1 //===- CFLAliasAnalysis.cpp - CFL-Based Alias Analysis Implementation ------==// 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 file implements a CFL-based context-insensitive alias analysis 11 // algorithm. It does not depend on types. The algorithm is a mixture of the one 12 // described in "Demand-driven alias analysis for C" by Xin Zheng and Radu 13 // Rugina, and "Fast algorithms for Dyck-CFL-reachability with applications to 14 // Alias Analysis" by Zhang Q, Lyu M R, Yuan H, and Su Z. -- to summarize the 15 // papers, we build a graph of the uses of a variable, where each node is a 16 // memory location, and each edge is an action that happened on that memory 17 // location. The "actions" can be one of Dereference, Reference, Assign, or 18 // Assign. 19 // 20 // Two variables are considered as aliasing iff you can reach one value's node 21 // from the other value's node and the language formed by concatenating all of 22 // the edge labels (actions) conforms to a context-free grammar. 23 // 24 // Because this algorithm requires a graph search on each query, we execute the 25 // algorithm outlined in "Fast algorithms..." (mentioned above) 26 // in order to transform the graph into sets of variables that may alias in 27 // ~nlogn time (n = number of variables.), which makes queries take constant 28 // time. 29 //===----------------------------------------------------------------------===// 30 31 #include "StratifiedSets.h" 32 #include "llvm/ADT/BitVector.h" 33 #include "llvm/ADT/DenseMap.h" 34 #include "llvm/ADT/None.h" 35 #include "llvm/ADT/Optional.h" 36 #include "llvm/Analysis/AliasAnalysis.h" 37 #include "llvm/Analysis/Passes.h" 38 #include "llvm/IR/Constants.h" 39 #include "llvm/IR/Function.h" 40 #include "llvm/IR/InstVisitor.h" 41 #include "llvm/IR/Instructions.h" 42 #include "llvm/IR/ValueHandle.h" 43 #include "llvm/Pass.h" 44 #include "llvm/Support/Allocator.h" 45 #include "llvm/Support/Compiler.h" 46 #include "llvm/Support/Debug.h" 47 #include "llvm/Support/ErrorHandling.h" 48 #include "llvm/Support/raw_ostream.h" 49 #include <algorithm> 50 #include <cassert> 51 #include <forward_list> 52 #include <memory> 53 #include <tuple> 54 55 using namespace llvm; 56 57 #define DEBUG_TYPE "cfl-aa" 58 59 // Try to go from a Value* to a Function*. Never returns nullptr. 60 static Optional<Function *> parentFunctionOfValue(Value *); 61 62 // Returns possible functions called by the Inst* into the given 63 // SmallVectorImpl. Returns true if targets found, false otherwise. 64 // This is templated because InvokeInst/CallInst give us the same 65 // set of functions that we care about, and I don't like repeating 66 // myself. 67 template <typename Inst> 68 static bool getPossibleTargets(Inst *, SmallVectorImpl<Function *> &); 69 70 // Some instructions need to have their users tracked. Instructions like 71 // `add` require you to get the users of the Instruction* itself, other 72 // instructions like `store` require you to get the users of the first 73 // operand. This function gets the "proper" value to track for each 74 // type of instruction we support. 75 static Optional<Value *> getTargetValue(Instruction *); 76 77 // There are certain instructions (i.e. FenceInst, etc.) that we ignore. 78 // This notes that we should ignore those. 79 static bool hasUsefulEdges(Instruction *); 80 81 const StratifiedIndex StratifiedLink::SetSentinel = 82 std::numeric_limits<StratifiedIndex>::max(); 83 84 namespace { 85 // StratifiedInfo Attribute things. 86 typedef unsigned StratifiedAttr; 87 LLVM_CONSTEXPR unsigned MaxStratifiedAttrIndex = NumStratifiedAttrs; 88 LLVM_CONSTEXPR unsigned AttrAllIndex = 0; 89 LLVM_CONSTEXPR unsigned AttrGlobalIndex = 1; 90 LLVM_CONSTEXPR unsigned AttrUnknownIndex = 2; 91 LLVM_CONSTEXPR unsigned AttrFirstArgIndex = 3; 92 LLVM_CONSTEXPR unsigned AttrLastArgIndex = MaxStratifiedAttrIndex; 93 LLVM_CONSTEXPR unsigned AttrMaxNumArgs = AttrLastArgIndex - AttrFirstArgIndex; 94 95 LLVM_CONSTEXPR StratifiedAttr AttrNone = 0; 96 LLVM_CONSTEXPR StratifiedAttr AttrUnknown = 1 << AttrUnknownIndex; 97 LLVM_CONSTEXPR StratifiedAttr AttrAll = ~AttrNone; 98 99 // \brief StratifiedSets call for knowledge of "direction", so this is how we 100 // represent that locally. 101 enum class Level { Same, Above, Below }; 102 103 // \brief Edges can be one of four "weights" -- each weight must have an inverse 104 // weight (Assign has Assign; Reference has Dereference). 105 enum class EdgeType { 106 // The weight assigned when assigning from or to a value. For example, in: 107 // %b = getelementptr %a, 0 108 // ...The relationships are %b assign %a, and %a assign %b. This used to be 109 // two edges, but having a distinction bought us nothing. 110 Assign, 111 112 // The edge used when we have an edge going from some handle to a Value. 113 // Examples of this include: 114 // %b = load %a (%b Dereference %a) 115 // %b = extractelement %a, 0 (%a Dereference %b) 116 Dereference, 117 118 // The edge used when our edge goes from a value to a handle that may have 119 // contained it at some point. Examples: 120 // %b = load %a (%a Reference %b) 121 // %b = extractelement %a, 0 (%b Reference %a) 122 Reference 123 }; 124 125 // \brief Encodes the notion of a "use" 126 struct Edge { 127 // \brief Which value the edge is coming from 128 Value *From; 129 130 // \brief Which value the edge is pointing to 131 Value *To; 132 133 // \brief Edge weight 134 EdgeType Weight; 135 136 // \brief Whether we aliased any external values along the way that may be 137 // invisible to the analysis (i.e. landingpad for exceptions, calls for 138 // interprocedural analysis, etc.) 139 StratifiedAttrs AdditionalAttrs; 140 141 Edge(Value *From, Value *To, EdgeType W, StratifiedAttrs A) 142 : From(From), To(To), Weight(W), AdditionalAttrs(A) {} 143 }; 144 145 // \brief Information we have about a function and would like to keep around 146 struct FunctionInfo { 147 StratifiedSets<Value *> Sets; 148 // Lots of functions have < 4 returns. Adjust as necessary. 149 SmallVector<Value *, 4> ReturnedValues; 150 151 FunctionInfo(StratifiedSets<Value *> &&S, SmallVector<Value *, 4> &&RV) 152 : Sets(std::move(S)), ReturnedValues(std::move(RV)) {} 153 }; 154 155 struct CFLAliasAnalysis; 156 157 struct FunctionHandle : public CallbackVH { 158 FunctionHandle(Function *Fn, CFLAliasAnalysis *CFLAA) 159 : CallbackVH(Fn), CFLAA(CFLAA) { 160 assert(Fn != nullptr); 161 assert(CFLAA != nullptr); 162 } 163 164 ~FunctionHandle() override {} 165 166 void deleted() override { removeSelfFromCache(); } 167 void allUsesReplacedWith(Value *) override { removeSelfFromCache(); } 168 169 private: 170 CFLAliasAnalysis *CFLAA; 171 172 void removeSelfFromCache(); 173 }; 174 175 struct CFLAliasAnalysis : public ImmutablePass, public AliasAnalysis { 176 private: 177 /// \brief Cached mapping of Functions to their StratifiedSets. 178 /// If a function's sets are currently being built, it is marked 179 /// in the cache as an Optional without a value. This way, if we 180 /// have any kind of recursion, it is discernable from a function 181 /// that simply has empty sets. 182 DenseMap<Function *, Optional<FunctionInfo>> Cache; 183 std::forward_list<FunctionHandle> Handles; 184 185 public: 186 static char ID; 187 188 CFLAliasAnalysis() : ImmutablePass(ID) { 189 initializeCFLAliasAnalysisPass(*PassRegistry::getPassRegistry()); 190 } 191 192 ~CFLAliasAnalysis() override {} 193 194 void getAnalysisUsage(AnalysisUsage &AU) const override { 195 AliasAnalysis::getAnalysisUsage(AU); 196 } 197 198 void *getAdjustedAnalysisPointer(const void *ID) override { 199 if (ID == &AliasAnalysis::ID) 200 return (AliasAnalysis *)this; 201 return this; 202 } 203 204 /// \brief Inserts the given Function into the cache. 205 void scan(Function *Fn); 206 207 void evict(Function *Fn) { Cache.erase(Fn); } 208 209 /// \brief Ensures that the given function is available in the cache. 210 /// Returns the appropriate entry from the cache. 211 const Optional<FunctionInfo> &ensureCached(Function *Fn) { 212 auto Iter = Cache.find(Fn); 213 if (Iter == Cache.end()) { 214 scan(Fn); 215 Iter = Cache.find(Fn); 216 assert(Iter != Cache.end()); 217 assert(Iter->second.hasValue()); 218 } 219 return Iter->second; 220 } 221 222 AliasResult query(const Location &LocA, const Location &LocB); 223 224 AliasResult alias(const Location &LocA, const Location &LocB) override { 225 if (LocA.Ptr == LocB.Ptr) { 226 if (LocA.Size == LocB.Size) { 227 return MustAlias; 228 } else { 229 return PartialAlias; 230 } 231 } 232 233 // Comparisons between global variables and other constants should be 234 // handled by BasicAA. 235 // TODO: ConstantExpr handling -- CFLAA may report NoAlias when comparing 236 // a GlobalValue and ConstantExpr, but every query needs to have at least 237 // one Value tied to a Function, and neither GlobalValues nor ConstantExprs 238 // are. 239 if (isa<Constant>(LocA.Ptr) && isa<Constant>(LocB.Ptr)) { 240 return AliasAnalysis::alias(LocA, LocB); 241 } 242 243 AliasResult QueryResult = query(LocA, LocB); 244 if (QueryResult == MayAlias) 245 return AliasAnalysis::alias(LocA, LocB); 246 247 return QueryResult; 248 } 249 250 bool doInitialization(Module &M) override; 251 }; 252 253 void FunctionHandle::removeSelfFromCache() { 254 assert(CFLAA != nullptr); 255 auto *Val = getValPtr(); 256 CFLAA->evict(cast<Function>(Val)); 257 setValPtr(nullptr); 258 } 259 260 // \brief Gets the edges our graph should have, based on an Instruction* 261 class GetEdgesVisitor : public InstVisitor<GetEdgesVisitor, void> { 262 CFLAliasAnalysis &AA; 263 SmallVectorImpl<Edge> &Output; 264 265 public: 266 GetEdgesVisitor(CFLAliasAnalysis &AA, SmallVectorImpl<Edge> &Output) 267 : AA(AA), Output(Output) {} 268 269 void visitInstruction(Instruction &) { 270 llvm_unreachable("Unsupported instruction encountered"); 271 } 272 273 void visitPtrToIntInst(PtrToIntInst &Inst) { 274 auto *Ptr = Inst.getOperand(0); 275 Output.push_back(Edge(Ptr, Ptr, EdgeType::Assign, AttrUnknown)); 276 } 277 278 void visitIntToPtrInst(IntToPtrInst &Inst) { 279 auto *Ptr = &Inst; 280 Output.push_back(Edge(Ptr, Ptr, EdgeType::Assign, AttrUnknown)); 281 } 282 283 void visitCastInst(CastInst &Inst) { 284 Output.push_back( 285 Edge(&Inst, Inst.getOperand(0), EdgeType::Assign, AttrNone)); 286 } 287 288 void visitBinaryOperator(BinaryOperator &Inst) { 289 auto *Op1 = Inst.getOperand(0); 290 auto *Op2 = Inst.getOperand(1); 291 Output.push_back(Edge(&Inst, Op1, EdgeType::Assign, AttrNone)); 292 Output.push_back(Edge(&Inst, Op2, EdgeType::Assign, AttrNone)); 293 } 294 295 void visitAtomicCmpXchgInst(AtomicCmpXchgInst &Inst) { 296 auto *Ptr = Inst.getPointerOperand(); 297 auto *Val = Inst.getNewValOperand(); 298 Output.push_back(Edge(Ptr, Val, EdgeType::Dereference, AttrNone)); 299 } 300 301 void visitAtomicRMWInst(AtomicRMWInst &Inst) { 302 auto *Ptr = Inst.getPointerOperand(); 303 auto *Val = Inst.getValOperand(); 304 Output.push_back(Edge(Ptr, Val, EdgeType::Dereference, AttrNone)); 305 } 306 307 void visitPHINode(PHINode &Inst) { 308 for (unsigned I = 0, E = Inst.getNumIncomingValues(); I != E; ++I) { 309 Value *Val = Inst.getIncomingValue(I); 310 Output.push_back(Edge(&Inst, Val, EdgeType::Assign, AttrNone)); 311 } 312 } 313 314 void visitGetElementPtrInst(GetElementPtrInst &Inst) { 315 auto *Op = Inst.getPointerOperand(); 316 Output.push_back(Edge(&Inst, Op, EdgeType::Assign, AttrNone)); 317 for (auto I = Inst.idx_begin(), E = Inst.idx_end(); I != E; ++I) 318 Output.push_back(Edge(&Inst, *I, EdgeType::Assign, AttrNone)); 319 } 320 321 void visitSelectInst(SelectInst &Inst) { 322 // Condition is not processed here (The actual statement producing 323 // the condition result is processed elsewhere). For select, the 324 // condition is evaluated, but not loaded, stored, or assigned 325 // simply as a result of being the condition of a select. 326 327 auto *TrueVal = Inst.getTrueValue(); 328 Output.push_back(Edge(&Inst, TrueVal, EdgeType::Assign, AttrNone)); 329 auto *FalseVal = Inst.getFalseValue(); 330 Output.push_back(Edge(&Inst, FalseVal, EdgeType::Assign, AttrNone)); 331 } 332 333 void visitAllocaInst(AllocaInst &) {} 334 335 void visitLoadInst(LoadInst &Inst) { 336 auto *Ptr = Inst.getPointerOperand(); 337 auto *Val = &Inst; 338 Output.push_back(Edge(Val, Ptr, EdgeType::Reference, AttrNone)); 339 } 340 341 void visitStoreInst(StoreInst &Inst) { 342 auto *Ptr = Inst.getPointerOperand(); 343 auto *Val = Inst.getValueOperand(); 344 Output.push_back(Edge(Ptr, Val, EdgeType::Dereference, AttrNone)); 345 } 346 347 void visitVAArgInst(VAArgInst &Inst) { 348 // We can't fully model va_arg here. For *Ptr = Inst.getOperand(0), it does 349 // two things: 350 // 1. Loads a value from *((T*)*Ptr). 351 // 2. Increments (stores to) *Ptr by some target-specific amount. 352 // For now, we'll handle this like a landingpad instruction (by placing the 353 // result in its own group, and having that group alias externals). 354 auto *Val = &Inst; 355 Output.push_back(Edge(Val, Val, EdgeType::Assign, AttrAll)); 356 } 357 358 static bool isFunctionExternal(Function *Fn) { 359 return Fn->isDeclaration() || !Fn->hasLocalLinkage(); 360 } 361 362 // Gets whether the sets at Index1 above, below, or equal to the sets at 363 // Index2. Returns None if they are not in the same set chain. 364 static Optional<Level> getIndexRelation(const StratifiedSets<Value *> &Sets, 365 StratifiedIndex Index1, 366 StratifiedIndex Index2) { 367 if (Index1 == Index2) 368 return Level::Same; 369 370 const auto *Current = &Sets.getLink(Index1); 371 while (Current->hasBelow()) { 372 if (Current->Below == Index2) 373 return Level::Below; 374 Current = &Sets.getLink(Current->Below); 375 } 376 377 Current = &Sets.getLink(Index1); 378 while (Current->hasAbove()) { 379 if (Current->Above == Index2) 380 return Level::Above; 381 Current = &Sets.getLink(Current->Above); 382 } 383 384 return NoneType(); 385 } 386 387 bool 388 tryInterproceduralAnalysis(const SmallVectorImpl<Function *> &Fns, 389 Value *FuncValue, 390 const iterator_range<User::op_iterator> &Args) { 391 const unsigned ExpectedMaxArgs = 8; 392 const unsigned MaxSupportedArgs = 50; 393 assert(Fns.size() > 0); 394 395 // I put this here to give us an upper bound on time taken by IPA. Is it 396 // really (realistically) needed? Keep in mind that we do have an n^2 algo. 397 if (std::distance(Args.begin(), Args.end()) > (int)MaxSupportedArgs) 398 return false; 399 400 // Exit early if we'll fail anyway 401 for (auto *Fn : Fns) { 402 if (isFunctionExternal(Fn) || Fn->isVarArg()) 403 return false; 404 auto &MaybeInfo = AA.ensureCached(Fn); 405 if (!MaybeInfo.hasValue()) 406 return false; 407 } 408 409 SmallVector<Value *, ExpectedMaxArgs> Arguments(Args.begin(), Args.end()); 410 SmallVector<StratifiedInfo, ExpectedMaxArgs> Parameters; 411 for (auto *Fn : Fns) { 412 auto &Info = *AA.ensureCached(Fn); 413 auto &Sets = Info.Sets; 414 auto &RetVals = Info.ReturnedValues; 415 416 Parameters.clear(); 417 for (auto &Param : Fn->args()) { 418 auto MaybeInfo = Sets.find(&Param); 419 // Did a new parameter somehow get added to the function/slip by? 420 if (!MaybeInfo.hasValue()) 421 return false; 422 Parameters.push_back(*MaybeInfo); 423 } 424 425 // Adding an edge from argument -> return value for each parameter that 426 // may alias the return value 427 for (unsigned I = 0, E = Parameters.size(); I != E; ++I) { 428 auto &ParamInfo = Parameters[I]; 429 auto &ArgVal = Arguments[I]; 430 bool AddEdge = false; 431 StratifiedAttrs Externals; 432 for (unsigned X = 0, XE = RetVals.size(); X != XE; ++X) { 433 auto MaybeInfo = Sets.find(RetVals[X]); 434 if (!MaybeInfo.hasValue()) 435 return false; 436 437 auto &RetInfo = *MaybeInfo; 438 auto RetAttrs = Sets.getLink(RetInfo.Index).Attrs; 439 auto ParamAttrs = Sets.getLink(ParamInfo.Index).Attrs; 440 auto MaybeRelation = 441 getIndexRelation(Sets, ParamInfo.Index, RetInfo.Index); 442 if (MaybeRelation.hasValue()) { 443 AddEdge = true; 444 Externals |= RetAttrs | ParamAttrs; 445 } 446 } 447 if (AddEdge) 448 Output.push_back(Edge(FuncValue, ArgVal, EdgeType::Assign, 449 StratifiedAttrs().flip())); 450 } 451 452 if (Parameters.size() != Arguments.size()) 453 return false; 454 455 // Adding edges between arguments for arguments that may end up aliasing 456 // each other. This is necessary for functions such as 457 // void foo(int** a, int** b) { *a = *b; } 458 // (Technically, the proper sets for this would be those below 459 // Arguments[I] and Arguments[X], but our algorithm will produce 460 // extremely similar, and equally correct, results either way) 461 for (unsigned I = 0, E = Arguments.size(); I != E; ++I) { 462 auto &MainVal = Arguments[I]; 463 auto &MainInfo = Parameters[I]; 464 auto &MainAttrs = Sets.getLink(MainInfo.Index).Attrs; 465 for (unsigned X = I + 1; X != E; ++X) { 466 auto &SubInfo = Parameters[X]; 467 auto &SubVal = Arguments[X]; 468 auto &SubAttrs = Sets.getLink(SubInfo.Index).Attrs; 469 auto MaybeRelation = 470 getIndexRelation(Sets, MainInfo.Index, SubInfo.Index); 471 472 if (!MaybeRelation.hasValue()) 473 continue; 474 475 auto NewAttrs = SubAttrs | MainAttrs; 476 Output.push_back(Edge(MainVal, SubVal, EdgeType::Assign, NewAttrs)); 477 } 478 } 479 } 480 return true; 481 } 482 483 template <typename InstT> void visitCallLikeInst(InstT &Inst) { 484 SmallVector<Function *, 4> Targets; 485 if (getPossibleTargets(&Inst, Targets)) { 486 if (tryInterproceduralAnalysis(Targets, &Inst, Inst.arg_operands())) 487 return; 488 // Cleanup from interprocedural analysis 489 Output.clear(); 490 } 491 492 for (Value *V : Inst.arg_operands()) 493 Output.push_back(Edge(&Inst, V, EdgeType::Assign, AttrAll)); 494 } 495 496 void visitCallInst(CallInst &Inst) { visitCallLikeInst(Inst); } 497 498 void visitInvokeInst(InvokeInst &Inst) { visitCallLikeInst(Inst); } 499 500 // Because vectors/aggregates are immutable and unaddressable, 501 // there's nothing we can do to coax a value out of them, other 502 // than calling Extract{Element,Value}. We can effectively treat 503 // them as pointers to arbitrary memory locations we can store in 504 // and load from. 505 void visitExtractElementInst(ExtractElementInst &Inst) { 506 auto *Ptr = Inst.getVectorOperand(); 507 auto *Val = &Inst; 508 Output.push_back(Edge(Val, Ptr, EdgeType::Reference, AttrNone)); 509 } 510 511 void visitInsertElementInst(InsertElementInst &Inst) { 512 auto *Vec = Inst.getOperand(0); 513 auto *Val = Inst.getOperand(1); 514 Output.push_back(Edge(&Inst, Vec, EdgeType::Assign, AttrNone)); 515 Output.push_back(Edge(&Inst, Val, EdgeType::Dereference, AttrNone)); 516 } 517 518 void visitLandingPadInst(LandingPadInst &Inst) { 519 // Exceptions come from "nowhere", from our analysis' perspective. 520 // So we place the instruction its own group, noting that said group may 521 // alias externals 522 Output.push_back(Edge(&Inst, &Inst, EdgeType::Assign, AttrAll)); 523 } 524 525 void visitInsertValueInst(InsertValueInst &Inst) { 526 auto *Agg = Inst.getOperand(0); 527 auto *Val = Inst.getOperand(1); 528 Output.push_back(Edge(&Inst, Agg, EdgeType::Assign, AttrNone)); 529 Output.push_back(Edge(&Inst, Val, EdgeType::Dereference, AttrNone)); 530 } 531 532 void visitExtractValueInst(ExtractValueInst &Inst) { 533 auto *Ptr = Inst.getAggregateOperand(); 534 Output.push_back(Edge(&Inst, Ptr, EdgeType::Reference, AttrNone)); 535 } 536 537 void visitShuffleVectorInst(ShuffleVectorInst &Inst) { 538 auto *From1 = Inst.getOperand(0); 539 auto *From2 = Inst.getOperand(1); 540 Output.push_back(Edge(&Inst, From1, EdgeType::Assign, AttrNone)); 541 Output.push_back(Edge(&Inst, From2, EdgeType::Assign, AttrNone)); 542 } 543 }; 544 545 // For a given instruction, we need to know which Value* to get the 546 // users of in order to build our graph. In some cases (i.e. add), 547 // we simply need the Instruction*. In other cases (i.e. store), 548 // finding the users of the Instruction* is useless; we need to find 549 // the users of the first operand. This handles determining which 550 // value to follow for us. 551 // 552 // Note: we *need* to keep this in sync with GetEdgesVisitor. Add 553 // something to GetEdgesVisitor, add it here -- remove something from 554 // GetEdgesVisitor, remove it here. 555 class GetTargetValueVisitor 556 : public InstVisitor<GetTargetValueVisitor, Value *> { 557 public: 558 Value *visitInstruction(Instruction &Inst) { return &Inst; } 559 560 Value *visitStoreInst(StoreInst &Inst) { return Inst.getPointerOperand(); } 561 562 Value *visitAtomicCmpXchgInst(AtomicCmpXchgInst &Inst) { 563 return Inst.getPointerOperand(); 564 } 565 566 Value *visitAtomicRMWInst(AtomicRMWInst &Inst) { 567 return Inst.getPointerOperand(); 568 } 569 570 Value *visitInsertElementInst(InsertElementInst &Inst) { 571 return Inst.getOperand(0); 572 } 573 574 Value *visitInsertValueInst(InsertValueInst &Inst) { 575 return Inst.getAggregateOperand(); 576 } 577 }; 578 579 // Set building requires a weighted bidirectional graph. 580 template <typename EdgeTypeT> class WeightedBidirectionalGraph { 581 public: 582 typedef std::size_t Node; 583 584 private: 585 const static Node StartNode = Node(0); 586 587 struct Edge { 588 EdgeTypeT Weight; 589 Node Other; 590 591 Edge(const EdgeTypeT &W, const Node &N) : Weight(W), Other(N) {} 592 593 bool operator==(const Edge &E) const { 594 return Weight == E.Weight && Other == E.Other; 595 } 596 597 bool operator!=(const Edge &E) const { return !operator==(E); } 598 }; 599 600 struct NodeImpl { 601 std::vector<Edge> Edges; 602 }; 603 604 std::vector<NodeImpl> NodeImpls; 605 606 bool inbounds(Node NodeIndex) const { return NodeIndex < NodeImpls.size(); } 607 608 const NodeImpl &getNode(Node N) const { return NodeImpls[N]; } 609 NodeImpl &getNode(Node N) { return NodeImpls[N]; } 610 611 public: 612 // ----- Various Edge iterators for the graph ----- // 613 614 // \brief Iterator for edges. Because this graph is bidirected, we don't 615 // allow modificaiton of the edges using this iterator. Additionally, the 616 // iterator becomes invalid if you add edges to or from the node you're 617 // getting the edges of. 618 struct EdgeIterator : public std::iterator<std::forward_iterator_tag, 619 std::tuple<EdgeTypeT, Node *>> { 620 EdgeIterator(const typename std::vector<Edge>::const_iterator &Iter) 621 : Current(Iter) {} 622 623 EdgeIterator(NodeImpl &Impl) : Current(Impl.begin()) {} 624 625 EdgeIterator &operator++() { 626 ++Current; 627 return *this; 628 } 629 630 EdgeIterator operator++(int) { 631 EdgeIterator Copy(Current); 632 operator++(); 633 return Copy; 634 } 635 636 std::tuple<EdgeTypeT, Node> &operator*() { 637 Store = std::make_tuple(Current->Weight, Current->Other); 638 return Store; 639 } 640 641 bool operator==(const EdgeIterator &Other) const { 642 return Current == Other.Current; 643 } 644 645 bool operator!=(const EdgeIterator &Other) const { 646 return !operator==(Other); 647 } 648 649 private: 650 typename std::vector<Edge>::const_iterator Current; 651 std::tuple<EdgeTypeT, Node> Store; 652 }; 653 654 // Wrapper for EdgeIterator with begin()/end() calls. 655 struct EdgeIterable { 656 EdgeIterable(const std::vector<Edge> &Edges) 657 : BeginIter(Edges.begin()), EndIter(Edges.end()) {} 658 659 EdgeIterator begin() { return EdgeIterator(BeginIter); } 660 661 EdgeIterator end() { return EdgeIterator(EndIter); } 662 663 private: 664 typename std::vector<Edge>::const_iterator BeginIter; 665 typename std::vector<Edge>::const_iterator EndIter; 666 }; 667 668 // ----- Actual graph-related things ----- // 669 670 WeightedBidirectionalGraph() {} 671 672 WeightedBidirectionalGraph(WeightedBidirectionalGraph<EdgeTypeT> &&Other) 673 : NodeImpls(std::move(Other.NodeImpls)) {} 674 675 WeightedBidirectionalGraph<EdgeTypeT> & 676 operator=(WeightedBidirectionalGraph<EdgeTypeT> &&Other) { 677 NodeImpls = std::move(Other.NodeImpls); 678 return *this; 679 } 680 681 Node addNode() { 682 auto Index = NodeImpls.size(); 683 auto NewNode = Node(Index); 684 NodeImpls.push_back(NodeImpl()); 685 return NewNode; 686 } 687 688 void addEdge(Node From, Node To, const EdgeTypeT &Weight, 689 const EdgeTypeT &ReverseWeight) { 690 assert(inbounds(From)); 691 assert(inbounds(To)); 692 auto &FromNode = getNode(From); 693 auto &ToNode = getNode(To); 694 FromNode.Edges.push_back(Edge(Weight, To)); 695 ToNode.Edges.push_back(Edge(ReverseWeight, From)); 696 } 697 698 EdgeIterable edgesFor(const Node &N) const { 699 const auto &Node = getNode(N); 700 return EdgeIterable(Node.Edges); 701 } 702 703 bool empty() const { return NodeImpls.empty(); } 704 std::size_t size() const { return NodeImpls.size(); } 705 706 // \brief Gets an arbitrary node in the graph as a starting point for 707 // traversal. 708 Node getEntryNode() { 709 assert(inbounds(StartNode)); 710 return StartNode; 711 } 712 }; 713 714 typedef WeightedBidirectionalGraph<std::pair<EdgeType, StratifiedAttrs>> GraphT; 715 typedef DenseMap<Value *, GraphT::Node> NodeMapT; 716 } 717 718 // -- Setting up/registering CFLAA pass -- // 719 char CFLAliasAnalysis::ID = 0; 720 721 INITIALIZE_AG_PASS(CFLAliasAnalysis, AliasAnalysis, "cfl-aa", 722 "CFL-Based AA implementation", false, true, false) 723 724 ImmutablePass *llvm::createCFLAliasAnalysisPass() { 725 return new CFLAliasAnalysis(); 726 } 727 728 //===----------------------------------------------------------------------===// 729 // Function declarations that require types defined in the namespace above 730 //===----------------------------------------------------------------------===// 731 732 // Given an argument number, returns the appropriate Attr index to set. 733 static StratifiedAttr argNumberToAttrIndex(StratifiedAttr); 734 735 // Given a Value, potentially return which AttrIndex it maps to. 736 static Optional<StratifiedAttr> valueToAttrIndex(Value *Val); 737 738 // Gets the inverse of a given EdgeType. 739 static EdgeType flipWeight(EdgeType); 740 741 // Gets edges of the given Instruction*, writing them to the SmallVector*. 742 static void argsToEdges(CFLAliasAnalysis &, Instruction *, 743 SmallVectorImpl<Edge> &); 744 745 // Gets the "Level" that one should travel in StratifiedSets 746 // given an EdgeType. 747 static Level directionOfEdgeType(EdgeType); 748 749 // Builds the graph needed for constructing the StratifiedSets for the 750 // given function 751 static void buildGraphFrom(CFLAliasAnalysis &, Function *, 752 SmallVectorImpl<Value *> &, NodeMapT &, GraphT &); 753 754 // Gets the edges of a ConstantExpr as if it was an Instruction. This 755 // function also acts on any nested ConstantExprs, adding the edges 756 // of those to the given SmallVector as well. 757 static void constexprToEdges(CFLAliasAnalysis &, ConstantExpr &, 758 SmallVectorImpl<Edge> &); 759 760 // Given an Instruction, this will add it to the graph, along with any 761 // Instructions that are potentially only available from said Instruction 762 // For example, given the following line: 763 // %0 = load i16* getelementptr ([1 x i16]* @a, 0, 0), align 2 764 // addInstructionToGraph would add both the `load` and `getelementptr` 765 // instructions to the graph appropriately. 766 static void addInstructionToGraph(CFLAliasAnalysis &, Instruction &, 767 SmallVectorImpl<Value *> &, NodeMapT &, 768 GraphT &); 769 770 // Notes whether it would be pointless to add the given Value to our sets. 771 static bool canSkipAddingToSets(Value *Val); 772 773 // Builds the graph + StratifiedSets for a function. 774 static FunctionInfo buildSetsFrom(CFLAliasAnalysis &, Function *); 775 776 static Optional<Function *> parentFunctionOfValue(Value *Val) { 777 if (auto *Inst = dyn_cast<Instruction>(Val)) { 778 auto *Bb = Inst->getParent(); 779 return Bb->getParent(); 780 } 781 782 if (auto *Arg = dyn_cast<Argument>(Val)) 783 return Arg->getParent(); 784 return NoneType(); 785 } 786 787 template <typename Inst> 788 static bool getPossibleTargets(Inst *Call, 789 SmallVectorImpl<Function *> &Output) { 790 if (auto *Fn = Call->getCalledFunction()) { 791 Output.push_back(Fn); 792 return true; 793 } 794 795 // TODO: If the call is indirect, we might be able to enumerate all potential 796 // targets of the call and return them, rather than just failing. 797 return false; 798 } 799 800 static Optional<Value *> getTargetValue(Instruction *Inst) { 801 GetTargetValueVisitor V; 802 return V.visit(Inst); 803 } 804 805 static bool hasUsefulEdges(Instruction *Inst) { 806 bool IsNonInvokeTerminator = 807 isa<TerminatorInst>(Inst) && !isa<InvokeInst>(Inst); 808 return !isa<CmpInst>(Inst) && !isa<FenceInst>(Inst) && !IsNonInvokeTerminator; 809 } 810 811 static Optional<StratifiedAttr> valueToAttrIndex(Value *Val) { 812 if (isa<GlobalValue>(Val)) 813 return AttrGlobalIndex; 814 815 if (auto *Arg = dyn_cast<Argument>(Val)) 816 // Only pointer arguments should have the argument attribute, 817 // because things can't escape through scalars without us seeing a 818 // cast, and thus, interaction with them doesn't matter. 819 if (!Arg->hasNoAliasAttr() && Arg->getType()->isPointerTy()) 820 return argNumberToAttrIndex(Arg->getArgNo()); 821 return NoneType(); 822 } 823 824 static StratifiedAttr argNumberToAttrIndex(unsigned ArgNum) { 825 if (ArgNum >= AttrMaxNumArgs) 826 return AttrAllIndex; 827 return ArgNum + AttrFirstArgIndex; 828 } 829 830 static EdgeType flipWeight(EdgeType Initial) { 831 switch (Initial) { 832 case EdgeType::Assign: 833 return EdgeType::Assign; 834 case EdgeType::Dereference: 835 return EdgeType::Reference; 836 case EdgeType::Reference: 837 return EdgeType::Dereference; 838 } 839 llvm_unreachable("Incomplete coverage of EdgeType enum"); 840 } 841 842 static void argsToEdges(CFLAliasAnalysis &Analysis, Instruction *Inst, 843 SmallVectorImpl<Edge> &Output) { 844 assert(hasUsefulEdges(Inst) && 845 "Expected instructions to have 'useful' edges"); 846 GetEdgesVisitor v(Analysis, Output); 847 v.visit(Inst); 848 } 849 850 static Level directionOfEdgeType(EdgeType Weight) { 851 switch (Weight) { 852 case EdgeType::Reference: 853 return Level::Above; 854 case EdgeType::Dereference: 855 return Level::Below; 856 case EdgeType::Assign: 857 return Level::Same; 858 } 859 llvm_unreachable("Incomplete switch coverage"); 860 } 861 862 static void constexprToEdges(CFLAliasAnalysis &Analysis, 863 ConstantExpr &CExprToCollapse, 864 SmallVectorImpl<Edge> &Results) { 865 SmallVector<ConstantExpr *, 4> Worklist; 866 Worklist.push_back(&CExprToCollapse); 867 868 SmallVector<Edge, 8> ConstexprEdges; 869 while (!Worklist.empty()) { 870 auto *CExpr = Worklist.pop_back_val(); 871 std::unique_ptr<Instruction> Inst(CExpr->getAsInstruction()); 872 873 if (!hasUsefulEdges(Inst.get())) 874 continue; 875 876 ConstexprEdges.clear(); 877 argsToEdges(Analysis, Inst.get(), ConstexprEdges); 878 for (auto &Edge : ConstexprEdges) { 879 if (Edge.From == Inst.get()) 880 Edge.From = CExpr; 881 else if (auto *Nested = dyn_cast<ConstantExpr>(Edge.From)) 882 Worklist.push_back(Nested); 883 884 if (Edge.To == Inst.get()) 885 Edge.To = CExpr; 886 else if (auto *Nested = dyn_cast<ConstantExpr>(Edge.To)) 887 Worklist.push_back(Nested); 888 } 889 890 Results.append(ConstexprEdges.begin(), ConstexprEdges.end()); 891 } 892 } 893 894 static void addInstructionToGraph(CFLAliasAnalysis &Analysis, Instruction &Inst, 895 SmallVectorImpl<Value *> &ReturnedValues, 896 NodeMapT &Map, GraphT &Graph) { 897 const auto findOrInsertNode = [&Map, &Graph](Value *Val) { 898 auto Pair = Map.insert(std::make_pair(Val, GraphT::Node())); 899 auto &Iter = Pair.first; 900 if (Pair.second) { 901 auto NewNode = Graph.addNode(); 902 Iter->second = NewNode; 903 } 904 return Iter->second; 905 }; 906 907 // We don't want the edges of most "return" instructions, but we *do* want 908 // to know what can be returned. 909 if (isa<ReturnInst>(&Inst)) 910 ReturnedValues.push_back(&Inst); 911 912 if (!hasUsefulEdges(&Inst)) 913 return; 914 915 SmallVector<Edge, 8> Edges; 916 argsToEdges(Analysis, &Inst, Edges); 917 918 // In the case of an unused alloca (or similar), edges may be empty. Note 919 // that it exists so we can potentially answer NoAlias. 920 if (Edges.empty()) { 921 auto MaybeVal = getTargetValue(&Inst); 922 assert(MaybeVal.hasValue()); 923 auto *Target = *MaybeVal; 924 findOrInsertNode(Target); 925 return; 926 } 927 928 const auto addEdgeToGraph = [&Graph, &findOrInsertNode](const Edge &E) { 929 auto To = findOrInsertNode(E.To); 930 auto From = findOrInsertNode(E.From); 931 auto FlippedWeight = flipWeight(E.Weight); 932 auto Attrs = E.AdditionalAttrs; 933 Graph.addEdge(From, To, std::make_pair(E.Weight, Attrs), 934 std::make_pair(FlippedWeight, Attrs)); 935 }; 936 937 SmallVector<ConstantExpr *, 4> ConstantExprs; 938 for (const Edge &E : Edges) { 939 addEdgeToGraph(E); 940 if (auto *Constexpr = dyn_cast<ConstantExpr>(E.To)) 941 ConstantExprs.push_back(Constexpr); 942 if (auto *Constexpr = dyn_cast<ConstantExpr>(E.From)) 943 ConstantExprs.push_back(Constexpr); 944 } 945 946 for (ConstantExpr *CE : ConstantExprs) { 947 Edges.clear(); 948 constexprToEdges(Analysis, *CE, Edges); 949 std::for_each(Edges.begin(), Edges.end(), addEdgeToGraph); 950 } 951 } 952 953 // Aside: We may remove graph construction entirely, because it doesn't really 954 // buy us much that we don't already have. I'd like to add interprocedural 955 // analysis prior to this however, in case that somehow requires the graph 956 // produced by this for efficient execution 957 static void buildGraphFrom(CFLAliasAnalysis &Analysis, Function *Fn, 958 SmallVectorImpl<Value *> &ReturnedValues, 959 NodeMapT &Map, GraphT &Graph) { 960 for (auto &Bb : Fn->getBasicBlockList()) 961 for (auto &Inst : Bb.getInstList()) 962 addInstructionToGraph(Analysis, Inst, ReturnedValues, Map, Graph); 963 } 964 965 static bool canSkipAddingToSets(Value *Val) { 966 // Constants can share instances, which may falsely unify multiple 967 // sets, e.g. in 968 // store i32* null, i32** %ptr1 969 // store i32* null, i32** %ptr2 970 // clearly ptr1 and ptr2 should not be unified into the same set, so 971 // we should filter out the (potentially shared) instance to 972 // i32* null. 973 if (isa<Constant>(Val)) { 974 bool Container = isa<ConstantVector>(Val) || isa<ConstantArray>(Val) || 975 isa<ConstantStruct>(Val); 976 // TODO: Because all of these things are constant, we can determine whether 977 // the data is *actually* mutable at graph building time. This will probably 978 // come for free/cheap with offset awareness. 979 bool CanStoreMutableData = 980 isa<GlobalValue>(Val) || isa<ConstantExpr>(Val) || Container; 981 return !CanStoreMutableData; 982 } 983 984 return false; 985 } 986 987 static FunctionInfo buildSetsFrom(CFLAliasAnalysis &Analysis, Function *Fn) { 988 NodeMapT Map; 989 GraphT Graph; 990 SmallVector<Value *, 4> ReturnedValues; 991 992 buildGraphFrom(Analysis, Fn, ReturnedValues, Map, Graph); 993 994 DenseMap<GraphT::Node, Value *> NodeValueMap; 995 NodeValueMap.resize(Map.size()); 996 for (const auto &Pair : Map) 997 NodeValueMap.insert(std::make_pair(Pair.second, Pair.first)); 998 999 const auto findValueOrDie = [&NodeValueMap](GraphT::Node Node) { 1000 auto ValIter = NodeValueMap.find(Node); 1001 assert(ValIter != NodeValueMap.end()); 1002 return ValIter->second; 1003 }; 1004 1005 StratifiedSetsBuilder<Value *> Builder; 1006 1007 SmallVector<GraphT::Node, 16> Worklist; 1008 for (auto &Pair : Map) { 1009 Worklist.clear(); 1010 1011 auto *Value = Pair.first; 1012 Builder.add(Value); 1013 auto InitialNode = Pair.second; 1014 Worklist.push_back(InitialNode); 1015 while (!Worklist.empty()) { 1016 auto Node = Worklist.pop_back_val(); 1017 auto *CurValue = findValueOrDie(Node); 1018 if (canSkipAddingToSets(CurValue)) 1019 continue; 1020 1021 for (const auto &EdgeTuple : Graph.edgesFor(Node)) { 1022 auto Weight = std::get<0>(EdgeTuple); 1023 auto Label = Weight.first; 1024 auto &OtherNode = std::get<1>(EdgeTuple); 1025 auto *OtherValue = findValueOrDie(OtherNode); 1026 1027 if (canSkipAddingToSets(OtherValue)) 1028 continue; 1029 1030 bool Added; 1031 switch (directionOfEdgeType(Label)) { 1032 case Level::Above: 1033 Added = Builder.addAbove(CurValue, OtherValue); 1034 break; 1035 case Level::Below: 1036 Added = Builder.addBelow(CurValue, OtherValue); 1037 break; 1038 case Level::Same: 1039 Added = Builder.addWith(CurValue, OtherValue); 1040 break; 1041 } 1042 1043 auto Aliasing = Weight.second; 1044 if (auto MaybeCurIndex = valueToAttrIndex(CurValue)) 1045 Aliasing.set(*MaybeCurIndex); 1046 if (auto MaybeOtherIndex = valueToAttrIndex(OtherValue)) 1047 Aliasing.set(*MaybeOtherIndex); 1048 Builder.noteAttributes(CurValue, Aliasing); 1049 Builder.noteAttributes(OtherValue, Aliasing); 1050 1051 if (Added) 1052 Worklist.push_back(OtherNode); 1053 } 1054 } 1055 } 1056 1057 // There are times when we end up with parameters not in our graph (i.e. if 1058 // it's only used as the condition of a branch). Other bits of code depend on 1059 // things that were present during construction being present in the graph. 1060 // So, we add all present arguments here. 1061 for (auto &Arg : Fn->args()) { 1062 if (!Builder.add(&Arg)) 1063 continue; 1064 1065 auto Attrs = valueToAttrIndex(&Arg); 1066 if (Attrs.hasValue()) 1067 Builder.noteAttributes(&Arg, *Attrs); 1068 } 1069 1070 return FunctionInfo(Builder.build(), std::move(ReturnedValues)); 1071 } 1072 1073 void CFLAliasAnalysis::scan(Function *Fn) { 1074 auto InsertPair = Cache.insert(std::make_pair(Fn, Optional<FunctionInfo>())); 1075 (void)InsertPair; 1076 assert(InsertPair.second && 1077 "Trying to scan a function that has already been cached"); 1078 1079 FunctionInfo Info(buildSetsFrom(*this, Fn)); 1080 Cache[Fn] = std::move(Info); 1081 Handles.push_front(FunctionHandle(Fn, this)); 1082 } 1083 1084 AliasAnalysis::AliasResult 1085 CFLAliasAnalysis::query(const AliasAnalysis::Location &LocA, 1086 const AliasAnalysis::Location &LocB) { 1087 auto *ValA = const_cast<Value *>(LocA.Ptr); 1088 auto *ValB = const_cast<Value *>(LocB.Ptr); 1089 1090 Function *Fn = nullptr; 1091 auto MaybeFnA = parentFunctionOfValue(ValA); 1092 auto MaybeFnB = parentFunctionOfValue(ValB); 1093 if (!MaybeFnA.hasValue() && !MaybeFnB.hasValue()) { 1094 // The only times this is known to happen are when globals + InlineAsm 1095 // are involved 1096 DEBUG(dbgs() << "CFLAA: could not extract parent function information.\n"); 1097 return AliasAnalysis::MayAlias; 1098 } 1099 1100 if (MaybeFnA.hasValue()) { 1101 Fn = *MaybeFnA; 1102 assert((!MaybeFnB.hasValue() || *MaybeFnB == *MaybeFnA) && 1103 "Interprocedural queries not supported"); 1104 } else { 1105 Fn = *MaybeFnB; 1106 } 1107 1108 assert(Fn != nullptr); 1109 auto &MaybeInfo = ensureCached(Fn); 1110 assert(MaybeInfo.hasValue()); 1111 1112 auto &Sets = MaybeInfo->Sets; 1113 auto MaybeA = Sets.find(ValA); 1114 if (!MaybeA.hasValue()) 1115 return AliasAnalysis::MayAlias; 1116 1117 auto MaybeB = Sets.find(ValB); 1118 if (!MaybeB.hasValue()) 1119 return AliasAnalysis::MayAlias; 1120 1121 auto SetA = *MaybeA; 1122 auto SetB = *MaybeB; 1123 auto AttrsA = Sets.getLink(SetA.Index).Attrs; 1124 auto AttrsB = Sets.getLink(SetB.Index).Attrs; 1125 1126 // Stratified set attributes are used as markets to signify whether a member 1127 // of a StratifiedSet (or a member of a set above the current set) has 1128 // interacted with either arguments or globals. "Interacted with" meaning 1129 // its value may be different depending on the value of an argument or 1130 // global. The thought behind this is that, because arguments and globals 1131 // may alias each other, if AttrsA and AttrsB have touched args/globals, 1132 // we must conservatively say that they alias. However, if at least one of 1133 // the sets has no values that could legally be altered by changing the value 1134 // of an argument or global, then we don't have to be as conservative. 1135 if (AttrsA.any() && AttrsB.any()) 1136 return AliasAnalysis::MayAlias; 1137 1138 // We currently unify things even if the accesses to them may not be in 1139 // bounds, so we can't return partial alias here because we don't 1140 // know whether the pointer is really within the object or not. 1141 // IE Given an out of bounds GEP and an alloca'd pointer, we may 1142 // unify the two. We can't return partial alias for this case. 1143 // Since we do not currently track enough information to 1144 // differentiate 1145 1146 if (SetA.Index == SetB.Index) 1147 return AliasAnalysis::MayAlias; 1148 1149 return AliasAnalysis::NoAlias; 1150 } 1151 1152 bool CFLAliasAnalysis::doInitialization(Module &M) { 1153 InitializeAliasAnalysis(this, &M.getDataLayout()); 1154 return true; 1155 } 1156