SimpleSValBuilder.cpp revision 112344ab7f96cf482bce80530676712c282756d5
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
18using namespace clang;
19using namespace ento;
20
21namespace {
22class SimpleSValBuilder : public SValBuilder {
23protected:
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
28public:
29  SimpleSValBuilder(llvm::BumpPtrAllocator &alloc, ASTContext &context,
30                    ProgramStateManager &stateMgr)
31                    : SValBuilder(alloc, context, stateMgr) {}
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
52SValBuilder *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
62SVal 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
68SVal 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
124SVal 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
161SVal 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
170SVal 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
183SVal 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
278SVal 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
510// FIXME: all this logic will change if/when we have MemRegion::getLocation().
511SVal SimpleSValBuilder::evalBinOpLL(ProgramStateRef state,
512                                  BinaryOperator::Opcode op,
513                                  Loc lhs, Loc rhs,
514                                  QualType resultTy) {
515  // Only comparisons and subtractions are valid operations on two pointers.
516  // See [C99 6.5.5 through 6.5.14] or [C++0x 5.6 through 5.15].
517  // However, if a pointer is casted to an integer, evalBinOpNN may end up
518  // calling this function with another operation (PR7527). We don't attempt to
519  // model this for now, but it could be useful, particularly when the
520  // "location" is actually an integer value that's been passed through a void*.
521  if (!(BinaryOperator::isComparisonOp(op) || op == BO_Sub))
522    return UnknownVal();
523
524  // Special cases for when both sides are identical.
525  if (lhs == rhs) {
526    switch (op) {
527    default:
528      llvm_unreachable("Unimplemented operation for two identical values");
529    case BO_Sub:
530      return makeZeroVal(resultTy);
531    case BO_EQ:
532    case BO_LE:
533    case BO_GE:
534      return makeTruthVal(true, resultTy);
535    case BO_NE:
536    case BO_LT:
537    case BO_GT:
538      return makeTruthVal(false, resultTy);
539    }
540  }
541
542  switch (lhs.getSubKind()) {
543  default:
544    llvm_unreachable("Ordering not implemented for this Loc.");
545
546  case loc::GotoLabelKind:
547    // The only thing we know about labels is that they're non-null.
548    if (rhs.isZeroConstant()) {
549      switch (op) {
550      default:
551        break;
552      case BO_Sub:
553        return evalCastFromLoc(lhs, resultTy);
554      case BO_EQ:
555      case BO_LE:
556      case BO_LT:
557        return makeTruthVal(false, resultTy);
558      case BO_NE:
559      case BO_GT:
560      case BO_GE:
561        return makeTruthVal(true, resultTy);
562      }
563    }
564    // There may be two labels for the same location, and a function region may
565    // have the same address as a label at the start of the function (depending
566    // on the ABI).
567    // FIXME: we can probably do a comparison against other MemRegions, though.
568    // FIXME: is there a way to tell if two labels refer to the same location?
569    return UnknownVal();
570
571  case loc::ConcreteIntKind: {
572    // If one of the operands is a symbol and the other is a constant,
573    // build an expression for use by the constraint manager.
574    if (SymbolRef rSym = rhs.getAsLocSymbol()) {
575      // We can only build expressions with symbols on the left,
576      // so we need a reversible operator.
577      if (!BinaryOperator::isComparisonOp(op))
578        return UnknownVal();
579
580      const llvm::APSInt &lVal = lhs.castAs<loc::ConcreteInt>().getValue();
581      op = BinaryOperator::reverseComparisonOp(op);
582      return makeNonLoc(rSym, op, lVal, resultTy);
583    }
584
585    // If both operands are constants, just perform the operation.
586    if (Optional<loc::ConcreteInt> rInt = rhs.getAs<loc::ConcreteInt>()) {
587      SVal ResultVal =
588          lhs.castAs<loc::ConcreteInt>().evalBinOp(BasicVals, op, *rInt);
589      if (Optional<NonLoc> Result = ResultVal.getAs<NonLoc>())
590        return evalCastFromNonLoc(*Result, resultTy);
591
592      assert(!ResultVal.getAs<Loc>() && "Loc-Loc ops should not produce Locs");
593      return UnknownVal();
594    }
595
596    // Special case comparisons against NULL.
597    // This must come after the test if the RHS is a symbol, which is used to
598    // build constraints. The address of any non-symbolic region is guaranteed
599    // to be non-NULL, as is any label.
600    assert(rhs.getAs<loc::MemRegionVal>() || rhs.getAs<loc::GotoLabel>());
601    if (lhs.isZeroConstant()) {
602      switch (op) {
603      default:
604        break;
605      case BO_EQ:
606      case BO_GT:
607      case BO_GE:
608        return makeTruthVal(false, resultTy);
609      case BO_NE:
610      case BO_LT:
611      case BO_LE:
612        return makeTruthVal(true, resultTy);
613      }
614    }
615
616    // Comparing an arbitrary integer to a region or label address is
617    // completely unknowable.
618    return UnknownVal();
619  }
620  case loc::MemRegionKind: {
621    if (Optional<loc::ConcreteInt> rInt = rhs.getAs<loc::ConcreteInt>()) {
622      // If one of the operands is a symbol and the other is a constant,
623      // build an expression for use by the constraint manager.
624      if (SymbolRef lSym = lhs.getAsLocSymbol())
625        return MakeSymIntVal(lSym, op, rInt->getValue(), resultTy);
626
627      // Special case comparisons to NULL.
628      // This must come after the test if the LHS is a symbol, which is used to
629      // build constraints. The address of any non-symbolic region is guaranteed
630      // to be non-NULL.
631      if (rInt->isZeroConstant()) {
632        switch (op) {
633        default:
634          break;
635        case BO_Sub:
636          return evalCastFromLoc(lhs, resultTy);
637        case BO_EQ:
638        case BO_LT:
639        case BO_LE:
640          return makeTruthVal(false, resultTy);
641        case BO_NE:
642        case BO_GT:
643        case BO_GE:
644          return makeTruthVal(true, resultTy);
645        }
646      }
647
648      // Comparing a region to an arbitrary integer is completely unknowable.
649      return UnknownVal();
650    }
651
652    // Get both values as regions, if possible.
653    const MemRegion *LeftMR = lhs.getAsRegion();
654    assert(LeftMR && "MemRegionKind SVal doesn't have a region!");
655
656    const MemRegion *RightMR = rhs.getAsRegion();
657    if (!RightMR)
658      // The RHS is probably a label, which in theory could address a region.
659      // FIXME: we can probably make a more useful statement about non-code
660      // regions, though.
661      return UnknownVal();
662
663    const MemRegion *LeftBase = LeftMR->getBaseRegion();
664    const MemRegion *RightBase = RightMR->getBaseRegion();
665    const MemSpaceRegion *LeftMS = LeftBase->getMemorySpace();
666    const MemSpaceRegion *RightMS = RightBase->getMemorySpace();
667    const MemSpaceRegion *UnknownMS = MemMgr.getUnknownRegion();
668
669    // If the two regions are from different known memory spaces they cannot be
670    // equal. Also, assume that no symbolic region (whose memory space is
671    // unknown) is on the stack.
672    if (LeftMS != RightMS &&
673        ((LeftMS != UnknownMS && RightMS != UnknownMS) ||
674         (isa<StackSpaceRegion>(LeftMS) || isa<StackSpaceRegion>(RightMS)))) {
675      switch (op) {
676      default:
677        return UnknownVal();
678      case BO_EQ:
679        return makeTruthVal(false, resultTy);
680      case BO_NE:
681        return makeTruthVal(true, resultTy);
682      }
683    }
684
685    // If both values wrap regions, see if they're from different base regions.
686    // Note, heap base symbolic regions are assumed to not alias with
687    // each other; for example, we assume that malloc returns different address
688    // on each invocation.
689    if (LeftBase != RightBase &&
690        ((!isa<SymbolicRegion>(LeftBase) && !isa<SymbolicRegion>(RightBase)) ||
691         (isa<HeapSpaceRegion>(LeftMS) || isa<HeapSpaceRegion>(RightMS))) ){
692      switch (op) {
693      default:
694        return UnknownVal();
695      case BO_EQ:
696        return makeTruthVal(false, resultTy);
697      case BO_NE:
698        return makeTruthVal(true, resultTy);
699      }
700    }
701
702    // FIXME: If/when there is a getAsRawOffset() for FieldRegions, this
703    // ElementRegion path and the FieldRegion path below should be unified.
704    if (const ElementRegion *LeftER = dyn_cast<ElementRegion>(LeftMR)) {
705      // First see if the right region is also an ElementRegion.
706      const ElementRegion *RightER = dyn_cast<ElementRegion>(RightMR);
707      if (!RightER)
708        return UnknownVal();
709
710      // Next, see if the two ERs have the same super-region and matching types.
711      // FIXME: This should do something useful even if the types don't match,
712      // though if both indexes are constant the RegionRawOffset path will
713      // give the correct answer.
714      if (LeftER->getSuperRegion() == RightER->getSuperRegion() &&
715          LeftER->getElementType() == RightER->getElementType()) {
716        // Get the left index and cast it to the correct type.
717        // If the index is unknown or undefined, bail out here.
718        SVal LeftIndexVal = LeftER->getIndex();
719        Optional<NonLoc> LeftIndex = LeftIndexVal.getAs<NonLoc>();
720        if (!LeftIndex)
721          return UnknownVal();
722        LeftIndexVal = evalCastFromNonLoc(*LeftIndex, ArrayIndexTy);
723        LeftIndex = LeftIndexVal.getAs<NonLoc>();
724        if (!LeftIndex)
725          return UnknownVal();
726
727        // Do the same for the right index.
728        SVal RightIndexVal = RightER->getIndex();
729        Optional<NonLoc> RightIndex = RightIndexVal.getAs<NonLoc>();
730        if (!RightIndex)
731          return UnknownVal();
732        RightIndexVal = evalCastFromNonLoc(*RightIndex, ArrayIndexTy);
733        RightIndex = RightIndexVal.getAs<NonLoc>();
734        if (!RightIndex)
735          return UnknownVal();
736
737        // Actually perform the operation.
738        // evalBinOpNN expects the two indexes to already be the right type.
739        return evalBinOpNN(state, op, *LeftIndex, *RightIndex, resultTy);
740      }
741
742      // If the element indexes aren't comparable, see if the raw offsets are.
743      RegionRawOffset LeftOffset = LeftER->getAsArrayOffset();
744      RegionRawOffset RightOffset = RightER->getAsArrayOffset();
745
746      if (LeftOffset.getRegion() != NULL &&
747          LeftOffset.getRegion() == RightOffset.getRegion()) {
748        CharUnits left = LeftOffset.getOffset();
749        CharUnits right = RightOffset.getOffset();
750
751        switch (op) {
752        default:
753          return UnknownVal();
754        case BO_LT:
755          return makeTruthVal(left < right, resultTy);
756        case BO_GT:
757          return makeTruthVal(left > right, resultTy);
758        case BO_LE:
759          return makeTruthVal(left <= right, resultTy);
760        case BO_GE:
761          return makeTruthVal(left >= right, resultTy);
762        case BO_EQ:
763          return makeTruthVal(left == right, resultTy);
764        case BO_NE:
765          return makeTruthVal(left != right, resultTy);
766        }
767      }
768
769      // If we get here, we have no way of comparing the ElementRegions.
770    }
771
772    // See if both regions are fields of the same structure.
773    // FIXME: This doesn't handle nesting, inheritance, or Objective-C ivars.
774    if (const FieldRegion *LeftFR = dyn_cast<FieldRegion>(LeftMR)) {
775      // Only comparisons are meaningful here!
776      if (!BinaryOperator::isComparisonOp(op))
777        return UnknownVal();
778
779      // First see if the right region is also a FieldRegion.
780      const FieldRegion *RightFR = dyn_cast<FieldRegion>(RightMR);
781      if (!RightFR)
782        return UnknownVal();
783
784      // Next, see if the two FRs have the same super-region.
785      // FIXME: This doesn't handle casts yet, and simply stripping the casts
786      // doesn't help.
787      if (LeftFR->getSuperRegion() != RightFR->getSuperRegion())
788        return UnknownVal();
789
790      const FieldDecl *LeftFD = LeftFR->getDecl();
791      const FieldDecl *RightFD = RightFR->getDecl();
792      const RecordDecl *RD = LeftFD->getParent();
793
794      // Make sure the two FRs are from the same kind of record. Just in case!
795      // FIXME: This is probably where inheritance would be a problem.
796      if (RD != RightFD->getParent())
797        return UnknownVal();
798
799      // We know for sure that the two fields are not the same, since that
800      // would have given us the same SVal.
801      if (op == BO_EQ)
802        return makeTruthVal(false, resultTy);
803      if (op == BO_NE)
804        return makeTruthVal(true, resultTy);
805
806      // Iterate through the fields and see which one comes first.
807      // [C99 6.7.2.1.13] "Within a structure object, the non-bit-field
808      // members and the units in which bit-fields reside have addresses that
809      // increase in the order in which they are declared."
810      bool leftFirst = (op == BO_LT || op == BO_LE);
811      for (RecordDecl::field_iterator I = RD->field_begin(),
812           E = RD->field_end(); I!=E; ++I) {
813        if (*I == LeftFD)
814          return makeTruthVal(leftFirst, resultTy);
815        if (*I == RightFD)
816          return makeTruthVal(!leftFirst, resultTy);
817      }
818
819      llvm_unreachable("Fields not found in parent record's definition");
820    }
821
822    // At this point we're not going to get a good answer, but we can try
823    // conjuring an expression instead.
824    SymbolRef LHSSym = lhs.getAsLocSymbol();
825    SymbolRef RHSSym = rhs.getAsLocSymbol();
826    if (LHSSym && RHSSym)
827      return makeNonLoc(LHSSym, op, RHSSym, resultTy);
828
829    // If we get here, we have no way of comparing the regions.
830    return UnknownVal();
831  }
832  }
833}
834
835SVal SimpleSValBuilder::evalBinOpLN(ProgramStateRef state,
836                                  BinaryOperator::Opcode op,
837                                  Loc lhs, NonLoc rhs, QualType resultTy) {
838
839  // Special case: rhs is a zero constant.
840  if (rhs.isZeroConstant())
841    return lhs;
842
843  // Special case: 'rhs' is an integer that has the same width as a pointer and
844  // we are using the integer location in a comparison.  Normally this cannot be
845  // triggered, but transfer functions like those for OSCompareAndSwapBarrier32
846  // can generate comparisons that trigger this code.
847  // FIXME: Are all locations guaranteed to have pointer width?
848  if (BinaryOperator::isComparisonOp(op)) {
849    if (Optional<nonloc::ConcreteInt> rhsInt =
850            rhs.getAs<nonloc::ConcreteInt>()) {
851      const llvm::APSInt *x = &rhsInt->getValue();
852      ASTContext &ctx = Context;
853      if (ctx.getTypeSize(ctx.VoidPtrTy) == x->getBitWidth()) {
854        // Convert the signedness of the integer (if necessary).
855        if (x->isSigned())
856          x = &getBasicValueFactory().getValue(*x, true);
857
858        return evalBinOpLL(state, op, lhs, loc::ConcreteInt(*x), resultTy);
859      }
860    }
861    return UnknownVal();
862  }
863
864  // We are dealing with pointer arithmetic.
865
866  // Handle pointer arithmetic on constant values.
867  if (Optional<nonloc::ConcreteInt> rhsInt = rhs.getAs<nonloc::ConcreteInt>()) {
868    if (Optional<loc::ConcreteInt> lhsInt = lhs.getAs<loc::ConcreteInt>()) {
869      const llvm::APSInt &leftI = lhsInt->getValue();
870      assert(leftI.isUnsigned());
871      llvm::APSInt rightI(rhsInt->getValue(), /* isUnsigned */ true);
872
873      // Convert the bitwidth of rightI.  This should deal with overflow
874      // since we are dealing with concrete values.
875      rightI = rightI.extOrTrunc(leftI.getBitWidth());
876
877      // Offset the increment by the pointer size.
878      llvm::APSInt Multiplicand(rightI.getBitWidth(), /* isUnsigned */ true);
879      rightI *= Multiplicand;
880
881      // Compute the adjusted pointer.
882      switch (op) {
883        case BO_Add:
884          rightI = leftI + rightI;
885          break;
886        case BO_Sub:
887          rightI = leftI - rightI;
888          break;
889        default:
890          llvm_unreachable("Invalid pointer arithmetic operation");
891      }
892      return loc::ConcreteInt(getBasicValueFactory().getValue(rightI));
893    }
894  }
895
896  // Handle cases where 'lhs' is a region.
897  if (const MemRegion *region = lhs.getAsRegion()) {
898    rhs = convertToArrayIndex(rhs).castAs<NonLoc>();
899    SVal index = UnknownVal();
900    const MemRegion *superR = 0;
901    QualType elementType;
902
903    if (const ElementRegion *elemReg = dyn_cast<ElementRegion>(region)) {
904      assert(op == BO_Add || op == BO_Sub);
905      index = evalBinOpNN(state, op, elemReg->getIndex(), rhs,
906                          getArrayIndexType());
907      superR = elemReg->getSuperRegion();
908      elementType = elemReg->getElementType();
909    }
910    else if (isa<SubRegion>(region)) {
911      superR = region;
912      index = rhs;
913      if (resultTy->isAnyPointerType())
914        elementType = resultTy->getPointeeType();
915    }
916
917    if (Optional<NonLoc> indexV = index.getAs<NonLoc>()) {
918      return loc::MemRegionVal(MemMgr.getElementRegion(elementType, *indexV,
919                                                       superR, getContext()));
920    }
921  }
922  return UnknownVal();
923}
924
925const llvm::APSInt *SimpleSValBuilder::getKnownValue(ProgramStateRef state,
926                                                   SVal V) {
927  if (V.isUnknownOrUndef())
928    return NULL;
929
930  if (Optional<loc::ConcreteInt> X = V.getAs<loc::ConcreteInt>())
931    return &X->getValue();
932
933  if (Optional<nonloc::ConcreteInt> X = V.getAs<nonloc::ConcreteInt>())
934    return &X->getValue();
935
936  if (SymbolRef Sym = V.getAsSymbol())
937    return state->getConstraintManager().getSymVal(state, Sym);
938
939  // FIXME: Add support for SymExprs.
940  return NULL;
941}
942