1 //===- EarlyCSE.cpp - Simple and fast CSE pass ----------------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This pass performs a simple dominator tree walk that eliminates trivially 11 // redundant instructions. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "llvm/Transforms/Scalar/EarlyCSE.h" 16 #include "llvm/ADT/Hashing.h" 17 #include "llvm/ADT/ScopedHashTable.h" 18 #include "llvm/ADT/Statistic.h" 19 #include "llvm/Analysis/GlobalsModRef.h" 20 #include "llvm/Analysis/AssumptionCache.h" 21 #include "llvm/Analysis/InstructionSimplify.h" 22 #include "llvm/Analysis/TargetLibraryInfo.h" 23 #include "llvm/Analysis/TargetTransformInfo.h" 24 #include "llvm/IR/DataLayout.h" 25 #include "llvm/IR/Dominators.h" 26 #include "llvm/IR/Instructions.h" 27 #include "llvm/IR/IntrinsicInst.h" 28 #include "llvm/IR/PatternMatch.h" 29 #include "llvm/Pass.h" 30 #include "llvm/Support/Debug.h" 31 #include "llvm/Support/RecyclingAllocator.h" 32 #include "llvm/Support/raw_ostream.h" 33 #include "llvm/Transforms/Scalar.h" 34 #include "llvm/Transforms/Utils/Local.h" 35 #include <deque> 36 using namespace llvm; 37 using namespace llvm::PatternMatch; 38 39 #define DEBUG_TYPE "early-cse" 40 41 STATISTIC(NumSimplify, "Number of instructions simplified or DCE'd"); 42 STATISTIC(NumCSE, "Number of instructions CSE'd"); 43 STATISTIC(NumCSELoad, "Number of load instructions CSE'd"); 44 STATISTIC(NumCSECall, "Number of call instructions CSE'd"); 45 STATISTIC(NumDSE, "Number of trivial dead stores removed"); 46 47 //===----------------------------------------------------------------------===// 48 // SimpleValue 49 //===----------------------------------------------------------------------===// 50 51 namespace { 52 /// \brief Struct representing the available values in the scoped hash table. 53 struct SimpleValue { 54 Instruction *Inst; 55 56 SimpleValue(Instruction *I) : Inst(I) { 57 assert((isSentinel() || canHandle(I)) && "Inst can't be handled!"); 58 } 59 60 bool isSentinel() const { 61 return Inst == DenseMapInfo<Instruction *>::getEmptyKey() || 62 Inst == DenseMapInfo<Instruction *>::getTombstoneKey(); 63 } 64 65 static bool canHandle(Instruction *Inst) { 66 // This can only handle non-void readnone functions. 67 if (CallInst *CI = dyn_cast<CallInst>(Inst)) 68 return CI->doesNotAccessMemory() && !CI->getType()->isVoidTy(); 69 return isa<CastInst>(Inst) || isa<BinaryOperator>(Inst) || 70 isa<GetElementPtrInst>(Inst) || isa<CmpInst>(Inst) || 71 isa<SelectInst>(Inst) || isa<ExtractElementInst>(Inst) || 72 isa<InsertElementInst>(Inst) || isa<ShuffleVectorInst>(Inst) || 73 isa<ExtractValueInst>(Inst) || isa<InsertValueInst>(Inst); 74 } 75 }; 76 } 77 78 namespace llvm { 79 template <> struct DenseMapInfo<SimpleValue> { 80 static inline SimpleValue getEmptyKey() { 81 return DenseMapInfo<Instruction *>::getEmptyKey(); 82 } 83 static inline SimpleValue getTombstoneKey() { 84 return DenseMapInfo<Instruction *>::getTombstoneKey(); 85 } 86 static unsigned getHashValue(SimpleValue Val); 87 static bool isEqual(SimpleValue LHS, SimpleValue RHS); 88 }; 89 } 90 91 unsigned DenseMapInfo<SimpleValue>::getHashValue(SimpleValue Val) { 92 Instruction *Inst = Val.Inst; 93 // Hash in all of the operands as pointers. 94 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst)) { 95 Value *LHS = BinOp->getOperand(0); 96 Value *RHS = BinOp->getOperand(1); 97 if (BinOp->isCommutative() && BinOp->getOperand(0) > BinOp->getOperand(1)) 98 std::swap(LHS, RHS); 99 100 if (isa<OverflowingBinaryOperator>(BinOp)) { 101 // Hash the overflow behavior 102 unsigned Overflow = 103 BinOp->hasNoSignedWrap() * OverflowingBinaryOperator::NoSignedWrap | 104 BinOp->hasNoUnsignedWrap() * 105 OverflowingBinaryOperator::NoUnsignedWrap; 106 return hash_combine(BinOp->getOpcode(), Overflow, LHS, RHS); 107 } 108 109 return hash_combine(BinOp->getOpcode(), LHS, RHS); 110 } 111 112 if (CmpInst *CI = dyn_cast<CmpInst>(Inst)) { 113 Value *LHS = CI->getOperand(0); 114 Value *RHS = CI->getOperand(1); 115 CmpInst::Predicate Pred = CI->getPredicate(); 116 if (Inst->getOperand(0) > Inst->getOperand(1)) { 117 std::swap(LHS, RHS); 118 Pred = CI->getSwappedPredicate(); 119 } 120 return hash_combine(Inst->getOpcode(), Pred, LHS, RHS); 121 } 122 123 if (CastInst *CI = dyn_cast<CastInst>(Inst)) 124 return hash_combine(CI->getOpcode(), CI->getType(), CI->getOperand(0)); 125 126 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(Inst)) 127 return hash_combine(EVI->getOpcode(), EVI->getOperand(0), 128 hash_combine_range(EVI->idx_begin(), EVI->idx_end())); 129 130 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(Inst)) 131 return hash_combine(IVI->getOpcode(), IVI->getOperand(0), 132 IVI->getOperand(1), 133 hash_combine_range(IVI->idx_begin(), IVI->idx_end())); 134 135 assert((isa<CallInst>(Inst) || isa<BinaryOperator>(Inst) || 136 isa<GetElementPtrInst>(Inst) || isa<SelectInst>(Inst) || 137 isa<ExtractElementInst>(Inst) || isa<InsertElementInst>(Inst) || 138 isa<ShuffleVectorInst>(Inst)) && 139 "Invalid/unknown instruction"); 140 141 // Mix in the opcode. 142 return hash_combine( 143 Inst->getOpcode(), 144 hash_combine_range(Inst->value_op_begin(), Inst->value_op_end())); 145 } 146 147 bool DenseMapInfo<SimpleValue>::isEqual(SimpleValue LHS, SimpleValue RHS) { 148 Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst; 149 150 if (LHS.isSentinel() || RHS.isSentinel()) 151 return LHSI == RHSI; 152 153 if (LHSI->getOpcode() != RHSI->getOpcode()) 154 return false; 155 if (LHSI->isIdenticalTo(RHSI)) 156 return true; 157 158 // If we're not strictly identical, we still might be a commutable instruction 159 if (BinaryOperator *LHSBinOp = dyn_cast<BinaryOperator>(LHSI)) { 160 if (!LHSBinOp->isCommutative()) 161 return false; 162 163 assert(isa<BinaryOperator>(RHSI) && 164 "same opcode, but different instruction type?"); 165 BinaryOperator *RHSBinOp = cast<BinaryOperator>(RHSI); 166 167 // Check overflow attributes 168 if (isa<OverflowingBinaryOperator>(LHSBinOp)) { 169 assert(isa<OverflowingBinaryOperator>(RHSBinOp) && 170 "same opcode, but different operator type?"); 171 if (LHSBinOp->hasNoUnsignedWrap() != RHSBinOp->hasNoUnsignedWrap() || 172 LHSBinOp->hasNoSignedWrap() != RHSBinOp->hasNoSignedWrap()) 173 return false; 174 } 175 176 // Commuted equality 177 return LHSBinOp->getOperand(0) == RHSBinOp->getOperand(1) && 178 LHSBinOp->getOperand(1) == RHSBinOp->getOperand(0); 179 } 180 if (CmpInst *LHSCmp = dyn_cast<CmpInst>(LHSI)) { 181 assert(isa<CmpInst>(RHSI) && 182 "same opcode, but different instruction type?"); 183 CmpInst *RHSCmp = cast<CmpInst>(RHSI); 184 // Commuted equality 185 return LHSCmp->getOperand(0) == RHSCmp->getOperand(1) && 186 LHSCmp->getOperand(1) == RHSCmp->getOperand(0) && 187 LHSCmp->getSwappedPredicate() == RHSCmp->getPredicate(); 188 } 189 190 return false; 191 } 192 193 //===----------------------------------------------------------------------===// 194 // CallValue 195 //===----------------------------------------------------------------------===// 196 197 namespace { 198 /// \brief Struct representing the available call values in the scoped hash 199 /// table. 200 struct CallValue { 201 Instruction *Inst; 202 203 CallValue(Instruction *I) : Inst(I) { 204 assert((isSentinel() || canHandle(I)) && "Inst can't be handled!"); 205 } 206 207 bool isSentinel() const { 208 return Inst == DenseMapInfo<Instruction *>::getEmptyKey() || 209 Inst == DenseMapInfo<Instruction *>::getTombstoneKey(); 210 } 211 212 static bool canHandle(Instruction *Inst) { 213 // Don't value number anything that returns void. 214 if (Inst->getType()->isVoidTy()) 215 return false; 216 217 CallInst *CI = dyn_cast<CallInst>(Inst); 218 if (!CI || !CI->onlyReadsMemory()) 219 return false; 220 return true; 221 } 222 }; 223 } 224 225 namespace llvm { 226 template <> struct DenseMapInfo<CallValue> { 227 static inline CallValue getEmptyKey() { 228 return DenseMapInfo<Instruction *>::getEmptyKey(); 229 } 230 static inline CallValue getTombstoneKey() { 231 return DenseMapInfo<Instruction *>::getTombstoneKey(); 232 } 233 static unsigned getHashValue(CallValue Val); 234 static bool isEqual(CallValue LHS, CallValue RHS); 235 }; 236 } 237 238 unsigned DenseMapInfo<CallValue>::getHashValue(CallValue Val) { 239 Instruction *Inst = Val.Inst; 240 // Hash all of the operands as pointers and mix in the opcode. 241 return hash_combine( 242 Inst->getOpcode(), 243 hash_combine_range(Inst->value_op_begin(), Inst->value_op_end())); 244 } 245 246 bool DenseMapInfo<CallValue>::isEqual(CallValue LHS, CallValue RHS) { 247 Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst; 248 if (LHS.isSentinel() || RHS.isSentinel()) 249 return LHSI == RHSI; 250 return LHSI->isIdenticalTo(RHSI); 251 } 252 253 //===----------------------------------------------------------------------===// 254 // EarlyCSE implementation 255 //===----------------------------------------------------------------------===// 256 257 namespace { 258 /// \brief A simple and fast domtree-based CSE pass. 259 /// 260 /// This pass does a simple depth-first walk over the dominator tree, 261 /// eliminating trivially redundant instructions and using instsimplify to 262 /// canonicalize things as it goes. It is intended to be fast and catch obvious 263 /// cases so that instcombine and other passes are more effective. It is 264 /// expected that a later pass of GVN will catch the interesting/hard cases. 265 class EarlyCSE { 266 public: 267 const TargetLibraryInfo &TLI; 268 const TargetTransformInfo &TTI; 269 DominatorTree &DT; 270 AssumptionCache &AC; 271 typedef RecyclingAllocator< 272 BumpPtrAllocator, ScopedHashTableVal<SimpleValue, Value *>> AllocatorTy; 273 typedef ScopedHashTable<SimpleValue, Value *, DenseMapInfo<SimpleValue>, 274 AllocatorTy> ScopedHTType; 275 276 /// \brief A scoped hash table of the current values of all of our simple 277 /// scalar expressions. 278 /// 279 /// As we walk down the domtree, we look to see if instructions are in this: 280 /// if so, we replace them with what we find, otherwise we insert them so 281 /// that dominated values can succeed in their lookup. 282 ScopedHTType AvailableValues; 283 284 /// A scoped hash table of the current values of previously encounted memory 285 /// locations. 286 /// 287 /// This allows us to get efficient access to dominating loads or stores when 288 /// we have a fully redundant load. In addition to the most recent load, we 289 /// keep track of a generation count of the read, which is compared against 290 /// the current generation count. The current generation count is incremented 291 /// after every possibly writing memory operation, which ensures that we only 292 /// CSE loads with other loads that have no intervening store. Ordering 293 /// events (such as fences or atomic instructions) increment the generation 294 /// count as well; essentially, we model these as writes to all possible 295 /// locations. Note that atomic and/or volatile loads and stores can be 296 /// present the table; it is the responsibility of the consumer to inspect 297 /// the atomicity/volatility if needed. 298 struct LoadValue { 299 Value *Data; 300 unsigned Generation; 301 int MatchingId; 302 bool IsAtomic; 303 LoadValue() 304 : Data(nullptr), Generation(0), MatchingId(-1), IsAtomic(false) {} 305 LoadValue(Value *Data, unsigned Generation, unsigned MatchingId, 306 bool IsAtomic) 307 : Data(Data), Generation(Generation), MatchingId(MatchingId), 308 IsAtomic(IsAtomic) {} 309 }; 310 typedef RecyclingAllocator<BumpPtrAllocator, 311 ScopedHashTableVal<Value *, LoadValue>> 312 LoadMapAllocator; 313 typedef ScopedHashTable<Value *, LoadValue, DenseMapInfo<Value *>, 314 LoadMapAllocator> LoadHTType; 315 LoadHTType AvailableLoads; 316 317 /// \brief A scoped hash table of the current values of read-only call 318 /// values. 319 /// 320 /// It uses the same generation count as loads. 321 typedef ScopedHashTable<CallValue, std::pair<Value *, unsigned>> CallHTType; 322 CallHTType AvailableCalls; 323 324 /// \brief This is the current generation of the memory value. 325 unsigned CurrentGeneration; 326 327 /// \brief Set up the EarlyCSE runner for a particular function. 328 EarlyCSE(const TargetLibraryInfo &TLI, const TargetTransformInfo &TTI, 329 DominatorTree &DT, AssumptionCache &AC) 330 : TLI(TLI), TTI(TTI), DT(DT), AC(AC), CurrentGeneration(0) {} 331 332 bool run(); 333 334 private: 335 // Almost a POD, but needs to call the constructors for the scoped hash 336 // tables so that a new scope gets pushed on. These are RAII so that the 337 // scope gets popped when the NodeScope is destroyed. 338 class NodeScope { 339 public: 340 NodeScope(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads, 341 CallHTType &AvailableCalls) 342 : Scope(AvailableValues), LoadScope(AvailableLoads), 343 CallScope(AvailableCalls) {} 344 345 private: 346 NodeScope(const NodeScope &) = delete; 347 void operator=(const NodeScope &) = delete; 348 349 ScopedHTType::ScopeTy Scope; 350 LoadHTType::ScopeTy LoadScope; 351 CallHTType::ScopeTy CallScope; 352 }; 353 354 // Contains all the needed information to create a stack for doing a depth 355 // first tranversal of the tree. This includes scopes for values, loads, and 356 // calls as well as the generation. There is a child iterator so that the 357 // children do not need to be store spearately. 358 class StackNode { 359 public: 360 StackNode(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads, 361 CallHTType &AvailableCalls, unsigned cg, DomTreeNode *n, 362 DomTreeNode::iterator child, DomTreeNode::iterator end) 363 : CurrentGeneration(cg), ChildGeneration(cg), Node(n), ChildIter(child), 364 EndIter(end), Scopes(AvailableValues, AvailableLoads, AvailableCalls), 365 Processed(false) {} 366 367 // Accessors. 368 unsigned currentGeneration() { return CurrentGeneration; } 369 unsigned childGeneration() { return ChildGeneration; } 370 void childGeneration(unsigned generation) { ChildGeneration = generation; } 371 DomTreeNode *node() { return Node; } 372 DomTreeNode::iterator childIter() { return ChildIter; } 373 DomTreeNode *nextChild() { 374 DomTreeNode *child = *ChildIter; 375 ++ChildIter; 376 return child; 377 } 378 DomTreeNode::iterator end() { return EndIter; } 379 bool isProcessed() { return Processed; } 380 void process() { Processed = true; } 381 382 private: 383 StackNode(const StackNode &) = delete; 384 void operator=(const StackNode &) = delete; 385 386 // Members. 387 unsigned CurrentGeneration; 388 unsigned ChildGeneration; 389 DomTreeNode *Node; 390 DomTreeNode::iterator ChildIter; 391 DomTreeNode::iterator EndIter; 392 NodeScope Scopes; 393 bool Processed; 394 }; 395 396 /// \brief Wrapper class to handle memory instructions, including loads, 397 /// stores and intrinsic loads and stores defined by the target. 398 class ParseMemoryInst { 399 public: 400 ParseMemoryInst(Instruction *Inst, const TargetTransformInfo &TTI) 401 : IsTargetMemInst(false), Inst(Inst) { 402 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) 403 if (TTI.getTgtMemIntrinsic(II, Info) && Info.NumMemRefs == 1) 404 IsTargetMemInst = true; 405 } 406 bool isLoad() const { 407 if (IsTargetMemInst) return Info.ReadMem; 408 return isa<LoadInst>(Inst); 409 } 410 bool isStore() const { 411 if (IsTargetMemInst) return Info.WriteMem; 412 return isa<StoreInst>(Inst); 413 } 414 bool isAtomic() const { 415 if (IsTargetMemInst) { 416 assert(Info.IsSimple && "need to refine IsSimple in TTI"); 417 return false; 418 } 419 return Inst->isAtomic(); 420 } 421 bool isUnordered() const { 422 if (IsTargetMemInst) { 423 assert(Info.IsSimple && "need to refine IsSimple in TTI"); 424 return true; 425 } 426 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 427 return LI->isUnordered(); 428 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 429 return SI->isUnordered(); 430 } 431 // Conservative answer 432 return !Inst->isAtomic(); 433 } 434 435 bool isVolatile() const { 436 if (IsTargetMemInst) { 437 assert(Info.IsSimple && "need to refine IsSimple in TTI"); 438 return false; 439 } 440 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 441 return LI->isVolatile(); 442 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 443 return SI->isVolatile(); 444 } 445 // Conservative answer 446 return true; 447 } 448 449 450 bool isMatchingMemLoc(const ParseMemoryInst &Inst) const { 451 return (getPointerOperand() == Inst.getPointerOperand() && 452 getMatchingId() == Inst.getMatchingId()); 453 } 454 bool isValid() const { return getPointerOperand() != nullptr; } 455 456 // For regular (non-intrinsic) loads/stores, this is set to -1. For 457 // intrinsic loads/stores, the id is retrieved from the corresponding 458 // field in the MemIntrinsicInfo structure. That field contains 459 // non-negative values only. 460 int getMatchingId() const { 461 if (IsTargetMemInst) return Info.MatchingId; 462 return -1; 463 } 464 Value *getPointerOperand() const { 465 if (IsTargetMemInst) return Info.PtrVal; 466 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 467 return LI->getPointerOperand(); 468 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 469 return SI->getPointerOperand(); 470 } 471 return nullptr; 472 } 473 bool mayReadFromMemory() const { 474 if (IsTargetMemInst) return Info.ReadMem; 475 return Inst->mayReadFromMemory(); 476 } 477 bool mayWriteToMemory() const { 478 if (IsTargetMemInst) return Info.WriteMem; 479 return Inst->mayWriteToMemory(); 480 } 481 482 private: 483 bool IsTargetMemInst; 484 MemIntrinsicInfo Info; 485 Instruction *Inst; 486 }; 487 488 bool processNode(DomTreeNode *Node); 489 490 Value *getOrCreateResult(Value *Inst, Type *ExpectedType) const { 491 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 492 return LI; 493 else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) 494 return SI->getValueOperand(); 495 assert(isa<IntrinsicInst>(Inst) && "Instruction not supported"); 496 return TTI.getOrCreateResultFromMemIntrinsic(cast<IntrinsicInst>(Inst), 497 ExpectedType); 498 } 499 }; 500 } 501 502 bool EarlyCSE::processNode(DomTreeNode *Node) { 503 BasicBlock *BB = Node->getBlock(); 504 505 // If this block has a single predecessor, then the predecessor is the parent 506 // of the domtree node and all of the live out memory values are still current 507 // in this block. If this block has multiple predecessors, then they could 508 // have invalidated the live-out memory values of our parent value. For now, 509 // just be conservative and invalidate memory if this block has multiple 510 // predecessors. 511 if (!BB->getSinglePredecessor()) 512 ++CurrentGeneration; 513 514 // If this node has a single predecessor which ends in a conditional branch, 515 // we can infer the value of the branch condition given that we took this 516 // path. We need the single predeccesor to ensure there's not another path 517 // which reaches this block where the condition might hold a different 518 // value. Since we're adding this to the scoped hash table (like any other 519 // def), it will have been popped if we encounter a future merge block. 520 if (BasicBlock *Pred = BB->getSinglePredecessor()) 521 if (auto *BI = dyn_cast<BranchInst>(Pred->getTerminator())) 522 if (BI->isConditional()) 523 if (auto *CondInst = dyn_cast<Instruction>(BI->getCondition())) 524 if (SimpleValue::canHandle(CondInst)) { 525 assert(BI->getSuccessor(0) == BB || BI->getSuccessor(1) == BB); 526 auto *ConditionalConstant = (BI->getSuccessor(0) == BB) ? 527 ConstantInt::getTrue(BB->getContext()) : 528 ConstantInt::getFalse(BB->getContext()); 529 AvailableValues.insert(CondInst, ConditionalConstant); 530 DEBUG(dbgs() << "EarlyCSE CVP: Add conditional value for '" 531 << CondInst->getName() << "' as " << *ConditionalConstant 532 << " in " << BB->getName() << "\n"); 533 // Replace all dominated uses with the known value 534 replaceDominatedUsesWith(CondInst, ConditionalConstant, DT, 535 BasicBlockEdge(Pred, BB)); 536 } 537 538 /// LastStore - Keep track of the last non-volatile store that we saw... for 539 /// as long as there in no instruction that reads memory. If we see a store 540 /// to the same location, we delete the dead store. This zaps trivial dead 541 /// stores which can occur in bitfield code among other things. 542 Instruction *LastStore = nullptr; 543 544 bool Changed = false; 545 const DataLayout &DL = BB->getModule()->getDataLayout(); 546 547 // See if any instructions in the block can be eliminated. If so, do it. If 548 // not, add them to AvailableValues. 549 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) { 550 Instruction *Inst = &*I++; 551 552 // Dead instructions should just be removed. 553 if (isInstructionTriviallyDead(Inst, &TLI)) { 554 DEBUG(dbgs() << "EarlyCSE DCE: " << *Inst << '\n'); 555 Inst->eraseFromParent(); 556 Changed = true; 557 ++NumSimplify; 558 continue; 559 } 560 561 // Skip assume intrinsics, they don't really have side effects (although 562 // they're marked as such to ensure preservation of control dependencies), 563 // and this pass will not disturb any of the assumption's control 564 // dependencies. 565 if (match(Inst, m_Intrinsic<Intrinsic::assume>())) { 566 DEBUG(dbgs() << "EarlyCSE skipping assumption: " << *Inst << '\n'); 567 continue; 568 } 569 570 // If the instruction can be simplified (e.g. X+0 = X) then replace it with 571 // its simpler value. 572 if (Value *V = SimplifyInstruction(Inst, DL, &TLI, &DT, &AC)) { 573 DEBUG(dbgs() << "EarlyCSE Simplify: " << *Inst << " to: " << *V << '\n'); 574 Inst->replaceAllUsesWith(V); 575 Inst->eraseFromParent(); 576 Changed = true; 577 ++NumSimplify; 578 continue; 579 } 580 581 // If this is a simple instruction that we can value number, process it. 582 if (SimpleValue::canHandle(Inst)) { 583 // See if the instruction has an available value. If so, use it. 584 if (Value *V = AvailableValues.lookup(Inst)) { 585 DEBUG(dbgs() << "EarlyCSE CSE: " << *Inst << " to: " << *V << '\n'); 586 Inst->replaceAllUsesWith(V); 587 Inst->eraseFromParent(); 588 Changed = true; 589 ++NumCSE; 590 continue; 591 } 592 593 // Otherwise, just remember that this value is available. 594 AvailableValues.insert(Inst, Inst); 595 continue; 596 } 597 598 ParseMemoryInst MemInst(Inst, TTI); 599 // If this is a non-volatile load, process it. 600 if (MemInst.isValid() && MemInst.isLoad()) { 601 // (conservatively) we can't peak past the ordering implied by this 602 // operation, but we can add this load to our set of available values 603 if (MemInst.isVolatile() || !MemInst.isUnordered()) { 604 LastStore = nullptr; 605 ++CurrentGeneration; 606 } 607 608 // If we have an available version of this load, and if it is the right 609 // generation, replace this instruction. 610 LoadValue InVal = AvailableLoads.lookup(MemInst.getPointerOperand()); 611 if (InVal.Data != nullptr && InVal.Generation == CurrentGeneration && 612 InVal.MatchingId == MemInst.getMatchingId() && 613 // We don't yet handle removing loads with ordering of any kind. 614 !MemInst.isVolatile() && MemInst.isUnordered() && 615 // We can't replace an atomic load with one which isn't also atomic. 616 InVal.IsAtomic >= MemInst.isAtomic()) { 617 Value *Op = getOrCreateResult(InVal.Data, Inst->getType()); 618 if (Op != nullptr) { 619 DEBUG(dbgs() << "EarlyCSE CSE LOAD: " << *Inst 620 << " to: " << *InVal.Data << '\n'); 621 if (!Inst->use_empty()) 622 Inst->replaceAllUsesWith(Op); 623 Inst->eraseFromParent(); 624 Changed = true; 625 ++NumCSELoad; 626 continue; 627 } 628 } 629 630 // Otherwise, remember that we have this instruction. 631 AvailableLoads.insert( 632 MemInst.getPointerOperand(), 633 LoadValue(Inst, CurrentGeneration, MemInst.getMatchingId(), 634 MemInst.isAtomic())); 635 LastStore = nullptr; 636 continue; 637 } 638 639 // If this instruction may read from memory, forget LastStore. 640 // Load/store intrinsics will indicate both a read and a write to 641 // memory. The target may override this (e.g. so that a store intrinsic 642 // does not read from memory, and thus will be treated the same as a 643 // regular store for commoning purposes). 644 if (Inst->mayReadFromMemory() && 645 !(MemInst.isValid() && !MemInst.mayReadFromMemory())) 646 LastStore = nullptr; 647 648 // If this is a read-only call, process it. 649 if (CallValue::canHandle(Inst)) { 650 // If we have an available version of this call, and if it is the right 651 // generation, replace this instruction. 652 std::pair<Value *, unsigned> InVal = AvailableCalls.lookup(Inst); 653 if (InVal.first != nullptr && InVal.second == CurrentGeneration) { 654 DEBUG(dbgs() << "EarlyCSE CSE CALL: " << *Inst 655 << " to: " << *InVal.first << '\n'); 656 if (!Inst->use_empty()) 657 Inst->replaceAllUsesWith(InVal.first); 658 Inst->eraseFromParent(); 659 Changed = true; 660 ++NumCSECall; 661 continue; 662 } 663 664 // Otherwise, remember that we have this instruction. 665 AvailableCalls.insert( 666 Inst, std::pair<Value *, unsigned>(Inst, CurrentGeneration)); 667 continue; 668 } 669 670 // A release fence requires that all stores complete before it, but does 671 // not prevent the reordering of following loads 'before' the fence. As a 672 // result, we don't need to consider it as writing to memory and don't need 673 // to advance the generation. We do need to prevent DSE across the fence, 674 // but that's handled above. 675 if (FenceInst *FI = dyn_cast<FenceInst>(Inst)) 676 if (FI->getOrdering() == Release) { 677 assert(Inst->mayReadFromMemory() && "relied on to prevent DSE above"); 678 continue; 679 } 680 681 // write back DSE - If we write back the same value we just loaded from 682 // the same location and haven't passed any intervening writes or ordering 683 // operations, we can remove the write. The primary benefit is in allowing 684 // the available load table to remain valid and value forward past where 685 // the store originally was. 686 if (MemInst.isValid() && MemInst.isStore()) { 687 LoadValue InVal = AvailableLoads.lookup(MemInst.getPointerOperand()); 688 if (InVal.Data && 689 InVal.Data == getOrCreateResult(Inst, InVal.Data->getType()) && 690 InVal.Generation == CurrentGeneration && 691 InVal.MatchingId == MemInst.getMatchingId() && 692 // We don't yet handle removing stores with ordering of any kind. 693 !MemInst.isVolatile() && MemInst.isUnordered()) { 694 assert((!LastStore || 695 ParseMemoryInst(LastStore, TTI).getPointerOperand() == 696 MemInst.getPointerOperand()) && 697 "can't have an intervening store!"); 698 DEBUG(dbgs() << "EarlyCSE DSE (writeback): " << *Inst << '\n'); 699 Inst->eraseFromParent(); 700 Changed = true; 701 ++NumDSE; 702 // We can avoid incrementing the generation count since we were able 703 // to eliminate this store. 704 continue; 705 } 706 } 707 708 // Okay, this isn't something we can CSE at all. Check to see if it is 709 // something that could modify memory. If so, our available memory values 710 // cannot be used so bump the generation count. 711 if (Inst->mayWriteToMemory()) { 712 ++CurrentGeneration; 713 714 if (MemInst.isValid() && MemInst.isStore()) { 715 // We do a trivial form of DSE if there are two stores to the same 716 // location with no intervening loads. Delete the earlier store. 717 // At the moment, we don't remove ordered stores, but do remove 718 // unordered atomic stores. There's no special requirement (for 719 // unordered atomics) about removing atomic stores only in favor of 720 // other atomic stores since we we're going to execute the non-atomic 721 // one anyway and the atomic one might never have become visible. 722 if (LastStore) { 723 ParseMemoryInst LastStoreMemInst(LastStore, TTI); 724 assert(LastStoreMemInst.isUnordered() && 725 !LastStoreMemInst.isVolatile() && 726 "Violated invariant"); 727 if (LastStoreMemInst.isMatchingMemLoc(MemInst)) { 728 DEBUG(dbgs() << "EarlyCSE DEAD STORE: " << *LastStore 729 << " due to: " << *Inst << '\n'); 730 LastStore->eraseFromParent(); 731 Changed = true; 732 ++NumDSE; 733 LastStore = nullptr; 734 } 735 // fallthrough - we can exploit information about this store 736 } 737 738 // Okay, we just invalidated anything we knew about loaded values. Try 739 // to salvage *something* by remembering that the stored value is a live 740 // version of the pointer. It is safe to forward from volatile stores 741 // to non-volatile loads, so we don't have to check for volatility of 742 // the store. 743 AvailableLoads.insert( 744 MemInst.getPointerOperand(), 745 LoadValue(Inst, CurrentGeneration, MemInst.getMatchingId(), 746 MemInst.isAtomic())); 747 748 // Remember that this was the last unordered store we saw for DSE. We 749 // don't yet handle DSE on ordered or volatile stores since we don't 750 // have a good way to model the ordering requirement for following 751 // passes once the store is removed. We could insert a fence, but 752 // since fences are slightly stronger than stores in their ordering, 753 // it's not clear this is a profitable transform. Another option would 754 // be to merge the ordering with that of the post dominating store. 755 if (MemInst.isUnordered() && !MemInst.isVolatile()) 756 LastStore = Inst; 757 else 758 LastStore = nullptr; 759 } 760 } 761 } 762 763 return Changed; 764 } 765 766 bool EarlyCSE::run() { 767 // Note, deque is being used here because there is significant performance 768 // gains over vector when the container becomes very large due to the 769 // specific access patterns. For more information see the mailing list 770 // discussion on this: 771 // http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20120116/135228.html 772 std::deque<StackNode *> nodesToProcess; 773 774 bool Changed = false; 775 776 // Process the root node. 777 nodesToProcess.push_back(new StackNode( 778 AvailableValues, AvailableLoads, AvailableCalls, CurrentGeneration, 779 DT.getRootNode(), DT.getRootNode()->begin(), DT.getRootNode()->end())); 780 781 // Save the current generation. 782 unsigned LiveOutGeneration = CurrentGeneration; 783 784 // Process the stack. 785 while (!nodesToProcess.empty()) { 786 // Grab the first item off the stack. Set the current generation, remove 787 // the node from the stack, and process it. 788 StackNode *NodeToProcess = nodesToProcess.back(); 789 790 // Initialize class members. 791 CurrentGeneration = NodeToProcess->currentGeneration(); 792 793 // Check if the node needs to be processed. 794 if (!NodeToProcess->isProcessed()) { 795 // Process the node. 796 Changed |= processNode(NodeToProcess->node()); 797 NodeToProcess->childGeneration(CurrentGeneration); 798 NodeToProcess->process(); 799 } else if (NodeToProcess->childIter() != NodeToProcess->end()) { 800 // Push the next child onto the stack. 801 DomTreeNode *child = NodeToProcess->nextChild(); 802 nodesToProcess.push_back( 803 new StackNode(AvailableValues, AvailableLoads, AvailableCalls, 804 NodeToProcess->childGeneration(), child, child->begin(), 805 child->end())); 806 } else { 807 // It has been processed, and there are no more children to process, 808 // so delete it and pop it off the stack. 809 delete NodeToProcess; 810 nodesToProcess.pop_back(); 811 } 812 } // while (!nodes...) 813 814 // Reset the current generation. 815 CurrentGeneration = LiveOutGeneration; 816 817 return Changed; 818 } 819 820 PreservedAnalyses EarlyCSEPass::run(Function &F, 821 AnalysisManager<Function> *AM) { 822 auto &TLI = AM->getResult<TargetLibraryAnalysis>(F); 823 auto &TTI = AM->getResult<TargetIRAnalysis>(F); 824 auto &DT = AM->getResult<DominatorTreeAnalysis>(F); 825 auto &AC = AM->getResult<AssumptionAnalysis>(F); 826 827 EarlyCSE CSE(TLI, TTI, DT, AC); 828 829 if (!CSE.run()) 830 return PreservedAnalyses::all(); 831 832 // CSE preserves the dominator tree because it doesn't mutate the CFG. 833 // FIXME: Bundle this with other CFG-preservation. 834 PreservedAnalyses PA; 835 PA.preserve<DominatorTreeAnalysis>(); 836 return PA; 837 } 838 839 namespace { 840 /// \brief A simple and fast domtree-based CSE pass. 841 /// 842 /// This pass does a simple depth-first walk over the dominator tree, 843 /// eliminating trivially redundant instructions and using instsimplify to 844 /// canonicalize things as it goes. It is intended to be fast and catch obvious 845 /// cases so that instcombine and other passes are more effective. It is 846 /// expected that a later pass of GVN will catch the interesting/hard cases. 847 class EarlyCSELegacyPass : public FunctionPass { 848 public: 849 static char ID; 850 851 EarlyCSELegacyPass() : FunctionPass(ID) { 852 initializeEarlyCSELegacyPassPass(*PassRegistry::getPassRegistry()); 853 } 854 855 bool runOnFunction(Function &F) override { 856 if (skipOptnoneFunction(F)) 857 return false; 858 859 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(); 860 auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 861 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 862 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 863 864 EarlyCSE CSE(TLI, TTI, DT, AC); 865 866 return CSE.run(); 867 } 868 869 void getAnalysisUsage(AnalysisUsage &AU) const override { 870 AU.addRequired<AssumptionCacheTracker>(); 871 AU.addRequired<DominatorTreeWrapperPass>(); 872 AU.addRequired<TargetLibraryInfoWrapperPass>(); 873 AU.addRequired<TargetTransformInfoWrapperPass>(); 874 AU.addPreserved<GlobalsAAWrapperPass>(); 875 AU.setPreservesCFG(); 876 } 877 }; 878 } 879 880 char EarlyCSELegacyPass::ID = 0; 881 882 FunctionPass *llvm::createEarlyCSEPass() { return new EarlyCSELegacyPass(); } 883 884 INITIALIZE_PASS_BEGIN(EarlyCSELegacyPass, "early-cse", "Early CSE", false, 885 false) 886 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 887 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 888 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 889 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 890 INITIALIZE_PASS_END(EarlyCSELegacyPass, "early-cse", "Early CSE", false, false) 891