• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SimpleSValBuilder.cpp - A basic SValBuilder -----------------------*- C++ -*-
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 //  This file defines SimpleSValBuilder, a basic implementation of SValBuilder.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "clang/StaticAnalyzer/Core/PathSensitive/SValBuilder.h"
15 #include "clang/StaticAnalyzer/Core/PathSensitive/APSIntType.h"
16 #include "clang/StaticAnalyzer/Core/PathSensitive/ProgramState.h"
17 
18 using namespace clang;
19 using namespace ento;
20 
21 namespace {
22 class SimpleSValBuilder : public SValBuilder {
23 protected:
24   virtual SVal dispatchCast(SVal val, QualType castTy);
25   virtual SVal evalCastFromNonLoc(NonLoc val, QualType castTy);
26   virtual SVal evalCastFromLoc(Loc val, QualType castTy);
27 
28 public:
SimpleSValBuilder(llvm::BumpPtrAllocator & alloc,ASTContext & context,ProgramStateManager & stateMgr)29   SimpleSValBuilder(llvm::BumpPtrAllocator &alloc, ASTContext &context,
30                     ProgramStateManager &stateMgr)
31                     : SValBuilder(alloc, context, stateMgr) {}
~SimpleSValBuilder()32   virtual ~SimpleSValBuilder() {}
33 
34   virtual SVal evalMinus(NonLoc val);
35   virtual SVal evalComplement(NonLoc val);
36   virtual SVal evalBinOpNN(ProgramStateRef state, BinaryOperator::Opcode op,
37                            NonLoc lhs, NonLoc rhs, QualType resultTy);
38   virtual SVal evalBinOpLL(ProgramStateRef state, BinaryOperator::Opcode op,
39                            Loc lhs, Loc rhs, QualType resultTy);
40   virtual SVal evalBinOpLN(ProgramStateRef state, BinaryOperator::Opcode op,
41                            Loc lhs, NonLoc rhs, QualType resultTy);
42 
43   /// getKnownValue - evaluates a given SVal. If the SVal has only one possible
44   ///  (integer) value, that value is returned. Otherwise, returns NULL.
45   virtual const llvm::APSInt *getKnownValue(ProgramStateRef state, SVal V);
46 
47   SVal MakeSymIntVal(const SymExpr *LHS, BinaryOperator::Opcode op,
48                      const llvm::APSInt &RHS, QualType resultTy);
49 };
50 } // end anonymous namespace
51 
createSimpleSValBuilder(llvm::BumpPtrAllocator & alloc,ASTContext & context,ProgramStateManager & stateMgr)52 SValBuilder *ento::createSimpleSValBuilder(llvm::BumpPtrAllocator &alloc,
53                                            ASTContext &context,
54                                            ProgramStateManager &stateMgr) {
55   return new SimpleSValBuilder(alloc, context, stateMgr);
56 }
57 
58 //===----------------------------------------------------------------------===//
59 // Transfer function for Casts.
60 //===----------------------------------------------------------------------===//
61 
dispatchCast(SVal Val,QualType CastTy)62 SVal SimpleSValBuilder::dispatchCast(SVal Val, QualType CastTy) {
63   assert(Val.getAs<Loc>() || Val.getAs<NonLoc>());
64   return Val.getAs<Loc>() ? evalCastFromLoc(Val.castAs<Loc>(), CastTy)
65                            : evalCastFromNonLoc(Val.castAs<NonLoc>(), CastTy);
66 }
67 
evalCastFromNonLoc(NonLoc val,QualType castTy)68 SVal SimpleSValBuilder::evalCastFromNonLoc(NonLoc val, QualType castTy) {
69 
70   bool isLocType = Loc::isLocType(castTy);
71 
72   if (Optional<nonloc::LocAsInteger> LI = val.getAs<nonloc::LocAsInteger>()) {
73     if (isLocType)
74       return LI->getLoc();
75 
76     // FIXME: Correctly support promotions/truncations.
77     unsigned castSize = Context.getTypeSize(castTy);
78     if (castSize == LI->getNumBits())
79       return val;
80     return makeLocAsInteger(LI->getLoc(), castSize);
81   }
82 
83   if (const SymExpr *se = val.getAsSymbolicExpression()) {
84     QualType T = Context.getCanonicalType(se->getType());
85     // If types are the same or both are integers, ignore the cast.
86     // FIXME: Remove this hack when we support symbolic truncation/extension.
87     // HACK: If both castTy and T are integers, ignore the cast.  This is
88     // not a permanent solution.  Eventually we want to precisely handle
89     // extension/truncation of symbolic integers.  This prevents us from losing
90     // precision when we assign 'x = y' and 'y' is symbolic and x and y are
91     // different integer types.
92    if (haveSameType(T, castTy))
93       return val;
94 
95     if (!isLocType)
96       return makeNonLoc(se, T, castTy);
97     return UnknownVal();
98   }
99 
100   // If value is a non integer constant, produce unknown.
101   if (!val.getAs<nonloc::ConcreteInt>())
102     return UnknownVal();
103 
104   // Handle casts to a boolean type.
105   if (castTy->isBooleanType()) {
106     bool b = val.castAs<nonloc::ConcreteInt>().getValue().getBoolValue();
107     return makeTruthVal(b, castTy);
108   }
109 
110   // Only handle casts from integers to integers - if val is an integer constant
111   // being cast to a non integer type, produce unknown.
112   if (!isLocType && !castTy->isIntegralOrEnumerationType())
113     return UnknownVal();
114 
115   llvm::APSInt i = val.castAs<nonloc::ConcreteInt>().getValue();
116   BasicVals.getAPSIntType(castTy).apply(i);
117 
118   if (isLocType)
119     return makeIntLocVal(i);
120   else
121     return makeIntVal(i);
122 }
123 
evalCastFromLoc(Loc val,QualType castTy)124 SVal SimpleSValBuilder::evalCastFromLoc(Loc val, QualType castTy) {
125 
126   // Casts from pointers -> pointers, just return the lval.
127   //
128   // Casts from pointers -> references, just return the lval.  These
129   //   can be introduced by the frontend for corner cases, e.g
130   //   casting from va_list* to __builtin_va_list&.
131   //
132   if (Loc::isLocType(castTy) || castTy->isReferenceType())
133     return val;
134 
135   // FIXME: Handle transparent unions where a value can be "transparently"
136   //  lifted into a union type.
137   if (castTy->isUnionType())
138     return UnknownVal();
139 
140   if (castTy->isIntegralOrEnumerationType()) {
141     unsigned BitWidth = Context.getTypeSize(castTy);
142 
143     if (!val.getAs<loc::ConcreteInt>())
144       return makeLocAsInteger(val, BitWidth);
145 
146     llvm::APSInt i = val.castAs<loc::ConcreteInt>().getValue();
147     BasicVals.getAPSIntType(castTy).apply(i);
148     return makeIntVal(i);
149   }
150 
151   // All other cases: return 'UnknownVal'.  This includes casting pointers
152   // to floats, which is probably badness it itself, but this is a good
153   // intermediate solution until we do something better.
154   return UnknownVal();
155 }
156 
157 //===----------------------------------------------------------------------===//
158 // Transfer function for unary operators.
159 //===----------------------------------------------------------------------===//
160 
evalMinus(NonLoc val)161 SVal SimpleSValBuilder::evalMinus(NonLoc val) {
162   switch (val.getSubKind()) {
163   case nonloc::ConcreteIntKind:
164     return val.castAs<nonloc::ConcreteInt>().evalMinus(*this);
165   default:
166     return UnknownVal();
167   }
168 }
169 
evalComplement(NonLoc X)170 SVal SimpleSValBuilder::evalComplement(NonLoc X) {
171   switch (X.getSubKind()) {
172   case nonloc::ConcreteIntKind:
173     return X.castAs<nonloc::ConcreteInt>().evalComplement(*this);
174   default:
175     return UnknownVal();
176   }
177 }
178 
179 //===----------------------------------------------------------------------===//
180 // Transfer function for binary operators.
181 //===----------------------------------------------------------------------===//
182 
MakeSymIntVal(const SymExpr * LHS,BinaryOperator::Opcode op,const llvm::APSInt & RHS,QualType resultTy)183 SVal SimpleSValBuilder::MakeSymIntVal(const SymExpr *LHS,
184                                     BinaryOperator::Opcode op,
185                                     const llvm::APSInt &RHS,
186                                     QualType resultTy) {
187   bool isIdempotent = false;
188 
189   // Check for a few special cases with known reductions first.
190   switch (op) {
191   default:
192     // We can't reduce this case; just treat it normally.
193     break;
194   case BO_Mul:
195     // a*0 and a*1
196     if (RHS == 0)
197       return makeIntVal(0, resultTy);
198     else if (RHS == 1)
199       isIdempotent = true;
200     break;
201   case BO_Div:
202     // a/0 and a/1
203     if (RHS == 0)
204       // This is also handled elsewhere.
205       return UndefinedVal();
206     else if (RHS == 1)
207       isIdempotent = true;
208     break;
209   case BO_Rem:
210     // a%0 and a%1
211     if (RHS == 0)
212       // This is also handled elsewhere.
213       return UndefinedVal();
214     else if (RHS == 1)
215       return makeIntVal(0, resultTy);
216     break;
217   case BO_Add:
218   case BO_Sub:
219   case BO_Shl:
220   case BO_Shr:
221   case BO_Xor:
222     // a+0, a-0, a<<0, a>>0, a^0
223     if (RHS == 0)
224       isIdempotent = true;
225     break;
226   case BO_And:
227     // a&0 and a&(~0)
228     if (RHS == 0)
229       return makeIntVal(0, resultTy);
230     else if (RHS.isAllOnesValue())
231       isIdempotent = true;
232     break;
233   case BO_Or:
234     // a|0 and a|(~0)
235     if (RHS == 0)
236       isIdempotent = true;
237     else if (RHS.isAllOnesValue()) {
238       const llvm::APSInt &Result = BasicVals.Convert(resultTy, RHS);
239       return nonloc::ConcreteInt(Result);
240     }
241     break;
242   }
243 
244   // Idempotent ops (like a*1) can still change the type of an expression.
245   // Wrap the LHS up in a NonLoc again and let evalCastFromNonLoc do the
246   // dirty work.
247   if (isIdempotent)
248       return evalCastFromNonLoc(nonloc::SymbolVal(LHS), resultTy);
249 
250   // If we reach this point, the expression cannot be simplified.
251   // Make a SymbolVal for the entire expression, after converting the RHS.
252   const llvm::APSInt *ConvertedRHS = &RHS;
253   if (BinaryOperator::isComparisonOp(op)) {
254     // We're looking for a type big enough to compare the symbolic value
255     // with the given constant.
256     // FIXME: This is an approximation of Sema::UsualArithmeticConversions.
257     ASTContext &Ctx = getContext();
258     QualType SymbolType = LHS->getType();
259     uint64_t ValWidth = RHS.getBitWidth();
260     uint64_t TypeWidth = Ctx.getTypeSize(SymbolType);
261 
262     if (ValWidth < TypeWidth) {
263       // If the value is too small, extend it.
264       ConvertedRHS = &BasicVals.Convert(SymbolType, RHS);
265     } else if (ValWidth == TypeWidth) {
266       // If the value is signed but the symbol is unsigned, do the comparison
267       // in unsigned space. [C99 6.3.1.8]
268       // (For the opposite case, the value is already unsigned.)
269       if (RHS.isSigned() && !SymbolType->isSignedIntegerOrEnumerationType())
270         ConvertedRHS = &BasicVals.Convert(SymbolType, RHS);
271     }
272   } else
273     ConvertedRHS = &BasicVals.Convert(resultTy, RHS);
274 
275   return makeNonLoc(LHS, op, *ConvertedRHS, resultTy);
276 }
277 
evalBinOpNN(ProgramStateRef state,BinaryOperator::Opcode op,NonLoc lhs,NonLoc rhs,QualType resultTy)278 SVal SimpleSValBuilder::evalBinOpNN(ProgramStateRef state,
279                                   BinaryOperator::Opcode op,
280                                   NonLoc lhs, NonLoc rhs,
281                                   QualType resultTy)  {
282   NonLoc InputLHS = lhs;
283   NonLoc InputRHS = rhs;
284 
285   // Handle trivial case where left-side and right-side are the same.
286   if (lhs == rhs)
287     switch (op) {
288       default:
289         break;
290       case BO_EQ:
291       case BO_LE:
292       case BO_GE:
293         return makeTruthVal(true, resultTy);
294       case BO_LT:
295       case BO_GT:
296       case BO_NE:
297         return makeTruthVal(false, resultTy);
298       case BO_Xor:
299       case BO_Sub:
300         if (resultTy->isIntegralOrEnumerationType())
301           return makeIntVal(0, resultTy);
302         return evalCastFromNonLoc(makeIntVal(0, /*Unsigned=*/false), resultTy);
303       case BO_Or:
304       case BO_And:
305         return evalCastFromNonLoc(lhs, resultTy);
306     }
307 
308   while (1) {
309     switch (lhs.getSubKind()) {
310     default:
311       return makeSymExprValNN(state, op, lhs, rhs, resultTy);
312     case nonloc::LocAsIntegerKind: {
313       Loc lhsL = lhs.castAs<nonloc::LocAsInteger>().getLoc();
314       switch (rhs.getSubKind()) {
315         case nonloc::LocAsIntegerKind:
316           return evalBinOpLL(state, op, lhsL,
317                              rhs.castAs<nonloc::LocAsInteger>().getLoc(),
318                              resultTy);
319         case nonloc::ConcreteIntKind: {
320           // Transform the integer into a location and compare.
321           llvm::APSInt i = rhs.castAs<nonloc::ConcreteInt>().getValue();
322           BasicVals.getAPSIntType(Context.VoidPtrTy).apply(i);
323           return evalBinOpLL(state, op, lhsL, makeLoc(i), resultTy);
324         }
325         default:
326           switch (op) {
327             case BO_EQ:
328               return makeTruthVal(false, resultTy);
329             case BO_NE:
330               return makeTruthVal(true, resultTy);
331             default:
332               // This case also handles pointer arithmetic.
333               return makeSymExprValNN(state, op, InputLHS, InputRHS, resultTy);
334           }
335       }
336     }
337     case nonloc::ConcreteIntKind: {
338       llvm::APSInt LHSValue = lhs.castAs<nonloc::ConcreteInt>().getValue();
339 
340       // If we're dealing with two known constants, just perform the operation.
341       if (const llvm::APSInt *KnownRHSValue = getKnownValue(state, rhs)) {
342         llvm::APSInt RHSValue = *KnownRHSValue;
343         if (BinaryOperator::isComparisonOp(op)) {
344           // We're looking for a type big enough to compare the two values.
345           // FIXME: This is not correct. char + short will result in a promotion
346           // to int. Unfortunately we have lost types by this point.
347           APSIntType CompareType = std::max(APSIntType(LHSValue),
348                                             APSIntType(RHSValue));
349           CompareType.apply(LHSValue);
350           CompareType.apply(RHSValue);
351         } else if (!BinaryOperator::isShiftOp(op)) {
352           APSIntType IntType = BasicVals.getAPSIntType(resultTy);
353           IntType.apply(LHSValue);
354           IntType.apply(RHSValue);
355         }
356 
357         const llvm::APSInt *Result =
358           BasicVals.evalAPSInt(op, LHSValue, RHSValue);
359         if (!Result)
360           return UndefinedVal();
361 
362         return nonloc::ConcreteInt(*Result);
363       }
364 
365       // Swap the left and right sides and flip the operator if doing so
366       // allows us to better reason about the expression (this is a form
367       // of expression canonicalization).
368       // While we're at it, catch some special cases for non-commutative ops.
369       switch (op) {
370       case BO_LT:
371       case BO_GT:
372       case BO_LE:
373       case BO_GE:
374         op = BinaryOperator::reverseComparisonOp(op);
375         // FALL-THROUGH
376       case BO_EQ:
377       case BO_NE:
378       case BO_Add:
379       case BO_Mul:
380       case BO_And:
381       case BO_Xor:
382       case BO_Or:
383         std::swap(lhs, rhs);
384         continue;
385       case BO_Shr:
386         // (~0)>>a
387         if (LHSValue.isAllOnesValue() && LHSValue.isSigned())
388           return evalCastFromNonLoc(lhs, resultTy);
389         // FALL-THROUGH
390       case BO_Shl:
391         // 0<<a and 0>>a
392         if (LHSValue == 0)
393           return evalCastFromNonLoc(lhs, resultTy);
394         return makeSymExprValNN(state, op, InputLHS, InputRHS, resultTy);
395       default:
396         return makeSymExprValNN(state, op, InputLHS, InputRHS, resultTy);
397       }
398     }
399     case nonloc::SymbolValKind: {
400       // We only handle LHS as simple symbols or SymIntExprs.
401       SymbolRef Sym = lhs.castAs<nonloc::SymbolVal>().getSymbol();
402 
403       // LHS is a symbolic expression.
404       if (const SymIntExpr *symIntExpr = dyn_cast<SymIntExpr>(Sym)) {
405 
406         // Is this a logical not? (!x is represented as x == 0.)
407         if (op == BO_EQ && rhs.isZeroConstant()) {
408           // We know how to negate certain expressions. Simplify them here.
409 
410           BinaryOperator::Opcode opc = symIntExpr->getOpcode();
411           switch (opc) {
412           default:
413             // We don't know how to negate this operation.
414             // Just handle it as if it were a normal comparison to 0.
415             break;
416           case BO_LAnd:
417           case BO_LOr:
418             llvm_unreachable("Logical operators handled by branching logic.");
419           case BO_Assign:
420           case BO_MulAssign:
421           case BO_DivAssign:
422           case BO_RemAssign:
423           case BO_AddAssign:
424           case BO_SubAssign:
425           case BO_ShlAssign:
426           case BO_ShrAssign:
427           case BO_AndAssign:
428           case BO_XorAssign:
429           case BO_OrAssign:
430           case BO_Comma:
431             llvm_unreachable("'=' and ',' operators handled by ExprEngine.");
432           case BO_PtrMemD:
433           case BO_PtrMemI:
434             llvm_unreachable("Pointer arithmetic not handled here.");
435           case BO_LT:
436           case BO_GT:
437           case BO_LE:
438           case BO_GE:
439           case BO_EQ:
440           case BO_NE:
441             assert(resultTy->isBooleanType() ||
442                    resultTy == getConditionType());
443             assert(symIntExpr->getType()->isBooleanType() ||
444                    getContext().hasSameUnqualifiedType(symIntExpr->getType(),
445                                                        getConditionType()));
446             // Negate the comparison and make a value.
447             opc = BinaryOperator::negateComparisonOp(opc);
448             return makeNonLoc(symIntExpr->getLHS(), opc,
449                 symIntExpr->getRHS(), resultTy);
450           }
451         }
452 
453         // For now, only handle expressions whose RHS is a constant.
454         if (const llvm::APSInt *RHSValue = getKnownValue(state, rhs)) {
455           // If both the LHS and the current expression are additive,
456           // fold their constants and try again.
457           if (BinaryOperator::isAdditiveOp(op)) {
458             BinaryOperator::Opcode lop = symIntExpr->getOpcode();
459             if (BinaryOperator::isAdditiveOp(lop)) {
460               // Convert the two constants to a common type, then combine them.
461 
462               // resultTy may not be the best type to convert to, but it's
463               // probably the best choice in expressions with mixed type
464               // (such as x+1U+2LL). The rules for implicit conversions should
465               // choose a reasonable type to preserve the expression, and will
466               // at least match how the value is going to be used.
467               APSIntType IntType = BasicVals.getAPSIntType(resultTy);
468               const llvm::APSInt &first = IntType.convert(symIntExpr->getRHS());
469               const llvm::APSInt &second = IntType.convert(*RHSValue);
470 
471               const llvm::APSInt *newRHS;
472               if (lop == op)
473                 newRHS = BasicVals.evalAPSInt(BO_Add, first, second);
474               else
475                 newRHS = BasicVals.evalAPSInt(BO_Sub, first, second);
476 
477               assert(newRHS && "Invalid operation despite common type!");
478               rhs = nonloc::ConcreteInt(*newRHS);
479               lhs = nonloc::SymbolVal(symIntExpr->getLHS());
480               op = lop;
481               continue;
482             }
483           }
484 
485           // Otherwise, make a SymIntExpr out of the expression.
486           return MakeSymIntVal(symIntExpr, op, *RHSValue, resultTy);
487         }
488       }
489 
490       // Does the symbolic expression simplify to a constant?
491       // If so, "fold" the constant by setting 'lhs' to a ConcreteInt
492       // and try again.
493       ConstraintManager &CMgr = state->getConstraintManager();
494       if (const llvm::APSInt *Constant = CMgr.getSymVal(state, Sym)) {
495         lhs = nonloc::ConcreteInt(*Constant);
496         continue;
497       }
498 
499       // Is the RHS a constant?
500       if (const llvm::APSInt *RHSValue = getKnownValue(state, rhs))
501         return MakeSymIntVal(Sym, op, *RHSValue, resultTy);
502 
503       // Give up -- this is not a symbolic expression we can handle.
504       return makeSymExprValNN(state, op, InputLHS, InputRHS, resultTy);
505     }
506     }
507   }
508 }
509 
evalBinOpFieldRegionFieldRegion(const FieldRegion * LeftFR,const FieldRegion * RightFR,BinaryOperator::Opcode op,QualType resultTy,SimpleSValBuilder & SVB)510 static SVal evalBinOpFieldRegionFieldRegion(const FieldRegion *LeftFR,
511                                             const FieldRegion *RightFR,
512                                             BinaryOperator::Opcode op,
513                                             QualType resultTy,
514                                             SimpleSValBuilder &SVB) {
515   // Only comparisons are meaningful here!
516   if (!BinaryOperator::isComparisonOp(op))
517     return UnknownVal();
518 
519   // Next, see if the two FRs have the same super-region.
520   // FIXME: This doesn't handle casts yet, and simply stripping the casts
521   // doesn't help.
522   if (LeftFR->getSuperRegion() != RightFR->getSuperRegion())
523     return UnknownVal();
524 
525   const FieldDecl *LeftFD = LeftFR->getDecl();
526   const FieldDecl *RightFD = RightFR->getDecl();
527   const RecordDecl *RD = LeftFD->getParent();
528 
529   // Make sure the two FRs are from the same kind of record. Just in case!
530   // FIXME: This is probably where inheritance would be a problem.
531   if (RD != RightFD->getParent())
532     return UnknownVal();
533 
534   // We know for sure that the two fields are not the same, since that
535   // would have given us the same SVal.
536   if (op == BO_EQ)
537     return SVB.makeTruthVal(false, resultTy);
538   if (op == BO_NE)
539     return SVB.makeTruthVal(true, resultTy);
540 
541   // Iterate through the fields and see which one comes first.
542   // [C99 6.7.2.1.13] "Within a structure object, the non-bit-field
543   // members and the units in which bit-fields reside have addresses that
544   // increase in the order in which they are declared."
545   bool leftFirst = (op == BO_LT || op == BO_LE);
546   for (RecordDecl::field_iterator I = RD->field_begin(),
547        E = RD->field_end(); I!=E; ++I) {
548     if (*I == LeftFD)
549       return SVB.makeTruthVal(leftFirst, resultTy);
550     if (*I == RightFD)
551       return SVB.makeTruthVal(!leftFirst, resultTy);
552   }
553 
554   llvm_unreachable("Fields not found in parent record's definition");
555 }
556 
557 // FIXME: all this logic will change if/when we have MemRegion::getLocation().
evalBinOpLL(ProgramStateRef state,BinaryOperator::Opcode op,Loc lhs,Loc rhs,QualType resultTy)558 SVal SimpleSValBuilder::evalBinOpLL(ProgramStateRef state,
559                                   BinaryOperator::Opcode op,
560                                   Loc lhs, Loc rhs,
561                                   QualType resultTy) {
562   // Only comparisons and subtractions are valid operations on two pointers.
563   // See [C99 6.5.5 through 6.5.14] or [C++0x 5.6 through 5.15].
564   // However, if a pointer is casted to an integer, evalBinOpNN may end up
565   // calling this function with another operation (PR7527). We don't attempt to
566   // model this for now, but it could be useful, particularly when the
567   // "location" is actually an integer value that's been passed through a void*.
568   if (!(BinaryOperator::isComparisonOp(op) || op == BO_Sub))
569     return UnknownVal();
570 
571   // Special cases for when both sides are identical.
572   if (lhs == rhs) {
573     switch (op) {
574     default:
575       llvm_unreachable("Unimplemented operation for two identical values");
576     case BO_Sub:
577       return makeZeroVal(resultTy);
578     case BO_EQ:
579     case BO_LE:
580     case BO_GE:
581       return makeTruthVal(true, resultTy);
582     case BO_NE:
583     case BO_LT:
584     case BO_GT:
585       return makeTruthVal(false, resultTy);
586     }
587   }
588 
589   switch (lhs.getSubKind()) {
590   default:
591     llvm_unreachable("Ordering not implemented for this Loc.");
592 
593   case loc::GotoLabelKind:
594     // The only thing we know about labels is that they're non-null.
595     if (rhs.isZeroConstant()) {
596       switch (op) {
597       default:
598         break;
599       case BO_Sub:
600         return evalCastFromLoc(lhs, resultTy);
601       case BO_EQ:
602       case BO_LE:
603       case BO_LT:
604         return makeTruthVal(false, resultTy);
605       case BO_NE:
606       case BO_GT:
607       case BO_GE:
608         return makeTruthVal(true, resultTy);
609       }
610     }
611     // There may be two labels for the same location, and a function region may
612     // have the same address as a label at the start of the function (depending
613     // on the ABI).
614     // FIXME: we can probably do a comparison against other MemRegions, though.
615     // FIXME: is there a way to tell if two labels refer to the same location?
616     return UnknownVal();
617 
618   case loc::ConcreteIntKind: {
619     // If one of the operands is a symbol and the other is a constant,
620     // build an expression for use by the constraint manager.
621     if (SymbolRef rSym = rhs.getAsLocSymbol()) {
622       // We can only build expressions with symbols on the left,
623       // so we need a reversible operator.
624       if (!BinaryOperator::isComparisonOp(op))
625         return UnknownVal();
626 
627       const llvm::APSInt &lVal = lhs.castAs<loc::ConcreteInt>().getValue();
628       op = BinaryOperator::reverseComparisonOp(op);
629       return makeNonLoc(rSym, op, lVal, resultTy);
630     }
631 
632     // If both operands are constants, just perform the operation.
633     if (Optional<loc::ConcreteInt> rInt = rhs.getAs<loc::ConcreteInt>()) {
634       SVal ResultVal =
635           lhs.castAs<loc::ConcreteInt>().evalBinOp(BasicVals, op, *rInt);
636       if (Optional<NonLoc> Result = ResultVal.getAs<NonLoc>())
637         return evalCastFromNonLoc(*Result, resultTy);
638 
639       assert(!ResultVal.getAs<Loc>() && "Loc-Loc ops should not produce Locs");
640       return UnknownVal();
641     }
642 
643     // Special case comparisons against NULL.
644     // This must come after the test if the RHS is a symbol, which is used to
645     // build constraints. The address of any non-symbolic region is guaranteed
646     // to be non-NULL, as is any label.
647     assert(rhs.getAs<loc::MemRegionVal>() || rhs.getAs<loc::GotoLabel>());
648     if (lhs.isZeroConstant()) {
649       switch (op) {
650       default:
651         break;
652       case BO_EQ:
653       case BO_GT:
654       case BO_GE:
655         return makeTruthVal(false, resultTy);
656       case BO_NE:
657       case BO_LT:
658       case BO_LE:
659         return makeTruthVal(true, resultTy);
660       }
661     }
662 
663     // Comparing an arbitrary integer to a region or label address is
664     // completely unknowable.
665     return UnknownVal();
666   }
667   case loc::MemRegionKind: {
668     if (Optional<loc::ConcreteInt> rInt = rhs.getAs<loc::ConcreteInt>()) {
669       // If one of the operands is a symbol and the other is a constant,
670       // build an expression for use by the constraint manager.
671       if (SymbolRef lSym = lhs.getAsLocSymbol())
672         return MakeSymIntVal(lSym, op, rInt->getValue(), resultTy);
673 
674       // Special case comparisons to NULL.
675       // This must come after the test if the LHS is a symbol, which is used to
676       // build constraints. The address of any non-symbolic region is guaranteed
677       // to be non-NULL.
678       if (rInt->isZeroConstant()) {
679         switch (op) {
680         default:
681           break;
682         case BO_Sub:
683           return evalCastFromLoc(lhs, resultTy);
684         case BO_EQ:
685         case BO_LT:
686         case BO_LE:
687           return makeTruthVal(false, resultTy);
688         case BO_NE:
689         case BO_GT:
690         case BO_GE:
691           return makeTruthVal(true, resultTy);
692         }
693       }
694 
695       // Comparing a region to an arbitrary integer is completely unknowable.
696       return UnknownVal();
697     }
698 
699     // Get both values as regions, if possible.
700     const MemRegion *LeftMR = lhs.getAsRegion();
701     assert(LeftMR && "MemRegionKind SVal doesn't have a region!");
702 
703     const MemRegion *RightMR = rhs.getAsRegion();
704     if (!RightMR)
705       // The RHS is probably a label, which in theory could address a region.
706       // FIXME: we can probably make a more useful statement about non-code
707       // regions, though.
708       return UnknownVal();
709 
710     const MemRegion *LeftBase = LeftMR->getBaseRegion();
711     const MemRegion *RightBase = RightMR->getBaseRegion();
712     const MemSpaceRegion *LeftMS = LeftBase->getMemorySpace();
713     const MemSpaceRegion *RightMS = RightBase->getMemorySpace();
714     const MemSpaceRegion *UnknownMS = MemMgr.getUnknownRegion();
715 
716     // If the two regions are from different known memory spaces they cannot be
717     // equal. Also, assume that no symbolic region (whose memory space is
718     // unknown) is on the stack.
719     if (LeftMS != RightMS &&
720         ((LeftMS != UnknownMS && RightMS != UnknownMS) ||
721          (isa<StackSpaceRegion>(LeftMS) || isa<StackSpaceRegion>(RightMS)))) {
722       switch (op) {
723       default:
724         return UnknownVal();
725       case BO_EQ:
726         return makeTruthVal(false, resultTy);
727       case BO_NE:
728         return makeTruthVal(true, resultTy);
729       }
730     }
731 
732     // If both values wrap regions, see if they're from different base regions.
733     // Note, heap base symbolic regions are assumed to not alias with
734     // each other; for example, we assume that malloc returns different address
735     // on each invocation.
736     if (LeftBase != RightBase &&
737         ((!isa<SymbolicRegion>(LeftBase) && !isa<SymbolicRegion>(RightBase)) ||
738          (isa<HeapSpaceRegion>(LeftMS) || isa<HeapSpaceRegion>(RightMS))) ){
739       switch (op) {
740       default:
741         return UnknownVal();
742       case BO_EQ:
743         return makeTruthVal(false, resultTy);
744       case BO_NE:
745         return makeTruthVal(true, resultTy);
746       }
747     }
748 
749     // Handle special cases for when both regions are element regions.
750     const ElementRegion *RightER = dyn_cast<ElementRegion>(RightMR);
751     const ElementRegion *LeftER = dyn_cast<ElementRegion>(LeftMR);
752     if (RightER && LeftER) {
753       // Next, see if the two ERs have the same super-region and matching types.
754       // FIXME: This should do something useful even if the types don't match,
755       // though if both indexes are constant the RegionRawOffset path will
756       // give the correct answer.
757       if (LeftER->getSuperRegion() == RightER->getSuperRegion() &&
758           LeftER->getElementType() == RightER->getElementType()) {
759         // Get the left index and cast it to the correct type.
760         // If the index is unknown or undefined, bail out here.
761         SVal LeftIndexVal = LeftER->getIndex();
762         Optional<NonLoc> LeftIndex = LeftIndexVal.getAs<NonLoc>();
763         if (!LeftIndex)
764           return UnknownVal();
765         LeftIndexVal = evalCastFromNonLoc(*LeftIndex, ArrayIndexTy);
766         LeftIndex = LeftIndexVal.getAs<NonLoc>();
767         if (!LeftIndex)
768           return UnknownVal();
769 
770         // Do the same for the right index.
771         SVal RightIndexVal = RightER->getIndex();
772         Optional<NonLoc> RightIndex = RightIndexVal.getAs<NonLoc>();
773         if (!RightIndex)
774           return UnknownVal();
775         RightIndexVal = evalCastFromNonLoc(*RightIndex, ArrayIndexTy);
776         RightIndex = RightIndexVal.getAs<NonLoc>();
777         if (!RightIndex)
778           return UnknownVal();
779 
780         // Actually perform the operation.
781         // evalBinOpNN expects the two indexes to already be the right type.
782         return evalBinOpNN(state, op, *LeftIndex, *RightIndex, resultTy);
783       }
784     }
785 
786     // Special handling of the FieldRegions, even with symbolic offsets.
787     const FieldRegion *RightFR = dyn_cast<FieldRegion>(RightMR);
788     const FieldRegion *LeftFR = dyn_cast<FieldRegion>(LeftMR);
789     if (RightFR && LeftFR) {
790       SVal R = evalBinOpFieldRegionFieldRegion(LeftFR, RightFR, op, resultTy,
791                                                *this);
792       if (!R.isUnknown())
793         return R;
794     }
795 
796     // Compare the regions using the raw offsets.
797     RegionOffset LeftOffset = LeftMR->getAsOffset();
798     RegionOffset RightOffset = RightMR->getAsOffset();
799 
800     if (LeftOffset.getRegion() != NULL &&
801         LeftOffset.getRegion() == RightOffset.getRegion() &&
802         !LeftOffset.hasSymbolicOffset() && !RightOffset.hasSymbolicOffset()) {
803       int64_t left = LeftOffset.getOffset();
804       int64_t right = RightOffset.getOffset();
805 
806       switch (op) {
807         default:
808           return UnknownVal();
809         case BO_LT:
810           return makeTruthVal(left < right, resultTy);
811         case BO_GT:
812           return makeTruthVal(left > right, resultTy);
813         case BO_LE:
814           return makeTruthVal(left <= right, resultTy);
815         case BO_GE:
816           return makeTruthVal(left >= right, resultTy);
817         case BO_EQ:
818           return makeTruthVal(left == right, resultTy);
819         case BO_NE:
820           return makeTruthVal(left != right, resultTy);
821       }
822     }
823 
824     // At this point we're not going to get a good answer, but we can try
825     // conjuring an expression instead.
826     SymbolRef LHSSym = lhs.getAsLocSymbol();
827     SymbolRef RHSSym = rhs.getAsLocSymbol();
828     if (LHSSym && RHSSym)
829       return makeNonLoc(LHSSym, op, RHSSym, resultTy);
830 
831     // If we get here, we have no way of comparing the regions.
832     return UnknownVal();
833   }
834   }
835 }
836 
evalBinOpLN(ProgramStateRef state,BinaryOperator::Opcode op,Loc lhs,NonLoc rhs,QualType resultTy)837 SVal SimpleSValBuilder::evalBinOpLN(ProgramStateRef state,
838                                   BinaryOperator::Opcode op,
839                                   Loc lhs, NonLoc rhs, QualType resultTy) {
840   assert(!BinaryOperator::isComparisonOp(op) &&
841          "arguments to comparison ops must be of the same type");
842 
843   // Special case: rhs is a zero constant.
844   if (rhs.isZeroConstant())
845     return lhs;
846 
847   // We are dealing with pointer arithmetic.
848 
849   // Handle pointer arithmetic on constant values.
850   if (Optional<nonloc::ConcreteInt> rhsInt = rhs.getAs<nonloc::ConcreteInt>()) {
851     if (Optional<loc::ConcreteInt> lhsInt = lhs.getAs<loc::ConcreteInt>()) {
852       const llvm::APSInt &leftI = lhsInt->getValue();
853       assert(leftI.isUnsigned());
854       llvm::APSInt rightI(rhsInt->getValue(), /* isUnsigned */ true);
855 
856       // Convert the bitwidth of rightI.  This should deal with overflow
857       // since we are dealing with concrete values.
858       rightI = rightI.extOrTrunc(leftI.getBitWidth());
859 
860       // Offset the increment by the pointer size.
861       llvm::APSInt Multiplicand(rightI.getBitWidth(), /* isUnsigned */ true);
862       rightI *= Multiplicand;
863 
864       // Compute the adjusted pointer.
865       switch (op) {
866         case BO_Add:
867           rightI = leftI + rightI;
868           break;
869         case BO_Sub:
870           rightI = leftI - rightI;
871           break;
872         default:
873           llvm_unreachable("Invalid pointer arithmetic operation");
874       }
875       return loc::ConcreteInt(getBasicValueFactory().getValue(rightI));
876     }
877   }
878 
879   // Handle cases where 'lhs' is a region.
880   if (const MemRegion *region = lhs.getAsRegion()) {
881     rhs = convertToArrayIndex(rhs).castAs<NonLoc>();
882     SVal index = UnknownVal();
883     const MemRegion *superR = 0;
884     QualType elementType;
885 
886     if (const ElementRegion *elemReg = dyn_cast<ElementRegion>(region)) {
887       assert(op == BO_Add || op == BO_Sub);
888       index = evalBinOpNN(state, op, elemReg->getIndex(), rhs,
889                           getArrayIndexType());
890       superR = elemReg->getSuperRegion();
891       elementType = elemReg->getElementType();
892     }
893     else if (isa<SubRegion>(region)) {
894       superR = region;
895       index = rhs;
896       if (resultTy->isAnyPointerType())
897         elementType = resultTy->getPointeeType();
898     }
899 
900     if (Optional<NonLoc> indexV = index.getAs<NonLoc>()) {
901       return loc::MemRegionVal(MemMgr.getElementRegion(elementType, *indexV,
902                                                        superR, getContext()));
903     }
904   }
905   return UnknownVal();
906 }
907 
getKnownValue(ProgramStateRef state,SVal V)908 const llvm::APSInt *SimpleSValBuilder::getKnownValue(ProgramStateRef state,
909                                                    SVal V) {
910   if (V.isUnknownOrUndef())
911     return NULL;
912 
913   if (Optional<loc::ConcreteInt> X = V.getAs<loc::ConcreteInt>())
914     return &X->getValue();
915 
916   if (Optional<nonloc::ConcreteInt> X = V.getAs<nonloc::ConcreteInt>())
917     return &X->getValue();
918 
919   if (SymbolRef Sym = V.getAsSymbol())
920     return state->getConstraintManager().getSymVal(state, Sym);
921 
922   // FIXME: Add support for SymExprs.
923   return NULL;
924 }
925