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/AST/StmtVisitor.h"
18 #include "clang/Analysis/Analyses/PostOrderCFGView.h"
19 #include "clang/Analysis/Analyses/UninitializedValues.h"
20 #include "clang/Analysis/AnalysisContext.h"
21 #include "clang/Analysis/CFG.h"
22 #include "clang/Analysis/DomainSpecific/ObjCNoReturn.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   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 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 None;
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 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 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.pop_back_val();
245 
246   // Next dequeue from the initial reverse post order.  This is the
247   // theoretical ideal in the presence of no back edges.
248   else if (PO_I != PO_E) {
249     B = *PO_I;
250     ++PO_I;
251   }
252   else {
253     return 0;
254   }
255 
256   assert(enqueuedBlocks[B->getBlockID()] == true);
257   enqueuedBlocks[B->getBlockID()] = false;
258   return B;
259 }
260 
261 //------------------------------------------------------------------------====//
262 // Classification of DeclRefExprs as use or initialization.
263 //====------------------------------------------------------------------------//
264 
265 namespace {
266 class FindVarResult {
267   const VarDecl *vd;
268   const DeclRefExpr *dr;
269 public:
270   FindVarResult(const VarDecl *vd, const DeclRefExpr *dr) : vd(vd), dr(dr) {}
271 
272   const DeclRefExpr *getDeclRefExpr() const { return dr; }
273   const VarDecl *getDecl() const { return vd; }
274 };
275 
276 static const Expr *stripCasts(ASTContext &C, const Expr *Ex) {
277   while (Ex) {
278     Ex = Ex->IgnoreParenNoopCasts(C);
279     if (const CastExpr *CE = dyn_cast<CastExpr>(Ex)) {
280       if (CE->getCastKind() == CK_LValueBitCast) {
281         Ex = CE->getSubExpr();
282         continue;
283       }
284     }
285     break;
286   }
287   return Ex;
288 }
289 
290 /// If E is an expression comprising a reference to a single variable, find that
291 /// variable.
292 static FindVarResult findVar(const Expr *E, const DeclContext *DC) {
293   if (const DeclRefExpr *DRE =
294         dyn_cast<DeclRefExpr>(stripCasts(DC->getParentASTContext(), E)))
295     if (const VarDecl *VD = dyn_cast<VarDecl>(DRE->getDecl()))
296       if (isTrackedVar(VD, DC))
297         return FindVarResult(VD, DRE);
298   return FindVarResult(0, 0);
299 }
300 
301 /// \brief Classify each DeclRefExpr as an initialization or a use. Any
302 /// DeclRefExpr which isn't explicitly classified will be assumed to have
303 /// escaped the analysis and will be treated as an initialization.
304 class ClassifyRefs : public StmtVisitor<ClassifyRefs> {
305 public:
306   enum Class {
307     Init,
308     Use,
309     SelfInit,
310     Ignore
311   };
312 
313 private:
314   const DeclContext *DC;
315   llvm::DenseMap<const DeclRefExpr*, Class> Classification;
316 
317   bool isTrackedVar(const VarDecl *VD) const {
318     return ::isTrackedVar(VD, DC);
319   }
320 
321   void classify(const Expr *E, Class C);
322 
323 public:
324   ClassifyRefs(AnalysisDeclContext &AC) : DC(cast<DeclContext>(AC.getDecl())) {}
325 
326   void VisitDeclStmt(DeclStmt *DS);
327   void VisitUnaryOperator(UnaryOperator *UO);
328   void VisitBinaryOperator(BinaryOperator *BO);
329   void VisitCallExpr(CallExpr *CE);
330   void VisitCastExpr(CastExpr *CE);
331 
332   void operator()(Stmt *S) { Visit(S); }
333 
334   Class get(const DeclRefExpr *DRE) const {
335     llvm::DenseMap<const DeclRefExpr*, Class>::const_iterator I
336         = Classification.find(DRE);
337     if (I != Classification.end())
338       return I->second;
339 
340     const VarDecl *VD = dyn_cast<VarDecl>(DRE->getDecl());
341     if (!VD || !isTrackedVar(VD))
342       return Ignore;
343 
344     return Init;
345   }
346 };
347 }
348 
349 static const DeclRefExpr *getSelfInitExpr(VarDecl *VD) {
350   if (Expr *Init = VD->getInit()) {
351     const DeclRefExpr *DRE
352       = dyn_cast<DeclRefExpr>(stripCasts(VD->getASTContext(), Init));
353     if (DRE && DRE->getDecl() == VD)
354       return DRE;
355   }
356   return 0;
357 }
358 
359 void ClassifyRefs::classify(const Expr *E, Class C) {
360   // The result of a ?: could also be an lvalue.
361   E = E->IgnoreParens();
362   if (const ConditionalOperator *CO = dyn_cast<ConditionalOperator>(E)) {
363     const Expr *TrueExpr = CO->getTrueExpr();
364     if (!isa<OpaqueValueExpr>(TrueExpr))
365       classify(TrueExpr, C);
366     classify(CO->getFalseExpr(), C);
367     return;
368   }
369 
370   FindVarResult Var = findVar(E, DC);
371   if (const DeclRefExpr *DRE = Var.getDeclRefExpr())
372     Classification[DRE] = std::max(Classification[DRE], C);
373 }
374 
375 void ClassifyRefs::VisitDeclStmt(DeclStmt *DS) {
376   for (DeclStmt::decl_iterator DI = DS->decl_begin(), DE = DS->decl_end();
377        DI != DE; ++DI) {
378     VarDecl *VD = dyn_cast<VarDecl>(*DI);
379     if (VD && isTrackedVar(VD))
380       if (const DeclRefExpr *DRE = getSelfInitExpr(VD))
381         Classification[DRE] = SelfInit;
382   }
383 }
384 
385 void ClassifyRefs::VisitBinaryOperator(BinaryOperator *BO) {
386   // Ignore the evaluation of a DeclRefExpr on the LHS of an assignment. If this
387   // is not a compound-assignment, we will treat it as initializing the variable
388   // when TransferFunctions visits it. A compound-assignment does not affect
389   // whether a variable is uninitialized, and there's no point counting it as a
390   // use.
391   if (BO->isCompoundAssignmentOp())
392     classify(BO->getLHS(), Use);
393   else if (BO->getOpcode() == BO_Assign)
394     classify(BO->getLHS(), Ignore);
395 }
396 
397 void ClassifyRefs::VisitUnaryOperator(UnaryOperator *UO) {
398   // Increment and decrement are uses despite there being no lvalue-to-rvalue
399   // conversion.
400   if (UO->isIncrementDecrementOp())
401     classify(UO->getSubExpr(), Use);
402 }
403 
404 void ClassifyRefs::VisitCallExpr(CallExpr *CE) {
405   // If a value is passed by const reference to a function, we should not assume
406   // that it is initialized by the call, and we conservatively do not assume
407   // that it is used.
408   for (CallExpr::arg_iterator I = CE->arg_begin(), E = CE->arg_end();
409        I != E; ++I)
410     if ((*I)->getType().isConstQualified() && (*I)->isGLValue())
411       classify(*I, Ignore);
412 }
413 
414 void ClassifyRefs::VisitCastExpr(CastExpr *CE) {
415   if (CE->getCastKind() == CK_LValueToRValue)
416     classify(CE->getSubExpr(), Use);
417   else if (CStyleCastExpr *CSE = dyn_cast<CStyleCastExpr>(CE)) {
418     if (CSE->getType()->isVoidType()) {
419       // Squelch any detected load of an uninitialized value if
420       // we cast it to void.
421       // e.g. (void) x;
422       classify(CSE->getSubExpr(), Ignore);
423     }
424   }
425 }
426 
427 //------------------------------------------------------------------------====//
428 // Transfer function for uninitialized values analysis.
429 //====------------------------------------------------------------------------//
430 
431 namespace {
432 class TransferFunctions : public StmtVisitor<TransferFunctions> {
433   CFGBlockValues &vals;
434   const CFG &cfg;
435   const CFGBlock *block;
436   AnalysisDeclContext &ac;
437   const ClassifyRefs &classification;
438   ObjCNoReturn objCNoRet;
439   UninitVariablesHandler &handler;
440 
441 public:
442   TransferFunctions(CFGBlockValues &vals, const CFG &cfg,
443                     const CFGBlock *block, AnalysisDeclContext &ac,
444                     const ClassifyRefs &classification,
445                     UninitVariablesHandler &handler)
446     : vals(vals), cfg(cfg), block(block), ac(ac),
447       classification(classification), objCNoRet(ac.getASTContext()),
448       handler(handler) {}
449 
450   void reportUse(const Expr *ex, const VarDecl *vd);
451 
452   void VisitBinaryOperator(BinaryOperator *bo);
453   void VisitBlockExpr(BlockExpr *be);
454   void VisitCallExpr(CallExpr *ce);
455   void VisitDeclRefExpr(DeclRefExpr *dr);
456   void VisitDeclStmt(DeclStmt *ds);
457   void VisitObjCForCollectionStmt(ObjCForCollectionStmt *FS);
458   void VisitObjCMessageExpr(ObjCMessageExpr *ME);
459 
460   bool isTrackedVar(const VarDecl *vd) {
461     return ::isTrackedVar(vd, cast<DeclContext>(ac.getDecl()));
462   }
463 
464   FindVarResult findVar(const Expr *ex) {
465     return ::findVar(ex, cast<DeclContext>(ac.getDecl()));
466   }
467 
468   UninitUse getUninitUse(const Expr *ex, const VarDecl *vd, Value v) {
469     UninitUse Use(ex, isAlwaysUninit(v));
470 
471     assert(isUninitialized(v));
472     if (Use.getKind() == UninitUse::Always)
473       return Use;
474 
475     // If an edge which leads unconditionally to this use did not initialize
476     // the variable, we can say something stronger than 'may be uninitialized':
477     // we can say 'either it's used uninitialized or you have dead code'.
478     //
479     // We track the number of successors of a node which have been visited, and
480     // visit a node once we have visited all of its successors. Only edges where
481     // the variable might still be uninitialized are followed. Since a variable
482     // can't transfer from being initialized to being uninitialized, this will
483     // trace out the subgraph which inevitably leads to the use and does not
484     // initialize the variable. We do not want to skip past loops, since their
485     // non-termination might be correlated with the initialization condition.
486     //
487     // For example:
488     //
489     //         void f(bool a, bool b) {
490     // block1:   int n;
491     //           if (a) {
492     // block2:     if (b)
493     // block3:       n = 1;
494     // block4:   } else if (b) {
495     // block5:     while (!a) {
496     // block6:       do_work(&a);
497     //               n = 2;
498     //             }
499     //           }
500     // block7:   if (a)
501     // block8:     g();
502     // block9:   return n;
503     //         }
504     //
505     // Starting from the maybe-uninitialized use in block 9:
506     //  * Block 7 is not visited because we have only visited one of its two
507     //    successors.
508     //  * Block 8 is visited because we've visited its only successor.
509     // From block 8:
510     //  * Block 7 is visited because we've now visited both of its successors.
511     // From block 7:
512     //  * Blocks 1, 2, 4, 5, and 6 are not visited because we didn't visit all
513     //    of their successors (we didn't visit 4, 3, 5, 6, and 5, respectively).
514     //  * Block 3 is not visited because it initializes 'n'.
515     // Now the algorithm terminates, having visited blocks 7 and 8, and having
516     // found the frontier is blocks 2, 4, and 5.
517     //
518     // 'n' is definitely uninitialized for two edges into block 7 (from blocks 2
519     // and 4), so we report that any time either of those edges is taken (in
520     // each case when 'b == false'), 'n' is used uninitialized.
521     SmallVector<const CFGBlock*, 32> Queue;
522     SmallVector<unsigned, 32> SuccsVisited(cfg.getNumBlockIDs(), 0);
523     Queue.push_back(block);
524     // Specify that we've already visited all successors of the starting block.
525     // This has the dual purpose of ensuring we never add it to the queue, and
526     // of marking it as not being a candidate element of the frontier.
527     SuccsVisited[block->getBlockID()] = block->succ_size();
528     while (!Queue.empty()) {
529       const CFGBlock *B = Queue.pop_back_val();
530       for (CFGBlock::const_pred_iterator I = B->pred_begin(), E = B->pred_end();
531            I != E; ++I) {
532         const CFGBlock *Pred = *I;
533         if (vals.getValue(Pred, B, vd) == Initialized)
534           // This block initializes the variable.
535           continue;
536 
537         unsigned &SV = SuccsVisited[Pred->getBlockID()];
538         if (!SV) {
539           // When visiting the first successor of a block, mark all NULL
540           // successors as having been visited.
541           for (CFGBlock::const_succ_iterator SI = Pred->succ_begin(),
542                                              SE = Pred->succ_end();
543                SI != SE; ++SI)
544             if (!*SI)
545               ++SV;
546         }
547 
548         if (++SV == Pred->succ_size())
549           // All paths from this block lead to the use and don't initialize the
550           // variable.
551           Queue.push_back(Pred);
552       }
553     }
554 
555     // Scan the frontier, looking for blocks where the variable was
556     // uninitialized.
557     for (CFG::const_iterator BI = cfg.begin(), BE = cfg.end(); BI != BE; ++BI) {
558       const CFGBlock *Block = *BI;
559       unsigned BlockID = Block->getBlockID();
560       const Stmt *Term = Block->getTerminator();
561       if (SuccsVisited[BlockID] && SuccsVisited[BlockID] < Block->succ_size() &&
562           Term) {
563         // This block inevitably leads to the use. If we have an edge from here
564         // to a post-dominator block, and the variable is uninitialized on that
565         // edge, we have found a bug.
566         for (CFGBlock::const_succ_iterator I = Block->succ_begin(),
567              E = Block->succ_end(); I != E; ++I) {
568           const CFGBlock *Succ = *I;
569           if (Succ && SuccsVisited[Succ->getBlockID()] >= Succ->succ_size() &&
570               vals.getValue(Block, Succ, vd) == Uninitialized) {
571             // Switch cases are a special case: report the label to the caller
572             // as the 'terminator', not the switch statement itself. Suppress
573             // situations where no label matched: we can't be sure that's
574             // possible.
575             if (isa<SwitchStmt>(Term)) {
576               const Stmt *Label = Succ->getLabel();
577               if (!Label || !isa<SwitchCase>(Label))
578                 // Might not be possible.
579                 continue;
580               UninitUse::Branch Branch;
581               Branch.Terminator = Label;
582               Branch.Output = 0; // Ignored.
583               Use.addUninitBranch(Branch);
584             } else {
585               UninitUse::Branch Branch;
586               Branch.Terminator = Term;
587               Branch.Output = I - Block->succ_begin();
588               Use.addUninitBranch(Branch);
589             }
590           }
591         }
592       }
593     }
594 
595     return Use;
596   }
597 };
598 }
599 
600 void TransferFunctions::reportUse(const Expr *ex, const VarDecl *vd) {
601   Value v = vals[vd];
602   if (isUninitialized(v))
603     handler.handleUseOfUninitVariable(vd, getUninitUse(ex, vd, v));
604 }
605 
606 void TransferFunctions::VisitObjCForCollectionStmt(ObjCForCollectionStmt *FS) {
607   // This represents an initialization of the 'element' value.
608   if (DeclStmt *DS = dyn_cast<DeclStmt>(FS->getElement())) {
609     const VarDecl *VD = cast<VarDecl>(DS->getSingleDecl());
610     if (isTrackedVar(VD))
611       vals[VD] = Initialized;
612   }
613 }
614 
615 void TransferFunctions::VisitBlockExpr(BlockExpr *be) {
616   const BlockDecl *bd = be->getBlockDecl();
617   for (BlockDecl::capture_const_iterator i = bd->capture_begin(),
618         e = bd->capture_end() ; i != e; ++i) {
619     const VarDecl *vd = i->getVariable();
620     if (!isTrackedVar(vd))
621       continue;
622     if (i->isByRef()) {
623       vals[vd] = Initialized;
624       continue;
625     }
626     reportUse(be, vd);
627   }
628 }
629 
630 void TransferFunctions::VisitCallExpr(CallExpr *ce) {
631   if (Decl *Callee = ce->getCalleeDecl()) {
632     if (Callee->hasAttr<ReturnsTwiceAttr>()) {
633       // After a call to a function like setjmp or vfork, any variable which is
634       // initialized anywhere within this function may now be initialized. For
635       // now, just assume such a call initializes all variables.  FIXME: Only
636       // mark variables as initialized if they have an initializer which is
637       // reachable from here.
638       vals.setAllScratchValues(Initialized);
639     }
640     else if (Callee->hasAttr<AnalyzerNoReturnAttr>()) {
641       // Functions labeled like "analyzer_noreturn" are often used to denote
642       // "panic" functions that in special debug situations can still return,
643       // but for the most part should not be treated as returning.  This is a
644       // useful annotation borrowed from the static analyzer that is useful for
645       // suppressing branch-specific false positives when we call one of these
646       // functions but keep pretending the path continues (when in reality the
647       // user doesn't care).
648       vals.setAllScratchValues(Unknown);
649     }
650   }
651 }
652 
653 void TransferFunctions::VisitDeclRefExpr(DeclRefExpr *dr) {
654   switch (classification.get(dr)) {
655   case ClassifyRefs::Ignore:
656     break;
657   case ClassifyRefs::Use:
658     reportUse(dr, cast<VarDecl>(dr->getDecl()));
659     break;
660   case ClassifyRefs::Init:
661     vals[cast<VarDecl>(dr->getDecl())] = Initialized;
662     break;
663   case ClassifyRefs::SelfInit:
664       handler.handleSelfInit(cast<VarDecl>(dr->getDecl()));
665     break;
666   }
667 }
668 
669 void TransferFunctions::VisitBinaryOperator(BinaryOperator *BO) {
670   if (BO->getOpcode() == BO_Assign) {
671     FindVarResult Var = findVar(BO->getLHS());
672     if (const VarDecl *VD = Var.getDecl())
673       vals[VD] = Initialized;
674   }
675 }
676 
677 void TransferFunctions::VisitDeclStmt(DeclStmt *DS) {
678   for (DeclStmt::decl_iterator DI = DS->decl_begin(), DE = DS->decl_end();
679        DI != DE; ++DI) {
680     VarDecl *VD = dyn_cast<VarDecl>(*DI);
681     if (VD && isTrackedVar(VD)) {
682       if (getSelfInitExpr(VD)) {
683         // If the initializer consists solely of a reference to itself, we
684         // explicitly mark the variable as uninitialized. This allows code
685         // like the following:
686         //
687         //   int x = x;
688         //
689         // to deliberately leave a variable uninitialized. Different analysis
690         // clients can detect this pattern and adjust their reporting
691         // appropriately, but we need to continue to analyze subsequent uses
692         // of the variable.
693         vals[VD] = Uninitialized;
694       } else if (VD->getInit()) {
695         // Treat the new variable as initialized.
696         vals[VD] = Initialized;
697       } else {
698         // No initializer: the variable is now uninitialized. This matters
699         // for cases like:
700         //   while (...) {
701         //     int n;
702         //     use(n);
703         //     n = 0;
704         //   }
705         // FIXME: Mark the variable as uninitialized whenever its scope is
706         // left, since its scope could be re-entered by a jump over the
707         // declaration.
708         vals[VD] = Uninitialized;
709       }
710     }
711   }
712 }
713 
714 void TransferFunctions::VisitObjCMessageExpr(ObjCMessageExpr *ME) {
715   // If the Objective-C message expression is an implicit no-return that
716   // is not modeled in the CFG, set the tracked dataflow values to Unknown.
717   if (objCNoRet.isImplicitNoReturn(ME)) {
718     vals.setAllScratchValues(Unknown);
719   }
720 }
721 
722 //------------------------------------------------------------------------====//
723 // High-level "driver" logic for uninitialized values analysis.
724 //====------------------------------------------------------------------------//
725 
726 static bool runOnBlock(const CFGBlock *block, const CFG &cfg,
727                        AnalysisDeclContext &ac, CFGBlockValues &vals,
728                        const ClassifyRefs &classification,
729                        llvm::BitVector &wasAnalyzed,
730                        UninitVariablesHandler &handler) {
731   wasAnalyzed[block->getBlockID()] = true;
732   vals.resetScratch();
733   // Merge in values of predecessor blocks.
734   bool isFirst = true;
735   for (CFGBlock::const_pred_iterator I = block->pred_begin(),
736        E = block->pred_end(); I != E; ++I) {
737     const CFGBlock *pred = *I;
738     if (wasAnalyzed[pred->getBlockID()]) {
739       vals.mergeIntoScratch(vals.getValueVector(pred), isFirst);
740       isFirst = false;
741     }
742   }
743   // Apply the transfer function.
744   TransferFunctions tf(vals, cfg, block, ac, classification, handler);
745   for (CFGBlock::const_iterator I = block->begin(), E = block->end();
746        I != E; ++I) {
747     if (Optional<CFGStmt> cs = I->getAs<CFGStmt>())
748       tf.Visit(const_cast<Stmt*>(cs->getStmt()));
749   }
750   return vals.updateValueVectorWithScratch(block);
751 }
752 
753 /// PruneBlocksHandler is a special UninitVariablesHandler that is used
754 /// to detect when a CFGBlock has any *potential* use of an uninitialized
755 /// variable.  It is mainly used to prune out work during the final
756 /// reporting pass.
757 namespace {
758 struct PruneBlocksHandler : public UninitVariablesHandler {
759   PruneBlocksHandler(unsigned numBlocks)
760     : hadUse(numBlocks, false), hadAnyUse(false),
761       currentBlock(0) {}
762 
763   virtual ~PruneBlocksHandler() {}
764 
765   /// Records if a CFGBlock had a potential use of an uninitialized variable.
766   llvm::BitVector hadUse;
767 
768   /// Records if any CFGBlock had a potential use of an uninitialized variable.
769   bool hadAnyUse;
770 
771   /// The current block to scribble use information.
772   unsigned currentBlock;
773 
774   virtual void handleUseOfUninitVariable(const VarDecl *vd,
775                                          const UninitUse &use) {
776     hadUse[currentBlock] = true;
777     hadAnyUse = true;
778   }
779 
780   /// Called when the uninitialized variable analysis detects the
781   /// idiom 'int x = x'.  All other uses of 'x' within the initializer
782   /// are handled by handleUseOfUninitVariable.
783   virtual void handleSelfInit(const VarDecl *vd) {
784     hadUse[currentBlock] = true;
785     hadAnyUse = true;
786   }
787 };
788 }
789 
790 void clang::runUninitializedVariablesAnalysis(
791     const DeclContext &dc,
792     const CFG &cfg,
793     AnalysisDeclContext &ac,
794     UninitVariablesHandler &handler,
795     UninitVariablesAnalysisStats &stats) {
796   CFGBlockValues vals(cfg);
797   vals.computeSetOfDeclarations(dc);
798   if (vals.hasNoDeclarations())
799     return;
800 
801   stats.NumVariablesAnalyzed = vals.getNumEntries();
802 
803   // Precompute which expressions are uses and which are initializations.
804   ClassifyRefs classification(ac);
805   cfg.VisitBlockStmts(classification);
806 
807   // Mark all variables uninitialized at the entry.
808   const CFGBlock &entry = cfg.getEntry();
809   ValueVector &vec = vals.getValueVector(&entry);
810   const unsigned n = vals.getNumEntries();
811   for (unsigned j = 0; j < n ; ++j) {
812     vec[j] = Uninitialized;
813   }
814 
815   // Proceed with the workist.
816   DataflowWorklist worklist(cfg, *ac.getAnalysis<PostOrderCFGView>());
817   llvm::BitVector previouslyVisited(cfg.getNumBlockIDs());
818   worklist.enqueueSuccessors(&cfg.getEntry());
819   llvm::BitVector wasAnalyzed(cfg.getNumBlockIDs(), false);
820   wasAnalyzed[cfg.getEntry().getBlockID()] = true;
821   PruneBlocksHandler PBH(cfg.getNumBlockIDs());
822 
823   while (const CFGBlock *block = worklist.dequeue()) {
824     PBH.currentBlock = block->getBlockID();
825 
826     // Did the block change?
827     bool changed = runOnBlock(block, cfg, ac, vals,
828                               classification, wasAnalyzed, PBH);
829     ++stats.NumBlockVisits;
830     if (changed || !previouslyVisited[block->getBlockID()])
831       worklist.enqueueSuccessors(block);
832     previouslyVisited[block->getBlockID()] = true;
833   }
834 
835   if (!PBH.hadAnyUse)
836     return;
837 
838   // Run through the blocks one more time, and report uninitialized variables.
839   for (CFG::const_iterator BI = cfg.begin(), BE = cfg.end(); BI != BE; ++BI) {
840     const CFGBlock *block = *BI;
841     if (PBH.hadUse[block->getBlockID()]) {
842       runOnBlock(block, cfg, ac, vals, classification, wasAnalyzed, handler);
843       ++stats.NumBlockVisits;
844     }
845   }
846 }
847 
848 UninitVariablesHandler::~UninitVariablesHandler() {}
849