SimpleSValBuilder.cpp revision dc84cd5efdd3430efb22546b4ac656aa0540b210
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->isIntegerType())
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->isIntegerType()) {
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
183static BinaryOperator::Opcode NegateComparison(BinaryOperator::Opcode op) {
184  switch (op) {
185  default:
186    llvm_unreachable("Invalid opcode.");
187  case BO_LT: return BO_GE;
188  case BO_GT: return BO_LE;
189  case BO_LE: return BO_GT;
190  case BO_GE: return BO_LT;
191  case BO_EQ: return BO_NE;
192  case BO_NE: return BO_EQ;
193  }
194}
195
196static BinaryOperator::Opcode ReverseComparison(BinaryOperator::Opcode op) {
197  switch (op) {
198  default:
199    llvm_unreachable("Invalid opcode.");
200  case BO_LT: return BO_GT;
201  case BO_GT: return BO_LT;
202  case BO_LE: return BO_GE;
203  case BO_GE: return BO_LE;
204  case BO_EQ:
205  case BO_NE:
206    return op;
207  }
208}
209
210SVal SimpleSValBuilder::MakeSymIntVal(const SymExpr *LHS,
211                                    BinaryOperator::Opcode op,
212                                    const llvm::APSInt &RHS,
213                                    QualType resultTy) {
214  bool isIdempotent = false;
215
216  // Check for a few special cases with known reductions first.
217  switch (op) {
218  default:
219    // We can't reduce this case; just treat it normally.
220    break;
221  case BO_Mul:
222    // a*0 and a*1
223    if (RHS == 0)
224      return makeIntVal(0, resultTy);
225    else if (RHS == 1)
226      isIdempotent = true;
227    break;
228  case BO_Div:
229    // a/0 and a/1
230    if (RHS == 0)
231      // This is also handled elsewhere.
232      return UndefinedVal();
233    else if (RHS == 1)
234      isIdempotent = true;
235    break;
236  case BO_Rem:
237    // a%0 and a%1
238    if (RHS == 0)
239      // This is also handled elsewhere.
240      return UndefinedVal();
241    else if (RHS == 1)
242      return makeIntVal(0, resultTy);
243    break;
244  case BO_Add:
245  case BO_Sub:
246  case BO_Shl:
247  case BO_Shr:
248  case BO_Xor:
249    // a+0, a-0, a<<0, a>>0, a^0
250    if (RHS == 0)
251      isIdempotent = true;
252    break;
253  case BO_And:
254    // a&0 and a&(~0)
255    if (RHS == 0)
256      return makeIntVal(0, resultTy);
257    else if (RHS.isAllOnesValue())
258      isIdempotent = true;
259    break;
260  case BO_Or:
261    // a|0 and a|(~0)
262    if (RHS == 0)
263      isIdempotent = true;
264    else if (RHS.isAllOnesValue()) {
265      const llvm::APSInt &Result = BasicVals.Convert(resultTy, RHS);
266      return nonloc::ConcreteInt(Result);
267    }
268    break;
269  }
270
271  // Idempotent ops (like a*1) can still change the type of an expression.
272  // Wrap the LHS up in a NonLoc again and let evalCastFromNonLoc do the
273  // dirty work.
274  if (isIdempotent)
275      return evalCastFromNonLoc(nonloc::SymbolVal(LHS), resultTy);
276
277  // If we reach this point, the expression cannot be simplified.
278  // Make a SymbolVal for the entire expression, after converting the RHS.
279  const llvm::APSInt *ConvertedRHS = &RHS;
280  if (BinaryOperator::isComparisonOp(op)) {
281    // We're looking for a type big enough to compare the symbolic value
282    // with the given constant.
283    // FIXME: This is an approximation of Sema::UsualArithmeticConversions.
284    ASTContext &Ctx = getContext();
285    QualType SymbolType = LHS->getType();
286    uint64_t ValWidth = RHS.getBitWidth();
287    uint64_t TypeWidth = Ctx.getTypeSize(SymbolType);
288
289    if (ValWidth < TypeWidth) {
290      // If the value is too small, extend it.
291      ConvertedRHS = &BasicVals.Convert(SymbolType, RHS);
292    } else if (ValWidth == TypeWidth) {
293      // If the value is signed but the symbol is unsigned, do the comparison
294      // in unsigned space. [C99 6.3.1.8]
295      // (For the opposite case, the value is already unsigned.)
296      if (RHS.isSigned() && !SymbolType->isSignedIntegerOrEnumerationType())
297        ConvertedRHS = &BasicVals.Convert(SymbolType, RHS);
298    }
299  } else
300    ConvertedRHS = &BasicVals.Convert(resultTy, RHS);
301
302  return makeNonLoc(LHS, op, *ConvertedRHS, resultTy);
303}
304
305SVal SimpleSValBuilder::evalBinOpNN(ProgramStateRef state,
306                                  BinaryOperator::Opcode op,
307                                  NonLoc lhs, NonLoc rhs,
308                                  QualType resultTy)  {
309  NonLoc InputLHS = lhs;
310  NonLoc InputRHS = rhs;
311
312  // Handle trivial case where left-side and right-side are the same.
313  if (lhs == rhs)
314    switch (op) {
315      default:
316        break;
317      case BO_EQ:
318      case BO_LE:
319      case BO_GE:
320        return makeTruthVal(true, resultTy);
321      case BO_LT:
322      case BO_GT:
323      case BO_NE:
324        return makeTruthVal(false, resultTy);
325      case BO_Xor:
326      case BO_Sub:
327        if (resultTy->isIntegralOrEnumerationType())
328          return makeIntVal(0, resultTy);
329        return evalCastFromNonLoc(makeIntVal(0, /*Unsigned=*/false), resultTy);
330      case BO_Or:
331      case BO_And:
332        return evalCastFromNonLoc(lhs, resultTy);
333    }
334
335  while (1) {
336    switch (lhs.getSubKind()) {
337    default:
338      return makeSymExprValNN(state, op, lhs, rhs, resultTy);
339    case nonloc::LocAsIntegerKind: {
340      Loc lhsL = lhs.castAs<nonloc::LocAsInteger>().getLoc();
341      switch (rhs.getSubKind()) {
342        case nonloc::LocAsIntegerKind:
343          return evalBinOpLL(state, op, lhsL,
344                             rhs.castAs<nonloc::LocAsInteger>().getLoc(),
345                             resultTy);
346        case nonloc::ConcreteIntKind: {
347          // Transform the integer into a location and compare.
348          llvm::APSInt i = rhs.castAs<nonloc::ConcreteInt>().getValue();
349          BasicVals.getAPSIntType(Context.VoidPtrTy).apply(i);
350          return evalBinOpLL(state, op, lhsL, makeLoc(i), resultTy);
351        }
352        default:
353          switch (op) {
354            case BO_EQ:
355              return makeTruthVal(false, resultTy);
356            case BO_NE:
357              return makeTruthVal(true, resultTy);
358            default:
359              // This case also handles pointer arithmetic.
360              return makeSymExprValNN(state, op, InputLHS, InputRHS, resultTy);
361          }
362      }
363    }
364    case nonloc::ConcreteIntKind: {
365      llvm::APSInt LHSValue = lhs.castAs<nonloc::ConcreteInt>().getValue();
366
367      // If we're dealing with two known constants, just perform the operation.
368      if (const llvm::APSInt *KnownRHSValue = getKnownValue(state, rhs)) {
369        llvm::APSInt RHSValue = *KnownRHSValue;
370        if (BinaryOperator::isComparisonOp(op)) {
371          // We're looking for a type big enough to compare the two values.
372          // FIXME: This is not correct. char + short will result in a promotion
373          // to int. Unfortunately we have lost types by this point.
374          APSIntType CompareType = std::max(APSIntType(LHSValue),
375                                            APSIntType(RHSValue));
376          CompareType.apply(LHSValue);
377          CompareType.apply(RHSValue);
378        } else if (!BinaryOperator::isShiftOp(op)) {
379          APSIntType IntType = BasicVals.getAPSIntType(resultTy);
380          IntType.apply(LHSValue);
381          IntType.apply(RHSValue);
382        }
383
384        const llvm::APSInt *Result =
385          BasicVals.evalAPSInt(op, LHSValue, RHSValue);
386        if (!Result)
387          return UndefinedVal();
388
389        return nonloc::ConcreteInt(*Result);
390      }
391
392      // Swap the left and right sides and flip the operator if doing so
393      // allows us to better reason about the expression (this is a form
394      // of expression canonicalization).
395      // While we're at it, catch some special cases for non-commutative ops.
396      switch (op) {
397      case BO_LT:
398      case BO_GT:
399      case BO_LE:
400      case BO_GE:
401        op = ReverseComparison(op);
402        // FALL-THROUGH
403      case BO_EQ:
404      case BO_NE:
405      case BO_Add:
406      case BO_Mul:
407      case BO_And:
408      case BO_Xor:
409      case BO_Or:
410        std::swap(lhs, rhs);
411        continue;
412      case BO_Shr:
413        // (~0)>>a
414        if (LHSValue.isAllOnesValue() && LHSValue.isSigned())
415          return evalCastFromNonLoc(lhs, resultTy);
416        // FALL-THROUGH
417      case BO_Shl:
418        // 0<<a and 0>>a
419        if (LHSValue == 0)
420          return evalCastFromNonLoc(lhs, resultTy);
421        return makeSymExprValNN(state, op, InputLHS, InputRHS, resultTy);
422      default:
423        return makeSymExprValNN(state, op, InputLHS, InputRHS, resultTy);
424      }
425    }
426    case nonloc::SymbolValKind: {
427      // We only handle LHS as simple symbols or SymIntExprs.
428      SymbolRef Sym = lhs.castAs<nonloc::SymbolVal>().getSymbol();
429
430      // LHS is a symbolic expression.
431      if (const SymIntExpr *symIntExpr = dyn_cast<SymIntExpr>(Sym)) {
432
433        // Is this a logical not? (!x is represented as x == 0.)
434        if (op == BO_EQ && rhs.isZeroConstant()) {
435          // We know how to negate certain expressions. Simplify them here.
436
437          BinaryOperator::Opcode opc = symIntExpr->getOpcode();
438          switch (opc) {
439          default:
440            // We don't know how to negate this operation.
441            // Just handle it as if it were a normal comparison to 0.
442            break;
443          case BO_LAnd:
444          case BO_LOr:
445            llvm_unreachable("Logical operators handled by branching logic.");
446          case BO_Assign:
447          case BO_MulAssign:
448          case BO_DivAssign:
449          case BO_RemAssign:
450          case BO_AddAssign:
451          case BO_SubAssign:
452          case BO_ShlAssign:
453          case BO_ShrAssign:
454          case BO_AndAssign:
455          case BO_XorAssign:
456          case BO_OrAssign:
457          case BO_Comma:
458            llvm_unreachable("'=' and ',' operators handled by ExprEngine.");
459          case BO_PtrMemD:
460          case BO_PtrMemI:
461            llvm_unreachable("Pointer arithmetic not handled here.");
462          case BO_LT:
463          case BO_GT:
464          case BO_LE:
465          case BO_GE:
466          case BO_EQ:
467          case BO_NE:
468            // Negate the comparison and make a value.
469            opc = NegateComparison(opc);
470            assert(symIntExpr->getType() == resultTy);
471            return makeNonLoc(symIntExpr->getLHS(), opc,
472                symIntExpr->getRHS(), resultTy);
473          }
474        }
475
476        // For now, only handle expressions whose RHS is a constant.
477        if (const llvm::APSInt *RHSValue = getKnownValue(state, rhs)) {
478          // If both the LHS and the current expression are additive,
479          // fold their constants and try again.
480          if (BinaryOperator::isAdditiveOp(op)) {
481            BinaryOperator::Opcode lop = symIntExpr->getOpcode();
482            if (BinaryOperator::isAdditiveOp(lop)) {
483              // Convert the two constants to a common type, then combine them.
484
485              // resultTy may not be the best type to convert to, but it's
486              // probably the best choice in expressions with mixed type
487              // (such as x+1U+2LL). The rules for implicit conversions should
488              // choose a reasonable type to preserve the expression, and will
489              // at least match how the value is going to be used.
490              APSIntType IntType = BasicVals.getAPSIntType(resultTy);
491              const llvm::APSInt &first = IntType.convert(symIntExpr->getRHS());
492              const llvm::APSInt &second = IntType.convert(*RHSValue);
493
494              const llvm::APSInt *newRHS;
495              if (lop == op)
496                newRHS = BasicVals.evalAPSInt(BO_Add, first, second);
497              else
498                newRHS = BasicVals.evalAPSInt(BO_Sub, first, second);
499
500              assert(newRHS && "Invalid operation despite common type!");
501              rhs = nonloc::ConcreteInt(*newRHS);
502              lhs = nonloc::SymbolVal(symIntExpr->getLHS());
503              op = lop;
504              continue;
505            }
506          }
507
508          // Otherwise, make a SymIntExpr out of the expression.
509          return MakeSymIntVal(symIntExpr, op, *RHSValue, resultTy);
510        }
511
512
513      } else if (isa<SymbolData>(Sym)) {
514        // Does the symbol simplify to a constant?  If so, "fold" the constant
515        // by setting 'lhs' to a ConcreteInt and try again.
516        if (const llvm::APSInt *Constant = state->getConstraintManager()
517                                                  .getSymVal(state, Sym)) {
518          lhs = nonloc::ConcreteInt(*Constant);
519          continue;
520        }
521
522        // Is the RHS a constant?
523        if (const llvm::APSInt *RHSValue = getKnownValue(state, rhs))
524          return MakeSymIntVal(Sym, op, *RHSValue, resultTy);
525      }
526
527      // Give up -- this is not a symbolic expression we can handle.
528      return makeSymExprValNN(state, op, InputLHS, InputRHS, resultTy);
529    }
530    }
531  }
532}
533
534// FIXME: all this logic will change if/when we have MemRegion::getLocation().
535SVal SimpleSValBuilder::evalBinOpLL(ProgramStateRef state,
536                                  BinaryOperator::Opcode op,
537                                  Loc lhs, Loc rhs,
538                                  QualType resultTy) {
539  // Only comparisons and subtractions are valid operations on two pointers.
540  // See [C99 6.5.5 through 6.5.14] or [C++0x 5.6 through 5.15].
541  // However, if a pointer is casted to an integer, evalBinOpNN may end up
542  // calling this function with another operation (PR7527). We don't attempt to
543  // model this for now, but it could be useful, particularly when the
544  // "location" is actually an integer value that's been passed through a void*.
545  if (!(BinaryOperator::isComparisonOp(op) || op == BO_Sub))
546    return UnknownVal();
547
548  // Special cases for when both sides are identical.
549  if (lhs == rhs) {
550    switch (op) {
551    default:
552      llvm_unreachable("Unimplemented operation for two identical values");
553    case BO_Sub:
554      return makeZeroVal(resultTy);
555    case BO_EQ:
556    case BO_LE:
557    case BO_GE:
558      return makeTruthVal(true, resultTy);
559    case BO_NE:
560    case BO_LT:
561    case BO_GT:
562      return makeTruthVal(false, resultTy);
563    }
564  }
565
566  switch (lhs.getSubKind()) {
567  default:
568    llvm_unreachable("Ordering not implemented for this Loc.");
569
570  case loc::GotoLabelKind:
571    // The only thing we know about labels is that they're non-null.
572    if (rhs.isZeroConstant()) {
573      switch (op) {
574      default:
575        break;
576      case BO_Sub:
577        return evalCastFromLoc(lhs, resultTy);
578      case BO_EQ:
579      case BO_LE:
580      case BO_LT:
581        return makeTruthVal(false, resultTy);
582      case BO_NE:
583      case BO_GT:
584      case BO_GE:
585        return makeTruthVal(true, resultTy);
586      }
587    }
588    // There may be two labels for the same location, and a function region may
589    // have the same address as a label at the start of the function (depending
590    // on the ABI).
591    // FIXME: we can probably do a comparison against other MemRegions, though.
592    // FIXME: is there a way to tell if two labels refer to the same location?
593    return UnknownVal();
594
595  case loc::ConcreteIntKind: {
596    // If one of the operands is a symbol and the other is a constant,
597    // build an expression for use by the constraint manager.
598    if (SymbolRef rSym = rhs.getAsLocSymbol()) {
599      // We can only build expressions with symbols on the left,
600      // so we need a reversible operator.
601      if (!BinaryOperator::isComparisonOp(op))
602        return UnknownVal();
603
604      const llvm::APSInt &lVal = lhs.castAs<loc::ConcreteInt>().getValue();
605      return makeNonLoc(rSym, ReverseComparison(op), lVal, resultTy);
606    }
607
608    // If both operands are constants, just perform the operation.
609    if (Optional<loc::ConcreteInt> rInt = rhs.getAs<loc::ConcreteInt>()) {
610      SVal ResultVal =
611          lhs.castAs<loc::ConcreteInt>().evalBinOp(BasicVals, op, *rInt);
612      if (Optional<Loc> Result = ResultVal.getAs<Loc>())
613        return evalCastFromLoc(*Result, resultTy);
614      else
615        return UnknownVal();
616    }
617
618    // Special case comparisons against NULL.
619    // This must come after the test if the RHS is a symbol, which is used to
620    // build constraints. The address of any non-symbolic region is guaranteed
621    // to be non-NULL, as is any label.
622    assert(rhs.getAs<loc::MemRegionVal>() || rhs.getAs<loc::GotoLabel>());
623    if (lhs.isZeroConstant()) {
624      switch (op) {
625      default:
626        break;
627      case BO_EQ:
628      case BO_GT:
629      case BO_GE:
630        return makeTruthVal(false, resultTy);
631      case BO_NE:
632      case BO_LT:
633      case BO_LE:
634        return makeTruthVal(true, resultTy);
635      }
636    }
637
638    // Comparing an arbitrary integer to a region or label address is
639    // completely unknowable.
640    return UnknownVal();
641  }
642  case loc::MemRegionKind: {
643    if (Optional<loc::ConcreteInt> rInt = rhs.getAs<loc::ConcreteInt>()) {
644      // If one of the operands is a symbol and the other is a constant,
645      // build an expression for use by the constraint manager.
646      if (SymbolRef lSym = lhs.getAsLocSymbol())
647        return MakeSymIntVal(lSym, op, rInt->getValue(), resultTy);
648
649      // Special case comparisons to NULL.
650      // This must come after the test if the LHS is a symbol, which is used to
651      // build constraints. The address of any non-symbolic region is guaranteed
652      // to be non-NULL.
653      if (rInt->isZeroConstant()) {
654        switch (op) {
655        default:
656          break;
657        case BO_Sub:
658          return evalCastFromLoc(lhs, resultTy);
659        case BO_EQ:
660        case BO_LT:
661        case BO_LE:
662          return makeTruthVal(false, resultTy);
663        case BO_NE:
664        case BO_GT:
665        case BO_GE:
666          return makeTruthVal(true, resultTy);
667        }
668      }
669
670      // Comparing a region to an arbitrary integer is completely unknowable.
671      return UnknownVal();
672    }
673
674    // Get both values as regions, if possible.
675    const MemRegion *LeftMR = lhs.getAsRegion();
676    assert(LeftMR && "MemRegionKind SVal doesn't have a region!");
677
678    const MemRegion *RightMR = rhs.getAsRegion();
679    if (!RightMR)
680      // The RHS is probably a label, which in theory could address a region.
681      // FIXME: we can probably make a more useful statement about non-code
682      // regions, though.
683      return UnknownVal();
684
685    const MemSpaceRegion *LeftMS = LeftMR->getMemorySpace();
686    const MemSpaceRegion *RightMS = RightMR->getMemorySpace();
687    const MemSpaceRegion *UnknownMS = MemMgr.getUnknownRegion();
688    const MemRegion *LeftBase = LeftMR->getBaseRegion();
689    const MemRegion *RightBase = RightMR->getBaseRegion();
690
691    // If the two regions are from different known memory spaces they cannot be
692    // equal. Also, assume that no symbolic region (whose memory space is
693    // unknown) is on the stack.
694    if (LeftMS != RightMS &&
695        ((LeftMS != UnknownMS && RightMS != UnknownMS) ||
696         (isa<StackSpaceRegion>(LeftMS) || isa<StackSpaceRegion>(RightMS)))) {
697      switch (op) {
698      default:
699        return UnknownVal();
700      case BO_EQ:
701        return makeTruthVal(false, resultTy);
702      case BO_NE:
703        return makeTruthVal(true, resultTy);
704      }
705    }
706
707    // If both values wrap regions, see if they're from different base regions.
708    // Note, heap base symbolic regions are assumed to not alias with
709    // each other; for example, we assume that malloc returns different address
710    // on each invocation.
711    if (LeftBase != RightBase &&
712        ((!isa<SymbolicRegion>(LeftBase) && !isa<SymbolicRegion>(RightBase)) ||
713         (isa<HeapSpaceRegion>(LeftMS) || isa<HeapSpaceRegion>(RightMS))) ){
714      switch (op) {
715      default:
716        return UnknownVal();
717      case BO_EQ:
718        return makeTruthVal(false, resultTy);
719      case BO_NE:
720        return makeTruthVal(true, resultTy);
721      }
722    }
723
724    // FIXME: If/when there is a getAsRawOffset() for FieldRegions, this
725    // ElementRegion path and the FieldRegion path below should be unified.
726    if (const ElementRegion *LeftER = dyn_cast<ElementRegion>(LeftMR)) {
727      // First see if the right region is also an ElementRegion.
728      const ElementRegion *RightER = dyn_cast<ElementRegion>(RightMR);
729      if (!RightER)
730        return UnknownVal();
731
732      // Next, see if the two ERs have the same super-region and matching types.
733      // FIXME: This should do something useful even if the types don't match,
734      // though if both indexes are constant the RegionRawOffset path will
735      // give the correct answer.
736      if (LeftER->getSuperRegion() == RightER->getSuperRegion() &&
737          LeftER->getElementType() == RightER->getElementType()) {
738        // Get the left index and cast it to the correct type.
739        // If the index is unknown or undefined, bail out here.
740        SVal LeftIndexVal = LeftER->getIndex();
741        Optional<NonLoc> LeftIndex = LeftIndexVal.getAs<NonLoc>();
742        if (!LeftIndex)
743          return UnknownVal();
744        LeftIndexVal = evalCastFromNonLoc(*LeftIndex, ArrayIndexTy);
745        LeftIndex = LeftIndexVal.getAs<NonLoc>();
746        if (!LeftIndex)
747          return UnknownVal();
748
749        // Do the same for the right index.
750        SVal RightIndexVal = RightER->getIndex();
751        Optional<NonLoc> RightIndex = RightIndexVal.getAs<NonLoc>();
752        if (!RightIndex)
753          return UnknownVal();
754        RightIndexVal = evalCastFromNonLoc(*RightIndex, ArrayIndexTy);
755        RightIndex = RightIndexVal.getAs<NonLoc>();
756        if (!RightIndex)
757          return UnknownVal();
758
759        // Actually perform the operation.
760        // evalBinOpNN expects the two indexes to already be the right type.
761        return evalBinOpNN(state, op, *LeftIndex, *RightIndex, resultTy);
762      }
763
764      // If the element indexes aren't comparable, see if the raw offsets are.
765      RegionRawOffset LeftOffset = LeftER->getAsArrayOffset();
766      RegionRawOffset RightOffset = RightER->getAsArrayOffset();
767
768      if (LeftOffset.getRegion() != NULL &&
769          LeftOffset.getRegion() == RightOffset.getRegion()) {
770        CharUnits left = LeftOffset.getOffset();
771        CharUnits right = RightOffset.getOffset();
772
773        switch (op) {
774        default:
775          return UnknownVal();
776        case BO_LT:
777          return makeTruthVal(left < right, resultTy);
778        case BO_GT:
779          return makeTruthVal(left > right, resultTy);
780        case BO_LE:
781          return makeTruthVal(left <= right, resultTy);
782        case BO_GE:
783          return makeTruthVal(left >= right, resultTy);
784        case BO_EQ:
785          return makeTruthVal(left == right, resultTy);
786        case BO_NE:
787          return makeTruthVal(left != right, resultTy);
788        }
789      }
790
791      // If we get here, we have no way of comparing the ElementRegions.
792      return UnknownVal();
793    }
794
795    // See if both regions are fields of the same structure.
796    // FIXME: This doesn't handle nesting, inheritance, or Objective-C ivars.
797    if (const FieldRegion *LeftFR = dyn_cast<FieldRegion>(LeftMR)) {
798      // Only comparisons are meaningful here!
799      if (!BinaryOperator::isComparisonOp(op))
800        return UnknownVal();
801
802      // First see if the right region is also a FieldRegion.
803      const FieldRegion *RightFR = dyn_cast<FieldRegion>(RightMR);
804      if (!RightFR)
805        return UnknownVal();
806
807      // Next, see if the two FRs have the same super-region.
808      // FIXME: This doesn't handle casts yet, and simply stripping the casts
809      // doesn't help.
810      if (LeftFR->getSuperRegion() != RightFR->getSuperRegion())
811        return UnknownVal();
812
813      const FieldDecl *LeftFD = LeftFR->getDecl();
814      const FieldDecl *RightFD = RightFR->getDecl();
815      const RecordDecl *RD = LeftFD->getParent();
816
817      // Make sure the two FRs are from the same kind of record. Just in case!
818      // FIXME: This is probably where inheritance would be a problem.
819      if (RD != RightFD->getParent())
820        return UnknownVal();
821
822      // We know for sure that the two fields are not the same, since that
823      // would have given us the same SVal.
824      if (op == BO_EQ)
825        return makeTruthVal(false, resultTy);
826      if (op == BO_NE)
827        return makeTruthVal(true, resultTy);
828
829      // Iterate through the fields and see which one comes first.
830      // [C99 6.7.2.1.13] "Within a structure object, the non-bit-field
831      // members and the units in which bit-fields reside have addresses that
832      // increase in the order in which they are declared."
833      bool leftFirst = (op == BO_LT || op == BO_LE);
834      for (RecordDecl::field_iterator I = RD->field_begin(),
835           E = RD->field_end(); I!=E; ++I) {
836        if (*I == LeftFD)
837          return makeTruthVal(leftFirst, resultTy);
838        if (*I == RightFD)
839          return makeTruthVal(!leftFirst, resultTy);
840      }
841
842      llvm_unreachable("Fields not found in parent record's definition");
843    }
844
845    // If we get here, we have no way of comparing the regions.
846    return UnknownVal();
847  }
848  }
849}
850
851SVal SimpleSValBuilder::evalBinOpLN(ProgramStateRef state,
852                                  BinaryOperator::Opcode op,
853                                  Loc lhs, NonLoc rhs, QualType resultTy) {
854
855  // Special case: rhs is a zero constant.
856  if (rhs.isZeroConstant())
857    return lhs;
858
859  // Special case: 'rhs' is an integer that has the same width as a pointer and
860  // we are using the integer location in a comparison.  Normally this cannot be
861  // triggered, but transfer functions like those for OSCompareAndSwapBarrier32
862  // can generate comparisons that trigger this code.
863  // FIXME: Are all locations guaranteed to have pointer width?
864  if (BinaryOperator::isComparisonOp(op)) {
865    if (Optional<nonloc::ConcreteInt> rhsInt =
866            rhs.getAs<nonloc::ConcreteInt>()) {
867      const llvm::APSInt *x = &rhsInt->getValue();
868      ASTContext &ctx = Context;
869      if (ctx.getTypeSize(ctx.VoidPtrTy) == x->getBitWidth()) {
870        // Convert the signedness of the integer (if necessary).
871        if (x->isSigned())
872          x = &getBasicValueFactory().getValue(*x, true);
873
874        return evalBinOpLL(state, op, lhs, loc::ConcreteInt(*x), resultTy);
875      }
876    }
877    return UnknownVal();
878  }
879
880  // We are dealing with pointer arithmetic.
881
882  // Handle pointer arithmetic on constant values.
883  if (Optional<nonloc::ConcreteInt> rhsInt = rhs.getAs<nonloc::ConcreteInt>()) {
884    if (Optional<loc::ConcreteInt> lhsInt = lhs.getAs<loc::ConcreteInt>()) {
885      const llvm::APSInt &leftI = lhsInt->getValue();
886      assert(leftI.isUnsigned());
887      llvm::APSInt rightI(rhsInt->getValue(), /* isUnsigned */ true);
888
889      // Convert the bitwidth of rightI.  This should deal with overflow
890      // since we are dealing with concrete values.
891      rightI = rightI.extOrTrunc(leftI.getBitWidth());
892
893      // Offset the increment by the pointer size.
894      llvm::APSInt Multiplicand(rightI.getBitWidth(), /* isUnsigned */ true);
895      rightI *= Multiplicand;
896
897      // Compute the adjusted pointer.
898      switch (op) {
899        case BO_Add:
900          rightI = leftI + rightI;
901          break;
902        case BO_Sub:
903          rightI = leftI - rightI;
904          break;
905        default:
906          llvm_unreachable("Invalid pointer arithmetic operation");
907      }
908      return loc::ConcreteInt(getBasicValueFactory().getValue(rightI));
909    }
910  }
911
912  // Handle cases where 'lhs' is a region.
913  if (const MemRegion *region = lhs.getAsRegion()) {
914    rhs = convertToArrayIndex(rhs).castAs<NonLoc>();
915    SVal index = UnknownVal();
916    const MemRegion *superR = 0;
917    QualType elementType;
918
919    if (const ElementRegion *elemReg = dyn_cast<ElementRegion>(region)) {
920      assert(op == BO_Add || op == BO_Sub);
921      index = evalBinOpNN(state, op, elemReg->getIndex(), rhs,
922                          getArrayIndexType());
923      superR = elemReg->getSuperRegion();
924      elementType = elemReg->getElementType();
925    }
926    else if (isa<SubRegion>(region)) {
927      superR = region;
928      index = rhs;
929      if (resultTy->isAnyPointerType())
930        elementType = resultTy->getPointeeType();
931    }
932
933    if (Optional<NonLoc> indexV = index.getAs<NonLoc>()) {
934      return loc::MemRegionVal(MemMgr.getElementRegion(elementType, *indexV,
935                                                       superR, getContext()));
936    }
937  }
938  return UnknownVal();
939}
940
941const llvm::APSInt *SimpleSValBuilder::getKnownValue(ProgramStateRef state,
942                                                   SVal V) {
943  if (V.isUnknownOrUndef())
944    return NULL;
945
946  if (Optional<loc::ConcreteInt> X = V.getAs<loc::ConcreteInt>())
947    return &X->getValue();
948
949  if (Optional<nonloc::ConcreteInt> X = V.getAs<nonloc::ConcreteInt>())
950    return &X->getValue();
951
952  if (SymbolRef Sym = V.getAsSymbol())
953    return state->getConstraintManager().getSymVal(state, Sym);
954
955  // FIXME: Add support for SymExprs.
956  return NULL;
957}
958