RegionStore.cpp revision e2b1246a24e8babf2f58c93713fba16b8edb8e2d
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 33using namespace clang; 34using namespace ento; 35 36//===----------------------------------------------------------------------===// 37// Representation of binding keys. 38//===----------------------------------------------------------------------===// 39 40namespace { 41class BindingKey { 42public: 43 enum Kind { Default = 0x0, Direct = 0x1 }; 44private: 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 } 65public: 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 LLVM_ATTRIBUTE_USED void dump() const; 108}; 109} // end anonymous namespace 110 111BindingKey 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 119namespace 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 136void BindingKey::dump() const { 137 llvm::errs() << *this; 138} 139 140//===----------------------------------------------------------------------===// 141// Actual Store type. 142//===----------------------------------------------------------------------===// 143 144typedef llvm::ImmutableMap<BindingKey, SVal> ClusterBindings; 145typedef llvm::ImmutableMapRef<BindingKey, SVal> ClusterBindingsRef; 146typedef std::pair<BindingKey, SVal> BindingPair; 147 148typedef llvm::ImmutableMap<const MemRegion *, ClusterBindings> 149 RegionBindings; 150 151namespace { 152class RegionBindingsRef : public llvm::ImmutableMapRef<const MemRegion *, 153 ClusterBindings> { 154 ClusterBindings::Factory &CBFactory; 155public: 156 typedef llvm::ImmutableMapRef<const MemRegion *, ClusterBindings> 157 ParentTy; 158 159 RegionBindingsRef(ClusterBindings::Factory &CBFactory, 160 const RegionBindings::TreeTy *T, 161 RegionBindings::TreeTy::Factory *F) 162 : llvm::ImmutableMapRef<const MemRegion *, ClusterBindings>(T, F), 163 CBFactory(CBFactory) {} 164 165 RegionBindingsRef(const ParentTy &P, ClusterBindings::Factory &CBFactory) 166 : llvm::ImmutableMapRef<const MemRegion *, ClusterBindings>(P), 167 CBFactory(CBFactory) {} 168 169 RegionBindingsRef add(key_type_ref K, data_type_ref D) const { 170 return RegionBindingsRef(static_cast<const ParentTy*>(this)->add(K, D), 171 CBFactory); 172 } 173 174 RegionBindingsRef remove(key_type_ref K) const { 175 return RegionBindingsRef(static_cast<const ParentTy*>(this)->remove(K), 176 CBFactory); 177 } 178 179 RegionBindingsRef addBinding(BindingKey K, SVal V) const; 180 181 RegionBindingsRef addBinding(const MemRegion *R, 182 BindingKey::Kind k, SVal V) const; 183 184 RegionBindingsRef &operator=(const RegionBindingsRef &X) { 185 *static_cast<ParentTy*>(this) = X; 186 return *this; 187 } 188 189 const SVal *lookup(BindingKey K) const; 190 const SVal *lookup(const MemRegion *R, BindingKey::Kind k) const; 191 const ClusterBindings *lookup(const MemRegion *R) const { 192 return static_cast<const ParentTy*>(this)->lookup(R); 193 } 194 195 RegionBindingsRef removeBinding(BindingKey K); 196 197 RegionBindingsRef removeBinding(const MemRegion *R, 198 BindingKey::Kind k); 199 200 RegionBindingsRef removeBinding(const MemRegion *R) { 201 return removeBinding(R, BindingKey::Direct). 202 removeBinding(R, BindingKey::Default); 203 } 204 205 Optional<SVal> getDirectBinding(const MemRegion *R) const; 206 207 /// getDefaultBinding - Returns an SVal* representing an optional default 208 /// binding associated with a region and its subregions. 209 Optional<SVal> getDefaultBinding(const MemRegion *R) const; 210 211 /// Return the internal tree as a Store. 212 Store asStore() const { 213 return asImmutableMap().getRootWithoutRetain(); 214 } 215 216 void dump(raw_ostream &OS, const char *nl) const { 217 for (iterator I = begin(), E = end(); I != E; ++I) { 218 const ClusterBindings &Cluster = I.getData(); 219 for (ClusterBindings::iterator CI = Cluster.begin(), CE = Cluster.end(); 220 CI != CE; ++CI) { 221 OS << ' ' << CI.getKey() << " : " << CI.getData() << nl; 222 } 223 OS << nl; 224 } 225 } 226 227 LLVM_ATTRIBUTE_USED void dump() const { 228 dump(llvm::errs(), "\n"); 229 } 230}; 231} // end anonymous namespace 232 233typedef const RegionBindingsRef& RegionBindingsConstRef; 234 235Optional<SVal> RegionBindingsRef::getDirectBinding(const MemRegion *R) const { 236 return Optional<SVal>::create(lookup(R, BindingKey::Direct)); 237} 238 239Optional<SVal> RegionBindingsRef::getDefaultBinding(const MemRegion *R) const { 240 if (R->isBoundable()) 241 if (const TypedValueRegion *TR = dyn_cast<TypedValueRegion>(R)) 242 if (TR->getValueType()->isUnionType()) 243 return UnknownVal(); 244 245 return Optional<SVal>::create(lookup(R, BindingKey::Default)); 246} 247 248RegionBindingsRef RegionBindingsRef::addBinding(BindingKey K, SVal V) const { 249 const MemRegion *Base = K.getBaseRegion(); 250 251 const ClusterBindings *ExistingCluster = lookup(Base); 252 ClusterBindings Cluster = (ExistingCluster ? *ExistingCluster 253 : CBFactory.getEmptyMap()); 254 255 ClusterBindings NewCluster = CBFactory.add(Cluster, K, V); 256 return add(Base, NewCluster); 257} 258 259 260RegionBindingsRef RegionBindingsRef::addBinding(const MemRegion *R, 261 BindingKey::Kind k, 262 SVal V) const { 263 return addBinding(BindingKey::Make(R, k), V); 264} 265 266const SVal *RegionBindingsRef::lookup(BindingKey K) const { 267 const ClusterBindings *Cluster = lookup(K.getBaseRegion()); 268 if (!Cluster) 269 return 0; 270 return Cluster->lookup(K); 271} 272 273const SVal *RegionBindingsRef::lookup(const MemRegion *R, 274 BindingKey::Kind k) const { 275 return lookup(BindingKey::Make(R, k)); 276} 277 278RegionBindingsRef RegionBindingsRef::removeBinding(BindingKey K) { 279 const MemRegion *Base = K.getBaseRegion(); 280 const ClusterBindings *Cluster = lookup(Base); 281 if (!Cluster) 282 return *this; 283 284 ClusterBindings NewCluster = CBFactory.remove(*Cluster, K); 285 if (NewCluster.isEmpty()) 286 return remove(Base); 287 return add(Base, NewCluster); 288} 289 290RegionBindingsRef RegionBindingsRef::removeBinding(const MemRegion *R, 291 BindingKey::Kind k){ 292 return removeBinding(BindingKey::Make(R, k)); 293} 294 295//===----------------------------------------------------------------------===// 296// Fine-grained control of RegionStoreManager. 297//===----------------------------------------------------------------------===// 298 299namespace { 300struct minimal_features_tag {}; 301struct maximal_features_tag {}; 302 303class RegionStoreFeatures { 304 bool SupportsFields; 305public: 306 RegionStoreFeatures(minimal_features_tag) : 307 SupportsFields(false) {} 308 309 RegionStoreFeatures(maximal_features_tag) : 310 SupportsFields(true) {} 311 312 void enableFields(bool t) { SupportsFields = t; } 313 314 bool supportsFields() const { return SupportsFields; } 315}; 316} 317 318//===----------------------------------------------------------------------===// 319// Main RegionStore logic. 320//===----------------------------------------------------------------------===// 321 322namespace { 323class invalidateRegionsWorker; 324 325class RegionStoreManager : public StoreManager { 326public: 327 const RegionStoreFeatures Features; 328 329 RegionBindings::Factory RBFactory; 330 mutable ClusterBindings::Factory CBFactory; 331 332 typedef std::vector<SVal> SValListTy; 333private: 334 typedef llvm::DenseMap<const LazyCompoundValData *, 335 SValListTy> LazyBindingsMapTy; 336 LazyBindingsMapTy LazyBindingsMap; 337 338 /// The largest number of fields a struct can have and still be 339 /// considered "small". 340 /// 341 /// This is currently used to decide whether or not it is worth "forcing" a 342 /// LazyCompoundVal on bind. 343 /// 344 /// This is controlled by 'region-store-small-struct-limit' option. 345 /// To disable all small-struct-dependent behavior, set the option to "0". 346 unsigned SmallStructLimit; 347 348 /// \brief A helper used to populate the work list with the given set of 349 /// regions. 350 void populateWorkList(invalidateRegionsWorker &W, 351 ArrayRef<SVal> Values, 352 bool IsArrayOfConstRegions, 353 InvalidatedRegions *TopLevelRegions); 354 355public: 356 RegionStoreManager(ProgramStateManager& mgr, const RegionStoreFeatures &f) 357 : StoreManager(mgr), Features(f), 358 RBFactory(mgr.getAllocator()), CBFactory(mgr.getAllocator()), 359 SmallStructLimit(0) { 360 if (SubEngine *Eng = StateMgr.getOwningEngine()) { 361 AnalyzerOptions &Options = Eng->getAnalysisManager().options; 362 SmallStructLimit = 363 Options.getOptionAsInteger("region-store-small-struct-limit", 2); 364 } 365 } 366 367 368 /// setImplicitDefaultValue - Set the default binding for the provided 369 /// MemRegion to the value implicitly defined for compound literals when 370 /// the value is not specified. 371 RegionBindingsRef setImplicitDefaultValue(RegionBindingsConstRef B, 372 const MemRegion *R, QualType T); 373 374 /// ArrayToPointer - Emulates the "decay" of an array to a pointer 375 /// type. 'Array' represents the lvalue of the array being decayed 376 /// to a pointer, and the returned SVal represents the decayed 377 /// version of that lvalue (i.e., a pointer to the first element of 378 /// the array). This is called by ExprEngine when evaluating 379 /// casts from arrays to pointers. 380 SVal ArrayToPointer(Loc Array); 381 382 StoreRef getInitialStore(const LocationContext *InitLoc) { 383 return StoreRef(RBFactory.getEmptyMap().getRootWithoutRetain(), *this); 384 } 385 386 //===-------------------------------------------------------------------===// 387 // Binding values to regions. 388 //===-------------------------------------------------------------------===// 389 RegionBindingsRef invalidateGlobalRegion(MemRegion::Kind K, 390 const Expr *Ex, 391 unsigned Count, 392 const LocationContext *LCtx, 393 RegionBindingsRef B, 394 InvalidatedRegions *Invalidated); 395 396 StoreRef invalidateRegions(Store store, 397 ArrayRef<SVal> Values, 398 ArrayRef<SVal> ConstValues, 399 const Expr *E, unsigned Count, 400 const LocationContext *LCtx, 401 const CallEvent *Call, 402 InvalidatedSymbols &IS, 403 InvalidatedSymbols &ConstIS, 404 InvalidatedRegions *Invalidated, 405 InvalidatedRegions *InvalidatedTopLevel, 406 InvalidatedRegions *InvalidatedTopLevelConst); 407 408 bool scanReachableSymbols(Store S, const MemRegion *R, 409 ScanReachableSymbols &Callbacks); 410 411 RegionBindingsRef removeSubRegionBindings(RegionBindingsConstRef B, 412 const SubRegion *R); 413 414public: // Part of public interface to class. 415 416 virtual StoreRef Bind(Store store, Loc LV, SVal V) { 417 return StoreRef(bind(getRegionBindings(store), LV, V).asStore(), *this); 418 } 419 420 RegionBindingsRef bind(RegionBindingsConstRef B, Loc LV, SVal V); 421 422 // BindDefault is only used to initialize a region with a default value. 423 StoreRef BindDefault(Store store, const MemRegion *R, SVal V) { 424 RegionBindingsRef B = getRegionBindings(store); 425 assert(!B.lookup(R, BindingKey::Default)); 426 assert(!B.lookup(R, BindingKey::Direct)); 427 return StoreRef(B.addBinding(R, BindingKey::Default, V) 428 .asImmutableMap() 429 .getRootWithoutRetain(), *this); 430 } 431 432 /// \brief Create a new store that binds a value to a compound literal. 433 /// 434 /// \param ST The original store whose bindings are the basis for the new 435 /// store. 436 /// 437 /// \param CL The compound literal to bind (the binding key). 438 /// 439 /// \param LC The LocationContext for the binding. 440 /// 441 /// \param V The value to bind to the compound literal. 442 StoreRef bindCompoundLiteral(Store ST, 443 const CompoundLiteralExpr *CL, 444 const LocationContext *LC, SVal V); 445 446 /// Attempt to extract the fields of \p LCV and bind them to the struct region 447 /// \p R. 448 /// 449 /// This path is used when it seems advantageous to "force" loading the values 450 /// within a LazyCompoundVal to bind memberwise to the struct region, rather 451 /// than using a Default binding at the base of the entire region. This is a 452 /// heuristic attempting to avoid building long chains of LazyCompoundVals. 453 /// 454 /// \returns The updated store bindings, or \c None if binding non-lazily 455 /// would be too expensive. 456 Optional<RegionBindingsRef> tryBindSmallStruct(RegionBindingsConstRef B, 457 const TypedValueRegion *R, 458 const RecordDecl *RD, 459 nonloc::LazyCompoundVal LCV); 460 461 /// BindStruct - Bind a compound value to a structure. 462 RegionBindingsRef bindStruct(RegionBindingsConstRef B, 463 const TypedValueRegion* R, SVal V); 464 465 /// BindVector - Bind a compound value to a vector. 466 RegionBindingsRef bindVector(RegionBindingsConstRef B, 467 const TypedValueRegion* R, SVal V); 468 469 RegionBindingsRef bindArray(RegionBindingsConstRef B, 470 const TypedValueRegion* R, 471 SVal V); 472 473 /// Clears out all bindings in the given region and assigns a new value 474 /// as a Default binding. 475 RegionBindingsRef bindAggregate(RegionBindingsConstRef B, 476 const TypedRegion *R, 477 SVal DefaultVal); 478 479 /// \brief Create a new store with the specified binding removed. 480 /// \param ST the original store, that is the basis for the new store. 481 /// \param L the location whose binding should be removed. 482 virtual StoreRef killBinding(Store ST, Loc L); 483 484 void incrementReferenceCount(Store store) { 485 getRegionBindings(store).manualRetain(); 486 } 487 488 /// If the StoreManager supports it, decrement the reference count of 489 /// the specified Store object. If the reference count hits 0, the memory 490 /// associated with the object is recycled. 491 void decrementReferenceCount(Store store) { 492 getRegionBindings(store).manualRelease(); 493 } 494 495 bool includedInBindings(Store store, const MemRegion *region) const; 496 497 /// \brief Return the value bound to specified location in a given state. 498 /// 499 /// The high level logic for this method is this: 500 /// getBinding (L) 501 /// if L has binding 502 /// return L's binding 503 /// else if L is in killset 504 /// return unknown 505 /// else 506 /// if L is on stack or heap 507 /// return undefined 508 /// else 509 /// return symbolic 510 virtual SVal getBinding(Store S, Loc L, QualType T) { 511 return getBinding(getRegionBindings(S), L, T); 512 } 513 514 SVal getBinding(RegionBindingsConstRef B, Loc L, QualType T = QualType()); 515 516 SVal getBindingForElement(RegionBindingsConstRef B, const ElementRegion *R); 517 518 SVal getBindingForField(RegionBindingsConstRef B, const FieldRegion *R); 519 520 SVal getBindingForObjCIvar(RegionBindingsConstRef B, const ObjCIvarRegion *R); 521 522 SVal getBindingForVar(RegionBindingsConstRef B, const VarRegion *R); 523 524 SVal getBindingForLazySymbol(const TypedValueRegion *R); 525 526 SVal getBindingForFieldOrElementCommon(RegionBindingsConstRef B, 527 const TypedValueRegion *R, 528 QualType Ty); 529 530 SVal getLazyBinding(const SubRegion *LazyBindingRegion, 531 RegionBindingsRef LazyBinding); 532 533 /// Get bindings for the values in a struct and return a CompoundVal, used 534 /// when doing struct copy: 535 /// struct s x, y; 536 /// x = y; 537 /// y's value is retrieved by this method. 538 SVal getBindingForStruct(RegionBindingsConstRef B, const TypedValueRegion *R); 539 SVal getBindingForArray(RegionBindingsConstRef B, const TypedValueRegion *R); 540 NonLoc createLazyBinding(RegionBindingsConstRef B, const TypedValueRegion *R); 541 542 /// Used to lazily generate derived symbols for bindings that are defined 543 /// implicitly by default bindings in a super region. 544 /// 545 /// Note that callers may need to specially handle LazyCompoundVals, which 546 /// are returned as is in case the caller needs to treat them differently. 547 Optional<SVal> getBindingForDerivedDefaultValue(RegionBindingsConstRef B, 548 const MemRegion *superR, 549 const TypedValueRegion *R, 550 QualType Ty); 551 552 /// Get the state and region whose binding this region \p R corresponds to. 553 /// 554 /// If there is no lazy binding for \p R, the returned value will have a null 555 /// \c second. Note that a null pointer can represents a valid Store. 556 std::pair<Store, const SubRegion *> 557 findLazyBinding(RegionBindingsConstRef B, const SubRegion *R, 558 const SubRegion *originalRegion); 559 560 /// Returns the cached set of interesting SVals contained within a lazy 561 /// binding. 562 /// 563 /// The precise value of "interesting" is determined for the purposes of 564 /// RegionStore's internal analysis. It must always contain all regions and 565 /// symbols, but may omit constants and other kinds of SVal. 566 const SValListTy &getInterestingValues(nonloc::LazyCompoundVal LCV); 567 568 //===------------------------------------------------------------------===// 569 // State pruning. 570 //===------------------------------------------------------------------===// 571 572 /// removeDeadBindings - Scans the RegionStore of 'state' for dead values. 573 /// It returns a new Store with these values removed. 574 StoreRef removeDeadBindings(Store store, const StackFrameContext *LCtx, 575 SymbolReaper& SymReaper); 576 577 //===------------------------------------------------------------------===// 578 // Region "extents". 579 //===------------------------------------------------------------------===// 580 581 // FIXME: This method will soon be eliminated; see the note in Store.h. 582 DefinedOrUnknownSVal getSizeInElements(ProgramStateRef state, 583 const MemRegion* R, QualType EleTy); 584 585 //===------------------------------------------------------------------===// 586 // Utility methods. 587 //===------------------------------------------------------------------===// 588 589 RegionBindingsRef getRegionBindings(Store store) const { 590 return RegionBindingsRef(CBFactory, 591 static_cast<const RegionBindings::TreeTy*>(store), 592 RBFactory.getTreeFactory()); 593 } 594 595 void print(Store store, raw_ostream &Out, const char* nl, 596 const char *sep); 597 598 void iterBindings(Store store, BindingsHandler& f) { 599 RegionBindingsRef B = getRegionBindings(store); 600 for (RegionBindingsRef::iterator I = B.begin(), E = B.end(); I != E; ++I) { 601 const ClusterBindings &Cluster = I.getData(); 602 for (ClusterBindings::iterator CI = Cluster.begin(), CE = Cluster.end(); 603 CI != CE; ++CI) { 604 const BindingKey &K = CI.getKey(); 605 if (!K.isDirect()) 606 continue; 607 if (const SubRegion *R = dyn_cast<SubRegion>(K.getRegion())) { 608 // FIXME: Possibly incorporate the offset? 609 if (!f.HandleBinding(*this, store, R, CI.getData())) 610 return; 611 } 612 } 613 } 614 } 615}; 616 617} // end anonymous namespace 618 619//===----------------------------------------------------------------------===// 620// RegionStore creation. 621//===----------------------------------------------------------------------===// 622 623StoreManager *ento::CreateRegionStoreManager(ProgramStateManager& StMgr) { 624 RegionStoreFeatures F = maximal_features_tag(); 625 return new RegionStoreManager(StMgr, F); 626} 627 628StoreManager * 629ento::CreateFieldsOnlyRegionStoreManager(ProgramStateManager &StMgr) { 630 RegionStoreFeatures F = minimal_features_tag(); 631 F.enableFields(true); 632 return new RegionStoreManager(StMgr, F); 633} 634 635 636//===----------------------------------------------------------------------===// 637// Region Cluster analysis. 638//===----------------------------------------------------------------------===// 639 640namespace { 641/// Used to determine which global regions are automatically included in the 642/// initial worklist of a ClusterAnalysis. 643enum GlobalsFilterKind { 644 /// Don't include any global regions. 645 GFK_None, 646 /// Only include system globals. 647 GFK_SystemOnly, 648 /// Include all global regions. 649 GFK_All 650}; 651 652template <typename DERIVED> 653class ClusterAnalysis { 654protected: 655 typedef llvm::DenseMap<const MemRegion *, const ClusterBindings *> ClusterMap; 656 typedef llvm::PointerIntPair<const MemRegion *, 1, bool> WorkListElement; 657 typedef SmallVector<WorkListElement, 10> WorkList; 658 659 llvm::SmallPtrSet<const ClusterBindings *, 16> Visited; 660 661 WorkList WL; 662 663 RegionStoreManager &RM; 664 ASTContext &Ctx; 665 SValBuilder &svalBuilder; 666 667 RegionBindingsRef B; 668 669private: 670 GlobalsFilterKind GlobalsFilter; 671 672protected: 673 const ClusterBindings *getCluster(const MemRegion *R) { 674 return B.lookup(R); 675 } 676 677 /// Returns true if the memory space of the given region is one of the global 678 /// regions specially included at the start of analysis. 679 bool isInitiallyIncludedGlobalRegion(const MemRegion *R) { 680 switch (GlobalsFilter) { 681 case GFK_None: 682 return false; 683 case GFK_SystemOnly: 684 return isa<GlobalSystemSpaceRegion>(R->getMemorySpace()); 685 case GFK_All: 686 return isa<NonStaticGlobalSpaceRegion>(R->getMemorySpace()); 687 } 688 689 llvm_unreachable("unknown globals filter"); 690 } 691 692public: 693 ClusterAnalysis(RegionStoreManager &rm, ProgramStateManager &StateMgr, 694 RegionBindingsRef b, GlobalsFilterKind GFK) 695 : RM(rm), Ctx(StateMgr.getContext()), 696 svalBuilder(StateMgr.getSValBuilder()), 697 B(b), GlobalsFilter(GFK) {} 698 699 RegionBindingsRef getRegionBindings() const { return B; } 700 701 bool isVisited(const MemRegion *R) { 702 return Visited.count(getCluster(R)); 703 } 704 705 void GenerateClusters() { 706 // Scan the entire set of bindings and record the region clusters. 707 for (RegionBindingsRef::iterator RI = B.begin(), RE = B.end(); 708 RI != RE; ++RI){ 709 const MemRegion *Base = RI.getKey(); 710 711 const ClusterBindings &Cluster = RI.getData(); 712 assert(!Cluster.isEmpty() && "Empty clusters should be removed"); 713 static_cast<DERIVED*>(this)->VisitAddedToCluster(Base, Cluster); 714 715 // If this is an interesting global region, add it the work list up front. 716 if (isInitiallyIncludedGlobalRegion(Base)) 717 AddToWorkList(WorkListElement(Base), &Cluster); 718 } 719 } 720 721 bool AddToWorkList(WorkListElement E, const ClusterBindings *C) { 722 if (C && !Visited.insert(C)) 723 return false; 724 WL.push_back(E); 725 return true; 726 } 727 728 bool AddToWorkList(const MemRegion *R, bool Flag = false) { 729 const MemRegion *BaseR = R->getBaseRegion(); 730 return AddToWorkList(WorkListElement(BaseR, Flag), getCluster(BaseR)); 731 } 732 733 void RunWorkList() { 734 while (!WL.empty()) { 735 WorkListElement E = WL.pop_back_val(); 736 const MemRegion *BaseR = E.getPointer(); 737 738 static_cast<DERIVED*>(this)->VisitCluster(BaseR, getCluster(BaseR), 739 E.getInt()); 740 } 741 } 742 743 void VisitAddedToCluster(const MemRegion *baseR, const ClusterBindings &C) {} 744 void VisitCluster(const MemRegion *baseR, const ClusterBindings *C) {} 745 746 void VisitCluster(const MemRegion *BaseR, const ClusterBindings *C, 747 bool Flag) { 748 static_cast<DERIVED*>(this)->VisitCluster(BaseR, C); 749 } 750}; 751} 752 753//===----------------------------------------------------------------------===// 754// Binding invalidation. 755//===----------------------------------------------------------------------===// 756 757bool RegionStoreManager::scanReachableSymbols(Store S, const MemRegion *R, 758 ScanReachableSymbols &Callbacks) { 759 assert(R == R->getBaseRegion() && "Should only be called for base regions"); 760 RegionBindingsRef B = getRegionBindings(S); 761 const ClusterBindings *Cluster = B.lookup(R); 762 763 if (!Cluster) 764 return true; 765 766 for (ClusterBindings::iterator RI = Cluster->begin(), RE = Cluster->end(); 767 RI != RE; ++RI) { 768 if (!Callbacks.scan(RI.getData())) 769 return false; 770 } 771 772 return true; 773} 774 775static inline bool isUnionField(const FieldRegion *FR) { 776 return FR->getDecl()->getParent()->isUnion(); 777} 778 779typedef SmallVector<const FieldDecl *, 8> FieldVector; 780 781void getSymbolicOffsetFields(BindingKey K, FieldVector &Fields) { 782 assert(K.hasSymbolicOffset() && "Not implemented for concrete offset keys"); 783 784 const MemRegion *Base = K.getConcreteOffsetRegion(); 785 const MemRegion *R = K.getRegion(); 786 787 while (R != Base) { 788 if (const FieldRegion *FR = dyn_cast<FieldRegion>(R)) 789 if (!isUnionField(FR)) 790 Fields.push_back(FR->getDecl()); 791 792 R = cast<SubRegion>(R)->getSuperRegion(); 793 } 794} 795 796static bool isCompatibleWithFields(BindingKey K, const FieldVector &Fields) { 797 assert(K.hasSymbolicOffset() && "Not implemented for concrete offset keys"); 798 799 if (Fields.empty()) 800 return true; 801 802 FieldVector FieldsInBindingKey; 803 getSymbolicOffsetFields(K, FieldsInBindingKey); 804 805 ptrdiff_t Delta = FieldsInBindingKey.size() - Fields.size(); 806 if (Delta >= 0) 807 return std::equal(FieldsInBindingKey.begin() + Delta, 808 FieldsInBindingKey.end(), 809 Fields.begin()); 810 else 811 return std::equal(FieldsInBindingKey.begin(), FieldsInBindingKey.end(), 812 Fields.begin() - Delta); 813} 814 815/// Collects all bindings in \p Cluster that may refer to bindings within 816/// \p Top. 817/// 818/// Each binding is a pair whose \c first is the key (a BindingKey) and whose 819/// \c second is the value (an SVal). 820/// 821/// The \p IncludeAllDefaultBindings parameter specifies whether to include 822/// default bindings that may extend beyond \p Top itself, e.g. if \p Top is 823/// an aggregate within a larger aggregate with a default binding. 824static void 825collectSubRegionBindings(SmallVectorImpl<BindingPair> &Bindings, 826 SValBuilder &SVB, const ClusterBindings &Cluster, 827 const SubRegion *Top, BindingKey TopKey, 828 bool IncludeAllDefaultBindings) { 829 FieldVector FieldsInSymbolicSubregions; 830 if (TopKey.hasSymbolicOffset()) { 831 getSymbolicOffsetFields(TopKey, FieldsInSymbolicSubregions); 832 Top = cast<SubRegion>(TopKey.getConcreteOffsetRegion()); 833 TopKey = BindingKey::Make(Top, BindingKey::Default); 834 } 835 836 // Find the length (in bits) of the region being invalidated. 837 uint64_t Length = UINT64_MAX; 838 SVal Extent = Top->getExtent(SVB); 839 if (Optional<nonloc::ConcreteInt> ExtentCI = 840 Extent.getAs<nonloc::ConcreteInt>()) { 841 const llvm::APSInt &ExtentInt = ExtentCI->getValue(); 842 assert(ExtentInt.isNonNegative() || ExtentInt.isUnsigned()); 843 // Extents are in bytes but region offsets are in bits. Be careful! 844 Length = ExtentInt.getLimitedValue() * SVB.getContext().getCharWidth(); 845 } else if (const FieldRegion *FR = dyn_cast<FieldRegion>(Top)) { 846 if (FR->getDecl()->isBitField()) 847 Length = FR->getDecl()->getBitWidthValue(SVB.getContext()); 848 } 849 850 for (ClusterBindings::iterator I = Cluster.begin(), E = Cluster.end(); 851 I != E; ++I) { 852 BindingKey NextKey = I.getKey(); 853 if (NextKey.getRegion() == TopKey.getRegion()) { 854 // FIXME: This doesn't catch the case where we're really invalidating a 855 // region with a symbolic offset. Example: 856 // R: points[i].y 857 // Next: points[0].x 858 859 if (NextKey.getOffset() > TopKey.getOffset() && 860 NextKey.getOffset() - TopKey.getOffset() < Length) { 861 // Case 1: The next binding is inside the region we're invalidating. 862 // Include it. 863 Bindings.push_back(*I); 864 865 } else if (NextKey.getOffset() == TopKey.getOffset()) { 866 // Case 2: The next binding is at the same offset as the region we're 867 // invalidating. In this case, we need to leave default bindings alone, 868 // since they may be providing a default value for a regions beyond what 869 // we're invalidating. 870 // FIXME: This is probably incorrect; consider invalidating an outer 871 // struct whose first field is bound to a LazyCompoundVal. 872 if (IncludeAllDefaultBindings || NextKey.isDirect()) 873 Bindings.push_back(*I); 874 } 875 876 } else if (NextKey.hasSymbolicOffset()) { 877 const MemRegion *Base = NextKey.getConcreteOffsetRegion(); 878 if (Top->isSubRegionOf(Base)) { 879 // Case 3: The next key is symbolic and we just changed something within 880 // its concrete region. We don't know if the binding is still valid, so 881 // we'll be conservative and include it. 882 if (IncludeAllDefaultBindings || NextKey.isDirect()) 883 if (isCompatibleWithFields(NextKey, FieldsInSymbolicSubregions)) 884 Bindings.push_back(*I); 885 } else if (const SubRegion *BaseSR = dyn_cast<SubRegion>(Base)) { 886 // Case 4: The next key is symbolic, but we changed a known 887 // super-region. In this case the binding is certainly included. 888 if (Top == Base || BaseSR->isSubRegionOf(Top)) 889 if (isCompatibleWithFields(NextKey, FieldsInSymbolicSubregions)) 890 Bindings.push_back(*I); 891 } 892 } 893 } 894} 895 896static void 897collectSubRegionBindings(SmallVectorImpl<BindingPair> &Bindings, 898 SValBuilder &SVB, const ClusterBindings &Cluster, 899 const SubRegion *Top, bool IncludeAllDefaultBindings) { 900 collectSubRegionBindings(Bindings, SVB, Cluster, Top, 901 BindingKey::Make(Top, BindingKey::Default), 902 IncludeAllDefaultBindings); 903} 904 905RegionBindingsRef 906RegionStoreManager::removeSubRegionBindings(RegionBindingsConstRef B, 907 const SubRegion *Top) { 908 BindingKey TopKey = BindingKey::Make(Top, BindingKey::Default); 909 const MemRegion *ClusterHead = TopKey.getBaseRegion(); 910 911 if (Top == ClusterHead) { 912 // We can remove an entire cluster's bindings all in one go. 913 return B.remove(Top); 914 } 915 916 const ClusterBindings *Cluster = B.lookup(ClusterHead); 917 if (!Cluster) { 918 // If we're invalidating a region with a symbolic offset, we need to make 919 // sure we don't treat the base region as uninitialized anymore. 920 if (TopKey.hasSymbolicOffset()) { 921 const SubRegion *Concrete = TopKey.getConcreteOffsetRegion(); 922 return B.addBinding(Concrete, BindingKey::Default, UnknownVal()); 923 } 924 return B; 925 } 926 927 SmallVector<BindingPair, 32> Bindings; 928 collectSubRegionBindings(Bindings, svalBuilder, *Cluster, Top, TopKey, 929 /*IncludeAllDefaultBindings=*/false); 930 931 ClusterBindingsRef Result(*Cluster, CBFactory); 932 for (SmallVectorImpl<BindingPair>::const_iterator I = Bindings.begin(), 933 E = Bindings.end(); 934 I != E; ++I) 935 Result = Result.remove(I->first); 936 937 // If we're invalidating a region with a symbolic offset, we need to make sure 938 // we don't treat the base region as uninitialized anymore. 939 // FIXME: This isn't very precise; see the example in 940 // collectSubRegionBindings. 941 if (TopKey.hasSymbolicOffset()) { 942 const SubRegion *Concrete = TopKey.getConcreteOffsetRegion(); 943 Result = Result.add(BindingKey::Make(Concrete, BindingKey::Default), 944 UnknownVal()); 945 } 946 947 if (Result.isEmpty()) 948 return B.remove(ClusterHead); 949 return B.add(ClusterHead, Result.asImmutableMap()); 950} 951 952namespace { 953class invalidateRegionsWorker : public ClusterAnalysis<invalidateRegionsWorker> 954{ 955 const Expr *Ex; 956 unsigned Count; 957 const LocationContext *LCtx; 958 InvalidatedSymbols &IS; 959 InvalidatedSymbols &ConstIS; 960 StoreManager::InvalidatedRegions *Regions; 961public: 962 invalidateRegionsWorker(RegionStoreManager &rm, 963 ProgramStateManager &stateMgr, 964 RegionBindingsRef b, 965 const Expr *ex, unsigned count, 966 const LocationContext *lctx, 967 InvalidatedSymbols &is, 968 InvalidatedSymbols &inConstIS, 969 StoreManager::InvalidatedRegions *r, 970 GlobalsFilterKind GFK) 971 : ClusterAnalysis<invalidateRegionsWorker>(rm, stateMgr, b, GFK), 972 Ex(ex), Count(count), LCtx(lctx), IS(is), ConstIS(inConstIS), Regions(r){} 973 974 /// \param IsConst Specifies if the region we are invalidating is constant. 975 /// If it is, we invalidate all subregions, but not the base region itself. 976 void VisitCluster(const MemRegion *baseR, const ClusterBindings *C, 977 bool IsConst); 978 void VisitBinding(SVal V); 979}; 980} 981 982void invalidateRegionsWorker::VisitBinding(SVal V) { 983 // A symbol? Mark it touched by the invalidation. 984 if (SymbolRef Sym = V.getAsSymbol()) 985 IS.insert(Sym); 986 987 if (const MemRegion *R = V.getAsRegion()) { 988 AddToWorkList(R); 989 return; 990 } 991 992 // Is it a LazyCompoundVal? All references get invalidated as well. 993 if (Optional<nonloc::LazyCompoundVal> LCS = 994 V.getAs<nonloc::LazyCompoundVal>()) { 995 996 const RegionStoreManager::SValListTy &Vals = RM.getInterestingValues(*LCS); 997 998 for (RegionStoreManager::SValListTy::const_iterator I = Vals.begin(), 999 E = Vals.end(); 1000 I != E; ++I) 1001 VisitBinding(*I); 1002 1003 return; 1004 } 1005} 1006 1007void invalidateRegionsWorker::VisitCluster(const MemRegion *baseR, 1008 const ClusterBindings *C, 1009 bool IsConst) { 1010 if (C) { 1011 for (ClusterBindings::iterator I = C->begin(), E = C->end(); I != E; ++I) 1012 VisitBinding(I.getData()); 1013 1014 // Invalidate the contents of a non-const base region. 1015 if (!IsConst) 1016 B = B.remove(baseR); 1017 } 1018 1019 // BlockDataRegion? If so, invalidate captured variables that are passed 1020 // by reference. 1021 if (const BlockDataRegion *BR = dyn_cast<BlockDataRegion>(baseR)) { 1022 for (BlockDataRegion::referenced_vars_iterator 1023 BI = BR->referenced_vars_begin(), BE = BR->referenced_vars_end() ; 1024 BI != BE; ++BI) { 1025 const VarRegion *VR = BI.getCapturedRegion(); 1026 const VarDecl *VD = VR->getDecl(); 1027 if (VD->getAttr<BlocksAttr>() || !VD->hasLocalStorage()) { 1028 AddToWorkList(VR); 1029 } 1030 else if (Loc::isLocType(VR->getValueType())) { 1031 // Map the current bindings to a Store to retrieve the value 1032 // of the binding. If that binding itself is a region, we should 1033 // invalidate that region. This is because a block may capture 1034 // a pointer value, but the thing pointed by that pointer may 1035 // get invalidated. 1036 SVal V = RM.getBinding(B, loc::MemRegionVal(VR)); 1037 if (Optional<Loc> L = V.getAs<Loc>()) { 1038 if (const MemRegion *LR = L->getAsRegion()) 1039 AddToWorkList(LR); 1040 } 1041 } 1042 } 1043 return; 1044 } 1045 1046 // Symbolic region? 1047 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(baseR)) { 1048 SymbolRef RegionSym = SR->getSymbol(); 1049 1050 // Mark that symbol touched by the invalidation. 1051 if (IsConst) 1052 ConstIS.insert(RegionSym); 1053 else 1054 IS.insert(RegionSym); 1055 } 1056 1057 // Nothing else should be done for a const region. 1058 if (IsConst) 1059 return; 1060 1061 // Otherwise, we have a normal data region. Record that we touched the region. 1062 if (Regions) 1063 Regions->push_back(baseR); 1064 1065 if (isa<AllocaRegion>(baseR) || isa<SymbolicRegion>(baseR)) { 1066 // Invalidate the region by setting its default value to 1067 // conjured symbol. The type of the symbol is irrelavant. 1068 DefinedOrUnknownSVal V = 1069 svalBuilder.conjureSymbolVal(baseR, Ex, LCtx, Ctx.IntTy, Count); 1070 B = B.addBinding(baseR, BindingKey::Default, V); 1071 return; 1072 } 1073 1074 if (!baseR->isBoundable()) 1075 return; 1076 1077 const TypedValueRegion *TR = cast<TypedValueRegion>(baseR); 1078 QualType T = TR->getValueType(); 1079 1080 if (isInitiallyIncludedGlobalRegion(baseR)) { 1081 // If the region is a global and we are invalidating all globals, 1082 // erasing the entry is good enough. This causes all globals to be lazily 1083 // symbolicated from the same base symbol. 1084 return; 1085 } 1086 1087 if (T->isStructureOrClassType()) { 1088 // Invalidate the region by setting its default value to 1089 // conjured symbol. The type of the symbol is irrelavant. 1090 DefinedOrUnknownSVal V = svalBuilder.conjureSymbolVal(baseR, Ex, LCtx, 1091 Ctx.IntTy, Count); 1092 B = B.addBinding(baseR, BindingKey::Default, V); 1093 return; 1094 } 1095 1096 if (const ArrayType *AT = Ctx.getAsArrayType(T)) { 1097 // Set the default value of the array to conjured symbol. 1098 DefinedOrUnknownSVal V = 1099 svalBuilder.conjureSymbolVal(baseR, Ex, LCtx, 1100 AT->getElementType(), Count); 1101 B = B.addBinding(baseR, BindingKey::Default, V); 1102 return; 1103 } 1104 1105 DefinedOrUnknownSVal V = svalBuilder.conjureSymbolVal(baseR, Ex, LCtx, 1106 T,Count); 1107 assert(SymbolManager::canSymbolicate(T) || V.isUnknown()); 1108 B = B.addBinding(baseR, BindingKey::Direct, V); 1109} 1110 1111RegionBindingsRef 1112RegionStoreManager::invalidateGlobalRegion(MemRegion::Kind K, 1113 const Expr *Ex, 1114 unsigned Count, 1115 const LocationContext *LCtx, 1116 RegionBindingsRef B, 1117 InvalidatedRegions *Invalidated) { 1118 // Bind the globals memory space to a new symbol that we will use to derive 1119 // the bindings for all globals. 1120 const GlobalsSpaceRegion *GS = MRMgr.getGlobalsRegion(K); 1121 SVal V = svalBuilder.conjureSymbolVal(/* SymbolTag = */ (const void*) GS, Ex, LCtx, 1122 /* type does not matter */ Ctx.IntTy, 1123 Count); 1124 1125 B = B.removeBinding(GS) 1126 .addBinding(BindingKey::Make(GS, BindingKey::Default), V); 1127 1128 // Even if there are no bindings in the global scope, we still need to 1129 // record that we touched it. 1130 if (Invalidated) 1131 Invalidated->push_back(GS); 1132 1133 return B; 1134} 1135 1136void RegionStoreManager::populateWorkList(invalidateRegionsWorker &W, 1137 ArrayRef<SVal> Values, 1138 bool IsArrayOfConstRegions, 1139 InvalidatedRegions *TopLevelRegions) { 1140 for (ArrayRef<SVal>::iterator I = Values.begin(), 1141 E = Values.end(); I != E; ++I) { 1142 SVal V = *I; 1143 if (Optional<nonloc::LazyCompoundVal> LCS = 1144 V.getAs<nonloc::LazyCompoundVal>()) { 1145 1146 const SValListTy &Vals = getInterestingValues(*LCS); 1147 1148 for (SValListTy::const_iterator I = Vals.begin(), 1149 E = Vals.end(); I != E; ++I) { 1150 // Note: the last argument is false here because these are 1151 // non-top-level regions. 1152 if (const MemRegion *R = (*I).getAsRegion()) 1153 W.AddToWorkList(R, /*IsConst=*/ false); 1154 } 1155 continue; 1156 } 1157 1158 if (const MemRegion *R = V.getAsRegion()) { 1159 if (TopLevelRegions) 1160 TopLevelRegions->push_back(R); 1161 W.AddToWorkList(R, /*IsConst=*/ IsArrayOfConstRegions); 1162 continue; 1163 } 1164 } 1165} 1166 1167StoreRef 1168RegionStoreManager::invalidateRegions(Store store, 1169 ArrayRef<SVal> Values, 1170 ArrayRef<SVal> ConstValues, 1171 const Expr *Ex, unsigned Count, 1172 const LocationContext *LCtx, 1173 const CallEvent *Call, 1174 InvalidatedSymbols &IS, 1175 InvalidatedSymbols &ConstIS, 1176 InvalidatedRegions *TopLevelRegions, 1177 InvalidatedRegions *TopLevelConstRegions, 1178 InvalidatedRegions *Invalidated) { 1179 GlobalsFilterKind GlobalsFilter; 1180 if (Call) { 1181 if (Call->isInSystemHeader()) 1182 GlobalsFilter = GFK_SystemOnly; 1183 else 1184 GlobalsFilter = GFK_All; 1185 } else { 1186 GlobalsFilter = GFK_None; 1187 } 1188 1189 RegionBindingsRef B = getRegionBindings(store); 1190 invalidateRegionsWorker W(*this, StateMgr, B, Ex, Count, LCtx, IS, ConstIS, 1191 Invalidated, GlobalsFilter); 1192 1193 // Scan the bindings and generate the clusters. 1194 W.GenerateClusters(); 1195 1196 // Add the regions to the worklist. 1197 populateWorkList(W, Values, /*IsArrayOfConstRegions*/ false, 1198 TopLevelRegions); 1199 populateWorkList(W, ConstValues, /*IsArrayOfConstRegions*/ true, 1200 TopLevelConstRegions); 1201 1202 W.RunWorkList(); 1203 1204 // Return the new bindings. 1205 B = W.getRegionBindings(); 1206 1207 // For calls, determine which global regions should be invalidated and 1208 // invalidate them. (Note that function-static and immutable globals are never 1209 // invalidated by this.) 1210 // TODO: This could possibly be more precise with modules. 1211 switch (GlobalsFilter) { 1212 case GFK_All: 1213 B = invalidateGlobalRegion(MemRegion::GlobalInternalSpaceRegionKind, 1214 Ex, Count, LCtx, B, Invalidated); 1215 // FALLTHROUGH 1216 case GFK_SystemOnly: 1217 B = invalidateGlobalRegion(MemRegion::GlobalSystemSpaceRegionKind, 1218 Ex, Count, LCtx, B, Invalidated); 1219 // FALLTHROUGH 1220 case GFK_None: 1221 break; 1222 } 1223 1224 return StoreRef(B.asStore(), *this); 1225} 1226 1227//===----------------------------------------------------------------------===// 1228// Extents for regions. 1229//===----------------------------------------------------------------------===// 1230 1231DefinedOrUnknownSVal 1232RegionStoreManager::getSizeInElements(ProgramStateRef state, 1233 const MemRegion *R, 1234 QualType EleTy) { 1235 SVal Size = cast<SubRegion>(R)->getExtent(svalBuilder); 1236 const llvm::APSInt *SizeInt = svalBuilder.getKnownValue(state, Size); 1237 if (!SizeInt) 1238 return UnknownVal(); 1239 1240 CharUnits RegionSize = CharUnits::fromQuantity(SizeInt->getSExtValue()); 1241 1242 if (Ctx.getAsVariableArrayType(EleTy)) { 1243 // FIXME: We need to track extra state to properly record the size 1244 // of VLAs. Returning UnknownVal here, however, is a stop-gap so that 1245 // we don't have a divide-by-zero below. 1246 return UnknownVal(); 1247 } 1248 1249 CharUnits EleSize = Ctx.getTypeSizeInChars(EleTy); 1250 1251 // If a variable is reinterpreted as a type that doesn't fit into a larger 1252 // type evenly, round it down. 1253 // This is a signed value, since it's used in arithmetic with signed indices. 1254 return svalBuilder.makeIntVal(RegionSize / EleSize, false); 1255} 1256 1257//===----------------------------------------------------------------------===// 1258// Location and region casting. 1259//===----------------------------------------------------------------------===// 1260 1261/// ArrayToPointer - Emulates the "decay" of an array to a pointer 1262/// type. 'Array' represents the lvalue of the array being decayed 1263/// to a pointer, and the returned SVal represents the decayed 1264/// version of that lvalue (i.e., a pointer to the first element of 1265/// the array). This is called by ExprEngine when evaluating casts 1266/// from arrays to pointers. 1267SVal RegionStoreManager::ArrayToPointer(Loc Array) { 1268 if (!Array.getAs<loc::MemRegionVal>()) 1269 return UnknownVal(); 1270 1271 const MemRegion* R = Array.castAs<loc::MemRegionVal>().getRegion(); 1272 const TypedValueRegion* ArrayR = dyn_cast<TypedValueRegion>(R); 1273 1274 if (!ArrayR) 1275 return UnknownVal(); 1276 1277 // Strip off typedefs from the ArrayRegion's ValueType. 1278 QualType T = ArrayR->getValueType().getDesugaredType(Ctx); 1279 const ArrayType *AT = cast<ArrayType>(T); 1280 T = AT->getElementType(); 1281 1282 NonLoc ZeroIdx = svalBuilder.makeZeroArrayIndex(); 1283 return loc::MemRegionVal(MRMgr.getElementRegion(T, ZeroIdx, ArrayR, Ctx)); 1284} 1285 1286//===----------------------------------------------------------------------===// 1287// Loading values from regions. 1288//===----------------------------------------------------------------------===// 1289 1290SVal RegionStoreManager::getBinding(RegionBindingsConstRef B, Loc L, QualType T) { 1291 assert(!L.getAs<UnknownVal>() && "location unknown"); 1292 assert(!L.getAs<UndefinedVal>() && "location undefined"); 1293 1294 // For access to concrete addresses, return UnknownVal. Checks 1295 // for null dereferences (and similar errors) are done by checkers, not 1296 // the Store. 1297 // FIXME: We can consider lazily symbolicating such memory, but we really 1298 // should defer this when we can reason easily about symbolicating arrays 1299 // of bytes. 1300 if (L.getAs<loc::ConcreteInt>()) { 1301 return UnknownVal(); 1302 } 1303 if (!L.getAs<loc::MemRegionVal>()) { 1304 return UnknownVal(); 1305 } 1306 1307 const MemRegion *MR = L.castAs<loc::MemRegionVal>().getRegion(); 1308 1309 if (isa<AllocaRegion>(MR) || 1310 isa<SymbolicRegion>(MR) || 1311 isa<CodeTextRegion>(MR)) { 1312 if (T.isNull()) { 1313 if (const TypedRegion *TR = dyn_cast<TypedRegion>(MR)) 1314 T = TR->getLocationType(); 1315 else { 1316 const SymbolicRegion *SR = cast<SymbolicRegion>(MR); 1317 T = SR->getSymbol()->getType(); 1318 } 1319 } 1320 MR = GetElementZeroRegion(MR, T); 1321 } 1322 1323 // FIXME: Perhaps this method should just take a 'const MemRegion*' argument 1324 // instead of 'Loc', and have the other Loc cases handled at a higher level. 1325 const TypedValueRegion *R = cast<TypedValueRegion>(MR); 1326 QualType RTy = R->getValueType(); 1327 1328 // FIXME: we do not yet model the parts of a complex type, so treat the 1329 // whole thing as "unknown". 1330 if (RTy->isAnyComplexType()) 1331 return UnknownVal(); 1332 1333 // FIXME: We should eventually handle funny addressing. e.g.: 1334 // 1335 // int x = ...; 1336 // int *p = &x; 1337 // char *q = (char*) p; 1338 // char c = *q; // returns the first byte of 'x'. 1339 // 1340 // Such funny addressing will occur due to layering of regions. 1341 if (RTy->isStructureOrClassType()) 1342 return getBindingForStruct(B, R); 1343 1344 // FIXME: Handle unions. 1345 if (RTy->isUnionType()) 1346 return UnknownVal(); 1347 1348 if (RTy->isArrayType()) { 1349 if (RTy->isConstantArrayType()) 1350 return getBindingForArray(B, R); 1351 else 1352 return UnknownVal(); 1353 } 1354 1355 // FIXME: handle Vector types. 1356 if (RTy->isVectorType()) 1357 return UnknownVal(); 1358 1359 if (const FieldRegion* FR = dyn_cast<FieldRegion>(R)) 1360 return CastRetrievedVal(getBindingForField(B, FR), FR, T, false); 1361 1362 if (const ElementRegion* ER = dyn_cast<ElementRegion>(R)) { 1363 // FIXME: Here we actually perform an implicit conversion from the loaded 1364 // value to the element type. Eventually we want to compose these values 1365 // more intelligently. For example, an 'element' can encompass multiple 1366 // bound regions (e.g., several bound bytes), or could be a subset of 1367 // a larger value. 1368 return CastRetrievedVal(getBindingForElement(B, ER), ER, T, false); 1369 } 1370 1371 if (const ObjCIvarRegion *IVR = dyn_cast<ObjCIvarRegion>(R)) { 1372 // FIXME: Here we actually perform an implicit conversion from the loaded 1373 // value to the ivar type. What we should model is stores to ivars 1374 // that blow past the extent of the ivar. If the address of the ivar is 1375 // reinterpretted, it is possible we stored a different value that could 1376 // fit within the ivar. Either we need to cast these when storing them 1377 // or reinterpret them lazily (as we do here). 1378 return CastRetrievedVal(getBindingForObjCIvar(B, IVR), IVR, T, false); 1379 } 1380 1381 if (const VarRegion *VR = dyn_cast<VarRegion>(R)) { 1382 // FIXME: Here we actually perform an implicit conversion from the loaded 1383 // value to the variable type. What we should model is stores to variables 1384 // that blow past the extent of the variable. If the address of the 1385 // variable is reinterpretted, it is possible we stored a different value 1386 // that could fit within the variable. Either we need to cast these when 1387 // storing them or reinterpret them lazily (as we do here). 1388 return CastRetrievedVal(getBindingForVar(B, VR), VR, T, false); 1389 } 1390 1391 const SVal *V = B.lookup(R, BindingKey::Direct); 1392 1393 // Check if the region has a binding. 1394 if (V) 1395 return *V; 1396 1397 // The location does not have a bound value. This means that it has 1398 // the value it had upon its creation and/or entry to the analyzed 1399 // function/method. These are either symbolic values or 'undefined'. 1400 if (R->hasStackNonParametersStorage()) { 1401 // All stack variables are considered to have undefined values 1402 // upon creation. All heap allocated blocks are considered to 1403 // have undefined values as well unless they are explicitly bound 1404 // to specific values. 1405 return UndefinedVal(); 1406 } 1407 1408 // All other values are symbolic. 1409 return svalBuilder.getRegionValueSymbolVal(R); 1410} 1411 1412static QualType getUnderlyingType(const SubRegion *R) { 1413 QualType RegionTy; 1414 if (const TypedValueRegion *TVR = dyn_cast<TypedValueRegion>(R)) 1415 RegionTy = TVR->getValueType(); 1416 1417 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(R)) 1418 RegionTy = SR->getSymbol()->getType(); 1419 1420 return RegionTy; 1421} 1422 1423/// Checks to see if store \p B has a lazy binding for region \p R. 1424/// 1425/// If \p AllowSubregionBindings is \c false, a lazy binding will be rejected 1426/// if there are additional bindings within \p R. 1427/// 1428/// Note that unlike RegionStoreManager::findLazyBinding, this will not search 1429/// for lazy bindings for super-regions of \p R. 1430static Optional<nonloc::LazyCompoundVal> 1431getExistingLazyBinding(SValBuilder &SVB, RegionBindingsConstRef B, 1432 const SubRegion *R, bool AllowSubregionBindings) { 1433 Optional<SVal> V = B.getDefaultBinding(R); 1434 if (!V) 1435 return None; 1436 1437 Optional<nonloc::LazyCompoundVal> LCV = V->getAs<nonloc::LazyCompoundVal>(); 1438 if (!LCV) 1439 return None; 1440 1441 // If the LCV is for a subregion, the types might not match, and we shouldn't 1442 // reuse the binding. 1443 QualType RegionTy = getUnderlyingType(R); 1444 if (!RegionTy.isNull() && 1445 !RegionTy->isVoidPointerType()) { 1446 QualType SourceRegionTy = LCV->getRegion()->getValueType(); 1447 if (!SVB.getContext().hasSameUnqualifiedType(RegionTy, SourceRegionTy)) 1448 return None; 1449 } 1450 1451 if (!AllowSubregionBindings) { 1452 // If there are any other bindings within this region, we shouldn't reuse 1453 // the top-level binding. 1454 SmallVector<BindingPair, 16> Bindings; 1455 collectSubRegionBindings(Bindings, SVB, *B.lookup(R->getBaseRegion()), R, 1456 /*IncludeAllDefaultBindings=*/true); 1457 if (Bindings.size() > 1) 1458 return None; 1459 } 1460 1461 return *LCV; 1462} 1463 1464 1465std::pair<Store, const SubRegion *> 1466RegionStoreManager::findLazyBinding(RegionBindingsConstRef B, 1467 const SubRegion *R, 1468 const SubRegion *originalRegion) { 1469 if (originalRegion != R) { 1470 if (Optional<nonloc::LazyCompoundVal> V = 1471 getExistingLazyBinding(svalBuilder, B, R, true)) 1472 return std::make_pair(V->getStore(), V->getRegion()); 1473 } 1474 1475 typedef std::pair<Store, const SubRegion *> StoreRegionPair; 1476 StoreRegionPair Result = StoreRegionPair(); 1477 1478 if (const ElementRegion *ER = dyn_cast<ElementRegion>(R)) { 1479 Result = findLazyBinding(B, cast<SubRegion>(ER->getSuperRegion()), 1480 originalRegion); 1481 1482 if (Result.second) 1483 Result.second = MRMgr.getElementRegionWithSuper(ER, Result.second); 1484 1485 } else if (const FieldRegion *FR = dyn_cast<FieldRegion>(R)) { 1486 Result = findLazyBinding(B, cast<SubRegion>(FR->getSuperRegion()), 1487 originalRegion); 1488 1489 if (Result.second) 1490 Result.second = MRMgr.getFieldRegionWithSuper(FR, Result.second); 1491 1492 } else if (const CXXBaseObjectRegion *BaseReg = 1493 dyn_cast<CXXBaseObjectRegion>(R)) { 1494 // C++ base object region is another kind of region that we should blast 1495 // through to look for lazy compound value. It is like a field region. 1496 Result = findLazyBinding(B, cast<SubRegion>(BaseReg->getSuperRegion()), 1497 originalRegion); 1498 1499 if (Result.second) 1500 Result.second = MRMgr.getCXXBaseObjectRegionWithSuper(BaseReg, 1501 Result.second); 1502 } 1503 1504 return Result; 1505} 1506 1507SVal RegionStoreManager::getBindingForElement(RegionBindingsConstRef B, 1508 const ElementRegion* R) { 1509 // We do not currently model bindings of the CompoundLiteralregion. 1510 if (isa<CompoundLiteralRegion>(R->getBaseRegion())) 1511 return UnknownVal(); 1512 1513 // Check if the region has a binding. 1514 if (const Optional<SVal> &V = B.getDirectBinding(R)) 1515 return *V; 1516 1517 const MemRegion* superR = R->getSuperRegion(); 1518 1519 // Check if the region is an element region of a string literal. 1520 if (const StringRegion *StrR=dyn_cast<StringRegion>(superR)) { 1521 // FIXME: Handle loads from strings where the literal is treated as 1522 // an integer, e.g., *((unsigned int*)"hello") 1523 QualType T = Ctx.getAsArrayType(StrR->getValueType())->getElementType(); 1524 if (T != Ctx.getCanonicalType(R->getElementType())) 1525 return UnknownVal(); 1526 1527 const StringLiteral *Str = StrR->getStringLiteral(); 1528 SVal Idx = R->getIndex(); 1529 if (Optional<nonloc::ConcreteInt> CI = Idx.getAs<nonloc::ConcreteInt>()) { 1530 int64_t i = CI->getValue().getSExtValue(); 1531 // Abort on string underrun. This can be possible by arbitrary 1532 // clients of getBindingForElement(). 1533 if (i < 0) 1534 return UndefinedVal(); 1535 int64_t length = Str->getLength(); 1536 // Technically, only i == length is guaranteed to be null. 1537 // However, such overflows should be caught before reaching this point; 1538 // the only time such an access would be made is if a string literal was 1539 // used to initialize a larger array. 1540 char c = (i >= length) ? '\0' : Str->getCodeUnit(i); 1541 return svalBuilder.makeIntVal(c, T); 1542 } 1543 } 1544 1545 // Check for loads from a code text region. For such loads, just give up. 1546 if (isa<CodeTextRegion>(superR)) 1547 return UnknownVal(); 1548 1549 // Handle the case where we are indexing into a larger scalar object. 1550 // For example, this handles: 1551 // int x = ... 1552 // char *y = &x; 1553 // return *y; 1554 // FIXME: This is a hack, and doesn't do anything really intelligent yet. 1555 const RegionRawOffset &O = R->getAsArrayOffset(); 1556 1557 // If we cannot reason about the offset, return an unknown value. 1558 if (!O.getRegion()) 1559 return UnknownVal(); 1560 1561 if (const TypedValueRegion *baseR = 1562 dyn_cast_or_null<TypedValueRegion>(O.getRegion())) { 1563 QualType baseT = baseR->getValueType(); 1564 if (baseT->isScalarType()) { 1565 QualType elemT = R->getElementType(); 1566 if (elemT->isScalarType()) { 1567 if (Ctx.getTypeSizeInChars(baseT) >= Ctx.getTypeSizeInChars(elemT)) { 1568 if (const Optional<SVal> &V = B.getDirectBinding(superR)) { 1569 if (SymbolRef parentSym = V->getAsSymbol()) 1570 return svalBuilder.getDerivedRegionValueSymbolVal(parentSym, R); 1571 1572 if (V->isUnknownOrUndef()) 1573 return *V; 1574 // Other cases: give up. We are indexing into a larger object 1575 // that has some value, but we don't know how to handle that yet. 1576 return UnknownVal(); 1577 } 1578 } 1579 } 1580 } 1581 } 1582 return getBindingForFieldOrElementCommon(B, R, R->getElementType()); 1583} 1584 1585SVal RegionStoreManager::getBindingForField(RegionBindingsConstRef B, 1586 const FieldRegion* R) { 1587 1588 // Check if the region has a binding. 1589 if (const Optional<SVal> &V = B.getDirectBinding(R)) 1590 return *V; 1591 1592 QualType Ty = R->getValueType(); 1593 return getBindingForFieldOrElementCommon(B, R, Ty); 1594} 1595 1596Optional<SVal> 1597RegionStoreManager::getBindingForDerivedDefaultValue(RegionBindingsConstRef B, 1598 const MemRegion *superR, 1599 const TypedValueRegion *R, 1600 QualType Ty) { 1601 1602 if (const Optional<SVal> &D = B.getDefaultBinding(superR)) { 1603 const SVal &val = D.getValue(); 1604 if (SymbolRef parentSym = val.getAsSymbol()) 1605 return svalBuilder.getDerivedRegionValueSymbolVal(parentSym, R); 1606 1607 if (val.isZeroConstant()) 1608 return svalBuilder.makeZeroVal(Ty); 1609 1610 if (val.isUnknownOrUndef()) 1611 return val; 1612 1613 // Lazy bindings are usually handled through getExistingLazyBinding(). 1614 // We should unify these two code paths at some point. 1615 if (val.getAs<nonloc::LazyCompoundVal>()) 1616 return val; 1617 1618 llvm_unreachable("Unknown default value"); 1619 } 1620 1621 return None; 1622} 1623 1624SVal RegionStoreManager::getLazyBinding(const SubRegion *LazyBindingRegion, 1625 RegionBindingsRef LazyBinding) { 1626 SVal Result; 1627 if (const ElementRegion *ER = dyn_cast<ElementRegion>(LazyBindingRegion)) 1628 Result = getBindingForElement(LazyBinding, ER); 1629 else 1630 Result = getBindingForField(LazyBinding, 1631 cast<FieldRegion>(LazyBindingRegion)); 1632 1633 // FIXME: This is a hack to deal with RegionStore's inability to distinguish a 1634 // default value for /part/ of an aggregate from a default value for the 1635 // /entire/ aggregate. The most common case of this is when struct Outer 1636 // has as its first member a struct Inner, which is copied in from a stack 1637 // variable. In this case, even if the Outer's default value is symbolic, 0, 1638 // or unknown, it gets overridden by the Inner's default value of undefined. 1639 // 1640 // This is a general problem -- if the Inner is zero-initialized, the Outer 1641 // will now look zero-initialized. The proper way to solve this is with a 1642 // new version of RegionStore that tracks the extent of a binding as well 1643 // as the offset. 1644 // 1645 // This hack only takes care of the undefined case because that can very 1646 // quickly result in a warning. 1647 if (Result.isUndef()) 1648 Result = UnknownVal(); 1649 1650 return Result; 1651} 1652 1653SVal 1654RegionStoreManager::getBindingForFieldOrElementCommon(RegionBindingsConstRef B, 1655 const TypedValueRegion *R, 1656 QualType Ty) { 1657 1658 // At this point we have already checked in either getBindingForElement or 1659 // getBindingForField if 'R' has a direct binding. 1660 1661 // Lazy binding? 1662 Store lazyBindingStore = NULL; 1663 const SubRegion *lazyBindingRegion = NULL; 1664 llvm::tie(lazyBindingStore, lazyBindingRegion) = findLazyBinding(B, R, R); 1665 if (lazyBindingRegion) 1666 return getLazyBinding(lazyBindingRegion, 1667 getRegionBindings(lazyBindingStore)); 1668 1669 // Record whether or not we see a symbolic index. That can completely 1670 // be out of scope of our lookup. 1671 bool hasSymbolicIndex = false; 1672 1673 // FIXME: This is a hack to deal with RegionStore's inability to distinguish a 1674 // default value for /part/ of an aggregate from a default value for the 1675 // /entire/ aggregate. The most common case of this is when struct Outer 1676 // has as its first member a struct Inner, which is copied in from a stack 1677 // variable. In this case, even if the Outer's default value is symbolic, 0, 1678 // or unknown, it gets overridden by the Inner's default value of undefined. 1679 // 1680 // This is a general problem -- if the Inner is zero-initialized, the Outer 1681 // will now look zero-initialized. The proper way to solve this is with a 1682 // new version of RegionStore that tracks the extent of a binding as well 1683 // as the offset. 1684 // 1685 // This hack only takes care of the undefined case because that can very 1686 // quickly result in a warning. 1687 bool hasPartialLazyBinding = false; 1688 1689 const SubRegion *SR = dyn_cast<SubRegion>(R); 1690 while (SR) { 1691 const MemRegion *Base = SR->getSuperRegion(); 1692 if (Optional<SVal> D = getBindingForDerivedDefaultValue(B, Base, R, Ty)) { 1693 if (D->getAs<nonloc::LazyCompoundVal>()) { 1694 hasPartialLazyBinding = true; 1695 break; 1696 } 1697 1698 return *D; 1699 } 1700 1701 if (const ElementRegion *ER = dyn_cast<ElementRegion>(Base)) { 1702 NonLoc index = ER->getIndex(); 1703 if (!index.isConstant()) 1704 hasSymbolicIndex = true; 1705 } 1706 1707 // If our super region is a field or element itself, walk up the region 1708 // hierarchy to see if there is a default value installed in an ancestor. 1709 SR = dyn_cast<SubRegion>(Base); 1710 } 1711 1712 if (R->hasStackNonParametersStorage()) { 1713 if (isa<ElementRegion>(R)) { 1714 // Currently we don't reason specially about Clang-style vectors. Check 1715 // if superR is a vector and if so return Unknown. 1716 if (const TypedValueRegion *typedSuperR = 1717 dyn_cast<TypedValueRegion>(R->getSuperRegion())) { 1718 if (typedSuperR->getValueType()->isVectorType()) 1719 return UnknownVal(); 1720 } 1721 } 1722 1723 // FIXME: We also need to take ElementRegions with symbolic indexes into 1724 // account. This case handles both directly accessing an ElementRegion 1725 // with a symbolic offset, but also fields within an element with 1726 // a symbolic offset. 1727 if (hasSymbolicIndex) 1728 return UnknownVal(); 1729 1730 if (!hasPartialLazyBinding) 1731 return UndefinedVal(); 1732 } 1733 1734 // All other values are symbolic. 1735 return svalBuilder.getRegionValueSymbolVal(R); 1736} 1737 1738SVal RegionStoreManager::getBindingForObjCIvar(RegionBindingsConstRef B, 1739 const ObjCIvarRegion* R) { 1740 // Check if the region has a binding. 1741 if (const Optional<SVal> &V = B.getDirectBinding(R)) 1742 return *V; 1743 1744 const MemRegion *superR = R->getSuperRegion(); 1745 1746 // Check if the super region has a default binding. 1747 if (const Optional<SVal> &V = B.getDefaultBinding(superR)) { 1748 if (SymbolRef parentSym = V->getAsSymbol()) 1749 return svalBuilder.getDerivedRegionValueSymbolVal(parentSym, R); 1750 1751 // Other cases: give up. 1752 return UnknownVal(); 1753 } 1754 1755 return getBindingForLazySymbol(R); 1756} 1757 1758SVal RegionStoreManager::getBindingForVar(RegionBindingsConstRef B, 1759 const VarRegion *R) { 1760 1761 // Check if the region has a binding. 1762 if (const Optional<SVal> &V = B.getDirectBinding(R)) 1763 return *V; 1764 1765 // Lazily derive a value for the VarRegion. 1766 const VarDecl *VD = R->getDecl(); 1767 const MemSpaceRegion *MS = R->getMemorySpace(); 1768 1769 // Arguments are always symbolic. 1770 if (isa<StackArgumentsSpaceRegion>(MS)) 1771 return svalBuilder.getRegionValueSymbolVal(R); 1772 1773 // Is 'VD' declared constant? If so, retrieve the constant value. 1774 if (VD->getType().isConstQualified()) 1775 if (const Expr *Init = VD->getInit()) 1776 if (Optional<SVal> V = svalBuilder.getConstantVal(Init)) 1777 return *V; 1778 1779 // This must come after the check for constants because closure-captured 1780 // constant variables may appear in UnknownSpaceRegion. 1781 if (isa<UnknownSpaceRegion>(MS)) 1782 return svalBuilder.getRegionValueSymbolVal(R); 1783 1784 if (isa<GlobalsSpaceRegion>(MS)) { 1785 QualType T = VD->getType(); 1786 1787 // Function-scoped static variables are default-initialized to 0; if they 1788 // have an initializer, it would have been processed by now. 1789 if (isa<StaticGlobalSpaceRegion>(MS)) 1790 return svalBuilder.makeZeroVal(T); 1791 1792 if (Optional<SVal> V = getBindingForDerivedDefaultValue(B, MS, R, T)) { 1793 assert(!V->getAs<nonloc::LazyCompoundVal>()); 1794 return V.getValue(); 1795 } 1796 1797 return svalBuilder.getRegionValueSymbolVal(R); 1798 } 1799 1800 return UndefinedVal(); 1801} 1802 1803SVal RegionStoreManager::getBindingForLazySymbol(const TypedValueRegion *R) { 1804 // All other values are symbolic. 1805 return svalBuilder.getRegionValueSymbolVal(R); 1806} 1807 1808const RegionStoreManager::SValListTy & 1809RegionStoreManager::getInterestingValues(nonloc::LazyCompoundVal LCV) { 1810 // First, check the cache. 1811 LazyBindingsMapTy::iterator I = LazyBindingsMap.find(LCV.getCVData()); 1812 if (I != LazyBindingsMap.end()) 1813 return I->second; 1814 1815 // If we don't have a list of values cached, start constructing it. 1816 SValListTy List; 1817 1818 const SubRegion *LazyR = LCV.getRegion(); 1819 RegionBindingsRef B = getRegionBindings(LCV.getStore()); 1820 1821 // If this region had /no/ bindings at the time, there are no interesting 1822 // values to return. 1823 const ClusterBindings *Cluster = B.lookup(LazyR->getBaseRegion()); 1824 if (!Cluster) 1825 return (LazyBindingsMap[LCV.getCVData()] = llvm_move(List)); 1826 1827 SmallVector<BindingPair, 32> Bindings; 1828 collectSubRegionBindings(Bindings, svalBuilder, *Cluster, LazyR, 1829 /*IncludeAllDefaultBindings=*/true); 1830 for (SmallVectorImpl<BindingPair>::const_iterator I = Bindings.begin(), 1831 E = Bindings.end(); 1832 I != E; ++I) { 1833 SVal V = I->second; 1834 if (V.isUnknownOrUndef() || V.isConstant()) 1835 continue; 1836 1837 if (Optional<nonloc::LazyCompoundVal> InnerLCV = 1838 V.getAs<nonloc::LazyCompoundVal>()) { 1839 const SValListTy &InnerList = getInterestingValues(*InnerLCV); 1840 List.insert(List.end(), InnerList.begin(), InnerList.end()); 1841 continue; 1842 } 1843 1844 List.push_back(V); 1845 } 1846 1847 return (LazyBindingsMap[LCV.getCVData()] = llvm_move(List)); 1848} 1849 1850NonLoc RegionStoreManager::createLazyBinding(RegionBindingsConstRef B, 1851 const TypedValueRegion *R) { 1852 if (Optional<nonloc::LazyCompoundVal> V = 1853 getExistingLazyBinding(svalBuilder, B, R, false)) 1854 return *V; 1855 1856 return svalBuilder.makeLazyCompoundVal(StoreRef(B.asStore(), *this), R); 1857} 1858 1859SVal RegionStoreManager::getBindingForStruct(RegionBindingsConstRef B, 1860 const TypedValueRegion *R) { 1861 const RecordDecl *RD = R->getValueType()->castAs<RecordType>()->getDecl(); 1862 if (RD->field_empty()) 1863 return UnknownVal(); 1864 1865 return createLazyBinding(B, R); 1866} 1867 1868SVal RegionStoreManager::getBindingForArray(RegionBindingsConstRef B, 1869 const TypedValueRegion *R) { 1870 assert(Ctx.getAsConstantArrayType(R->getValueType()) && 1871 "Only constant array types can have compound bindings."); 1872 1873 return createLazyBinding(B, R); 1874} 1875 1876bool RegionStoreManager::includedInBindings(Store store, 1877 const MemRegion *region) const { 1878 RegionBindingsRef B = getRegionBindings(store); 1879 region = region->getBaseRegion(); 1880 1881 // Quick path: if the base is the head of a cluster, the region is live. 1882 if (B.lookup(region)) 1883 return true; 1884 1885 // Slow path: if the region is the VALUE of any binding, it is live. 1886 for (RegionBindingsRef::iterator RI = B.begin(), RE = B.end(); RI != RE; ++RI) { 1887 const ClusterBindings &Cluster = RI.getData(); 1888 for (ClusterBindings::iterator CI = Cluster.begin(), CE = Cluster.end(); 1889 CI != CE; ++CI) { 1890 const SVal &D = CI.getData(); 1891 if (const MemRegion *R = D.getAsRegion()) 1892 if (R->getBaseRegion() == region) 1893 return true; 1894 } 1895 } 1896 1897 return false; 1898} 1899 1900//===----------------------------------------------------------------------===// 1901// Binding values to regions. 1902//===----------------------------------------------------------------------===// 1903 1904StoreRef RegionStoreManager::killBinding(Store ST, Loc L) { 1905 if (Optional<loc::MemRegionVal> LV = L.getAs<loc::MemRegionVal>()) 1906 if (const MemRegion* R = LV->getRegion()) 1907 return StoreRef(getRegionBindings(ST).removeBinding(R) 1908 .asImmutableMap() 1909 .getRootWithoutRetain(), 1910 *this); 1911 1912 return StoreRef(ST, *this); 1913} 1914 1915RegionBindingsRef 1916RegionStoreManager::bind(RegionBindingsConstRef B, Loc L, SVal V) { 1917 if (L.getAs<loc::ConcreteInt>()) 1918 return B; 1919 1920 // If we get here, the location should be a region. 1921 const MemRegion *R = L.castAs<loc::MemRegionVal>().getRegion(); 1922 1923 // Check if the region is a struct region. 1924 if (const TypedValueRegion* TR = dyn_cast<TypedValueRegion>(R)) { 1925 QualType Ty = TR->getValueType(); 1926 if (Ty->isArrayType()) 1927 return bindArray(B, TR, V); 1928 if (Ty->isStructureOrClassType()) 1929 return bindStruct(B, TR, V); 1930 if (Ty->isVectorType()) 1931 return bindVector(B, TR, V); 1932 } 1933 1934 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(R)) { 1935 // Binding directly to a symbolic region should be treated as binding 1936 // to element 0. 1937 QualType T = SR->getSymbol()->getType(); 1938 if (T->isAnyPointerType() || T->isReferenceType()) 1939 T = T->getPointeeType(); 1940 1941 R = GetElementZeroRegion(SR, T); 1942 } 1943 1944 // Clear out bindings that may overlap with this binding. 1945 RegionBindingsRef NewB = removeSubRegionBindings(B, cast<SubRegion>(R)); 1946 return NewB.addBinding(BindingKey::Make(R, BindingKey::Direct), V); 1947} 1948 1949// FIXME: this method should be merged into Bind(). 1950StoreRef RegionStoreManager::bindCompoundLiteral(Store ST, 1951 const CompoundLiteralExpr *CL, 1952 const LocationContext *LC, 1953 SVal V) { 1954 return Bind(ST, loc::MemRegionVal(MRMgr.getCompoundLiteralRegion(CL, LC)), V); 1955} 1956 1957RegionBindingsRef 1958RegionStoreManager::setImplicitDefaultValue(RegionBindingsConstRef B, 1959 const MemRegion *R, 1960 QualType T) { 1961 SVal V; 1962 1963 if (Loc::isLocType(T)) 1964 V = svalBuilder.makeNull(); 1965 else if (T->isIntegralOrEnumerationType()) 1966 V = svalBuilder.makeZeroVal(T); 1967 else if (T->isStructureOrClassType() || T->isArrayType()) { 1968 // Set the default value to a zero constant when it is a structure 1969 // or array. The type doesn't really matter. 1970 V = svalBuilder.makeZeroVal(Ctx.IntTy); 1971 } 1972 else { 1973 // We can't represent values of this type, but we still need to set a value 1974 // to record that the region has been initialized. 1975 // If this assertion ever fires, a new case should be added above -- we 1976 // should know how to default-initialize any value we can symbolicate. 1977 assert(!SymbolManager::canSymbolicate(T) && "This type is representable"); 1978 V = UnknownVal(); 1979 } 1980 1981 return B.addBinding(R, BindingKey::Default, V); 1982} 1983 1984RegionBindingsRef 1985RegionStoreManager::bindArray(RegionBindingsConstRef B, 1986 const TypedValueRegion* R, 1987 SVal Init) { 1988 1989 const ArrayType *AT =cast<ArrayType>(Ctx.getCanonicalType(R->getValueType())); 1990 QualType ElementTy = AT->getElementType(); 1991 Optional<uint64_t> Size; 1992 1993 if (const ConstantArrayType* CAT = dyn_cast<ConstantArrayType>(AT)) 1994 Size = CAT->getSize().getZExtValue(); 1995 1996 // Check if the init expr is a string literal. 1997 if (Optional<loc::MemRegionVal> MRV = Init.getAs<loc::MemRegionVal>()) { 1998 const StringRegion *S = cast<StringRegion>(MRV->getRegion()); 1999 2000 // Treat the string as a lazy compound value. 2001 StoreRef store(B.asStore(), *this); 2002 nonloc::LazyCompoundVal LCV = svalBuilder.makeLazyCompoundVal(store, S) 2003 .castAs<nonloc::LazyCompoundVal>(); 2004 return bindAggregate(B, R, LCV); 2005 } 2006 2007 // Handle lazy compound values. 2008 if (Init.getAs<nonloc::LazyCompoundVal>()) 2009 return bindAggregate(B, R, Init); 2010 2011 // Remaining case: explicit compound values. 2012 2013 if (Init.isUnknown()) 2014 return setImplicitDefaultValue(B, R, ElementTy); 2015 2016 const nonloc::CompoundVal& CV = Init.castAs<nonloc::CompoundVal>(); 2017 nonloc::CompoundVal::iterator VI = CV.begin(), VE = CV.end(); 2018 uint64_t i = 0; 2019 2020 RegionBindingsRef NewB(B); 2021 2022 for (; Size.hasValue() ? i < Size.getValue() : true ; ++i, ++VI) { 2023 // The init list might be shorter than the array length. 2024 if (VI == VE) 2025 break; 2026 2027 const NonLoc &Idx = svalBuilder.makeArrayIndex(i); 2028 const ElementRegion *ER = MRMgr.getElementRegion(ElementTy, Idx, R, Ctx); 2029 2030 if (ElementTy->isStructureOrClassType()) 2031 NewB = bindStruct(NewB, ER, *VI); 2032 else if (ElementTy->isArrayType()) 2033 NewB = bindArray(NewB, ER, *VI); 2034 else 2035 NewB = bind(NewB, loc::MemRegionVal(ER), *VI); 2036 } 2037 2038 // If the init list is shorter than the array length, set the 2039 // array default value. 2040 if (Size.hasValue() && i < Size.getValue()) 2041 NewB = setImplicitDefaultValue(NewB, R, ElementTy); 2042 2043 return NewB; 2044} 2045 2046RegionBindingsRef RegionStoreManager::bindVector(RegionBindingsConstRef B, 2047 const TypedValueRegion* R, 2048 SVal V) { 2049 QualType T = R->getValueType(); 2050 assert(T->isVectorType()); 2051 const VectorType *VT = T->getAs<VectorType>(); // Use getAs for typedefs. 2052 2053 // Handle lazy compound values and symbolic values. 2054 if (V.getAs<nonloc::LazyCompoundVal>() || V.getAs<nonloc::SymbolVal>()) 2055 return bindAggregate(B, R, V); 2056 2057 // We may get non-CompoundVal accidentally due to imprecise cast logic or 2058 // that we are binding symbolic struct value. Kill the field values, and if 2059 // the value is symbolic go and bind it as a "default" binding. 2060 if (!V.getAs<nonloc::CompoundVal>()) { 2061 return bindAggregate(B, R, UnknownVal()); 2062 } 2063 2064 QualType ElemType = VT->getElementType(); 2065 nonloc::CompoundVal CV = V.castAs<nonloc::CompoundVal>(); 2066 nonloc::CompoundVal::iterator VI = CV.begin(), VE = CV.end(); 2067 unsigned index = 0, numElements = VT->getNumElements(); 2068 RegionBindingsRef NewB(B); 2069 2070 for ( ; index != numElements ; ++index) { 2071 if (VI == VE) 2072 break; 2073 2074 NonLoc Idx = svalBuilder.makeArrayIndex(index); 2075 const ElementRegion *ER = MRMgr.getElementRegion(ElemType, Idx, R, Ctx); 2076 2077 if (ElemType->isArrayType()) 2078 NewB = bindArray(NewB, ER, *VI); 2079 else if (ElemType->isStructureOrClassType()) 2080 NewB = bindStruct(NewB, ER, *VI); 2081 else 2082 NewB = bind(NewB, loc::MemRegionVal(ER), *VI); 2083 } 2084 return NewB; 2085} 2086 2087Optional<RegionBindingsRef> 2088RegionStoreManager::tryBindSmallStruct(RegionBindingsConstRef B, 2089 const TypedValueRegion *R, 2090 const RecordDecl *RD, 2091 nonloc::LazyCompoundVal LCV) { 2092 FieldVector Fields; 2093 2094 if (const CXXRecordDecl *Class = dyn_cast<CXXRecordDecl>(RD)) 2095 if (Class->getNumBases() != 0 || Class->getNumVBases() != 0) 2096 return None; 2097 2098 for (RecordDecl::field_iterator I = RD->field_begin(), E = RD->field_end(); 2099 I != E; ++I) { 2100 const FieldDecl *FD = *I; 2101 if (FD->isUnnamedBitfield()) 2102 continue; 2103 2104 // If there are too many fields, or if any of the fields are aggregates, 2105 // just use the LCV as a default binding. 2106 if (Fields.size() == SmallStructLimit) 2107 return None; 2108 2109 QualType Ty = FD->getType(); 2110 if (!(Ty->isScalarType() || Ty->isReferenceType())) 2111 return None; 2112 2113 Fields.push_back(*I); 2114 } 2115 2116 RegionBindingsRef NewB = B; 2117 2118 for (FieldVector::iterator I = Fields.begin(), E = Fields.end(); I != E; ++I){ 2119 const FieldRegion *SourceFR = MRMgr.getFieldRegion(*I, LCV.getRegion()); 2120 SVal V = getBindingForField(getRegionBindings(LCV.getStore()), SourceFR); 2121 2122 const FieldRegion *DestFR = MRMgr.getFieldRegion(*I, R); 2123 NewB = bind(NewB, loc::MemRegionVal(DestFR), V); 2124 } 2125 2126 return NewB; 2127} 2128 2129RegionBindingsRef RegionStoreManager::bindStruct(RegionBindingsConstRef B, 2130 const TypedValueRegion* R, 2131 SVal V) { 2132 if (!Features.supportsFields()) 2133 return B; 2134 2135 QualType T = R->getValueType(); 2136 assert(T->isStructureOrClassType()); 2137 2138 const RecordType* RT = T->getAs<RecordType>(); 2139 const RecordDecl *RD = RT->getDecl(); 2140 2141 if (!RD->isCompleteDefinition()) 2142 return B; 2143 2144 // Handle lazy compound values and symbolic values. 2145 if (Optional<nonloc::LazyCompoundVal> LCV = 2146 V.getAs<nonloc::LazyCompoundVal>()) { 2147 if (Optional<RegionBindingsRef> NewB = tryBindSmallStruct(B, R, RD, *LCV)) 2148 return *NewB; 2149 return bindAggregate(B, R, V); 2150 } 2151 if (V.getAs<nonloc::SymbolVal>()) 2152 return bindAggregate(B, R, V); 2153 2154 // We may get non-CompoundVal accidentally due to imprecise cast logic or 2155 // that we are binding symbolic struct value. Kill the field values, and if 2156 // the value is symbolic go and bind it as a "default" binding. 2157 if (V.isUnknown() || !V.getAs<nonloc::CompoundVal>()) 2158 return bindAggregate(B, R, UnknownVal()); 2159 2160 const nonloc::CompoundVal& CV = V.castAs<nonloc::CompoundVal>(); 2161 nonloc::CompoundVal::iterator VI = CV.begin(), VE = CV.end(); 2162 2163 RecordDecl::field_iterator FI, FE; 2164 RegionBindingsRef NewB(B); 2165 2166 for (FI = RD->field_begin(), FE = RD->field_end(); FI != FE; ++FI) { 2167 2168 if (VI == VE) 2169 break; 2170 2171 // Skip any unnamed bitfields to stay in sync with the initializers. 2172 if (FI->isUnnamedBitfield()) 2173 continue; 2174 2175 QualType FTy = FI->getType(); 2176 const FieldRegion* FR = MRMgr.getFieldRegion(*FI, R); 2177 2178 if (FTy->isArrayType()) 2179 NewB = bindArray(NewB, FR, *VI); 2180 else if (FTy->isStructureOrClassType()) 2181 NewB = bindStruct(NewB, FR, *VI); 2182 else 2183 NewB = bind(NewB, loc::MemRegionVal(FR), *VI); 2184 ++VI; 2185 } 2186 2187 // There may be fewer values in the initialize list than the fields of struct. 2188 if (FI != FE) { 2189 NewB = NewB.addBinding(R, BindingKey::Default, 2190 svalBuilder.makeIntVal(0, false)); 2191 } 2192 2193 return NewB; 2194} 2195 2196RegionBindingsRef 2197RegionStoreManager::bindAggregate(RegionBindingsConstRef B, 2198 const TypedRegion *R, 2199 SVal Val) { 2200 // Remove the old bindings, using 'R' as the root of all regions 2201 // we will invalidate. Then add the new binding. 2202 return removeSubRegionBindings(B, R).addBinding(R, BindingKey::Default, Val); 2203} 2204 2205//===----------------------------------------------------------------------===// 2206// State pruning. 2207//===----------------------------------------------------------------------===// 2208 2209namespace { 2210class removeDeadBindingsWorker : 2211 public ClusterAnalysis<removeDeadBindingsWorker> { 2212 SmallVector<const SymbolicRegion*, 12> Postponed; 2213 SymbolReaper &SymReaper; 2214 const StackFrameContext *CurrentLCtx; 2215 2216public: 2217 removeDeadBindingsWorker(RegionStoreManager &rm, 2218 ProgramStateManager &stateMgr, 2219 RegionBindingsRef b, SymbolReaper &symReaper, 2220 const StackFrameContext *LCtx) 2221 : ClusterAnalysis<removeDeadBindingsWorker>(rm, stateMgr, b, GFK_None), 2222 SymReaper(symReaper), CurrentLCtx(LCtx) {} 2223 2224 // Called by ClusterAnalysis. 2225 void VisitAddedToCluster(const MemRegion *baseR, const ClusterBindings &C); 2226 void VisitCluster(const MemRegion *baseR, const ClusterBindings *C); 2227 using ClusterAnalysis<removeDeadBindingsWorker>::VisitCluster; 2228 2229 bool UpdatePostponed(); 2230 void VisitBinding(SVal V); 2231}; 2232} 2233 2234void removeDeadBindingsWorker::VisitAddedToCluster(const MemRegion *baseR, 2235 const ClusterBindings &C) { 2236 2237 if (const VarRegion *VR = dyn_cast<VarRegion>(baseR)) { 2238 if (SymReaper.isLive(VR)) 2239 AddToWorkList(baseR, &C); 2240 2241 return; 2242 } 2243 2244 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(baseR)) { 2245 if (SymReaper.isLive(SR->getSymbol())) 2246 AddToWorkList(SR, &C); 2247 else 2248 Postponed.push_back(SR); 2249 2250 return; 2251 } 2252 2253 if (isa<NonStaticGlobalSpaceRegion>(baseR)) { 2254 AddToWorkList(baseR, &C); 2255 return; 2256 } 2257 2258 // CXXThisRegion in the current or parent location context is live. 2259 if (const CXXThisRegion *TR = dyn_cast<CXXThisRegion>(baseR)) { 2260 const StackArgumentsSpaceRegion *StackReg = 2261 cast<StackArgumentsSpaceRegion>(TR->getSuperRegion()); 2262 const StackFrameContext *RegCtx = StackReg->getStackFrame(); 2263 if (CurrentLCtx && 2264 (RegCtx == CurrentLCtx || RegCtx->isParentOf(CurrentLCtx))) 2265 AddToWorkList(TR, &C); 2266 } 2267} 2268 2269void removeDeadBindingsWorker::VisitCluster(const MemRegion *baseR, 2270 const ClusterBindings *C) { 2271 if (!C) 2272 return; 2273 2274 // Mark the symbol for any SymbolicRegion with live bindings as live itself. 2275 // This means we should continue to track that symbol. 2276 if (const SymbolicRegion *SymR = dyn_cast<SymbolicRegion>(baseR)) 2277 SymReaper.markLive(SymR->getSymbol()); 2278 2279 for (ClusterBindings::iterator I = C->begin(), E = C->end(); I != E; ++I) 2280 VisitBinding(I.getData()); 2281} 2282 2283void removeDeadBindingsWorker::VisitBinding(SVal V) { 2284 // Is it a LazyCompoundVal? All referenced regions are live as well. 2285 if (Optional<nonloc::LazyCompoundVal> LCS = 2286 V.getAs<nonloc::LazyCompoundVal>()) { 2287 2288 const RegionStoreManager::SValListTy &Vals = RM.getInterestingValues(*LCS); 2289 2290 for (RegionStoreManager::SValListTy::const_iterator I = Vals.begin(), 2291 E = Vals.end(); 2292 I != E; ++I) 2293 VisitBinding(*I); 2294 2295 return; 2296 } 2297 2298 // If V is a region, then add it to the worklist. 2299 if (const MemRegion *R = V.getAsRegion()) { 2300 AddToWorkList(R); 2301 2302 // All regions captured by a block are also live. 2303 if (const BlockDataRegion *BR = dyn_cast<BlockDataRegion>(R)) { 2304 BlockDataRegion::referenced_vars_iterator I = BR->referenced_vars_begin(), 2305 E = BR->referenced_vars_end(); 2306 for ( ; I != E; ++I) 2307 AddToWorkList(I.getCapturedRegion()); 2308 } 2309 } 2310 2311 2312 // Update the set of live symbols. 2313 for (SymExpr::symbol_iterator SI = V.symbol_begin(), SE = V.symbol_end(); 2314 SI!=SE; ++SI) 2315 SymReaper.markLive(*SI); 2316} 2317 2318bool removeDeadBindingsWorker::UpdatePostponed() { 2319 // See if any postponed SymbolicRegions are actually live now, after 2320 // having done a scan. 2321 bool changed = false; 2322 2323 for (SmallVectorImpl<const SymbolicRegion*>::iterator 2324 I = Postponed.begin(), E = Postponed.end() ; I != E ; ++I) { 2325 if (const SymbolicRegion *SR = *I) { 2326 if (SymReaper.isLive(SR->getSymbol())) { 2327 changed |= AddToWorkList(SR); 2328 *I = NULL; 2329 } 2330 } 2331 } 2332 2333 return changed; 2334} 2335 2336StoreRef RegionStoreManager::removeDeadBindings(Store store, 2337 const StackFrameContext *LCtx, 2338 SymbolReaper& SymReaper) { 2339 RegionBindingsRef B = getRegionBindings(store); 2340 removeDeadBindingsWorker W(*this, StateMgr, B, SymReaper, LCtx); 2341 W.GenerateClusters(); 2342 2343 // Enqueue the region roots onto the worklist. 2344 for (SymbolReaper::region_iterator I = SymReaper.region_begin(), 2345 E = SymReaper.region_end(); I != E; ++I) { 2346 W.AddToWorkList(*I); 2347 } 2348 2349 do W.RunWorkList(); while (W.UpdatePostponed()); 2350 2351 // We have now scanned the store, marking reachable regions and symbols 2352 // as live. We now remove all the regions that are dead from the store 2353 // as well as update DSymbols with the set symbols that are now dead. 2354 for (RegionBindingsRef::iterator I = B.begin(), E = B.end(); I != E; ++I) { 2355 const MemRegion *Base = I.getKey(); 2356 2357 // If the cluster has been visited, we know the region has been marked. 2358 if (W.isVisited(Base)) 2359 continue; 2360 2361 // Remove the dead entry. 2362 B = B.remove(Base); 2363 2364 if (const SymbolicRegion *SymR = dyn_cast<SymbolicRegion>(Base)) 2365 SymReaper.maybeDead(SymR->getSymbol()); 2366 2367 // Mark all non-live symbols that this binding references as dead. 2368 const ClusterBindings &Cluster = I.getData(); 2369 for (ClusterBindings::iterator CI = Cluster.begin(), CE = Cluster.end(); 2370 CI != CE; ++CI) { 2371 SVal X = CI.getData(); 2372 SymExpr::symbol_iterator SI = X.symbol_begin(), SE = X.symbol_end(); 2373 for (; SI != SE; ++SI) 2374 SymReaper.maybeDead(*SI); 2375 } 2376 } 2377 2378 return StoreRef(B.asStore(), *this); 2379} 2380 2381//===----------------------------------------------------------------------===// 2382// Utility methods. 2383//===----------------------------------------------------------------------===// 2384 2385void RegionStoreManager::print(Store store, raw_ostream &OS, 2386 const char* nl, const char *sep) { 2387 RegionBindingsRef B = getRegionBindings(store); 2388 OS << "Store (direct and default bindings), " 2389 << B.asStore() 2390 << " :" << nl; 2391 B.dump(OS, nl); 2392} 2393