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