1 //===- MachineSink.cpp - Sinking for machine instructions -----------------===// 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 // This pass moves instructions into successor blocks when possible, so that 10 // they aren't executed on paths where their results aren't needed. 11 // 12 // This pass is not intended to be a replacement or a complete alternative 13 // for an LLVM-IR-level sinking pass. It is only designed to sink simple 14 // constructs that are not exposed before lowering and instruction selection. 15 // 16 //===----------------------------------------------------------------------===// 17 18 #include "llvm/ADT/DenseSet.h" 19 #include "llvm/ADT/DepthFirstIterator.h" 20 #include "llvm/ADT/MapVector.h" 21 #include "llvm/ADT/PointerIntPair.h" 22 #include "llvm/ADT/PostOrderIterator.h" 23 #include "llvm/ADT/SetVector.h" 24 #include "llvm/ADT/SmallSet.h" 25 #include "llvm/ADT/SmallVector.h" 26 #include "llvm/ADT/Statistic.h" 27 #include "llvm/Analysis/AliasAnalysis.h" 28 #include "llvm/Analysis/CFG.h" 29 #include "llvm/CodeGen/MachineBasicBlock.h" 30 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h" 31 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h" 32 #include "llvm/CodeGen/MachineCycleAnalysis.h" 33 #include "llvm/CodeGen/MachineDominators.h" 34 #include "llvm/CodeGen/MachineFunction.h" 35 #include "llvm/CodeGen/MachineFunctionPass.h" 36 #include "llvm/CodeGen/MachineInstr.h" 37 #include "llvm/CodeGen/MachineLoopInfo.h" 38 #include "llvm/CodeGen/MachineOperand.h" 39 #include "llvm/CodeGen/MachinePostDominators.h" 40 #include "llvm/CodeGen/MachineRegisterInfo.h" 41 #include "llvm/CodeGen/RegisterClassInfo.h" 42 #include "llvm/CodeGen/RegisterPressure.h" 43 #include "llvm/CodeGen/TargetInstrInfo.h" 44 #include "llvm/CodeGen/TargetPassConfig.h" 45 #include "llvm/CodeGen/TargetRegisterInfo.h" 46 #include "llvm/CodeGen/TargetSubtargetInfo.h" 47 #include "llvm/IR/BasicBlock.h" 48 #include "llvm/IR/DebugInfoMetadata.h" 49 #include "llvm/IR/LLVMContext.h" 50 #include "llvm/InitializePasses.h" 51 #include "llvm/MC/MCRegisterInfo.h" 52 #include "llvm/Pass.h" 53 #include "llvm/Support/BranchProbability.h" 54 #include "llvm/Support/CommandLine.h" 55 #include "llvm/Support/Debug.h" 56 #include "llvm/Support/raw_ostream.h" 57 #include <algorithm> 58 #include <cassert> 59 #include <cstdint> 60 #include <utility> 61 #include <vector> 62 63 using namespace llvm; 64 65 #define DEBUG_TYPE "machine-sink" 66 67 static cl::opt<bool> 68 SplitEdges("machine-sink-split", 69 cl::desc("Split critical edges during machine sinking"), 70 cl::init(true), cl::Hidden); 71 72 static cl::opt<bool> 73 UseBlockFreqInfo("machine-sink-bfi", 74 cl::desc("Use block frequency info to find successors to sink"), 75 cl::init(true), cl::Hidden); 76 77 static cl::opt<unsigned> SplitEdgeProbabilityThreshold( 78 "machine-sink-split-probability-threshold", 79 cl::desc( 80 "Percentage threshold for splitting single-instruction critical edge. " 81 "If the branch threshold is higher than this threshold, we allow " 82 "speculative execution of up to 1 instruction to avoid branching to " 83 "splitted critical edge"), 84 cl::init(40), cl::Hidden); 85 86 static cl::opt<unsigned> SinkLoadInstsPerBlockThreshold( 87 "machine-sink-load-instrs-threshold", 88 cl::desc("Do not try to find alias store for a load if there is a in-path " 89 "block whose instruction number is higher than this threshold."), 90 cl::init(2000), cl::Hidden); 91 92 static cl::opt<unsigned> SinkLoadBlocksThreshold( 93 "machine-sink-load-blocks-threshold", 94 cl::desc("Do not try to find alias store for a load if the block number in " 95 "the straight line is higher than this threshold."), 96 cl::init(20), cl::Hidden); 97 98 static cl::opt<bool> 99 SinkInstsIntoCycle("sink-insts-to-avoid-spills", 100 cl::desc("Sink instructions into cycles to avoid " 101 "register spills"), 102 cl::init(false), cl::Hidden); 103 104 static cl::opt<unsigned> SinkIntoCycleLimit( 105 "machine-sink-cycle-limit", 106 cl::desc("The maximum number of instructions considered for cycle sinking."), 107 cl::init(50), cl::Hidden); 108 109 STATISTIC(NumSunk, "Number of machine instructions sunk"); 110 STATISTIC(NumCycleSunk, "Number of machine instructions sunk into a cycle"); 111 STATISTIC(NumSplit, "Number of critical edges split"); 112 STATISTIC(NumCoalesces, "Number of copies coalesced"); 113 STATISTIC(NumPostRACopySink, "Number of copies sunk after RA"); 114 115 namespace { 116 117 class MachineSinking : public MachineFunctionPass { 118 const TargetSubtargetInfo *STI = nullptr; 119 const TargetInstrInfo *TII = nullptr; 120 const TargetRegisterInfo *TRI = nullptr; 121 MachineRegisterInfo *MRI = nullptr; // Machine register information 122 MachineDominatorTree *DT = nullptr; // Machine dominator tree 123 MachinePostDominatorTree *PDT = nullptr; // Machine post dominator tree 124 MachineCycleInfo *CI = nullptr; 125 MachineBlockFrequencyInfo *MBFI = nullptr; 126 const MachineBranchProbabilityInfo *MBPI = nullptr; 127 AliasAnalysis *AA = nullptr; 128 RegisterClassInfo RegClassInfo; 129 130 // Remember which edges have been considered for breaking. 131 SmallSet<std::pair<MachineBasicBlock*, MachineBasicBlock*>, 8> 132 CEBCandidates; 133 // Memorize the register that also wanted to sink into the same block along 134 // a different critical edge. 135 // {register to sink, sink-to block} -> the first sink-from block. 136 // We're recording the first sink-from block because that (critical) edge 137 // was deferred until we see another register that's going to sink into the 138 // same block. 139 DenseMap<std::pair<Register, MachineBasicBlock *>, MachineBasicBlock *> 140 CEMergeCandidates; 141 // Remember which edges we are about to split. 142 // This is different from CEBCandidates since those edges 143 // will be split. 144 SetVector<std::pair<MachineBasicBlock *, MachineBasicBlock *>> ToSplit; 145 146 DenseSet<Register> RegsToClearKillFlags; 147 148 using AllSuccsCache = 149 SmallDenseMap<MachineBasicBlock *, SmallVector<MachineBasicBlock *, 4>>; 150 151 /// DBG_VALUE pointer and flag. The flag is true if this DBG_VALUE is 152 /// post-dominated by another DBG_VALUE of the same variable location. 153 /// This is necessary to detect sequences such as: 154 /// %0 = someinst 155 /// DBG_VALUE %0, !123, !DIExpression() 156 /// %1 = anotherinst 157 /// DBG_VALUE %1, !123, !DIExpression() 158 /// Where if %0 were to sink, the DBG_VAUE should not sink with it, as that 159 /// would re-order assignments. 160 using SeenDbgUser = PointerIntPair<MachineInstr *, 1>; 161 162 /// Record of DBG_VALUE uses of vregs in a block, so that we can identify 163 /// debug instructions to sink. 164 SmallDenseMap<unsigned, TinyPtrVector<SeenDbgUser>> SeenDbgUsers; 165 166 /// Record of debug variables that have had their locations set in the 167 /// current block. 168 DenseSet<DebugVariable> SeenDbgVars; 169 170 DenseMap<std::pair<MachineBasicBlock *, MachineBasicBlock *>, bool> 171 HasStoreCache; 172 173 DenseMap<std::pair<MachineBasicBlock *, MachineBasicBlock *>, 174 SmallVector<MachineInstr *>> 175 StoreInstrCache; 176 177 /// Cached BB's register pressure. 178 DenseMap<const MachineBasicBlock *, std::vector<unsigned>> 179 CachedRegisterPressure; 180 181 bool EnableSinkAndFold; 182 183 public: 184 static char ID; // Pass identification 185 186 MachineSinking() : MachineFunctionPass(ID) { 187 initializeMachineSinkingPass(*PassRegistry::getPassRegistry()); 188 } 189 190 bool runOnMachineFunction(MachineFunction &MF) override; 191 192 void getAnalysisUsage(AnalysisUsage &AU) const override { 193 MachineFunctionPass::getAnalysisUsage(AU); 194 AU.addRequired<AAResultsWrapperPass>(); 195 AU.addRequired<MachineDominatorTreeWrapperPass>(); 196 AU.addRequired<MachinePostDominatorTreeWrapperPass>(); 197 AU.addRequired<MachineCycleInfoWrapperPass>(); 198 AU.addRequired<MachineBranchProbabilityInfoWrapperPass>(); 199 AU.addPreserved<MachineCycleInfoWrapperPass>(); 200 AU.addPreserved<MachineLoopInfoWrapperPass>(); 201 if (UseBlockFreqInfo) 202 AU.addRequired<MachineBlockFrequencyInfoWrapperPass>(); 203 AU.addRequired<TargetPassConfig>(); 204 } 205 206 void releaseMemory() override { 207 CEBCandidates.clear(); 208 CEMergeCandidates.clear(); 209 } 210 211 private: 212 bool ProcessBlock(MachineBasicBlock &MBB); 213 void ProcessDbgInst(MachineInstr &MI); 214 bool isLegalToBreakCriticalEdge(MachineInstr &MI, MachineBasicBlock *From, 215 MachineBasicBlock *To, bool BreakPHIEdge); 216 bool isWorthBreakingCriticalEdge(MachineInstr &MI, MachineBasicBlock *From, 217 MachineBasicBlock *To, 218 MachineBasicBlock *&DeferredFromBlock); 219 220 bool hasStoreBetween(MachineBasicBlock *From, MachineBasicBlock *To, 221 MachineInstr &MI); 222 223 /// Postpone the splitting of the given critical 224 /// edge (\p From, \p To). 225 /// 226 /// We do not split the edges on the fly. Indeed, this invalidates 227 /// the dominance information and thus triggers a lot of updates 228 /// of that information underneath. 229 /// Instead, we postpone all the splits after each iteration of 230 /// the main loop. That way, the information is at least valid 231 /// for the lifetime of an iteration. 232 /// 233 /// \return True if the edge is marked as toSplit, false otherwise. 234 /// False can be returned if, for instance, this is not profitable. 235 bool PostponeSplitCriticalEdge(MachineInstr &MI, 236 MachineBasicBlock *From, 237 MachineBasicBlock *To, 238 bool BreakPHIEdge); 239 bool SinkInstruction(MachineInstr &MI, bool &SawStore, 240 AllSuccsCache &AllSuccessors); 241 242 /// If we sink a COPY inst, some debug users of it's destination may no 243 /// longer be dominated by the COPY, and will eventually be dropped. 244 /// This is easily rectified by forwarding the non-dominated debug uses 245 /// to the copy source. 246 void SalvageUnsunkDebugUsersOfCopy(MachineInstr &, 247 MachineBasicBlock *TargetBlock); 248 bool AllUsesDominatedByBlock(Register Reg, MachineBasicBlock *MBB, 249 MachineBasicBlock *DefMBB, bool &BreakPHIEdge, 250 bool &LocalUse) const; 251 MachineBasicBlock *FindSuccToSinkTo(MachineInstr &MI, MachineBasicBlock *MBB, 252 bool &BreakPHIEdge, AllSuccsCache &AllSuccessors); 253 254 void FindCycleSinkCandidates(MachineCycle *Cycle, MachineBasicBlock *BB, 255 SmallVectorImpl<MachineInstr *> &Candidates); 256 bool SinkIntoCycle(MachineCycle *Cycle, MachineInstr &I); 257 258 bool isProfitableToSinkTo(Register Reg, MachineInstr &MI, 259 MachineBasicBlock *MBB, 260 MachineBasicBlock *SuccToSinkTo, 261 AllSuccsCache &AllSuccessors); 262 263 bool PerformTrivialForwardCoalescing(MachineInstr &MI, 264 MachineBasicBlock *MBB); 265 266 bool PerformSinkAndFold(MachineInstr &MI, MachineBasicBlock *MBB); 267 268 SmallVector<MachineBasicBlock *, 4> & 269 GetAllSortedSuccessors(MachineInstr &MI, MachineBasicBlock *MBB, 270 AllSuccsCache &AllSuccessors) const; 271 272 std::vector<unsigned> &getBBRegisterPressure(const MachineBasicBlock &MBB); 273 274 bool registerPressureSetExceedsLimit(unsigned NRegs, 275 const TargetRegisterClass *RC, 276 const MachineBasicBlock &MBB); 277 }; 278 279 } // end anonymous namespace 280 281 char MachineSinking::ID = 0; 282 283 char &llvm::MachineSinkingID = MachineSinking::ID; 284 285 INITIALIZE_PASS_BEGIN(MachineSinking, DEBUG_TYPE, 286 "Machine code sinking", false, false) 287 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfoWrapperPass) 288 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass) 289 INITIALIZE_PASS_DEPENDENCY(MachineCycleInfoWrapperPass) 290 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 291 INITIALIZE_PASS_END(MachineSinking, DEBUG_TYPE, 292 "Machine code sinking", false, false) 293 294 /// Return true if a target defined block prologue instruction interferes 295 /// with a sink candidate. 296 static bool blockPrologueInterferes(const MachineBasicBlock *BB, 297 MachineBasicBlock::const_iterator End, 298 const MachineInstr &MI, 299 const TargetRegisterInfo *TRI, 300 const TargetInstrInfo *TII, 301 const MachineRegisterInfo *MRI) { 302 for (MachineBasicBlock::const_iterator PI = BB->getFirstNonPHI(); PI != End; 303 ++PI) { 304 // Only check target defined prologue instructions 305 if (!TII->isBasicBlockPrologue(*PI)) 306 continue; 307 for (auto &MO : MI.operands()) { 308 if (!MO.isReg()) 309 continue; 310 Register Reg = MO.getReg(); 311 if (!Reg) 312 continue; 313 if (MO.isUse()) { 314 if (Reg.isPhysical() && 315 (TII->isIgnorableUse(MO) || (MRI && MRI->isConstantPhysReg(Reg)))) 316 continue; 317 if (PI->modifiesRegister(Reg, TRI)) 318 return true; 319 } else { 320 if (PI->readsRegister(Reg, TRI)) 321 return true; 322 // Check for interference with non-dead defs 323 auto *DefOp = PI->findRegisterDefOperand(Reg, TRI, false, true); 324 if (DefOp && !DefOp->isDead()) 325 return true; 326 } 327 } 328 } 329 330 return false; 331 } 332 333 bool MachineSinking::PerformTrivialForwardCoalescing(MachineInstr &MI, 334 MachineBasicBlock *MBB) { 335 if (!MI.isCopy()) 336 return false; 337 338 Register SrcReg = MI.getOperand(1).getReg(); 339 Register DstReg = MI.getOperand(0).getReg(); 340 if (!SrcReg.isVirtual() || !DstReg.isVirtual() || 341 !MRI->hasOneNonDBGUse(SrcReg)) 342 return false; 343 344 const TargetRegisterClass *SRC = MRI->getRegClass(SrcReg); 345 const TargetRegisterClass *DRC = MRI->getRegClass(DstReg); 346 if (SRC != DRC) 347 return false; 348 349 MachineInstr *DefMI = MRI->getVRegDef(SrcReg); 350 if (DefMI->isCopyLike()) 351 return false; 352 LLVM_DEBUG(dbgs() << "Coalescing: " << *DefMI); 353 LLVM_DEBUG(dbgs() << "*** to: " << MI); 354 MRI->replaceRegWith(DstReg, SrcReg); 355 MI.eraseFromParent(); 356 357 // Conservatively, clear any kill flags, since it's possible that they are no 358 // longer correct. 359 MRI->clearKillFlags(SrcReg); 360 361 ++NumCoalesces; 362 return true; 363 } 364 365 bool MachineSinking::PerformSinkAndFold(MachineInstr &MI, 366 MachineBasicBlock *MBB) { 367 if (MI.isCopy() || MI.mayLoadOrStore() || 368 MI.getOpcode() == TargetOpcode::REG_SEQUENCE) 369 return false; 370 371 // Don't sink instructions that the target prefers not to sink. 372 if (!TII->shouldSink(MI)) 373 return false; 374 375 // Check if it's safe to move the instruction. 376 bool SawStore = true; 377 if (!MI.isSafeToMove(AA, SawStore)) 378 return false; 379 380 // Convergent operations may not be made control-dependent on additional 381 // values. 382 if (MI.isConvergent()) 383 return false; 384 385 // Don't sink defs/uses of hard registers or if the instruction defines more 386 // than one register. 387 // Don't sink more than two register uses - it'll cover most of the cases and 388 // greatly simplifies the register pressure checks. 389 Register DefReg; 390 Register UsedRegA, UsedRegB; 391 for (const MachineOperand &MO : MI.operands()) { 392 if (MO.isImm() || MO.isRegMask() || MO.isRegLiveOut() || MO.isMetadata() || 393 MO.isMCSymbol() || MO.isDbgInstrRef() || MO.isCFIIndex() || 394 MO.isIntrinsicID() || MO.isPredicate() || MO.isShuffleMask()) 395 continue; 396 if (!MO.isReg()) 397 return false; 398 399 Register Reg = MO.getReg(); 400 if (Reg == 0) 401 continue; 402 403 if (Reg.isVirtual()) { 404 if (MO.isDef()) { 405 if (DefReg) 406 return false; 407 DefReg = Reg; 408 continue; 409 } 410 411 if (UsedRegA == 0) 412 UsedRegA = Reg; 413 else if (UsedRegB == 0) 414 UsedRegB = Reg; 415 else 416 return false; 417 continue; 418 } 419 420 if (Reg.isPhysical() && MO.isUse() && 421 (MRI->isConstantPhysReg(Reg) || TII->isIgnorableUse(MO))) 422 continue; 423 424 return false; 425 } 426 427 // Scan uses of the destination register. Every use, except the last, must be 428 // a copy, with a chain of copies terminating with either a copy into a hard 429 // register, or a load/store instruction where the use is part of the 430 // address (*not* the stored value). 431 using SinkInfo = std::pair<MachineInstr *, ExtAddrMode>; 432 SmallVector<SinkInfo> SinkInto; 433 SmallVector<Register> Worklist; 434 435 const TargetRegisterClass *RC = MRI->getRegClass(DefReg); 436 const TargetRegisterClass *RCA = 437 UsedRegA == 0 ? nullptr : MRI->getRegClass(UsedRegA); 438 const TargetRegisterClass *RCB = 439 UsedRegB == 0 ? nullptr : MRI->getRegClass(UsedRegB); 440 441 Worklist.push_back(DefReg); 442 while (!Worklist.empty()) { 443 Register Reg = Worklist.pop_back_val(); 444 445 for (MachineOperand &MO : MRI->use_nodbg_operands(Reg)) { 446 ExtAddrMode MaybeAM; 447 MachineInstr &UseInst = *MO.getParent(); 448 if (UseInst.isCopy()) { 449 Register DstReg; 450 if (const MachineOperand &O = UseInst.getOperand(0); O.isReg()) 451 DstReg = O.getReg(); 452 if (DstReg == 0) 453 return false; 454 if (DstReg.isVirtual()) { 455 Worklist.push_back(DstReg); 456 continue; 457 } 458 // If we are going to replace a copy, the original instruction must be 459 // as cheap as a copy. 460 if (!TII->isAsCheapAsAMove(MI)) 461 return false; 462 // The hard register must be in the register class of the original 463 // instruction's destination register. 464 if (!RC->contains(DstReg)) 465 return false; 466 } else if (UseInst.mayLoadOrStore()) { 467 ExtAddrMode AM; 468 if (!TII->canFoldIntoAddrMode(UseInst, Reg, MI, AM)) 469 return false; 470 MaybeAM = AM; 471 } else { 472 return false; 473 } 474 475 if (UseInst.getParent() != MI.getParent()) { 476 // If the register class of the register we are replacing is a superset 477 // of any of the register classes of the operands of the materialized 478 // instruction don't consider that live range extended. 479 const TargetRegisterClass *RCS = MRI->getRegClass(Reg); 480 if (RCA && RCA->hasSuperClassEq(RCS)) 481 RCA = nullptr; 482 else if (RCB && RCB->hasSuperClassEq(RCS)) 483 RCB = nullptr; 484 if (RCA || RCB) { 485 if (RCA == nullptr) { 486 RCA = RCB; 487 RCB = nullptr; 488 } 489 490 unsigned NRegs = !!RCA + !!RCB; 491 if (RCA == RCB) 492 RCB = nullptr; 493 494 // Check we don't exceed register pressure at the destination. 495 const MachineBasicBlock &MBB = *UseInst.getParent(); 496 if (RCB == nullptr) { 497 if (registerPressureSetExceedsLimit(NRegs, RCA, MBB)) 498 return false; 499 } else if (registerPressureSetExceedsLimit(1, RCA, MBB) || 500 registerPressureSetExceedsLimit(1, RCB, MBB)) { 501 return false; 502 } 503 } 504 } 505 506 SinkInto.emplace_back(&UseInst, MaybeAM); 507 } 508 } 509 510 if (SinkInto.empty()) 511 return false; 512 513 // Now we know we can fold the instruction in all its users. 514 for (auto &[SinkDst, MaybeAM] : SinkInto) { 515 MachineInstr *New = nullptr; 516 LLVM_DEBUG(dbgs() << "Sinking copy of"; MI.dump(); dbgs() << "into"; 517 SinkDst->dump()); 518 if (SinkDst->isCopy()) { 519 // TODO: After performing the sink-and-fold, the original instruction is 520 // deleted. Its value is still available (in a hard register), so if there 521 // are debug instructions which refer to the (now deleted) virtual 522 // register they could be updated to refer to the hard register, in 523 // principle. However, it's not clear how to do that, moreover in some 524 // cases the debug instructions may need to be replicated proportionally 525 // to the number of the COPY instructions replaced and in some extreme 526 // cases we can end up with quadratic increase in the number of debug 527 // instructions. 528 529 // Sink a copy of the instruction, replacing a COPY instruction. 530 MachineBasicBlock::iterator InsertPt = SinkDst->getIterator(); 531 Register DstReg = SinkDst->getOperand(0).getReg(); 532 TII->reMaterialize(*SinkDst->getParent(), InsertPt, DstReg, 0, MI, *TRI); 533 New = &*std::prev(InsertPt); 534 if (!New->getDebugLoc()) 535 New->setDebugLoc(SinkDst->getDebugLoc()); 536 537 // The operand registers of the "sunk" instruction have their live range 538 // extended and their kill flags may no longer be correct. Conservatively 539 // clear the kill flags. 540 if (UsedRegA) 541 MRI->clearKillFlags(UsedRegA); 542 if (UsedRegB) 543 MRI->clearKillFlags(UsedRegB); 544 } else { 545 // Fold instruction into the addressing mode of a memory instruction. 546 New = TII->emitLdStWithAddr(*SinkDst, MaybeAM); 547 548 // The registers of the addressing mode may have their live range extended 549 // and their kill flags may no longer be correct. Conservatively clear the 550 // kill flags. 551 if (Register R = MaybeAM.BaseReg; R.isValid() && R.isVirtual()) 552 MRI->clearKillFlags(R); 553 if (Register R = MaybeAM.ScaledReg; R.isValid() && R.isVirtual()) 554 MRI->clearKillFlags(R); 555 } 556 LLVM_DEBUG(dbgs() << "yielding"; New->dump()); 557 // Clear the StoreInstrCache, since we may invalidate it by erasing. 558 if (SinkDst->mayStore() && !SinkDst->hasOrderedMemoryRef()) 559 StoreInstrCache.clear(); 560 SinkDst->eraseFromParent(); 561 } 562 563 // Collect operands that need to be cleaned up because the registers no longer 564 // exist (in COPYs and debug instructions). We cannot delete instructions or 565 // clear operands while traversing register uses. 566 SmallVector<MachineOperand *> Cleanup; 567 Worklist.push_back(DefReg); 568 while (!Worklist.empty()) { 569 Register Reg = Worklist.pop_back_val(); 570 for (MachineOperand &MO : MRI->use_operands(Reg)) { 571 MachineInstr *U = MO.getParent(); 572 assert((U->isCopy() || U->isDebugInstr()) && 573 "Only debug uses and copies must remain"); 574 if (U->isCopy()) 575 Worklist.push_back(U->getOperand(0).getReg()); 576 Cleanup.push_back(&MO); 577 } 578 } 579 580 // Delete the dead COPYs and clear operands in debug instructions 581 for (MachineOperand *MO : Cleanup) { 582 MachineInstr *I = MO->getParent(); 583 if (I->isCopy()) { 584 I->eraseFromParent(); 585 } else { 586 MO->setReg(0); 587 MO->setSubReg(0); 588 } 589 } 590 591 MI.eraseFromParent(); 592 return true; 593 } 594 595 /// AllUsesDominatedByBlock - Return true if all uses of the specified register 596 /// occur in blocks dominated by the specified block. If any use is in the 597 /// definition block, then return false since it is never legal to move def 598 /// after uses. 599 bool MachineSinking::AllUsesDominatedByBlock(Register Reg, 600 MachineBasicBlock *MBB, 601 MachineBasicBlock *DefMBB, 602 bool &BreakPHIEdge, 603 bool &LocalUse) const { 604 assert(Reg.isVirtual() && "Only makes sense for vregs"); 605 606 // Ignore debug uses because debug info doesn't affect the code. 607 if (MRI->use_nodbg_empty(Reg)) 608 return true; 609 610 // BreakPHIEdge is true if all the uses are in the successor MBB being sunken 611 // into and they are all PHI nodes. In this case, machine-sink must break 612 // the critical edge first. e.g. 613 // 614 // %bb.1: 615 // Predecessors according to CFG: %bb.0 616 // ... 617 // %def = DEC64_32r %x, implicit-def dead %eflags 618 // ... 619 // JE_4 <%bb.37>, implicit %eflags 620 // Successors according to CFG: %bb.37 %bb.2 621 // 622 // %bb.2: 623 // %p = PHI %y, %bb.0, %def, %bb.1 624 if (all_of(MRI->use_nodbg_operands(Reg), [&](MachineOperand &MO) { 625 MachineInstr *UseInst = MO.getParent(); 626 unsigned OpNo = MO.getOperandNo(); 627 MachineBasicBlock *UseBlock = UseInst->getParent(); 628 return UseBlock == MBB && UseInst->isPHI() && 629 UseInst->getOperand(OpNo + 1).getMBB() == DefMBB; 630 })) { 631 BreakPHIEdge = true; 632 return true; 633 } 634 635 for (MachineOperand &MO : MRI->use_nodbg_operands(Reg)) { 636 // Determine the block of the use. 637 MachineInstr *UseInst = MO.getParent(); 638 unsigned OpNo = &MO - &UseInst->getOperand(0); 639 MachineBasicBlock *UseBlock = UseInst->getParent(); 640 if (UseInst->isPHI()) { 641 // PHI nodes use the operand in the predecessor block, not the block with 642 // the PHI. 643 UseBlock = UseInst->getOperand(OpNo+1).getMBB(); 644 } else if (UseBlock == DefMBB) { 645 LocalUse = true; 646 return false; 647 } 648 649 // Check that it dominates. 650 if (!DT->dominates(MBB, UseBlock)) 651 return false; 652 } 653 654 return true; 655 } 656 657 /// Return true if this machine instruction loads from global offset table or 658 /// constant pool. 659 static bool mayLoadFromGOTOrConstantPool(MachineInstr &MI) { 660 assert(MI.mayLoad() && "Expected MI that loads!"); 661 662 // If we lost memory operands, conservatively assume that the instruction 663 // reads from everything.. 664 if (MI.memoperands_empty()) 665 return true; 666 667 for (MachineMemOperand *MemOp : MI.memoperands()) 668 if (const PseudoSourceValue *PSV = MemOp->getPseudoValue()) 669 if (PSV->isGOT() || PSV->isConstantPool()) 670 return true; 671 672 return false; 673 } 674 675 void MachineSinking::FindCycleSinkCandidates( 676 MachineCycle *Cycle, MachineBasicBlock *BB, 677 SmallVectorImpl<MachineInstr *> &Candidates) { 678 for (auto &MI : *BB) { 679 LLVM_DEBUG(dbgs() << "CycleSink: Analysing candidate: " << MI); 680 if (!TII->shouldSink(MI)) { 681 LLVM_DEBUG(dbgs() << "CycleSink: Instruction not a candidate for this " 682 "target\n"); 683 continue; 684 } 685 if (!isCycleInvariant(Cycle, MI)) { 686 LLVM_DEBUG(dbgs() << "CycleSink: Instruction is not cycle invariant\n"); 687 continue; 688 } 689 bool DontMoveAcrossStore = true; 690 if (!MI.isSafeToMove(AA, DontMoveAcrossStore)) { 691 LLVM_DEBUG(dbgs() << "CycleSink: Instruction not safe to move.\n"); 692 continue; 693 } 694 if (MI.mayLoad() && !mayLoadFromGOTOrConstantPool(MI)) { 695 LLVM_DEBUG(dbgs() << "CycleSink: Dont sink GOT or constant pool loads\n"); 696 continue; 697 } 698 if (MI.isConvergent()) 699 continue; 700 701 const MachineOperand &MO = MI.getOperand(0); 702 if (!MO.isReg() || !MO.getReg() || !MO.isDef()) 703 continue; 704 if (!MRI->hasOneDef(MO.getReg())) 705 continue; 706 707 LLVM_DEBUG(dbgs() << "CycleSink: Instruction added as candidate.\n"); 708 Candidates.push_back(&MI); 709 } 710 } 711 712 bool MachineSinking::runOnMachineFunction(MachineFunction &MF) { 713 if (skipFunction(MF.getFunction())) 714 return false; 715 716 LLVM_DEBUG(dbgs() << "******** Machine Sinking ********\n"); 717 718 STI = &MF.getSubtarget(); 719 TII = STI->getInstrInfo(); 720 TRI = STI->getRegisterInfo(); 721 MRI = &MF.getRegInfo(); 722 DT = &getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree(); 723 PDT = &getAnalysis<MachinePostDominatorTreeWrapperPass>().getPostDomTree(); 724 CI = &getAnalysis<MachineCycleInfoWrapperPass>().getCycleInfo(); 725 MBFI = UseBlockFreqInfo 726 ? &getAnalysis<MachineBlockFrequencyInfoWrapperPass>().getMBFI() 727 : nullptr; 728 MBPI = &getAnalysis<MachineBranchProbabilityInfoWrapperPass>().getMBPI(); 729 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); 730 RegClassInfo.runOnMachineFunction(MF); 731 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>(); 732 EnableSinkAndFold = PassConfig->getEnableSinkAndFold(); 733 734 bool EverMadeChange = false; 735 736 while (true) { 737 bool MadeChange = false; 738 739 // Process all basic blocks. 740 CEBCandidates.clear(); 741 CEMergeCandidates.clear(); 742 ToSplit.clear(); 743 for (auto &MBB: MF) 744 MadeChange |= ProcessBlock(MBB); 745 746 // If we have anything we marked as toSplit, split it now. 747 for (const auto &Pair : ToSplit) { 748 auto NewSucc = Pair.first->SplitCriticalEdge(Pair.second, *this); 749 if (NewSucc != nullptr) { 750 LLVM_DEBUG(dbgs() << " *** Splitting critical edge: " 751 << printMBBReference(*Pair.first) << " -- " 752 << printMBBReference(*NewSucc) << " -- " 753 << printMBBReference(*Pair.second) << '\n'); 754 if (MBFI) 755 MBFI->onEdgeSplit(*Pair.first, *NewSucc, *MBPI); 756 757 MadeChange = true; 758 ++NumSplit; 759 CI->splitCriticalEdge(Pair.first, Pair.second, NewSucc); 760 } else 761 LLVM_DEBUG(dbgs() << " *** Not legal to break critical edge\n"); 762 } 763 // If this iteration over the code changed anything, keep iterating. 764 if (!MadeChange) break; 765 EverMadeChange = true; 766 } 767 768 if (SinkInstsIntoCycle) { 769 SmallVector<MachineCycle *, 8> Cycles(CI->toplevel_begin(), 770 CI->toplevel_end()); 771 for (auto *Cycle : Cycles) { 772 MachineBasicBlock *Preheader = Cycle->getCyclePreheader(); 773 if (!Preheader) { 774 LLVM_DEBUG(dbgs() << "CycleSink: Can't find preheader\n"); 775 continue; 776 } 777 SmallVector<MachineInstr *, 8> Candidates; 778 FindCycleSinkCandidates(Cycle, Preheader, Candidates); 779 780 // Walk the candidates in reverse order so that we start with the use 781 // of a def-use chain, if there is any. 782 // TODO: Sort the candidates using a cost-model. 783 unsigned i = 0; 784 for (MachineInstr *I : llvm::reverse(Candidates)) { 785 if (i++ == SinkIntoCycleLimit) { 786 LLVM_DEBUG(dbgs() << "CycleSink: Limit reached of instructions to " 787 "be analysed."); 788 break; 789 } 790 791 if (!SinkIntoCycle(Cycle, *I)) 792 break; 793 EverMadeChange = true; 794 ++NumCycleSunk; 795 } 796 } 797 } 798 799 HasStoreCache.clear(); 800 StoreInstrCache.clear(); 801 802 // Now clear any kill flags for recorded registers. 803 for (auto I : RegsToClearKillFlags) 804 MRI->clearKillFlags(I); 805 RegsToClearKillFlags.clear(); 806 807 return EverMadeChange; 808 } 809 810 bool MachineSinking::ProcessBlock(MachineBasicBlock &MBB) { 811 if ((!EnableSinkAndFold && MBB.succ_size() <= 1) || MBB.empty()) 812 return false; 813 814 // Don't bother sinking code out of unreachable blocks. In addition to being 815 // unprofitable, it can also lead to infinite looping, because in an 816 // unreachable cycle there may be nowhere to stop. 817 if (!DT->isReachableFromEntry(&MBB)) return false; 818 819 bool MadeChange = false; 820 821 // Cache all successors, sorted by frequency info and cycle depth. 822 AllSuccsCache AllSuccessors; 823 824 // Walk the basic block bottom-up. Remember if we saw a store. 825 MachineBasicBlock::iterator I = MBB.end(); 826 --I; 827 bool ProcessedBegin, SawStore = false; 828 do { 829 MachineInstr &MI = *I; // The instruction to sink. 830 831 // Predecrement I (if it's not begin) so that it isn't invalidated by 832 // sinking. 833 ProcessedBegin = I == MBB.begin(); 834 if (!ProcessedBegin) 835 --I; 836 837 if (MI.isDebugOrPseudoInstr()) { 838 if (MI.isDebugValue()) 839 ProcessDbgInst(MI); 840 continue; 841 } 842 843 if (EnableSinkAndFold && PerformSinkAndFold(MI, &MBB)) { 844 MadeChange = true; 845 continue; 846 } 847 848 // Can't sink anything out of a block that has less than two successors. 849 if (MBB.succ_size() <= 1) 850 continue; 851 852 if (PerformTrivialForwardCoalescing(MI, &MBB)) { 853 MadeChange = true; 854 continue; 855 } 856 857 if (SinkInstruction(MI, SawStore, AllSuccessors)) { 858 ++NumSunk; 859 MadeChange = true; 860 } 861 862 // If we just processed the first instruction in the block, we're done. 863 } while (!ProcessedBegin); 864 865 SeenDbgUsers.clear(); 866 SeenDbgVars.clear(); 867 // recalculate the bb register pressure after sinking one BB. 868 CachedRegisterPressure.clear(); 869 return MadeChange; 870 } 871 872 void MachineSinking::ProcessDbgInst(MachineInstr &MI) { 873 // When we see DBG_VALUEs for registers, record any vreg it reads, so that 874 // we know what to sink if the vreg def sinks. 875 assert(MI.isDebugValue() && "Expected DBG_VALUE for processing"); 876 877 DebugVariable Var(MI.getDebugVariable(), MI.getDebugExpression(), 878 MI.getDebugLoc()->getInlinedAt()); 879 bool SeenBefore = SeenDbgVars.contains(Var); 880 881 for (MachineOperand &MO : MI.debug_operands()) { 882 if (MO.isReg() && MO.getReg().isVirtual()) 883 SeenDbgUsers[MO.getReg()].push_back(SeenDbgUser(&MI, SeenBefore)); 884 } 885 886 // Record the variable for any DBG_VALUE, to avoid re-ordering any of them. 887 SeenDbgVars.insert(Var); 888 } 889 890 bool MachineSinking::isWorthBreakingCriticalEdge( 891 MachineInstr &MI, MachineBasicBlock *From, MachineBasicBlock *To, 892 MachineBasicBlock *&DeferredFromBlock) { 893 // FIXME: Need much better heuristics. 894 895 // If the pass has already considered breaking this edge (during this pass 896 // through the function), then let's go ahead and break it. This means 897 // sinking multiple "cheap" instructions into the same block. 898 if (!CEBCandidates.insert(std::make_pair(From, To)).second) 899 return true; 900 901 if (!MI.isCopy() && !TII->isAsCheapAsAMove(MI)) 902 return true; 903 904 // Check and record the register and the destination block we want to sink 905 // into. Note that we want to do the following before the next check on branch 906 // probability. Because we want to record the initial candidate even if it's 907 // on hot edge, so that other candidates that might not on hot edges can be 908 // sinked as well. 909 for (const auto &MO : MI.all_defs()) { 910 Register Reg = MO.getReg(); 911 if (!Reg) 912 continue; 913 Register SrcReg = Reg.isVirtual() ? TRI->lookThruCopyLike(Reg, MRI) : Reg; 914 auto Key = std::make_pair(SrcReg, To); 915 auto Res = CEMergeCandidates.try_emplace(Key, From); 916 // We wanted to sink the same register into the same block, consider it to 917 // be profitable. 918 if (!Res.second) { 919 // Return the source block that was previously held off. 920 DeferredFromBlock = Res.first->second; 921 return true; 922 } 923 } 924 925 if (From->isSuccessor(To) && MBPI->getEdgeProbability(From, To) <= 926 BranchProbability(SplitEdgeProbabilityThreshold, 100)) 927 return true; 928 929 // MI is cheap, we probably don't want to break the critical edge for it. 930 // However, if this would allow some definitions of its source operands 931 // to be sunk then it's probably worth it. 932 for (const MachineOperand &MO : MI.all_uses()) { 933 Register Reg = MO.getReg(); 934 if (Reg == 0) 935 continue; 936 937 // We don't move live definitions of physical registers, 938 // so sinking their uses won't enable any opportunities. 939 if (Reg.isPhysical()) 940 continue; 941 942 // If this instruction is the only user of a virtual register, 943 // check if breaking the edge will enable sinking 944 // both this instruction and the defining instruction. 945 if (MRI->hasOneNonDBGUse(Reg)) { 946 // If the definition resides in same MBB, 947 // claim it's likely we can sink these together. 948 // If definition resides elsewhere, we aren't 949 // blocking it from being sunk so don't break the edge. 950 MachineInstr *DefMI = MRI->getVRegDef(Reg); 951 if (DefMI->getParent() == MI.getParent()) 952 return true; 953 } 954 } 955 956 return false; 957 } 958 959 bool MachineSinking::isLegalToBreakCriticalEdge(MachineInstr &MI, 960 MachineBasicBlock *FromBB, 961 MachineBasicBlock *ToBB, 962 bool BreakPHIEdge) { 963 // Avoid breaking back edge. From == To means backedge for single BB cycle. 964 if (!SplitEdges || FromBB == ToBB || !FromBB->isSuccessor(ToBB)) 965 return false; 966 967 MachineCycle *FromCycle = CI->getCycle(FromBB); 968 MachineCycle *ToCycle = CI->getCycle(ToBB); 969 970 // Check for backedges of more "complex" cycles. 971 if (FromCycle == ToCycle && FromCycle && 972 (!FromCycle->isReducible() || FromCycle->getHeader() == ToBB)) 973 return false; 974 975 // It's not always legal to break critical edges and sink the computation 976 // to the edge. 977 // 978 // %bb.1: 979 // v1024 980 // Beq %bb.3 981 // <fallthrough> 982 // %bb.2: 983 // ... no uses of v1024 984 // <fallthrough> 985 // %bb.3: 986 // ... 987 // = v1024 988 // 989 // If %bb.1 -> %bb.3 edge is broken and computation of v1024 is inserted: 990 // 991 // %bb.1: 992 // ... 993 // Bne %bb.2 994 // %bb.4: 995 // v1024 = 996 // B %bb.3 997 // %bb.2: 998 // ... no uses of v1024 999 // <fallthrough> 1000 // %bb.3: 1001 // ... 1002 // = v1024 1003 // 1004 // This is incorrect since v1024 is not computed along the %bb.1->%bb.2->%bb.3 1005 // flow. We need to ensure the new basic block where the computation is 1006 // sunk to dominates all the uses. 1007 // It's only legal to break critical edge and sink the computation to the 1008 // new block if all the predecessors of "To", except for "From", are 1009 // not dominated by "From". Given SSA property, this means these 1010 // predecessors are dominated by "To". 1011 // 1012 // There is no need to do this check if all the uses are PHI nodes. PHI 1013 // sources are only defined on the specific predecessor edges. 1014 if (!BreakPHIEdge) { 1015 for (MachineBasicBlock *Pred : ToBB->predecessors()) 1016 if (Pred != FromBB && !DT->dominates(ToBB, Pred)) 1017 return false; 1018 } 1019 1020 return true; 1021 } 1022 1023 bool MachineSinking::PostponeSplitCriticalEdge(MachineInstr &MI, 1024 MachineBasicBlock *FromBB, 1025 MachineBasicBlock *ToBB, 1026 bool BreakPHIEdge) { 1027 bool Status = false; 1028 MachineBasicBlock *DeferredFromBB = nullptr; 1029 if (isWorthBreakingCriticalEdge(MI, FromBB, ToBB, DeferredFromBB)) { 1030 // If there is a DeferredFromBB, we consider FromBB only if _both_ 1031 // of them are legal to split. 1032 if ((!DeferredFromBB || 1033 ToSplit.count(std::make_pair(DeferredFromBB, ToBB)) || 1034 isLegalToBreakCriticalEdge(MI, DeferredFromBB, ToBB, BreakPHIEdge)) && 1035 isLegalToBreakCriticalEdge(MI, FromBB, ToBB, BreakPHIEdge)) { 1036 ToSplit.insert(std::make_pair(FromBB, ToBB)); 1037 if (DeferredFromBB) 1038 ToSplit.insert(std::make_pair(DeferredFromBB, ToBB)); 1039 Status = true; 1040 } 1041 } 1042 1043 return Status; 1044 } 1045 1046 std::vector<unsigned> & 1047 MachineSinking::getBBRegisterPressure(const MachineBasicBlock &MBB) { 1048 // Currently to save compiling time, MBB's register pressure will not change 1049 // in one ProcessBlock iteration because of CachedRegisterPressure. but MBB's 1050 // register pressure is changed after sinking any instructions into it. 1051 // FIXME: need a accurate and cheap register pressure estiminate model here. 1052 auto RP = CachedRegisterPressure.find(&MBB); 1053 if (RP != CachedRegisterPressure.end()) 1054 return RP->second; 1055 1056 RegionPressure Pressure; 1057 RegPressureTracker RPTracker(Pressure); 1058 1059 // Initialize the register pressure tracker. 1060 RPTracker.init(MBB.getParent(), &RegClassInfo, nullptr, &MBB, MBB.end(), 1061 /*TrackLaneMasks*/ false, /*TrackUntiedDefs=*/true); 1062 1063 for (MachineBasicBlock::const_iterator MII = MBB.instr_end(), 1064 MIE = MBB.instr_begin(); 1065 MII != MIE; --MII) { 1066 const MachineInstr &MI = *std::prev(MII); 1067 if (MI.isDebugInstr() || MI.isPseudoProbe()) 1068 continue; 1069 RegisterOperands RegOpers; 1070 RegOpers.collect(MI, *TRI, *MRI, false, false); 1071 RPTracker.recedeSkipDebugValues(); 1072 assert(&*RPTracker.getPos() == &MI && "RPTracker sync error!"); 1073 RPTracker.recede(RegOpers); 1074 } 1075 1076 RPTracker.closeRegion(); 1077 auto It = CachedRegisterPressure.insert( 1078 std::make_pair(&MBB, RPTracker.getPressure().MaxSetPressure)); 1079 return It.first->second; 1080 } 1081 1082 bool MachineSinking::registerPressureSetExceedsLimit( 1083 unsigned NRegs, const TargetRegisterClass *RC, 1084 const MachineBasicBlock &MBB) { 1085 unsigned Weight = NRegs * TRI->getRegClassWeight(RC).RegWeight; 1086 const int *PS = TRI->getRegClassPressureSets(RC); 1087 std::vector<unsigned> BBRegisterPressure = getBBRegisterPressure(MBB); 1088 for (; *PS != -1; PS++) 1089 if (Weight + BBRegisterPressure[*PS] >= 1090 TRI->getRegPressureSetLimit(*MBB.getParent(), *PS)) 1091 return true; 1092 return false; 1093 } 1094 1095 /// isProfitableToSinkTo - Return true if it is profitable to sink MI. 1096 bool MachineSinking::isProfitableToSinkTo(Register Reg, MachineInstr &MI, 1097 MachineBasicBlock *MBB, 1098 MachineBasicBlock *SuccToSinkTo, 1099 AllSuccsCache &AllSuccessors) { 1100 assert (SuccToSinkTo && "Invalid SinkTo Candidate BB"); 1101 1102 if (MBB == SuccToSinkTo) 1103 return false; 1104 1105 // It is profitable if SuccToSinkTo does not post dominate current block. 1106 if (!PDT->dominates(SuccToSinkTo, MBB)) 1107 return true; 1108 1109 // It is profitable to sink an instruction from a deeper cycle to a shallower 1110 // cycle, even if the latter post-dominates the former (PR21115). 1111 if (CI->getCycleDepth(MBB) > CI->getCycleDepth(SuccToSinkTo)) 1112 return true; 1113 1114 // Check if only use in post dominated block is PHI instruction. 1115 bool NonPHIUse = false; 1116 for (MachineInstr &UseInst : MRI->use_nodbg_instructions(Reg)) { 1117 MachineBasicBlock *UseBlock = UseInst.getParent(); 1118 if (UseBlock == SuccToSinkTo && !UseInst.isPHI()) 1119 NonPHIUse = true; 1120 } 1121 if (!NonPHIUse) 1122 return true; 1123 1124 // If SuccToSinkTo post dominates then also it may be profitable if MI 1125 // can further profitably sinked into another block in next round. 1126 bool BreakPHIEdge = false; 1127 // FIXME - If finding successor is compile time expensive then cache results. 1128 if (MachineBasicBlock *MBB2 = 1129 FindSuccToSinkTo(MI, SuccToSinkTo, BreakPHIEdge, AllSuccessors)) 1130 return isProfitableToSinkTo(Reg, MI, SuccToSinkTo, MBB2, AllSuccessors); 1131 1132 MachineCycle *MCycle = CI->getCycle(MBB); 1133 1134 // If the instruction is not inside a cycle, it is not profitable to sink MI to 1135 // a post dominate block SuccToSinkTo. 1136 if (!MCycle) 1137 return false; 1138 1139 // If this instruction is inside a Cycle and sinking this instruction can make 1140 // more registers live range shorten, it is still prifitable. 1141 for (const MachineOperand &MO : MI.operands()) { 1142 // Ignore non-register operands. 1143 if (!MO.isReg()) 1144 continue; 1145 Register Reg = MO.getReg(); 1146 if (Reg == 0) 1147 continue; 1148 1149 if (Reg.isPhysical()) { 1150 // Don't handle non-constant and non-ignorable physical register uses. 1151 if (MO.isUse() && !MRI->isConstantPhysReg(Reg) && !TII->isIgnorableUse(MO)) 1152 return false; 1153 continue; 1154 } 1155 1156 // Users for the defs are all dominated by SuccToSinkTo. 1157 if (MO.isDef()) { 1158 // This def register's live range is shortened after sinking. 1159 bool LocalUse = false; 1160 if (!AllUsesDominatedByBlock(Reg, SuccToSinkTo, MBB, BreakPHIEdge, 1161 LocalUse)) 1162 return false; 1163 } else { 1164 MachineInstr *DefMI = MRI->getVRegDef(Reg); 1165 if (!DefMI) 1166 continue; 1167 MachineCycle *Cycle = CI->getCycle(DefMI->getParent()); 1168 // DefMI is defined outside of cycle. There should be no live range 1169 // impact for this operand. Defination outside of cycle means: 1170 // 1: defination is outside of cycle. 1171 // 2: defination is in this cycle, but it is a PHI in the cycle header. 1172 if (Cycle != MCycle || (DefMI->isPHI() && Cycle && Cycle->isReducible() && 1173 Cycle->getHeader() == DefMI->getParent())) 1174 continue; 1175 // The DefMI is defined inside the cycle. 1176 // If sinking this operand makes some register pressure set exceed limit, 1177 // it is not profitable. 1178 if (registerPressureSetExceedsLimit(1, MRI->getRegClass(Reg), 1179 *SuccToSinkTo)) { 1180 LLVM_DEBUG(dbgs() << "register pressure exceed limit, not profitable."); 1181 return false; 1182 } 1183 } 1184 } 1185 1186 // If MI is in cycle and all its operands are alive across the whole cycle or 1187 // if no operand sinking make register pressure set exceed limit, it is 1188 // profitable to sink MI. 1189 return true; 1190 } 1191 1192 /// Get the sorted sequence of successors for this MachineBasicBlock, possibly 1193 /// computing it if it was not already cached. 1194 SmallVector<MachineBasicBlock *, 4> & 1195 MachineSinking::GetAllSortedSuccessors(MachineInstr &MI, MachineBasicBlock *MBB, 1196 AllSuccsCache &AllSuccessors) const { 1197 // Do we have the sorted successors in cache ? 1198 auto Succs = AllSuccessors.find(MBB); 1199 if (Succs != AllSuccessors.end()) 1200 return Succs->second; 1201 1202 SmallVector<MachineBasicBlock *, 4> AllSuccs(MBB->successors()); 1203 1204 // Handle cases where sinking can happen but where the sink point isn't a 1205 // successor. For example: 1206 // 1207 // x = computation 1208 // if () {} else {} 1209 // use x 1210 // 1211 for (MachineDomTreeNode *DTChild : DT->getNode(MBB)->children()) { 1212 // DomTree children of MBB that have MBB as immediate dominator are added. 1213 if (DTChild->getIDom()->getBlock() == MI.getParent() && 1214 // Skip MBBs already added to the AllSuccs vector above. 1215 !MBB->isSuccessor(DTChild->getBlock())) 1216 AllSuccs.push_back(DTChild->getBlock()); 1217 } 1218 1219 // Sort Successors according to their cycle depth or block frequency info. 1220 llvm::stable_sort( 1221 AllSuccs, [this](const MachineBasicBlock *L, const MachineBasicBlock *R) { 1222 uint64_t LHSFreq = MBFI ? MBFI->getBlockFreq(L).getFrequency() : 0; 1223 uint64_t RHSFreq = MBFI ? MBFI->getBlockFreq(R).getFrequency() : 0; 1224 bool HasBlockFreq = LHSFreq != 0 || RHSFreq != 0; 1225 return HasBlockFreq ? LHSFreq < RHSFreq 1226 : CI->getCycleDepth(L) < CI->getCycleDepth(R); 1227 }); 1228 1229 auto it = AllSuccessors.insert(std::make_pair(MBB, AllSuccs)); 1230 1231 return it.first->second; 1232 } 1233 1234 /// FindSuccToSinkTo - Find a successor to sink this instruction to. 1235 MachineBasicBlock * 1236 MachineSinking::FindSuccToSinkTo(MachineInstr &MI, MachineBasicBlock *MBB, 1237 bool &BreakPHIEdge, 1238 AllSuccsCache &AllSuccessors) { 1239 assert (MBB && "Invalid MachineBasicBlock!"); 1240 1241 // loop over all the operands of the specified instruction. If there is 1242 // anything we can't handle, bail out. 1243 1244 // SuccToSinkTo - This is the successor to sink this instruction to, once we 1245 // decide. 1246 MachineBasicBlock *SuccToSinkTo = nullptr; 1247 for (const MachineOperand &MO : MI.operands()) { 1248 if (!MO.isReg()) continue; // Ignore non-register operands. 1249 1250 Register Reg = MO.getReg(); 1251 if (Reg == 0) continue; 1252 1253 if (Reg.isPhysical()) { 1254 if (MO.isUse()) { 1255 // If the physreg has no defs anywhere, it's just an ambient register 1256 // and we can freely move its uses. Alternatively, if it's allocatable, 1257 // it could get allocated to something with a def during allocation. 1258 if (!MRI->isConstantPhysReg(Reg) && !TII->isIgnorableUse(MO)) 1259 return nullptr; 1260 } else if (!MO.isDead()) { 1261 // A def that isn't dead. We can't move it. 1262 return nullptr; 1263 } 1264 } else { 1265 // Virtual register uses are always safe to sink. 1266 if (MO.isUse()) continue; 1267 1268 // If it's not safe to move defs of the register class, then abort. 1269 if (!TII->isSafeToMoveRegClassDefs(MRI->getRegClass(Reg))) 1270 return nullptr; 1271 1272 // Virtual register defs can only be sunk if all their uses are in blocks 1273 // dominated by one of the successors. 1274 if (SuccToSinkTo) { 1275 // If a previous operand picked a block to sink to, then this operand 1276 // must be sinkable to the same block. 1277 bool LocalUse = false; 1278 if (!AllUsesDominatedByBlock(Reg, SuccToSinkTo, MBB, 1279 BreakPHIEdge, LocalUse)) 1280 return nullptr; 1281 1282 continue; 1283 } 1284 1285 // Otherwise, we should look at all the successors and decide which one 1286 // we should sink to. If we have reliable block frequency information 1287 // (frequency != 0) available, give successors with smaller frequencies 1288 // higher priority, otherwise prioritize smaller cycle depths. 1289 for (MachineBasicBlock *SuccBlock : 1290 GetAllSortedSuccessors(MI, MBB, AllSuccessors)) { 1291 bool LocalUse = false; 1292 if (AllUsesDominatedByBlock(Reg, SuccBlock, MBB, 1293 BreakPHIEdge, LocalUse)) { 1294 SuccToSinkTo = SuccBlock; 1295 break; 1296 } 1297 if (LocalUse) 1298 // Def is used locally, it's never safe to move this def. 1299 return nullptr; 1300 } 1301 1302 // If we couldn't find a block to sink to, ignore this instruction. 1303 if (!SuccToSinkTo) 1304 return nullptr; 1305 if (!isProfitableToSinkTo(Reg, MI, MBB, SuccToSinkTo, AllSuccessors)) 1306 return nullptr; 1307 } 1308 } 1309 1310 // It is not possible to sink an instruction into its own block. This can 1311 // happen with cycles. 1312 if (MBB == SuccToSinkTo) 1313 return nullptr; 1314 1315 // It's not safe to sink instructions to EH landing pad. Control flow into 1316 // landing pad is implicitly defined. 1317 if (SuccToSinkTo && SuccToSinkTo->isEHPad()) 1318 return nullptr; 1319 1320 // It ought to be okay to sink instructions into an INLINEASM_BR target, but 1321 // only if we make sure that MI occurs _before_ an INLINEASM_BR instruction in 1322 // the source block (which this code does not yet do). So for now, forbid 1323 // doing so. 1324 if (SuccToSinkTo && SuccToSinkTo->isInlineAsmBrIndirectTarget()) 1325 return nullptr; 1326 1327 if (SuccToSinkTo && !TII->isSafeToSink(MI, SuccToSinkTo, CI)) 1328 return nullptr; 1329 1330 return SuccToSinkTo; 1331 } 1332 1333 /// Return true if MI is likely to be usable as a memory operation by the 1334 /// implicit null check optimization. 1335 /// 1336 /// This is a "best effort" heuristic, and should not be relied upon for 1337 /// correctness. This returning true does not guarantee that the implicit null 1338 /// check optimization is legal over MI, and this returning false does not 1339 /// guarantee MI cannot possibly be used to do a null check. 1340 static bool SinkingPreventsImplicitNullCheck(MachineInstr &MI, 1341 const TargetInstrInfo *TII, 1342 const TargetRegisterInfo *TRI) { 1343 using MachineBranchPredicate = TargetInstrInfo::MachineBranchPredicate; 1344 1345 auto *MBB = MI.getParent(); 1346 if (MBB->pred_size() != 1) 1347 return false; 1348 1349 auto *PredMBB = *MBB->pred_begin(); 1350 auto *PredBB = PredMBB->getBasicBlock(); 1351 1352 // Frontends that don't use implicit null checks have no reason to emit 1353 // branches with make.implicit metadata, and this function should always 1354 // return false for them. 1355 if (!PredBB || 1356 !PredBB->getTerminator()->getMetadata(LLVMContext::MD_make_implicit)) 1357 return false; 1358 1359 const MachineOperand *BaseOp; 1360 int64_t Offset; 1361 bool OffsetIsScalable; 1362 if (!TII->getMemOperandWithOffset(MI, BaseOp, Offset, OffsetIsScalable, TRI)) 1363 return false; 1364 1365 if (!BaseOp->isReg()) 1366 return false; 1367 1368 if (!(MI.mayLoad() && !MI.isPredicable())) 1369 return false; 1370 1371 MachineBranchPredicate MBP; 1372 if (TII->analyzeBranchPredicate(*PredMBB, MBP, false)) 1373 return false; 1374 1375 return MBP.LHS.isReg() && MBP.RHS.isImm() && MBP.RHS.getImm() == 0 && 1376 (MBP.Predicate == MachineBranchPredicate::PRED_NE || 1377 MBP.Predicate == MachineBranchPredicate::PRED_EQ) && 1378 MBP.LHS.getReg() == BaseOp->getReg(); 1379 } 1380 1381 /// If the sunk instruction is a copy, try to forward the copy instead of 1382 /// leaving an 'undef' DBG_VALUE in the original location. Don't do this if 1383 /// there's any subregister weirdness involved. Returns true if copy 1384 /// propagation occurred. 1385 static bool attemptDebugCopyProp(MachineInstr &SinkInst, MachineInstr &DbgMI, 1386 Register Reg) { 1387 const MachineRegisterInfo &MRI = SinkInst.getMF()->getRegInfo(); 1388 const TargetInstrInfo &TII = *SinkInst.getMF()->getSubtarget().getInstrInfo(); 1389 1390 // Copy DBG_VALUE operand and set the original to undef. We then check to 1391 // see whether this is something that can be copy-forwarded. If it isn't, 1392 // continue around the loop. 1393 1394 const MachineOperand *SrcMO = nullptr, *DstMO = nullptr; 1395 auto CopyOperands = TII.isCopyInstr(SinkInst); 1396 if (!CopyOperands) 1397 return false; 1398 SrcMO = CopyOperands->Source; 1399 DstMO = CopyOperands->Destination; 1400 1401 // Check validity of forwarding this copy. 1402 bool PostRA = MRI.getNumVirtRegs() == 0; 1403 1404 // Trying to forward between physical and virtual registers is too hard. 1405 if (Reg.isVirtual() != SrcMO->getReg().isVirtual()) 1406 return false; 1407 1408 // Only try virtual register copy-forwarding before regalloc, and physical 1409 // register copy-forwarding after regalloc. 1410 bool arePhysRegs = !Reg.isVirtual(); 1411 if (arePhysRegs != PostRA) 1412 return false; 1413 1414 // Pre-regalloc, only forward if all subregisters agree (or there are no 1415 // subregs at all). More analysis might recover some forwardable copies. 1416 if (!PostRA) 1417 for (auto &DbgMO : DbgMI.getDebugOperandsForReg(Reg)) 1418 if (DbgMO.getSubReg() != SrcMO->getSubReg() || 1419 DbgMO.getSubReg() != DstMO->getSubReg()) 1420 return false; 1421 1422 // Post-regalloc, we may be sinking a DBG_VALUE of a sub or super-register 1423 // of this copy. Only forward the copy if the DBG_VALUE operand exactly 1424 // matches the copy destination. 1425 if (PostRA && Reg != DstMO->getReg()) 1426 return false; 1427 1428 for (auto &DbgMO : DbgMI.getDebugOperandsForReg(Reg)) { 1429 DbgMO.setReg(SrcMO->getReg()); 1430 DbgMO.setSubReg(SrcMO->getSubReg()); 1431 } 1432 return true; 1433 } 1434 1435 using MIRegs = std::pair<MachineInstr *, SmallVector<unsigned, 2>>; 1436 /// Sink an instruction and its associated debug instructions. 1437 static void performSink(MachineInstr &MI, MachineBasicBlock &SuccToSinkTo, 1438 MachineBasicBlock::iterator InsertPos, 1439 ArrayRef<MIRegs> DbgValuesToSink) { 1440 // If we cannot find a location to use (merge with), then we erase the debug 1441 // location to prevent debug-info driven tools from potentially reporting 1442 // wrong location information. 1443 if (!SuccToSinkTo.empty() && InsertPos != SuccToSinkTo.end()) 1444 MI.setDebugLoc(DILocation::getMergedLocation(MI.getDebugLoc(), 1445 InsertPos->getDebugLoc())); 1446 else 1447 MI.setDebugLoc(DebugLoc()); 1448 1449 // Move the instruction. 1450 MachineBasicBlock *ParentBlock = MI.getParent(); 1451 SuccToSinkTo.splice(InsertPos, ParentBlock, MI, 1452 ++MachineBasicBlock::iterator(MI)); 1453 1454 // Sink a copy of debug users to the insert position. Mark the original 1455 // DBG_VALUE location as 'undef', indicating that any earlier variable 1456 // location should be terminated as we've optimised away the value at this 1457 // point. 1458 for (const auto &DbgValueToSink : DbgValuesToSink) { 1459 MachineInstr *DbgMI = DbgValueToSink.first; 1460 MachineInstr *NewDbgMI = DbgMI->getMF()->CloneMachineInstr(DbgMI); 1461 SuccToSinkTo.insert(InsertPos, NewDbgMI); 1462 1463 bool PropagatedAllSunkOps = true; 1464 for (unsigned Reg : DbgValueToSink.second) { 1465 if (DbgMI->hasDebugOperandForReg(Reg)) { 1466 if (!attemptDebugCopyProp(MI, *DbgMI, Reg)) { 1467 PropagatedAllSunkOps = false; 1468 break; 1469 } 1470 } 1471 } 1472 if (!PropagatedAllSunkOps) 1473 DbgMI->setDebugValueUndef(); 1474 } 1475 } 1476 1477 /// hasStoreBetween - check if there is store betweeen straight line blocks From 1478 /// and To. 1479 bool MachineSinking::hasStoreBetween(MachineBasicBlock *From, 1480 MachineBasicBlock *To, MachineInstr &MI) { 1481 // Make sure From and To are in straight line which means From dominates To 1482 // and To post dominates From. 1483 if (!DT->dominates(From, To) || !PDT->dominates(To, From)) 1484 return true; 1485 1486 auto BlockPair = std::make_pair(From, To); 1487 1488 // Does these two blocks pair be queried before and have a definite cached 1489 // result? 1490 if (auto It = HasStoreCache.find(BlockPair); It != HasStoreCache.end()) 1491 return It->second; 1492 1493 if (auto It = StoreInstrCache.find(BlockPair); It != StoreInstrCache.end()) 1494 return llvm::any_of(It->second, [&](MachineInstr *I) { 1495 return I->mayAlias(AA, MI, false); 1496 }); 1497 1498 bool SawStore = false; 1499 bool HasAliasedStore = false; 1500 DenseSet<MachineBasicBlock *> HandledBlocks; 1501 DenseSet<MachineBasicBlock *> HandledDomBlocks; 1502 // Go through all reachable blocks from From. 1503 for (MachineBasicBlock *BB : depth_first(From)) { 1504 // We insert the instruction at the start of block To, so no need to worry 1505 // about stores inside To. 1506 // Store in block From should be already considered when just enter function 1507 // SinkInstruction. 1508 if (BB == To || BB == From) 1509 continue; 1510 1511 // We already handle this BB in previous iteration. 1512 if (HandledBlocks.count(BB)) 1513 continue; 1514 1515 HandledBlocks.insert(BB); 1516 // To post dominates BB, it must be a path from block From. 1517 if (PDT->dominates(To, BB)) { 1518 if (!HandledDomBlocks.count(BB)) 1519 HandledDomBlocks.insert(BB); 1520 1521 // If this BB is too big or the block number in straight line between From 1522 // and To is too big, stop searching to save compiling time. 1523 if (BB->sizeWithoutDebugLargerThan(SinkLoadInstsPerBlockThreshold) || 1524 HandledDomBlocks.size() > SinkLoadBlocksThreshold) { 1525 for (auto *DomBB : HandledDomBlocks) { 1526 if (DomBB != BB && DT->dominates(DomBB, BB)) 1527 HasStoreCache[std::make_pair(DomBB, To)] = true; 1528 else if(DomBB != BB && DT->dominates(BB, DomBB)) 1529 HasStoreCache[std::make_pair(From, DomBB)] = true; 1530 } 1531 HasStoreCache[BlockPair] = true; 1532 return true; 1533 } 1534 1535 for (MachineInstr &I : *BB) { 1536 // Treat as alias conservatively for a call or an ordered memory 1537 // operation. 1538 if (I.isCall() || I.hasOrderedMemoryRef()) { 1539 for (auto *DomBB : HandledDomBlocks) { 1540 if (DomBB != BB && DT->dominates(DomBB, BB)) 1541 HasStoreCache[std::make_pair(DomBB, To)] = true; 1542 else if(DomBB != BB && DT->dominates(BB, DomBB)) 1543 HasStoreCache[std::make_pair(From, DomBB)] = true; 1544 } 1545 HasStoreCache[BlockPair] = true; 1546 return true; 1547 } 1548 1549 if (I.mayStore()) { 1550 SawStore = true; 1551 // We still have chance to sink MI if all stores between are not 1552 // aliased to MI. 1553 // Cache all store instructions, so that we don't need to go through 1554 // all From reachable blocks for next load instruction. 1555 if (I.mayAlias(AA, MI, false)) 1556 HasAliasedStore = true; 1557 StoreInstrCache[BlockPair].push_back(&I); 1558 } 1559 } 1560 } 1561 } 1562 // If there is no store at all, cache the result. 1563 if (!SawStore) 1564 HasStoreCache[BlockPair] = false; 1565 return HasAliasedStore; 1566 } 1567 1568 /// Sink instructions into cycles if profitable. This especially tries to 1569 /// prevent register spills caused by register pressure if there is little to no 1570 /// overhead moving instructions into cycles. 1571 bool MachineSinking::SinkIntoCycle(MachineCycle *Cycle, MachineInstr &I) { 1572 LLVM_DEBUG(dbgs() << "CycleSink: Finding sink block for: " << I); 1573 MachineBasicBlock *Preheader = Cycle->getCyclePreheader(); 1574 assert(Preheader && "Cycle sink needs a preheader block"); 1575 MachineBasicBlock *SinkBlock = nullptr; 1576 bool CanSink = true; 1577 const MachineOperand &MO = I.getOperand(0); 1578 1579 for (MachineInstr &MI : MRI->use_instructions(MO.getReg())) { 1580 LLVM_DEBUG(dbgs() << "CycleSink: Analysing use: " << MI); 1581 if (!Cycle->contains(MI.getParent())) { 1582 LLVM_DEBUG(dbgs() << "CycleSink: Use not in cycle, can't sink.\n"); 1583 CanSink = false; 1584 break; 1585 } 1586 1587 // FIXME: Come up with a proper cost model that estimates whether sinking 1588 // the instruction (and thus possibly executing it on every cycle 1589 // iteration) is more expensive than a register. 1590 // For now assumes that copies are cheap and thus almost always worth it. 1591 if (!MI.isCopy()) { 1592 LLVM_DEBUG(dbgs() << "CycleSink: Use is not a copy\n"); 1593 CanSink = false; 1594 break; 1595 } 1596 if (!SinkBlock) { 1597 SinkBlock = MI.getParent(); 1598 LLVM_DEBUG(dbgs() << "CycleSink: Setting sink block to: " 1599 << printMBBReference(*SinkBlock) << "\n"); 1600 continue; 1601 } 1602 SinkBlock = DT->findNearestCommonDominator(SinkBlock, MI.getParent()); 1603 if (!SinkBlock) { 1604 LLVM_DEBUG(dbgs() << "CycleSink: Can't find nearest dominator\n"); 1605 CanSink = false; 1606 break; 1607 } 1608 LLVM_DEBUG(dbgs() << "CycleSink: Setting nearest common dom block: " << 1609 printMBBReference(*SinkBlock) << "\n"); 1610 } 1611 1612 if (!CanSink) { 1613 LLVM_DEBUG(dbgs() << "CycleSink: Can't sink instruction.\n"); 1614 return false; 1615 } 1616 if (!SinkBlock) { 1617 LLVM_DEBUG(dbgs() << "CycleSink: Not sinking, can't find sink block.\n"); 1618 return false; 1619 } 1620 if (SinkBlock == Preheader) { 1621 LLVM_DEBUG( 1622 dbgs() << "CycleSink: Not sinking, sink block is the preheader\n"); 1623 return false; 1624 } 1625 if (SinkBlock->sizeWithoutDebugLargerThan(SinkLoadInstsPerBlockThreshold)) { 1626 LLVM_DEBUG( 1627 dbgs() << "CycleSink: Not Sinking, block too large to analyse.\n"); 1628 return false; 1629 } 1630 1631 LLVM_DEBUG(dbgs() << "CycleSink: Sinking instruction!\n"); 1632 SinkBlock->splice(SinkBlock->SkipPHIsAndLabels(SinkBlock->begin()), Preheader, 1633 I); 1634 1635 // Conservatively clear any kill flags on uses of sunk instruction 1636 for (MachineOperand &MO : I.operands()) { 1637 if (MO.isReg() && MO.readsReg()) 1638 RegsToClearKillFlags.insert(MO.getReg()); 1639 } 1640 1641 // The instruction is moved from its basic block, so do not retain the 1642 // debug information. 1643 assert(!I.isDebugInstr() && "Should not sink debug inst"); 1644 I.setDebugLoc(DebugLoc()); 1645 return true; 1646 } 1647 1648 /// SinkInstruction - Determine whether it is safe to sink the specified machine 1649 /// instruction out of its current block into a successor. 1650 bool MachineSinking::SinkInstruction(MachineInstr &MI, bool &SawStore, 1651 AllSuccsCache &AllSuccessors) { 1652 // Don't sink instructions that the target prefers not to sink. 1653 if (!TII->shouldSink(MI)) 1654 return false; 1655 1656 // Check if it's safe to move the instruction. 1657 if (!MI.isSafeToMove(AA, SawStore)) 1658 return false; 1659 1660 // Convergent operations may not be made control-dependent on additional 1661 // values. 1662 if (MI.isConvergent()) 1663 return false; 1664 1665 // Don't break implicit null checks. This is a performance heuristic, and not 1666 // required for correctness. 1667 if (SinkingPreventsImplicitNullCheck(MI, TII, TRI)) 1668 return false; 1669 1670 // FIXME: This should include support for sinking instructions within the 1671 // block they are currently in to shorten the live ranges. We often get 1672 // instructions sunk into the top of a large block, but it would be better to 1673 // also sink them down before their first use in the block. This xform has to 1674 // be careful not to *increase* register pressure though, e.g. sinking 1675 // "x = y + z" down if it kills y and z would increase the live ranges of y 1676 // and z and only shrink the live range of x. 1677 1678 bool BreakPHIEdge = false; 1679 MachineBasicBlock *ParentBlock = MI.getParent(); 1680 MachineBasicBlock *SuccToSinkTo = 1681 FindSuccToSinkTo(MI, ParentBlock, BreakPHIEdge, AllSuccessors); 1682 1683 // If there are no outputs, it must have side-effects. 1684 if (!SuccToSinkTo) 1685 return false; 1686 1687 // If the instruction to move defines a dead physical register which is live 1688 // when leaving the basic block, don't move it because it could turn into a 1689 // "zombie" define of that preg. E.g., EFLAGS. 1690 for (const MachineOperand &MO : MI.all_defs()) { 1691 Register Reg = MO.getReg(); 1692 if (Reg == 0 || !Reg.isPhysical()) 1693 continue; 1694 if (SuccToSinkTo->isLiveIn(Reg)) 1695 return false; 1696 } 1697 1698 LLVM_DEBUG(dbgs() << "Sink instr " << MI << "\tinto block " << *SuccToSinkTo); 1699 1700 // If the block has multiple predecessors, this is a critical edge. 1701 // Decide if we can sink along it or need to break the edge. 1702 if (SuccToSinkTo->pred_size() > 1) { 1703 // We cannot sink a load across a critical edge - there may be stores in 1704 // other code paths. 1705 bool TryBreak = false; 1706 bool Store = 1707 MI.mayLoad() ? hasStoreBetween(ParentBlock, SuccToSinkTo, MI) : true; 1708 if (!MI.isSafeToMove(AA, Store)) { 1709 LLVM_DEBUG(dbgs() << " *** NOTE: Won't sink load along critical edge.\n"); 1710 TryBreak = true; 1711 } 1712 1713 // We don't want to sink across a critical edge if we don't dominate the 1714 // successor. We could be introducing calculations to new code paths. 1715 if (!TryBreak && !DT->dominates(ParentBlock, SuccToSinkTo)) { 1716 LLVM_DEBUG(dbgs() << " *** NOTE: Critical edge found\n"); 1717 TryBreak = true; 1718 } 1719 1720 // Don't sink instructions into a cycle. 1721 if (!TryBreak && CI->getCycle(SuccToSinkTo) && 1722 (!CI->getCycle(SuccToSinkTo)->isReducible() || 1723 CI->getCycle(SuccToSinkTo)->getHeader() == SuccToSinkTo)) { 1724 LLVM_DEBUG(dbgs() << " *** NOTE: cycle header found\n"); 1725 TryBreak = true; 1726 } 1727 1728 // Otherwise we are OK with sinking along a critical edge. 1729 if (!TryBreak) 1730 LLVM_DEBUG(dbgs() << "Sinking along critical edge.\n"); 1731 else { 1732 // Mark this edge as to be split. 1733 // If the edge can actually be split, the next iteration of the main loop 1734 // will sink MI in the newly created block. 1735 bool Status = 1736 PostponeSplitCriticalEdge(MI, ParentBlock, SuccToSinkTo, BreakPHIEdge); 1737 if (!Status) 1738 LLVM_DEBUG(dbgs() << " *** PUNTING: Not legal or profitable to " 1739 "break critical edge\n"); 1740 // The instruction will not be sunk this time. 1741 return false; 1742 } 1743 } 1744 1745 if (BreakPHIEdge) { 1746 // BreakPHIEdge is true if all the uses are in the successor MBB being 1747 // sunken into and they are all PHI nodes. In this case, machine-sink must 1748 // break the critical edge first. 1749 bool Status = PostponeSplitCriticalEdge(MI, ParentBlock, 1750 SuccToSinkTo, BreakPHIEdge); 1751 if (!Status) 1752 LLVM_DEBUG(dbgs() << " *** PUNTING: Not legal or profitable to " 1753 "break critical edge\n"); 1754 // The instruction will not be sunk this time. 1755 return false; 1756 } 1757 1758 // Determine where to insert into. Skip phi nodes. 1759 MachineBasicBlock::iterator InsertPos = 1760 SuccToSinkTo->SkipPHIsAndLabels(SuccToSinkTo->begin()); 1761 if (blockPrologueInterferes(SuccToSinkTo, InsertPos, MI, TRI, TII, MRI)) { 1762 LLVM_DEBUG(dbgs() << " *** Not sinking: prologue interference\n"); 1763 return false; 1764 } 1765 1766 // Collect debug users of any vreg that this inst defines. 1767 SmallVector<MIRegs, 4> DbgUsersToSink; 1768 for (auto &MO : MI.all_defs()) { 1769 if (!MO.getReg().isVirtual()) 1770 continue; 1771 if (!SeenDbgUsers.count(MO.getReg())) 1772 continue; 1773 1774 // Sink any users that don't pass any other DBG_VALUEs for this variable. 1775 auto &Users = SeenDbgUsers[MO.getReg()]; 1776 for (auto &User : Users) { 1777 MachineInstr *DbgMI = User.getPointer(); 1778 if (User.getInt()) { 1779 // This DBG_VALUE would re-order assignments. If we can't copy-propagate 1780 // it, it can't be recovered. Set it undef. 1781 if (!attemptDebugCopyProp(MI, *DbgMI, MO.getReg())) 1782 DbgMI->setDebugValueUndef(); 1783 } else { 1784 DbgUsersToSink.push_back( 1785 {DbgMI, SmallVector<unsigned, 2>(1, MO.getReg())}); 1786 } 1787 } 1788 } 1789 1790 // After sinking, some debug users may not be dominated any more. If possible, 1791 // copy-propagate their operands. As it's expensive, don't do this if there's 1792 // no debuginfo in the program. 1793 if (MI.getMF()->getFunction().getSubprogram() && MI.isCopy()) 1794 SalvageUnsunkDebugUsersOfCopy(MI, SuccToSinkTo); 1795 1796 performSink(MI, *SuccToSinkTo, InsertPos, DbgUsersToSink); 1797 1798 // Conservatively, clear any kill flags, since it's possible that they are no 1799 // longer correct. 1800 // Note that we have to clear the kill flags for any register this instruction 1801 // uses as we may sink over another instruction which currently kills the 1802 // used registers. 1803 for (MachineOperand &MO : MI.all_uses()) 1804 RegsToClearKillFlags.insert(MO.getReg()); // Remember to clear kill flags. 1805 1806 return true; 1807 } 1808 1809 void MachineSinking::SalvageUnsunkDebugUsersOfCopy( 1810 MachineInstr &MI, MachineBasicBlock *TargetBlock) { 1811 assert(MI.isCopy()); 1812 assert(MI.getOperand(1).isReg()); 1813 1814 // Enumerate all users of vreg operands that are def'd. Skip those that will 1815 // be sunk. For the rest, if they are not dominated by the block we will sink 1816 // MI into, propagate the copy source to them. 1817 SmallVector<MachineInstr *, 4> DbgDefUsers; 1818 SmallVector<Register, 4> DbgUseRegs; 1819 const MachineRegisterInfo &MRI = MI.getMF()->getRegInfo(); 1820 for (auto &MO : MI.all_defs()) { 1821 if (!MO.getReg().isVirtual()) 1822 continue; 1823 DbgUseRegs.push_back(MO.getReg()); 1824 for (auto &User : MRI.use_instructions(MO.getReg())) { 1825 if (!User.isDebugValue() || DT->dominates(TargetBlock, User.getParent())) 1826 continue; 1827 1828 // If is in same block, will either sink or be use-before-def. 1829 if (User.getParent() == MI.getParent()) 1830 continue; 1831 1832 assert(User.hasDebugOperandForReg(MO.getReg()) && 1833 "DBG_VALUE user of vreg, but has no operand for it?"); 1834 DbgDefUsers.push_back(&User); 1835 } 1836 } 1837 1838 // Point the users of this copy that are no longer dominated, at the source 1839 // of the copy. 1840 for (auto *User : DbgDefUsers) { 1841 for (auto &Reg : DbgUseRegs) { 1842 for (auto &DbgOp : User->getDebugOperandsForReg(Reg)) { 1843 DbgOp.setReg(MI.getOperand(1).getReg()); 1844 DbgOp.setSubReg(MI.getOperand(1).getSubReg()); 1845 } 1846 } 1847 } 1848 } 1849 1850 //===----------------------------------------------------------------------===// 1851 // This pass is not intended to be a replacement or a complete alternative 1852 // for the pre-ra machine sink pass. It is only designed to sink COPY 1853 // instructions which should be handled after RA. 1854 // 1855 // This pass sinks COPY instructions into a successor block, if the COPY is not 1856 // used in the current block and the COPY is live-in to a single successor 1857 // (i.e., doesn't require the COPY to be duplicated). This avoids executing the 1858 // copy on paths where their results aren't needed. This also exposes 1859 // additional opportunites for dead copy elimination and shrink wrapping. 1860 // 1861 // These copies were either not handled by or are inserted after the MachineSink 1862 // pass. As an example of the former case, the MachineSink pass cannot sink 1863 // COPY instructions with allocatable source registers; for AArch64 these type 1864 // of copy instructions are frequently used to move function parameters (PhyReg) 1865 // into virtual registers in the entry block. 1866 // 1867 // For the machine IR below, this pass will sink %w19 in the entry into its 1868 // successor (%bb.1) because %w19 is only live-in in %bb.1. 1869 // %bb.0: 1870 // %wzr = SUBSWri %w1, 1 1871 // %w19 = COPY %w0 1872 // Bcc 11, %bb.2 1873 // %bb.1: 1874 // Live Ins: %w19 1875 // BL @fun 1876 // %w0 = ADDWrr %w0, %w19 1877 // RET %w0 1878 // %bb.2: 1879 // %w0 = COPY %wzr 1880 // RET %w0 1881 // As we sink %w19 (CSR in AArch64) into %bb.1, the shrink-wrapping pass will be 1882 // able to see %bb.0 as a candidate. 1883 //===----------------------------------------------------------------------===// 1884 namespace { 1885 1886 class PostRAMachineSinking : public MachineFunctionPass { 1887 public: 1888 bool runOnMachineFunction(MachineFunction &MF) override; 1889 1890 static char ID; 1891 PostRAMachineSinking() : MachineFunctionPass(ID) {} 1892 StringRef getPassName() const override { return "PostRA Machine Sink"; } 1893 1894 void getAnalysisUsage(AnalysisUsage &AU) const override { 1895 AU.setPreservesCFG(); 1896 MachineFunctionPass::getAnalysisUsage(AU); 1897 } 1898 1899 MachineFunctionProperties getRequiredProperties() const override { 1900 return MachineFunctionProperties().set( 1901 MachineFunctionProperties::Property::NoVRegs); 1902 } 1903 1904 private: 1905 /// Track which register units have been modified and used. 1906 LiveRegUnits ModifiedRegUnits, UsedRegUnits; 1907 1908 /// Track DBG_VALUEs of (unmodified) register units. Each DBG_VALUE has an 1909 /// entry in this map for each unit it touches. The DBG_VALUE's entry 1910 /// consists of a pointer to the instruction itself, and a vector of registers 1911 /// referred to by the instruction that overlap the key register unit. 1912 DenseMap<unsigned, SmallVector<MIRegs, 2>> SeenDbgInstrs; 1913 1914 /// Sink Copy instructions unused in the same block close to their uses in 1915 /// successors. 1916 bool tryToSinkCopy(MachineBasicBlock &BB, MachineFunction &MF, 1917 const TargetRegisterInfo *TRI, const TargetInstrInfo *TII); 1918 }; 1919 } // namespace 1920 1921 char PostRAMachineSinking::ID = 0; 1922 char &llvm::PostRAMachineSinkingID = PostRAMachineSinking::ID; 1923 1924 INITIALIZE_PASS(PostRAMachineSinking, "postra-machine-sink", 1925 "PostRA Machine Sink", false, false) 1926 1927 static bool aliasWithRegsInLiveIn(MachineBasicBlock &MBB, unsigned Reg, 1928 const TargetRegisterInfo *TRI) { 1929 LiveRegUnits LiveInRegUnits(*TRI); 1930 LiveInRegUnits.addLiveIns(MBB); 1931 return !LiveInRegUnits.available(Reg); 1932 } 1933 1934 static MachineBasicBlock * 1935 getSingleLiveInSuccBB(MachineBasicBlock &CurBB, 1936 const SmallPtrSetImpl<MachineBasicBlock *> &SinkableBBs, 1937 unsigned Reg, const TargetRegisterInfo *TRI) { 1938 // Try to find a single sinkable successor in which Reg is live-in. 1939 MachineBasicBlock *BB = nullptr; 1940 for (auto *SI : SinkableBBs) { 1941 if (aliasWithRegsInLiveIn(*SI, Reg, TRI)) { 1942 // If BB is set here, Reg is live-in to at least two sinkable successors, 1943 // so quit. 1944 if (BB) 1945 return nullptr; 1946 BB = SI; 1947 } 1948 } 1949 // Reg is not live-in to any sinkable successors. 1950 if (!BB) 1951 return nullptr; 1952 1953 // Check if any register aliased with Reg is live-in in other successors. 1954 for (auto *SI : CurBB.successors()) { 1955 if (!SinkableBBs.count(SI) && aliasWithRegsInLiveIn(*SI, Reg, TRI)) 1956 return nullptr; 1957 } 1958 return BB; 1959 } 1960 1961 static MachineBasicBlock * 1962 getSingleLiveInSuccBB(MachineBasicBlock &CurBB, 1963 const SmallPtrSetImpl<MachineBasicBlock *> &SinkableBBs, 1964 ArrayRef<unsigned> DefedRegsInCopy, 1965 const TargetRegisterInfo *TRI) { 1966 MachineBasicBlock *SingleBB = nullptr; 1967 for (auto DefReg : DefedRegsInCopy) { 1968 MachineBasicBlock *BB = 1969 getSingleLiveInSuccBB(CurBB, SinkableBBs, DefReg, TRI); 1970 if (!BB || (SingleBB && SingleBB != BB)) 1971 return nullptr; 1972 SingleBB = BB; 1973 } 1974 return SingleBB; 1975 } 1976 1977 static void clearKillFlags(MachineInstr *MI, MachineBasicBlock &CurBB, 1978 SmallVectorImpl<unsigned> &UsedOpsInCopy, 1979 LiveRegUnits &UsedRegUnits, 1980 const TargetRegisterInfo *TRI) { 1981 for (auto U : UsedOpsInCopy) { 1982 MachineOperand &MO = MI->getOperand(U); 1983 Register SrcReg = MO.getReg(); 1984 if (!UsedRegUnits.available(SrcReg)) { 1985 MachineBasicBlock::iterator NI = std::next(MI->getIterator()); 1986 for (MachineInstr &UI : make_range(NI, CurBB.end())) { 1987 if (UI.killsRegister(SrcReg, TRI)) { 1988 UI.clearRegisterKills(SrcReg, TRI); 1989 MO.setIsKill(true); 1990 break; 1991 } 1992 } 1993 } 1994 } 1995 } 1996 1997 static void updateLiveIn(MachineInstr *MI, MachineBasicBlock *SuccBB, 1998 SmallVectorImpl<unsigned> &UsedOpsInCopy, 1999 SmallVectorImpl<unsigned> &DefedRegsInCopy) { 2000 MachineFunction &MF = *SuccBB->getParent(); 2001 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 2002 for (unsigned DefReg : DefedRegsInCopy) 2003 for (MCPhysReg S : TRI->subregs_inclusive(DefReg)) 2004 SuccBB->removeLiveIn(S); 2005 for (auto U : UsedOpsInCopy) 2006 SuccBB->addLiveIn(MI->getOperand(U).getReg()); 2007 SuccBB->sortUniqueLiveIns(); 2008 } 2009 2010 static bool hasRegisterDependency(MachineInstr *MI, 2011 SmallVectorImpl<unsigned> &UsedOpsInCopy, 2012 SmallVectorImpl<unsigned> &DefedRegsInCopy, 2013 LiveRegUnits &ModifiedRegUnits, 2014 LiveRegUnits &UsedRegUnits) { 2015 bool HasRegDependency = false; 2016 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 2017 MachineOperand &MO = MI->getOperand(i); 2018 if (!MO.isReg()) 2019 continue; 2020 Register Reg = MO.getReg(); 2021 if (!Reg) 2022 continue; 2023 if (MO.isDef()) { 2024 if (!ModifiedRegUnits.available(Reg) || !UsedRegUnits.available(Reg)) { 2025 HasRegDependency = true; 2026 break; 2027 } 2028 DefedRegsInCopy.push_back(Reg); 2029 2030 // FIXME: instead of isUse(), readsReg() would be a better fix here, 2031 // For example, we can ignore modifications in reg with undef. However, 2032 // it's not perfectly clear if skipping the internal read is safe in all 2033 // other targets. 2034 } else if (MO.isUse()) { 2035 if (!ModifiedRegUnits.available(Reg)) { 2036 HasRegDependency = true; 2037 break; 2038 } 2039 UsedOpsInCopy.push_back(i); 2040 } 2041 } 2042 return HasRegDependency; 2043 } 2044 2045 bool PostRAMachineSinking::tryToSinkCopy(MachineBasicBlock &CurBB, 2046 MachineFunction &MF, 2047 const TargetRegisterInfo *TRI, 2048 const TargetInstrInfo *TII) { 2049 SmallPtrSet<MachineBasicBlock *, 2> SinkableBBs; 2050 // FIXME: For now, we sink only to a successor which has a single predecessor 2051 // so that we can directly sink COPY instructions to the successor without 2052 // adding any new block or branch instruction. 2053 for (MachineBasicBlock *SI : CurBB.successors()) 2054 if (!SI->livein_empty() && SI->pred_size() == 1) 2055 SinkableBBs.insert(SI); 2056 2057 if (SinkableBBs.empty()) 2058 return false; 2059 2060 bool Changed = false; 2061 2062 // Track which registers have been modified and used between the end of the 2063 // block and the current instruction. 2064 ModifiedRegUnits.clear(); 2065 UsedRegUnits.clear(); 2066 SeenDbgInstrs.clear(); 2067 2068 for (MachineInstr &MI : llvm::make_early_inc_range(llvm::reverse(CurBB))) { 2069 // Track the operand index for use in Copy. 2070 SmallVector<unsigned, 2> UsedOpsInCopy; 2071 // Track the register number defed in Copy. 2072 SmallVector<unsigned, 2> DefedRegsInCopy; 2073 2074 // We must sink this DBG_VALUE if its operand is sunk. To avoid searching 2075 // for DBG_VALUEs later, record them when they're encountered. 2076 if (MI.isDebugValue() && !MI.isDebugRef()) { 2077 SmallDenseMap<MCRegister, SmallVector<unsigned, 2>, 4> MIUnits; 2078 bool IsValid = true; 2079 for (MachineOperand &MO : MI.debug_operands()) { 2080 if (MO.isReg() && MO.getReg().isPhysical()) { 2081 // Bail if we can already tell the sink would be rejected, rather 2082 // than needlessly accumulating lots of DBG_VALUEs. 2083 if (hasRegisterDependency(&MI, UsedOpsInCopy, DefedRegsInCopy, 2084 ModifiedRegUnits, UsedRegUnits)) { 2085 IsValid = false; 2086 break; 2087 } 2088 2089 // Record debug use of each reg unit. 2090 for (MCRegUnit Unit : TRI->regunits(MO.getReg())) 2091 MIUnits[Unit].push_back(MO.getReg()); 2092 } 2093 } 2094 if (IsValid) { 2095 for (auto &RegOps : MIUnits) 2096 SeenDbgInstrs[RegOps.first].emplace_back(&MI, 2097 std::move(RegOps.second)); 2098 } 2099 continue; 2100 } 2101 2102 if (MI.isDebugOrPseudoInstr()) 2103 continue; 2104 2105 // Do not move any instruction across function call. 2106 if (MI.isCall()) 2107 return false; 2108 2109 if (!MI.isCopy() || !MI.getOperand(0).isRenamable()) { 2110 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits, 2111 TRI); 2112 continue; 2113 } 2114 2115 // Don't sink the COPY if it would violate a register dependency. 2116 if (hasRegisterDependency(&MI, UsedOpsInCopy, DefedRegsInCopy, 2117 ModifiedRegUnits, UsedRegUnits)) { 2118 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits, 2119 TRI); 2120 continue; 2121 } 2122 assert((!UsedOpsInCopy.empty() && !DefedRegsInCopy.empty()) && 2123 "Unexpect SrcReg or DefReg"); 2124 MachineBasicBlock *SuccBB = 2125 getSingleLiveInSuccBB(CurBB, SinkableBBs, DefedRegsInCopy, TRI); 2126 // Don't sink if we cannot find a single sinkable successor in which Reg 2127 // is live-in. 2128 if (!SuccBB) { 2129 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits, 2130 TRI); 2131 continue; 2132 } 2133 assert((SuccBB->pred_size() == 1 && *SuccBB->pred_begin() == &CurBB) && 2134 "Unexpected predecessor"); 2135 2136 // Collect DBG_VALUEs that must sink with this copy. We've previously 2137 // recorded which reg units that DBG_VALUEs read, if this instruction 2138 // writes any of those units then the corresponding DBG_VALUEs must sink. 2139 MapVector<MachineInstr *, MIRegs::second_type> DbgValsToSinkMap; 2140 for (auto &MO : MI.all_defs()) { 2141 for (MCRegUnit Unit : TRI->regunits(MO.getReg())) { 2142 for (const auto &MIRegs : SeenDbgInstrs.lookup(Unit)) { 2143 auto &Regs = DbgValsToSinkMap[MIRegs.first]; 2144 for (unsigned Reg : MIRegs.second) 2145 Regs.push_back(Reg); 2146 } 2147 } 2148 } 2149 auto DbgValsToSink = DbgValsToSinkMap.takeVector(); 2150 2151 LLVM_DEBUG(dbgs() << "Sink instr " << MI << "\tinto block " << *SuccBB); 2152 2153 MachineBasicBlock::iterator InsertPos = 2154 SuccBB->SkipPHIsAndLabels(SuccBB->begin()); 2155 if (blockPrologueInterferes(SuccBB, InsertPos, MI, TRI, TII, nullptr)) { 2156 LLVM_DEBUG( 2157 dbgs() << " *** Not sinking: prologue interference\n"); 2158 continue; 2159 } 2160 2161 // Clear the kill flag if SrcReg is killed between MI and the end of the 2162 // block. 2163 clearKillFlags(&MI, CurBB, UsedOpsInCopy, UsedRegUnits, TRI); 2164 performSink(MI, *SuccBB, InsertPos, DbgValsToSink); 2165 updateLiveIn(&MI, SuccBB, UsedOpsInCopy, DefedRegsInCopy); 2166 2167 Changed = true; 2168 ++NumPostRACopySink; 2169 } 2170 return Changed; 2171 } 2172 2173 bool PostRAMachineSinking::runOnMachineFunction(MachineFunction &MF) { 2174 if (skipFunction(MF.getFunction())) 2175 return false; 2176 2177 bool Changed = false; 2178 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); 2179 const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo(); 2180 2181 ModifiedRegUnits.init(*TRI); 2182 UsedRegUnits.init(*TRI); 2183 for (auto &BB : MF) 2184 Changed |= tryToSinkCopy(BB, MF, TRI, TII); 2185 2186 return Changed; 2187 } 2188