1 //== RegionStore.cpp - Field-sensitive store model --------------*- 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 // This file defines a basic region store model. In this model, we do have field 11 // sensitivity. But we assume nothing about the heap shape. So recursive data 12 // structures are largely ignored. Basically we do 1-limiting analysis. 13 // Parameter pointers are assumed with no aliasing. Pointee objects of 14 // parameters are created lazily. 15 // 16 //===----------------------------------------------------------------------===// 17 #include "clang/AST/Attr.h" 18 #include "clang/AST/CharUnits.h" 19 #include "clang/Analysis/Analyses/LiveVariables.h" 20 #include "clang/Analysis/AnalysisContext.h" 21 #include "clang/Basic/TargetInfo.h" 22 #include "clang/StaticAnalyzer/Core/PathSensitive/AnalysisManager.h" 23 #include "clang/StaticAnalyzer/Core/PathSensitive/CallEvent.h" 24 #include "clang/StaticAnalyzer/Core/PathSensitive/MemRegion.h" 25 #include "clang/StaticAnalyzer/Core/PathSensitive/ProgramState.h" 26 #include "clang/StaticAnalyzer/Core/PathSensitive/ProgramStateTrait.h" 27 #include "clang/StaticAnalyzer/Core/PathSensitive/SubEngine.h" 28 #include "llvm/ADT/ImmutableList.h" 29 #include "llvm/ADT/ImmutableMap.h" 30 #include "llvm/ADT/Optional.h" 31 #include "llvm/Support/raw_ostream.h" 32 33 using namespace clang; 34 using namespace ento; 35 36 //===----------------------------------------------------------------------===// 37 // Representation of binding keys. 38 //===----------------------------------------------------------------------===// 39 40 namespace { 41 class BindingKey { 42 public: 43 enum Kind { Default = 0x0, Direct = 0x1 }; 44 private: 45 enum { Symbolic = 0x2 }; 46 47 llvm::PointerIntPair<const MemRegion *, 2> P; 48 uint64_t Data; 49 50 /// Create a key for a binding to region \p r, which has a symbolic offset 51 /// from region \p Base. 52 explicit BindingKey(const SubRegion *r, const SubRegion *Base, Kind k) 53 : P(r, k | Symbolic), Data(reinterpret_cast<uintptr_t>(Base)) { 54 assert(r && Base && "Must have known regions."); 55 assert(getConcreteOffsetRegion() == Base && "Failed to store base region"); 56 } 57 58 /// Create a key for a binding at \p offset from base region \p r. 59 explicit BindingKey(const MemRegion *r, uint64_t offset, Kind k) 60 : P(r, k), Data(offset) { 61 assert(r && "Must have known regions."); 62 assert(getOffset() == offset && "Failed to store offset"); 63 assert((r == r->getBaseRegion() || isa<ObjCIvarRegion>(r)) && "Not a base"); 64 } 65 public: 66 67 bool isDirect() const { return P.getInt() & Direct; } 68 bool hasSymbolicOffset() const { return P.getInt() & Symbolic; } 69 70 const MemRegion *getRegion() const { return P.getPointer(); } 71 uint64_t getOffset() const { 72 assert(!hasSymbolicOffset()); 73 return Data; 74 } 75 76 const SubRegion *getConcreteOffsetRegion() const { 77 assert(hasSymbolicOffset()); 78 return reinterpret_cast<const SubRegion *>(static_cast<uintptr_t>(Data)); 79 } 80 81 const MemRegion *getBaseRegion() const { 82 if (hasSymbolicOffset()) 83 return getConcreteOffsetRegion()->getBaseRegion(); 84 return getRegion()->getBaseRegion(); 85 } 86 87 void Profile(llvm::FoldingSetNodeID& ID) const { 88 ID.AddPointer(P.getOpaqueValue()); 89 ID.AddInteger(Data); 90 } 91 92 static BindingKey Make(const MemRegion *R, Kind k); 93 94 bool operator<(const BindingKey &X) const { 95 if (P.getOpaqueValue() < X.P.getOpaqueValue()) 96 return true; 97 if (P.getOpaqueValue() > X.P.getOpaqueValue()) 98 return false; 99 return Data < X.Data; 100 } 101 102 bool operator==(const BindingKey &X) const { 103 return P.getOpaqueValue() == X.P.getOpaqueValue() && 104 Data == X.Data; 105 } 106 107 void dump() const; 108 }; 109 } // end anonymous namespace 110 111 BindingKey BindingKey::Make(const MemRegion *R, Kind k) { 112 const RegionOffset &RO = R->getAsOffset(); 113 if (RO.hasSymbolicOffset()) 114 return BindingKey(cast<SubRegion>(R), cast<SubRegion>(RO.getRegion()), k); 115 116 return BindingKey(RO.getRegion(), RO.getOffset(), k); 117 } 118 119 namespace llvm { 120 static inline 121 raw_ostream &operator<<(raw_ostream &os, BindingKey K) { 122 os << '(' << K.getRegion(); 123 if (!K.hasSymbolicOffset()) 124 os << ',' << K.getOffset(); 125 os << ',' << (K.isDirect() ? "direct" : "default") 126 << ')'; 127 return os; 128 } 129 130 template <typename T> struct isPodLike; 131 template <> struct isPodLike<BindingKey> { 132 static const bool value = true; 133 }; 134 } // end llvm namespace 135 136 LLVM_DUMP_METHOD void BindingKey::dump() const { llvm::errs() << *this; } 137 138 //===----------------------------------------------------------------------===// 139 // Actual Store type. 140 //===----------------------------------------------------------------------===// 141 142 typedef llvm::ImmutableMap<BindingKey, SVal> ClusterBindings; 143 typedef llvm::ImmutableMapRef<BindingKey, SVal> ClusterBindingsRef; 144 typedef std::pair<BindingKey, SVal> BindingPair; 145 146 typedef llvm::ImmutableMap<const MemRegion *, ClusterBindings> 147 RegionBindings; 148 149 namespace { 150 class RegionBindingsRef : public llvm::ImmutableMapRef<const MemRegion *, 151 ClusterBindings> { 152 ClusterBindings::Factory &CBFactory; 153 public: 154 typedef llvm::ImmutableMapRef<const MemRegion *, ClusterBindings> 155 ParentTy; 156 157 RegionBindingsRef(ClusterBindings::Factory &CBFactory, 158 const RegionBindings::TreeTy *T, 159 RegionBindings::TreeTy::Factory *F) 160 : llvm::ImmutableMapRef<const MemRegion *, ClusterBindings>(T, F), 161 CBFactory(CBFactory) {} 162 163 RegionBindingsRef(const ParentTy &P, ClusterBindings::Factory &CBFactory) 164 : llvm::ImmutableMapRef<const MemRegion *, ClusterBindings>(P), 165 CBFactory(CBFactory) {} 166 167 RegionBindingsRef add(key_type_ref K, data_type_ref D) const { 168 return RegionBindingsRef(static_cast<const ParentTy*>(this)->add(K, D), 169 CBFactory); 170 } 171 172 RegionBindingsRef remove(key_type_ref K) const { 173 return RegionBindingsRef(static_cast<const ParentTy*>(this)->remove(K), 174 CBFactory); 175 } 176 177 RegionBindingsRef addBinding(BindingKey K, SVal V) const; 178 179 RegionBindingsRef addBinding(const MemRegion *R, 180 BindingKey::Kind k, SVal V) const; 181 182 RegionBindingsRef &operator=(const RegionBindingsRef &X) { 183 *static_cast<ParentTy*>(this) = X; 184 return *this; 185 } 186 187 const SVal *lookup(BindingKey K) const; 188 const SVal *lookup(const MemRegion *R, BindingKey::Kind k) const; 189 const ClusterBindings *lookup(const MemRegion *R) const { 190 return static_cast<const ParentTy*>(this)->lookup(R); 191 } 192 193 RegionBindingsRef removeBinding(BindingKey K); 194 195 RegionBindingsRef removeBinding(const MemRegion *R, 196 BindingKey::Kind k); 197 198 RegionBindingsRef removeBinding(const MemRegion *R) { 199 return removeBinding(R, BindingKey::Direct). 200 removeBinding(R, BindingKey::Default); 201 } 202 203 Optional<SVal> getDirectBinding(const MemRegion *R) const; 204 205 /// getDefaultBinding - Returns an SVal* representing an optional default 206 /// binding associated with a region and its subregions. 207 Optional<SVal> getDefaultBinding(const MemRegion *R) const; 208 209 /// Return the internal tree as a Store. 210 Store asStore() const { 211 return asImmutableMap().getRootWithoutRetain(); 212 } 213 214 void dump(raw_ostream &OS, const char *nl) const { 215 for (iterator I = begin(), E = end(); I != E; ++I) { 216 const ClusterBindings &Cluster = I.getData(); 217 for (ClusterBindings::iterator CI = Cluster.begin(), CE = Cluster.end(); 218 CI != CE; ++CI) { 219 OS << ' ' << CI.getKey() << " : " << CI.getData() << nl; 220 } 221 OS << nl; 222 } 223 } 224 225 LLVM_DUMP_METHOD void dump() const { dump(llvm::errs(), "\n"); } 226 }; 227 } // end anonymous namespace 228 229 typedef const RegionBindingsRef& RegionBindingsConstRef; 230 231 Optional<SVal> RegionBindingsRef::getDirectBinding(const MemRegion *R) const { 232 return Optional<SVal>::create(lookup(R, BindingKey::Direct)); 233 } 234 235 Optional<SVal> RegionBindingsRef::getDefaultBinding(const MemRegion *R) const { 236 if (R->isBoundable()) 237 if (const TypedValueRegion *TR = dyn_cast<TypedValueRegion>(R)) 238 if (TR->getValueType()->isUnionType()) 239 return UnknownVal(); 240 241 return Optional<SVal>::create(lookup(R, BindingKey::Default)); 242 } 243 244 RegionBindingsRef RegionBindingsRef::addBinding(BindingKey K, SVal V) const { 245 const MemRegion *Base = K.getBaseRegion(); 246 247 const ClusterBindings *ExistingCluster = lookup(Base); 248 ClusterBindings Cluster = (ExistingCluster ? *ExistingCluster 249 : CBFactory.getEmptyMap()); 250 251 ClusterBindings NewCluster = CBFactory.add(Cluster, K, V); 252 return add(Base, NewCluster); 253 } 254 255 256 RegionBindingsRef RegionBindingsRef::addBinding(const MemRegion *R, 257 BindingKey::Kind k, 258 SVal V) const { 259 return addBinding(BindingKey::Make(R, k), V); 260 } 261 262 const SVal *RegionBindingsRef::lookup(BindingKey K) const { 263 const ClusterBindings *Cluster = lookup(K.getBaseRegion()); 264 if (!Cluster) 265 return nullptr; 266 return Cluster->lookup(K); 267 } 268 269 const SVal *RegionBindingsRef::lookup(const MemRegion *R, 270 BindingKey::Kind k) const { 271 return lookup(BindingKey::Make(R, k)); 272 } 273 274 RegionBindingsRef RegionBindingsRef::removeBinding(BindingKey K) { 275 const MemRegion *Base = K.getBaseRegion(); 276 const ClusterBindings *Cluster = lookup(Base); 277 if (!Cluster) 278 return *this; 279 280 ClusterBindings NewCluster = CBFactory.remove(*Cluster, K); 281 if (NewCluster.isEmpty()) 282 return remove(Base); 283 return add(Base, NewCluster); 284 } 285 286 RegionBindingsRef RegionBindingsRef::removeBinding(const MemRegion *R, 287 BindingKey::Kind k){ 288 return removeBinding(BindingKey::Make(R, k)); 289 } 290 291 //===----------------------------------------------------------------------===// 292 // Fine-grained control of RegionStoreManager. 293 //===----------------------------------------------------------------------===// 294 295 namespace { 296 struct minimal_features_tag {}; 297 struct maximal_features_tag {}; 298 299 class RegionStoreFeatures { 300 bool SupportsFields; 301 public: 302 RegionStoreFeatures(minimal_features_tag) : 303 SupportsFields(false) {} 304 305 RegionStoreFeatures(maximal_features_tag) : 306 SupportsFields(true) {} 307 308 void enableFields(bool t) { SupportsFields = t; } 309 310 bool supportsFields() const { return SupportsFields; } 311 }; 312 } 313 314 //===----------------------------------------------------------------------===// 315 // Main RegionStore logic. 316 //===----------------------------------------------------------------------===// 317 318 namespace { 319 class invalidateRegionsWorker; 320 321 class RegionStoreManager : public StoreManager { 322 public: 323 const RegionStoreFeatures Features; 324 325 RegionBindings::Factory RBFactory; 326 mutable ClusterBindings::Factory CBFactory; 327 328 typedef std::vector<SVal> SValListTy; 329 private: 330 typedef llvm::DenseMap<const LazyCompoundValData *, 331 SValListTy> LazyBindingsMapTy; 332 LazyBindingsMapTy LazyBindingsMap; 333 334 /// The largest number of fields a struct can have and still be 335 /// considered "small". 336 /// 337 /// This is currently used to decide whether or not it is worth "forcing" a 338 /// LazyCompoundVal on bind. 339 /// 340 /// This is controlled by 'region-store-small-struct-limit' option. 341 /// To disable all small-struct-dependent behavior, set the option to "0". 342 unsigned SmallStructLimit; 343 344 /// \brief A helper used to populate the work list with the given set of 345 /// regions. 346 void populateWorkList(invalidateRegionsWorker &W, 347 ArrayRef<SVal> Values, 348 InvalidatedRegions *TopLevelRegions); 349 350 public: 351 RegionStoreManager(ProgramStateManager& mgr, const RegionStoreFeatures &f) 352 : StoreManager(mgr), Features(f), 353 RBFactory(mgr.getAllocator()), CBFactory(mgr.getAllocator()), 354 SmallStructLimit(0) { 355 if (SubEngine *Eng = StateMgr.getOwningEngine()) { 356 AnalyzerOptions &Options = Eng->getAnalysisManager().options; 357 SmallStructLimit = 358 Options.getOptionAsInteger("region-store-small-struct-limit", 2); 359 } 360 } 361 362 363 /// setImplicitDefaultValue - Set the default binding for the provided 364 /// MemRegion to the value implicitly defined for compound literals when 365 /// the value is not specified. 366 RegionBindingsRef setImplicitDefaultValue(RegionBindingsConstRef B, 367 const MemRegion *R, QualType T); 368 369 /// ArrayToPointer - Emulates the "decay" of an array to a pointer 370 /// type. 'Array' represents the lvalue of the array being decayed 371 /// to a pointer, and the returned SVal represents the decayed 372 /// version of that lvalue (i.e., a pointer to the first element of 373 /// the array). This is called by ExprEngine when evaluating 374 /// casts from arrays to pointers. 375 SVal ArrayToPointer(Loc Array, QualType ElementTy) override; 376 377 StoreRef getInitialStore(const LocationContext *InitLoc) override { 378 return StoreRef(RBFactory.getEmptyMap().getRootWithoutRetain(), *this); 379 } 380 381 //===-------------------------------------------------------------------===// 382 // Binding values to regions. 383 //===-------------------------------------------------------------------===// 384 RegionBindingsRef invalidateGlobalRegion(MemRegion::Kind K, 385 const Expr *Ex, 386 unsigned Count, 387 const LocationContext *LCtx, 388 RegionBindingsRef B, 389 InvalidatedRegions *Invalidated); 390 391 StoreRef invalidateRegions(Store store, 392 ArrayRef<SVal> Values, 393 const Expr *E, unsigned Count, 394 const LocationContext *LCtx, 395 const CallEvent *Call, 396 InvalidatedSymbols &IS, 397 RegionAndSymbolInvalidationTraits &ITraits, 398 InvalidatedRegions *Invalidated, 399 InvalidatedRegions *InvalidatedTopLevel) override; 400 401 bool scanReachableSymbols(Store S, const MemRegion *R, 402 ScanReachableSymbols &Callbacks) override; 403 404 RegionBindingsRef removeSubRegionBindings(RegionBindingsConstRef B, 405 const SubRegion *R); 406 407 public: // Part of public interface to class. 408 409 StoreRef Bind(Store store, Loc LV, SVal V) override { 410 return StoreRef(bind(getRegionBindings(store), LV, V).asStore(), *this); 411 } 412 413 RegionBindingsRef bind(RegionBindingsConstRef B, Loc LV, SVal V); 414 415 // BindDefault is only used to initialize a region with a default value. 416 StoreRef BindDefault(Store store, const MemRegion *R, SVal V) override { 417 RegionBindingsRef B = getRegionBindings(store); 418 assert(!B.lookup(R, BindingKey::Direct)); 419 420 BindingKey Key = BindingKey::Make(R, BindingKey::Default); 421 if (B.lookup(Key)) { 422 const SubRegion *SR = cast<SubRegion>(R); 423 assert(SR->getAsOffset().getOffset() == 424 SR->getSuperRegion()->getAsOffset().getOffset() && 425 "A default value must come from a super-region"); 426 B = removeSubRegionBindings(B, SR); 427 } else { 428 B = B.addBinding(Key, V); 429 } 430 431 return StoreRef(B.asImmutableMap().getRootWithoutRetain(), *this); 432 } 433 434 /// Attempt to extract the fields of \p LCV and bind them to the struct region 435 /// \p R. 436 /// 437 /// This path is used when it seems advantageous to "force" loading the values 438 /// within a LazyCompoundVal to bind memberwise to the struct region, rather 439 /// than using a Default binding at the base of the entire region. This is a 440 /// heuristic attempting to avoid building long chains of LazyCompoundVals. 441 /// 442 /// \returns The updated store bindings, or \c None if binding non-lazily 443 /// would be too expensive. 444 Optional<RegionBindingsRef> tryBindSmallStruct(RegionBindingsConstRef B, 445 const TypedValueRegion *R, 446 const RecordDecl *RD, 447 nonloc::LazyCompoundVal LCV); 448 449 /// BindStruct - Bind a compound value to a structure. 450 RegionBindingsRef bindStruct(RegionBindingsConstRef B, 451 const TypedValueRegion* R, SVal V); 452 453 /// BindVector - Bind a compound value to a vector. 454 RegionBindingsRef bindVector(RegionBindingsConstRef B, 455 const TypedValueRegion* R, SVal V); 456 457 RegionBindingsRef bindArray(RegionBindingsConstRef B, 458 const TypedValueRegion* R, 459 SVal V); 460 461 /// Clears out all bindings in the given region and assigns a new value 462 /// as a Default binding. 463 RegionBindingsRef bindAggregate(RegionBindingsConstRef B, 464 const TypedRegion *R, 465 SVal DefaultVal); 466 467 /// \brief Create a new store with the specified binding removed. 468 /// \param ST the original store, that is the basis for the new store. 469 /// \param L the location whose binding should be removed. 470 StoreRef killBinding(Store ST, Loc L) override; 471 472 void incrementReferenceCount(Store store) override { 473 getRegionBindings(store).manualRetain(); 474 } 475 476 /// If the StoreManager supports it, decrement the reference count of 477 /// the specified Store object. If the reference count hits 0, the memory 478 /// associated with the object is recycled. 479 void decrementReferenceCount(Store store) override { 480 getRegionBindings(store).manualRelease(); 481 } 482 483 bool includedInBindings(Store store, const MemRegion *region) const override; 484 485 /// \brief Return the value bound to specified location in a given state. 486 /// 487 /// The high level logic for this method is this: 488 /// getBinding (L) 489 /// if L has binding 490 /// return L's binding 491 /// else if L is in killset 492 /// return unknown 493 /// else 494 /// if L is on stack or heap 495 /// return undefined 496 /// else 497 /// return symbolic 498 SVal getBinding(Store S, Loc L, QualType T) override { 499 return getBinding(getRegionBindings(S), L, T); 500 } 501 502 SVal getBinding(RegionBindingsConstRef B, Loc L, QualType T = QualType()); 503 504 SVal getBindingForElement(RegionBindingsConstRef B, const ElementRegion *R); 505 506 SVal getBindingForField(RegionBindingsConstRef B, const FieldRegion *R); 507 508 SVal getBindingForObjCIvar(RegionBindingsConstRef B, const ObjCIvarRegion *R); 509 510 SVal getBindingForVar(RegionBindingsConstRef B, const VarRegion *R); 511 512 SVal getBindingForLazySymbol(const TypedValueRegion *R); 513 514 SVal getBindingForFieldOrElementCommon(RegionBindingsConstRef B, 515 const TypedValueRegion *R, 516 QualType Ty); 517 518 SVal getLazyBinding(const SubRegion *LazyBindingRegion, 519 RegionBindingsRef LazyBinding); 520 521 /// Get bindings for the values in a struct and return a CompoundVal, used 522 /// when doing struct copy: 523 /// struct s x, y; 524 /// x = y; 525 /// y's value is retrieved by this method. 526 SVal getBindingForStruct(RegionBindingsConstRef B, const TypedValueRegion *R); 527 SVal getBindingForArray(RegionBindingsConstRef B, const TypedValueRegion *R); 528 NonLoc createLazyBinding(RegionBindingsConstRef B, const TypedValueRegion *R); 529 530 /// Used to lazily generate derived symbols for bindings that are defined 531 /// implicitly by default bindings in a super region. 532 /// 533 /// Note that callers may need to specially handle LazyCompoundVals, which 534 /// are returned as is in case the caller needs to treat them differently. 535 Optional<SVal> getBindingForDerivedDefaultValue(RegionBindingsConstRef B, 536 const MemRegion *superR, 537 const TypedValueRegion *R, 538 QualType Ty); 539 540 /// Get the state and region whose binding this region \p R corresponds to. 541 /// 542 /// If there is no lazy binding for \p R, the returned value will have a null 543 /// \c second. Note that a null pointer can represents a valid Store. 544 std::pair<Store, const SubRegion *> 545 findLazyBinding(RegionBindingsConstRef B, const SubRegion *R, 546 const SubRegion *originalRegion); 547 548 /// Returns the cached set of interesting SVals contained within a lazy 549 /// binding. 550 /// 551 /// The precise value of "interesting" is determined for the purposes of 552 /// RegionStore's internal analysis. It must always contain all regions and 553 /// symbols, but may omit constants and other kinds of SVal. 554 const SValListTy &getInterestingValues(nonloc::LazyCompoundVal LCV); 555 556 //===------------------------------------------------------------------===// 557 // State pruning. 558 //===------------------------------------------------------------------===// 559 560 /// removeDeadBindings - Scans the RegionStore of 'state' for dead values. 561 /// It returns a new Store with these values removed. 562 StoreRef removeDeadBindings(Store store, const StackFrameContext *LCtx, 563 SymbolReaper& SymReaper) override; 564 565 //===------------------------------------------------------------------===// 566 // Region "extents". 567 //===------------------------------------------------------------------===// 568 569 // FIXME: This method will soon be eliminated; see the note in Store.h. 570 DefinedOrUnknownSVal getSizeInElements(ProgramStateRef state, 571 const MemRegion* R, 572 QualType EleTy) override; 573 574 //===------------------------------------------------------------------===// 575 // Utility methods. 576 //===------------------------------------------------------------------===// 577 578 RegionBindingsRef getRegionBindings(Store store) const { 579 return RegionBindingsRef(CBFactory, 580 static_cast<const RegionBindings::TreeTy*>(store), 581 RBFactory.getTreeFactory()); 582 } 583 584 void print(Store store, raw_ostream &Out, const char* nl, 585 const char *sep) override; 586 587 void iterBindings(Store store, BindingsHandler& f) override { 588 RegionBindingsRef B = getRegionBindings(store); 589 for (RegionBindingsRef::iterator I = B.begin(), E = B.end(); I != E; ++I) { 590 const ClusterBindings &Cluster = I.getData(); 591 for (ClusterBindings::iterator CI = Cluster.begin(), CE = Cluster.end(); 592 CI != CE; ++CI) { 593 const BindingKey &K = CI.getKey(); 594 if (!K.isDirect()) 595 continue; 596 if (const SubRegion *R = dyn_cast<SubRegion>(K.getRegion())) { 597 // FIXME: Possibly incorporate the offset? 598 if (!f.HandleBinding(*this, store, R, CI.getData())) 599 return; 600 } 601 } 602 } 603 } 604 }; 605 606 } // end anonymous namespace 607 608 //===----------------------------------------------------------------------===// 609 // RegionStore creation. 610 //===----------------------------------------------------------------------===// 611 612 std::unique_ptr<StoreManager> 613 ento::CreateRegionStoreManager(ProgramStateManager &StMgr) { 614 RegionStoreFeatures F = maximal_features_tag(); 615 return llvm::make_unique<RegionStoreManager>(StMgr, F); 616 } 617 618 std::unique_ptr<StoreManager> 619 ento::CreateFieldsOnlyRegionStoreManager(ProgramStateManager &StMgr) { 620 RegionStoreFeatures F = minimal_features_tag(); 621 F.enableFields(true); 622 return llvm::make_unique<RegionStoreManager>(StMgr, F); 623 } 624 625 626 //===----------------------------------------------------------------------===// 627 // Region Cluster analysis. 628 //===----------------------------------------------------------------------===// 629 630 namespace { 631 /// Used to determine which global regions are automatically included in the 632 /// initial worklist of a ClusterAnalysis. 633 enum GlobalsFilterKind { 634 /// Don't include any global regions. 635 GFK_None, 636 /// Only include system globals. 637 GFK_SystemOnly, 638 /// Include all global regions. 639 GFK_All 640 }; 641 642 template <typename DERIVED> 643 class ClusterAnalysis { 644 protected: 645 typedef llvm::DenseMap<const MemRegion *, const ClusterBindings *> ClusterMap; 646 typedef const MemRegion * WorkListElement; 647 typedef SmallVector<WorkListElement, 10> WorkList; 648 649 llvm::SmallPtrSet<const ClusterBindings *, 16> Visited; 650 651 WorkList WL; 652 653 RegionStoreManager &RM; 654 ASTContext &Ctx; 655 SValBuilder &svalBuilder; 656 657 RegionBindingsRef B; 658 659 private: 660 GlobalsFilterKind GlobalsFilter; 661 662 protected: 663 const ClusterBindings *getCluster(const MemRegion *R) { 664 return B.lookup(R); 665 } 666 667 /// Returns true if the memory space of the given region is one of the global 668 /// regions specially included at the start of analysis. 669 bool isInitiallyIncludedGlobalRegion(const MemRegion *R) { 670 switch (GlobalsFilter) { 671 case GFK_None: 672 return false; 673 case GFK_SystemOnly: 674 return isa<GlobalSystemSpaceRegion>(R->getMemorySpace()); 675 case GFK_All: 676 return isa<NonStaticGlobalSpaceRegion>(R->getMemorySpace()); 677 } 678 679 llvm_unreachable("unknown globals filter"); 680 } 681 682 public: 683 ClusterAnalysis(RegionStoreManager &rm, ProgramStateManager &StateMgr, 684 RegionBindingsRef b, GlobalsFilterKind GFK) 685 : RM(rm), Ctx(StateMgr.getContext()), 686 svalBuilder(StateMgr.getSValBuilder()), 687 B(b), GlobalsFilter(GFK) {} 688 689 RegionBindingsRef getRegionBindings() const { return B; } 690 691 bool isVisited(const MemRegion *R) { 692 return Visited.count(getCluster(R)); 693 } 694 695 void GenerateClusters() { 696 // Scan the entire set of bindings and record the region clusters. 697 for (RegionBindingsRef::iterator RI = B.begin(), RE = B.end(); 698 RI != RE; ++RI){ 699 const MemRegion *Base = RI.getKey(); 700 701 const ClusterBindings &Cluster = RI.getData(); 702 assert(!Cluster.isEmpty() && "Empty clusters should be removed"); 703 static_cast<DERIVED*>(this)->VisitAddedToCluster(Base, Cluster); 704 705 // If this is an interesting global region, add it the work list up front. 706 if (isInitiallyIncludedGlobalRegion(Base)) 707 AddToWorkList(WorkListElement(Base), &Cluster); 708 } 709 } 710 711 bool AddToWorkList(WorkListElement E, const ClusterBindings *C) { 712 if (C && !Visited.insert(C).second) 713 return false; 714 WL.push_back(E); 715 return true; 716 } 717 718 bool AddToWorkList(const MemRegion *R) { 719 const MemRegion *BaseR = R->getBaseRegion(); 720 return AddToWorkList(WorkListElement(BaseR), getCluster(BaseR)); 721 } 722 723 void RunWorkList() { 724 while (!WL.empty()) { 725 WorkListElement E = WL.pop_back_val(); 726 const MemRegion *BaseR = E; 727 728 static_cast<DERIVED*>(this)->VisitCluster(BaseR, getCluster(BaseR)); 729 } 730 } 731 732 void VisitAddedToCluster(const MemRegion *baseR, const ClusterBindings &C) {} 733 void VisitCluster(const MemRegion *baseR, const ClusterBindings *C) {} 734 735 void VisitCluster(const MemRegion *BaseR, const ClusterBindings *C, 736 bool Flag) { 737 static_cast<DERIVED*>(this)->VisitCluster(BaseR, C); 738 } 739 }; 740 } 741 742 //===----------------------------------------------------------------------===// 743 // Binding invalidation. 744 //===----------------------------------------------------------------------===// 745 746 bool RegionStoreManager::scanReachableSymbols(Store S, const MemRegion *R, 747 ScanReachableSymbols &Callbacks) { 748 assert(R == R->getBaseRegion() && "Should only be called for base regions"); 749 RegionBindingsRef B = getRegionBindings(S); 750 const ClusterBindings *Cluster = B.lookup(R); 751 752 if (!Cluster) 753 return true; 754 755 for (ClusterBindings::iterator RI = Cluster->begin(), RE = Cluster->end(); 756 RI != RE; ++RI) { 757 if (!Callbacks.scan(RI.getData())) 758 return false; 759 } 760 761 return true; 762 } 763 764 static inline bool isUnionField(const FieldRegion *FR) { 765 return FR->getDecl()->getParent()->isUnion(); 766 } 767 768 typedef SmallVector<const FieldDecl *, 8> FieldVector; 769 770 static void getSymbolicOffsetFields(BindingKey K, FieldVector &Fields) { 771 assert(K.hasSymbolicOffset() && "Not implemented for concrete offset keys"); 772 773 const MemRegion *Base = K.getConcreteOffsetRegion(); 774 const MemRegion *R = K.getRegion(); 775 776 while (R != Base) { 777 if (const FieldRegion *FR = dyn_cast<FieldRegion>(R)) 778 if (!isUnionField(FR)) 779 Fields.push_back(FR->getDecl()); 780 781 R = cast<SubRegion>(R)->getSuperRegion(); 782 } 783 } 784 785 static bool isCompatibleWithFields(BindingKey K, const FieldVector &Fields) { 786 assert(K.hasSymbolicOffset() && "Not implemented for concrete offset keys"); 787 788 if (Fields.empty()) 789 return true; 790 791 FieldVector FieldsInBindingKey; 792 getSymbolicOffsetFields(K, FieldsInBindingKey); 793 794 ptrdiff_t Delta = FieldsInBindingKey.size() - Fields.size(); 795 if (Delta >= 0) 796 return std::equal(FieldsInBindingKey.begin() + Delta, 797 FieldsInBindingKey.end(), 798 Fields.begin()); 799 else 800 return std::equal(FieldsInBindingKey.begin(), FieldsInBindingKey.end(), 801 Fields.begin() - Delta); 802 } 803 804 /// Collects all bindings in \p Cluster that may refer to bindings within 805 /// \p Top. 806 /// 807 /// Each binding is a pair whose \c first is the key (a BindingKey) and whose 808 /// \c second is the value (an SVal). 809 /// 810 /// The \p IncludeAllDefaultBindings parameter specifies whether to include 811 /// default bindings that may extend beyond \p Top itself, e.g. if \p Top is 812 /// an aggregate within a larger aggregate with a default binding. 813 static void 814 collectSubRegionBindings(SmallVectorImpl<BindingPair> &Bindings, 815 SValBuilder &SVB, const ClusterBindings &Cluster, 816 const SubRegion *Top, BindingKey TopKey, 817 bool IncludeAllDefaultBindings) { 818 FieldVector FieldsInSymbolicSubregions; 819 if (TopKey.hasSymbolicOffset()) { 820 getSymbolicOffsetFields(TopKey, FieldsInSymbolicSubregions); 821 Top = cast<SubRegion>(TopKey.getConcreteOffsetRegion()); 822 TopKey = BindingKey::Make(Top, BindingKey::Default); 823 } 824 825 // Find the length (in bits) of the region being invalidated. 826 uint64_t Length = UINT64_MAX; 827 SVal Extent = Top->getExtent(SVB); 828 if (Optional<nonloc::ConcreteInt> ExtentCI = 829 Extent.getAs<nonloc::ConcreteInt>()) { 830 const llvm::APSInt &ExtentInt = ExtentCI->getValue(); 831 assert(ExtentInt.isNonNegative() || ExtentInt.isUnsigned()); 832 // Extents are in bytes but region offsets are in bits. Be careful! 833 Length = ExtentInt.getLimitedValue() * SVB.getContext().getCharWidth(); 834 } else if (const FieldRegion *FR = dyn_cast<FieldRegion>(Top)) { 835 if (FR->getDecl()->isBitField()) 836 Length = FR->getDecl()->getBitWidthValue(SVB.getContext()); 837 } 838 839 for (ClusterBindings::iterator I = Cluster.begin(), E = Cluster.end(); 840 I != E; ++I) { 841 BindingKey NextKey = I.getKey(); 842 if (NextKey.getRegion() == TopKey.getRegion()) { 843 // FIXME: This doesn't catch the case where we're really invalidating a 844 // region with a symbolic offset. Example: 845 // R: points[i].y 846 // Next: points[0].x 847 848 if (NextKey.getOffset() > TopKey.getOffset() && 849 NextKey.getOffset() - TopKey.getOffset() < Length) { 850 // Case 1: The next binding is inside the region we're invalidating. 851 // Include it. 852 Bindings.push_back(*I); 853 854 } else if (NextKey.getOffset() == TopKey.getOffset()) { 855 // Case 2: The next binding is at the same offset as the region we're 856 // invalidating. In this case, we need to leave default bindings alone, 857 // since they may be providing a default value for a regions beyond what 858 // we're invalidating. 859 // FIXME: This is probably incorrect; consider invalidating an outer 860 // struct whose first field is bound to a LazyCompoundVal. 861 if (IncludeAllDefaultBindings || NextKey.isDirect()) 862 Bindings.push_back(*I); 863 } 864 865 } else if (NextKey.hasSymbolicOffset()) { 866 const MemRegion *Base = NextKey.getConcreteOffsetRegion(); 867 if (Top->isSubRegionOf(Base)) { 868 // Case 3: The next key is symbolic and we just changed something within 869 // its concrete region. We don't know if the binding is still valid, so 870 // we'll be conservative and include it. 871 if (IncludeAllDefaultBindings || NextKey.isDirect()) 872 if (isCompatibleWithFields(NextKey, FieldsInSymbolicSubregions)) 873 Bindings.push_back(*I); 874 } else if (const SubRegion *BaseSR = dyn_cast<SubRegion>(Base)) { 875 // Case 4: The next key is symbolic, but we changed a known 876 // super-region. In this case the binding is certainly included. 877 if (Top == Base || BaseSR->isSubRegionOf(Top)) 878 if (isCompatibleWithFields(NextKey, FieldsInSymbolicSubregions)) 879 Bindings.push_back(*I); 880 } 881 } 882 } 883 } 884 885 static void 886 collectSubRegionBindings(SmallVectorImpl<BindingPair> &Bindings, 887 SValBuilder &SVB, const ClusterBindings &Cluster, 888 const SubRegion *Top, bool IncludeAllDefaultBindings) { 889 collectSubRegionBindings(Bindings, SVB, Cluster, Top, 890 BindingKey::Make(Top, BindingKey::Default), 891 IncludeAllDefaultBindings); 892 } 893 894 RegionBindingsRef 895 RegionStoreManager::removeSubRegionBindings(RegionBindingsConstRef B, 896 const SubRegion *Top) { 897 BindingKey TopKey = BindingKey::Make(Top, BindingKey::Default); 898 const MemRegion *ClusterHead = TopKey.getBaseRegion(); 899 900 if (Top == ClusterHead) { 901 // We can remove an entire cluster's bindings all in one go. 902 return B.remove(Top); 903 } 904 905 const ClusterBindings *Cluster = B.lookup(ClusterHead); 906 if (!Cluster) { 907 // If we're invalidating a region with a symbolic offset, we need to make 908 // sure we don't treat the base region as uninitialized anymore. 909 if (TopKey.hasSymbolicOffset()) { 910 const SubRegion *Concrete = TopKey.getConcreteOffsetRegion(); 911 return B.addBinding(Concrete, BindingKey::Default, UnknownVal()); 912 } 913 return B; 914 } 915 916 SmallVector<BindingPair, 32> Bindings; 917 collectSubRegionBindings(Bindings, svalBuilder, *Cluster, Top, TopKey, 918 /*IncludeAllDefaultBindings=*/false); 919 920 ClusterBindingsRef Result(*Cluster, CBFactory); 921 for (SmallVectorImpl<BindingPair>::const_iterator I = Bindings.begin(), 922 E = Bindings.end(); 923 I != E; ++I) 924 Result = Result.remove(I->first); 925 926 // If we're invalidating a region with a symbolic offset, we need to make sure 927 // we don't treat the base region as uninitialized anymore. 928 // FIXME: This isn't very precise; see the example in 929 // collectSubRegionBindings. 930 if (TopKey.hasSymbolicOffset()) { 931 const SubRegion *Concrete = TopKey.getConcreteOffsetRegion(); 932 Result = Result.add(BindingKey::Make(Concrete, BindingKey::Default), 933 UnknownVal()); 934 } 935 936 if (Result.isEmpty()) 937 return B.remove(ClusterHead); 938 return B.add(ClusterHead, Result.asImmutableMap()); 939 } 940 941 namespace { 942 class invalidateRegionsWorker : public ClusterAnalysis<invalidateRegionsWorker> 943 { 944 const Expr *Ex; 945 unsigned Count; 946 const LocationContext *LCtx; 947 InvalidatedSymbols &IS; 948 RegionAndSymbolInvalidationTraits &ITraits; 949 StoreManager::InvalidatedRegions *Regions; 950 public: 951 invalidateRegionsWorker(RegionStoreManager &rm, 952 ProgramStateManager &stateMgr, 953 RegionBindingsRef b, 954 const Expr *ex, unsigned count, 955 const LocationContext *lctx, 956 InvalidatedSymbols &is, 957 RegionAndSymbolInvalidationTraits &ITraitsIn, 958 StoreManager::InvalidatedRegions *r, 959 GlobalsFilterKind GFK) 960 : ClusterAnalysis<invalidateRegionsWorker>(rm, stateMgr, b, GFK), 961 Ex(ex), Count(count), LCtx(lctx), IS(is), ITraits(ITraitsIn), Regions(r){} 962 963 void VisitCluster(const MemRegion *baseR, const ClusterBindings *C); 964 void VisitBinding(SVal V); 965 }; 966 } 967 968 void invalidateRegionsWorker::VisitBinding(SVal V) { 969 // A symbol? Mark it touched by the invalidation. 970 if (SymbolRef Sym = V.getAsSymbol()) 971 IS.insert(Sym); 972 973 if (const MemRegion *R = V.getAsRegion()) { 974 AddToWorkList(R); 975 return; 976 } 977 978 // Is it a LazyCompoundVal? All references get invalidated as well. 979 if (Optional<nonloc::LazyCompoundVal> LCS = 980 V.getAs<nonloc::LazyCompoundVal>()) { 981 982 const RegionStoreManager::SValListTy &Vals = RM.getInterestingValues(*LCS); 983 984 for (RegionStoreManager::SValListTy::const_iterator I = Vals.begin(), 985 E = Vals.end(); 986 I != E; ++I) 987 VisitBinding(*I); 988 989 return; 990 } 991 } 992 993 void invalidateRegionsWorker::VisitCluster(const MemRegion *baseR, 994 const ClusterBindings *C) { 995 996 bool PreserveRegionsContents = 997 ITraits.hasTrait(baseR, 998 RegionAndSymbolInvalidationTraits::TK_PreserveContents); 999 1000 if (C) { 1001 for (ClusterBindings::iterator I = C->begin(), E = C->end(); I != E; ++I) 1002 VisitBinding(I.getData()); 1003 1004 // Invalidate regions contents. 1005 if (!PreserveRegionsContents) 1006 B = B.remove(baseR); 1007 } 1008 1009 // BlockDataRegion? If so, invalidate captured variables that are passed 1010 // by reference. 1011 if (const BlockDataRegion *BR = dyn_cast<BlockDataRegion>(baseR)) { 1012 for (BlockDataRegion::referenced_vars_iterator 1013 BI = BR->referenced_vars_begin(), BE = BR->referenced_vars_end() ; 1014 BI != BE; ++BI) { 1015 const VarRegion *VR = BI.getCapturedRegion(); 1016 const VarDecl *VD = VR->getDecl(); 1017 if (VD->hasAttr<BlocksAttr>() || !VD->hasLocalStorage()) { 1018 AddToWorkList(VR); 1019 } 1020 else if (Loc::isLocType(VR->getValueType())) { 1021 // Map the current bindings to a Store to retrieve the value 1022 // of the binding. If that binding itself is a region, we should 1023 // invalidate that region. This is because a block may capture 1024 // a pointer value, but the thing pointed by that pointer may 1025 // get invalidated. 1026 SVal V = RM.getBinding(B, loc::MemRegionVal(VR)); 1027 if (Optional<Loc> L = V.getAs<Loc>()) { 1028 if (const MemRegion *LR = L->getAsRegion()) 1029 AddToWorkList(LR); 1030 } 1031 } 1032 } 1033 return; 1034 } 1035 1036 // Symbolic region? 1037 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(baseR)) 1038 IS.insert(SR->getSymbol()); 1039 1040 // Nothing else should be done in the case when we preserve regions context. 1041 if (PreserveRegionsContents) 1042 return; 1043 1044 // Otherwise, we have a normal data region. Record that we touched the region. 1045 if (Regions) 1046 Regions->push_back(baseR); 1047 1048 if (isa<AllocaRegion>(baseR) || isa<SymbolicRegion>(baseR)) { 1049 // Invalidate the region by setting its default value to 1050 // conjured symbol. The type of the symbol is irrelevant. 1051 DefinedOrUnknownSVal V = 1052 svalBuilder.conjureSymbolVal(baseR, Ex, LCtx, Ctx.IntTy, Count); 1053 B = B.addBinding(baseR, BindingKey::Default, V); 1054 return; 1055 } 1056 1057 if (!baseR->isBoundable()) 1058 return; 1059 1060 const TypedValueRegion *TR = cast<TypedValueRegion>(baseR); 1061 QualType T = TR->getValueType(); 1062 1063 if (isInitiallyIncludedGlobalRegion(baseR)) { 1064 // If the region is a global and we are invalidating all globals, 1065 // erasing the entry is good enough. This causes all globals to be lazily 1066 // symbolicated from the same base symbol. 1067 return; 1068 } 1069 1070 if (T->isStructureOrClassType()) { 1071 // Invalidate the region by setting its default value to 1072 // conjured symbol. The type of the symbol is irrelevant. 1073 DefinedOrUnknownSVal V = svalBuilder.conjureSymbolVal(baseR, Ex, LCtx, 1074 Ctx.IntTy, Count); 1075 B = B.addBinding(baseR, BindingKey::Default, V); 1076 return; 1077 } 1078 1079 if (const ArrayType *AT = Ctx.getAsArrayType(T)) { 1080 // Set the default value of the array to conjured symbol. 1081 DefinedOrUnknownSVal V = 1082 svalBuilder.conjureSymbolVal(baseR, Ex, LCtx, 1083 AT->getElementType(), Count); 1084 B = B.addBinding(baseR, BindingKey::Default, V); 1085 return; 1086 } 1087 1088 DefinedOrUnknownSVal V = svalBuilder.conjureSymbolVal(baseR, Ex, LCtx, 1089 T,Count); 1090 assert(SymbolManager::canSymbolicate(T) || V.isUnknown()); 1091 B = B.addBinding(baseR, BindingKey::Direct, V); 1092 } 1093 1094 RegionBindingsRef 1095 RegionStoreManager::invalidateGlobalRegion(MemRegion::Kind K, 1096 const Expr *Ex, 1097 unsigned Count, 1098 const LocationContext *LCtx, 1099 RegionBindingsRef B, 1100 InvalidatedRegions *Invalidated) { 1101 // Bind the globals memory space to a new symbol that we will use to derive 1102 // the bindings for all globals. 1103 const GlobalsSpaceRegion *GS = MRMgr.getGlobalsRegion(K); 1104 SVal V = svalBuilder.conjureSymbolVal(/* SymbolTag = */ (const void*) GS, Ex, LCtx, 1105 /* type does not matter */ Ctx.IntTy, 1106 Count); 1107 1108 B = B.removeBinding(GS) 1109 .addBinding(BindingKey::Make(GS, BindingKey::Default), V); 1110 1111 // Even if there are no bindings in the global scope, we still need to 1112 // record that we touched it. 1113 if (Invalidated) 1114 Invalidated->push_back(GS); 1115 1116 return B; 1117 } 1118 1119 void RegionStoreManager::populateWorkList(invalidateRegionsWorker &W, 1120 ArrayRef<SVal> Values, 1121 InvalidatedRegions *TopLevelRegions) { 1122 for (ArrayRef<SVal>::iterator I = Values.begin(), 1123 E = Values.end(); I != E; ++I) { 1124 SVal V = *I; 1125 if (Optional<nonloc::LazyCompoundVal> LCS = 1126 V.getAs<nonloc::LazyCompoundVal>()) { 1127 1128 const SValListTy &Vals = getInterestingValues(*LCS); 1129 1130 for (SValListTy::const_iterator I = Vals.begin(), 1131 E = Vals.end(); I != E; ++I) { 1132 // Note: the last argument is false here because these are 1133 // non-top-level regions. 1134 if (const MemRegion *R = (*I).getAsRegion()) 1135 W.AddToWorkList(R); 1136 } 1137 continue; 1138 } 1139 1140 if (const MemRegion *R = V.getAsRegion()) { 1141 if (TopLevelRegions) 1142 TopLevelRegions->push_back(R); 1143 W.AddToWorkList(R); 1144 continue; 1145 } 1146 } 1147 } 1148 1149 StoreRef 1150 RegionStoreManager::invalidateRegions(Store store, 1151 ArrayRef<SVal> Values, 1152 const Expr *Ex, unsigned Count, 1153 const LocationContext *LCtx, 1154 const CallEvent *Call, 1155 InvalidatedSymbols &IS, 1156 RegionAndSymbolInvalidationTraits &ITraits, 1157 InvalidatedRegions *TopLevelRegions, 1158 InvalidatedRegions *Invalidated) { 1159 GlobalsFilterKind GlobalsFilter; 1160 if (Call) { 1161 if (Call->isInSystemHeader()) 1162 GlobalsFilter = GFK_SystemOnly; 1163 else 1164 GlobalsFilter = GFK_All; 1165 } else { 1166 GlobalsFilter = GFK_None; 1167 } 1168 1169 RegionBindingsRef B = getRegionBindings(store); 1170 invalidateRegionsWorker W(*this, StateMgr, B, Ex, Count, LCtx, IS, ITraits, 1171 Invalidated, GlobalsFilter); 1172 1173 // Scan the bindings and generate the clusters. 1174 W.GenerateClusters(); 1175 1176 // Add the regions to the worklist. 1177 populateWorkList(W, Values, TopLevelRegions); 1178 1179 W.RunWorkList(); 1180 1181 // Return the new bindings. 1182 B = W.getRegionBindings(); 1183 1184 // For calls, determine which global regions should be invalidated and 1185 // invalidate them. (Note that function-static and immutable globals are never 1186 // invalidated by this.) 1187 // TODO: This could possibly be more precise with modules. 1188 switch (GlobalsFilter) { 1189 case GFK_All: 1190 B = invalidateGlobalRegion(MemRegion::GlobalInternalSpaceRegionKind, 1191 Ex, Count, LCtx, B, Invalidated); 1192 // FALLTHROUGH 1193 case GFK_SystemOnly: 1194 B = invalidateGlobalRegion(MemRegion::GlobalSystemSpaceRegionKind, 1195 Ex, Count, LCtx, B, Invalidated); 1196 // FALLTHROUGH 1197 case GFK_None: 1198 break; 1199 } 1200 1201 return StoreRef(B.asStore(), *this); 1202 } 1203 1204 //===----------------------------------------------------------------------===// 1205 // Extents for regions. 1206 //===----------------------------------------------------------------------===// 1207 1208 DefinedOrUnknownSVal 1209 RegionStoreManager::getSizeInElements(ProgramStateRef state, 1210 const MemRegion *R, 1211 QualType EleTy) { 1212 SVal Size = cast<SubRegion>(R)->getExtent(svalBuilder); 1213 const llvm::APSInt *SizeInt = svalBuilder.getKnownValue(state, Size); 1214 if (!SizeInt) 1215 return UnknownVal(); 1216 1217 CharUnits RegionSize = CharUnits::fromQuantity(SizeInt->getSExtValue()); 1218 1219 if (Ctx.getAsVariableArrayType(EleTy)) { 1220 // FIXME: We need to track extra state to properly record the size 1221 // of VLAs. Returning UnknownVal here, however, is a stop-gap so that 1222 // we don't have a divide-by-zero below. 1223 return UnknownVal(); 1224 } 1225 1226 CharUnits EleSize = Ctx.getTypeSizeInChars(EleTy); 1227 1228 // If a variable is reinterpreted as a type that doesn't fit into a larger 1229 // type evenly, round it down. 1230 // This is a signed value, since it's used in arithmetic with signed indices. 1231 return svalBuilder.makeIntVal(RegionSize / EleSize, false); 1232 } 1233 1234 //===----------------------------------------------------------------------===// 1235 // Location and region casting. 1236 //===----------------------------------------------------------------------===// 1237 1238 /// ArrayToPointer - Emulates the "decay" of an array to a pointer 1239 /// type. 'Array' represents the lvalue of the array being decayed 1240 /// to a pointer, and the returned SVal represents the decayed 1241 /// version of that lvalue (i.e., a pointer to the first element of 1242 /// the array). This is called by ExprEngine when evaluating casts 1243 /// from arrays to pointers. 1244 SVal RegionStoreManager::ArrayToPointer(Loc Array, QualType T) { 1245 if (!Array.getAs<loc::MemRegionVal>()) 1246 return UnknownVal(); 1247 1248 const MemRegion* R = Array.castAs<loc::MemRegionVal>().getRegion(); 1249 NonLoc ZeroIdx = svalBuilder.makeZeroArrayIndex(); 1250 return loc::MemRegionVal(MRMgr.getElementRegion(T, ZeroIdx, R, Ctx)); 1251 } 1252 1253 //===----------------------------------------------------------------------===// 1254 // Loading values from regions. 1255 //===----------------------------------------------------------------------===// 1256 1257 SVal RegionStoreManager::getBinding(RegionBindingsConstRef B, Loc L, QualType T) { 1258 assert(!L.getAs<UnknownVal>() && "location unknown"); 1259 assert(!L.getAs<UndefinedVal>() && "location undefined"); 1260 1261 // For access to concrete addresses, return UnknownVal. Checks 1262 // for null dereferences (and similar errors) are done by checkers, not 1263 // the Store. 1264 // FIXME: We can consider lazily symbolicating such memory, but we really 1265 // should defer this when we can reason easily about symbolicating arrays 1266 // of bytes. 1267 if (L.getAs<loc::ConcreteInt>()) { 1268 return UnknownVal(); 1269 } 1270 if (!L.getAs<loc::MemRegionVal>()) { 1271 return UnknownVal(); 1272 } 1273 1274 const MemRegion *MR = L.castAs<loc::MemRegionVal>().getRegion(); 1275 1276 if (isa<AllocaRegion>(MR) || 1277 isa<SymbolicRegion>(MR) || 1278 isa<CodeTextRegion>(MR)) { 1279 if (T.isNull()) { 1280 if (const TypedRegion *TR = dyn_cast<TypedRegion>(MR)) 1281 T = TR->getLocationType(); 1282 else { 1283 const SymbolicRegion *SR = cast<SymbolicRegion>(MR); 1284 T = SR->getSymbol()->getType(); 1285 } 1286 } 1287 MR = GetElementZeroRegion(MR, T); 1288 } 1289 1290 // FIXME: Perhaps this method should just take a 'const MemRegion*' argument 1291 // instead of 'Loc', and have the other Loc cases handled at a higher level. 1292 const TypedValueRegion *R = cast<TypedValueRegion>(MR); 1293 QualType RTy = R->getValueType(); 1294 1295 // FIXME: we do not yet model the parts of a complex type, so treat the 1296 // whole thing as "unknown". 1297 if (RTy->isAnyComplexType()) 1298 return UnknownVal(); 1299 1300 // FIXME: We should eventually handle funny addressing. e.g.: 1301 // 1302 // int x = ...; 1303 // int *p = &x; 1304 // char *q = (char*) p; 1305 // char c = *q; // returns the first byte of 'x'. 1306 // 1307 // Such funny addressing will occur due to layering of regions. 1308 if (RTy->isStructureOrClassType()) 1309 return getBindingForStruct(B, R); 1310 1311 // FIXME: Handle unions. 1312 if (RTy->isUnionType()) 1313 return createLazyBinding(B, R); 1314 1315 if (RTy->isArrayType()) { 1316 if (RTy->isConstantArrayType()) 1317 return getBindingForArray(B, R); 1318 else 1319 return UnknownVal(); 1320 } 1321 1322 // FIXME: handle Vector types. 1323 if (RTy->isVectorType()) 1324 return UnknownVal(); 1325 1326 if (const FieldRegion* FR = dyn_cast<FieldRegion>(R)) 1327 return CastRetrievedVal(getBindingForField(B, FR), FR, T, false); 1328 1329 if (const ElementRegion* ER = dyn_cast<ElementRegion>(R)) { 1330 // FIXME: Here we actually perform an implicit conversion from the loaded 1331 // value to the element type. Eventually we want to compose these values 1332 // more intelligently. For example, an 'element' can encompass multiple 1333 // bound regions (e.g., several bound bytes), or could be a subset of 1334 // a larger value. 1335 return CastRetrievedVal(getBindingForElement(B, ER), ER, T, false); 1336 } 1337 1338 if (const ObjCIvarRegion *IVR = dyn_cast<ObjCIvarRegion>(R)) { 1339 // FIXME: Here we actually perform an implicit conversion from the loaded 1340 // value to the ivar type. What we should model is stores to ivars 1341 // that blow past the extent of the ivar. If the address of the ivar is 1342 // reinterpretted, it is possible we stored a different value that could 1343 // fit within the ivar. Either we need to cast these when storing them 1344 // or reinterpret them lazily (as we do here). 1345 return CastRetrievedVal(getBindingForObjCIvar(B, IVR), IVR, T, false); 1346 } 1347 1348 if (const VarRegion *VR = dyn_cast<VarRegion>(R)) { 1349 // FIXME: Here we actually perform an implicit conversion from the loaded 1350 // value to the variable type. What we should model is stores to variables 1351 // that blow past the extent of the variable. If the address of the 1352 // variable is reinterpretted, it is possible we stored a different value 1353 // that could fit within the variable. Either we need to cast these when 1354 // storing them or reinterpret them lazily (as we do here). 1355 return CastRetrievedVal(getBindingForVar(B, VR), VR, T, false); 1356 } 1357 1358 const SVal *V = B.lookup(R, BindingKey::Direct); 1359 1360 // Check if the region has a binding. 1361 if (V) 1362 return *V; 1363 1364 // The location does not have a bound value. This means that it has 1365 // the value it had upon its creation and/or entry to the analyzed 1366 // function/method. These are either symbolic values or 'undefined'. 1367 if (R->hasStackNonParametersStorage()) { 1368 // All stack variables are considered to have undefined values 1369 // upon creation. All heap allocated blocks are considered to 1370 // have undefined values as well unless they are explicitly bound 1371 // to specific values. 1372 return UndefinedVal(); 1373 } 1374 1375 // All other values are symbolic. 1376 return svalBuilder.getRegionValueSymbolVal(R); 1377 } 1378 1379 static QualType getUnderlyingType(const SubRegion *R) { 1380 QualType RegionTy; 1381 if (const TypedValueRegion *TVR = dyn_cast<TypedValueRegion>(R)) 1382 RegionTy = TVR->getValueType(); 1383 1384 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(R)) 1385 RegionTy = SR->getSymbol()->getType(); 1386 1387 return RegionTy; 1388 } 1389 1390 /// Checks to see if store \p B has a lazy binding for region \p R. 1391 /// 1392 /// If \p AllowSubregionBindings is \c false, a lazy binding will be rejected 1393 /// if there are additional bindings within \p R. 1394 /// 1395 /// Note that unlike RegionStoreManager::findLazyBinding, this will not search 1396 /// for lazy bindings for super-regions of \p R. 1397 static Optional<nonloc::LazyCompoundVal> 1398 getExistingLazyBinding(SValBuilder &SVB, RegionBindingsConstRef B, 1399 const SubRegion *R, bool AllowSubregionBindings) { 1400 Optional<SVal> V = B.getDefaultBinding(R); 1401 if (!V) 1402 return None; 1403 1404 Optional<nonloc::LazyCompoundVal> LCV = V->getAs<nonloc::LazyCompoundVal>(); 1405 if (!LCV) 1406 return None; 1407 1408 // If the LCV is for a subregion, the types might not match, and we shouldn't 1409 // reuse the binding. 1410 QualType RegionTy = getUnderlyingType(R); 1411 if (!RegionTy.isNull() && 1412 !RegionTy->isVoidPointerType()) { 1413 QualType SourceRegionTy = LCV->getRegion()->getValueType(); 1414 if (!SVB.getContext().hasSameUnqualifiedType(RegionTy, SourceRegionTy)) 1415 return None; 1416 } 1417 1418 if (!AllowSubregionBindings) { 1419 // If there are any other bindings within this region, we shouldn't reuse 1420 // the top-level binding. 1421 SmallVector<BindingPair, 16> Bindings; 1422 collectSubRegionBindings(Bindings, SVB, *B.lookup(R->getBaseRegion()), R, 1423 /*IncludeAllDefaultBindings=*/true); 1424 if (Bindings.size() > 1) 1425 return None; 1426 } 1427 1428 return *LCV; 1429 } 1430 1431 1432 std::pair<Store, const SubRegion *> 1433 RegionStoreManager::findLazyBinding(RegionBindingsConstRef B, 1434 const SubRegion *R, 1435 const SubRegion *originalRegion) { 1436 if (originalRegion != R) { 1437 if (Optional<nonloc::LazyCompoundVal> V = 1438 getExistingLazyBinding(svalBuilder, B, R, true)) 1439 return std::make_pair(V->getStore(), V->getRegion()); 1440 } 1441 1442 typedef std::pair<Store, const SubRegion *> StoreRegionPair; 1443 StoreRegionPair Result = StoreRegionPair(); 1444 1445 if (const ElementRegion *ER = dyn_cast<ElementRegion>(R)) { 1446 Result = findLazyBinding(B, cast<SubRegion>(ER->getSuperRegion()), 1447 originalRegion); 1448 1449 if (Result.second) 1450 Result.second = MRMgr.getElementRegionWithSuper(ER, Result.second); 1451 1452 } else if (const FieldRegion *FR = dyn_cast<FieldRegion>(R)) { 1453 Result = findLazyBinding(B, cast<SubRegion>(FR->getSuperRegion()), 1454 originalRegion); 1455 1456 if (Result.second) 1457 Result.second = MRMgr.getFieldRegionWithSuper(FR, Result.second); 1458 1459 } else if (const CXXBaseObjectRegion *BaseReg = 1460 dyn_cast<CXXBaseObjectRegion>(R)) { 1461 // C++ base object region is another kind of region that we should blast 1462 // through to look for lazy compound value. It is like a field region. 1463 Result = findLazyBinding(B, cast<SubRegion>(BaseReg->getSuperRegion()), 1464 originalRegion); 1465 1466 if (Result.second) 1467 Result.second = MRMgr.getCXXBaseObjectRegionWithSuper(BaseReg, 1468 Result.second); 1469 } 1470 1471 return Result; 1472 } 1473 1474 SVal RegionStoreManager::getBindingForElement(RegionBindingsConstRef B, 1475 const ElementRegion* R) { 1476 // We do not currently model bindings of the CompoundLiteralregion. 1477 if (isa<CompoundLiteralRegion>(R->getBaseRegion())) 1478 return UnknownVal(); 1479 1480 // Check if the region has a binding. 1481 if (const Optional<SVal> &V = B.getDirectBinding(R)) 1482 return *V; 1483 1484 const MemRegion* superR = R->getSuperRegion(); 1485 1486 // Check if the region is an element region of a string literal. 1487 if (const StringRegion *StrR=dyn_cast<StringRegion>(superR)) { 1488 // FIXME: Handle loads from strings where the literal is treated as 1489 // an integer, e.g., *((unsigned int*)"hello") 1490 QualType T = Ctx.getAsArrayType(StrR->getValueType())->getElementType(); 1491 if (!Ctx.hasSameUnqualifiedType(T, R->getElementType())) 1492 return UnknownVal(); 1493 1494 const StringLiteral *Str = StrR->getStringLiteral(); 1495 SVal Idx = R->getIndex(); 1496 if (Optional<nonloc::ConcreteInt> CI = Idx.getAs<nonloc::ConcreteInt>()) { 1497 int64_t i = CI->getValue().getSExtValue(); 1498 // Abort on string underrun. This can be possible by arbitrary 1499 // clients of getBindingForElement(). 1500 if (i < 0) 1501 return UndefinedVal(); 1502 int64_t length = Str->getLength(); 1503 // Technically, only i == length is guaranteed to be null. 1504 // However, such overflows should be caught before reaching this point; 1505 // the only time such an access would be made is if a string literal was 1506 // used to initialize a larger array. 1507 char c = (i >= length) ? '\0' : Str->getCodeUnit(i); 1508 return svalBuilder.makeIntVal(c, T); 1509 } 1510 } 1511 1512 // Check for loads from a code text region. For such loads, just give up. 1513 if (isa<CodeTextRegion>(superR)) 1514 return UnknownVal(); 1515 1516 // Handle the case where we are indexing into a larger scalar object. 1517 // For example, this handles: 1518 // int x = ... 1519 // char *y = &x; 1520 // return *y; 1521 // FIXME: This is a hack, and doesn't do anything really intelligent yet. 1522 const RegionRawOffset &O = R->getAsArrayOffset(); 1523 1524 // If we cannot reason about the offset, return an unknown value. 1525 if (!O.getRegion()) 1526 return UnknownVal(); 1527 1528 if (const TypedValueRegion *baseR = 1529 dyn_cast_or_null<TypedValueRegion>(O.getRegion())) { 1530 QualType baseT = baseR->getValueType(); 1531 if (baseT->isScalarType()) { 1532 QualType elemT = R->getElementType(); 1533 if (elemT->isScalarType()) { 1534 if (Ctx.getTypeSizeInChars(baseT) >= Ctx.getTypeSizeInChars(elemT)) { 1535 if (const Optional<SVal> &V = B.getDirectBinding(superR)) { 1536 if (SymbolRef parentSym = V->getAsSymbol()) 1537 return svalBuilder.getDerivedRegionValueSymbolVal(parentSym, R); 1538 1539 if (V->isUnknownOrUndef()) 1540 return *V; 1541 // Other cases: give up. We are indexing into a larger object 1542 // that has some value, but we don't know how to handle that yet. 1543 return UnknownVal(); 1544 } 1545 } 1546 } 1547 } 1548 } 1549 return getBindingForFieldOrElementCommon(B, R, R->getElementType()); 1550 } 1551 1552 SVal RegionStoreManager::getBindingForField(RegionBindingsConstRef B, 1553 const FieldRegion* R) { 1554 1555 // Check if the region has a binding. 1556 if (const Optional<SVal> &V = B.getDirectBinding(R)) 1557 return *V; 1558 1559 QualType Ty = R->getValueType(); 1560 return getBindingForFieldOrElementCommon(B, R, Ty); 1561 } 1562 1563 Optional<SVal> 1564 RegionStoreManager::getBindingForDerivedDefaultValue(RegionBindingsConstRef B, 1565 const MemRegion *superR, 1566 const TypedValueRegion *R, 1567 QualType Ty) { 1568 1569 if (const Optional<SVal> &D = B.getDefaultBinding(superR)) { 1570 const SVal &val = D.getValue(); 1571 if (SymbolRef parentSym = val.getAsSymbol()) 1572 return svalBuilder.getDerivedRegionValueSymbolVal(parentSym, R); 1573 1574 if (val.isZeroConstant()) 1575 return svalBuilder.makeZeroVal(Ty); 1576 1577 if (val.isUnknownOrUndef()) 1578 return val; 1579 1580 // Lazy bindings are usually handled through getExistingLazyBinding(). 1581 // We should unify these two code paths at some point. 1582 if (val.getAs<nonloc::LazyCompoundVal>()) 1583 return val; 1584 1585 llvm_unreachable("Unknown default value"); 1586 } 1587 1588 return None; 1589 } 1590 1591 SVal RegionStoreManager::getLazyBinding(const SubRegion *LazyBindingRegion, 1592 RegionBindingsRef LazyBinding) { 1593 SVal Result; 1594 if (const ElementRegion *ER = dyn_cast<ElementRegion>(LazyBindingRegion)) 1595 Result = getBindingForElement(LazyBinding, ER); 1596 else 1597 Result = getBindingForField(LazyBinding, 1598 cast<FieldRegion>(LazyBindingRegion)); 1599 1600 // FIXME: This is a hack to deal with RegionStore's inability to distinguish a 1601 // default value for /part/ of an aggregate from a default value for the 1602 // /entire/ aggregate. The most common case of this is when struct Outer 1603 // has as its first member a struct Inner, which is copied in from a stack 1604 // variable. In this case, even if the Outer's default value is symbolic, 0, 1605 // or unknown, it gets overridden by the Inner's default value of undefined. 1606 // 1607 // This is a general problem -- if the Inner is zero-initialized, the Outer 1608 // will now look zero-initialized. The proper way to solve this is with a 1609 // new version of RegionStore that tracks the extent of a binding as well 1610 // as the offset. 1611 // 1612 // This hack only takes care of the undefined case because that can very 1613 // quickly result in a warning. 1614 if (Result.isUndef()) 1615 Result = UnknownVal(); 1616 1617 return Result; 1618 } 1619 1620 SVal 1621 RegionStoreManager::getBindingForFieldOrElementCommon(RegionBindingsConstRef B, 1622 const TypedValueRegion *R, 1623 QualType Ty) { 1624 1625 // At this point we have already checked in either getBindingForElement or 1626 // getBindingForField if 'R' has a direct binding. 1627 1628 // Lazy binding? 1629 Store lazyBindingStore = nullptr; 1630 const SubRegion *lazyBindingRegion = nullptr; 1631 std::tie(lazyBindingStore, lazyBindingRegion) = findLazyBinding(B, R, R); 1632 if (lazyBindingRegion) 1633 return getLazyBinding(lazyBindingRegion, 1634 getRegionBindings(lazyBindingStore)); 1635 1636 // Record whether or not we see a symbolic index. That can completely 1637 // be out of scope of our lookup. 1638 bool hasSymbolicIndex = false; 1639 1640 // FIXME: This is a hack to deal with RegionStore's inability to distinguish a 1641 // default value for /part/ of an aggregate from a default value for the 1642 // /entire/ aggregate. The most common case of this is when struct Outer 1643 // has as its first member a struct Inner, which is copied in from a stack 1644 // variable. In this case, even if the Outer's default value is symbolic, 0, 1645 // or unknown, it gets overridden by the Inner's default value of undefined. 1646 // 1647 // This is a general problem -- if the Inner is zero-initialized, the Outer 1648 // will now look zero-initialized. The proper way to solve this is with a 1649 // new version of RegionStore that tracks the extent of a binding as well 1650 // as the offset. 1651 // 1652 // This hack only takes care of the undefined case because that can very 1653 // quickly result in a warning. 1654 bool hasPartialLazyBinding = false; 1655 1656 const SubRegion *SR = dyn_cast<SubRegion>(R); 1657 while (SR) { 1658 const MemRegion *Base = SR->getSuperRegion(); 1659 if (Optional<SVal> D = getBindingForDerivedDefaultValue(B, Base, R, Ty)) { 1660 if (D->getAs<nonloc::LazyCompoundVal>()) { 1661 hasPartialLazyBinding = true; 1662 break; 1663 } 1664 1665 return *D; 1666 } 1667 1668 if (const ElementRegion *ER = dyn_cast<ElementRegion>(Base)) { 1669 NonLoc index = ER->getIndex(); 1670 if (!index.isConstant()) 1671 hasSymbolicIndex = true; 1672 } 1673 1674 // If our super region is a field or element itself, walk up the region 1675 // hierarchy to see if there is a default value installed in an ancestor. 1676 SR = dyn_cast<SubRegion>(Base); 1677 } 1678 1679 if (R->hasStackNonParametersStorage()) { 1680 if (isa<ElementRegion>(R)) { 1681 // Currently we don't reason specially about Clang-style vectors. Check 1682 // if superR is a vector and if so return Unknown. 1683 if (const TypedValueRegion *typedSuperR = 1684 dyn_cast<TypedValueRegion>(R->getSuperRegion())) { 1685 if (typedSuperR->getValueType()->isVectorType()) 1686 return UnknownVal(); 1687 } 1688 } 1689 1690 // FIXME: We also need to take ElementRegions with symbolic indexes into 1691 // account. This case handles both directly accessing an ElementRegion 1692 // with a symbolic offset, but also fields within an element with 1693 // a symbolic offset. 1694 if (hasSymbolicIndex) 1695 return UnknownVal(); 1696 1697 if (!hasPartialLazyBinding) 1698 return UndefinedVal(); 1699 } 1700 1701 // All other values are symbolic. 1702 return svalBuilder.getRegionValueSymbolVal(R); 1703 } 1704 1705 SVal RegionStoreManager::getBindingForObjCIvar(RegionBindingsConstRef B, 1706 const ObjCIvarRegion* R) { 1707 // Check if the region has a binding. 1708 if (const Optional<SVal> &V = B.getDirectBinding(R)) 1709 return *V; 1710 1711 const MemRegion *superR = R->getSuperRegion(); 1712 1713 // Check if the super region has a default binding. 1714 if (const Optional<SVal> &V = B.getDefaultBinding(superR)) { 1715 if (SymbolRef parentSym = V->getAsSymbol()) 1716 return svalBuilder.getDerivedRegionValueSymbolVal(parentSym, R); 1717 1718 // Other cases: give up. 1719 return UnknownVal(); 1720 } 1721 1722 return getBindingForLazySymbol(R); 1723 } 1724 1725 SVal RegionStoreManager::getBindingForVar(RegionBindingsConstRef B, 1726 const VarRegion *R) { 1727 1728 // Check if the region has a binding. 1729 if (const Optional<SVal> &V = B.getDirectBinding(R)) 1730 return *V; 1731 1732 // Lazily derive a value for the VarRegion. 1733 const VarDecl *VD = R->getDecl(); 1734 const MemSpaceRegion *MS = R->getMemorySpace(); 1735 1736 // Arguments are always symbolic. 1737 if (isa<StackArgumentsSpaceRegion>(MS)) 1738 return svalBuilder.getRegionValueSymbolVal(R); 1739 1740 // Is 'VD' declared constant? If so, retrieve the constant value. 1741 if (VD->getType().isConstQualified()) 1742 if (const Expr *Init = VD->getInit()) 1743 if (Optional<SVal> V = svalBuilder.getConstantVal(Init)) 1744 return *V; 1745 1746 // This must come after the check for constants because closure-captured 1747 // constant variables may appear in UnknownSpaceRegion. 1748 if (isa<UnknownSpaceRegion>(MS)) 1749 return svalBuilder.getRegionValueSymbolVal(R); 1750 1751 if (isa<GlobalsSpaceRegion>(MS)) { 1752 QualType T = VD->getType(); 1753 1754 // Function-scoped static variables are default-initialized to 0; if they 1755 // have an initializer, it would have been processed by now. 1756 if (isa<StaticGlobalSpaceRegion>(MS)) 1757 return svalBuilder.makeZeroVal(T); 1758 1759 if (Optional<SVal> V = getBindingForDerivedDefaultValue(B, MS, R, T)) { 1760 assert(!V->getAs<nonloc::LazyCompoundVal>()); 1761 return V.getValue(); 1762 } 1763 1764 return svalBuilder.getRegionValueSymbolVal(R); 1765 } 1766 1767 return UndefinedVal(); 1768 } 1769 1770 SVal RegionStoreManager::getBindingForLazySymbol(const TypedValueRegion *R) { 1771 // All other values are symbolic. 1772 return svalBuilder.getRegionValueSymbolVal(R); 1773 } 1774 1775 const RegionStoreManager::SValListTy & 1776 RegionStoreManager::getInterestingValues(nonloc::LazyCompoundVal LCV) { 1777 // First, check the cache. 1778 LazyBindingsMapTy::iterator I = LazyBindingsMap.find(LCV.getCVData()); 1779 if (I != LazyBindingsMap.end()) 1780 return I->second; 1781 1782 // If we don't have a list of values cached, start constructing it. 1783 SValListTy List; 1784 1785 const SubRegion *LazyR = LCV.getRegion(); 1786 RegionBindingsRef B = getRegionBindings(LCV.getStore()); 1787 1788 // If this region had /no/ bindings at the time, there are no interesting 1789 // values to return. 1790 const ClusterBindings *Cluster = B.lookup(LazyR->getBaseRegion()); 1791 if (!Cluster) 1792 return (LazyBindingsMap[LCV.getCVData()] = std::move(List)); 1793 1794 SmallVector<BindingPair, 32> Bindings; 1795 collectSubRegionBindings(Bindings, svalBuilder, *Cluster, LazyR, 1796 /*IncludeAllDefaultBindings=*/true); 1797 for (SmallVectorImpl<BindingPair>::const_iterator I = Bindings.begin(), 1798 E = Bindings.end(); 1799 I != E; ++I) { 1800 SVal V = I->second; 1801 if (V.isUnknownOrUndef() || V.isConstant()) 1802 continue; 1803 1804 if (Optional<nonloc::LazyCompoundVal> InnerLCV = 1805 V.getAs<nonloc::LazyCompoundVal>()) { 1806 const SValListTy &InnerList = getInterestingValues(*InnerLCV); 1807 List.insert(List.end(), InnerList.begin(), InnerList.end()); 1808 continue; 1809 } 1810 1811 List.push_back(V); 1812 } 1813 1814 return (LazyBindingsMap[LCV.getCVData()] = std::move(List)); 1815 } 1816 1817 NonLoc RegionStoreManager::createLazyBinding(RegionBindingsConstRef B, 1818 const TypedValueRegion *R) { 1819 if (Optional<nonloc::LazyCompoundVal> V = 1820 getExistingLazyBinding(svalBuilder, B, R, false)) 1821 return *V; 1822 1823 return svalBuilder.makeLazyCompoundVal(StoreRef(B.asStore(), *this), R); 1824 } 1825 1826 static bool isRecordEmpty(const RecordDecl *RD) { 1827 if (!RD->field_empty()) 1828 return false; 1829 if (const CXXRecordDecl *CRD = dyn_cast<CXXRecordDecl>(RD)) 1830 return CRD->getNumBases() == 0; 1831 return true; 1832 } 1833 1834 SVal RegionStoreManager::getBindingForStruct(RegionBindingsConstRef B, 1835 const TypedValueRegion *R) { 1836 const RecordDecl *RD = R->getValueType()->castAs<RecordType>()->getDecl(); 1837 if (!RD->getDefinition() || isRecordEmpty(RD)) 1838 return UnknownVal(); 1839 1840 return createLazyBinding(B, R); 1841 } 1842 1843 SVal RegionStoreManager::getBindingForArray(RegionBindingsConstRef B, 1844 const TypedValueRegion *R) { 1845 assert(Ctx.getAsConstantArrayType(R->getValueType()) && 1846 "Only constant array types can have compound bindings."); 1847 1848 return createLazyBinding(B, R); 1849 } 1850 1851 bool RegionStoreManager::includedInBindings(Store store, 1852 const MemRegion *region) const { 1853 RegionBindingsRef B = getRegionBindings(store); 1854 region = region->getBaseRegion(); 1855 1856 // Quick path: if the base is the head of a cluster, the region is live. 1857 if (B.lookup(region)) 1858 return true; 1859 1860 // Slow path: if the region is the VALUE of any binding, it is live. 1861 for (RegionBindingsRef::iterator RI = B.begin(), RE = B.end(); RI != RE; ++RI) { 1862 const ClusterBindings &Cluster = RI.getData(); 1863 for (ClusterBindings::iterator CI = Cluster.begin(), CE = Cluster.end(); 1864 CI != CE; ++CI) { 1865 const SVal &D = CI.getData(); 1866 if (const MemRegion *R = D.getAsRegion()) 1867 if (R->getBaseRegion() == region) 1868 return true; 1869 } 1870 } 1871 1872 return false; 1873 } 1874 1875 //===----------------------------------------------------------------------===// 1876 // Binding values to regions. 1877 //===----------------------------------------------------------------------===// 1878 1879 StoreRef RegionStoreManager::killBinding(Store ST, Loc L) { 1880 if (Optional<loc::MemRegionVal> LV = L.getAs<loc::MemRegionVal>()) 1881 if (const MemRegion* R = LV->getRegion()) 1882 return StoreRef(getRegionBindings(ST).removeBinding(R) 1883 .asImmutableMap() 1884 .getRootWithoutRetain(), 1885 *this); 1886 1887 return StoreRef(ST, *this); 1888 } 1889 1890 RegionBindingsRef 1891 RegionStoreManager::bind(RegionBindingsConstRef B, Loc L, SVal V) { 1892 if (L.getAs<loc::ConcreteInt>()) 1893 return B; 1894 1895 // If we get here, the location should be a region. 1896 const MemRegion *R = L.castAs<loc::MemRegionVal>().getRegion(); 1897 1898 // Check if the region is a struct region. 1899 if (const TypedValueRegion* TR = dyn_cast<TypedValueRegion>(R)) { 1900 QualType Ty = TR->getValueType(); 1901 if (Ty->isArrayType()) 1902 return bindArray(B, TR, V); 1903 if (Ty->isStructureOrClassType()) 1904 return bindStruct(B, TR, V); 1905 if (Ty->isVectorType()) 1906 return bindVector(B, TR, V); 1907 if (Ty->isUnionType()) 1908 return bindAggregate(B, TR, V); 1909 } 1910 1911 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(R)) { 1912 // Binding directly to a symbolic region should be treated as binding 1913 // to element 0. 1914 QualType T = SR->getSymbol()->getType(); 1915 if (T->isAnyPointerType() || T->isReferenceType()) 1916 T = T->getPointeeType(); 1917 1918 R = GetElementZeroRegion(SR, T); 1919 } 1920 1921 // Clear out bindings that may overlap with this binding. 1922 RegionBindingsRef NewB = removeSubRegionBindings(B, cast<SubRegion>(R)); 1923 return NewB.addBinding(BindingKey::Make(R, BindingKey::Direct), V); 1924 } 1925 1926 RegionBindingsRef 1927 RegionStoreManager::setImplicitDefaultValue(RegionBindingsConstRef B, 1928 const MemRegion *R, 1929 QualType T) { 1930 SVal V; 1931 1932 if (Loc::isLocType(T)) 1933 V = svalBuilder.makeNull(); 1934 else if (T->isIntegralOrEnumerationType()) 1935 V = svalBuilder.makeZeroVal(T); 1936 else if (T->isStructureOrClassType() || T->isArrayType()) { 1937 // Set the default value to a zero constant when it is a structure 1938 // or array. The type doesn't really matter. 1939 V = svalBuilder.makeZeroVal(Ctx.IntTy); 1940 } 1941 else { 1942 // We can't represent values of this type, but we still need to set a value 1943 // to record that the region has been initialized. 1944 // If this assertion ever fires, a new case should be added above -- we 1945 // should know how to default-initialize any value we can symbolicate. 1946 assert(!SymbolManager::canSymbolicate(T) && "This type is representable"); 1947 V = UnknownVal(); 1948 } 1949 1950 return B.addBinding(R, BindingKey::Default, V); 1951 } 1952 1953 RegionBindingsRef 1954 RegionStoreManager::bindArray(RegionBindingsConstRef B, 1955 const TypedValueRegion* R, 1956 SVal Init) { 1957 1958 const ArrayType *AT =cast<ArrayType>(Ctx.getCanonicalType(R->getValueType())); 1959 QualType ElementTy = AT->getElementType(); 1960 Optional<uint64_t> Size; 1961 1962 if (const ConstantArrayType* CAT = dyn_cast<ConstantArrayType>(AT)) 1963 Size = CAT->getSize().getZExtValue(); 1964 1965 // Check if the init expr is a string literal. 1966 if (Optional<loc::MemRegionVal> MRV = Init.getAs<loc::MemRegionVal>()) { 1967 const StringRegion *S = cast<StringRegion>(MRV->getRegion()); 1968 1969 // Treat the string as a lazy compound value. 1970 StoreRef store(B.asStore(), *this); 1971 nonloc::LazyCompoundVal LCV = svalBuilder.makeLazyCompoundVal(store, S) 1972 .castAs<nonloc::LazyCompoundVal>(); 1973 return bindAggregate(B, R, LCV); 1974 } 1975 1976 // Handle lazy compound values. 1977 if (Init.getAs<nonloc::LazyCompoundVal>()) 1978 return bindAggregate(B, R, Init); 1979 1980 // Remaining case: explicit compound values. 1981 1982 if (Init.isUnknown()) 1983 return setImplicitDefaultValue(B, R, ElementTy); 1984 1985 const nonloc::CompoundVal& CV = Init.castAs<nonloc::CompoundVal>(); 1986 nonloc::CompoundVal::iterator VI = CV.begin(), VE = CV.end(); 1987 uint64_t i = 0; 1988 1989 RegionBindingsRef NewB(B); 1990 1991 for (; Size.hasValue() ? i < Size.getValue() : true ; ++i, ++VI) { 1992 // The init list might be shorter than the array length. 1993 if (VI == VE) 1994 break; 1995 1996 const NonLoc &Idx = svalBuilder.makeArrayIndex(i); 1997 const ElementRegion *ER = MRMgr.getElementRegion(ElementTy, Idx, R, Ctx); 1998 1999 if (ElementTy->isStructureOrClassType()) 2000 NewB = bindStruct(NewB, ER, *VI); 2001 else if (ElementTy->isArrayType()) 2002 NewB = bindArray(NewB, ER, *VI); 2003 else 2004 NewB = bind(NewB, loc::MemRegionVal(ER), *VI); 2005 } 2006 2007 // If the init list is shorter than the array length, set the 2008 // array default value. 2009 if (Size.hasValue() && i < Size.getValue()) 2010 NewB = setImplicitDefaultValue(NewB, R, ElementTy); 2011 2012 return NewB; 2013 } 2014 2015 RegionBindingsRef RegionStoreManager::bindVector(RegionBindingsConstRef B, 2016 const TypedValueRegion* R, 2017 SVal V) { 2018 QualType T = R->getValueType(); 2019 assert(T->isVectorType()); 2020 const VectorType *VT = T->getAs<VectorType>(); // Use getAs for typedefs. 2021 2022 // Handle lazy compound values and symbolic values. 2023 if (V.getAs<nonloc::LazyCompoundVal>() || V.getAs<nonloc::SymbolVal>()) 2024 return bindAggregate(B, R, V); 2025 2026 // We may get non-CompoundVal accidentally due to imprecise cast logic or 2027 // that we are binding symbolic struct value. Kill the field values, and if 2028 // the value is symbolic go and bind it as a "default" binding. 2029 if (!V.getAs<nonloc::CompoundVal>()) { 2030 return bindAggregate(B, R, UnknownVal()); 2031 } 2032 2033 QualType ElemType = VT->getElementType(); 2034 nonloc::CompoundVal CV = V.castAs<nonloc::CompoundVal>(); 2035 nonloc::CompoundVal::iterator VI = CV.begin(), VE = CV.end(); 2036 unsigned index = 0, numElements = VT->getNumElements(); 2037 RegionBindingsRef NewB(B); 2038 2039 for ( ; index != numElements ; ++index) { 2040 if (VI == VE) 2041 break; 2042 2043 NonLoc Idx = svalBuilder.makeArrayIndex(index); 2044 const ElementRegion *ER = MRMgr.getElementRegion(ElemType, Idx, R, Ctx); 2045 2046 if (ElemType->isArrayType()) 2047 NewB = bindArray(NewB, ER, *VI); 2048 else if (ElemType->isStructureOrClassType()) 2049 NewB = bindStruct(NewB, ER, *VI); 2050 else 2051 NewB = bind(NewB, loc::MemRegionVal(ER), *VI); 2052 } 2053 return NewB; 2054 } 2055 2056 Optional<RegionBindingsRef> 2057 RegionStoreManager::tryBindSmallStruct(RegionBindingsConstRef B, 2058 const TypedValueRegion *R, 2059 const RecordDecl *RD, 2060 nonloc::LazyCompoundVal LCV) { 2061 FieldVector Fields; 2062 2063 if (const CXXRecordDecl *Class = dyn_cast<CXXRecordDecl>(RD)) 2064 if (Class->getNumBases() != 0 || Class->getNumVBases() != 0) 2065 return None; 2066 2067 for (const auto *FD : RD->fields()) { 2068 if (FD->isUnnamedBitfield()) 2069 continue; 2070 2071 // If there are too many fields, or if any of the fields are aggregates, 2072 // just use the LCV as a default binding. 2073 if (Fields.size() == SmallStructLimit) 2074 return None; 2075 2076 QualType Ty = FD->getType(); 2077 if (!(Ty->isScalarType() || Ty->isReferenceType())) 2078 return None; 2079 2080 Fields.push_back(FD); 2081 } 2082 2083 RegionBindingsRef NewB = B; 2084 2085 for (FieldVector::iterator I = Fields.begin(), E = Fields.end(); I != E; ++I){ 2086 const FieldRegion *SourceFR = MRMgr.getFieldRegion(*I, LCV.getRegion()); 2087 SVal V = getBindingForField(getRegionBindings(LCV.getStore()), SourceFR); 2088 2089 const FieldRegion *DestFR = MRMgr.getFieldRegion(*I, R); 2090 NewB = bind(NewB, loc::MemRegionVal(DestFR), V); 2091 } 2092 2093 return NewB; 2094 } 2095 2096 RegionBindingsRef RegionStoreManager::bindStruct(RegionBindingsConstRef B, 2097 const TypedValueRegion* R, 2098 SVal V) { 2099 if (!Features.supportsFields()) 2100 return B; 2101 2102 QualType T = R->getValueType(); 2103 assert(T->isStructureOrClassType()); 2104 2105 const RecordType* RT = T->getAs<RecordType>(); 2106 const RecordDecl *RD = RT->getDecl(); 2107 2108 if (!RD->isCompleteDefinition()) 2109 return B; 2110 2111 // Handle lazy compound values and symbolic values. 2112 if (Optional<nonloc::LazyCompoundVal> LCV = 2113 V.getAs<nonloc::LazyCompoundVal>()) { 2114 if (Optional<RegionBindingsRef> NewB = tryBindSmallStruct(B, R, RD, *LCV)) 2115 return *NewB; 2116 return bindAggregate(B, R, V); 2117 } 2118 if (V.getAs<nonloc::SymbolVal>()) 2119 return bindAggregate(B, R, V); 2120 2121 // We may get non-CompoundVal accidentally due to imprecise cast logic or 2122 // that we are binding symbolic struct value. Kill the field values, and if 2123 // the value is symbolic go and bind it as a "default" binding. 2124 if (V.isUnknown() || !V.getAs<nonloc::CompoundVal>()) 2125 return bindAggregate(B, R, UnknownVal()); 2126 2127 const nonloc::CompoundVal& CV = V.castAs<nonloc::CompoundVal>(); 2128 nonloc::CompoundVal::iterator VI = CV.begin(), VE = CV.end(); 2129 2130 RecordDecl::field_iterator FI, FE; 2131 RegionBindingsRef NewB(B); 2132 2133 for (FI = RD->field_begin(), FE = RD->field_end(); FI != FE; ++FI) { 2134 2135 if (VI == VE) 2136 break; 2137 2138 // Skip any unnamed bitfields to stay in sync with the initializers. 2139 if (FI->isUnnamedBitfield()) 2140 continue; 2141 2142 QualType FTy = FI->getType(); 2143 const FieldRegion* FR = MRMgr.getFieldRegion(*FI, R); 2144 2145 if (FTy->isArrayType()) 2146 NewB = bindArray(NewB, FR, *VI); 2147 else if (FTy->isStructureOrClassType()) 2148 NewB = bindStruct(NewB, FR, *VI); 2149 else 2150 NewB = bind(NewB, loc::MemRegionVal(FR), *VI); 2151 ++VI; 2152 } 2153 2154 // There may be fewer values in the initialize list than the fields of struct. 2155 if (FI != FE) { 2156 NewB = NewB.addBinding(R, BindingKey::Default, 2157 svalBuilder.makeIntVal(0, false)); 2158 } 2159 2160 return NewB; 2161 } 2162 2163 RegionBindingsRef 2164 RegionStoreManager::bindAggregate(RegionBindingsConstRef B, 2165 const TypedRegion *R, 2166 SVal Val) { 2167 // Remove the old bindings, using 'R' as the root of all regions 2168 // we will invalidate. Then add the new binding. 2169 return removeSubRegionBindings(B, R).addBinding(R, BindingKey::Default, Val); 2170 } 2171 2172 //===----------------------------------------------------------------------===// 2173 // State pruning. 2174 //===----------------------------------------------------------------------===// 2175 2176 namespace { 2177 class removeDeadBindingsWorker : 2178 public ClusterAnalysis<removeDeadBindingsWorker> { 2179 SmallVector<const SymbolicRegion*, 12> Postponed; 2180 SymbolReaper &SymReaper; 2181 const StackFrameContext *CurrentLCtx; 2182 2183 public: 2184 removeDeadBindingsWorker(RegionStoreManager &rm, 2185 ProgramStateManager &stateMgr, 2186 RegionBindingsRef b, SymbolReaper &symReaper, 2187 const StackFrameContext *LCtx) 2188 : ClusterAnalysis<removeDeadBindingsWorker>(rm, stateMgr, b, GFK_None), 2189 SymReaper(symReaper), CurrentLCtx(LCtx) {} 2190 2191 // Called by ClusterAnalysis. 2192 void VisitAddedToCluster(const MemRegion *baseR, const ClusterBindings &C); 2193 void VisitCluster(const MemRegion *baseR, const ClusterBindings *C); 2194 using ClusterAnalysis<removeDeadBindingsWorker>::VisitCluster; 2195 2196 bool UpdatePostponed(); 2197 void VisitBinding(SVal V); 2198 }; 2199 } 2200 2201 void removeDeadBindingsWorker::VisitAddedToCluster(const MemRegion *baseR, 2202 const ClusterBindings &C) { 2203 2204 if (const VarRegion *VR = dyn_cast<VarRegion>(baseR)) { 2205 if (SymReaper.isLive(VR)) 2206 AddToWorkList(baseR, &C); 2207 2208 return; 2209 } 2210 2211 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(baseR)) { 2212 if (SymReaper.isLive(SR->getSymbol())) 2213 AddToWorkList(SR, &C); 2214 else 2215 Postponed.push_back(SR); 2216 2217 return; 2218 } 2219 2220 if (isa<NonStaticGlobalSpaceRegion>(baseR)) { 2221 AddToWorkList(baseR, &C); 2222 return; 2223 } 2224 2225 // CXXThisRegion in the current or parent location context is live. 2226 if (const CXXThisRegion *TR = dyn_cast<CXXThisRegion>(baseR)) { 2227 const StackArgumentsSpaceRegion *StackReg = 2228 cast<StackArgumentsSpaceRegion>(TR->getSuperRegion()); 2229 const StackFrameContext *RegCtx = StackReg->getStackFrame(); 2230 if (CurrentLCtx && 2231 (RegCtx == CurrentLCtx || RegCtx->isParentOf(CurrentLCtx))) 2232 AddToWorkList(TR, &C); 2233 } 2234 } 2235 2236 void removeDeadBindingsWorker::VisitCluster(const MemRegion *baseR, 2237 const ClusterBindings *C) { 2238 if (!C) 2239 return; 2240 2241 // Mark the symbol for any SymbolicRegion with live bindings as live itself. 2242 // This means we should continue to track that symbol. 2243 if (const SymbolicRegion *SymR = dyn_cast<SymbolicRegion>(baseR)) 2244 SymReaper.markLive(SymR->getSymbol()); 2245 2246 for (ClusterBindings::iterator I = C->begin(), E = C->end(); I != E; ++I) 2247 VisitBinding(I.getData()); 2248 } 2249 2250 void removeDeadBindingsWorker::VisitBinding(SVal V) { 2251 // Is it a LazyCompoundVal? All referenced regions are live as well. 2252 if (Optional<nonloc::LazyCompoundVal> LCS = 2253 V.getAs<nonloc::LazyCompoundVal>()) { 2254 2255 const RegionStoreManager::SValListTy &Vals = RM.getInterestingValues(*LCS); 2256 2257 for (RegionStoreManager::SValListTy::const_iterator I = Vals.begin(), 2258 E = Vals.end(); 2259 I != E; ++I) 2260 VisitBinding(*I); 2261 2262 return; 2263 } 2264 2265 // If V is a region, then add it to the worklist. 2266 if (const MemRegion *R = V.getAsRegion()) { 2267 AddToWorkList(R); 2268 2269 // All regions captured by a block are also live. 2270 if (const BlockDataRegion *BR = dyn_cast<BlockDataRegion>(R)) { 2271 BlockDataRegion::referenced_vars_iterator I = BR->referenced_vars_begin(), 2272 E = BR->referenced_vars_end(); 2273 for ( ; I != E; ++I) 2274 AddToWorkList(I.getCapturedRegion()); 2275 } 2276 } 2277 2278 2279 // Update the set of live symbols. 2280 for (SymExpr::symbol_iterator SI = V.symbol_begin(), SE = V.symbol_end(); 2281 SI!=SE; ++SI) 2282 SymReaper.markLive(*SI); 2283 } 2284 2285 bool removeDeadBindingsWorker::UpdatePostponed() { 2286 // See if any postponed SymbolicRegions are actually live now, after 2287 // having done a scan. 2288 bool changed = false; 2289 2290 for (SmallVectorImpl<const SymbolicRegion*>::iterator 2291 I = Postponed.begin(), E = Postponed.end() ; I != E ; ++I) { 2292 if (const SymbolicRegion *SR = *I) { 2293 if (SymReaper.isLive(SR->getSymbol())) { 2294 changed |= AddToWorkList(SR); 2295 *I = nullptr; 2296 } 2297 } 2298 } 2299 2300 return changed; 2301 } 2302 2303 StoreRef RegionStoreManager::removeDeadBindings(Store store, 2304 const StackFrameContext *LCtx, 2305 SymbolReaper& SymReaper) { 2306 RegionBindingsRef B = getRegionBindings(store); 2307 removeDeadBindingsWorker W(*this, StateMgr, B, SymReaper, LCtx); 2308 W.GenerateClusters(); 2309 2310 // Enqueue the region roots onto the worklist. 2311 for (SymbolReaper::region_iterator I = SymReaper.region_begin(), 2312 E = SymReaper.region_end(); I != E; ++I) { 2313 W.AddToWorkList(*I); 2314 } 2315 2316 do W.RunWorkList(); while (W.UpdatePostponed()); 2317 2318 // We have now scanned the store, marking reachable regions and symbols 2319 // as live. We now remove all the regions that are dead from the store 2320 // as well as update DSymbols with the set symbols that are now dead. 2321 for (RegionBindingsRef::iterator I = B.begin(), E = B.end(); I != E; ++I) { 2322 const MemRegion *Base = I.getKey(); 2323 2324 // If the cluster has been visited, we know the region has been marked. 2325 if (W.isVisited(Base)) 2326 continue; 2327 2328 // Remove the dead entry. 2329 B = B.remove(Base); 2330 2331 if (const SymbolicRegion *SymR = dyn_cast<SymbolicRegion>(Base)) 2332 SymReaper.maybeDead(SymR->getSymbol()); 2333 2334 // Mark all non-live symbols that this binding references as dead. 2335 const ClusterBindings &Cluster = I.getData(); 2336 for (ClusterBindings::iterator CI = Cluster.begin(), CE = Cluster.end(); 2337 CI != CE; ++CI) { 2338 SVal X = CI.getData(); 2339 SymExpr::symbol_iterator SI = X.symbol_begin(), SE = X.symbol_end(); 2340 for (; SI != SE; ++SI) 2341 SymReaper.maybeDead(*SI); 2342 } 2343 } 2344 2345 return StoreRef(B.asStore(), *this); 2346 } 2347 2348 //===----------------------------------------------------------------------===// 2349 // Utility methods. 2350 //===----------------------------------------------------------------------===// 2351 2352 void RegionStoreManager::print(Store store, raw_ostream &OS, 2353 const char* nl, const char *sep) { 2354 RegionBindingsRef B = getRegionBindings(store); 2355 OS << "Store (direct and default bindings), " 2356 << B.asStore() 2357 << " :" << nl; 2358 B.dump(OS, nl); 2359 } 2360