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 0; 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); 376 377 StoreRef getInitialStore(const LocationContext *InitLoc) { 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); 400 401 bool scanReachableSymbols(Store S, const MemRegion *R, 402 ScanReachableSymbols &Callbacks); 403 404 RegionBindingsRef removeSubRegionBindings(RegionBindingsConstRef B, 405 const SubRegion *R); 406 407 public: // Part of public interface to class. 408 409 virtual StoreRef Bind(Store store, Loc LV, SVal V) { 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) { 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 virtual StoreRef killBinding(Store ST, Loc L); 471 472 void incrementReferenceCount(Store store) { 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) { 480 getRegionBindings(store).manualRelease(); 481 } 482 483 bool includedInBindings(Store store, const MemRegion *region) const; 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 virtual SVal getBinding(Store S, Loc L, QualType T) { 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); 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, QualType EleTy); 572 573 //===------------------------------------------------------------------===// 574 // Utility methods. 575 //===------------------------------------------------------------------===// 576 577 RegionBindingsRef getRegionBindings(Store store) const { 578 return RegionBindingsRef(CBFactory, 579 static_cast<const RegionBindings::TreeTy*>(store), 580 RBFactory.getTreeFactory()); 581 } 582 583 void print(Store store, raw_ostream &Out, const char* nl, 584 const char *sep); 585 586 void iterBindings(Store store, BindingsHandler& f) { 587 RegionBindingsRef B = getRegionBindings(store); 588 for (RegionBindingsRef::iterator I = B.begin(), E = B.end(); I != E; ++I) { 589 const ClusterBindings &Cluster = I.getData(); 590 for (ClusterBindings::iterator CI = Cluster.begin(), CE = Cluster.end(); 591 CI != CE; ++CI) { 592 const BindingKey &K = CI.getKey(); 593 if (!K.isDirect()) 594 continue; 595 if (const SubRegion *R = dyn_cast<SubRegion>(K.getRegion())) { 596 // FIXME: Possibly incorporate the offset? 597 if (!f.HandleBinding(*this, store, R, CI.getData())) 598 return; 599 } 600 } 601 } 602 } 603 }; 604 605 } // end anonymous namespace 606 607 //===----------------------------------------------------------------------===// 608 // RegionStore creation. 609 //===----------------------------------------------------------------------===// 610 611 StoreManager *ento::CreateRegionStoreManager(ProgramStateManager& StMgr) { 612 RegionStoreFeatures F = maximal_features_tag(); 613 return new RegionStoreManager(StMgr, F); 614 } 615 616 StoreManager * 617 ento::CreateFieldsOnlyRegionStoreManager(ProgramStateManager &StMgr) { 618 RegionStoreFeatures F = minimal_features_tag(); 619 F.enableFields(true); 620 return new RegionStoreManager(StMgr, F); 621 } 622 623 624 //===----------------------------------------------------------------------===// 625 // Region Cluster analysis. 626 //===----------------------------------------------------------------------===// 627 628 namespace { 629 /// Used to determine which global regions are automatically included in the 630 /// initial worklist of a ClusterAnalysis. 631 enum GlobalsFilterKind { 632 /// Don't include any global regions. 633 GFK_None, 634 /// Only include system globals. 635 GFK_SystemOnly, 636 /// Include all global regions. 637 GFK_All 638 }; 639 640 template <typename DERIVED> 641 class ClusterAnalysis { 642 protected: 643 typedef llvm::DenseMap<const MemRegion *, const ClusterBindings *> ClusterMap; 644 typedef const MemRegion * WorkListElement; 645 typedef SmallVector<WorkListElement, 10> WorkList; 646 647 llvm::SmallPtrSet<const ClusterBindings *, 16> Visited; 648 649 WorkList WL; 650 651 RegionStoreManager &RM; 652 ASTContext &Ctx; 653 SValBuilder &svalBuilder; 654 655 RegionBindingsRef B; 656 657 private: 658 GlobalsFilterKind GlobalsFilter; 659 660 protected: 661 const ClusterBindings *getCluster(const MemRegion *R) { 662 return B.lookup(R); 663 } 664 665 /// Returns true if the memory space of the given region is one of the global 666 /// regions specially included at the start of analysis. 667 bool isInitiallyIncludedGlobalRegion(const MemRegion *R) { 668 switch (GlobalsFilter) { 669 case GFK_None: 670 return false; 671 case GFK_SystemOnly: 672 return isa<GlobalSystemSpaceRegion>(R->getMemorySpace()); 673 case GFK_All: 674 return isa<NonStaticGlobalSpaceRegion>(R->getMemorySpace()); 675 } 676 677 llvm_unreachable("unknown globals filter"); 678 } 679 680 public: 681 ClusterAnalysis(RegionStoreManager &rm, ProgramStateManager &StateMgr, 682 RegionBindingsRef b, GlobalsFilterKind GFK) 683 : RM(rm), Ctx(StateMgr.getContext()), 684 svalBuilder(StateMgr.getSValBuilder()), 685 B(b), GlobalsFilter(GFK) {} 686 687 RegionBindingsRef getRegionBindings() const { return B; } 688 689 bool isVisited(const MemRegion *R) { 690 return Visited.count(getCluster(R)); 691 } 692 693 void GenerateClusters() { 694 // Scan the entire set of bindings and record the region clusters. 695 for (RegionBindingsRef::iterator RI = B.begin(), RE = B.end(); 696 RI != RE; ++RI){ 697 const MemRegion *Base = RI.getKey(); 698 699 const ClusterBindings &Cluster = RI.getData(); 700 assert(!Cluster.isEmpty() && "Empty clusters should be removed"); 701 static_cast<DERIVED*>(this)->VisitAddedToCluster(Base, Cluster); 702 703 // If this is an interesting global region, add it the work list up front. 704 if (isInitiallyIncludedGlobalRegion(Base)) 705 AddToWorkList(WorkListElement(Base), &Cluster); 706 } 707 } 708 709 bool AddToWorkList(WorkListElement E, const ClusterBindings *C) { 710 if (C && !Visited.insert(C)) 711 return false; 712 WL.push_back(E); 713 return true; 714 } 715 716 bool AddToWorkList(const MemRegion *R) { 717 const MemRegion *BaseR = R->getBaseRegion(); 718 return AddToWorkList(WorkListElement(BaseR), getCluster(BaseR)); 719 } 720 721 void RunWorkList() { 722 while (!WL.empty()) { 723 WorkListElement E = WL.pop_back_val(); 724 const MemRegion *BaseR = E; 725 726 static_cast<DERIVED*>(this)->VisitCluster(BaseR, getCluster(BaseR)); 727 } 728 } 729 730 void VisitAddedToCluster(const MemRegion *baseR, const ClusterBindings &C) {} 731 void VisitCluster(const MemRegion *baseR, const ClusterBindings *C) {} 732 733 void VisitCluster(const MemRegion *BaseR, const ClusterBindings *C, 734 bool Flag) { 735 static_cast<DERIVED*>(this)->VisitCluster(BaseR, C); 736 } 737 }; 738 } 739 740 //===----------------------------------------------------------------------===// 741 // Binding invalidation. 742 //===----------------------------------------------------------------------===// 743 744 bool RegionStoreManager::scanReachableSymbols(Store S, const MemRegion *R, 745 ScanReachableSymbols &Callbacks) { 746 assert(R == R->getBaseRegion() && "Should only be called for base regions"); 747 RegionBindingsRef B = getRegionBindings(S); 748 const ClusterBindings *Cluster = B.lookup(R); 749 750 if (!Cluster) 751 return true; 752 753 for (ClusterBindings::iterator RI = Cluster->begin(), RE = Cluster->end(); 754 RI != RE; ++RI) { 755 if (!Callbacks.scan(RI.getData())) 756 return false; 757 } 758 759 return true; 760 } 761 762 static inline bool isUnionField(const FieldRegion *FR) { 763 return FR->getDecl()->getParent()->isUnion(); 764 } 765 766 typedef SmallVector<const FieldDecl *, 8> FieldVector; 767 768 void getSymbolicOffsetFields(BindingKey K, FieldVector &Fields) { 769 assert(K.hasSymbolicOffset() && "Not implemented for concrete offset keys"); 770 771 const MemRegion *Base = K.getConcreteOffsetRegion(); 772 const MemRegion *R = K.getRegion(); 773 774 while (R != Base) { 775 if (const FieldRegion *FR = dyn_cast<FieldRegion>(R)) 776 if (!isUnionField(FR)) 777 Fields.push_back(FR->getDecl()); 778 779 R = cast<SubRegion>(R)->getSuperRegion(); 780 } 781 } 782 783 static bool isCompatibleWithFields(BindingKey K, const FieldVector &Fields) { 784 assert(K.hasSymbolicOffset() && "Not implemented for concrete offset keys"); 785 786 if (Fields.empty()) 787 return true; 788 789 FieldVector FieldsInBindingKey; 790 getSymbolicOffsetFields(K, FieldsInBindingKey); 791 792 ptrdiff_t Delta = FieldsInBindingKey.size() - Fields.size(); 793 if (Delta >= 0) 794 return std::equal(FieldsInBindingKey.begin() + Delta, 795 FieldsInBindingKey.end(), 796 Fields.begin()); 797 else 798 return std::equal(FieldsInBindingKey.begin(), FieldsInBindingKey.end(), 799 Fields.begin() - Delta); 800 } 801 802 /// Collects all bindings in \p Cluster that may refer to bindings within 803 /// \p Top. 804 /// 805 /// Each binding is a pair whose \c first is the key (a BindingKey) and whose 806 /// \c second is the value (an SVal). 807 /// 808 /// The \p IncludeAllDefaultBindings parameter specifies whether to include 809 /// default bindings that may extend beyond \p Top itself, e.g. if \p Top is 810 /// an aggregate within a larger aggregate with a default binding. 811 static void 812 collectSubRegionBindings(SmallVectorImpl<BindingPair> &Bindings, 813 SValBuilder &SVB, const ClusterBindings &Cluster, 814 const SubRegion *Top, BindingKey TopKey, 815 bool IncludeAllDefaultBindings) { 816 FieldVector FieldsInSymbolicSubregions; 817 if (TopKey.hasSymbolicOffset()) { 818 getSymbolicOffsetFields(TopKey, FieldsInSymbolicSubregions); 819 Top = cast<SubRegion>(TopKey.getConcreteOffsetRegion()); 820 TopKey = BindingKey::Make(Top, BindingKey::Default); 821 } 822 823 // Find the length (in bits) of the region being invalidated. 824 uint64_t Length = UINT64_MAX; 825 SVal Extent = Top->getExtent(SVB); 826 if (Optional<nonloc::ConcreteInt> ExtentCI = 827 Extent.getAs<nonloc::ConcreteInt>()) { 828 const llvm::APSInt &ExtentInt = ExtentCI->getValue(); 829 assert(ExtentInt.isNonNegative() || ExtentInt.isUnsigned()); 830 // Extents are in bytes but region offsets are in bits. Be careful! 831 Length = ExtentInt.getLimitedValue() * SVB.getContext().getCharWidth(); 832 } else if (const FieldRegion *FR = dyn_cast<FieldRegion>(Top)) { 833 if (FR->getDecl()->isBitField()) 834 Length = FR->getDecl()->getBitWidthValue(SVB.getContext()); 835 } 836 837 for (ClusterBindings::iterator I = Cluster.begin(), E = Cluster.end(); 838 I != E; ++I) { 839 BindingKey NextKey = I.getKey(); 840 if (NextKey.getRegion() == TopKey.getRegion()) { 841 // FIXME: This doesn't catch the case where we're really invalidating a 842 // region with a symbolic offset. Example: 843 // R: points[i].y 844 // Next: points[0].x 845 846 if (NextKey.getOffset() > TopKey.getOffset() && 847 NextKey.getOffset() - TopKey.getOffset() < Length) { 848 // Case 1: The next binding is inside the region we're invalidating. 849 // Include it. 850 Bindings.push_back(*I); 851 852 } else if (NextKey.getOffset() == TopKey.getOffset()) { 853 // Case 2: The next binding is at the same offset as the region we're 854 // invalidating. In this case, we need to leave default bindings alone, 855 // since they may be providing a default value for a regions beyond what 856 // we're invalidating. 857 // FIXME: This is probably incorrect; consider invalidating an outer 858 // struct whose first field is bound to a LazyCompoundVal. 859 if (IncludeAllDefaultBindings || NextKey.isDirect()) 860 Bindings.push_back(*I); 861 } 862 863 } else if (NextKey.hasSymbolicOffset()) { 864 const MemRegion *Base = NextKey.getConcreteOffsetRegion(); 865 if (Top->isSubRegionOf(Base)) { 866 // Case 3: The next key is symbolic and we just changed something within 867 // its concrete region. We don't know if the binding is still valid, so 868 // we'll be conservative and include it. 869 if (IncludeAllDefaultBindings || NextKey.isDirect()) 870 if (isCompatibleWithFields(NextKey, FieldsInSymbolicSubregions)) 871 Bindings.push_back(*I); 872 } else if (const SubRegion *BaseSR = dyn_cast<SubRegion>(Base)) { 873 // Case 4: The next key is symbolic, but we changed a known 874 // super-region. In this case the binding is certainly included. 875 if (Top == Base || BaseSR->isSubRegionOf(Top)) 876 if (isCompatibleWithFields(NextKey, FieldsInSymbolicSubregions)) 877 Bindings.push_back(*I); 878 } 879 } 880 } 881 } 882 883 static void 884 collectSubRegionBindings(SmallVectorImpl<BindingPair> &Bindings, 885 SValBuilder &SVB, const ClusterBindings &Cluster, 886 const SubRegion *Top, bool IncludeAllDefaultBindings) { 887 collectSubRegionBindings(Bindings, SVB, Cluster, Top, 888 BindingKey::Make(Top, BindingKey::Default), 889 IncludeAllDefaultBindings); 890 } 891 892 RegionBindingsRef 893 RegionStoreManager::removeSubRegionBindings(RegionBindingsConstRef B, 894 const SubRegion *Top) { 895 BindingKey TopKey = BindingKey::Make(Top, BindingKey::Default); 896 const MemRegion *ClusterHead = TopKey.getBaseRegion(); 897 898 if (Top == ClusterHead) { 899 // We can remove an entire cluster's bindings all in one go. 900 return B.remove(Top); 901 } 902 903 const ClusterBindings *Cluster = B.lookup(ClusterHead); 904 if (!Cluster) { 905 // If we're invalidating a region with a symbolic offset, we need to make 906 // sure we don't treat the base region as uninitialized anymore. 907 if (TopKey.hasSymbolicOffset()) { 908 const SubRegion *Concrete = TopKey.getConcreteOffsetRegion(); 909 return B.addBinding(Concrete, BindingKey::Default, UnknownVal()); 910 } 911 return B; 912 } 913 914 SmallVector<BindingPair, 32> Bindings; 915 collectSubRegionBindings(Bindings, svalBuilder, *Cluster, Top, TopKey, 916 /*IncludeAllDefaultBindings=*/false); 917 918 ClusterBindingsRef Result(*Cluster, CBFactory); 919 for (SmallVectorImpl<BindingPair>::const_iterator I = Bindings.begin(), 920 E = Bindings.end(); 921 I != E; ++I) 922 Result = Result.remove(I->first); 923 924 // If we're invalidating a region with a symbolic offset, we need to make sure 925 // we don't treat the base region as uninitialized anymore. 926 // FIXME: This isn't very precise; see the example in 927 // collectSubRegionBindings. 928 if (TopKey.hasSymbolicOffset()) { 929 const SubRegion *Concrete = TopKey.getConcreteOffsetRegion(); 930 Result = Result.add(BindingKey::Make(Concrete, BindingKey::Default), 931 UnknownVal()); 932 } 933 934 if (Result.isEmpty()) 935 return B.remove(ClusterHead); 936 return B.add(ClusterHead, Result.asImmutableMap()); 937 } 938 939 namespace { 940 class invalidateRegionsWorker : public ClusterAnalysis<invalidateRegionsWorker> 941 { 942 const Expr *Ex; 943 unsigned Count; 944 const LocationContext *LCtx; 945 InvalidatedSymbols &IS; 946 RegionAndSymbolInvalidationTraits &ITraits; 947 StoreManager::InvalidatedRegions *Regions; 948 public: 949 invalidateRegionsWorker(RegionStoreManager &rm, 950 ProgramStateManager &stateMgr, 951 RegionBindingsRef b, 952 const Expr *ex, unsigned count, 953 const LocationContext *lctx, 954 InvalidatedSymbols &is, 955 RegionAndSymbolInvalidationTraits &ITraitsIn, 956 StoreManager::InvalidatedRegions *r, 957 GlobalsFilterKind GFK) 958 : ClusterAnalysis<invalidateRegionsWorker>(rm, stateMgr, b, GFK), 959 Ex(ex), Count(count), LCtx(lctx), IS(is), ITraits(ITraitsIn), Regions(r){} 960 961 void VisitCluster(const MemRegion *baseR, const ClusterBindings *C); 962 void VisitBinding(SVal V); 963 }; 964 } 965 966 void invalidateRegionsWorker::VisitBinding(SVal V) { 967 // A symbol? Mark it touched by the invalidation. 968 if (SymbolRef Sym = V.getAsSymbol()) 969 IS.insert(Sym); 970 971 if (const MemRegion *R = V.getAsRegion()) { 972 AddToWorkList(R); 973 return; 974 } 975 976 // Is it a LazyCompoundVal? All references get invalidated as well. 977 if (Optional<nonloc::LazyCompoundVal> LCS = 978 V.getAs<nonloc::LazyCompoundVal>()) { 979 980 const RegionStoreManager::SValListTy &Vals = RM.getInterestingValues(*LCS); 981 982 for (RegionStoreManager::SValListTy::const_iterator I = Vals.begin(), 983 E = Vals.end(); 984 I != E; ++I) 985 VisitBinding(*I); 986 987 return; 988 } 989 } 990 991 void invalidateRegionsWorker::VisitCluster(const MemRegion *baseR, 992 const ClusterBindings *C) { 993 994 bool PreserveRegionsContents = 995 ITraits.hasTrait(baseR, 996 RegionAndSymbolInvalidationTraits::TK_PreserveContents); 997 998 if (C) { 999 for (ClusterBindings::iterator I = C->begin(), E = C->end(); I != E; ++I) 1000 VisitBinding(I.getData()); 1001 1002 // Invalidate regions contents. 1003 if (!PreserveRegionsContents) 1004 B = B.remove(baseR); 1005 } 1006 1007 // BlockDataRegion? If so, invalidate captured variables that are passed 1008 // by reference. 1009 if (const BlockDataRegion *BR = dyn_cast<BlockDataRegion>(baseR)) { 1010 for (BlockDataRegion::referenced_vars_iterator 1011 BI = BR->referenced_vars_begin(), BE = BR->referenced_vars_end() ; 1012 BI != BE; ++BI) { 1013 const VarRegion *VR = BI.getCapturedRegion(); 1014 const VarDecl *VD = VR->getDecl(); 1015 if (VD->hasAttr<BlocksAttr>() || !VD->hasLocalStorage()) { 1016 AddToWorkList(VR); 1017 } 1018 else if (Loc::isLocType(VR->getValueType())) { 1019 // Map the current bindings to a Store to retrieve the value 1020 // of the binding. If that binding itself is a region, we should 1021 // invalidate that region. This is because a block may capture 1022 // a pointer value, but the thing pointed by that pointer may 1023 // get invalidated. 1024 SVal V = RM.getBinding(B, loc::MemRegionVal(VR)); 1025 if (Optional<Loc> L = V.getAs<Loc>()) { 1026 if (const MemRegion *LR = L->getAsRegion()) 1027 AddToWorkList(LR); 1028 } 1029 } 1030 } 1031 return; 1032 } 1033 1034 // Symbolic region? 1035 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(baseR)) 1036 IS.insert(SR->getSymbol()); 1037 1038 // Nothing else should be done in the case when we preserve regions context. 1039 if (PreserveRegionsContents) 1040 return; 1041 1042 // Otherwise, we have a normal data region. Record that we touched the region. 1043 if (Regions) 1044 Regions->push_back(baseR); 1045 1046 if (isa<AllocaRegion>(baseR) || isa<SymbolicRegion>(baseR)) { 1047 // Invalidate the region by setting its default value to 1048 // conjured symbol. The type of the symbol is irrelevant. 1049 DefinedOrUnknownSVal V = 1050 svalBuilder.conjureSymbolVal(baseR, Ex, LCtx, Ctx.IntTy, Count); 1051 B = B.addBinding(baseR, BindingKey::Default, V); 1052 return; 1053 } 1054 1055 if (!baseR->isBoundable()) 1056 return; 1057 1058 const TypedValueRegion *TR = cast<TypedValueRegion>(baseR); 1059 QualType T = TR->getValueType(); 1060 1061 if (isInitiallyIncludedGlobalRegion(baseR)) { 1062 // If the region is a global and we are invalidating all globals, 1063 // erasing the entry is good enough. This causes all globals to be lazily 1064 // symbolicated from the same base symbol. 1065 return; 1066 } 1067 1068 if (T->isStructureOrClassType()) { 1069 // Invalidate the region by setting its default value to 1070 // conjured symbol. The type of the symbol is irrelevant. 1071 DefinedOrUnknownSVal V = svalBuilder.conjureSymbolVal(baseR, Ex, LCtx, 1072 Ctx.IntTy, Count); 1073 B = B.addBinding(baseR, BindingKey::Default, V); 1074 return; 1075 } 1076 1077 if (const ArrayType *AT = Ctx.getAsArrayType(T)) { 1078 // Set the default value of the array to conjured symbol. 1079 DefinedOrUnknownSVal V = 1080 svalBuilder.conjureSymbolVal(baseR, Ex, LCtx, 1081 AT->getElementType(), Count); 1082 B = B.addBinding(baseR, BindingKey::Default, V); 1083 return; 1084 } 1085 1086 DefinedOrUnknownSVal V = svalBuilder.conjureSymbolVal(baseR, Ex, LCtx, 1087 T,Count); 1088 assert(SymbolManager::canSymbolicate(T) || V.isUnknown()); 1089 B = B.addBinding(baseR, BindingKey::Direct, V); 1090 } 1091 1092 RegionBindingsRef 1093 RegionStoreManager::invalidateGlobalRegion(MemRegion::Kind K, 1094 const Expr *Ex, 1095 unsigned Count, 1096 const LocationContext *LCtx, 1097 RegionBindingsRef B, 1098 InvalidatedRegions *Invalidated) { 1099 // Bind the globals memory space to a new symbol that we will use to derive 1100 // the bindings for all globals. 1101 const GlobalsSpaceRegion *GS = MRMgr.getGlobalsRegion(K); 1102 SVal V = svalBuilder.conjureSymbolVal(/* SymbolTag = */ (const void*) GS, Ex, LCtx, 1103 /* type does not matter */ Ctx.IntTy, 1104 Count); 1105 1106 B = B.removeBinding(GS) 1107 .addBinding(BindingKey::Make(GS, BindingKey::Default), V); 1108 1109 // Even if there are no bindings in the global scope, we still need to 1110 // record that we touched it. 1111 if (Invalidated) 1112 Invalidated->push_back(GS); 1113 1114 return B; 1115 } 1116 1117 void RegionStoreManager::populateWorkList(invalidateRegionsWorker &W, 1118 ArrayRef<SVal> Values, 1119 InvalidatedRegions *TopLevelRegions) { 1120 for (ArrayRef<SVal>::iterator I = Values.begin(), 1121 E = Values.end(); I != E; ++I) { 1122 SVal V = *I; 1123 if (Optional<nonloc::LazyCompoundVal> LCS = 1124 V.getAs<nonloc::LazyCompoundVal>()) { 1125 1126 const SValListTy &Vals = getInterestingValues(*LCS); 1127 1128 for (SValListTy::const_iterator I = Vals.begin(), 1129 E = Vals.end(); I != E; ++I) { 1130 // Note: the last argument is false here because these are 1131 // non-top-level regions. 1132 if (const MemRegion *R = (*I).getAsRegion()) 1133 W.AddToWorkList(R); 1134 } 1135 continue; 1136 } 1137 1138 if (const MemRegion *R = V.getAsRegion()) { 1139 if (TopLevelRegions) 1140 TopLevelRegions->push_back(R); 1141 W.AddToWorkList(R); 1142 continue; 1143 } 1144 } 1145 } 1146 1147 StoreRef 1148 RegionStoreManager::invalidateRegions(Store store, 1149 ArrayRef<SVal> Values, 1150 const Expr *Ex, unsigned Count, 1151 const LocationContext *LCtx, 1152 const CallEvent *Call, 1153 InvalidatedSymbols &IS, 1154 RegionAndSymbolInvalidationTraits &ITraits, 1155 InvalidatedRegions *TopLevelRegions, 1156 InvalidatedRegions *Invalidated) { 1157 GlobalsFilterKind GlobalsFilter; 1158 if (Call) { 1159 if (Call->isInSystemHeader()) 1160 GlobalsFilter = GFK_SystemOnly; 1161 else 1162 GlobalsFilter = GFK_All; 1163 } else { 1164 GlobalsFilter = GFK_None; 1165 } 1166 1167 RegionBindingsRef B = getRegionBindings(store); 1168 invalidateRegionsWorker W(*this, StateMgr, B, Ex, Count, LCtx, IS, ITraits, 1169 Invalidated, GlobalsFilter); 1170 1171 // Scan the bindings and generate the clusters. 1172 W.GenerateClusters(); 1173 1174 // Add the regions to the worklist. 1175 populateWorkList(W, Values, TopLevelRegions); 1176 1177 W.RunWorkList(); 1178 1179 // Return the new bindings. 1180 B = W.getRegionBindings(); 1181 1182 // For calls, determine which global regions should be invalidated and 1183 // invalidate them. (Note that function-static and immutable globals are never 1184 // invalidated by this.) 1185 // TODO: This could possibly be more precise with modules. 1186 switch (GlobalsFilter) { 1187 case GFK_All: 1188 B = invalidateGlobalRegion(MemRegion::GlobalInternalSpaceRegionKind, 1189 Ex, Count, LCtx, B, Invalidated); 1190 // FALLTHROUGH 1191 case GFK_SystemOnly: 1192 B = invalidateGlobalRegion(MemRegion::GlobalSystemSpaceRegionKind, 1193 Ex, Count, LCtx, B, Invalidated); 1194 // FALLTHROUGH 1195 case GFK_None: 1196 break; 1197 } 1198 1199 return StoreRef(B.asStore(), *this); 1200 } 1201 1202 //===----------------------------------------------------------------------===// 1203 // Extents for regions. 1204 //===----------------------------------------------------------------------===// 1205 1206 DefinedOrUnknownSVal 1207 RegionStoreManager::getSizeInElements(ProgramStateRef state, 1208 const MemRegion *R, 1209 QualType EleTy) { 1210 SVal Size = cast<SubRegion>(R)->getExtent(svalBuilder); 1211 const llvm::APSInt *SizeInt = svalBuilder.getKnownValue(state, Size); 1212 if (!SizeInt) 1213 return UnknownVal(); 1214 1215 CharUnits RegionSize = CharUnits::fromQuantity(SizeInt->getSExtValue()); 1216 1217 if (Ctx.getAsVariableArrayType(EleTy)) { 1218 // FIXME: We need to track extra state to properly record the size 1219 // of VLAs. Returning UnknownVal here, however, is a stop-gap so that 1220 // we don't have a divide-by-zero below. 1221 return UnknownVal(); 1222 } 1223 1224 CharUnits EleSize = Ctx.getTypeSizeInChars(EleTy); 1225 1226 // If a variable is reinterpreted as a type that doesn't fit into a larger 1227 // type evenly, round it down. 1228 // This is a signed value, since it's used in arithmetic with signed indices. 1229 return svalBuilder.makeIntVal(RegionSize / EleSize, false); 1230 } 1231 1232 //===----------------------------------------------------------------------===// 1233 // Location and region casting. 1234 //===----------------------------------------------------------------------===// 1235 1236 /// ArrayToPointer - Emulates the "decay" of an array to a pointer 1237 /// type. 'Array' represents the lvalue of the array being decayed 1238 /// to a pointer, and the returned SVal represents the decayed 1239 /// version of that lvalue (i.e., a pointer to the first element of 1240 /// the array). This is called by ExprEngine when evaluating casts 1241 /// from arrays to pointers. 1242 SVal RegionStoreManager::ArrayToPointer(Loc Array, QualType T) { 1243 if (!Array.getAs<loc::MemRegionVal>()) 1244 return UnknownVal(); 1245 1246 const MemRegion* R = Array.castAs<loc::MemRegionVal>().getRegion(); 1247 NonLoc ZeroIdx = svalBuilder.makeZeroArrayIndex(); 1248 return loc::MemRegionVal(MRMgr.getElementRegion(T, ZeroIdx, R, Ctx)); 1249 } 1250 1251 //===----------------------------------------------------------------------===// 1252 // Loading values from regions. 1253 //===----------------------------------------------------------------------===// 1254 1255 SVal RegionStoreManager::getBinding(RegionBindingsConstRef B, Loc L, QualType T) { 1256 assert(!L.getAs<UnknownVal>() && "location unknown"); 1257 assert(!L.getAs<UndefinedVal>() && "location undefined"); 1258 1259 // For access to concrete addresses, return UnknownVal. Checks 1260 // for null dereferences (and similar errors) are done by checkers, not 1261 // the Store. 1262 // FIXME: We can consider lazily symbolicating such memory, but we really 1263 // should defer this when we can reason easily about symbolicating arrays 1264 // of bytes. 1265 if (L.getAs<loc::ConcreteInt>()) { 1266 return UnknownVal(); 1267 } 1268 if (!L.getAs<loc::MemRegionVal>()) { 1269 return UnknownVal(); 1270 } 1271 1272 const MemRegion *MR = L.castAs<loc::MemRegionVal>().getRegion(); 1273 1274 if (isa<AllocaRegion>(MR) || 1275 isa<SymbolicRegion>(MR) || 1276 isa<CodeTextRegion>(MR)) { 1277 if (T.isNull()) { 1278 if (const TypedRegion *TR = dyn_cast<TypedRegion>(MR)) 1279 T = TR->getLocationType(); 1280 else { 1281 const SymbolicRegion *SR = cast<SymbolicRegion>(MR); 1282 T = SR->getSymbol()->getType(); 1283 } 1284 } 1285 MR = GetElementZeroRegion(MR, T); 1286 } 1287 1288 // FIXME: Perhaps this method should just take a 'const MemRegion*' argument 1289 // instead of 'Loc', and have the other Loc cases handled at a higher level. 1290 const TypedValueRegion *R = cast<TypedValueRegion>(MR); 1291 QualType RTy = R->getValueType(); 1292 1293 // FIXME: we do not yet model the parts of a complex type, so treat the 1294 // whole thing as "unknown". 1295 if (RTy->isAnyComplexType()) 1296 return UnknownVal(); 1297 1298 // FIXME: We should eventually handle funny addressing. e.g.: 1299 // 1300 // int x = ...; 1301 // int *p = &x; 1302 // char *q = (char*) p; 1303 // char c = *q; // returns the first byte of 'x'. 1304 // 1305 // Such funny addressing will occur due to layering of regions. 1306 if (RTy->isStructureOrClassType()) 1307 return getBindingForStruct(B, R); 1308 1309 // FIXME: Handle unions. 1310 if (RTy->isUnionType()) 1311 return createLazyBinding(B, R); 1312 1313 if (RTy->isArrayType()) { 1314 if (RTy->isConstantArrayType()) 1315 return getBindingForArray(B, R); 1316 else 1317 return UnknownVal(); 1318 } 1319 1320 // FIXME: handle Vector types. 1321 if (RTy->isVectorType()) 1322 return UnknownVal(); 1323 1324 if (const FieldRegion* FR = dyn_cast<FieldRegion>(R)) 1325 return CastRetrievedVal(getBindingForField(B, FR), FR, T, false); 1326 1327 if (const ElementRegion* ER = dyn_cast<ElementRegion>(R)) { 1328 // FIXME: Here we actually perform an implicit conversion from the loaded 1329 // value to the element type. Eventually we want to compose these values 1330 // more intelligently. For example, an 'element' can encompass multiple 1331 // bound regions (e.g., several bound bytes), or could be a subset of 1332 // a larger value. 1333 return CastRetrievedVal(getBindingForElement(B, ER), ER, T, false); 1334 } 1335 1336 if (const ObjCIvarRegion *IVR = dyn_cast<ObjCIvarRegion>(R)) { 1337 // FIXME: Here we actually perform an implicit conversion from the loaded 1338 // value to the ivar type. What we should model is stores to ivars 1339 // that blow past the extent of the ivar. If the address of the ivar is 1340 // reinterpretted, it is possible we stored a different value that could 1341 // fit within the ivar. Either we need to cast these when storing them 1342 // or reinterpret them lazily (as we do here). 1343 return CastRetrievedVal(getBindingForObjCIvar(B, IVR), IVR, T, false); 1344 } 1345 1346 if (const VarRegion *VR = dyn_cast<VarRegion>(R)) { 1347 // FIXME: Here we actually perform an implicit conversion from the loaded 1348 // value to the variable type. What we should model is stores to variables 1349 // that blow past the extent of the variable. If the address of the 1350 // variable is reinterpretted, it is possible we stored a different value 1351 // that could fit within the variable. Either we need to cast these when 1352 // storing them or reinterpret them lazily (as we do here). 1353 return CastRetrievedVal(getBindingForVar(B, VR), VR, T, false); 1354 } 1355 1356 const SVal *V = B.lookup(R, BindingKey::Direct); 1357 1358 // Check if the region has a binding. 1359 if (V) 1360 return *V; 1361 1362 // The location does not have a bound value. This means that it has 1363 // the value it had upon its creation and/or entry to the analyzed 1364 // function/method. These are either symbolic values or 'undefined'. 1365 if (R->hasStackNonParametersStorage()) { 1366 // All stack variables are considered to have undefined values 1367 // upon creation. All heap allocated blocks are considered to 1368 // have undefined values as well unless they are explicitly bound 1369 // to specific values. 1370 return UndefinedVal(); 1371 } 1372 1373 // All other values are symbolic. 1374 return svalBuilder.getRegionValueSymbolVal(R); 1375 } 1376 1377 static QualType getUnderlyingType(const SubRegion *R) { 1378 QualType RegionTy; 1379 if (const TypedValueRegion *TVR = dyn_cast<TypedValueRegion>(R)) 1380 RegionTy = TVR->getValueType(); 1381 1382 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(R)) 1383 RegionTy = SR->getSymbol()->getType(); 1384 1385 return RegionTy; 1386 } 1387 1388 /// Checks to see if store \p B has a lazy binding for region \p R. 1389 /// 1390 /// If \p AllowSubregionBindings is \c false, a lazy binding will be rejected 1391 /// if there are additional bindings within \p R. 1392 /// 1393 /// Note that unlike RegionStoreManager::findLazyBinding, this will not search 1394 /// for lazy bindings for super-regions of \p R. 1395 static Optional<nonloc::LazyCompoundVal> 1396 getExistingLazyBinding(SValBuilder &SVB, RegionBindingsConstRef B, 1397 const SubRegion *R, bool AllowSubregionBindings) { 1398 Optional<SVal> V = B.getDefaultBinding(R); 1399 if (!V) 1400 return None; 1401 1402 Optional<nonloc::LazyCompoundVal> LCV = V->getAs<nonloc::LazyCompoundVal>(); 1403 if (!LCV) 1404 return None; 1405 1406 // If the LCV is for a subregion, the types might not match, and we shouldn't 1407 // reuse the binding. 1408 QualType RegionTy = getUnderlyingType(R); 1409 if (!RegionTy.isNull() && 1410 !RegionTy->isVoidPointerType()) { 1411 QualType SourceRegionTy = LCV->getRegion()->getValueType(); 1412 if (!SVB.getContext().hasSameUnqualifiedType(RegionTy, SourceRegionTy)) 1413 return None; 1414 } 1415 1416 if (!AllowSubregionBindings) { 1417 // If there are any other bindings within this region, we shouldn't reuse 1418 // the top-level binding. 1419 SmallVector<BindingPair, 16> Bindings; 1420 collectSubRegionBindings(Bindings, SVB, *B.lookup(R->getBaseRegion()), R, 1421 /*IncludeAllDefaultBindings=*/true); 1422 if (Bindings.size() > 1) 1423 return None; 1424 } 1425 1426 return *LCV; 1427 } 1428 1429 1430 std::pair<Store, const SubRegion *> 1431 RegionStoreManager::findLazyBinding(RegionBindingsConstRef B, 1432 const SubRegion *R, 1433 const SubRegion *originalRegion) { 1434 if (originalRegion != R) { 1435 if (Optional<nonloc::LazyCompoundVal> V = 1436 getExistingLazyBinding(svalBuilder, B, R, true)) 1437 return std::make_pair(V->getStore(), V->getRegion()); 1438 } 1439 1440 typedef std::pair<Store, const SubRegion *> StoreRegionPair; 1441 StoreRegionPair Result = StoreRegionPair(); 1442 1443 if (const ElementRegion *ER = dyn_cast<ElementRegion>(R)) { 1444 Result = findLazyBinding(B, cast<SubRegion>(ER->getSuperRegion()), 1445 originalRegion); 1446 1447 if (Result.second) 1448 Result.second = MRMgr.getElementRegionWithSuper(ER, Result.second); 1449 1450 } else if (const FieldRegion *FR = dyn_cast<FieldRegion>(R)) { 1451 Result = findLazyBinding(B, cast<SubRegion>(FR->getSuperRegion()), 1452 originalRegion); 1453 1454 if (Result.second) 1455 Result.second = MRMgr.getFieldRegionWithSuper(FR, Result.second); 1456 1457 } else if (const CXXBaseObjectRegion *BaseReg = 1458 dyn_cast<CXXBaseObjectRegion>(R)) { 1459 // C++ base object region is another kind of region that we should blast 1460 // through to look for lazy compound value. It is like a field region. 1461 Result = findLazyBinding(B, cast<SubRegion>(BaseReg->getSuperRegion()), 1462 originalRegion); 1463 1464 if (Result.second) 1465 Result.second = MRMgr.getCXXBaseObjectRegionWithSuper(BaseReg, 1466 Result.second); 1467 } 1468 1469 return Result; 1470 } 1471 1472 SVal RegionStoreManager::getBindingForElement(RegionBindingsConstRef B, 1473 const ElementRegion* R) { 1474 // We do not currently model bindings of the CompoundLiteralregion. 1475 if (isa<CompoundLiteralRegion>(R->getBaseRegion())) 1476 return UnknownVal(); 1477 1478 // Check if the region has a binding. 1479 if (const Optional<SVal> &V = B.getDirectBinding(R)) 1480 return *V; 1481 1482 const MemRegion* superR = R->getSuperRegion(); 1483 1484 // Check if the region is an element region of a string literal. 1485 if (const StringRegion *StrR=dyn_cast<StringRegion>(superR)) { 1486 // FIXME: Handle loads from strings where the literal is treated as 1487 // an integer, e.g., *((unsigned int*)"hello") 1488 QualType T = Ctx.getAsArrayType(StrR->getValueType())->getElementType(); 1489 if (!Ctx.hasSameUnqualifiedType(T, R->getElementType())) 1490 return UnknownVal(); 1491 1492 const StringLiteral *Str = StrR->getStringLiteral(); 1493 SVal Idx = R->getIndex(); 1494 if (Optional<nonloc::ConcreteInt> CI = Idx.getAs<nonloc::ConcreteInt>()) { 1495 int64_t i = CI->getValue().getSExtValue(); 1496 // Abort on string underrun. This can be possible by arbitrary 1497 // clients of getBindingForElement(). 1498 if (i < 0) 1499 return UndefinedVal(); 1500 int64_t length = Str->getLength(); 1501 // Technically, only i == length is guaranteed to be null. 1502 // However, such overflows should be caught before reaching this point; 1503 // the only time such an access would be made is if a string literal was 1504 // used to initialize a larger array. 1505 char c = (i >= length) ? '\0' : Str->getCodeUnit(i); 1506 return svalBuilder.makeIntVal(c, T); 1507 } 1508 } 1509 1510 // Check for loads from a code text region. For such loads, just give up. 1511 if (isa<CodeTextRegion>(superR)) 1512 return UnknownVal(); 1513 1514 // Handle the case where we are indexing into a larger scalar object. 1515 // For example, this handles: 1516 // int x = ... 1517 // char *y = &x; 1518 // return *y; 1519 // FIXME: This is a hack, and doesn't do anything really intelligent yet. 1520 const RegionRawOffset &O = R->getAsArrayOffset(); 1521 1522 // If we cannot reason about the offset, return an unknown value. 1523 if (!O.getRegion()) 1524 return UnknownVal(); 1525 1526 if (const TypedValueRegion *baseR = 1527 dyn_cast_or_null<TypedValueRegion>(O.getRegion())) { 1528 QualType baseT = baseR->getValueType(); 1529 if (baseT->isScalarType()) { 1530 QualType elemT = R->getElementType(); 1531 if (elemT->isScalarType()) { 1532 if (Ctx.getTypeSizeInChars(baseT) >= Ctx.getTypeSizeInChars(elemT)) { 1533 if (const Optional<SVal> &V = B.getDirectBinding(superR)) { 1534 if (SymbolRef parentSym = V->getAsSymbol()) 1535 return svalBuilder.getDerivedRegionValueSymbolVal(parentSym, R); 1536 1537 if (V->isUnknownOrUndef()) 1538 return *V; 1539 // Other cases: give up. We are indexing into a larger object 1540 // that has some value, but we don't know how to handle that yet. 1541 return UnknownVal(); 1542 } 1543 } 1544 } 1545 } 1546 } 1547 return getBindingForFieldOrElementCommon(B, R, R->getElementType()); 1548 } 1549 1550 SVal RegionStoreManager::getBindingForField(RegionBindingsConstRef B, 1551 const FieldRegion* R) { 1552 1553 // Check if the region has a binding. 1554 if (const Optional<SVal> &V = B.getDirectBinding(R)) 1555 return *V; 1556 1557 QualType Ty = R->getValueType(); 1558 return getBindingForFieldOrElementCommon(B, R, Ty); 1559 } 1560 1561 Optional<SVal> 1562 RegionStoreManager::getBindingForDerivedDefaultValue(RegionBindingsConstRef B, 1563 const MemRegion *superR, 1564 const TypedValueRegion *R, 1565 QualType Ty) { 1566 1567 if (const Optional<SVal> &D = B.getDefaultBinding(superR)) { 1568 const SVal &val = D.getValue(); 1569 if (SymbolRef parentSym = val.getAsSymbol()) 1570 return svalBuilder.getDerivedRegionValueSymbolVal(parentSym, R); 1571 1572 if (val.isZeroConstant()) 1573 return svalBuilder.makeZeroVal(Ty); 1574 1575 if (val.isUnknownOrUndef()) 1576 return val; 1577 1578 // Lazy bindings are usually handled through getExistingLazyBinding(). 1579 // We should unify these two code paths at some point. 1580 if (val.getAs<nonloc::LazyCompoundVal>()) 1581 return val; 1582 1583 llvm_unreachable("Unknown default value"); 1584 } 1585 1586 return None; 1587 } 1588 1589 SVal RegionStoreManager::getLazyBinding(const SubRegion *LazyBindingRegion, 1590 RegionBindingsRef LazyBinding) { 1591 SVal Result; 1592 if (const ElementRegion *ER = dyn_cast<ElementRegion>(LazyBindingRegion)) 1593 Result = getBindingForElement(LazyBinding, ER); 1594 else 1595 Result = getBindingForField(LazyBinding, 1596 cast<FieldRegion>(LazyBindingRegion)); 1597 1598 // FIXME: This is a hack to deal with RegionStore's inability to distinguish a 1599 // default value for /part/ of an aggregate from a default value for the 1600 // /entire/ aggregate. The most common case of this is when struct Outer 1601 // has as its first member a struct Inner, which is copied in from a stack 1602 // variable. In this case, even if the Outer's default value is symbolic, 0, 1603 // or unknown, it gets overridden by the Inner's default value of undefined. 1604 // 1605 // This is a general problem -- if the Inner is zero-initialized, the Outer 1606 // will now look zero-initialized. The proper way to solve this is with a 1607 // new version of RegionStore that tracks the extent of a binding as well 1608 // as the offset. 1609 // 1610 // This hack only takes care of the undefined case because that can very 1611 // quickly result in a warning. 1612 if (Result.isUndef()) 1613 Result = UnknownVal(); 1614 1615 return Result; 1616 } 1617 1618 SVal 1619 RegionStoreManager::getBindingForFieldOrElementCommon(RegionBindingsConstRef B, 1620 const TypedValueRegion *R, 1621 QualType Ty) { 1622 1623 // At this point we have already checked in either getBindingForElement or 1624 // getBindingForField if 'R' has a direct binding. 1625 1626 // Lazy binding? 1627 Store lazyBindingStore = NULL; 1628 const SubRegion *lazyBindingRegion = NULL; 1629 llvm::tie(lazyBindingStore, lazyBindingRegion) = findLazyBinding(B, R, R); 1630 if (lazyBindingRegion) 1631 return getLazyBinding(lazyBindingRegion, 1632 getRegionBindings(lazyBindingStore)); 1633 1634 // Record whether or not we see a symbolic index. That can completely 1635 // be out of scope of our lookup. 1636 bool hasSymbolicIndex = false; 1637 1638 // FIXME: This is a hack to deal with RegionStore's inability to distinguish a 1639 // default value for /part/ of an aggregate from a default value for the 1640 // /entire/ aggregate. The most common case of this is when struct Outer 1641 // has as its first member a struct Inner, which is copied in from a stack 1642 // variable. In this case, even if the Outer's default value is symbolic, 0, 1643 // or unknown, it gets overridden by the Inner's default value of undefined. 1644 // 1645 // This is a general problem -- if the Inner is zero-initialized, the Outer 1646 // will now look zero-initialized. The proper way to solve this is with a 1647 // new version of RegionStore that tracks the extent of a binding as well 1648 // as the offset. 1649 // 1650 // This hack only takes care of the undefined case because that can very 1651 // quickly result in a warning. 1652 bool hasPartialLazyBinding = false; 1653 1654 const SubRegion *SR = dyn_cast<SubRegion>(R); 1655 while (SR) { 1656 const MemRegion *Base = SR->getSuperRegion(); 1657 if (Optional<SVal> D = getBindingForDerivedDefaultValue(B, Base, R, Ty)) { 1658 if (D->getAs<nonloc::LazyCompoundVal>()) { 1659 hasPartialLazyBinding = true; 1660 break; 1661 } 1662 1663 return *D; 1664 } 1665 1666 if (const ElementRegion *ER = dyn_cast<ElementRegion>(Base)) { 1667 NonLoc index = ER->getIndex(); 1668 if (!index.isConstant()) 1669 hasSymbolicIndex = true; 1670 } 1671 1672 // If our super region is a field or element itself, walk up the region 1673 // hierarchy to see if there is a default value installed in an ancestor. 1674 SR = dyn_cast<SubRegion>(Base); 1675 } 1676 1677 if (R->hasStackNonParametersStorage()) { 1678 if (isa<ElementRegion>(R)) { 1679 // Currently we don't reason specially about Clang-style vectors. Check 1680 // if superR is a vector and if so return Unknown. 1681 if (const TypedValueRegion *typedSuperR = 1682 dyn_cast<TypedValueRegion>(R->getSuperRegion())) { 1683 if (typedSuperR->getValueType()->isVectorType()) 1684 return UnknownVal(); 1685 } 1686 } 1687 1688 // FIXME: We also need to take ElementRegions with symbolic indexes into 1689 // account. This case handles both directly accessing an ElementRegion 1690 // with a symbolic offset, but also fields within an element with 1691 // a symbolic offset. 1692 if (hasSymbolicIndex) 1693 return UnknownVal(); 1694 1695 if (!hasPartialLazyBinding) 1696 return UndefinedVal(); 1697 } 1698 1699 // All other values are symbolic. 1700 return svalBuilder.getRegionValueSymbolVal(R); 1701 } 1702 1703 SVal RegionStoreManager::getBindingForObjCIvar(RegionBindingsConstRef B, 1704 const ObjCIvarRegion* R) { 1705 // Check if the region has a binding. 1706 if (const Optional<SVal> &V = B.getDirectBinding(R)) 1707 return *V; 1708 1709 const MemRegion *superR = R->getSuperRegion(); 1710 1711 // Check if the super region has a default binding. 1712 if (const Optional<SVal> &V = B.getDefaultBinding(superR)) { 1713 if (SymbolRef parentSym = V->getAsSymbol()) 1714 return svalBuilder.getDerivedRegionValueSymbolVal(parentSym, R); 1715 1716 // Other cases: give up. 1717 return UnknownVal(); 1718 } 1719 1720 return getBindingForLazySymbol(R); 1721 } 1722 1723 SVal RegionStoreManager::getBindingForVar(RegionBindingsConstRef B, 1724 const VarRegion *R) { 1725 1726 // Check if the region has a binding. 1727 if (const Optional<SVal> &V = B.getDirectBinding(R)) 1728 return *V; 1729 1730 // Lazily derive a value for the VarRegion. 1731 const VarDecl *VD = R->getDecl(); 1732 const MemSpaceRegion *MS = R->getMemorySpace(); 1733 1734 // Arguments are always symbolic. 1735 if (isa<StackArgumentsSpaceRegion>(MS)) 1736 return svalBuilder.getRegionValueSymbolVal(R); 1737 1738 // Is 'VD' declared constant? If so, retrieve the constant value. 1739 if (VD->getType().isConstQualified()) 1740 if (const Expr *Init = VD->getInit()) 1741 if (Optional<SVal> V = svalBuilder.getConstantVal(Init)) 1742 return *V; 1743 1744 // This must come after the check for constants because closure-captured 1745 // constant variables may appear in UnknownSpaceRegion. 1746 if (isa<UnknownSpaceRegion>(MS)) 1747 return svalBuilder.getRegionValueSymbolVal(R); 1748 1749 if (isa<GlobalsSpaceRegion>(MS)) { 1750 QualType T = VD->getType(); 1751 1752 // Function-scoped static variables are default-initialized to 0; if they 1753 // have an initializer, it would have been processed by now. 1754 if (isa<StaticGlobalSpaceRegion>(MS)) 1755 return svalBuilder.makeZeroVal(T); 1756 1757 if (Optional<SVal> V = getBindingForDerivedDefaultValue(B, MS, R, T)) { 1758 assert(!V->getAs<nonloc::LazyCompoundVal>()); 1759 return V.getValue(); 1760 } 1761 1762 return svalBuilder.getRegionValueSymbolVal(R); 1763 } 1764 1765 return UndefinedVal(); 1766 } 1767 1768 SVal RegionStoreManager::getBindingForLazySymbol(const TypedValueRegion *R) { 1769 // All other values are symbolic. 1770 return svalBuilder.getRegionValueSymbolVal(R); 1771 } 1772 1773 const RegionStoreManager::SValListTy & 1774 RegionStoreManager::getInterestingValues(nonloc::LazyCompoundVal LCV) { 1775 // First, check the cache. 1776 LazyBindingsMapTy::iterator I = LazyBindingsMap.find(LCV.getCVData()); 1777 if (I != LazyBindingsMap.end()) 1778 return I->second; 1779 1780 // If we don't have a list of values cached, start constructing it. 1781 SValListTy List; 1782 1783 const SubRegion *LazyR = LCV.getRegion(); 1784 RegionBindingsRef B = getRegionBindings(LCV.getStore()); 1785 1786 // If this region had /no/ bindings at the time, there are no interesting 1787 // values to return. 1788 const ClusterBindings *Cluster = B.lookup(LazyR->getBaseRegion()); 1789 if (!Cluster) 1790 return (LazyBindingsMap[LCV.getCVData()] = llvm_move(List)); 1791 1792 SmallVector<BindingPair, 32> Bindings; 1793 collectSubRegionBindings(Bindings, svalBuilder, *Cluster, LazyR, 1794 /*IncludeAllDefaultBindings=*/true); 1795 for (SmallVectorImpl<BindingPair>::const_iterator I = Bindings.begin(), 1796 E = Bindings.end(); 1797 I != E; ++I) { 1798 SVal V = I->second; 1799 if (V.isUnknownOrUndef() || V.isConstant()) 1800 continue; 1801 1802 if (Optional<nonloc::LazyCompoundVal> InnerLCV = 1803 V.getAs<nonloc::LazyCompoundVal>()) { 1804 const SValListTy &InnerList = getInterestingValues(*InnerLCV); 1805 List.insert(List.end(), InnerList.begin(), InnerList.end()); 1806 continue; 1807 } 1808 1809 List.push_back(V); 1810 } 1811 1812 return (LazyBindingsMap[LCV.getCVData()] = llvm_move(List)); 1813 } 1814 1815 NonLoc RegionStoreManager::createLazyBinding(RegionBindingsConstRef B, 1816 const TypedValueRegion *R) { 1817 if (Optional<nonloc::LazyCompoundVal> V = 1818 getExistingLazyBinding(svalBuilder, B, R, false)) 1819 return *V; 1820 1821 return svalBuilder.makeLazyCompoundVal(StoreRef(B.asStore(), *this), R); 1822 } 1823 1824 static bool isRecordEmpty(const RecordDecl *RD) { 1825 if (!RD->field_empty()) 1826 return false; 1827 if (const CXXRecordDecl *CRD = dyn_cast<CXXRecordDecl>(RD)) 1828 return CRD->getNumBases() == 0; 1829 return true; 1830 } 1831 1832 SVal RegionStoreManager::getBindingForStruct(RegionBindingsConstRef B, 1833 const TypedValueRegion *R) { 1834 const RecordDecl *RD = R->getValueType()->castAs<RecordType>()->getDecl(); 1835 if (!RD->getDefinition() || isRecordEmpty(RD)) 1836 return UnknownVal(); 1837 1838 return createLazyBinding(B, R); 1839 } 1840 1841 SVal RegionStoreManager::getBindingForArray(RegionBindingsConstRef B, 1842 const TypedValueRegion *R) { 1843 assert(Ctx.getAsConstantArrayType(R->getValueType()) && 1844 "Only constant array types can have compound bindings."); 1845 1846 return createLazyBinding(B, R); 1847 } 1848 1849 bool RegionStoreManager::includedInBindings(Store store, 1850 const MemRegion *region) const { 1851 RegionBindingsRef B = getRegionBindings(store); 1852 region = region->getBaseRegion(); 1853 1854 // Quick path: if the base is the head of a cluster, the region is live. 1855 if (B.lookup(region)) 1856 return true; 1857 1858 // Slow path: if the region is the VALUE of any binding, it is live. 1859 for (RegionBindingsRef::iterator RI = B.begin(), RE = B.end(); RI != RE; ++RI) { 1860 const ClusterBindings &Cluster = RI.getData(); 1861 for (ClusterBindings::iterator CI = Cluster.begin(), CE = Cluster.end(); 1862 CI != CE; ++CI) { 1863 const SVal &D = CI.getData(); 1864 if (const MemRegion *R = D.getAsRegion()) 1865 if (R->getBaseRegion() == region) 1866 return true; 1867 } 1868 } 1869 1870 return false; 1871 } 1872 1873 //===----------------------------------------------------------------------===// 1874 // Binding values to regions. 1875 //===----------------------------------------------------------------------===// 1876 1877 StoreRef RegionStoreManager::killBinding(Store ST, Loc L) { 1878 if (Optional<loc::MemRegionVal> LV = L.getAs<loc::MemRegionVal>()) 1879 if (const MemRegion* R = LV->getRegion()) 1880 return StoreRef(getRegionBindings(ST).removeBinding(R) 1881 .asImmutableMap() 1882 .getRootWithoutRetain(), 1883 *this); 1884 1885 return StoreRef(ST, *this); 1886 } 1887 1888 RegionBindingsRef 1889 RegionStoreManager::bind(RegionBindingsConstRef B, Loc L, SVal V) { 1890 if (L.getAs<loc::ConcreteInt>()) 1891 return B; 1892 1893 // If we get here, the location should be a region. 1894 const MemRegion *R = L.castAs<loc::MemRegionVal>().getRegion(); 1895 1896 // Check if the region is a struct region. 1897 if (const TypedValueRegion* TR = dyn_cast<TypedValueRegion>(R)) { 1898 QualType Ty = TR->getValueType(); 1899 if (Ty->isArrayType()) 1900 return bindArray(B, TR, V); 1901 if (Ty->isStructureOrClassType()) 1902 return bindStruct(B, TR, V); 1903 if (Ty->isVectorType()) 1904 return bindVector(B, TR, V); 1905 if (Ty->isUnionType()) 1906 return bindAggregate(B, TR, V); 1907 } 1908 1909 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(R)) { 1910 // Binding directly to a symbolic region should be treated as binding 1911 // to element 0. 1912 QualType T = SR->getSymbol()->getType(); 1913 if (T->isAnyPointerType() || T->isReferenceType()) 1914 T = T->getPointeeType(); 1915 1916 R = GetElementZeroRegion(SR, T); 1917 } 1918 1919 // Clear out bindings that may overlap with this binding. 1920 RegionBindingsRef NewB = removeSubRegionBindings(B, cast<SubRegion>(R)); 1921 return NewB.addBinding(BindingKey::Make(R, BindingKey::Direct), V); 1922 } 1923 1924 RegionBindingsRef 1925 RegionStoreManager::setImplicitDefaultValue(RegionBindingsConstRef B, 1926 const MemRegion *R, 1927 QualType T) { 1928 SVal V; 1929 1930 if (Loc::isLocType(T)) 1931 V = svalBuilder.makeNull(); 1932 else if (T->isIntegralOrEnumerationType()) 1933 V = svalBuilder.makeZeroVal(T); 1934 else if (T->isStructureOrClassType() || T->isArrayType()) { 1935 // Set the default value to a zero constant when it is a structure 1936 // or array. The type doesn't really matter. 1937 V = svalBuilder.makeZeroVal(Ctx.IntTy); 1938 } 1939 else { 1940 // We can't represent values of this type, but we still need to set a value 1941 // to record that the region has been initialized. 1942 // If this assertion ever fires, a new case should be added above -- we 1943 // should know how to default-initialize any value we can symbolicate. 1944 assert(!SymbolManager::canSymbolicate(T) && "This type is representable"); 1945 V = UnknownVal(); 1946 } 1947 1948 return B.addBinding(R, BindingKey::Default, V); 1949 } 1950 1951 RegionBindingsRef 1952 RegionStoreManager::bindArray(RegionBindingsConstRef B, 1953 const TypedValueRegion* R, 1954 SVal Init) { 1955 1956 const ArrayType *AT =cast<ArrayType>(Ctx.getCanonicalType(R->getValueType())); 1957 QualType ElementTy = AT->getElementType(); 1958 Optional<uint64_t> Size; 1959 1960 if (const ConstantArrayType* CAT = dyn_cast<ConstantArrayType>(AT)) 1961 Size = CAT->getSize().getZExtValue(); 1962 1963 // Check if the init expr is a string literal. 1964 if (Optional<loc::MemRegionVal> MRV = Init.getAs<loc::MemRegionVal>()) { 1965 const StringRegion *S = cast<StringRegion>(MRV->getRegion()); 1966 1967 // Treat the string as a lazy compound value. 1968 StoreRef store(B.asStore(), *this); 1969 nonloc::LazyCompoundVal LCV = svalBuilder.makeLazyCompoundVal(store, S) 1970 .castAs<nonloc::LazyCompoundVal>(); 1971 return bindAggregate(B, R, LCV); 1972 } 1973 1974 // Handle lazy compound values. 1975 if (Init.getAs<nonloc::LazyCompoundVal>()) 1976 return bindAggregate(B, R, Init); 1977 1978 // Remaining case: explicit compound values. 1979 1980 if (Init.isUnknown()) 1981 return setImplicitDefaultValue(B, R, ElementTy); 1982 1983 const nonloc::CompoundVal& CV = Init.castAs<nonloc::CompoundVal>(); 1984 nonloc::CompoundVal::iterator VI = CV.begin(), VE = CV.end(); 1985 uint64_t i = 0; 1986 1987 RegionBindingsRef NewB(B); 1988 1989 for (; Size.hasValue() ? i < Size.getValue() : true ; ++i, ++VI) { 1990 // The init list might be shorter than the array length. 1991 if (VI == VE) 1992 break; 1993 1994 const NonLoc &Idx = svalBuilder.makeArrayIndex(i); 1995 const ElementRegion *ER = MRMgr.getElementRegion(ElementTy, Idx, R, Ctx); 1996 1997 if (ElementTy->isStructureOrClassType()) 1998 NewB = bindStruct(NewB, ER, *VI); 1999 else if (ElementTy->isArrayType()) 2000 NewB = bindArray(NewB, ER, *VI); 2001 else 2002 NewB = bind(NewB, loc::MemRegionVal(ER), *VI); 2003 } 2004 2005 // If the init list is shorter than the array length, set the 2006 // array default value. 2007 if (Size.hasValue() && i < Size.getValue()) 2008 NewB = setImplicitDefaultValue(NewB, R, ElementTy); 2009 2010 return NewB; 2011 } 2012 2013 RegionBindingsRef RegionStoreManager::bindVector(RegionBindingsConstRef B, 2014 const TypedValueRegion* R, 2015 SVal V) { 2016 QualType T = R->getValueType(); 2017 assert(T->isVectorType()); 2018 const VectorType *VT = T->getAs<VectorType>(); // Use getAs for typedefs. 2019 2020 // Handle lazy compound values and symbolic values. 2021 if (V.getAs<nonloc::LazyCompoundVal>() || V.getAs<nonloc::SymbolVal>()) 2022 return bindAggregate(B, R, V); 2023 2024 // We may get non-CompoundVal accidentally due to imprecise cast logic or 2025 // that we are binding symbolic struct value. Kill the field values, and if 2026 // the value is symbolic go and bind it as a "default" binding. 2027 if (!V.getAs<nonloc::CompoundVal>()) { 2028 return bindAggregate(B, R, UnknownVal()); 2029 } 2030 2031 QualType ElemType = VT->getElementType(); 2032 nonloc::CompoundVal CV = V.castAs<nonloc::CompoundVal>(); 2033 nonloc::CompoundVal::iterator VI = CV.begin(), VE = CV.end(); 2034 unsigned index = 0, numElements = VT->getNumElements(); 2035 RegionBindingsRef NewB(B); 2036 2037 for ( ; index != numElements ; ++index) { 2038 if (VI == VE) 2039 break; 2040 2041 NonLoc Idx = svalBuilder.makeArrayIndex(index); 2042 const ElementRegion *ER = MRMgr.getElementRegion(ElemType, Idx, R, Ctx); 2043 2044 if (ElemType->isArrayType()) 2045 NewB = bindArray(NewB, ER, *VI); 2046 else if (ElemType->isStructureOrClassType()) 2047 NewB = bindStruct(NewB, ER, *VI); 2048 else 2049 NewB = bind(NewB, loc::MemRegionVal(ER), *VI); 2050 } 2051 return NewB; 2052 } 2053 2054 Optional<RegionBindingsRef> 2055 RegionStoreManager::tryBindSmallStruct(RegionBindingsConstRef B, 2056 const TypedValueRegion *R, 2057 const RecordDecl *RD, 2058 nonloc::LazyCompoundVal LCV) { 2059 FieldVector Fields; 2060 2061 if (const CXXRecordDecl *Class = dyn_cast<CXXRecordDecl>(RD)) 2062 if (Class->getNumBases() != 0 || Class->getNumVBases() != 0) 2063 return None; 2064 2065 for (RecordDecl::field_iterator I = RD->field_begin(), E = RD->field_end(); 2066 I != E; ++I) { 2067 const FieldDecl *FD = *I; 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(*I); 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 = NULL; 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