1 //== RegionStore.cpp - Field-sensitive store model --------------*- C++ -*--==// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file defines a basic region store model. In this model, we do have field 10 // sensitivity. But we assume nothing about the heap shape. So recursive data 11 // structures are largely ignored. Basically we do 1-limiting analysis. 12 // Parameter pointers are assumed with no aliasing. Pointee objects of 13 // parameters are created lazily. 14 // 15 //===----------------------------------------------------------------------===// 16 17 #include "clang/AST/Attr.h" 18 #include "clang/AST/CharUnits.h" 19 #include "clang/ASTMatchers/ASTMatchFinder.h" 20 #include "clang/Analysis/Analyses/LiveVariables.h" 21 #include "clang/Analysis/AnalysisDeclContext.h" 22 #include "clang/Basic/JsonSupport.h" 23 #include "clang/Basic/TargetInfo.h" 24 #include "clang/StaticAnalyzer/Core/PathSensitive/AnalysisManager.h" 25 #include "clang/StaticAnalyzer/Core/PathSensitive/CallEvent.h" 26 #include "clang/StaticAnalyzer/Core/PathSensitive/MemRegion.h" 27 #include "clang/StaticAnalyzer/Core/PathSensitive/ProgramState.h" 28 #include "clang/StaticAnalyzer/Core/PathSensitive/ProgramStateTrait.h" 29 #include "clang/StaticAnalyzer/Core/PathSensitive/SubEngine.h" 30 #include "llvm/ADT/ImmutableMap.h" 31 #include "llvm/ADT/Optional.h" 32 #include "llvm/Support/raw_ostream.h" 33 #include <utility> 34 35 using namespace clang; 36 using namespace ento; 37 38 //===----------------------------------------------------------------------===// 39 // Representation of binding keys. 40 //===----------------------------------------------------------------------===// 41 42 namespace { 43 class BindingKey { 44 public: 45 enum Kind { Default = 0x0, Direct = 0x1 }; 46 private: 47 enum { Symbolic = 0x2 }; 48 49 llvm::PointerIntPair<const MemRegion *, 2> P; 50 uint64_t Data; 51 52 /// Create a key for a binding to region \p r, which has a symbolic offset 53 /// from region \p Base. 54 explicit BindingKey(const SubRegion *r, const SubRegion *Base, Kind k) 55 : P(r, k | Symbolic), Data(reinterpret_cast<uintptr_t>(Base)) { 56 assert(r && Base && "Must have known regions."); 57 assert(getConcreteOffsetRegion() == Base && "Failed to store base region"); 58 } 59 60 /// Create a key for a binding at \p offset from base region \p r. 61 explicit BindingKey(const MemRegion *r, uint64_t offset, Kind k) 62 : P(r, k), Data(offset) { 63 assert(r && "Must have known regions."); 64 assert(getOffset() == offset && "Failed to store offset"); 65 assert((r == r->getBaseRegion() || isa<ObjCIvarRegion>(r) || 66 isa <CXXDerivedObjectRegion>(r)) && 67 "Not a base"); 68 } 69 public: 70 71 bool isDirect() const { return P.getInt() & Direct; } 72 bool hasSymbolicOffset() const { return P.getInt() & Symbolic; } 73 74 const MemRegion *getRegion() const { return P.getPointer(); } 75 uint64_t getOffset() const { 76 assert(!hasSymbolicOffset()); 77 return Data; 78 } 79 80 const SubRegion *getConcreteOffsetRegion() const { 81 assert(hasSymbolicOffset()); 82 return reinterpret_cast<const SubRegion *>(static_cast<uintptr_t>(Data)); 83 } 84 85 const MemRegion *getBaseRegion() const { 86 if (hasSymbolicOffset()) 87 return getConcreteOffsetRegion()->getBaseRegion(); 88 return getRegion()->getBaseRegion(); 89 } 90 91 void Profile(llvm::FoldingSetNodeID& ID) const { 92 ID.AddPointer(P.getOpaqueValue()); 93 ID.AddInteger(Data); 94 } 95 96 static BindingKey Make(const MemRegion *R, Kind k); 97 98 bool operator<(const BindingKey &X) const { 99 if (P.getOpaqueValue() < X.P.getOpaqueValue()) 100 return true; 101 if (P.getOpaqueValue() > X.P.getOpaqueValue()) 102 return false; 103 return Data < X.Data; 104 } 105 106 bool operator==(const BindingKey &X) const { 107 return P.getOpaqueValue() == X.P.getOpaqueValue() && 108 Data == X.Data; 109 } 110 111 void dump() const; 112 }; 113 } // end anonymous namespace 114 115 BindingKey BindingKey::Make(const MemRegion *R, Kind k) { 116 const RegionOffset &RO = R->getAsOffset(); 117 if (RO.hasSymbolicOffset()) 118 return BindingKey(cast<SubRegion>(R), cast<SubRegion>(RO.getRegion()), k); 119 120 return BindingKey(RO.getRegion(), RO.getOffset(), k); 121 } 122 123 namespace llvm { 124 static inline raw_ostream &operator<<(raw_ostream &Out, BindingKey K) { 125 Out << "\"kind\": \"" << (K.isDirect() ? "Direct" : "Default") 126 << "\", \"offset\": "; 127 128 if (!K.hasSymbolicOffset()) 129 Out << K.getOffset(); 130 else 131 Out << "null"; 132 133 return Out; 134 } 135 136 } // namespace llvm 137 138 LLVM_DUMP_METHOD void BindingKey::dump() const { llvm::errs() << *this; } 139 140 //===----------------------------------------------------------------------===// 141 // Actual Store type. 142 //===----------------------------------------------------------------------===// 143 144 typedef llvm::ImmutableMap<BindingKey, SVal> ClusterBindings; 145 typedef llvm::ImmutableMapRef<BindingKey, SVal> ClusterBindingsRef; 146 typedef std::pair<BindingKey, SVal> BindingPair; 147 148 typedef llvm::ImmutableMap<const MemRegion *, ClusterBindings> 149 RegionBindings; 150 151 namespace { 152 class RegionBindingsRef : public llvm::ImmutableMapRef<const MemRegion *, 153 ClusterBindings> { 154 ClusterBindings::Factory *CBFactory; 155 156 public: 157 typedef llvm::ImmutableMapRef<const MemRegion *, ClusterBindings> 158 ParentTy; 159 160 RegionBindingsRef(ClusterBindings::Factory &CBFactory, 161 const RegionBindings::TreeTy *T, 162 RegionBindings::TreeTy::Factory *F) 163 : llvm::ImmutableMapRef<const MemRegion *, ClusterBindings>(T, F), 164 CBFactory(&CBFactory) {} 165 166 RegionBindingsRef(const ParentTy &P, ClusterBindings::Factory &CBFactory) 167 : llvm::ImmutableMapRef<const MemRegion *, ClusterBindings>(P), 168 CBFactory(&CBFactory) {} 169 170 RegionBindingsRef add(key_type_ref K, data_type_ref D) const { 171 return RegionBindingsRef(static_cast<const ParentTy *>(this)->add(K, D), 172 *CBFactory); 173 } 174 175 RegionBindingsRef remove(key_type_ref K) const { 176 return RegionBindingsRef(static_cast<const ParentTy *>(this)->remove(K), 177 *CBFactory); 178 } 179 180 RegionBindingsRef addBinding(BindingKey K, SVal V) const; 181 182 RegionBindingsRef addBinding(const MemRegion *R, 183 BindingKey::Kind k, SVal V) const; 184 185 const SVal *lookup(BindingKey K) const; 186 const SVal *lookup(const MemRegion *R, BindingKey::Kind k) const; 187 using llvm::ImmutableMapRef<const MemRegion *, ClusterBindings>::lookup; 188 189 RegionBindingsRef removeBinding(BindingKey K); 190 191 RegionBindingsRef removeBinding(const MemRegion *R, 192 BindingKey::Kind k); 193 194 RegionBindingsRef removeBinding(const MemRegion *R) { 195 return removeBinding(R, BindingKey::Direct). 196 removeBinding(R, BindingKey::Default); 197 } 198 199 Optional<SVal> getDirectBinding(const MemRegion *R) const; 200 201 /// getDefaultBinding - Returns an SVal* representing an optional default 202 /// binding associated with a region and its subregions. 203 Optional<SVal> getDefaultBinding(const MemRegion *R) const; 204 205 /// Return the internal tree as a Store. 206 Store asStore() const { 207 return asImmutableMap().getRootWithoutRetain(); 208 } 209 210 void printJson(raw_ostream &Out, const char *NL = "\n", 211 unsigned int Space = 0, bool IsDot = false) const { 212 for (iterator I = begin(); I != end(); ++I) { 213 Indent(Out, Space, IsDot) 214 << "{ \"cluster\": \"" << I.getKey() << "\", \"items\": [" << NL; 215 216 ++Space; 217 const ClusterBindings &CB = I.getData(); 218 for (ClusterBindings::iterator CI = CB.begin(); CI != CB.end(); ++CI) { 219 Indent(Out, Space, IsDot) << "{ " << CI.getKey() << ", \"value\": \"" 220 << CI.getData() << "\" }"; 221 if (std::next(CI) != CB.end()) 222 Out << ','; 223 Out << NL; 224 } 225 226 --Space; 227 Indent(Out, Space, IsDot) << "]}"; 228 if (std::next(I) != end()) 229 Out << ','; 230 Out << NL; 231 } 232 } 233 234 LLVM_DUMP_METHOD void dump() const { printJson(llvm::errs()); } 235 }; 236 } // end anonymous namespace 237 238 typedef const RegionBindingsRef& RegionBindingsConstRef; 239 240 Optional<SVal> RegionBindingsRef::getDirectBinding(const MemRegion *R) const { 241 return Optional<SVal>::create(lookup(R, BindingKey::Direct)); 242 } 243 244 Optional<SVal> RegionBindingsRef::getDefaultBinding(const MemRegion *R) const { 245 return Optional<SVal>::create(lookup(R, BindingKey::Default)); 246 } 247 248 RegionBindingsRef RegionBindingsRef::addBinding(BindingKey K, SVal V) const { 249 const MemRegion *Base = K.getBaseRegion(); 250 251 const ClusterBindings *ExistingCluster = lookup(Base); 252 ClusterBindings Cluster = 253 (ExistingCluster ? *ExistingCluster : CBFactory->getEmptyMap()); 254 255 ClusterBindings NewCluster = CBFactory->add(Cluster, K, V); 256 return add(Base, NewCluster); 257 } 258 259 260 RegionBindingsRef RegionBindingsRef::addBinding(const MemRegion *R, 261 BindingKey::Kind k, 262 SVal V) const { 263 return addBinding(BindingKey::Make(R, k), V); 264 } 265 266 const SVal *RegionBindingsRef::lookup(BindingKey K) const { 267 const ClusterBindings *Cluster = lookup(K.getBaseRegion()); 268 if (!Cluster) 269 return nullptr; 270 return Cluster->lookup(K); 271 } 272 273 const SVal *RegionBindingsRef::lookup(const MemRegion *R, 274 BindingKey::Kind k) const { 275 return lookup(BindingKey::Make(R, k)); 276 } 277 278 RegionBindingsRef RegionBindingsRef::removeBinding(BindingKey K) { 279 const MemRegion *Base = K.getBaseRegion(); 280 const ClusterBindings *Cluster = lookup(Base); 281 if (!Cluster) 282 return *this; 283 284 ClusterBindings NewCluster = CBFactory->remove(*Cluster, K); 285 if (NewCluster.isEmpty()) 286 return remove(Base); 287 return add(Base, NewCluster); 288 } 289 290 RegionBindingsRef RegionBindingsRef::removeBinding(const MemRegion *R, 291 BindingKey::Kind k){ 292 return removeBinding(BindingKey::Make(R, k)); 293 } 294 295 //===----------------------------------------------------------------------===// 296 // Fine-grained control of RegionStoreManager. 297 //===----------------------------------------------------------------------===// 298 299 namespace { 300 struct minimal_features_tag {}; 301 struct maximal_features_tag {}; 302 303 class RegionStoreFeatures { 304 bool SupportsFields; 305 public: 306 RegionStoreFeatures(minimal_features_tag) : 307 SupportsFields(false) {} 308 309 RegionStoreFeatures(maximal_features_tag) : 310 SupportsFields(true) {} 311 312 void enableFields(bool t) { SupportsFields = t; } 313 314 bool supportsFields() const { return SupportsFields; } 315 }; 316 } 317 318 //===----------------------------------------------------------------------===// 319 // Main RegionStore logic. 320 //===----------------------------------------------------------------------===// 321 322 namespace { 323 class InvalidateRegionsWorker; 324 325 class RegionStoreManager : public StoreManager { 326 public: 327 const RegionStoreFeatures Features; 328 329 RegionBindings::Factory RBFactory; 330 mutable ClusterBindings::Factory CBFactory; 331 332 typedef std::vector<SVal> SValListTy; 333 private: 334 typedef llvm::DenseMap<const LazyCompoundValData *, 335 SValListTy> LazyBindingsMapTy; 336 LazyBindingsMapTy LazyBindingsMap; 337 338 /// The largest number of fields a struct can have and still be 339 /// considered "small". 340 /// 341 /// This is currently used to decide whether or not it is worth "forcing" a 342 /// LazyCompoundVal on bind. 343 /// 344 /// This is controlled by 'region-store-small-struct-limit' option. 345 /// To disable all small-struct-dependent behavior, set the option to "0". 346 unsigned SmallStructLimit; 347 348 /// A helper used to populate the work list with the given set of 349 /// regions. 350 void populateWorkList(InvalidateRegionsWorker &W, 351 ArrayRef<SVal> Values, 352 InvalidatedRegions *TopLevelRegions); 353 354 public: 355 RegionStoreManager(ProgramStateManager& mgr, const RegionStoreFeatures &f) 356 : StoreManager(mgr), Features(f), 357 RBFactory(mgr.getAllocator()), CBFactory(mgr.getAllocator()), 358 SmallStructLimit(0) { 359 SubEngine &Eng = StateMgr.getOwningEngine(); 360 AnalyzerOptions &Options = Eng.getAnalysisManager().options; 361 SmallStructLimit = Options.RegionStoreSmallStructLimit; 362 } 363 364 365 /// setImplicitDefaultValue - Set the default binding for the provided 366 /// MemRegion to the value implicitly defined for compound literals when 367 /// the value is not specified. 368 RegionBindingsRef setImplicitDefaultValue(RegionBindingsConstRef B, 369 const MemRegion *R, QualType T); 370 371 /// ArrayToPointer - Emulates the "decay" of an array to a pointer 372 /// type. 'Array' represents the lvalue of the array being decayed 373 /// to a pointer, and the returned SVal represents the decayed 374 /// version of that lvalue (i.e., a pointer to the first element of 375 /// the array). This is called by ExprEngine when evaluating 376 /// casts from arrays to pointers. 377 SVal ArrayToPointer(Loc Array, QualType ElementTy) override; 378 379 StoreRef getInitialStore(const LocationContext *InitLoc) override { 380 return StoreRef(RBFactory.getEmptyMap().getRootWithoutRetain(), *this); 381 } 382 383 //===-------------------------------------------------------------------===// 384 // Binding values to regions. 385 //===-------------------------------------------------------------------===// 386 RegionBindingsRef invalidateGlobalRegion(MemRegion::Kind K, 387 const Expr *Ex, 388 unsigned Count, 389 const LocationContext *LCtx, 390 RegionBindingsRef B, 391 InvalidatedRegions *Invalidated); 392 393 StoreRef invalidateRegions(Store store, 394 ArrayRef<SVal> Values, 395 const Expr *E, unsigned Count, 396 const LocationContext *LCtx, 397 const CallEvent *Call, 398 InvalidatedSymbols &IS, 399 RegionAndSymbolInvalidationTraits &ITraits, 400 InvalidatedRegions *Invalidated, 401 InvalidatedRegions *InvalidatedTopLevel) override; 402 403 bool scanReachableSymbols(Store S, const MemRegion *R, 404 ScanReachableSymbols &Callbacks) override; 405 406 RegionBindingsRef removeSubRegionBindings(RegionBindingsConstRef B, 407 const SubRegion *R); 408 409 public: // Part of public interface to class. 410 411 StoreRef Bind(Store store, Loc LV, SVal V) override { 412 return StoreRef(bind(getRegionBindings(store), LV, V).asStore(), *this); 413 } 414 415 RegionBindingsRef bind(RegionBindingsConstRef B, Loc LV, SVal V); 416 417 // BindDefaultInitial is only used to initialize a region with 418 // a default value. 419 StoreRef BindDefaultInitial(Store store, const MemRegion *R, 420 SVal V) override { 421 RegionBindingsRef B = getRegionBindings(store); 422 // Use other APIs when you have to wipe the region that was initialized 423 // earlier. 424 assert(!(B.getDefaultBinding(R) || B.getDirectBinding(R)) && 425 "Double initialization!"); 426 B = B.addBinding(BindingKey::Make(R, BindingKey::Default), V); 427 return StoreRef(B.asImmutableMap().getRootWithoutRetain(), *this); 428 } 429 430 // BindDefaultZero is used for zeroing constructors that may accidentally 431 // overwrite existing bindings. 432 StoreRef BindDefaultZero(Store store, const MemRegion *R) override { 433 // FIXME: The offsets of empty bases can be tricky because of 434 // of the so called "empty base class optimization". 435 // If a base class has been optimized out 436 // we should not try to create a binding, otherwise we should. 437 // Unfortunately, at the moment ASTRecordLayout doesn't expose 438 // the actual sizes of the empty bases 439 // and trying to infer them from offsets/alignments 440 // seems to be error-prone and non-trivial because of the trailing padding. 441 // As a temporary mitigation we don't create bindings for empty bases. 442 if (const auto *BR = dyn_cast<CXXBaseObjectRegion>(R)) 443 if (BR->getDecl()->isEmpty()) 444 return StoreRef(store, *this); 445 446 RegionBindingsRef B = getRegionBindings(store); 447 SVal V = svalBuilder.makeZeroVal(Ctx.CharTy); 448 B = removeSubRegionBindings(B, cast<SubRegion>(R)); 449 B = B.addBinding(BindingKey::Make(R, BindingKey::Default), V); 450 return StoreRef(B.asImmutableMap().getRootWithoutRetain(), *this); 451 } 452 453 /// Attempt to extract the fields of \p LCV and bind them to the struct region 454 /// \p R. 455 /// 456 /// This path is used when it seems advantageous to "force" loading the values 457 /// within a LazyCompoundVal to bind memberwise to the struct region, rather 458 /// than using a Default binding at the base of the entire region. This is a 459 /// heuristic attempting to avoid building long chains of LazyCompoundVals. 460 /// 461 /// \returns The updated store bindings, or \c None if binding non-lazily 462 /// would be too expensive. 463 Optional<RegionBindingsRef> tryBindSmallStruct(RegionBindingsConstRef B, 464 const TypedValueRegion *R, 465 const RecordDecl *RD, 466 nonloc::LazyCompoundVal LCV); 467 468 /// BindStruct - Bind a compound value to a structure. 469 RegionBindingsRef bindStruct(RegionBindingsConstRef B, 470 const TypedValueRegion* R, SVal V); 471 472 /// BindVector - Bind a compound value to a vector. 473 RegionBindingsRef bindVector(RegionBindingsConstRef B, 474 const TypedValueRegion* R, SVal V); 475 476 RegionBindingsRef bindArray(RegionBindingsConstRef B, 477 const TypedValueRegion* R, 478 SVal V); 479 480 /// Clears out all bindings in the given region and assigns a new value 481 /// as a Default binding. 482 RegionBindingsRef bindAggregate(RegionBindingsConstRef B, 483 const TypedRegion *R, 484 SVal DefaultVal); 485 486 /// Create a new store with the specified binding removed. 487 /// \param ST the original store, that is the basis for the new store. 488 /// \param L the location whose binding should be removed. 489 StoreRef killBinding(Store ST, Loc L) override; 490 491 void incrementReferenceCount(Store store) override { 492 getRegionBindings(store).manualRetain(); 493 } 494 495 /// If the StoreManager supports it, decrement the reference count of 496 /// the specified Store object. If the reference count hits 0, the memory 497 /// associated with the object is recycled. 498 void decrementReferenceCount(Store store) override { 499 getRegionBindings(store).manualRelease(); 500 } 501 502 bool includedInBindings(Store store, const MemRegion *region) const override; 503 504 /// Return the value bound to specified location in a given state. 505 /// 506 /// The high level logic for this method is this: 507 /// getBinding (L) 508 /// if L has binding 509 /// return L's binding 510 /// else if L is in killset 511 /// return unknown 512 /// else 513 /// if L is on stack or heap 514 /// return undefined 515 /// else 516 /// return symbolic 517 SVal getBinding(Store S, Loc L, QualType T) override { 518 return getBinding(getRegionBindings(S), L, T); 519 } 520 521 Optional<SVal> getDefaultBinding(Store S, const MemRegion *R) override { 522 RegionBindingsRef B = getRegionBindings(S); 523 // Default bindings are always applied over a base region so look up the 524 // base region's default binding, otherwise the lookup will fail when R 525 // is at an offset from R->getBaseRegion(). 526 return B.getDefaultBinding(R->getBaseRegion()); 527 } 528 529 SVal getBinding(RegionBindingsConstRef B, Loc L, QualType T = QualType()); 530 531 SVal getBindingForElement(RegionBindingsConstRef B, const ElementRegion *R); 532 533 SVal getBindingForField(RegionBindingsConstRef B, const FieldRegion *R); 534 535 SVal getBindingForObjCIvar(RegionBindingsConstRef B, const ObjCIvarRegion *R); 536 537 SVal getBindingForVar(RegionBindingsConstRef B, const VarRegion *R); 538 539 SVal getBindingForLazySymbol(const TypedValueRegion *R); 540 541 SVal getBindingForFieldOrElementCommon(RegionBindingsConstRef B, 542 const TypedValueRegion *R, 543 QualType Ty); 544 545 SVal getLazyBinding(const SubRegion *LazyBindingRegion, 546 RegionBindingsRef LazyBinding); 547 548 /// Get bindings for the values in a struct and return a CompoundVal, used 549 /// when doing struct copy: 550 /// struct s x, y; 551 /// x = y; 552 /// y's value is retrieved by this method. 553 SVal getBindingForStruct(RegionBindingsConstRef B, const TypedValueRegion *R); 554 SVal getBindingForArray(RegionBindingsConstRef B, const TypedValueRegion *R); 555 NonLoc createLazyBinding(RegionBindingsConstRef B, const TypedValueRegion *R); 556 557 /// Used to lazily generate derived symbols for bindings that are defined 558 /// implicitly by default bindings in a super region. 559 /// 560 /// Note that callers may need to specially handle LazyCompoundVals, which 561 /// are returned as is in case the caller needs to treat them differently. 562 Optional<SVal> getBindingForDerivedDefaultValue(RegionBindingsConstRef B, 563 const MemRegion *superR, 564 const TypedValueRegion *R, 565 QualType Ty); 566 567 /// Get the state and region whose binding this region \p R corresponds to. 568 /// 569 /// If there is no lazy binding for \p R, the returned value will have a null 570 /// \c second. Note that a null pointer can represents a valid Store. 571 std::pair<Store, const SubRegion *> 572 findLazyBinding(RegionBindingsConstRef B, const SubRegion *R, 573 const SubRegion *originalRegion); 574 575 /// Returns the cached set of interesting SVals contained within a lazy 576 /// binding. 577 /// 578 /// The precise value of "interesting" is determined for the purposes of 579 /// RegionStore's internal analysis. It must always contain all regions and 580 /// symbols, but may omit constants and other kinds of SVal. 581 const SValListTy &getInterestingValues(nonloc::LazyCompoundVal LCV); 582 583 //===------------------------------------------------------------------===// 584 // State pruning. 585 //===------------------------------------------------------------------===// 586 587 /// removeDeadBindings - Scans the RegionStore of 'state' for dead values. 588 /// It returns a new Store with these values removed. 589 StoreRef removeDeadBindings(Store store, const StackFrameContext *LCtx, 590 SymbolReaper& SymReaper) override; 591 592 //===------------------------------------------------------------------===// 593 // Region "extents". 594 //===------------------------------------------------------------------===// 595 596 // FIXME: This method will soon be eliminated; see the note in Store.h. 597 DefinedOrUnknownSVal getSizeInElements(ProgramStateRef state, 598 const MemRegion* R, 599 QualType EleTy) override; 600 601 //===------------------------------------------------------------------===// 602 // Utility methods. 603 //===------------------------------------------------------------------===// 604 605 RegionBindingsRef getRegionBindings(Store store) const { 606 return RegionBindingsRef(CBFactory, 607 static_cast<const RegionBindings::TreeTy*>(store), 608 RBFactory.getTreeFactory()); 609 } 610 611 void printJson(raw_ostream &Out, Store S, const char *NL = "\n", 612 unsigned int Space = 0, bool IsDot = false) const override; 613 614 void iterBindings(Store store, BindingsHandler& f) override { 615 RegionBindingsRef B = getRegionBindings(store); 616 for (RegionBindingsRef::iterator I = B.begin(), E = B.end(); I != E; ++I) { 617 const ClusterBindings &Cluster = I.getData(); 618 for (ClusterBindings::iterator CI = Cluster.begin(), CE = Cluster.end(); 619 CI != CE; ++CI) { 620 const BindingKey &K = CI.getKey(); 621 if (!K.isDirect()) 622 continue; 623 if (const SubRegion *R = dyn_cast<SubRegion>(K.getRegion())) { 624 // FIXME: Possibly incorporate the offset? 625 if (!f.HandleBinding(*this, store, R, CI.getData())) 626 return; 627 } 628 } 629 } 630 } 631 }; 632 633 } // end anonymous namespace 634 635 //===----------------------------------------------------------------------===// 636 // RegionStore creation. 637 //===----------------------------------------------------------------------===// 638 639 std::unique_ptr<StoreManager> 640 ento::CreateRegionStoreManager(ProgramStateManager &StMgr) { 641 RegionStoreFeatures F = maximal_features_tag(); 642 return llvm::make_unique<RegionStoreManager>(StMgr, F); 643 } 644 645 std::unique_ptr<StoreManager> 646 ento::CreateFieldsOnlyRegionStoreManager(ProgramStateManager &StMgr) { 647 RegionStoreFeatures F = minimal_features_tag(); 648 F.enableFields(true); 649 return llvm::make_unique<RegionStoreManager>(StMgr, F); 650 } 651 652 653 //===----------------------------------------------------------------------===// 654 // Region Cluster analysis. 655 //===----------------------------------------------------------------------===// 656 657 namespace { 658 /// Used to determine which global regions are automatically included in the 659 /// initial worklist of a ClusterAnalysis. 660 enum GlobalsFilterKind { 661 /// Don't include any global regions. 662 GFK_None, 663 /// Only include system globals. 664 GFK_SystemOnly, 665 /// Include all global regions. 666 GFK_All 667 }; 668 669 template <typename DERIVED> 670 class ClusterAnalysis { 671 protected: 672 typedef llvm::DenseMap<const MemRegion *, const ClusterBindings *> ClusterMap; 673 typedef const MemRegion * WorkListElement; 674 typedef SmallVector<WorkListElement, 10> WorkList; 675 676 llvm::SmallPtrSet<const ClusterBindings *, 16> Visited; 677 678 WorkList WL; 679 680 RegionStoreManager &RM; 681 ASTContext &Ctx; 682 SValBuilder &svalBuilder; 683 684 RegionBindingsRef B; 685 686 687 protected: 688 const ClusterBindings *getCluster(const MemRegion *R) { 689 return B.lookup(R); 690 } 691 692 /// Returns true if all clusters in the given memspace should be initially 693 /// included in the cluster analysis. Subclasses may provide their 694 /// own implementation. 695 bool includeEntireMemorySpace(const MemRegion *Base) { 696 return false; 697 } 698 699 public: 700 ClusterAnalysis(RegionStoreManager &rm, ProgramStateManager &StateMgr, 701 RegionBindingsRef b) 702 : RM(rm), Ctx(StateMgr.getContext()), 703 svalBuilder(StateMgr.getSValBuilder()), B(std::move(b)) {} 704 705 RegionBindingsRef getRegionBindings() const { return B; } 706 707 bool isVisited(const MemRegion *R) { 708 return Visited.count(getCluster(R)); 709 } 710 711 void GenerateClusters() { 712 // Scan the entire set of bindings and record the region clusters. 713 for (RegionBindingsRef::iterator RI = B.begin(), RE = B.end(); 714 RI != RE; ++RI){ 715 const MemRegion *Base = RI.getKey(); 716 717 const ClusterBindings &Cluster = RI.getData(); 718 assert(!Cluster.isEmpty() && "Empty clusters should be removed"); 719 static_cast<DERIVED*>(this)->VisitAddedToCluster(Base, Cluster); 720 721 // If the base's memspace should be entirely invalidated, add the cluster 722 // to the workspace up front. 723 if (static_cast<DERIVED*>(this)->includeEntireMemorySpace(Base)) 724 AddToWorkList(WorkListElement(Base), &Cluster); 725 } 726 } 727 728 bool AddToWorkList(WorkListElement E, const ClusterBindings *C) { 729 if (C && !Visited.insert(C).second) 730 return false; 731 WL.push_back(E); 732 return true; 733 } 734 735 bool AddToWorkList(const MemRegion *R) { 736 return static_cast<DERIVED*>(this)->AddToWorkList(R); 737 } 738 739 void RunWorkList() { 740 while (!WL.empty()) { 741 WorkListElement E = WL.pop_back_val(); 742 const MemRegion *BaseR = E; 743 744 static_cast<DERIVED*>(this)->VisitCluster(BaseR, getCluster(BaseR)); 745 } 746 } 747 748 void VisitAddedToCluster(const MemRegion *baseR, const ClusterBindings &C) {} 749 void VisitCluster(const MemRegion *baseR, const ClusterBindings *C) {} 750 751 void VisitCluster(const MemRegion *BaseR, const ClusterBindings *C, 752 bool Flag) { 753 static_cast<DERIVED*>(this)->VisitCluster(BaseR, C); 754 } 755 }; 756 } 757 758 //===----------------------------------------------------------------------===// 759 // Binding invalidation. 760 //===----------------------------------------------------------------------===// 761 762 bool RegionStoreManager::scanReachableSymbols(Store S, const MemRegion *R, 763 ScanReachableSymbols &Callbacks) { 764 assert(R == R->getBaseRegion() && "Should only be called for base regions"); 765 RegionBindingsRef B = getRegionBindings(S); 766 const ClusterBindings *Cluster = B.lookup(R); 767 768 if (!Cluster) 769 return true; 770 771 for (ClusterBindings::iterator RI = Cluster->begin(), RE = Cluster->end(); 772 RI != RE; ++RI) { 773 if (!Callbacks.scan(RI.getData())) 774 return false; 775 } 776 777 return true; 778 } 779 780 static inline bool isUnionField(const FieldRegion *FR) { 781 return FR->getDecl()->getParent()->isUnion(); 782 } 783 784 typedef SmallVector<const FieldDecl *, 8> FieldVector; 785 786 static void getSymbolicOffsetFields(BindingKey K, FieldVector &Fields) { 787 assert(K.hasSymbolicOffset() && "Not implemented for concrete offset keys"); 788 789 const MemRegion *Base = K.getConcreteOffsetRegion(); 790 const MemRegion *R = K.getRegion(); 791 792 while (R != Base) { 793 if (const FieldRegion *FR = dyn_cast<FieldRegion>(R)) 794 if (!isUnionField(FR)) 795 Fields.push_back(FR->getDecl()); 796 797 R = cast<SubRegion>(R)->getSuperRegion(); 798 } 799 } 800 801 static bool isCompatibleWithFields(BindingKey K, const FieldVector &Fields) { 802 assert(K.hasSymbolicOffset() && "Not implemented for concrete offset keys"); 803 804 if (Fields.empty()) 805 return true; 806 807 FieldVector FieldsInBindingKey; 808 getSymbolicOffsetFields(K, FieldsInBindingKey); 809 810 ptrdiff_t Delta = FieldsInBindingKey.size() - Fields.size(); 811 if (Delta >= 0) 812 return std::equal(FieldsInBindingKey.begin() + Delta, 813 FieldsInBindingKey.end(), 814 Fields.begin()); 815 else 816 return std::equal(FieldsInBindingKey.begin(), FieldsInBindingKey.end(), 817 Fields.begin() - Delta); 818 } 819 820 /// Collects all bindings in \p Cluster that may refer to bindings within 821 /// \p Top. 822 /// 823 /// Each binding is a pair whose \c first is the key (a BindingKey) and whose 824 /// \c second is the value (an SVal). 825 /// 826 /// The \p IncludeAllDefaultBindings parameter specifies whether to include 827 /// default bindings that may extend beyond \p Top itself, e.g. if \p Top is 828 /// an aggregate within a larger aggregate with a default binding. 829 static void 830 collectSubRegionBindings(SmallVectorImpl<BindingPair> &Bindings, 831 SValBuilder &SVB, const ClusterBindings &Cluster, 832 const SubRegion *Top, BindingKey TopKey, 833 bool IncludeAllDefaultBindings) { 834 FieldVector FieldsInSymbolicSubregions; 835 if (TopKey.hasSymbolicOffset()) { 836 getSymbolicOffsetFields(TopKey, FieldsInSymbolicSubregions); 837 Top = TopKey.getConcreteOffsetRegion(); 838 TopKey = BindingKey::Make(Top, BindingKey::Default); 839 } 840 841 // Find the length (in bits) of the region being invalidated. 842 uint64_t Length = UINT64_MAX; 843 SVal Extent = Top->getExtent(SVB); 844 if (Optional<nonloc::ConcreteInt> ExtentCI = 845 Extent.getAs<nonloc::ConcreteInt>()) { 846 const llvm::APSInt &ExtentInt = ExtentCI->getValue(); 847 assert(ExtentInt.isNonNegative() || ExtentInt.isUnsigned()); 848 // Extents are in bytes but region offsets are in bits. Be careful! 849 Length = ExtentInt.getLimitedValue() * SVB.getContext().getCharWidth(); 850 } else if (const FieldRegion *FR = dyn_cast<FieldRegion>(Top)) { 851 if (FR->getDecl()->isBitField()) 852 Length = FR->getDecl()->getBitWidthValue(SVB.getContext()); 853 } 854 855 for (ClusterBindings::iterator I = Cluster.begin(), E = Cluster.end(); 856 I != E; ++I) { 857 BindingKey NextKey = I.getKey(); 858 if (NextKey.getRegion() == TopKey.getRegion()) { 859 // FIXME: This doesn't catch the case where we're really invalidating a 860 // region with a symbolic offset. Example: 861 // R: points[i].y 862 // Next: points[0].x 863 864 if (NextKey.getOffset() > TopKey.getOffset() && 865 NextKey.getOffset() - TopKey.getOffset() < Length) { 866 // Case 1: The next binding is inside the region we're invalidating. 867 // Include it. 868 Bindings.push_back(*I); 869 870 } else if (NextKey.getOffset() == TopKey.getOffset()) { 871 // Case 2: The next binding is at the same offset as the region we're 872 // invalidating. In this case, we need to leave default bindings alone, 873 // since they may be providing a default value for a regions beyond what 874 // we're invalidating. 875 // FIXME: This is probably incorrect; consider invalidating an outer 876 // struct whose first field is bound to a LazyCompoundVal. 877 if (IncludeAllDefaultBindings || NextKey.isDirect()) 878 Bindings.push_back(*I); 879 } 880 881 } else if (NextKey.hasSymbolicOffset()) { 882 const MemRegion *Base = NextKey.getConcreteOffsetRegion(); 883 if (Top->isSubRegionOf(Base) && Top != Base) { 884 // Case 3: The next key is symbolic and we just changed something within 885 // its concrete region. We don't know if the binding is still valid, so 886 // we'll be conservative and include it. 887 if (IncludeAllDefaultBindings || NextKey.isDirect()) 888 if (isCompatibleWithFields(NextKey, FieldsInSymbolicSubregions)) 889 Bindings.push_back(*I); 890 } else if (const SubRegion *BaseSR = dyn_cast<SubRegion>(Base)) { 891 // Case 4: The next key is symbolic, but we changed a known 892 // super-region. In this case the binding is certainly included. 893 if (BaseSR->isSubRegionOf(Top)) 894 if (isCompatibleWithFields(NextKey, FieldsInSymbolicSubregions)) 895 Bindings.push_back(*I); 896 } 897 } 898 } 899 } 900 901 static void 902 collectSubRegionBindings(SmallVectorImpl<BindingPair> &Bindings, 903 SValBuilder &SVB, const ClusterBindings &Cluster, 904 const SubRegion *Top, bool IncludeAllDefaultBindings) { 905 collectSubRegionBindings(Bindings, SVB, Cluster, Top, 906 BindingKey::Make(Top, BindingKey::Default), 907 IncludeAllDefaultBindings); 908 } 909 910 RegionBindingsRef 911 RegionStoreManager::removeSubRegionBindings(RegionBindingsConstRef B, 912 const SubRegion *Top) { 913 BindingKey TopKey = BindingKey::Make(Top, BindingKey::Default); 914 const MemRegion *ClusterHead = TopKey.getBaseRegion(); 915 916 if (Top == ClusterHead) { 917 // We can remove an entire cluster's bindings all in one go. 918 return B.remove(Top); 919 } 920 921 const ClusterBindings *Cluster = B.lookup(ClusterHead); 922 if (!Cluster) { 923 // If we're invalidating a region with a symbolic offset, we need to make 924 // sure we don't treat the base region as uninitialized anymore. 925 if (TopKey.hasSymbolicOffset()) { 926 const SubRegion *Concrete = TopKey.getConcreteOffsetRegion(); 927 return B.addBinding(Concrete, BindingKey::Default, UnknownVal()); 928 } 929 return B; 930 } 931 932 SmallVector<BindingPair, 32> Bindings; 933 collectSubRegionBindings(Bindings, svalBuilder, *Cluster, Top, TopKey, 934 /*IncludeAllDefaultBindings=*/false); 935 936 ClusterBindingsRef Result(*Cluster, CBFactory); 937 for (SmallVectorImpl<BindingPair>::const_iterator I = Bindings.begin(), 938 E = Bindings.end(); 939 I != E; ++I) 940 Result = Result.remove(I->first); 941 942 // If we're invalidating a region with a symbolic offset, we need to make sure 943 // we don't treat the base region as uninitialized anymore. 944 // FIXME: This isn't very precise; see the example in 945 // collectSubRegionBindings. 946 if (TopKey.hasSymbolicOffset()) { 947 const SubRegion *Concrete = TopKey.getConcreteOffsetRegion(); 948 Result = Result.add(BindingKey::Make(Concrete, BindingKey::Default), 949 UnknownVal()); 950 } 951 952 if (Result.isEmpty()) 953 return B.remove(ClusterHead); 954 return B.add(ClusterHead, Result.asImmutableMap()); 955 } 956 957 namespace { 958 class InvalidateRegionsWorker : public ClusterAnalysis<InvalidateRegionsWorker> 959 { 960 const Expr *Ex; 961 unsigned Count; 962 const LocationContext *LCtx; 963 InvalidatedSymbols &IS; 964 RegionAndSymbolInvalidationTraits &ITraits; 965 StoreManager::InvalidatedRegions *Regions; 966 GlobalsFilterKind GlobalsFilter; 967 public: 968 InvalidateRegionsWorker(RegionStoreManager &rm, 969 ProgramStateManager &stateMgr, 970 RegionBindingsRef b, 971 const Expr *ex, unsigned count, 972 const LocationContext *lctx, 973 InvalidatedSymbols &is, 974 RegionAndSymbolInvalidationTraits &ITraitsIn, 975 StoreManager::InvalidatedRegions *r, 976 GlobalsFilterKind GFK) 977 : ClusterAnalysis<InvalidateRegionsWorker>(rm, stateMgr, b), 978 Ex(ex), Count(count), LCtx(lctx), IS(is), ITraits(ITraitsIn), Regions(r), 979 GlobalsFilter(GFK) {} 980 981 void VisitCluster(const MemRegion *baseR, const ClusterBindings *C); 982 void VisitBinding(SVal V); 983 984 using ClusterAnalysis::AddToWorkList; 985 986 bool AddToWorkList(const MemRegion *R); 987 988 /// Returns true if all clusters in the memory space for \p Base should be 989 /// be invalidated. 990 bool includeEntireMemorySpace(const MemRegion *Base); 991 992 /// Returns true if the memory space of the given region is one of the global 993 /// regions specially included at the start of invalidation. 994 bool isInitiallyIncludedGlobalRegion(const MemRegion *R); 995 }; 996 } 997 998 bool InvalidateRegionsWorker::AddToWorkList(const MemRegion *R) { 999 bool doNotInvalidateSuperRegion = ITraits.hasTrait( 1000 R, RegionAndSymbolInvalidationTraits::TK_DoNotInvalidateSuperRegion); 1001 const MemRegion *BaseR = doNotInvalidateSuperRegion ? R : R->getBaseRegion(); 1002 return AddToWorkList(WorkListElement(BaseR), getCluster(BaseR)); 1003 } 1004 1005 void InvalidateRegionsWorker::VisitBinding(SVal V) { 1006 // A symbol? Mark it touched by the invalidation. 1007 if (SymbolRef Sym = V.getAsSymbol()) 1008 IS.insert(Sym); 1009 1010 if (const MemRegion *R = V.getAsRegion()) { 1011 AddToWorkList(R); 1012 return; 1013 } 1014 1015 // Is it a LazyCompoundVal? All references get invalidated as well. 1016 if (Optional<nonloc::LazyCompoundVal> LCS = 1017 V.getAs<nonloc::LazyCompoundVal>()) { 1018 1019 const RegionStoreManager::SValListTy &Vals = RM.getInterestingValues(*LCS); 1020 1021 for (RegionStoreManager::SValListTy::const_iterator I = Vals.begin(), 1022 E = Vals.end(); 1023 I != E; ++I) 1024 VisitBinding(*I); 1025 1026 return; 1027 } 1028 } 1029 1030 void InvalidateRegionsWorker::VisitCluster(const MemRegion *baseR, 1031 const ClusterBindings *C) { 1032 1033 bool PreserveRegionsContents = 1034 ITraits.hasTrait(baseR, 1035 RegionAndSymbolInvalidationTraits::TK_PreserveContents); 1036 1037 if (C) { 1038 for (ClusterBindings::iterator I = C->begin(), E = C->end(); I != E; ++I) 1039 VisitBinding(I.getData()); 1040 1041 // Invalidate regions contents. 1042 if (!PreserveRegionsContents) 1043 B = B.remove(baseR); 1044 } 1045 1046 if (const auto *TO = dyn_cast<TypedValueRegion>(baseR)) { 1047 if (const auto *RD = TO->getValueType()->getAsCXXRecordDecl()) { 1048 1049 // Lambdas can affect all static local variables without explicitly 1050 // capturing those. 1051 // We invalidate all static locals referenced inside the lambda body. 1052 if (RD->isLambda() && RD->getLambdaCallOperator()->getBody()) { 1053 using namespace ast_matchers; 1054 1055 const char *DeclBind = "DeclBind"; 1056 StatementMatcher RefToStatic = stmt(hasDescendant(declRefExpr( 1057 to(varDecl(hasStaticStorageDuration()).bind(DeclBind))))); 1058 auto Matches = 1059 match(RefToStatic, *RD->getLambdaCallOperator()->getBody(), 1060 RD->getASTContext()); 1061 1062 for (BoundNodes &Match : Matches) { 1063 auto *VD = Match.getNodeAs<VarDecl>(DeclBind); 1064 const VarRegion *ToInvalidate = 1065 RM.getRegionManager().getVarRegion(VD, LCtx); 1066 AddToWorkList(ToInvalidate); 1067 } 1068 } 1069 } 1070 } 1071 1072 // BlockDataRegion? If so, invalidate captured variables that are passed 1073 // by reference. 1074 if (const BlockDataRegion *BR = dyn_cast<BlockDataRegion>(baseR)) { 1075 for (BlockDataRegion::referenced_vars_iterator 1076 BI = BR->referenced_vars_begin(), BE = BR->referenced_vars_end() ; 1077 BI != BE; ++BI) { 1078 const VarRegion *VR = BI.getCapturedRegion(); 1079 const VarDecl *VD = VR->getDecl(); 1080 if (VD->hasAttr<BlocksAttr>() || !VD->hasLocalStorage()) { 1081 AddToWorkList(VR); 1082 } 1083 else if (Loc::isLocType(VR->getValueType())) { 1084 // Map the current bindings to a Store to retrieve the value 1085 // of the binding. If that binding itself is a region, we should 1086 // invalidate that region. This is because a block may capture 1087 // a pointer value, but the thing pointed by that pointer may 1088 // get invalidated. 1089 SVal V = RM.getBinding(B, loc::MemRegionVal(VR)); 1090 if (Optional<Loc> L = V.getAs<Loc>()) { 1091 if (const MemRegion *LR = L->getAsRegion()) 1092 AddToWorkList(LR); 1093 } 1094 } 1095 } 1096 return; 1097 } 1098 1099 // Symbolic region? 1100 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(baseR)) 1101 IS.insert(SR->getSymbol()); 1102 1103 // Nothing else should be done in the case when we preserve regions context. 1104 if (PreserveRegionsContents) 1105 return; 1106 1107 // Otherwise, we have a normal data region. Record that we touched the region. 1108 if (Regions) 1109 Regions->push_back(baseR); 1110 1111 if (isa<AllocaRegion>(baseR) || isa<SymbolicRegion>(baseR)) { 1112 // Invalidate the region by setting its default value to 1113 // conjured symbol. The type of the symbol is irrelevant. 1114 DefinedOrUnknownSVal V = 1115 svalBuilder.conjureSymbolVal(baseR, Ex, LCtx, Ctx.IntTy, Count); 1116 B = B.addBinding(baseR, BindingKey::Default, V); 1117 return; 1118 } 1119 1120 if (!baseR->isBoundable()) 1121 return; 1122 1123 const TypedValueRegion *TR = cast<TypedValueRegion>(baseR); 1124 QualType T = TR->getValueType(); 1125 1126 if (isInitiallyIncludedGlobalRegion(baseR)) { 1127 // If the region is a global and we are invalidating all globals, 1128 // erasing the entry is good enough. This causes all globals to be lazily 1129 // symbolicated from the same base symbol. 1130 return; 1131 } 1132 1133 if (T->isRecordType()) { 1134 // Invalidate the region by setting its default value to 1135 // conjured symbol. The type of the symbol is irrelevant. 1136 DefinedOrUnknownSVal V = svalBuilder.conjureSymbolVal(baseR, Ex, LCtx, 1137 Ctx.IntTy, Count); 1138 B = B.addBinding(baseR, BindingKey::Default, V); 1139 return; 1140 } 1141 1142 if (const ArrayType *AT = Ctx.getAsArrayType(T)) { 1143 bool doNotInvalidateSuperRegion = ITraits.hasTrait( 1144 baseR, 1145 RegionAndSymbolInvalidationTraits::TK_DoNotInvalidateSuperRegion); 1146 1147 if (doNotInvalidateSuperRegion) { 1148 // We are not doing blank invalidation of the whole array region so we 1149 // have to manually invalidate each elements. 1150 Optional<uint64_t> NumElements; 1151 1152 // Compute lower and upper offsets for region within array. 1153 if (const ConstantArrayType *CAT = dyn_cast<ConstantArrayType>(AT)) 1154 NumElements = CAT->getSize().getZExtValue(); 1155 if (!NumElements) // We are not dealing with a constant size array 1156 goto conjure_default; 1157 QualType ElementTy = AT->getElementType(); 1158 uint64_t ElemSize = Ctx.getTypeSize(ElementTy); 1159 const RegionOffset &RO = baseR->getAsOffset(); 1160 const MemRegion *SuperR = baseR->getBaseRegion(); 1161 if (RO.hasSymbolicOffset()) { 1162 // If base region has a symbolic offset, 1163 // we revert to invalidating the super region. 1164 if (SuperR) 1165 AddToWorkList(SuperR); 1166 goto conjure_default; 1167 } 1168 1169 uint64_t LowerOffset = RO.getOffset(); 1170 uint64_t UpperOffset = LowerOffset + *NumElements * ElemSize; 1171 bool UpperOverflow = UpperOffset < LowerOffset; 1172 1173 // Invalidate regions which are within array boundaries, 1174 // or have a symbolic offset. 1175 if (!SuperR) 1176 goto conjure_default; 1177 1178 const ClusterBindings *C = B.lookup(SuperR); 1179 if (!C) 1180 goto conjure_default; 1181 1182 for (ClusterBindings::iterator I = C->begin(), E = C->end(); I != E; 1183 ++I) { 1184 const BindingKey &BK = I.getKey(); 1185 Optional<uint64_t> ROffset = 1186 BK.hasSymbolicOffset() ? Optional<uint64_t>() : BK.getOffset(); 1187 1188 // Check offset is not symbolic and within array's boundaries. 1189 // Handles arrays of 0 elements and of 0-sized elements as well. 1190 if (!ROffset || 1191 ((*ROffset >= LowerOffset && *ROffset < UpperOffset) || 1192 (UpperOverflow && 1193 (*ROffset >= LowerOffset || *ROffset < UpperOffset)) || 1194 (LowerOffset == UpperOffset && *ROffset == LowerOffset))) { 1195 B = B.removeBinding(I.getKey()); 1196 // Bound symbolic regions need to be invalidated for dead symbol 1197 // detection. 1198 SVal V = I.getData(); 1199 const MemRegion *R = V.getAsRegion(); 1200 if (R && isa<SymbolicRegion>(R)) 1201 VisitBinding(V); 1202 } 1203 } 1204 } 1205 conjure_default: 1206 // Set the default value of the array to conjured symbol. 1207 DefinedOrUnknownSVal V = 1208 svalBuilder.conjureSymbolVal(baseR, Ex, LCtx, 1209 AT->getElementType(), Count); 1210 B = B.addBinding(baseR, BindingKey::Default, V); 1211 return; 1212 } 1213 1214 DefinedOrUnknownSVal V = svalBuilder.conjureSymbolVal(baseR, Ex, LCtx, 1215 T,Count); 1216 assert(SymbolManager::canSymbolicate(T) || V.isUnknown()); 1217 B = B.addBinding(baseR, BindingKey::Direct, V); 1218 } 1219 1220 bool InvalidateRegionsWorker::isInitiallyIncludedGlobalRegion( 1221 const MemRegion *R) { 1222 switch (GlobalsFilter) { 1223 case GFK_None: 1224 return false; 1225 case GFK_SystemOnly: 1226 return isa<GlobalSystemSpaceRegion>(R->getMemorySpace()); 1227 case GFK_All: 1228 return isa<NonStaticGlobalSpaceRegion>(R->getMemorySpace()); 1229 } 1230 1231 llvm_unreachable("unknown globals filter"); 1232 } 1233 1234 bool InvalidateRegionsWorker::includeEntireMemorySpace(const MemRegion *Base) { 1235 if (isInitiallyIncludedGlobalRegion(Base)) 1236 return true; 1237 1238 const MemSpaceRegion *MemSpace = Base->getMemorySpace(); 1239 return ITraits.hasTrait(MemSpace, 1240 RegionAndSymbolInvalidationTraits::TK_EntireMemSpace); 1241 } 1242 1243 RegionBindingsRef 1244 RegionStoreManager::invalidateGlobalRegion(MemRegion::Kind K, 1245 const Expr *Ex, 1246 unsigned Count, 1247 const LocationContext *LCtx, 1248 RegionBindingsRef B, 1249 InvalidatedRegions *Invalidated) { 1250 // Bind the globals memory space to a new symbol that we will use to derive 1251 // the bindings for all globals. 1252 const GlobalsSpaceRegion *GS = MRMgr.getGlobalsRegion(K); 1253 SVal V = svalBuilder.conjureSymbolVal(/* SymbolTag = */ (const void*) GS, Ex, LCtx, 1254 /* type does not matter */ Ctx.IntTy, 1255 Count); 1256 1257 B = B.removeBinding(GS) 1258 .addBinding(BindingKey::Make(GS, BindingKey::Default), V); 1259 1260 // Even if there are no bindings in the global scope, we still need to 1261 // record that we touched it. 1262 if (Invalidated) 1263 Invalidated->push_back(GS); 1264 1265 return B; 1266 } 1267 1268 void RegionStoreManager::populateWorkList(InvalidateRegionsWorker &W, 1269 ArrayRef<SVal> Values, 1270 InvalidatedRegions *TopLevelRegions) { 1271 for (ArrayRef<SVal>::iterator I = Values.begin(), 1272 E = Values.end(); I != E; ++I) { 1273 SVal V = *I; 1274 if (Optional<nonloc::LazyCompoundVal> LCS = 1275 V.getAs<nonloc::LazyCompoundVal>()) { 1276 1277 const SValListTy &Vals = getInterestingValues(*LCS); 1278 1279 for (SValListTy::const_iterator I = Vals.begin(), 1280 E = Vals.end(); I != E; ++I) { 1281 // Note: the last argument is false here because these are 1282 // non-top-level regions. 1283 if (const MemRegion *R = (*I).getAsRegion()) 1284 W.AddToWorkList(R); 1285 } 1286 continue; 1287 } 1288 1289 if (const MemRegion *R = V.getAsRegion()) { 1290 if (TopLevelRegions) 1291 TopLevelRegions->push_back(R); 1292 W.AddToWorkList(R); 1293 continue; 1294 } 1295 } 1296 } 1297 1298 StoreRef 1299 RegionStoreManager::invalidateRegions(Store store, 1300 ArrayRef<SVal> Values, 1301 const Expr *Ex, unsigned Count, 1302 const LocationContext *LCtx, 1303 const CallEvent *Call, 1304 InvalidatedSymbols &IS, 1305 RegionAndSymbolInvalidationTraits &ITraits, 1306 InvalidatedRegions *TopLevelRegions, 1307 InvalidatedRegions *Invalidated) { 1308 GlobalsFilterKind GlobalsFilter; 1309 if (Call) { 1310 if (Call->isInSystemHeader()) 1311 GlobalsFilter = GFK_SystemOnly; 1312 else 1313 GlobalsFilter = GFK_All; 1314 } else { 1315 GlobalsFilter = GFK_None; 1316 } 1317 1318 RegionBindingsRef B = getRegionBindings(store); 1319 InvalidateRegionsWorker W(*this, StateMgr, B, Ex, Count, LCtx, IS, ITraits, 1320 Invalidated, GlobalsFilter); 1321 1322 // Scan the bindings and generate the clusters. 1323 W.GenerateClusters(); 1324 1325 // Add the regions to the worklist. 1326 populateWorkList(W, Values, TopLevelRegions); 1327 1328 W.RunWorkList(); 1329 1330 // Return the new bindings. 1331 B = W.getRegionBindings(); 1332 1333 // For calls, determine which global regions should be invalidated and 1334 // invalidate them. (Note that function-static and immutable globals are never 1335 // invalidated by this.) 1336 // TODO: This could possibly be more precise with modules. 1337 switch (GlobalsFilter) { 1338 case GFK_All: 1339 B = invalidateGlobalRegion(MemRegion::GlobalInternalSpaceRegionKind, 1340 Ex, Count, LCtx, B, Invalidated); 1341 LLVM_FALLTHROUGH; 1342 case GFK_SystemOnly: 1343 B = invalidateGlobalRegion(MemRegion::GlobalSystemSpaceRegionKind, 1344 Ex, Count, LCtx, B, Invalidated); 1345 LLVM_FALLTHROUGH; 1346 case GFK_None: 1347 break; 1348 } 1349 1350 return StoreRef(B.asStore(), *this); 1351 } 1352 1353 //===----------------------------------------------------------------------===// 1354 // Extents for regions. 1355 //===----------------------------------------------------------------------===// 1356 1357 DefinedOrUnknownSVal 1358 RegionStoreManager::getSizeInElements(ProgramStateRef state, 1359 const MemRegion *R, 1360 QualType EleTy) { 1361 SVal Size = cast<SubRegion>(R)->getExtent(svalBuilder); 1362 const llvm::APSInt *SizeInt = svalBuilder.getKnownValue(state, Size); 1363 if (!SizeInt) 1364 return UnknownVal(); 1365 1366 CharUnits RegionSize = CharUnits::fromQuantity(SizeInt->getSExtValue()); 1367 1368 if (Ctx.getAsVariableArrayType(EleTy)) { 1369 // FIXME: We need to track extra state to properly record the size 1370 // of VLAs. Returning UnknownVal here, however, is a stop-gap so that 1371 // we don't have a divide-by-zero below. 1372 return UnknownVal(); 1373 } 1374 1375 CharUnits EleSize = Ctx.getTypeSizeInChars(EleTy); 1376 1377 // If a variable is reinterpreted as a type that doesn't fit into a larger 1378 // type evenly, round it down. 1379 // This is a signed value, since it's used in arithmetic with signed indices. 1380 return svalBuilder.makeIntVal(RegionSize / EleSize, 1381 svalBuilder.getArrayIndexType()); 1382 } 1383 1384 //===----------------------------------------------------------------------===// 1385 // Location and region casting. 1386 //===----------------------------------------------------------------------===// 1387 1388 /// ArrayToPointer - Emulates the "decay" of an array to a pointer 1389 /// type. 'Array' represents the lvalue of the array being decayed 1390 /// to a pointer, and the returned SVal represents the decayed 1391 /// version of that lvalue (i.e., a pointer to the first element of 1392 /// the array). This is called by ExprEngine when evaluating casts 1393 /// from arrays to pointers. 1394 SVal RegionStoreManager::ArrayToPointer(Loc Array, QualType T) { 1395 if (Array.getAs<loc::ConcreteInt>()) 1396 return Array; 1397 1398 if (!Array.getAs<loc::MemRegionVal>()) 1399 return UnknownVal(); 1400 1401 const SubRegion *R = 1402 cast<SubRegion>(Array.castAs<loc::MemRegionVal>().getRegion()); 1403 NonLoc ZeroIdx = svalBuilder.makeZeroArrayIndex(); 1404 return loc::MemRegionVal(MRMgr.getElementRegion(T, ZeroIdx, R, Ctx)); 1405 } 1406 1407 //===----------------------------------------------------------------------===// 1408 // Loading values from regions. 1409 //===----------------------------------------------------------------------===// 1410 1411 SVal RegionStoreManager::getBinding(RegionBindingsConstRef B, Loc L, QualType T) { 1412 assert(!L.getAs<UnknownVal>() && "location unknown"); 1413 assert(!L.getAs<UndefinedVal>() && "location undefined"); 1414 1415 // For access to concrete addresses, return UnknownVal. Checks 1416 // for null dereferences (and similar errors) are done by checkers, not 1417 // the Store. 1418 // FIXME: We can consider lazily symbolicating such memory, but we really 1419 // should defer this when we can reason easily about symbolicating arrays 1420 // of bytes. 1421 if (L.getAs<loc::ConcreteInt>()) { 1422 return UnknownVal(); 1423 } 1424 if (!L.getAs<loc::MemRegionVal>()) { 1425 return UnknownVal(); 1426 } 1427 1428 const MemRegion *MR = L.castAs<loc::MemRegionVal>().getRegion(); 1429 1430 if (isa<BlockDataRegion>(MR)) { 1431 return UnknownVal(); 1432 } 1433 1434 if (!isa<TypedValueRegion>(MR)) { 1435 if (T.isNull()) { 1436 if (const TypedRegion *TR = dyn_cast<TypedRegion>(MR)) 1437 T = TR->getLocationType()->getPointeeType(); 1438 else if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(MR)) 1439 T = SR->getSymbol()->getType()->getPointeeType(); 1440 } 1441 assert(!T.isNull() && "Unable to auto-detect binding type!"); 1442 assert(!T->isVoidType() && "Attempting to dereference a void pointer!"); 1443 MR = GetElementZeroRegion(cast<SubRegion>(MR), T); 1444 } else { 1445 T = cast<TypedValueRegion>(MR)->getValueType(); 1446 } 1447 1448 // FIXME: Perhaps this method should just take a 'const MemRegion*' argument 1449 // instead of 'Loc', and have the other Loc cases handled at a higher level. 1450 const TypedValueRegion *R = cast<TypedValueRegion>(MR); 1451 QualType RTy = R->getValueType(); 1452 1453 // FIXME: we do not yet model the parts of a complex type, so treat the 1454 // whole thing as "unknown". 1455 if (RTy->isAnyComplexType()) 1456 return UnknownVal(); 1457 1458 // FIXME: We should eventually handle funny addressing. e.g.: 1459 // 1460 // int x = ...; 1461 // int *p = &x; 1462 // char *q = (char*) p; 1463 // char c = *q; // returns the first byte of 'x'. 1464 // 1465 // Such funny addressing will occur due to layering of regions. 1466 if (RTy->isStructureOrClassType()) 1467 return getBindingForStruct(B, R); 1468 1469 // FIXME: Handle unions. 1470 if (RTy->isUnionType()) 1471 return createLazyBinding(B, R); 1472 1473 if (RTy->isArrayType()) { 1474 if (RTy->isConstantArrayType()) 1475 return getBindingForArray(B, R); 1476 else 1477 return UnknownVal(); 1478 } 1479 1480 // FIXME: handle Vector types. 1481 if (RTy->isVectorType()) 1482 return UnknownVal(); 1483 1484 if (const FieldRegion* FR = dyn_cast<FieldRegion>(R)) 1485 return CastRetrievedVal(getBindingForField(B, FR), FR, T); 1486 1487 if (const ElementRegion* ER = dyn_cast<ElementRegion>(R)) { 1488 // FIXME: Here we actually perform an implicit conversion from the loaded 1489 // value to the element type. Eventually we want to compose these values 1490 // more intelligently. For example, an 'element' can encompass multiple 1491 // bound regions (e.g., several bound bytes), or could be a subset of 1492 // a larger value. 1493 return CastRetrievedVal(getBindingForElement(B, ER), ER, T); 1494 } 1495 1496 if (const ObjCIvarRegion *IVR = dyn_cast<ObjCIvarRegion>(R)) { 1497 // FIXME: Here we actually perform an implicit conversion from the loaded 1498 // value to the ivar type. What we should model is stores to ivars 1499 // that blow past the extent of the ivar. If the address of the ivar is 1500 // reinterpretted, it is possible we stored a different value that could 1501 // fit within the ivar. Either we need to cast these when storing them 1502 // or reinterpret them lazily (as we do here). 1503 return CastRetrievedVal(getBindingForObjCIvar(B, IVR), IVR, T); 1504 } 1505 1506 if (const VarRegion *VR = dyn_cast<VarRegion>(R)) { 1507 // FIXME: Here we actually perform an implicit conversion from the loaded 1508 // value to the variable type. What we should model is stores to variables 1509 // that blow past the extent of the variable. If the address of the 1510 // variable is reinterpretted, it is possible we stored a different value 1511 // that could fit within the variable. Either we need to cast these when 1512 // storing them or reinterpret them lazily (as we do here). 1513 return CastRetrievedVal(getBindingForVar(B, VR), VR, T); 1514 } 1515 1516 const SVal *V = B.lookup(R, BindingKey::Direct); 1517 1518 // Check if the region has a binding. 1519 if (V) 1520 return *V; 1521 1522 // The location does not have a bound value. This means that it has 1523 // the value it had upon its creation and/or entry to the analyzed 1524 // function/method. These are either symbolic values or 'undefined'. 1525 if (R->hasStackNonParametersStorage()) { 1526 // All stack variables are considered to have undefined values 1527 // upon creation. All heap allocated blocks are considered to 1528 // have undefined values as well unless they are explicitly bound 1529 // to specific values. 1530 return UndefinedVal(); 1531 } 1532 1533 // All other values are symbolic. 1534 return svalBuilder.getRegionValueSymbolVal(R); 1535 } 1536 1537 static QualType getUnderlyingType(const SubRegion *R) { 1538 QualType RegionTy; 1539 if (const TypedValueRegion *TVR = dyn_cast<TypedValueRegion>(R)) 1540 RegionTy = TVR->getValueType(); 1541 1542 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(R)) 1543 RegionTy = SR->getSymbol()->getType(); 1544 1545 return RegionTy; 1546 } 1547 1548 /// Checks to see if store \p B has a lazy binding for region \p R. 1549 /// 1550 /// If \p AllowSubregionBindings is \c false, a lazy binding will be rejected 1551 /// if there are additional bindings within \p R. 1552 /// 1553 /// Note that unlike RegionStoreManager::findLazyBinding, this will not search 1554 /// for lazy bindings for super-regions of \p R. 1555 static Optional<nonloc::LazyCompoundVal> 1556 getExistingLazyBinding(SValBuilder &SVB, RegionBindingsConstRef B, 1557 const SubRegion *R, bool AllowSubregionBindings) { 1558 Optional<SVal> V = B.getDefaultBinding(R); 1559 if (!V) 1560 return None; 1561 1562 Optional<nonloc::LazyCompoundVal> LCV = V->getAs<nonloc::LazyCompoundVal>(); 1563 if (!LCV) 1564 return None; 1565 1566 // If the LCV is for a subregion, the types might not match, and we shouldn't 1567 // reuse the binding. 1568 QualType RegionTy = getUnderlyingType(R); 1569 if (!RegionTy.isNull() && 1570 !RegionTy->isVoidPointerType()) { 1571 QualType SourceRegionTy = LCV->getRegion()->getValueType(); 1572 if (!SVB.getContext().hasSameUnqualifiedType(RegionTy, SourceRegionTy)) 1573 return None; 1574 } 1575 1576 if (!AllowSubregionBindings) { 1577 // If there are any other bindings within this region, we shouldn't reuse 1578 // the top-level binding. 1579 SmallVector<BindingPair, 16> Bindings; 1580 collectSubRegionBindings(Bindings, SVB, *B.lookup(R->getBaseRegion()), R, 1581 /*IncludeAllDefaultBindings=*/true); 1582 if (Bindings.size() > 1) 1583 return None; 1584 } 1585 1586 return *LCV; 1587 } 1588 1589 1590 std::pair<Store, const SubRegion *> 1591 RegionStoreManager::findLazyBinding(RegionBindingsConstRef B, 1592 const SubRegion *R, 1593 const SubRegion *originalRegion) { 1594 if (originalRegion != R) { 1595 if (Optional<nonloc::LazyCompoundVal> V = 1596 getExistingLazyBinding(svalBuilder, B, R, true)) 1597 return std::make_pair(V->getStore(), V->getRegion()); 1598 } 1599 1600 typedef std::pair<Store, const SubRegion *> StoreRegionPair; 1601 StoreRegionPair Result = StoreRegionPair(); 1602 1603 if (const ElementRegion *ER = dyn_cast<ElementRegion>(R)) { 1604 Result = findLazyBinding(B, cast<SubRegion>(ER->getSuperRegion()), 1605 originalRegion); 1606 1607 if (Result.second) 1608 Result.second = MRMgr.getElementRegionWithSuper(ER, Result.second); 1609 1610 } else if (const FieldRegion *FR = dyn_cast<FieldRegion>(R)) { 1611 Result = findLazyBinding(B, cast<SubRegion>(FR->getSuperRegion()), 1612 originalRegion); 1613 1614 if (Result.second) 1615 Result.second = MRMgr.getFieldRegionWithSuper(FR, Result.second); 1616 1617 } else if (const CXXBaseObjectRegion *BaseReg = 1618 dyn_cast<CXXBaseObjectRegion>(R)) { 1619 // C++ base object region is another kind of region that we should blast 1620 // through to look for lazy compound value. It is like a field region. 1621 Result = findLazyBinding(B, cast<SubRegion>(BaseReg->getSuperRegion()), 1622 originalRegion); 1623 1624 if (Result.second) 1625 Result.second = MRMgr.getCXXBaseObjectRegionWithSuper(BaseReg, 1626 Result.second); 1627 } 1628 1629 return Result; 1630 } 1631 1632 SVal RegionStoreManager::getBindingForElement(RegionBindingsConstRef B, 1633 const ElementRegion* R) { 1634 // We do not currently model bindings of the CompoundLiteralregion. 1635 if (isa<CompoundLiteralRegion>(R->getBaseRegion())) 1636 return UnknownVal(); 1637 1638 // Check if the region has a binding. 1639 if (const Optional<SVal> &V = B.getDirectBinding(R)) 1640 return *V; 1641 1642 const MemRegion* superR = R->getSuperRegion(); 1643 1644 // Check if the region is an element region of a string literal. 1645 if (const StringRegion *StrR = dyn_cast<StringRegion>(superR)) { 1646 // FIXME: Handle loads from strings where the literal is treated as 1647 // an integer, e.g., *((unsigned int*)"hello") 1648 QualType T = Ctx.getAsArrayType(StrR->getValueType())->getElementType(); 1649 if (!Ctx.hasSameUnqualifiedType(T, R->getElementType())) 1650 return UnknownVal(); 1651 1652 const StringLiteral *Str = StrR->getStringLiteral(); 1653 SVal Idx = R->getIndex(); 1654 if (Optional<nonloc::ConcreteInt> CI = Idx.getAs<nonloc::ConcreteInt>()) { 1655 int64_t i = CI->getValue().getSExtValue(); 1656 // Abort on string underrun. This can be possible by arbitrary 1657 // clients of getBindingForElement(). 1658 if (i < 0) 1659 return UndefinedVal(); 1660 int64_t length = Str->getLength(); 1661 // Technically, only i == length is guaranteed to be null. 1662 // However, such overflows should be caught before reaching this point; 1663 // the only time such an access would be made is if a string literal was 1664 // used to initialize a larger array. 1665 char c = (i >= length) ? '\0' : Str->getCodeUnit(i); 1666 return svalBuilder.makeIntVal(c, T); 1667 } 1668 } else if (const VarRegion *VR = dyn_cast<VarRegion>(superR)) { 1669 // Check if the containing array is const and has an initialized value. 1670 const VarDecl *VD = VR->getDecl(); 1671 // Either the array or the array element has to be const. 1672 if (VD->getType().isConstQualified() || R->getElementType().isConstQualified()) { 1673 if (const Expr *Init = VD->getAnyInitializer()) { 1674 if (const auto *InitList = dyn_cast<InitListExpr>(Init)) { 1675 // The array index has to be known. 1676 if (auto CI = R->getIndex().getAs<nonloc::ConcreteInt>()) { 1677 int64_t i = CI->getValue().getSExtValue(); 1678 // If it is known that the index is out of bounds, we can return 1679 // an undefined value. 1680 if (i < 0) 1681 return UndefinedVal(); 1682 1683 if (auto CAT = Ctx.getAsConstantArrayType(VD->getType())) 1684 if (CAT->getSize().sle(i)) 1685 return UndefinedVal(); 1686 1687 // If there is a list, but no init, it must be zero. 1688 if (i >= InitList->getNumInits()) 1689 return svalBuilder.makeZeroVal(R->getElementType()); 1690 1691 if (const Expr *ElemInit = InitList->getInit(i)) 1692 if (Optional<SVal> V = svalBuilder.getConstantVal(ElemInit)) 1693 return *V; 1694 } 1695 } 1696 } 1697 } 1698 } 1699 1700 // Check for loads from a code text region. For such loads, just give up. 1701 if (isa<CodeTextRegion>(superR)) 1702 return UnknownVal(); 1703 1704 // Handle the case where we are indexing into a larger scalar object. 1705 // For example, this handles: 1706 // int x = ... 1707 // char *y = &x; 1708 // return *y; 1709 // FIXME: This is a hack, and doesn't do anything really intelligent yet. 1710 const RegionRawOffset &O = R->getAsArrayOffset(); 1711 1712 // If we cannot reason about the offset, return an unknown value. 1713 if (!O.getRegion()) 1714 return UnknownVal(); 1715 1716 if (const TypedValueRegion *baseR = 1717 dyn_cast_or_null<TypedValueRegion>(O.getRegion())) { 1718 QualType baseT = baseR->getValueType(); 1719 if (baseT->isScalarType()) { 1720 QualType elemT = R->getElementType(); 1721 if (elemT->isScalarType()) { 1722 if (Ctx.getTypeSizeInChars(baseT) >= Ctx.getTypeSizeInChars(elemT)) { 1723 if (const Optional<SVal> &V = B.getDirectBinding(superR)) { 1724 if (SymbolRef parentSym = V->getAsSymbol()) 1725 return svalBuilder.getDerivedRegionValueSymbolVal(parentSym, R); 1726 1727 if (V->isUnknownOrUndef()) 1728 return *V; 1729 // Other cases: give up. We are indexing into a larger object 1730 // that has some value, but we don't know how to handle that yet. 1731 return UnknownVal(); 1732 } 1733 } 1734 } 1735 } 1736 } 1737 return getBindingForFieldOrElementCommon(B, R, R->getElementType()); 1738 } 1739 1740 SVal RegionStoreManager::getBindingForField(RegionBindingsConstRef B, 1741 const FieldRegion* R) { 1742 1743 // Check if the region has a binding. 1744 if (const Optional<SVal> &V = B.getDirectBinding(R)) 1745 return *V; 1746 1747 // Is the field declared constant and has an in-class initializer? 1748 const FieldDecl *FD = R->getDecl(); 1749 QualType Ty = FD->getType(); 1750 if (Ty.isConstQualified()) 1751 if (const Expr *Init = FD->getInClassInitializer()) 1752 if (Optional<SVal> V = svalBuilder.getConstantVal(Init)) 1753 return *V; 1754 1755 // If the containing record was initialized, try to get its constant value. 1756 const MemRegion* superR = R->getSuperRegion(); 1757 if (const auto *VR = dyn_cast<VarRegion>(superR)) { 1758 const VarDecl *VD = VR->getDecl(); 1759 QualType RecordVarTy = VD->getType(); 1760 unsigned Index = FD->getFieldIndex(); 1761 // Either the record variable or the field has to be const qualified. 1762 if (RecordVarTy.isConstQualified() || Ty.isConstQualified()) 1763 if (const Expr *Init = VD->getAnyInitializer()) 1764 if (const auto *InitList = dyn_cast<InitListExpr>(Init)) { 1765 if (Index < InitList->getNumInits()) { 1766 if (const Expr *FieldInit = InitList->getInit(Index)) 1767 if (Optional<SVal> V = svalBuilder.getConstantVal(FieldInit)) 1768 return *V; 1769 } else { 1770 return svalBuilder.makeZeroVal(Ty); 1771 } 1772 } 1773 } 1774 1775 return getBindingForFieldOrElementCommon(B, R, Ty); 1776 } 1777 1778 Optional<SVal> 1779 RegionStoreManager::getBindingForDerivedDefaultValue(RegionBindingsConstRef B, 1780 const MemRegion *superR, 1781 const TypedValueRegion *R, 1782 QualType Ty) { 1783 1784 if (const Optional<SVal> &D = B.getDefaultBinding(superR)) { 1785 const SVal &val = D.getValue(); 1786 if (SymbolRef parentSym = val.getAsSymbol()) 1787 return svalBuilder.getDerivedRegionValueSymbolVal(parentSym, R); 1788 1789 if (val.isZeroConstant()) 1790 return svalBuilder.makeZeroVal(Ty); 1791 1792 if (val.isUnknownOrUndef()) 1793 return val; 1794 1795 // Lazy bindings are usually handled through getExistingLazyBinding(). 1796 // We should unify these two code paths at some point. 1797 if (val.getAs<nonloc::LazyCompoundVal>() || 1798 val.getAs<nonloc::CompoundVal>()) 1799 return val; 1800 1801 llvm_unreachable("Unknown default value"); 1802 } 1803 1804 return None; 1805 } 1806 1807 SVal RegionStoreManager::getLazyBinding(const SubRegion *LazyBindingRegion, 1808 RegionBindingsRef LazyBinding) { 1809 SVal Result; 1810 if (const ElementRegion *ER = dyn_cast<ElementRegion>(LazyBindingRegion)) 1811 Result = getBindingForElement(LazyBinding, ER); 1812 else 1813 Result = getBindingForField(LazyBinding, 1814 cast<FieldRegion>(LazyBindingRegion)); 1815 1816 // FIXME: This is a hack to deal with RegionStore's inability to distinguish a 1817 // default value for /part/ of an aggregate from a default value for the 1818 // /entire/ aggregate. The most common case of this is when struct Outer 1819 // has as its first member a struct Inner, which is copied in from a stack 1820 // variable. In this case, even if the Outer's default value is symbolic, 0, 1821 // or unknown, it gets overridden by the Inner's default value of undefined. 1822 // 1823 // This is a general problem -- if the Inner is zero-initialized, the Outer 1824 // will now look zero-initialized. The proper way to solve this is with a 1825 // new version of RegionStore that tracks the extent of a binding as well 1826 // as the offset. 1827 // 1828 // This hack only takes care of the undefined case because that can very 1829 // quickly result in a warning. 1830 if (Result.isUndef()) 1831 Result = UnknownVal(); 1832 1833 return Result; 1834 } 1835 1836 SVal 1837 RegionStoreManager::getBindingForFieldOrElementCommon(RegionBindingsConstRef B, 1838 const TypedValueRegion *R, 1839 QualType Ty) { 1840 1841 // At this point we have already checked in either getBindingForElement or 1842 // getBindingForField if 'R' has a direct binding. 1843 1844 // Lazy binding? 1845 Store lazyBindingStore = nullptr; 1846 const SubRegion *lazyBindingRegion = nullptr; 1847 std::tie(lazyBindingStore, lazyBindingRegion) = findLazyBinding(B, R, R); 1848 if (lazyBindingRegion) 1849 return getLazyBinding(lazyBindingRegion, 1850 getRegionBindings(lazyBindingStore)); 1851 1852 // Record whether or not we see a symbolic index. That can completely 1853 // be out of scope of our lookup. 1854 bool hasSymbolicIndex = false; 1855 1856 // FIXME: This is a hack to deal with RegionStore's inability to distinguish a 1857 // default value for /part/ of an aggregate from a default value for the 1858 // /entire/ aggregate. The most common case of this is when struct Outer 1859 // has as its first member a struct Inner, which is copied in from a stack 1860 // variable. In this case, even if the Outer's default value is symbolic, 0, 1861 // or unknown, it gets overridden by the Inner's default value of undefined. 1862 // 1863 // This is a general problem -- if the Inner is zero-initialized, the Outer 1864 // will now look zero-initialized. The proper way to solve this is with a 1865 // new version of RegionStore that tracks the extent of a binding as well 1866 // as the offset. 1867 // 1868 // This hack only takes care of the undefined case because that can very 1869 // quickly result in a warning. 1870 bool hasPartialLazyBinding = false; 1871 1872 const SubRegion *SR = R; 1873 while (SR) { 1874 const MemRegion *Base = SR->getSuperRegion(); 1875 if (Optional<SVal> D = getBindingForDerivedDefaultValue(B, Base, R, Ty)) { 1876 if (D->getAs<nonloc::LazyCompoundVal>()) { 1877 hasPartialLazyBinding = true; 1878 break; 1879 } 1880 1881 return *D; 1882 } 1883 1884 if (const ElementRegion *ER = dyn_cast<ElementRegion>(Base)) { 1885 NonLoc index = ER->getIndex(); 1886 if (!index.isConstant()) 1887 hasSymbolicIndex = true; 1888 } 1889 1890 // If our super region is a field or element itself, walk up the region 1891 // hierarchy to see if there is a default value installed in an ancestor. 1892 SR = dyn_cast<SubRegion>(Base); 1893 } 1894 1895 if (R->hasStackNonParametersStorage()) { 1896 if (isa<ElementRegion>(R)) { 1897 // Currently we don't reason specially about Clang-style vectors. Check 1898 // if superR is a vector and if so return Unknown. 1899 if (const TypedValueRegion *typedSuperR = 1900 dyn_cast<TypedValueRegion>(R->getSuperRegion())) { 1901 if (typedSuperR->getValueType()->isVectorType()) 1902 return UnknownVal(); 1903 } 1904 } 1905 1906 // FIXME: We also need to take ElementRegions with symbolic indexes into 1907 // account. This case handles both directly accessing an ElementRegion 1908 // with a symbolic offset, but also fields within an element with 1909 // a symbolic offset. 1910 if (hasSymbolicIndex) 1911 return UnknownVal(); 1912 1913 if (!hasPartialLazyBinding) 1914 return UndefinedVal(); 1915 } 1916 1917 // All other values are symbolic. 1918 return svalBuilder.getRegionValueSymbolVal(R); 1919 } 1920 1921 SVal RegionStoreManager::getBindingForObjCIvar(RegionBindingsConstRef B, 1922 const ObjCIvarRegion* R) { 1923 // Check if the region has a binding. 1924 if (const Optional<SVal> &V = B.getDirectBinding(R)) 1925 return *V; 1926 1927 const MemRegion *superR = R->getSuperRegion(); 1928 1929 // Check if the super region has a default binding. 1930 if (const Optional<SVal> &V = B.getDefaultBinding(superR)) { 1931 if (SymbolRef parentSym = V->getAsSymbol()) 1932 return svalBuilder.getDerivedRegionValueSymbolVal(parentSym, R); 1933 1934 // Other cases: give up. 1935 return UnknownVal(); 1936 } 1937 1938 return getBindingForLazySymbol(R); 1939 } 1940 1941 SVal RegionStoreManager::getBindingForVar(RegionBindingsConstRef B, 1942 const VarRegion *R) { 1943 1944 // Check if the region has a binding. 1945 if (Optional<SVal> V = B.getDirectBinding(R)) 1946 return *V; 1947 1948 if (Optional<SVal> V = B.getDefaultBinding(R)) 1949 return *V; 1950 1951 // Lazily derive a value for the VarRegion. 1952 const VarDecl *VD = R->getDecl(); 1953 const MemSpaceRegion *MS = R->getMemorySpace(); 1954 1955 // Arguments are always symbolic. 1956 if (isa<StackArgumentsSpaceRegion>(MS)) 1957 return svalBuilder.getRegionValueSymbolVal(R); 1958 1959 // Is 'VD' declared constant? If so, retrieve the constant value. 1960 if (VD->getType().isConstQualified()) { 1961 if (const Expr *Init = VD->getAnyInitializer()) { 1962 if (Optional<SVal> V = svalBuilder.getConstantVal(Init)) 1963 return *V; 1964 1965 // If the variable is const qualified and has an initializer but 1966 // we couldn't evaluate initializer to a value, treat the value as 1967 // unknown. 1968 return UnknownVal(); 1969 } 1970 } 1971 1972 // This must come after the check for constants because closure-captured 1973 // constant variables may appear in UnknownSpaceRegion. 1974 if (isa<UnknownSpaceRegion>(MS)) 1975 return svalBuilder.getRegionValueSymbolVal(R); 1976 1977 if (isa<GlobalsSpaceRegion>(MS)) { 1978 QualType T = VD->getType(); 1979 1980 // Function-scoped static variables are default-initialized to 0; if they 1981 // have an initializer, it would have been processed by now. 1982 // FIXME: This is only true when we're starting analysis from main(). 1983 // We're losing a lot of coverage here. 1984 if (isa<StaticGlobalSpaceRegion>(MS)) 1985 return svalBuilder.makeZeroVal(T); 1986 1987 if (Optional<SVal> V = getBindingForDerivedDefaultValue(B, MS, R, T)) { 1988 assert(!V->getAs<nonloc::LazyCompoundVal>()); 1989 return V.getValue(); 1990 } 1991 1992 return svalBuilder.getRegionValueSymbolVal(R); 1993 } 1994 1995 return UndefinedVal(); 1996 } 1997 1998 SVal RegionStoreManager::getBindingForLazySymbol(const TypedValueRegion *R) { 1999 // All other values are symbolic. 2000 return svalBuilder.getRegionValueSymbolVal(R); 2001 } 2002 2003 const RegionStoreManager::SValListTy & 2004 RegionStoreManager::getInterestingValues(nonloc::LazyCompoundVal LCV) { 2005 // First, check the cache. 2006 LazyBindingsMapTy::iterator I = LazyBindingsMap.find(LCV.getCVData()); 2007 if (I != LazyBindingsMap.end()) 2008 return I->second; 2009 2010 // If we don't have a list of values cached, start constructing it. 2011 SValListTy List; 2012 2013 const SubRegion *LazyR = LCV.getRegion(); 2014 RegionBindingsRef B = getRegionBindings(LCV.getStore()); 2015 2016 // If this region had /no/ bindings at the time, there are no interesting 2017 // values to return. 2018 const ClusterBindings *Cluster = B.lookup(LazyR->getBaseRegion()); 2019 if (!Cluster) 2020 return (LazyBindingsMap[LCV.getCVData()] = std::move(List)); 2021 2022 SmallVector<BindingPair, 32> Bindings; 2023 collectSubRegionBindings(Bindings, svalBuilder, *Cluster, LazyR, 2024 /*IncludeAllDefaultBindings=*/true); 2025 for (SmallVectorImpl<BindingPair>::const_iterator I = Bindings.begin(), 2026 E = Bindings.end(); 2027 I != E; ++I) { 2028 SVal V = I->second; 2029 if (V.isUnknownOrUndef() || V.isConstant()) 2030 continue; 2031 2032 if (Optional<nonloc::LazyCompoundVal> InnerLCV = 2033 V.getAs<nonloc::LazyCompoundVal>()) { 2034 const SValListTy &InnerList = getInterestingValues(*InnerLCV); 2035 List.insert(List.end(), InnerList.begin(), InnerList.end()); 2036 continue; 2037 } 2038 2039 List.push_back(V); 2040 } 2041 2042 return (LazyBindingsMap[LCV.getCVData()] = std::move(List)); 2043 } 2044 2045 NonLoc RegionStoreManager::createLazyBinding(RegionBindingsConstRef B, 2046 const TypedValueRegion *R) { 2047 if (Optional<nonloc::LazyCompoundVal> V = 2048 getExistingLazyBinding(svalBuilder, B, R, false)) 2049 return *V; 2050 2051 return svalBuilder.makeLazyCompoundVal(StoreRef(B.asStore(), *this), R); 2052 } 2053 2054 static bool isRecordEmpty(const RecordDecl *RD) { 2055 if (!RD->field_empty()) 2056 return false; 2057 if (const CXXRecordDecl *CRD = dyn_cast<CXXRecordDecl>(RD)) 2058 return CRD->getNumBases() == 0; 2059 return true; 2060 } 2061 2062 SVal RegionStoreManager::getBindingForStruct(RegionBindingsConstRef B, 2063 const TypedValueRegion *R) { 2064 const RecordDecl *RD = R->getValueType()->castAs<RecordType>()->getDecl(); 2065 if (!RD->getDefinition() || isRecordEmpty(RD)) 2066 return UnknownVal(); 2067 2068 return createLazyBinding(B, R); 2069 } 2070 2071 SVal RegionStoreManager::getBindingForArray(RegionBindingsConstRef B, 2072 const TypedValueRegion *R) { 2073 assert(Ctx.getAsConstantArrayType(R->getValueType()) && 2074 "Only constant array types can have compound bindings."); 2075 2076 return createLazyBinding(B, R); 2077 } 2078 2079 bool RegionStoreManager::includedInBindings(Store store, 2080 const MemRegion *region) const { 2081 RegionBindingsRef B = getRegionBindings(store); 2082 region = region->getBaseRegion(); 2083 2084 // Quick path: if the base is the head of a cluster, the region is live. 2085 if (B.lookup(region)) 2086 return true; 2087 2088 // Slow path: if the region is the VALUE of any binding, it is live. 2089 for (RegionBindingsRef::iterator RI = B.begin(), RE = B.end(); RI != RE; ++RI) { 2090 const ClusterBindings &Cluster = RI.getData(); 2091 for (ClusterBindings::iterator CI = Cluster.begin(), CE = Cluster.end(); 2092 CI != CE; ++CI) { 2093 const SVal &D = CI.getData(); 2094 if (const MemRegion *R = D.getAsRegion()) 2095 if (R->getBaseRegion() == region) 2096 return true; 2097 } 2098 } 2099 2100 return false; 2101 } 2102 2103 //===----------------------------------------------------------------------===// 2104 // Binding values to regions. 2105 //===----------------------------------------------------------------------===// 2106 2107 StoreRef RegionStoreManager::killBinding(Store ST, Loc L) { 2108 if (Optional<loc::MemRegionVal> LV = L.getAs<loc::MemRegionVal>()) 2109 if (const MemRegion* R = LV->getRegion()) 2110 return StoreRef(getRegionBindings(ST).removeBinding(R) 2111 .asImmutableMap() 2112 .getRootWithoutRetain(), 2113 *this); 2114 2115 return StoreRef(ST, *this); 2116 } 2117 2118 RegionBindingsRef 2119 RegionStoreManager::bind(RegionBindingsConstRef B, Loc L, SVal V) { 2120 if (L.getAs<loc::ConcreteInt>()) 2121 return B; 2122 2123 // If we get here, the location should be a region. 2124 const MemRegion *R = L.castAs<loc::MemRegionVal>().getRegion(); 2125 2126 // Check if the region is a struct region. 2127 if (const TypedValueRegion* TR = dyn_cast<TypedValueRegion>(R)) { 2128 QualType Ty = TR->getValueType(); 2129 if (Ty->isArrayType()) 2130 return bindArray(B, TR, V); 2131 if (Ty->isStructureOrClassType()) 2132 return bindStruct(B, TR, V); 2133 if (Ty->isVectorType()) 2134 return bindVector(B, TR, V); 2135 if (Ty->isUnionType()) 2136 return bindAggregate(B, TR, V); 2137 } 2138 2139 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(R)) { 2140 // Binding directly to a symbolic region should be treated as binding 2141 // to element 0. 2142 QualType T = SR->getSymbol()->getType(); 2143 if (T->isAnyPointerType() || T->isReferenceType()) 2144 T = T->getPointeeType(); 2145 2146 R = GetElementZeroRegion(SR, T); 2147 } 2148 2149 assert((!isa<CXXThisRegion>(R) || !B.lookup(R)) && 2150 "'this' pointer is not an l-value and is not assignable"); 2151 2152 // Clear out bindings that may overlap with this binding. 2153 RegionBindingsRef NewB = removeSubRegionBindings(B, cast<SubRegion>(R)); 2154 return NewB.addBinding(BindingKey::Make(R, BindingKey::Direct), V); 2155 } 2156 2157 RegionBindingsRef 2158 RegionStoreManager::setImplicitDefaultValue(RegionBindingsConstRef B, 2159 const MemRegion *R, 2160 QualType T) { 2161 SVal V; 2162 2163 if (Loc::isLocType(T)) 2164 V = svalBuilder.makeNull(); 2165 else if (T->isIntegralOrEnumerationType()) 2166 V = svalBuilder.makeZeroVal(T); 2167 else if (T->isStructureOrClassType() || T->isArrayType()) { 2168 // Set the default value to a zero constant when it is a structure 2169 // or array. The type doesn't really matter. 2170 V = svalBuilder.makeZeroVal(Ctx.IntTy); 2171 } 2172 else { 2173 // We can't represent values of this type, but we still need to set a value 2174 // to record that the region has been initialized. 2175 // If this assertion ever fires, a new case should be added above -- we 2176 // should know how to default-initialize any value we can symbolicate. 2177 assert(!SymbolManager::canSymbolicate(T) && "This type is representable"); 2178 V = UnknownVal(); 2179 } 2180 2181 return B.addBinding(R, BindingKey::Default, V); 2182 } 2183 2184 RegionBindingsRef 2185 RegionStoreManager::bindArray(RegionBindingsConstRef B, 2186 const TypedValueRegion* R, 2187 SVal Init) { 2188 2189 const ArrayType *AT =cast<ArrayType>(Ctx.getCanonicalType(R->getValueType())); 2190 QualType ElementTy = AT->getElementType(); 2191 Optional<uint64_t> Size; 2192 2193 if (const ConstantArrayType* CAT = dyn_cast<ConstantArrayType>(AT)) 2194 Size = CAT->getSize().getZExtValue(); 2195 2196 // Check if the init expr is a literal. If so, bind the rvalue instead. 2197 // FIXME: It's not responsibility of the Store to transform this lvalue 2198 // to rvalue. ExprEngine or maybe even CFG should do this before binding. 2199 if (Optional<loc::MemRegionVal> MRV = Init.getAs<loc::MemRegionVal>()) { 2200 SVal V = getBinding(B.asStore(), *MRV, R->getValueType()); 2201 return bindAggregate(B, R, V); 2202 } 2203 2204 // Handle lazy compound values. 2205 if (Init.getAs<nonloc::LazyCompoundVal>()) 2206 return bindAggregate(B, R, Init); 2207 2208 if (Init.isUnknown()) 2209 return bindAggregate(B, R, UnknownVal()); 2210 2211 // Remaining case: explicit compound values. 2212 const nonloc::CompoundVal& CV = Init.castAs<nonloc::CompoundVal>(); 2213 nonloc::CompoundVal::iterator VI = CV.begin(), VE = CV.end(); 2214 uint64_t i = 0; 2215 2216 RegionBindingsRef NewB(B); 2217 2218 for (; Size.hasValue() ? i < Size.getValue() : true ; ++i, ++VI) { 2219 // The init list might be shorter than the array length. 2220 if (VI == VE) 2221 break; 2222 2223 const NonLoc &Idx = svalBuilder.makeArrayIndex(i); 2224 const ElementRegion *ER = MRMgr.getElementRegion(ElementTy, Idx, R, Ctx); 2225 2226 if (ElementTy->isStructureOrClassType()) 2227 NewB = bindStruct(NewB, ER, *VI); 2228 else if (ElementTy->isArrayType()) 2229 NewB = bindArray(NewB, ER, *VI); 2230 else 2231 NewB = bind(NewB, loc::MemRegionVal(ER), *VI); 2232 } 2233 2234 // If the init list is shorter than the array length (or the array has 2235 // variable length), set the array default value. Values that are already set 2236 // are not overwritten. 2237 if (!Size.hasValue() || i < Size.getValue()) 2238 NewB = setImplicitDefaultValue(NewB, R, ElementTy); 2239 2240 return NewB; 2241 } 2242 2243 RegionBindingsRef RegionStoreManager::bindVector(RegionBindingsConstRef B, 2244 const TypedValueRegion* R, 2245 SVal V) { 2246 QualType T = R->getValueType(); 2247 assert(T->isVectorType()); 2248 const VectorType *VT = T->getAs<VectorType>(); // Use getAs for typedefs. 2249 2250 // Handle lazy compound values and symbolic values. 2251 if (V.getAs<nonloc::LazyCompoundVal>() || V.getAs<nonloc::SymbolVal>()) 2252 return bindAggregate(B, R, V); 2253 2254 // We may get non-CompoundVal accidentally due to imprecise cast logic or 2255 // that we are binding symbolic struct value. Kill the field values, and if 2256 // the value is symbolic go and bind it as a "default" binding. 2257 if (!V.getAs<nonloc::CompoundVal>()) { 2258 return bindAggregate(B, R, UnknownVal()); 2259 } 2260 2261 QualType ElemType = VT->getElementType(); 2262 nonloc::CompoundVal CV = V.castAs<nonloc::CompoundVal>(); 2263 nonloc::CompoundVal::iterator VI = CV.begin(), VE = CV.end(); 2264 unsigned index = 0, numElements = VT->getNumElements(); 2265 RegionBindingsRef NewB(B); 2266 2267 for ( ; index != numElements ; ++index) { 2268 if (VI == VE) 2269 break; 2270 2271 NonLoc Idx = svalBuilder.makeArrayIndex(index); 2272 const ElementRegion *ER = MRMgr.getElementRegion(ElemType, Idx, R, Ctx); 2273 2274 if (ElemType->isArrayType()) 2275 NewB = bindArray(NewB, ER, *VI); 2276 else if (ElemType->isStructureOrClassType()) 2277 NewB = bindStruct(NewB, ER, *VI); 2278 else 2279 NewB = bind(NewB, loc::MemRegionVal(ER), *VI); 2280 } 2281 return NewB; 2282 } 2283 2284 Optional<RegionBindingsRef> 2285 RegionStoreManager::tryBindSmallStruct(RegionBindingsConstRef B, 2286 const TypedValueRegion *R, 2287 const RecordDecl *RD, 2288 nonloc::LazyCompoundVal LCV) { 2289 FieldVector Fields; 2290 2291 if (const CXXRecordDecl *Class = dyn_cast<CXXRecordDecl>(RD)) 2292 if (Class->getNumBases() != 0 || Class->getNumVBases() != 0) 2293 return None; 2294 2295 for (const auto *FD : RD->fields()) { 2296 if (FD->isUnnamedBitfield()) 2297 continue; 2298 2299 // If there are too many fields, or if any of the fields are aggregates, 2300 // just use the LCV as a default binding. 2301 if (Fields.size() == SmallStructLimit) 2302 return None; 2303 2304 QualType Ty = FD->getType(); 2305 if (!(Ty->isScalarType() || Ty->isReferenceType())) 2306 return None; 2307 2308 Fields.push_back(FD); 2309 } 2310 2311 RegionBindingsRef NewB = B; 2312 2313 for (FieldVector::iterator I = Fields.begin(), E = Fields.end(); I != E; ++I){ 2314 const FieldRegion *SourceFR = MRMgr.getFieldRegion(*I, LCV.getRegion()); 2315 SVal V = getBindingForField(getRegionBindings(LCV.getStore()), SourceFR); 2316 2317 const FieldRegion *DestFR = MRMgr.getFieldRegion(*I, R); 2318 NewB = bind(NewB, loc::MemRegionVal(DestFR), V); 2319 } 2320 2321 return NewB; 2322 } 2323 2324 RegionBindingsRef RegionStoreManager::bindStruct(RegionBindingsConstRef B, 2325 const TypedValueRegion* R, 2326 SVal V) { 2327 if (!Features.supportsFields()) 2328 return B; 2329 2330 QualType T = R->getValueType(); 2331 assert(T->isStructureOrClassType()); 2332 2333 const RecordType* RT = T->getAs<RecordType>(); 2334 const RecordDecl *RD = RT->getDecl(); 2335 2336 if (!RD->isCompleteDefinition()) 2337 return B; 2338 2339 // Handle lazy compound values and symbolic values. 2340 if (Optional<nonloc::LazyCompoundVal> LCV = 2341 V.getAs<nonloc::LazyCompoundVal>()) { 2342 if (Optional<RegionBindingsRef> NewB = tryBindSmallStruct(B, R, RD, *LCV)) 2343 return *NewB; 2344 return bindAggregate(B, R, V); 2345 } 2346 if (V.getAs<nonloc::SymbolVal>()) 2347 return bindAggregate(B, R, V); 2348 2349 // We may get non-CompoundVal accidentally due to imprecise cast logic or 2350 // that we are binding symbolic struct value. Kill the field values, and if 2351 // the value is symbolic go and bind it as a "default" binding. 2352 if (V.isUnknown() || !V.getAs<nonloc::CompoundVal>()) 2353 return bindAggregate(B, R, UnknownVal()); 2354 2355 // The raw CompoundVal is essentially a symbolic InitListExpr: an (immutable) 2356 // list of other values. It appears pretty much only when there's an actual 2357 // initializer list expression in the program, and the analyzer tries to 2358 // unwrap it as soon as possible. 2359 // This code is where such unwrap happens: when the compound value is put into 2360 // the object that it was supposed to initialize (it's an *initializer* list, 2361 // after all), instead of binding the whole value to the whole object, we bind 2362 // sub-values to sub-objects. Sub-values may themselves be compound values, 2363 // and in this case the procedure becomes recursive. 2364 // FIXME: The annoying part about compound values is that they don't carry 2365 // any sort of information about which value corresponds to which sub-object. 2366 // It's simply a list of values in the middle of nowhere; we expect to match 2367 // them to sub-objects, essentially, "by index": first value binds to 2368 // the first field, second value binds to the second field, etc. 2369 // It would have been much safer to organize non-lazy compound values as 2370 // a mapping from fields/bases to values. 2371 const nonloc::CompoundVal& CV = V.castAs<nonloc::CompoundVal>(); 2372 nonloc::CompoundVal::iterator VI = CV.begin(), VE = CV.end(); 2373 2374 RegionBindingsRef NewB(B); 2375 2376 // In C++17 aggregates may have base classes, handle those as well. 2377 // They appear before fields in the initializer list / compound value. 2378 if (const auto *CRD = dyn_cast<CXXRecordDecl>(RD)) { 2379 // If the object was constructed with a constructor, its value is a 2380 // LazyCompoundVal. If it's a raw CompoundVal, it means that we're 2381 // performing aggregate initialization. The only exception from this 2382 // rule is sending an Objective-C++ message that returns a C++ object 2383 // to a nil receiver; in this case the semantics is to return a 2384 // zero-initialized object even if it's a C++ object that doesn't have 2385 // this sort of constructor; the CompoundVal is empty in this case. 2386 assert((CRD->isAggregate() || (Ctx.getLangOpts().ObjC && VI == VE)) && 2387 "Non-aggregates are constructed with a constructor!"); 2388 2389 for (const auto &B : CRD->bases()) { 2390 // (Multiple inheritance is fine though.) 2391 assert(!B.isVirtual() && "Aggregates cannot have virtual base classes!"); 2392 2393 if (VI == VE) 2394 break; 2395 2396 QualType BTy = B.getType(); 2397 assert(BTy->isStructureOrClassType() && "Base classes must be classes!"); 2398 2399 const CXXRecordDecl *BRD = BTy->getAsCXXRecordDecl(); 2400 assert(BRD && "Base classes must be C++ classes!"); 2401 2402 const CXXBaseObjectRegion *BR = 2403 MRMgr.getCXXBaseObjectRegion(BRD, R, /*IsVirtual=*/false); 2404 2405 NewB = bindStruct(NewB, BR, *VI); 2406 2407 ++VI; 2408 } 2409 } 2410 2411 RecordDecl::field_iterator FI, FE; 2412 2413 for (FI = RD->field_begin(), FE = RD->field_end(); FI != FE; ++FI) { 2414 2415 if (VI == VE) 2416 break; 2417 2418 // Skip any unnamed bitfields to stay in sync with the initializers. 2419 if (FI->isUnnamedBitfield()) 2420 continue; 2421 2422 QualType FTy = FI->getType(); 2423 const FieldRegion* FR = MRMgr.getFieldRegion(*FI, R); 2424 2425 if (FTy->isArrayType()) 2426 NewB = bindArray(NewB, FR, *VI); 2427 else if (FTy->isStructureOrClassType()) 2428 NewB = bindStruct(NewB, FR, *VI); 2429 else 2430 NewB = bind(NewB, loc::MemRegionVal(FR), *VI); 2431 ++VI; 2432 } 2433 2434 // There may be fewer values in the initialize list than the fields of struct. 2435 if (FI != FE) { 2436 NewB = NewB.addBinding(R, BindingKey::Default, 2437 svalBuilder.makeIntVal(0, false)); 2438 } 2439 2440 return NewB; 2441 } 2442 2443 RegionBindingsRef 2444 RegionStoreManager::bindAggregate(RegionBindingsConstRef B, 2445 const TypedRegion *R, 2446 SVal Val) { 2447 // Remove the old bindings, using 'R' as the root of all regions 2448 // we will invalidate. Then add the new binding. 2449 return removeSubRegionBindings(B, R).addBinding(R, BindingKey::Default, Val); 2450 } 2451 2452 //===----------------------------------------------------------------------===// 2453 // State pruning. 2454 //===----------------------------------------------------------------------===// 2455 2456 namespace { 2457 class RemoveDeadBindingsWorker 2458 : public ClusterAnalysis<RemoveDeadBindingsWorker> { 2459 SmallVector<const SymbolicRegion *, 12> Postponed; 2460 SymbolReaper &SymReaper; 2461 const StackFrameContext *CurrentLCtx; 2462 2463 public: 2464 RemoveDeadBindingsWorker(RegionStoreManager &rm, 2465 ProgramStateManager &stateMgr, 2466 RegionBindingsRef b, SymbolReaper &symReaper, 2467 const StackFrameContext *LCtx) 2468 : ClusterAnalysis<RemoveDeadBindingsWorker>(rm, stateMgr, b), 2469 SymReaper(symReaper), CurrentLCtx(LCtx) {} 2470 2471 // Called by ClusterAnalysis. 2472 void VisitAddedToCluster(const MemRegion *baseR, const ClusterBindings &C); 2473 void VisitCluster(const MemRegion *baseR, const ClusterBindings *C); 2474 using ClusterAnalysis<RemoveDeadBindingsWorker>::VisitCluster; 2475 2476 using ClusterAnalysis::AddToWorkList; 2477 2478 bool AddToWorkList(const MemRegion *R); 2479 2480 bool UpdatePostponed(); 2481 void VisitBinding(SVal V); 2482 }; 2483 } 2484 2485 bool RemoveDeadBindingsWorker::AddToWorkList(const MemRegion *R) { 2486 const MemRegion *BaseR = R->getBaseRegion(); 2487 return AddToWorkList(WorkListElement(BaseR), getCluster(BaseR)); 2488 } 2489 2490 void RemoveDeadBindingsWorker::VisitAddedToCluster(const MemRegion *baseR, 2491 const ClusterBindings &C) { 2492 2493 if (const VarRegion *VR = dyn_cast<VarRegion>(baseR)) { 2494 if (SymReaper.isLive(VR)) 2495 AddToWorkList(baseR, &C); 2496 2497 return; 2498 } 2499 2500 if (const SymbolicRegion *SR = dyn_cast<SymbolicRegion>(baseR)) { 2501 if (SymReaper.isLive(SR->getSymbol())) 2502 AddToWorkList(SR, &C); 2503 else 2504 Postponed.push_back(SR); 2505 2506 return; 2507 } 2508 2509 if (isa<NonStaticGlobalSpaceRegion>(baseR)) { 2510 AddToWorkList(baseR, &C); 2511 return; 2512 } 2513 2514 // CXXThisRegion in the current or parent location context is live. 2515 if (const CXXThisRegion *TR = dyn_cast<CXXThisRegion>(baseR)) { 2516 const auto *StackReg = 2517 cast<StackArgumentsSpaceRegion>(TR->getSuperRegion()); 2518 const StackFrameContext *RegCtx = StackReg->getStackFrame(); 2519 if (CurrentLCtx && 2520 (RegCtx == CurrentLCtx || RegCtx->isParentOf(CurrentLCtx))) 2521 AddToWorkList(TR, &C); 2522 } 2523 } 2524 2525 void RemoveDeadBindingsWorker::VisitCluster(const MemRegion *baseR, 2526 const ClusterBindings *C) { 2527 if (!C) 2528 return; 2529 2530 // Mark the symbol for any SymbolicRegion with live bindings as live itself. 2531 // This means we should continue to track that symbol. 2532 if (const SymbolicRegion *SymR = dyn_cast<SymbolicRegion>(baseR)) 2533 SymReaper.markLive(SymR->getSymbol()); 2534 2535 for (ClusterBindings::iterator I = C->begin(), E = C->end(); I != E; ++I) { 2536 // Element index of a binding key is live. 2537 SymReaper.markElementIndicesLive(I.getKey().getRegion()); 2538 2539 VisitBinding(I.getData()); 2540 } 2541 } 2542 2543 void RemoveDeadBindingsWorker::VisitBinding(SVal V) { 2544 // Is it a LazyCompoundVal? All referenced regions are live as well. 2545 if (Optional<nonloc::LazyCompoundVal> LCS = 2546 V.getAs<nonloc::LazyCompoundVal>()) { 2547 2548 const RegionStoreManager::SValListTy &Vals = RM.getInterestingValues(*LCS); 2549 2550 for (RegionStoreManager::SValListTy::const_iterator I = Vals.begin(), 2551 E = Vals.end(); 2552 I != E; ++I) 2553 VisitBinding(*I); 2554 2555 return; 2556 } 2557 2558 // If V is a region, then add it to the worklist. 2559 if (const MemRegion *R = V.getAsRegion()) { 2560 AddToWorkList(R); 2561 SymReaper.markLive(R); 2562 2563 // All regions captured by a block are also live. 2564 if (const BlockDataRegion *BR = dyn_cast<BlockDataRegion>(R)) { 2565 BlockDataRegion::referenced_vars_iterator I = BR->referenced_vars_begin(), 2566 E = BR->referenced_vars_end(); 2567 for ( ; I != E; ++I) 2568 AddToWorkList(I.getCapturedRegion()); 2569 } 2570 } 2571 2572 2573 // Update the set of live symbols. 2574 for (auto SI = V.symbol_begin(), SE = V.symbol_end(); SI!=SE; ++SI) 2575 SymReaper.markLive(*SI); 2576 } 2577 2578 bool RemoveDeadBindingsWorker::UpdatePostponed() { 2579 // See if any postponed SymbolicRegions are actually live now, after 2580 // having done a scan. 2581 bool Changed = false; 2582 2583 for (auto I = Postponed.begin(), E = Postponed.end(); I != E; ++I) { 2584 if (const SymbolicRegion *SR = *I) { 2585 if (SymReaper.isLive(SR->getSymbol())) { 2586 Changed |= AddToWorkList(SR); 2587 *I = nullptr; 2588 } 2589 } 2590 } 2591 2592 return Changed; 2593 } 2594 2595 StoreRef RegionStoreManager::removeDeadBindings(Store store, 2596 const StackFrameContext *LCtx, 2597 SymbolReaper& SymReaper) { 2598 RegionBindingsRef B = getRegionBindings(store); 2599 RemoveDeadBindingsWorker W(*this, StateMgr, B, SymReaper, LCtx); 2600 W.GenerateClusters(); 2601 2602 // Enqueue the region roots onto the worklist. 2603 for (SymbolReaper::region_iterator I = SymReaper.region_begin(), 2604 E = SymReaper.region_end(); I != E; ++I) { 2605 W.AddToWorkList(*I); 2606 } 2607 2608 do W.RunWorkList(); while (W.UpdatePostponed()); 2609 2610 // We have now scanned the store, marking reachable regions and symbols 2611 // as live. We now remove all the regions that are dead from the store 2612 // as well as update DSymbols with the set symbols that are now dead. 2613 for (RegionBindingsRef::iterator I = B.begin(), E = B.end(); I != E; ++I) { 2614 const MemRegion *Base = I.getKey(); 2615 2616 // If the cluster has been visited, we know the region has been marked. 2617 // Otherwise, remove the dead entry. 2618 if (!W.isVisited(Base)) 2619 B = B.remove(Base); 2620 } 2621 2622 return StoreRef(B.asStore(), *this); 2623 } 2624 2625 //===----------------------------------------------------------------------===// 2626 // Utility methods. 2627 //===----------------------------------------------------------------------===// 2628 2629 void RegionStoreManager::printJson(raw_ostream &Out, Store S, const char *NL, 2630 unsigned int Space, bool IsDot) const { 2631 RegionBindingsRef Bindings = getRegionBindings(S); 2632 2633 Indent(Out, Space, IsDot) << "\"store\": "; 2634 2635 if (Bindings.isEmpty()) { 2636 Out << "null," << NL; 2637 return; 2638 } 2639 2640 Out << '[' << NL; 2641 Bindings.printJson(Out, NL, ++Space, IsDot); 2642 Indent(Out, --Space, IsDot) << "]," << NL; 2643 } 2644