RegionStore.cpp revision e0208ff84598f48e0aafecf5b543afeff8574045
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/CallEvent.h"
23#include "clang/StaticAnalyzer/Core/PathSensitive/MemRegion.h"
24#include "clang/StaticAnalyzer/Core/PathSensitive/ProgramState.h"
25#include "clang/StaticAnalyzer/Core/PathSensitive/ProgramStateTrait.h"
26#include "llvm/ADT/ImmutableList.h"
27#include "llvm/ADT/ImmutableMap.h"
28#include "llvm/ADT/Optional.h"
29#include "llvm/Support/raw_ostream.h"
30
31using namespace clang;
32using namespace ento;
33
34//===----------------------------------------------------------------------===//
35// Representation of binding keys.
36//===----------------------------------------------------------------------===//
37
38namespace {
39class BindingKey {
40public:
41  enum Kind { Default = 0x0, Direct = 0x1 };
42private:
43  enum { Symbolic = 0x2 };
44
45  llvm::PointerIntPair<const MemRegion *, 2> P;
46  uint64_t Data;
47
48  /// Create a key for a binding to region \p r, which has a symbolic offset
49  /// from region \p Base.
50  explicit BindingKey(const SubRegion *r, const SubRegion *Base, Kind k)
51    : P(r, k | Symbolic), Data(reinterpret_cast<uintptr_t>(Base)) {
52    assert(r && Base && "Must have known regions.");
53    assert(getConcreteOffsetRegion() == Base && "Failed to store base region");
54  }
55
56  /// Create a key for a binding at \p offset from base region \p r.
57  explicit BindingKey(const MemRegion *r, uint64_t offset, Kind k)
58    : P(r, k), Data(offset) {
59    assert(r && "Must have known regions.");
60    assert(getOffset() == offset && "Failed to store offset");
61    assert((r == r->getBaseRegion() || isa<ObjCIvarRegion>(r)) && "Not a base");
62  }
63public:
64
65  bool isDirect() const { return P.getInt() & Direct; }
66  bool hasSymbolicOffset() const { return P.getInt() & Symbolic; }
67
68  const MemRegion *getRegion() const { return P.getPointer(); }
69  uint64_t getOffset() const {
70    assert(!hasSymbolicOffset());
71    return Data;
72  }
73
74  const SubRegion *getConcreteOffsetRegion() const {
75    assert(hasSymbolicOffset());
76    return reinterpret_cast<const SubRegion *>(static_cast<uintptr_t>(Data));
77  }
78
79  const MemRegion *getBaseRegion() const {
80    if (hasSymbolicOffset())
81      return getConcreteOffsetRegion()->getBaseRegion();
82    return getRegion()->getBaseRegion();
83  }
84
85  void Profile(llvm::FoldingSetNodeID& ID) const {
86    ID.AddPointer(P.getOpaqueValue());
87    ID.AddInteger(Data);
88  }
89
90  static BindingKey Make(const MemRegion *R, Kind k);
91
92  bool operator<(const BindingKey &X) const {
93    if (P.getOpaqueValue() < X.P.getOpaqueValue())
94      return true;
95    if (P.getOpaqueValue() > X.P.getOpaqueValue())
96      return false;
97    return Data < X.Data;
98  }
99
100  bool operator==(const BindingKey &X) const {
101    return P.getOpaqueValue() == X.P.getOpaqueValue() &&
102           Data == X.Data;
103  }
104
105  LLVM_ATTRIBUTE_USED void dump() const;
106};
107} // end anonymous namespace
108
109BindingKey BindingKey::Make(const MemRegion *R, Kind k) {
110  const RegionOffset &RO = R->getAsOffset();
111  if (RO.hasSymbolicOffset())
112    return BindingKey(cast<SubRegion>(R), cast<SubRegion>(RO.getRegion()), k);
113
114  return BindingKey(RO.getRegion(), RO.getOffset(), k);
115}
116
117namespace llvm {
118  static inline
119  raw_ostream &operator<<(raw_ostream &os, BindingKey K) {
120    os << '(' << K.getRegion();
121    if (!K.hasSymbolicOffset())
122      os << ',' << K.getOffset();
123    os << ',' << (K.isDirect() ? "direct" : "default")
124       << ')';
125    return os;
126  }
127
128  template <typename T> struct isPodLike;
129  template <> struct isPodLike<BindingKey> {
130    static const bool value = true;
131  };
132} // end llvm namespace
133
134void BindingKey::dump() const {
135  llvm::errs() << *this;
136}
137
138//===----------------------------------------------------------------------===//
139// Actual Store type.
140//===----------------------------------------------------------------------===//
141
142typedef llvm::ImmutableMap<BindingKey, SVal>    ClusterBindings;
143typedef llvm::ImmutableMapRef<BindingKey, SVal> ClusterBindingsRef;
144typedef std::pair<BindingKey, SVal> BindingPair;
145
146typedef llvm::ImmutableMap<const MemRegion *, ClusterBindings>
147        RegionBindings;
148
149namespace {
150class RegionBindingsRef : public llvm::ImmutableMapRef<const MemRegion *,
151                                 ClusterBindings> {
152 ClusterBindings::Factory &CBFactory;
153public:
154  typedef llvm::ImmutableMapRef<const MemRegion *, ClusterBindings>
155          ParentTy;
156
157  RegionBindingsRef(ClusterBindings::Factory &CBFactory,
158                    const RegionBindings::TreeTy *T,
159                    RegionBindings::TreeTy::Factory *F)
160    : llvm::ImmutableMapRef<const MemRegion *, ClusterBindings>(T, F),
161      CBFactory(CBFactory) {}
162
163  RegionBindingsRef(const ParentTy &P, ClusterBindings::Factory &CBFactory)
164    : llvm::ImmutableMapRef<const MemRegion *, ClusterBindings>(P),
165      CBFactory(CBFactory) {}
166
167  RegionBindingsRef add(key_type_ref K, data_type_ref D) const {
168    return RegionBindingsRef(static_cast<const ParentTy*>(this)->add(K, D),
169                             CBFactory);
170  }
171
172  RegionBindingsRef remove(key_type_ref K) const {
173    return RegionBindingsRef(static_cast<const ParentTy*>(this)->remove(K),
174                             CBFactory);
175  }
176
177  RegionBindingsRef addBinding(BindingKey K, SVal V) const;
178
179  RegionBindingsRef addBinding(const MemRegion *R,
180                               BindingKey::Kind k, SVal V) const;
181
182  RegionBindingsRef &operator=(const RegionBindingsRef &X) {
183    *static_cast<ParentTy*>(this) = X;
184    return *this;
185  }
186
187  const SVal *lookup(BindingKey K) const;
188  const SVal *lookup(const MemRegion *R, BindingKey::Kind k) const;
189  const ClusterBindings *lookup(const MemRegion *R) const {
190    return static_cast<const ParentTy*>(this)->lookup(R);
191  }
192
193  RegionBindingsRef removeBinding(BindingKey K);
194
195  RegionBindingsRef removeBinding(const MemRegion *R,
196                                  BindingKey::Kind k);
197
198  RegionBindingsRef removeBinding(const MemRegion *R) {
199    return removeBinding(R, BindingKey::Direct).
200           removeBinding(R, BindingKey::Default);
201  }
202
203  Optional<SVal> getDirectBinding(const MemRegion *R) const;
204
205  /// getDefaultBinding - Returns an SVal* representing an optional default
206  ///  binding associated with a region and its subregions.
207  Optional<SVal> getDefaultBinding(const MemRegion *R) const;
208
209  /// Return the internal tree as a Store.
210  Store asStore() const {
211    return asImmutableMap().getRootWithoutRetain();
212  }
213
214  void dump(raw_ostream &OS, const char *nl) const {
215   for (iterator I = begin(), E = end(); I != E; ++I) {
216     const ClusterBindings &Cluster = I.getData();
217     for (ClusterBindings::iterator CI = Cluster.begin(), CE = Cluster.end();
218          CI != CE; ++CI) {
219       OS << ' ' << CI.getKey() << " : " << CI.getData() << nl;
220     }
221     OS << nl;
222   }
223  }
224
225  LLVM_ATTRIBUTE_USED void dump() const {
226    dump(llvm::errs(), "\n");
227  }
228};
229} // end anonymous namespace
230
231typedef const RegionBindingsRef& RegionBindingsConstRef;
232
233Optional<SVal> RegionBindingsRef::getDirectBinding(const MemRegion *R) const {
234  return Optional<SVal>::create(lookup(R, BindingKey::Direct));
235}
236
237Optional<SVal> RegionBindingsRef::getDefaultBinding(const MemRegion *R) const {
238  if (R->isBoundable())
239    if (const TypedValueRegion *TR = dyn_cast<TypedValueRegion>(R))
240      if (TR->getValueType()->isUnionType())
241        return UnknownVal();
242
243  return Optional<SVal>::create(lookup(R, BindingKey::Default));
244}
245
246RegionBindingsRef RegionBindingsRef::addBinding(BindingKey K, SVal V) const {
247  const MemRegion *Base = K.getBaseRegion();
248
249  const ClusterBindings *ExistingCluster = lookup(Base);
250  ClusterBindings Cluster = (ExistingCluster ? *ExistingCluster
251                             : CBFactory.getEmptyMap());
252
253  ClusterBindings NewCluster = CBFactory.add(Cluster, K, V);
254  return add(Base, NewCluster);
255}
256
257
258RegionBindingsRef RegionBindingsRef::addBinding(const MemRegion *R,
259                                                BindingKey::Kind k,
260                                                SVal V) const {
261  return addBinding(BindingKey::Make(R, k), V);
262}
263
264const SVal *RegionBindingsRef::lookup(BindingKey K) const {
265  const ClusterBindings *Cluster = lookup(K.getBaseRegion());
266  if (!Cluster)
267    return 0;
268  return Cluster->lookup(K);
269}
270
271const SVal *RegionBindingsRef::lookup(const MemRegion *R,
272                                      BindingKey::Kind k) const {
273  return lookup(BindingKey::Make(R, k));
274}
275
276RegionBindingsRef RegionBindingsRef::removeBinding(BindingKey K) {
277  const MemRegion *Base = K.getBaseRegion();
278  const ClusterBindings *Cluster = lookup(Base);
279  if (!Cluster)
280    return *this;
281
282  ClusterBindings NewCluster = CBFactory.remove(*Cluster, K);
283  if (NewCluster.isEmpty())
284    return remove(Base);
285  return add(Base, NewCluster);
286}
287
288RegionBindingsRef RegionBindingsRef::removeBinding(const MemRegion *R,
289                                                BindingKey::Kind k){
290  return removeBinding(BindingKey::Make(R, k));
291}
292
293//===----------------------------------------------------------------------===//
294// Fine-grained control of RegionStoreManager.
295//===----------------------------------------------------------------------===//
296
297namespace {
298struct minimal_features_tag {};
299struct maximal_features_tag {};
300
301class RegionStoreFeatures {
302  bool SupportsFields;
303public:
304  RegionStoreFeatures(minimal_features_tag) :
305    SupportsFields(false) {}
306
307  RegionStoreFeatures(maximal_features_tag) :
308    SupportsFields(true) {}
309
310  void enableFields(bool t) { SupportsFields = t; }
311
312  bool supportsFields() const { return SupportsFields; }
313};
314}
315
316//===----------------------------------------------------------------------===//
317// Main RegionStore logic.
318//===----------------------------------------------------------------------===//
319
320namespace {
321class invalidateRegionsWorker;
322
323class RegionStoreManager : public StoreManager {
324public:
325  const RegionStoreFeatures Features;
326  RegionBindings::Factory RBFactory;
327  mutable ClusterBindings::Factory CBFactory;
328
329  typedef std::vector<SVal> SValListTy;
330private:
331  typedef llvm::DenseMap<const LazyCompoundValData *,
332                         SValListTy> LazyBindingsMapTy;
333  LazyBindingsMapTy LazyBindingsMap;
334
335  /// \brief A helper used to populate the work list with the given set of
336  /// regions.
337  void populateWorkList(invalidateRegionsWorker &W,
338                        ArrayRef<SVal> Values,
339                        bool IsArrayOfConstRegions,
340                        InvalidatedRegions *TopLevelRegions);
341
342public:
343  RegionStoreManager(ProgramStateManager& mgr, const RegionStoreFeatures &f)
344    : StoreManager(mgr), Features(f),
345      RBFactory(mgr.getAllocator()), CBFactory(mgr.getAllocator()) {}
346
347
348  /// setImplicitDefaultValue - Set the default binding for the provided
349  ///  MemRegion to the value implicitly defined for compound literals when
350  ///  the value is not specified.
351  RegionBindingsRef setImplicitDefaultValue(RegionBindingsConstRef B,
352                                            const MemRegion *R, QualType T);
353
354  /// ArrayToPointer - Emulates the "decay" of an array to a pointer
355  ///  type.  'Array' represents the lvalue of the array being decayed
356  ///  to a pointer, and the returned SVal represents the decayed
357  ///  version of that lvalue (i.e., a pointer to the first element of
358  ///  the array).  This is called by ExprEngine when evaluating
359  ///  casts from arrays to pointers.
360  SVal ArrayToPointer(Loc Array);
361
362  StoreRef getInitialStore(const LocationContext *InitLoc) {
363    return StoreRef(RBFactory.getEmptyMap().getRootWithoutRetain(), *this);
364  }
365
366  //===-------------------------------------------------------------------===//
367  // Binding values to regions.
368  //===-------------------------------------------------------------------===//
369  RegionBindingsRef invalidateGlobalRegion(MemRegion::Kind K,
370                                           const Expr *Ex,
371                                           unsigned Count,
372                                           const LocationContext *LCtx,
373                                           RegionBindingsRef B,
374                                           InvalidatedRegions *Invalidated);
375
376  StoreRef invalidateRegions(Store store,
377                             ArrayRef<SVal> Values,
378                             ArrayRef<SVal> ConstValues,
379                             const Expr *E, unsigned Count,
380                             const LocationContext *LCtx,
381                             const CallEvent *Call,
382                             InvalidatedSymbols &IS,
383                             InvalidatedSymbols &ConstIS,
384                             InvalidatedRegions *Invalidated,
385                             InvalidatedRegions *InvalidatedTopLevel,
386                             InvalidatedRegions *InvalidatedTopLevelConst);
387
388  bool scanReachableSymbols(Store S, const MemRegion *R,
389                            ScanReachableSymbols &Callbacks);
390
391  RegionBindingsRef removeSubRegionBindings(RegionBindingsConstRef B,
392                                            const SubRegion *R);
393
394public: // Part of public interface to class.
395
396  virtual StoreRef Bind(Store store, Loc LV, SVal V) {
397    return StoreRef(bind(getRegionBindings(store), LV, V).asStore(), *this);
398  }
399
400  RegionBindingsRef bind(RegionBindingsConstRef B, Loc LV, SVal V);
401
402  // BindDefault is only used to initialize a region with a default value.
403  StoreRef BindDefault(Store store, const MemRegion *R, SVal V) {
404    RegionBindingsRef B = getRegionBindings(store);
405    assert(!B.lookup(R, BindingKey::Default));
406    assert(!B.lookup(R, BindingKey::Direct));
407    return StoreRef(B.addBinding(R, BindingKey::Default, V)
408                     .asImmutableMap()
409                     .getRootWithoutRetain(), *this);
410  }
411
412  /// \brief Create a new store that binds a value to a compound literal.
413  ///
414  /// \param ST The original store whose bindings are the basis for the new
415  ///        store.
416  ///
417  /// \param CL The compound literal to bind (the binding key).
418  ///
419  /// \param LC The LocationContext for the binding.
420  ///
421  /// \param V The value to bind to the compound literal.
422  StoreRef bindCompoundLiteral(Store ST,
423                               const CompoundLiteralExpr *CL,
424                               const LocationContext *LC, SVal V);
425
426  /// BindStruct - Bind a compound value to a structure.
427  RegionBindingsRef bindStruct(RegionBindingsConstRef B,
428                               const TypedValueRegion* R, SVal V);
429
430  /// BindVector - Bind a compound value to a vector.
431  RegionBindingsRef bindVector(RegionBindingsConstRef B,
432                               const TypedValueRegion* R, SVal V);
433
434  RegionBindingsRef bindArray(RegionBindingsConstRef B,
435                              const TypedValueRegion* R,
436                              SVal V);
437
438  /// Clears out all bindings in the given region and assigns a new value
439  /// as a Default binding.
440  RegionBindingsRef bindAggregate(RegionBindingsConstRef B,
441                                  const TypedRegion *R,
442                                  SVal DefaultVal);
443
444  /// \brief Create a new store with the specified binding removed.
445  /// \param ST the original store, that is the basis for the new store.
446  /// \param L the location whose binding should be removed.
447  virtual StoreRef killBinding(Store ST, Loc L);
448
449  void incrementReferenceCount(Store store) {
450    getRegionBindings(store).manualRetain();
451  }
452
453  /// If the StoreManager supports it, decrement the reference count of
454  /// the specified Store object.  If the reference count hits 0, the memory
455  /// associated with the object is recycled.
456  void decrementReferenceCount(Store store) {
457    getRegionBindings(store).manualRelease();
458  }
459
460  bool includedInBindings(Store store, const MemRegion *region) const;
461
462  /// \brief Return the value bound to specified location in a given state.
463  ///
464  /// The high level logic for this method is this:
465  /// getBinding (L)
466  ///   if L has binding
467  ///     return L's binding
468  ///   else if L is in killset
469  ///     return unknown
470  ///   else
471  ///     if L is on stack or heap
472  ///       return undefined
473  ///     else
474  ///       return symbolic
475  virtual SVal getBinding(Store S, Loc L, QualType T) {
476    return getBinding(getRegionBindings(S), L, T);
477  }
478
479  SVal getBinding(RegionBindingsConstRef B, Loc L, QualType T = QualType());
480
481  SVal getBindingForElement(RegionBindingsConstRef B, const ElementRegion *R);
482
483  SVal getBindingForField(RegionBindingsConstRef B, const FieldRegion *R);
484
485  SVal getBindingForObjCIvar(RegionBindingsConstRef B, const ObjCIvarRegion *R);
486
487  SVal getBindingForVar(RegionBindingsConstRef B, const VarRegion *R);
488
489  SVal getBindingForLazySymbol(const TypedValueRegion *R);
490
491  SVal getBindingForFieldOrElementCommon(RegionBindingsConstRef B,
492                                         const TypedValueRegion *R,
493                                         QualType Ty,
494                                         const MemRegion *superR);
495
496  SVal getLazyBinding(const SubRegion *LazyBindingRegion,
497                      RegionBindingsRef LazyBinding);
498
499  /// Get bindings for the values in a struct and return a CompoundVal, used
500  /// when doing struct copy:
501  /// struct s x, y;
502  /// x = y;
503  /// y's value is retrieved by this method.
504  SVal getBindingForStruct(RegionBindingsConstRef B, const TypedValueRegion *R);
505  SVal getBindingForArray(RegionBindingsConstRef B, const TypedValueRegion *R);
506  NonLoc createLazyBinding(RegionBindingsConstRef B, const TypedValueRegion *R);
507
508  /// Used to lazily generate derived symbols for bindings that are defined
509  /// implicitly by default bindings in a super region.
510  ///
511  /// Note that callers may need to specially handle LazyCompoundVals, which
512  /// are returned as is in case the caller needs to treat them differently.
513  Optional<SVal> getBindingForDerivedDefaultValue(RegionBindingsConstRef B,
514                                                  const MemRegion *superR,
515                                                  const TypedValueRegion *R,
516                                                  QualType Ty);
517
518  /// Get the state and region whose binding this region \p R corresponds to.
519  ///
520  /// If there is no lazy binding for \p R, the returned value will have a null
521  /// \c second. Note that a null pointer can represents a valid Store.
522  std::pair<Store, const SubRegion *>
523  findLazyBinding(RegionBindingsConstRef B, const SubRegion *R,
524                  const SubRegion *originalRegion);
525
526  /// Returns the cached set of interesting SVals contained within a lazy
527  /// binding.
528  ///
529  /// The precise value of "interesting" is determined for the purposes of
530  /// RegionStore's internal analysis. It must always contain all regions and
531  /// symbols, but may omit constants and other kinds of SVal.
532  const SValListTy &getInterestingValues(nonloc::LazyCompoundVal LCV);
533
534  //===------------------------------------------------------------------===//
535  // State pruning.
536  //===------------------------------------------------------------------===//
537
538  /// removeDeadBindings - Scans the RegionStore of 'state' for dead values.
539  ///  It returns a new Store with these values removed.
540  StoreRef removeDeadBindings(Store store, const StackFrameContext *LCtx,
541                              SymbolReaper& SymReaper);
542
543  //===------------------------------------------------------------------===//
544  // Region "extents".
545  //===------------------------------------------------------------------===//
546
547  // FIXME: This method will soon be eliminated; see the note in Store.h.
548  DefinedOrUnknownSVal getSizeInElements(ProgramStateRef state,
549                                         const MemRegion* R, QualType EleTy);
550
551  //===------------------------------------------------------------------===//
552  // Utility methods.
553  //===------------------------------------------------------------------===//
554
555  RegionBindingsRef getRegionBindings(Store store) const {
556    return RegionBindingsRef(CBFactory,
557                             static_cast<const RegionBindings::TreeTy*>(store),
558                             RBFactory.getTreeFactory());
559  }
560
561  void print(Store store, raw_ostream &Out, const char* nl,
562             const char *sep);
563
564  void iterBindings(Store store, BindingsHandler& f) {
565    RegionBindingsRef B = getRegionBindings(store);
566    for (RegionBindingsRef::iterator I = B.begin(), E = B.end(); I != E; ++I) {
567      const ClusterBindings &Cluster = I.getData();
568      for (ClusterBindings::iterator CI = Cluster.begin(), CE = Cluster.end();
569           CI != CE; ++CI) {
570        const BindingKey &K = CI.getKey();
571        if (!K.isDirect())
572          continue;
573        if (const SubRegion *R = dyn_cast<SubRegion>(K.getRegion())) {
574          // FIXME: Possibly incorporate the offset?
575          if (!f.HandleBinding(*this, store, R, CI.getData()))
576            return;
577        }
578      }
579    }
580  }
581};
582
583} // end anonymous namespace
584
585//===----------------------------------------------------------------------===//
586// RegionStore creation.
587//===----------------------------------------------------------------------===//
588
589StoreManager *ento::CreateRegionStoreManager(ProgramStateManager& StMgr) {
590  RegionStoreFeatures F = maximal_features_tag();
591  return new RegionStoreManager(StMgr, F);
592}
593
594StoreManager *
595ento::CreateFieldsOnlyRegionStoreManager(ProgramStateManager &StMgr) {
596  RegionStoreFeatures F = minimal_features_tag();
597  F.enableFields(true);
598  return new RegionStoreManager(StMgr, F);
599}
600
601
602//===----------------------------------------------------------------------===//
603// Region Cluster analysis.
604//===----------------------------------------------------------------------===//
605
606namespace {
607/// Used to determine which global regions are automatically included in the
608/// initial worklist of a ClusterAnalysis.
609enum GlobalsFilterKind {
610  /// Don't include any global regions.
611  GFK_None,
612  /// Only include system globals.
613  GFK_SystemOnly,
614  /// Include all global regions.
615  GFK_All
616};
617
618template <typename DERIVED>
619class ClusterAnalysis  {
620protected:
621  typedef llvm::DenseMap<const MemRegion *, const ClusterBindings *> ClusterMap;
622  typedef llvm::PointerIntPair<const MemRegion *, 1, bool> WorkListElement;
623  typedef SmallVector<WorkListElement, 10> WorkList;
624
625  llvm::SmallPtrSet<const ClusterBindings *, 16> Visited;
626
627  WorkList WL;
628
629  RegionStoreManager &RM;
630  ASTContext &Ctx;
631  SValBuilder &svalBuilder;
632
633  RegionBindingsRef B;
634
635private:
636  GlobalsFilterKind GlobalsFilter;
637
638protected:
639  const ClusterBindings *getCluster(const MemRegion *R) {
640    return B.lookup(R);
641  }
642
643  /// Returns true if the memory space of the given region is one of the global
644  /// regions specially included at the start of analysis.
645  bool isInitiallyIncludedGlobalRegion(const MemRegion *R) {
646    switch (GlobalsFilter) {
647    case GFK_None:
648      return false;
649    case GFK_SystemOnly:
650      return isa<GlobalSystemSpaceRegion>(R->getMemorySpace());
651    case GFK_All:
652      return isa<NonStaticGlobalSpaceRegion>(R->getMemorySpace());
653    }
654
655    llvm_unreachable("unknown globals filter");
656  }
657
658public:
659  ClusterAnalysis(RegionStoreManager &rm, ProgramStateManager &StateMgr,
660                  RegionBindingsRef b, GlobalsFilterKind GFK)
661    : RM(rm), Ctx(StateMgr.getContext()),
662      svalBuilder(StateMgr.getSValBuilder()),
663      B(b), GlobalsFilter(GFK) {}
664
665  RegionBindingsRef getRegionBindings() const { return B; }
666
667  bool isVisited(const MemRegion *R) {
668    return Visited.count(getCluster(R));
669  }
670
671  void GenerateClusters() {
672    // Scan the entire set of bindings and record the region clusters.
673    for (RegionBindingsRef::iterator RI = B.begin(), RE = B.end();
674         RI != RE; ++RI){
675      const MemRegion *Base = RI.getKey();
676
677      const ClusterBindings &Cluster = RI.getData();
678      assert(!Cluster.isEmpty() && "Empty clusters should be removed");
679      static_cast<DERIVED*>(this)->VisitAddedToCluster(Base, Cluster);
680
681      // If this is an interesting global region, add it the work list up front.
682      if (isInitiallyIncludedGlobalRegion(Base))
683        AddToWorkList(WorkListElement(Base), &Cluster);
684    }
685  }
686
687  bool AddToWorkList(WorkListElement E, const ClusterBindings *C) {
688    if (C && !Visited.insert(C))
689      return false;
690    WL.push_back(E);
691    return true;
692  }
693
694  bool AddToWorkList(const MemRegion *R, bool Flag = false) {
695    const MemRegion *BaseR = R->getBaseRegion();
696    return AddToWorkList(WorkListElement(BaseR, Flag), getCluster(BaseR));
697  }
698
699  void RunWorkList() {
700    while (!WL.empty()) {
701      WorkListElement E = WL.pop_back_val();
702      const MemRegion *BaseR = E.getPointer();
703
704      static_cast<DERIVED*>(this)->VisitCluster(BaseR, getCluster(BaseR),
705                                                E.getInt());
706    }
707  }
708
709  void VisitAddedToCluster(const MemRegion *baseR, const ClusterBindings &C) {}
710  void VisitCluster(const MemRegion *baseR, const ClusterBindings *C) {}
711
712  void VisitCluster(const MemRegion *BaseR, const ClusterBindings *C,
713                    bool Flag) {
714    static_cast<DERIVED*>(this)->VisitCluster(BaseR, C);
715  }
716};
717}
718
719//===----------------------------------------------------------------------===//
720// Binding invalidation.
721//===----------------------------------------------------------------------===//
722
723bool RegionStoreManager::scanReachableSymbols(Store S, const MemRegion *R,
724                                              ScanReachableSymbols &Callbacks) {
725  assert(R == R->getBaseRegion() && "Should only be called for base regions");
726  RegionBindingsRef B = getRegionBindings(S);
727  const ClusterBindings *Cluster = B.lookup(R);
728
729  if (!Cluster)
730    return true;
731
732  for (ClusterBindings::iterator RI = Cluster->begin(), RE = Cluster->end();
733       RI != RE; ++RI) {
734    if (!Callbacks.scan(RI.getData()))
735      return false;
736  }
737
738  return true;
739}
740
741static inline bool isUnionField(const FieldRegion *FR) {
742  return FR->getDecl()->getParent()->isUnion();
743}
744
745typedef SmallVector<const FieldDecl *, 8> FieldVector;
746
747void getSymbolicOffsetFields(BindingKey K, FieldVector &Fields) {
748  assert(K.hasSymbolicOffset() && "Not implemented for concrete offset keys");
749
750  const MemRegion *Base = K.getConcreteOffsetRegion();
751  const MemRegion *R = K.getRegion();
752
753  while (R != Base) {
754    if (const FieldRegion *FR = dyn_cast<FieldRegion>(R))
755      if (!isUnionField(FR))
756        Fields.push_back(FR->getDecl());
757
758    R = cast<SubRegion>(R)->getSuperRegion();
759  }
760}
761
762static bool isCompatibleWithFields(BindingKey K, const FieldVector &Fields) {
763  assert(K.hasSymbolicOffset() && "Not implemented for concrete offset keys");
764
765  if (Fields.empty())
766    return true;
767
768  FieldVector FieldsInBindingKey;
769  getSymbolicOffsetFields(K, FieldsInBindingKey);
770
771  ptrdiff_t Delta = FieldsInBindingKey.size() - Fields.size();
772  if (Delta >= 0)
773    return std::equal(FieldsInBindingKey.begin() + Delta,
774                      FieldsInBindingKey.end(),
775                      Fields.begin());
776  else
777    return std::equal(FieldsInBindingKey.begin(), FieldsInBindingKey.end(),
778                      Fields.begin() - Delta);
779}
780
781/// Collects all bindings in \p Cluster that may refer to bindings within
782/// \p Top.
783///
784/// Each binding is a pair whose \c first is the key (a BindingKey) and whose
785/// \c second is the value (an SVal).
786///
787/// The \p IncludeAllDefaultBindings parameter specifies whether to include
788/// default bindings that may extend beyond \p Top itself, e.g. if \p Top is
789/// an aggregate within a larger aggregate with a default binding.
790static void
791collectSubRegionBindings(SmallVectorImpl<BindingPair> &Bindings,
792                         SValBuilder &SVB, const ClusterBindings &Cluster,
793                         const SubRegion *Top, BindingKey TopKey,
794                         bool IncludeAllDefaultBindings) {
795  FieldVector FieldsInSymbolicSubregions;
796  if (TopKey.hasSymbolicOffset()) {
797    getSymbolicOffsetFields(TopKey, FieldsInSymbolicSubregions);
798    Top = cast<SubRegion>(TopKey.getConcreteOffsetRegion());
799    TopKey = BindingKey::Make(Top, BindingKey::Default);
800  }
801
802  // Find the length (in bits) of the region being invalidated.
803  uint64_t Length = UINT64_MAX;
804  SVal Extent = Top->getExtent(SVB);
805  if (Optional<nonloc::ConcreteInt> ExtentCI =
806          Extent.getAs<nonloc::ConcreteInt>()) {
807    const llvm::APSInt &ExtentInt = ExtentCI->getValue();
808    assert(ExtentInt.isNonNegative() || ExtentInt.isUnsigned());
809    // Extents are in bytes but region offsets are in bits. Be careful!
810    Length = ExtentInt.getLimitedValue() * SVB.getContext().getCharWidth();
811  } else if (const FieldRegion *FR = dyn_cast<FieldRegion>(Top)) {
812    if (FR->getDecl()->isBitField())
813      Length = FR->getDecl()->getBitWidthValue(SVB.getContext());
814  }
815
816  for (ClusterBindings::iterator I = Cluster.begin(), E = Cluster.end();
817       I != E; ++I) {
818    BindingKey NextKey = I.getKey();
819    if (NextKey.getRegion() == TopKey.getRegion()) {
820      // FIXME: This doesn't catch the case where we're really invalidating a
821      // region with a symbolic offset. Example:
822      //      R: points[i].y
823      //   Next: points[0].x
824
825      if (NextKey.getOffset() > TopKey.getOffset() &&
826          NextKey.getOffset() - TopKey.getOffset() < Length) {
827        // Case 1: The next binding is inside the region we're invalidating.
828        // Include it.
829        Bindings.push_back(*I);
830
831      } else if (NextKey.getOffset() == TopKey.getOffset()) {
832        // Case 2: The next binding is at the same offset as the region we're
833        // invalidating. In this case, we need to leave default bindings alone,
834        // since they may be providing a default value for a regions beyond what
835        // we're invalidating.
836        // FIXME: This is probably incorrect; consider invalidating an outer
837        // struct whose first field is bound to a LazyCompoundVal.
838        if (IncludeAllDefaultBindings || NextKey.isDirect())
839          Bindings.push_back(*I);
840      }
841
842    } else if (NextKey.hasSymbolicOffset()) {
843      const MemRegion *Base = NextKey.getConcreteOffsetRegion();
844      if (Top->isSubRegionOf(Base)) {
845        // Case 3: The next key is symbolic and we just changed something within
846        // its concrete region. We don't know if the binding is still valid, so
847        // we'll be conservative and include it.
848        if (IncludeAllDefaultBindings || NextKey.isDirect())
849          if (isCompatibleWithFields(NextKey, FieldsInSymbolicSubregions))
850            Bindings.push_back(*I);
851      } else if (const SubRegion *BaseSR = dyn_cast<SubRegion>(Base)) {
852        // Case 4: The next key is symbolic, but we changed a known
853        // super-region. In this case the binding is certainly included.
854        if (Top == Base || BaseSR->isSubRegionOf(Top))
855          if (isCompatibleWithFields(NextKey, FieldsInSymbolicSubregions))
856            Bindings.push_back(*I);
857      }
858    }
859  }
860}
861
862static void
863collectSubRegionBindings(SmallVectorImpl<BindingPair> &Bindings,
864                         SValBuilder &SVB, const ClusterBindings &Cluster,
865                         const SubRegion *Top, bool IncludeAllDefaultBindings) {
866  collectSubRegionBindings(Bindings, SVB, Cluster, Top,
867                           BindingKey::Make(Top, BindingKey::Default),
868                           IncludeAllDefaultBindings);
869}
870
871RegionBindingsRef
872RegionStoreManager::removeSubRegionBindings(RegionBindingsConstRef B,
873                                            const SubRegion *Top) {
874  BindingKey TopKey = BindingKey::Make(Top, BindingKey::Default);
875  const MemRegion *ClusterHead = TopKey.getBaseRegion();
876
877  if (Top == ClusterHead) {
878    // We can remove an entire cluster's bindings all in one go.
879    return B.remove(Top);
880  }
881
882  const ClusterBindings *Cluster = B.lookup(ClusterHead);
883  if (!Cluster) {
884    // If we're invalidating a region with a symbolic offset, we need to make
885    // sure we don't treat the base region as uninitialized anymore.
886    if (TopKey.hasSymbolicOffset()) {
887      const SubRegion *Concrete = TopKey.getConcreteOffsetRegion();
888      return B.addBinding(Concrete, BindingKey::Default, UnknownVal());
889    }
890    return B;
891  }
892
893  SmallVector<BindingPair, 32> Bindings;
894  collectSubRegionBindings(Bindings, svalBuilder, *Cluster, Top, TopKey,
895                           /*IncludeAllDefaultBindings=*/false);
896
897  ClusterBindingsRef Result(*Cluster, CBFactory);
898  for (SmallVectorImpl<BindingPair>::const_iterator I = Bindings.begin(),
899                                                    E = Bindings.end();
900       I != E; ++I)
901    Result = Result.remove(I->first);
902
903  // If we're invalidating a region with a symbolic offset, we need to make sure
904  // we don't treat the base region as uninitialized anymore.
905  // FIXME: This isn't very precise; see the example in
906  // collectSubRegionBindings.
907  if (TopKey.hasSymbolicOffset()) {
908    const SubRegion *Concrete = TopKey.getConcreteOffsetRegion();
909    Result = Result.add(BindingKey::Make(Concrete, BindingKey::Default),
910                        UnknownVal());
911  }
912
913  if (Result.isEmpty())
914    return B.remove(ClusterHead);
915  return B.add(ClusterHead, Result.asImmutableMap());
916}
917
918namespace {
919class invalidateRegionsWorker : public ClusterAnalysis<invalidateRegionsWorker>
920{
921  const Expr *Ex;
922  unsigned Count;
923  const LocationContext *LCtx;
924  InvalidatedSymbols &IS;
925  InvalidatedSymbols &ConstIS;
926  StoreManager::InvalidatedRegions *Regions;
927public:
928  invalidateRegionsWorker(RegionStoreManager &rm,
929                          ProgramStateManager &stateMgr,
930                          RegionBindingsRef b,
931                          const Expr *ex, unsigned count,
932                          const LocationContext *lctx,
933                          InvalidatedSymbols &is,
934                          InvalidatedSymbols &inConstIS,
935                          StoreManager::InvalidatedRegions *r,
936                          GlobalsFilterKind GFK)
937    : ClusterAnalysis<invalidateRegionsWorker>(rm, stateMgr, b, GFK),
938      Ex(ex), Count(count), LCtx(lctx), IS(is), ConstIS(inConstIS), Regions(r){}
939
940  /// \param IsConst Specifies if the region we are invalidating is constant.
941  /// If it is, we invalidate all subregions, but not the base region itself.
942  void VisitCluster(const MemRegion *baseR, const ClusterBindings *C,
943                    bool IsConst);
944  void VisitBinding(SVal V);
945};
946}
947
948void invalidateRegionsWorker::VisitBinding(SVal V) {
949  // A symbol?  Mark it touched by the invalidation.
950  if (SymbolRef Sym = V.getAsSymbol())
951    IS.insert(Sym);
952
953  if (const MemRegion *R = V.getAsRegion()) {
954    AddToWorkList(R);
955    return;
956  }
957
958  // Is it a LazyCompoundVal?  All references get invalidated as well.
959  if (Optional<nonloc::LazyCompoundVal> LCS =
960          V.getAs<nonloc::LazyCompoundVal>()) {
961
962    const RegionStoreManager::SValListTy &Vals = RM.getInterestingValues(*LCS);
963
964    for (RegionStoreManager::SValListTy::const_iterator I = Vals.begin(),
965                                                        E = Vals.end();
966         I != E; ++I)
967      VisitBinding(*I);
968
969    return;
970  }
971}
972
973void invalidateRegionsWorker::VisitCluster(const MemRegion *baseR,
974                                           const ClusterBindings *C,
975                                           bool IsConst) {
976  if (C) {
977    for (ClusterBindings::iterator I = C->begin(), E = C->end(); I != E; ++I)
978      VisitBinding(I.getData());
979
980    if (!IsConst)
981      B = B.remove(baseR);
982  }
983
984  // BlockDataRegion?  If so, invalidate captured variables that are passed
985  // by reference.
986  if (const BlockDataRegion *BR = dyn_cast<BlockDataRegion>(baseR)) {
987    for (BlockDataRegion::referenced_vars_iterator
988         BI = BR->referenced_vars_begin(), BE = BR->referenced_vars_end() ;
989         BI != BE; ++BI) {
990      const VarRegion *VR = BI.getCapturedRegion();
991      const VarDecl *VD = VR->getDecl();
992      if (VD->getAttr<BlocksAttr>() || !VD->hasLocalStorage()) {
993        AddToWorkList(VR);
994      }
995      else if (Loc::isLocType(VR->getValueType())) {
996        // Map the current bindings to a Store to retrieve the value
997        // of the binding.  If that binding itself is a region, we should
998        // invalidate that region.  This is because a block may capture
999        // a pointer value, but the thing pointed by that pointer may
1000        // get invalidated.
1001        SVal V = RM.getBinding(B, loc::MemRegionVal(VR));
1002        if (Optional<Loc> L = V.getAs<Loc>()) {
1003          if (const MemRegion *LR = L->getAsRegion())
1004            AddToWorkList(LR);
1005        }
1006      }
1007    }
1008    return;
1009  }
1010
1011  // Symbolic region?
1012  SymbolRef RegionSym = 0;
1013  if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(baseR))
1014    RegionSym = SR->getSymbol();
1015
1016  if (IsConst) {
1017    // Mark that symbol touched by the invalidation.
1018    ConstIS.insert(RegionSym);
1019    return;
1020  }
1021
1022  // Mark that symbol touched by the invalidation.
1023  IS.insert(RegionSym);
1024
1025  // Otherwise, we have a normal data region. Record that we touched the region.
1026  if (Regions)
1027    Regions->push_back(baseR);
1028
1029  if (isa<AllocaRegion>(baseR) || isa<SymbolicRegion>(baseR)) {
1030    // Invalidate the region by setting its default value to
1031    // conjured symbol. The type of the symbol is irrelavant.
1032    DefinedOrUnknownSVal V =
1033      svalBuilder.conjureSymbolVal(baseR, Ex, LCtx, Ctx.IntTy, Count);
1034    B = B.addBinding(baseR, BindingKey::Default, V);
1035    return;
1036  }
1037
1038  if (!baseR->isBoundable())
1039    return;
1040
1041  const TypedValueRegion *TR = cast<TypedValueRegion>(baseR);
1042  QualType T = TR->getValueType();
1043
1044    // Invalidate the binding.
1045  if (T->isStructureOrClassType()) {
1046    // Invalidate the region by setting its default value to
1047    // conjured symbol. The type of the symbol is irrelavant.
1048    DefinedOrUnknownSVal V = svalBuilder.conjureSymbolVal(baseR, Ex, LCtx,
1049                                                          Ctx.IntTy, Count);
1050    B = B.addBinding(baseR, BindingKey::Default, V);
1051    return;
1052  }
1053
1054  if (const ArrayType *AT = Ctx.getAsArrayType(T)) {
1055      // Set the default value of the array to conjured symbol.
1056    DefinedOrUnknownSVal V =
1057    svalBuilder.conjureSymbolVal(baseR, Ex, LCtx,
1058                                     AT->getElementType(), Count);
1059    B = B.addBinding(baseR, BindingKey::Default, V);
1060    return;
1061  }
1062
1063  if (isInitiallyIncludedGlobalRegion(baseR)) {
1064    // If the region is a global and we are invalidating all globals,
1065    // just erase the entry.  This causes all globals to be lazily
1066    // symbolicated from the same base symbol.
1067    B = B.removeBinding(baseR);
1068    return;
1069  }
1070
1071
1072  DefinedOrUnknownSVal V = svalBuilder.conjureSymbolVal(baseR, Ex, LCtx,
1073                                                        T,Count);
1074  assert(SymbolManager::canSymbolicate(T) || V.isUnknown());
1075  B = B.addBinding(baseR, BindingKey::Direct, V);
1076}
1077
1078RegionBindingsRef
1079RegionStoreManager::invalidateGlobalRegion(MemRegion::Kind K,
1080                                           const Expr *Ex,
1081                                           unsigned Count,
1082                                           const LocationContext *LCtx,
1083                                           RegionBindingsRef B,
1084                                           InvalidatedRegions *Invalidated) {
1085  // Bind the globals memory space to a new symbol that we will use to derive
1086  // the bindings for all globals.
1087  const GlobalsSpaceRegion *GS = MRMgr.getGlobalsRegion(K);
1088  SVal V = svalBuilder.conjureSymbolVal(/* SymbolTag = */ (const void*) GS, Ex, LCtx,
1089                                        /* type does not matter */ Ctx.IntTy,
1090                                        Count);
1091
1092  B = B.removeBinding(GS)
1093       .addBinding(BindingKey::Make(GS, BindingKey::Default), V);
1094
1095  // Even if there are no bindings in the global scope, we still need to
1096  // record that we touched it.
1097  if (Invalidated)
1098    Invalidated->push_back(GS);
1099
1100  return B;
1101}
1102
1103void RegionStoreManager::populateWorkList(invalidateRegionsWorker &W,
1104                                          ArrayRef<SVal> Values,
1105                                          bool IsArrayOfConstRegions,
1106                                          InvalidatedRegions *TopLevelRegions) {
1107  for (ArrayRef<SVal>::iterator I = Values.begin(),
1108                                E = Values.end(); I != E; ++I) {
1109    SVal V = *I;
1110    if (Optional<nonloc::LazyCompoundVal> LCS =
1111        V.getAs<nonloc::LazyCompoundVal>()) {
1112
1113      const SValListTy &Vals = getInterestingValues(*LCS);
1114
1115      for (SValListTy::const_iterator I = Vals.begin(),
1116                                      E = Vals.end(); I != E; ++I) {
1117        // Note: the last argument is false here because these are
1118        // non-top-level regions.
1119        if (const MemRegion *R = (*I).getAsRegion())
1120          W.AddToWorkList(R, /*IsConst=*/ false);
1121      }
1122      continue;
1123    }
1124
1125    if (const MemRegion *R = V.getAsRegion()) {
1126      if (TopLevelRegions)
1127        TopLevelRegions->push_back(R);
1128      W.AddToWorkList(R, /*IsConst=*/ IsArrayOfConstRegions);
1129      continue;
1130    }
1131  }
1132}
1133
1134StoreRef
1135RegionStoreManager::invalidateRegions(Store store,
1136                                      ArrayRef<SVal> Values,
1137                                      ArrayRef<SVal> ConstValues,
1138                                      const Expr *Ex, unsigned Count,
1139                                      const LocationContext *LCtx,
1140                                      const CallEvent *Call,
1141                                      InvalidatedSymbols &IS,
1142                                      InvalidatedSymbols &ConstIS,
1143                                      InvalidatedRegions *TopLevelRegions,
1144                                      InvalidatedRegions *TopLevelConstRegions,
1145                                      InvalidatedRegions *Invalidated) {
1146  GlobalsFilterKind GlobalsFilter;
1147  if (Call) {
1148    if (Call->isInSystemHeader())
1149      GlobalsFilter = GFK_SystemOnly;
1150    else
1151      GlobalsFilter = GFK_All;
1152  } else {
1153    GlobalsFilter = GFK_None;
1154  }
1155
1156  RegionBindingsRef B = getRegionBindings(store);
1157  invalidateRegionsWorker W(*this, StateMgr, B, Ex, Count, LCtx, IS, ConstIS,
1158                            Invalidated, GlobalsFilter);
1159
1160  // Scan the bindings and generate the clusters.
1161  W.GenerateClusters();
1162
1163  // Add the regions to the worklist.
1164  populateWorkList(W, Values, /*IsArrayOfConstRegions*/ false,
1165                   TopLevelRegions);
1166  populateWorkList(W, ConstValues, /*IsArrayOfConstRegions*/ true,
1167                   TopLevelConstRegions);
1168
1169  W.RunWorkList();
1170
1171  // Return the new bindings.
1172  B = W.getRegionBindings();
1173
1174  // For calls, determine which global regions should be invalidated and
1175  // invalidate them. (Note that function-static and immutable globals are never
1176  // invalidated by this.)
1177  // TODO: This could possibly be more precise with modules.
1178  switch (GlobalsFilter) {
1179  case GFK_All:
1180    B = invalidateGlobalRegion(MemRegion::GlobalInternalSpaceRegionKind,
1181                               Ex, Count, LCtx, B, Invalidated);
1182    // FALLTHROUGH
1183  case GFK_SystemOnly:
1184    B = invalidateGlobalRegion(MemRegion::GlobalSystemSpaceRegionKind,
1185                               Ex, Count, LCtx, B, Invalidated);
1186    // FALLTHROUGH
1187  case GFK_None:
1188    break;
1189  }
1190
1191  return StoreRef(B.asStore(), *this);
1192}
1193
1194//===----------------------------------------------------------------------===//
1195// Extents for regions.
1196//===----------------------------------------------------------------------===//
1197
1198DefinedOrUnknownSVal
1199RegionStoreManager::getSizeInElements(ProgramStateRef state,
1200                                      const MemRegion *R,
1201                                      QualType EleTy) {
1202  SVal Size = cast<SubRegion>(R)->getExtent(svalBuilder);
1203  const llvm::APSInt *SizeInt = svalBuilder.getKnownValue(state, Size);
1204  if (!SizeInt)
1205    return UnknownVal();
1206
1207  CharUnits RegionSize = CharUnits::fromQuantity(SizeInt->getSExtValue());
1208
1209  if (Ctx.getAsVariableArrayType(EleTy)) {
1210    // FIXME: We need to track extra state to properly record the size
1211    // of VLAs.  Returning UnknownVal here, however, is a stop-gap so that
1212    // we don't have a divide-by-zero below.
1213    return UnknownVal();
1214  }
1215
1216  CharUnits EleSize = Ctx.getTypeSizeInChars(EleTy);
1217
1218  // If a variable is reinterpreted as a type that doesn't fit into a larger
1219  // type evenly, round it down.
1220  // This is a signed value, since it's used in arithmetic with signed indices.
1221  return svalBuilder.makeIntVal(RegionSize / EleSize, false);
1222}
1223
1224//===----------------------------------------------------------------------===//
1225// Location and region casting.
1226//===----------------------------------------------------------------------===//
1227
1228/// ArrayToPointer - Emulates the "decay" of an array to a pointer
1229///  type.  'Array' represents the lvalue of the array being decayed
1230///  to a pointer, and the returned SVal represents the decayed
1231///  version of that lvalue (i.e., a pointer to the first element of
1232///  the array).  This is called by ExprEngine when evaluating casts
1233///  from arrays to pointers.
1234SVal RegionStoreManager::ArrayToPointer(Loc Array) {
1235  if (!Array.getAs<loc::MemRegionVal>())
1236    return UnknownVal();
1237
1238  const MemRegion* R = Array.castAs<loc::MemRegionVal>().getRegion();
1239  const TypedValueRegion* ArrayR = dyn_cast<TypedValueRegion>(R);
1240
1241  if (!ArrayR)
1242    return UnknownVal();
1243
1244  // Strip off typedefs from the ArrayRegion's ValueType.
1245  QualType T = ArrayR->getValueType().getDesugaredType(Ctx);
1246  const ArrayType *AT = cast<ArrayType>(T);
1247  T = AT->getElementType();
1248
1249  NonLoc ZeroIdx = svalBuilder.makeZeroArrayIndex();
1250  return loc::MemRegionVal(MRMgr.getElementRegion(T, ZeroIdx, ArrayR, Ctx));
1251}
1252
1253//===----------------------------------------------------------------------===//
1254// Loading values from regions.
1255//===----------------------------------------------------------------------===//
1256
1257SVal RegionStoreManager::getBinding(RegionBindingsConstRef B, Loc L, QualType T) {
1258  assert(!L.getAs<UnknownVal>() && "location unknown");
1259  assert(!L.getAs<UndefinedVal>() && "location undefined");
1260
1261  // For access to concrete addresses, return UnknownVal.  Checks
1262  // for null dereferences (and similar errors) are done by checkers, not
1263  // the Store.
1264  // FIXME: We can consider lazily symbolicating such memory, but we really
1265  // should defer this when we can reason easily about symbolicating arrays
1266  // of bytes.
1267  if (L.getAs<loc::ConcreteInt>()) {
1268    return UnknownVal();
1269  }
1270  if (!L.getAs<loc::MemRegionVal>()) {
1271    return UnknownVal();
1272  }
1273
1274  const MemRegion *MR = L.castAs<loc::MemRegionVal>().getRegion();
1275
1276  if (isa<AllocaRegion>(MR) ||
1277      isa<SymbolicRegion>(MR) ||
1278      isa<CodeTextRegion>(MR)) {
1279    if (T.isNull()) {
1280      if (const TypedRegion *TR = dyn_cast<TypedRegion>(MR))
1281        T = TR->getLocationType();
1282      else {
1283        const SymbolicRegion *SR = cast<SymbolicRegion>(MR);
1284        T = SR->getSymbol()->getType();
1285      }
1286    }
1287    MR = GetElementZeroRegion(MR, T);
1288  }
1289
1290  // FIXME: Perhaps this method should just take a 'const MemRegion*' argument
1291  //  instead of 'Loc', and have the other Loc cases handled at a higher level.
1292  const TypedValueRegion *R = cast<TypedValueRegion>(MR);
1293  QualType RTy = R->getValueType();
1294
1295  // FIXME: we do not yet model the parts of a complex type, so treat the
1296  // whole thing as "unknown".
1297  if (RTy->isAnyComplexType())
1298    return UnknownVal();
1299
1300  // FIXME: We should eventually handle funny addressing.  e.g.:
1301  //
1302  //   int x = ...;
1303  //   int *p = &x;
1304  //   char *q = (char*) p;
1305  //   char c = *q;  // returns the first byte of 'x'.
1306  //
1307  // Such funny addressing will occur due to layering of regions.
1308  if (RTy->isStructureOrClassType())
1309    return getBindingForStruct(B, R);
1310
1311  // FIXME: Handle unions.
1312  if (RTy->isUnionType())
1313    return UnknownVal();
1314
1315  if (RTy->isArrayType()) {
1316    if (RTy->isConstantArrayType())
1317      return getBindingForArray(B, R);
1318    else
1319      return UnknownVal();
1320  }
1321
1322  // FIXME: handle Vector types.
1323  if (RTy->isVectorType())
1324    return UnknownVal();
1325
1326  if (const FieldRegion* FR = dyn_cast<FieldRegion>(R))
1327    return CastRetrievedVal(getBindingForField(B, FR), FR, T, false);
1328
1329  if (const ElementRegion* ER = dyn_cast<ElementRegion>(R)) {
1330    // FIXME: Here we actually perform an implicit conversion from the loaded
1331    // value to the element type.  Eventually we want to compose these values
1332    // more intelligently.  For example, an 'element' can encompass multiple
1333    // bound regions (e.g., several bound bytes), or could be a subset of
1334    // a larger value.
1335    return CastRetrievedVal(getBindingForElement(B, ER), ER, T, false);
1336  }
1337
1338  if (const ObjCIvarRegion *IVR = dyn_cast<ObjCIvarRegion>(R)) {
1339    // FIXME: Here we actually perform an implicit conversion from the loaded
1340    // value to the ivar type.  What we should model is stores to ivars
1341    // that blow past the extent of the ivar.  If the address of the ivar is
1342    // reinterpretted, it is possible we stored a different value that could
1343    // fit within the ivar.  Either we need to cast these when storing them
1344    // or reinterpret them lazily (as we do here).
1345    return CastRetrievedVal(getBindingForObjCIvar(B, IVR), IVR, T, false);
1346  }
1347
1348  if (const VarRegion *VR = dyn_cast<VarRegion>(R)) {
1349    // FIXME: Here we actually perform an implicit conversion from the loaded
1350    // value to the variable type.  What we should model is stores to variables
1351    // that blow past the extent of the variable.  If the address of the
1352    // variable is reinterpretted, it is possible we stored a different value
1353    // that could fit within the variable.  Either we need to cast these when
1354    // storing them or reinterpret them lazily (as we do here).
1355    return CastRetrievedVal(getBindingForVar(B, VR), VR, T, false);
1356  }
1357
1358  const SVal *V = B.lookup(R, BindingKey::Direct);
1359
1360  // Check if the region has a binding.
1361  if (V)
1362    return *V;
1363
1364  // The location does not have a bound value.  This means that it has
1365  // the value it had upon its creation and/or entry to the analyzed
1366  // function/method.  These are either symbolic values or 'undefined'.
1367  if (R->hasStackNonParametersStorage()) {
1368    // All stack variables are considered to have undefined values
1369    // upon creation.  All heap allocated blocks are considered to
1370    // have undefined values as well unless they are explicitly bound
1371    // to specific values.
1372    return UndefinedVal();
1373  }
1374
1375  // All other values are symbolic.
1376  return svalBuilder.getRegionValueSymbolVal(R);
1377}
1378
1379static QualType getUnderlyingType(const SubRegion *R) {
1380  QualType RegionTy;
1381  if (const TypedValueRegion *TVR = dyn_cast<TypedValueRegion>(R))
1382    RegionTy = TVR->getValueType();
1383
1384  if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(R))
1385    RegionTy = SR->getSymbol()->getType();
1386
1387  return RegionTy;
1388}
1389
1390/// Checks to see if store \p B has a lazy binding for region \p R.
1391///
1392/// If \p AllowSubregionBindings is \c false, a lazy binding will be rejected
1393/// if there are additional bindings within \p R.
1394///
1395/// Note that unlike RegionStoreManager::findLazyBinding, this will not search
1396/// for lazy bindings for super-regions of \p R.
1397static Optional<nonloc::LazyCompoundVal>
1398getExistingLazyBinding(SValBuilder &SVB, RegionBindingsConstRef B,
1399                       const SubRegion *R, bool AllowSubregionBindings) {
1400  Optional<SVal> V = B.getDefaultBinding(R);
1401  if (!V)
1402    return None;
1403
1404  Optional<nonloc::LazyCompoundVal> LCV = V->getAs<nonloc::LazyCompoundVal>();
1405  if (!LCV)
1406    return None;
1407
1408  // If the LCV is for a subregion, the types might not match, and we shouldn't
1409  // reuse the binding.
1410  QualType RegionTy = getUnderlyingType(R);
1411  if (!RegionTy.isNull() &&
1412      !RegionTy->isVoidPointerType()) {
1413    QualType SourceRegionTy = LCV->getRegion()->getValueType();
1414    if (!SVB.getContext().hasSameUnqualifiedType(RegionTy, SourceRegionTy))
1415      return None;
1416  }
1417
1418  if (!AllowSubregionBindings) {
1419    // If there are any other bindings within this region, we shouldn't reuse
1420    // the top-level binding.
1421    SmallVector<BindingPair, 16> Bindings;
1422    collectSubRegionBindings(Bindings, SVB, *B.lookup(R->getBaseRegion()), R,
1423                             /*IncludeAllDefaultBindings=*/true);
1424    if (Bindings.size() > 1)
1425      return None;
1426  }
1427
1428  return *LCV;
1429}
1430
1431
1432std::pair<Store, const SubRegion *>
1433RegionStoreManager::findLazyBinding(RegionBindingsConstRef B,
1434                                   const SubRegion *R,
1435                                   const SubRegion *originalRegion) {
1436  if (originalRegion != R) {
1437    if (Optional<nonloc::LazyCompoundVal> V =
1438          getExistingLazyBinding(svalBuilder, B, R, true))
1439      return std::make_pair(V->getStore(), V->getRegion());
1440  }
1441
1442  typedef std::pair<Store, const SubRegion *> StoreRegionPair;
1443  StoreRegionPair Result = StoreRegionPair();
1444
1445  if (const ElementRegion *ER = dyn_cast<ElementRegion>(R)) {
1446    Result = findLazyBinding(B, cast<SubRegion>(ER->getSuperRegion()),
1447                             originalRegion);
1448
1449    if (Result.second)
1450      Result.second = MRMgr.getElementRegionWithSuper(ER, Result.second);
1451
1452  } else if (const FieldRegion *FR = dyn_cast<FieldRegion>(R)) {
1453    Result = findLazyBinding(B, cast<SubRegion>(FR->getSuperRegion()),
1454                                       originalRegion);
1455
1456    if (Result.second)
1457      Result.second = MRMgr.getFieldRegionWithSuper(FR, Result.second);
1458
1459  } else if (const CXXBaseObjectRegion *BaseReg =
1460               dyn_cast<CXXBaseObjectRegion>(R)) {
1461    // C++ base object region is another kind of region that we should blast
1462    // through to look for lazy compound value. It is like a field region.
1463    Result = findLazyBinding(B, cast<SubRegion>(BaseReg->getSuperRegion()),
1464                             originalRegion);
1465
1466    if (Result.second)
1467      Result.second = MRMgr.getCXXBaseObjectRegionWithSuper(BaseReg,
1468                                                            Result.second);
1469  }
1470
1471  return Result;
1472}
1473
1474SVal RegionStoreManager::getBindingForElement(RegionBindingsConstRef B,
1475                                              const ElementRegion* R) {
1476  // We do not currently model bindings of the CompoundLiteralregion.
1477  if (isa<CompoundLiteralRegion>(R->getBaseRegion()))
1478    return UnknownVal();
1479
1480  // Check if the region has a binding.
1481  if (const Optional<SVal> &V = B.getDirectBinding(R))
1482    return *V;
1483
1484  const MemRegion* superR = R->getSuperRegion();
1485
1486  // Check if the region is an element region of a string literal.
1487  if (const StringRegion *StrR=dyn_cast<StringRegion>(superR)) {
1488    // FIXME: Handle loads from strings where the literal is treated as
1489    // an integer, e.g., *((unsigned int*)"hello")
1490    QualType T = Ctx.getAsArrayType(StrR->getValueType())->getElementType();
1491    if (T != Ctx.getCanonicalType(R->getElementType()))
1492      return UnknownVal();
1493
1494    const StringLiteral *Str = StrR->getStringLiteral();
1495    SVal Idx = R->getIndex();
1496    if (Optional<nonloc::ConcreteInt> CI = Idx.getAs<nonloc::ConcreteInt>()) {
1497      int64_t i = CI->getValue().getSExtValue();
1498      // Abort on string underrun.  This can be possible by arbitrary
1499      // clients of getBindingForElement().
1500      if (i < 0)
1501        return UndefinedVal();
1502      int64_t length = Str->getLength();
1503      // Technically, only i == length is guaranteed to be null.
1504      // However, such overflows should be caught before reaching this point;
1505      // the only time such an access would be made is if a string literal was
1506      // used to initialize a larger array.
1507      char c = (i >= length) ? '\0' : Str->getCodeUnit(i);
1508      return svalBuilder.makeIntVal(c, T);
1509    }
1510  }
1511
1512  // Check for loads from a code text region.  For such loads, just give up.
1513  if (isa<CodeTextRegion>(superR))
1514    return UnknownVal();
1515
1516  // Handle the case where we are indexing into a larger scalar object.
1517  // For example, this handles:
1518  //   int x = ...
1519  //   char *y = &x;
1520  //   return *y;
1521  // FIXME: This is a hack, and doesn't do anything really intelligent yet.
1522  const RegionRawOffset &O = R->getAsArrayOffset();
1523
1524  // If we cannot reason about the offset, return an unknown value.
1525  if (!O.getRegion())
1526    return UnknownVal();
1527
1528  if (const TypedValueRegion *baseR =
1529        dyn_cast_or_null<TypedValueRegion>(O.getRegion())) {
1530    QualType baseT = baseR->getValueType();
1531    if (baseT->isScalarType()) {
1532      QualType elemT = R->getElementType();
1533      if (elemT->isScalarType()) {
1534        if (Ctx.getTypeSizeInChars(baseT) >= Ctx.getTypeSizeInChars(elemT)) {
1535          if (const Optional<SVal> &V = B.getDirectBinding(superR)) {
1536            if (SymbolRef parentSym = V->getAsSymbol())
1537              return svalBuilder.getDerivedRegionValueSymbolVal(parentSym, R);
1538
1539            if (V->isUnknownOrUndef())
1540              return *V;
1541            // Other cases: give up.  We are indexing into a larger object
1542            // that has some value, but we don't know how to handle that yet.
1543            return UnknownVal();
1544          }
1545        }
1546      }
1547    }
1548  }
1549  return getBindingForFieldOrElementCommon(B, R, R->getElementType(),superR);
1550}
1551
1552SVal RegionStoreManager::getBindingForField(RegionBindingsConstRef B,
1553                                            const FieldRegion* R) {
1554
1555  // Check if the region has a binding.
1556  if (const Optional<SVal> &V = B.getDirectBinding(R))
1557    return *V;
1558
1559  QualType Ty = R->getValueType();
1560  return getBindingForFieldOrElementCommon(B, R, Ty, R->getSuperRegion());
1561}
1562
1563Optional<SVal>
1564RegionStoreManager::getBindingForDerivedDefaultValue(RegionBindingsConstRef B,
1565                                                     const MemRegion *superR,
1566                                                     const TypedValueRegion *R,
1567                                                     QualType Ty) {
1568
1569  if (const Optional<SVal> &D = B.getDefaultBinding(superR)) {
1570    const SVal &val = D.getValue();
1571    if (SymbolRef parentSym = val.getAsSymbol())
1572      return svalBuilder.getDerivedRegionValueSymbolVal(parentSym, R);
1573
1574    if (val.isZeroConstant())
1575      return svalBuilder.makeZeroVal(Ty);
1576
1577    if (val.isUnknownOrUndef())
1578      return val;
1579
1580    // Lazy bindings are usually handled through getExistingLazyBinding().
1581    // We should unify these two code paths at some point.
1582    if (val.getAs<nonloc::LazyCompoundVal>())
1583      return val;
1584
1585    llvm_unreachable("Unknown default value");
1586  }
1587
1588  return None;
1589}
1590
1591SVal RegionStoreManager::getLazyBinding(const SubRegion *LazyBindingRegion,
1592                                        RegionBindingsRef LazyBinding) {
1593  SVal Result;
1594  if (const ElementRegion *ER = dyn_cast<ElementRegion>(LazyBindingRegion))
1595    Result = getBindingForElement(LazyBinding, ER);
1596  else
1597    Result = getBindingForField(LazyBinding,
1598                                cast<FieldRegion>(LazyBindingRegion));
1599
1600  // FIXME: This is a hack to deal with RegionStore's inability to distinguish a
1601  // default value for /part/ of an aggregate from a default value for the
1602  // /entire/ aggregate. The most common case of this is when struct Outer
1603  // has as its first member a struct Inner, which is copied in from a stack
1604  // variable. In this case, even if the Outer's default value is symbolic, 0,
1605  // or unknown, it gets overridden by the Inner's default value of undefined.
1606  //
1607  // This is a general problem -- if the Inner is zero-initialized, the Outer
1608  // will now look zero-initialized. The proper way to solve this is with a
1609  // new version of RegionStore that tracks the extent of a binding as well
1610  // as the offset.
1611  //
1612  // This hack only takes care of the undefined case because that can very
1613  // quickly result in a warning.
1614  if (Result.isUndef())
1615    Result = UnknownVal();
1616
1617  return Result;
1618}
1619
1620SVal
1621RegionStoreManager::getBindingForFieldOrElementCommon(RegionBindingsConstRef B,
1622                                                      const TypedValueRegion *R,
1623                                                      QualType Ty,
1624                                                      const MemRegion *superR) {
1625
1626  // At this point we have already checked in either getBindingForElement or
1627  // getBindingForField if 'R' has a direct binding.
1628
1629  // Lazy binding?
1630  Store lazyBindingStore = NULL;
1631  const SubRegion *lazyBindingRegion = NULL;
1632  llvm::tie(lazyBindingStore, lazyBindingRegion) = findLazyBinding(B, R, R);
1633  if (lazyBindingRegion)
1634    return getLazyBinding(lazyBindingRegion,
1635                          getRegionBindings(lazyBindingStore));
1636
1637  // Record whether or not we see a symbolic index.  That can completely
1638  // be out of scope of our lookup.
1639  bool hasSymbolicIndex = false;
1640
1641  // FIXME: This is a hack to deal with RegionStore's inability to distinguish a
1642  // default value for /part/ of an aggregate from a default value for the
1643  // /entire/ aggregate. The most common case of this is when struct Outer
1644  // has as its first member a struct Inner, which is copied in from a stack
1645  // variable. In this case, even if the Outer's default value is symbolic, 0,
1646  // or unknown, it gets overridden by the Inner's default value of undefined.
1647  //
1648  // This is a general problem -- if the Inner is zero-initialized, the Outer
1649  // will now look zero-initialized. The proper way to solve this is with a
1650  // new version of RegionStore that tracks the extent of a binding as well
1651  // as the offset.
1652  //
1653  // This hack only takes care of the undefined case because that can very
1654  // quickly result in a warning.
1655  bool hasPartialLazyBinding = false;
1656
1657  const SubRegion *Base = dyn_cast<SubRegion>(superR);
1658  while (Base) {
1659    if (Optional<SVal> D = getBindingForDerivedDefaultValue(B, Base, R, Ty)) {
1660      if (D->getAs<nonloc::LazyCompoundVal>()) {
1661        hasPartialLazyBinding = true;
1662        break;
1663      }
1664
1665      return *D;
1666    }
1667
1668    if (const ElementRegion *ER = dyn_cast<ElementRegion>(Base)) {
1669      NonLoc index = ER->getIndex();
1670      if (!index.isConstant())
1671        hasSymbolicIndex = true;
1672    }
1673
1674    // If our super region is a field or element itself, walk up the region
1675    // hierarchy to see if there is a default value installed in an ancestor.
1676    Base = dyn_cast<SubRegion>(Base->getSuperRegion());
1677  }
1678
1679  if (R->hasStackNonParametersStorage()) {
1680    if (isa<ElementRegion>(R)) {
1681      // Currently we don't reason specially about Clang-style vectors.  Check
1682      // if superR is a vector and if so return Unknown.
1683      if (const TypedValueRegion *typedSuperR =
1684            dyn_cast<TypedValueRegion>(superR)) {
1685        if (typedSuperR->getValueType()->isVectorType())
1686          return UnknownVal();
1687      }
1688    }
1689
1690    // FIXME: We also need to take ElementRegions with symbolic indexes into
1691    // account.  This case handles both directly accessing an ElementRegion
1692    // with a symbolic offset, but also fields within an element with
1693    // a symbolic offset.
1694    if (hasSymbolicIndex)
1695      return UnknownVal();
1696
1697    if (!hasPartialLazyBinding)
1698      return UndefinedVal();
1699  }
1700
1701  // All other values are symbolic.
1702  return svalBuilder.getRegionValueSymbolVal(R);
1703}
1704
1705SVal RegionStoreManager::getBindingForObjCIvar(RegionBindingsConstRef B,
1706                                               const ObjCIvarRegion* R) {
1707  // Check if the region has a binding.
1708  if (const Optional<SVal> &V = B.getDirectBinding(R))
1709    return *V;
1710
1711  const MemRegion *superR = R->getSuperRegion();
1712
1713  // Check if the super region has a default binding.
1714  if (const Optional<SVal> &V = B.getDefaultBinding(superR)) {
1715    if (SymbolRef parentSym = V->getAsSymbol())
1716      return svalBuilder.getDerivedRegionValueSymbolVal(parentSym, R);
1717
1718    // Other cases: give up.
1719    return UnknownVal();
1720  }
1721
1722  return getBindingForLazySymbol(R);
1723}
1724
1725static Optional<SVal> getConstValue(SValBuilder &SVB, const VarDecl *VD) {
1726  ASTContext &Ctx = SVB.getContext();
1727  if (!VD->getType().isConstQualified())
1728    return None;
1729
1730  const Expr *Init = VD->getInit();
1731  if (!Init)
1732    return None;
1733
1734  llvm::APSInt Result;
1735  if (!Init->isGLValue() && Init->EvaluateAsInt(Result, Ctx))
1736    return SVB.makeIntVal(Result);
1737
1738  if (Init->isNullPointerConstant(Ctx, Expr::NPC_ValueDependentIsNotNull))
1739    return SVB.makeNull();
1740
1741  // FIXME: Handle other possible constant expressions.
1742  return None;
1743}
1744
1745SVal RegionStoreManager::getBindingForVar(RegionBindingsConstRef B,
1746                                          const VarRegion *R) {
1747
1748  // Check if the region has a binding.
1749  if (const Optional<SVal> &V = B.getDirectBinding(R))
1750    return *V;
1751
1752  // Lazily derive a value for the VarRegion.
1753  const VarDecl *VD = R->getDecl();
1754  const MemSpaceRegion *MS = R->getMemorySpace();
1755
1756  // Arguments are always symbolic.
1757  if (isa<StackArgumentsSpaceRegion>(MS))
1758    return svalBuilder.getRegionValueSymbolVal(R);
1759
1760  // Is 'VD' declared constant?  If so, retrieve the constant value.
1761  if (Optional<SVal> V = getConstValue(svalBuilder, VD))
1762    return *V;
1763
1764  // This must come after the check for constants because closure-captured
1765  // constant variables may appear in UnknownSpaceRegion.
1766  if (isa<UnknownSpaceRegion>(MS))
1767    return svalBuilder.getRegionValueSymbolVal(R);
1768
1769  if (isa<GlobalsSpaceRegion>(MS)) {
1770    QualType T = VD->getType();
1771
1772    // Function-scoped static variables are default-initialized to 0; if they
1773    // have an initializer, it would have been processed by now.
1774    if (isa<StaticGlobalSpaceRegion>(MS))
1775      return svalBuilder.makeZeroVal(T);
1776
1777    if (Optional<SVal> V = getBindingForDerivedDefaultValue(B, MS, R, T)) {
1778      assert(!V->getAs<nonloc::LazyCompoundVal>());
1779      return V.getValue();
1780    }
1781
1782    return svalBuilder.getRegionValueSymbolVal(R);
1783  }
1784
1785  return UndefinedVal();
1786}
1787
1788SVal RegionStoreManager::getBindingForLazySymbol(const TypedValueRegion *R) {
1789  // All other values are symbolic.
1790  return svalBuilder.getRegionValueSymbolVal(R);
1791}
1792
1793const RegionStoreManager::SValListTy &
1794RegionStoreManager::getInterestingValues(nonloc::LazyCompoundVal LCV) {
1795  // First, check the cache.
1796  LazyBindingsMapTy::iterator I = LazyBindingsMap.find(LCV.getCVData());
1797  if (I != LazyBindingsMap.end())
1798    return I->second;
1799
1800  // If we don't have a list of values cached, start constructing it.
1801  SValListTy List;
1802
1803  const SubRegion *LazyR = LCV.getRegion();
1804  RegionBindingsRef B = getRegionBindings(LCV.getStore());
1805
1806  // If this region had /no/ bindings at the time, there are no interesting
1807  // values to return.
1808  const ClusterBindings *Cluster = B.lookup(LazyR->getBaseRegion());
1809  if (!Cluster)
1810    return (LazyBindingsMap[LCV.getCVData()] = llvm_move(List));
1811
1812  SmallVector<BindingPair, 32> Bindings;
1813  collectSubRegionBindings(Bindings, svalBuilder, *Cluster, LazyR,
1814                           /*IncludeAllDefaultBindings=*/true);
1815  for (SmallVectorImpl<BindingPair>::const_iterator I = Bindings.begin(),
1816                                                    E = Bindings.end();
1817       I != E; ++I) {
1818    SVal V = I->second;
1819    if (V.isUnknownOrUndef() || V.isConstant())
1820      continue;
1821
1822    if (Optional<nonloc::LazyCompoundVal> InnerLCV =
1823            V.getAs<nonloc::LazyCompoundVal>()) {
1824      const SValListTy &InnerList = getInterestingValues(*InnerLCV);
1825      List.insert(List.end(), InnerList.begin(), InnerList.end());
1826      continue;
1827    }
1828
1829    List.push_back(V);
1830  }
1831
1832  return (LazyBindingsMap[LCV.getCVData()] = llvm_move(List));
1833}
1834
1835NonLoc RegionStoreManager::createLazyBinding(RegionBindingsConstRef B,
1836                                             const TypedValueRegion *R) {
1837  if (Optional<nonloc::LazyCompoundVal> V =
1838        getExistingLazyBinding(svalBuilder, B, R, false))
1839    return *V;
1840
1841  return svalBuilder.makeLazyCompoundVal(StoreRef(B.asStore(), *this), R);
1842}
1843
1844SVal RegionStoreManager::getBindingForStruct(RegionBindingsConstRef B,
1845                                             const TypedValueRegion *R) {
1846  const RecordDecl *RD = R->getValueType()->castAs<RecordType>()->getDecl();
1847  if (RD->field_empty())
1848    return UnknownVal();
1849
1850  return createLazyBinding(B, R);
1851}
1852
1853SVal RegionStoreManager::getBindingForArray(RegionBindingsConstRef B,
1854                                            const TypedValueRegion *R) {
1855  assert(Ctx.getAsConstantArrayType(R->getValueType()) &&
1856         "Only constant array types can have compound bindings.");
1857
1858  return createLazyBinding(B, R);
1859}
1860
1861bool RegionStoreManager::includedInBindings(Store store,
1862                                            const MemRegion *region) const {
1863  RegionBindingsRef B = getRegionBindings(store);
1864  region = region->getBaseRegion();
1865
1866  // Quick path: if the base is the head of a cluster, the region is live.
1867  if (B.lookup(region))
1868    return true;
1869
1870  // Slow path: if the region is the VALUE of any binding, it is live.
1871  for (RegionBindingsRef::iterator RI = B.begin(), RE = B.end(); RI != RE; ++RI) {
1872    const ClusterBindings &Cluster = RI.getData();
1873    for (ClusterBindings::iterator CI = Cluster.begin(), CE = Cluster.end();
1874         CI != CE; ++CI) {
1875      const SVal &D = CI.getData();
1876      if (const MemRegion *R = D.getAsRegion())
1877        if (R->getBaseRegion() == region)
1878          return true;
1879    }
1880  }
1881
1882  return false;
1883}
1884
1885//===----------------------------------------------------------------------===//
1886// Binding values to regions.
1887//===----------------------------------------------------------------------===//
1888
1889StoreRef RegionStoreManager::killBinding(Store ST, Loc L) {
1890  if (Optional<loc::MemRegionVal> LV = L.getAs<loc::MemRegionVal>())
1891    if (const MemRegion* R = LV->getRegion())
1892      return StoreRef(getRegionBindings(ST).removeBinding(R)
1893                                           .asImmutableMap()
1894                                           .getRootWithoutRetain(),
1895                      *this);
1896
1897  return StoreRef(ST, *this);
1898}
1899
1900RegionBindingsRef
1901RegionStoreManager::bind(RegionBindingsConstRef B, Loc L, SVal V) {
1902  if (L.getAs<loc::ConcreteInt>())
1903    return B;
1904
1905  // If we get here, the location should be a region.
1906  const MemRegion *R = L.castAs<loc::MemRegionVal>().getRegion();
1907
1908  // Check if the region is a struct region.
1909  if (const TypedValueRegion* TR = dyn_cast<TypedValueRegion>(R)) {
1910    QualType Ty = TR->getValueType();
1911    if (Ty->isArrayType())
1912      return bindArray(B, TR, V);
1913    if (Ty->isStructureOrClassType())
1914      return bindStruct(B, TR, V);
1915    if (Ty->isVectorType())
1916      return bindVector(B, TR, V);
1917  }
1918
1919  if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(R)) {
1920    // Binding directly to a symbolic region should be treated as binding
1921    // to element 0.
1922    QualType T = SR->getSymbol()->getType();
1923    if (T->isAnyPointerType() || T->isReferenceType())
1924      T = T->getPointeeType();
1925
1926    R = GetElementZeroRegion(SR, T);
1927  }
1928
1929  // Clear out bindings that may overlap with this binding.
1930  RegionBindingsRef NewB = removeSubRegionBindings(B, cast<SubRegion>(R));
1931  return NewB.addBinding(BindingKey::Make(R, BindingKey::Direct), V);
1932}
1933
1934// FIXME: this method should be merged into Bind().
1935StoreRef RegionStoreManager::bindCompoundLiteral(Store ST,
1936                                                 const CompoundLiteralExpr *CL,
1937                                                 const LocationContext *LC,
1938                                                 SVal V) {
1939  return Bind(ST, loc::MemRegionVal(MRMgr.getCompoundLiteralRegion(CL, LC)), V);
1940}
1941
1942RegionBindingsRef
1943RegionStoreManager::setImplicitDefaultValue(RegionBindingsConstRef B,
1944                                            const MemRegion *R,
1945                                            QualType T) {
1946  SVal V;
1947
1948  if (Loc::isLocType(T))
1949    V = svalBuilder.makeNull();
1950  else if (T->isIntegralOrEnumerationType())
1951    V = svalBuilder.makeZeroVal(T);
1952  else if (T->isStructureOrClassType() || T->isArrayType()) {
1953    // Set the default value to a zero constant when it is a structure
1954    // or array.  The type doesn't really matter.
1955    V = svalBuilder.makeZeroVal(Ctx.IntTy);
1956  }
1957  else {
1958    // We can't represent values of this type, but we still need to set a value
1959    // to record that the region has been initialized.
1960    // If this assertion ever fires, a new case should be added above -- we
1961    // should know how to default-initialize any value we can symbolicate.
1962    assert(!SymbolManager::canSymbolicate(T) && "This type is representable");
1963    V = UnknownVal();
1964  }
1965
1966  return B.addBinding(R, BindingKey::Default, V);
1967}
1968
1969RegionBindingsRef
1970RegionStoreManager::bindArray(RegionBindingsConstRef B,
1971                              const TypedValueRegion* R,
1972                              SVal Init) {
1973
1974  const ArrayType *AT =cast<ArrayType>(Ctx.getCanonicalType(R->getValueType()));
1975  QualType ElementTy = AT->getElementType();
1976  Optional<uint64_t> Size;
1977
1978  if (const ConstantArrayType* CAT = dyn_cast<ConstantArrayType>(AT))
1979    Size = CAT->getSize().getZExtValue();
1980
1981  // Check if the init expr is a string literal.
1982  if (Optional<loc::MemRegionVal> MRV = Init.getAs<loc::MemRegionVal>()) {
1983    const StringRegion *S = cast<StringRegion>(MRV->getRegion());
1984
1985    // Treat the string as a lazy compound value.
1986    StoreRef store(B.asStore(), *this);
1987    nonloc::LazyCompoundVal LCV = svalBuilder.makeLazyCompoundVal(store, S)
1988        .castAs<nonloc::LazyCompoundVal>();
1989    return bindAggregate(B, R, LCV);
1990  }
1991
1992  // Handle lazy compound values.
1993  if (Init.getAs<nonloc::LazyCompoundVal>())
1994    return bindAggregate(B, R, Init);
1995
1996  // Remaining case: explicit compound values.
1997
1998  if (Init.isUnknown())
1999    return setImplicitDefaultValue(B, R, ElementTy);
2000
2001  const nonloc::CompoundVal& CV = Init.castAs<nonloc::CompoundVal>();
2002  nonloc::CompoundVal::iterator VI = CV.begin(), VE = CV.end();
2003  uint64_t i = 0;
2004
2005  RegionBindingsRef NewB(B);
2006
2007  for (; Size.hasValue() ? i < Size.getValue() : true ; ++i, ++VI) {
2008    // The init list might be shorter than the array length.
2009    if (VI == VE)
2010      break;
2011
2012    const NonLoc &Idx = svalBuilder.makeArrayIndex(i);
2013    const ElementRegion *ER = MRMgr.getElementRegion(ElementTy, Idx, R, Ctx);
2014
2015    if (ElementTy->isStructureOrClassType())
2016      NewB = bindStruct(NewB, ER, *VI);
2017    else if (ElementTy->isArrayType())
2018      NewB = bindArray(NewB, ER, *VI);
2019    else
2020      NewB = bind(NewB, svalBuilder.makeLoc(ER), *VI);
2021  }
2022
2023  // If the init list is shorter than the array length, set the
2024  // array default value.
2025  if (Size.hasValue() && i < Size.getValue())
2026    NewB = setImplicitDefaultValue(NewB, R, ElementTy);
2027
2028  return NewB;
2029}
2030
2031RegionBindingsRef RegionStoreManager::bindVector(RegionBindingsConstRef B,
2032                                                 const TypedValueRegion* R,
2033                                                 SVal V) {
2034  QualType T = R->getValueType();
2035  assert(T->isVectorType());
2036  const VectorType *VT = T->getAs<VectorType>(); // Use getAs for typedefs.
2037
2038  // Handle lazy compound values and symbolic values.
2039  if (V.getAs<nonloc::LazyCompoundVal>() || V.getAs<nonloc::SymbolVal>())
2040    return bindAggregate(B, R, V);
2041
2042  // We may get non-CompoundVal accidentally due to imprecise cast logic or
2043  // that we are binding symbolic struct value. Kill the field values, and if
2044  // the value is symbolic go and bind it as a "default" binding.
2045  if (!V.getAs<nonloc::CompoundVal>()) {
2046    return bindAggregate(B, R, UnknownVal());
2047  }
2048
2049  QualType ElemType = VT->getElementType();
2050  nonloc::CompoundVal CV = V.castAs<nonloc::CompoundVal>();
2051  nonloc::CompoundVal::iterator VI = CV.begin(), VE = CV.end();
2052  unsigned index = 0, numElements = VT->getNumElements();
2053  RegionBindingsRef NewB(B);
2054
2055  for ( ; index != numElements ; ++index) {
2056    if (VI == VE)
2057      break;
2058
2059    NonLoc Idx = svalBuilder.makeArrayIndex(index);
2060    const ElementRegion *ER = MRMgr.getElementRegion(ElemType, Idx, R, Ctx);
2061
2062    if (ElemType->isArrayType())
2063      NewB = bindArray(NewB, ER, *VI);
2064    else if (ElemType->isStructureOrClassType())
2065      NewB = bindStruct(NewB, ER, *VI);
2066    else
2067      NewB = bind(NewB, svalBuilder.makeLoc(ER), *VI);
2068  }
2069  return NewB;
2070}
2071
2072RegionBindingsRef RegionStoreManager::bindStruct(RegionBindingsConstRef B,
2073                                                 const TypedValueRegion* R,
2074                                                 SVal V) {
2075  if (!Features.supportsFields())
2076    return B;
2077
2078  QualType T = R->getValueType();
2079  assert(T->isStructureOrClassType());
2080
2081  const RecordType* RT = T->getAs<RecordType>();
2082  RecordDecl *RD = RT->getDecl();
2083
2084  if (!RD->isCompleteDefinition())
2085    return B;
2086
2087  // Handle lazy compound values and symbolic values.
2088  if (V.getAs<nonloc::LazyCompoundVal>() || V.getAs<nonloc::SymbolVal>())
2089    return bindAggregate(B, R, V);
2090
2091  // We may get non-CompoundVal accidentally due to imprecise cast logic or
2092  // that we are binding symbolic struct value. Kill the field values, and if
2093  // the value is symbolic go and bind it as a "default" binding.
2094  if (V.isUnknown() || !V.getAs<nonloc::CompoundVal>())
2095    return bindAggregate(B, R, UnknownVal());
2096
2097  const nonloc::CompoundVal& CV = V.castAs<nonloc::CompoundVal>();
2098  nonloc::CompoundVal::iterator VI = CV.begin(), VE = CV.end();
2099
2100  RecordDecl::field_iterator FI, FE;
2101  RegionBindingsRef NewB(B);
2102
2103  for (FI = RD->field_begin(), FE = RD->field_end(); FI != FE; ++FI) {
2104
2105    if (VI == VE)
2106      break;
2107
2108    // Skip any unnamed bitfields to stay in sync with the initializers.
2109    if (FI->isUnnamedBitfield())
2110      continue;
2111
2112    QualType FTy = FI->getType();
2113    const FieldRegion* FR = MRMgr.getFieldRegion(*FI, R);
2114
2115    if (FTy->isArrayType())
2116      NewB = bindArray(NewB, FR, *VI);
2117    else if (FTy->isStructureOrClassType())
2118      NewB = bindStruct(NewB, FR, *VI);
2119    else
2120      NewB = bind(NewB, svalBuilder.makeLoc(FR), *VI);
2121    ++VI;
2122  }
2123
2124  // There may be fewer values in the initialize list than the fields of struct.
2125  if (FI != FE) {
2126    NewB = NewB.addBinding(R, BindingKey::Default,
2127                           svalBuilder.makeIntVal(0, false));
2128  }
2129
2130  return NewB;
2131}
2132
2133RegionBindingsRef
2134RegionStoreManager::bindAggregate(RegionBindingsConstRef B,
2135                                  const TypedRegion *R,
2136                                  SVal Val) {
2137  // Remove the old bindings, using 'R' as the root of all regions
2138  // we will invalidate. Then add the new binding.
2139  return removeSubRegionBindings(B, R).addBinding(R, BindingKey::Default, Val);
2140}
2141
2142//===----------------------------------------------------------------------===//
2143// State pruning.
2144//===----------------------------------------------------------------------===//
2145
2146namespace {
2147class removeDeadBindingsWorker :
2148  public ClusterAnalysis<removeDeadBindingsWorker> {
2149  SmallVector<const SymbolicRegion*, 12> Postponed;
2150  SymbolReaper &SymReaper;
2151  const StackFrameContext *CurrentLCtx;
2152
2153public:
2154  removeDeadBindingsWorker(RegionStoreManager &rm,
2155                           ProgramStateManager &stateMgr,
2156                           RegionBindingsRef b, SymbolReaper &symReaper,
2157                           const StackFrameContext *LCtx)
2158    : ClusterAnalysis<removeDeadBindingsWorker>(rm, stateMgr, b, GFK_None),
2159      SymReaper(symReaper), CurrentLCtx(LCtx) {}
2160
2161  // Called by ClusterAnalysis.
2162  void VisitAddedToCluster(const MemRegion *baseR, const ClusterBindings &C);
2163  void VisitCluster(const MemRegion *baseR, const ClusterBindings *C);
2164  using ClusterAnalysis<removeDeadBindingsWorker>::VisitCluster;
2165
2166  bool UpdatePostponed();
2167  void VisitBinding(SVal V);
2168};
2169}
2170
2171void removeDeadBindingsWorker::VisitAddedToCluster(const MemRegion *baseR,
2172                                                   const ClusterBindings &C) {
2173
2174  if (const VarRegion *VR = dyn_cast<VarRegion>(baseR)) {
2175    if (SymReaper.isLive(VR))
2176      AddToWorkList(baseR, &C);
2177
2178    return;
2179  }
2180
2181  if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(baseR)) {
2182    if (SymReaper.isLive(SR->getSymbol()))
2183      AddToWorkList(SR, &C);
2184    else
2185      Postponed.push_back(SR);
2186
2187    return;
2188  }
2189
2190  if (isa<NonStaticGlobalSpaceRegion>(baseR)) {
2191    AddToWorkList(baseR, &C);
2192    return;
2193  }
2194
2195  // CXXThisRegion in the current or parent location context is live.
2196  if (const CXXThisRegion *TR = dyn_cast<CXXThisRegion>(baseR)) {
2197    const StackArgumentsSpaceRegion *StackReg =
2198      cast<StackArgumentsSpaceRegion>(TR->getSuperRegion());
2199    const StackFrameContext *RegCtx = StackReg->getStackFrame();
2200    if (CurrentLCtx &&
2201        (RegCtx == CurrentLCtx || RegCtx->isParentOf(CurrentLCtx)))
2202      AddToWorkList(TR, &C);
2203  }
2204}
2205
2206void removeDeadBindingsWorker::VisitCluster(const MemRegion *baseR,
2207                                            const ClusterBindings *C) {
2208  if (!C)
2209    return;
2210
2211  // Mark the symbol for any SymbolicRegion with live bindings as live itself.
2212  // This means we should continue to track that symbol.
2213  if (const SymbolicRegion *SymR = dyn_cast<SymbolicRegion>(baseR))
2214    SymReaper.markLive(SymR->getSymbol());
2215
2216  for (ClusterBindings::iterator I = C->begin(), E = C->end(); I != E; ++I)
2217    VisitBinding(I.getData());
2218}
2219
2220void removeDeadBindingsWorker::VisitBinding(SVal V) {
2221  // Is it a LazyCompoundVal?  All referenced regions are live as well.
2222  if (Optional<nonloc::LazyCompoundVal> LCS =
2223          V.getAs<nonloc::LazyCompoundVal>()) {
2224
2225    const RegionStoreManager::SValListTy &Vals = RM.getInterestingValues(*LCS);
2226
2227    for (RegionStoreManager::SValListTy::const_iterator I = Vals.begin(),
2228                                                        E = Vals.end();
2229         I != E; ++I)
2230      VisitBinding(*I);
2231
2232    return;
2233  }
2234
2235  // If V is a region, then add it to the worklist.
2236  if (const MemRegion *R = V.getAsRegion()) {
2237    AddToWorkList(R);
2238
2239    // All regions captured by a block are also live.
2240    if (const BlockDataRegion *BR = dyn_cast<BlockDataRegion>(R)) {
2241      BlockDataRegion::referenced_vars_iterator I = BR->referenced_vars_begin(),
2242                                                E = BR->referenced_vars_end();
2243      for ( ; I != E; ++I)
2244        AddToWorkList(I.getCapturedRegion());
2245    }
2246  }
2247
2248
2249  // Update the set of live symbols.
2250  for (SymExpr::symbol_iterator SI = V.symbol_begin(), SE = V.symbol_end();
2251       SI!=SE; ++SI)
2252    SymReaper.markLive(*SI);
2253}
2254
2255bool removeDeadBindingsWorker::UpdatePostponed() {
2256  // See if any postponed SymbolicRegions are actually live now, after
2257  // having done a scan.
2258  bool changed = false;
2259
2260  for (SmallVectorImpl<const SymbolicRegion*>::iterator
2261        I = Postponed.begin(), E = Postponed.end() ; I != E ; ++I) {
2262    if (const SymbolicRegion *SR = *I) {
2263      if (SymReaper.isLive(SR->getSymbol())) {
2264        changed |= AddToWorkList(SR);
2265        *I = NULL;
2266      }
2267    }
2268  }
2269
2270  return changed;
2271}
2272
2273StoreRef RegionStoreManager::removeDeadBindings(Store store,
2274                                                const StackFrameContext *LCtx,
2275                                                SymbolReaper& SymReaper) {
2276  RegionBindingsRef B = getRegionBindings(store);
2277  removeDeadBindingsWorker W(*this, StateMgr, B, SymReaper, LCtx);
2278  W.GenerateClusters();
2279
2280  // Enqueue the region roots onto the worklist.
2281  for (SymbolReaper::region_iterator I = SymReaper.region_begin(),
2282       E = SymReaper.region_end(); I != E; ++I) {
2283    W.AddToWorkList(*I);
2284  }
2285
2286  do W.RunWorkList(); while (W.UpdatePostponed());
2287
2288  // We have now scanned the store, marking reachable regions and symbols
2289  // as live.  We now remove all the regions that are dead from the store
2290  // as well as update DSymbols with the set symbols that are now dead.
2291  for (RegionBindingsRef::iterator I = B.begin(), E = B.end(); I != E; ++I) {
2292    const MemRegion *Base = I.getKey();
2293
2294    // If the cluster has been visited, we know the region has been marked.
2295    if (W.isVisited(Base))
2296      continue;
2297
2298    // Remove the dead entry.
2299    B = B.remove(Base);
2300
2301    if (const SymbolicRegion *SymR = dyn_cast<SymbolicRegion>(Base))
2302      SymReaper.maybeDead(SymR->getSymbol());
2303
2304    // Mark all non-live symbols that this binding references as dead.
2305    const ClusterBindings &Cluster = I.getData();
2306    for (ClusterBindings::iterator CI = Cluster.begin(), CE = Cluster.end();
2307         CI != CE; ++CI) {
2308      SVal X = CI.getData();
2309      SymExpr::symbol_iterator SI = X.symbol_begin(), SE = X.symbol_end();
2310      for (; SI != SE; ++SI)
2311        SymReaper.maybeDead(*SI);
2312    }
2313  }
2314
2315  return StoreRef(B.asStore(), *this);
2316}
2317
2318//===----------------------------------------------------------------------===//
2319// Utility methods.
2320//===----------------------------------------------------------------------===//
2321
2322void RegionStoreManager::print(Store store, raw_ostream &OS,
2323                               const char* nl, const char *sep) {
2324  RegionBindingsRef B = getRegionBindings(store);
2325  OS << "Store (direct and default bindings), "
2326     << B.asStore()
2327     << " :" << nl;
2328  B.dump(OS, nl);
2329}
2330