1 //===- MemorySSA.h - Build Memory SSA ---------------------------*- C++ -*-===// 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 // \file 11 // \brief This file exposes an interface to building/using memory SSA to 12 // walk memory instructions using a use/def graph. 13 // 14 // Memory SSA class builds an SSA form that links together memory access 15 // instructions such as loads, stores, atomics, and calls. Additionally, it does 16 // a trivial form of "heap versioning" Every time the memory state changes in 17 // the program, we generate a new heap version. It generates MemoryDef/Uses/Phis 18 // that are overlayed on top of the existing instructions. 19 // 20 // As a trivial example, 21 // define i32 @main() #0 { 22 // entry: 23 // %call = call noalias i8* @_Znwm(i64 4) #2 24 // %0 = bitcast i8* %call to i32* 25 // %call1 = call noalias i8* @_Znwm(i64 4) #2 26 // %1 = bitcast i8* %call1 to i32* 27 // store i32 5, i32* %0, align 4 28 // store i32 7, i32* %1, align 4 29 // %2 = load i32* %0, align 4 30 // %3 = load i32* %1, align 4 31 // %add = add nsw i32 %2, %3 32 // ret i32 %add 33 // } 34 // 35 // Will become 36 // define i32 @main() #0 { 37 // entry: 38 // ; 1 = MemoryDef(0) 39 // %call = call noalias i8* @_Znwm(i64 4) #3 40 // %2 = bitcast i8* %call to i32* 41 // ; 2 = MemoryDef(1) 42 // %call1 = call noalias i8* @_Znwm(i64 4) #3 43 // %4 = bitcast i8* %call1 to i32* 44 // ; 3 = MemoryDef(2) 45 // store i32 5, i32* %2, align 4 46 // ; 4 = MemoryDef(3) 47 // store i32 7, i32* %4, align 4 48 // ; MemoryUse(3) 49 // %7 = load i32* %2, align 4 50 // ; MemoryUse(4) 51 // %8 = load i32* %4, align 4 52 // %add = add nsw i32 %7, %8 53 // ret i32 %add 54 // } 55 // 56 // Given this form, all the stores that could ever effect the load at %8 can be 57 // gotten by using the MemoryUse associated with it, and walking from use to def 58 // until you hit the top of the function. 59 // 60 // Each def also has a list of users associated with it, so you can walk from 61 // both def to users, and users to defs. Note that we disambiguate MemoryUses, 62 // but not the RHS of MemoryDefs. You can see this above at %7, which would 63 // otherwise be a MemoryUse(4). Being disambiguated means that for a given 64 // store, all the MemoryUses on its use lists are may-aliases of that store (but 65 // the MemoryDefs on its use list may not be). 66 // 67 // MemoryDefs are not disambiguated because it would require multiple reaching 68 // definitions, which would require multiple phis, and multiple memoryaccesses 69 // per instruction. 70 //===----------------------------------------------------------------------===// 71 72 #ifndef LLVM_TRANSFORMS_UTILS_MEMORYSSA_H 73 #define LLVM_TRANSFORMS_UTILS_MEMORYSSA_H 74 75 #include "llvm/ADT/DenseMap.h" 76 #include "llvm/ADT/GraphTraits.h" 77 #include "llvm/ADT/SmallPtrSet.h" 78 #include "llvm/ADT/SmallVector.h" 79 #include "llvm/ADT/ilist.h" 80 #include "llvm/ADT/ilist_node.h" 81 #include "llvm/ADT/iterator.h" 82 #include "llvm/Analysis/AliasAnalysis.h" 83 #include "llvm/Analysis/MemoryLocation.h" 84 #include "llvm/Analysis/PHITransAddr.h" 85 #include "llvm/IR/BasicBlock.h" 86 #include "llvm/IR/Dominators.h" 87 #include "llvm/IR/Module.h" 88 #include "llvm/IR/OperandTraits.h" 89 #include "llvm/IR/Type.h" 90 #include "llvm/IR/Use.h" 91 #include "llvm/IR/User.h" 92 #include "llvm/IR/Value.h" 93 #include "llvm/Pass.h" 94 #include "llvm/PassAnalysisSupport.h" 95 #include "llvm/Support/Casting.h" 96 #include "llvm/Support/Compiler.h" 97 #include "llvm/Support/ErrorHandling.h" 98 #include <algorithm> 99 #include <cassert> 100 #include <cstddef> 101 #include <iterator> 102 #include <memory> 103 #include <utility> 104 105 namespace llvm { 106 107 class DominatorTree; 108 class Function; 109 class Instruction; 110 class MemoryAccess; 111 class LLVMContext; 112 class raw_ostream; 113 114 template <class T> class memoryaccess_def_iterator_base; 115 using memoryaccess_def_iterator = memoryaccess_def_iterator_base<MemoryAccess>; 116 using const_memoryaccess_def_iterator = 117 memoryaccess_def_iterator_base<const MemoryAccess>; 118 119 // \brief The base for all memory accesses. All memory accesses in a block are 120 // linked together using an intrusive list. 121 class MemoryAccess : public User, public ilist_node<MemoryAccess> { 122 void *operator new(size_t, unsigned) = delete; 123 void *operator new(size_t) = delete; 124 125 public: 126 // Methods for support type inquiry through isa, cast, and 127 // dyn_cast classof(const MemoryAccess *)128 static inline bool classof(const MemoryAccess *) { return true; } classof(const Value * V)129 static inline bool classof(const Value *V) { 130 unsigned ID = V->getValueID(); 131 return ID == MemoryUseVal || ID == MemoryPhiVal || ID == MemoryDefVal; 132 } 133 134 ~MemoryAccess() override; 135 getBlock()136 BasicBlock *getBlock() const { return Block; } 137 138 virtual void print(raw_ostream &OS) const = 0; 139 virtual void dump() const; 140 141 /// \brief The user iterators for a memory access 142 typedef user_iterator iterator; 143 typedef const_user_iterator const_iterator; 144 145 /// \brief This iterator walks over all of the defs in a given 146 /// MemoryAccess. For MemoryPhi nodes, this walks arguments. For 147 /// MemoryUse/MemoryDef, this walks the defining access. 148 memoryaccess_def_iterator defs_begin(); 149 const_memoryaccess_def_iterator defs_begin() const; 150 memoryaccess_def_iterator defs_end(); 151 const_memoryaccess_def_iterator defs_end() const; 152 153 protected: 154 friend class MemorySSA; 155 friend class MemoryUseOrDef; 156 friend class MemoryUse; 157 friend class MemoryDef; 158 friend class MemoryPhi; 159 160 /// \brief Used internally to give IDs to MemoryAccesses for printing 161 virtual unsigned getID() const = 0; 162 MemoryAccess(LLVMContext & C,unsigned Vty,BasicBlock * BB,unsigned NumOperands)163 MemoryAccess(LLVMContext &C, unsigned Vty, BasicBlock *BB, 164 unsigned NumOperands) 165 : User(Type::getVoidTy(C), Vty, nullptr, NumOperands), Block(BB) {} 166 167 private: 168 MemoryAccess(const MemoryAccess &); 169 void operator=(const MemoryAccess &); 170 BasicBlock *Block; 171 }; 172 173 template <> 174 struct ilist_traits<MemoryAccess> : public ilist_default_traits<MemoryAccess> { 175 /// See details of the instruction class for why this trick works 176 // FIXME: This downcast is UB. See llvm.org/PR26753. 177 LLVM_NO_SANITIZE("object-size") 178 MemoryAccess *createSentinel() const { 179 return static_cast<MemoryAccess *>(&Sentinel); 180 } 181 182 static void destroySentinel(MemoryAccess *) {} 183 184 MemoryAccess *provideInitialHead() const { return createSentinel(); } 185 MemoryAccess *ensureHead(MemoryAccess *) const { return createSentinel(); } 186 static void noteHead(MemoryAccess *, MemoryAccess *) {} 187 188 private: 189 mutable ilist_half_node<MemoryAccess> Sentinel; 190 }; 191 192 inline raw_ostream &operator<<(raw_ostream &OS, const MemoryAccess &MA) { 193 MA.print(OS); 194 return OS; 195 } 196 197 /// \brief Class that has the common methods + fields of memory uses/defs. It's 198 /// a little awkward to have, but there are many cases where we want either a 199 /// use or def, and there are many cases where uses are needed (defs aren't 200 /// acceptable), and vice-versa. 201 /// 202 /// This class should never be instantiated directly; make a MemoryUse or 203 /// MemoryDef instead. 204 class MemoryUseOrDef : public MemoryAccess { 205 void *operator new(size_t, unsigned) = delete; 206 void *operator new(size_t) = delete; 207 208 public: 209 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(MemoryAccess); 210 211 /// \brief Get the instruction that this MemoryUse represents. 212 Instruction *getMemoryInst() const { return MemoryInst; } 213 214 /// \brief Get the access that produces the memory state used by this Use. 215 MemoryAccess *getDefiningAccess() const { return getOperand(0); } 216 217 static inline bool classof(const MemoryUseOrDef *) { return true; } 218 static inline bool classof(const Value *MA) { 219 return MA->getValueID() == MemoryUseVal || MA->getValueID() == MemoryDefVal; 220 } 221 222 protected: 223 friend class MemorySSA; 224 225 MemoryUseOrDef(LLVMContext &C, MemoryAccess *DMA, unsigned Vty, 226 Instruction *MI, BasicBlock *BB) 227 : MemoryAccess(C, Vty, BB, 1), MemoryInst(MI) { 228 setDefiningAccess(DMA); 229 } 230 231 void setDefiningAccess(MemoryAccess *DMA) { setOperand(0, DMA); } 232 233 private: 234 Instruction *MemoryInst; 235 }; 236 237 template <> 238 struct OperandTraits<MemoryUseOrDef> 239 : public FixedNumOperandTraits<MemoryUseOrDef, 1> {}; 240 DEFINE_TRANSPARENT_OPERAND_ACCESSORS(MemoryUseOrDef, MemoryAccess) 241 242 /// \brief Represents read-only accesses to memory 243 /// 244 /// In particular, the set of Instructions that will be represented by 245 /// MemoryUse's is exactly the set of Instructions for which 246 /// AliasAnalysis::getModRefInfo returns "Ref". 247 class MemoryUse final : public MemoryUseOrDef { 248 void *operator new(size_t, unsigned) = delete; 249 250 public: 251 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(MemoryAccess); 252 253 // allocate space for exactly one operand 254 void *operator new(size_t s) { return User::operator new(s, 1); } 255 256 MemoryUse(LLVMContext &C, MemoryAccess *DMA, Instruction *MI, BasicBlock *BB) 257 : MemoryUseOrDef(C, DMA, MemoryUseVal, MI, BB) {} 258 259 static inline bool classof(const MemoryUse *) { return true; } 260 static inline bool classof(const Value *MA) { 261 return MA->getValueID() == MemoryUseVal; 262 } 263 264 void print(raw_ostream &OS) const override; 265 266 protected: 267 friend class MemorySSA; 268 269 unsigned getID() const override { 270 llvm_unreachable("MemoryUses do not have IDs"); 271 } 272 }; 273 274 template <> 275 struct OperandTraits<MemoryUse> : public FixedNumOperandTraits<MemoryUse, 1> {}; 276 DEFINE_TRANSPARENT_OPERAND_ACCESSORS(MemoryUse, MemoryAccess) 277 278 /// \brief Represents a read-write access to memory, whether it is a must-alias, 279 /// or a may-alias. 280 /// 281 /// In particular, the set of Instructions that will be represented by 282 /// MemoryDef's is exactly the set of Instructions for which 283 /// AliasAnalysis::getModRefInfo returns "Mod" or "ModRef". 284 /// Note that, in order to provide def-def chains, all defs also have a use 285 /// associated with them. This use points to the nearest reaching 286 /// MemoryDef/MemoryPhi. 287 class MemoryDef final : public MemoryUseOrDef { 288 void *operator new(size_t, unsigned) = delete; 289 290 public: 291 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(MemoryAccess); 292 293 // allocate space for exactly one operand 294 void *operator new(size_t s) { return User::operator new(s, 1); } 295 296 MemoryDef(LLVMContext &C, MemoryAccess *DMA, Instruction *MI, BasicBlock *BB, 297 unsigned Ver) 298 : MemoryUseOrDef(C, DMA, MemoryDefVal, MI, BB), ID(Ver) {} 299 300 static inline bool classof(const MemoryDef *) { return true; } 301 static inline bool classof(const Value *MA) { 302 return MA->getValueID() == MemoryDefVal; 303 } 304 305 void print(raw_ostream &OS) const override; 306 307 protected: 308 friend class MemorySSA; 309 310 // For debugging only. This gets used to give memory accesses pretty numbers 311 // when printing them out 312 unsigned getID() const override { return ID; } 313 314 private: 315 const unsigned ID; 316 }; 317 318 template <> 319 struct OperandTraits<MemoryDef> : public FixedNumOperandTraits<MemoryDef, 1> {}; 320 DEFINE_TRANSPARENT_OPERAND_ACCESSORS(MemoryDef, MemoryAccess) 321 322 /// \brief Represents phi nodes for memory accesses. 323 /// 324 /// These have the same semantic as regular phi nodes, with the exception that 325 /// only one phi will ever exist in a given basic block. 326 /// Guaranteeing one phi per block means guaranteeing there is only ever one 327 /// valid reaching MemoryDef/MemoryPHI along each path to the phi node. 328 /// This is ensured by not allowing disambiguation of the RHS of a MemoryDef or 329 /// a MemoryPhi's operands. 330 /// That is, given 331 /// if (a) { 332 /// store %a 333 /// store %b 334 /// } 335 /// it *must* be transformed into 336 /// if (a) { 337 /// 1 = MemoryDef(liveOnEntry) 338 /// store %a 339 /// 2 = MemoryDef(1) 340 /// store %b 341 /// } 342 /// and *not* 343 /// if (a) { 344 /// 1 = MemoryDef(liveOnEntry) 345 /// store %a 346 /// 2 = MemoryDef(liveOnEntry) 347 /// store %b 348 /// } 349 /// even if the two stores do not conflict. Otherwise, both 1 and 2 reach the 350 /// end of the branch, and if there are not two phi nodes, one will be 351 /// disconnected completely from the SSA graph below that point. 352 /// Because MemoryUse's do not generate new definitions, they do not have this 353 /// issue. 354 class MemoryPhi final : public MemoryAccess { 355 void *operator new(size_t, unsigned) = delete; 356 // allocate space for exactly zero operands 357 void *operator new(size_t s) { return User::operator new(s); } 358 359 public: 360 /// Provide fast operand accessors 361 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(MemoryAccess); 362 363 MemoryPhi(LLVMContext &C, BasicBlock *BB, unsigned Ver, unsigned NumPreds = 0) 364 : MemoryAccess(C, MemoryPhiVal, BB, 0), ID(Ver), ReservedSpace(NumPreds) { 365 allocHungoffUses(ReservedSpace); 366 } 367 368 // Block iterator interface. This provides access to the list of incoming 369 // basic blocks, which parallels the list of incoming values. 370 typedef BasicBlock **block_iterator; 371 typedef BasicBlock *const *const_block_iterator; 372 373 block_iterator block_begin() { 374 auto *Ref = reinterpret_cast<Use::UserRef *>(op_begin() + ReservedSpace); 375 return reinterpret_cast<block_iterator>(Ref + 1); 376 } 377 378 const_block_iterator block_begin() const { 379 const auto *Ref = 380 reinterpret_cast<const Use::UserRef *>(op_begin() + ReservedSpace); 381 return reinterpret_cast<const_block_iterator>(Ref + 1); 382 } 383 384 block_iterator block_end() { return block_begin() + getNumOperands(); } 385 386 const_block_iterator block_end() const { 387 return block_begin() + getNumOperands(); 388 } 389 390 op_range incoming_values() { return operands(); } 391 392 const_op_range incoming_values() const { return operands(); } 393 394 /// \brief Return the number of incoming edges 395 unsigned getNumIncomingValues() const { return getNumOperands(); } 396 397 /// \brief Return incoming value number x 398 MemoryAccess *getIncomingValue(unsigned I) const { return getOperand(I); } 399 void setIncomingValue(unsigned I, MemoryAccess *V) { 400 assert(V && "PHI node got a null value!"); 401 setOperand(I, V); 402 } 403 static unsigned getOperandNumForIncomingValue(unsigned I) { return I; } 404 static unsigned getIncomingValueNumForOperand(unsigned I) { return I; } 405 406 /// \brief Return incoming basic block number @p i. 407 BasicBlock *getIncomingBlock(unsigned I) const { return block_begin()[I]; } 408 409 /// \brief Return incoming basic block corresponding 410 /// to an operand of the PHI. 411 BasicBlock *getIncomingBlock(const Use &U) const { 412 assert(this == U.getUser() && "Iterator doesn't point to PHI's Uses?"); 413 return getIncomingBlock(unsigned(&U - op_begin())); 414 } 415 416 /// \brief Return incoming basic block corresponding 417 /// to value use iterator. 418 BasicBlock *getIncomingBlock(MemoryAccess::const_user_iterator I) const { 419 return getIncomingBlock(I.getUse()); 420 } 421 422 void setIncomingBlock(unsigned I, BasicBlock *BB) { 423 assert(BB && "PHI node got a null basic block!"); 424 block_begin()[I] = BB; 425 } 426 427 /// \brief Add an incoming value to the end of the PHI list 428 void addIncoming(MemoryAccess *V, BasicBlock *BB) { 429 if (getNumOperands() == ReservedSpace) 430 growOperands(); // Get more space! 431 // Initialize some new operands. 432 setNumHungOffUseOperands(getNumOperands() + 1); 433 setIncomingValue(getNumOperands() - 1, V); 434 setIncomingBlock(getNumOperands() - 1, BB); 435 } 436 437 /// \brief Return the first index of the specified basic 438 /// block in the value list for this PHI. Returns -1 if no instance. 439 int getBasicBlockIndex(const BasicBlock *BB) const { 440 for (unsigned I = 0, E = getNumOperands(); I != E; ++I) 441 if (block_begin()[I] == BB) 442 return I; 443 return -1; 444 } 445 446 Value *getIncomingValueForBlock(const BasicBlock *BB) const { 447 int Idx = getBasicBlockIndex(BB); 448 assert(Idx >= 0 && "Invalid basic block argument!"); 449 return getIncomingValue(Idx); 450 } 451 452 static inline bool classof(const MemoryPhi *) { return true; } 453 static inline bool classof(const Value *V) { 454 return V->getValueID() == MemoryPhiVal; 455 } 456 457 void print(raw_ostream &OS) const override; 458 459 protected: 460 friend class MemorySSA; 461 /// \brief this is more complicated than the generic 462 /// User::allocHungoffUses, because we have to allocate Uses for the incoming 463 /// values and pointers to the incoming blocks, all in one allocation. 464 void allocHungoffUses(unsigned N) { 465 User::allocHungoffUses(N, /* IsPhi */ true); 466 } 467 468 /// For debugging only. This gets used to give memory accesses pretty numbers 469 /// when printing them out 470 unsigned getID() const final { return ID; } 471 472 private: 473 // For debugging only 474 const unsigned ID; 475 unsigned ReservedSpace; 476 477 /// \brief This grows the operand list in response to a push_back style of 478 /// operation. This grows the number of ops by 1.5 times. 479 void growOperands() { 480 unsigned E = getNumOperands(); 481 // 2 op PHI nodes are VERY common, so reserve at least enough for that. 482 ReservedSpace = std::max(E + E / 2, 2u); 483 growHungoffUses(ReservedSpace, /* IsPhi */ true); 484 } 485 }; 486 487 template <> struct OperandTraits<MemoryPhi> : public HungoffOperandTraits<2> {}; 488 DEFINE_TRANSPARENT_OPERAND_ACCESSORS(MemoryPhi, MemoryAccess) 489 490 class MemorySSAWalker; 491 492 /// \brief Encapsulates MemorySSA, including all data associated with memory 493 /// accesses. 494 class MemorySSA { 495 public: 496 MemorySSA(Function &, AliasAnalysis *, DominatorTree *); 497 MemorySSA(MemorySSA &&); 498 ~MemorySSA(); 499 500 MemorySSAWalker *getWalker(); 501 502 /// \brief Given a memory Mod/Ref'ing instruction, get the MemorySSA 503 /// access associated with it. If passed a basic block gets the memory phi 504 /// node that exists for that block, if there is one. Otherwise, this will get 505 /// a MemoryUseOrDef. 506 MemoryAccess *getMemoryAccess(const Value *) const; 507 MemoryPhi *getMemoryAccess(const BasicBlock *BB) const; 508 509 void dump() const; 510 void print(raw_ostream &) const; 511 512 /// \brief Return true if \p MA represents the live on entry value 513 /// 514 /// Loads and stores from pointer arguments and other global values may be 515 /// defined by memory operations that do not occur in the current function, so 516 /// they may be live on entry to the function. MemorySSA represents such 517 /// memory state by the live on entry definition, which is guaranteed to occur 518 /// before any other memory access in the function. 519 inline bool isLiveOnEntryDef(const MemoryAccess *MA) const { 520 return MA == LiveOnEntryDef.get(); 521 } 522 523 inline MemoryAccess *getLiveOnEntryDef() const { 524 return LiveOnEntryDef.get(); 525 } 526 527 using AccessList = iplist<MemoryAccess>; 528 529 /// \brief Return the list of MemoryAccess's for a given basic block. 530 /// 531 /// This list is not modifiable by the user. 532 const AccessList *getBlockAccesses(const BasicBlock *BB) const { 533 auto It = PerBlockAccesses.find(BB); 534 return It == PerBlockAccesses.end() ? nullptr : It->second.get(); 535 } 536 537 /// \brief Create an empty MemoryPhi in MemorySSA 538 MemoryPhi *createMemoryPhi(BasicBlock *BB); 539 540 enum InsertionPlace { Beginning, End }; 541 542 /// \brief Create a MemoryAccess in MemorySSA at a specified point in a block, 543 /// with a specified clobbering definition. 544 /// 545 /// Returns the new MemoryAccess. 546 /// This should be called when a memory instruction is created that is being 547 /// used to replace an existing memory instruction. It will *not* create PHI 548 /// nodes, or verify the clobbering definition. The insertion place is used 549 /// solely to determine where in the memoryssa access lists the instruction 550 /// will be placed. The caller is expected to keep ordering the same as 551 /// instructions. 552 /// It will return the new MemoryAccess. 553 MemoryAccess *createMemoryAccessInBB(Instruction *I, MemoryAccess *Definition, 554 const BasicBlock *BB, 555 InsertionPlace Point); 556 /// \brief Create a MemoryAccess in MemorySSA before or after an existing 557 /// MemoryAccess. 558 /// 559 /// Returns the new MemoryAccess. 560 /// This should be called when a memory instruction is created that is being 561 /// used to replace an existing memory instruction. It will *not* create PHI 562 /// nodes, or verify the clobbering definition. The clobbering definition 563 /// must be non-null. 564 MemoryAccess *createMemoryAccessBefore(Instruction *I, 565 MemoryAccess *Definition, 566 MemoryAccess *InsertPt); 567 MemoryAccess *createMemoryAccessAfter(Instruction *I, 568 MemoryAccess *Definition, 569 MemoryAccess *InsertPt); 570 571 /// \brief Remove a MemoryAccess from MemorySSA, including updating all 572 /// definitions and uses. 573 /// This should be called when a memory instruction that has a MemoryAccess 574 /// associated with it is erased from the program. For example, if a store or 575 /// load is simply erased (not replaced), removeMemoryAccess should be called 576 /// on the MemoryAccess for that store/load. 577 void removeMemoryAccess(MemoryAccess *); 578 579 /// \brief Given two memory accesses in the same basic block, determine 580 /// whether MemoryAccess \p A dominates MemoryAccess \p B. 581 bool locallyDominates(const MemoryAccess *A, const MemoryAccess *B) const; 582 583 /// \brief Verify that MemorySSA is self consistent (IE definitions dominate 584 /// all uses, uses appear in the right places). This is used by unit tests. 585 void verifyMemorySSA() const; 586 587 protected: 588 // Used by Memory SSA annotater, dumpers, and wrapper pass 589 friend class MemorySSAAnnotatedWriter; 590 friend class MemorySSAPrinterLegacyPass; 591 void verifyDefUses(Function &F) const; 592 void verifyDomination(Function &F) const; 593 void verifyOrdering(Function &F) const; 594 595 private: 596 class CachingWalker; 597 void buildMemorySSA(); 598 void verifyUseInDefs(MemoryAccess *, MemoryAccess *) const; 599 using AccessMap = DenseMap<const BasicBlock *, std::unique_ptr<AccessList>>; 600 601 void 602 determineInsertionPoint(const SmallPtrSetImpl<BasicBlock *> &DefiningBlocks); 603 void computeDomLevels(DenseMap<DomTreeNode *, unsigned> &DomLevels); 604 void markUnreachableAsLiveOnEntry(BasicBlock *BB); 605 bool dominatesUse(const MemoryAccess *, const MemoryAccess *) const; 606 MemoryUseOrDef *createNewAccess(Instruction *); 607 MemoryUseOrDef *createDefinedAccess(Instruction *, MemoryAccess *); 608 MemoryAccess *findDominatingDef(BasicBlock *, enum InsertionPlace); 609 void removeFromLookups(MemoryAccess *); 610 611 MemoryAccess *renameBlock(BasicBlock *, MemoryAccess *); 612 void renamePass(DomTreeNode *, MemoryAccess *IncomingVal, 613 SmallPtrSet<BasicBlock *, 16> &Visited); 614 AccessList *getOrCreateAccessList(const BasicBlock *); 615 AliasAnalysis *AA; 616 DominatorTree *DT; 617 Function &F; 618 619 // Memory SSA mappings 620 DenseMap<const Value *, MemoryAccess *> ValueToMemoryAccess; 621 AccessMap PerBlockAccesses; 622 std::unique_ptr<MemoryAccess> LiveOnEntryDef; 623 624 // Memory SSA building info 625 std::unique_ptr<CachingWalker> Walker; 626 unsigned NextID; 627 }; 628 629 // This pass does eager building and then printing of MemorySSA. It is used by 630 // the tests to be able to build, dump, and verify Memory SSA. 631 class MemorySSAPrinterLegacyPass : public FunctionPass { 632 public: 633 MemorySSAPrinterLegacyPass(); 634 635 static char ID; 636 bool runOnFunction(Function &) override; 637 void getAnalysisUsage(AnalysisUsage &AU) const override; 638 }; 639 640 /// An analysis that produces \c MemorySSA for a function. 641 /// 642 class MemorySSAAnalysis : public AnalysisInfoMixin<MemorySSAAnalysis> { 643 friend AnalysisInfoMixin<MemorySSAAnalysis>; 644 static char PassID; 645 646 public: 647 typedef MemorySSA Result; 648 649 MemorySSA run(Function &F, AnalysisManager<Function> &AM); 650 }; 651 652 /// \brief Printer pass for \c MemorySSA. 653 class MemorySSAPrinterPass : public PassInfoMixin<MemorySSAPrinterPass> { 654 raw_ostream &OS; 655 656 public: 657 explicit MemorySSAPrinterPass(raw_ostream &OS) : OS(OS) {} 658 PreservedAnalyses run(Function &F, AnalysisManager<Function> &AM); 659 }; 660 661 /// \brief Verifier pass for \c MemorySSA. 662 struct MemorySSAVerifierPass : PassInfoMixin<MemorySSAVerifierPass> { 663 PreservedAnalyses run(Function &F, AnalysisManager<Function> &AM); 664 }; 665 666 /// \brief Legacy analysis pass which computes \c MemorySSA. 667 class MemorySSAWrapperPass : public FunctionPass { 668 public: 669 MemorySSAWrapperPass(); 670 671 static char ID; 672 bool runOnFunction(Function &) override; 673 void releaseMemory() override; 674 MemorySSA &getMSSA() { return *MSSA; } 675 const MemorySSA &getMSSA() const { return *MSSA; } 676 677 void getAnalysisUsage(AnalysisUsage &AU) const override; 678 679 void verifyAnalysis() const override; 680 void print(raw_ostream &OS, const Module *M = nullptr) const override; 681 682 private: 683 std::unique_ptr<MemorySSA> MSSA; 684 }; 685 686 /// \brief This is the generic walker interface for walkers of MemorySSA. 687 /// Walkers are used to be able to further disambiguate the def-use chains 688 /// MemorySSA gives you, or otherwise produce better info than MemorySSA gives 689 /// you. 690 /// In particular, while the def-use chains provide basic information, and are 691 /// guaranteed to give, for example, the nearest may-aliasing MemoryDef for a 692 /// MemoryUse as AliasAnalysis considers it, a user mant want better or other 693 /// information. In particular, they may want to use SCEV info to further 694 /// disambiguate memory accesses, or they may want the nearest dominating 695 /// may-aliasing MemoryDef for a call or a store. This API enables a 696 /// standardized interface to getting and using that info. 697 class MemorySSAWalker { 698 public: 699 MemorySSAWalker(MemorySSA *); 700 virtual ~MemorySSAWalker() {} 701 702 using MemoryAccessSet = SmallVector<MemoryAccess *, 8>; 703 704 /// \brief Given a memory Mod/Ref/ModRef'ing instruction, calling this 705 /// will give you the nearest dominating MemoryAccess that Mod's the location 706 /// the instruction accesses (by skipping any def which AA can prove does not 707 /// alias the location(s) accessed by the instruction given). 708 /// 709 /// Note that this will return a single access, and it must dominate the 710 /// Instruction, so if an operand of a MemoryPhi node Mod's the instruction, 711 /// this will return the MemoryPhi, not the operand. This means that 712 /// given: 713 /// if (a) { 714 /// 1 = MemoryDef(liveOnEntry) 715 /// store %a 716 /// } else { 717 /// 2 = MemoryDef(liveOnEntry) 718 /// store %b 719 /// } 720 /// 3 = MemoryPhi(2, 1) 721 /// MemoryUse(3) 722 /// load %a 723 /// 724 /// calling this API on load(%a) will return the MemoryPhi, not the MemoryDef 725 /// in the if (a) branch. 726 virtual MemoryAccess *getClobberingMemoryAccess(const Instruction *) = 0; 727 728 /// \brief Given a potentially clobbering memory access and a new location, 729 /// calling this will give you the nearest dominating clobbering MemoryAccess 730 /// (by skipping non-aliasing def links). 731 /// 732 /// This version of the function is mainly used to disambiguate phi translated 733 /// pointers, where the value of a pointer may have changed from the initial 734 /// memory access. Note that this expects to be handed either a MemoryUse, 735 /// or an already potentially clobbering access. Unlike the above API, if 736 /// given a MemoryDef that clobbers the pointer as the starting access, it 737 /// will return that MemoryDef, whereas the above would return the clobber 738 /// starting from the use side of the memory def. 739 virtual MemoryAccess *getClobberingMemoryAccess(MemoryAccess *, 740 MemoryLocation &) = 0; 741 742 /// \brief Given a memory access, invalidate anything this walker knows about 743 /// that access. 744 /// This API is used by walkers that store information to perform basic cache 745 /// invalidation. This will be called by MemorySSA at appropriate times for 746 /// the walker it uses or returns. 747 virtual void invalidateInfo(MemoryAccess *) {} 748 749 protected: 750 friend class MemorySSA; // For updating MSSA pointer in MemorySSA move 751 // constructor. 752 MemorySSA *MSSA; 753 }; 754 755 /// \brief A MemorySSAWalker that does no alias queries, or anything else. It 756 /// simply returns the links as they were constructed by the builder. 757 class DoNothingMemorySSAWalker final : public MemorySSAWalker { 758 public: 759 MemoryAccess *getClobberingMemoryAccess(const Instruction *) override; 760 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *, 761 MemoryLocation &) override; 762 }; 763 764 using MemoryAccessPair = std::pair<MemoryAccess *, MemoryLocation>; 765 using ConstMemoryAccessPair = std::pair<const MemoryAccess *, MemoryLocation>; 766 767 /// \brief Iterator base class used to implement const and non-const iterators 768 /// over the defining accesses of a MemoryAccess. 769 template <class T> 770 class memoryaccess_def_iterator_base 771 : public iterator_facade_base<memoryaccess_def_iterator_base<T>, 772 std::forward_iterator_tag, T, ptrdiff_t, T *, 773 T *> { 774 using BaseT = typename memoryaccess_def_iterator_base::iterator_facade_base; 775 776 public: 777 memoryaccess_def_iterator_base(T *Start) : Access(Start), ArgNo(0) {} 778 memoryaccess_def_iterator_base() : Access(nullptr), ArgNo(0) {} 779 bool operator==(const memoryaccess_def_iterator_base &Other) const { 780 return Access == Other.Access && (!Access || ArgNo == Other.ArgNo); 781 } 782 783 // This is a bit ugly, but for MemoryPHI's, unlike PHINodes, you can't get the 784 // block from the operand in constant time (In a PHINode, the uselist has 785 // both, so it's just subtraction). We provide it as part of the 786 // iterator to avoid callers having to linear walk to get the block. 787 // If the operation becomes constant time on MemoryPHI's, this bit of 788 // abstraction breaking should be removed. 789 BasicBlock *getPhiArgBlock() const { 790 MemoryPhi *MP = dyn_cast<MemoryPhi>(Access); 791 assert(MP && "Tried to get phi arg block when not iterating over a PHI"); 792 return MP->getIncomingBlock(ArgNo); 793 } 794 typename BaseT::iterator::pointer operator*() const { 795 assert(Access && "Tried to access past the end of our iterator"); 796 // Go to the first argument for phis, and the defining access for everything 797 // else. 798 if (MemoryPhi *MP = dyn_cast<MemoryPhi>(Access)) 799 return MP->getIncomingValue(ArgNo); 800 return cast<MemoryUseOrDef>(Access)->getDefiningAccess(); 801 } 802 using BaseT::operator++; 803 memoryaccess_def_iterator &operator++() { 804 assert(Access && "Hit end of iterator"); 805 if (MemoryPhi *MP = dyn_cast<MemoryPhi>(Access)) { 806 if (++ArgNo >= MP->getNumIncomingValues()) { 807 ArgNo = 0; 808 Access = nullptr; 809 } 810 } else { 811 Access = nullptr; 812 } 813 return *this; 814 } 815 816 private: 817 T *Access; 818 unsigned ArgNo; 819 }; 820 821 inline memoryaccess_def_iterator MemoryAccess::defs_begin() { 822 return memoryaccess_def_iterator(this); 823 } 824 825 inline const_memoryaccess_def_iterator MemoryAccess::defs_begin() const { 826 return const_memoryaccess_def_iterator(this); 827 } 828 829 inline memoryaccess_def_iterator MemoryAccess::defs_end() { 830 return memoryaccess_def_iterator(); 831 } 832 833 inline const_memoryaccess_def_iterator MemoryAccess::defs_end() const { 834 return const_memoryaccess_def_iterator(); 835 } 836 837 /// \brief GraphTraits for a MemoryAccess, which walks defs in the normal case, 838 /// and uses in the inverse case. 839 template <> struct GraphTraits<MemoryAccess *> { 840 using NodeType = MemoryAccess; 841 using ChildIteratorType = memoryaccess_def_iterator; 842 843 static NodeType *getEntryNode(NodeType *N) { return N; } 844 static inline ChildIteratorType child_begin(NodeType *N) { 845 return N->defs_begin(); 846 } 847 static inline ChildIteratorType child_end(NodeType *N) { 848 return N->defs_end(); 849 } 850 }; 851 852 template <> struct GraphTraits<Inverse<MemoryAccess *>> { 853 using NodeType = MemoryAccess; 854 using ChildIteratorType = MemoryAccess::iterator; 855 856 static NodeType *getEntryNode(NodeType *N) { return N; } 857 static inline ChildIteratorType child_begin(NodeType *N) { 858 return N->user_begin(); 859 } 860 static inline ChildIteratorType child_end(NodeType *N) { 861 return N->user_end(); 862 } 863 }; 864 865 /// \brief Provide an iterator that walks defs, giving both the memory access, 866 /// and the current pointer location, updating the pointer location as it 867 /// changes due to phi node translation. 868 /// 869 /// This iterator, while somewhat specialized, is what most clients actually 870 /// want when walking upwards through MemorySSA def chains. It takes a pair of 871 /// <MemoryAccess,MemoryLocation>, and walks defs, properly translating the 872 /// memory location through phi nodes for the user. 873 class upward_defs_iterator 874 : public iterator_facade_base<upward_defs_iterator, 875 std::forward_iterator_tag, 876 const MemoryAccessPair> { 877 using BaseT = upward_defs_iterator::iterator_facade_base; 878 879 public: 880 upward_defs_iterator(const MemoryAccessPair &Info) 881 : DefIterator(Info.first), Location(Info.second), 882 OriginalAccess(Info.first) { 883 CurrentPair.first = nullptr; 884 885 WalkingPhi = Info.first && isa<MemoryPhi>(Info.first); 886 fillInCurrentPair(); 887 } 888 889 upward_defs_iterator() 890 : DefIterator(), Location(), OriginalAccess(), WalkingPhi(false) { 891 CurrentPair.first = nullptr; 892 } 893 894 bool operator==(const upward_defs_iterator &Other) const { 895 return DefIterator == Other.DefIterator; 896 } 897 898 BaseT::iterator::reference operator*() const { 899 assert(DefIterator != OriginalAccess->defs_end() && 900 "Tried to access past the end of our iterator"); 901 return CurrentPair; 902 } 903 904 using BaseT::operator++; 905 upward_defs_iterator &operator++() { 906 assert(DefIterator != OriginalAccess->defs_end() && 907 "Tried to access past the end of the iterator"); 908 ++DefIterator; 909 if (DefIterator != OriginalAccess->defs_end()) 910 fillInCurrentPair(); 911 return *this; 912 } 913 914 BasicBlock *getPhiArgBlock() const { return DefIterator.getPhiArgBlock(); } 915 916 private: 917 void fillInCurrentPair() { 918 CurrentPair.first = *DefIterator; 919 if (WalkingPhi && Location.Ptr) { 920 PHITransAddr Translator( 921 const_cast<Value *>(Location.Ptr), 922 OriginalAccess->getBlock()->getModule()->getDataLayout(), nullptr); 923 if (!Translator.PHITranslateValue(OriginalAccess->getBlock(), 924 DefIterator.getPhiArgBlock(), nullptr, 925 false)) 926 if (Translator.getAddr() != Location.Ptr) { 927 CurrentPair.second = Location.getWithNewPtr(Translator.getAddr()); 928 return; 929 } 930 } 931 CurrentPair.second = Location; 932 } 933 934 MemoryAccessPair CurrentPair; 935 memoryaccess_def_iterator DefIterator; 936 MemoryLocation Location; 937 MemoryAccess *OriginalAccess; 938 bool WalkingPhi; 939 }; 940 941 inline upward_defs_iterator upward_defs_begin(const MemoryAccessPair &Pair) { 942 return upward_defs_iterator(Pair); 943 } 944 945 inline upward_defs_iterator upward_defs_end() { return upward_defs_iterator(); } 946 947 } // end namespace llvm 948 949 #endif // LLVM_TRANSFORMS_UTILS_MEMORYSSA_H 950