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/CodeGen/MachinePipeliner.h" 33 #include "llvm/ADT/ArrayRef.h" 34 #include "llvm/ADT/BitVector.h" 35 #include "llvm/ADT/DenseMap.h" 36 #include "llvm/ADT/MapVector.h" 37 #include "llvm/ADT/PriorityQueue.h" 38 #include "llvm/ADT/SetOperations.h" 39 #include "llvm/ADT/SetVector.h" 40 #include "llvm/ADT/SmallPtrSet.h" 41 #include "llvm/ADT/SmallSet.h" 42 #include "llvm/ADT/SmallVector.h" 43 #include "llvm/ADT/Statistic.h" 44 #include "llvm/ADT/iterator_range.h" 45 #include "llvm/Analysis/AliasAnalysis.h" 46 #include "llvm/Analysis/MemoryLocation.h" 47 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 48 #include "llvm/Analysis/ValueTracking.h" 49 #include "llvm/CodeGen/DFAPacketizer.h" 50 #include "llvm/CodeGen/LiveIntervals.h" 51 #include "llvm/CodeGen/MachineBasicBlock.h" 52 #include "llvm/CodeGen/MachineDominators.h" 53 #include "llvm/CodeGen/MachineFunction.h" 54 #include "llvm/CodeGen/MachineFunctionPass.h" 55 #include "llvm/CodeGen/MachineInstr.h" 56 #include "llvm/CodeGen/MachineInstrBuilder.h" 57 #include "llvm/CodeGen/MachineLoopInfo.h" 58 #include "llvm/CodeGen/MachineMemOperand.h" 59 #include "llvm/CodeGen/MachineOperand.h" 60 #include "llvm/CodeGen/MachineRegisterInfo.h" 61 #include "llvm/CodeGen/ModuloSchedule.h" 62 #include "llvm/CodeGen/RegisterPressure.h" 63 #include "llvm/CodeGen/ScheduleDAG.h" 64 #include "llvm/CodeGen/ScheduleDAGMutation.h" 65 #include "llvm/CodeGen/TargetOpcodes.h" 66 #include "llvm/CodeGen/TargetRegisterInfo.h" 67 #include "llvm/CodeGen/TargetSubtargetInfo.h" 68 #include "llvm/Config/llvm-config.h" 69 #include "llvm/IR/Attributes.h" 70 #include "llvm/IR/Function.h" 71 #include "llvm/MC/LaneBitmask.h" 72 #include "llvm/MC/MCInstrDesc.h" 73 #include "llvm/MC/MCInstrItineraries.h" 74 #include "llvm/MC/MCRegisterInfo.h" 75 #include "llvm/Pass.h" 76 #include "llvm/Support/CommandLine.h" 77 #include "llvm/Support/Compiler.h" 78 #include "llvm/Support/Debug.h" 79 #include "llvm/Support/MathExtras.h" 80 #include "llvm/Support/raw_ostream.h" 81 #include <algorithm> 82 #include <cassert> 83 #include <climits> 84 #include <cstdint> 85 #include <deque> 86 #include <functional> 87 #include <iterator> 88 #include <map> 89 #include <memory> 90 #include <tuple> 91 #include <utility> 92 #include <vector> 93 94 using namespace llvm; 95 96 #define DEBUG_TYPE "pipeliner" 97 98 STATISTIC(NumTrytoPipeline, "Number of loops that we attempt to pipeline"); 99 STATISTIC(NumPipelined, "Number of loops software pipelined"); 100 STATISTIC(NumNodeOrderIssues, "Number of node order issues found"); 101 STATISTIC(NumFailBranch, "Pipeliner abort due to unknown branch"); 102 STATISTIC(NumFailLoop, "Pipeliner abort due to unsupported loop"); 103 STATISTIC(NumFailPreheader, "Pipeliner abort due to missing preheader"); 104 STATISTIC(NumFailLargeMaxMII, "Pipeliner abort due to MaxMII too large"); 105 STATISTIC(NumFailZeroMII, "Pipeliner abort due to zero MII"); 106 STATISTIC(NumFailNoSchedule, "Pipeliner abort due to no schedule found"); 107 STATISTIC(NumFailZeroStage, "Pipeliner abort due to zero stage"); 108 STATISTIC(NumFailLargeMaxStage, "Pipeliner abort due to too many stages"); 109 110 /// A command line option to turn software pipelining on or off. 111 static cl::opt<bool> EnableSWP("enable-pipeliner", cl::Hidden, cl::init(true), 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, 150 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> SwpEnableCopyToPhi("pipeliner-enable-copytophi", cl::ReallyHidden, 172 cl::init(true), 173 cl::desc("Enable CopyToPhi DAG Mutation")); 174 175 } // end namespace llvm 176 177 unsigned SwingSchedulerDAG::Circuits::MaxPaths = 5; 178 char MachinePipeliner::ID = 0; 179 #ifndef NDEBUG 180 int MachinePipeliner::NumTries = 0; 181 #endif 182 char &llvm::MachinePipelinerID = MachinePipeliner::ID; 183 184 INITIALIZE_PASS_BEGIN(MachinePipeliner, DEBUG_TYPE, 185 "Modulo Software Pipelining", false, false) 186 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 187 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) 188 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) 189 INITIALIZE_PASS_DEPENDENCY(LiveIntervals) 190 INITIALIZE_PASS_END(MachinePipeliner, DEBUG_TYPE, 191 "Modulo Software Pipelining", false, false) 192 193 /// The "main" function for implementing Swing Modulo Scheduling. 194 bool MachinePipeliner::runOnMachineFunction(MachineFunction &mf) { 195 if (skipFunction(mf.getFunction())) 196 return false; 197 198 if (!EnableSWP) 199 return false; 200 201 if (mf.getFunction().getAttributes().hasFnAttr(Attribute::OptimizeForSize) && 202 !EnableSWPOptSize.getPosition()) 203 return false; 204 205 if (!mf.getSubtarget().enableMachinePipeliner()) 206 return false; 207 208 // Cannot pipeline loops without instruction itineraries if we are using 209 // DFA for the pipeliner. 210 if (mf.getSubtarget().useDFAforSMS() && 211 (!mf.getSubtarget().getInstrItineraryData() || 212 mf.getSubtarget().getInstrItineraryData()->isEmpty())) 213 return false; 214 215 MF = &mf; 216 MLI = &getAnalysis<MachineLoopInfo>(); 217 MDT = &getAnalysis<MachineDominatorTree>(); 218 ORE = &getAnalysis<MachineOptimizationRemarkEmitterPass>().getORE(); 219 TII = MF->getSubtarget().getInstrInfo(); 220 RegClassInfo.runOnMachineFunction(*MF); 221 222 for (const auto &L : *MLI) 223 scheduleLoop(*L); 224 225 return false; 226 } 227 228 /// Attempt to perform the SMS algorithm on the specified loop. This function is 229 /// the main entry point for the algorithm. The function identifies candidate 230 /// loops, calculates the minimum initiation interval, and attempts to schedule 231 /// the loop. 232 bool MachinePipeliner::scheduleLoop(MachineLoop &L) { 233 bool Changed = false; 234 for (const auto &InnerLoop : L) 235 Changed |= scheduleLoop(*InnerLoop); 236 237 #ifndef NDEBUG 238 // Stop trying after reaching the limit (if any). 239 int Limit = SwpLoopLimit; 240 if (Limit >= 0) { 241 if (NumTries >= SwpLoopLimit) 242 return Changed; 243 NumTries++; 244 } 245 #endif 246 247 setPragmaPipelineOptions(L); 248 if (!canPipelineLoop(L)) { 249 LLVM_DEBUG(dbgs() << "\n!!! Can not pipeline loop.\n"); 250 ORE->emit([&]() { 251 return MachineOptimizationRemarkMissed(DEBUG_TYPE, "canPipelineLoop", 252 L.getStartLoc(), L.getHeader()) 253 << "Failed to pipeline loop"; 254 }); 255 256 LI.LoopPipelinerInfo.reset(); 257 return Changed; 258 } 259 260 ++NumTrytoPipeline; 261 262 Changed = swingModuloScheduler(L); 263 264 LI.LoopPipelinerInfo.reset(); 265 return Changed; 266 } 267 268 void MachinePipeliner::setPragmaPipelineOptions(MachineLoop &L) { 269 // Reset the pragma for the next loop in iteration. 270 disabledByPragma = false; 271 II_setByPragma = 0; 272 273 MachineBasicBlock *LBLK = L.getTopBlock(); 274 275 if (LBLK == nullptr) 276 return; 277 278 const BasicBlock *BBLK = LBLK->getBasicBlock(); 279 if (BBLK == nullptr) 280 return; 281 282 const Instruction *TI = BBLK->getTerminator(); 283 if (TI == nullptr) 284 return; 285 286 MDNode *LoopID = TI->getMetadata(LLVMContext::MD_loop); 287 if (LoopID == nullptr) 288 return; 289 290 assert(LoopID->getNumOperands() > 0 && "requires atleast one operand"); 291 assert(LoopID->getOperand(0) == LoopID && "invalid loop"); 292 293 for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) { 294 MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i)); 295 296 if (MD == nullptr) 297 continue; 298 299 MDString *S = dyn_cast<MDString>(MD->getOperand(0)); 300 301 if (S == nullptr) 302 continue; 303 304 if (S->getString() == "llvm.loop.pipeline.initiationinterval") { 305 assert(MD->getNumOperands() == 2 && 306 "Pipeline initiation interval hint metadata should have two operands."); 307 II_setByPragma = 308 mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue(); 309 assert(II_setByPragma >= 1 && "Pipeline initiation interval must be positive."); 310 } else if (S->getString() == "llvm.loop.pipeline.disable") { 311 disabledByPragma = true; 312 } 313 } 314 } 315 316 /// Return true if the loop can be software pipelined. The algorithm is 317 /// restricted to loops with a single basic block. Make sure that the 318 /// branch in the loop can be analyzed. 319 bool MachinePipeliner::canPipelineLoop(MachineLoop &L) { 320 if (L.getNumBlocks() != 1) { 321 ORE->emit([&]() { 322 return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop", 323 L.getStartLoc(), L.getHeader()) 324 << "Not a single basic block: " 325 << ore::NV("NumBlocks", L.getNumBlocks()); 326 }); 327 return false; 328 } 329 330 if (disabledByPragma) { 331 ORE->emit([&]() { 332 return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop", 333 L.getStartLoc(), L.getHeader()) 334 << "Disabled by Pragma."; 335 }); 336 return false; 337 } 338 339 // Check if the branch can't be understood because we can't do pipelining 340 // if that's the case. 341 LI.TBB = nullptr; 342 LI.FBB = nullptr; 343 LI.BrCond.clear(); 344 if (TII->analyzeBranch(*L.getHeader(), LI.TBB, LI.FBB, LI.BrCond)) { 345 LLVM_DEBUG(dbgs() << "Unable to analyzeBranch, can NOT pipeline Loop\n"); 346 NumFailBranch++; 347 ORE->emit([&]() { 348 return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop", 349 L.getStartLoc(), L.getHeader()) 350 << "The branch can't be understood"; 351 }); 352 return false; 353 } 354 355 LI.LoopInductionVar = nullptr; 356 LI.LoopCompare = nullptr; 357 LI.LoopPipelinerInfo = TII->analyzeLoopForPipelining(L.getTopBlock()); 358 if (!LI.LoopPipelinerInfo) { 359 LLVM_DEBUG(dbgs() << "Unable to analyzeLoop, can NOT pipeline Loop\n"); 360 NumFailLoop++; 361 ORE->emit([&]() { 362 return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop", 363 L.getStartLoc(), L.getHeader()) 364 << "The loop structure is not supported"; 365 }); 366 return false; 367 } 368 369 if (!L.getLoopPreheader()) { 370 LLVM_DEBUG(dbgs() << "Preheader not found, can NOT pipeline Loop\n"); 371 NumFailPreheader++; 372 ORE->emit([&]() { 373 return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop", 374 L.getStartLoc(), L.getHeader()) 375 << "No loop preheader found"; 376 }); 377 return false; 378 } 379 380 // Remove any subregisters from inputs to phi nodes. 381 preprocessPhiNodes(*L.getHeader()); 382 return true; 383 } 384 385 void MachinePipeliner::preprocessPhiNodes(MachineBasicBlock &B) { 386 MachineRegisterInfo &MRI = MF->getRegInfo(); 387 SlotIndexes &Slots = *getAnalysis<LiveIntervals>().getSlotIndexes(); 388 389 for (MachineInstr &PI : B.phis()) { 390 MachineOperand &DefOp = PI.getOperand(0); 391 assert(DefOp.getSubReg() == 0); 392 auto *RC = MRI.getRegClass(DefOp.getReg()); 393 394 for (unsigned i = 1, n = PI.getNumOperands(); i != n; i += 2) { 395 MachineOperand &RegOp = PI.getOperand(i); 396 if (RegOp.getSubReg() == 0) 397 continue; 398 399 // If the operand uses a subregister, replace it with a new register 400 // without subregisters, and generate a copy to the new register. 401 Register NewReg = MRI.createVirtualRegister(RC); 402 MachineBasicBlock &PredB = *PI.getOperand(i+1).getMBB(); 403 MachineBasicBlock::iterator At = PredB.getFirstTerminator(); 404 const DebugLoc &DL = PredB.findDebugLoc(At); 405 auto Copy = BuildMI(PredB, At, DL, TII->get(TargetOpcode::COPY), NewReg) 406 .addReg(RegOp.getReg(), getRegState(RegOp), 407 RegOp.getSubReg()); 408 Slots.insertMachineInstrInMaps(*Copy); 409 RegOp.setReg(NewReg); 410 RegOp.setSubReg(0); 411 } 412 } 413 } 414 415 /// The SMS algorithm consists of the following main steps: 416 /// 1. Computation and analysis of the dependence graph. 417 /// 2. Ordering of the nodes (instructions). 418 /// 3. Attempt to Schedule the loop. 419 bool MachinePipeliner::swingModuloScheduler(MachineLoop &L) { 420 assert(L.getBlocks().size() == 1 && "SMS works on single blocks only."); 421 422 SwingSchedulerDAG SMS(*this, L, getAnalysis<LiveIntervals>(), RegClassInfo, 423 II_setByPragma, LI.LoopPipelinerInfo.get()); 424 425 MachineBasicBlock *MBB = L.getHeader(); 426 // The kernel should not include any terminator instructions. These 427 // will be added back later. 428 SMS.startBlock(MBB); 429 430 // Compute the number of 'real' instructions in the basic block by 431 // ignoring terminators. 432 unsigned size = MBB->size(); 433 for (MachineBasicBlock::iterator I = MBB->getFirstTerminator(), 434 E = MBB->instr_end(); 435 I != E; ++I, --size) 436 ; 437 438 SMS.enterRegion(MBB, MBB->begin(), MBB->getFirstTerminator(), size); 439 SMS.schedule(); 440 SMS.exitRegion(); 441 442 SMS.finishBlock(); 443 return SMS.hasNewSchedule(); 444 } 445 446 void MachinePipeliner::getAnalysisUsage(AnalysisUsage &AU) const { 447 AU.addRequired<AAResultsWrapperPass>(); 448 AU.addPreserved<AAResultsWrapperPass>(); 449 AU.addRequired<MachineLoopInfo>(); 450 AU.addRequired<MachineDominatorTree>(); 451 AU.addRequired<LiveIntervals>(); 452 AU.addRequired<MachineOptimizationRemarkEmitterPass>(); 453 MachineFunctionPass::getAnalysisUsage(AU); 454 } 455 456 void SwingSchedulerDAG::setMII(unsigned ResMII, unsigned RecMII) { 457 if (II_setByPragma > 0) 458 MII = II_setByPragma; 459 else 460 MII = std::max(ResMII, RecMII); 461 } 462 463 void SwingSchedulerDAG::setMAX_II() { 464 if (II_setByPragma > 0) 465 MAX_II = II_setByPragma; 466 else 467 MAX_II = MII + 10; 468 } 469 470 /// We override the schedule function in ScheduleDAGInstrs to implement the 471 /// scheduling part of the Swing Modulo Scheduling algorithm. 472 void SwingSchedulerDAG::schedule() { 473 AliasAnalysis *AA = &Pass.getAnalysis<AAResultsWrapperPass>().getAAResults(); 474 buildSchedGraph(AA); 475 addLoopCarriedDependences(AA); 476 updatePhiDependences(); 477 Topo.InitDAGTopologicalSorting(); 478 changeDependences(); 479 postprocessDAG(); 480 LLVM_DEBUG(dump()); 481 482 NodeSetType NodeSets; 483 findCircuits(NodeSets); 484 NodeSetType Circuits = NodeSets; 485 486 // Calculate the MII. 487 unsigned ResMII = calculateResMII(); 488 unsigned RecMII = calculateRecMII(NodeSets); 489 490 fuseRecs(NodeSets); 491 492 // This flag is used for testing and can cause correctness problems. 493 if (SwpIgnoreRecMII) 494 RecMII = 0; 495 496 setMII(ResMII, RecMII); 497 setMAX_II(); 498 499 LLVM_DEBUG(dbgs() << "MII = " << MII << " MAX_II = " << MAX_II 500 << " (rec=" << RecMII << ", res=" << ResMII << ")\n"); 501 502 // Can't schedule a loop without a valid MII. 503 if (MII == 0) { 504 LLVM_DEBUG(dbgs() << "Invalid Minimal Initiation Interval: 0\n"); 505 NumFailZeroMII++; 506 Pass.ORE->emit([&]() { 507 return MachineOptimizationRemarkAnalysis( 508 DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader()) 509 << "Invalid Minimal Initiation Interval: 0"; 510 }); 511 return; 512 } 513 514 // Don't pipeline large loops. 515 if (SwpMaxMii != -1 && (int)MII > SwpMaxMii) { 516 LLVM_DEBUG(dbgs() << "MII > " << SwpMaxMii 517 << ", we don't pipeline large loops\n"); 518 NumFailLargeMaxMII++; 519 Pass.ORE->emit([&]() { 520 return MachineOptimizationRemarkAnalysis( 521 DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader()) 522 << "Minimal Initiation Interval too large: " 523 << ore::NV("MII", (int)MII) << " > " 524 << ore::NV("SwpMaxMii", SwpMaxMii) << "." 525 << "Refer to -pipeliner-max-mii."; 526 }); 527 return; 528 } 529 530 computeNodeFunctions(NodeSets); 531 532 registerPressureFilter(NodeSets); 533 534 colocateNodeSets(NodeSets); 535 536 checkNodeSets(NodeSets); 537 538 LLVM_DEBUG({ 539 for (auto &I : NodeSets) { 540 dbgs() << " Rec NodeSet "; 541 I.dump(); 542 } 543 }); 544 545 llvm::stable_sort(NodeSets, std::greater<NodeSet>()); 546 547 groupRemainingNodes(NodeSets); 548 549 removeDuplicateNodes(NodeSets); 550 551 LLVM_DEBUG({ 552 for (auto &I : NodeSets) { 553 dbgs() << " NodeSet "; 554 I.dump(); 555 } 556 }); 557 558 computeNodeOrder(NodeSets); 559 560 // check for node order issues 561 checkValidNodeOrder(Circuits); 562 563 SMSchedule Schedule(Pass.MF); 564 Scheduled = schedulePipeline(Schedule); 565 566 if (!Scheduled){ 567 LLVM_DEBUG(dbgs() << "No schedule found, return\n"); 568 NumFailNoSchedule++; 569 Pass.ORE->emit([&]() { 570 return MachineOptimizationRemarkAnalysis( 571 DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader()) 572 << "Unable to find schedule"; 573 }); 574 return; 575 } 576 577 unsigned numStages = Schedule.getMaxStageCount(); 578 // No need to generate pipeline if there are no overlapped iterations. 579 if (numStages == 0) { 580 LLVM_DEBUG(dbgs() << "No overlapped iterations, skip.\n"); 581 NumFailZeroStage++; 582 Pass.ORE->emit([&]() { 583 return MachineOptimizationRemarkAnalysis( 584 DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader()) 585 << "No need to pipeline - no overlapped iterations in schedule."; 586 }); 587 return; 588 } 589 // Check that the maximum stage count is less than user-defined limit. 590 if (SwpMaxStages > -1 && (int)numStages > SwpMaxStages) { 591 LLVM_DEBUG(dbgs() << "numStages:" << numStages << ">" << SwpMaxStages 592 << " : too many stages, abort\n"); 593 NumFailLargeMaxStage++; 594 Pass.ORE->emit([&]() { 595 return MachineOptimizationRemarkAnalysis( 596 DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader()) 597 << "Too many stages in schedule: " 598 << ore::NV("numStages", (int)numStages) << " > " 599 << ore::NV("SwpMaxStages", SwpMaxStages) 600 << ". Refer to -pipeliner-max-stages."; 601 }); 602 return; 603 } 604 605 Pass.ORE->emit([&]() { 606 return MachineOptimizationRemark(DEBUG_TYPE, "schedule", Loop.getStartLoc(), 607 Loop.getHeader()) 608 << "Pipelined succesfully!"; 609 }); 610 611 // Generate the schedule as a ModuloSchedule. 612 DenseMap<MachineInstr *, int> Cycles, Stages; 613 std::vector<MachineInstr *> OrderedInsts; 614 for (int Cycle = Schedule.getFirstCycle(); Cycle <= Schedule.getFinalCycle(); 615 ++Cycle) { 616 for (SUnit *SU : Schedule.getInstructions(Cycle)) { 617 OrderedInsts.push_back(SU->getInstr()); 618 Cycles[SU->getInstr()] = Cycle; 619 Stages[SU->getInstr()] = Schedule.stageScheduled(SU); 620 } 621 } 622 DenseMap<MachineInstr *, std::pair<unsigned, int64_t>> NewInstrChanges; 623 for (auto &KV : NewMIs) { 624 Cycles[KV.first] = Cycles[KV.second]; 625 Stages[KV.first] = Stages[KV.second]; 626 NewInstrChanges[KV.first] = InstrChanges[getSUnit(KV.first)]; 627 } 628 629 ModuloSchedule MS(MF, &Loop, std::move(OrderedInsts), std::move(Cycles), 630 std::move(Stages)); 631 if (EmitTestAnnotations) { 632 assert(NewInstrChanges.empty() && 633 "Cannot serialize a schedule with InstrChanges!"); 634 ModuloScheduleTestAnnotater MSTI(MF, MS); 635 MSTI.annotate(); 636 return; 637 } 638 // The experimental code generator can't work if there are InstChanges. 639 if (ExperimentalCodeGen && NewInstrChanges.empty()) { 640 PeelingModuloScheduleExpander MSE(MF, MS, &LIS); 641 MSE.expand(); 642 } else { 643 ModuloScheduleExpander MSE(MF, MS, LIS, std::move(NewInstrChanges)); 644 MSE.expand(); 645 MSE.cleanup(); 646 } 647 ++NumPipelined; 648 } 649 650 /// Clean up after the software pipeliner runs. 651 void SwingSchedulerDAG::finishBlock() { 652 for (auto &KV : NewMIs) 653 MF.deleteMachineInstr(KV.second); 654 NewMIs.clear(); 655 656 // Call the superclass. 657 ScheduleDAGInstrs::finishBlock(); 658 } 659 660 /// Return the register values for the operands of a Phi instruction. 661 /// This function assume the instruction is a Phi. 662 static void getPhiRegs(MachineInstr &Phi, MachineBasicBlock *Loop, 663 unsigned &InitVal, unsigned &LoopVal) { 664 assert(Phi.isPHI() && "Expecting a Phi."); 665 666 InitVal = 0; 667 LoopVal = 0; 668 for (unsigned i = 1, e = Phi.getNumOperands(); i != e; i += 2) 669 if (Phi.getOperand(i + 1).getMBB() != Loop) 670 InitVal = Phi.getOperand(i).getReg(); 671 else 672 LoopVal = Phi.getOperand(i).getReg(); 673 674 assert(InitVal != 0 && LoopVal != 0 && "Unexpected Phi structure."); 675 } 676 677 /// Return the Phi register value that comes the loop block. 678 static unsigned getLoopPhiReg(MachineInstr &Phi, MachineBasicBlock *LoopBB) { 679 for (unsigned i = 1, e = Phi.getNumOperands(); i != e; i += 2) 680 if (Phi.getOperand(i + 1).getMBB() == LoopBB) 681 return Phi.getOperand(i).getReg(); 682 return 0; 683 } 684 685 /// Return true if SUb can be reached from SUa following the chain edges. 686 static bool isSuccOrder(SUnit *SUa, SUnit *SUb) { 687 SmallPtrSet<SUnit *, 8> Visited; 688 SmallVector<SUnit *, 8> Worklist; 689 Worklist.push_back(SUa); 690 while (!Worklist.empty()) { 691 const SUnit *SU = Worklist.pop_back_val(); 692 for (const auto &SI : SU->Succs) { 693 SUnit *SuccSU = SI.getSUnit(); 694 if (SI.getKind() == SDep::Order) { 695 if (Visited.count(SuccSU)) 696 continue; 697 if (SuccSU == SUb) 698 return true; 699 Worklist.push_back(SuccSU); 700 Visited.insert(SuccSU); 701 } 702 } 703 } 704 return false; 705 } 706 707 /// Return true if the instruction causes a chain between memory 708 /// references before and after it. 709 static bool isDependenceBarrier(MachineInstr &MI) { 710 return MI.isCall() || MI.mayRaiseFPException() || 711 MI.hasUnmodeledSideEffects() || 712 (MI.hasOrderedMemoryRef() && 713 (!MI.mayLoad() || !MI.isDereferenceableInvariantLoad())); 714 } 715 716 /// Return the underlying objects for the memory references of an instruction. 717 /// This function calls the code in ValueTracking, but first checks that the 718 /// instruction has a memory operand. 719 static void getUnderlyingObjects(const MachineInstr *MI, 720 SmallVectorImpl<const Value *> &Objs) { 721 if (!MI->hasOneMemOperand()) 722 return; 723 MachineMemOperand *MM = *MI->memoperands_begin(); 724 if (!MM->getValue()) 725 return; 726 getUnderlyingObjects(MM->getValue(), Objs); 727 for (const Value *V : Objs) { 728 if (!isIdentifiedObject(V)) { 729 Objs.clear(); 730 return; 731 } 732 Objs.push_back(V); 733 } 734 } 735 736 /// Add a chain edge between a load and store if the store can be an 737 /// alias of the load on a subsequent iteration, i.e., a loop carried 738 /// dependence. This code is very similar to the code in ScheduleDAGInstrs 739 /// but that code doesn't create loop carried dependences. 740 void SwingSchedulerDAG::addLoopCarriedDependences(AliasAnalysis *AA) { 741 MapVector<const Value *, SmallVector<SUnit *, 4>> PendingLoads; 742 Value *UnknownValue = 743 UndefValue::get(Type::getVoidTy(MF.getFunction().getContext())); 744 for (auto &SU : SUnits) { 745 MachineInstr &MI = *SU.getInstr(); 746 if (isDependenceBarrier(MI)) 747 PendingLoads.clear(); 748 else if (MI.mayLoad()) { 749 SmallVector<const Value *, 4> Objs; 750 ::getUnderlyingObjects(&MI, Objs); 751 if (Objs.empty()) 752 Objs.push_back(UnknownValue); 753 for (const auto *V : Objs) { 754 SmallVector<SUnit *, 4> &SUs = PendingLoads[V]; 755 SUs.push_back(&SU); 756 } 757 } else if (MI.mayStore()) { 758 SmallVector<const Value *, 4> Objs; 759 ::getUnderlyingObjects(&MI, Objs); 760 if (Objs.empty()) 761 Objs.push_back(UnknownValue); 762 for (const auto *V : Objs) { 763 MapVector<const Value *, SmallVector<SUnit *, 4>>::iterator I = 764 PendingLoads.find(V); 765 if (I == PendingLoads.end()) 766 continue; 767 for (auto *Load : I->second) { 768 if (isSuccOrder(Load, &SU)) 769 continue; 770 MachineInstr &LdMI = *Load->getInstr(); 771 // First, perform the cheaper check that compares the base register. 772 // If they are the same and the load offset is less than the store 773 // offset, then mark the dependence as loop carried potentially. 774 const MachineOperand *BaseOp1, *BaseOp2; 775 int64_t Offset1, Offset2; 776 bool Offset1IsScalable, Offset2IsScalable; 777 if (TII->getMemOperandWithOffset(LdMI, BaseOp1, Offset1, 778 Offset1IsScalable, TRI) && 779 TII->getMemOperandWithOffset(MI, BaseOp2, Offset2, 780 Offset2IsScalable, TRI)) { 781 if (BaseOp1->isIdenticalTo(*BaseOp2) && 782 Offset1IsScalable == Offset2IsScalable && 783 (int)Offset1 < (int)Offset2) { 784 assert(TII->areMemAccessesTriviallyDisjoint(LdMI, MI) && 785 "What happened to the chain edge?"); 786 SDep Dep(Load, SDep::Barrier); 787 Dep.setLatency(1); 788 SU.addPred(Dep); 789 continue; 790 } 791 } 792 // Second, the more expensive check that uses alias analysis on the 793 // base registers. If they alias, and the load offset is less than 794 // the store offset, the mark the dependence as loop carried. 795 if (!AA) { 796 SDep Dep(Load, SDep::Barrier); 797 Dep.setLatency(1); 798 SU.addPred(Dep); 799 continue; 800 } 801 MachineMemOperand *MMO1 = *LdMI.memoperands_begin(); 802 MachineMemOperand *MMO2 = *MI.memoperands_begin(); 803 if (!MMO1->getValue() || !MMO2->getValue()) { 804 SDep Dep(Load, SDep::Barrier); 805 Dep.setLatency(1); 806 SU.addPred(Dep); 807 continue; 808 } 809 if (MMO1->getValue() == MMO2->getValue() && 810 MMO1->getOffset() <= MMO2->getOffset()) { 811 SDep Dep(Load, SDep::Barrier); 812 Dep.setLatency(1); 813 SU.addPred(Dep); 814 continue; 815 } 816 if (!AA->isNoAlias( 817 MemoryLocation::getAfter(MMO1->getValue(), MMO1->getAAInfo()), 818 MemoryLocation::getAfter(MMO2->getValue(), 819 MMO2->getAAInfo()))) { 820 SDep Dep(Load, SDep::Barrier); 821 Dep.setLatency(1); 822 SU.addPred(Dep); 823 } 824 } 825 } 826 } 827 } 828 } 829 830 /// Update the phi dependences to the DAG because ScheduleDAGInstrs no longer 831 /// processes dependences for PHIs. This function adds true dependences 832 /// from a PHI to a use, and a loop carried dependence from the use to the 833 /// PHI. The loop carried dependence is represented as an anti dependence 834 /// edge. This function also removes chain dependences between unrelated 835 /// PHIs. 836 void SwingSchedulerDAG::updatePhiDependences() { 837 SmallVector<SDep, 4> RemoveDeps; 838 const TargetSubtargetInfo &ST = MF.getSubtarget<TargetSubtargetInfo>(); 839 840 // Iterate over each DAG node. 841 for (SUnit &I : SUnits) { 842 RemoveDeps.clear(); 843 // Set to true if the instruction has an operand defined by a Phi. 844 unsigned HasPhiUse = 0; 845 unsigned HasPhiDef = 0; 846 MachineInstr *MI = I.getInstr(); 847 // Iterate over each operand, and we process the definitions. 848 for (MachineInstr::mop_iterator MOI = MI->operands_begin(), 849 MOE = MI->operands_end(); 850 MOI != MOE; ++MOI) { 851 if (!MOI->isReg()) 852 continue; 853 Register Reg = MOI->getReg(); 854 if (MOI->isDef()) { 855 // If the register is used by a Phi, then create an anti dependence. 856 for (MachineRegisterInfo::use_instr_iterator 857 UI = MRI.use_instr_begin(Reg), 858 UE = MRI.use_instr_end(); 859 UI != UE; ++UI) { 860 MachineInstr *UseMI = &*UI; 861 SUnit *SU = getSUnit(UseMI); 862 if (SU != nullptr && UseMI->isPHI()) { 863 if (!MI->isPHI()) { 864 SDep Dep(SU, SDep::Anti, Reg); 865 Dep.setLatency(1); 866 I.addPred(Dep); 867 } else { 868 HasPhiDef = Reg; 869 // Add a chain edge to a dependent Phi that isn't an existing 870 // predecessor. 871 if (SU->NodeNum < I.NodeNum && !I.isPred(SU)) 872 I.addPred(SDep(SU, SDep::Barrier)); 873 } 874 } 875 } 876 } else if (MOI->isUse()) { 877 // If the register is defined by a Phi, then create a true dependence. 878 MachineInstr *DefMI = MRI.getUniqueVRegDef(Reg); 879 if (DefMI == nullptr) 880 continue; 881 SUnit *SU = getSUnit(DefMI); 882 if (SU != nullptr && DefMI->isPHI()) { 883 if (!MI->isPHI()) { 884 SDep Dep(SU, SDep::Data, Reg); 885 Dep.setLatency(0); 886 ST.adjustSchedDependency(SU, 0, &I, MI->getOperandNo(MOI), Dep); 887 I.addPred(Dep); 888 } else { 889 HasPhiUse = Reg; 890 // Add a chain edge to a dependent Phi that isn't an existing 891 // predecessor. 892 if (SU->NodeNum < I.NodeNum && !I.isPred(SU)) 893 I.addPred(SDep(SU, SDep::Barrier)); 894 } 895 } 896 } 897 } 898 // Remove order dependences from an unrelated Phi. 899 if (!SwpPruneDeps) 900 continue; 901 for (auto &PI : I.Preds) { 902 MachineInstr *PMI = PI.getSUnit()->getInstr(); 903 if (PMI->isPHI() && PI.getKind() == SDep::Order) { 904 if (I.getInstr()->isPHI()) { 905 if (PMI->getOperand(0).getReg() == HasPhiUse) 906 continue; 907 if (getLoopPhiReg(*PMI, PMI->getParent()) == HasPhiDef) 908 continue; 909 } 910 RemoveDeps.push_back(PI); 911 } 912 } 913 for (int i = 0, e = RemoveDeps.size(); i != e; ++i) 914 I.removePred(RemoveDeps[i]); 915 } 916 } 917 918 /// Iterate over each DAG node and see if we can change any dependences 919 /// in order to reduce the recurrence MII. 920 void SwingSchedulerDAG::changeDependences() { 921 // See if an instruction can use a value from the previous iteration. 922 // If so, we update the base and offset of the instruction and change 923 // the dependences. 924 for (SUnit &I : SUnits) { 925 unsigned BasePos = 0, OffsetPos = 0, NewBase = 0; 926 int64_t NewOffset = 0; 927 if (!canUseLastOffsetValue(I.getInstr(), BasePos, OffsetPos, NewBase, 928 NewOffset)) 929 continue; 930 931 // Get the MI and SUnit for the instruction that defines the original base. 932 Register OrigBase = I.getInstr()->getOperand(BasePos).getReg(); 933 MachineInstr *DefMI = MRI.getUniqueVRegDef(OrigBase); 934 if (!DefMI) 935 continue; 936 SUnit *DefSU = getSUnit(DefMI); 937 if (!DefSU) 938 continue; 939 // Get the MI and SUnit for the instruction that defins the new base. 940 MachineInstr *LastMI = MRI.getUniqueVRegDef(NewBase); 941 if (!LastMI) 942 continue; 943 SUnit *LastSU = getSUnit(LastMI); 944 if (!LastSU) 945 continue; 946 947 if (Topo.IsReachable(&I, LastSU)) 948 continue; 949 950 // Remove the dependence. The value now depends on a prior iteration. 951 SmallVector<SDep, 4> Deps; 952 for (const SDep &P : I.Preds) 953 if (P.getSUnit() == DefSU) 954 Deps.push_back(P); 955 for (int i = 0, e = Deps.size(); i != e; i++) { 956 Topo.RemovePred(&I, Deps[i].getSUnit()); 957 I.removePred(Deps[i]); 958 } 959 // Remove the chain dependence between the instructions. 960 Deps.clear(); 961 for (auto &P : LastSU->Preds) 962 if (P.getSUnit() == &I && P.getKind() == SDep::Order) 963 Deps.push_back(P); 964 for (int i = 0, e = Deps.size(); i != e; i++) { 965 Topo.RemovePred(LastSU, Deps[i].getSUnit()); 966 LastSU->removePred(Deps[i]); 967 } 968 969 // Add a dependence between the new instruction and the instruction 970 // that defines the new base. 971 SDep Dep(&I, SDep::Anti, NewBase); 972 Topo.AddPred(LastSU, &I); 973 LastSU->addPred(Dep); 974 975 // Remember the base and offset information so that we can update the 976 // instruction during code generation. 977 InstrChanges[&I] = std::make_pair(NewBase, NewOffset); 978 } 979 } 980 981 namespace { 982 983 // FuncUnitSorter - Comparison operator used to sort instructions by 984 // the number of functional unit choices. 985 struct FuncUnitSorter { 986 const InstrItineraryData *InstrItins; 987 const MCSubtargetInfo *STI; 988 DenseMap<InstrStage::FuncUnits, unsigned> Resources; 989 990 FuncUnitSorter(const TargetSubtargetInfo &TSI) 991 : InstrItins(TSI.getInstrItineraryData()), STI(&TSI) {} 992 993 // Compute the number of functional unit alternatives needed 994 // at each stage, and take the minimum value. We prioritize the 995 // instructions by the least number of choices first. 996 unsigned minFuncUnits(const MachineInstr *Inst, 997 InstrStage::FuncUnits &F) const { 998 unsigned SchedClass = Inst->getDesc().getSchedClass(); 999 unsigned min = UINT_MAX; 1000 if (InstrItins && !InstrItins->isEmpty()) { 1001 for (const InstrStage &IS : 1002 make_range(InstrItins->beginStage(SchedClass), 1003 InstrItins->endStage(SchedClass))) { 1004 InstrStage::FuncUnits funcUnits = IS.getUnits(); 1005 unsigned numAlternatives = countPopulation(funcUnits); 1006 if (numAlternatives < min) { 1007 min = numAlternatives; 1008 F = funcUnits; 1009 } 1010 } 1011 return min; 1012 } 1013 if (STI && STI->getSchedModel().hasInstrSchedModel()) { 1014 const MCSchedClassDesc *SCDesc = 1015 STI->getSchedModel().getSchedClassDesc(SchedClass); 1016 if (!SCDesc->isValid()) 1017 // No valid Schedule Class Desc for schedClass, should be 1018 // Pseudo/PostRAPseudo 1019 return min; 1020 1021 for (const MCWriteProcResEntry &PRE : 1022 make_range(STI->getWriteProcResBegin(SCDesc), 1023 STI->getWriteProcResEnd(SCDesc))) { 1024 if (!PRE.Cycles) 1025 continue; 1026 const MCProcResourceDesc *ProcResource = 1027 STI->getSchedModel().getProcResource(PRE.ProcResourceIdx); 1028 unsigned NumUnits = ProcResource->NumUnits; 1029 if (NumUnits < min) { 1030 min = NumUnits; 1031 F = PRE.ProcResourceIdx; 1032 } 1033 } 1034 return min; 1035 } 1036 llvm_unreachable("Should have non-empty InstrItins or hasInstrSchedModel!"); 1037 } 1038 1039 // Compute the critical resources needed by the instruction. This 1040 // function records the functional units needed by instructions that 1041 // must use only one functional unit. We use this as a tie breaker 1042 // for computing the resource MII. The instrutions that require 1043 // the same, highly used, functional unit have high priority. 1044 void calcCriticalResources(MachineInstr &MI) { 1045 unsigned SchedClass = MI.getDesc().getSchedClass(); 1046 if (InstrItins && !InstrItins->isEmpty()) { 1047 for (const InstrStage &IS : 1048 make_range(InstrItins->beginStage(SchedClass), 1049 InstrItins->endStage(SchedClass))) { 1050 InstrStage::FuncUnits FuncUnits = IS.getUnits(); 1051 if (countPopulation(FuncUnits) == 1) 1052 Resources[FuncUnits]++; 1053 } 1054 return; 1055 } 1056 if (STI && STI->getSchedModel().hasInstrSchedModel()) { 1057 const MCSchedClassDesc *SCDesc = 1058 STI->getSchedModel().getSchedClassDesc(SchedClass); 1059 if (!SCDesc->isValid()) 1060 // No valid Schedule Class Desc for schedClass, should be 1061 // Pseudo/PostRAPseudo 1062 return; 1063 1064 for (const MCWriteProcResEntry &PRE : 1065 make_range(STI->getWriteProcResBegin(SCDesc), 1066 STI->getWriteProcResEnd(SCDesc))) { 1067 if (!PRE.Cycles) 1068 continue; 1069 Resources[PRE.ProcResourceIdx]++; 1070 } 1071 return; 1072 } 1073 llvm_unreachable("Should have non-empty InstrItins or hasInstrSchedModel!"); 1074 } 1075 1076 /// Return true if IS1 has less priority than IS2. 1077 bool operator()(const MachineInstr *IS1, const MachineInstr *IS2) const { 1078 InstrStage::FuncUnits F1 = 0, F2 = 0; 1079 unsigned MFUs1 = minFuncUnits(IS1, F1); 1080 unsigned MFUs2 = minFuncUnits(IS2, F2); 1081 if (MFUs1 == MFUs2) 1082 return Resources.lookup(F1) < Resources.lookup(F2); 1083 return MFUs1 > MFUs2; 1084 } 1085 }; 1086 1087 } // end anonymous namespace 1088 1089 /// Calculate the resource constrained minimum initiation interval for the 1090 /// specified loop. We use the DFA to model the resources needed for 1091 /// each instruction, and we ignore dependences. A different DFA is created 1092 /// for each cycle that is required. When adding a new instruction, we attempt 1093 /// to add it to each existing DFA, until a legal space is found. If the 1094 /// instruction cannot be reserved in an existing DFA, we create a new one. 1095 unsigned SwingSchedulerDAG::calculateResMII() { 1096 1097 LLVM_DEBUG(dbgs() << "calculateResMII:\n"); 1098 SmallVector<ResourceManager*, 8> Resources; 1099 MachineBasicBlock *MBB = Loop.getHeader(); 1100 Resources.push_back(new ResourceManager(&MF.getSubtarget())); 1101 1102 // Sort the instructions by the number of available choices for scheduling, 1103 // least to most. Use the number of critical resources as the tie breaker. 1104 FuncUnitSorter FUS = FuncUnitSorter(MF.getSubtarget()); 1105 for (MachineInstr &MI : 1106 llvm::make_range(MBB->getFirstNonPHI(), MBB->getFirstTerminator())) 1107 FUS.calcCriticalResources(MI); 1108 PriorityQueue<MachineInstr *, std::vector<MachineInstr *>, FuncUnitSorter> 1109 FuncUnitOrder(FUS); 1110 1111 for (MachineInstr &MI : 1112 llvm::make_range(MBB->getFirstNonPHI(), MBB->getFirstTerminator())) 1113 FuncUnitOrder.push(&MI); 1114 1115 while (!FuncUnitOrder.empty()) { 1116 MachineInstr *MI = FuncUnitOrder.top(); 1117 FuncUnitOrder.pop(); 1118 if (TII->isZeroCost(MI->getOpcode())) 1119 continue; 1120 // Attempt to reserve the instruction in an existing DFA. At least one 1121 // DFA is needed for each cycle. 1122 unsigned NumCycles = getSUnit(MI)->Latency; 1123 unsigned ReservedCycles = 0; 1124 SmallVectorImpl<ResourceManager *>::iterator RI = Resources.begin(); 1125 SmallVectorImpl<ResourceManager *>::iterator RE = Resources.end(); 1126 LLVM_DEBUG({ 1127 dbgs() << "Trying to reserve resource for " << NumCycles 1128 << " cycles for \n"; 1129 MI->dump(); 1130 }); 1131 for (unsigned C = 0; C < NumCycles; ++C) 1132 while (RI != RE) { 1133 if ((*RI)->canReserveResources(*MI)) { 1134 (*RI)->reserveResources(*MI); 1135 ++ReservedCycles; 1136 break; 1137 } 1138 RI++; 1139 } 1140 LLVM_DEBUG(dbgs() << "ReservedCycles:" << ReservedCycles 1141 << ", NumCycles:" << NumCycles << "\n"); 1142 // Add new DFAs, if needed, to reserve resources. 1143 for (unsigned C = ReservedCycles; C < NumCycles; ++C) { 1144 LLVM_DEBUG(if (SwpDebugResource) dbgs() 1145 << "NewResource created to reserve resources" 1146 << "\n"); 1147 ResourceManager *NewResource = new ResourceManager(&MF.getSubtarget()); 1148 assert(NewResource->canReserveResources(*MI) && "Reserve error."); 1149 NewResource->reserveResources(*MI); 1150 Resources.push_back(NewResource); 1151 } 1152 } 1153 int Resmii = Resources.size(); 1154 LLVM_DEBUG(dbgs() << "Return Res MII:" << Resmii << "\n"); 1155 // Delete the memory for each of the DFAs that were created earlier. 1156 for (ResourceManager *RI : Resources) { 1157 ResourceManager *D = RI; 1158 delete D; 1159 } 1160 Resources.clear(); 1161 return Resmii; 1162 } 1163 1164 /// Calculate the recurrence-constrainted minimum initiation interval. 1165 /// Iterate over each circuit. Compute the delay(c) and distance(c) 1166 /// for each circuit. The II needs to satisfy the inequality 1167 /// delay(c) - II*distance(c) <= 0. For each circuit, choose the smallest 1168 /// II that satisfies the inequality, and the RecMII is the maximum 1169 /// of those values. 1170 unsigned SwingSchedulerDAG::calculateRecMII(NodeSetType &NodeSets) { 1171 unsigned RecMII = 0; 1172 1173 for (NodeSet &Nodes : NodeSets) { 1174 if (Nodes.empty()) 1175 continue; 1176 1177 unsigned Delay = Nodes.getLatency(); 1178 unsigned Distance = 1; 1179 1180 // ii = ceil(delay / distance) 1181 unsigned CurMII = (Delay + Distance - 1) / Distance; 1182 Nodes.setRecMII(CurMII); 1183 if (CurMII > RecMII) 1184 RecMII = CurMII; 1185 } 1186 1187 return RecMII; 1188 } 1189 1190 /// Swap all the anti dependences in the DAG. That means it is no longer a DAG, 1191 /// but we do this to find the circuits, and then change them back. 1192 static void swapAntiDependences(std::vector<SUnit> &SUnits) { 1193 SmallVector<std::pair<SUnit *, SDep>, 8> DepsAdded; 1194 for (SUnit &SU : SUnits) { 1195 for (SDep &Pred : SU.Preds) 1196 if (Pred.getKind() == SDep::Anti) 1197 DepsAdded.push_back(std::make_pair(&SU, Pred)); 1198 } 1199 for (std::pair<SUnit *, SDep> &P : DepsAdded) { 1200 // Remove this anti dependency and add one in the reverse direction. 1201 SUnit *SU = P.first; 1202 SDep &D = P.second; 1203 SUnit *TargetSU = D.getSUnit(); 1204 unsigned Reg = D.getReg(); 1205 unsigned Lat = D.getLatency(); 1206 SU->removePred(D); 1207 SDep Dep(SU, SDep::Anti, Reg); 1208 Dep.setLatency(Lat); 1209 TargetSU->addPred(Dep); 1210 } 1211 } 1212 1213 /// Create the adjacency structure of the nodes in the graph. 1214 void SwingSchedulerDAG::Circuits::createAdjacencyStructure( 1215 SwingSchedulerDAG *DAG) { 1216 BitVector Added(SUnits.size()); 1217 DenseMap<int, int> OutputDeps; 1218 for (int i = 0, e = SUnits.size(); i != e; ++i) { 1219 Added.reset(); 1220 // Add any successor to the adjacency matrix and exclude duplicates. 1221 for (auto &SI : SUnits[i].Succs) { 1222 // Only create a back-edge on the first and last nodes of a dependence 1223 // chain. This records any chains and adds them later. 1224 if (SI.getKind() == SDep::Output) { 1225 int N = SI.getSUnit()->NodeNum; 1226 int BackEdge = i; 1227 auto Dep = OutputDeps.find(BackEdge); 1228 if (Dep != OutputDeps.end()) { 1229 BackEdge = Dep->second; 1230 OutputDeps.erase(Dep); 1231 } 1232 OutputDeps[N] = BackEdge; 1233 } 1234 // Do not process a boundary node, an artificial node. 1235 // A back-edge is processed only if it goes to a Phi. 1236 if (SI.getSUnit()->isBoundaryNode() || SI.isArtificial() || 1237 (SI.getKind() == SDep::Anti && !SI.getSUnit()->getInstr()->isPHI())) 1238 continue; 1239 int N = SI.getSUnit()->NodeNum; 1240 if (!Added.test(N)) { 1241 AdjK[i].push_back(N); 1242 Added.set(N); 1243 } 1244 } 1245 // A chain edge between a store and a load is treated as a back-edge in the 1246 // adjacency matrix. 1247 for (auto &PI : SUnits[i].Preds) { 1248 if (!SUnits[i].getInstr()->mayStore() || 1249 !DAG->isLoopCarriedDep(&SUnits[i], PI, false)) 1250 continue; 1251 if (PI.getKind() == SDep::Order && PI.getSUnit()->getInstr()->mayLoad()) { 1252 int N = PI.getSUnit()->NodeNum; 1253 if (!Added.test(N)) { 1254 AdjK[i].push_back(N); 1255 Added.set(N); 1256 } 1257 } 1258 } 1259 } 1260 // Add back-edges in the adjacency matrix for the output dependences. 1261 for (auto &OD : OutputDeps) 1262 if (!Added.test(OD.second)) { 1263 AdjK[OD.first].push_back(OD.second); 1264 Added.set(OD.second); 1265 } 1266 } 1267 1268 /// Identify an elementary circuit in the dependence graph starting at the 1269 /// specified node. 1270 bool SwingSchedulerDAG::Circuits::circuit(int V, int S, NodeSetType &NodeSets, 1271 bool HasBackedge) { 1272 SUnit *SV = &SUnits[V]; 1273 bool F = false; 1274 Stack.insert(SV); 1275 Blocked.set(V); 1276 1277 for (auto W : AdjK[V]) { 1278 if (NumPaths > MaxPaths) 1279 break; 1280 if (W < S) 1281 continue; 1282 if (W == S) { 1283 if (!HasBackedge) 1284 NodeSets.push_back(NodeSet(Stack.begin(), Stack.end())); 1285 F = true; 1286 ++NumPaths; 1287 break; 1288 } else if (!Blocked.test(W)) { 1289 if (circuit(W, S, NodeSets, 1290 Node2Idx->at(W) < Node2Idx->at(V) ? true : HasBackedge)) 1291 F = true; 1292 } 1293 } 1294 1295 if (F) 1296 unblock(V); 1297 else { 1298 for (auto W : AdjK[V]) { 1299 if (W < S) 1300 continue; 1301 B[W].insert(SV); 1302 } 1303 } 1304 Stack.pop_back(); 1305 return F; 1306 } 1307 1308 /// Unblock a node in the circuit finding algorithm. 1309 void SwingSchedulerDAG::Circuits::unblock(int U) { 1310 Blocked.reset(U); 1311 SmallPtrSet<SUnit *, 4> &BU = B[U]; 1312 while (!BU.empty()) { 1313 SmallPtrSet<SUnit *, 4>::iterator SI = BU.begin(); 1314 assert(SI != BU.end() && "Invalid B set."); 1315 SUnit *W = *SI; 1316 BU.erase(W); 1317 if (Blocked.test(W->NodeNum)) 1318 unblock(W->NodeNum); 1319 } 1320 } 1321 1322 /// Identify all the elementary circuits in the dependence graph using 1323 /// Johnson's circuit algorithm. 1324 void SwingSchedulerDAG::findCircuits(NodeSetType &NodeSets) { 1325 // Swap all the anti dependences in the DAG. That means it is no longer a DAG, 1326 // but we do this to find the circuits, and then change them back. 1327 swapAntiDependences(SUnits); 1328 1329 Circuits Cir(SUnits, Topo); 1330 // Create the adjacency structure. 1331 Cir.createAdjacencyStructure(this); 1332 for (int i = 0, e = SUnits.size(); i != e; ++i) { 1333 Cir.reset(); 1334 Cir.circuit(i, i, NodeSets); 1335 } 1336 1337 // Change the dependences back so that we've created a DAG again. 1338 swapAntiDependences(SUnits); 1339 } 1340 1341 // Create artificial dependencies between the source of COPY/REG_SEQUENCE that 1342 // is loop-carried to the USE in next iteration. This will help pipeliner avoid 1343 // additional copies that are needed across iterations. An artificial dependence 1344 // edge is added from USE to SOURCE of COPY/REG_SEQUENCE. 1345 1346 // PHI-------Anti-Dep-----> COPY/REG_SEQUENCE (loop-carried) 1347 // SRCOfCopY------True-Dep---> COPY/REG_SEQUENCE 1348 // PHI-------True-Dep------> USEOfPhi 1349 1350 // The mutation creates 1351 // USEOfPHI -------Artificial-Dep---> SRCOfCopy 1352 1353 // This overall will ensure, the USEOfPHI is scheduled before SRCOfCopy 1354 // (since USE is a predecessor), implies, the COPY/ REG_SEQUENCE is scheduled 1355 // late to avoid additional copies across iterations. The possible scheduling 1356 // order would be 1357 // USEOfPHI --- SRCOfCopy--- COPY/REG_SEQUENCE. 1358 1359 void SwingSchedulerDAG::CopyToPhiMutation::apply(ScheduleDAGInstrs *DAG) { 1360 for (SUnit &SU : DAG->SUnits) { 1361 // Find the COPY/REG_SEQUENCE instruction. 1362 if (!SU.getInstr()->isCopy() && !SU.getInstr()->isRegSequence()) 1363 continue; 1364 1365 // Record the loop carried PHIs. 1366 SmallVector<SUnit *, 4> PHISUs; 1367 // Record the SrcSUs that feed the COPY/REG_SEQUENCE instructions. 1368 SmallVector<SUnit *, 4> SrcSUs; 1369 1370 for (auto &Dep : SU.Preds) { 1371 SUnit *TmpSU = Dep.getSUnit(); 1372 MachineInstr *TmpMI = TmpSU->getInstr(); 1373 SDep::Kind DepKind = Dep.getKind(); 1374 // Save the loop carried PHI. 1375 if (DepKind == SDep::Anti && TmpMI->isPHI()) 1376 PHISUs.push_back(TmpSU); 1377 // Save the source of COPY/REG_SEQUENCE. 1378 // If the source has no pre-decessors, we will end up creating cycles. 1379 else if (DepKind == SDep::Data && !TmpMI->isPHI() && TmpSU->NumPreds > 0) 1380 SrcSUs.push_back(TmpSU); 1381 } 1382 1383 if (PHISUs.size() == 0 || SrcSUs.size() == 0) 1384 continue; 1385 1386 // Find the USEs of PHI. If the use is a PHI or REG_SEQUENCE, push back this 1387 // SUnit to the container. 1388 SmallVector<SUnit *, 8> UseSUs; 1389 // Do not use iterator based loop here as we are updating the container. 1390 for (size_t Index = 0; Index < PHISUs.size(); ++Index) { 1391 for (auto &Dep : PHISUs[Index]->Succs) { 1392 if (Dep.getKind() != SDep::Data) 1393 continue; 1394 1395 SUnit *TmpSU = Dep.getSUnit(); 1396 MachineInstr *TmpMI = TmpSU->getInstr(); 1397 if (TmpMI->isPHI() || TmpMI->isRegSequence()) { 1398 PHISUs.push_back(TmpSU); 1399 continue; 1400 } 1401 UseSUs.push_back(TmpSU); 1402 } 1403 } 1404 1405 if (UseSUs.size() == 0) 1406 continue; 1407 1408 SwingSchedulerDAG *SDAG = cast<SwingSchedulerDAG>(DAG); 1409 // Add the artificial dependencies if it does not form a cycle. 1410 for (auto *I : UseSUs) { 1411 for (auto *Src : SrcSUs) { 1412 if (!SDAG->Topo.IsReachable(I, Src) && Src != I) { 1413 Src->addPred(SDep(I, SDep::Artificial)); 1414 SDAG->Topo.AddPred(Src, I); 1415 } 1416 } 1417 } 1418 } 1419 } 1420 1421 /// Return true for DAG nodes that we ignore when computing the cost functions. 1422 /// We ignore the back-edge recurrence in order to avoid unbounded recursion 1423 /// in the calculation of the ASAP, ALAP, etc functions. 1424 static bool ignoreDependence(const SDep &D, bool isPred) { 1425 if (D.isArtificial() || D.getSUnit()->isBoundaryNode()) 1426 return true; 1427 return D.getKind() == SDep::Anti && isPred; 1428 } 1429 1430 /// Compute several functions need to order the nodes for scheduling. 1431 /// ASAP - Earliest time to schedule a node. 1432 /// ALAP - Latest time to schedule a node. 1433 /// MOV - Mobility function, difference between ALAP and ASAP. 1434 /// D - Depth of each node. 1435 /// H - Height of each node. 1436 void SwingSchedulerDAG::computeNodeFunctions(NodeSetType &NodeSets) { 1437 ScheduleInfo.resize(SUnits.size()); 1438 1439 LLVM_DEBUG({ 1440 for (int I : Topo) { 1441 const SUnit &SU = SUnits[I]; 1442 dumpNode(SU); 1443 } 1444 }); 1445 1446 int maxASAP = 0; 1447 // Compute ASAP and ZeroLatencyDepth. 1448 for (int I : Topo) { 1449 int asap = 0; 1450 int zeroLatencyDepth = 0; 1451 SUnit *SU = &SUnits[I]; 1452 for (const SDep &P : SU->Preds) { 1453 SUnit *pred = P.getSUnit(); 1454 if (P.getLatency() == 0) 1455 zeroLatencyDepth = 1456 std::max(zeroLatencyDepth, getZeroLatencyDepth(pred) + 1); 1457 if (ignoreDependence(P, true)) 1458 continue; 1459 asap = std::max(asap, (int)(getASAP(pred) + P.getLatency() - 1460 getDistance(pred, SU, P) * MII)); 1461 } 1462 maxASAP = std::max(maxASAP, asap); 1463 ScheduleInfo[I].ASAP = asap; 1464 ScheduleInfo[I].ZeroLatencyDepth = zeroLatencyDepth; 1465 } 1466 1467 // Compute ALAP, ZeroLatencyHeight, and MOV. 1468 for (int I : llvm::reverse(Topo)) { 1469 int alap = maxASAP; 1470 int zeroLatencyHeight = 0; 1471 SUnit *SU = &SUnits[I]; 1472 for (const SDep &S : SU->Succs) { 1473 SUnit *succ = S.getSUnit(); 1474 if (succ->isBoundaryNode()) 1475 continue; 1476 if (S.getLatency() == 0) 1477 zeroLatencyHeight = 1478 std::max(zeroLatencyHeight, getZeroLatencyHeight(succ) + 1); 1479 if (ignoreDependence(S, true)) 1480 continue; 1481 alap = std::min(alap, (int)(getALAP(succ) - S.getLatency() + 1482 getDistance(SU, succ, S) * MII)); 1483 } 1484 1485 ScheduleInfo[I].ALAP = alap; 1486 ScheduleInfo[I].ZeroLatencyHeight = zeroLatencyHeight; 1487 } 1488 1489 // After computing the node functions, compute the summary for each node set. 1490 for (NodeSet &I : NodeSets) 1491 I.computeNodeSetInfo(this); 1492 1493 LLVM_DEBUG({ 1494 for (unsigned i = 0; i < SUnits.size(); i++) { 1495 dbgs() << "\tNode " << i << ":\n"; 1496 dbgs() << "\t ASAP = " << getASAP(&SUnits[i]) << "\n"; 1497 dbgs() << "\t ALAP = " << getALAP(&SUnits[i]) << "\n"; 1498 dbgs() << "\t MOV = " << getMOV(&SUnits[i]) << "\n"; 1499 dbgs() << "\t D = " << getDepth(&SUnits[i]) << "\n"; 1500 dbgs() << "\t H = " << getHeight(&SUnits[i]) << "\n"; 1501 dbgs() << "\t ZLD = " << getZeroLatencyDepth(&SUnits[i]) << "\n"; 1502 dbgs() << "\t ZLH = " << getZeroLatencyHeight(&SUnits[i]) << "\n"; 1503 } 1504 }); 1505 } 1506 1507 /// Compute the Pred_L(O) set, as defined in the paper. The set is defined 1508 /// as the predecessors of the elements of NodeOrder that are not also in 1509 /// NodeOrder. 1510 static bool pred_L(SetVector<SUnit *> &NodeOrder, 1511 SmallSetVector<SUnit *, 8> &Preds, 1512 const NodeSet *S = nullptr) { 1513 Preds.clear(); 1514 for (const SUnit *SU : NodeOrder) { 1515 for (const SDep &Pred : SU->Preds) { 1516 if (S && S->count(Pred.getSUnit()) == 0) 1517 continue; 1518 if (ignoreDependence(Pred, true)) 1519 continue; 1520 if (NodeOrder.count(Pred.getSUnit()) == 0) 1521 Preds.insert(Pred.getSUnit()); 1522 } 1523 // Back-edges are predecessors with an anti-dependence. 1524 for (const SDep &Succ : SU->Succs) { 1525 if (Succ.getKind() != SDep::Anti) 1526 continue; 1527 if (S && S->count(Succ.getSUnit()) == 0) 1528 continue; 1529 if (NodeOrder.count(Succ.getSUnit()) == 0) 1530 Preds.insert(Succ.getSUnit()); 1531 } 1532 } 1533 return !Preds.empty(); 1534 } 1535 1536 /// Compute the Succ_L(O) set, as defined in the paper. The set is defined 1537 /// as the successors of the elements of NodeOrder that are not also in 1538 /// NodeOrder. 1539 static bool succ_L(SetVector<SUnit *> &NodeOrder, 1540 SmallSetVector<SUnit *, 8> &Succs, 1541 const NodeSet *S = nullptr) { 1542 Succs.clear(); 1543 for (const SUnit *SU : NodeOrder) { 1544 for (const SDep &Succ : SU->Succs) { 1545 if (S && S->count(Succ.getSUnit()) == 0) 1546 continue; 1547 if (ignoreDependence(Succ, false)) 1548 continue; 1549 if (NodeOrder.count(Succ.getSUnit()) == 0) 1550 Succs.insert(Succ.getSUnit()); 1551 } 1552 for (const SDep &Pred : SU->Preds) { 1553 if (Pred.getKind() != SDep::Anti) 1554 continue; 1555 if (S && S->count(Pred.getSUnit()) == 0) 1556 continue; 1557 if (NodeOrder.count(Pred.getSUnit()) == 0) 1558 Succs.insert(Pred.getSUnit()); 1559 } 1560 } 1561 return !Succs.empty(); 1562 } 1563 1564 /// Return true if there is a path from the specified node to any of the nodes 1565 /// in DestNodes. Keep track and return the nodes in any path. 1566 static bool computePath(SUnit *Cur, SetVector<SUnit *> &Path, 1567 SetVector<SUnit *> &DestNodes, 1568 SetVector<SUnit *> &Exclude, 1569 SmallPtrSet<SUnit *, 8> &Visited) { 1570 if (Cur->isBoundaryNode()) 1571 return false; 1572 if (Exclude.contains(Cur)) 1573 return false; 1574 if (DestNodes.contains(Cur)) 1575 return true; 1576 if (!Visited.insert(Cur).second) 1577 return Path.contains(Cur); 1578 bool FoundPath = false; 1579 for (auto &SI : Cur->Succs) 1580 if (!ignoreDependence(SI, false)) 1581 FoundPath |= 1582 computePath(SI.getSUnit(), Path, DestNodes, Exclude, Visited); 1583 for (auto &PI : Cur->Preds) 1584 if (PI.getKind() == SDep::Anti) 1585 FoundPath |= 1586 computePath(PI.getSUnit(), Path, DestNodes, Exclude, Visited); 1587 if (FoundPath) 1588 Path.insert(Cur); 1589 return FoundPath; 1590 } 1591 1592 /// Compute the live-out registers for the instructions in a node-set. 1593 /// The live-out registers are those that are defined in the node-set, 1594 /// but not used. Except for use operands of Phis. 1595 static void computeLiveOuts(MachineFunction &MF, RegPressureTracker &RPTracker, 1596 NodeSet &NS) { 1597 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 1598 MachineRegisterInfo &MRI = MF.getRegInfo(); 1599 SmallVector<RegisterMaskPair, 8> LiveOutRegs; 1600 SmallSet<unsigned, 4> Uses; 1601 for (SUnit *SU : NS) { 1602 const MachineInstr *MI = SU->getInstr(); 1603 if (MI->isPHI()) 1604 continue; 1605 for (const MachineOperand &MO : MI->operands()) 1606 if (MO.isReg() && MO.isUse()) { 1607 Register Reg = MO.getReg(); 1608 if (Register::isVirtualRegister(Reg)) 1609 Uses.insert(Reg); 1610 else if (MRI.isAllocatable(Reg)) 1611 for (MCRegUnitIterator Units(Reg.asMCReg(), TRI); Units.isValid(); 1612 ++Units) 1613 Uses.insert(*Units); 1614 } 1615 } 1616 for (SUnit *SU : NS) 1617 for (const MachineOperand &MO : SU->getInstr()->operands()) 1618 if (MO.isReg() && MO.isDef() && !MO.isDead()) { 1619 Register Reg = MO.getReg(); 1620 if (Register::isVirtualRegister(Reg)) { 1621 if (!Uses.count(Reg)) 1622 LiveOutRegs.push_back(RegisterMaskPair(Reg, 1623 LaneBitmask::getNone())); 1624 } else if (MRI.isAllocatable(Reg)) { 1625 for (MCRegUnitIterator Units(Reg.asMCReg(), TRI); Units.isValid(); 1626 ++Units) 1627 if (!Uses.count(*Units)) 1628 LiveOutRegs.push_back(RegisterMaskPair(*Units, 1629 LaneBitmask::getNone())); 1630 } 1631 } 1632 RPTracker.addLiveRegs(LiveOutRegs); 1633 } 1634 1635 /// A heuristic to filter nodes in recurrent node-sets if the register 1636 /// pressure of a set is too high. 1637 void SwingSchedulerDAG::registerPressureFilter(NodeSetType &NodeSets) { 1638 for (auto &NS : NodeSets) { 1639 // Skip small node-sets since they won't cause register pressure problems. 1640 if (NS.size() <= 2) 1641 continue; 1642 IntervalPressure RecRegPressure; 1643 RegPressureTracker RecRPTracker(RecRegPressure); 1644 RecRPTracker.init(&MF, &RegClassInfo, &LIS, BB, BB->end(), false, true); 1645 computeLiveOuts(MF, RecRPTracker, NS); 1646 RecRPTracker.closeBottom(); 1647 1648 std::vector<SUnit *> SUnits(NS.begin(), NS.end()); 1649 llvm::sort(SUnits, [](const SUnit *A, const SUnit *B) { 1650 return A->NodeNum > B->NodeNum; 1651 }); 1652 1653 for (auto &SU : SUnits) { 1654 // Since we're computing the register pressure for a subset of the 1655 // instructions in a block, we need to set the tracker for each 1656 // instruction in the node-set. The tracker is set to the instruction 1657 // just after the one we're interested in. 1658 MachineBasicBlock::const_iterator CurInstI = SU->getInstr(); 1659 RecRPTracker.setPos(std::next(CurInstI)); 1660 1661 RegPressureDelta RPDelta; 1662 ArrayRef<PressureChange> CriticalPSets; 1663 RecRPTracker.getMaxUpwardPressureDelta(SU->getInstr(), nullptr, RPDelta, 1664 CriticalPSets, 1665 RecRegPressure.MaxSetPressure); 1666 if (RPDelta.Excess.isValid()) { 1667 LLVM_DEBUG( 1668 dbgs() << "Excess register pressure: SU(" << SU->NodeNum << ") " 1669 << TRI->getRegPressureSetName(RPDelta.Excess.getPSet()) 1670 << ":" << RPDelta.Excess.getUnitInc() << "\n"); 1671 NS.setExceedPressure(SU); 1672 break; 1673 } 1674 RecRPTracker.recede(); 1675 } 1676 } 1677 } 1678 1679 /// A heuristic to colocate node sets that have the same set of 1680 /// successors. 1681 void SwingSchedulerDAG::colocateNodeSets(NodeSetType &NodeSets) { 1682 unsigned Colocate = 0; 1683 for (int i = 0, e = NodeSets.size(); i < e; ++i) { 1684 NodeSet &N1 = NodeSets[i]; 1685 SmallSetVector<SUnit *, 8> S1; 1686 if (N1.empty() || !succ_L(N1, S1)) 1687 continue; 1688 for (int j = i + 1; j < e; ++j) { 1689 NodeSet &N2 = NodeSets[j]; 1690 if (N1.compareRecMII(N2) != 0) 1691 continue; 1692 SmallSetVector<SUnit *, 8> S2; 1693 if (N2.empty() || !succ_L(N2, S2)) 1694 continue; 1695 if (llvm::set_is_subset(S1, S2) && S1.size() == S2.size()) { 1696 N1.setColocate(++Colocate); 1697 N2.setColocate(Colocate); 1698 break; 1699 } 1700 } 1701 } 1702 } 1703 1704 /// Check if the existing node-sets are profitable. If not, then ignore the 1705 /// recurrent node-sets, and attempt to schedule all nodes together. This is 1706 /// a heuristic. If the MII is large and all the recurrent node-sets are small, 1707 /// then it's best to try to schedule all instructions together instead of 1708 /// starting with the recurrent node-sets. 1709 void SwingSchedulerDAG::checkNodeSets(NodeSetType &NodeSets) { 1710 // Look for loops with a large MII. 1711 if (MII < 17) 1712 return; 1713 // Check if the node-set contains only a simple add recurrence. 1714 for (auto &NS : NodeSets) { 1715 if (NS.getRecMII() > 2) 1716 return; 1717 if (NS.getMaxDepth() > MII) 1718 return; 1719 } 1720 NodeSets.clear(); 1721 LLVM_DEBUG(dbgs() << "Clear recurrence node-sets\n"); 1722 } 1723 1724 /// Add the nodes that do not belong to a recurrence set into groups 1725 /// based upon connected components. 1726 void SwingSchedulerDAG::groupRemainingNodes(NodeSetType &NodeSets) { 1727 SetVector<SUnit *> NodesAdded; 1728 SmallPtrSet<SUnit *, 8> Visited; 1729 // Add the nodes that are on a path between the previous node sets and 1730 // the current node set. 1731 for (NodeSet &I : NodeSets) { 1732 SmallSetVector<SUnit *, 8> N; 1733 // Add the nodes from the current node set to the previous node set. 1734 if (succ_L(I, N)) { 1735 SetVector<SUnit *> Path; 1736 for (SUnit *NI : N) { 1737 Visited.clear(); 1738 computePath(NI, Path, NodesAdded, I, Visited); 1739 } 1740 if (!Path.empty()) 1741 I.insert(Path.begin(), Path.end()); 1742 } 1743 // Add the nodes from the previous node set to the current node set. 1744 N.clear(); 1745 if (succ_L(NodesAdded, N)) { 1746 SetVector<SUnit *> Path; 1747 for (SUnit *NI : N) { 1748 Visited.clear(); 1749 computePath(NI, Path, I, NodesAdded, Visited); 1750 } 1751 if (!Path.empty()) 1752 I.insert(Path.begin(), Path.end()); 1753 } 1754 NodesAdded.insert(I.begin(), I.end()); 1755 } 1756 1757 // Create a new node set with the connected nodes of any successor of a node 1758 // in a recurrent set. 1759 NodeSet NewSet; 1760 SmallSetVector<SUnit *, 8> N; 1761 if (succ_L(NodesAdded, N)) 1762 for (SUnit *I : N) 1763 addConnectedNodes(I, NewSet, NodesAdded); 1764 if (!NewSet.empty()) 1765 NodeSets.push_back(NewSet); 1766 1767 // Create a new node set with the connected nodes of any predecessor of a node 1768 // in a recurrent set. 1769 NewSet.clear(); 1770 if (pred_L(NodesAdded, N)) 1771 for (SUnit *I : N) 1772 addConnectedNodes(I, NewSet, NodesAdded); 1773 if (!NewSet.empty()) 1774 NodeSets.push_back(NewSet); 1775 1776 // Create new nodes sets with the connected nodes any remaining node that 1777 // has no predecessor. 1778 for (SUnit &SU : SUnits) { 1779 if (NodesAdded.count(&SU) == 0) { 1780 NewSet.clear(); 1781 addConnectedNodes(&SU, NewSet, NodesAdded); 1782 if (!NewSet.empty()) 1783 NodeSets.push_back(NewSet); 1784 } 1785 } 1786 } 1787 1788 /// Add the node to the set, and add all of its connected nodes to the set. 1789 void SwingSchedulerDAG::addConnectedNodes(SUnit *SU, NodeSet &NewSet, 1790 SetVector<SUnit *> &NodesAdded) { 1791 NewSet.insert(SU); 1792 NodesAdded.insert(SU); 1793 for (auto &SI : SU->Succs) { 1794 SUnit *Successor = SI.getSUnit(); 1795 if (!SI.isArtificial() && !Successor->isBoundaryNode() && 1796 NodesAdded.count(Successor) == 0) 1797 addConnectedNodes(Successor, NewSet, NodesAdded); 1798 } 1799 for (auto &PI : SU->Preds) { 1800 SUnit *Predecessor = PI.getSUnit(); 1801 if (!PI.isArtificial() && NodesAdded.count(Predecessor) == 0) 1802 addConnectedNodes(Predecessor, NewSet, NodesAdded); 1803 } 1804 } 1805 1806 /// Return true if Set1 contains elements in Set2. The elements in common 1807 /// are returned in a different container. 1808 static bool isIntersect(SmallSetVector<SUnit *, 8> &Set1, const NodeSet &Set2, 1809 SmallSetVector<SUnit *, 8> &Result) { 1810 Result.clear(); 1811 for (SUnit *SU : Set1) { 1812 if (Set2.count(SU) != 0) 1813 Result.insert(SU); 1814 } 1815 return !Result.empty(); 1816 } 1817 1818 /// Merge the recurrence node sets that have the same initial node. 1819 void SwingSchedulerDAG::fuseRecs(NodeSetType &NodeSets) { 1820 for (NodeSetType::iterator I = NodeSets.begin(), E = NodeSets.end(); I != E; 1821 ++I) { 1822 NodeSet &NI = *I; 1823 for (NodeSetType::iterator J = I + 1; J != E;) { 1824 NodeSet &NJ = *J; 1825 if (NI.getNode(0)->NodeNum == NJ.getNode(0)->NodeNum) { 1826 if (NJ.compareRecMII(NI) > 0) 1827 NI.setRecMII(NJ.getRecMII()); 1828 for (SUnit *SU : *J) 1829 I->insert(SU); 1830 NodeSets.erase(J); 1831 E = NodeSets.end(); 1832 } else { 1833 ++J; 1834 } 1835 } 1836 } 1837 } 1838 1839 /// Remove nodes that have been scheduled in previous NodeSets. 1840 void SwingSchedulerDAG::removeDuplicateNodes(NodeSetType &NodeSets) { 1841 for (NodeSetType::iterator I = NodeSets.begin(), E = NodeSets.end(); I != E; 1842 ++I) 1843 for (NodeSetType::iterator J = I + 1; J != E;) { 1844 J->remove_if([&](SUnit *SUJ) { return I->count(SUJ); }); 1845 1846 if (J->empty()) { 1847 NodeSets.erase(J); 1848 E = NodeSets.end(); 1849 } else { 1850 ++J; 1851 } 1852 } 1853 } 1854 1855 /// Compute an ordered list of the dependence graph nodes, which 1856 /// indicates the order that the nodes will be scheduled. This is a 1857 /// two-level algorithm. First, a partial order is created, which 1858 /// consists of a list of sets ordered from highest to lowest priority. 1859 void SwingSchedulerDAG::computeNodeOrder(NodeSetType &NodeSets) { 1860 SmallSetVector<SUnit *, 8> R; 1861 NodeOrder.clear(); 1862 1863 for (auto &Nodes : NodeSets) { 1864 LLVM_DEBUG(dbgs() << "NodeSet size " << Nodes.size() << "\n"); 1865 OrderKind Order; 1866 SmallSetVector<SUnit *, 8> N; 1867 if (pred_L(NodeOrder, N) && llvm::set_is_subset(N, Nodes)) { 1868 R.insert(N.begin(), N.end()); 1869 Order = BottomUp; 1870 LLVM_DEBUG(dbgs() << " Bottom up (preds) "); 1871 } else if (succ_L(NodeOrder, N) && llvm::set_is_subset(N, Nodes)) { 1872 R.insert(N.begin(), N.end()); 1873 Order = TopDown; 1874 LLVM_DEBUG(dbgs() << " Top down (succs) "); 1875 } else if (isIntersect(N, Nodes, R)) { 1876 // If some of the successors are in the existing node-set, then use the 1877 // top-down ordering. 1878 Order = TopDown; 1879 LLVM_DEBUG(dbgs() << " Top down (intersect) "); 1880 } else if (NodeSets.size() == 1) { 1881 for (const auto &N : Nodes) 1882 if (N->Succs.size() == 0) 1883 R.insert(N); 1884 Order = BottomUp; 1885 LLVM_DEBUG(dbgs() << " Bottom up (all) "); 1886 } else { 1887 // Find the node with the highest ASAP. 1888 SUnit *maxASAP = nullptr; 1889 for (SUnit *SU : Nodes) { 1890 if (maxASAP == nullptr || getASAP(SU) > getASAP(maxASAP) || 1891 (getASAP(SU) == getASAP(maxASAP) && SU->NodeNum > maxASAP->NodeNum)) 1892 maxASAP = SU; 1893 } 1894 R.insert(maxASAP); 1895 Order = BottomUp; 1896 LLVM_DEBUG(dbgs() << " Bottom up (default) "); 1897 } 1898 1899 while (!R.empty()) { 1900 if (Order == TopDown) { 1901 // Choose the node with the maximum height. If more than one, choose 1902 // the node wiTH the maximum ZeroLatencyHeight. If still more than one, 1903 // choose the node with the lowest MOV. 1904 while (!R.empty()) { 1905 SUnit *maxHeight = nullptr; 1906 for (SUnit *I : R) { 1907 if (maxHeight == nullptr || getHeight(I) > getHeight(maxHeight)) 1908 maxHeight = I; 1909 else if (getHeight(I) == getHeight(maxHeight) && 1910 getZeroLatencyHeight(I) > getZeroLatencyHeight(maxHeight)) 1911 maxHeight = I; 1912 else if (getHeight(I) == getHeight(maxHeight) && 1913 getZeroLatencyHeight(I) == 1914 getZeroLatencyHeight(maxHeight) && 1915 getMOV(I) < getMOV(maxHeight)) 1916 maxHeight = I; 1917 } 1918 NodeOrder.insert(maxHeight); 1919 LLVM_DEBUG(dbgs() << maxHeight->NodeNum << " "); 1920 R.remove(maxHeight); 1921 for (const auto &I : maxHeight->Succs) { 1922 if (Nodes.count(I.getSUnit()) == 0) 1923 continue; 1924 if (NodeOrder.contains(I.getSUnit())) 1925 continue; 1926 if (ignoreDependence(I, false)) 1927 continue; 1928 R.insert(I.getSUnit()); 1929 } 1930 // Back-edges are predecessors with an anti-dependence. 1931 for (const auto &I : maxHeight->Preds) { 1932 if (I.getKind() != SDep::Anti) 1933 continue; 1934 if (Nodes.count(I.getSUnit()) == 0) 1935 continue; 1936 if (NodeOrder.contains(I.getSUnit())) 1937 continue; 1938 R.insert(I.getSUnit()); 1939 } 1940 } 1941 Order = BottomUp; 1942 LLVM_DEBUG(dbgs() << "\n Switching order to bottom up "); 1943 SmallSetVector<SUnit *, 8> N; 1944 if (pred_L(NodeOrder, N, &Nodes)) 1945 R.insert(N.begin(), N.end()); 1946 } else { 1947 // Choose the node with the maximum depth. If more than one, choose 1948 // the node with the maximum ZeroLatencyDepth. If still more than one, 1949 // choose the node with the lowest MOV. 1950 while (!R.empty()) { 1951 SUnit *maxDepth = nullptr; 1952 for (SUnit *I : R) { 1953 if (maxDepth == nullptr || getDepth(I) > getDepth(maxDepth)) 1954 maxDepth = I; 1955 else if (getDepth(I) == getDepth(maxDepth) && 1956 getZeroLatencyDepth(I) > getZeroLatencyDepth(maxDepth)) 1957 maxDepth = I; 1958 else if (getDepth(I) == getDepth(maxDepth) && 1959 getZeroLatencyDepth(I) == getZeroLatencyDepth(maxDepth) && 1960 getMOV(I) < getMOV(maxDepth)) 1961 maxDepth = I; 1962 } 1963 NodeOrder.insert(maxDepth); 1964 LLVM_DEBUG(dbgs() << maxDepth->NodeNum << " "); 1965 R.remove(maxDepth); 1966 if (Nodes.isExceedSU(maxDepth)) { 1967 Order = TopDown; 1968 R.clear(); 1969 R.insert(Nodes.getNode(0)); 1970 break; 1971 } 1972 for (const auto &I : maxDepth->Preds) { 1973 if (Nodes.count(I.getSUnit()) == 0) 1974 continue; 1975 if (NodeOrder.contains(I.getSUnit())) 1976 continue; 1977 R.insert(I.getSUnit()); 1978 } 1979 // Back-edges are predecessors with an anti-dependence. 1980 for (const auto &I : maxDepth->Succs) { 1981 if (I.getKind() != SDep::Anti) 1982 continue; 1983 if (Nodes.count(I.getSUnit()) == 0) 1984 continue; 1985 if (NodeOrder.contains(I.getSUnit())) 1986 continue; 1987 R.insert(I.getSUnit()); 1988 } 1989 } 1990 Order = TopDown; 1991 LLVM_DEBUG(dbgs() << "\n Switching order to top down "); 1992 SmallSetVector<SUnit *, 8> N; 1993 if (succ_L(NodeOrder, N, &Nodes)) 1994 R.insert(N.begin(), N.end()); 1995 } 1996 } 1997 LLVM_DEBUG(dbgs() << "\nDone with Nodeset\n"); 1998 } 1999 2000 LLVM_DEBUG({ 2001 dbgs() << "Node order: "; 2002 for (SUnit *I : NodeOrder) 2003 dbgs() << " " << I->NodeNum << " "; 2004 dbgs() << "\n"; 2005 }); 2006 } 2007 2008 /// Process the nodes in the computed order and create the pipelined schedule 2009 /// of the instructions, if possible. Return true if a schedule is found. 2010 bool SwingSchedulerDAG::schedulePipeline(SMSchedule &Schedule) { 2011 2012 if (NodeOrder.empty()){ 2013 LLVM_DEBUG(dbgs() << "NodeOrder is empty! abort scheduling\n" ); 2014 return false; 2015 } 2016 2017 bool scheduleFound = false; 2018 // Keep increasing II until a valid schedule is found. 2019 for (unsigned II = MII; II <= MAX_II && !scheduleFound; ++II) { 2020 Schedule.reset(); 2021 Schedule.setInitiationInterval(II); 2022 LLVM_DEBUG(dbgs() << "Try to schedule with " << II << "\n"); 2023 2024 SetVector<SUnit *>::iterator NI = NodeOrder.begin(); 2025 SetVector<SUnit *>::iterator NE = NodeOrder.end(); 2026 do { 2027 SUnit *SU = *NI; 2028 2029 // Compute the schedule time for the instruction, which is based 2030 // upon the scheduled time for any predecessors/successors. 2031 int EarlyStart = INT_MIN; 2032 int LateStart = INT_MAX; 2033 // These values are set when the size of the schedule window is limited 2034 // due to chain dependences. 2035 int SchedEnd = INT_MAX; 2036 int SchedStart = INT_MIN; 2037 Schedule.computeStart(SU, &EarlyStart, &LateStart, &SchedEnd, &SchedStart, 2038 II, this); 2039 LLVM_DEBUG({ 2040 dbgs() << "\n"; 2041 dbgs() << "Inst (" << SU->NodeNum << ") "; 2042 SU->getInstr()->dump(); 2043 dbgs() << "\n"; 2044 }); 2045 LLVM_DEBUG({ 2046 dbgs() << format("\tes: %8x ls: %8x me: %8x ms: %8x\n", EarlyStart, 2047 LateStart, SchedEnd, SchedStart); 2048 }); 2049 2050 if (EarlyStart > LateStart || SchedEnd < EarlyStart || 2051 SchedStart > LateStart) 2052 scheduleFound = false; 2053 else if (EarlyStart != INT_MIN && LateStart == INT_MAX) { 2054 SchedEnd = std::min(SchedEnd, EarlyStart + (int)II - 1); 2055 scheduleFound = Schedule.insert(SU, EarlyStart, SchedEnd, II); 2056 } else if (EarlyStart == INT_MIN && LateStart != INT_MAX) { 2057 SchedStart = std::max(SchedStart, LateStart - (int)II + 1); 2058 scheduleFound = Schedule.insert(SU, LateStart, SchedStart, II); 2059 } else if (EarlyStart != INT_MIN && LateStart != INT_MAX) { 2060 SchedEnd = 2061 std::min(SchedEnd, std::min(LateStart, EarlyStart + (int)II - 1)); 2062 // When scheduling a Phi it is better to start at the late cycle and go 2063 // backwards. The default order may insert the Phi too far away from 2064 // its first dependence. 2065 if (SU->getInstr()->isPHI()) 2066 scheduleFound = Schedule.insert(SU, SchedEnd, EarlyStart, II); 2067 else 2068 scheduleFound = Schedule.insert(SU, EarlyStart, SchedEnd, II); 2069 } else { 2070 int FirstCycle = Schedule.getFirstCycle(); 2071 scheduleFound = Schedule.insert(SU, FirstCycle + getASAP(SU), 2072 FirstCycle + getASAP(SU) + II - 1, II); 2073 } 2074 // Even if we find a schedule, make sure the schedule doesn't exceed the 2075 // allowable number of stages. We keep trying if this happens. 2076 if (scheduleFound) 2077 if (SwpMaxStages > -1 && 2078 Schedule.getMaxStageCount() > (unsigned)SwpMaxStages) 2079 scheduleFound = false; 2080 2081 LLVM_DEBUG({ 2082 if (!scheduleFound) 2083 dbgs() << "\tCan't schedule\n"; 2084 }); 2085 } while (++NI != NE && scheduleFound); 2086 2087 // If a schedule is found, ensure non-pipelined instructions are in stage 0 2088 if (scheduleFound) 2089 scheduleFound = 2090 Schedule.normalizeNonPipelinedInstructions(this, LoopPipelinerInfo); 2091 2092 // If a schedule is found, check if it is a valid schedule too. 2093 if (scheduleFound) 2094 scheduleFound = Schedule.isValidSchedule(this); 2095 } 2096 2097 LLVM_DEBUG(dbgs() << "Schedule Found? " << scheduleFound 2098 << " (II=" << Schedule.getInitiationInterval() 2099 << ")\n"); 2100 2101 if (scheduleFound) { 2102 Schedule.finalizeSchedule(this); 2103 Pass.ORE->emit([&]() { 2104 return MachineOptimizationRemarkAnalysis( 2105 DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader()) 2106 << "Schedule found with Initiation Interval: " 2107 << ore::NV("II", Schedule.getInitiationInterval()) 2108 << ", MaxStageCount: " 2109 << ore::NV("MaxStageCount", Schedule.getMaxStageCount()); 2110 }); 2111 } else 2112 Schedule.reset(); 2113 2114 return scheduleFound && Schedule.getMaxStageCount() > 0; 2115 } 2116 2117 /// Return true if we can compute the amount the instruction changes 2118 /// during each iteration. Set Delta to the amount of the change. 2119 bool SwingSchedulerDAG::computeDelta(MachineInstr &MI, unsigned &Delta) { 2120 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 2121 const MachineOperand *BaseOp; 2122 int64_t Offset; 2123 bool OffsetIsScalable; 2124 if (!TII->getMemOperandWithOffset(MI, BaseOp, Offset, OffsetIsScalable, TRI)) 2125 return false; 2126 2127 // FIXME: This algorithm assumes instructions have fixed-size offsets. 2128 if (OffsetIsScalable) 2129 return false; 2130 2131 if (!BaseOp->isReg()) 2132 return false; 2133 2134 Register BaseReg = BaseOp->getReg(); 2135 2136 MachineRegisterInfo &MRI = MF.getRegInfo(); 2137 // Check if there is a Phi. If so, get the definition in the loop. 2138 MachineInstr *BaseDef = MRI.getVRegDef(BaseReg); 2139 if (BaseDef && BaseDef->isPHI()) { 2140 BaseReg = getLoopPhiReg(*BaseDef, MI.getParent()); 2141 BaseDef = MRI.getVRegDef(BaseReg); 2142 } 2143 if (!BaseDef) 2144 return false; 2145 2146 int D = 0; 2147 if (!TII->getIncrementValue(*BaseDef, D) && D >= 0) 2148 return false; 2149 2150 Delta = D; 2151 return true; 2152 } 2153 2154 /// Check if we can change the instruction to use an offset value from the 2155 /// previous iteration. If so, return true and set the base and offset values 2156 /// so that we can rewrite the load, if necessary. 2157 /// v1 = Phi(v0, v3) 2158 /// v2 = load v1, 0 2159 /// v3 = post_store v1, 4, x 2160 /// This function enables the load to be rewritten as v2 = load v3, 4. 2161 bool SwingSchedulerDAG::canUseLastOffsetValue(MachineInstr *MI, 2162 unsigned &BasePos, 2163 unsigned &OffsetPos, 2164 unsigned &NewBase, 2165 int64_t &Offset) { 2166 // Get the load instruction. 2167 if (TII->isPostIncrement(*MI)) 2168 return false; 2169 unsigned BasePosLd, OffsetPosLd; 2170 if (!TII->getBaseAndOffsetPosition(*MI, BasePosLd, OffsetPosLd)) 2171 return false; 2172 Register BaseReg = MI->getOperand(BasePosLd).getReg(); 2173 2174 // Look for the Phi instruction. 2175 MachineRegisterInfo &MRI = MI->getMF()->getRegInfo(); 2176 MachineInstr *Phi = MRI.getVRegDef(BaseReg); 2177 if (!Phi || !Phi->isPHI()) 2178 return false; 2179 // Get the register defined in the loop block. 2180 unsigned PrevReg = getLoopPhiReg(*Phi, MI->getParent()); 2181 if (!PrevReg) 2182 return false; 2183 2184 // Check for the post-increment load/store instruction. 2185 MachineInstr *PrevDef = MRI.getVRegDef(PrevReg); 2186 if (!PrevDef || PrevDef == MI) 2187 return false; 2188 2189 if (!TII->isPostIncrement(*PrevDef)) 2190 return false; 2191 2192 unsigned BasePos1 = 0, OffsetPos1 = 0; 2193 if (!TII->getBaseAndOffsetPosition(*PrevDef, BasePos1, OffsetPos1)) 2194 return false; 2195 2196 // Make sure that the instructions do not access the same memory location in 2197 // the next iteration. 2198 int64_t LoadOffset = MI->getOperand(OffsetPosLd).getImm(); 2199 int64_t StoreOffset = PrevDef->getOperand(OffsetPos1).getImm(); 2200 MachineInstr *NewMI = MF.CloneMachineInstr(MI); 2201 NewMI->getOperand(OffsetPosLd).setImm(LoadOffset + StoreOffset); 2202 bool Disjoint = TII->areMemAccessesTriviallyDisjoint(*NewMI, *PrevDef); 2203 MF.deleteMachineInstr(NewMI); 2204 if (!Disjoint) 2205 return false; 2206 2207 // Set the return value once we determine that we return true. 2208 BasePos = BasePosLd; 2209 OffsetPos = OffsetPosLd; 2210 NewBase = PrevReg; 2211 Offset = StoreOffset; 2212 return true; 2213 } 2214 2215 /// Apply changes to the instruction if needed. The changes are need 2216 /// to improve the scheduling and depend up on the final schedule. 2217 void SwingSchedulerDAG::applyInstrChange(MachineInstr *MI, 2218 SMSchedule &Schedule) { 2219 SUnit *SU = getSUnit(MI); 2220 DenseMap<SUnit *, std::pair<unsigned, int64_t>>::iterator It = 2221 InstrChanges.find(SU); 2222 if (It != InstrChanges.end()) { 2223 std::pair<unsigned, int64_t> RegAndOffset = It->second; 2224 unsigned BasePos, OffsetPos; 2225 if (!TII->getBaseAndOffsetPosition(*MI, BasePos, OffsetPos)) 2226 return; 2227 Register BaseReg = MI->getOperand(BasePos).getReg(); 2228 MachineInstr *LoopDef = findDefInLoop(BaseReg); 2229 int DefStageNum = Schedule.stageScheduled(getSUnit(LoopDef)); 2230 int DefCycleNum = Schedule.cycleScheduled(getSUnit(LoopDef)); 2231 int BaseStageNum = Schedule.stageScheduled(SU); 2232 int BaseCycleNum = Schedule.cycleScheduled(SU); 2233 if (BaseStageNum < DefStageNum) { 2234 MachineInstr *NewMI = MF.CloneMachineInstr(MI); 2235 int OffsetDiff = DefStageNum - BaseStageNum; 2236 if (DefCycleNum < BaseCycleNum) { 2237 NewMI->getOperand(BasePos).setReg(RegAndOffset.first); 2238 if (OffsetDiff > 0) 2239 --OffsetDiff; 2240 } 2241 int64_t NewOffset = 2242 MI->getOperand(OffsetPos).getImm() + RegAndOffset.second * OffsetDiff; 2243 NewMI->getOperand(OffsetPos).setImm(NewOffset); 2244 SU->setInstr(NewMI); 2245 MISUnitMap[NewMI] = SU; 2246 NewMIs[MI] = NewMI; 2247 } 2248 } 2249 } 2250 2251 /// Return the instruction in the loop that defines the register. 2252 /// If the definition is a Phi, then follow the Phi operand to 2253 /// the instruction in the loop. 2254 MachineInstr *SwingSchedulerDAG::findDefInLoop(Register Reg) { 2255 SmallPtrSet<MachineInstr *, 8> Visited; 2256 MachineInstr *Def = MRI.getVRegDef(Reg); 2257 while (Def->isPHI()) { 2258 if (!Visited.insert(Def).second) 2259 break; 2260 for (unsigned i = 1, e = Def->getNumOperands(); i < e; i += 2) 2261 if (Def->getOperand(i + 1).getMBB() == BB) { 2262 Def = MRI.getVRegDef(Def->getOperand(i).getReg()); 2263 break; 2264 } 2265 } 2266 return Def; 2267 } 2268 2269 /// Return true for an order or output dependence that is loop carried 2270 /// potentially. A dependence is loop carried if the destination defines a valu 2271 /// that may be used or defined by the source in a subsequent iteration. 2272 bool SwingSchedulerDAG::isLoopCarriedDep(SUnit *Source, const SDep &Dep, 2273 bool isSucc) { 2274 if ((Dep.getKind() != SDep::Order && Dep.getKind() != SDep::Output) || 2275 Dep.isArtificial() || Dep.getSUnit()->isBoundaryNode()) 2276 return false; 2277 2278 if (!SwpPruneLoopCarried) 2279 return true; 2280 2281 if (Dep.getKind() == SDep::Output) 2282 return true; 2283 2284 MachineInstr *SI = Source->getInstr(); 2285 MachineInstr *DI = Dep.getSUnit()->getInstr(); 2286 if (!isSucc) 2287 std::swap(SI, DI); 2288 assert(SI != nullptr && DI != nullptr && "Expecting SUnit with an MI."); 2289 2290 // Assume ordered loads and stores may have a loop carried dependence. 2291 if (SI->hasUnmodeledSideEffects() || DI->hasUnmodeledSideEffects() || 2292 SI->mayRaiseFPException() || DI->mayRaiseFPException() || 2293 SI->hasOrderedMemoryRef() || DI->hasOrderedMemoryRef()) 2294 return true; 2295 2296 // Only chain dependences between a load and store can be loop carried. 2297 if (!DI->mayStore() || !SI->mayLoad()) 2298 return false; 2299 2300 unsigned DeltaS, DeltaD; 2301 if (!computeDelta(*SI, DeltaS) || !computeDelta(*DI, DeltaD)) 2302 return true; 2303 2304 const MachineOperand *BaseOpS, *BaseOpD; 2305 int64_t OffsetS, OffsetD; 2306 bool OffsetSIsScalable, OffsetDIsScalable; 2307 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 2308 if (!TII->getMemOperandWithOffset(*SI, BaseOpS, OffsetS, OffsetSIsScalable, 2309 TRI) || 2310 !TII->getMemOperandWithOffset(*DI, BaseOpD, OffsetD, OffsetDIsScalable, 2311 TRI)) 2312 return true; 2313 2314 assert(!OffsetSIsScalable && !OffsetDIsScalable && 2315 "Expected offsets to be byte offsets"); 2316 2317 if (!BaseOpS->isIdenticalTo(*BaseOpD)) 2318 return true; 2319 2320 // Check that the base register is incremented by a constant value for each 2321 // iteration. 2322 MachineInstr *Def = MRI.getVRegDef(BaseOpS->getReg()); 2323 if (!Def || !Def->isPHI()) 2324 return true; 2325 unsigned InitVal = 0; 2326 unsigned LoopVal = 0; 2327 getPhiRegs(*Def, BB, InitVal, LoopVal); 2328 MachineInstr *LoopDef = MRI.getVRegDef(LoopVal); 2329 int D = 0; 2330 if (!LoopDef || !TII->getIncrementValue(*LoopDef, D)) 2331 return true; 2332 2333 uint64_t AccessSizeS = (*SI->memoperands_begin())->getSize(); 2334 uint64_t AccessSizeD = (*DI->memoperands_begin())->getSize(); 2335 2336 // This is the main test, which checks the offset values and the loop 2337 // increment value to determine if the accesses may be loop carried. 2338 if (AccessSizeS == MemoryLocation::UnknownSize || 2339 AccessSizeD == MemoryLocation::UnknownSize) 2340 return true; 2341 2342 if (DeltaS != DeltaD || DeltaS < AccessSizeS || DeltaD < AccessSizeD) 2343 return true; 2344 2345 return (OffsetS + (int64_t)AccessSizeS < OffsetD + (int64_t)AccessSizeD); 2346 } 2347 2348 void SwingSchedulerDAG::postprocessDAG() { 2349 for (auto &M : Mutations) 2350 M->apply(this); 2351 } 2352 2353 /// Try to schedule the node at the specified StartCycle and continue 2354 /// until the node is schedule or the EndCycle is reached. This function 2355 /// returns true if the node is scheduled. This routine may search either 2356 /// forward or backward for a place to insert the instruction based upon 2357 /// the relative values of StartCycle and EndCycle. 2358 bool SMSchedule::insert(SUnit *SU, int StartCycle, int EndCycle, int II) { 2359 bool forward = true; 2360 LLVM_DEBUG({ 2361 dbgs() << "Trying to insert node between " << StartCycle << " and " 2362 << EndCycle << " II: " << II << "\n"; 2363 }); 2364 if (StartCycle > EndCycle) 2365 forward = false; 2366 2367 // The terminating condition depends on the direction. 2368 int termCycle = forward ? EndCycle + 1 : EndCycle - 1; 2369 for (int curCycle = StartCycle; curCycle != termCycle; 2370 forward ? ++curCycle : --curCycle) { 2371 2372 // Add the already scheduled instructions at the specified cycle to the 2373 // DFA. 2374 ProcItinResources.clearResources(); 2375 for (int checkCycle = FirstCycle + ((curCycle - FirstCycle) % II); 2376 checkCycle <= LastCycle; checkCycle += II) { 2377 std::deque<SUnit *> &cycleInstrs = ScheduledInstrs[checkCycle]; 2378 2379 for (SUnit *CI : cycleInstrs) { 2380 if (ST.getInstrInfo()->isZeroCost(CI->getInstr()->getOpcode())) 2381 continue; 2382 assert(ProcItinResources.canReserveResources(*CI->getInstr()) && 2383 "These instructions have already been scheduled."); 2384 ProcItinResources.reserveResources(*CI->getInstr()); 2385 } 2386 } 2387 if (ST.getInstrInfo()->isZeroCost(SU->getInstr()->getOpcode()) || 2388 ProcItinResources.canReserveResources(*SU->getInstr())) { 2389 LLVM_DEBUG({ 2390 dbgs() << "\tinsert at cycle " << curCycle << " "; 2391 SU->getInstr()->dump(); 2392 }); 2393 2394 ScheduledInstrs[curCycle].push_back(SU); 2395 InstrToCycle.insert(std::make_pair(SU, curCycle)); 2396 if (curCycle > LastCycle) 2397 LastCycle = curCycle; 2398 if (curCycle < FirstCycle) 2399 FirstCycle = curCycle; 2400 return true; 2401 } 2402 LLVM_DEBUG({ 2403 dbgs() << "\tfailed to insert at cycle " << curCycle << " "; 2404 SU->getInstr()->dump(); 2405 }); 2406 } 2407 return false; 2408 } 2409 2410 // Return the cycle of the earliest scheduled instruction in the chain. 2411 int SMSchedule::earliestCycleInChain(const SDep &Dep) { 2412 SmallPtrSet<SUnit *, 8> Visited; 2413 SmallVector<SDep, 8> Worklist; 2414 Worklist.push_back(Dep); 2415 int EarlyCycle = INT_MAX; 2416 while (!Worklist.empty()) { 2417 const SDep &Cur = Worklist.pop_back_val(); 2418 SUnit *PrevSU = Cur.getSUnit(); 2419 if (Visited.count(PrevSU)) 2420 continue; 2421 std::map<SUnit *, int>::const_iterator it = InstrToCycle.find(PrevSU); 2422 if (it == InstrToCycle.end()) 2423 continue; 2424 EarlyCycle = std::min(EarlyCycle, it->second); 2425 for (const auto &PI : PrevSU->Preds) 2426 if (PI.getKind() == SDep::Order || PI.getKind() == SDep::Output) 2427 Worklist.push_back(PI); 2428 Visited.insert(PrevSU); 2429 } 2430 return EarlyCycle; 2431 } 2432 2433 // Return the cycle of the latest scheduled instruction in the chain. 2434 int SMSchedule::latestCycleInChain(const SDep &Dep) { 2435 SmallPtrSet<SUnit *, 8> Visited; 2436 SmallVector<SDep, 8> Worklist; 2437 Worklist.push_back(Dep); 2438 int LateCycle = INT_MIN; 2439 while (!Worklist.empty()) { 2440 const SDep &Cur = Worklist.pop_back_val(); 2441 SUnit *SuccSU = Cur.getSUnit(); 2442 if (Visited.count(SuccSU) || SuccSU->isBoundaryNode()) 2443 continue; 2444 std::map<SUnit *, int>::const_iterator it = InstrToCycle.find(SuccSU); 2445 if (it == InstrToCycle.end()) 2446 continue; 2447 LateCycle = std::max(LateCycle, it->second); 2448 for (const auto &SI : SuccSU->Succs) 2449 if (SI.getKind() == SDep::Order || SI.getKind() == SDep::Output) 2450 Worklist.push_back(SI); 2451 Visited.insert(SuccSU); 2452 } 2453 return LateCycle; 2454 } 2455 2456 /// If an instruction has a use that spans multiple iterations, then 2457 /// return true. These instructions are characterized by having a back-ege 2458 /// to a Phi, which contains a reference to another Phi. 2459 static SUnit *multipleIterations(SUnit *SU, SwingSchedulerDAG *DAG) { 2460 for (auto &P : SU->Preds) 2461 if (DAG->isBackedge(SU, P) && P.getSUnit()->getInstr()->isPHI()) 2462 for (auto &S : P.getSUnit()->Succs) 2463 if (S.getKind() == SDep::Data && S.getSUnit()->getInstr()->isPHI()) 2464 return P.getSUnit(); 2465 return nullptr; 2466 } 2467 2468 /// Compute the scheduling start slot for the instruction. The start slot 2469 /// depends on any predecessor or successor nodes scheduled already. 2470 void SMSchedule::computeStart(SUnit *SU, int *MaxEarlyStart, int *MinLateStart, 2471 int *MinEnd, int *MaxStart, int II, 2472 SwingSchedulerDAG *DAG) { 2473 // Iterate over each instruction that has been scheduled already. The start 2474 // slot computation depends on whether the previously scheduled instruction 2475 // is a predecessor or successor of the specified instruction. 2476 for (int cycle = getFirstCycle(); cycle <= LastCycle; ++cycle) { 2477 2478 // Iterate over each instruction in the current cycle. 2479 for (SUnit *I : getInstructions(cycle)) { 2480 // Because we're processing a DAG for the dependences, we recognize 2481 // the back-edge in recurrences by anti dependences. 2482 for (unsigned i = 0, e = (unsigned)SU->Preds.size(); i != e; ++i) { 2483 const SDep &Dep = SU->Preds[i]; 2484 if (Dep.getSUnit() == I) { 2485 if (!DAG->isBackedge(SU, Dep)) { 2486 int EarlyStart = cycle + Dep.getLatency() - 2487 DAG->getDistance(Dep.getSUnit(), SU, Dep) * II; 2488 *MaxEarlyStart = std::max(*MaxEarlyStart, EarlyStart); 2489 if (DAG->isLoopCarriedDep(SU, Dep, false)) { 2490 int End = earliestCycleInChain(Dep) + (II - 1); 2491 *MinEnd = std::min(*MinEnd, End); 2492 } 2493 } else { 2494 int LateStart = cycle - Dep.getLatency() + 2495 DAG->getDistance(SU, Dep.getSUnit(), Dep) * II; 2496 *MinLateStart = std::min(*MinLateStart, LateStart); 2497 } 2498 } 2499 // For instruction that requires multiple iterations, make sure that 2500 // the dependent instruction is not scheduled past the definition. 2501 SUnit *BE = multipleIterations(I, DAG); 2502 if (BE && Dep.getSUnit() == BE && !SU->getInstr()->isPHI() && 2503 !SU->isPred(I)) 2504 *MinLateStart = std::min(*MinLateStart, cycle); 2505 } 2506 for (unsigned i = 0, e = (unsigned)SU->Succs.size(); i != e; ++i) { 2507 if (SU->Succs[i].getSUnit() == I) { 2508 const SDep &Dep = SU->Succs[i]; 2509 if (!DAG->isBackedge(SU, Dep)) { 2510 int LateStart = cycle - Dep.getLatency() + 2511 DAG->getDistance(SU, Dep.getSUnit(), Dep) * II; 2512 *MinLateStart = std::min(*MinLateStart, LateStart); 2513 if (DAG->isLoopCarriedDep(SU, Dep)) { 2514 int Start = latestCycleInChain(Dep) + 1 - II; 2515 *MaxStart = std::max(*MaxStart, Start); 2516 } 2517 } else { 2518 int EarlyStart = cycle + Dep.getLatency() - 2519 DAG->getDistance(Dep.getSUnit(), SU, Dep) * II; 2520 *MaxEarlyStart = std::max(*MaxEarlyStart, EarlyStart); 2521 } 2522 } 2523 } 2524 } 2525 } 2526 } 2527 2528 /// Order the instructions within a cycle so that the definitions occur 2529 /// before the uses. Returns true if the instruction is added to the start 2530 /// of the list, or false if added to the end. 2531 void SMSchedule::orderDependence(SwingSchedulerDAG *SSD, SUnit *SU, 2532 std::deque<SUnit *> &Insts) { 2533 MachineInstr *MI = SU->getInstr(); 2534 bool OrderBeforeUse = false; 2535 bool OrderAfterDef = false; 2536 bool OrderBeforeDef = false; 2537 unsigned MoveDef = 0; 2538 unsigned MoveUse = 0; 2539 int StageInst1 = stageScheduled(SU); 2540 2541 unsigned Pos = 0; 2542 for (std::deque<SUnit *>::iterator I = Insts.begin(), E = Insts.end(); I != E; 2543 ++I, ++Pos) { 2544 for (MachineOperand &MO : MI->operands()) { 2545 if (!MO.isReg() || !Register::isVirtualRegister(MO.getReg())) 2546 continue; 2547 2548 Register Reg = MO.getReg(); 2549 unsigned BasePos, OffsetPos; 2550 if (ST.getInstrInfo()->getBaseAndOffsetPosition(*MI, BasePos, OffsetPos)) 2551 if (MI->getOperand(BasePos).getReg() == Reg) 2552 if (unsigned NewReg = SSD->getInstrBaseReg(SU)) 2553 Reg = NewReg; 2554 bool Reads, Writes; 2555 std::tie(Reads, Writes) = 2556 (*I)->getInstr()->readsWritesVirtualRegister(Reg); 2557 if (MO.isDef() && Reads && stageScheduled(*I) <= StageInst1) { 2558 OrderBeforeUse = true; 2559 if (MoveUse == 0) 2560 MoveUse = Pos; 2561 } else if (MO.isDef() && Reads && stageScheduled(*I) > StageInst1) { 2562 // Add the instruction after the scheduled instruction. 2563 OrderAfterDef = true; 2564 MoveDef = Pos; 2565 } else if (MO.isUse() && Writes && stageScheduled(*I) == StageInst1) { 2566 if (cycleScheduled(*I) == cycleScheduled(SU) && !(*I)->isSucc(SU)) { 2567 OrderBeforeUse = true; 2568 if (MoveUse == 0) 2569 MoveUse = Pos; 2570 } else { 2571 OrderAfterDef = true; 2572 MoveDef = Pos; 2573 } 2574 } else if (MO.isUse() && Writes && stageScheduled(*I) > StageInst1) { 2575 OrderBeforeUse = true; 2576 if (MoveUse == 0) 2577 MoveUse = Pos; 2578 if (MoveUse != 0) { 2579 OrderAfterDef = true; 2580 MoveDef = Pos - 1; 2581 } 2582 } else if (MO.isUse() && Writes && stageScheduled(*I) < StageInst1) { 2583 // Add the instruction before the scheduled instruction. 2584 OrderBeforeUse = true; 2585 if (MoveUse == 0) 2586 MoveUse = Pos; 2587 } else if (MO.isUse() && stageScheduled(*I) == StageInst1 && 2588 isLoopCarriedDefOfUse(SSD, (*I)->getInstr(), MO)) { 2589 if (MoveUse == 0) { 2590 OrderBeforeDef = true; 2591 MoveUse = Pos; 2592 } 2593 } 2594 } 2595 // Check for order dependences between instructions. Make sure the source 2596 // is ordered before the destination. 2597 for (auto &S : SU->Succs) { 2598 if (S.getSUnit() != *I) 2599 continue; 2600 if (S.getKind() == SDep::Order && stageScheduled(*I) == StageInst1) { 2601 OrderBeforeUse = true; 2602 if (Pos < MoveUse) 2603 MoveUse = Pos; 2604 } 2605 // We did not handle HW dependences in previous for loop, 2606 // and we normally set Latency = 0 for Anti deps, 2607 // so may have nodes in same cycle with Anti denpendent on HW regs. 2608 else if (S.getKind() == SDep::Anti && stageScheduled(*I) == StageInst1) { 2609 OrderBeforeUse = true; 2610 if ((MoveUse == 0) || (Pos < MoveUse)) 2611 MoveUse = Pos; 2612 } 2613 } 2614 for (auto &P : SU->Preds) { 2615 if (P.getSUnit() != *I) 2616 continue; 2617 if (P.getKind() == SDep::Order && stageScheduled(*I) == StageInst1) { 2618 OrderAfterDef = true; 2619 MoveDef = Pos; 2620 } 2621 } 2622 } 2623 2624 // A circular dependence. 2625 if (OrderAfterDef && OrderBeforeUse && MoveUse == MoveDef) 2626 OrderBeforeUse = false; 2627 2628 // OrderAfterDef takes precedences over OrderBeforeDef. The latter is due 2629 // to a loop-carried dependence. 2630 if (OrderBeforeDef) 2631 OrderBeforeUse = !OrderAfterDef || (MoveUse > MoveDef); 2632 2633 // The uncommon case when the instruction order needs to be updated because 2634 // there is both a use and def. 2635 if (OrderBeforeUse && OrderAfterDef) { 2636 SUnit *UseSU = Insts.at(MoveUse); 2637 SUnit *DefSU = Insts.at(MoveDef); 2638 if (MoveUse > MoveDef) { 2639 Insts.erase(Insts.begin() + MoveUse); 2640 Insts.erase(Insts.begin() + MoveDef); 2641 } else { 2642 Insts.erase(Insts.begin() + MoveDef); 2643 Insts.erase(Insts.begin() + MoveUse); 2644 } 2645 orderDependence(SSD, UseSU, Insts); 2646 orderDependence(SSD, SU, Insts); 2647 orderDependence(SSD, DefSU, Insts); 2648 return; 2649 } 2650 // Put the new instruction first if there is a use in the list. Otherwise, 2651 // put it at the end of the list. 2652 if (OrderBeforeUse) 2653 Insts.push_front(SU); 2654 else 2655 Insts.push_back(SU); 2656 } 2657 2658 /// Return true if the scheduled Phi has a loop carried operand. 2659 bool SMSchedule::isLoopCarried(SwingSchedulerDAG *SSD, MachineInstr &Phi) { 2660 if (!Phi.isPHI()) 2661 return false; 2662 assert(Phi.isPHI() && "Expecting a Phi."); 2663 SUnit *DefSU = SSD->getSUnit(&Phi); 2664 unsigned DefCycle = cycleScheduled(DefSU); 2665 int DefStage = stageScheduled(DefSU); 2666 2667 unsigned InitVal = 0; 2668 unsigned LoopVal = 0; 2669 getPhiRegs(Phi, Phi.getParent(), InitVal, LoopVal); 2670 SUnit *UseSU = SSD->getSUnit(MRI.getVRegDef(LoopVal)); 2671 if (!UseSU) 2672 return true; 2673 if (UseSU->getInstr()->isPHI()) 2674 return true; 2675 unsigned LoopCycle = cycleScheduled(UseSU); 2676 int LoopStage = stageScheduled(UseSU); 2677 return (LoopCycle > DefCycle) || (LoopStage <= DefStage); 2678 } 2679 2680 /// Return true if the instruction is a definition that is loop carried 2681 /// and defines the use on the next iteration. 2682 /// v1 = phi(v2, v3) 2683 /// (Def) v3 = op v1 2684 /// (MO) = v1 2685 /// If MO appears before Def, then then v1 and v3 may get assigned to the same 2686 /// register. 2687 bool SMSchedule::isLoopCarriedDefOfUse(SwingSchedulerDAG *SSD, 2688 MachineInstr *Def, MachineOperand &MO) { 2689 if (!MO.isReg()) 2690 return false; 2691 if (Def->isPHI()) 2692 return false; 2693 MachineInstr *Phi = MRI.getVRegDef(MO.getReg()); 2694 if (!Phi || !Phi->isPHI() || Phi->getParent() != Def->getParent()) 2695 return false; 2696 if (!isLoopCarried(SSD, *Phi)) 2697 return false; 2698 unsigned LoopReg = getLoopPhiReg(*Phi, Phi->getParent()); 2699 for (unsigned i = 0, e = Def->getNumOperands(); i != e; ++i) { 2700 MachineOperand &DMO = Def->getOperand(i); 2701 if (!DMO.isReg() || !DMO.isDef()) 2702 continue; 2703 if (DMO.getReg() == LoopReg) 2704 return true; 2705 } 2706 return false; 2707 } 2708 2709 /// Determine transitive dependences of unpipelineable instructions 2710 SmallSet<SUnit *, 8> SMSchedule::computeUnpipelineableNodes( 2711 SwingSchedulerDAG *SSD, TargetInstrInfo::PipelinerLoopInfo *PLI) { 2712 SmallSet<SUnit *, 8> DoNotPipeline; 2713 SmallVector<SUnit *, 8> Worklist; 2714 2715 for (auto &SU : SSD->SUnits) 2716 if (SU.isInstr() && PLI->shouldIgnoreForPipelining(SU.getInstr())) 2717 Worklist.push_back(&SU); 2718 2719 while (!Worklist.empty()) { 2720 auto SU = Worklist.pop_back_val(); 2721 if (DoNotPipeline.count(SU)) 2722 continue; 2723 LLVM_DEBUG(dbgs() << "Do not pipeline SU(" << SU->NodeNum << ")\n"); 2724 DoNotPipeline.insert(SU); 2725 for (auto &Dep : SU->Preds) 2726 Worklist.push_back(Dep.getSUnit()); 2727 if (SU->getInstr()->isPHI()) 2728 for (auto &Dep : SU->Succs) 2729 if (Dep.getKind() == SDep::Anti) 2730 Worklist.push_back(Dep.getSUnit()); 2731 } 2732 return DoNotPipeline; 2733 } 2734 2735 // Determine all instructions upon which any unpipelineable instruction depends 2736 // and ensure that they are in stage 0. If unable to do so, return false. 2737 bool SMSchedule::normalizeNonPipelinedInstructions( 2738 SwingSchedulerDAG *SSD, TargetInstrInfo::PipelinerLoopInfo *PLI) { 2739 SmallSet<SUnit *, 8> DNP = computeUnpipelineableNodes(SSD, PLI); 2740 2741 int NewLastCycle = INT_MIN; 2742 for (SUnit &SU : SSD->SUnits) { 2743 if (!SU.isInstr()) 2744 continue; 2745 if (!DNP.contains(&SU) || stageScheduled(&SU) == 0) { 2746 NewLastCycle = std::max(NewLastCycle, InstrToCycle[&SU]); 2747 continue; 2748 } 2749 2750 // Put the non-pipelined instruction as early as possible in the schedule 2751 int NewCycle = getFirstCycle(); 2752 for (auto &Dep : SU.Preds) 2753 NewCycle = std::max(InstrToCycle[Dep.getSUnit()], NewCycle); 2754 2755 int OldCycle = InstrToCycle[&SU]; 2756 if (OldCycle != NewCycle) { 2757 InstrToCycle[&SU] = NewCycle; 2758 auto &OldS = getInstructions(OldCycle); 2759 llvm::erase_value(OldS, &SU); 2760 getInstructions(NewCycle).emplace_back(&SU); 2761 LLVM_DEBUG(dbgs() << "SU(" << SU.NodeNum 2762 << ") is not pipelined; moving from cycle " << OldCycle 2763 << " to " << NewCycle << " Instr:" << *SU.getInstr()); 2764 } 2765 NewLastCycle = std::max(NewLastCycle, NewCycle); 2766 } 2767 LastCycle = NewLastCycle; 2768 return true; 2769 } 2770 2771 // Check if the generated schedule is valid. This function checks if 2772 // an instruction that uses a physical register is scheduled in a 2773 // different stage than the definition. The pipeliner does not handle 2774 // physical register values that may cross a basic block boundary. 2775 // Furthermore, if a physical def/use pair is assigned to the same 2776 // cycle, orderDependence does not guarantee def/use ordering, so that 2777 // case should be considered invalid. (The test checks for both 2778 // earlier and same-cycle use to be more robust.) 2779 bool SMSchedule::isValidSchedule(SwingSchedulerDAG *SSD) { 2780 for (SUnit &SU : SSD->SUnits) { 2781 if (!SU.hasPhysRegDefs) 2782 continue; 2783 int StageDef = stageScheduled(&SU); 2784 int CycleDef = InstrToCycle[&SU]; 2785 assert(StageDef != -1 && "Instruction should have been scheduled."); 2786 for (auto &SI : SU.Succs) 2787 if (SI.isAssignedRegDep() && !SI.getSUnit()->isBoundaryNode()) 2788 if (Register::isPhysicalRegister(SI.getReg())) { 2789 if (stageScheduled(SI.getSUnit()) != StageDef) 2790 return false; 2791 if (InstrToCycle[SI.getSUnit()] <= CycleDef) 2792 return false; 2793 } 2794 } 2795 return true; 2796 } 2797 2798 /// A property of the node order in swing-modulo-scheduling is 2799 /// that for nodes outside circuits the following holds: 2800 /// none of them is scheduled after both a successor and a 2801 /// predecessor. 2802 /// The method below checks whether the property is met. 2803 /// If not, debug information is printed and statistics information updated. 2804 /// Note that we do not use an assert statement. 2805 /// The reason is that although an invalid node oder may prevent 2806 /// the pipeliner from finding a pipelined schedule for arbitrary II, 2807 /// it does not lead to the generation of incorrect code. 2808 void SwingSchedulerDAG::checkValidNodeOrder(const NodeSetType &Circuits) const { 2809 2810 // a sorted vector that maps each SUnit to its index in the NodeOrder 2811 typedef std::pair<SUnit *, unsigned> UnitIndex; 2812 std::vector<UnitIndex> Indices(NodeOrder.size(), std::make_pair(nullptr, 0)); 2813 2814 for (unsigned i = 0, s = NodeOrder.size(); i < s; ++i) 2815 Indices.push_back(std::make_pair(NodeOrder[i], i)); 2816 2817 auto CompareKey = [](UnitIndex i1, UnitIndex i2) { 2818 return std::get<0>(i1) < std::get<0>(i2); 2819 }; 2820 2821 // sort, so that we can perform a binary search 2822 llvm::sort(Indices, CompareKey); 2823 2824 bool Valid = true; 2825 (void)Valid; 2826 // for each SUnit in the NodeOrder, check whether 2827 // it appears after both a successor and a predecessor 2828 // of the SUnit. If this is the case, and the SUnit 2829 // is not part of circuit, then the NodeOrder is not 2830 // valid. 2831 for (unsigned i = 0, s = NodeOrder.size(); i < s; ++i) { 2832 SUnit *SU = NodeOrder[i]; 2833 unsigned Index = i; 2834 2835 bool PredBefore = false; 2836 bool SuccBefore = false; 2837 2838 SUnit *Succ; 2839 SUnit *Pred; 2840 (void)Succ; 2841 (void)Pred; 2842 2843 for (SDep &PredEdge : SU->Preds) { 2844 SUnit *PredSU = PredEdge.getSUnit(); 2845 unsigned PredIndex = std::get<1>( 2846 *llvm::lower_bound(Indices, std::make_pair(PredSU, 0), CompareKey)); 2847 if (!PredSU->getInstr()->isPHI() && PredIndex < Index) { 2848 PredBefore = true; 2849 Pred = PredSU; 2850 break; 2851 } 2852 } 2853 2854 for (SDep &SuccEdge : SU->Succs) { 2855 SUnit *SuccSU = SuccEdge.getSUnit(); 2856 // Do not process a boundary node, it was not included in NodeOrder, 2857 // hence not in Indices either, call to std::lower_bound() below will 2858 // return Indices.end(). 2859 if (SuccSU->isBoundaryNode()) 2860 continue; 2861 unsigned SuccIndex = std::get<1>( 2862 *llvm::lower_bound(Indices, std::make_pair(SuccSU, 0), CompareKey)); 2863 if (!SuccSU->getInstr()->isPHI() && SuccIndex < Index) { 2864 SuccBefore = true; 2865 Succ = SuccSU; 2866 break; 2867 } 2868 } 2869 2870 if (PredBefore && SuccBefore && !SU->getInstr()->isPHI()) { 2871 // instructions in circuits are allowed to be scheduled 2872 // after both a successor and predecessor. 2873 bool InCircuit = llvm::any_of( 2874 Circuits, [SU](const NodeSet &Circuit) { return Circuit.count(SU); }); 2875 if (InCircuit) 2876 LLVM_DEBUG(dbgs() << "In a circuit, predecessor ";); 2877 else { 2878 Valid = false; 2879 NumNodeOrderIssues++; 2880 LLVM_DEBUG(dbgs() << "Predecessor ";); 2881 } 2882 LLVM_DEBUG(dbgs() << Pred->NodeNum << " and successor " << Succ->NodeNum 2883 << " are scheduled before node " << SU->NodeNum 2884 << "\n";); 2885 } 2886 } 2887 2888 LLVM_DEBUG({ 2889 if (!Valid) 2890 dbgs() << "Invalid node order found!\n"; 2891 }); 2892 } 2893 2894 /// Attempt to fix the degenerate cases when the instruction serialization 2895 /// causes the register lifetimes to overlap. For example, 2896 /// p' = store_pi(p, b) 2897 /// = load p, offset 2898 /// In this case p and p' overlap, which means that two registers are needed. 2899 /// Instead, this function changes the load to use p' and updates the offset. 2900 void SwingSchedulerDAG::fixupRegisterOverlaps(std::deque<SUnit *> &Instrs) { 2901 unsigned OverlapReg = 0; 2902 unsigned NewBaseReg = 0; 2903 for (SUnit *SU : Instrs) { 2904 MachineInstr *MI = SU->getInstr(); 2905 for (unsigned i = 0, e = MI->getNumOperands(); i < e; ++i) { 2906 const MachineOperand &MO = MI->getOperand(i); 2907 // Look for an instruction that uses p. The instruction occurs in the 2908 // same cycle but occurs later in the serialized order. 2909 if (MO.isReg() && MO.isUse() && MO.getReg() == OverlapReg) { 2910 // Check that the instruction appears in the InstrChanges structure, 2911 // which contains instructions that can have the offset updated. 2912 DenseMap<SUnit *, std::pair<unsigned, int64_t>>::iterator It = 2913 InstrChanges.find(SU); 2914 if (It != InstrChanges.end()) { 2915 unsigned BasePos, OffsetPos; 2916 // Update the base register and adjust the offset. 2917 if (TII->getBaseAndOffsetPosition(*MI, BasePos, OffsetPos)) { 2918 MachineInstr *NewMI = MF.CloneMachineInstr(MI); 2919 NewMI->getOperand(BasePos).setReg(NewBaseReg); 2920 int64_t NewOffset = 2921 MI->getOperand(OffsetPos).getImm() - It->second.second; 2922 NewMI->getOperand(OffsetPos).setImm(NewOffset); 2923 SU->setInstr(NewMI); 2924 MISUnitMap[NewMI] = SU; 2925 NewMIs[MI] = NewMI; 2926 } 2927 } 2928 OverlapReg = 0; 2929 NewBaseReg = 0; 2930 break; 2931 } 2932 // Look for an instruction of the form p' = op(p), which uses and defines 2933 // two virtual registers that get allocated to the same physical register. 2934 unsigned TiedUseIdx = 0; 2935 if (MI->isRegTiedToUseOperand(i, &TiedUseIdx)) { 2936 // OverlapReg is p in the example above. 2937 OverlapReg = MI->getOperand(TiedUseIdx).getReg(); 2938 // NewBaseReg is p' in the example above. 2939 NewBaseReg = MI->getOperand(i).getReg(); 2940 break; 2941 } 2942 } 2943 } 2944 } 2945 2946 /// After the schedule has been formed, call this function to combine 2947 /// the instructions from the different stages/cycles. That is, this 2948 /// function creates a schedule that represents a single iteration. 2949 void SMSchedule::finalizeSchedule(SwingSchedulerDAG *SSD) { 2950 // Move all instructions to the first stage from later stages. 2951 for (int cycle = getFirstCycle(); cycle <= getFinalCycle(); ++cycle) { 2952 for (int stage = 1, lastStage = getMaxStageCount(); stage <= lastStage; 2953 ++stage) { 2954 std::deque<SUnit *> &cycleInstrs = 2955 ScheduledInstrs[cycle + (stage * InitiationInterval)]; 2956 for (SUnit *SU : llvm::reverse(cycleInstrs)) 2957 ScheduledInstrs[cycle].push_front(SU); 2958 } 2959 } 2960 2961 // Erase all the elements in the later stages. Only one iteration should 2962 // remain in the scheduled list, and it contains all the instructions. 2963 for (int cycle = getFinalCycle() + 1; cycle <= LastCycle; ++cycle) 2964 ScheduledInstrs.erase(cycle); 2965 2966 // Change the registers in instruction as specified in the InstrChanges 2967 // map. We need to use the new registers to create the correct order. 2968 for (const SUnit &SU : SSD->SUnits) 2969 SSD->applyInstrChange(SU.getInstr(), *this); 2970 2971 // Reorder the instructions in each cycle to fix and improve the 2972 // generated code. 2973 for (int Cycle = getFirstCycle(), E = getFinalCycle(); Cycle <= E; ++Cycle) { 2974 std::deque<SUnit *> &cycleInstrs = ScheduledInstrs[Cycle]; 2975 std::deque<SUnit *> newOrderPhi; 2976 for (SUnit *SU : cycleInstrs) { 2977 if (SU->getInstr()->isPHI()) 2978 newOrderPhi.push_back(SU); 2979 } 2980 std::deque<SUnit *> newOrderI; 2981 for (SUnit *SU : cycleInstrs) { 2982 if (!SU->getInstr()->isPHI()) 2983 orderDependence(SSD, SU, newOrderI); 2984 } 2985 // Replace the old order with the new order. 2986 cycleInstrs.swap(newOrderPhi); 2987 llvm::append_range(cycleInstrs, newOrderI); 2988 SSD->fixupRegisterOverlaps(cycleInstrs); 2989 } 2990 2991 LLVM_DEBUG(dump();); 2992 } 2993 2994 void NodeSet::print(raw_ostream &os) const { 2995 os << "Num nodes " << size() << " rec " << RecMII << " mov " << MaxMOV 2996 << " depth " << MaxDepth << " col " << Colocate << "\n"; 2997 for (const auto &I : Nodes) 2998 os << " SU(" << I->NodeNum << ") " << *(I->getInstr()); 2999 os << "\n"; 3000 } 3001 3002 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 3003 /// Print the schedule information to the given output. 3004 void SMSchedule::print(raw_ostream &os) const { 3005 // Iterate over each cycle. 3006 for (int cycle = getFirstCycle(); cycle <= getFinalCycle(); ++cycle) { 3007 // Iterate over each instruction in the cycle. 3008 const_sched_iterator cycleInstrs = ScheduledInstrs.find(cycle); 3009 for (SUnit *CI : cycleInstrs->second) { 3010 os << "cycle " << cycle << " (" << stageScheduled(CI) << ") "; 3011 os << "(" << CI->NodeNum << ") "; 3012 CI->getInstr()->print(os); 3013 os << "\n"; 3014 } 3015 } 3016 } 3017 3018 /// Utility function used for debugging to print the schedule. 3019 LLVM_DUMP_METHOD void SMSchedule::dump() const { print(dbgs()); } 3020 LLVM_DUMP_METHOD void NodeSet::dump() const { print(dbgs()); } 3021 3022 #endif 3023 3024 void ResourceManager::initProcResourceVectors( 3025 const MCSchedModel &SM, SmallVectorImpl<uint64_t> &Masks) { 3026 unsigned ProcResourceID = 0; 3027 3028 // We currently limit the resource kinds to 64 and below so that we can use 3029 // uint64_t for Masks 3030 assert(SM.getNumProcResourceKinds() < 64 && 3031 "Too many kinds of resources, unsupported"); 3032 // Create a unique bitmask for every processor resource unit. 3033 // Skip resource at index 0, since it always references 'InvalidUnit'. 3034 Masks.resize(SM.getNumProcResourceKinds()); 3035 for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) { 3036 const MCProcResourceDesc &Desc = *SM.getProcResource(I); 3037 if (Desc.SubUnitsIdxBegin) 3038 continue; 3039 Masks[I] = 1ULL << ProcResourceID; 3040 ProcResourceID++; 3041 } 3042 // Create a unique bitmask for every processor resource group. 3043 for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) { 3044 const MCProcResourceDesc &Desc = *SM.getProcResource(I); 3045 if (!Desc.SubUnitsIdxBegin) 3046 continue; 3047 Masks[I] = 1ULL << ProcResourceID; 3048 for (unsigned U = 0; U < Desc.NumUnits; ++U) 3049 Masks[I] |= Masks[Desc.SubUnitsIdxBegin[U]]; 3050 ProcResourceID++; 3051 } 3052 LLVM_DEBUG({ 3053 if (SwpShowResMask) { 3054 dbgs() << "ProcResourceDesc:\n"; 3055 for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) { 3056 const MCProcResourceDesc *ProcResource = SM.getProcResource(I); 3057 dbgs() << format(" %16s(%2d): Mask: 0x%08x, NumUnits:%2d\n", 3058 ProcResource->Name, I, Masks[I], 3059 ProcResource->NumUnits); 3060 } 3061 dbgs() << " -----------------\n"; 3062 } 3063 }); 3064 } 3065 3066 bool ResourceManager::canReserveResources(const MCInstrDesc *MID) const { 3067 3068 LLVM_DEBUG({ 3069 if (SwpDebugResource) 3070 dbgs() << "canReserveResources:\n"; 3071 }); 3072 if (UseDFA) 3073 return DFAResources->canReserveResources(MID); 3074 3075 unsigned InsnClass = MID->getSchedClass(); 3076 const MCSchedClassDesc *SCDesc = SM.getSchedClassDesc(InsnClass); 3077 if (!SCDesc->isValid()) { 3078 LLVM_DEBUG({ 3079 dbgs() << "No valid Schedule Class Desc for schedClass!\n"; 3080 dbgs() << "isPseudo:" << MID->isPseudo() << "\n"; 3081 }); 3082 return true; 3083 } 3084 3085 const MCWriteProcResEntry *I = STI->getWriteProcResBegin(SCDesc); 3086 const MCWriteProcResEntry *E = STI->getWriteProcResEnd(SCDesc); 3087 for (; I != E; ++I) { 3088 if (!I->Cycles) 3089 continue; 3090 const MCProcResourceDesc *ProcResource = 3091 SM.getProcResource(I->ProcResourceIdx); 3092 unsigned NumUnits = ProcResource->NumUnits; 3093 LLVM_DEBUG({ 3094 if (SwpDebugResource) 3095 dbgs() << format(" %16s(%2d): Count: %2d, NumUnits:%2d, Cycles:%2d\n", 3096 ProcResource->Name, I->ProcResourceIdx, 3097 ProcResourceCount[I->ProcResourceIdx], NumUnits, 3098 I->Cycles); 3099 }); 3100 if (ProcResourceCount[I->ProcResourceIdx] >= NumUnits) 3101 return false; 3102 } 3103 LLVM_DEBUG(if (SwpDebugResource) dbgs() << "return true\n\n";); 3104 return true; 3105 } 3106 3107 void ResourceManager::reserveResources(const MCInstrDesc *MID) { 3108 LLVM_DEBUG({ 3109 if (SwpDebugResource) 3110 dbgs() << "reserveResources:\n"; 3111 }); 3112 if (UseDFA) 3113 return DFAResources->reserveResources(MID); 3114 3115 unsigned InsnClass = MID->getSchedClass(); 3116 const MCSchedClassDesc *SCDesc = SM.getSchedClassDesc(InsnClass); 3117 if (!SCDesc->isValid()) { 3118 LLVM_DEBUG({ 3119 dbgs() << "No valid Schedule Class Desc for schedClass!\n"; 3120 dbgs() << "isPseudo:" << MID->isPseudo() << "\n"; 3121 }); 3122 return; 3123 } 3124 for (const MCWriteProcResEntry &PRE : 3125 make_range(STI->getWriteProcResBegin(SCDesc), 3126 STI->getWriteProcResEnd(SCDesc))) { 3127 if (!PRE.Cycles) 3128 continue; 3129 ++ProcResourceCount[PRE.ProcResourceIdx]; 3130 LLVM_DEBUG({ 3131 if (SwpDebugResource) { 3132 const MCProcResourceDesc *ProcResource = 3133 SM.getProcResource(PRE.ProcResourceIdx); 3134 dbgs() << format(" %16s(%2d): Count: %2d, NumUnits:%2d, Cycles:%2d\n", 3135 ProcResource->Name, PRE.ProcResourceIdx, 3136 ProcResourceCount[PRE.ProcResourceIdx], 3137 ProcResource->NumUnits, PRE.Cycles); 3138 } 3139 }); 3140 } 3141 LLVM_DEBUG({ 3142 if (SwpDebugResource) 3143 dbgs() << "reserveResources: done!\n\n"; 3144 }); 3145 } 3146 3147 bool ResourceManager::canReserveResources(const MachineInstr &MI) const { 3148 return canReserveResources(&MI.getDesc()); 3149 } 3150 3151 void ResourceManager::reserveResources(const MachineInstr &MI) { 3152 return reserveResources(&MI.getDesc()); 3153 } 3154 3155 void ResourceManager::clearResources() { 3156 if (UseDFA) 3157 return DFAResources->clearResources(); 3158 std::fill(ProcResourceCount.begin(), ProcResourceCount.end(), 0); 3159 } 3160