1 | //===-- DataflowEnvironment.cpp ---------------------------------*- 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 an Environment class that is used by dataflow analyses |
10 | // that run over Control-Flow Graphs (CFGs) to keep track of the state of the |
11 | // program at given program points. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #include "clang/Analysis/FlowSensitive/DataflowEnvironment.h" |
16 | #include "clang/AST/Decl.h" |
17 | #include "clang/AST/DeclCXX.h" |
18 | #include "clang/AST/RecursiveASTVisitor.h" |
19 | #include "clang/AST/Type.h" |
20 | #include "clang/Analysis/FlowSensitive/ASTOps.h" |
21 | #include "clang/Analysis/FlowSensitive/DataflowLattice.h" |
22 | #include "clang/Analysis/FlowSensitive/Value.h" |
23 | #include "llvm/ADT/DenseMap.h" |
24 | #include "llvm/ADT/DenseSet.h" |
25 | #include "llvm/ADT/MapVector.h" |
26 | #include "llvm/ADT/STLExtras.h" |
27 | #include "llvm/ADT/ScopeExit.h" |
28 | #include "llvm/Support/ErrorHandling.h" |
29 | #include <cassert> |
30 | #include <utility> |
31 | |
32 | #define DEBUG_TYPE "dataflow" |
33 | |
34 | namespace clang { |
35 | namespace dataflow { |
36 | |
37 | // FIXME: convert these to parameters of the analysis or environment. Current |
38 | // settings have been experimentaly validated, but only for a particular |
39 | // analysis. |
40 | static constexpr int MaxCompositeValueDepth = 3; |
41 | static constexpr int MaxCompositeValueSize = 1000; |
42 | |
43 | /// Returns a map consisting of key-value entries that are present in both maps. |
44 | static llvm::DenseMap<const ValueDecl *, StorageLocation *> intersectDeclToLoc( |
45 | const llvm::DenseMap<const ValueDecl *, StorageLocation *> &DeclToLoc1, |
46 | const llvm::DenseMap<const ValueDecl *, StorageLocation *> &DeclToLoc2) { |
47 | llvm::DenseMap<const ValueDecl *, StorageLocation *> Result; |
48 | for (auto &Entry : DeclToLoc1) { |
49 | auto It = DeclToLoc2.find(Val: Entry.first); |
50 | if (It != DeclToLoc2.end() && Entry.second == It->second) |
51 | Result.insert(KV: {Entry.first, Entry.second}); |
52 | } |
53 | return Result; |
54 | } |
55 | |
56 | // Performs a join on either `ExprToLoc` or `ExprToVal`. |
57 | // The maps must be consistent in the sense that any entries for the same |
58 | // expression must map to the same location / value. This is the case if we are |
59 | // performing a join for control flow within a full-expression (which is the |
60 | // only case when this function should be used). |
61 | template <typename MapT> MapT joinExprMaps(const MapT &Map1, const MapT &Map2) { |
62 | MapT Result = Map1; |
63 | |
64 | for (const auto &Entry : Map2) { |
65 | [[maybe_unused]] auto [It, Inserted] = Result.insert(Entry); |
66 | // If there was an existing entry, its value should be the same as for the |
67 | // entry we were trying to insert. |
68 | assert(It->second == Entry.second); |
69 | } |
70 | |
71 | return Result; |
72 | } |
73 | |
74 | // Whether to consider equivalent two values with an unknown relation. |
75 | // |
76 | // FIXME: this function is a hack enabling unsoundness to support |
77 | // convergence. Once we have widening support for the reference/pointer and |
78 | // struct built-in models, this should be unconditionally `false` (and inlined |
79 | // as such at its call sites). |
80 | static bool equateUnknownValues(Value::Kind K) { |
81 | switch (K) { |
82 | case Value::Kind::Integer: |
83 | case Value::Kind::Pointer: |
84 | return true; |
85 | default: |
86 | return false; |
87 | } |
88 | } |
89 | |
90 | static bool compareDistinctValues(QualType Type, Value &Val1, |
91 | const Environment &Env1, Value &Val2, |
92 | const Environment &Env2, |
93 | Environment::ValueModel &Model) { |
94 | // Note: Potentially costly, but, for booleans, we could check whether both |
95 | // can be proven equivalent in their respective environments. |
96 | |
97 | // FIXME: move the reference/pointers logic from `areEquivalentValues` to here |
98 | // and implement separate, join/widen specific handling for |
99 | // reference/pointers. |
100 | switch (Model.compare(Type, Val1, Env1, Val2, Env2)) { |
101 | case ComparisonResult::Same: |
102 | return true; |
103 | case ComparisonResult::Different: |
104 | return false; |
105 | case ComparisonResult::Unknown: |
106 | return equateUnknownValues(K: Val1.getKind()); |
107 | } |
108 | llvm_unreachable("All cases covered in switch" ); |
109 | } |
110 | |
111 | /// Attempts to join distinct values `Val1` and `Val2` in `Env1` and `Env2`, |
112 | /// respectively, of the same type `Type`. Joining generally produces a single |
113 | /// value that (soundly) approximates the two inputs, although the actual |
114 | /// meaning depends on `Model`. |
115 | static Value *joinDistinctValues(QualType Type, Value &Val1, |
116 | const Environment &Env1, Value &Val2, |
117 | const Environment &Env2, |
118 | Environment &JoinedEnv, |
119 | Environment::ValueModel &Model) { |
120 | // Join distinct boolean values preserving information about the constraints |
121 | // in the respective path conditions. |
122 | if (isa<BoolValue>(Val: &Val1) && isa<BoolValue>(Val: &Val2)) { |
123 | // FIXME: Checking both values should be unnecessary, since they should have |
124 | // a consistent shape. However, right now we can end up with BoolValue's in |
125 | // integer-typed variables due to our incorrect handling of |
126 | // boolean-to-integer casts (we just propagate the BoolValue to the result |
127 | // of the cast). So, a join can encounter an integer in one branch but a |
128 | // bool in the other. |
129 | // For example: |
130 | // ``` |
131 | // std::optional<bool> o; |
132 | // int x; |
133 | // if (o.has_value()) |
134 | // x = o.value(); |
135 | // ``` |
136 | auto &Expr1 = cast<BoolValue>(Val&: Val1).formula(); |
137 | auto &Expr2 = cast<BoolValue>(Val&: Val2).formula(); |
138 | auto &A = JoinedEnv.arena(); |
139 | auto &JoinedVal = A.makeAtomRef(A: A.makeAtom()); |
140 | JoinedEnv.assume( |
141 | A.makeOr(LHS: A.makeAnd(LHS: A.makeAtomRef(A: Env1.getFlowConditionToken()), |
142 | RHS: A.makeEquals(LHS: JoinedVal, RHS: Expr1)), |
143 | RHS: A.makeAnd(LHS: A.makeAtomRef(A: Env2.getFlowConditionToken()), |
144 | RHS: A.makeEquals(LHS: JoinedVal, RHS: Expr2)))); |
145 | return &A.makeBoolValue(JoinedVal); |
146 | } |
147 | |
148 | Value *JoinedVal = JoinedEnv.createValue(Type); |
149 | if (JoinedVal) |
150 | Model.join(Type, Val1, Env1, Val2, Env2, JoinedVal&: *JoinedVal, JoinedEnv); |
151 | |
152 | return JoinedVal; |
153 | } |
154 | |
155 | static WidenResult widenDistinctValues(QualType Type, Value &Prev, |
156 | const Environment &PrevEnv, |
157 | Value &Current, Environment &CurrentEnv, |
158 | Environment::ValueModel &Model) { |
159 | // Boolean-model widening. |
160 | if (auto *PrevBool = dyn_cast<BoolValue>(Val: &Prev)) { |
161 | if (isa<TopBoolValue>(Val: Prev)) |
162 | // Safe to return `Prev` here, because Top is never dependent on the |
163 | // environment. |
164 | return {.V: &Prev, .Effect: LatticeEffect::Unchanged}; |
165 | |
166 | // We may need to widen to Top, but before we do so, check whether both |
167 | // values are implied to be either true or false in the current environment. |
168 | // In that case, we can simply return a literal instead. |
169 | auto &CurBool = cast<BoolValue>(Val&: Current); |
170 | bool TruePrev = PrevEnv.proves(PrevBool->formula()); |
171 | bool TrueCur = CurrentEnv.proves(CurBool.formula()); |
172 | if (TruePrev && TrueCur) |
173 | return {.V: &CurrentEnv.getBoolLiteralValue(Value: true), .Effect: LatticeEffect::Unchanged}; |
174 | if (!TruePrev && !TrueCur && |
175 | PrevEnv.proves(PrevEnv.arena().makeNot(Val: PrevBool->formula())) && |
176 | CurrentEnv.proves(CurrentEnv.arena().makeNot(Val: CurBool.formula()))) |
177 | return {.V: &CurrentEnv.getBoolLiteralValue(Value: false), .Effect: LatticeEffect::Unchanged}; |
178 | |
179 | return {.V: &CurrentEnv.makeTopBoolValue(), .Effect: LatticeEffect::Changed}; |
180 | } |
181 | |
182 | // FIXME: Add other built-in model widening. |
183 | |
184 | // Custom-model widening. |
185 | if (auto Result = Model.widen(Type, Prev, PrevEnv, Current, CurrentEnv)) |
186 | return *Result; |
187 | |
188 | return {.V: &Current, .Effect: equateUnknownValues(K: Prev.getKind()) |
189 | ? LatticeEffect::Unchanged |
190 | : LatticeEffect::Changed}; |
191 | } |
192 | |
193 | // Returns whether the values in `Map1` and `Map2` compare equal for those |
194 | // keys that `Map1` and `Map2` have in common. |
195 | template <typename Key> |
196 | bool compareKeyToValueMaps(const llvm::MapVector<Key, Value *> &Map1, |
197 | const llvm::MapVector<Key, Value *> &Map2, |
198 | const Environment &Env1, const Environment &Env2, |
199 | Environment::ValueModel &Model) { |
200 | for (auto &Entry : Map1) { |
201 | Key K = Entry.first; |
202 | assert(K != nullptr); |
203 | |
204 | Value *Val = Entry.second; |
205 | assert(Val != nullptr); |
206 | |
207 | auto It = Map2.find(K); |
208 | if (It == Map2.end()) |
209 | continue; |
210 | assert(It->second != nullptr); |
211 | |
212 | if (!areEquivalentValues(*Val, *It->second) && |
213 | !compareDistinctValues(K->getType(), *Val, Env1, *It->second, Env2, |
214 | Model)) |
215 | return false; |
216 | } |
217 | |
218 | return true; |
219 | } |
220 | |
221 | // Perform a join on two `LocToVal` maps. |
222 | static llvm::MapVector<const StorageLocation *, Value *> |
223 | joinLocToVal(const llvm::MapVector<const StorageLocation *, Value *> &LocToVal, |
224 | const llvm::MapVector<const StorageLocation *, Value *> &LocToVal2, |
225 | const Environment &Env1, const Environment &Env2, |
226 | Environment &JoinedEnv, Environment::ValueModel &Model) { |
227 | llvm::MapVector<const StorageLocation *, Value *> Result; |
228 | for (auto &Entry : LocToVal) { |
229 | const StorageLocation *Loc = Entry.first; |
230 | assert(Loc != nullptr); |
231 | |
232 | Value *Val = Entry.second; |
233 | assert(Val != nullptr); |
234 | |
235 | auto It = LocToVal2.find(Key: Loc); |
236 | if (It == LocToVal2.end()) |
237 | continue; |
238 | assert(It->second != nullptr); |
239 | |
240 | if (Value *JoinedVal = Environment::joinValues( |
241 | Ty: Loc->getType(), Val1: Val, Env1, Val2: It->second, Env2, JoinedEnv, Model)) { |
242 | Result.insert(KV: {Loc, JoinedVal}); |
243 | } |
244 | } |
245 | |
246 | return Result; |
247 | } |
248 | |
249 | // Perform widening on either `LocToVal` or `ExprToVal`. `Key` must be either |
250 | // `const StorageLocation *` or `const Expr *`. |
251 | template <typename Key> |
252 | llvm::MapVector<Key, Value *> |
253 | widenKeyToValueMap(const llvm::MapVector<Key, Value *> &CurMap, |
254 | const llvm::MapVector<Key, Value *> &PrevMap, |
255 | Environment &CurEnv, const Environment &PrevEnv, |
256 | Environment::ValueModel &Model, LatticeEffect &Effect) { |
257 | llvm::MapVector<Key, Value *> WidenedMap; |
258 | for (auto &Entry : CurMap) { |
259 | Key K = Entry.first; |
260 | assert(K != nullptr); |
261 | |
262 | Value *Val = Entry.second; |
263 | assert(Val != nullptr); |
264 | |
265 | auto PrevIt = PrevMap.find(K); |
266 | if (PrevIt == PrevMap.end()) |
267 | continue; |
268 | assert(PrevIt->second != nullptr); |
269 | |
270 | if (areEquivalentValues(*Val, *PrevIt->second)) { |
271 | WidenedMap.insert({K, Val}); |
272 | continue; |
273 | } |
274 | |
275 | auto [WidenedVal, ValEffect] = widenDistinctValues( |
276 | K->getType(), *PrevIt->second, PrevEnv, *Val, CurEnv, Model); |
277 | WidenedMap.insert({K, WidenedVal}); |
278 | if (ValEffect == LatticeEffect::Changed) |
279 | Effect = LatticeEffect::Changed; |
280 | } |
281 | |
282 | return WidenedMap; |
283 | } |
284 | |
285 | namespace { |
286 | |
287 | // Visitor that builds a map from record prvalues to result objects. |
288 | // This traverses the body of the function to be analyzed; for each result |
289 | // object that it encounters, it propagates the storage location of the result |
290 | // object to all record prvalues that can initialize it. |
291 | class ResultObjectVisitor : public RecursiveASTVisitor<ResultObjectVisitor> { |
292 | public: |
293 | // `ResultObjectMap` will be filled with a map from record prvalues to result |
294 | // object. If the function being analyzed returns a record by value, |
295 | // `LocForRecordReturnVal` is the location to which this record should be |
296 | // written; otherwise, it is null. |
297 | explicit ResultObjectVisitor( |
298 | llvm::DenseMap<const Expr *, RecordStorageLocation *> &ResultObjectMap, |
299 | RecordStorageLocation *LocForRecordReturnVal, |
300 | DataflowAnalysisContext &DACtx) |
301 | : ResultObjectMap(ResultObjectMap), |
302 | LocForRecordReturnVal(LocForRecordReturnVal), DACtx(DACtx) {} |
303 | |
304 | bool shouldVisitImplicitCode() { return true; } |
305 | |
306 | bool shouldVisitLambdaBody() const { return false; } |
307 | |
308 | // Traverse all member and base initializers of `Ctor`. This function is not |
309 | // called by `RecursiveASTVisitor`; it should be called manually if we are |
310 | // analyzing a constructor. `ThisPointeeLoc` is the storage location that |
311 | // `this` points to. |
312 | void TraverseConstructorInits(const CXXConstructorDecl *Ctor, |
313 | RecordStorageLocation *ThisPointeeLoc) { |
314 | assert(ThisPointeeLoc != nullptr); |
315 | for (const CXXCtorInitializer *Init : Ctor->inits()) { |
316 | Expr *InitExpr = Init->getInit(); |
317 | if (FieldDecl *Field = Init->getMember(); |
318 | Field != nullptr && Field->getType()->isRecordType()) { |
319 | PropagateResultObject(E: InitExpr, Loc: cast<RecordStorageLocation>( |
320 | Val: ThisPointeeLoc->getChild(*Field))); |
321 | } else if (Init->getBaseClass()) { |
322 | PropagateResultObject(E: InitExpr, Loc: ThisPointeeLoc); |
323 | } |
324 | |
325 | // Ensure that any result objects within `InitExpr` (e.g. temporaries) |
326 | // are also propagated to the prvalues that initialize them. |
327 | TraverseStmt(InitExpr); |
328 | |
329 | // If this is a `CXXDefaultInitExpr`, also propagate any result objects |
330 | // within the default expression. |
331 | if (auto *DefaultInit = dyn_cast<CXXDefaultInitExpr>(Val: InitExpr)) |
332 | TraverseStmt(DefaultInit->getExpr()); |
333 | } |
334 | } |
335 | |
336 | bool TraverseBindingDecl(BindingDecl *BD) { |
337 | // `RecursiveASTVisitor` doesn't traverse holding variables for |
338 | // `BindingDecl`s by itself, so we need to tell it to. |
339 | if (VarDecl *HoldingVar = BD->getHoldingVar()) |
340 | TraverseDecl(HoldingVar); |
341 | return RecursiveASTVisitor<ResultObjectVisitor>::TraverseBindingDecl(BD); |
342 | } |
343 | |
344 | bool VisitVarDecl(VarDecl *VD) { |
345 | if (VD->getType()->isRecordType() && VD->hasInit()) |
346 | PropagateResultObject( |
347 | E: VD->getInit(), |
348 | Loc: &cast<RecordStorageLocation>(Val&: DACtx.getStableStorageLocation(*VD))); |
349 | return true; |
350 | } |
351 | |
352 | bool VisitMaterializeTemporaryExpr(MaterializeTemporaryExpr *MTE) { |
353 | if (MTE->getType()->isRecordType()) |
354 | PropagateResultObject( |
355 | E: MTE->getSubExpr(), |
356 | Loc: &cast<RecordStorageLocation>(Val&: DACtx.getStableStorageLocation(*MTE))); |
357 | return true; |
358 | } |
359 | |
360 | bool VisitReturnStmt(ReturnStmt *Return) { |
361 | Expr *RetValue = Return->getRetValue(); |
362 | if (RetValue != nullptr && RetValue->getType()->isRecordType() && |
363 | RetValue->isPRValue()) |
364 | PropagateResultObject(E: RetValue, Loc: LocForRecordReturnVal); |
365 | return true; |
366 | } |
367 | |
368 | bool VisitExpr(Expr *E) { |
369 | // Clang's AST can have record-type prvalues without a result object -- for |
370 | // example as full-expressions contained in a compound statement or as |
371 | // arguments of call expressions. We notice this if we get here and a |
372 | // storage location has not yet been associated with `E`. In this case, |
373 | // treat this as if it was a `MaterializeTemporaryExpr`. |
374 | if (E->isPRValue() && E->getType()->isRecordType() && |
375 | !ResultObjectMap.contains(Val: E)) |
376 | PropagateResultObject( |
377 | E, Loc: &cast<RecordStorageLocation>(Val&: DACtx.getStableStorageLocation(E: *E))); |
378 | return true; |
379 | } |
380 | |
381 | void |
382 | PropagateResultObjectToRecordInitList(const RecordInitListHelper &InitList, |
383 | RecordStorageLocation *Loc) { |
384 | for (auto [Base, Init] : InitList.base_inits()) { |
385 | assert(Base->getType().getCanonicalType() == |
386 | Init->getType().getCanonicalType()); |
387 | |
388 | // Storage location for the base class is the same as that of the |
389 | // derived class because we "flatten" the object hierarchy and put all |
390 | // fields in `RecordStorageLocation` of the derived class. |
391 | PropagateResultObject(E: Init, Loc); |
392 | } |
393 | |
394 | for (auto [Field, Init] : InitList.field_inits()) { |
395 | // Fields of non-record type are handled in |
396 | // `TransferVisitor::VisitInitListExpr()`. |
397 | if (Field->getType()->isRecordType()) |
398 | PropagateResultObject( |
399 | E: Init, Loc: cast<RecordStorageLocation>(Val: Loc->getChild(*Field))); |
400 | } |
401 | } |
402 | |
403 | // Assigns `Loc` as the result object location of `E`, then propagates the |
404 | // location to all lower-level prvalues that initialize the same object as |
405 | // `E` (or one of its base classes or member variables). |
406 | void PropagateResultObject(Expr *E, RecordStorageLocation *Loc) { |
407 | if (!E->isPRValue() || !E->getType()->isRecordType()) { |
408 | assert(false); |
409 | // Ensure we don't propagate the result object if we hit this in a |
410 | // release build. |
411 | return; |
412 | } |
413 | |
414 | ResultObjectMap[E] = Loc; |
415 | |
416 | // The following AST node kinds are "original initializers": They are the |
417 | // lowest-level AST node that initializes a given object, and nothing |
418 | // below them can initialize the same object (or part of it). |
419 | if (isa<CXXConstructExpr>(Val: E) || isa<CallExpr>(Val: E) || isa<LambdaExpr>(Val: E) || |
420 | isa<CXXDefaultArgExpr>(Val: E) || isa<CXXDefaultInitExpr>(Val: E) || |
421 | isa<CXXStdInitializerListExpr>(Val: E) || |
422 | // We treat `BuiltinBitCastExpr` as an "original initializer" too as |
423 | // it may not even be casting from a record type -- and even if it is, |
424 | // the two objects are in general of unrelated type. |
425 | isa<BuiltinBitCastExpr>(Val: E)) { |
426 | return; |
427 | } |
428 | if (auto *Op = dyn_cast<BinaryOperator>(Val: E); |
429 | Op && Op->getOpcode() == BO_Cmp) { |
430 | // Builtin `<=>` returns a `std::strong_ordering` object. |
431 | return; |
432 | } |
433 | |
434 | if (auto *InitList = dyn_cast<InitListExpr>(Val: E)) { |
435 | if (!InitList->isSemanticForm()) |
436 | return; |
437 | if (InitList->isTransparent()) { |
438 | PropagateResultObject(E: InitList->getInit(Init: 0), Loc); |
439 | return; |
440 | } |
441 | |
442 | PropagateResultObjectToRecordInitList(InitList: RecordInitListHelper(InitList), |
443 | Loc); |
444 | return; |
445 | } |
446 | |
447 | if (auto *ParenInitList = dyn_cast<CXXParenListInitExpr>(Val: E)) { |
448 | PropagateResultObjectToRecordInitList(InitList: RecordInitListHelper(ParenInitList), |
449 | Loc); |
450 | return; |
451 | } |
452 | |
453 | if (auto *Op = dyn_cast<BinaryOperator>(Val: E); Op && Op->isCommaOp()) { |
454 | PropagateResultObject(E: Op->getRHS(), Loc); |
455 | return; |
456 | } |
457 | |
458 | if (auto *Cond = dyn_cast<AbstractConditionalOperator>(Val: E)) { |
459 | PropagateResultObject(E: Cond->getTrueExpr(), Loc); |
460 | PropagateResultObject(E: Cond->getFalseExpr(), Loc); |
461 | return; |
462 | } |
463 | |
464 | if (auto *SE = dyn_cast<StmtExpr>(Val: E)) { |
465 | PropagateResultObject(E: cast<Expr>(Val: SE->getSubStmt()->body_back()), Loc); |
466 | return; |
467 | } |
468 | |
469 | // All other expression nodes that propagate a record prvalue should have |
470 | // exactly one child. |
471 | SmallVector<Stmt *, 1> Children(E->child_begin(), E->child_end()); |
472 | LLVM_DEBUG({ |
473 | if (Children.size() != 1) |
474 | E->dump(); |
475 | }); |
476 | assert(Children.size() == 1); |
477 | for (Stmt *S : Children) |
478 | PropagateResultObject(cast<Expr>(S), Loc); |
479 | } |
480 | |
481 | private: |
482 | llvm::DenseMap<const Expr *, RecordStorageLocation *> &ResultObjectMap; |
483 | RecordStorageLocation *LocForRecordReturnVal; |
484 | DataflowAnalysisContext &DACtx; |
485 | }; |
486 | |
487 | } // namespace |
488 | |
489 | Environment::Environment(DataflowAnalysisContext &DACtx) |
490 | : DACtx(&DACtx), |
491 | FlowConditionToken(DACtx.arena().makeFlowConditionToken()) {} |
492 | |
493 | Environment::Environment(DataflowAnalysisContext &DACtx, |
494 | const DeclContext &DeclCtx) |
495 | : Environment(DACtx) { |
496 | CallStack.push_back(x: &DeclCtx); |
497 | } |
498 | |
499 | void Environment::initialize() { |
500 | const DeclContext *DeclCtx = getDeclCtx(); |
501 | if (DeclCtx == nullptr) |
502 | return; |
503 | |
504 | const auto *FuncDecl = dyn_cast<FunctionDecl>(Val: DeclCtx); |
505 | if (FuncDecl == nullptr) |
506 | return; |
507 | |
508 | assert(FuncDecl->doesThisDeclarationHaveABody()); |
509 | |
510 | initFieldsGlobalsAndFuncs(FuncDecl); |
511 | |
512 | for (const auto *ParamDecl : FuncDecl->parameters()) { |
513 | assert(ParamDecl != nullptr); |
514 | setStorageLocation(*ParamDecl, createObject(*ParamDecl, nullptr)); |
515 | } |
516 | |
517 | if (FuncDecl->getReturnType()->isRecordType()) |
518 | LocForRecordReturnVal = &cast<RecordStorageLocation>( |
519 | Val&: createStorageLocation(Type: FuncDecl->getReturnType())); |
520 | |
521 | if (const auto *MethodDecl = dyn_cast<CXXMethodDecl>(Val: DeclCtx)) { |
522 | auto *Parent = MethodDecl->getParent(); |
523 | assert(Parent != nullptr); |
524 | |
525 | if (Parent->isLambda()) { |
526 | for (const auto &Capture : Parent->captures()) { |
527 | if (Capture.capturesVariable()) { |
528 | const auto *VarDecl = Capture.getCapturedVar(); |
529 | assert(VarDecl != nullptr); |
530 | setStorageLocation(D: *VarDecl, Loc&: createObject(D: *VarDecl, InitExpr: nullptr)); |
531 | } else if (Capture.capturesThis()) { |
532 | const auto *SurroundingMethodDecl = |
533 | cast<CXXMethodDecl>(Val: DeclCtx->getNonClosureAncestor()); |
534 | QualType ThisPointeeType = |
535 | SurroundingMethodDecl->getFunctionObjectParameterType(); |
536 | setThisPointeeStorageLocation( |
537 | cast<RecordStorageLocation>(Val&: createObject(Ty: ThisPointeeType))); |
538 | } |
539 | } |
540 | } else if (MethodDecl->isImplicitObjectMemberFunction()) { |
541 | QualType ThisPointeeType = MethodDecl->getFunctionObjectParameterType(); |
542 | auto &ThisLoc = |
543 | cast<RecordStorageLocation>(Val&: createStorageLocation(Type: ThisPointeeType)); |
544 | setThisPointeeStorageLocation(ThisLoc); |
545 | // Initialize fields of `*this` with values, but only if we're not |
546 | // analyzing a constructor; after all, it's the constructor's job to do |
547 | // this (and we want to be able to test that). |
548 | if (!isa<CXXConstructorDecl>(Val: MethodDecl)) |
549 | initializeFieldsWithValues(Loc&: ThisLoc); |
550 | } |
551 | } |
552 | |
553 | // We do this below the handling of `CXXMethodDecl` above so that we can |
554 | // be sure that the storage location for `this` has been set. |
555 | ResultObjectMap = std::make_shared<PrValueToResultObject>( |
556 | args: buildResultObjectMap(DACtx, FuncDecl, ThisPointeeLoc: getThisPointeeStorageLocation(), |
557 | LocForRecordReturnVal)); |
558 | } |
559 | |
560 | // FIXME: Add support for resetting globals after function calls to enable |
561 | // the implementation of sound analyses. |
562 | void Environment::initFieldsGlobalsAndFuncs(const FunctionDecl *FuncDecl) { |
563 | assert(FuncDecl->doesThisDeclarationHaveABody()); |
564 | |
565 | ReferencedDecls Referenced = getReferencedDecls(FD: *FuncDecl); |
566 | |
567 | // These have to be added before the lines that follow to ensure that |
568 | // `create*` work correctly for structs. |
569 | DACtx->addModeledFields(Fields: Referenced.Fields); |
570 | |
571 | for (const VarDecl *D : Referenced.Globals) { |
572 | if (getStorageLocation(*D) != nullptr) |
573 | continue; |
574 | |
575 | // We don't run transfer functions on the initializers of global variables, |
576 | // so they won't be associated with a value or storage location. We |
577 | // therefore intentionally don't pass an initializer to `createObject()`; |
578 | // in particular, this ensures that `createObject()` will initialize the |
579 | // fields of record-type variables with values. |
580 | setStorageLocation(*D, createObject(*D, nullptr)); |
581 | } |
582 | |
583 | for (const FunctionDecl *FD : Referenced.Functions) { |
584 | if (getStorageLocation(*FD) != nullptr) |
585 | continue; |
586 | auto &Loc = createStorageLocation(*FD); |
587 | setStorageLocation(*FD, Loc); |
588 | } |
589 | } |
590 | |
591 | Environment Environment::fork() const { |
592 | Environment Copy(*this); |
593 | Copy.FlowConditionToken = DACtx->forkFlowCondition(Token: FlowConditionToken); |
594 | return Copy; |
595 | } |
596 | |
597 | bool Environment::canDescend(unsigned MaxDepth, |
598 | const DeclContext *Callee) const { |
599 | return CallStack.size() <= MaxDepth && !llvm::is_contained(Range: CallStack, Element: Callee); |
600 | } |
601 | |
602 | Environment Environment::pushCall(const CallExpr *Call) const { |
603 | Environment Env(*this); |
604 | |
605 | if (const auto *MethodCall = dyn_cast<CXXMemberCallExpr>(Val: Call)) { |
606 | if (const Expr *Arg = MethodCall->getImplicitObjectArgument()) { |
607 | if (!isa<CXXThisExpr>(Val: Arg)) |
608 | Env.ThisPointeeLoc = |
609 | cast<RecordStorageLocation>(Val: getStorageLocation(E: *Arg)); |
610 | // Otherwise (when the argument is `this`), retain the current |
611 | // environment's `ThisPointeeLoc`. |
612 | } |
613 | } |
614 | |
615 | if (Call->getType()->isRecordType() && Call->isPRValue()) |
616 | Env.LocForRecordReturnVal = &Env.getResultObjectLocation(*Call); |
617 | |
618 | Env.pushCallInternal(FuncDecl: Call->getDirectCallee(), |
619 | Args: llvm::ArrayRef(Call->getArgs(), Call->getNumArgs())); |
620 | |
621 | return Env; |
622 | } |
623 | |
624 | Environment Environment::pushCall(const CXXConstructExpr *Call) const { |
625 | Environment Env(*this); |
626 | |
627 | Env.ThisPointeeLoc = &Env.getResultObjectLocation(*Call); |
628 | Env.LocForRecordReturnVal = &Env.getResultObjectLocation(*Call); |
629 | |
630 | Env.pushCallInternal(Call->getConstructor(), |
631 | llvm::ArrayRef(Call->getArgs(), Call->getNumArgs())); |
632 | |
633 | return Env; |
634 | } |
635 | |
636 | void Environment::pushCallInternal(const FunctionDecl *FuncDecl, |
637 | ArrayRef<const Expr *> Args) { |
638 | // Canonicalize to the definition of the function. This ensures that we're |
639 | // putting arguments into the same `ParamVarDecl`s` that the callee will later |
640 | // be retrieving them from. |
641 | assert(FuncDecl->getDefinition() != nullptr); |
642 | FuncDecl = FuncDecl->getDefinition(); |
643 | |
644 | CallStack.push_back(FuncDecl); |
645 | |
646 | initFieldsGlobalsAndFuncs(FuncDecl); |
647 | |
648 | const auto *ParamIt = FuncDecl->param_begin(); |
649 | |
650 | // FIXME: Parameters don't always map to arguments 1:1; examples include |
651 | // overloaded operators implemented as member functions, and parameter packs. |
652 | for (unsigned ArgIndex = 0; ArgIndex < Args.size(); ++ParamIt, ++ArgIndex) { |
653 | assert(ParamIt != FuncDecl->param_end()); |
654 | const VarDecl *Param = *ParamIt; |
655 | setStorageLocation(*Param, createObject(*Param, Args[ArgIndex])); |
656 | } |
657 | |
658 | ResultObjectMap = std::make_shared<PrValueToResultObject>( |
659 | args: buildResultObjectMap(DACtx, FuncDecl, ThisPointeeLoc: getThisPointeeStorageLocation(), |
660 | LocForRecordReturnVal)); |
661 | } |
662 | |
663 | void Environment::popCall(const CallExpr *Call, const Environment &CalleeEnv) { |
664 | // We ignore some entries of `CalleeEnv`: |
665 | // - `DACtx` because is already the same in both |
666 | // - We don't want the callee's `DeclCtx`, `ReturnVal`, `ReturnLoc` or |
667 | // `ThisPointeeLoc` because they don't apply to us. |
668 | // - `DeclToLoc`, `ExprToLoc`, and `ExprToVal` capture information from the |
669 | // callee's local scope, so when popping that scope, we do not propagate |
670 | // the maps. |
671 | this->LocToVal = std::move(CalleeEnv.LocToVal); |
672 | this->FlowConditionToken = std::move(CalleeEnv.FlowConditionToken); |
673 | |
674 | if (Call->isGLValue()) { |
675 | if (CalleeEnv.ReturnLoc != nullptr) |
676 | setStorageLocation(*Call, *CalleeEnv.ReturnLoc); |
677 | } else if (!Call->getType()->isVoidType()) { |
678 | if (CalleeEnv.ReturnVal != nullptr) |
679 | setValue(*Call, *CalleeEnv.ReturnVal); |
680 | } |
681 | } |
682 | |
683 | void Environment::popCall(const CXXConstructExpr *Call, |
684 | const Environment &CalleeEnv) { |
685 | // See also comment in `popCall(const CallExpr *, const Environment &)` above. |
686 | this->LocToVal = std::move(CalleeEnv.LocToVal); |
687 | this->FlowConditionToken = std::move(CalleeEnv.FlowConditionToken); |
688 | } |
689 | |
690 | bool Environment::equivalentTo(const Environment &Other, |
691 | Environment::ValueModel &Model) const { |
692 | assert(DACtx == Other.DACtx); |
693 | |
694 | if (ReturnVal != Other.ReturnVal) |
695 | return false; |
696 | |
697 | if (ReturnLoc != Other.ReturnLoc) |
698 | return false; |
699 | |
700 | if (LocForRecordReturnVal != Other.LocForRecordReturnVal) |
701 | return false; |
702 | |
703 | if (ThisPointeeLoc != Other.ThisPointeeLoc) |
704 | return false; |
705 | |
706 | if (DeclToLoc != Other.DeclToLoc) |
707 | return false; |
708 | |
709 | if (ExprToLoc != Other.ExprToLoc) |
710 | return false; |
711 | |
712 | if (!compareKeyToValueMaps(Map1: ExprToVal, Map2: Other.ExprToVal, Env1: *this, Env2: Other, Model)) |
713 | return false; |
714 | |
715 | if (!compareKeyToValueMaps(Map1: LocToVal, Map2: Other.LocToVal, Env1: *this, Env2: Other, Model)) |
716 | return false; |
717 | |
718 | return true; |
719 | } |
720 | |
721 | LatticeEffect Environment::widen(const Environment &PrevEnv, |
722 | Environment::ValueModel &Model) { |
723 | assert(DACtx == PrevEnv.DACtx); |
724 | assert(ReturnVal == PrevEnv.ReturnVal); |
725 | assert(ReturnLoc == PrevEnv.ReturnLoc); |
726 | assert(LocForRecordReturnVal == PrevEnv.LocForRecordReturnVal); |
727 | assert(ThisPointeeLoc == PrevEnv.ThisPointeeLoc); |
728 | assert(CallStack == PrevEnv.CallStack); |
729 | assert(ResultObjectMap == PrevEnv.ResultObjectMap); |
730 | |
731 | auto Effect = LatticeEffect::Unchanged; |
732 | |
733 | // By the API, `PrevEnv` is a previous version of the environment for the same |
734 | // block, so we have some guarantees about its shape. In particular, it will |
735 | // be the result of a join or widen operation on previous values for this |
736 | // block. For `DeclToLoc`, `ExprToVal`, and `ExprToLoc`, join guarantees that |
737 | // these maps are subsets of the maps in `PrevEnv`. So, as long as we maintain |
738 | // this property here, we don't need change their current values to widen. |
739 | assert(DeclToLoc.size() <= PrevEnv.DeclToLoc.size()); |
740 | assert(ExprToVal.size() <= PrevEnv.ExprToVal.size()); |
741 | assert(ExprToLoc.size() <= PrevEnv.ExprToLoc.size()); |
742 | |
743 | ExprToVal = widenKeyToValueMap(CurMap: ExprToVal, PrevMap: PrevEnv.ExprToVal, CurEnv&: *this, PrevEnv, |
744 | Model, Effect); |
745 | |
746 | LocToVal = widenKeyToValueMap(CurMap: LocToVal, PrevMap: PrevEnv.LocToVal, CurEnv&: *this, PrevEnv, |
747 | Model, Effect); |
748 | if (DeclToLoc.size() != PrevEnv.DeclToLoc.size() || |
749 | ExprToLoc.size() != PrevEnv.ExprToLoc.size() || |
750 | ExprToVal.size() != PrevEnv.ExprToVal.size() || |
751 | LocToVal.size() != PrevEnv.LocToVal.size()) |
752 | Effect = LatticeEffect::Changed; |
753 | |
754 | return Effect; |
755 | } |
756 | |
757 | Environment Environment::join(const Environment &EnvA, const Environment &EnvB, |
758 | Environment::ValueModel &Model, |
759 | ExprJoinBehavior ExprBehavior) { |
760 | assert(EnvA.DACtx == EnvB.DACtx); |
761 | assert(EnvA.LocForRecordReturnVal == EnvB.LocForRecordReturnVal); |
762 | assert(EnvA.ThisPointeeLoc == EnvB.ThisPointeeLoc); |
763 | assert(EnvA.CallStack == EnvB.CallStack); |
764 | assert(EnvA.ResultObjectMap == EnvB.ResultObjectMap); |
765 | |
766 | Environment JoinedEnv(*EnvA.DACtx); |
767 | |
768 | JoinedEnv.CallStack = EnvA.CallStack; |
769 | JoinedEnv.ResultObjectMap = EnvA.ResultObjectMap; |
770 | JoinedEnv.LocForRecordReturnVal = EnvA.LocForRecordReturnVal; |
771 | JoinedEnv.ThisPointeeLoc = EnvA.ThisPointeeLoc; |
772 | |
773 | if (EnvA.CallStack.empty()) { |
774 | JoinedEnv.ReturnVal = nullptr; |
775 | } else { |
776 | // FIXME: Make `CallStack` a vector of `FunctionDecl` so we don't need this |
777 | // cast. |
778 | auto *Func = dyn_cast<FunctionDecl>(Val: EnvA.CallStack.back()); |
779 | assert(Func != nullptr); |
780 | JoinedEnv.ReturnVal = |
781 | joinValues(Ty: Func->getReturnType(), Val1: EnvA.ReturnVal, Env1: EnvA, Val2: EnvB.ReturnVal, |
782 | Env2: EnvB, JoinedEnv, Model); |
783 | } |
784 | |
785 | if (EnvA.ReturnLoc == EnvB.ReturnLoc) |
786 | JoinedEnv.ReturnLoc = EnvA.ReturnLoc; |
787 | else |
788 | JoinedEnv.ReturnLoc = nullptr; |
789 | |
790 | JoinedEnv.DeclToLoc = intersectDeclToLoc(DeclToLoc1: EnvA.DeclToLoc, DeclToLoc2: EnvB.DeclToLoc); |
791 | |
792 | // FIXME: update join to detect backedges and simplify the flow condition |
793 | // accordingly. |
794 | JoinedEnv.FlowConditionToken = EnvA.DACtx->joinFlowConditions( |
795 | FirstToken: EnvA.FlowConditionToken, SecondToken: EnvB.FlowConditionToken); |
796 | |
797 | JoinedEnv.LocToVal = |
798 | joinLocToVal(LocToVal: EnvA.LocToVal, LocToVal2: EnvB.LocToVal, Env1: EnvA, Env2: EnvB, JoinedEnv, Model); |
799 | |
800 | if (ExprBehavior == KeepExprState) { |
801 | JoinedEnv.ExprToVal = joinExprMaps(Map1: EnvA.ExprToVal, Map2: EnvB.ExprToVal); |
802 | JoinedEnv.ExprToLoc = joinExprMaps(Map1: EnvA.ExprToLoc, Map2: EnvB.ExprToLoc); |
803 | } |
804 | |
805 | return JoinedEnv; |
806 | } |
807 | |
808 | Value *Environment::joinValues(QualType Ty, Value *Val1, |
809 | const Environment &Env1, Value *Val2, |
810 | const Environment &Env2, Environment &JoinedEnv, |
811 | Environment::ValueModel &Model) { |
812 | if (Val1 == nullptr || Val2 == nullptr) |
813 | // We can't say anything about the joined value -- even if one of the values |
814 | // is non-null, we don't want to simply propagate it, because it would be |
815 | // too specific: Because the other value is null, that means we have no |
816 | // information at all about the value (i.e. the value is unconstrained). |
817 | return nullptr; |
818 | |
819 | if (areEquivalentValues(Val1: *Val1, Val2: *Val2)) |
820 | // Arbitrarily return one of the two values. |
821 | return Val1; |
822 | |
823 | return joinDistinctValues(Type: Ty, Val1&: *Val1, Env1, Val2&: *Val2, Env2, JoinedEnv, Model); |
824 | } |
825 | |
826 | StorageLocation &Environment::createStorageLocation(QualType Type) { |
827 | return DACtx->createStorageLocation(Type); |
828 | } |
829 | |
830 | StorageLocation &Environment::createStorageLocation(const ValueDecl &D) { |
831 | // Evaluated declarations are always assigned the same storage locations to |
832 | // ensure that the environment stabilizes across loop iterations. Storage |
833 | // locations for evaluated declarations are stored in the analysis context. |
834 | return DACtx->getStableStorageLocation(D); |
835 | } |
836 | |
837 | StorageLocation &Environment::createStorageLocation(const Expr &E) { |
838 | // Evaluated expressions are always assigned the same storage locations to |
839 | // ensure that the environment stabilizes across loop iterations. Storage |
840 | // locations for evaluated expressions are stored in the analysis context. |
841 | return DACtx->getStableStorageLocation(E); |
842 | } |
843 | |
844 | void Environment::setStorageLocation(const ValueDecl &D, StorageLocation &Loc) { |
845 | assert(!DeclToLoc.contains(&D)); |
846 | // The only kinds of declarations that may have a "variable" storage location |
847 | // are declarations of reference type and `BindingDecl`. For all other |
848 | // declaration, the storage location should be the stable storage location |
849 | // returned by `createStorageLocation()`. |
850 | assert(D.getType()->isReferenceType() || isa<BindingDecl>(D) || |
851 | &Loc == &createStorageLocation(D)); |
852 | DeclToLoc[&D] = &Loc; |
853 | } |
854 | |
855 | StorageLocation *Environment::getStorageLocation(const ValueDecl &D) const { |
856 | auto It = DeclToLoc.find(Val: &D); |
857 | if (It == DeclToLoc.end()) |
858 | return nullptr; |
859 | |
860 | StorageLocation *Loc = It->second; |
861 | |
862 | return Loc; |
863 | } |
864 | |
865 | void Environment::removeDecl(const ValueDecl &D) { DeclToLoc.erase(Val: &D); } |
866 | |
867 | void Environment::setStorageLocation(const Expr &E, StorageLocation &Loc) { |
868 | // `DeclRefExpr`s to builtin function types aren't glvalues, for some reason, |
869 | // but we still want to be able to associate a `StorageLocation` with them, |
870 | // so allow these as an exception. |
871 | assert(E.isGLValue() || |
872 | E.getType()->isSpecificBuiltinType(BuiltinType::BuiltinFn)); |
873 | const Expr &CanonE = ignoreCFGOmittedNodes(E); |
874 | assert(!ExprToLoc.contains(&CanonE)); |
875 | ExprToLoc[&CanonE] = &Loc; |
876 | } |
877 | |
878 | StorageLocation *Environment::getStorageLocation(const Expr &E) const { |
879 | // See comment in `setStorageLocation()`. |
880 | assert(E.isGLValue() || |
881 | E.getType()->isSpecificBuiltinType(BuiltinType::BuiltinFn)); |
882 | auto It = ExprToLoc.find(Val: &ignoreCFGOmittedNodes(E)); |
883 | return It == ExprToLoc.end() ? nullptr : &*It->second; |
884 | } |
885 | |
886 | RecordStorageLocation & |
887 | Environment::getResultObjectLocation(const Expr &RecordPRValue) const { |
888 | assert(RecordPRValue.getType()->isRecordType()); |
889 | assert(RecordPRValue.isPRValue()); |
890 | |
891 | assert(ResultObjectMap != nullptr); |
892 | RecordStorageLocation *Loc = ResultObjectMap->lookup(Val: &RecordPRValue); |
893 | assert(Loc != nullptr); |
894 | // In release builds, use the "stable" storage location if the map lookup |
895 | // failed. |
896 | if (Loc == nullptr) |
897 | return cast<RecordStorageLocation>( |
898 | Val&: DACtx->getStableStorageLocation(E: RecordPRValue)); |
899 | return *Loc; |
900 | } |
901 | |
902 | PointerValue &Environment::getOrCreateNullPointerValue(QualType PointeeType) { |
903 | return DACtx->getOrCreateNullPointerValue(PointeeType); |
904 | } |
905 | |
906 | void Environment::initializeFieldsWithValues(RecordStorageLocation &Loc, |
907 | QualType Type) { |
908 | llvm::DenseSet<QualType> Visited; |
909 | int CreatedValuesCount = 0; |
910 | initializeFieldsWithValues(Loc, Type, Visited, Depth: 0, CreatedValuesCount); |
911 | if (CreatedValuesCount > MaxCompositeValueSize) { |
912 | llvm::errs() << "Attempting to initialize a huge value of type: " << Type |
913 | << '\n'; |
914 | } |
915 | } |
916 | |
917 | void Environment::setValue(const StorageLocation &Loc, Value &Val) { |
918 | // Records should not be associated with values. |
919 | assert(!isa<RecordStorageLocation>(Loc)); |
920 | LocToVal[&Loc] = &Val; |
921 | } |
922 | |
923 | void Environment::setValue(const Expr &E, Value &Val) { |
924 | const Expr &CanonE = ignoreCFGOmittedNodes(E); |
925 | |
926 | assert(CanonE.isPRValue()); |
927 | // Records should not be associated with values. |
928 | assert(!CanonE.getType()->isRecordType()); |
929 | ExprToVal[&CanonE] = &Val; |
930 | } |
931 | |
932 | Value *Environment::getValue(const StorageLocation &Loc) const { |
933 | // Records should not be associated with values. |
934 | assert(!isa<RecordStorageLocation>(Loc)); |
935 | return LocToVal.lookup(Key: &Loc); |
936 | } |
937 | |
938 | Value *Environment::getValue(const ValueDecl &D) const { |
939 | auto *Loc = getStorageLocation(D); |
940 | if (Loc == nullptr) |
941 | return nullptr; |
942 | return getValue(Loc: *Loc); |
943 | } |
944 | |
945 | Value *Environment::getValue(const Expr &E) const { |
946 | // Records should not be associated with values. |
947 | assert(!E.getType()->isRecordType()); |
948 | |
949 | if (E.isPRValue()) { |
950 | auto It = ExprToVal.find(Key: &ignoreCFGOmittedNodes(E)); |
951 | return It == ExprToVal.end() ? nullptr : It->second; |
952 | } |
953 | |
954 | auto It = ExprToLoc.find(Val: &ignoreCFGOmittedNodes(E)); |
955 | if (It == ExprToLoc.end()) |
956 | return nullptr; |
957 | return getValue(Loc: *It->second); |
958 | } |
959 | |
960 | Value *Environment::createValue(QualType Type) { |
961 | llvm::DenseSet<QualType> Visited; |
962 | int CreatedValuesCount = 0; |
963 | Value *Val = createValueUnlessSelfReferential(Type, Visited, /*Depth=*/0, |
964 | CreatedValuesCount); |
965 | if (CreatedValuesCount > MaxCompositeValueSize) { |
966 | llvm::errs() << "Attempting to initialize a huge value of type: " << Type |
967 | << '\n'; |
968 | } |
969 | return Val; |
970 | } |
971 | |
972 | Value *Environment::createValueUnlessSelfReferential( |
973 | QualType Type, llvm::DenseSet<QualType> &Visited, int Depth, |
974 | int &CreatedValuesCount) { |
975 | assert(!Type.isNull()); |
976 | assert(!Type->isReferenceType()); |
977 | assert(!Type->isRecordType()); |
978 | |
979 | // Allow unlimited fields at depth 1; only cap at deeper nesting levels. |
980 | if ((Depth > 1 && CreatedValuesCount > MaxCompositeValueSize) || |
981 | Depth > MaxCompositeValueDepth) |
982 | return nullptr; |
983 | |
984 | if (Type->isBooleanType()) { |
985 | CreatedValuesCount++; |
986 | return &makeAtomicBoolValue(); |
987 | } |
988 | |
989 | if (Type->isIntegerType()) { |
990 | // FIXME: consider instead `return nullptr`, given that we do nothing useful |
991 | // with integers, and so distinguishing them serves no purpose, but could |
992 | // prevent convergence. |
993 | CreatedValuesCount++; |
994 | return &arena().create<IntegerValue>(); |
995 | } |
996 | |
997 | if (Type->isPointerType()) { |
998 | CreatedValuesCount++; |
999 | QualType PointeeType = Type->getPointeeType(); |
1000 | StorageLocation &PointeeLoc = |
1001 | createLocAndMaybeValue(Ty: PointeeType, Visited, Depth, CreatedValuesCount); |
1002 | |
1003 | return &arena().create<PointerValue>(args&: PointeeLoc); |
1004 | } |
1005 | |
1006 | return nullptr; |
1007 | } |
1008 | |
1009 | StorageLocation & |
1010 | Environment::createLocAndMaybeValue(QualType Ty, |
1011 | llvm::DenseSet<QualType> &Visited, |
1012 | int Depth, int &CreatedValuesCount) { |
1013 | if (!Visited.insert(V: Ty.getCanonicalType()).second) |
1014 | return createStorageLocation(Type: Ty.getNonReferenceType()); |
1015 | auto EraseVisited = llvm::make_scope_exit( |
1016 | [&Visited, Ty] { Visited.erase(V: Ty.getCanonicalType()); }); |
1017 | |
1018 | Ty = Ty.getNonReferenceType(); |
1019 | |
1020 | if (Ty->isRecordType()) { |
1021 | auto &Loc = cast<RecordStorageLocation>(Val&: createStorageLocation(Type: Ty)); |
1022 | initializeFieldsWithValues(Loc, Type: Ty, Visited, Depth, CreatedValuesCount); |
1023 | return Loc; |
1024 | } |
1025 | |
1026 | StorageLocation &Loc = createStorageLocation(Type: Ty); |
1027 | |
1028 | if (Value *Val = createValueUnlessSelfReferential(Type: Ty, Visited, Depth, |
1029 | CreatedValuesCount)) |
1030 | setValue(Loc, Val&: *Val); |
1031 | |
1032 | return Loc; |
1033 | } |
1034 | |
1035 | void Environment::initializeFieldsWithValues(RecordStorageLocation &Loc, |
1036 | QualType Type, |
1037 | llvm::DenseSet<QualType> &Visited, |
1038 | int Depth, |
1039 | int &CreatedValuesCount) { |
1040 | auto initField = [&](QualType FieldType, StorageLocation &FieldLoc) { |
1041 | if (FieldType->isRecordType()) { |
1042 | auto &FieldRecordLoc = cast<RecordStorageLocation>(Val&: FieldLoc); |
1043 | initializeFieldsWithValues(FieldRecordLoc, FieldRecordLoc.getType(), |
1044 | Visited, Depth + 1, CreatedValuesCount); |
1045 | } else { |
1046 | if (getValue(Loc: FieldLoc) != nullptr) |
1047 | return; |
1048 | if (!Visited.insert(V: FieldType.getCanonicalType()).second) |
1049 | return; |
1050 | if (Value *Val = createValueUnlessSelfReferential( |
1051 | Type: FieldType, Visited, Depth: Depth + 1, CreatedValuesCount)) |
1052 | setValue(Loc: FieldLoc, Val&: *Val); |
1053 | Visited.erase(V: FieldType.getCanonicalType()); |
1054 | } |
1055 | }; |
1056 | |
1057 | for (const FieldDecl *Field : DACtx->getModeledFields(Type)) { |
1058 | assert(Field != nullptr); |
1059 | QualType FieldType = Field->getType(); |
1060 | |
1061 | if (FieldType->isReferenceType()) { |
1062 | Loc.setChild(*Field, |
1063 | &createLocAndMaybeValue(Ty: FieldType, Visited, Depth: Depth + 1, |
1064 | CreatedValuesCount)); |
1065 | } else { |
1066 | StorageLocation *FieldLoc = Loc.getChild(*Field); |
1067 | assert(FieldLoc != nullptr); |
1068 | initField(FieldType, *FieldLoc); |
1069 | } |
1070 | } |
1071 | for (const auto &[FieldName, FieldType] : DACtx->getSyntheticFields(Type)) { |
1072 | // Synthetic fields cannot have reference type, so we don't need to deal |
1073 | // with this case. |
1074 | assert(!FieldType->isReferenceType()); |
1075 | initField(FieldType, Loc.getSyntheticField(Name: FieldName)); |
1076 | } |
1077 | } |
1078 | |
1079 | StorageLocation &Environment::createObjectInternal(const ValueDecl *D, |
1080 | QualType Ty, |
1081 | const Expr *InitExpr) { |
1082 | if (Ty->isReferenceType()) { |
1083 | // Although variables of reference type always need to be initialized, it |
1084 | // can happen that we can't see the initializer, so `InitExpr` may still |
1085 | // be null. |
1086 | if (InitExpr) { |
1087 | if (auto *InitExprLoc = getStorageLocation(E: *InitExpr)) |
1088 | return *InitExprLoc; |
1089 | } |
1090 | |
1091 | // Even though we have an initializer, we might not get an |
1092 | // InitExprLoc, for example if the InitExpr is a CallExpr for which we |
1093 | // don't have a function body. In this case, we just invent a storage |
1094 | // location and value -- it's the best we can do. |
1095 | return createObjectInternal(D, Ty: Ty.getNonReferenceType(), InitExpr: nullptr); |
1096 | } |
1097 | |
1098 | StorageLocation &Loc = |
1099 | D ? createStorageLocation(D: *D) : createStorageLocation(Type: Ty); |
1100 | |
1101 | if (Ty->isRecordType()) { |
1102 | auto &RecordLoc = cast<RecordStorageLocation>(Val&: Loc); |
1103 | if (!InitExpr) |
1104 | initializeFieldsWithValues(Loc&: RecordLoc); |
1105 | } else { |
1106 | Value *Val = nullptr; |
1107 | if (InitExpr) |
1108 | // In the (few) cases where an expression is intentionally |
1109 | // "uninterpreted", `InitExpr` is not associated with a value. There are |
1110 | // two ways to handle this situation: propagate the status, so that |
1111 | // uninterpreted initializers result in uninterpreted variables, or |
1112 | // provide a default value. We choose the latter so that later refinements |
1113 | // of the variable can be used for reasoning about the surrounding code. |
1114 | // For this reason, we let this case be handled by the `createValue()` |
1115 | // call below. |
1116 | // |
1117 | // FIXME. If and when we interpret all language cases, change this to |
1118 | // assert that `InitExpr` is interpreted, rather than supplying a |
1119 | // default value (assuming we don't update the environment API to return |
1120 | // references). |
1121 | Val = getValue(E: *InitExpr); |
1122 | if (!Val) |
1123 | Val = createValue(Type: Ty); |
1124 | if (Val) |
1125 | setValue(Loc, Val&: *Val); |
1126 | } |
1127 | |
1128 | return Loc; |
1129 | } |
1130 | |
1131 | void Environment::assume(const Formula &F) { |
1132 | DACtx->addFlowConditionConstraint(Token: FlowConditionToken, Constraint: F); |
1133 | } |
1134 | |
1135 | bool Environment::proves(const Formula &F) const { |
1136 | return DACtx->flowConditionImplies(Token: FlowConditionToken, F); |
1137 | } |
1138 | |
1139 | bool Environment::allows(const Formula &F) const { |
1140 | return DACtx->flowConditionAllows(Token: FlowConditionToken, F); |
1141 | } |
1142 | |
1143 | void Environment::dump(raw_ostream &OS) const { |
1144 | llvm::DenseMap<const StorageLocation *, std::string> LocToName; |
1145 | if (LocForRecordReturnVal != nullptr) |
1146 | LocToName[LocForRecordReturnVal] = "(returned record)" ; |
1147 | if (ThisPointeeLoc != nullptr) |
1148 | LocToName[ThisPointeeLoc] = "this" ; |
1149 | |
1150 | OS << "DeclToLoc:\n" ; |
1151 | for (auto [D, L] : DeclToLoc) { |
1152 | auto Iter = LocToName.insert({L, D->getNameAsString()}).first; |
1153 | OS << " [" << Iter->second << ", " << L << "]\n" ; |
1154 | } |
1155 | OS << "ExprToLoc:\n" ; |
1156 | for (auto [E, L] : ExprToLoc) |
1157 | OS << " [" << E << ", " << L << "]\n" ; |
1158 | |
1159 | OS << "ExprToVal:\n" ; |
1160 | for (auto [E, V] : ExprToVal) |
1161 | OS << " [" << E << ", " << V << ": " << *V << "]\n" ; |
1162 | |
1163 | OS << "LocToVal:\n" ; |
1164 | for (auto [L, V] : LocToVal) { |
1165 | OS << " [" << L; |
1166 | if (auto Iter = LocToName.find(Val: L); Iter != LocToName.end()) |
1167 | OS << " (" << Iter->second << ")" ; |
1168 | OS << ", " << V << ": " << *V << "]\n" ; |
1169 | } |
1170 | |
1171 | if (const FunctionDecl *Func = getCurrentFunc()) { |
1172 | if (Func->getReturnType()->isReferenceType()) { |
1173 | OS << "ReturnLoc: " << ReturnLoc; |
1174 | if (auto Iter = LocToName.find(Val: ReturnLoc); Iter != LocToName.end()) |
1175 | OS << " (" << Iter->second << ")" ; |
1176 | OS << "\n" ; |
1177 | } else if (Func->getReturnType()->isRecordType() || |
1178 | isa<CXXConstructorDecl>(Val: Func)) { |
1179 | OS << "LocForRecordReturnVal: " << LocForRecordReturnVal << "\n" ; |
1180 | } else if (!Func->getReturnType()->isVoidType()) { |
1181 | if (ReturnVal == nullptr) |
1182 | OS << "ReturnVal: nullptr\n" ; |
1183 | else |
1184 | OS << "ReturnVal: " << *ReturnVal << "\n" ; |
1185 | } |
1186 | |
1187 | if (isa<CXXMethodDecl>(Val: Func)) { |
1188 | OS << "ThisPointeeLoc: " << ThisPointeeLoc << "\n" ; |
1189 | } |
1190 | } |
1191 | |
1192 | OS << "\n" ; |
1193 | DACtx->dumpFlowCondition(Token: FlowConditionToken, OS); |
1194 | } |
1195 | |
1196 | void Environment::dump() const { dump(OS&: llvm::dbgs()); } |
1197 | |
1198 | Environment::PrValueToResultObject Environment::buildResultObjectMap( |
1199 | DataflowAnalysisContext *DACtx, const FunctionDecl *FuncDecl, |
1200 | RecordStorageLocation *ThisPointeeLoc, |
1201 | RecordStorageLocation *LocForRecordReturnVal) { |
1202 | assert(FuncDecl->doesThisDeclarationHaveABody()); |
1203 | |
1204 | PrValueToResultObject Map; |
1205 | |
1206 | ResultObjectVisitor Visitor(Map, LocForRecordReturnVal, *DACtx); |
1207 | if (const auto *Ctor = dyn_cast<CXXConstructorDecl>(Val: FuncDecl)) |
1208 | Visitor.TraverseConstructorInits(Ctor, ThisPointeeLoc); |
1209 | Visitor.TraverseStmt(S: FuncDecl->getBody()); |
1210 | |
1211 | return Map; |
1212 | } |
1213 | |
1214 | RecordStorageLocation *getImplicitObjectLocation(const CXXMemberCallExpr &MCE, |
1215 | const Environment &Env) { |
1216 | Expr *ImplicitObject = MCE.getImplicitObjectArgument(); |
1217 | if (ImplicitObject == nullptr) |
1218 | return nullptr; |
1219 | if (ImplicitObject->getType()->isPointerType()) { |
1220 | if (auto *Val = Env.get<PointerValue>(E: *ImplicitObject)) |
1221 | return &cast<RecordStorageLocation>(Val&: Val->getPointeeLoc()); |
1222 | return nullptr; |
1223 | } |
1224 | return cast_or_null<RecordStorageLocation>( |
1225 | Val: Env.getStorageLocation(E: *ImplicitObject)); |
1226 | } |
1227 | |
1228 | RecordStorageLocation *getBaseObjectLocation(const MemberExpr &ME, |
1229 | const Environment &Env) { |
1230 | Expr *Base = ME.getBase(); |
1231 | if (Base == nullptr) |
1232 | return nullptr; |
1233 | if (ME.isArrow()) { |
1234 | if (auto *Val = Env.get<PointerValue>(E: *Base)) |
1235 | return &cast<RecordStorageLocation>(Val&: Val->getPointeeLoc()); |
1236 | return nullptr; |
1237 | } |
1238 | return Env.get<RecordStorageLocation>(E: *Base); |
1239 | } |
1240 | |
1241 | } // namespace dataflow |
1242 | } // namespace clang |
1243 | |