1 //===- MachineCSE.cpp - Machine Common Subexpression Elimination 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 // This pass performs global common subexpression elimination on machine 10 // instructions using a scoped hash table based value numbering scheme. It 11 // must be run while the machine function is still in SSA form. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "llvm/ADT/DenseMap.h" 16 #include "llvm/ADT/ScopedHashTable.h" 17 #include "llvm/ADT/SmallPtrSet.h" 18 #include "llvm/ADT/SmallSet.h" 19 #include "llvm/ADT/SmallVector.h" 20 #include "llvm/ADT/Statistic.h" 21 #include "llvm/Analysis/AliasAnalysis.h" 22 #include "llvm/Analysis/CFG.h" 23 #include "llvm/CodeGen/MachineBasicBlock.h" 24 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h" 25 #include "llvm/CodeGen/MachineDominators.h" 26 #include "llvm/CodeGen/MachineFunction.h" 27 #include "llvm/CodeGen/MachineFunctionPass.h" 28 #include "llvm/CodeGen/MachineInstr.h" 29 #include "llvm/CodeGen/MachineOperand.h" 30 #include "llvm/CodeGen/MachineRegisterInfo.h" 31 #include "llvm/CodeGen/Passes.h" 32 #include "llvm/CodeGen/TargetInstrInfo.h" 33 #include "llvm/CodeGen/TargetOpcodes.h" 34 #include "llvm/CodeGen/TargetRegisterInfo.h" 35 #include "llvm/CodeGen/TargetSubtargetInfo.h" 36 #include "llvm/InitializePasses.h" 37 #include "llvm/MC/MCRegister.h" 38 #include "llvm/MC/MCRegisterInfo.h" 39 #include "llvm/Pass.h" 40 #include "llvm/Support/Allocator.h" 41 #include "llvm/Support/Debug.h" 42 #include "llvm/Support/RecyclingAllocator.h" 43 #include "llvm/Support/raw_ostream.h" 44 #include <cassert> 45 #include <iterator> 46 #include <utility> 47 #include <vector> 48 49 using namespace llvm; 50 51 #define DEBUG_TYPE "machine-cse" 52 53 STATISTIC(NumCoalesces, "Number of copies coalesced"); 54 STATISTIC(NumCSEs, "Number of common subexpression eliminated"); 55 STATISTIC(NumPREs, "Number of partial redundant expression" 56 " transformed to fully redundant"); 57 STATISTIC(NumPhysCSEs, 58 "Number of physreg referencing common subexpr eliminated"); 59 STATISTIC(NumCrossBBCSEs, 60 "Number of cross-MBB physreg referencing CS eliminated"); 61 STATISTIC(NumCommutes, "Number of copies coalesced after commuting"); 62 63 // Threshold to avoid excessive cost to compute isProfitableToCSE. 64 static cl::opt<int> 65 CSUsesThreshold("csuses-threshold", cl::Hidden, cl::init(1024), 66 cl::desc("Threshold for the size of CSUses")); 67 68 namespace { 69 70 class MachineCSE : public MachineFunctionPass { 71 const TargetInstrInfo *TII; 72 const TargetRegisterInfo *TRI; 73 AliasAnalysis *AA; 74 MachineDominatorTree *DT; 75 MachineRegisterInfo *MRI; 76 MachineBlockFrequencyInfo *MBFI; 77 78 public: 79 static char ID; // Pass identification 80 81 MachineCSE() : MachineFunctionPass(ID) { 82 initializeMachineCSEPass(*PassRegistry::getPassRegistry()); 83 } 84 85 bool runOnMachineFunction(MachineFunction &MF) override; 86 87 void getAnalysisUsage(AnalysisUsage &AU) const override { 88 AU.setPreservesCFG(); 89 MachineFunctionPass::getAnalysisUsage(AU); 90 AU.addRequired<AAResultsWrapperPass>(); 91 AU.addPreservedID(MachineLoopInfoID); 92 AU.addRequired<MachineDominatorTree>(); 93 AU.addPreserved<MachineDominatorTree>(); 94 AU.addRequired<MachineBlockFrequencyInfo>(); 95 AU.addPreserved<MachineBlockFrequencyInfo>(); 96 } 97 98 MachineFunctionProperties getRequiredProperties() const override { 99 return MachineFunctionProperties() 100 .set(MachineFunctionProperties::Property::IsSSA); 101 } 102 103 void releaseMemory() override { 104 ScopeMap.clear(); 105 PREMap.clear(); 106 Exps.clear(); 107 } 108 109 private: 110 using AllocatorTy = RecyclingAllocator<BumpPtrAllocator, 111 ScopedHashTableVal<MachineInstr *, unsigned>>; 112 using ScopedHTType = 113 ScopedHashTable<MachineInstr *, unsigned, MachineInstrExpressionTrait, 114 AllocatorTy>; 115 using ScopeType = ScopedHTType::ScopeTy; 116 using PhysDefVector = SmallVector<std::pair<unsigned, unsigned>, 2>; 117 118 unsigned LookAheadLimit = 0; 119 DenseMap<MachineBasicBlock *, ScopeType *> ScopeMap; 120 DenseMap<MachineInstr *, MachineBasicBlock *, MachineInstrExpressionTrait> 121 PREMap; 122 ScopedHTType VNT; 123 SmallVector<MachineInstr *, 64> Exps; 124 unsigned CurrVN = 0; 125 126 bool PerformTrivialCopyPropagation(MachineInstr *MI, 127 MachineBasicBlock *MBB); 128 bool isPhysDefTriviallyDead(MCRegister Reg, 129 MachineBasicBlock::const_iterator I, 130 MachineBasicBlock::const_iterator E) const; 131 bool hasLivePhysRegDefUses(const MachineInstr *MI, 132 const MachineBasicBlock *MBB, 133 SmallSet<MCRegister, 8> &PhysRefs, 134 PhysDefVector &PhysDefs, bool &PhysUseDef) const; 135 bool PhysRegDefsReach(MachineInstr *CSMI, MachineInstr *MI, 136 SmallSet<MCRegister, 8> &PhysRefs, 137 PhysDefVector &PhysDefs, bool &NonLocal) const; 138 bool isCSECandidate(MachineInstr *MI); 139 bool isProfitableToCSE(Register CSReg, Register Reg, 140 MachineBasicBlock *CSBB, MachineInstr *MI); 141 void EnterScope(MachineBasicBlock *MBB); 142 void ExitScope(MachineBasicBlock *MBB); 143 bool ProcessBlockCSE(MachineBasicBlock *MBB); 144 void ExitScopeIfDone(MachineDomTreeNode *Node, 145 DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren); 146 bool PerformCSE(MachineDomTreeNode *Node); 147 148 bool isPRECandidate(MachineInstr *MI, SmallSet<MCRegister, 8> &PhysRefs); 149 bool ProcessBlockPRE(MachineDominatorTree *MDT, MachineBasicBlock *MBB); 150 bool PerformSimplePRE(MachineDominatorTree *DT); 151 /// Heuristics to see if it's profitable to move common computations of MBB 152 /// and MBB1 to CandidateBB. 153 bool isProfitableToHoistInto(MachineBasicBlock *CandidateBB, 154 MachineBasicBlock *MBB, 155 MachineBasicBlock *MBB1); 156 }; 157 158 } // end anonymous namespace 159 160 char MachineCSE::ID = 0; 161 162 char &llvm::MachineCSEID = MachineCSE::ID; 163 164 INITIALIZE_PASS_BEGIN(MachineCSE, DEBUG_TYPE, 165 "Machine Common Subexpression Elimination", false, false) 166 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) 167 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 168 INITIALIZE_PASS_END(MachineCSE, DEBUG_TYPE, 169 "Machine Common Subexpression Elimination", false, false) 170 171 /// The source register of a COPY machine instruction can be propagated to all 172 /// its users, and this propagation could increase the probability of finding 173 /// common subexpressions. If the COPY has only one user, the COPY itself can 174 /// be removed. 175 bool MachineCSE::PerformTrivialCopyPropagation(MachineInstr *MI, 176 MachineBasicBlock *MBB) { 177 bool Changed = false; 178 for (MachineOperand &MO : MI->operands()) { 179 if (!MO.isReg() || !MO.isUse()) 180 continue; 181 Register Reg = MO.getReg(); 182 if (!Reg.isVirtual()) 183 continue; 184 bool OnlyOneUse = MRI->hasOneNonDBGUse(Reg); 185 MachineInstr *DefMI = MRI->getVRegDef(Reg); 186 if (!DefMI->isCopy()) 187 continue; 188 Register SrcReg = DefMI->getOperand(1).getReg(); 189 if (!SrcReg.isVirtual()) 190 continue; 191 if (DefMI->getOperand(0).getSubReg()) 192 continue; 193 // FIXME: We should trivially coalesce subregister copies to expose CSE 194 // opportunities on instructions with truncated operands (see 195 // cse-add-with-overflow.ll). This can be done here as follows: 196 // if (SrcSubReg) 197 // RC = TRI->getMatchingSuperRegClass(MRI->getRegClass(SrcReg), RC, 198 // SrcSubReg); 199 // MO.substVirtReg(SrcReg, SrcSubReg, *TRI); 200 // 201 // The 2-addr pass has been updated to handle coalesced subregs. However, 202 // some machine-specific code still can't handle it. 203 // To handle it properly we also need a way find a constrained subregister 204 // class given a super-reg class and subreg index. 205 if (DefMI->getOperand(1).getSubReg()) 206 continue; 207 if (!MRI->constrainRegAttrs(SrcReg, Reg)) 208 continue; 209 LLVM_DEBUG(dbgs() << "Coalescing: " << *DefMI); 210 LLVM_DEBUG(dbgs() << "*** to: " << *MI); 211 212 // Propagate SrcReg of copies to MI. 213 MO.setReg(SrcReg); 214 MRI->clearKillFlags(SrcReg); 215 // Coalesce single use copies. 216 if (OnlyOneUse) { 217 // If (and only if) we've eliminated all uses of the copy, also 218 // copy-propagate to any debug-users of MI, or they'll be left using 219 // an undefined value. 220 DefMI->changeDebugValuesDefReg(SrcReg); 221 222 DefMI->eraseFromParent(); 223 ++NumCoalesces; 224 } 225 Changed = true; 226 } 227 228 return Changed; 229 } 230 231 bool MachineCSE::isPhysDefTriviallyDead( 232 MCRegister Reg, MachineBasicBlock::const_iterator I, 233 MachineBasicBlock::const_iterator E) const { 234 unsigned LookAheadLeft = LookAheadLimit; 235 while (LookAheadLeft) { 236 // Skip over dbg_value's. 237 I = skipDebugInstructionsForward(I, E); 238 239 if (I == E) 240 // Reached end of block, we don't know if register is dead or not. 241 return false; 242 243 bool SeenDef = false; 244 for (const MachineOperand &MO : I->operands()) { 245 if (MO.isRegMask() && MO.clobbersPhysReg(Reg)) 246 SeenDef = true; 247 if (!MO.isReg() || !MO.getReg()) 248 continue; 249 if (!TRI->regsOverlap(MO.getReg(), Reg)) 250 continue; 251 if (MO.isUse()) 252 // Found a use! 253 return false; 254 SeenDef = true; 255 } 256 if (SeenDef) 257 // See a def of Reg (or an alias) before encountering any use, it's 258 // trivially dead. 259 return true; 260 261 --LookAheadLeft; 262 ++I; 263 } 264 return false; 265 } 266 267 static bool isCallerPreservedOrConstPhysReg(MCRegister Reg, 268 const MachineOperand &MO, 269 const MachineFunction &MF, 270 const TargetRegisterInfo &TRI, 271 const TargetInstrInfo &TII) { 272 // MachineRegisterInfo::isConstantPhysReg directly called by 273 // MachineRegisterInfo::isCallerPreservedOrConstPhysReg expects the 274 // reserved registers to be frozen. That doesn't cause a problem post-ISel as 275 // most (if not all) targets freeze reserved registers right after ISel. 276 // 277 // It does cause issues mid-GlobalISel, however, hence the additional 278 // reservedRegsFrozen check. 279 const MachineRegisterInfo &MRI = MF.getRegInfo(); 280 return TRI.isCallerPreservedPhysReg(Reg, MF) || TII.isIgnorableUse(MO) || 281 (MRI.reservedRegsFrozen() && MRI.isConstantPhysReg(Reg)); 282 } 283 284 /// hasLivePhysRegDefUses - Return true if the specified instruction read/write 285 /// physical registers (except for dead defs of physical registers). It also 286 /// returns the physical register def by reference if it's the only one and the 287 /// instruction does not uses a physical register. 288 bool MachineCSE::hasLivePhysRegDefUses(const MachineInstr *MI, 289 const MachineBasicBlock *MBB, 290 SmallSet<MCRegister, 8> &PhysRefs, 291 PhysDefVector &PhysDefs, 292 bool &PhysUseDef) const { 293 // First, add all uses to PhysRefs. 294 for (const MachineOperand &MO : MI->operands()) { 295 if (!MO.isReg() || MO.isDef()) 296 continue; 297 Register Reg = MO.getReg(); 298 if (!Reg) 299 continue; 300 if (Reg.isVirtual()) 301 continue; 302 // Reading either caller preserved or constant physregs is ok. 303 if (!isCallerPreservedOrConstPhysReg(Reg.asMCReg(), MO, *MI->getMF(), *TRI, 304 *TII)) 305 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) 306 PhysRefs.insert(*AI); 307 } 308 309 // Next, collect all defs into PhysDefs. If any is already in PhysRefs 310 // (which currently contains only uses), set the PhysUseDef flag. 311 PhysUseDef = false; 312 MachineBasicBlock::const_iterator I = MI; I = std::next(I); 313 for (const auto &MOP : llvm::enumerate(MI->operands())) { 314 const MachineOperand &MO = MOP.value(); 315 if (!MO.isReg() || !MO.isDef()) 316 continue; 317 Register Reg = MO.getReg(); 318 if (!Reg) 319 continue; 320 if (Reg.isVirtual()) 321 continue; 322 // Check against PhysRefs even if the def is "dead". 323 if (PhysRefs.count(Reg.asMCReg())) 324 PhysUseDef = true; 325 // If the def is dead, it's ok. But the def may not marked "dead". That's 326 // common since this pass is run before livevariables. We can scan 327 // forward a few instructions and check if it is obviously dead. 328 if (!MO.isDead() && !isPhysDefTriviallyDead(Reg.asMCReg(), I, MBB->end())) 329 PhysDefs.push_back(std::make_pair(MOP.index(), Reg)); 330 } 331 332 // Finally, add all defs to PhysRefs as well. 333 for (unsigned i = 0, e = PhysDefs.size(); i != e; ++i) 334 for (MCRegAliasIterator AI(PhysDefs[i].second, TRI, true); AI.isValid(); 335 ++AI) 336 PhysRefs.insert(*AI); 337 338 return !PhysRefs.empty(); 339 } 340 341 bool MachineCSE::PhysRegDefsReach(MachineInstr *CSMI, MachineInstr *MI, 342 SmallSet<MCRegister, 8> &PhysRefs, 343 PhysDefVector &PhysDefs, 344 bool &NonLocal) const { 345 // For now conservatively returns false if the common subexpression is 346 // not in the same basic block as the given instruction. The only exception 347 // is if the common subexpression is in the sole predecessor block. 348 const MachineBasicBlock *MBB = MI->getParent(); 349 const MachineBasicBlock *CSMBB = CSMI->getParent(); 350 351 bool CrossMBB = false; 352 if (CSMBB != MBB) { 353 if (MBB->pred_size() != 1 || *MBB->pred_begin() != CSMBB) 354 return false; 355 356 for (unsigned i = 0, e = PhysDefs.size(); i != e; ++i) { 357 if (MRI->isAllocatable(PhysDefs[i].second) || 358 MRI->isReserved(PhysDefs[i].second)) 359 // Avoid extending live range of physical registers if they are 360 //allocatable or reserved. 361 return false; 362 } 363 CrossMBB = true; 364 } 365 MachineBasicBlock::const_iterator I = CSMI; I = std::next(I); 366 MachineBasicBlock::const_iterator E = MI; 367 MachineBasicBlock::const_iterator EE = CSMBB->end(); 368 unsigned LookAheadLeft = LookAheadLimit; 369 while (LookAheadLeft) { 370 // Skip over dbg_value's. 371 while (I != E && I != EE && I->isDebugInstr()) 372 ++I; 373 374 if (I == EE) { 375 assert(CrossMBB && "Reaching end-of-MBB without finding MI?"); 376 (void)CrossMBB; 377 CrossMBB = false; 378 NonLocal = true; 379 I = MBB->begin(); 380 EE = MBB->end(); 381 continue; 382 } 383 384 if (I == E) 385 return true; 386 387 for (const MachineOperand &MO : I->operands()) { 388 // RegMasks go on instructions like calls that clobber lots of physregs. 389 // Don't attempt to CSE across such an instruction. 390 if (MO.isRegMask()) 391 return false; 392 if (!MO.isReg() || !MO.isDef()) 393 continue; 394 Register MOReg = MO.getReg(); 395 if (MOReg.isVirtual()) 396 continue; 397 if (PhysRefs.count(MOReg.asMCReg())) 398 return false; 399 } 400 401 --LookAheadLeft; 402 ++I; 403 } 404 405 return false; 406 } 407 408 bool MachineCSE::isCSECandidate(MachineInstr *MI) { 409 if (MI->isPosition() || MI->isPHI() || MI->isImplicitDef() || MI->isKill() || 410 MI->isInlineAsm() || MI->isDebugInstr()) 411 return false; 412 413 // Ignore copies. 414 if (MI->isCopyLike()) 415 return false; 416 417 // Ignore stuff that we obviously can't move. 418 if (MI->mayStore() || MI->isCall() || MI->isTerminator() || 419 MI->mayRaiseFPException() || MI->hasUnmodeledSideEffects()) 420 return false; 421 422 if (MI->mayLoad()) { 423 // Okay, this instruction does a load. As a refinement, we allow the target 424 // to decide whether the loaded value is actually a constant. If so, we can 425 // actually use it as a load. 426 if (!MI->isDereferenceableInvariantLoad()) 427 // FIXME: we should be able to hoist loads with no other side effects if 428 // there are no other instructions which can change memory in this loop. 429 // This is a trivial form of alias analysis. 430 return false; 431 } 432 433 // Ignore stack guard loads, otherwise the register that holds CSEed value may 434 // be spilled and get loaded back with corrupted data. 435 if (MI->getOpcode() == TargetOpcode::LOAD_STACK_GUARD) 436 return false; 437 438 return true; 439 } 440 441 /// isProfitableToCSE - Return true if it's profitable to eliminate MI with a 442 /// common expression that defines Reg. CSBB is basic block where CSReg is 443 /// defined. 444 bool MachineCSE::isProfitableToCSE(Register CSReg, Register Reg, 445 MachineBasicBlock *CSBB, MachineInstr *MI) { 446 // FIXME: Heuristics that works around the lack the live range splitting. 447 448 // If CSReg is used at all uses of Reg, CSE should not increase register 449 // pressure of CSReg. 450 bool MayIncreasePressure = true; 451 if (CSReg.isVirtual() && Reg.isVirtual()) { 452 MayIncreasePressure = false; 453 SmallPtrSet<MachineInstr*, 8> CSUses; 454 int NumOfUses = 0; 455 for (MachineInstr &MI : MRI->use_nodbg_instructions(CSReg)) { 456 CSUses.insert(&MI); 457 // Too costly to compute if NumOfUses is very large. Conservatively assume 458 // MayIncreasePressure to avoid spending too much time here. 459 if (++NumOfUses > CSUsesThreshold) { 460 MayIncreasePressure = true; 461 break; 462 } 463 } 464 if (!MayIncreasePressure) 465 for (MachineInstr &MI : MRI->use_nodbg_instructions(Reg)) { 466 if (!CSUses.count(&MI)) { 467 MayIncreasePressure = true; 468 break; 469 } 470 } 471 } 472 if (!MayIncreasePressure) return true; 473 474 // Heuristics #1: Don't CSE "cheap" computation if the def is not local or in 475 // an immediate predecessor. We don't want to increase register pressure and 476 // end up causing other computation to be spilled. 477 if (TII->isAsCheapAsAMove(*MI)) { 478 MachineBasicBlock *BB = MI->getParent(); 479 if (CSBB != BB && !CSBB->isSuccessor(BB)) 480 return false; 481 } 482 483 // Heuristics #2: If the expression doesn't not use a vr and the only use 484 // of the redundant computation are copies, do not cse. 485 bool HasVRegUse = false; 486 for (const MachineOperand &MO : MI->operands()) { 487 if (MO.isReg() && MO.isUse() && MO.getReg().isVirtual()) { 488 HasVRegUse = true; 489 break; 490 } 491 } 492 if (!HasVRegUse) { 493 bool HasNonCopyUse = false; 494 for (MachineInstr &MI : MRI->use_nodbg_instructions(Reg)) { 495 // Ignore copies. 496 if (!MI.isCopyLike()) { 497 HasNonCopyUse = true; 498 break; 499 } 500 } 501 if (!HasNonCopyUse) 502 return false; 503 } 504 505 // Heuristics #3: If the common subexpression is used by PHIs, do not reuse 506 // it unless the defined value is already used in the BB of the new use. 507 bool HasPHI = false; 508 for (MachineInstr &UseMI : MRI->use_nodbg_instructions(CSReg)) { 509 HasPHI |= UseMI.isPHI(); 510 if (UseMI.getParent() == MI->getParent()) 511 return true; 512 } 513 514 return !HasPHI; 515 } 516 517 void MachineCSE::EnterScope(MachineBasicBlock *MBB) { 518 LLVM_DEBUG(dbgs() << "Entering: " << MBB->getName() << '\n'); 519 ScopeType *Scope = new ScopeType(VNT); 520 ScopeMap[MBB] = Scope; 521 } 522 523 void MachineCSE::ExitScope(MachineBasicBlock *MBB) { 524 LLVM_DEBUG(dbgs() << "Exiting: " << MBB->getName() << '\n'); 525 DenseMap<MachineBasicBlock*, ScopeType*>::iterator SI = ScopeMap.find(MBB); 526 assert(SI != ScopeMap.end()); 527 delete SI->second; 528 ScopeMap.erase(SI); 529 } 530 531 bool MachineCSE::ProcessBlockCSE(MachineBasicBlock *MBB) { 532 bool Changed = false; 533 534 SmallVector<std::pair<unsigned, unsigned>, 8> CSEPairs; 535 SmallVector<unsigned, 2> ImplicitDefsToUpdate; 536 SmallVector<unsigned, 2> ImplicitDefs; 537 for (MachineInstr &MI : llvm::make_early_inc_range(*MBB)) { 538 if (!isCSECandidate(&MI)) 539 continue; 540 541 bool FoundCSE = VNT.count(&MI); 542 if (!FoundCSE) { 543 // Using trivial copy propagation to find more CSE opportunities. 544 if (PerformTrivialCopyPropagation(&MI, MBB)) { 545 Changed = true; 546 547 // After coalescing MI itself may become a copy. 548 if (MI.isCopyLike()) 549 continue; 550 551 // Try again to see if CSE is possible. 552 FoundCSE = VNT.count(&MI); 553 } 554 } 555 556 // Commute commutable instructions. 557 bool Commuted = false; 558 if (!FoundCSE && MI.isCommutable()) { 559 if (MachineInstr *NewMI = TII->commuteInstruction(MI)) { 560 Commuted = true; 561 FoundCSE = VNT.count(NewMI); 562 if (NewMI != &MI) { 563 // New instruction. It doesn't need to be kept. 564 NewMI->eraseFromParent(); 565 Changed = true; 566 } else if (!FoundCSE) 567 // MI was changed but it didn't help, commute it back! 568 (void)TII->commuteInstruction(MI); 569 } 570 } 571 572 // If the instruction defines physical registers and the values *may* be 573 // used, then it's not safe to replace it with a common subexpression. 574 // It's also not safe if the instruction uses physical registers. 575 bool CrossMBBPhysDef = false; 576 SmallSet<MCRegister, 8> PhysRefs; 577 PhysDefVector PhysDefs; 578 bool PhysUseDef = false; 579 if (FoundCSE && 580 hasLivePhysRegDefUses(&MI, MBB, PhysRefs, PhysDefs, PhysUseDef)) { 581 FoundCSE = false; 582 583 // ... Unless the CS is local or is in the sole predecessor block 584 // and it also defines the physical register which is not clobbered 585 // in between and the physical register uses were not clobbered. 586 // This can never be the case if the instruction both uses and 587 // defines the same physical register, which was detected above. 588 if (!PhysUseDef) { 589 unsigned CSVN = VNT.lookup(&MI); 590 MachineInstr *CSMI = Exps[CSVN]; 591 if (PhysRegDefsReach(CSMI, &MI, PhysRefs, PhysDefs, CrossMBBPhysDef)) 592 FoundCSE = true; 593 } 594 } 595 596 if (!FoundCSE) { 597 VNT.insert(&MI, CurrVN++); 598 Exps.push_back(&MI); 599 continue; 600 } 601 602 // Found a common subexpression, eliminate it. 603 unsigned CSVN = VNT.lookup(&MI); 604 MachineInstr *CSMI = Exps[CSVN]; 605 LLVM_DEBUG(dbgs() << "Examining: " << MI); 606 LLVM_DEBUG(dbgs() << "*** Found a common subexpression: " << *CSMI); 607 608 // Prevent CSE-ing non-local convergent instructions. 609 // LLVM's current definition of `isConvergent` does not necessarily prove 610 // that non-local CSE is illegal. The following check extends the definition 611 // of `isConvergent` to assume a convergent instruction is dependent not 612 // only on additional conditions, but also on fewer conditions. LLVM does 613 // not have a MachineInstr attribute which expresses this extended 614 // definition, so it's necessary to use `isConvergent` to prevent illegally 615 // CSE-ing the subset of `isConvergent` instructions which do fall into this 616 // extended definition. 617 if (MI.isConvergent() && MI.getParent() != CSMI->getParent()) { 618 LLVM_DEBUG(dbgs() << "*** Convergent MI and subexpression exist in " 619 "different BBs, avoid CSE!\n"); 620 VNT.insert(&MI, CurrVN++); 621 Exps.push_back(&MI); 622 continue; 623 } 624 625 // Check if it's profitable to perform this CSE. 626 bool DoCSE = true; 627 unsigned NumDefs = MI.getNumDefs(); 628 629 for (unsigned i = 0, e = MI.getNumOperands(); NumDefs && i != e; ++i) { 630 MachineOperand &MO = MI.getOperand(i); 631 if (!MO.isReg() || !MO.isDef()) 632 continue; 633 Register OldReg = MO.getReg(); 634 Register NewReg = CSMI->getOperand(i).getReg(); 635 636 // Go through implicit defs of CSMI and MI, if a def is not dead at MI, 637 // we should make sure it is not dead at CSMI. 638 if (MO.isImplicit() && !MO.isDead() && CSMI->getOperand(i).isDead()) 639 ImplicitDefsToUpdate.push_back(i); 640 641 // Keep track of implicit defs of CSMI and MI, to clear possibly 642 // made-redundant kill flags. 643 if (MO.isImplicit() && !MO.isDead() && OldReg == NewReg) 644 ImplicitDefs.push_back(OldReg); 645 646 if (OldReg == NewReg) { 647 --NumDefs; 648 continue; 649 } 650 651 assert(OldReg.isVirtual() && NewReg.isVirtual() && 652 "Do not CSE physical register defs!"); 653 654 if (!isProfitableToCSE(NewReg, OldReg, CSMI->getParent(), &MI)) { 655 LLVM_DEBUG(dbgs() << "*** Not profitable, avoid CSE!\n"); 656 DoCSE = false; 657 break; 658 } 659 660 // Don't perform CSE if the result of the new instruction cannot exist 661 // within the constraints (register class, bank, or low-level type) of 662 // the old instruction. 663 if (!MRI->constrainRegAttrs(NewReg, OldReg)) { 664 LLVM_DEBUG( 665 dbgs() << "*** Not the same register constraints, avoid CSE!\n"); 666 DoCSE = false; 667 break; 668 } 669 670 CSEPairs.push_back(std::make_pair(OldReg, NewReg)); 671 --NumDefs; 672 } 673 674 // Actually perform the elimination. 675 if (DoCSE) { 676 for (const std::pair<unsigned, unsigned> &CSEPair : CSEPairs) { 677 unsigned OldReg = CSEPair.first; 678 unsigned NewReg = CSEPair.second; 679 // OldReg may have been unused but is used now, clear the Dead flag 680 MachineInstr *Def = MRI->getUniqueVRegDef(NewReg); 681 assert(Def != nullptr && "CSEd register has no unique definition?"); 682 Def->clearRegisterDeads(NewReg); 683 // Replace with NewReg and clear kill flags which may be wrong now. 684 MRI->replaceRegWith(OldReg, NewReg); 685 MRI->clearKillFlags(NewReg); 686 } 687 688 // Go through implicit defs of CSMI and MI, if a def is not dead at MI, 689 // we should make sure it is not dead at CSMI. 690 for (unsigned ImplicitDefToUpdate : ImplicitDefsToUpdate) 691 CSMI->getOperand(ImplicitDefToUpdate).setIsDead(false); 692 for (const auto &PhysDef : PhysDefs) 693 if (!MI.getOperand(PhysDef.first).isDead()) 694 CSMI->getOperand(PhysDef.first).setIsDead(false); 695 696 // Go through implicit defs of CSMI and MI, and clear the kill flags on 697 // their uses in all the instructions between CSMI and MI. 698 // We might have made some of the kill flags redundant, consider: 699 // subs ... implicit-def %nzcv <- CSMI 700 // csinc ... implicit killed %nzcv <- this kill flag isn't valid anymore 701 // subs ... implicit-def %nzcv <- MI, to be eliminated 702 // csinc ... implicit killed %nzcv 703 // Since we eliminated MI, and reused a register imp-def'd by CSMI 704 // (here %nzcv), that register, if it was killed before MI, should have 705 // that kill flag removed, because it's lifetime was extended. 706 if (CSMI->getParent() == MI.getParent()) { 707 for (MachineBasicBlock::iterator II = CSMI, IE = &MI; II != IE; ++II) 708 for (auto ImplicitDef : ImplicitDefs) 709 if (MachineOperand *MO = II->findRegisterUseOperand( 710 ImplicitDef, /*isKill=*/true, TRI)) 711 MO->setIsKill(false); 712 } else { 713 // If the instructions aren't in the same BB, bail out and clear the 714 // kill flag on all uses of the imp-def'd register. 715 for (auto ImplicitDef : ImplicitDefs) 716 MRI->clearKillFlags(ImplicitDef); 717 } 718 719 if (CrossMBBPhysDef) { 720 // Add physical register defs now coming in from a predecessor to MBB 721 // livein list. 722 while (!PhysDefs.empty()) { 723 auto LiveIn = PhysDefs.pop_back_val(); 724 if (!MBB->isLiveIn(LiveIn.second)) 725 MBB->addLiveIn(LiveIn.second); 726 } 727 ++NumCrossBBCSEs; 728 } 729 730 MI.eraseFromParent(); 731 ++NumCSEs; 732 if (!PhysRefs.empty()) 733 ++NumPhysCSEs; 734 if (Commuted) 735 ++NumCommutes; 736 Changed = true; 737 } else { 738 VNT.insert(&MI, CurrVN++); 739 Exps.push_back(&MI); 740 } 741 CSEPairs.clear(); 742 ImplicitDefsToUpdate.clear(); 743 ImplicitDefs.clear(); 744 } 745 746 return Changed; 747 } 748 749 /// ExitScopeIfDone - Destroy scope for the MBB that corresponds to the given 750 /// dominator tree node if its a leaf or all of its children are done. Walk 751 /// up the dominator tree to destroy ancestors which are now done. 752 void 753 MachineCSE::ExitScopeIfDone(MachineDomTreeNode *Node, 754 DenseMap<MachineDomTreeNode*, unsigned> &OpenChildren) { 755 if (OpenChildren[Node]) 756 return; 757 758 // Pop scope. 759 ExitScope(Node->getBlock()); 760 761 // Now traverse upwards to pop ancestors whose offsprings are all done. 762 while (MachineDomTreeNode *Parent = Node->getIDom()) { 763 unsigned Left = --OpenChildren[Parent]; 764 if (Left != 0) 765 break; 766 ExitScope(Parent->getBlock()); 767 Node = Parent; 768 } 769 } 770 771 bool MachineCSE::PerformCSE(MachineDomTreeNode *Node) { 772 SmallVector<MachineDomTreeNode*, 32> Scopes; 773 SmallVector<MachineDomTreeNode*, 8> WorkList; 774 DenseMap<MachineDomTreeNode*, unsigned> OpenChildren; 775 776 CurrVN = 0; 777 778 // Perform a DFS walk to determine the order of visit. 779 WorkList.push_back(Node); 780 do { 781 Node = WorkList.pop_back_val(); 782 Scopes.push_back(Node); 783 OpenChildren[Node] = Node->getNumChildren(); 784 append_range(WorkList, Node->children()); 785 } while (!WorkList.empty()); 786 787 // Now perform CSE. 788 bool Changed = false; 789 for (MachineDomTreeNode *Node : Scopes) { 790 MachineBasicBlock *MBB = Node->getBlock(); 791 EnterScope(MBB); 792 Changed |= ProcessBlockCSE(MBB); 793 // If it's a leaf node, it's done. Traverse upwards to pop ancestors. 794 ExitScopeIfDone(Node, OpenChildren); 795 } 796 797 return Changed; 798 } 799 800 // We use stronger checks for PRE candidate rather than for CSE ones to embrace 801 // checks inside ProcessBlockCSE(), not only inside isCSECandidate(). This helps 802 // to exclude instrs created by PRE that won't be CSEed later. 803 bool MachineCSE::isPRECandidate(MachineInstr *MI, 804 SmallSet<MCRegister, 8> &PhysRefs) { 805 if (!isCSECandidate(MI) || 806 MI->isNotDuplicable() || 807 MI->mayLoad() || 808 TII->isAsCheapAsAMove(*MI) || 809 MI->getNumDefs() != 1 || 810 MI->getNumExplicitDefs() != 1) 811 return false; 812 813 for (const MachineOperand &MO : MI->operands()) { 814 if (MO.isReg() && !MO.getReg().isVirtual()) { 815 if (MO.isDef()) 816 return false; 817 else 818 PhysRefs.insert(MO.getReg()); 819 } 820 } 821 822 return true; 823 } 824 825 bool MachineCSE::ProcessBlockPRE(MachineDominatorTree *DT, 826 MachineBasicBlock *MBB) { 827 bool Changed = false; 828 for (MachineInstr &MI : llvm::make_early_inc_range(*MBB)) { 829 SmallSet<MCRegister, 8> PhysRefs; 830 if (!isPRECandidate(&MI, PhysRefs)) 831 continue; 832 833 if (!PREMap.count(&MI)) { 834 PREMap[&MI] = MBB; 835 continue; 836 } 837 838 auto MBB1 = PREMap[&MI]; 839 assert( 840 !DT->properlyDominates(MBB, MBB1) && 841 "MBB cannot properly dominate MBB1 while DFS through dominators tree!"); 842 auto CMBB = DT->findNearestCommonDominator(MBB, MBB1); 843 if (!CMBB->isLegalToHoistInto()) 844 continue; 845 846 if (!isProfitableToHoistInto(CMBB, MBB, MBB1)) 847 continue; 848 849 // Two instrs are partial redundant if their basic blocks are reachable 850 // from one to another but one doesn't dominate another. 851 if (CMBB != MBB1) { 852 auto BB = MBB->getBasicBlock(), BB1 = MBB1->getBasicBlock(); 853 if (BB != nullptr && BB1 != nullptr && 854 (isPotentiallyReachable(BB1, BB) || 855 isPotentiallyReachable(BB, BB1))) { 856 // The following check extends the definition of `isConvergent` to 857 // assume a convergent instruction is dependent not only on additional 858 // conditions, but also on fewer conditions. LLVM does not have a 859 // MachineInstr attribute which expresses this extended definition, so 860 // it's necessary to use `isConvergent` to prevent illegally PRE-ing the 861 // subset of `isConvergent` instructions which do fall into this 862 // extended definition. 863 if (MI.isConvergent() && CMBB != MBB) 864 continue; 865 866 // If this instruction uses physical registers then we can only do PRE 867 // if it's using the value that is live at the place we're hoisting to. 868 bool NonLocal; 869 PhysDefVector PhysDefs; 870 if (!PhysRefs.empty() && 871 !PhysRegDefsReach(&*(CMBB->getFirstTerminator()), &MI, PhysRefs, 872 PhysDefs, NonLocal)) 873 continue; 874 875 assert(MI.getOperand(0).isDef() && 876 "First operand of instr with one explicit def must be this def"); 877 Register VReg = MI.getOperand(0).getReg(); 878 Register NewReg = MRI->cloneVirtualRegister(VReg); 879 if (!isProfitableToCSE(NewReg, VReg, CMBB, &MI)) 880 continue; 881 MachineInstr &NewMI = 882 TII->duplicate(*CMBB, CMBB->getFirstTerminator(), MI); 883 884 // When hoisting, make sure we don't carry the debug location of 885 // the original instruction, as that's not correct and can cause 886 // unexpected jumps when debugging optimized code. 887 auto EmptyDL = DebugLoc(); 888 NewMI.setDebugLoc(EmptyDL); 889 890 NewMI.getOperand(0).setReg(NewReg); 891 892 PREMap[&MI] = CMBB; 893 ++NumPREs; 894 Changed = true; 895 } 896 } 897 } 898 return Changed; 899 } 900 901 // This simple PRE (partial redundancy elimination) pass doesn't actually 902 // eliminate partial redundancy but transforms it to full redundancy, 903 // anticipating that the next CSE step will eliminate this created redundancy. 904 // If CSE doesn't eliminate this, than created instruction will remain dead 905 // and eliminated later by Remove Dead Machine Instructions pass. 906 bool MachineCSE::PerformSimplePRE(MachineDominatorTree *DT) { 907 SmallVector<MachineDomTreeNode *, 32> BBs; 908 909 PREMap.clear(); 910 bool Changed = false; 911 BBs.push_back(DT->getRootNode()); 912 do { 913 auto Node = BBs.pop_back_val(); 914 append_range(BBs, Node->children()); 915 916 MachineBasicBlock *MBB = Node->getBlock(); 917 Changed |= ProcessBlockPRE(DT, MBB); 918 919 } while (!BBs.empty()); 920 921 return Changed; 922 } 923 924 bool MachineCSE::isProfitableToHoistInto(MachineBasicBlock *CandidateBB, 925 MachineBasicBlock *MBB, 926 MachineBasicBlock *MBB1) { 927 if (CandidateBB->getParent()->getFunction().hasMinSize()) 928 return true; 929 assert(DT->dominates(CandidateBB, MBB) && "CandidateBB should dominate MBB"); 930 assert(DT->dominates(CandidateBB, MBB1) && 931 "CandidateBB should dominate MBB1"); 932 return MBFI->getBlockFreq(CandidateBB) <= 933 MBFI->getBlockFreq(MBB) + MBFI->getBlockFreq(MBB1); 934 } 935 936 bool MachineCSE::runOnMachineFunction(MachineFunction &MF) { 937 if (skipFunction(MF.getFunction())) 938 return false; 939 940 TII = MF.getSubtarget().getInstrInfo(); 941 TRI = MF.getSubtarget().getRegisterInfo(); 942 MRI = &MF.getRegInfo(); 943 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); 944 DT = &getAnalysis<MachineDominatorTree>(); 945 MBFI = &getAnalysis<MachineBlockFrequencyInfo>(); 946 LookAheadLimit = TII->getMachineCSELookAheadLimit(); 947 bool ChangedPRE, ChangedCSE; 948 ChangedPRE = PerformSimplePRE(DT); 949 ChangedCSE = PerformCSE(DT->getRootNode()); 950 return ChangedPRE || ChangedCSE; 951 } 952