI.memoperands_begin(); 718 MachineMemOperand *MMO2 = *MI.memoperands_begin(); 719 if (!MMO1->getValue() || !MMO2->getValue()) { 720 SDep Dep(Load, SDep::Barrier); 721 Dep.setLatency(1); 722 SU.addPred(Dep); 723 continue; 724 } 725 if (MMO1->getValue() == MMO2->getValue() && 726 MMO1->getOffset() <= MMO2->getOffset()) { 727 SDep Dep(Load, SDep::Barrier); 728 Dep.setLatency(1); 729 SU.addPred(Dep); 730 continue; 731 } 732 AliasResult AAResult = AA->alias( 733 MemoryLocation(MMO1->getValue(), LocationSize::unknown(), 734 MMO1->getAAInfo()), 735 MemoryLocation(MMO2->getValue(), LocationSize::unknown(), 736 MMO2->getAAInfo())); 737 738 if (AAResult != NoAlias) { 739 SDep Dep(Load, SDep::Barrier); 740 Dep.setLatency(1); 741 SU.addPred(Dep); 742 } 743 } 744 } 745 } 746 } 747 } 748 749 /// Update the phi dependences to the DAG because ScheduleDAGInstrs no longer 750 /// processes dependences for PHIs. This function adds true dependences 751 /// from a PHI to a use, and a loop carried dependence from the use to the 752 /// PHI. The loop carried dependence is represented as an anti dependence 753 /// edge. This function also removes chain dependences between unrelated 754 /// PHIs. 755 void SwingSchedulerDAG::updatePhiDependences() { 756 SmallVector<SDep, 4> RemoveDeps; 757 const TargetSubtargetInfo &ST = MF.getSubtarget<TargetSubtargetInfo>(); 758 759 // Iterate over each DAG node. 760 for (SUnit &I : SUnits) { 761 RemoveDeps.clear(); 762 // Set to true if the instruction has an operand defined by a Phi. 763 unsigned HasPhiUse = 0; 764 unsigned HasPhiDef = 0; 765 MachineInstr *MI = I.getInstr(); 766 // Iterate over each operand, and we process the definitions. 767 for (MachineInstr::mop_iterator MOI = MI->operands_begin(), 768 MOE = MI->operands_end(); 769 MOI != MOE; ++MOI) { 770 if (!MOI->isReg()) 771 continue; 772 Register Reg = MOI->getReg(); 773 if (MOI->isDef()) { 774 // If the register is used by a Phi, then create an anti dependence. 775 for (MachineRegisterInfo::use_instr_iterator 776 UI = MRI.use_instr_begin(Reg), 777 UE = MRI.use_instr_end(); 778 UI != UE; ++UI) { 779 MachineInstr *UseMI = &*UI; 780 SUnit *SU = getSUnit(UseMI); 781 if (SU != nullptr && UseMI->isPHI()) { 782 if (!MI->isPHI()) { 783 SDep Dep(SU, SDep::Anti, Reg); 784 Dep.setLatency(1); 785 I.addPred(Dep); 786 } else { 787 HasPhiDef = Reg; 788 // Add a chain edge to a dependent Phi that isn't an existing 789 // predecessor. 790 if (SU->NodeNum < I.NodeNum && !I.isPred(SU)) 791 I.addPred(SDep(SU, SDep::Barrier)); 792 } 793 } 794 } 795 } else if (MOI->isUse()) { 796 // If the register is defined by a Phi, then create a true dependence. 797 MachineInstr *DefMI = MRI.getUniqueVRegDef(Reg); 798 if (DefMI == nullptr) 799 continue; 800 SUnit *SU = getSUnit(DefMI); 801 if (SU != nullptr && DefMI->isPHI()) { 802 if (!MI->isPHI()) { 803 SDep Dep(SU, SDep::Data, Reg); 804 Dep.setLatency(0); 805 ST.adjustSchedDependency(SU, &I, Dep); 806 I.addPred(Dep); 807 } else { 808 HasPhiUse = Reg; 809 // Add a chain edge to a dependent Phi that isn't an existing 810 // predecessor. 811 if (SU->NodeNum < I.NodeNum && !I.isPred(SU)) 812 I.addPred(SDep(SU, SDep::Barrier)); 813 } 814 } 815 } 816 } 817 // Remove order dependences from an unrelated Phi. 818 if (!SwpPruneDeps) 819 continue; 820 for (auto &PI : I.Preds) { 821 MachineInstr *PMI = PI.getSUnit()->getInstr(); 822 if (PMI->isPHI() && PI.getKind() == SDep::Order) { 823 if (I.getInstr()->isPHI()) { 824 if (PMI->getOperand(0).getReg() == HasPhiUse) 825 continue; 826 if (getLoopPhiReg(*PMI, PMI->getParent()) == HasPhiDef) 827 continue; 828 } 829 RemoveDeps.push_back(PI); 830 } 831 } 832 for (int i = 0, e = RemoveDeps.size(); i != e; ++i) 833 I.removePred(RemoveDeps[i]); 834 } 835 } 836 837 /// Iterate over each DAG node and see if we can change any dependences 838 /// in order to reduce the recurrence MII. 839 void SwingSchedulerDAG::changeDependences() { 840 // See if an instruction can use a value from the previous iteration. 841 // If so, we update the base and offset of the instruction and change 842 // the dependences. 843 for (SUnit &I : SUnits) { 844 unsigned BasePos = 0, OffsetPos = 0, NewBase = 0; 845 int64_t NewOffset = 0; 846 if (!canUseLastOffsetValue(I.getInstr(), BasePos, OffsetPos, NewBase, 847 NewOffset)) 848 continue; 849 850 // Get the MI and SUnit for the instruction that defines the original base. 851 Register OrigBase = I.getInstr()->getOperand(BasePos).getReg(); 852 MachineInstr *DefMI = MRI.getUniqueVRegDef(OrigBase); 853 if (!DefMI) 854 continue; 855 SUnit *DefSU = getSUnit(DefMI); 856 if (!DefSU) 857 continue; 858 // Get the MI and SUnit for the instruction that defins the new base. 859 MachineInstr *LastMI = MRI.getUniqueVRegDef(NewBase); 860 if (!LastMI) 861 continue; 862 SUnit *LastSU = getSUnit(LastMI); 863 if (!LastSU) 864 continue; 865 866 if (Topo.IsReachable(&I, LastSU)) 867 continue; 868 869 // Remove the dependence. The value now depends on a prior iteration. 870 SmallVector<SDep, 4> Deps; 871 for (SUnit::pred_iterator P = I.Preds.begin(), E = I.Preds.end(); P != E; 872 ++P) 873 if (P->getSUnit() == DefSU) 874 Deps.push_back(*P); 875 for (int i = 0, e = Deps.size(); i != e; i++) { 876 Topo.RemovePred(&I, Deps[i].getSUnit()); 877 I.removePred(Deps[i]); 878 } 879 // Remove the chain dependence between the instructions. 880 Deps.clear(); 881 for (auto &P : LastSU->Preds) 882 if (P.getSUnit() == &I && P.getKind() == SDep::Order) 883 Deps.push_back(P); 884 for (int i = 0, e = Deps.size(); i != e; i++) { 885 Topo.RemovePred(LastSU, Deps[i].getSUnit()); 886 LastSU->removePred(Deps[i]); 887 } 888 889 // Add a dependence between the new instruction and the instruction 890 // that defines the new base. 891 SDep Dep(&I, SDep::Anti, NewBase); 892 Topo.AddPred(LastSU, &I); 893 LastSU->addPred(Dep); 894 895 // Remember the base and offset information so that we can update the 896 // instruction during code generation. 897 InstrChanges[&I] = std::make_pair(NewBase, NewOffset); 898 } 899 } 900 901 namespace { 902 903 // FuncUnitSorter - Comparison operator used to sort instructions by 904 // the number of functional unit choices. 905 struct FuncUnitSorter { 906 const InstrItineraryData *InstrItins; 907 const MCSubtargetInfo *STI; 908 DenseMap<unsigned, unsigned> Resources; 909 910 FuncUnitSorter(const TargetSubtargetInfo &TSI) 911 : InstrItins(TSI.getInstrItineraryData()), STI(&TSI) {} 912 913 // Compute the number of functional unit alternatives needed 914 // at each stage, and take the minimum value. We prioritize the 915 // instructions by the least number of choices first. 916 unsigned minFuncUnits(const MachineInstr *Inst, unsigned &F) const { 917 unsigned SchedClass = Inst->getDesc().getSchedClass(); 918 unsigned min = UINT_MAX; 919 if (InstrItins && !InstrItins->isEmpty()) { 920 for (const InstrStage &IS : 921 make_range(InstrItins->beginStage(SchedClass), 922 InstrItins->endStage(SchedClass))) { 923 unsigned funcUnits = IS.getUnits(); 924 unsigned numAlternatives = countPopulation(funcUnits); 925 if (numAlternatives < min) { 926 min = numAlternatives; 927 F = funcUnits; 928 } 929 } 930 return min; 931 } 932 if (STI && STI->getSchedModel().hasInstrSchedModel()) { 933 const MCSchedClassDesc *SCDesc = 934 STI->getSchedModel().getSchedClassDesc(SchedClass); 935 if (!SCDesc->isValid()) 936 // No valid Schedule Class Desc for schedClass, should be 937 // Pseudo/PostRAPseudo 938 return min; 939 940 for (const MCWriteProcResEntry &PRE : 941 make_range(STI->getWriteProcResBegin(SCDesc), 942 STI->getWriteProcResEnd(SCDesc))) { 943 if (!PRE.Cycles) 944 continue; 945 const MCProcResourceDesc *ProcResource = 946 STI->getSchedModel().getProcResource(PRE.ProcResourceIdx); 947 unsigned NumUnits = ProcResource->NumUnits; 948 if (NumUnits < min) { 949 min = NumUnits; 950 F = PRE.ProcResourceIdx; 951 } 952 } 953 return min; 954 } 955 llvm_unreachable("Should have non-empty InstrItins or hasInstrSchedModel!"); 956 } 957 958 // Compute the critical resources needed by the instruction. This 959 // function records the functional units needed by instructions that 960 // must use only one functional unit. We use this as a tie breaker 961 // for computing the resource MII. The instrutions that require 962 // the same, highly used, functional unit have high priority. 963 void calcCriticalResources(MachineInstr &MI) { 964 unsigned SchedClass = MI.getDesc().getSchedClass(); 965 if (InstrItins && !InstrItins->isEmpty()) { 966 for (const InstrStage &IS : 967 make_range(InstrItins->beginStage(SchedClass), 968 InstrItins->endStage(SchedClass))) { 969 unsigned FuncUnits = IS.getUnits(); 970 if (countPopulation(FuncUnits) == 1) 971 Resources[FuncUnits]++; 972 } 973 return; 974 } 975 if (STI && STI->getSchedModel().hasInstrSchedModel()) { 976 const MCSchedClassDesc *SCDesc = 977 STI->getSchedModel().getSchedClassDesc(SchedClass); 978 if (!SCDesc->isValid()) 979 // No valid Schedule Class Desc for schedClass, should be 980 // Pseudo/PostRAPseudo 981 return; 982 983 for (const MCWriteProcResEntry &PRE : 984 make_range(STI->getWriteProcResBegin(SCDesc), 985 STI->getWriteProcResEnd(SCDesc))) { 986 if (!PRE.Cycles) 987 continue; 988 Resources[PRE.ProcResourceIdx]++; 989 } 990 return; 991 } 992 llvm_unreachable("Should have non-empty InstrItins or hasInstrSchedModel!"); 993 } 994 995 /// Return true if IS1 has less priority than IS2. 996 bool operator()(const MachineInstr *IS1, const MachineInstr *IS2) const { 997 unsigned F1 = 0, F2 = 0; 998 unsigned MFUs1 = minFuncUnits(IS1, F1); 999 unsigned MFUs2 = minFuncUnits(IS2, F2); 1000 if (MFUs1 == MFUs2) 1001 return Resources.lookup(F1) < Resources.lookup(F2); 1002 return MFUs1 > MFUs2; 1003 } 1004 }; 1005 1006 } // end anonymous namespace 1007 1008 /// Calculate the resource constrained minimum initiation interval for the 1009 /// specified loop. We use the DFA to model the resources needed for 1010 /// each instruction, and we ignore dependences. A different DFA is created 1011 /// for each cycle that is required. When adding a new instruction, we attempt 1012 /// to add it to each existing DFA, until a legal space is found. If the 1013 /// instruction cannot be reserved in an existing DFA, we create a new one. 1014 unsigned SwingSchedulerDAG::calculateResMII() { 1015 1016 LLVM_DEBUG(dbgs() << "calculateResMII:\n"); 1017 SmallVector<ResourceManager*, 8> Resources; 1018 MachineBasicBlock *MBB = Loop.getHeader(); 1019 Resources.push_back(new ResourceManager(&MF.getSubtarget())); 1020 1021 // Sort the instructions by the number of available choices for scheduling, 1022 // least to most. Use the number of critical resources as the tie breaker. 1023 FuncUnitSorter FUS = FuncUnitSorter(MF.getSubtarget()); 1024 for (MachineBasicBlock::iterator I = MBB->getFirstNonPHI(), 1025 E = MBB->getFirstTerminator(); 1026 I != E; ++I) 1027 FUS.calcCriticalResources(*I); 1028 PriorityQueue<MachineInstr *, std::vector<MachineInstr *>, FuncUnitSorter> 1029 FuncUnitOrder(FUS); 1030 1031 for (MachineBasicBlock::iterator I = MBB->getFirstNonPHI(), 1032 E = MBB->getFirstTerminator(); 1033 I != E; ++I) 1034 FuncUnitOrder.push(&*I); 1035 1036 while (!FuncUnitOrder.empty()) { 1037 MachineInstr *MI = FuncUnitOrder.top(); 1038 FuncUnitOrder.pop(); 1039 if (TII->isZeroCost(MI->getOpcode())) 1040 continue; 1041 // Attempt to reserve the instruction in an existing DFA. At least one 1042 // DFA is needed for each cycle. 1043 unsigned NumCycles = getSUnit(MI)->Latency; 1044 unsigned ReservedCycles = 0; 1045 SmallVectorImpl<ResourceManager *>::iterator RI = Resources.begin(); 1046 SmallVectorImpl<ResourceManager *>::iterator RE = Resources.end(); 1047 LLVM_DEBUG({ 1048 dbgs() << "Trying to reserve resource for " << NumCycles 1049 << " cycles for \n"; 1050 MI->dump(); 1051 }); 1052 for (unsigned C = 0; C < NumCycles; ++C) 1053 while (RI != RE) { 1054 if ((*RI)->canReserveResources(*MI)) { 1055 (*RI)->reserveResources(*MI); 1056 ++ReservedCycles; 1057 break; 1058 } 1059 RI++; 1060 } 1061 LLVM_DEBUG(dbgs() << "ReservedCycles:" << ReservedCycles 1062 << ", NumCycles:" << NumCycles << "\n"); 1063 // Add new DFAs, if needed, to reserve resources. 1064 for (unsigned C = ReservedCycles; C < NumCycles; ++C) { 1065 LLVM_DEBUG(if (SwpDebugResource) dbgs() 1066 << "NewResource created to reserve resources" 1067 << "\n"); 1068 ResourceManager *NewResource = new ResourceManager(&MF.getSubtarget()); 1069 assert(NewResource->canReserveResources(*MI) && "Reserve error."); 1070 NewResource->reserveResources(*MI); 1071 Resources.push_back(NewResource); 1072 } 1073 } 1074 int Resmii = Resources.size(); 1075 LLVM_DEBUG(dbgs() << "Retrun Res MII:" << Resmii << "\n"); 1076 // Delete the memory for each of the DFAs that were created earlier. 1077 for (ResourceManager *RI : Resources) { 1078 ResourceManager *D = RI; 1079 delete D; 1080 } 1081 Resources.clear(); 1082 return Resmii; 1083 } 1084 1085 /// Calculate the recurrence-constrainted minimum initiation interval. 1086 /// Iterate over each circuit. Compute the delay(c) and distance(c) 1087 /// for each circuit. The II needs to satisfy the inequality 1088 /// delay(c) - II*distance(c) <= 0. For each circuit, choose the smallest 1089 /// II that satisfies the inequality, and the RecMII is the maximum 1090 /// of those values. 1091 unsigned SwingSchedulerDAG::calculateRecMII(NodeSetType &NodeSets) { 1092 unsigned RecMII = 0; 1093 1094 for (NodeSet &Nodes : NodeSets) { 1095 if (Nodes.empty()) 1096 continue; 1097 1098 unsigned Delay = Nodes.getLatency(); 1099 unsigned Distance = 1; 1100 1101 // ii = ceil(delay / distance) 1102 unsigned CurMII = (Delay + Distance - 1) / Distance; 1103 Nodes.setRecMII(CurMII); 1104 if (CurMII > RecMII) 1105 RecMII = CurMII; 1106 } 1107 1108 return RecMII; 1109 } 1110 1111 /// Swap all the anti dependences in the DAG. That means it is no longer a DAG, 1112 /// but we do this to find the circuits, and then change them back. 1113 static void swapAntiDependences(std::vector<SUnit> &SUnits) { 1114 SmallVector<std::pair<SUnit *, SDep>, 8> DepsAdded; 1115 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) { 1116 SUnit *SU = &SUnits[i]; 1117 for (SUnit::pred_iterator IP = SU->Preds.begin(), EP = SU->Preds.end(); 1118 IP != EP; ++IP) { 1119 if (IP->getKind() != SDep::Anti) 1120 continue; 1121 DepsAdded.push_back(std::make_pair(SU, *IP)); 1122 } 1123 } 1124 for (SmallVector<std::pair<SUnit *, SDep>, 8>::iterator I = DepsAdded.begin(), 1125 E = DepsAdded.end(); 1126 I != E; ++I) { 1127 // Remove this anti dependency and add one in the reverse direction. 1128 SUnit *SU = I->first; 1129 SDep &D = I->second; 1130 SUnit *TargetSU = D.getSUnit(); 1131 unsigned Reg = D.getReg(); 1132 unsigned Lat = D.getLatency(); 1133 SU->removePred(D); 1134 SDep Dep(SU, SDep::Anti, Reg); 1135 Dep.setLatency(Lat); 1136 TargetSU->addPred(Dep); 1137 } 1138 } 1139 1140 /// Create the adjacency structure of the nodes in the graph. 1141 void SwingSchedulerDAG::Circuits::createAdjacencyStructure( 1142 SwingSchedulerDAG *DAG) { 1143 BitVector Added(SUnits.size()); 1144 DenseMap<int, int> OutputDeps; 1145 for (int i = 0, e = SUnits.size(); i != e; ++i) { 1146 Added.reset(); 1147 // Add any successor to the adjacency matrix and exclude duplicates. 1148 for (auto &SI : SUnits[i].Succs) { 1149 // Only create a back-edge on the first and last nodes of a dependence 1150 // chain. This records any chains and adds them later. 1151 if (SI.getKind() == SDep::Output) { 1152 int N = SI.getSUnit()->NodeNum; 1153 int BackEdge = i; 1154 auto Dep = OutputDeps.find(BackEdge); 1155 if (Dep != OutputDeps.end()) { 1156 BackEdge = Dep->second; 1157 OutputDeps.erase(Dep); 1158 } 1159 OutputDeps[N] = BackEdge; 1160 } 1161 // Do not process a boundary node, an artificial node. 1162 // A back-edge is processed only if it goes to a Phi. 1163 if (SI.getSUnit()->isBoundaryNode() || SI.isArtificial() || 1164 (SI.getKind() == SDep::Anti && !SI.getSUnit()->getInstr()->isPHI())) 1165 continue; 1166 int N = SI.getSUnit()->NodeNum; 1167 if (!Added.test(N)) { 1168 AdjK[i].push_back(N); 1169 Added.set(N); 1170 } 1171 } 1172 // A chain edge between a store and a load is treated as a back-edge in the 1173 // adjacency matrix. 1174 for (auto &PI : SUnits[i].Preds) { 1175 if (!SUnits[i].getInstr()->mayStore() || 1176 !DAG->isLoopCarriedDep(&SUnits[i], PI, false)) 1177 continue; 1178 if (PI.getKind() == SDep::Order && PI.getSUnit()->getInstr()->mayLoad()) { 1179 int N = PI.getSUnit()->NodeNum; 1180 if (!Added.test(N)) { 1181 AdjK[i].push_back(N); 1182 Added.set(N); 1183 } 1184 } 1185 } 1186 } 1187 // Add back-edges in the adjacency matrix for the output dependences. 1188 for (auto &OD : OutputDeps) 1189 if (!Added.test(OD.second)) { 1190 AdjK[OD.first].push_back(OD.second); 1191 Added.set(OD.second); 1192 } 1193 } 1194 1195 /// Identify an elementary circuit in the dependence graph starting at the 1196 /// specified node. 1197 bool SwingSchedulerDAG::Circuits::circuit(int V, int S, NodeSetType &NodeSets, 1198 bool HasBackedge) { 1199 SUnit *SV = &SUnits[V]; 1200 bool F = false; 1201 Stack.insert(SV); 1202 Blocked.set(V); 1203 1204 for (auto W : AdjK[V]) { 1205 if (NumPaths > MaxPaths) 1206 break; 1207 if (W < S) 1208 continue; 1209 if (W == S) { 1210 if (!HasBackedge) 1211 NodeSets.push_back(NodeSet(Stack.begin(), Stack.end())); 1212 F = true; 1213 ++NumPaths; 1214 break; 1215 } else if (!Blocked.test(W)) { 1216 if (circuit(W, S, NodeSets, 1217 Node2Idx->at(W) < Node2Idx->at(V) ? true : HasBackedge)) 1218 F = true; 1219 } 1220 } 1221 1222 if (F) 1223 unblock(V); 1224 else { 1225 for (auto W : AdjK[V]) { 1226 if (W < S) 1227 continue; 1228 if (B[W].count(SV) == 0) 1229 B[W].insert(SV); 1230 } 1231 } 1232 Stack.pop_back(); 1233 return F; 1234 } 1235 1236 /// Unblock a node in the circuit finding algorithm. 1237 void SwingSchedulerDAG::Circuits::unblock(int U) { 1238 Blocked.reset(U); 1239 SmallPtrSet<SUnit *, 4> &BU = B[U]; 1240 while (!BU.empty()) { 1241 SmallPtrSet<SUnit *, 4>::iterator SI = BU.begin(); 1242 assert(SI != BU.end() && "Invalid B set."); 1243 SUnit *W = *SI; 1244 BU.erase(W); 1245 if (Blocked.test(W->NodeNum)) 1246 unblock(W->NodeNum); 1247 } 1248 } 1249 1250 /// Identify all the elementary circuits in the dependence graph using 1251 /// Johnson's circuit algorithm. 1252 void SwingSchedulerDAG::findCircuits(NodeSetType &NodeSets) { 1253 // Swap all the anti dependences in the DAG. That means it is no longer a DAG, 1254 // but we do this to find the circuits, and then change them back. 1255 swapAntiDependences(SUnits); 1256 1257 Circuits Cir(SUnits, Topo); 1258 // Create the adjacency structure. 1259 Cir.createAdjacencyStructure(this); 1260 for (int i = 0, e = SUnits.size(); i != e; ++i) { 1261 Cir.reset(); 1262 Cir.circuit(i, i, NodeSets); 1263 } 1264 1265 // Change the dependences back so that we've created a DAG again. 1266 swapAntiDependences(SUnits); 1267 } 1268 1269 // Create artificial dependencies between the source of COPY/REG_SEQUENCE that 1270 // is loop-carried to the USE in next iteration. This will help pipeliner avoid 1271 // additional copies that are needed across iterations. An artificial dependence 1272 // edge is added from USE to SOURCE of COPY/REG_SEQUENCE. 1273 1274 // PHI-------Anti-Dep-----> COPY/REG_SEQUENCE (loop-carried) 1275 // SRCOfCopY------True-Dep---> COPY/REG_SEQUENCE 1276 // PHI-------True-Dep------> USEOfPhi 1277 1278 // The mutation creates 1279 // USEOfPHI -------Artificial-Dep---> SRCOfCopy 1280 1281 // This overall will ensure, the USEOfPHI is scheduled before SRCOfCopy 1282 // (since USE is a predecessor), implies, the COPY/ REG_SEQUENCE is scheduled 1283 // late to avoid additional copies across iterations. The possible scheduling 1284 // order would be 1285 // USEOfPHI --- SRCOfCopy--- COPY/REG_SEQUENCE. 1286 1287 void SwingSchedulerDAG::CopyToPhiMutation::apply(ScheduleDAGInstrs *DAG) { 1288 for (SUnit &SU : DAG->SUnits) { 1289 // Find the COPY/REG_SEQUENCE instruction. 1290 if (!SU.getInstr()->isCopy() && !SU.getInstr()->isRegSequence()) 1291 continue; 1292 1293 // Record the loop carried PHIs. 1294 SmallVector<SUnit *, 4> PHISUs; 1295 // Record the SrcSUs that feed the COPY/REG_SEQUENCE instructions. 1296 SmallVector<SUnit *, 4> SrcSUs; 1297 1298 for (auto &Dep : SU.Preds) { 1299 SUnit *TmpSU = Dep.getSUnit(); 1300 MachineInstr *TmpMI = TmpSU->getInstr(); 1301 SDep::Kind DepKind = Dep.getKind(); 1302 // Save the loop carried PHI. 1303 if (DepKind == SDep::Anti && TmpMI->isPHI()) 1304 PHISUs.push_back(TmpSU); 1305 // Save the source of COPY/REG_SEQUENCE. 1306 // If the source has no pre-decessors, we will end up creating cycles. 1307 else if (DepKind == SDep::Data && !TmpMI->isPHI() && TmpSU->NumPreds > 0) 1308 SrcSUs.push_back(TmpSU); 1309 } 1310 1311 if (PHISUs.size() == 0 || SrcSUs.size() == 0) 1312 continue; 1313 1314 // Find the USEs of PHI. If the use is a PHI or REG_SEQUENCE, push back this 1315 // SUnit to the container. 1316 SmallVector<SUnit *, 8> UseSUs; 1317 // Do not use iterator based loop here as we are updating the container. 1318 for (size_t Index = 0; Index < PHISUs.size(); ++Index) { 1319 for (auto &Dep : PHISUs[Index]->Succs) { 1320 if (Dep.getKind() != SDep::Data) 1321 continue; 1322 1323 SUnit *TmpSU = Dep.getSUnit(); 1324 MachineInstr *TmpMI = TmpSU->getInstr(); 1325 if (TmpMI->isPHI() || TmpMI->isRegSequence()) { 1326 PHISUs.push_back(TmpSU); 1327 continue; 1328 } 1329 UseSUs.push_back(TmpSU); 1330 } 1331 } 1332 1333 if (UseSUs.size() == 0) 1334 continue; 1335 1336 SwingSchedulerDAG *SDAG = cast<SwingSchedulerDAG>(DAG); 1337 // Add the artificial dependencies if it does not form a cycle. 1338 for (auto I : UseSUs) { 1339 for (auto Src : SrcSUs) { 1340 if (!SDAG->Topo.IsReachable(I, Src) && Src != I) { 1341 Src->addPred(SDep(I, SDep::Artificial)); 1342 SDAG->Topo.AddPred(Src, I); 1343 } 1344 } 1345 } 1346 } 1347 } 1348 1349 /// Return true for DAG nodes that we ignore when computing the cost functions. 1350 /// We ignore the back-edge recurrence in order to avoid unbounded recursion 1351 /// in the calculation of the ASAP, ALAP, etc functions. 1352 static bool ignoreDependence(const SDep &D, bool isPred) { 1353 if (D.isArtificial()) 1354 return true; 1355 return D.getKind() == SDep::Anti && isPred; 1356 } 1357 1358 /// Compute several functions need to order the nodes for scheduling. 1359 /// ASAP - Earliest time to schedule a node. 1360 /// ALAP - Latest time to schedule a node. 1361 /// MOV - Mobility function, difference between ALAP and ASAP. 1362 /// D - Depth of each node. 1363 /// H - Height of each node. 1364 void SwingSchedulerDAG::computeNodeFunctions(NodeSetType &NodeSets) { 1365 ScheduleInfo.resize(SUnits.size()); 1366 1367 LLVM_DEBUG({ 1368 for (ScheduleDAGTopologicalSort::const_iterator I = Topo.begin(), 1369 E = Topo.end(); 1370 I != E; ++I) { 1371 const SUnit &SU = SUnits[*I]; 1372 dumpNode(SU); 1373 } 1374 }); 1375 1376 int maxASAP = 0; 1377 // Compute ASAP and ZeroLatencyDepth. 1378 for (ScheduleDAGTopologicalSort::const_iterator I = Topo.begin(), 1379 E = Topo.end(); 1380 I != E; ++I) { 1381 int asap = 0; 1382 int zeroLatencyDepth = 0; 1383 SUnit *SU = &SUnits[*I]; 1384 for (SUnit::const_pred_iterator IP = SU->Preds.begin(), 1385 EP = SU->Preds.end(); 1386 IP != EP; ++IP) { 1387 SUnit *pred = IP->getSUnit(); 1388 if (IP->getLatency() == 0) 1389 zeroLatencyDepth = 1390 std::max(zeroLatencyDepth, getZeroLatencyDepth(pred) + 1); 1391 if (ignoreDependence(*IP, true)) 1392 continue; 1393 asap = std::max(asap, (int)(getASAP(pred) + IP->getLatency() - 1394 getDistance(pred, SU, *IP) * MII)); 1395 } 1396 maxASAP = std::max(maxASAP, asap); 1397 ScheduleInfo[*I].ASAP = asap; 1398 ScheduleInfo[*I].ZeroLatencyDepth = zeroLatencyDepth; 1399 } 1400 1401 // Compute ALAP, ZeroLatencyHeight, and MOV. 1402 for (ScheduleDAGTopologicalSort::const_reverse_iterator I = Topo.rbegin(), 1403 E = Topo.rend(); 1404 I != E; ++I) { 1405 int alap = maxASAP; 1406 int zeroLatencyHeight = 0; 1407 SUnit *SU = &SUnits[*I]; 1408 for (SUnit::const_succ_iterator IS = SU->Succs.begin(), 1409 ES = SU->Succs.end(); 1410 IS != ES; ++IS) { 1411 SUnit *succ = IS->getSUnit(); 1412 if (IS->getLatency() == 0) 1413 zeroLatencyHeight = 1414 std::max(zeroLatencyHeight, getZeroLatencyHeight(succ) + 1); 1415 if (ignoreDependence(*IS, true)) 1416 continue; 1417 alap = std::min(alap, (int)(getALAP(succ) - IS->getLatency() + 1418 getDistance(SU, succ, *IS) * MII)); 1419 } 1420 1421 ScheduleInfo[*I].ALAP = alap; 1422 ScheduleInfo[*I].ZeroLatencyHeight = zeroLatencyHeight; 1423 } 1424 1425 // After computing the node functions, compute the summary for each node set. 1426 for (NodeSet &I : NodeSets) 1427 I.computeNodeSetInfo(this); 1428 1429 LLVM_DEBUG({ 1430 for (unsigned i = 0; i < SUnits.size(); i++) { 1431 dbgs() << "\tNode " << i << ":\n"; 1432 dbgs() << "\t ASAP = " << getASAP(&SUnits[i]) << "\n"; 1433 dbgs() << "\t ALAP = " << getALAP(&SUnits[i]) << "\n"; 1434 dbgs() << "\t MOV = " << getMOV(&SUnits[i]) << "\n"; 1435 dbgs() << "\t D = " << getDepth(&SUnits[i]) << "\n"; 1436 dbgs() << "\t H = " << getHeight(&SUnits[i]) << "\n"; 1437 dbgs() << "\t ZLD = " << getZeroLatencyDepth(&SUnits[i]) << "\n"; 1438 dbgs() << "\t ZLH = " << getZeroLatencyHeight(&SUnits[i]) << "\n"; 1439 } 1440 }); 1441 } 1442 1443 /// Compute the Pred_L(O) set, as defined in the paper. The set is defined 1444 /// as the predecessors of the elements of NodeOrder that are not also in 1445 /// NodeOrder. 1446 static bool pred_L(SetVector<SUnit *> &NodeOrder, 1447 SmallSetVector<SUnit *, 8> &Preds, 1448 const NodeSet *S = nullptr) { 1449 Preds.clear(); 1450 for (SetVector<SUnit *>::iterator I = NodeOrder.begin(), E = NodeOrder.end(); 1451 I != E; ++I) { 1452 for (SUnit::pred_iterator PI = (*I)->Preds.begin(), PE = (*I)->Preds.end(); 1453 PI != PE; ++PI) { 1454 if (S && S->count(PI->getSUnit()) == 0) 1455 continue; 1456 if (ignoreDependence(*PI, true)) 1457 continue; 1458 if (NodeOrder.count(PI->getSUnit()) == 0) 1459 Preds.insert(PI->getSUnit()); 1460 } 1461 // Back-edges are predecessors with an anti-dependence. 1462 for (SUnit::const_succ_iterator IS = (*I)->Succs.begin(), 1463 ES = (*I)->Succs.end(); 1464 IS != ES; ++IS) { 1465 if (IS->getKind() != SDep::Anti) 1466 continue; 1467 if (S && S->count(IS->getSUnit()) == 0) 1468 continue; 1469 if (NodeOrder.count(IS->getSUnit()) == 0) 1470 Preds.insert(IS->getSUnit()); 1471 } 1472 } 1473 return !Preds.empty(); 1474 } 1475 1476 /// Compute the Succ_L(O) set, as defined in the paper. The set is defined 1477 /// as the successors of the elements of NodeOrder that are not also in 1478 /// NodeOrder. 1479 static bool succ_L(SetVector<SUnit *> &NodeOrder, 1480 SmallSetVector<SUnit *, 8> &Succs, 1481 const NodeSet *S = nullptr) { 1482 Succs.clear(); 1483 for (SetVector<SUnit *>::iterator I = NodeOrder.begin(), E = NodeOrder.end(); 1484 I != E; ++I) { 1485 for (SUnit::succ_iterator SI = (*I)->Succs.begin(), SE = (*I)->Succs.end(); 1486 SI != SE; ++SI) { 1487 if (S && S->count(SI->getSUnit()) == 0) 1488 continue; 1489 if (ignoreDependence(*SI, false)) 1490 continue; 1491 if (NodeOrder.count(SI->getSUnit()) == 0) 1492 Succs.insert(SI->getSUnit()); 1493 } 1494 for (SUnit::const_pred_iterator PI = (*I)->Preds.begin(), 1495 PE = (*I)->Preds.end(); 1496 PI != PE; ++PI) { 1497 if (PI->getKind() != SDep::Anti) 1498 continue; 1499 if (S && S->count(PI->getSUnit()) == 0) 1500 continue; 1501 if (NodeOrder.count(PI->getSUnit()) == 0) 1502 Succs.insert(PI->getSUnit()); 1503 } 1504 } 1505 return !Succs.empty(); 1506 } 1507 1508 /// Return true if there is a path from the specified node to any of the nodes 1509 /// in DestNodes. Keep track and return the nodes in any path. 1510 static bool computePath(SUnit *Cur, SetVector<SUnit *> &Path, 1511 SetVector<SUnit *> &DestNodes, 1512 SetVector<SUnit *> &Exclude, 1513 SmallPtrSet<SUnit *, 8> &Visited) { 1514 if (Cur->isBoundaryNode()) 1515 return false; 1516 if (Exclude.count(Cur) != 0) 1517 return false; 1518 if (DestNodes.count(Cur) != 0) 1519 return true; 1520 if (!Visited.insert(Cur).second) 1521 return Path.count(Cur) != 0; 1522 bool FoundPath = false; 1523 for (auto &SI : Cur->Succs) 1524 FoundPath |= computePath(SI.getSUnit(), Path, DestNodes, Exclude, Visited); 1525 for (auto &PI : Cur->Preds) 1526 if (PI.getKind() == SDep::Anti) 1527 FoundPath |= 1528 computePath(PI.getSUnit(), Path, DestNodes, Exclude, Visited); 1529 if (FoundPath) 1530 Path.insert(Cur); 1531 return FoundPath; 1532 } 1533 1534 /// Return true if Set1 is a subset of Set2. 1535 template <class S1Ty, class S2Ty> static bool isSubset(S1Ty &Set1, S2Ty &Set2) { 1536 for (typename S1Ty::iterator I = Set1.begin(), E = Set1.end(); I != E; ++I) 1537 if (Set2.count(*I) == 0) 1538 return false; 1539 return true; 1540 } 1541 1542 /// Compute the live-out registers for the instructions in a node-set. 1543 /// The live-out registers are those that are defined in the node-set, 1544 /// but not used. Except for use operands of Phis. 1545 static void computeLiveOuts(MachineFunction &MF, RegPressureTracker &RPTracker, 1546 NodeSet &NS) { 1547 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 1548 MachineRegisterInfo &MRI = MF.getRegInfo(); 1549 SmallVector<RegisterMaskPair, 8> LiveOutRegs; 1550 SmallSet<unsigned, 4> Uses; 1551 for (SUnit *SU : NS) { 1552 const MachineInstr *MI = SU->getInstr(); 1553 if (MI->isPHI()) 1554 continue; 1555 for (const MachineOperand &MO : MI->operands()) 1556 if (MO.isReg() && MO.isUse()) { 1557 Register Reg = MO.getReg(); 1558 if (Register::isVirtualRegister(Reg)) 1559 Uses.insert(Reg); 1560 else if (MRI.isAllocatable(Reg)) 1561 for (MCRegUnitIterator Units(Reg, TRI); Units.isValid(); ++Units) 1562 Uses.insert(*Units); 1563 } 1564 } 1565 for (SUnit *SU : NS) 1566 for (const MachineOperand &MO : SU->getInstr()->operands()) 1567 if (MO.isReg() && MO.isDef() && !MO.isDead()) { 1568 Register Reg = MO.getReg(); 1569 if (Register::isVirtualRegister(Reg)) { 1570 if (!Uses.count(Reg)) 1571 LiveOutRegs.push_back(RegisterMaskPair(Reg, 1572 LaneBitmask::getNone())); 1573 } else if (MRI.isAllocatable(Reg)) { 1574 for (MCRegUnitIterator Units(Reg, TRI); Units.isValid(); ++Units) 1575 if (!Uses.count(*Units)) 1576 LiveOutRegs.push_back(RegisterMaskPair(*Units, 1577 LaneBitmask::getNone())); 1578 } 1579 } 1580 RPTracker.addLiveRegs(LiveOutRegs); 1581 } 1582 1583 /// A heuristic to filter nodes in recurrent node-sets if the register 1584 /// pressure of a set is too high. 1585 void SwingSchedulerDAG::registerPressureFilter(NodeSetType &NodeSets) { 1586 for (auto &NS : NodeSets) { 1587 // Skip small node-sets since they won't cause register pressure problems. 1588 if (NS.size() <= 2) 1589 continue; 1590 IntervalPressure RecRegPressure; 1591 RegPressureTracker RecRPTracker(RecRegPressure); 1592 RecRPTracker.init(&MF, &RegClassInfo, &LIS, BB, BB->end(), false, true); 1593 computeLiveOuts(MF, RecRPTracker, NS); 1594 RecRPTracker.closeBottom(); 1595 1596 std::vector<SUnit *> SUnits(NS.begin(), NS.end()); 1597 llvm::sort(SUnits, [](const SUnit *A, const SUnit *B) { 1598 return A->NodeNum > B->NodeNum; 1599 }); 1600 1601 for (auto &SU : SUnits) { 1602 // Since we're computing the register pressure for a subset of the 1603 // instructions in a block, we need to set the tracker for each 1604 // instruction in the node-set. The tracker is set to the instruction 1605 // just after the one we're interested in. 1606 MachineBasicBlock::const_iterator CurInstI = SU->getInstr(); 1607 RecRPTracker.setPos(std::next(CurInstI)); 1608 1609 RegPressureDelta RPDelta; 1610 ArrayRef<PressureChange> CriticalPSets; 1611 RecRPTracker.getMaxUpwardPressureDelta(SU->getInstr(), nullptr, RPDelta, 1612 CriticalPSets, 1613 RecRegPressure.MaxSetPressure); 1614 if (RPDelta.Excess.isValid()) { 1615 LLVM_DEBUG( 1616 dbgs() << "Excess register pressure: SU(" << SU->NodeNum << ") " 1617 << TRI->getRegPressureSetName(RPDelta.Excess.getPSet()) 1618 << ":" << RPDelta.Excess.getUnitInc()); 1619 NS.setExceedPressure(SU); 1620 break; 1621 } 1622 RecRPTracker.recede(); 1623 } 1624 } 1625 } 1626 1627 /// A heuristic to colocate node sets that have the same set of 1628 /// successors. 1629 void SwingSchedulerDAG::colocateNodeSets(NodeSetType &NodeSets) { 1630 unsigned Colocate = 0; 1631 for (int i = 0, e = NodeSets.size(); i < e; ++i) { 1632 NodeSet &N1 = NodeSets[i]; 1633 SmallSetVector<SUnit *, 8> S1; 1634 if (N1.empty() || !succ_L(N1, S1)) 1635 continue; 1636 for (int j = i + 1; j < e; ++j) { 1637 NodeSet &N2 = NodeSets[j]; 1638 if (N1.compareRecMII(N2) != 0) 1639 continue; 1640 SmallSetVector<SUnit *, 8> S2; 1641 if (N2.empty() || !succ_L(N2, S2)) 1642 continue; 1643 if (isSubset(S1, S2) && S1.size() == S2.size()) { 1644 N1.setColocate(++Colocate); 1645 N2.setColocate(Colocate); 1646 break; 1647 } 1648 } 1649 } 1650 } 1651 1652 /// Check if the existing node-sets are profitable. If not, then ignore the 1653 /// recurrent node-sets, and attempt to schedule all nodes together. This is 1654 /// a heuristic. If the MII is large and all the recurrent node-sets are small, 1655 /// then it's best to try to schedule all instructions together instead of 1656 /// starting with the recurrent node-sets. 1657 void SwingSchedulerDAG::checkNodeSets(NodeSetType &NodeSets) { 1658 // Look for loops with a large MII. 1659 if (MII < 17) 1660 return; 1661 // Check if the node-set contains only a simple add recurrence. 1662 for (auto &NS : NodeSets) { 1663 if (NS.getRecMII() > 2) 1664 return; 1665 if (NS.getMaxDepth() > MII) 1666 return; 1667 } 1668 NodeSets.clear(); 1669 LLVM_DEBUG(dbgs() << "Clear recurrence node-sets\n"); 1670 return; 1671 } 1672 1673 /// Add the nodes that do not belong to a recurrence set into groups 1674 /// based upon connected componenets. 1675 void SwingSchedulerDAG::groupRemainingNodes(NodeSetType &NodeSets) { 1676 SetVector<SUnit *> NodesAdded; 1677 SmallPtrSet<SUnit *, 8> Visited; 1678 // Add the nodes that are on a path between the previous node sets and 1679 // the current node set. 1680 for (NodeSet &I : NodeSets) { 1681 SmallSetVector<SUnit *, 8> N; 1682 // Add the nodes from the current node set to the previous node set. 1683 if (succ_L(I, N)) { 1684 SetVector<SUnit *> Path; 1685 for (SUnit *NI : N) { 1686 Visited.clear(); 1687 computePath(NI, Path, NodesAdded, I, Visited); 1688 } 1689 if (!Path.empty()) 1690 I.insert(Path.begin(), Path.end()); 1691 } 1692 // Add the nodes from the previous node set to the current node set. 1693 N.clear(); 1694 if (succ_L(NodesAdded, N)) { 1695 SetVector<SUnit *> Path; 1696 for (SUnit *NI : N) { 1697 Visited.clear(); 1698 computePath(NI, Path, I, NodesAdded, Visited); 1699 } 1700 if (!Path.empty()) 1701 I.insert(Path.begin(), Path.end()); 1702 } 1703 NodesAdded.insert(I.begin(), I.end()); 1704 } 1705 1706 // Create a new node set with the connected nodes of any successor of a node 1707 // in a recurrent set. 1708 NodeSet NewSet; 1709 SmallSetVector<SUnit *, 8> N; 1710 if (succ_L(NodesAdded, N)) 1711 for (SUnit *I : N) 1712 addConnectedNodes(I, NewSet, NodesAdded); 1713 if (!NewSet.empty()) 1714 NodeSets.push_back(NewSet); 1715 1716 // Create a new node set with the connected nodes of any predecessor of a node 1717 // in a recurrent set. 1718 NewSet.clear(); 1719 if (pred_L(NodesAdded, N)) 1720 for (SUnit *I : N) 1721 addConnectedNodes(I, NewSet, NodesAdded); 1722 if (!NewSet.empty()) 1723 NodeSets.push_back(NewSet); 1724 1725 // Create new nodes sets with the connected nodes any remaining node that 1726 // has no predecessor. 1727 for (unsigned i = 0; i < SUnits.size(); ++i) { 1728 SUnit *SU = &SUnits[i]; 1729 if (NodesAdded.count(SU) == 0) { 1730 NewSet.clear(); 1731 addConnectedNodes(SU, NewSet, NodesAdded); 1732 if (!NewSet.empty()) 1733 NodeSets.push_back(NewSet); 1734 } 1735 } 1736 } 1737 1738 /// Add the node to the set, and add all of its connected nodes to the set. 1739 void SwingSchedulerDAG::addConnectedNodes(SUnit *SU, NodeSet &NewSet, 1740 SetVector<SUnit *> &NodesAdded) { 1741 NewSet.insert(SU); 1742 NodesAdded.insert(SU); 1743 for (auto &SI : SU->Succs) { 1744 SUnit *Successor = SI.getSUnit(); 1745 if (!SI.isArtificial() && NodesAdded.count(Successor) == 0) 1746 addConnectedNodes(Successor, NewSet, NodesAdded); 1747 } 1748 for (auto &PI : SU->Preds) { 1749 SUnit *Predecessor = PI.getSUnit(); 1750 if (!PI.isArtificial() && NodesAdded.count(Predecessor) == 0) 1751 addConnectedNodes(Predecessor, NewSet, NodesAdded); 1752 } 1753 } 1754 1755 /// Return true if Set1 contains elements in Set2. The elements in common 1756 /// are returned in a different container. 1757 static bool isIntersect(SmallSetVector<SUnit *, 8> &Set1, const NodeSet &Set2, 1758 SmallSetVector<SUnit *, 8> &Result) { 1759 Result.clear(); 1760 for (unsigned i = 0, e = Set1.size(); i != e; ++i) { 1761 SUnit *SU = Set1[i]; 1762 if (Set2.count(SU) != 0) 1763 Result.insert(SU); 1764 } 1765 return !Result.empty(); 1766 } 1767 1768 /// Merge the recurrence node sets that have the same initial node. 1769 void SwingSchedulerDAG::fuseRecs(NodeSetType &NodeSets) { 1770 for (NodeSetType::iterator I = NodeSets.begin(), E = NodeSets.end(); I != E; 1771 ++I) { 1772 NodeSet &NI = *I; 1773 for (NodeSetType::iterator J = I + 1; J != E;) { 1774 NodeSet &NJ = *J; 1775 if (NI.getNode(0)->NodeNum == NJ.getNode(0)->NodeNum) { 1776 if (NJ.compareRecMII(NI) > 0) 1777 NI.setRecMII(NJ.getRecMII()); 1778 for (NodeSet::iterator NII = J->begin(), ENI = J->end(); NII != ENI; 1779 ++NII) 1780 I->insert(*NII); 1781 NodeSets.erase(J); 1782 E = NodeSets.end(); 1783 } else { 1784 ++J; 1785 } 1786 } 1787 } 1788 } 1789 1790 /// Remove nodes that have been scheduled in previous NodeSets. 1791 void SwingSchedulerDAG::removeDuplicateNodes(NodeSetType &NodeSets) { 1792 for (NodeSetType::iterator I = NodeSets.begin(), E = NodeSets.end(); I != E; 1793 ++I) 1794 for (NodeSetType::iterator J = I + 1; J != E;) { 1795 J->remove_if([&](SUnit *SUJ) { return I->count(SUJ); }); 1796 1797 if (J->empty()) { 1798 NodeSets.erase(J); 1799 E = NodeSets.end(); 1800 } else { 1801 ++J; 1802 } 1803 } 1804 } 1805 1806 /// Compute an ordered list of the dependence graph nodes, which 1807 /// indicates the order that the nodes will be scheduled. This is a 1808 /// two-level algorithm. First, a partial order is created, which 1809 /// consists of a list of sets ordered from highest to lowest priority. 1810 void SwingSchedulerDAG::computeNodeOrder(NodeSetType &NodeSets) { 1811 SmallSetVector<SUnit *, 8> R; 1812 NodeOrder.clear(); 1813 1814 for (auto &Nodes : NodeSets) { 1815 LLVM_DEBUG(dbgs() << "NodeSet size " << Nodes.size() << "\n"); 1816 OrderKind Order; 1817 SmallSetVector<SUnit *, 8> N; 1818 if (pred_L(NodeOrder, N) && isSubset(N, Nodes)) { 1819 R.insert(N.begin(), N.end()); 1820 Order = BottomUp; 1821 LLVM_DEBUG(dbgs() << " Bottom up (preds) "); 1822 } else if (succ_L(NodeOrder, N) && isSubset(N, Nodes)) { 1823 R.insert(N.begin(), N.end()); 1824 Order = TopDown; 1825 LLVM_DEBUG(dbgs() << " Top down (succs) "); 1826 } else if (isIntersect(N, Nodes, R)) { 1827 // If some of the successors are in the existing node-set, then use the 1828 // top-down ordering. 1829 Order = TopDown; 1830 LLVM_DEBUG(dbgs() << " Top down (intersect) "); 1831 } else if (NodeSets.size() == 1) { 1832 for (auto &N : Nodes) 1833 if (N->Succs.size() == 0) 1834 R.insert(N); 1835 Order = BottomUp; 1836 LLVM_DEBUG(dbgs() << " Bottom up (all) "); 1837 } else { 1838 // Find the node with the highest ASAP. 1839 SUnit *maxASAP = nullptr; 1840 for (SUnit *SU : Nodes) { 1841 if (maxASAP == nullptr || getASAP(SU) > getASAP(maxASAP) || 1842 (getASAP(SU) == getASAP(maxASAP) && SU->NodeNum > maxASAP->NodeNum)) 1843 maxASAP = SU; 1844 } 1845 R.insert(maxASAP); 1846 Order = BottomUp; 1847 LLVM_DEBUG(dbgs() << " Bottom up (default) "); 1848 } 1849 1850 while (!R.empty()) { 1851 if (Order == TopDown) { 1852 // Choose the node with the maximum height. If more than one, choose 1853 // the node wiTH the maximum ZeroLatencyHeight. If still more than one, 1854 // choose the node with the lowest MOV. 1855 while (!R.empty()) { 1856 SUnit *maxHeight = nullptr; 1857 for (SUnit *I : R) { 1858 if (maxHeight == nullptr || getHeight(I) > getHeight(maxHeight)) 1859 maxHeight = I; 1860 else if (getHeight(I) == getHeight(maxHeight) && 1861 getZeroLatencyHeight(I) > getZeroLatencyHeight(maxHeight)) 1862 maxHeight = I; 1863 else if (getHeight(I) == getHeight(maxHeight) && 1864 getZeroLatencyHeight(I) == 1865 getZeroLatencyHeight(maxHeight) && 1866 getMOV(I) < getMOV(maxHeight)) 1867 maxHeight = I; 1868 } 1869 NodeOrder.insert(maxHeight); 1870 LLVM_DEBUG(dbgs() << maxHeight->NodeNum << " "); 1871 R.remove(maxHeight); 1872 for (const auto &I : maxHeight->Succs) { 1873 if (Nodes.count(I.getSUnit()) == 0) 1874 continue; 1875 if (NodeOrder.count(I.getSUnit()) != 0) 1876 continue; 1877 if (ignoreDependence(I, false)) 1878 continue; 1879 R.insert(I.getSUnit()); 1880 } 1881 // Back-edges are predecessors with an anti-dependence. 1882 for (const auto &I : maxHeight->Preds) { 1883 if (I.getKind() != SDep::Anti) 1884 continue; 1885 if (Nodes.count(I.getSUnit()) == 0) 1886 continue; 1887 if (NodeOrder.count(I.getSUnit()) != 0) 1888 continue; 1889 R.insert(I.getSUnit()); 1890 } 1891 } 1892 Order = BottomUp; 1893 LLVM_DEBUG(dbgs() << "\n Switching order to bottom up "); 1894 SmallSetVector<SUnit *, 8> N; 1895 if (pred_L(NodeOrder, N, &Nodes)) 1896 R.insert(N.begin(), N.end()); 1897 } else { 1898 // Choose the node with the maximum depth. If more than one, choose 1899 // the node with the maximum ZeroLatencyDepth. If still more than one, 1900 // choose the node with the lowest MOV. 1901 while (!R.empty()) { 1902 SUnit *maxDepth = nullptr; 1903 for (SUnit *I : R) { 1904 if (maxDepth == nullptr || getDepth(I) > getDepth(maxDepth)) 1905 maxDepth = I; 1906 else if (getDepth(I) == getDepth(maxDepth) && 1907 getZeroLatencyDepth(I) > getZeroLatencyDepth(maxDepth)) 1908 maxDepth = I; 1909 else if (getDepth(I) == getDepth(maxDepth) && 1910 getZeroLatencyDepth(I) == getZeroLatencyDepth(maxDepth) && 1911 getMOV(I) < getMOV(maxDepth)) 1912 maxDepth = I; 1913 } 1914 NodeOrder.insert(maxDepth); 1915 LLVM_DEBUG(dbgs() << maxDepth->NodeNum << " "); 1916 R.remove(maxDepth); 1917 if (Nodes.isExceedSU(maxDepth)) { 1918 Order = TopDown; 1919 R.clear(); 1920 R.insert(Nodes.getNode(0)); 1921 break; 1922 } 1923 for (const auto &I : maxDepth->Preds) { 1924 if (Nodes.count(I.getSUnit()) == 0) 1925 continue; 1926 if (NodeOrder.count(I.getSUnit()) != 0) 1927 continue; 1928 R.insert(I.getSUnit()); 1929 } 1930 // Back-edges are predecessors with an anti-dependence. 1931 for (const auto &I : maxDepth->Succs) { 1932 if (I.getKind() != SDep::Anti) 1933 continue; 1934 if (Nodes.count(I.getSUnit()) == 0) 1935 continue; 1936 if (NodeOrder.count(I.getSUnit()) != 0) 1937 continue; 1938 R.insert(I.getSUnit()); 1939 } 1940 } 1941 Order = TopDown; 1942 LLVM_DEBUG(dbgs() << "\n Switching order to top down "); 1943 SmallSetVector<SUnit *, 8> N; 1944 if (succ_L(NodeOrder, N, &Nodes)) 1945 R.insert(N.begin(), N.end()); 1946 } 1947 } 1948 LLVM_DEBUG(dbgs() << "\nDone with Nodeset\n"); 1949 } 1950 1951 LLVM_DEBUG({ 1952 dbgs() << "Node order: "; 1953 for (SUnit *I : NodeOrder) 1954 dbgs() << " " << I->NodeNum << " "; 1955 dbgs() << "\n"; 1956 }); 1957 } 1958 1959 /// Process the nodes in the computed order and create the pipelined schedule 1960 /// of the instructions, if possible. Return true if a schedule is found. 1961 bool SwingSchedulerDAG::schedulePipeline(SMSchedule &Schedule) { 1962 1963 if (NodeOrder.empty()){ 1964 LLVM_DEBUG(dbgs() << "NodeOrder is empty! abort scheduling\n" ); 1965 return false; 1966 } 1967 1968 bool scheduleFound = false; 1969 unsigned II = 0; 1970 // Keep increasing II until a valid schedule is found. 1971 for (II = MII; II <= MAX_II && !scheduleFound; ++II) { 1972 Schedule.reset(); 1973 Schedule.setInitiationInterval(II); 1974 LLVM_DEBUG(dbgs() << "Try to schedule with " << II << "\n"); 1975 1976 SetVector<SUnit *>::iterator NI = NodeOrder.begin(); 1977 SetVector<SUnit *>::iterator NE = NodeOrder.end(); 1978 do { 1979 SUnit *SU = *NI; 1980 1981 // Compute the schedule time for the instruction, which is based 1982 // upon the scheduled time for any predecessors/successors. 1983 int EarlyStart = INT_MIN; 1984 int LateStart = INT_MAX; 1985 // These values are set when the size of the schedule window is limited 1986 // due to chain dependences. 1987 int SchedEnd = INT_MAX; 1988 int SchedStart = INT_MIN; 1989 Schedule.computeStart(SU, &EarlyStart, &LateStart, &SchedEnd, &SchedStart, 1990 II, this); 1991 LLVM_DEBUG({ 1992 dbgs() << "\n"; 1993 dbgs() << "Inst (" << SU->NodeNum << ") "; 1994 SU->getInstr()->dump(); 1995 dbgs() << "\n"; 1996 }); 1997 LLVM_DEBUG({ 1998 dbgs() << format("\tes: %8x ls: %8x me: %8x ms: %8x\n", EarlyStart, 1999 LateStart, SchedEnd, SchedStart); 2000 }); 2001 2002 if (EarlyStart > LateStart || SchedEnd < EarlyStart || 2003 SchedStart > LateStart) 2004 scheduleFound = false; 2005 else if (EarlyStart != INT_MIN && LateStart == INT_MAX) { 2006 SchedEnd = std::min(SchedEnd, EarlyStart + (int)II - 1); 2007 scheduleFound = Schedule.insert(SU, EarlyStart, SchedEnd, II); 2008 } else if (EarlyStart == INT_MIN && LateStart != INT_MAX) { 2009 SchedStart = std::max(SchedStart, LateStart - (int)II + 1); 2010 scheduleFound = Schedule.insert(SU, LateStart, SchedStart, II); 2011 } else if (EarlyStart != INT_MIN && LateStart != INT_MAX) { 2012 SchedEnd = 2013 std::min(SchedEnd, std::min(LateStart, EarlyStart + (int)II - 1)); 2014 // When scheduling a Phi it is better to start at the late cycle and go 2015 // backwards. The default order may insert the Phi too far away from 2016 // its first dependence. 2017 if (SU->getInstr()->isPHI()) 2018 scheduleFound = Schedule.insert(SU, SchedEnd, EarlyStart, II); 2019 else 2020 scheduleFound = Schedule.insert(SU, EarlyStart, SchedEnd, II); 2021 } else { 2022 int FirstCycle = Schedule.getFirstCycle(); 2023 scheduleFound = Schedule.insert(SU, FirstCycle + getASAP(SU), 2024 FirstCycle + getASAP(SU) + II - 1, II); 2025 } 2026 // Even if we find a schedule, make sure the schedule doesn't exceed the 2027 // allowable number of stages. We keep trying if this happens. 2028 if (scheduleFound) 2029 if (SwpMaxStages > -1 && 2030 Schedule.getMaxStageCount() > (unsigned)SwpMaxStages) 2031 scheduleFound = false; 2032 2033 LLVM_DEBUG({ 2034 if (!scheduleFound) 2035 dbgs() << "\tCan't schedule\n"; 2036 }); 2037 } while (++NI != NE && scheduleFound); 2038 2039 // If a schedule is found, check if it is a valid schedule too. 2040 if (scheduleFound) 2041 scheduleFound = Schedule.isValidSchedule(this); 2042 } 2043 2044 LLVM_DEBUG(dbgs() << "Schedule Found? " << scheduleFound << " (II=" << II 2045 << ")\n"); 2046 2047 if (scheduleFound) 2048 Schedule.finalizeSchedule(this); 2049 else 2050 Schedule.reset(); 2051 2052 return scheduleFound && Schedule.getMaxStageCount() > 0; 2053 } 2054 2055 /// Return true if we can compute the amount the instruction changes 2056 /// during each iteration. Set Delta to the amount of the change. 2057 bool SwingSchedulerDAG::computeDelta(MachineInstr &MI, unsigned &Delta) { 2058 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 2059 const MachineOperand *BaseOp; 2060 int64_t Offset; 2061 if (!TII->getMemOperandWithOffset(MI, BaseOp, Offset, TRI)) 2062 return false; 2063 2064 if (!BaseOp->isReg()) 2065 return false; 2066 2067 Register BaseReg = BaseOp->getReg(); 2068 2069 MachineRegisterInfo &MRI = MF.getRegInfo(); 2070 // Check if there is a Phi. If so, get the definition in the loop. 2071 MachineInstr *BaseDef = MRI.getVRegDef(BaseReg); 2072 if (BaseDef && BaseDef->isPHI()) { 2073 BaseReg = getLoopPhiReg(*BaseDef, MI.getParent()); 2074 BaseDef = MRI.getVRegDef(BaseReg); 2075 } 2076 if (!BaseDef) 2077 return false; 2078 2079 int D = 0; 2080 if (!TII->getIncrementValue(*BaseDef, D) && D >= 0) 2081 return false; 2082 2083 Delta = D; 2084 return true; 2085 } 2086 2087 /// Check if we can change the instruction to use an offset value from the 2088 /// previous iteration. If so, return true and set the base and offset values 2089 /// so that we can rewrite the load, if necessary. 2090 /// v1 = Phi(v0, v3) 2091 /// v2 = load v1, 0 2092 /// v3 = post_store v1, 4, x 2093 /// This function enables the load to be rewritten as v2 = load v3, 4. 2094 bool SwingSchedulerDAG::canUseLastOffsetValue(MachineInstr *MI, 2095 unsigned &BasePos, 2096 unsigned &OffsetPos, 2097 unsigned &NewBase, 2098 int64_t &Offset) { 2099 // Get the load instruction. 2100 if (TII->isPostIncrement(*MI)) 2101 return false; 2102 unsigned BasePosLd, OffsetPosLd; 2103 if (!TII->getBaseAndOffsetPosition(*MI, BasePosLd, OffsetPosLd)) 2104 return false; 2105 Register BaseReg = MI->getOperand(BasePosLd).getReg(); 2106 2107 // Look for the Phi instruction. 2108 MachineRegisterInfo &MRI = MI->getMF()->getRegInfo(); 2109 MachineInstr *Phi = MRI.getVRegDef(BaseReg); 2110 if (!Phi || !Phi->isPHI()) 2111 return false; 2112 // Get the register defined in the loop block. 2113 unsigned PrevReg = getLoopPhiReg(*Phi, MI->getParent()); 2114 if (!PrevReg) 2115 return false; 2116 2117 // Check for the post-increment load/store instruction. 2118 MachineInstr *PrevDef = MRI.getVRegDef(PrevReg); 2119 if (!PrevDef || PrevDef == MI) 2120 return false; 2121 2122 if (!TII->isPostIncrement(*PrevDef)) 2123 return false; 2124 2125 unsigned BasePos1 = 0, OffsetPos1 = 0; 2126 if (!TII->getBaseAndOffsetPosition(*PrevDef, BasePos1, OffsetPos1)) 2127 return false; 2128 2129 // Make sure that the instructions do not access the same memory location in 2130 // the next iteration. 2131 int64_t LoadOffset = MI->getOperand(OffsetPosLd).getImm(); 2132 int64_t StoreOffset = PrevDef->getOperand(OffsetPos1).getImm(); 2133 MachineInstr *NewMI = MF.CloneMachineInstr(MI); 2134 NewMI->getOperand(OffsetPosLd).setImm(LoadOffset + StoreOffset); 2135 bool Disjoint = TII->areMemAccessesTriviallyDisjoint(*NewMI, *PrevDef); 2136 MF.DeleteMachineInstr(NewMI); 2137 if (!Disjoint) 2138 return false; 2139 2140 // Set the return value once we determine that we return true. 2141 BasePos = BasePosLd; 2142 OffsetPos = OffsetPosLd; 2143 NewBase = PrevReg; 2144 Offset = StoreOffset; 2145 return true; 2146 } 2147 2148 /// Apply changes to the instruction if needed. The changes are need 2149 /// to improve the scheduling and depend up on the final schedule. 2150 void SwingSchedulerDAG::applyInstrChange(MachineInstr *MI, 2151 SMSchedule &Schedule) { 2152 SUnit *SU = getSUnit(MI); 2153 DenseMap<SUnit *, std::pair<unsigned, int64_t>>::iterator It = 2154 InstrChanges.find(SU); 2155 if (It != InstrChanges.end()) { 2156 std::pair<unsigned, int64_t> RegAndOffset = It->second; 2157 unsigned BasePos, OffsetPos; 2158 if (!TII->getBaseAndOffsetPosition(*MI, BasePos, OffsetPos)) 2159 return; 2160 Register BaseReg = MI->getOperand(BasePos).getReg(); 2161 MachineInstr *LoopDef = findDefInLoop(BaseReg); 2162 int DefStageNum = Schedule.stageScheduled(getSUnit(LoopDef)); 2163 int DefCycleNum = Schedule.cycleScheduled(getSUnit(LoopDef)); 2164 int BaseStageNum = Schedule.stageScheduled(SU); 2165 int BaseCycleNum = Schedule.cycleScheduled(SU); 2166 if (BaseStageNum < DefStageNum) { 2167 MachineInstr *NewMI = MF.CloneMachineInstr(MI); 2168 int OffsetDiff = DefStageNum - BaseStageNum; 2169 if (DefCycleNum < BaseCycleNum) { 2170 NewMI->getOperand(BasePos).setReg(RegAndOffset.first); 2171 if (OffsetDiff > 0) 2172 --OffsetDiff; 2173 } 2174 int64_t NewOffset = 2175 MI->getOperand(OffsetPos).getImm() + RegAndOffset.second * OffsetDiff; 2176 NewMI->getOperand(OffsetPos).setImm(NewOffset); 2177 SU->setInstr(NewMI); 2178 MISUnitMap[NewMI] = SU; 2179 NewMIs[MI] = NewMI; 2180 } 2181 } 2182 } 2183 2184 /// Return the instruction in the loop that defines the register. 2185 /// If the definition is a Phi, then follow the Phi operand to 2186 /// the instruction in the loop. 2187 MachineInstr *SwingSchedulerDAG::findDefInLoop(unsigned Reg) { 2188 SmallPtrSet<MachineInstr *, 8> Visited; 2189 MachineInstr *Def = MRI.getVRegDef(Reg); 2190 while (Def->isPHI()) { 2191 if (!Visited.insert(Def).second) 2192 break; 2193 for (unsigned i = 1, e = Def->getNumOperands(); i < e; i += 2) 2194 if (Def->getOperand(i + 1).getMBB() == BB) { 2195 Def = MRI.getVRegDef(Def->getOperand(i).getReg()); 2196 break; 2197 } 2198 } 2199 return Def; 2200 } 2201 2202 /// Return true for an order or output dependence that is loop carried 2203 /// potentially. A dependence is loop carried if the destination defines a valu 2204 /// that may be used or defined by the source in a subsequent iteration. 2205 bool SwingSchedulerDAG::isLoopCarriedDep(SUnit *Source, const SDep &Dep, 2206 bool isSucc) { 2207 if ((Dep.getKind() != SDep::Order && Dep.getKind() != SDep::Output) || 2208 Dep.isArtificial()) 2209 return false; 2210 2211 if (!SwpPruneLoopCarried) 2212 return true; 2213 2214 if (Dep.getKind() == SDep::Output) 2215 return true; 2216 2217 MachineInstr *SI = Source->getInstr(); 2218 MachineInstr *DI = Dep.getSUnit()->getInstr(); 2219 if (!isSucc) 2220 std::swap(SI, DI); 2221 assert(SI != nullptr && DI != nullptr && "Expecting SUnit with an MI."); 2222 2223 // Assume ordered loads and stores may have a loop carried dependence. 2224 if (SI->hasUnmodeledSideEffects() || DI->hasUnmodeledSideEffects() || 2225 SI->mayRaiseFPException() || DI->mayRaiseFPException() || 2226 SI->hasOrderedMemoryRef() || DI->hasOrderedMemoryRef()) 2227 return true; 2228 2229 // Only chain dependences between a load and store can be loop carried. 2230 if (!DI->mayStore() || !SI->mayLoad()) 2231 return false; 2232 2233 unsigned DeltaS, DeltaD; 2234 if (!computeDelta(*SI, DeltaS) || !computeDelta(*DI, DeltaD)) 2235 return true; 2236 2237 const MachineOperand *BaseOpS, *BaseOpD; 2238 int64_t OffsetS, OffsetD; 2239 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 2240 if (!TII->getMemOperandWithOffset(*SI, BaseOpS, OffsetS, TRI) || 2241 !TII->getMemOperandWithOffset(*DI, BaseOpD, OffsetD, TRI)) 2242 return true; 2243 2244 if (!BaseOpS->isIdenticalTo(*BaseOpD)) 2245 return true; 2246 2247 // Check that the base register is incremented by a constant value for each 2248 // iteration. 2249 MachineInstr *Def = MRI.getVRegDef(BaseOpS->getReg()); 2250 if (!Def || !Def->isPHI()) 2251 return true; 2252 unsigned InitVal = 0; 2253 unsigned LoopVal = 0; 2254 getPhiRegs(*Def, BB, InitVal, LoopVal); 2255 MachineInstr *LoopDef = MRI.getVRegDef(LoopVal); 2256 int D = 0; 2257 if (!LoopDef || !TII->getIncrementValue(*LoopDef, D)) 2258 return true; 2259 2260 uint64_t AccessSizeS = (*SI->memoperands_begin())->getSize(); 2261 uint64_t AccessSizeD = (*DI->memoperands_begin())->getSize(); 2262 2263 // This is the main test, which checks the offset values and the loop 2264 // increment value to determine if the accesses may be loop carried. 2265 if (AccessSizeS == MemoryLocation::UnknownSize || 2266 AccessSizeD == MemoryLocation::UnknownSize) 2267 return true; 2268 2269 if (DeltaS != DeltaD || DeltaS < AccessSizeS || DeltaD < AccessSizeD) 2270 return true; 2271 2272 return (OffsetS + (int64_t)AccessSizeS < OffsetD + (int64_t)AccessSizeD); 2273 } 2274 2275 void SwingSchedulerDAG::postprocessDAG() { 2276 for (auto &M : Mutations) 2277 M->apply(this); 2278 } 2279 2280 /// Try to schedule the node at the specified StartCycle and continue 2281 /// until the node is schedule or the EndCycle is reached. This function 2282 /// returns true if the node is scheduled. This routine may search either 2283 /// forward or backward for a place to insert the instruction based upon 2284 /// the relative values of StartCycle and EndCycle. 2285 bool SMSchedule::insert(SUnit *SU, int StartCycle, int EndCycle, int II) { 2286 bool forward = true; 2287 LLVM_DEBUG({ 2288 dbgs() << "Trying to insert node between " << StartCycle << " and " 2289 << EndCycle << " II: " << II << "\n"; 2290 }); 2291 if (StartCycle > EndCycle) 2292 forward = false; 2293 2294 // The terminating condition depends on the direction. 2295 int termCycle = forward ? EndCycle + 1 : EndCycle - 1; 2296 for (int curCycle = StartCycle; curCycle != termCycle; 2297 forward ? ++curCycle : --curCycle) { 2298 2299 // Add the already scheduled instructions at the specified cycle to the 2300 // DFA. 2301 ProcItinResources.clearResources(); 2302 for (int checkCycle = FirstCycle + ((curCycle - FirstCycle) % II); 2303 checkCycle <= LastCycle; checkCycle += II) { 2304 std::deque<SUnit *> &cycleInstrs = ScheduledInstrs[checkCycle]; 2305 2306 for (std::deque<SUnit *>::iterator I = cycleInstrs.begin(), 2307 E = cycleInstrs.end(); 2308 I != E; ++I) { 2309 if (ST.getInstrInfo()->isZeroCost((*I)->getInstr()->getOpcode())) 2310 continue; 2311 assert(ProcItinResources.canReserveResources(*(*I)->getInstr()) && 2312 "These instructions have already been scheduled."); 2313 ProcItinResources.reserveResources(*(*I)->getInstr()); 2314 } 2315 } 2316 if (ST.getInstrInfo()->isZeroCost(SU->getInstr()->getOpcode()) || 2317 ProcItinResources.canReserveResources(*SU->getInstr())) { 2318 LLVM_DEBUG({ 2319 dbgs() << "\tinsert at cycle " << curCycle << " "; 2320 SU->getInstr()->dump(); 2321 }); 2322 2323 ScheduledInstrs[curCycle].push_back(SU); 2324 InstrToCycle.insert(std::make_pair(SU, curCycle)); 2325 if (curCycle > LastCycle) 2326 LastCycle = curCycle; 2327 if (curCycle < FirstCycle) 2328 FirstCycle = curCycle; 2329 return true; 2330 } 2331 LLVM_DEBUG({ 2332 dbgs() << "\tfailed to insert at cycle " << curCycle << " "; 2333 SU->getInstr()->dump(); 2334 }); 2335 } 2336 return false; 2337 } 2338 2339 // Return the cycle of the earliest scheduled instruction in the chain. 2340 int SMSchedule::earliestCycleInChain(const SDep &Dep) { 2341 SmallPtrSet<SUnit *, 8> Visited; 2342 SmallVector<SDep, 8> Worklist; 2343 Worklist.push_back(Dep); 2344 int EarlyCycle = INT_MAX; 2345 while (!Worklist.empty()) { 2346 const SDep &Cur = Worklist.pop_back_val(); 2347 SUnit *PrevSU = Cur.getSUnit(); 2348 if (Visited.count(PrevSU)) 2349 continue; 2350 std::map<SUnit *, int>::const_iterator it = InstrToCycle.find(PrevSU); 2351 if (it == InstrToCycle.end()) 2352 continue; 2353 EarlyCycle = std::min(EarlyCycle, it->second); 2354 for (const auto &PI : PrevSU->Preds) 2355 if (PI.getKind() == SDep::Order || Dep.getKind() == SDep::Output) 2356 Worklist.push_back(PI); 2357 Visited.insert(PrevSU); 2358 } 2359 return EarlyCycle; 2360 } 2361 2362 // Return the cycle of the latest scheduled instruction in the chain. 2363 int SMSchedule::latestCycleInChain(const SDep &Dep) { 2364 SmallPtrSet<SUnit *, 8> Visited; 2365 SmallVector<SDep, 8> Worklist; 2366 Worklist.push_back(Dep); 2367 int LateCycle = INT_MIN; 2368 while (!Worklist.empty()) { 2369 const SDep &Cur = Worklist.pop_back_val(); 2370 SUnit *SuccSU = Cur.getSUnit(); 2371 if (Visited.count(SuccSU)) 2372 continue; 2373 std::map<SUnit *, int>::const_iterator it = InstrToCycle.find(SuccSU); 2374 if (it == InstrToCycle.end()) 2375 continue; 2376 LateCycle = std::max(LateCycle, it->second); 2377 for (const auto &SI : SuccSU->Succs) 2378 if (SI.getKind() == SDep::Order || Dep.getKind() == SDep::Output) 2379 Worklist.push_back(SI); 2380 Visited.insert(SuccSU); 2381 } 2382 return LateCycle; 2383 } 2384 2385 /// If an instruction has a use that spans multiple iterations, then 2386 /// return true. These instructions are characterized by having a back-ege 2387 /// to a Phi, which contains a reference to another Phi. 2388 static SUnit *multipleIterations(SUnit *SU, SwingSchedulerDAG *DAG) { 2389 for (auto &P : SU->Preds) 2390 if (DAG->isBackedge(SU, P) && P.getSUnit()->getInstr()->isPHI()) 2391 for (auto &S : P.getSUnit()->Succs) 2392 if (S.getKind() == SDep::Data && S.getSUnit()->getInstr()->isPHI()) 2393 return P.getSUnit(); 2394 return nullptr; 2395 } 2396 2397 /// Compute the scheduling start slot for the instruction. The start slot 2398 /// depends on any predecessor or successor nodes scheduled already. 2399 void SMSchedule::computeStart(SUnit *SU, int *MaxEarlyStart, int *MinLateStart, 2400 int *MinEnd, int *MaxStart, int II, 2401 SwingSchedulerDAG *DAG) { 2402 // Iterate over each instruction that has been scheduled already. The start 2403 // slot computation depends on whether the previously scheduled instruction 2404 // is a predecessor or successor of the specified instruction. 2405 for (int cycle = getFirstCycle(); cycle <= LastCycle; ++cycle) { 2406 2407 // Iterate over each instruction in the current cycle. 2408 for (SUnit *I : getInstructions(cycle)) { 2409 // Because we're processing a DAG for the dependences, we recognize 2410 // the back-edge in recurrences by anti dependences. 2411 for (unsigned i = 0, e = (unsigned)SU->Preds.size(); i != e; ++i) { 2412 const SDep &Dep = SU->Preds[i]; 2413 if (Dep.getSUnit() == I) { 2414 if (!DAG->isBackedge(SU, Dep)) { 2415 int EarlyStart = cycle + Dep.getLatency() - 2416 DAG->getDistance(Dep.getSUnit(), SU, Dep) * II; 2417 *MaxEarlyStart = std::max(*MaxEarlyStart, EarlyStart); 2418 if (DAG->isLoopCarriedDep(SU, Dep, false)) { 2419 int End = earliestCycleInChain(Dep) + (II - 1); 2420 *MinEnd = std::min(*MinEnd, End); 2421 } 2422 } else { 2423 int LateStart = cycle - Dep.getLatency() + 2424 DAG->getDistance(SU, Dep.getSUnit(), Dep) * II; 2425 *MinLateStart = std::min(*MinLateStart, LateStart); 2426 } 2427 } 2428 // For instruction that requires multiple iterations, make sure that 2429 // the dependent instruction is not scheduled past the definition. 2430 SUnit *BE = multipleIterations(I, DAG); 2431 if (BE && Dep.getSUnit() == BE && !SU->getInstr()->isPHI() && 2432 !SU->isPred(I)) 2433 *MinLateStart = std::min(*MinLateStart, cycle); 2434 } 2435 for (unsigned i = 0, e = (unsigned)SU->Succs.size(); i != e; ++i) { 2436 if (SU->Succs[i].getSUnit() == I) { 2437 const SDep &Dep = SU->Succs[i]; 2438 if (!DAG->isBackedge(SU, Dep)) { 2439 int LateStart = cycle - Dep.getLatency() + 2440 DAG->getDistance(SU, Dep.getSUnit(), Dep) * II; 2441 *MinLateStart = std::min(*MinLateStart, LateStart); 2442 if (DAG->isLoopCarriedDep(SU, Dep)) { 2443 int Start = latestCycleInChain(Dep) + 1 - II; 2444 *MaxStart = std::max(*MaxStart, Start); 2445 } 2446 } else { 2447 int EarlyStart = cycle + Dep.getLatency() - 2448 DAG->getDistance(Dep.getSUnit(), SU, Dep) * II; 2449 *MaxEarlyStart = std::max(*MaxEarlyStart, EarlyStart); 2450 } 2451 } 2452 } 2453 } 2454 } 2455 } 2456 2457 /// Order the instructions within a cycle so that the definitions occur 2458 /// before the uses. Returns true if the instruction is added to the start 2459 /// of the list, or false if added to the end. 2460 void SMSchedule::orderDependence(SwingSchedulerDAG *SSD, SUnit *SU, 2461 std::deque<SUnit *> &Insts) { 2462 MachineInstr *MI = SU->getInstr(); 2463 bool OrderBeforeUse = false; 2464 bool OrderAfterDef = false; 2465 bool OrderBeforeDef = false; 2466 unsigned MoveDef = 0; 2467 unsigned MoveUse = 0; 2468 int StageInst1 = stageScheduled(SU); 2469 2470 unsigned Pos = 0; 2471 for (std::deque<SUnit *>::iterator I = Insts.begin(), E = Insts.end(); I != E; 2472 ++I, ++Pos) { 2473 for (unsigned i = 0, e = MI->getNumOperands(); i < e; ++i) { 2474 MachineOperand &MO = MI->getOperand(i); 2475 if (!MO.isReg() || !Register::isVirtualRegister(MO.getReg())) 2476 continue; 2477 2478 Register Reg = MO.getReg(); 2479 unsigned BasePos, OffsetPos; 2480 if (ST.getInstrInfo()->getBaseAndOffsetPosition(*MI, BasePos, OffsetPos)) 2481 if (MI->getOperand(BasePos).getReg() == Reg) 2482 if (unsigned NewReg = SSD->getInstrBaseReg(SU)) 2483 Reg = NewReg; 2484 bool Reads, Writes; 2485 std::tie(Reads, Writes) = 2486 (*I)->getInstr()->readsWritesVirtualRegister(Reg); 2487 if (MO.isDef() && Reads && stageScheduled(*I) <= StageInst1) { 2488 OrderBeforeUse = true; 2489 if (MoveUse == 0) 2490 MoveUse = Pos; 2491 } else if (MO.isDef() && Reads && stageScheduled(*I) > StageInst1) { 2492 // Add the instruction after the scheduled instruction. 2493 OrderAfterDef = true; 2494 MoveDef = Pos; 2495 } else if (MO.isUse() && Writes && stageScheduled(*I) == StageInst1) { 2496 if (cycleScheduled(*I) == cycleScheduled(SU) && !(*I)->isSucc(SU)) { 2497 OrderBeforeUse = true; 2498 if (MoveUse == 0) 2499 MoveUse = Pos; 2500 } else { 2501 OrderAfterDef = true; 2502 MoveDef = Pos; 2503 } 2504 } else if (MO.isUse() && Writes && stageScheduled(*I) > StageInst1) { 2505 OrderBeforeUse = true; 2506 if (MoveUse == 0) 2507 MoveUse = Pos; 2508 if (MoveUse != 0) { 2509 OrderAfterDef = true; 2510 MoveDef = Pos - 1; 2511 } 2512 } else if (MO.isUse() && Writes && stageScheduled(*I) < StageInst1) { 2513 // Add the instruction before the scheduled instruction. 2514 OrderBeforeUse = true; 2515 if (MoveUse == 0) 2516 MoveUse = Pos; 2517 } else if (MO.isUse() && stageScheduled(*I) == StageInst1 && 2518 isLoopCarriedDefOfUse(SSD, (*I)->getInstr(), MO)) { 2519 if (MoveUse == 0) { 2520 OrderBeforeDef = true; 2521 MoveUse = Pos; 2522 } 2523 } 2524 } 2525 // Check for order dependences between instructions. Make sure the source 2526 // is ordered before the destination. 2527 for (auto &S : SU->Succs) { 2528 if (S.getSUnit() != *I) 2529 continue; 2530 if (S.getKind() == SDep::Order && stageScheduled(*I) == StageInst1) { 2531 OrderBeforeUse = true; 2532 if (Pos < MoveUse) 2533 MoveUse = Pos; 2534 } 2535 // We did not handle HW dependences in previous for loop, 2536 // and we normally set Latency = 0 for Anti deps, 2537 // so may have nodes in same cycle with Anti denpendent on HW regs. 2538 else if (S.getKind() == SDep::Anti && stageScheduled(*I) == StageInst1) { 2539 OrderBeforeUse = true; 2540 if ((MoveUse == 0) || (Pos < MoveUse)) 2541 MoveUse = Pos; 2542 } 2543 } 2544 for (auto &P : SU->Preds) { 2545 if (P.getSUnit() != *I) 2546 continue; 2547 if (P.getKind() == SDep::Order && stageScheduled(*I) == StageInst1) { 2548 OrderAfterDef = true; 2549 MoveDef = Pos; 2550 } 2551 } 2552 } 2553 2554 // A circular dependence. 2555 if (OrderAfterDef && OrderBeforeUse && MoveUse == MoveDef) 2556 OrderBeforeUse = false; 2557 2558 // OrderAfterDef takes precedences over OrderBeforeDef. The latter is due 2559 // to a loop-carried dependence. 2560 if (OrderBeforeDef) 2561 OrderBeforeUse = !OrderAfterDef || (MoveUse > MoveDef); 2562 2563 // The uncommon case when the instruction order needs to be updated because 2564 // there is both a use and def. 2565 if (OrderBeforeUse && OrderAfterDef) { 2566 SUnit *UseSU = Insts.at(MoveUse); 2567 SUnit *DefSU = Insts.at(MoveDef); 2568 if (MoveUse > MoveDef) { 2569 Insts.erase(Insts.begin() + MoveUse); 2570 Insts.erase(Insts.begin() + MoveDef); 2571 } else { 2572 Insts.erase(Insts.begin() + MoveDef); 2573 Insts.erase(Insts.begin() + MoveUse); 2574 } 2575 orderDependence(SSD, UseSU, Insts); 2576 orderDependence(SSD, SU, Insts); 2577 orderDependence(SSD, DefSU, Insts); 2578 return; 2579 } 2580 // Put the new instruction first if there is a use in the list. Otherwise, 2581 // put it at the end of the list. 2582 if (OrderBeforeUse) 2583 Insts.push_front(SU); 2584 else 2585 Insts.push_back(SU); 2586 } 2587 2588 /// Return true if the scheduled Phi has a loop carried operand. 2589 bool SMSchedule::isLoopCarried(SwingSchedulerDAG *SSD, MachineInstr &Phi) { 2590 if (!Phi.isPHI()) 2591 return false; 2592 assert(Phi.isPHI() && "Expecting a Phi."); 2593 SUnit *DefSU = SSD->getSUnit(&Phi); 2594 unsigned DefCycle = cycleScheduled(DefSU); 2595 int DefStage = stageScheduled(DefSU); 2596 2597 unsigned InitVal = 0; 2598 unsigned LoopVal = 0; 2599 getPhiRegs(Phi, Phi.getParent(), InitVal, LoopVal); 2600 SUnit *UseSU = SSD->getSUnit(MRI.getVRegDef(LoopVal)); 2601 if (!UseSU) 2602 return true; 2603 if (UseSU->getInstr()->isPHI()) 2604 return true; 2605 unsigned LoopCycle = cycleScheduled(UseSU); 2606 int LoopStage = stageScheduled(UseSU); 2607 return (LoopCycle > DefCycle) || (LoopStage <= DefStage); 2608 } 2609 2610 /// Return true if the instruction is a definition that is loop carried 2611 /// and defines the use on the next iteration. 2612 /// v1 = phi(v2, v3) 2613 /// (Def) v3 = op v1 2614 /// (MO) = v1 2615 /// If MO appears before Def, then then v1 and v3 may get assigned to the same 2616 /// register. 2617 bool SMSchedule::isLoopCarriedDefOfUse(SwingSchedulerDAG *SSD, 2618 MachineInstr *Def, MachineOperand &MO) { 2619 if (!MO.isReg()) 2620 return false; 2621 if (Def->isPHI()) 2622 return false; 2623 MachineInstr *Phi = MRI.getVRegDef(MO.getReg()); 2624 if (!Phi || !Phi->isPHI() || Phi->getParent() != Def->getParent()) 2625 return false; 2626 if (!isLoopCarried(SSD, *Phi)) 2627 return false; 2628 unsigned LoopReg = getLoopPhiReg(*Phi, Phi->getParent()); 2629 for (unsigned i = 0, e = Def->getNumOperands(); i != e; ++i) { 2630 MachineOperand &DMO = Def->getOperand(i); 2631 if (!DMO.isReg() || !DMO.isDef()) 2632 continue; 2633 if (DMO.getReg() == LoopReg) 2634 return true; 2635 } 2636 return false; 2637 } 2638 2639 // Check if the generated schedule is valid. This function checks if 2640 // an instruction that uses a physical register is scheduled in a 2641 // different stage than the definition. The pipeliner does not handle 2642 // physical register values that may cross a basic block boundary. 2643 bool SMSchedule::isValidSchedule(SwingSchedulerDAG *SSD) { 2644 for (int i = 0, e = SSD->SUnits.size(); i < e; ++i) { 2645 SUnit &SU = SSD->SUnits[i]; 2646 if (!SU.hasPhysRegDefs) 2647 continue; 2648 int StageDef = stageScheduled(&SU); 2649 assert(StageDef != -1 && "Instruction should have been scheduled."); 2650 for (auto &SI : SU.Succs) 2651 if (SI.isAssignedRegDep()) 2652 if (Register::isPhysicalRegister(SI.getReg())) 2653 if (stageScheduled(SI.getSUnit()) != StageDef) 2654 return false; 2655 } 2656 return true; 2657 } 2658 2659 /// A property of the node order in swing-modulo-scheduling is 2660 /// that for nodes outside circuits the following holds: 2661 /// none of them is scheduled after both a successor and a 2662 /// predecessor. 2663 /// The method below checks whether the property is met. 2664 /// If not, debug information is printed and statistics information updated. 2665 /// Note that we do not use an assert statement. 2666 /// The reason is that although an invalid node oder may prevent 2667 /// the pipeliner from finding a pipelined schedule for arbitrary II, 2668 /// it does not lead to the generation of incorrect code. 2669 void SwingSchedulerDAG::checkValidNodeOrder(const NodeSetType &Circuits) const { 2670 2671 // a sorted vector that maps each SUnit to its index in the NodeOrder 2672 typedef std::pair<SUnit *, unsigned> UnitIndex; 2673 std::vector<UnitIndex> Indices(NodeOrder.size(), std::make_pair(nullptr, 0)); 2674 2675 for (unsigned i = 0, s = NodeOrder.size(); i < s; ++i) 2676 Indices.push_back(std::make_pair(NodeOrder[i], i)); 2677 2678 auto CompareKey = [](UnitIndex i1, UnitIndex i2) { 2679 return std::get<0>(i1) < std::get<0>(i2); 2680 }; 2681 2682 // sort, so that we can perform a binary search 2683 llvm::sort(Indices, CompareKey); 2684 2685 bool Valid = true; 2686 (void)Valid; 2687 // for each SUnit in the NodeOrder, check whether 2688 // it appears after both a successor and a predecessor 2689 // of the SUnit. If this is the case, and the SUnit 2690 // is not part of circuit, then the NodeOrder is not 2691 // valid. 2692 for (unsigned i = 0, s = NodeOrder.size(); i < s; ++i) { 2693 SUnit *SU = NodeOrder[i]; 2694 unsigned Index = i; 2695 2696 bool PredBefore = false; 2697 bool SuccBefore = false; 2698 2699 SUnit *Succ; 2700 SUnit *Pred; 2701 (void)Succ; 2702 (void)Pred; 2703 2704 for (SDep &PredEdge : SU->Preds) { 2705 SUnit *PredSU = PredEdge.getSUnit(); 2706 unsigned PredIndex = std::get<1>( 2707 *llvm::lower_bound(Indices, std::make_pair(PredSU, 0), CompareKey)); 2708 if (!PredSU->getInstr()->isPHI() && PredIndex < Index) { 2709 PredBefore = true; 2710 Pred = PredSU; 2711 break; 2712 } 2713 } 2714 2715 for (SDep &SuccEdge : SU->Succs) { 2716 SUnit *SuccSU = SuccEdge.getSUnit(); 2717 // Do not process a boundary node, it was not included in NodeOrder, 2718 // hence not in Indices either, call to std::lower_bound() below will 2719 // return Indices.end(). 2720 if (SuccSU->isBoundaryNode()) 2721 continue; 2722 unsigned SuccIndex = std::get<1>( 2723 *llvm::lower_bound(Indices, std::make_pair(SuccSU, 0), CompareKey)); 2724 if (!SuccSU->getInstr()->isPHI() && SuccIndex < Index) { 2725 SuccBefore = true; 2726 Succ = SuccSU; 2727 break; 2728 } 2729 } 2730 2731 if (PredBefore && SuccBefore && !SU->getInstr()->isPHI()) { 2732 // instructions in circuits are allowed to be scheduled 2733 // after both a successor and predecessor. 2734 bool InCircuit = llvm::any_of( 2735 Circuits, [SU](const NodeSet &Circuit) { return Circuit.count(SU); }); 2736 if (InCircuit) 2737 LLVM_DEBUG(dbgs() << "In a circuit, predecessor ";); 2738 else { 2739 Valid = false; 2740 NumNodeOrderIssues++; 2741 LLVM_DEBUG(dbgs() << "Predecessor ";); 2742 } 2743 LLVM_DEBUG(dbgs() << Pred->NodeNum << " and successor " << Succ->NodeNum 2744 << " are scheduled before node " << SU->NodeNum 2745 << "\n";); 2746 } 2747 } 2748 2749 LLVM_DEBUG({ 2750 if (!Valid) 2751 dbgs() << "Invalid node order found!\n"; 2752 }); 2753 } 2754 2755 /// Attempt to fix the degenerate cases when the instruction serialization 2756 /// causes the register lifetimes to overlap. For example, 2757 /// p' = store_pi(p, b) 2758 /// = load p, offset 2759 /// In this case p and p' overlap, which means that two registers are needed. 2760 /// Instead, this function changes the load to use p' and updates the offset. 2761 void SwingSchedulerDAG::fixupRegisterOverlaps(std::deque<SUnit *> &Instrs) { 2762 unsigned OverlapReg = 0; 2763 unsigned NewBaseReg = 0; 2764 for (SUnit *SU : Instrs) { 2765 MachineInstr *MI = SU->getInstr(); 2766 for (unsigned i = 0, e = MI->getNumOperands(); i < e; ++i) { 2767 const MachineOperand &MO = MI->getOperand(i); 2768 // Look for an instruction that uses p. The instruction occurs in the 2769 // same cycle but occurs later in the serialized order. 2770 if (MO.isReg() && MO.isUse() && MO.getReg() == OverlapReg) { 2771 // Check that the instruction appears in the InstrChanges structure, 2772 // which contains instructions that can have the offset updated. 2773 DenseMap<SUnit *, std::pair<unsigned, int64_t>>::iterator It = 2774 InstrChanges.find(SU); 2775 if (It != InstrChanges.end()) { 2776 unsigned BasePos, OffsetPos; 2777 // Update the base register and adjust the offset. 2778 if (TII->getBaseAndOffsetPosition(*MI, BasePos, OffsetPos)) { 2779 MachineInstr *NewMI = MF.CloneMachineInstr(MI); 2780 NewMI->getOperand(BasePos).setReg(NewBaseReg); 2781 int64_t NewOffset = 2782 MI->getOperand(OffsetPos).getImm() - It->second.second; 2783 NewMI->getOperand(OffsetPos).setImm(NewOffset); 2784 SU->setInstr(NewMI); 2785 MISUnitMap[NewMI] = SU; 2786 NewMIs[MI] = NewMI; 2787 } 2788 } 2789 OverlapReg = 0; 2790 NewBaseReg = 0; 2791 break; 2792 } 2793 // Look for an instruction of the form p' = op(p), which uses and defines 2794 // two virtual registers that get allocated to the same physical register. 2795 unsigned TiedUseIdx = 0; 2796 if (MI->isRegTiedToUseOperand(i, &TiedUseIdx)) { 2797 // OverlapReg is p in the example above. 2798 OverlapReg = MI->getOperand(TiedUseIdx).getReg(); 2799 // NewBaseReg is p' in the example above. 2800 NewBaseReg = MI->getOperand(i).getReg(); 2801 break; 2802 } 2803 } 2804 } 2805 } 2806 2807 /// After the schedule has been formed, call this function to combine 2808 /// the instructions from the different stages/cycles. That is, this 2809 /// function creates a schedule that represents a single iteration. 2810 void SMSchedule::finalizeSchedule(SwingSchedulerDAG *SSD) { 2811 // Move all instructions to the first stage from later stages. 2812 for (int cycle = getFirstCycle(); cycle <= getFinalCycle(); ++cycle) { 2813 for (int stage = 1, lastStage = getMaxStageCount(); stage <= lastStage; 2814 ++stage) { 2815 std::deque<SUnit *> &cycleInstrs = 2816 ScheduledInstrs[cycle + (stage * InitiationInterval)]; 2817 for (std::deque<SUnit *>::reverse_iterator I = cycleInstrs.rbegin(), 2818 E = cycleInstrs.rend(); 2819 I != E; ++I) 2820 ScheduledInstrs[cycle].push_front(*I); 2821 } 2822 } 2823 2824 // Erase all the elements in the later stages. Only one iteration should 2825 // remain in the scheduled list, and it contains all the instructions. 2826 for (int cycle = getFinalCycle() + 1; cycle <= LastCycle; ++cycle) 2827 ScheduledInstrs.erase(cycle); 2828 2829 // Change the registers in instruction as specified in the InstrChanges 2830 // map. We need to use the new registers to create the correct order. 2831 for (int i = 0, e = SSD->SUnits.size(); i != e; ++i) { 2832 SUnit *SU = &SSD->SUnits[i]; 2833 SSD->applyInstrChange(SU->getInstr(), *this); 2834 } 2835 2836 // Reorder the instructions in each cycle to fix and improve the 2837 // generated code. 2838 for (int Cycle = getFirstCycle(), E = getFinalCycle(); Cycle <= E; ++Cycle) { 2839 std::deque<SUnit *> &cycleInstrs = ScheduledInstrs[Cycle]; 2840 std::deque<SUnit *> newOrderPhi; 2841 for (unsigned i = 0, e = cycleInstrs.size(); i < e; ++i) { 2842 SUnit *SU = cycleInstrs[i]; 2843 if (SU->getInstr()->isPHI()) 2844 newOrderPhi.push_back(SU); 2845 } 2846 std::deque<SUnit *> newOrderI; 2847 for (unsigned i = 0, e = cycleInstrs.size(); i < e; ++i) { 2848 SUnit *SU = cycleInstrs[i]; 2849 if (!SU->getInstr()->isPHI()) 2850 orderDependence(SSD, SU, newOrderI); 2851 } 2852 // Replace the old order with the new order. 2853 cycleInstrs.swap(newOrderPhi); 2854 cycleInstrs.insert(cycleInstrs.end(), newOrderI.begin(), newOrderI.end()); 2855 SSD->fixupRegisterOverlaps(cycleInstrs); 2856 } 2857 2858 LLVM_DEBUG(dump();); 2859 } 2860 2861 void NodeSet::print(raw_ostream &os) const { 2862 os << "Num nodes " << size() << " rec " << RecMII << " mov " << MaxMOV 2863 << " depth " << MaxDepth << " col " << Colocate << "\n"; 2864 for (const auto &I : Nodes) 2865 os << " SU(" << I->NodeNum << ") " << *(I->getInstr()); 2866 os << "\n"; 2867 } 2868 2869 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 2870 /// Print the schedule information to the given output. 2871 void SMSchedule::print(raw_ostream &os) const { 2872 // Iterate over each cycle. 2873 for (int cycle = getFirstCycle(); cycle <= getFinalCycle(); ++cycle) { 2874 // Iterate over each instruction in the cycle. 2875 const_sched_iterator cycleInstrs = ScheduledInstrs.find(cycle); 2876 for (SUnit *CI : cycleInstrs->second) { 2877 os << "cycle " << cycle << " (" << stageScheduled(CI) << ") "; 2878 os << "(" << CI->NodeNum << ") "; 2879 CI->getInstr()->print(os); 2880 os << "\n"; 2881 } 2882 } 2883 } 2884 2885 /// Utility function used for debugging to print the schedule. 2886 LLVM_DUMP_METHOD void SMSchedule::dump() const { print(dbgs()); } 2887 LLVM_DUMP_METHOD void NodeSet::dump() const { print(dbgs()); } 2888 2889 #endif 2890 2891 void ResourceManager::initProcResourceVectors( 2892 const MCSchedModel &SM, SmallVectorImpl<uint64_t> &Masks) { 2893 unsigned ProcResourceID = 0; 2894 2895 // We currently limit the resource kinds to 64 and below so that we can use 2896 // uint64_t for Masks 2897 assert(SM.getNumProcResourceKinds() < 64 && 2898 "Too many kinds of resources, unsupported"); 2899 // Create a unique bitmask for every processor resource unit. 2900 // Skip resource at index 0, since it always references 'InvalidUnit'. 2901 Masks.resize(SM.getNumProcResourceKinds()); 2902 for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) { 2903 const MCProcResourceDesc &Desc = *SM.getProcResource(I); 2904 if (Desc.SubUnitsIdxBegin) 2905 continue; 2906 Masks[I] = 1ULL << ProcResourceID; 2907 ProcResourceID++; 2908 } 2909 // Create a unique bitmask for every processor resource group. 2910 for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) { 2911 const MCProcResourceDesc &Desc = *SM.getProcResource(I); 2912 if (!Desc.SubUnitsIdxBegin) 2913 continue; 2914 Masks[I] = 1ULL << ProcResourceID; 2915 for (unsigned U = 0; U < Desc.NumUnits; ++U) 2916 Masks[I] |= Masks[Desc.SubUnitsIdxBegin[U]]; 2917 ProcResourceID++; 2918 } 2919 LLVM_DEBUG({ 2920 if (SwpShowResMask) { 2921 dbgs() << "ProcResourceDesc:\n"; 2922 for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) { 2923 const MCProcResourceDesc *ProcResource = SM.getProcResource(I); 2924 dbgs() << format(" %16s(%2d): Mask: 0x%08x, NumUnits:%2d\n", 2925 ProcResource->Name, I, Masks[I], 2926 ProcResource->NumUnits); 2927 } 2928 dbgs() << " -----------------\n"; 2929 } 2930 }); 2931 } 2932 2933 bool ResourceManager::canReserveResources(const MCInstrDesc *MID) const { 2934 2935 LLVM_DEBUG({ 2936 if (SwpDebugResource) 2937 dbgs() << "canReserveResources:\n"; 2938 }); 2939 if (UseDFA) 2940 return DFAResources->canReserveResources(MID); 2941 2942 unsigned InsnClass = MID->getSchedClass(); 2943 const MCSchedClassDesc *SCDesc = SM.getSchedClassDesc(InsnClass); 2944 if (!SCDesc->isValid()) { 2945 LLVM_DEBUG({ 2946 dbgs() << "No valid Schedule Class Desc for schedClass!\n"; 2947 dbgs() << "isPseduo:" << MID->isPseudo() << "\n"; 2948 }); 2949 return true; 2950 } 2951 2952 const MCWriteProcResEntry *I = STI->getWriteProcResBegin(SCDesc); 2953 const MCWriteProcResEntry *E = STI->getWriteProcResEnd(SCDesc); 2954 for (; I != E; ++I) { 2955 if (!I->Cycles) 2956 continue; 2957 const MCProcResourceDesc *ProcResource = 2958 SM.getProcResource(I->ProcResourceIdx); 2959 unsigned NumUnits = ProcResource->NumUnits; 2960 LLVM_DEBUG({ 2961 if (SwpDebugResource) 2962 dbgs() << format(" %16s(%2d): Count: %2d, NumUnits:%2d, Cycles:%2d\n", 2963 ProcResource->Name, I->ProcResourceIdx, 2964 ProcResourceCount[I->ProcResourceIdx], NumUnits, 2965 I->Cycles); 2966 }); 2967 if (ProcResourceCount[I->ProcResourceIdx] >= NumUnits) 2968 return false; 2969 } 2970 LLVM_DEBUG(if (SwpDebugResource) dbgs() << "return true\n\n";); 2971 return true; 2972 } 2973 2974 void ResourceManager::reserveResources(const MCInstrDesc *MID) { 2975 LLVM_DEBUG({ 2976 if (SwpDebugResource) 2977 dbgs() << "reserveResources:\n"; 2978 }); 2979 if (UseDFA) 2980 return DFAResources->reserveResources(MID); 2981 2982 unsigned InsnClass = MID->getSchedClass(); 2983 const MCSchedClassDesc *SCDesc = SM.getSchedClassDesc(InsnClass); 2984 if (!SCDesc->isValid()) { 2985 LLVM_DEBUG({ 2986 dbgs() << "No valid Schedule Class Desc for schedClass!\n"; 2987 dbgs() << "isPseduo:" << MID->isPseudo() << "\n"; 2988 }); 2989 return; 2990 } 2991 for (const MCWriteProcResEntry &PRE : 2992 make_range(STI->getWriteProcResBegin(SCDesc), 2993 STI->getWriteProcResEnd(SCDesc))) { 2994 if (!PRE.Cycles) 2995 continue; 2996 ++ProcResourceCount[PRE.ProcResourceIdx]; 2997 LLVM_DEBUG({ 2998 if (SwpDebugResource) { 2999 const MCProcResourceDesc *ProcResource = 3000 SM.getProcResource(PRE.ProcResourceIdx); 3001 dbgs() << format(" %16s(%2d): Count: %2d, NumUnits:%2d, Cycles:%2d\n", 3002 ProcResource->Name, PRE.ProcResourceIdx, 3003 ProcResourceCount[PRE.ProcResourceIdx], 3004 ProcResource->NumUnits, PRE.Cycles); 3005 } 3006 }); 3007 } 3008 LLVM_DEBUG({ 3009 if (SwpDebugResource) 3010 dbgs() << "reserveResources: done!\n\n"; 3011 }); 3012 } 3013 3014 bool ResourceManager::canReserveResources(const MachineInstr &MI) const { 3015 return canReserveResources(&MI.getDesc()); 3016 } 3017 3018 void ResourceManager::reserveResources(const MachineInstr &MI) { 3019 return reserveResources(&MI.getDesc()); 3020 } 3021 3022 void ResourceManager::clearResources() { 3023 if (UseDFA) 3024 return DFAResources->clearResources(); 3025 std::fill(ProcResourceCount.begin(), ProcResourceCount.end(), 0); 3026 } 3027