xref: /netbsd-src/external/apache2/llvm/dist/clang/lib/StaticAnalyzer/Core/RangeConstraintManager.cpp (revision e038c9c4676b0f19b1b7dd08a940c6ed64a6d5ae)
1 //== RangeConstraintManager.cpp - Manage range constraints.------*- C++ -*--==//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 //  This file defines RangeConstraintManager, a class that tracks simple
10 //  equality and inequality constraints on symbolic values of ProgramState.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "clang/Basic/JsonSupport.h"
15 #include "clang/StaticAnalyzer/Core/PathSensitive/APSIntType.h"
16 #include "clang/StaticAnalyzer/Core/PathSensitive/ProgramState.h"
17 #include "clang/StaticAnalyzer/Core/PathSensitive/ProgramStateTrait.h"
18 #include "clang/StaticAnalyzer/Core/PathSensitive/RangedConstraintManager.h"
19 #include "clang/StaticAnalyzer/Core/PathSensitive/SValVisitor.h"
20 #include "llvm/ADT/FoldingSet.h"
21 #include "llvm/ADT/ImmutableSet.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/Support/Compiler.h"
24 #include "llvm/Support/raw_ostream.h"
25 #include <algorithm>
26 #include <iterator>
27 
28 using namespace clang;
29 using namespace ento;
30 
31 // This class can be extended with other tables which will help to reason
32 // about ranges more precisely.
33 class OperatorRelationsTable {
34   static_assert(BO_LT < BO_GT && BO_GT < BO_LE && BO_LE < BO_GE &&
35                     BO_GE < BO_EQ && BO_EQ < BO_NE,
36                 "This class relies on operators order. Rework it otherwise.");
37 
38 public:
39   enum TriStateKind {
40     False = 0,
41     True,
42     Unknown,
43   };
44 
45 private:
46   // CmpOpTable holds states which represent the corresponding range for
47   // branching an exploded graph. We can reason about the branch if there is
48   // a previously known fact of the existence of a comparison expression with
49   // operands used in the current expression.
50   // E.g. assuming (x < y) is true that means (x != y) is surely true.
51   // if (x previous_operation y)  // <    | !=      | >
52   //   if (x operation y)         // !=   | >       | <
53   //     tristate                 // True | Unknown | False
54   //
55   // CmpOpTable represents next:
56   // __|< |> |<=|>=|==|!=|UnknownX2|
57   // < |1 |0 |* |0 |0 |* |1        |
58   // > |0 |1 |0 |* |0 |* |1        |
59   // <=|1 |0 |1 |* |1 |* |0        |
60   // >=|0 |1 |* |1 |1 |* |0        |
61   // ==|0 |0 |* |* |1 |0 |1        |
62   // !=|1 |1 |* |* |0 |1 |0        |
63   //
64   // Columns stands for a previous operator.
65   // Rows stands for a current operator.
66   // Each row has exactly two `Unknown` cases.
67   // UnknownX2 means that both `Unknown` previous operators are met in code,
68   // and there is a special column for that, for example:
69   // if (x >= y)
70   //   if (x != y)
71   //     if (x <= y)
72   //       False only
73   static constexpr size_t CmpOpCount = BO_NE - BO_LT + 1;
74   const TriStateKind CmpOpTable[CmpOpCount][CmpOpCount + 1] = {
75       // <      >      <=     >=     ==     !=    UnknownX2
76       {True, False, Unknown, False, False, Unknown, True}, // <
77       {False, True, False, Unknown, False, Unknown, True}, // >
78       {True, False, True, Unknown, True, Unknown, False},  // <=
79       {False, True, Unknown, True, True, Unknown, False},  // >=
80       {False, False, Unknown, Unknown, True, False, True}, // ==
81       {True, True, Unknown, Unknown, False, True, False},  // !=
82   };
83 
getIndexFromOp(BinaryOperatorKind OP)84   static size_t getIndexFromOp(BinaryOperatorKind OP) {
85     return static_cast<size_t>(OP - BO_LT);
86   }
87 
88 public:
getCmpOpCount() const89   constexpr size_t getCmpOpCount() const { return CmpOpCount; }
90 
getOpFromIndex(size_t Index)91   static BinaryOperatorKind getOpFromIndex(size_t Index) {
92     return static_cast<BinaryOperatorKind>(Index + BO_LT);
93   }
94 
getCmpOpState(BinaryOperatorKind CurrentOP,BinaryOperatorKind QueriedOP) const95   TriStateKind getCmpOpState(BinaryOperatorKind CurrentOP,
96                              BinaryOperatorKind QueriedOP) const {
97     return CmpOpTable[getIndexFromOp(CurrentOP)][getIndexFromOp(QueriedOP)];
98   }
99 
getCmpOpStateForUnknownX2(BinaryOperatorKind CurrentOP) const100   TriStateKind getCmpOpStateForUnknownX2(BinaryOperatorKind CurrentOP) const {
101     return CmpOpTable[getIndexFromOp(CurrentOP)][CmpOpCount];
102   }
103 };
104 
105 //===----------------------------------------------------------------------===//
106 //                           RangeSet implementation
107 //===----------------------------------------------------------------------===//
108 
109 RangeSet::ContainerType RangeSet::Factory::EmptySet{};
110 
add(RangeSet Original,Range Element)111 RangeSet RangeSet::Factory::add(RangeSet Original, Range Element) {
112   ContainerType Result;
113   Result.reserve(Original.size() + 1);
114 
115   const_iterator Lower = llvm::lower_bound(Original, Element);
116   Result.insert(Result.end(), Original.begin(), Lower);
117   Result.push_back(Element);
118   Result.insert(Result.end(), Lower, Original.end());
119 
120   return makePersistent(std::move(Result));
121 }
122 
add(RangeSet Original,const llvm::APSInt & Point)123 RangeSet RangeSet::Factory::add(RangeSet Original, const llvm::APSInt &Point) {
124   return add(Original, Range(Point));
125 }
126 
getRangeSet(Range From)127 RangeSet RangeSet::Factory::getRangeSet(Range From) {
128   ContainerType Result;
129   Result.push_back(From);
130   return makePersistent(std::move(Result));
131 }
132 
makePersistent(ContainerType && From)133 RangeSet RangeSet::Factory::makePersistent(ContainerType &&From) {
134   llvm::FoldingSetNodeID ID;
135   void *InsertPos;
136 
137   From.Profile(ID);
138   ContainerType *Result = Cache.FindNodeOrInsertPos(ID, InsertPos);
139 
140   if (!Result) {
141     // It is cheaper to fully construct the resulting range on stack
142     // and move it to the freshly allocated buffer if we don't have
143     // a set like this already.
144     Result = construct(std::move(From));
145     Cache.InsertNode(Result, InsertPos);
146   }
147 
148   return Result;
149 }
150 
construct(ContainerType && From)151 RangeSet::ContainerType *RangeSet::Factory::construct(ContainerType &&From) {
152   void *Buffer = Arena.Allocate();
153   return new (Buffer) ContainerType(std::move(From));
154 }
155 
add(RangeSet LHS,RangeSet RHS)156 RangeSet RangeSet::Factory::add(RangeSet LHS, RangeSet RHS) {
157   ContainerType Result;
158   std::merge(LHS.begin(), LHS.end(), RHS.begin(), RHS.end(),
159              std::back_inserter(Result));
160   return makePersistent(std::move(Result));
161 }
162 
getMinValue() const163 const llvm::APSInt &RangeSet::getMinValue() const {
164   assert(!isEmpty());
165   return begin()->From();
166 }
167 
getMaxValue() const168 const llvm::APSInt &RangeSet::getMaxValue() const {
169   assert(!isEmpty());
170   return std::prev(end())->To();
171 }
172 
containsImpl(llvm::APSInt & Point) const173 bool RangeSet::containsImpl(llvm::APSInt &Point) const {
174   if (isEmpty() || !pin(Point))
175     return false;
176 
177   Range Dummy(Point);
178   const_iterator It = llvm::upper_bound(*this, Dummy);
179   if (It == begin())
180     return false;
181 
182   return std::prev(It)->Includes(Point);
183 }
184 
pin(llvm::APSInt & Point) const185 bool RangeSet::pin(llvm::APSInt &Point) const {
186   APSIntType Type(getMinValue());
187   if (Type.testInRange(Point, true) != APSIntType::RTR_Within)
188     return false;
189 
190   Type.apply(Point);
191   return true;
192 }
193 
pin(llvm::APSInt & Lower,llvm::APSInt & Upper) const194 bool RangeSet::pin(llvm::APSInt &Lower, llvm::APSInt &Upper) const {
195   // This function has nine cases, the cartesian product of range-testing
196   // both the upper and lower bounds against the symbol's type.
197   // Each case requires a different pinning operation.
198   // The function returns false if the described range is entirely outside
199   // the range of values for the associated symbol.
200   APSIntType Type(getMinValue());
201   APSIntType::RangeTestResultKind LowerTest = Type.testInRange(Lower, true);
202   APSIntType::RangeTestResultKind UpperTest = Type.testInRange(Upper, true);
203 
204   switch (LowerTest) {
205   case APSIntType::RTR_Below:
206     switch (UpperTest) {
207     case APSIntType::RTR_Below:
208       // The entire range is outside the symbol's set of possible values.
209       // If this is a conventionally-ordered range, the state is infeasible.
210       if (Lower <= Upper)
211         return false;
212 
213       // However, if the range wraps around, it spans all possible values.
214       Lower = Type.getMinValue();
215       Upper = Type.getMaxValue();
216       break;
217     case APSIntType::RTR_Within:
218       // The range starts below what's possible but ends within it. Pin.
219       Lower = Type.getMinValue();
220       Type.apply(Upper);
221       break;
222     case APSIntType::RTR_Above:
223       // The range spans all possible values for the symbol. Pin.
224       Lower = Type.getMinValue();
225       Upper = Type.getMaxValue();
226       break;
227     }
228     break;
229   case APSIntType::RTR_Within:
230     switch (UpperTest) {
231     case APSIntType::RTR_Below:
232       // The range wraps around, but all lower values are not possible.
233       Type.apply(Lower);
234       Upper = Type.getMaxValue();
235       break;
236     case APSIntType::RTR_Within:
237       // The range may or may not wrap around, but both limits are valid.
238       Type.apply(Lower);
239       Type.apply(Upper);
240       break;
241     case APSIntType::RTR_Above:
242       // The range starts within what's possible but ends above it. Pin.
243       Type.apply(Lower);
244       Upper = Type.getMaxValue();
245       break;
246     }
247     break;
248   case APSIntType::RTR_Above:
249     switch (UpperTest) {
250     case APSIntType::RTR_Below:
251       // The range wraps but is outside the symbol's set of possible values.
252       return false;
253     case APSIntType::RTR_Within:
254       // The range starts above what's possible but ends within it (wrap).
255       Lower = Type.getMinValue();
256       Type.apply(Upper);
257       break;
258     case APSIntType::RTR_Above:
259       // The entire range is outside the symbol's set of possible values.
260       // If this is a conventionally-ordered range, the state is infeasible.
261       if (Lower <= Upper)
262         return false;
263 
264       // However, if the range wraps around, it spans all possible values.
265       Lower = Type.getMinValue();
266       Upper = Type.getMaxValue();
267       break;
268     }
269     break;
270   }
271 
272   return true;
273 }
274 
intersect(RangeSet What,llvm::APSInt Lower,llvm::APSInt Upper)275 RangeSet RangeSet::Factory::intersect(RangeSet What, llvm::APSInt Lower,
276                                       llvm::APSInt Upper) {
277   if (What.isEmpty() || !What.pin(Lower, Upper))
278     return getEmptySet();
279 
280   ContainerType DummyContainer;
281 
282   if (Lower <= Upper) {
283     // [Lower, Upper] is a regular range.
284     //
285     // Shortcut: check that there is even a possibility of the intersection
286     //           by checking the two following situations:
287     //
288     //               <---[  What  ]---[------]------>
289     //                              Lower  Upper
290     //                            -or-
291     //               <----[------]----[  What  ]---->
292     //                  Lower  Upper
293     if (What.getMaxValue() < Lower || Upper < What.getMinValue())
294       return getEmptySet();
295 
296     DummyContainer.push_back(
297         Range(ValueFactory.getValue(Lower), ValueFactory.getValue(Upper)));
298   } else {
299     // [Lower, Upper] is an inverted range, i.e. [MIN, Upper] U [Lower, MAX]
300     //
301     // Shortcut: check that there is even a possibility of the intersection
302     //           by checking the following situation:
303     //
304     //               <------]---[  What  ]---[------>
305     //                    Upper             Lower
306     if (What.getMaxValue() < Lower && Upper < What.getMinValue())
307       return getEmptySet();
308 
309     DummyContainer.push_back(
310         Range(ValueFactory.getMinValue(Upper), ValueFactory.getValue(Upper)));
311     DummyContainer.push_back(
312         Range(ValueFactory.getValue(Lower), ValueFactory.getMaxValue(Lower)));
313   }
314 
315   return intersect(*What.Impl, DummyContainer);
316 }
317 
intersect(const RangeSet::ContainerType & LHS,const RangeSet::ContainerType & RHS)318 RangeSet RangeSet::Factory::intersect(const RangeSet::ContainerType &LHS,
319                                       const RangeSet::ContainerType &RHS) {
320   ContainerType Result;
321   Result.reserve(std::max(LHS.size(), RHS.size()));
322 
323   const_iterator First = LHS.begin(), Second = RHS.begin(),
324                  FirstEnd = LHS.end(), SecondEnd = RHS.end();
325 
326   const auto SwapIterators = [&First, &FirstEnd, &Second, &SecondEnd]() {
327     std::swap(First, Second);
328     std::swap(FirstEnd, SecondEnd);
329   };
330 
331   // If we ran out of ranges in one set, but not in the other,
332   // it means that those elements are definitely not in the
333   // intersection.
334   while (First != FirstEnd && Second != SecondEnd) {
335     // We want to keep the following invariant at all times:
336     //
337     //    ----[ First ---------------------->
338     //    --------[ Second ----------------->
339     if (Second->From() < First->From())
340       SwapIterators();
341 
342     // Loop where the invariant holds:
343     do {
344       // Check for the following situation:
345       //
346       //    ----[ First ]--------------------->
347       //    ---------------[ Second ]--------->
348       //
349       // which means that...
350       if (Second->From() > First->To()) {
351         // ...First is not in the intersection.
352         //
353         // We should move on to the next range after First and break out of the
354         // loop because the invariant might not be true.
355         ++First;
356         break;
357       }
358 
359       // We have a guaranteed intersection at this point!
360       // And this is the current situation:
361       //
362       //    ----[   First   ]----------------->
363       //    -------[ Second ------------------>
364       //
365       // Additionally, it definitely starts with Second->From().
366       const llvm::APSInt &IntersectionStart = Second->From();
367 
368       // It is important to know which of the two ranges' ends
369       // is greater.  That "longer" range might have some other
370       // intersections, while the "shorter" range might not.
371       if (Second->To() > First->To()) {
372         // Here we make a decision to keep First as the "longer"
373         // range.
374         SwapIterators();
375       }
376 
377       // At this point, we have the following situation:
378       //
379       //    ---- First      ]-------------------->
380       //    ---- Second ]--[  Second+1 ---------->
381       //
382       // We don't know the relationship between First->From and
383       // Second->From and we don't know whether Second+1 intersects
384       // with First.
385       //
386       // However, we know that [IntersectionStart, Second->To] is
387       // a part of the intersection...
388       Result.push_back(Range(IntersectionStart, Second->To()));
389       ++Second;
390       // ...and that the invariant will hold for a valid Second+1
391       // because First->From <= Second->To < (Second+1)->From.
392     } while (Second != SecondEnd);
393   }
394 
395   if (Result.empty())
396     return getEmptySet();
397 
398   return makePersistent(std::move(Result));
399 }
400 
intersect(RangeSet LHS,RangeSet RHS)401 RangeSet RangeSet::Factory::intersect(RangeSet LHS, RangeSet RHS) {
402   // Shortcut: let's see if the intersection is even possible.
403   if (LHS.isEmpty() || RHS.isEmpty() || LHS.getMaxValue() < RHS.getMinValue() ||
404       RHS.getMaxValue() < LHS.getMinValue())
405     return getEmptySet();
406 
407   return intersect(*LHS.Impl, *RHS.Impl);
408 }
409 
intersect(RangeSet LHS,llvm::APSInt Point)410 RangeSet RangeSet::Factory::intersect(RangeSet LHS, llvm::APSInt Point) {
411   if (LHS.containsImpl(Point))
412     return getRangeSet(ValueFactory.getValue(Point));
413 
414   return getEmptySet();
415 }
416 
negate(RangeSet What)417 RangeSet RangeSet::Factory::negate(RangeSet What) {
418   if (What.isEmpty())
419     return getEmptySet();
420 
421   const llvm::APSInt SampleValue = What.getMinValue();
422   const llvm::APSInt &MIN = ValueFactory.getMinValue(SampleValue);
423   const llvm::APSInt &MAX = ValueFactory.getMaxValue(SampleValue);
424 
425   ContainerType Result;
426   Result.reserve(What.size() + (SampleValue == MIN));
427 
428   // Handle a special case for MIN value.
429   const_iterator It = What.begin();
430   const_iterator End = What.end();
431 
432   const llvm::APSInt &From = It->From();
433   const llvm::APSInt &To = It->To();
434 
435   if (From == MIN) {
436     // If the range [From, To] is [MIN, MAX], then result is also [MIN, MAX].
437     if (To == MAX) {
438       return What;
439     }
440 
441     const_iterator Last = std::prev(End);
442 
443     // Try to find and unite the following ranges:
444     // [MIN, MIN] & [MIN + 1, N] => [MIN, N].
445     if (Last->To() == MAX) {
446       // It means that in the original range we have ranges
447       //   [MIN, A], ... , [B, MAX]
448       // And the result should be [MIN, -B], ..., [-A, MAX]
449       Result.emplace_back(MIN, ValueFactory.getValue(-Last->From()));
450       // We already negated Last, so we can skip it.
451       End = Last;
452     } else {
453       // Add a separate range for the lowest value.
454       Result.emplace_back(MIN, MIN);
455     }
456 
457     // Skip adding the second range in case when [From, To] are [MIN, MIN].
458     if (To != MIN) {
459       Result.emplace_back(ValueFactory.getValue(-To), MAX);
460     }
461 
462     // Skip the first range in the loop.
463     ++It;
464   }
465 
466   // Negate all other ranges.
467   for (; It != End; ++It) {
468     // Negate int values.
469     const llvm::APSInt &NewFrom = ValueFactory.getValue(-It->To());
470     const llvm::APSInt &NewTo = ValueFactory.getValue(-It->From());
471 
472     // Add a negated range.
473     Result.emplace_back(NewFrom, NewTo);
474   }
475 
476   llvm::sort(Result);
477   return makePersistent(std::move(Result));
478 }
479 
deletePoint(RangeSet From,const llvm::APSInt & Point)480 RangeSet RangeSet::Factory::deletePoint(RangeSet From,
481                                         const llvm::APSInt &Point) {
482   if (!From.contains(Point))
483     return From;
484 
485   llvm::APSInt Upper = Point;
486   llvm::APSInt Lower = Point;
487 
488   ++Upper;
489   --Lower;
490 
491   // Notice that the lower bound is greater than the upper bound.
492   return intersect(From, Upper, Lower);
493 }
494 
dump(raw_ostream & OS) const495 void Range::dump(raw_ostream &OS) const {
496   OS << '[' << From().toString(10) << ", " << To().toString(10) << ']';
497 }
498 
dump(raw_ostream & OS) const499 void RangeSet::dump(raw_ostream &OS) const {
500   OS << "{ ";
501   llvm::interleaveComma(*this, OS, [&OS](const Range &R) { R.dump(OS); });
502   OS << " }";
503 }
504 
505 REGISTER_SET_FACTORY_WITH_PROGRAMSTATE(SymbolSet, SymbolRef)
506 
507 namespace {
508 class EquivalenceClass;
509 } // end anonymous namespace
510 
511 REGISTER_MAP_WITH_PROGRAMSTATE(ClassMap, SymbolRef, EquivalenceClass)
512 REGISTER_MAP_WITH_PROGRAMSTATE(ClassMembers, EquivalenceClass, SymbolSet)
513 REGISTER_MAP_WITH_PROGRAMSTATE(ConstraintRange, EquivalenceClass, RangeSet)
514 
515 REGISTER_SET_FACTORY_WITH_PROGRAMSTATE(ClassSet, EquivalenceClass)
516 REGISTER_MAP_WITH_PROGRAMSTATE(DisequalityMap, EquivalenceClass, ClassSet)
517 
518 namespace {
519 /// This class encapsulates a set of symbols equal to each other.
520 ///
521 /// The main idea of the approach requiring such classes is in narrowing
522 /// and sharing constraints between symbols within the class.  Also we can
523 /// conclude that there is no practical need in storing constraints for
524 /// every member of the class separately.
525 ///
526 /// Main terminology:
527 ///
528 ///   * "Equivalence class" is an object of this class, which can be efficiently
529 ///     compared to other classes.  It represents the whole class without
530 ///     storing the actual in it.  The members of the class however can be
531 ///     retrieved from the state.
532 ///
533 ///   * "Class members" are the symbols corresponding to the class.  This means
534 ///     that A == B for every member symbols A and B from the class.  Members of
535 ///     each class are stored in the state.
536 ///
537 ///   * "Trivial class" is a class that has and ever had only one same symbol.
538 ///
539 ///   * "Merge operation" merges two classes into one.  It is the main operation
540 ///     to produce non-trivial classes.
541 ///     If, at some point, we can assume that two symbols from two distinct
542 ///     classes are equal, we can merge these classes.
543 class EquivalenceClass : public llvm::FoldingSetNode {
544 public:
545   /// Find equivalence class for the given symbol in the given state.
546   LLVM_NODISCARD static inline EquivalenceClass find(ProgramStateRef State,
547                                                      SymbolRef Sym);
548 
549   /// Merge classes for the given symbols and return a new state.
550   LLVM_NODISCARD static inline ProgramStateRef
551   merge(BasicValueFactory &BV, RangeSet::Factory &F, ProgramStateRef State,
552         SymbolRef First, SymbolRef Second);
553   // Merge this class with the given class and return a new state.
554   LLVM_NODISCARD inline ProgramStateRef merge(BasicValueFactory &BV,
555                                               RangeSet::Factory &F,
556                                               ProgramStateRef State,
557                                               EquivalenceClass Other);
558 
559   /// Return a set of class members for the given state.
560   LLVM_NODISCARD inline SymbolSet getClassMembers(ProgramStateRef State) const;
561   /// Return true if the current class is trivial in the given state.
562   LLVM_NODISCARD inline bool isTrivial(ProgramStateRef State) const;
563   /// Return true if the current class is trivial and its only member is dead.
564   LLVM_NODISCARD inline bool isTriviallyDead(ProgramStateRef State,
565                                              SymbolReaper &Reaper) const;
566 
567   LLVM_NODISCARD static inline ProgramStateRef
568   markDisequal(BasicValueFactory &BV, RangeSet::Factory &F,
569                ProgramStateRef State, SymbolRef First, SymbolRef Second);
570   LLVM_NODISCARD static inline ProgramStateRef
571   markDisequal(BasicValueFactory &BV, RangeSet::Factory &F,
572                ProgramStateRef State, EquivalenceClass First,
573                EquivalenceClass Second);
574   LLVM_NODISCARD inline ProgramStateRef
575   markDisequal(BasicValueFactory &BV, RangeSet::Factory &F,
576                ProgramStateRef State, EquivalenceClass Other) const;
577   LLVM_NODISCARD static inline ClassSet
578   getDisequalClasses(ProgramStateRef State, SymbolRef Sym);
579   LLVM_NODISCARD inline ClassSet
580   getDisequalClasses(ProgramStateRef State) const;
581   LLVM_NODISCARD inline ClassSet
582   getDisequalClasses(DisequalityMapTy Map, ClassSet::Factory &Factory) const;
583 
584   LLVM_NODISCARD static inline Optional<bool>
585   areEqual(ProgramStateRef State, SymbolRef First, SymbolRef Second);
586 
587   /// Check equivalence data for consistency.
588   LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED static bool
589   isClassDataConsistent(ProgramStateRef State);
590 
getType() const591   LLVM_NODISCARD QualType getType() const {
592     return getRepresentativeSymbol()->getType();
593   }
594 
595   EquivalenceClass() = delete;
596   EquivalenceClass(const EquivalenceClass &) = default;
597   EquivalenceClass &operator=(const EquivalenceClass &) = delete;
598   EquivalenceClass(EquivalenceClass &&) = default;
599   EquivalenceClass &operator=(EquivalenceClass &&) = delete;
600 
operator ==(const EquivalenceClass & Other) const601   bool operator==(const EquivalenceClass &Other) const {
602     return ID == Other.ID;
603   }
operator <(const EquivalenceClass & Other) const604   bool operator<(const EquivalenceClass &Other) const { return ID < Other.ID; }
operator !=(const EquivalenceClass & Other) const605   bool operator!=(const EquivalenceClass &Other) const {
606     return !operator==(Other);
607   }
608 
Profile(llvm::FoldingSetNodeID & ID,uintptr_t CID)609   static void Profile(llvm::FoldingSetNodeID &ID, uintptr_t CID) {
610     ID.AddInteger(CID);
611   }
612 
Profile(llvm::FoldingSetNodeID & ID) const613   void Profile(llvm::FoldingSetNodeID &ID) const { Profile(ID, this->ID); }
614 
615 private:
EquivalenceClass(SymbolRef Sym)616   /* implicit */ EquivalenceClass(SymbolRef Sym)
617       : ID(reinterpret_cast<uintptr_t>(Sym)) {}
618 
619   /// This function is intended to be used ONLY within the class.
620   /// The fact that ID is a pointer to a symbol is an implementation detail
621   /// and should stay that way.
622   /// In the current implementation, we use it to retrieve the only member
623   /// of the trivial class.
getRepresentativeSymbol() const624   SymbolRef getRepresentativeSymbol() const {
625     return reinterpret_cast<SymbolRef>(ID);
626   }
627   static inline SymbolSet::Factory &getMembersFactory(ProgramStateRef State);
628 
629   inline ProgramStateRef mergeImpl(BasicValueFactory &BV, RangeSet::Factory &F,
630                                    ProgramStateRef State, SymbolSet Members,
631                                    EquivalenceClass Other,
632                                    SymbolSet OtherMembers);
633   static inline bool
634   addToDisequalityInfo(DisequalityMapTy &Info, ConstraintRangeTy &Constraints,
635                        BasicValueFactory &BV, RangeSet::Factory &F,
636                        ProgramStateRef State, EquivalenceClass First,
637                        EquivalenceClass Second);
638 
639   /// This is a unique identifier of the class.
640   uintptr_t ID;
641 };
642 
643 //===----------------------------------------------------------------------===//
644 //                             Constraint functions
645 //===----------------------------------------------------------------------===//
646 
647 LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED bool
areFeasible(ConstraintRangeTy Constraints)648 areFeasible(ConstraintRangeTy Constraints) {
649   return llvm::none_of(
650       Constraints,
651       [](const std::pair<EquivalenceClass, RangeSet> &ClassConstraint) {
652         return ClassConstraint.second.isEmpty();
653       });
654 }
655 
getConstraint(ProgramStateRef State,EquivalenceClass Class)656 LLVM_NODISCARD inline const RangeSet *getConstraint(ProgramStateRef State,
657                                                     EquivalenceClass Class) {
658   return State->get<ConstraintRange>(Class);
659 }
660 
getConstraint(ProgramStateRef State,SymbolRef Sym)661 LLVM_NODISCARD inline const RangeSet *getConstraint(ProgramStateRef State,
662                                                     SymbolRef Sym) {
663   return getConstraint(State, EquivalenceClass::find(State, Sym));
664 }
665 
666 //===----------------------------------------------------------------------===//
667 //                       Equality/diseqiality abstraction
668 //===----------------------------------------------------------------------===//
669 
670 /// A small helper structure representing symbolic equality.
671 ///
672 /// Equality check can have different forms (like a == b or a - b) and this
673 /// class encapsulates those away if the only thing the user wants to check -
674 /// whether it's equality/diseqiality or not and have an easy access to the
675 /// compared symbols.
676 struct EqualityInfo {
677 public:
678   SymbolRef Left, Right;
679   // true for equality and false for disequality.
680   bool IsEquality = true;
681 
invert__anon0bb35a870411::EqualityInfo682   void invert() { IsEquality = !IsEquality; }
683   /// Extract equality information from the given symbol and the constants.
684   ///
685   /// This function assumes the following expression Sym + Adjustment != Int.
686   /// It is a default because the most widespread case of the equality check
687   /// is (A == B) + 0 != 0.
extract__anon0bb35a870411::EqualityInfo688   static Optional<EqualityInfo> extract(SymbolRef Sym, const llvm::APSInt &Int,
689                                         const llvm::APSInt &Adjustment) {
690     // As of now, the only equality form supported is Sym + 0 != 0.
691     if (!Int.isNullValue() || !Adjustment.isNullValue())
692       return llvm::None;
693 
694     return extract(Sym);
695   }
696   /// Extract equality information from the given symbol.
extract__anon0bb35a870411::EqualityInfo697   static Optional<EqualityInfo> extract(SymbolRef Sym) {
698     return EqualityExtractor().Visit(Sym);
699   }
700 
701 private:
702   class EqualityExtractor
703       : public SymExprVisitor<EqualityExtractor, Optional<EqualityInfo>> {
704   public:
VisitSymSymExpr(const SymSymExpr * Sym) const705     Optional<EqualityInfo> VisitSymSymExpr(const SymSymExpr *Sym) const {
706       switch (Sym->getOpcode()) {
707       case BO_Sub:
708         // This case is: A - B != 0 -> disequality check.
709         return EqualityInfo{Sym->getLHS(), Sym->getRHS(), false};
710       case BO_EQ:
711         // This case is: A == B != 0 -> equality check.
712         return EqualityInfo{Sym->getLHS(), Sym->getRHS(), true};
713       case BO_NE:
714         // This case is: A != B != 0 -> diseqiality check.
715         return EqualityInfo{Sym->getLHS(), Sym->getRHS(), false};
716       default:
717         return llvm::None;
718       }
719     }
720   };
721 };
722 
723 //===----------------------------------------------------------------------===//
724 //                            Intersection functions
725 //===----------------------------------------------------------------------===//
726 
727 template <class SecondTy, class... RestTy>
728 LLVM_NODISCARD inline RangeSet intersect(BasicValueFactory &BV,
729                                          RangeSet::Factory &F, RangeSet Head,
730                                          SecondTy Second, RestTy... Tail);
731 
732 template <class... RangeTy> struct IntersectionTraits;
733 
734 template <class... TailTy> struct IntersectionTraits<RangeSet, TailTy...> {
735   // Found RangeSet, no need to check any further
736   using Type = RangeSet;
737 };
738 
739 template <> struct IntersectionTraits<> {
740   // We ran out of types, and we didn't find any RangeSet, so the result should
741   // be optional.
742   using Type = Optional<RangeSet>;
743 };
744 
745 template <class OptionalOrPointer, class... TailTy>
746 struct IntersectionTraits<OptionalOrPointer, TailTy...> {
747   // If current type is Optional or a raw pointer, we should keep looking.
748   using Type = typename IntersectionTraits<TailTy...>::Type;
749 };
750 
751 template <class EndTy>
intersect(BasicValueFactory & BV,RangeSet::Factory & F,EndTy End)752 LLVM_NODISCARD inline EndTy intersect(BasicValueFactory &BV,
753                                       RangeSet::Factory &F, EndTy End) {
754   // If the list contains only RangeSet or Optional<RangeSet>, simply return
755   // that range set.
756   return End;
757 }
758 
759 LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED inline Optional<RangeSet>
intersect(BasicValueFactory & BV,RangeSet::Factory & F,const RangeSet * End)760 intersect(BasicValueFactory &BV, RangeSet::Factory &F, const RangeSet *End) {
761   // This is an extraneous conversion from a raw pointer into Optional<RangeSet>
762   if (End) {
763     return *End;
764   }
765   return llvm::None;
766 }
767 
768 template <class... RestTy>
intersect(BasicValueFactory & BV,RangeSet::Factory & F,RangeSet Head,RangeSet Second,RestTy...Tail)769 LLVM_NODISCARD inline RangeSet intersect(BasicValueFactory &BV,
770                                          RangeSet::Factory &F, RangeSet Head,
771                                          RangeSet Second, RestTy... Tail) {
772   // Here we call either the <RangeSet,RangeSet,...> or <RangeSet,...> version
773   // of the function and can be sure that the result is RangeSet.
774   return intersect(BV, F, F.intersect(Head, Second), Tail...);
775 }
776 
777 template <class SecondTy, class... RestTy>
intersect(BasicValueFactory & BV,RangeSet::Factory & F,RangeSet Head,SecondTy Second,RestTy...Tail)778 LLVM_NODISCARD inline RangeSet intersect(BasicValueFactory &BV,
779                                          RangeSet::Factory &F, RangeSet Head,
780                                          SecondTy Second, RestTy... Tail) {
781   if (Second) {
782     // Here we call the <RangeSet,RangeSet,...> version of the function...
783     return intersect(BV, F, Head, *Second, Tail...);
784   }
785   // ...and here it is either <RangeSet,RangeSet,...> or <RangeSet,...>, which
786   // means that the result is definitely RangeSet.
787   return intersect(BV, F, Head, Tail...);
788 }
789 
790 /// Main generic intersect function.
791 /// It intersects all of the given range sets.  If some of the given arguments
792 /// don't hold a range set (nullptr or llvm::None), the function will skip them.
793 ///
794 /// Available representations for the arguments are:
795 ///   * RangeSet
796 ///   * Optional<RangeSet>
797 ///   * RangeSet *
798 /// Pointer to a RangeSet is automatically assumed to be nullable and will get
799 /// checked as well as the optional version.  If this behaviour is undesired,
800 /// please dereference the pointer in the call.
801 ///
802 /// Return type depends on the arguments' types.  If we can be sure in compile
803 /// time that there will be a range set as a result, the returning type is
804 /// simply RangeSet, in other cases we have to back off to Optional<RangeSet>.
805 ///
806 /// Please, prefer optional range sets to raw pointers.  If the last argument is
807 /// a raw pointer and all previous arguments are None, it will cost one
808 /// additional check to convert RangeSet * into Optional<RangeSet>.
809 template <class HeadTy, class SecondTy, class... RestTy>
810 LLVM_NODISCARD inline
811     typename IntersectionTraits<HeadTy, SecondTy, RestTy...>::Type
intersect(BasicValueFactory & BV,RangeSet::Factory & F,HeadTy Head,SecondTy Second,RestTy...Tail)812     intersect(BasicValueFactory &BV, RangeSet::Factory &F, HeadTy Head,
813               SecondTy Second, RestTy... Tail) {
814   if (Head) {
815     return intersect(BV, F, *Head, Second, Tail...);
816   }
817   return intersect(BV, F, Second, Tail...);
818 }
819 
820 //===----------------------------------------------------------------------===//
821 //                           Symbolic reasoning logic
822 //===----------------------------------------------------------------------===//
823 
824 /// A little component aggregating all of the reasoning we have about
825 /// the ranges of symbolic expressions.
826 ///
827 /// Even when we don't know the exact values of the operands, we still
828 /// can get a pretty good estimate of the result's range.
829 class SymbolicRangeInferrer
830     : public SymExprVisitor<SymbolicRangeInferrer, RangeSet> {
831 public:
832   template <class SourceType>
inferRange(BasicValueFactory & BV,RangeSet::Factory & F,ProgramStateRef State,SourceType Origin)833   static RangeSet inferRange(BasicValueFactory &BV, RangeSet::Factory &F,
834                              ProgramStateRef State, SourceType Origin) {
835     SymbolicRangeInferrer Inferrer(BV, F, State);
836     return Inferrer.infer(Origin);
837   }
838 
VisitSymExpr(SymbolRef Sym)839   RangeSet VisitSymExpr(SymbolRef Sym) {
840     // If we got to this function, the actual type of the symbolic
841     // expression is not supported for advanced inference.
842     // In this case, we simply backoff to the default "let's simply
843     // infer the range from the expression's type".
844     return infer(Sym->getType());
845   }
846 
VisitSymIntExpr(const SymIntExpr * Sym)847   RangeSet VisitSymIntExpr(const SymIntExpr *Sym) {
848     return VisitBinaryOperator(Sym);
849   }
850 
VisitIntSymExpr(const IntSymExpr * Sym)851   RangeSet VisitIntSymExpr(const IntSymExpr *Sym) {
852     return VisitBinaryOperator(Sym);
853   }
854 
VisitSymSymExpr(const SymSymExpr * Sym)855   RangeSet VisitSymSymExpr(const SymSymExpr *Sym) {
856     return VisitBinaryOperator(Sym);
857   }
858 
859 private:
SymbolicRangeInferrer(BasicValueFactory & BV,RangeSet::Factory & F,ProgramStateRef S)860   SymbolicRangeInferrer(BasicValueFactory &BV, RangeSet::Factory &F,
861                         ProgramStateRef S)
862       : ValueFactory(BV), RangeFactory(F), State(S) {}
863 
864   /// Infer range information from the given integer constant.
865   ///
866   /// It's not a real "inference", but is here for operating with
867   /// sub-expressions in a more polymorphic manner.
inferAs(const llvm::APSInt & Val,QualType)868   RangeSet inferAs(const llvm::APSInt &Val, QualType) {
869     return {RangeFactory, Val};
870   }
871 
872   /// Infer range information from symbol in the context of the given type.
inferAs(SymbolRef Sym,QualType DestType)873   RangeSet inferAs(SymbolRef Sym, QualType DestType) {
874     QualType ActualType = Sym->getType();
875     // Check that we can reason about the symbol at all.
876     if (ActualType->isIntegralOrEnumerationType() ||
877         Loc::isLocType(ActualType)) {
878       return infer(Sym);
879     }
880     // Otherwise, let's simply infer from the destination type.
881     // We couldn't figure out nothing else about that expression.
882     return infer(DestType);
883   }
884 
infer(SymbolRef Sym)885   RangeSet infer(SymbolRef Sym) {
886     if (Optional<RangeSet> ConstraintBasedRange = intersect(
887             ValueFactory, RangeFactory, getConstraint(State, Sym),
888             // If Sym is a difference of symbols A - B, then maybe we have range
889             // set stored for B - A.
890             //
891             // If we have range set stored for both A - B and B - A then
892             // calculate the effective range set by intersecting the range set
893             // for A - B and the negated range set of B - A.
894             getRangeForNegatedSub(Sym), getRangeForEqualities(Sym))) {
895       return *ConstraintBasedRange;
896     }
897 
898     // If Sym is a comparison expression (except <=>),
899     // find any other comparisons with the same operands.
900     // See function description.
901     if (Optional<RangeSet> CmpRangeSet = getRangeForComparisonSymbol(Sym)) {
902       return *CmpRangeSet;
903     }
904 
905     return Visit(Sym);
906   }
907 
infer(EquivalenceClass Class)908   RangeSet infer(EquivalenceClass Class) {
909     if (const RangeSet *AssociatedConstraint = getConstraint(State, Class))
910       return *AssociatedConstraint;
911 
912     return infer(Class.getType());
913   }
914 
915   /// Infer range information solely from the type.
infer(QualType T)916   RangeSet infer(QualType T) {
917     // Lazily generate a new RangeSet representing all possible values for the
918     // given symbol type.
919     RangeSet Result(RangeFactory, ValueFactory.getMinValue(T),
920                     ValueFactory.getMaxValue(T));
921 
922     // References are known to be non-zero.
923     if (T->isReferenceType())
924       return assumeNonZero(Result, T);
925 
926     return Result;
927   }
928 
929   template <class BinarySymExprTy>
VisitBinaryOperator(const BinarySymExprTy * Sym)930   RangeSet VisitBinaryOperator(const BinarySymExprTy *Sym) {
931     // TODO #1: VisitBinaryOperator implementation might not make a good
932     // use of the inferred ranges.  In this case, we might be calculating
933     // everything for nothing.  This being said, we should introduce some
934     // sort of laziness mechanism here.
935     //
936     // TODO #2: We didn't go into the nested expressions before, so it
937     // might cause us spending much more time doing the inference.
938     // This can be a problem for deeply nested expressions that are
939     // involved in conditions and get tested continuously.  We definitely
940     // need to address this issue and introduce some sort of caching
941     // in here.
942     QualType ResultType = Sym->getType();
943     return VisitBinaryOperator(inferAs(Sym->getLHS(), ResultType),
944                                Sym->getOpcode(),
945                                inferAs(Sym->getRHS(), ResultType), ResultType);
946   }
947 
VisitBinaryOperator(RangeSet LHS,BinaryOperator::Opcode Op,RangeSet RHS,QualType T)948   RangeSet VisitBinaryOperator(RangeSet LHS, BinaryOperator::Opcode Op,
949                                RangeSet RHS, QualType T) {
950     switch (Op) {
951     case BO_Or:
952       return VisitBinaryOperator<BO_Or>(LHS, RHS, T);
953     case BO_And:
954       return VisitBinaryOperator<BO_And>(LHS, RHS, T);
955     case BO_Rem:
956       return VisitBinaryOperator<BO_Rem>(LHS, RHS, T);
957     default:
958       return infer(T);
959     }
960   }
961 
962   //===----------------------------------------------------------------------===//
963   //                         Ranges and operators
964   //===----------------------------------------------------------------------===//
965 
966   /// Return a rough approximation of the given range set.
967   ///
968   /// For the range set:
969   ///   { [x_0, y_0], [x_1, y_1], ... , [x_N, y_N] }
970   /// it will return the range [x_0, y_N].
fillGaps(RangeSet Origin)971   static Range fillGaps(RangeSet Origin) {
972     assert(!Origin.isEmpty());
973     return {Origin.getMinValue(), Origin.getMaxValue()};
974   }
975 
976   /// Try to convert given range into the given type.
977   ///
978   /// It will return llvm::None only when the trivial conversion is possible.
convert(const Range & Origin,APSIntType To)979   llvm::Optional<Range> convert(const Range &Origin, APSIntType To) {
980     if (To.testInRange(Origin.From(), false) != APSIntType::RTR_Within ||
981         To.testInRange(Origin.To(), false) != APSIntType::RTR_Within) {
982       return llvm::None;
983     }
984     return Range(ValueFactory.Convert(To, Origin.From()),
985                  ValueFactory.Convert(To, Origin.To()));
986   }
987 
988   template <BinaryOperator::Opcode Op>
VisitBinaryOperator(RangeSet LHS,RangeSet RHS,QualType T)989   RangeSet VisitBinaryOperator(RangeSet LHS, RangeSet RHS, QualType T) {
990     // We should propagate information about unfeasbility of one of the
991     // operands to the resulting range.
992     if (LHS.isEmpty() || RHS.isEmpty()) {
993       return RangeFactory.getEmptySet();
994     }
995 
996     Range CoarseLHS = fillGaps(LHS);
997     Range CoarseRHS = fillGaps(RHS);
998 
999     APSIntType ResultType = ValueFactory.getAPSIntType(T);
1000 
1001     // We need to convert ranges to the resulting type, so we can compare values
1002     // and combine them in a meaningful (in terms of the given operation) way.
1003     auto ConvertedCoarseLHS = convert(CoarseLHS, ResultType);
1004     auto ConvertedCoarseRHS = convert(CoarseRHS, ResultType);
1005 
1006     // It is hard to reason about ranges when conversion changes
1007     // borders of the ranges.
1008     if (!ConvertedCoarseLHS || !ConvertedCoarseRHS) {
1009       return infer(T);
1010     }
1011 
1012     return VisitBinaryOperator<Op>(*ConvertedCoarseLHS, *ConvertedCoarseRHS, T);
1013   }
1014 
1015   template <BinaryOperator::Opcode Op>
VisitBinaryOperator(Range LHS,Range RHS,QualType T)1016   RangeSet VisitBinaryOperator(Range LHS, Range RHS, QualType T) {
1017     return infer(T);
1018   }
1019 
1020   /// Return a symmetrical range for the given range and type.
1021   ///
1022   /// If T is signed, return the smallest range [-x..x] that covers the original
1023   /// range, or [-min(T), max(T)] if the aforementioned symmetric range doesn't
1024   /// exist due to original range covering min(T)).
1025   ///
1026   /// If T is unsigned, return the smallest range [0..x] that covers the
1027   /// original range.
getSymmetricalRange(Range Origin,QualType T)1028   Range getSymmetricalRange(Range Origin, QualType T) {
1029     APSIntType RangeType = ValueFactory.getAPSIntType(T);
1030 
1031     if (RangeType.isUnsigned()) {
1032       return Range(ValueFactory.getMinValue(RangeType), Origin.To());
1033     }
1034 
1035     if (Origin.From().isMinSignedValue()) {
1036       // If mini is a minimal signed value, absolute value of it is greater
1037       // than the maximal signed value.  In order to avoid these
1038       // complications, we simply return the whole range.
1039       return {ValueFactory.getMinValue(RangeType),
1040               ValueFactory.getMaxValue(RangeType)};
1041     }
1042 
1043     // At this point, we are sure that the type is signed and we can safely
1044     // use unary - operator.
1045     //
1046     // While calculating absolute maximum, we can use the following formula
1047     // because of these reasons:
1048     //   * If From >= 0 then To >= From and To >= -From.
1049     //     AbsMax == To == max(To, -From)
1050     //   * If To <= 0 then -From >= -To and -From >= From.
1051     //     AbsMax == -From == max(-From, To)
1052     //   * Otherwise, From <= 0, To >= 0, and
1053     //     AbsMax == max(abs(From), abs(To))
1054     llvm::APSInt AbsMax = std::max(-Origin.From(), Origin.To());
1055 
1056     // Intersection is guaranteed to be non-empty.
1057     return {ValueFactory.getValue(-AbsMax), ValueFactory.getValue(AbsMax)};
1058   }
1059 
1060   /// Return a range set subtracting zero from \p Domain.
assumeNonZero(RangeSet Domain,QualType T)1061   RangeSet assumeNonZero(RangeSet Domain, QualType T) {
1062     APSIntType IntType = ValueFactory.getAPSIntType(T);
1063     return RangeFactory.deletePoint(Domain, IntType.getZeroValue());
1064   }
1065 
1066   // FIXME: Once SValBuilder supports unary minus, we should use SValBuilder to
1067   //        obtain the negated symbolic expression instead of constructing the
1068   //        symbol manually. This will allow us to support finding ranges of not
1069   //        only negated SymSymExpr-type expressions, but also of other, simpler
1070   //        expressions which we currently do not know how to negate.
getRangeForNegatedSub(SymbolRef Sym)1071   Optional<RangeSet> getRangeForNegatedSub(SymbolRef Sym) {
1072     if (const SymSymExpr *SSE = dyn_cast<SymSymExpr>(Sym)) {
1073       if (SSE->getOpcode() == BO_Sub) {
1074         QualType T = Sym->getType();
1075 
1076         // Do not negate unsigned ranges
1077         if (!T->isUnsignedIntegerOrEnumerationType() &&
1078             !T->isSignedIntegerOrEnumerationType())
1079           return llvm::None;
1080 
1081         SymbolManager &SymMgr = State->getSymbolManager();
1082         SymbolRef NegatedSym =
1083             SymMgr.getSymSymExpr(SSE->getRHS(), BO_Sub, SSE->getLHS(), T);
1084 
1085         if (const RangeSet *NegatedRange = getConstraint(State, NegatedSym)) {
1086           return RangeFactory.negate(*NegatedRange);
1087         }
1088       }
1089     }
1090     return llvm::None;
1091   }
1092 
1093   // Returns ranges only for binary comparison operators (except <=>)
1094   // when left and right operands are symbolic values.
1095   // Finds any other comparisons with the same operands.
1096   // Then do logical calculations and refuse impossible branches.
1097   // E.g. (x < y) and (x > y) at the same time are impossible.
1098   // E.g. (x >= y) and (x != y) at the same time makes (x > y) true only.
1099   // E.g. (x == y) and (y == x) are just reversed but the same.
1100   // It covers all possible combinations (see CmpOpTable description).
1101   // Note that `x` and `y` can also stand for subexpressions,
1102   // not only for actual symbols.
getRangeForComparisonSymbol(SymbolRef Sym)1103   Optional<RangeSet> getRangeForComparisonSymbol(SymbolRef Sym) {
1104     const auto *SSE = dyn_cast<SymSymExpr>(Sym);
1105     if (!SSE)
1106       return llvm::None;
1107 
1108     BinaryOperatorKind CurrentOP = SSE->getOpcode();
1109 
1110     // We currently do not support <=> (C++20).
1111     if (!BinaryOperator::isComparisonOp(CurrentOP) || (CurrentOP == BO_Cmp))
1112       return llvm::None;
1113 
1114     static const OperatorRelationsTable CmpOpTable{};
1115 
1116     const SymExpr *LHS = SSE->getLHS();
1117     const SymExpr *RHS = SSE->getRHS();
1118     QualType T = SSE->getType();
1119 
1120     SymbolManager &SymMgr = State->getSymbolManager();
1121 
1122     int UnknownStates = 0;
1123 
1124     // Loop goes through all of the columns exept the last one ('UnknownX2').
1125     // We treat `UnknownX2` column separately at the end of the loop body.
1126     for (size_t i = 0; i < CmpOpTable.getCmpOpCount(); ++i) {
1127 
1128       // Let's find an expression e.g. (x < y).
1129       BinaryOperatorKind QueriedOP = OperatorRelationsTable::getOpFromIndex(i);
1130       const SymSymExpr *SymSym = SymMgr.getSymSymExpr(LHS, QueriedOP, RHS, T);
1131       const RangeSet *QueriedRangeSet = getConstraint(State, SymSym);
1132 
1133       // If ranges were not previously found,
1134       // try to find a reversed expression (y > x).
1135       if (!QueriedRangeSet) {
1136         const BinaryOperatorKind ROP =
1137             BinaryOperator::reverseComparisonOp(QueriedOP);
1138         SymSym = SymMgr.getSymSymExpr(RHS, ROP, LHS, T);
1139         QueriedRangeSet = getConstraint(State, SymSym);
1140       }
1141 
1142       if (!QueriedRangeSet || QueriedRangeSet->isEmpty())
1143         continue;
1144 
1145       const llvm::APSInt *ConcreteValue = QueriedRangeSet->getConcreteValue();
1146       const bool isInFalseBranch =
1147           ConcreteValue ? (*ConcreteValue == 0) : false;
1148 
1149       // If it is a false branch, we shall be guided by opposite operator,
1150       // because the table is made assuming we are in the true branch.
1151       // E.g. when (x <= y) is false, then (x > y) is true.
1152       if (isInFalseBranch)
1153         QueriedOP = BinaryOperator::negateComparisonOp(QueriedOP);
1154 
1155       OperatorRelationsTable::TriStateKind BranchState =
1156           CmpOpTable.getCmpOpState(CurrentOP, QueriedOP);
1157 
1158       if (BranchState == OperatorRelationsTable::Unknown) {
1159         if (++UnknownStates == 2)
1160           // If we met both Unknown states.
1161           // if (x <= y)    // assume true
1162           //   if (x != y)  // assume true
1163           //     if (x < y) // would be also true
1164           // Get a state from `UnknownX2` column.
1165           BranchState = CmpOpTable.getCmpOpStateForUnknownX2(CurrentOP);
1166         else
1167           continue;
1168       }
1169 
1170       return (BranchState == OperatorRelationsTable::True) ? getTrueRange(T)
1171                                                            : getFalseRange(T);
1172     }
1173 
1174     return llvm::None;
1175   }
1176 
getRangeForEqualities(SymbolRef Sym)1177   Optional<RangeSet> getRangeForEqualities(SymbolRef Sym) {
1178     Optional<EqualityInfo> Equality = EqualityInfo::extract(Sym);
1179 
1180     if (!Equality)
1181       return llvm::None;
1182 
1183     if (Optional<bool> AreEqual = EquivalenceClass::areEqual(
1184             State, Equality->Left, Equality->Right)) {
1185       if (*AreEqual == Equality->IsEquality) {
1186         return getTrueRange(Sym->getType());
1187       }
1188       return getFalseRange(Sym->getType());
1189     }
1190 
1191     return llvm::None;
1192   }
1193 
getTrueRange(QualType T)1194   RangeSet getTrueRange(QualType T) {
1195     RangeSet TypeRange = infer(T);
1196     return assumeNonZero(TypeRange, T);
1197   }
1198 
getFalseRange(QualType T)1199   RangeSet getFalseRange(QualType T) {
1200     const llvm::APSInt &Zero = ValueFactory.getValue(0, T);
1201     return RangeSet(RangeFactory, Zero);
1202   }
1203 
1204   BasicValueFactory &ValueFactory;
1205   RangeSet::Factory &RangeFactory;
1206   ProgramStateRef State;
1207 };
1208 
1209 //===----------------------------------------------------------------------===//
1210 //               Range-based reasoning about symbolic operations
1211 //===----------------------------------------------------------------------===//
1212 
1213 template <>
VisitBinaryOperator(Range LHS,Range RHS,QualType T)1214 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_Or>(Range LHS, Range RHS,
1215                                                            QualType T) {
1216   APSIntType ResultType = ValueFactory.getAPSIntType(T);
1217   llvm::APSInt Zero = ResultType.getZeroValue();
1218 
1219   bool IsLHSPositiveOrZero = LHS.From() >= Zero;
1220   bool IsRHSPositiveOrZero = RHS.From() >= Zero;
1221 
1222   bool IsLHSNegative = LHS.To() < Zero;
1223   bool IsRHSNegative = RHS.To() < Zero;
1224 
1225   // Check if both ranges have the same sign.
1226   if ((IsLHSPositiveOrZero && IsRHSPositiveOrZero) ||
1227       (IsLHSNegative && IsRHSNegative)) {
1228     // The result is definitely greater or equal than any of the operands.
1229     const llvm::APSInt &Min = std::max(LHS.From(), RHS.From());
1230 
1231     // We estimate maximal value for positives as the maximal value for the
1232     // given type.  For negatives, we estimate it with -1 (e.g. 0x11111111).
1233     //
1234     // TODO: We basically, limit the resulting range from below, but don't do
1235     //       anything with the upper bound.
1236     //
1237     //       For positive operands, it can be done as follows: for the upper
1238     //       bound of LHS and RHS we calculate the most significant bit set.
1239     //       Let's call it the N-th bit.  Then we can estimate the maximal
1240     //       number to be 2^(N+1)-1, i.e. the number with all the bits up to
1241     //       the N-th bit set.
1242     const llvm::APSInt &Max = IsLHSNegative
1243                                   ? ValueFactory.getValue(--Zero)
1244                                   : ValueFactory.getMaxValue(ResultType);
1245 
1246     return {RangeFactory, ValueFactory.getValue(Min), Max};
1247   }
1248 
1249   // Otherwise, let's check if at least one of the operands is negative.
1250   if (IsLHSNegative || IsRHSNegative) {
1251     // This means that the result is definitely negative as well.
1252     return {RangeFactory, ValueFactory.getMinValue(ResultType),
1253             ValueFactory.getValue(--Zero)};
1254   }
1255 
1256   RangeSet DefaultRange = infer(T);
1257 
1258   // It is pretty hard to reason about operands with different signs
1259   // (and especially with possibly different signs).  We simply check if it
1260   // can be zero.  In order to conclude that the result could not be zero,
1261   // at least one of the operands should be definitely not zero itself.
1262   if (!LHS.Includes(Zero) || !RHS.Includes(Zero)) {
1263     return assumeNonZero(DefaultRange, T);
1264   }
1265 
1266   // Nothing much else to do here.
1267   return DefaultRange;
1268 }
1269 
1270 template <>
VisitBinaryOperator(Range LHS,Range RHS,QualType T)1271 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_And>(Range LHS,
1272                                                             Range RHS,
1273                                                             QualType T) {
1274   APSIntType ResultType = ValueFactory.getAPSIntType(T);
1275   llvm::APSInt Zero = ResultType.getZeroValue();
1276 
1277   bool IsLHSPositiveOrZero = LHS.From() >= Zero;
1278   bool IsRHSPositiveOrZero = RHS.From() >= Zero;
1279 
1280   bool IsLHSNegative = LHS.To() < Zero;
1281   bool IsRHSNegative = RHS.To() < Zero;
1282 
1283   // Check if both ranges have the same sign.
1284   if ((IsLHSPositiveOrZero && IsRHSPositiveOrZero) ||
1285       (IsLHSNegative && IsRHSNegative)) {
1286     // The result is definitely less or equal than any of the operands.
1287     const llvm::APSInt &Max = std::min(LHS.To(), RHS.To());
1288 
1289     // We conservatively estimate lower bound to be the smallest positive
1290     // or negative value corresponding to the sign of the operands.
1291     const llvm::APSInt &Min = IsLHSNegative
1292                                   ? ValueFactory.getMinValue(ResultType)
1293                                   : ValueFactory.getValue(Zero);
1294 
1295     return {RangeFactory, Min, Max};
1296   }
1297 
1298   // Otherwise, let's check if at least one of the operands is positive.
1299   if (IsLHSPositiveOrZero || IsRHSPositiveOrZero) {
1300     // This makes result definitely positive.
1301     //
1302     // We can also reason about a maximal value by finding the maximal
1303     // value of the positive operand.
1304     const llvm::APSInt &Max = IsLHSPositiveOrZero ? LHS.To() : RHS.To();
1305 
1306     // The minimal value on the other hand is much harder to reason about.
1307     // The only thing we know for sure is that the result is positive.
1308     return {RangeFactory, ValueFactory.getValue(Zero),
1309             ValueFactory.getValue(Max)};
1310   }
1311 
1312   // Nothing much else to do here.
1313   return infer(T);
1314 }
1315 
1316 template <>
VisitBinaryOperator(Range LHS,Range RHS,QualType T)1317 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_Rem>(Range LHS,
1318                                                             Range RHS,
1319                                                             QualType T) {
1320   llvm::APSInt Zero = ValueFactory.getAPSIntType(T).getZeroValue();
1321 
1322   Range ConservativeRange = getSymmetricalRange(RHS, T);
1323 
1324   llvm::APSInt Max = ConservativeRange.To();
1325   llvm::APSInt Min = ConservativeRange.From();
1326 
1327   if (Max == Zero) {
1328     // It's an undefined behaviour to divide by 0 and it seems like we know
1329     // for sure that RHS is 0.  Let's say that the resulting range is
1330     // simply infeasible for that matter.
1331     return RangeFactory.getEmptySet();
1332   }
1333 
1334   // At this point, our conservative range is closed.  The result, however,
1335   // couldn't be greater than the RHS' maximal absolute value.  Because of
1336   // this reason, we turn the range into open (or half-open in case of
1337   // unsigned integers).
1338   //
1339   // While we operate on integer values, an open interval (a, b) can be easily
1340   // represented by the closed interval [a + 1, b - 1].  And this is exactly
1341   // what we do next.
1342   //
1343   // If we are dealing with unsigned case, we shouldn't move the lower bound.
1344   if (Min.isSigned()) {
1345     ++Min;
1346   }
1347   --Max;
1348 
1349   bool IsLHSPositiveOrZero = LHS.From() >= Zero;
1350   bool IsRHSPositiveOrZero = RHS.From() >= Zero;
1351 
1352   // Remainder operator results with negative operands is implementation
1353   // defined.  Positive cases are much easier to reason about though.
1354   if (IsLHSPositiveOrZero && IsRHSPositiveOrZero) {
1355     // If maximal value of LHS is less than maximal value of RHS,
1356     // the result won't get greater than LHS.To().
1357     Max = std::min(LHS.To(), Max);
1358     // We want to check if it is a situation similar to the following:
1359     //
1360     // <------------|---[  LHS  ]--------[  RHS  ]----->
1361     //  -INF        0                              +INF
1362     //
1363     // In this situation, we can conclude that (LHS / RHS) == 0 and
1364     // (LHS % RHS) == LHS.
1365     Min = LHS.To() < RHS.From() ? LHS.From() : Zero;
1366   }
1367 
1368   // Nevertheless, the symmetrical range for RHS is a conservative estimate
1369   // for any sign of either LHS, or RHS.
1370   return {RangeFactory, ValueFactory.getValue(Min), ValueFactory.getValue(Max)};
1371 }
1372 
1373 //===----------------------------------------------------------------------===//
1374 //                  Constraint manager implementation details
1375 //===----------------------------------------------------------------------===//
1376 
1377 class RangeConstraintManager : public RangedConstraintManager {
1378 public:
RangeConstraintManager(ExprEngine * EE,SValBuilder & SVB)1379   RangeConstraintManager(ExprEngine *EE, SValBuilder &SVB)
1380       : RangedConstraintManager(EE, SVB), F(getBasicVals()) {}
1381 
1382   //===------------------------------------------------------------------===//
1383   // Implementation for interface from ConstraintManager.
1384   //===------------------------------------------------------------------===//
1385 
haveEqualConstraints(ProgramStateRef S1,ProgramStateRef S2) const1386   bool haveEqualConstraints(ProgramStateRef S1,
1387                             ProgramStateRef S2) const override {
1388     // NOTE: ClassMembers are as simple as back pointers for ClassMap,
1389     //       so comparing constraint ranges and class maps should be
1390     //       sufficient.
1391     return S1->get<ConstraintRange>() == S2->get<ConstraintRange>() &&
1392            S1->get<ClassMap>() == S2->get<ClassMap>();
1393   }
1394 
1395   bool canReasonAbout(SVal X) const override;
1396 
1397   ConditionTruthVal checkNull(ProgramStateRef State, SymbolRef Sym) override;
1398 
1399   const llvm::APSInt *getSymVal(ProgramStateRef State,
1400                                 SymbolRef Sym) const override;
1401 
1402   ProgramStateRef removeDeadBindings(ProgramStateRef State,
1403                                      SymbolReaper &SymReaper) override;
1404 
1405   void printJson(raw_ostream &Out, ProgramStateRef State, const char *NL = "\n",
1406                  unsigned int Space = 0, bool IsDot = false) const override;
1407 
1408   //===------------------------------------------------------------------===//
1409   // Implementation for interface from RangedConstraintManager.
1410   //===------------------------------------------------------------------===//
1411 
1412   ProgramStateRef assumeSymNE(ProgramStateRef State, SymbolRef Sym,
1413                               const llvm::APSInt &V,
1414                               const llvm::APSInt &Adjustment) override;
1415 
1416   ProgramStateRef assumeSymEQ(ProgramStateRef State, SymbolRef Sym,
1417                               const llvm::APSInt &V,
1418                               const llvm::APSInt &Adjustment) override;
1419 
1420   ProgramStateRef assumeSymLT(ProgramStateRef State, SymbolRef Sym,
1421                               const llvm::APSInt &V,
1422                               const llvm::APSInt &Adjustment) override;
1423 
1424   ProgramStateRef assumeSymGT(ProgramStateRef State, SymbolRef Sym,
1425                               const llvm::APSInt &V,
1426                               const llvm::APSInt &Adjustment) override;
1427 
1428   ProgramStateRef assumeSymLE(ProgramStateRef State, SymbolRef Sym,
1429                               const llvm::APSInt &V,
1430                               const llvm::APSInt &Adjustment) override;
1431 
1432   ProgramStateRef assumeSymGE(ProgramStateRef State, SymbolRef Sym,
1433                               const llvm::APSInt &V,
1434                               const llvm::APSInt &Adjustment) override;
1435 
1436   ProgramStateRef assumeSymWithinInclusiveRange(
1437       ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From,
1438       const llvm::APSInt &To, const llvm::APSInt &Adjustment) override;
1439 
1440   ProgramStateRef assumeSymOutsideInclusiveRange(
1441       ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From,
1442       const llvm::APSInt &To, const llvm::APSInt &Adjustment) override;
1443 
1444 private:
1445   RangeSet::Factory F;
1446 
1447   RangeSet getRange(ProgramStateRef State, SymbolRef Sym);
1448   RangeSet getRange(ProgramStateRef State, EquivalenceClass Class);
1449 
1450   RangeSet getSymLTRange(ProgramStateRef St, SymbolRef Sym,
1451                          const llvm::APSInt &Int,
1452                          const llvm::APSInt &Adjustment);
1453   RangeSet getSymGTRange(ProgramStateRef St, SymbolRef Sym,
1454                          const llvm::APSInt &Int,
1455                          const llvm::APSInt &Adjustment);
1456   RangeSet getSymLERange(ProgramStateRef St, SymbolRef Sym,
1457                          const llvm::APSInt &Int,
1458                          const llvm::APSInt &Adjustment);
1459   RangeSet getSymLERange(llvm::function_ref<RangeSet()> RS,
1460                          const llvm::APSInt &Int,
1461                          const llvm::APSInt &Adjustment);
1462   RangeSet getSymGERange(ProgramStateRef St, SymbolRef Sym,
1463                          const llvm::APSInt &Int,
1464                          const llvm::APSInt &Adjustment);
1465 
1466   //===------------------------------------------------------------------===//
1467   // Equality tracking implementation
1468   //===------------------------------------------------------------------===//
1469 
trackEQ(RangeSet NewConstraint,ProgramStateRef State,SymbolRef Sym,const llvm::APSInt & Int,const llvm::APSInt & Adjustment)1470   ProgramStateRef trackEQ(RangeSet NewConstraint, ProgramStateRef State,
1471                           SymbolRef Sym, const llvm::APSInt &Int,
1472                           const llvm::APSInt &Adjustment) {
1473     return track<true>(NewConstraint, State, Sym, Int, Adjustment);
1474   }
1475 
trackNE(RangeSet NewConstraint,ProgramStateRef State,SymbolRef Sym,const llvm::APSInt & Int,const llvm::APSInt & Adjustment)1476   ProgramStateRef trackNE(RangeSet NewConstraint, ProgramStateRef State,
1477                           SymbolRef Sym, const llvm::APSInt &Int,
1478                           const llvm::APSInt &Adjustment) {
1479     return track<false>(NewConstraint, State, Sym, Int, Adjustment);
1480   }
1481 
1482   template <bool EQ>
track(RangeSet NewConstraint,ProgramStateRef State,SymbolRef Sym,const llvm::APSInt & Int,const llvm::APSInt & Adjustment)1483   ProgramStateRef track(RangeSet NewConstraint, ProgramStateRef State,
1484                         SymbolRef Sym, const llvm::APSInt &Int,
1485                         const llvm::APSInt &Adjustment) {
1486     if (NewConstraint.isEmpty())
1487       // This is an infeasible assumption.
1488       return nullptr;
1489 
1490     if (ProgramStateRef NewState = setConstraint(State, Sym, NewConstraint)) {
1491       if (auto Equality = EqualityInfo::extract(Sym, Int, Adjustment)) {
1492         // If the original assumption is not Sym + Adjustment !=/</> Int,
1493         // we should invert IsEquality flag.
1494         Equality->IsEquality = Equality->IsEquality != EQ;
1495         return track(NewState, *Equality);
1496       }
1497 
1498       return NewState;
1499     }
1500 
1501     return nullptr;
1502   }
1503 
track(ProgramStateRef State,EqualityInfo ToTrack)1504   ProgramStateRef track(ProgramStateRef State, EqualityInfo ToTrack) {
1505     if (ToTrack.IsEquality) {
1506       return trackEquality(State, ToTrack.Left, ToTrack.Right);
1507     }
1508     return trackDisequality(State, ToTrack.Left, ToTrack.Right);
1509   }
1510 
trackDisequality(ProgramStateRef State,SymbolRef LHS,SymbolRef RHS)1511   ProgramStateRef trackDisequality(ProgramStateRef State, SymbolRef LHS,
1512                                    SymbolRef RHS) {
1513     return EquivalenceClass::markDisequal(getBasicVals(), F, State, LHS, RHS);
1514   }
1515 
trackEquality(ProgramStateRef State,SymbolRef LHS,SymbolRef RHS)1516   ProgramStateRef trackEquality(ProgramStateRef State, SymbolRef LHS,
1517                                 SymbolRef RHS) {
1518     return EquivalenceClass::merge(getBasicVals(), F, State, LHS, RHS);
1519   }
1520 
setConstraint(ProgramStateRef State,EquivalenceClass Class,RangeSet Constraint)1521   LLVM_NODISCARD ProgramStateRef setConstraint(ProgramStateRef State,
1522                                                EquivalenceClass Class,
1523                                                RangeSet Constraint) {
1524     ConstraintRangeTy Constraints = State->get<ConstraintRange>();
1525     ConstraintRangeTy::Factory &CF = State->get_context<ConstraintRange>();
1526 
1527     assert(!Constraint.isEmpty() && "New constraint should not be empty");
1528 
1529     // Add new constraint.
1530     Constraints = CF.add(Constraints, Class, Constraint);
1531 
1532     // There is a chance that we might need to update constraints for the
1533     // classes that are known to be disequal to Class.
1534     //
1535     // In order for this to be even possible, the new constraint should
1536     // be simply a constant because we can't reason about range disequalities.
1537     if (const llvm::APSInt *Point = Constraint.getConcreteValue())
1538       for (EquivalenceClass DisequalClass : Class.getDisequalClasses(State)) {
1539         RangeSet UpdatedConstraint = getRange(State, DisequalClass);
1540         UpdatedConstraint = F.deletePoint(UpdatedConstraint, *Point);
1541 
1542         // If we end up with at least one of the disequal classes to be
1543         // constrained with an empty range-set, the state is infeasible.
1544         if (UpdatedConstraint.isEmpty())
1545           return nullptr;
1546 
1547         Constraints = CF.add(Constraints, DisequalClass, UpdatedConstraint);
1548       }
1549 
1550     assert(areFeasible(Constraints) && "Constraint manager shouldn't produce "
1551                                        "a state with infeasible constraints");
1552 
1553     return State->set<ConstraintRange>(Constraints);
1554   }
1555 
1556   LLVM_NODISCARD inline ProgramStateRef
setConstraint(ProgramStateRef State,SymbolRef Sym,RangeSet Constraint)1557   setConstraint(ProgramStateRef State, SymbolRef Sym, RangeSet Constraint) {
1558     return setConstraint(State, EquivalenceClass::find(State, Sym), Constraint);
1559   }
1560 };
1561 
1562 } // end anonymous namespace
1563 
1564 std::unique_ptr<ConstraintManager>
CreateRangeConstraintManager(ProgramStateManager & StMgr,ExprEngine * Eng)1565 ento::CreateRangeConstraintManager(ProgramStateManager &StMgr,
1566                                    ExprEngine *Eng) {
1567   return std::make_unique<RangeConstraintManager>(Eng, StMgr.getSValBuilder());
1568 }
1569 
getConstraintMap(ProgramStateRef State)1570 ConstraintMap ento::getConstraintMap(ProgramStateRef State) {
1571   ConstraintMap::Factory &F = State->get_context<ConstraintMap>();
1572   ConstraintMap Result = F.getEmptyMap();
1573 
1574   ConstraintRangeTy Constraints = State->get<ConstraintRange>();
1575   for (std::pair<EquivalenceClass, RangeSet> ClassConstraint : Constraints) {
1576     EquivalenceClass Class = ClassConstraint.first;
1577     SymbolSet ClassMembers = Class.getClassMembers(State);
1578     assert(!ClassMembers.isEmpty() &&
1579            "Class must always have at least one member!");
1580 
1581     SymbolRef Representative = *ClassMembers.begin();
1582     Result = F.add(Result, Representative, ClassConstraint.second);
1583   }
1584 
1585   return Result;
1586 }
1587 
1588 //===----------------------------------------------------------------------===//
1589 //                     EqualityClass implementation details
1590 //===----------------------------------------------------------------------===//
1591 
find(ProgramStateRef State,SymbolRef Sym)1592 inline EquivalenceClass EquivalenceClass::find(ProgramStateRef State,
1593                                                SymbolRef Sym) {
1594   // We store far from all Symbol -> Class mappings
1595   if (const EquivalenceClass *NontrivialClass = State->get<ClassMap>(Sym))
1596     return *NontrivialClass;
1597 
1598   // This is a trivial class of Sym.
1599   return Sym;
1600 }
1601 
merge(BasicValueFactory & BV,RangeSet::Factory & F,ProgramStateRef State,SymbolRef First,SymbolRef Second)1602 inline ProgramStateRef EquivalenceClass::merge(BasicValueFactory &BV,
1603                                                RangeSet::Factory &F,
1604                                                ProgramStateRef State,
1605                                                SymbolRef First,
1606                                                SymbolRef Second) {
1607   EquivalenceClass FirstClass = find(State, First);
1608   EquivalenceClass SecondClass = find(State, Second);
1609 
1610   return FirstClass.merge(BV, F, State, SecondClass);
1611 }
1612 
merge(BasicValueFactory & BV,RangeSet::Factory & F,ProgramStateRef State,EquivalenceClass Other)1613 inline ProgramStateRef EquivalenceClass::merge(BasicValueFactory &BV,
1614                                                RangeSet::Factory &F,
1615                                                ProgramStateRef State,
1616                                                EquivalenceClass Other) {
1617   // It is already the same class.
1618   if (*this == Other)
1619     return State;
1620 
1621   // FIXME: As of now, we support only equivalence classes of the same type.
1622   //        This limitation is connected to the lack of explicit casts in
1623   //        our symbolic expression model.
1624   //
1625   //        That means that for `int x` and `char y` we don't distinguish
1626   //        between these two very different cases:
1627   //          * `x == y`
1628   //          * `(char)x == y`
1629   //
1630   //        The moment we introduce symbolic casts, this restriction can be
1631   //        lifted.
1632   if (getType() != Other.getType())
1633     return State;
1634 
1635   SymbolSet Members = getClassMembers(State);
1636   SymbolSet OtherMembers = Other.getClassMembers(State);
1637 
1638   // We estimate the size of the class by the height of tree containing
1639   // its members.  Merging is not a trivial operation, so it's easier to
1640   // merge the smaller class into the bigger one.
1641   if (Members.getHeight() >= OtherMembers.getHeight()) {
1642     return mergeImpl(BV, F, State, Members, Other, OtherMembers);
1643   } else {
1644     return Other.mergeImpl(BV, F, State, OtherMembers, *this, Members);
1645   }
1646 }
1647 
1648 inline ProgramStateRef
mergeImpl(BasicValueFactory & ValueFactory,RangeSet::Factory & RangeFactory,ProgramStateRef State,SymbolSet MyMembers,EquivalenceClass Other,SymbolSet OtherMembers)1649 EquivalenceClass::mergeImpl(BasicValueFactory &ValueFactory,
1650                             RangeSet::Factory &RangeFactory,
1651                             ProgramStateRef State, SymbolSet MyMembers,
1652                             EquivalenceClass Other, SymbolSet OtherMembers) {
1653   // Essentially what we try to recreate here is some kind of union-find
1654   // data structure.  It does have certain limitations due to persistence
1655   // and the need to remove elements from classes.
1656   //
1657   // In this setting, EquialityClass object is the representative of the class
1658   // or the parent element.  ClassMap is a mapping of class members to their
1659   // parent. Unlike the union-find structure, they all point directly to the
1660   // class representative because we don't have an opportunity to actually do
1661   // path compression when dealing with immutability.  This means that we
1662   // compress paths every time we do merges.  It also means that we lose
1663   // the main amortized complexity benefit from the original data structure.
1664   ConstraintRangeTy Constraints = State->get<ConstraintRange>();
1665   ConstraintRangeTy::Factory &CRF = State->get_context<ConstraintRange>();
1666 
1667   // 1. If the merged classes have any constraints associated with them, we
1668   //    need to transfer them to the class we have left.
1669   //
1670   // Intersection here makes perfect sense because both of these constraints
1671   // must hold for the whole new class.
1672   if (Optional<RangeSet> NewClassConstraint =
1673           intersect(ValueFactory, RangeFactory, getConstraint(State, *this),
1674                     getConstraint(State, Other))) {
1675     // NOTE: Essentially, NewClassConstraint should NEVER be infeasible because
1676     //       range inferrer shouldn't generate ranges incompatible with
1677     //       equivalence classes. However, at the moment, due to imperfections
1678     //       in the solver, it is possible and the merge function can also
1679     //       return infeasible states aka null states.
1680     if (NewClassConstraint->isEmpty())
1681       // Infeasible state
1682       return nullptr;
1683 
1684     // No need in tracking constraints of a now-dissolved class.
1685     Constraints = CRF.remove(Constraints, Other);
1686     // Assign new constraints for this class.
1687     Constraints = CRF.add(Constraints, *this, *NewClassConstraint);
1688 
1689     assert(areFeasible(Constraints) && "Constraint manager shouldn't produce "
1690                                        "a state with infeasible constraints");
1691 
1692     State = State->set<ConstraintRange>(Constraints);
1693   }
1694 
1695   // 2. Get ALL equivalence-related maps
1696   ClassMapTy Classes = State->get<ClassMap>();
1697   ClassMapTy::Factory &CMF = State->get_context<ClassMap>();
1698 
1699   ClassMembersTy Members = State->get<ClassMembers>();
1700   ClassMembersTy::Factory &MF = State->get_context<ClassMembers>();
1701 
1702   DisequalityMapTy DisequalityInfo = State->get<DisequalityMap>();
1703   DisequalityMapTy::Factory &DF = State->get_context<DisequalityMap>();
1704 
1705   ClassSet::Factory &CF = State->get_context<ClassSet>();
1706   SymbolSet::Factory &F = getMembersFactory(State);
1707 
1708   // 2. Merge members of the Other class into the current class.
1709   SymbolSet NewClassMembers = MyMembers;
1710   for (SymbolRef Sym : OtherMembers) {
1711     NewClassMembers = F.add(NewClassMembers, Sym);
1712     // *this is now the class for all these new symbols.
1713     Classes = CMF.add(Classes, Sym, *this);
1714   }
1715 
1716   // 3. Adjust member mapping.
1717   //
1718   // No need in tracking members of a now-dissolved class.
1719   Members = MF.remove(Members, Other);
1720   // Now only the current class is mapped to all the symbols.
1721   Members = MF.add(Members, *this, NewClassMembers);
1722 
1723   // 4. Update disequality relations
1724   ClassSet DisequalToOther = Other.getDisequalClasses(DisequalityInfo, CF);
1725   if (!DisequalToOther.isEmpty()) {
1726     ClassSet DisequalToThis = getDisequalClasses(DisequalityInfo, CF);
1727     DisequalityInfo = DF.remove(DisequalityInfo, Other);
1728 
1729     for (EquivalenceClass DisequalClass : DisequalToOther) {
1730       DisequalToThis = CF.add(DisequalToThis, DisequalClass);
1731 
1732       // Disequality is a symmetric relation meaning that if
1733       // DisequalToOther not null then the set for DisequalClass is not
1734       // empty and has at least Other.
1735       ClassSet OriginalSetLinkedToOther =
1736           *DisequalityInfo.lookup(DisequalClass);
1737 
1738       // Other will be eliminated and we should replace it with the bigger
1739       // united class.
1740       ClassSet NewSet = CF.remove(OriginalSetLinkedToOther, Other);
1741       NewSet = CF.add(NewSet, *this);
1742 
1743       DisequalityInfo = DF.add(DisequalityInfo, DisequalClass, NewSet);
1744     }
1745 
1746     DisequalityInfo = DF.add(DisequalityInfo, *this, DisequalToThis);
1747     State = State->set<DisequalityMap>(DisequalityInfo);
1748   }
1749 
1750   // 5. Update the state
1751   State = State->set<ClassMap>(Classes);
1752   State = State->set<ClassMembers>(Members);
1753 
1754   return State;
1755 }
1756 
1757 inline SymbolSet::Factory &
getMembersFactory(ProgramStateRef State)1758 EquivalenceClass::getMembersFactory(ProgramStateRef State) {
1759   return State->get_context<SymbolSet>();
1760 }
1761 
getClassMembers(ProgramStateRef State) const1762 SymbolSet EquivalenceClass::getClassMembers(ProgramStateRef State) const {
1763   if (const SymbolSet *Members = State->get<ClassMembers>(*this))
1764     return *Members;
1765 
1766   // This class is trivial, so we need to construct a set
1767   // with just that one symbol from the class.
1768   SymbolSet::Factory &F = getMembersFactory(State);
1769   return F.add(F.getEmptySet(), getRepresentativeSymbol());
1770 }
1771 
isTrivial(ProgramStateRef State) const1772 bool EquivalenceClass::isTrivial(ProgramStateRef State) const {
1773   return State->get<ClassMembers>(*this) == nullptr;
1774 }
1775 
isTriviallyDead(ProgramStateRef State,SymbolReaper & Reaper) const1776 bool EquivalenceClass::isTriviallyDead(ProgramStateRef State,
1777                                        SymbolReaper &Reaper) const {
1778   return isTrivial(State) && Reaper.isDead(getRepresentativeSymbol());
1779 }
1780 
markDisequal(BasicValueFactory & VF,RangeSet::Factory & RF,ProgramStateRef State,SymbolRef First,SymbolRef Second)1781 inline ProgramStateRef EquivalenceClass::markDisequal(BasicValueFactory &VF,
1782                                                       RangeSet::Factory &RF,
1783                                                       ProgramStateRef State,
1784                                                       SymbolRef First,
1785                                                       SymbolRef Second) {
1786   return markDisequal(VF, RF, State, find(State, First), find(State, Second));
1787 }
1788 
markDisequal(BasicValueFactory & VF,RangeSet::Factory & RF,ProgramStateRef State,EquivalenceClass First,EquivalenceClass Second)1789 inline ProgramStateRef EquivalenceClass::markDisequal(BasicValueFactory &VF,
1790                                                       RangeSet::Factory &RF,
1791                                                       ProgramStateRef State,
1792                                                       EquivalenceClass First,
1793                                                       EquivalenceClass Second) {
1794   return First.markDisequal(VF, RF, State, Second);
1795 }
1796 
1797 inline ProgramStateRef
markDisequal(BasicValueFactory & VF,RangeSet::Factory & RF,ProgramStateRef State,EquivalenceClass Other) const1798 EquivalenceClass::markDisequal(BasicValueFactory &VF, RangeSet::Factory &RF,
1799                                ProgramStateRef State,
1800                                EquivalenceClass Other) const {
1801   // If we know that two classes are equal, we can only produce an infeasible
1802   // state.
1803   if (*this == Other) {
1804     return nullptr;
1805   }
1806 
1807   DisequalityMapTy DisequalityInfo = State->get<DisequalityMap>();
1808   ConstraintRangeTy Constraints = State->get<ConstraintRange>();
1809 
1810   // Disequality is a symmetric relation, so if we mark A as disequal to B,
1811   // we should also mark B as disequalt to A.
1812   if (!addToDisequalityInfo(DisequalityInfo, Constraints, VF, RF, State, *this,
1813                             Other) ||
1814       !addToDisequalityInfo(DisequalityInfo, Constraints, VF, RF, State, Other,
1815                             *this))
1816     return nullptr;
1817 
1818   assert(areFeasible(Constraints) && "Constraint manager shouldn't produce "
1819                                      "a state with infeasible constraints");
1820 
1821   State = State->set<DisequalityMap>(DisequalityInfo);
1822   State = State->set<ConstraintRange>(Constraints);
1823 
1824   return State;
1825 }
1826 
addToDisequalityInfo(DisequalityMapTy & Info,ConstraintRangeTy & Constraints,BasicValueFactory & VF,RangeSet::Factory & RF,ProgramStateRef State,EquivalenceClass First,EquivalenceClass Second)1827 inline bool EquivalenceClass::addToDisequalityInfo(
1828     DisequalityMapTy &Info, ConstraintRangeTy &Constraints,
1829     BasicValueFactory &VF, RangeSet::Factory &RF, ProgramStateRef State,
1830     EquivalenceClass First, EquivalenceClass Second) {
1831 
1832   // 1. Get all of the required factories.
1833   DisequalityMapTy::Factory &F = State->get_context<DisequalityMap>();
1834   ClassSet::Factory &CF = State->get_context<ClassSet>();
1835   ConstraintRangeTy::Factory &CRF = State->get_context<ConstraintRange>();
1836 
1837   // 2. Add Second to the set of classes disequal to First.
1838   const ClassSet *CurrentSet = Info.lookup(First);
1839   ClassSet NewSet = CurrentSet ? *CurrentSet : CF.getEmptySet();
1840   NewSet = CF.add(NewSet, Second);
1841 
1842   Info = F.add(Info, First, NewSet);
1843 
1844   // 3. If Second is known to be a constant, we can delete this point
1845   //    from the constraint asociated with First.
1846   //
1847   //    So, if Second == 10, it means that First != 10.
1848   //    At the same time, the same logic does not apply to ranges.
1849   if (const RangeSet *SecondConstraint = Constraints.lookup(Second))
1850     if (const llvm::APSInt *Point = SecondConstraint->getConcreteValue()) {
1851 
1852       RangeSet FirstConstraint = SymbolicRangeInferrer::inferRange(
1853           VF, RF, State, First.getRepresentativeSymbol());
1854 
1855       FirstConstraint = RF.deletePoint(FirstConstraint, *Point);
1856 
1857       // If the First class is about to be constrained with an empty
1858       // range-set, the state is infeasible.
1859       if (FirstConstraint.isEmpty())
1860         return false;
1861 
1862       Constraints = CRF.add(Constraints, First, FirstConstraint);
1863     }
1864 
1865   return true;
1866 }
1867 
areEqual(ProgramStateRef State,SymbolRef FirstSym,SymbolRef SecondSym)1868 inline Optional<bool> EquivalenceClass::areEqual(ProgramStateRef State,
1869                                                  SymbolRef FirstSym,
1870                                                  SymbolRef SecondSym) {
1871   EquivalenceClass First = find(State, FirstSym);
1872   EquivalenceClass Second = find(State, SecondSym);
1873 
1874   // The same equivalence class => symbols are equal.
1875   if (First == Second)
1876     return true;
1877 
1878   // Let's check if we know anything about these two classes being not equal to
1879   // each other.
1880   ClassSet DisequalToFirst = First.getDisequalClasses(State);
1881   if (DisequalToFirst.contains(Second))
1882     return false;
1883 
1884   // It is not clear.
1885   return llvm::None;
1886 }
1887 
getDisequalClasses(ProgramStateRef State,SymbolRef Sym)1888 inline ClassSet EquivalenceClass::getDisequalClasses(ProgramStateRef State,
1889                                                      SymbolRef Sym) {
1890   return find(State, Sym).getDisequalClasses(State);
1891 }
1892 
1893 inline ClassSet
getDisequalClasses(ProgramStateRef State) const1894 EquivalenceClass::getDisequalClasses(ProgramStateRef State) const {
1895   return getDisequalClasses(State->get<DisequalityMap>(),
1896                             State->get_context<ClassSet>());
1897 }
1898 
1899 inline ClassSet
getDisequalClasses(DisequalityMapTy Map,ClassSet::Factory & Factory) const1900 EquivalenceClass::getDisequalClasses(DisequalityMapTy Map,
1901                                      ClassSet::Factory &Factory) const {
1902   if (const ClassSet *DisequalClasses = Map.lookup(*this))
1903     return *DisequalClasses;
1904 
1905   return Factory.getEmptySet();
1906 }
1907 
isClassDataConsistent(ProgramStateRef State)1908 bool EquivalenceClass::isClassDataConsistent(ProgramStateRef State) {
1909   ClassMembersTy Members = State->get<ClassMembers>();
1910 
1911   for (std::pair<EquivalenceClass, SymbolSet> ClassMembersPair : Members) {
1912     for (SymbolRef Member : ClassMembersPair.second) {
1913       // Every member of the class should have a mapping back to the class.
1914       if (find(State, Member) == ClassMembersPair.first) {
1915         continue;
1916       }
1917 
1918       return false;
1919     }
1920   }
1921 
1922   DisequalityMapTy Disequalities = State->get<DisequalityMap>();
1923   for (std::pair<EquivalenceClass, ClassSet> DisequalityInfo : Disequalities) {
1924     EquivalenceClass Class = DisequalityInfo.first;
1925     ClassSet DisequalClasses = DisequalityInfo.second;
1926 
1927     // There is no use in keeping empty sets in the map.
1928     if (DisequalClasses.isEmpty())
1929       return false;
1930 
1931     // Disequality is symmetrical, i.e. for every Class A and B that A != B,
1932     // B != A should also be true.
1933     for (EquivalenceClass DisequalClass : DisequalClasses) {
1934       const ClassSet *DisequalToDisequalClasses =
1935           Disequalities.lookup(DisequalClass);
1936 
1937       // It should be a set of at least one element: Class
1938       if (!DisequalToDisequalClasses ||
1939           !DisequalToDisequalClasses->contains(Class))
1940         return false;
1941     }
1942   }
1943 
1944   return true;
1945 }
1946 
1947 //===----------------------------------------------------------------------===//
1948 //                    RangeConstraintManager implementation
1949 //===----------------------------------------------------------------------===//
1950 
canReasonAbout(SVal X) const1951 bool RangeConstraintManager::canReasonAbout(SVal X) const {
1952   Optional<nonloc::SymbolVal> SymVal = X.getAs<nonloc::SymbolVal>();
1953   if (SymVal && SymVal->isExpression()) {
1954     const SymExpr *SE = SymVal->getSymbol();
1955 
1956     if (const SymIntExpr *SIE = dyn_cast<SymIntExpr>(SE)) {
1957       switch (SIE->getOpcode()) {
1958       // We don't reason yet about bitwise-constraints on symbolic values.
1959       case BO_And:
1960       case BO_Or:
1961       case BO_Xor:
1962         return false;
1963       // We don't reason yet about these arithmetic constraints on
1964       // symbolic values.
1965       case BO_Mul:
1966       case BO_Div:
1967       case BO_Rem:
1968       case BO_Shl:
1969       case BO_Shr:
1970         return false;
1971       // All other cases.
1972       default:
1973         return true;
1974       }
1975     }
1976 
1977     if (const SymSymExpr *SSE = dyn_cast<SymSymExpr>(SE)) {
1978       // FIXME: Handle <=> here.
1979       if (BinaryOperator::isEqualityOp(SSE->getOpcode()) ||
1980           BinaryOperator::isRelationalOp(SSE->getOpcode())) {
1981         // We handle Loc <> Loc comparisons, but not (yet) NonLoc <> NonLoc.
1982         // We've recently started producing Loc <> NonLoc comparisons (that
1983         // result from casts of one of the operands between eg. intptr_t and
1984         // void *), but we can't reason about them yet.
1985         if (Loc::isLocType(SSE->getLHS()->getType())) {
1986           return Loc::isLocType(SSE->getRHS()->getType());
1987         }
1988       }
1989     }
1990 
1991     return false;
1992   }
1993 
1994   return true;
1995 }
1996 
checkNull(ProgramStateRef State,SymbolRef Sym)1997 ConditionTruthVal RangeConstraintManager::checkNull(ProgramStateRef State,
1998                                                     SymbolRef Sym) {
1999   const RangeSet *Ranges = getConstraint(State, Sym);
2000 
2001   // If we don't have any information about this symbol, it's underconstrained.
2002   if (!Ranges)
2003     return ConditionTruthVal();
2004 
2005   // If we have a concrete value, see if it's zero.
2006   if (const llvm::APSInt *Value = Ranges->getConcreteValue())
2007     return *Value == 0;
2008 
2009   BasicValueFactory &BV = getBasicVals();
2010   APSIntType IntType = BV.getAPSIntType(Sym->getType());
2011   llvm::APSInt Zero = IntType.getZeroValue();
2012 
2013   // Check if zero is in the set of possible values.
2014   if (!Ranges->contains(Zero))
2015     return false;
2016 
2017   // Zero is a possible value, but it is not the /only/ possible value.
2018   return ConditionTruthVal();
2019 }
2020 
getSymVal(ProgramStateRef St,SymbolRef Sym) const2021 const llvm::APSInt *RangeConstraintManager::getSymVal(ProgramStateRef St,
2022                                                       SymbolRef Sym) const {
2023   const RangeSet *T = getConstraint(St, Sym);
2024   return T ? T->getConcreteValue() : nullptr;
2025 }
2026 
2027 //===----------------------------------------------------------------------===//
2028 //                Remove dead symbols from existing constraints
2029 //===----------------------------------------------------------------------===//
2030 
2031 /// Scan all symbols referenced by the constraints. If the symbol is not alive
2032 /// as marked in LSymbols, mark it as dead in DSymbols.
2033 ProgramStateRef
removeDeadBindings(ProgramStateRef State,SymbolReaper & SymReaper)2034 RangeConstraintManager::removeDeadBindings(ProgramStateRef State,
2035                                            SymbolReaper &SymReaper) {
2036   ClassMembersTy ClassMembersMap = State->get<ClassMembers>();
2037   ClassMembersTy NewClassMembersMap = ClassMembersMap;
2038   ClassMembersTy::Factory &EMFactory = State->get_context<ClassMembers>();
2039   SymbolSet::Factory &SetFactory = State->get_context<SymbolSet>();
2040 
2041   ConstraintRangeTy Constraints = State->get<ConstraintRange>();
2042   ConstraintRangeTy NewConstraints = Constraints;
2043   ConstraintRangeTy::Factory &ConstraintFactory =
2044       State->get_context<ConstraintRange>();
2045 
2046   ClassMapTy Map = State->get<ClassMap>();
2047   ClassMapTy NewMap = Map;
2048   ClassMapTy::Factory &ClassFactory = State->get_context<ClassMap>();
2049 
2050   DisequalityMapTy Disequalities = State->get<DisequalityMap>();
2051   DisequalityMapTy::Factory &DisequalityFactory =
2052       State->get_context<DisequalityMap>();
2053   ClassSet::Factory &ClassSetFactory = State->get_context<ClassSet>();
2054 
2055   bool ClassMapChanged = false;
2056   bool MembersMapChanged = false;
2057   bool ConstraintMapChanged = false;
2058   bool DisequalitiesChanged = false;
2059 
2060   auto removeDeadClass = [&](EquivalenceClass Class) {
2061     // Remove associated constraint ranges.
2062     Constraints = ConstraintFactory.remove(Constraints, Class);
2063     ConstraintMapChanged = true;
2064 
2065     // Update disequality information to not hold any information on the
2066     // removed class.
2067     ClassSet DisequalClasses =
2068         Class.getDisequalClasses(Disequalities, ClassSetFactory);
2069     if (!DisequalClasses.isEmpty()) {
2070       for (EquivalenceClass DisequalClass : DisequalClasses) {
2071         ClassSet DisequalToDisequalSet =
2072             DisequalClass.getDisequalClasses(Disequalities, ClassSetFactory);
2073         // DisequalToDisequalSet is guaranteed to be non-empty for consistent
2074         // disequality info.
2075         assert(!DisequalToDisequalSet.isEmpty());
2076         ClassSet NewSet = ClassSetFactory.remove(DisequalToDisequalSet, Class);
2077 
2078         // No need in keeping an empty set.
2079         if (NewSet.isEmpty()) {
2080           Disequalities =
2081               DisequalityFactory.remove(Disequalities, DisequalClass);
2082         } else {
2083           Disequalities =
2084               DisequalityFactory.add(Disequalities, DisequalClass, NewSet);
2085         }
2086       }
2087       // Remove the data for the class
2088       Disequalities = DisequalityFactory.remove(Disequalities, Class);
2089       DisequalitiesChanged = true;
2090     }
2091   };
2092 
2093   // 1. Let's see if dead symbols are trivial and have associated constraints.
2094   for (std::pair<EquivalenceClass, RangeSet> ClassConstraintPair :
2095        Constraints) {
2096     EquivalenceClass Class = ClassConstraintPair.first;
2097     if (Class.isTriviallyDead(State, SymReaper)) {
2098       // If this class is trivial, we can remove its constraints right away.
2099       removeDeadClass(Class);
2100     }
2101   }
2102 
2103   // 2. We don't need to track classes for dead symbols.
2104   for (std::pair<SymbolRef, EquivalenceClass> SymbolClassPair : Map) {
2105     SymbolRef Sym = SymbolClassPair.first;
2106 
2107     if (SymReaper.isDead(Sym)) {
2108       ClassMapChanged = true;
2109       NewMap = ClassFactory.remove(NewMap, Sym);
2110     }
2111   }
2112 
2113   // 3. Remove dead members from classes and remove dead non-trivial classes
2114   //    and their constraints.
2115   for (std::pair<EquivalenceClass, SymbolSet> ClassMembersPair :
2116        ClassMembersMap) {
2117     EquivalenceClass Class = ClassMembersPair.first;
2118     SymbolSet LiveMembers = ClassMembersPair.second;
2119     bool MembersChanged = false;
2120 
2121     for (SymbolRef Member : ClassMembersPair.second) {
2122       if (SymReaper.isDead(Member)) {
2123         MembersChanged = true;
2124         LiveMembers = SetFactory.remove(LiveMembers, Member);
2125       }
2126     }
2127 
2128     // Check if the class changed.
2129     if (!MembersChanged)
2130       continue;
2131 
2132     MembersMapChanged = true;
2133 
2134     if (LiveMembers.isEmpty()) {
2135       // The class is dead now, we need to wipe it out of the members map...
2136       NewClassMembersMap = EMFactory.remove(NewClassMembersMap, Class);
2137 
2138       // ...and remove all of its constraints.
2139       removeDeadClass(Class);
2140     } else {
2141       // We need to change the members associated with the class.
2142       NewClassMembersMap =
2143           EMFactory.add(NewClassMembersMap, Class, LiveMembers);
2144     }
2145   }
2146 
2147   // 4. Update the state with new maps.
2148   //
2149   // Here we try to be humble and update a map only if it really changed.
2150   if (ClassMapChanged)
2151     State = State->set<ClassMap>(NewMap);
2152 
2153   if (MembersMapChanged)
2154     State = State->set<ClassMembers>(NewClassMembersMap);
2155 
2156   if (ConstraintMapChanged)
2157     State = State->set<ConstraintRange>(Constraints);
2158 
2159   if (DisequalitiesChanged)
2160     State = State->set<DisequalityMap>(Disequalities);
2161 
2162   assert(EquivalenceClass::isClassDataConsistent(State));
2163 
2164   return State;
2165 }
2166 
getRange(ProgramStateRef State,SymbolRef Sym)2167 RangeSet RangeConstraintManager::getRange(ProgramStateRef State,
2168                                           SymbolRef Sym) {
2169   return SymbolicRangeInferrer::inferRange(getBasicVals(), F, State, Sym);
2170 }
2171 
getRange(ProgramStateRef State,EquivalenceClass Class)2172 RangeSet RangeConstraintManager::getRange(ProgramStateRef State,
2173                                           EquivalenceClass Class) {
2174   return SymbolicRangeInferrer::inferRange(getBasicVals(), F, State, Class);
2175 }
2176 
2177 //===------------------------------------------------------------------------===
2178 // assumeSymX methods: protected interface for RangeConstraintManager.
2179 //===------------------------------------------------------------------------===/
2180 
2181 // The syntax for ranges below is mathematical, using [x, y] for closed ranges
2182 // and (x, y) for open ranges. These ranges are modular, corresponding with
2183 // a common treatment of C integer overflow. This means that these methods
2184 // do not have to worry about overflow; RangeSet::Intersect can handle such a
2185 // "wraparound" range.
2186 // As an example, the range [UINT_MAX-1, 3) contains five values: UINT_MAX-1,
2187 // UINT_MAX, 0, 1, and 2.
2188 
2189 ProgramStateRef
assumeSymNE(ProgramStateRef St,SymbolRef Sym,const llvm::APSInt & Int,const llvm::APSInt & Adjustment)2190 RangeConstraintManager::assumeSymNE(ProgramStateRef St, SymbolRef Sym,
2191                                     const llvm::APSInt &Int,
2192                                     const llvm::APSInt &Adjustment) {
2193   // Before we do any real work, see if the value can even show up.
2194   APSIntType AdjustmentType(Adjustment);
2195   if (AdjustmentType.testInRange(Int, true) != APSIntType::RTR_Within)
2196     return St;
2197 
2198   llvm::APSInt Point = AdjustmentType.convert(Int) - Adjustment;
2199 
2200   RangeSet New = getRange(St, Sym);
2201   New = F.deletePoint(New, Point);
2202 
2203   return trackNE(New, St, Sym, Int, Adjustment);
2204 }
2205 
2206 ProgramStateRef
assumeSymEQ(ProgramStateRef St,SymbolRef Sym,const llvm::APSInt & Int,const llvm::APSInt & Adjustment)2207 RangeConstraintManager::assumeSymEQ(ProgramStateRef St, SymbolRef Sym,
2208                                     const llvm::APSInt &Int,
2209                                     const llvm::APSInt &Adjustment) {
2210   // Before we do any real work, see if the value can even show up.
2211   APSIntType AdjustmentType(Adjustment);
2212   if (AdjustmentType.testInRange(Int, true) != APSIntType::RTR_Within)
2213     return nullptr;
2214 
2215   // [Int-Adjustment, Int-Adjustment]
2216   llvm::APSInt AdjInt = AdjustmentType.convert(Int) - Adjustment;
2217   RangeSet New = getRange(St, Sym);
2218   New = F.intersect(New, AdjInt);
2219 
2220   return trackEQ(New, St, Sym, Int, Adjustment);
2221 }
2222 
getSymLTRange(ProgramStateRef St,SymbolRef Sym,const llvm::APSInt & Int,const llvm::APSInt & Adjustment)2223 RangeSet RangeConstraintManager::getSymLTRange(ProgramStateRef St,
2224                                                SymbolRef Sym,
2225                                                const llvm::APSInt &Int,
2226                                                const llvm::APSInt &Adjustment) {
2227   // Before we do any real work, see if the value can even show up.
2228   APSIntType AdjustmentType(Adjustment);
2229   switch (AdjustmentType.testInRange(Int, true)) {
2230   case APSIntType::RTR_Below:
2231     return F.getEmptySet();
2232   case APSIntType::RTR_Within:
2233     break;
2234   case APSIntType::RTR_Above:
2235     return getRange(St, Sym);
2236   }
2237 
2238   // Special case for Int == Min. This is always false.
2239   llvm::APSInt ComparisonVal = AdjustmentType.convert(Int);
2240   llvm::APSInt Min = AdjustmentType.getMinValue();
2241   if (ComparisonVal == Min)
2242     return F.getEmptySet();
2243 
2244   llvm::APSInt Lower = Min - Adjustment;
2245   llvm::APSInt Upper = ComparisonVal - Adjustment;
2246   --Upper;
2247 
2248   RangeSet Result = getRange(St, Sym);
2249   return F.intersect(Result, Lower, Upper);
2250 }
2251 
2252 ProgramStateRef
assumeSymLT(ProgramStateRef St,SymbolRef Sym,const llvm::APSInt & Int,const llvm::APSInt & Adjustment)2253 RangeConstraintManager::assumeSymLT(ProgramStateRef St, SymbolRef Sym,
2254                                     const llvm::APSInt &Int,
2255                                     const llvm::APSInt &Adjustment) {
2256   RangeSet New = getSymLTRange(St, Sym, Int, Adjustment);
2257   return trackNE(New, St, Sym, Int, Adjustment);
2258 }
2259 
getSymGTRange(ProgramStateRef St,SymbolRef Sym,const llvm::APSInt & Int,const llvm::APSInt & Adjustment)2260 RangeSet RangeConstraintManager::getSymGTRange(ProgramStateRef St,
2261                                                SymbolRef Sym,
2262                                                const llvm::APSInt &Int,
2263                                                const llvm::APSInt &Adjustment) {
2264   // Before we do any real work, see if the value can even show up.
2265   APSIntType AdjustmentType(Adjustment);
2266   switch (AdjustmentType.testInRange(Int, true)) {
2267   case APSIntType::RTR_Below:
2268     return getRange(St, Sym);
2269   case APSIntType::RTR_Within:
2270     break;
2271   case APSIntType::RTR_Above:
2272     return F.getEmptySet();
2273   }
2274 
2275   // Special case for Int == Max. This is always false.
2276   llvm::APSInt ComparisonVal = AdjustmentType.convert(Int);
2277   llvm::APSInt Max = AdjustmentType.getMaxValue();
2278   if (ComparisonVal == Max)
2279     return F.getEmptySet();
2280 
2281   llvm::APSInt Lower = ComparisonVal - Adjustment;
2282   llvm::APSInt Upper = Max - Adjustment;
2283   ++Lower;
2284 
2285   RangeSet SymRange = getRange(St, Sym);
2286   return F.intersect(SymRange, Lower, Upper);
2287 }
2288 
2289 ProgramStateRef
assumeSymGT(ProgramStateRef St,SymbolRef Sym,const llvm::APSInt & Int,const llvm::APSInt & Adjustment)2290 RangeConstraintManager::assumeSymGT(ProgramStateRef St, SymbolRef Sym,
2291                                     const llvm::APSInt &Int,
2292                                     const llvm::APSInt &Adjustment) {
2293   RangeSet New = getSymGTRange(St, Sym, Int, Adjustment);
2294   return trackNE(New, St, Sym, Int, Adjustment);
2295 }
2296 
getSymGERange(ProgramStateRef St,SymbolRef Sym,const llvm::APSInt & Int,const llvm::APSInt & Adjustment)2297 RangeSet RangeConstraintManager::getSymGERange(ProgramStateRef St,
2298                                                SymbolRef Sym,
2299                                                const llvm::APSInt &Int,
2300                                                const llvm::APSInt &Adjustment) {
2301   // Before we do any real work, see if the value can even show up.
2302   APSIntType AdjustmentType(Adjustment);
2303   switch (AdjustmentType.testInRange(Int, true)) {
2304   case APSIntType::RTR_Below:
2305     return getRange(St, Sym);
2306   case APSIntType::RTR_Within:
2307     break;
2308   case APSIntType::RTR_Above:
2309     return F.getEmptySet();
2310   }
2311 
2312   // Special case for Int == Min. This is always feasible.
2313   llvm::APSInt ComparisonVal = AdjustmentType.convert(Int);
2314   llvm::APSInt Min = AdjustmentType.getMinValue();
2315   if (ComparisonVal == Min)
2316     return getRange(St, Sym);
2317 
2318   llvm::APSInt Max = AdjustmentType.getMaxValue();
2319   llvm::APSInt Lower = ComparisonVal - Adjustment;
2320   llvm::APSInt Upper = Max - Adjustment;
2321 
2322   RangeSet SymRange = getRange(St, Sym);
2323   return F.intersect(SymRange, Lower, Upper);
2324 }
2325 
2326 ProgramStateRef
assumeSymGE(ProgramStateRef St,SymbolRef Sym,const llvm::APSInt & Int,const llvm::APSInt & Adjustment)2327 RangeConstraintManager::assumeSymGE(ProgramStateRef St, SymbolRef Sym,
2328                                     const llvm::APSInt &Int,
2329                                     const llvm::APSInt &Adjustment) {
2330   RangeSet New = getSymGERange(St, Sym, Int, Adjustment);
2331   return New.isEmpty() ? nullptr : setConstraint(St, Sym, New);
2332 }
2333 
2334 RangeSet
getSymLERange(llvm::function_ref<RangeSet ()> RS,const llvm::APSInt & Int,const llvm::APSInt & Adjustment)2335 RangeConstraintManager::getSymLERange(llvm::function_ref<RangeSet()> RS,
2336                                       const llvm::APSInt &Int,
2337                                       const llvm::APSInt &Adjustment) {
2338   // Before we do any real work, see if the value can even show up.
2339   APSIntType AdjustmentType(Adjustment);
2340   switch (AdjustmentType.testInRange(Int, true)) {
2341   case APSIntType::RTR_Below:
2342     return F.getEmptySet();
2343   case APSIntType::RTR_Within:
2344     break;
2345   case APSIntType::RTR_Above:
2346     return RS();
2347   }
2348 
2349   // Special case for Int == Max. This is always feasible.
2350   llvm::APSInt ComparisonVal = AdjustmentType.convert(Int);
2351   llvm::APSInt Max = AdjustmentType.getMaxValue();
2352   if (ComparisonVal == Max)
2353     return RS();
2354 
2355   llvm::APSInt Min = AdjustmentType.getMinValue();
2356   llvm::APSInt Lower = Min - Adjustment;
2357   llvm::APSInt Upper = ComparisonVal - Adjustment;
2358 
2359   RangeSet Default = RS();
2360   return F.intersect(Default, Lower, Upper);
2361 }
2362 
getSymLERange(ProgramStateRef St,SymbolRef Sym,const llvm::APSInt & Int,const llvm::APSInt & Adjustment)2363 RangeSet RangeConstraintManager::getSymLERange(ProgramStateRef St,
2364                                                SymbolRef Sym,
2365                                                const llvm::APSInt &Int,
2366                                                const llvm::APSInt &Adjustment) {
2367   return getSymLERange([&] { return getRange(St, Sym); }, Int, Adjustment);
2368 }
2369 
2370 ProgramStateRef
assumeSymLE(ProgramStateRef St,SymbolRef Sym,const llvm::APSInt & Int,const llvm::APSInt & Adjustment)2371 RangeConstraintManager::assumeSymLE(ProgramStateRef St, SymbolRef Sym,
2372                                     const llvm::APSInt &Int,
2373                                     const llvm::APSInt &Adjustment) {
2374   RangeSet New = getSymLERange(St, Sym, Int, Adjustment);
2375   return New.isEmpty() ? nullptr : setConstraint(St, Sym, New);
2376 }
2377 
assumeSymWithinInclusiveRange(ProgramStateRef State,SymbolRef Sym,const llvm::APSInt & From,const llvm::APSInt & To,const llvm::APSInt & Adjustment)2378 ProgramStateRef RangeConstraintManager::assumeSymWithinInclusiveRange(
2379     ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From,
2380     const llvm::APSInt &To, const llvm::APSInt &Adjustment) {
2381   RangeSet New = getSymGERange(State, Sym, From, Adjustment);
2382   if (New.isEmpty())
2383     return nullptr;
2384   RangeSet Out = getSymLERange([&] { return New; }, To, Adjustment);
2385   return Out.isEmpty() ? nullptr : setConstraint(State, Sym, Out);
2386 }
2387 
assumeSymOutsideInclusiveRange(ProgramStateRef State,SymbolRef Sym,const llvm::APSInt & From,const llvm::APSInt & To,const llvm::APSInt & Adjustment)2388 ProgramStateRef RangeConstraintManager::assumeSymOutsideInclusiveRange(
2389     ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From,
2390     const llvm::APSInt &To, const llvm::APSInt &Adjustment) {
2391   RangeSet RangeLT = getSymLTRange(State, Sym, From, Adjustment);
2392   RangeSet RangeGT = getSymGTRange(State, Sym, To, Adjustment);
2393   RangeSet New(F.add(RangeLT, RangeGT));
2394   return New.isEmpty() ? nullptr : setConstraint(State, Sym, New);
2395 }
2396 
2397 //===----------------------------------------------------------------------===//
2398 // Pretty-printing.
2399 //===----------------------------------------------------------------------===//
2400 
printJson(raw_ostream & Out,ProgramStateRef State,const char * NL,unsigned int Space,bool IsDot) const2401 void RangeConstraintManager::printJson(raw_ostream &Out, ProgramStateRef State,
2402                                        const char *NL, unsigned int Space,
2403                                        bool IsDot) const {
2404   ConstraintRangeTy Constraints = State->get<ConstraintRange>();
2405 
2406   Indent(Out, Space, IsDot) << "\"constraints\": ";
2407   if (Constraints.isEmpty()) {
2408     Out << "null," << NL;
2409     return;
2410   }
2411 
2412   ++Space;
2413   Out << '[' << NL;
2414   bool First = true;
2415   for (std::pair<EquivalenceClass, RangeSet> P : Constraints) {
2416     SymbolSet ClassMembers = P.first.getClassMembers(State);
2417 
2418     // We can print the same constraint for every class member.
2419     for (SymbolRef ClassMember : ClassMembers) {
2420       if (First) {
2421         First = false;
2422       } else {
2423         Out << ',';
2424         Out << NL;
2425       }
2426       Indent(Out, Space, IsDot)
2427           << "{ \"symbol\": \"" << ClassMember << "\", \"range\": \"";
2428       P.second.dump(Out);
2429       Out << "\" }";
2430     }
2431   }
2432   Out << NL;
2433 
2434   --Space;
2435   Indent(Out, Space, IsDot) << "]," << NL;
2436 }
2437