SimpleSValBuilder.cpp revision c838fd2ab889ffbb82c90da0cd634ef75b614b2c
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/ProgramState.h"
16
17using namespace clang;
18using namespace ento;
19
20namespace {
21class SimpleSValBuilder : public SValBuilder {
22protected:
23  virtual SVal dispatchCast(SVal val, QualType castTy);
24  virtual SVal evalCastFromNonLoc(NonLoc val, QualType castTy);
25  virtual SVal evalCastFromLoc(Loc val, QualType castTy);
26
27public:
28  SimpleSValBuilder(llvm::BumpPtrAllocator &alloc, ASTContext &context,
29                    ProgramStateManager &stateMgr)
30                    : SValBuilder(alloc, context, stateMgr) {}
31  virtual ~SimpleSValBuilder() {}
32
33  virtual SVal evalMinus(NonLoc val);
34  virtual SVal evalComplement(NonLoc val);
35  virtual SVal evalBinOpNN(ProgramStateRef state, BinaryOperator::Opcode op,
36                           NonLoc lhs, NonLoc rhs, QualType resultTy);
37  virtual SVal evalBinOpLL(ProgramStateRef state, BinaryOperator::Opcode op,
38                           Loc lhs, Loc rhs, QualType resultTy);
39  virtual SVal evalBinOpLN(ProgramStateRef state, BinaryOperator::Opcode op,
40                           Loc lhs, NonLoc rhs, QualType resultTy);
41
42  /// getKnownValue - evaluates a given SVal. If the SVal has only one possible
43  ///  (integer) value, that value is returned. Otherwise, returns NULL.
44  virtual const llvm::APSInt *getKnownValue(ProgramStateRef state, SVal V);
45
46  SVal MakeSymIntVal(const SymExpr *LHS, BinaryOperator::Opcode op,
47                     const llvm::APSInt &RHS, QualType resultTy);
48};
49} // end anonymous namespace
50
51SValBuilder *ento::createSimpleSValBuilder(llvm::BumpPtrAllocator &alloc,
52                                           ASTContext &context,
53                                           ProgramStateManager &stateMgr) {
54  return new SimpleSValBuilder(alloc, context, stateMgr);
55}
56
57//===----------------------------------------------------------------------===//
58// Transfer function for Casts.
59//===----------------------------------------------------------------------===//
60
61SVal SimpleSValBuilder::dispatchCast(SVal Val, QualType CastTy) {
62  assert(isa<Loc>(&Val) || isa<NonLoc>(&Val));
63  return isa<Loc>(Val) ? evalCastFromLoc(cast<Loc>(Val), CastTy)
64                       : evalCastFromNonLoc(cast<NonLoc>(Val), CastTy);
65}
66
67SVal SimpleSValBuilder::evalCastFromNonLoc(NonLoc val, QualType castTy) {
68
69  bool isLocType = Loc::isLocType(castTy);
70
71  if (nonloc::LocAsInteger *LI = dyn_cast<nonloc::LocAsInteger>(&val)) {
72    if (isLocType)
73      return LI->getLoc();
74
75    // FIXME: Correctly support promotions/truncations.
76    unsigned castSize = Context.getTypeSize(castTy);
77    if (castSize == LI->getNumBits())
78      return val;
79    return makeLocAsInteger(LI->getLoc(), castSize);
80  }
81
82  if (const SymExpr *se = val.getAsSymbolicExpression()) {
83    QualType T = Context.getCanonicalType(se->getType(Context));
84    // If types are the same or both are integers, ignore the cast.
85    // FIXME: Remove this hack when we support symbolic truncation/extension.
86    // HACK: If both castTy and T are integers, ignore the cast.  This is
87    // not a permanent solution.  Eventually we want to precisely handle
88    // extension/truncation of symbolic integers.  This prevents us from losing
89    // precision when we assign 'x = y' and 'y' is symbolic and x and y are
90    // different integer types.
91   if (haveSameType(T, castTy))
92      return val;
93
94    if (!isLocType)
95      return makeNonLoc(se, T, castTy);
96    return UnknownVal();
97  }
98
99  // If value is a non integer constant, produce unknown.
100  if (!isa<nonloc::ConcreteInt>(val))
101    return UnknownVal();
102
103  // Only handle casts from integers to integers - if val is an integer constant
104  // being cast to a non integer type, produce unknown.
105  if (!isLocType && !castTy->isIntegerType())
106    return UnknownVal();
107
108  llvm::APSInt i = cast<nonloc::ConcreteInt>(val).getValue();
109  i.setIsUnsigned(castTy->isUnsignedIntegerOrEnumerationType() ||
110                  Loc::isLocType(castTy));
111  i = i.extOrTrunc(Context.getTypeSize(castTy));
112
113  if (isLocType)
114    return makeIntLocVal(i);
115  else
116    return makeIntVal(i);
117}
118
119SVal SimpleSValBuilder::evalCastFromLoc(Loc val, QualType castTy) {
120
121  // Casts from pointers -> pointers, just return the lval.
122  //
123  // Casts from pointers -> references, just return the lval.  These
124  //   can be introduced by the frontend for corner cases, e.g
125  //   casting from va_list* to __builtin_va_list&.
126  //
127  if (Loc::isLocType(castTy) || castTy->isReferenceType())
128    return val;
129
130  // FIXME: Handle transparent unions where a value can be "transparently"
131  //  lifted into a union type.
132  if (castTy->isUnionType())
133    return UnknownVal();
134
135  if (castTy->isIntegerType()) {
136    unsigned BitWidth = Context.getTypeSize(castTy);
137
138    if (!isa<loc::ConcreteInt>(val))
139      return makeLocAsInteger(val, BitWidth);
140
141    llvm::APSInt i = cast<loc::ConcreteInt>(val).getValue();
142    i.setIsUnsigned(castTy->isUnsignedIntegerOrEnumerationType() ||
143                    Loc::isLocType(castTy));
144    i = i.extOrTrunc(BitWidth);
145    return makeIntVal(i);
146  }
147
148  // All other cases: return 'UnknownVal'.  This includes casting pointers
149  // to floats, which is probably badness it itself, but this is a good
150  // intermediate solution until we do something better.
151  return UnknownVal();
152}
153
154//===----------------------------------------------------------------------===//
155// Transfer function for unary operators.
156//===----------------------------------------------------------------------===//
157
158SVal SimpleSValBuilder::evalMinus(NonLoc val) {
159  switch (val.getSubKind()) {
160  case nonloc::ConcreteIntKind:
161    return cast<nonloc::ConcreteInt>(val).evalMinus(*this);
162  default:
163    return UnknownVal();
164  }
165}
166
167SVal SimpleSValBuilder::evalComplement(NonLoc X) {
168  switch (X.getSubKind()) {
169  case nonloc::ConcreteIntKind:
170    return cast<nonloc::ConcreteInt>(X).evalComplement(*this);
171  default:
172    return UnknownVal();
173  }
174}
175
176//===----------------------------------------------------------------------===//
177// Transfer function for binary operators.
178//===----------------------------------------------------------------------===//
179
180static BinaryOperator::Opcode NegateComparison(BinaryOperator::Opcode op) {
181  switch (op) {
182  default:
183    llvm_unreachable("Invalid opcode.");
184  case BO_LT: return BO_GE;
185  case BO_GT: return BO_LE;
186  case BO_LE: return BO_GT;
187  case BO_GE: return BO_LT;
188  case BO_EQ: return BO_NE;
189  case BO_NE: return BO_EQ;
190  }
191}
192
193static BinaryOperator::Opcode ReverseComparison(BinaryOperator::Opcode op) {
194  switch (op) {
195  default:
196    llvm_unreachable("Invalid opcode.");
197  case BO_LT: return BO_GT;
198  case BO_GT: return BO_LT;
199  case BO_LE: return BO_GE;
200  case BO_GE: return BO_LE;
201  case BO_EQ:
202  case BO_NE:
203    return op;
204  }
205}
206
207SVal SimpleSValBuilder::MakeSymIntVal(const SymExpr *LHS,
208                                    BinaryOperator::Opcode op,
209                                    const llvm::APSInt &RHS,
210                                    QualType resultTy) {
211  bool isIdempotent = false;
212
213  // Check for a few special cases with known reductions first.
214  switch (op) {
215  default:
216    // We can't reduce this case; just treat it normally.
217    break;
218  case BO_Mul:
219    // a*0 and a*1
220    if (RHS == 0)
221      return makeIntVal(0, resultTy);
222    else if (RHS == 1)
223      isIdempotent = true;
224    break;
225  case BO_Div:
226    // a/0 and a/1
227    if (RHS == 0)
228      // This is also handled elsewhere.
229      return UndefinedVal();
230    else if (RHS == 1)
231      isIdempotent = true;
232    break;
233  case BO_Rem:
234    // a%0 and a%1
235    if (RHS == 0)
236      // This is also handled elsewhere.
237      return UndefinedVal();
238    else if (RHS == 1)
239      return makeIntVal(0, resultTy);
240    break;
241  case BO_Add:
242  case BO_Sub:
243  case BO_Shl:
244  case BO_Shr:
245  case BO_Xor:
246    // a+0, a-0, a<<0, a>>0, a^0
247    if (RHS == 0)
248      isIdempotent = true;
249    break;
250  case BO_And:
251    // a&0 and a&(~0)
252    if (RHS == 0)
253      return makeIntVal(0, resultTy);
254    else if (RHS.isAllOnesValue())
255      isIdempotent = true;
256    break;
257  case BO_Or:
258    // a|0 and a|(~0)
259    if (RHS == 0)
260      isIdempotent = true;
261    else if (RHS.isAllOnesValue()) {
262      const llvm::APSInt &Result = BasicVals.Convert(resultTy, RHS);
263      return nonloc::ConcreteInt(Result);
264    }
265    break;
266  }
267
268  // Idempotent ops (like a*1) can still change the type of an expression.
269  // Wrap the LHS up in a NonLoc again and let evalCastFromNonLoc do the
270  // dirty work.
271  if (isIdempotent)
272      return evalCastFromNonLoc(nonloc::SymbolVal(LHS), resultTy);
273
274  // If we reach this point, the expression cannot be simplified.
275  // Make a SymbolVal for the entire expression, after converting the RHS.
276  const llvm::APSInt *ConvertedRHS = &RHS;
277  if (BinaryOperator::isComparisonOp(op)) {
278    // We're looking for a type big enough to compare the symbolic value
279    // with the given constant.
280    // FIXME: This is an approximation of Sema::UsualArithmeticConversions.
281    ASTContext &Ctx = getContext();
282    QualType SymbolType = LHS->getType(Ctx);
283    uint64_t ValWidth = RHS.getBitWidth();
284    uint64_t TypeWidth = Ctx.getTypeSize(SymbolType);
285
286    if (ValWidth < TypeWidth) {
287      // If the value is too small, extend it.
288      ConvertedRHS = &BasicVals.Convert(SymbolType, RHS);
289    } else if (ValWidth == TypeWidth) {
290      // If the value is signed but the symbol is unsigned, do the comparison
291      // in unsigned space. [C99 6.3.1.8]
292      // (For the opposite case, the value is already unsigned.)
293      if (RHS.isSigned() && !SymbolType->isSignedIntegerOrEnumerationType())
294        ConvertedRHS = &BasicVals.Convert(SymbolType, RHS);
295    }
296  } else
297    ConvertedRHS = &BasicVals.Convert(resultTy, RHS);
298
299  return makeNonLoc(LHS, op, *ConvertedRHS, resultTy);
300}
301
302SVal SimpleSValBuilder::evalBinOpNN(ProgramStateRef state,
303                                  BinaryOperator::Opcode op,
304                                  NonLoc lhs, NonLoc rhs,
305                                  QualType resultTy)  {
306  NonLoc InputLHS = lhs;
307  NonLoc InputRHS = rhs;
308
309  // Handle trivial case where left-side and right-side are the same.
310  if (lhs == rhs)
311    switch (op) {
312      default:
313        break;
314      case BO_EQ:
315      case BO_LE:
316      case BO_GE:
317        return makeTruthVal(true, resultTy);
318      case BO_LT:
319      case BO_GT:
320      case BO_NE:
321        return makeTruthVal(false, resultTy);
322      case BO_Xor:
323      case BO_Sub:
324        return makeIntVal(0, resultTy);
325      case BO_Or:
326      case BO_And:
327        return evalCastFromNonLoc(lhs, resultTy);
328    }
329
330  while (1) {
331    switch (lhs.getSubKind()) {
332    default:
333      return makeSymExprValNN(state, op, lhs, rhs, resultTy);
334    case nonloc::LocAsIntegerKind: {
335      Loc lhsL = cast<nonloc::LocAsInteger>(lhs).getLoc();
336      switch (rhs.getSubKind()) {
337        case nonloc::LocAsIntegerKind:
338          return evalBinOpLL(state, op, lhsL,
339                             cast<nonloc::LocAsInteger>(rhs).getLoc(),
340                             resultTy);
341        case nonloc::ConcreteIntKind: {
342          // Transform the integer into a location and compare.
343          llvm::APSInt i = cast<nonloc::ConcreteInt>(rhs).getValue();
344          i.setIsUnsigned(true);
345          i = i.extOrTrunc(Context.getTypeSize(Context.VoidPtrTy));
346          return evalBinOpLL(state, op, lhsL, makeLoc(i), resultTy);
347        }
348        default:
349          switch (op) {
350            case BO_EQ:
351              return makeTruthVal(false, resultTy);
352            case BO_NE:
353              return makeTruthVal(true, resultTy);
354            default:
355              // This case also handles pointer arithmetic.
356              return makeSymExprValNN(state, op, InputLHS, InputRHS, resultTy);
357          }
358      }
359    }
360    case nonloc::ConcreteIntKind: {
361      llvm::APSInt LHSValue = cast<nonloc::ConcreteInt>(lhs).getValue();
362
363      // If we're dealing with two known constants, just perform the operation.
364      if (const llvm::APSInt *KnownRHSValue = getKnownValue(state, rhs)) {
365        llvm::APSInt RHSValue = *KnownRHSValue;
366        if (BinaryOperator::isComparisonOp(op)) {
367          // We're looking for a type big enough to compare the two values.
368          uint32_t LeftWidth = LHSValue.getBitWidth();
369          uint32_t RightWidth = RHSValue.getBitWidth();
370
371          // Based on the conversion rules of [C99 6.3.1.8] and the example
372          // in SemaExpr's handleIntegerConversion().
373          if (LeftWidth > RightWidth)
374            RHSValue = RHSValue.extend(LeftWidth);
375          else if (LeftWidth < RightWidth)
376            LHSValue = LHSValue.extend(RightWidth);
377          else if (LHSValue.isUnsigned() != RHSValue.isUnsigned()) {
378            LHSValue.setIsUnsigned(true);
379            RHSValue.setIsUnsigned(true);
380          }
381        } else if (!BinaryOperator::isShiftOp(op)) {
382          // FIXME: These values don't need to be persistent.
383          LHSValue = BasicVals.Convert(resultTy, LHSValue);
384          RHSValue = BasicVals.Convert(resultTy, RHSValue);
385        }
386
387        const llvm::APSInt *Result =
388          BasicVals.evalAPSInt(op, LHSValue, RHSValue);
389        if (!Result)
390          return UndefinedVal();
391
392        return nonloc::ConcreteInt(*Result);
393      }
394
395      // Swap the left and right sides and flip the operator if doing so
396      // allows us to better reason about the expression (this is a form
397      // of expression canonicalization).
398      // While we're at it, catch some special cases for non-commutative ops.
399      switch (op) {
400      case BO_LT:
401      case BO_GT:
402      case BO_LE:
403      case BO_GE:
404        op = ReverseComparison(op);
405        // FALL-THROUGH
406      case BO_EQ:
407      case BO_NE:
408      case BO_Add:
409      case BO_Mul:
410      case BO_And:
411      case BO_Xor:
412      case BO_Or:
413        std::swap(lhs, rhs);
414        continue;
415      case BO_Shr:
416        // (~0)>>a
417        if (LHSValue.isAllOnesValue() && LHSValue.isSigned())
418          return evalCastFromNonLoc(lhs, resultTy);
419        // FALL-THROUGH
420      case BO_Shl:
421        // 0<<a and 0>>a
422        if (LHSValue == 0)
423          return evalCastFromNonLoc(lhs, resultTy);
424        return makeSymExprValNN(state, op, InputLHS, InputRHS, resultTy);
425      default:
426        return makeSymExprValNN(state, op, InputLHS, InputRHS, resultTy);
427      }
428    }
429    case nonloc::SymbolValKind: {
430      // We only handle LHS as simple symbols or SymIntExprs.
431      SymbolRef Sym = cast<nonloc::SymbolVal>(lhs).getSymbol();
432
433      // LHS is a symbolic expression.
434      if (const SymIntExpr *symIntExpr = dyn_cast<SymIntExpr>(Sym)) {
435
436        // Is this a logical not? (!x is represented as x == 0.)
437        if (op == BO_EQ && rhs.isZeroConstant()) {
438          // We know how to negate certain expressions. Simplify them here.
439
440          BinaryOperator::Opcode opc = symIntExpr->getOpcode();
441          switch (opc) {
442          default:
443            // We don't know how to negate this operation.
444            // Just handle it as if it were a normal comparison to 0.
445            break;
446          case BO_LAnd:
447          case BO_LOr:
448            llvm_unreachable("Logical operators handled by branching logic.");
449          case BO_Assign:
450          case BO_MulAssign:
451          case BO_DivAssign:
452          case BO_RemAssign:
453          case BO_AddAssign:
454          case BO_SubAssign:
455          case BO_ShlAssign:
456          case BO_ShrAssign:
457          case BO_AndAssign:
458          case BO_XorAssign:
459          case BO_OrAssign:
460          case BO_Comma:
461            llvm_unreachable("'=' and ',' operators handled by ExprEngine.");
462          case BO_PtrMemD:
463          case BO_PtrMemI:
464            llvm_unreachable("Pointer arithmetic not handled here.");
465          case BO_LT:
466          case BO_GT:
467          case BO_LE:
468          case BO_GE:
469          case BO_EQ:
470          case BO_NE:
471            // Negate the comparison and make a value.
472            opc = NegateComparison(opc);
473            assert(symIntExpr->getType(Context) == resultTy);
474            return makeNonLoc(symIntExpr->getLHS(), opc,
475                symIntExpr->getRHS(), resultTy);
476          }
477        }
478
479        // For now, only handle expressions whose RHS is a constant.
480        if (const llvm::APSInt *RHSValue = getKnownValue(state, rhs)) {
481          // If both the LHS and the current expression are additive,
482          // fold their constants and try again.
483          if (BinaryOperator::isAdditiveOp(op)) {
484            BinaryOperator::Opcode lop = symIntExpr->getOpcode();
485            if (BinaryOperator::isAdditiveOp(lop)) {
486              // Convert the two constants to a common type, then combine them.
487
488              // resultTy may not be the best type to convert to, but it's
489              // probably the best choice in expressions with mixed type
490              // (such as x+1U+2LL). The rules for implicit conversions should
491              // choose a reasonable type to preserve the expression, and will
492              // at least match how the value is going to be used.
493
494              // FIXME: These values don't need to be persistent.
495              const llvm::APSInt &first =
496                  BasicVals.Convert(resultTy, symIntExpr->getRHS());
497              const llvm::APSInt &second =
498                  BasicVals.Convert(resultTy, *RHSValue);
499
500              const llvm::APSInt *newRHS;
501              if (lop == op)
502                newRHS = BasicVals.evalAPSInt(BO_Add, first, second);
503              else
504                newRHS = BasicVals.evalAPSInt(BO_Sub, first, second);
505
506              assert(newRHS && "Invalid operation despite common type!");
507              rhs = nonloc::ConcreteInt(*newRHS);
508              lhs = nonloc::SymbolVal(symIntExpr->getLHS());
509              op = lop;
510              continue;
511            }
512          }
513
514          // Otherwise, make a SymIntExpr out of the expression.
515          return MakeSymIntVal(symIntExpr, op, *RHSValue, resultTy);
516        }
517
518
519      } else if (isa<SymbolData>(Sym)) {
520        // LHS is a simple symbol (not a symbolic expression).
521        QualType lhsType = Sym->getType(Context);
522
523        // Does the symbol simplify to a constant?  If so, "fold" the constant
524        // by setting 'lhs' to a ConcreteInt and try again.
525        if (const llvm::APSInt *Constant = state->getSymVal(Sym)) {
526          lhs = nonloc::ConcreteInt(*Constant);
527          continue;
528        }
529
530        // Is the RHS a constant?
531        if (const llvm::APSInt *RHSValue = getKnownValue(state, rhs))
532          return MakeSymIntVal(Sym, op, *RHSValue, resultTy);
533      }
534
535      // Give up -- this is not a symbolic expression we can handle.
536      return makeSymExprValNN(state, op, InputLHS, InputRHS, resultTy);
537    }
538    }
539  }
540}
541
542// FIXME: all this logic will change if/when we have MemRegion::getLocation().
543SVal SimpleSValBuilder::evalBinOpLL(ProgramStateRef state,
544                                  BinaryOperator::Opcode op,
545                                  Loc lhs, Loc rhs,
546                                  QualType resultTy) {
547  // Only comparisons and subtractions are valid operations on two pointers.
548  // See [C99 6.5.5 through 6.5.14] or [C++0x 5.6 through 5.15].
549  // However, if a pointer is casted to an integer, evalBinOpNN may end up
550  // calling this function with another operation (PR7527). We don't attempt to
551  // model this for now, but it could be useful, particularly when the
552  // "location" is actually an integer value that's been passed through a void*.
553  if (!(BinaryOperator::isComparisonOp(op) || op == BO_Sub))
554    return UnknownVal();
555
556  // Special cases for when both sides are identical.
557  if (lhs == rhs) {
558    switch (op) {
559    default:
560      llvm_unreachable("Unimplemented operation for two identical values");
561    case BO_Sub:
562      return makeZeroVal(resultTy);
563    case BO_EQ:
564    case BO_LE:
565    case BO_GE:
566      return makeTruthVal(true, resultTy);
567    case BO_NE:
568    case BO_LT:
569    case BO_GT:
570      return makeTruthVal(false, resultTy);
571    }
572  }
573
574  switch (lhs.getSubKind()) {
575  default:
576    llvm_unreachable("Ordering not implemented for this Loc.");
577
578  case loc::GotoLabelKind:
579    // The only thing we know about labels is that they're non-null.
580    if (rhs.isZeroConstant()) {
581      switch (op) {
582      default:
583        break;
584      case BO_Sub:
585        return evalCastFromLoc(lhs, resultTy);
586      case BO_EQ:
587      case BO_LE:
588      case BO_LT:
589        return makeTruthVal(false, resultTy);
590      case BO_NE:
591      case BO_GT:
592      case BO_GE:
593        return makeTruthVal(true, resultTy);
594      }
595    }
596    // There may be two labels for the same location, and a function region may
597    // have the same address as a label at the start of the function (depending
598    // on the ABI).
599    // FIXME: we can probably do a comparison against other MemRegions, though.
600    // FIXME: is there a way to tell if two labels refer to the same location?
601    return UnknownVal();
602
603  case loc::ConcreteIntKind: {
604    // If one of the operands is a symbol and the other is a constant,
605    // build an expression for use by the constraint manager.
606    if (SymbolRef rSym = rhs.getAsLocSymbol()) {
607      // We can only build expressions with symbols on the left,
608      // so we need a reversible operator.
609      if (!BinaryOperator::isComparisonOp(op))
610        return UnknownVal();
611
612      const llvm::APSInt &lVal = cast<loc::ConcreteInt>(lhs).getValue();
613      return makeNonLoc(rSym, ReverseComparison(op), lVal, resultTy);
614    }
615
616    // If both operands are constants, just perform the operation.
617    if (loc::ConcreteInt *rInt = dyn_cast<loc::ConcreteInt>(&rhs)) {
618      SVal ResultVal = cast<loc::ConcreteInt>(lhs).evalBinOp(BasicVals, op,
619                                                             *rInt);
620      if (Loc *Result = dyn_cast<Loc>(&ResultVal))
621        return evalCastFromLoc(*Result, resultTy);
622      else
623        return UnknownVal();
624    }
625
626    // Special case comparisons against NULL.
627    // This must come after the test if the RHS is a symbol, which is used to
628    // build constraints. The address of any non-symbolic region is guaranteed
629    // to be non-NULL, as is any label.
630    assert(isa<loc::MemRegionVal>(rhs) || isa<loc::GotoLabel>(rhs));
631    if (lhs.isZeroConstant()) {
632      switch (op) {
633      default:
634        break;
635      case BO_EQ:
636      case BO_GT:
637      case BO_GE:
638        return makeTruthVal(false, resultTy);
639      case BO_NE:
640      case BO_LT:
641      case BO_LE:
642        return makeTruthVal(true, resultTy);
643      }
644    }
645
646    // Comparing an arbitrary integer to a region or label address is
647    // completely unknowable.
648    return UnknownVal();
649  }
650  case loc::MemRegionKind: {
651    if (loc::ConcreteInt *rInt = dyn_cast<loc::ConcreteInt>(&rhs)) {
652      // If one of the operands is a symbol and the other is a constant,
653      // build an expression for use by the constraint manager.
654      if (SymbolRef lSym = lhs.getAsLocSymbol())
655        return MakeSymIntVal(lSym, op, rInt->getValue(), resultTy);
656
657      // Special case comparisons to NULL.
658      // This must come after the test if the LHS is a symbol, which is used to
659      // build constraints. The address of any non-symbolic region is guaranteed
660      // to be non-NULL.
661      if (rInt->isZeroConstant()) {
662        switch (op) {
663        default:
664          break;
665        case BO_Sub:
666          return evalCastFromLoc(lhs, resultTy);
667        case BO_EQ:
668        case BO_LT:
669        case BO_LE:
670          return makeTruthVal(false, resultTy);
671        case BO_NE:
672        case BO_GT:
673        case BO_GE:
674          return makeTruthVal(true, resultTy);
675        }
676      }
677
678      // Comparing a region to an arbitrary integer is completely unknowable.
679      return UnknownVal();
680    }
681
682    // Get both values as regions, if possible.
683    const MemRegion *LeftMR = lhs.getAsRegion();
684    assert(LeftMR && "MemRegionKind SVal doesn't have a region!");
685
686    const MemRegion *RightMR = rhs.getAsRegion();
687    if (!RightMR)
688      // The RHS is probably a label, which in theory could address a region.
689      // FIXME: we can probably make a more useful statement about non-code
690      // regions, though.
691      return UnknownVal();
692
693    // If both values wrap regions, see if they're from different base regions.
694    const MemRegion *LeftBase = LeftMR->getBaseRegion();
695    const MemRegion *RightBase = RightMR->getBaseRegion();
696    if (LeftBase != RightBase &&
697        !isa<SymbolicRegion>(LeftBase) && !isa<SymbolicRegion>(RightBase)) {
698      switch (op) {
699      default:
700        return UnknownVal();
701      case BO_EQ:
702        return makeTruthVal(false, resultTy);
703      case BO_NE:
704        return makeTruthVal(true, resultTy);
705      }
706    }
707
708    // The two regions are from the same base region. See if they're both a
709    // type of region we know how to compare.
710    const MemSpaceRegion *LeftMS = LeftBase->getMemorySpace();
711    const MemSpaceRegion *RightMS = RightBase->getMemorySpace();
712
713    // Heuristic: assume that no symbolic region (whose memory space is
714    // unknown) is on the stack.
715    // FIXME: we should be able to be more precise once we can do better
716    // aliasing constraints for symbolic regions, but this is a reasonable,
717    // albeit unsound, assumption that holds most of the time.
718    if (isa<StackSpaceRegion>(LeftMS) ^ isa<StackSpaceRegion>(RightMS)) {
719      switch (op) {
720        default:
721          break;
722        case BO_EQ:
723          return makeTruthVal(false, resultTy);
724        case BO_NE:
725          return makeTruthVal(true, resultTy);
726      }
727    }
728
729    // FIXME: If/when there is a getAsRawOffset() for FieldRegions, this
730    // ElementRegion path and the FieldRegion path below should be unified.
731    if (const ElementRegion *LeftER = dyn_cast<ElementRegion>(LeftMR)) {
732      // First see if the right region is also an ElementRegion.
733      const ElementRegion *RightER = dyn_cast<ElementRegion>(RightMR);
734      if (!RightER)
735        return UnknownVal();
736
737      // Next, see if the two ERs have the same super-region and matching types.
738      // FIXME: This should do something useful even if the types don't match,
739      // though if both indexes are constant the RegionRawOffset path will
740      // give the correct answer.
741      if (LeftER->getSuperRegion() == RightER->getSuperRegion() &&
742          LeftER->getElementType() == RightER->getElementType()) {
743        // Get the left index and cast it to the correct type.
744        // If the index is unknown or undefined, bail out here.
745        SVal LeftIndexVal = LeftER->getIndex();
746        NonLoc *LeftIndex = dyn_cast<NonLoc>(&LeftIndexVal);
747        if (!LeftIndex)
748          return UnknownVal();
749        LeftIndexVal = evalCastFromNonLoc(*LeftIndex, resultTy);
750        LeftIndex = dyn_cast<NonLoc>(&LeftIndexVal);
751        if (!LeftIndex)
752          return UnknownVal();
753
754        // Do the same for the right index.
755        SVal RightIndexVal = RightER->getIndex();
756        NonLoc *RightIndex = dyn_cast<NonLoc>(&RightIndexVal);
757        if (!RightIndex)
758          return UnknownVal();
759        RightIndexVal = evalCastFromNonLoc(*RightIndex, resultTy);
760        RightIndex = dyn_cast<NonLoc>(&RightIndexVal);
761        if (!RightIndex)
762          return UnknownVal();
763
764        // Actually perform the operation.
765        // evalBinOpNN expects the two indexes to already be the right type.
766        return evalBinOpNN(state, op, *LeftIndex, *RightIndex, resultTy);
767      }
768
769      // If the element indexes aren't comparable, see if the raw offsets are.
770      RegionRawOffset LeftOffset = LeftER->getAsArrayOffset();
771      RegionRawOffset RightOffset = RightER->getAsArrayOffset();
772
773      if (LeftOffset.getRegion() != NULL &&
774          LeftOffset.getRegion() == RightOffset.getRegion()) {
775        CharUnits left = LeftOffset.getOffset();
776        CharUnits right = RightOffset.getOffset();
777
778        switch (op) {
779        default:
780          return UnknownVal();
781        case BO_LT:
782          return makeTruthVal(left < right, resultTy);
783        case BO_GT:
784          return makeTruthVal(left > right, resultTy);
785        case BO_LE:
786          return makeTruthVal(left <= right, resultTy);
787        case BO_GE:
788          return makeTruthVal(left >= right, resultTy);
789        case BO_EQ:
790          return makeTruthVal(left == right, resultTy);
791        case BO_NE:
792          return makeTruthVal(left != right, resultTy);
793        }
794      }
795
796      // If we get here, we have no way of comparing the ElementRegions.
797      return UnknownVal();
798    }
799
800    // See if both regions are fields of the same structure.
801    // FIXME: This doesn't handle nesting, inheritance, or Objective-C ivars.
802    if (const FieldRegion *LeftFR = dyn_cast<FieldRegion>(LeftMR)) {
803      // Only comparisons are meaningful here!
804      if (!BinaryOperator::isComparisonOp(op))
805        return UnknownVal();
806
807      // First see if the right region is also a FieldRegion.
808      const FieldRegion *RightFR = dyn_cast<FieldRegion>(RightMR);
809      if (!RightFR)
810        return UnknownVal();
811
812      // Next, see if the two FRs have the same super-region.
813      // FIXME: This doesn't handle casts yet, and simply stripping the casts
814      // doesn't help.
815      if (LeftFR->getSuperRegion() != RightFR->getSuperRegion())
816        return UnknownVal();
817
818      const FieldDecl *LeftFD = LeftFR->getDecl();
819      const FieldDecl *RightFD = RightFR->getDecl();
820      const RecordDecl *RD = LeftFD->getParent();
821
822      // Make sure the two FRs are from the same kind of record. Just in case!
823      // FIXME: This is probably where inheritance would be a problem.
824      if (RD != RightFD->getParent())
825        return UnknownVal();
826
827      // We know for sure that the two fields are not the same, since that
828      // would have given us the same SVal.
829      if (op == BO_EQ)
830        return makeTruthVal(false, resultTy);
831      if (op == BO_NE)
832        return makeTruthVal(true, resultTy);
833
834      // Iterate through the fields and see which one comes first.
835      // [C99 6.7.2.1.13] "Within a structure object, the non-bit-field
836      // members and the units in which bit-fields reside have addresses that
837      // increase in the order in which they are declared."
838      bool leftFirst = (op == BO_LT || op == BO_LE);
839      for (RecordDecl::field_iterator I = RD->field_begin(),
840           E = RD->field_end(); I!=E; ++I) {
841        if (&*I == LeftFD)
842          return makeTruthVal(leftFirst, resultTy);
843        if (&*I == RightFD)
844          return makeTruthVal(!leftFirst, resultTy);
845      }
846
847      llvm_unreachable("Fields not found in parent record's definition");
848    }
849
850    // If we get here, we have no way of comparing the regions.
851    return UnknownVal();
852  }
853  }
854}
855
856SVal SimpleSValBuilder::evalBinOpLN(ProgramStateRef state,
857                                  BinaryOperator::Opcode op,
858                                  Loc lhs, NonLoc rhs, QualType resultTy) {
859
860  // Special case: rhs is a zero constant.
861  if (rhs.isZeroConstant())
862    return lhs;
863
864  // Special case: 'rhs' is an integer that has the same width as a pointer and
865  // we are using the integer location in a comparison.  Normally this cannot be
866  // triggered, but transfer functions like those for OSCommpareAndSwapBarrier32
867  // can generate comparisons that trigger this code.
868  // FIXME: Are all locations guaranteed to have pointer width?
869  if (BinaryOperator::isComparisonOp(op)) {
870    if (nonloc::ConcreteInt *rhsInt = dyn_cast<nonloc::ConcreteInt>(&rhs)) {
871      const llvm::APSInt *x = &rhsInt->getValue();
872      ASTContext &ctx = Context;
873      if (ctx.getTypeSize(ctx.VoidPtrTy) == x->getBitWidth()) {
874        // Convert the signedness of the integer (if necessary).
875        if (x->isSigned())
876          x = &getBasicValueFactory().getValue(*x, true);
877
878        return evalBinOpLL(state, op, lhs, loc::ConcreteInt(*x), resultTy);
879      }
880    }
881  }
882
883  // We are dealing with pointer arithmetic.
884
885  // Handle pointer arithmetic on constant values.
886  if (nonloc::ConcreteInt *rhsInt = dyn_cast<nonloc::ConcreteInt>(&rhs)) {
887    if (loc::ConcreteInt *lhsInt = dyn_cast<loc::ConcreteInt>(&lhs)) {
888      const llvm::APSInt &leftI = lhsInt->getValue();
889      assert(leftI.isUnsigned());
890      llvm::APSInt rightI(rhsInt->getValue(), /* isUnsigned */ true);
891
892      // Convert the bitwidth of rightI.  This should deal with overflow
893      // since we are dealing with concrete values.
894      rightI = rightI.extOrTrunc(leftI.getBitWidth());
895
896      // Offset the increment by the pointer size.
897      llvm::APSInt Multiplicand(rightI.getBitWidth(), /* isUnsigned */ true);
898      rightI *= Multiplicand;
899
900      // Compute the adjusted pointer.
901      switch (op) {
902        case BO_Add:
903          rightI = leftI + rightI;
904          break;
905        case BO_Sub:
906          rightI = leftI - rightI;
907          break;
908        default:
909          llvm_unreachable("Invalid pointer arithmetic operation");
910      }
911      return loc::ConcreteInt(getBasicValueFactory().getValue(rightI));
912    }
913  }
914
915  // Handle cases where 'lhs' is a region.
916  if (const MemRegion *region = lhs.getAsRegion()) {
917    rhs = cast<NonLoc>(convertToArrayIndex(rhs));
918    SVal index = UnknownVal();
919    const MemRegion *superR = 0;
920    QualType elementType;
921
922    if (const ElementRegion *elemReg = dyn_cast<ElementRegion>(region)) {
923      assert(op == BO_Add || op == BO_Sub);
924      index = evalBinOpNN(state, op, elemReg->getIndex(), rhs,
925                          getArrayIndexType());
926      superR = elemReg->getSuperRegion();
927      elementType = elemReg->getElementType();
928    }
929    else if (isa<SubRegion>(region)) {
930      superR = region;
931      index = rhs;
932      if (const PointerType *PT = resultTy->getAs<PointerType>()) {
933        elementType = PT->getPointeeType();
934      }
935      else {
936        const ObjCObjectPointerType *OT =
937          resultTy->getAs<ObjCObjectPointerType>();
938        elementType = OT->getPointeeType();
939      }
940    }
941
942    if (NonLoc *indexV = dyn_cast<NonLoc>(&index)) {
943      return loc::MemRegionVal(MemMgr.getElementRegion(elementType, *indexV,
944                                                       superR, getContext()));
945    }
946  }
947  return UnknownVal();
948}
949
950const llvm::APSInt *SimpleSValBuilder::getKnownValue(ProgramStateRef state,
951                                                   SVal V) {
952  if (V.isUnknownOrUndef())
953    return NULL;
954
955  if (loc::ConcreteInt* X = dyn_cast<loc::ConcreteInt>(&V))
956    return &X->getValue();
957
958  if (nonloc::ConcreteInt* X = dyn_cast<nonloc::ConcreteInt>(&V))
959    return &X->getValue();
960
961  if (SymbolRef Sym = V.getAsSymbol())
962    return state->getSymVal(Sym);
963
964  // FIXME: Add support for SymExprs.
965  return NULL;
966}
967