1 //==- UninitializedValues.cpp - Find Uninitialized Values -------*- C++ --*-==//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements uninitialized values analysis for source-level CFGs.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "clang/AST/ASTContext.h"
15 #include "clang/AST/Attr.h"
16 #include "clang/AST/Decl.h"
17 #include "clang/Analysis/Analyses/PostOrderCFGView.h"
18 #include "clang/Analysis/Analyses/UninitializedValues.h"
19 #include "clang/Analysis/AnalysisContext.h"
20 #include "clang/Analysis/CFG.h"
21 #include "clang/Analysis/DomainSpecific/ObjCNoReturn.h"
22 #include "clang/Analysis/Visitors/CFGRecStmtDeclVisitor.h"
23 #include "llvm/ADT/DenseMap.h"
24 #include "llvm/ADT/Optional.h"
25 #include "llvm/ADT/PackedVector.h"
26 #include "llvm/ADT/SmallBitVector.h"
27 #include "llvm/ADT/SmallVector.h"
28 #include "llvm/Support/SaveAndRestore.h"
29 #include <utility>
30 
31 using namespace clang;
32 
33 #define DEBUG_LOGGING 0
34 
35 static bool isTrackedVar(const VarDecl *vd, const DeclContext *dc) {
36   if (vd->isLocalVarDecl() && !vd->hasGlobalStorage() &&
37       !vd->isExceptionVariable() &&
38       vd->getDeclContext() == dc) {
39     QualType ty = vd->getType();
40     return ty->isScalarType() || ty->isVectorType();
41   }
42   return false;
43 }
44 
45 //------------------------------------------------------------------------====//
46 // DeclToIndex: a mapping from Decls we track to value indices.
47 //====------------------------------------------------------------------------//
48 
49 namespace {
50 class DeclToIndex {
51   llvm::DenseMap<const VarDecl *, unsigned> map;
52 public:
53   DeclToIndex() {}
54 
55   /// Compute the actual mapping from declarations to bits.
56   void computeMap(const DeclContext &dc);
57 
58   /// Return the number of declarations in the map.
59   unsigned size() const { return map.size(); }
60 
61   /// Returns the bit vector index for a given declaration.
62   llvm::Optional<unsigned> getValueIndex(const VarDecl *d) const;
63 };
64 }
65 
66 void DeclToIndex::computeMap(const DeclContext &dc) {
67   unsigned count = 0;
68   DeclContext::specific_decl_iterator<VarDecl> I(dc.decls_begin()),
69                                                E(dc.decls_end());
70   for ( ; I != E; ++I) {
71     const VarDecl *vd = *I;
72     if (isTrackedVar(vd, &dc))
73       map[vd] = count++;
74   }
75 }
76 
77 llvm::Optional<unsigned> DeclToIndex::getValueIndex(const VarDecl *d) const {
78   llvm::DenseMap<const VarDecl *, unsigned>::const_iterator I = map.find(d);
79   if (I == map.end())
80     return llvm::Optional<unsigned>();
81   return I->second;
82 }
83 
84 //------------------------------------------------------------------------====//
85 // CFGBlockValues: dataflow values for CFG blocks.
86 //====------------------------------------------------------------------------//
87 
88 // These values are defined in such a way that a merge can be done using
89 // a bitwise OR.
90 enum Value { Unknown = 0x0,         /* 00 */
91              Initialized = 0x1,     /* 01 */
92              Uninitialized = 0x2,   /* 10 */
93              MayUninitialized = 0x3 /* 11 */ };
94 
95 static bool isUninitialized(const Value v) {
96   return v >= Uninitialized;
97 }
98 static bool isAlwaysUninit(const Value v) {
99   return v == Uninitialized;
100 }
101 
102 namespace {
103 
104 typedef llvm::PackedVector<Value, 2, llvm::SmallBitVector> ValueVector;
105 
106 class CFGBlockValues {
107   const CFG &cfg;
108   SmallVector<ValueVector, 8> vals;
109   ValueVector scratch;
110   DeclToIndex declToIndex;
111 public:
112   CFGBlockValues(const CFG &cfg);
113 
114   unsigned getNumEntries() const { return declToIndex.size(); }
115 
116   void computeSetOfDeclarations(const DeclContext &dc);
117   ValueVector &getValueVector(const CFGBlock *block) {
118     return vals[block->getBlockID()];
119   }
120 
121   void setAllScratchValues(Value V);
122   void mergeIntoScratch(ValueVector const &source, bool isFirst);
123   bool updateValueVectorWithScratch(const CFGBlock *block);
124 
125   bool hasNoDeclarations() const {
126     return declToIndex.size() == 0;
127   }
128 
129   void resetScratch();
130 
131   ValueVector::reference operator[](const VarDecl *vd);
132 
133   Value getValue(const CFGBlock *block, const CFGBlock *dstBlock,
134                  const VarDecl *vd) {
135     const llvm::Optional<unsigned> &idx = declToIndex.getValueIndex(vd);
136     assert(idx.hasValue());
137     return getValueVector(block)[idx.getValue()];
138   }
139 };
140 } // end anonymous namespace
141 
142 CFGBlockValues::CFGBlockValues(const CFG &c) : cfg(c), vals(0) {}
143 
144 void CFGBlockValues::computeSetOfDeclarations(const DeclContext &dc) {
145   declToIndex.computeMap(dc);
146   unsigned decls = declToIndex.size();
147   scratch.resize(decls);
148   unsigned n = cfg.getNumBlockIDs();
149   if (!n)
150     return;
151   vals.resize(n);
152   for (unsigned i = 0; i < n; ++i)
153     vals[i].resize(decls);
154 }
155 
156 #if DEBUG_LOGGING
157 static void printVector(const CFGBlock *block, ValueVector &bv,
158                         unsigned num) {
159   llvm::errs() << block->getBlockID() << " :";
160   for (unsigned i = 0; i < bv.size(); ++i) {
161     llvm::errs() << ' ' << bv[i];
162   }
163   llvm::errs() << " : " << num << '\n';
164 }
165 #endif
166 
167 void CFGBlockValues::setAllScratchValues(Value V) {
168   for (unsigned I = 0, E = scratch.size(); I != E; ++I)
169     scratch[I] = V;
170 }
171 
172 void CFGBlockValues::mergeIntoScratch(ValueVector const &source,
173                                       bool isFirst) {
174   if (isFirst)
175     scratch = source;
176   else
177     scratch |= source;
178 }
179 
180 bool CFGBlockValues::updateValueVectorWithScratch(const CFGBlock *block) {
181   ValueVector &dst = getValueVector(block);
182   bool changed = (dst != scratch);
183   if (changed)
184     dst = scratch;
185 #if DEBUG_LOGGING
186   printVector(block, scratch, 0);
187 #endif
188   return changed;
189 }
190 
191 void CFGBlockValues::resetScratch() {
192   scratch.reset();
193 }
194 
195 ValueVector::reference CFGBlockValues::operator[](const VarDecl *vd) {
196   const llvm::Optional<unsigned> &idx = declToIndex.getValueIndex(vd);
197   assert(idx.hasValue());
198   return scratch[idx.getValue()];
199 }
200 
201 //------------------------------------------------------------------------====//
202 // Worklist: worklist for dataflow analysis.
203 //====------------------------------------------------------------------------//
204 
205 namespace {
206 class DataflowWorklist {
207   PostOrderCFGView::iterator PO_I, PO_E;
208   SmallVector<const CFGBlock *, 20> worklist;
209   llvm::BitVector enqueuedBlocks;
210 public:
211   DataflowWorklist(const CFG &cfg, PostOrderCFGView &view)
212     : PO_I(view.begin()), PO_E(view.end()),
213       enqueuedBlocks(cfg.getNumBlockIDs(), true) {
214         // Treat the first block as already analyzed.
215         if (PO_I != PO_E) {
216           assert(*PO_I == &cfg.getEntry());
217           enqueuedBlocks[(*PO_I)->getBlockID()] = false;
218           ++PO_I;
219         }
220       }
221 
222   void enqueueSuccessors(const CFGBlock *block);
223   const CFGBlock *dequeue();
224 };
225 }
226 
227 void DataflowWorklist::enqueueSuccessors(const clang::CFGBlock *block) {
228   for (CFGBlock::const_succ_iterator I = block->succ_begin(),
229        E = block->succ_end(); I != E; ++I) {
230     const CFGBlock *Successor = *I;
231     if (!Successor || enqueuedBlocks[Successor->getBlockID()])
232       continue;
233     worklist.push_back(Successor);
234     enqueuedBlocks[Successor->getBlockID()] = true;
235   }
236 }
237 
238 const CFGBlock *DataflowWorklist::dequeue() {
239   const CFGBlock *B = 0;
240 
241   // First dequeue from the worklist.  This can represent
242   // updates along backedges that we want propagated as quickly as possible.
243   if (!worklist.empty()) {
244     B = worklist.back();
245     worklist.pop_back();
246   }
247   // Next dequeue from the initial reverse post order.  This is the
248   // theoretical ideal in the presence of no back edges.
249   else if (PO_I != PO_E) {
250     B = *PO_I;
251     ++PO_I;
252   }
253   else {
254     return 0;
255   }
256 
257   assert(enqueuedBlocks[B->getBlockID()] == true);
258   enqueuedBlocks[B->getBlockID()] = false;
259   return B;
260 }
261 
262 //------------------------------------------------------------------------====//
263 // Classification of DeclRefExprs as use or initialization.
264 //====------------------------------------------------------------------------//
265 
266 namespace {
267 class FindVarResult {
268   const VarDecl *vd;
269   const DeclRefExpr *dr;
270 public:
271   FindVarResult(const VarDecl *vd, const DeclRefExpr *dr) : vd(vd), dr(dr) {}
272 
273   const DeclRefExpr *getDeclRefExpr() const { return dr; }
274   const VarDecl *getDecl() const { return vd; }
275 };
276 
277 static const Expr *stripCasts(ASTContext &C, const Expr *Ex) {
278   while (Ex) {
279     Ex = Ex->IgnoreParenNoopCasts(C);
280     if (const CastExpr *CE = dyn_cast<CastExpr>(Ex)) {
281       if (CE->getCastKind() == CK_LValueBitCast) {
282         Ex = CE->getSubExpr();
283         continue;
284       }
285     }
286     break;
287   }
288   return Ex;
289 }
290 
291 /// If E is an expression comprising a reference to a single variable, find that
292 /// variable.
293 static FindVarResult findVar(const Expr *E, const DeclContext *DC) {
294   if (const DeclRefExpr *DRE =
295         dyn_cast<DeclRefExpr>(stripCasts(DC->getParentASTContext(), E)))
296     if (const VarDecl *VD = dyn_cast<VarDecl>(DRE->getDecl()))
297       if (isTrackedVar(VD, DC))
298         return FindVarResult(VD, DRE);
299   return FindVarResult(0, 0);
300 }
301 
302 /// \brief Classify each DeclRefExpr as an initialization or a use. Any
303 /// DeclRefExpr which isn't explicitly classified will be assumed to have
304 /// escaped the analysis and will be treated as an initialization.
305 class ClassifyRefs : public StmtVisitor<ClassifyRefs> {
306 public:
307   enum Class {
308     Init,
309     Use,
310     SelfInit,
311     Ignore
312   };
313 
314 private:
315   const DeclContext *DC;
316   llvm::DenseMap<const DeclRefExpr*, Class> Classification;
317 
318   bool isTrackedVar(const VarDecl *VD) const {
319     return ::isTrackedVar(VD, DC);
320   }
321 
322   void classify(const Expr *E, Class C);
323 
324 public:
325   ClassifyRefs(AnalysisDeclContext &AC) : DC(cast<DeclContext>(AC.getDecl())) {}
326 
327   void VisitDeclStmt(DeclStmt *DS);
328   void VisitUnaryOperator(UnaryOperator *UO);
329   void VisitBinaryOperator(BinaryOperator *BO);
330   void VisitCallExpr(CallExpr *CE);
331   void VisitCastExpr(CastExpr *CE);
332 
333   void operator()(Stmt *S) { Visit(S); }
334 
335   Class get(const DeclRefExpr *DRE) const {
336     llvm::DenseMap<const DeclRefExpr*, Class>::const_iterator I
337         = Classification.find(DRE);
338     if (I != Classification.end())
339       return I->second;
340 
341     const VarDecl *VD = dyn_cast<VarDecl>(DRE->getDecl());
342     if (!VD || !isTrackedVar(VD))
343       return Ignore;
344 
345     return Init;
346   }
347 };
348 }
349 
350 static const DeclRefExpr *getSelfInitExpr(VarDecl *VD) {
351   if (Expr *Init = VD->getInit()) {
352     const DeclRefExpr *DRE
353       = dyn_cast<DeclRefExpr>(stripCasts(VD->getASTContext(), Init));
354     if (DRE && DRE->getDecl() == VD)
355       return DRE;
356   }
357   return 0;
358 }
359 
360 void ClassifyRefs::classify(const Expr *E, Class C) {
361   FindVarResult Var = findVar(E, DC);
362   if (const DeclRefExpr *DRE = Var.getDeclRefExpr())
363     Classification[DRE] = std::max(Classification[DRE], C);
364 }
365 
366 void ClassifyRefs::VisitDeclStmt(DeclStmt *DS) {
367   for (DeclStmt::decl_iterator DI = DS->decl_begin(), DE = DS->decl_end();
368        DI != DE; ++DI) {
369     VarDecl *VD = dyn_cast<VarDecl>(*DI);
370     if (VD && isTrackedVar(VD))
371       if (const DeclRefExpr *DRE = getSelfInitExpr(VD))
372         Classification[DRE] = SelfInit;
373   }
374 }
375 
376 void ClassifyRefs::VisitBinaryOperator(BinaryOperator *BO) {
377   // Ignore the evaluation of a DeclRefExpr on the LHS of an assignment. If this
378   // is not a compound-assignment, we will treat it as initializing the variable
379   // when TransferFunctions visits it. A compound-assignment does not affect
380   // whether a variable is uninitialized, and there's no point counting it as a
381   // use.
382   if (BO->isCompoundAssignmentOp())
383     classify(BO->getLHS(), Use);
384   else if (BO->getOpcode() == BO_Assign)
385     classify(BO->getLHS(), Ignore);
386 }
387 
388 void ClassifyRefs::VisitUnaryOperator(UnaryOperator *UO) {
389   // Increment and decrement are uses despite there being no lvalue-to-rvalue
390   // conversion.
391   if (UO->isIncrementDecrementOp())
392     classify(UO->getSubExpr(), Use);
393 }
394 
395 void ClassifyRefs::VisitCallExpr(CallExpr *CE) {
396   // If a value is passed by const reference to a function, we should not assume
397   // that it is initialized by the call, and we conservatively do not assume
398   // that it is used.
399   for (CallExpr::arg_iterator I = CE->arg_begin(), E = CE->arg_end();
400        I != E; ++I)
401     if ((*I)->getType().isConstQualified() && (*I)->isGLValue())
402       classify(*I, Ignore);
403 }
404 
405 void ClassifyRefs::VisitCastExpr(CastExpr *CE) {
406   if (CE->getCastKind() == CK_LValueToRValue)
407     classify(CE->getSubExpr(), Use);
408   else if (CStyleCastExpr *CSE = dyn_cast<CStyleCastExpr>(CE)) {
409     if (CSE->getType()->isVoidType()) {
410       // Squelch any detected load of an uninitialized value if
411       // we cast it to void.
412       // e.g. (void) x;
413       classify(CSE->getSubExpr(), Ignore);
414     }
415   }
416 }
417 
418 //------------------------------------------------------------------------====//
419 // Transfer function for uninitialized values analysis.
420 //====------------------------------------------------------------------------//
421 
422 namespace {
423 class TransferFunctions : public StmtVisitor<TransferFunctions> {
424   CFGBlockValues &vals;
425   const CFG &cfg;
426   const CFGBlock *block;
427   AnalysisDeclContext &ac;
428   const ClassifyRefs &classification;
429   ObjCNoReturn objCNoRet;
430   UninitVariablesHandler &handler;
431 
432 public:
433   TransferFunctions(CFGBlockValues &vals, const CFG &cfg,
434                     const CFGBlock *block, AnalysisDeclContext &ac,
435                     const ClassifyRefs &classification,
436                     UninitVariablesHandler &handler)
437     : vals(vals), cfg(cfg), block(block), ac(ac),
438       classification(classification), objCNoRet(ac.getASTContext()),
439       handler(handler) {}
440 
441   void reportUse(const Expr *ex, const VarDecl *vd);
442 
443   void VisitBinaryOperator(BinaryOperator *bo);
444   void VisitBlockExpr(BlockExpr *be);
445   void VisitCallExpr(CallExpr *ce);
446   void VisitDeclRefExpr(DeclRefExpr *dr);
447   void VisitDeclStmt(DeclStmt *ds);
448   void VisitObjCForCollectionStmt(ObjCForCollectionStmt *FS);
449   void VisitObjCMessageExpr(ObjCMessageExpr *ME);
450 
451   bool isTrackedVar(const VarDecl *vd) {
452     return ::isTrackedVar(vd, cast<DeclContext>(ac.getDecl()));
453   }
454 
455   FindVarResult findVar(const Expr *ex) {
456     return ::findVar(ex, cast<DeclContext>(ac.getDecl()));
457   }
458 
459   UninitUse getUninitUse(const Expr *ex, const VarDecl *vd, Value v) {
460     UninitUse Use(ex, isAlwaysUninit(v));
461 
462     assert(isUninitialized(v));
463     if (Use.getKind() == UninitUse::Always)
464       return Use;
465 
466     // If an edge which leads unconditionally to this use did not initialize
467     // the variable, we can say something stronger than 'may be uninitialized':
468     // we can say 'either it's used uninitialized or you have dead code'.
469     //
470     // We track the number of successors of a node which have been visited, and
471     // visit a node once we have visited all of its successors. Only edges where
472     // the variable might still be uninitialized are followed. Since a variable
473     // can't transfer from being initialized to being uninitialized, this will
474     // trace out the subgraph which inevitably leads to the use and does not
475     // initialize the variable. We do not want to skip past loops, since their
476     // non-termination might be correlated with the initialization condition.
477     //
478     // For example:
479     //
480     //         void f(bool a, bool b) {
481     // block1:   int n;
482     //           if (a) {
483     // block2:     if (b)
484     // block3:       n = 1;
485     // block4:   } else if (b) {
486     // block5:     while (!a) {
487     // block6:       do_work(&a);
488     //               n = 2;
489     //             }
490     //           }
491     // block7:   if (a)
492     // block8:     g();
493     // block9:   return n;
494     //         }
495     //
496     // Starting from the maybe-uninitialized use in block 9:
497     //  * Block 7 is not visited because we have only visited one of its two
498     //    successors.
499     //  * Block 8 is visited because we've visited its only successor.
500     // From block 8:
501     //  * Block 7 is visited because we've now visited both of its successors.
502     // From block 7:
503     //  * Blocks 1, 2, 4, 5, and 6 are not visited because we didn't visit all
504     //    of their successors (we didn't visit 4, 3, 5, 6, and 5, respectively).
505     //  * Block 3 is not visited because it initializes 'n'.
506     // Now the algorithm terminates, having visited blocks 7 and 8, and having
507     // found the frontier is blocks 2, 4, and 5.
508     //
509     // 'n' is definitely uninitialized for two edges into block 7 (from blocks 2
510     // and 4), so we report that any time either of those edges is taken (in
511     // each case when 'b == false'), 'n' is used uninitialized.
512     SmallVector<const CFGBlock*, 32> Queue;
513     SmallVector<unsigned, 32> SuccsVisited(cfg.getNumBlockIDs(), 0);
514     Queue.push_back(block);
515     // Specify that we've already visited all successors of the starting block.
516     // This has the dual purpose of ensuring we never add it to the queue, and
517     // of marking it as not being a candidate element of the frontier.
518     SuccsVisited[block->getBlockID()] = block->succ_size();
519     while (!Queue.empty()) {
520       const CFGBlock *B = Queue.back();
521       Queue.pop_back();
522       for (CFGBlock::const_pred_iterator I = B->pred_begin(), E = B->pred_end();
523            I != E; ++I) {
524         const CFGBlock *Pred = *I;
525         if (vals.getValue(Pred, B, vd) == Initialized)
526           // This block initializes the variable.
527           continue;
528 
529         unsigned &SV = SuccsVisited[Pred->getBlockID()];
530         if (!SV) {
531           // When visiting the first successor of a block, mark all NULL
532           // successors as having been visited.
533           for (CFGBlock::const_succ_iterator SI = Pred->succ_begin(),
534                                              SE = Pred->succ_end();
535                SI != SE; ++SI)
536             if (!*SI)
537               ++SV;
538         }
539 
540         if (++SV == Pred->succ_size())
541           // All paths from this block lead to the use and don't initialize the
542           // variable.
543           Queue.push_back(Pred);
544       }
545     }
546 
547     // Scan the frontier, looking for blocks where the variable was
548     // uninitialized.
549     for (CFG::const_iterator BI = cfg.begin(), BE = cfg.end(); BI != BE; ++BI) {
550       const CFGBlock *Block = *BI;
551       unsigned BlockID = Block->getBlockID();
552       const Stmt *Term = Block->getTerminator();
553       if (SuccsVisited[BlockID] && SuccsVisited[BlockID] < Block->succ_size() &&
554           Term) {
555         // This block inevitably leads to the use. If we have an edge from here
556         // to a post-dominator block, and the variable is uninitialized on that
557         // edge, we have found a bug.
558         for (CFGBlock::const_succ_iterator I = Block->succ_begin(),
559              E = Block->succ_end(); I != E; ++I) {
560           const CFGBlock *Succ = *I;
561           if (Succ && SuccsVisited[Succ->getBlockID()] >= Succ->succ_size() &&
562               vals.getValue(Block, Succ, vd) == Uninitialized) {
563             // Switch cases are a special case: report the label to the caller
564             // as the 'terminator', not the switch statement itself. Suppress
565             // situations where no label matched: we can't be sure that's
566             // possible.
567             if (isa<SwitchStmt>(Term)) {
568               const Stmt *Label = Succ->getLabel();
569               if (!Label || !isa<SwitchCase>(Label))
570                 // Might not be possible.
571                 continue;
572               UninitUse::Branch Branch;
573               Branch.Terminator = Label;
574               Branch.Output = 0; // Ignored.
575               Use.addUninitBranch(Branch);
576             } else {
577               UninitUse::Branch Branch;
578               Branch.Terminator = Term;
579               Branch.Output = I - Block->succ_begin();
580               Use.addUninitBranch(Branch);
581             }
582           }
583         }
584       }
585     }
586 
587     return Use;
588   }
589 };
590 }
591 
592 void TransferFunctions::reportUse(const Expr *ex, const VarDecl *vd) {
593   Value v = vals[vd];
594   if (isUninitialized(v))
595     handler.handleUseOfUninitVariable(vd, getUninitUse(ex, vd, v));
596 }
597 
598 void TransferFunctions::VisitObjCForCollectionStmt(ObjCForCollectionStmt *FS) {
599   // This represents an initialization of the 'element' value.
600   if (DeclStmt *DS = dyn_cast<DeclStmt>(FS->getElement())) {
601     const VarDecl *VD = cast<VarDecl>(DS->getSingleDecl());
602     if (isTrackedVar(VD))
603       vals[VD] = Initialized;
604   }
605 }
606 
607 void TransferFunctions::VisitBlockExpr(BlockExpr *be) {
608   const BlockDecl *bd = be->getBlockDecl();
609   for (BlockDecl::capture_const_iterator i = bd->capture_begin(),
610         e = bd->capture_end() ; i != e; ++i) {
611     const VarDecl *vd = i->getVariable();
612     if (!isTrackedVar(vd))
613       continue;
614     if (i->isByRef()) {
615       vals[vd] = Initialized;
616       continue;
617     }
618     reportUse(be, vd);
619   }
620 }
621 
622 void TransferFunctions::VisitCallExpr(CallExpr *ce) {
623   if (Decl *Callee = ce->getCalleeDecl()) {
624     if (Callee->hasAttr<ReturnsTwiceAttr>()) {
625       // After a call to a function like setjmp or vfork, any variable which is
626       // initialized anywhere within this function may now be initialized. For
627       // now, just assume such a call initializes all variables.  FIXME: Only
628       // mark variables as initialized if they have an initializer which is
629       // reachable from here.
630       vals.setAllScratchValues(Initialized);
631     }
632     else if (Callee->hasAttr<AnalyzerNoReturnAttr>()) {
633       // Functions labeled like "analyzer_noreturn" are often used to denote
634       // "panic" functions that in special debug situations can still return,
635       // but for the most part should not be treated as returning.  This is a
636       // useful annotation borrowed from the static analyzer that is useful for
637       // suppressing branch-specific false positives when we call one of these
638       // functions but keep pretending the path continues (when in reality the
639       // user doesn't care).
640       vals.setAllScratchValues(Unknown);
641     }
642   }
643 }
644 
645 void TransferFunctions::VisitDeclRefExpr(DeclRefExpr *dr) {
646   switch (classification.get(dr)) {
647   case ClassifyRefs::Ignore:
648     break;
649   case ClassifyRefs::Use:
650     reportUse(dr, cast<VarDecl>(dr->getDecl()));
651     break;
652   case ClassifyRefs::Init:
653     vals[cast<VarDecl>(dr->getDecl())] = Initialized;
654     break;
655   case ClassifyRefs::SelfInit:
656       handler.handleSelfInit(cast<VarDecl>(dr->getDecl()));
657     break;
658   }
659 }
660 
661 void TransferFunctions::VisitBinaryOperator(BinaryOperator *BO) {
662   if (BO->getOpcode() == BO_Assign) {
663     FindVarResult Var = findVar(BO->getLHS());
664     if (const VarDecl *VD = Var.getDecl())
665       vals[VD] = Initialized;
666   }
667 }
668 
669 void TransferFunctions::VisitDeclStmt(DeclStmt *DS) {
670   for (DeclStmt::decl_iterator DI = DS->decl_begin(), DE = DS->decl_end();
671        DI != DE; ++DI) {
672     VarDecl *VD = dyn_cast<VarDecl>(*DI);
673     if (VD && isTrackedVar(VD)) {
674       if (getSelfInitExpr(VD)) {
675         // If the initializer consists solely of a reference to itself, we
676         // explicitly mark the variable as uninitialized. This allows code
677         // like the following:
678         //
679         //   int x = x;
680         //
681         // to deliberately leave a variable uninitialized. Different analysis
682         // clients can detect this pattern and adjust their reporting
683         // appropriately, but we need to continue to analyze subsequent uses
684         // of the variable.
685         vals[VD] = Uninitialized;
686       } else if (VD->getInit()) {
687         // Treat the new variable as initialized.
688         vals[VD] = Initialized;
689       } else {
690         // No initializer: the variable is now uninitialized. This matters
691         // for cases like:
692         //   while (...) {
693         //     int n;
694         //     use(n);
695         //     n = 0;
696         //   }
697         // FIXME: Mark the variable as uninitialized whenever its scope is
698         // left, since its scope could be re-entered by a jump over the
699         // declaration.
700         vals[VD] = Uninitialized;
701       }
702     }
703   }
704 }
705 
706 void TransferFunctions::VisitObjCMessageExpr(ObjCMessageExpr *ME) {
707   // If the Objective-C message expression is an implicit no-return that
708   // is not modeled in the CFG, set the tracked dataflow values to Unknown.
709   if (objCNoRet.isImplicitNoReturn(ME)) {
710     vals.setAllScratchValues(Unknown);
711   }
712 }
713 
714 //------------------------------------------------------------------------====//
715 // High-level "driver" logic for uninitialized values analysis.
716 //====------------------------------------------------------------------------//
717 
718 static bool runOnBlock(const CFGBlock *block, const CFG &cfg,
719                        AnalysisDeclContext &ac, CFGBlockValues &vals,
720                        const ClassifyRefs &classification,
721                        llvm::BitVector &wasAnalyzed,
722                        UninitVariablesHandler &handler) {
723   wasAnalyzed[block->getBlockID()] = true;
724   vals.resetScratch();
725   // Merge in values of predecessor blocks.
726   bool isFirst = true;
727   for (CFGBlock::const_pred_iterator I = block->pred_begin(),
728        E = block->pred_end(); I != E; ++I) {
729     const CFGBlock *pred = *I;
730     if (wasAnalyzed[pred->getBlockID()]) {
731       vals.mergeIntoScratch(vals.getValueVector(pred), isFirst);
732       isFirst = false;
733     }
734   }
735   // Apply the transfer function.
736   TransferFunctions tf(vals, cfg, block, ac, classification, handler);
737   for (CFGBlock::const_iterator I = block->begin(), E = block->end();
738        I != E; ++I) {
739     if (const CFGStmt *cs = dyn_cast<CFGStmt>(&*I)) {
740       tf.Visit(const_cast<Stmt*>(cs->getStmt()));
741     }
742   }
743   return vals.updateValueVectorWithScratch(block);
744 }
745 
746 /// PruneBlocksHandler is a special UninitVariablesHandler that is used
747 /// to detect when a CFGBlock has any *potential* use of an uninitialized
748 /// variable.  It is mainly used to prune out work during the final
749 /// reporting pass.
750 namespace {
751 struct PruneBlocksHandler : public UninitVariablesHandler {
752   PruneBlocksHandler(unsigned numBlocks)
753     : hadUse(numBlocks, false), hadAnyUse(false),
754       currentBlock(0) {}
755 
756   virtual ~PruneBlocksHandler() {}
757 
758   /// Records if a CFGBlock had a potential use of an uninitialized variable.
759   llvm::BitVector hadUse;
760 
761   /// Records if any CFGBlock had a potential use of an uninitialized variable.
762   bool hadAnyUse;
763 
764   /// The current block to scribble use information.
765   unsigned currentBlock;
766 
767   virtual void handleUseOfUninitVariable(const VarDecl *vd,
768                                          const UninitUse &use) {
769     hadUse[currentBlock] = true;
770     hadAnyUse = true;
771   }
772 
773   /// Called when the uninitialized variable analysis detects the
774   /// idiom 'int x = x'.  All other uses of 'x' within the initializer
775   /// are handled by handleUseOfUninitVariable.
776   virtual void handleSelfInit(const VarDecl *vd) {
777     hadUse[currentBlock] = true;
778     hadAnyUse = true;
779   }
780 };
781 }
782 
783 void clang::runUninitializedVariablesAnalysis(
784     const DeclContext &dc,
785     const CFG &cfg,
786     AnalysisDeclContext &ac,
787     UninitVariablesHandler &handler,
788     UninitVariablesAnalysisStats &stats) {
789   CFGBlockValues vals(cfg);
790   vals.computeSetOfDeclarations(dc);
791   if (vals.hasNoDeclarations())
792     return;
793 
794   stats.NumVariablesAnalyzed = vals.getNumEntries();
795 
796   // Precompute which expressions are uses and which are initializations.
797   ClassifyRefs classification(ac);
798   cfg.VisitBlockStmts(classification);
799 
800   // Mark all variables uninitialized at the entry.
801   const CFGBlock &entry = cfg.getEntry();
802   ValueVector &vec = vals.getValueVector(&entry);
803   const unsigned n = vals.getNumEntries();
804   for (unsigned j = 0; j < n ; ++j) {
805     vec[j] = Uninitialized;
806   }
807 
808   // Proceed with the workist.
809   DataflowWorklist worklist(cfg, *ac.getAnalysis<PostOrderCFGView>());
810   llvm::BitVector previouslyVisited(cfg.getNumBlockIDs());
811   worklist.enqueueSuccessors(&cfg.getEntry());
812   llvm::BitVector wasAnalyzed(cfg.getNumBlockIDs(), false);
813   wasAnalyzed[cfg.getEntry().getBlockID()] = true;
814   PruneBlocksHandler PBH(cfg.getNumBlockIDs());
815 
816   while (const CFGBlock *block = worklist.dequeue()) {
817     PBH.currentBlock = block->getBlockID();
818 
819     // Did the block change?
820     bool changed = runOnBlock(block, cfg, ac, vals,
821                               classification, wasAnalyzed, PBH);
822     ++stats.NumBlockVisits;
823     if (changed || !previouslyVisited[block->getBlockID()])
824       worklist.enqueueSuccessors(block);
825     previouslyVisited[block->getBlockID()] = true;
826   }
827 
828   if (!PBH.hadAnyUse)
829     return;
830 
831   // Run through the blocks one more time, and report uninitialized variables.
832   for (CFG::const_iterator BI = cfg.begin(), BE = cfg.end(); BI != BE; ++BI) {
833     const CFGBlock *block = *BI;
834     if (PBH.hadUse[block->getBlockID()]) {
835       runOnBlock(block, cfg, ac, vals, classification, wasAnalyzed, handler);
836       ++stats.NumBlockVisits;
837     }
838   }
839 }
840 
841 UninitVariablesHandler::~UninitVariablesHandler() {}
842