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/ADT/StringExtras.h"
24 #include "llvm/ADT/SmallSet.h"
25 #include "llvm/Support/Compiler.h"
26 #include "llvm/Support/raw_ostream.h"
27 #include <algorithm>
28 #include <iterator>
29 
30 using namespace clang;
31 using namespace ento;
32 
33 // This class can be extended with other tables which will help to reason
34 // about ranges more precisely.
35 class OperatorRelationsTable {
36   static_assert(BO_LT < BO_GT && BO_GT < BO_LE && BO_LE < BO_GE &&
37                     BO_GE < BO_EQ && BO_EQ < BO_NE,
38                 "This class relies on operators order. Rework it otherwise.");
39 
40 public:
41   enum TriStateKind {
42     False = 0,
43     True,
44     Unknown,
45   };
46 
47 private:
48   // CmpOpTable holds states which represent the corresponding range for
49   // branching an exploded graph. We can reason about the branch if there is
50   // a previously known fact of the existence of a comparison expression with
51   // operands used in the current expression.
52   // E.g. assuming (x < y) is true that means (x != y) is surely true.
53   // if (x previous_operation y)  // <    | !=      | >
54   //   if (x operation y)         // !=   | >       | <
55   //     tristate                 // True | Unknown | False
56   //
57   // CmpOpTable represents next:
58   // __|< |> |<=|>=|==|!=|UnknownX2|
59   // < |1 |0 |* |0 |0 |* |1        |
60   // > |0 |1 |0 |* |0 |* |1        |
61   // <=|1 |0 |1 |* |1 |* |0        |
62   // >=|0 |1 |* |1 |1 |* |0        |
63   // ==|0 |0 |* |* |1 |0 |1        |
64   // !=|1 |1 |* |* |0 |1 |0        |
65   //
66   // Columns stands for a previous operator.
67   // Rows stands for a current operator.
68   // Each row has exactly two `Unknown` cases.
69   // UnknownX2 means that both `Unknown` previous operators are met in code,
70   // and there is a special column for that, for example:
71   // if (x >= y)
72   //   if (x != y)
73   //     if (x <= y)
74   //       False only
75   static constexpr size_t CmpOpCount = BO_NE - BO_LT + 1;
76   const TriStateKind CmpOpTable[CmpOpCount][CmpOpCount + 1] = {
77       // <      >      <=     >=     ==     !=    UnknownX2
78       {True, False, Unknown, False, False, Unknown, True}, // <
79       {False, True, False, Unknown, False, Unknown, True}, // >
80       {True, False, True, Unknown, True, Unknown, False},  // <=
81       {False, True, Unknown, True, True, Unknown, False},  // >=
82       {False, False, Unknown, Unknown, True, False, True}, // ==
83       {True, True, Unknown, Unknown, False, True, False},  // !=
84   };
85 
86   static size_t getIndexFromOp(BinaryOperatorKind OP) {
87     return static_cast<size_t>(OP - BO_LT);
88   }
89 
90 public:
91   constexpr size_t getCmpOpCount() const { return CmpOpCount; }
92 
93   static BinaryOperatorKind getOpFromIndex(size_t Index) {
94     return static_cast<BinaryOperatorKind>(Index + BO_LT);
95   }
96 
97   TriStateKind getCmpOpState(BinaryOperatorKind CurrentOP,
98                              BinaryOperatorKind QueriedOP) const {
99     return CmpOpTable[getIndexFromOp(CurrentOP)][getIndexFromOp(QueriedOP)];
100   }
101 
102   TriStateKind getCmpOpStateForUnknownX2(BinaryOperatorKind CurrentOP) const {
103     return CmpOpTable[getIndexFromOp(CurrentOP)][CmpOpCount];
104   }
105 };
106 
107 //===----------------------------------------------------------------------===//
108 //                           RangeSet implementation
109 //===----------------------------------------------------------------------===//
110 
111 RangeSet::ContainerType RangeSet::Factory::EmptySet{};
112 
113 RangeSet RangeSet::Factory::add(RangeSet Original, Range Element) {
114   ContainerType Result;
115   Result.reserve(Original.size() + 1);
116 
117   const_iterator Lower = llvm::lower_bound(Original, Element);
118   Result.insert(Result.end(), Original.begin(), Lower);
119   Result.push_back(Element);
120   Result.insert(Result.end(), Lower, Original.end());
121 
122   return makePersistent(std::move(Result));
123 }
124 
125 RangeSet RangeSet::Factory::add(RangeSet Original, const llvm::APSInt &Point) {
126   return add(Original, Range(Point));
127 }
128 
129 RangeSet RangeSet::Factory::getRangeSet(Range From) {
130   ContainerType Result;
131   Result.push_back(From);
132   return makePersistent(std::move(Result));
133 }
134 
135 RangeSet RangeSet::Factory::makePersistent(ContainerType &&From) {
136   llvm::FoldingSetNodeID ID;
137   void *InsertPos;
138 
139   From.Profile(ID);
140   ContainerType *Result = Cache.FindNodeOrInsertPos(ID, InsertPos);
141 
142   if (!Result) {
143     // It is cheaper to fully construct the resulting range on stack
144     // and move it to the freshly allocated buffer if we don't have
145     // a set like this already.
146     Result = construct(std::move(From));
147     Cache.InsertNode(Result, InsertPos);
148   }
149 
150   return Result;
151 }
152 
153 RangeSet::ContainerType *RangeSet::Factory::construct(ContainerType &&From) {
154   void *Buffer = Arena.Allocate();
155   return new (Buffer) ContainerType(std::move(From));
156 }
157 
158 RangeSet RangeSet::Factory::add(RangeSet LHS, RangeSet RHS) {
159   ContainerType Result;
160   std::merge(LHS.begin(), LHS.end(), RHS.begin(), RHS.end(),
161              std::back_inserter(Result));
162   return makePersistent(std::move(Result));
163 }
164 
165 const llvm::APSInt &RangeSet::getMinValue() const {
166   assert(!isEmpty());
167   return begin()->From();
168 }
169 
170 const llvm::APSInt &RangeSet::getMaxValue() const {
171   assert(!isEmpty());
172   return std::prev(end())->To();
173 }
174 
175 bool RangeSet::containsImpl(llvm::APSInt &Point) const {
176   if (isEmpty() || !pin(Point))
177     return false;
178 
179   Range Dummy(Point);
180   const_iterator It = llvm::upper_bound(*this, Dummy);
181   if (It == begin())
182     return false;
183 
184   return std::prev(It)->Includes(Point);
185 }
186 
187 bool RangeSet::pin(llvm::APSInt &Point) const {
188   APSIntType Type(getMinValue());
189   if (Type.testInRange(Point, true) != APSIntType::RTR_Within)
190     return false;
191 
192   Type.apply(Point);
193   return true;
194 }
195 
196 bool RangeSet::pin(llvm::APSInt &Lower, llvm::APSInt &Upper) const {
197   // This function has nine cases, the cartesian product of range-testing
198   // both the upper and lower bounds against the symbol's type.
199   // Each case requires a different pinning operation.
200   // The function returns false if the described range is entirely outside
201   // the range of values for the associated symbol.
202   APSIntType Type(getMinValue());
203   APSIntType::RangeTestResultKind LowerTest = Type.testInRange(Lower, true);
204   APSIntType::RangeTestResultKind UpperTest = Type.testInRange(Upper, true);
205 
206   switch (LowerTest) {
207   case APSIntType::RTR_Below:
208     switch (UpperTest) {
209     case APSIntType::RTR_Below:
210       // The entire range is outside the symbol's set of possible values.
211       // If this is a conventionally-ordered range, the state is infeasible.
212       if (Lower <= Upper)
213         return false;
214 
215       // However, if the range wraps around, it spans all possible values.
216       Lower = Type.getMinValue();
217       Upper = Type.getMaxValue();
218       break;
219     case APSIntType::RTR_Within:
220       // The range starts below what's possible but ends within it. Pin.
221       Lower = Type.getMinValue();
222       Type.apply(Upper);
223       break;
224     case APSIntType::RTR_Above:
225       // The range spans all possible values for the symbol. Pin.
226       Lower = Type.getMinValue();
227       Upper = Type.getMaxValue();
228       break;
229     }
230     break;
231   case APSIntType::RTR_Within:
232     switch (UpperTest) {
233     case APSIntType::RTR_Below:
234       // The range wraps around, but all lower values are not possible.
235       Type.apply(Lower);
236       Upper = Type.getMaxValue();
237       break;
238     case APSIntType::RTR_Within:
239       // The range may or may not wrap around, but both limits are valid.
240       Type.apply(Lower);
241       Type.apply(Upper);
242       break;
243     case APSIntType::RTR_Above:
244       // The range starts within what's possible but ends above it. Pin.
245       Type.apply(Lower);
246       Upper = Type.getMaxValue();
247       break;
248     }
249     break;
250   case APSIntType::RTR_Above:
251     switch (UpperTest) {
252     case APSIntType::RTR_Below:
253       // The range wraps but is outside the symbol's set of possible values.
254       return false;
255     case APSIntType::RTR_Within:
256       // The range starts above what's possible but ends within it (wrap).
257       Lower = Type.getMinValue();
258       Type.apply(Upper);
259       break;
260     case APSIntType::RTR_Above:
261       // The entire range is outside the symbol's set of possible values.
262       // If this is a conventionally-ordered range, the state is infeasible.
263       if (Lower <= Upper)
264         return false;
265 
266       // However, if the range wraps around, it spans all possible values.
267       Lower = Type.getMinValue();
268       Upper = Type.getMaxValue();
269       break;
270     }
271     break;
272   }
273 
274   return true;
275 }
276 
277 RangeSet RangeSet::Factory::intersect(RangeSet What, llvm::APSInt Lower,
278                                       llvm::APSInt Upper) {
279   if (What.isEmpty() || !What.pin(Lower, Upper))
280     return getEmptySet();
281 
282   ContainerType DummyContainer;
283 
284   if (Lower <= Upper) {
285     // [Lower, Upper] is a regular range.
286     //
287     // Shortcut: check that there is even a possibility of the intersection
288     //           by checking the two following situations:
289     //
290     //               <---[  What  ]---[------]------>
291     //                              Lower  Upper
292     //                            -or-
293     //               <----[------]----[  What  ]---->
294     //                  Lower  Upper
295     if (What.getMaxValue() < Lower || Upper < What.getMinValue())
296       return getEmptySet();
297 
298     DummyContainer.push_back(
299         Range(ValueFactory.getValue(Lower), ValueFactory.getValue(Upper)));
300   } else {
301     // [Lower, Upper] is an inverted range, i.e. [MIN, Upper] U [Lower, MAX]
302     //
303     // Shortcut: check that there is even a possibility of the intersection
304     //           by checking the following situation:
305     //
306     //               <------]---[  What  ]---[------>
307     //                    Upper             Lower
308     if (What.getMaxValue() < Lower && Upper < What.getMinValue())
309       return getEmptySet();
310 
311     DummyContainer.push_back(
312         Range(ValueFactory.getMinValue(Upper), ValueFactory.getValue(Upper)));
313     DummyContainer.push_back(
314         Range(ValueFactory.getValue(Lower), ValueFactory.getMaxValue(Lower)));
315   }
316 
317   return intersect(*What.Impl, DummyContainer);
318 }
319 
320 RangeSet RangeSet::Factory::intersect(const RangeSet::ContainerType &LHS,
321                                       const RangeSet::ContainerType &RHS) {
322   ContainerType Result;
323   Result.reserve(std::max(LHS.size(), RHS.size()));
324 
325   const_iterator First = LHS.begin(), Second = RHS.begin(),
326                  FirstEnd = LHS.end(), SecondEnd = RHS.end();
327 
328   const auto SwapIterators = [&First, &FirstEnd, &Second, &SecondEnd]() {
329     std::swap(First, Second);
330     std::swap(FirstEnd, SecondEnd);
331   };
332 
333   // If we ran out of ranges in one set, but not in the other,
334   // it means that those elements are definitely not in the
335   // intersection.
336   while (First != FirstEnd && Second != SecondEnd) {
337     // We want to keep the following invariant at all times:
338     //
339     //    ----[ First ---------------------->
340     //    --------[ Second ----------------->
341     if (Second->From() < First->From())
342       SwapIterators();
343 
344     // Loop where the invariant holds:
345     do {
346       // Check for the following situation:
347       //
348       //    ----[ First ]--------------------->
349       //    ---------------[ Second ]--------->
350       //
351       // which means that...
352       if (Second->From() > First->To()) {
353         // ...First is not in the intersection.
354         //
355         // We should move on to the next range after First and break out of the
356         // loop because the invariant might not be true.
357         ++First;
358         break;
359       }
360 
361       // We have a guaranteed intersection at this point!
362       // And this is the current situation:
363       //
364       //    ----[   First   ]----------------->
365       //    -------[ Second ------------------>
366       //
367       // Additionally, it definitely starts with Second->From().
368       const llvm::APSInt &IntersectionStart = Second->From();
369 
370       // It is important to know which of the two ranges' ends
371       // is greater.  That "longer" range might have some other
372       // intersections, while the "shorter" range might not.
373       if (Second->To() > First->To()) {
374         // Here we make a decision to keep First as the "longer"
375         // range.
376         SwapIterators();
377       }
378 
379       // At this point, we have the following situation:
380       //
381       //    ---- First      ]-------------------->
382       //    ---- Second ]--[  Second+1 ---------->
383       //
384       // We don't know the relationship between First->From and
385       // Second->From and we don't know whether Second+1 intersects
386       // with First.
387       //
388       // However, we know that [IntersectionStart, Second->To] is
389       // a part of the intersection...
390       Result.push_back(Range(IntersectionStart, Second->To()));
391       ++Second;
392       // ...and that the invariant will hold for a valid Second+1
393       // because First->From <= Second->To < (Second+1)->From.
394     } while (Second != SecondEnd);
395   }
396 
397   if (Result.empty())
398     return getEmptySet();
399 
400   return makePersistent(std::move(Result));
401 }
402 
403 RangeSet RangeSet::Factory::intersect(RangeSet LHS, RangeSet RHS) {
404   // Shortcut: let's see if the intersection is even possible.
405   if (LHS.isEmpty() || RHS.isEmpty() || LHS.getMaxValue() < RHS.getMinValue() ||
406       RHS.getMaxValue() < LHS.getMinValue())
407     return getEmptySet();
408 
409   return intersect(*LHS.Impl, *RHS.Impl);
410 }
411 
412 RangeSet RangeSet::Factory::intersect(RangeSet LHS, llvm::APSInt Point) {
413   if (LHS.containsImpl(Point))
414     return getRangeSet(ValueFactory.getValue(Point));
415 
416   return getEmptySet();
417 }
418 
419 RangeSet RangeSet::Factory::negate(RangeSet What) {
420   if (What.isEmpty())
421     return getEmptySet();
422 
423   const llvm::APSInt SampleValue = What.getMinValue();
424   const llvm::APSInt &MIN = ValueFactory.getMinValue(SampleValue);
425   const llvm::APSInt &MAX = ValueFactory.getMaxValue(SampleValue);
426 
427   ContainerType Result;
428   Result.reserve(What.size() + (SampleValue == MIN));
429 
430   // Handle a special case for MIN value.
431   const_iterator It = What.begin();
432   const_iterator End = What.end();
433 
434   const llvm::APSInt &From = It->From();
435   const llvm::APSInt &To = It->To();
436 
437   if (From == MIN) {
438     // If the range [From, To] is [MIN, MAX], then result is also [MIN, MAX].
439     if (To == MAX) {
440       return What;
441     }
442 
443     const_iterator Last = std::prev(End);
444 
445     // Try to find and unite the following ranges:
446     // [MIN, MIN] & [MIN + 1, N] => [MIN, N].
447     if (Last->To() == MAX) {
448       // It means that in the original range we have ranges
449       //   [MIN, A], ... , [B, MAX]
450       // And the result should be [MIN, -B], ..., [-A, MAX]
451       Result.emplace_back(MIN, ValueFactory.getValue(-Last->From()));
452       // We already negated Last, so we can skip it.
453       End = Last;
454     } else {
455       // Add a separate range for the lowest value.
456       Result.emplace_back(MIN, MIN);
457     }
458 
459     // Skip adding the second range in case when [From, To] are [MIN, MIN].
460     if (To != MIN) {
461       Result.emplace_back(ValueFactory.getValue(-To), MAX);
462     }
463 
464     // Skip the first range in the loop.
465     ++It;
466   }
467 
468   // Negate all other ranges.
469   for (; It != End; ++It) {
470     // Negate int values.
471     const llvm::APSInt &NewFrom = ValueFactory.getValue(-It->To());
472     const llvm::APSInt &NewTo = ValueFactory.getValue(-It->From());
473 
474     // Add a negated range.
475     Result.emplace_back(NewFrom, NewTo);
476   }
477 
478   llvm::sort(Result);
479   return makePersistent(std::move(Result));
480 }
481 
482 RangeSet RangeSet::Factory::deletePoint(RangeSet From,
483                                         const llvm::APSInt &Point) {
484   if (!From.contains(Point))
485     return From;
486 
487   llvm::APSInt Upper = Point;
488   llvm::APSInt Lower = Point;
489 
490   ++Upper;
491   --Lower;
492 
493   // Notice that the lower bound is greater than the upper bound.
494   return intersect(From, Upper, Lower);
495 }
496 
497 LLVM_DUMP_METHOD void Range::dump(raw_ostream &OS) const {
498   OS << '[' << toString(From(), 10) << ", " << toString(To(), 10) << ']';
499 }
500 LLVM_DUMP_METHOD void Range::dump() const { dump(llvm::errs()); }
501 
502 LLVM_DUMP_METHOD void RangeSet::dump(raw_ostream &OS) const {
503   OS << "{ ";
504   llvm::interleaveComma(*this, OS, [&OS](const Range &R) { R.dump(OS); });
505   OS << " }";
506 }
507 LLVM_DUMP_METHOD void RangeSet::dump() const { dump(llvm::errs()); }
508 
509 REGISTER_SET_FACTORY_WITH_PROGRAMSTATE(SymbolSet, SymbolRef)
510 
511 namespace {
512 class EquivalenceClass;
513 } // end anonymous namespace
514 
515 REGISTER_MAP_WITH_PROGRAMSTATE(ClassMap, SymbolRef, EquivalenceClass)
516 REGISTER_MAP_WITH_PROGRAMSTATE(ClassMembers, EquivalenceClass, SymbolSet)
517 REGISTER_MAP_WITH_PROGRAMSTATE(ConstraintRange, EquivalenceClass, RangeSet)
518 
519 REGISTER_SET_FACTORY_WITH_PROGRAMSTATE(ClassSet, EquivalenceClass)
520 REGISTER_MAP_WITH_PROGRAMSTATE(DisequalityMap, EquivalenceClass, ClassSet)
521 
522 namespace {
523 /// This class encapsulates a set of symbols equal to each other.
524 ///
525 /// The main idea of the approach requiring such classes is in narrowing
526 /// and sharing constraints between symbols within the class.  Also we can
527 /// conclude that there is no practical need in storing constraints for
528 /// every member of the class separately.
529 ///
530 /// Main terminology:
531 ///
532 ///   * "Equivalence class" is an object of this class, which can be efficiently
533 ///     compared to other classes.  It represents the whole class without
534 ///     storing the actual in it.  The members of the class however can be
535 ///     retrieved from the state.
536 ///
537 ///   * "Class members" are the symbols corresponding to the class.  This means
538 ///     that A == B for every member symbols A and B from the class.  Members of
539 ///     each class are stored in the state.
540 ///
541 ///   * "Trivial class" is a class that has and ever had only one same symbol.
542 ///
543 ///   * "Merge operation" merges two classes into one.  It is the main operation
544 ///     to produce non-trivial classes.
545 ///     If, at some point, we can assume that two symbols from two distinct
546 ///     classes are equal, we can merge these classes.
547 class EquivalenceClass : public llvm::FoldingSetNode {
548 public:
549   /// Find equivalence class for the given symbol in the given state.
550   LLVM_NODISCARD static inline EquivalenceClass find(ProgramStateRef State,
551                                                      SymbolRef Sym);
552 
553   /// Merge classes for the given symbols and return a new state.
554   LLVM_NODISCARD static inline ProgramStateRef merge(RangeSet::Factory &F,
555                                                      ProgramStateRef State,
556                                                      SymbolRef First,
557                                                      SymbolRef Second);
558   // Merge this class with the given class and return a new state.
559   LLVM_NODISCARD inline ProgramStateRef
560   merge(RangeSet::Factory &F, ProgramStateRef State, EquivalenceClass Other);
561 
562   /// Return a set of class members for the given state.
563   LLVM_NODISCARD inline SymbolSet getClassMembers(ProgramStateRef State) const;
564 
565   /// Return true if the current class is trivial in the given state.
566   /// A class is trivial if and only if there is not any member relations stored
567   /// to it in State/ClassMembers.
568   /// An equivalence class with one member might seem as it does not hold any
569   /// meaningful information, i.e. that is a tautology. However, during the
570   /// removal of dead symbols we do not remove classes with one member for
571   /// resource and performance reasons. Consequently, a class with one member is
572   /// not necessarily trivial. It could happen that we have a class with two
573   /// members and then during the removal of dead symbols we remove one of its
574   /// members. In this case, the class is still non-trivial (it still has the
575   /// mappings in ClassMembers), even though it has only one member.
576   LLVM_NODISCARD inline bool isTrivial(ProgramStateRef State) const;
577 
578   /// Return true if the current class is trivial and its only member is dead.
579   LLVM_NODISCARD inline bool isTriviallyDead(ProgramStateRef State,
580                                              SymbolReaper &Reaper) const;
581 
582   LLVM_NODISCARD static inline ProgramStateRef
583   markDisequal(RangeSet::Factory &F, ProgramStateRef State, SymbolRef First,
584                SymbolRef Second);
585   LLVM_NODISCARD static inline ProgramStateRef
586   markDisequal(RangeSet::Factory &F, ProgramStateRef State,
587                EquivalenceClass First, EquivalenceClass Second);
588   LLVM_NODISCARD inline ProgramStateRef
589   markDisequal(RangeSet::Factory &F, ProgramStateRef State,
590                EquivalenceClass Other) const;
591   LLVM_NODISCARD static inline ClassSet
592   getDisequalClasses(ProgramStateRef State, SymbolRef Sym);
593   LLVM_NODISCARD inline ClassSet
594   getDisequalClasses(ProgramStateRef State) const;
595   LLVM_NODISCARD inline ClassSet
596   getDisequalClasses(DisequalityMapTy Map, ClassSet::Factory &Factory) const;
597 
598   LLVM_NODISCARD static inline Optional<bool> areEqual(ProgramStateRef State,
599                                                        EquivalenceClass First,
600                                                        EquivalenceClass Second);
601   LLVM_NODISCARD static inline Optional<bool>
602   areEqual(ProgramStateRef State, SymbolRef First, SymbolRef Second);
603 
604   /// Iterate over all symbols and try to simplify them.
605   LLVM_NODISCARD static inline ProgramStateRef simplify(SValBuilder &SVB,
606                                                         RangeSet::Factory &F,
607                                                         ProgramStateRef State,
608                                                         EquivalenceClass Class);
609 
610   void dumpToStream(ProgramStateRef State, raw_ostream &os) const;
611   LLVM_DUMP_METHOD void dump(ProgramStateRef State) const {
612     dumpToStream(State, llvm::errs());
613   }
614 
615   /// Check equivalence data for consistency.
616   LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED static bool
617   isClassDataConsistent(ProgramStateRef State);
618 
619   LLVM_NODISCARD QualType getType() const {
620     return getRepresentativeSymbol()->getType();
621   }
622 
623   EquivalenceClass() = delete;
624   EquivalenceClass(const EquivalenceClass &) = default;
625   EquivalenceClass &operator=(const EquivalenceClass &) = delete;
626   EquivalenceClass(EquivalenceClass &&) = default;
627   EquivalenceClass &operator=(EquivalenceClass &&) = delete;
628 
629   bool operator==(const EquivalenceClass &Other) const {
630     return ID == Other.ID;
631   }
632   bool operator<(const EquivalenceClass &Other) const { return ID < Other.ID; }
633   bool operator!=(const EquivalenceClass &Other) const {
634     return !operator==(Other);
635   }
636 
637   static void Profile(llvm::FoldingSetNodeID &ID, uintptr_t CID) {
638     ID.AddInteger(CID);
639   }
640 
641   void Profile(llvm::FoldingSetNodeID &ID) const { Profile(ID, this->ID); }
642 
643 private:
644   /* implicit */ EquivalenceClass(SymbolRef Sym)
645       : ID(reinterpret_cast<uintptr_t>(Sym)) {}
646 
647   /// This function is intended to be used ONLY within the class.
648   /// The fact that ID is a pointer to a symbol is an implementation detail
649   /// and should stay that way.
650   /// In the current implementation, we use it to retrieve the only member
651   /// of the trivial class.
652   SymbolRef getRepresentativeSymbol() const {
653     return reinterpret_cast<SymbolRef>(ID);
654   }
655   static inline SymbolSet::Factory &getMembersFactory(ProgramStateRef State);
656 
657   inline ProgramStateRef mergeImpl(RangeSet::Factory &F, ProgramStateRef State,
658                                    SymbolSet Members, EquivalenceClass Other,
659                                    SymbolSet OtherMembers);
660   static inline bool
661   addToDisequalityInfo(DisequalityMapTy &Info, ConstraintRangeTy &Constraints,
662                        RangeSet::Factory &F, ProgramStateRef State,
663                        EquivalenceClass First, EquivalenceClass Second);
664 
665   /// This is a unique identifier of the class.
666   uintptr_t ID;
667 };
668 
669 //===----------------------------------------------------------------------===//
670 //                             Constraint functions
671 //===----------------------------------------------------------------------===//
672 
673 LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED bool
674 areFeasible(ConstraintRangeTy Constraints) {
675   return llvm::none_of(
676       Constraints,
677       [](const std::pair<EquivalenceClass, RangeSet> &ClassConstraint) {
678         return ClassConstraint.second.isEmpty();
679       });
680 }
681 
682 LLVM_NODISCARD inline const RangeSet *getConstraint(ProgramStateRef State,
683                                                     EquivalenceClass Class) {
684   return State->get<ConstraintRange>(Class);
685 }
686 
687 LLVM_NODISCARD inline const RangeSet *getConstraint(ProgramStateRef State,
688                                                     SymbolRef Sym) {
689   return getConstraint(State, EquivalenceClass::find(State, Sym));
690 }
691 
692 LLVM_NODISCARD ProgramStateRef setConstraint(ProgramStateRef State,
693                                              EquivalenceClass Class,
694                                              RangeSet Constraint) {
695   return State->set<ConstraintRange>(Class, Constraint);
696 }
697 
698 LLVM_NODISCARD ProgramStateRef setConstraints(ProgramStateRef State,
699                                               ConstraintRangeTy Constraints) {
700   return State->set<ConstraintRange>(Constraints);
701 }
702 
703 //===----------------------------------------------------------------------===//
704 //                       Equality/diseqiality abstraction
705 //===----------------------------------------------------------------------===//
706 
707 /// A small helper function for detecting symbolic (dis)equality.
708 ///
709 /// Equality check can have different forms (like a == b or a - b) and this
710 /// class encapsulates those away if the only thing the user wants to check -
711 /// whether it's equality/diseqiality or not.
712 ///
713 /// \returns true if assuming this Sym to be true means equality of operands
714 ///          false if it means disequality of operands
715 ///          None otherwise
716 Optional<bool> meansEquality(const SymSymExpr *Sym) {
717   switch (Sym->getOpcode()) {
718   case BO_Sub:
719     // This case is: A - B != 0 -> disequality check.
720     return false;
721   case BO_EQ:
722     // This case is: A == B != 0 -> equality check.
723     return true;
724   case BO_NE:
725     // This case is: A != B != 0 -> diseqiality check.
726     return false;
727   default:
728     return llvm::None;
729   }
730 }
731 
732 //===----------------------------------------------------------------------===//
733 //                            Intersection functions
734 //===----------------------------------------------------------------------===//
735 
736 template <class SecondTy, class... RestTy>
737 LLVM_NODISCARD inline RangeSet intersect(RangeSet::Factory &F, RangeSet Head,
738                                          SecondTy Second, RestTy... Tail);
739 
740 template <class... RangeTy> struct IntersectionTraits;
741 
742 template <class... TailTy> struct IntersectionTraits<RangeSet, TailTy...> {
743   // Found RangeSet, no need to check any further
744   using Type = RangeSet;
745 };
746 
747 template <> struct IntersectionTraits<> {
748   // We ran out of types, and we didn't find any RangeSet, so the result should
749   // be optional.
750   using Type = Optional<RangeSet>;
751 };
752 
753 template <class OptionalOrPointer, class... TailTy>
754 struct IntersectionTraits<OptionalOrPointer, TailTy...> {
755   // If current type is Optional or a raw pointer, we should keep looking.
756   using Type = typename IntersectionTraits<TailTy...>::Type;
757 };
758 
759 template <class EndTy>
760 LLVM_NODISCARD inline EndTy intersect(RangeSet::Factory &F, EndTy End) {
761   // If the list contains only RangeSet or Optional<RangeSet>, simply return
762   // that range set.
763   return End;
764 }
765 
766 LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED inline Optional<RangeSet>
767 intersect(RangeSet::Factory &F, const RangeSet *End) {
768   // This is an extraneous conversion from a raw pointer into Optional<RangeSet>
769   if (End) {
770     return *End;
771   }
772   return llvm::None;
773 }
774 
775 template <class... RestTy>
776 LLVM_NODISCARD inline RangeSet intersect(RangeSet::Factory &F, RangeSet Head,
777                                          RangeSet Second, RestTy... Tail) {
778   // Here we call either the <RangeSet,RangeSet,...> or <RangeSet,...> version
779   // of the function and can be sure that the result is RangeSet.
780   return intersect(F, F.intersect(Head, Second), Tail...);
781 }
782 
783 template <class SecondTy, class... RestTy>
784 LLVM_NODISCARD inline RangeSet intersect(RangeSet::Factory &F, RangeSet Head,
785                                          SecondTy Second, RestTy... Tail) {
786   if (Second) {
787     // Here we call the <RangeSet,RangeSet,...> version of the function...
788     return intersect(F, Head, *Second, Tail...);
789   }
790   // ...and here it is either <RangeSet,RangeSet,...> or <RangeSet,...>, which
791   // means that the result is definitely RangeSet.
792   return intersect(F, Head, Tail...);
793 }
794 
795 /// Main generic intersect function.
796 /// It intersects all of the given range sets.  If some of the given arguments
797 /// don't hold a range set (nullptr or llvm::None), the function will skip them.
798 ///
799 /// Available representations for the arguments are:
800 ///   * RangeSet
801 ///   * Optional<RangeSet>
802 ///   * RangeSet *
803 /// Pointer to a RangeSet is automatically assumed to be nullable and will get
804 /// checked as well as the optional version.  If this behaviour is undesired,
805 /// please dereference the pointer in the call.
806 ///
807 /// Return type depends on the arguments' types.  If we can be sure in compile
808 /// time that there will be a range set as a result, the returning type is
809 /// simply RangeSet, in other cases we have to back off to Optional<RangeSet>.
810 ///
811 /// Please, prefer optional range sets to raw pointers.  If the last argument is
812 /// a raw pointer and all previous arguments are None, it will cost one
813 /// additional check to convert RangeSet * into Optional<RangeSet>.
814 template <class HeadTy, class SecondTy, class... RestTy>
815 LLVM_NODISCARD inline
816     typename IntersectionTraits<HeadTy, SecondTy, RestTy...>::Type
817     intersect(RangeSet::Factory &F, HeadTy Head, SecondTy Second,
818               RestTy... Tail) {
819   if (Head) {
820     return intersect(F, *Head, Second, Tail...);
821   }
822   return intersect(F, Second, Tail...);
823 }
824 
825 //===----------------------------------------------------------------------===//
826 //                           Symbolic reasoning logic
827 //===----------------------------------------------------------------------===//
828 
829 /// A little component aggregating all of the reasoning we have about
830 /// the ranges of symbolic expressions.
831 ///
832 /// Even when we don't know the exact values of the operands, we still
833 /// can get a pretty good estimate of the result's range.
834 class SymbolicRangeInferrer
835     : public SymExprVisitor<SymbolicRangeInferrer, RangeSet> {
836 public:
837   template <class SourceType>
838   static RangeSet inferRange(RangeSet::Factory &F, ProgramStateRef State,
839                              SourceType Origin) {
840     SymbolicRangeInferrer Inferrer(F, State);
841     return Inferrer.infer(Origin);
842   }
843 
844   RangeSet VisitSymExpr(SymbolRef Sym) {
845     // If we got to this function, the actual type of the symbolic
846     // expression is not supported for advanced inference.
847     // In this case, we simply backoff to the default "let's simply
848     // infer the range from the expression's type".
849     return infer(Sym->getType());
850   }
851 
852   RangeSet VisitSymIntExpr(const SymIntExpr *Sym) {
853     return VisitBinaryOperator(Sym);
854   }
855 
856   RangeSet VisitIntSymExpr(const IntSymExpr *Sym) {
857     return VisitBinaryOperator(Sym);
858   }
859 
860   RangeSet VisitSymSymExpr(const SymSymExpr *Sym) {
861     return intersect(
862         RangeFactory,
863         // If Sym is (dis)equality, we might have some information
864         // on that in our equality classes data structure.
865         getRangeForEqualities(Sym),
866         // And we should always check what we can get from the operands.
867         VisitBinaryOperator(Sym));
868   }
869 
870 private:
871   SymbolicRangeInferrer(RangeSet::Factory &F, ProgramStateRef S)
872       : ValueFactory(F.getValueFactory()), RangeFactory(F), State(S) {}
873 
874   /// Infer range information from the given integer constant.
875   ///
876   /// It's not a real "inference", but is here for operating with
877   /// sub-expressions in a more polymorphic manner.
878   RangeSet inferAs(const llvm::APSInt &Val, QualType) {
879     return {RangeFactory, Val};
880   }
881 
882   /// Infer range information from symbol in the context of the given type.
883   RangeSet inferAs(SymbolRef Sym, QualType DestType) {
884     QualType ActualType = Sym->getType();
885     // Check that we can reason about the symbol at all.
886     if (ActualType->isIntegralOrEnumerationType() ||
887         Loc::isLocType(ActualType)) {
888       return infer(Sym);
889     }
890     // Otherwise, let's simply infer from the destination type.
891     // We couldn't figure out nothing else about that expression.
892     return infer(DestType);
893   }
894 
895   RangeSet infer(SymbolRef Sym) {
896     return intersect(
897         RangeFactory,
898         // Of course, we should take the constraint directly associated with
899         // this symbol into consideration.
900         getConstraint(State, Sym),
901         // If Sym is a difference of symbols A - B, then maybe we have range
902         // set stored for B - A.
903         //
904         // If we have range set stored for both A - B and B - A then
905         // calculate the effective range set by intersecting the range set
906         // for A - B and the negated range set of B - A.
907         getRangeForNegatedSub(Sym),
908         // If Sym is a comparison expression (except <=>),
909         // find any other comparisons with the same operands.
910         // See function description.
911         getRangeForComparisonSymbol(Sym),
912         // Apart from the Sym itself, we can infer quite a lot if we look
913         // into subexpressions of Sym.
914         Visit(Sym));
915   }
916 
917   RangeSet infer(EquivalenceClass Class) {
918     if (const RangeSet *AssociatedConstraint = getConstraint(State, Class))
919       return *AssociatedConstraint;
920 
921     return infer(Class.getType());
922   }
923 
924   /// Infer range information solely from the type.
925   RangeSet infer(QualType T) {
926     // Lazily generate a new RangeSet representing all possible values for the
927     // given symbol type.
928     RangeSet Result(RangeFactory, ValueFactory.getMinValue(T),
929                     ValueFactory.getMaxValue(T));
930 
931     // References are known to be non-zero.
932     if (T->isReferenceType())
933       return assumeNonZero(Result, T);
934 
935     return Result;
936   }
937 
938   template <class BinarySymExprTy>
939   RangeSet VisitBinaryOperator(const BinarySymExprTy *Sym) {
940     // TODO #1: VisitBinaryOperator implementation might not make a good
941     // use of the inferred ranges.  In this case, we might be calculating
942     // everything for nothing.  This being said, we should introduce some
943     // sort of laziness mechanism here.
944     //
945     // TODO #2: We didn't go into the nested expressions before, so it
946     // might cause us spending much more time doing the inference.
947     // This can be a problem for deeply nested expressions that are
948     // involved in conditions and get tested continuously.  We definitely
949     // need to address this issue and introduce some sort of caching
950     // in here.
951     QualType ResultType = Sym->getType();
952     return VisitBinaryOperator(inferAs(Sym->getLHS(), ResultType),
953                                Sym->getOpcode(),
954                                inferAs(Sym->getRHS(), ResultType), ResultType);
955   }
956 
957   RangeSet VisitBinaryOperator(RangeSet LHS, BinaryOperator::Opcode Op,
958                                RangeSet RHS, QualType T) {
959     switch (Op) {
960     case BO_Or:
961       return VisitBinaryOperator<BO_Or>(LHS, RHS, T);
962     case BO_And:
963       return VisitBinaryOperator<BO_And>(LHS, RHS, T);
964     case BO_Rem:
965       return VisitBinaryOperator<BO_Rem>(LHS, RHS, T);
966     default:
967       return infer(T);
968     }
969   }
970 
971   //===----------------------------------------------------------------------===//
972   //                         Ranges and operators
973   //===----------------------------------------------------------------------===//
974 
975   /// Return a rough approximation of the given range set.
976   ///
977   /// For the range set:
978   ///   { [x_0, y_0], [x_1, y_1], ... , [x_N, y_N] }
979   /// it will return the range [x_0, y_N].
980   static Range fillGaps(RangeSet Origin) {
981     assert(!Origin.isEmpty());
982     return {Origin.getMinValue(), Origin.getMaxValue()};
983   }
984 
985   /// Try to convert given range into the given type.
986   ///
987   /// It will return llvm::None only when the trivial conversion is possible.
988   llvm::Optional<Range> convert(const Range &Origin, APSIntType To) {
989     if (To.testInRange(Origin.From(), false) != APSIntType::RTR_Within ||
990         To.testInRange(Origin.To(), false) != APSIntType::RTR_Within) {
991       return llvm::None;
992     }
993     return Range(ValueFactory.Convert(To, Origin.From()),
994                  ValueFactory.Convert(To, Origin.To()));
995   }
996 
997   template <BinaryOperator::Opcode Op>
998   RangeSet VisitBinaryOperator(RangeSet LHS, RangeSet RHS, QualType T) {
999     // We should propagate information about unfeasbility of one of the
1000     // operands to the resulting range.
1001     if (LHS.isEmpty() || RHS.isEmpty()) {
1002       return RangeFactory.getEmptySet();
1003     }
1004 
1005     Range CoarseLHS = fillGaps(LHS);
1006     Range CoarseRHS = fillGaps(RHS);
1007 
1008     APSIntType ResultType = ValueFactory.getAPSIntType(T);
1009 
1010     // We need to convert ranges to the resulting type, so we can compare values
1011     // and combine them in a meaningful (in terms of the given operation) way.
1012     auto ConvertedCoarseLHS = convert(CoarseLHS, ResultType);
1013     auto ConvertedCoarseRHS = convert(CoarseRHS, ResultType);
1014 
1015     // It is hard to reason about ranges when conversion changes
1016     // borders of the ranges.
1017     if (!ConvertedCoarseLHS || !ConvertedCoarseRHS) {
1018       return infer(T);
1019     }
1020 
1021     return VisitBinaryOperator<Op>(*ConvertedCoarseLHS, *ConvertedCoarseRHS, T);
1022   }
1023 
1024   template <BinaryOperator::Opcode Op>
1025   RangeSet VisitBinaryOperator(Range LHS, Range RHS, QualType T) {
1026     return infer(T);
1027   }
1028 
1029   /// Return a symmetrical range for the given range and type.
1030   ///
1031   /// If T is signed, return the smallest range [-x..x] that covers the original
1032   /// range, or [-min(T), max(T)] if the aforementioned symmetric range doesn't
1033   /// exist due to original range covering min(T)).
1034   ///
1035   /// If T is unsigned, return the smallest range [0..x] that covers the
1036   /// original range.
1037   Range getSymmetricalRange(Range Origin, QualType T) {
1038     APSIntType RangeType = ValueFactory.getAPSIntType(T);
1039 
1040     if (RangeType.isUnsigned()) {
1041       return Range(ValueFactory.getMinValue(RangeType), Origin.To());
1042     }
1043 
1044     if (Origin.From().isMinSignedValue()) {
1045       // If mini is a minimal signed value, absolute value of it is greater
1046       // than the maximal signed value.  In order to avoid these
1047       // complications, we simply return the whole range.
1048       return {ValueFactory.getMinValue(RangeType),
1049               ValueFactory.getMaxValue(RangeType)};
1050     }
1051 
1052     // At this point, we are sure that the type is signed and we can safely
1053     // use unary - operator.
1054     //
1055     // While calculating absolute maximum, we can use the following formula
1056     // because of these reasons:
1057     //   * If From >= 0 then To >= From and To >= -From.
1058     //     AbsMax == To == max(To, -From)
1059     //   * If To <= 0 then -From >= -To and -From >= From.
1060     //     AbsMax == -From == max(-From, To)
1061     //   * Otherwise, From <= 0, To >= 0, and
1062     //     AbsMax == max(abs(From), abs(To))
1063     llvm::APSInt AbsMax = std::max(-Origin.From(), Origin.To());
1064 
1065     // Intersection is guaranteed to be non-empty.
1066     return {ValueFactory.getValue(-AbsMax), ValueFactory.getValue(AbsMax)};
1067   }
1068 
1069   /// Return a range set subtracting zero from \p Domain.
1070   RangeSet assumeNonZero(RangeSet Domain, QualType T) {
1071     APSIntType IntType = ValueFactory.getAPSIntType(T);
1072     return RangeFactory.deletePoint(Domain, IntType.getZeroValue());
1073   }
1074 
1075   // FIXME: Once SValBuilder supports unary minus, we should use SValBuilder to
1076   //        obtain the negated symbolic expression instead of constructing the
1077   //        symbol manually. This will allow us to support finding ranges of not
1078   //        only negated SymSymExpr-type expressions, but also of other, simpler
1079   //        expressions which we currently do not know how to negate.
1080   Optional<RangeSet> getRangeForNegatedSub(SymbolRef Sym) {
1081     if (const SymSymExpr *SSE = dyn_cast<SymSymExpr>(Sym)) {
1082       if (SSE->getOpcode() == BO_Sub) {
1083         QualType T = Sym->getType();
1084 
1085         // Do not negate unsigned ranges
1086         if (!T->isUnsignedIntegerOrEnumerationType() &&
1087             !T->isSignedIntegerOrEnumerationType())
1088           return llvm::None;
1089 
1090         SymbolManager &SymMgr = State->getSymbolManager();
1091         SymbolRef NegatedSym =
1092             SymMgr.getSymSymExpr(SSE->getRHS(), BO_Sub, SSE->getLHS(), T);
1093 
1094         if (const RangeSet *NegatedRange = getConstraint(State, NegatedSym)) {
1095           return RangeFactory.negate(*NegatedRange);
1096         }
1097       }
1098     }
1099     return llvm::None;
1100   }
1101 
1102   // Returns ranges only for binary comparison operators (except <=>)
1103   // when left and right operands are symbolic values.
1104   // Finds any other comparisons with the same operands.
1105   // Then do logical calculations and refuse impossible branches.
1106   // E.g. (x < y) and (x > y) at the same time are impossible.
1107   // E.g. (x >= y) and (x != y) at the same time makes (x > y) true only.
1108   // E.g. (x == y) and (y == x) are just reversed but the same.
1109   // It covers all possible combinations (see CmpOpTable description).
1110   // Note that `x` and `y` can also stand for subexpressions,
1111   // not only for actual symbols.
1112   Optional<RangeSet> getRangeForComparisonSymbol(SymbolRef Sym) {
1113     const auto *SSE = dyn_cast<SymSymExpr>(Sym);
1114     if (!SSE)
1115       return llvm::None;
1116 
1117     const BinaryOperatorKind CurrentOP = SSE->getOpcode();
1118 
1119     // We currently do not support <=> (C++20).
1120     if (!BinaryOperator::isComparisonOp(CurrentOP) || (CurrentOP == BO_Cmp))
1121       return llvm::None;
1122 
1123     static const OperatorRelationsTable CmpOpTable{};
1124 
1125     const SymExpr *LHS = SSE->getLHS();
1126     const SymExpr *RHS = SSE->getRHS();
1127     QualType T = SSE->getType();
1128 
1129     SymbolManager &SymMgr = State->getSymbolManager();
1130 
1131     // We use this variable to store the last queried operator (`QueriedOP`)
1132     // for which the `getCmpOpState` returned with `Unknown`. If there are two
1133     // different OPs that returned `Unknown` then we have to query the special
1134     // `UnknownX2` column. We assume that `getCmpOpState(CurrentOP, CurrentOP)`
1135     // never returns `Unknown`, so `CurrentOP` is a good initial value.
1136     BinaryOperatorKind LastQueriedOpToUnknown = CurrentOP;
1137 
1138     // Loop goes through all of the columns exept the last one ('UnknownX2').
1139     // We treat `UnknownX2` column separately at the end of the loop body.
1140     for (size_t i = 0; i < CmpOpTable.getCmpOpCount(); ++i) {
1141 
1142       // Let's find an expression e.g. (x < y).
1143       BinaryOperatorKind QueriedOP = OperatorRelationsTable::getOpFromIndex(i);
1144       const SymSymExpr *SymSym = SymMgr.getSymSymExpr(LHS, QueriedOP, RHS, T);
1145       const RangeSet *QueriedRangeSet = getConstraint(State, SymSym);
1146 
1147       // If ranges were not previously found,
1148       // try to find a reversed expression (y > x).
1149       if (!QueriedRangeSet) {
1150         const BinaryOperatorKind ROP =
1151             BinaryOperator::reverseComparisonOp(QueriedOP);
1152         SymSym = SymMgr.getSymSymExpr(RHS, ROP, LHS, T);
1153         QueriedRangeSet = getConstraint(State, SymSym);
1154       }
1155 
1156       if (!QueriedRangeSet || QueriedRangeSet->isEmpty())
1157         continue;
1158 
1159       const llvm::APSInt *ConcreteValue = QueriedRangeSet->getConcreteValue();
1160       const bool isInFalseBranch =
1161           ConcreteValue ? (*ConcreteValue == 0) : false;
1162 
1163       // If it is a false branch, we shall be guided by opposite operator,
1164       // because the table is made assuming we are in the true branch.
1165       // E.g. when (x <= y) is false, then (x > y) is true.
1166       if (isInFalseBranch)
1167         QueriedOP = BinaryOperator::negateComparisonOp(QueriedOP);
1168 
1169       OperatorRelationsTable::TriStateKind BranchState =
1170           CmpOpTable.getCmpOpState(CurrentOP, QueriedOP);
1171 
1172       if (BranchState == OperatorRelationsTable::Unknown) {
1173         if (LastQueriedOpToUnknown != CurrentOP &&
1174             LastQueriedOpToUnknown != QueriedOP) {
1175           // If we got the Unknown state for both different operators.
1176           // if (x <= y)    // assume true
1177           //   if (x != y)  // assume true
1178           //     if (x < y) // would be also true
1179           // Get a state from `UnknownX2` column.
1180           BranchState = CmpOpTable.getCmpOpStateForUnknownX2(CurrentOP);
1181         } else {
1182           LastQueriedOpToUnknown = QueriedOP;
1183           continue;
1184         }
1185       }
1186 
1187       return (BranchState == OperatorRelationsTable::True) ? getTrueRange(T)
1188                                                            : getFalseRange(T);
1189     }
1190 
1191     return llvm::None;
1192   }
1193 
1194   Optional<RangeSet> getRangeForEqualities(const SymSymExpr *Sym) {
1195     Optional<bool> Equality = meansEquality(Sym);
1196 
1197     if (!Equality)
1198       return llvm::None;
1199 
1200     if (Optional<bool> AreEqual =
1201             EquivalenceClass::areEqual(State, Sym->getLHS(), Sym->getRHS())) {
1202       // Here we cover two cases at once:
1203       //   * if Sym is equality and its operands are known to be equal -> true
1204       //   * if Sym is disequality and its operands are disequal -> true
1205       if (*AreEqual == *Equality) {
1206         return getTrueRange(Sym->getType());
1207       }
1208       // Opposite combinations result in false.
1209       return getFalseRange(Sym->getType());
1210     }
1211 
1212     return llvm::None;
1213   }
1214 
1215   RangeSet getTrueRange(QualType T) {
1216     RangeSet TypeRange = infer(T);
1217     return assumeNonZero(TypeRange, T);
1218   }
1219 
1220   RangeSet getFalseRange(QualType T) {
1221     const llvm::APSInt &Zero = ValueFactory.getValue(0, T);
1222     return RangeSet(RangeFactory, Zero);
1223   }
1224 
1225   BasicValueFactory &ValueFactory;
1226   RangeSet::Factory &RangeFactory;
1227   ProgramStateRef State;
1228 };
1229 
1230 //===----------------------------------------------------------------------===//
1231 //               Range-based reasoning about symbolic operations
1232 //===----------------------------------------------------------------------===//
1233 
1234 template <>
1235 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_Or>(Range LHS, Range RHS,
1236                                                            QualType T) {
1237   APSIntType ResultType = ValueFactory.getAPSIntType(T);
1238   llvm::APSInt Zero = ResultType.getZeroValue();
1239 
1240   bool IsLHSPositiveOrZero = LHS.From() >= Zero;
1241   bool IsRHSPositiveOrZero = RHS.From() >= Zero;
1242 
1243   bool IsLHSNegative = LHS.To() < Zero;
1244   bool IsRHSNegative = RHS.To() < Zero;
1245 
1246   // Check if both ranges have the same sign.
1247   if ((IsLHSPositiveOrZero && IsRHSPositiveOrZero) ||
1248       (IsLHSNegative && IsRHSNegative)) {
1249     // The result is definitely greater or equal than any of the operands.
1250     const llvm::APSInt &Min = std::max(LHS.From(), RHS.From());
1251 
1252     // We estimate maximal value for positives as the maximal value for the
1253     // given type.  For negatives, we estimate it with -1 (e.g. 0x11111111).
1254     //
1255     // TODO: We basically, limit the resulting range from below, but don't do
1256     //       anything with the upper bound.
1257     //
1258     //       For positive operands, it can be done as follows: for the upper
1259     //       bound of LHS and RHS we calculate the most significant bit set.
1260     //       Let's call it the N-th bit.  Then we can estimate the maximal
1261     //       number to be 2^(N+1)-1, i.e. the number with all the bits up to
1262     //       the N-th bit set.
1263     const llvm::APSInt &Max = IsLHSNegative
1264                                   ? ValueFactory.getValue(--Zero)
1265                                   : ValueFactory.getMaxValue(ResultType);
1266 
1267     return {RangeFactory, ValueFactory.getValue(Min), Max};
1268   }
1269 
1270   // Otherwise, let's check if at least one of the operands is negative.
1271   if (IsLHSNegative || IsRHSNegative) {
1272     // This means that the result is definitely negative as well.
1273     return {RangeFactory, ValueFactory.getMinValue(ResultType),
1274             ValueFactory.getValue(--Zero)};
1275   }
1276 
1277   RangeSet DefaultRange = infer(T);
1278 
1279   // It is pretty hard to reason about operands with different signs
1280   // (and especially with possibly different signs).  We simply check if it
1281   // can be zero.  In order to conclude that the result could not be zero,
1282   // at least one of the operands should be definitely not zero itself.
1283   if (!LHS.Includes(Zero) || !RHS.Includes(Zero)) {
1284     return assumeNonZero(DefaultRange, T);
1285   }
1286 
1287   // Nothing much else to do here.
1288   return DefaultRange;
1289 }
1290 
1291 template <>
1292 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_And>(Range LHS,
1293                                                             Range RHS,
1294                                                             QualType T) {
1295   APSIntType ResultType = ValueFactory.getAPSIntType(T);
1296   llvm::APSInt Zero = ResultType.getZeroValue();
1297 
1298   bool IsLHSPositiveOrZero = LHS.From() >= Zero;
1299   bool IsRHSPositiveOrZero = RHS.From() >= Zero;
1300 
1301   bool IsLHSNegative = LHS.To() < Zero;
1302   bool IsRHSNegative = RHS.To() < Zero;
1303 
1304   // Check if both ranges have the same sign.
1305   if ((IsLHSPositiveOrZero && IsRHSPositiveOrZero) ||
1306       (IsLHSNegative && IsRHSNegative)) {
1307     // The result is definitely less or equal than any of the operands.
1308     const llvm::APSInt &Max = std::min(LHS.To(), RHS.To());
1309 
1310     // We conservatively estimate lower bound to be the smallest positive
1311     // or negative value corresponding to the sign of the operands.
1312     const llvm::APSInt &Min = IsLHSNegative
1313                                   ? ValueFactory.getMinValue(ResultType)
1314                                   : ValueFactory.getValue(Zero);
1315 
1316     return {RangeFactory, Min, Max};
1317   }
1318 
1319   // Otherwise, let's check if at least one of the operands is positive.
1320   if (IsLHSPositiveOrZero || IsRHSPositiveOrZero) {
1321     // This makes result definitely positive.
1322     //
1323     // We can also reason about a maximal value by finding the maximal
1324     // value of the positive operand.
1325     const llvm::APSInt &Max = IsLHSPositiveOrZero ? LHS.To() : RHS.To();
1326 
1327     // The minimal value on the other hand is much harder to reason about.
1328     // The only thing we know for sure is that the result is positive.
1329     return {RangeFactory, ValueFactory.getValue(Zero),
1330             ValueFactory.getValue(Max)};
1331   }
1332 
1333   // Nothing much else to do here.
1334   return infer(T);
1335 }
1336 
1337 template <>
1338 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_Rem>(Range LHS,
1339                                                             Range RHS,
1340                                                             QualType T) {
1341   llvm::APSInt Zero = ValueFactory.getAPSIntType(T).getZeroValue();
1342 
1343   Range ConservativeRange = getSymmetricalRange(RHS, T);
1344 
1345   llvm::APSInt Max = ConservativeRange.To();
1346   llvm::APSInt Min = ConservativeRange.From();
1347 
1348   if (Max == Zero) {
1349     // It's an undefined behaviour to divide by 0 and it seems like we know
1350     // for sure that RHS is 0.  Let's say that the resulting range is
1351     // simply infeasible for that matter.
1352     return RangeFactory.getEmptySet();
1353   }
1354 
1355   // At this point, our conservative range is closed.  The result, however,
1356   // couldn't be greater than the RHS' maximal absolute value.  Because of
1357   // this reason, we turn the range into open (or half-open in case of
1358   // unsigned integers).
1359   //
1360   // While we operate on integer values, an open interval (a, b) can be easily
1361   // represented by the closed interval [a + 1, b - 1].  And this is exactly
1362   // what we do next.
1363   //
1364   // If we are dealing with unsigned case, we shouldn't move the lower bound.
1365   if (Min.isSigned()) {
1366     ++Min;
1367   }
1368   --Max;
1369 
1370   bool IsLHSPositiveOrZero = LHS.From() >= Zero;
1371   bool IsRHSPositiveOrZero = RHS.From() >= Zero;
1372 
1373   // Remainder operator results with negative operands is implementation
1374   // defined.  Positive cases are much easier to reason about though.
1375   if (IsLHSPositiveOrZero && IsRHSPositiveOrZero) {
1376     // If maximal value of LHS is less than maximal value of RHS,
1377     // the result won't get greater than LHS.To().
1378     Max = std::min(LHS.To(), Max);
1379     // We want to check if it is a situation similar to the following:
1380     //
1381     // <------------|---[  LHS  ]--------[  RHS  ]----->
1382     //  -INF        0                              +INF
1383     //
1384     // In this situation, we can conclude that (LHS / RHS) == 0 and
1385     // (LHS % RHS) == LHS.
1386     Min = LHS.To() < RHS.From() ? LHS.From() : Zero;
1387   }
1388 
1389   // Nevertheless, the symmetrical range for RHS is a conservative estimate
1390   // for any sign of either LHS, or RHS.
1391   return {RangeFactory, ValueFactory.getValue(Min), ValueFactory.getValue(Max)};
1392 }
1393 
1394 //===----------------------------------------------------------------------===//
1395 //                         Constraint assignment logic
1396 //===----------------------------------------------------------------------===//
1397 
1398 /// ConstraintAssignorBase is a small utility class that unifies visitor
1399 /// for ranges with a visitor for constraints (rangeset/range/constant).
1400 ///
1401 /// It is designed to have one derived class, but generally it can have more.
1402 /// Derived class can control which types we handle by defining methods of the
1403 /// following form:
1404 ///
1405 ///   bool handle${SYMBOL}To${CONSTRAINT}(const SYMBOL *Sym,
1406 ///                                       CONSTRAINT Constraint);
1407 ///
1408 /// where SYMBOL is the type of the symbol (e.g. SymSymExpr, SymbolCast, etc.)
1409 ///       CONSTRAINT is the type of constraint (RangeSet/Range/Const)
1410 ///       return value signifies whether we should try other handle methods
1411 ///          (i.e. false would mean to stop right after calling this method)
1412 template <class Derived> class ConstraintAssignorBase {
1413 public:
1414   using Const = const llvm::APSInt &;
1415 
1416 #define DISPATCH(CLASS) return assign##CLASS##Impl(cast<CLASS>(Sym), Constraint)
1417 
1418 #define ASSIGN(CLASS, TO, SYM, CONSTRAINT)                                     \
1419   if (!static_cast<Derived *>(this)->assign##CLASS##To##TO(SYM, CONSTRAINT))   \
1420   return false
1421 
1422   void assign(SymbolRef Sym, RangeSet Constraint) {
1423     assignImpl(Sym, Constraint);
1424   }
1425 
1426   bool assignImpl(SymbolRef Sym, RangeSet Constraint) {
1427     switch (Sym->getKind()) {
1428 #define SYMBOL(Id, Parent)                                                     \
1429   case SymExpr::Id##Kind:                                                      \
1430     DISPATCH(Id);
1431 #include "clang/StaticAnalyzer/Core/PathSensitive/Symbols.def"
1432     }
1433     llvm_unreachable("Unknown SymExpr kind!");
1434   }
1435 
1436 #define DEFAULT_ASSIGN(Id)                                                     \
1437   bool assign##Id##To##RangeSet(const Id *Sym, RangeSet Constraint) {          \
1438     return true;                                                               \
1439   }                                                                            \
1440   bool assign##Id##To##Range(const Id *Sym, Range Constraint) { return true; } \
1441   bool assign##Id##To##Const(const Id *Sym, Const Constraint) { return true; }
1442 
1443   // When we dispatch for constraint types, we first try to check
1444   // if the new constraint is the constant and try the corresponding
1445   // assignor methods.  If it didn't interrupt, we can proceed to the
1446   // range, and finally to the range set.
1447 #define CONSTRAINT_DISPATCH(Id)                                                \
1448   if (const llvm::APSInt *Const = Constraint.getConcreteValue()) {             \
1449     ASSIGN(Id, Const, Sym, *Const);                                            \
1450   }                                                                            \
1451   if (Constraint.size() == 1) {                                                \
1452     ASSIGN(Id, Range, Sym, *Constraint.begin());                               \
1453   }                                                                            \
1454   ASSIGN(Id, RangeSet, Sym, Constraint)
1455 
1456   // Our internal assign method first tries to call assignor methods for all
1457   // constraint types that apply.  And if not interrupted, continues with its
1458   // parent class.
1459 #define SYMBOL(Id, Parent)                                                     \
1460   bool assign##Id##Impl(const Id *Sym, RangeSet Constraint) {                  \
1461     CONSTRAINT_DISPATCH(Id);                                                   \
1462     DISPATCH(Parent);                                                          \
1463   }                                                                            \
1464   DEFAULT_ASSIGN(Id)
1465 #define ABSTRACT_SYMBOL(Id, Parent) SYMBOL(Id, Parent)
1466 #include "clang/StaticAnalyzer/Core/PathSensitive/Symbols.def"
1467 
1468   // Default implementations for the top class that doesn't have parents.
1469   bool assignSymExprImpl(const SymExpr *Sym, RangeSet Constraint) {
1470     CONSTRAINT_DISPATCH(SymExpr);
1471     return true;
1472   }
1473   DEFAULT_ASSIGN(SymExpr);
1474 
1475 #undef DISPATCH
1476 #undef CONSTRAINT_DISPATCH
1477 #undef DEFAULT_ASSIGN
1478 #undef ASSIGN
1479 };
1480 
1481 /// A little component aggregating all of the reasoning we have about
1482 /// assigning new constraints to symbols.
1483 ///
1484 /// The main purpose of this class is to associate constraints to symbols,
1485 /// and impose additional constraints on other symbols, when we can imply
1486 /// them.
1487 ///
1488 /// It has a nice symmetry with SymbolicRangeInferrer.  When the latter
1489 /// can provide more precise ranges by looking into the operands of the
1490 /// expression in question, ConstraintAssignor looks into the operands
1491 /// to see if we can imply more from the new constraint.
1492 class ConstraintAssignor : public ConstraintAssignorBase<ConstraintAssignor> {
1493 public:
1494   template <class ClassOrSymbol>
1495   LLVM_NODISCARD static ProgramStateRef
1496   assign(ProgramStateRef State, SValBuilder &Builder, RangeSet::Factory &F,
1497          ClassOrSymbol CoS, RangeSet NewConstraint) {
1498     if (!State || NewConstraint.isEmpty())
1499       return nullptr;
1500 
1501     ConstraintAssignor Assignor{State, Builder, F};
1502     return Assignor.assign(CoS, NewConstraint);
1503   }
1504 
1505   inline bool assignSymExprToConst(const SymExpr *Sym, Const Constraint);
1506   inline bool assignSymSymExprToRangeSet(const SymSymExpr *Sym,
1507                                          RangeSet Constraint);
1508 
1509 private:
1510   ConstraintAssignor(ProgramStateRef State, SValBuilder &Builder,
1511                      RangeSet::Factory &F)
1512       : State(State), Builder(Builder), RangeFactory(F) {}
1513   using Base = ConstraintAssignorBase<ConstraintAssignor>;
1514 
1515   /// Base method for handling new constraints for symbols.
1516   LLVM_NODISCARD ProgramStateRef assign(SymbolRef Sym, RangeSet NewConstraint) {
1517     // All constraints are actually associated with equivalence classes, and
1518     // that's what we are going to do first.
1519     State = assign(EquivalenceClass::find(State, Sym), NewConstraint);
1520     if (!State)
1521       return nullptr;
1522 
1523     // And after that we can check what other things we can get from this
1524     // constraint.
1525     Base::assign(Sym, NewConstraint);
1526     return State;
1527   }
1528 
1529   /// Base method for handling new constraints for classes.
1530   LLVM_NODISCARD ProgramStateRef assign(EquivalenceClass Class,
1531                                         RangeSet NewConstraint) {
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 = NewConstraint.getConcreteValue()) {
1538 
1539       ConstraintRangeTy Constraints = State->get<ConstraintRange>();
1540       ConstraintRangeTy::Factory &CF = State->get_context<ConstraintRange>();
1541 
1542       // Add new constraint.
1543       Constraints = CF.add(Constraints, Class, NewConstraint);
1544 
1545       for (EquivalenceClass DisequalClass : Class.getDisequalClasses(State)) {
1546         RangeSet UpdatedConstraint = SymbolicRangeInferrer::inferRange(
1547             RangeFactory, State, DisequalClass);
1548 
1549         UpdatedConstraint = RangeFactory.deletePoint(UpdatedConstraint, *Point);
1550 
1551         // If we end up with at least one of the disequal classes to be
1552         // constrained with an empty range-set, the state is infeasible.
1553         if (UpdatedConstraint.isEmpty())
1554           return nullptr;
1555 
1556         Constraints = CF.add(Constraints, DisequalClass, UpdatedConstraint);
1557       }
1558       assert(areFeasible(Constraints) && "Constraint manager shouldn't produce "
1559                                          "a state with infeasible constraints");
1560 
1561       return setConstraints(State, Constraints);
1562     }
1563 
1564     return setConstraint(State, Class, NewConstraint);
1565   }
1566 
1567   ProgramStateRef trackDisequality(ProgramStateRef State, SymbolRef LHS,
1568                                    SymbolRef RHS) {
1569     return EquivalenceClass::markDisequal(RangeFactory, State, LHS, RHS);
1570   }
1571 
1572   ProgramStateRef trackEquality(ProgramStateRef State, SymbolRef LHS,
1573                                 SymbolRef RHS) {
1574     return EquivalenceClass::merge(RangeFactory, State, LHS, RHS);
1575   }
1576 
1577   LLVM_NODISCARD Optional<bool> interpreteAsBool(RangeSet Constraint) {
1578     assert(!Constraint.isEmpty() && "Empty ranges shouldn't get here");
1579 
1580     if (Constraint.getConcreteValue())
1581       return !Constraint.getConcreteValue()->isZero();
1582 
1583     APSIntType T{Constraint.getMinValue()};
1584     Const Zero = T.getZeroValue();
1585     if (!Constraint.contains(Zero))
1586       return true;
1587 
1588     return llvm::None;
1589   }
1590 
1591   ProgramStateRef State;
1592   SValBuilder &Builder;
1593   RangeSet::Factory &RangeFactory;
1594 };
1595 
1596 //===----------------------------------------------------------------------===//
1597 //                  Constraint manager implementation details
1598 //===----------------------------------------------------------------------===//
1599 
1600 class RangeConstraintManager : public RangedConstraintManager {
1601 public:
1602   RangeConstraintManager(ExprEngine *EE, SValBuilder &SVB)
1603       : RangedConstraintManager(EE, SVB), F(getBasicVals()) {}
1604 
1605   //===------------------------------------------------------------------===//
1606   // Implementation for interface from ConstraintManager.
1607   //===------------------------------------------------------------------===//
1608 
1609   bool haveEqualConstraints(ProgramStateRef S1,
1610                             ProgramStateRef S2) const override {
1611     // NOTE: ClassMembers are as simple as back pointers for ClassMap,
1612     //       so comparing constraint ranges and class maps should be
1613     //       sufficient.
1614     return S1->get<ConstraintRange>() == S2->get<ConstraintRange>() &&
1615            S1->get<ClassMap>() == S2->get<ClassMap>();
1616   }
1617 
1618   bool canReasonAbout(SVal X) const override;
1619 
1620   ConditionTruthVal checkNull(ProgramStateRef State, SymbolRef Sym) override;
1621 
1622   const llvm::APSInt *getSymVal(ProgramStateRef State,
1623                                 SymbolRef Sym) const override;
1624 
1625   ProgramStateRef removeDeadBindings(ProgramStateRef State,
1626                                      SymbolReaper &SymReaper) override;
1627 
1628   void printJson(raw_ostream &Out, ProgramStateRef State, const char *NL = "\n",
1629                  unsigned int Space = 0, bool IsDot = false) const override;
1630   void printConstraints(raw_ostream &Out, ProgramStateRef State,
1631                         const char *NL = "\n", unsigned int Space = 0,
1632                         bool IsDot = false) const;
1633   void printEquivalenceClasses(raw_ostream &Out, ProgramStateRef State,
1634                                const char *NL = "\n", unsigned int Space = 0,
1635                                bool IsDot = false) const;
1636   void printDisequalities(raw_ostream &Out, ProgramStateRef State,
1637                           const char *NL = "\n", unsigned int Space = 0,
1638                           bool IsDot = false) const;
1639 
1640   //===------------------------------------------------------------------===//
1641   // Implementation for interface from RangedConstraintManager.
1642   //===------------------------------------------------------------------===//
1643 
1644   ProgramStateRef assumeSymNE(ProgramStateRef State, SymbolRef Sym,
1645                               const llvm::APSInt &V,
1646                               const llvm::APSInt &Adjustment) override;
1647 
1648   ProgramStateRef assumeSymEQ(ProgramStateRef State, SymbolRef Sym,
1649                               const llvm::APSInt &V,
1650                               const llvm::APSInt &Adjustment) override;
1651 
1652   ProgramStateRef assumeSymLT(ProgramStateRef State, SymbolRef Sym,
1653                               const llvm::APSInt &V,
1654                               const llvm::APSInt &Adjustment) override;
1655 
1656   ProgramStateRef assumeSymGT(ProgramStateRef State, SymbolRef Sym,
1657                               const llvm::APSInt &V,
1658                               const llvm::APSInt &Adjustment) override;
1659 
1660   ProgramStateRef assumeSymLE(ProgramStateRef State, SymbolRef Sym,
1661                               const llvm::APSInt &V,
1662                               const llvm::APSInt &Adjustment) override;
1663 
1664   ProgramStateRef assumeSymGE(ProgramStateRef State, SymbolRef Sym,
1665                               const llvm::APSInt &V,
1666                               const llvm::APSInt &Adjustment) override;
1667 
1668   ProgramStateRef assumeSymWithinInclusiveRange(
1669       ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From,
1670       const llvm::APSInt &To, const llvm::APSInt &Adjustment) override;
1671 
1672   ProgramStateRef assumeSymOutsideInclusiveRange(
1673       ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From,
1674       const llvm::APSInt &To, const llvm::APSInt &Adjustment) override;
1675 
1676 private:
1677   RangeSet::Factory F;
1678 
1679   RangeSet getRange(ProgramStateRef State, SymbolRef Sym);
1680   RangeSet getRange(ProgramStateRef State, EquivalenceClass Class);
1681   ProgramStateRef setRange(ProgramStateRef State, SymbolRef Sym,
1682                            RangeSet Range);
1683   ProgramStateRef setRange(ProgramStateRef State, EquivalenceClass Class,
1684                            RangeSet Range);
1685 
1686   RangeSet getSymLTRange(ProgramStateRef St, SymbolRef Sym,
1687                          const llvm::APSInt &Int,
1688                          const llvm::APSInt &Adjustment);
1689   RangeSet getSymGTRange(ProgramStateRef St, SymbolRef Sym,
1690                          const llvm::APSInt &Int,
1691                          const llvm::APSInt &Adjustment);
1692   RangeSet getSymLERange(ProgramStateRef St, SymbolRef Sym,
1693                          const llvm::APSInt &Int,
1694                          const llvm::APSInt &Adjustment);
1695   RangeSet getSymLERange(llvm::function_ref<RangeSet()> RS,
1696                          const llvm::APSInt &Int,
1697                          const llvm::APSInt &Adjustment);
1698   RangeSet getSymGERange(ProgramStateRef St, SymbolRef Sym,
1699                          const llvm::APSInt &Int,
1700                          const llvm::APSInt &Adjustment);
1701 };
1702 
1703 bool ConstraintAssignor::assignSymExprToConst(const SymExpr *Sym,
1704                                               const llvm::APSInt &Constraint) {
1705   llvm::SmallSet<EquivalenceClass, 4> SimplifiedClasses;
1706   // Iterate over all equivalence classes and try to simplify them.
1707   ClassMembersTy Members = State->get<ClassMembers>();
1708   for (std::pair<EquivalenceClass, SymbolSet> ClassToSymbolSet : Members) {
1709     EquivalenceClass Class = ClassToSymbolSet.first;
1710     State = EquivalenceClass::simplify(Builder, RangeFactory, State, Class);
1711     if (!State)
1712       return false;
1713     SimplifiedClasses.insert(Class);
1714   }
1715 
1716   // Trivial equivalence classes (those that have only one symbol member) are
1717   // not stored in the State. Thus, we must skim through the constraints as
1718   // well. And we try to simplify symbols in the constraints.
1719   ConstraintRangeTy Constraints = State->get<ConstraintRange>();
1720   for (std::pair<EquivalenceClass, RangeSet> ClassConstraint : Constraints) {
1721     EquivalenceClass Class = ClassConstraint.first;
1722     if (SimplifiedClasses.count(Class)) // Already simplified.
1723       continue;
1724     State = EquivalenceClass::simplify(Builder, RangeFactory, State, Class);
1725     if (!State)
1726       return false;
1727   }
1728 
1729   return true;
1730 }
1731 
1732 bool ConstraintAssignor::assignSymSymExprToRangeSet(const SymSymExpr *Sym,
1733                                                     RangeSet Constraint) {
1734   Optional<bool> ConstraintAsBool = interpreteAsBool(Constraint);
1735 
1736   if (!ConstraintAsBool)
1737     return true;
1738 
1739   if (Optional<bool> Equality = meansEquality(Sym)) {
1740     // Here we cover two cases:
1741     //   * if Sym is equality and the new constraint is true -> Sym's operands
1742     //     should be marked as equal
1743     //   * if Sym is disequality and the new constraint is false -> Sym's
1744     //     operands should be also marked as equal
1745     if (*Equality == *ConstraintAsBool) {
1746       State = trackEquality(State, Sym->getLHS(), Sym->getRHS());
1747     } else {
1748       // Other combinations leave as with disequal operands.
1749       State = trackDisequality(State, Sym->getLHS(), Sym->getRHS());
1750     }
1751 
1752     if (!State)
1753       return false;
1754   }
1755 
1756   return true;
1757 }
1758 
1759 } // end anonymous namespace
1760 
1761 std::unique_ptr<ConstraintManager>
1762 ento::CreateRangeConstraintManager(ProgramStateManager &StMgr,
1763                                    ExprEngine *Eng) {
1764   return std::make_unique<RangeConstraintManager>(Eng, StMgr.getSValBuilder());
1765 }
1766 
1767 ConstraintMap ento::getConstraintMap(ProgramStateRef State) {
1768   ConstraintMap::Factory &F = State->get_context<ConstraintMap>();
1769   ConstraintMap Result = F.getEmptyMap();
1770 
1771   ConstraintRangeTy Constraints = State->get<ConstraintRange>();
1772   for (std::pair<EquivalenceClass, RangeSet> ClassConstraint : Constraints) {
1773     EquivalenceClass Class = ClassConstraint.first;
1774     SymbolSet ClassMembers = Class.getClassMembers(State);
1775     assert(!ClassMembers.isEmpty() &&
1776            "Class must always have at least one member!");
1777 
1778     SymbolRef Representative = *ClassMembers.begin();
1779     Result = F.add(Result, Representative, ClassConstraint.second);
1780   }
1781 
1782   return Result;
1783 }
1784 
1785 //===----------------------------------------------------------------------===//
1786 //                     EqualityClass implementation details
1787 //===----------------------------------------------------------------------===//
1788 
1789 LLVM_DUMP_METHOD void EquivalenceClass::dumpToStream(ProgramStateRef State,
1790                                                      raw_ostream &os) const {
1791   SymbolSet ClassMembers = getClassMembers(State);
1792   for (const SymbolRef &MemberSym : ClassMembers) {
1793     MemberSym->dump();
1794     os << "\n";
1795   }
1796 }
1797 
1798 inline EquivalenceClass EquivalenceClass::find(ProgramStateRef State,
1799                                                SymbolRef Sym) {
1800   assert(State && "State should not be null");
1801   assert(Sym && "Symbol should not be null");
1802   // We store far from all Symbol -> Class mappings
1803   if (const EquivalenceClass *NontrivialClass = State->get<ClassMap>(Sym))
1804     return *NontrivialClass;
1805 
1806   // This is a trivial class of Sym.
1807   return Sym;
1808 }
1809 
1810 inline ProgramStateRef EquivalenceClass::merge(RangeSet::Factory &F,
1811                                                ProgramStateRef State,
1812                                                SymbolRef First,
1813                                                SymbolRef Second) {
1814   EquivalenceClass FirstClass = find(State, First);
1815   EquivalenceClass SecondClass = find(State, Second);
1816 
1817   return FirstClass.merge(F, State, SecondClass);
1818 }
1819 
1820 inline ProgramStateRef EquivalenceClass::merge(RangeSet::Factory &F,
1821                                                ProgramStateRef State,
1822                                                EquivalenceClass Other) {
1823   // It is already the same class.
1824   if (*this == Other)
1825     return State;
1826 
1827   // FIXME: As of now, we support only equivalence classes of the same type.
1828   //        This limitation is connected to the lack of explicit casts in
1829   //        our symbolic expression model.
1830   //
1831   //        That means that for `int x` and `char y` we don't distinguish
1832   //        between these two very different cases:
1833   //          * `x == y`
1834   //          * `(char)x == y`
1835   //
1836   //        The moment we introduce symbolic casts, this restriction can be
1837   //        lifted.
1838   if (getType() != Other.getType())
1839     return State;
1840 
1841   SymbolSet Members = getClassMembers(State);
1842   SymbolSet OtherMembers = Other.getClassMembers(State);
1843 
1844   // We estimate the size of the class by the height of tree containing
1845   // its members.  Merging is not a trivial operation, so it's easier to
1846   // merge the smaller class into the bigger one.
1847   if (Members.getHeight() >= OtherMembers.getHeight()) {
1848     return mergeImpl(F, State, Members, Other, OtherMembers);
1849   } else {
1850     return Other.mergeImpl(F, State, OtherMembers, *this, Members);
1851   }
1852 }
1853 
1854 inline ProgramStateRef
1855 EquivalenceClass::mergeImpl(RangeSet::Factory &RangeFactory,
1856                             ProgramStateRef State, SymbolSet MyMembers,
1857                             EquivalenceClass Other, SymbolSet OtherMembers) {
1858   // Essentially what we try to recreate here is some kind of union-find
1859   // data structure.  It does have certain limitations due to persistence
1860   // and the need to remove elements from classes.
1861   //
1862   // In this setting, EquialityClass object is the representative of the class
1863   // or the parent element.  ClassMap is a mapping of class members to their
1864   // parent. Unlike the union-find structure, they all point directly to the
1865   // class representative because we don't have an opportunity to actually do
1866   // path compression when dealing with immutability.  This means that we
1867   // compress paths every time we do merges.  It also means that we lose
1868   // the main amortized complexity benefit from the original data structure.
1869   ConstraintRangeTy Constraints = State->get<ConstraintRange>();
1870   ConstraintRangeTy::Factory &CRF = State->get_context<ConstraintRange>();
1871 
1872   // 1. If the merged classes have any constraints associated with them, we
1873   //    need to transfer them to the class we have left.
1874   //
1875   // Intersection here makes perfect sense because both of these constraints
1876   // must hold for the whole new class.
1877   if (Optional<RangeSet> NewClassConstraint =
1878           intersect(RangeFactory, getConstraint(State, *this),
1879                     getConstraint(State, Other))) {
1880     // NOTE: Essentially, NewClassConstraint should NEVER be infeasible because
1881     //       range inferrer shouldn't generate ranges incompatible with
1882     //       equivalence classes. However, at the moment, due to imperfections
1883     //       in the solver, it is possible and the merge function can also
1884     //       return infeasible states aka null states.
1885     if (NewClassConstraint->isEmpty())
1886       // Infeasible state
1887       return nullptr;
1888 
1889     // No need in tracking constraints of a now-dissolved class.
1890     Constraints = CRF.remove(Constraints, Other);
1891     // Assign new constraints for this class.
1892     Constraints = CRF.add(Constraints, *this, *NewClassConstraint);
1893 
1894     assert(areFeasible(Constraints) && "Constraint manager shouldn't produce "
1895                                        "a state with infeasible constraints");
1896 
1897     State = State->set<ConstraintRange>(Constraints);
1898   }
1899 
1900   // 2. Get ALL equivalence-related maps
1901   ClassMapTy Classes = State->get<ClassMap>();
1902   ClassMapTy::Factory &CMF = State->get_context<ClassMap>();
1903 
1904   ClassMembersTy Members = State->get<ClassMembers>();
1905   ClassMembersTy::Factory &MF = State->get_context<ClassMembers>();
1906 
1907   DisequalityMapTy DisequalityInfo = State->get<DisequalityMap>();
1908   DisequalityMapTy::Factory &DF = State->get_context<DisequalityMap>();
1909 
1910   ClassSet::Factory &CF = State->get_context<ClassSet>();
1911   SymbolSet::Factory &F = getMembersFactory(State);
1912 
1913   // 2. Merge members of the Other class into the current class.
1914   SymbolSet NewClassMembers = MyMembers;
1915   for (SymbolRef Sym : OtherMembers) {
1916     NewClassMembers = F.add(NewClassMembers, Sym);
1917     // *this is now the class for all these new symbols.
1918     Classes = CMF.add(Classes, Sym, *this);
1919   }
1920 
1921   // 3. Adjust member mapping.
1922   //
1923   // No need in tracking members of a now-dissolved class.
1924   Members = MF.remove(Members, Other);
1925   // Now only the current class is mapped to all the symbols.
1926   Members = MF.add(Members, *this, NewClassMembers);
1927 
1928   // 4. Update disequality relations
1929   ClassSet DisequalToOther = Other.getDisequalClasses(DisequalityInfo, CF);
1930   // We are about to merge two classes but they are already known to be
1931   // non-equal. This is a contradiction.
1932   if (DisequalToOther.contains(*this))
1933     return nullptr;
1934 
1935   if (!DisequalToOther.isEmpty()) {
1936     ClassSet DisequalToThis = getDisequalClasses(DisequalityInfo, CF);
1937     DisequalityInfo = DF.remove(DisequalityInfo, Other);
1938 
1939     for (EquivalenceClass DisequalClass : DisequalToOther) {
1940       DisequalToThis = CF.add(DisequalToThis, DisequalClass);
1941 
1942       // Disequality is a symmetric relation meaning that if
1943       // DisequalToOther not null then the set for DisequalClass is not
1944       // empty and has at least Other.
1945       ClassSet OriginalSetLinkedToOther =
1946           *DisequalityInfo.lookup(DisequalClass);
1947 
1948       // Other will be eliminated and we should replace it with the bigger
1949       // united class.
1950       ClassSet NewSet = CF.remove(OriginalSetLinkedToOther, Other);
1951       NewSet = CF.add(NewSet, *this);
1952 
1953       DisequalityInfo = DF.add(DisequalityInfo, DisequalClass, NewSet);
1954     }
1955 
1956     DisequalityInfo = DF.add(DisequalityInfo, *this, DisequalToThis);
1957     State = State->set<DisequalityMap>(DisequalityInfo);
1958   }
1959 
1960   // 5. Update the state
1961   State = State->set<ClassMap>(Classes);
1962   State = State->set<ClassMembers>(Members);
1963 
1964   return State;
1965 }
1966 
1967 inline SymbolSet::Factory &
1968 EquivalenceClass::getMembersFactory(ProgramStateRef State) {
1969   return State->get_context<SymbolSet>();
1970 }
1971 
1972 SymbolSet EquivalenceClass::getClassMembers(ProgramStateRef State) const {
1973   if (const SymbolSet *Members = State->get<ClassMembers>(*this))
1974     return *Members;
1975 
1976   // This class is trivial, so we need to construct a set
1977   // with just that one symbol from the class.
1978   SymbolSet::Factory &F = getMembersFactory(State);
1979   return F.add(F.getEmptySet(), getRepresentativeSymbol());
1980 }
1981 
1982 bool EquivalenceClass::isTrivial(ProgramStateRef State) const {
1983   return State->get<ClassMembers>(*this) == nullptr;
1984 }
1985 
1986 bool EquivalenceClass::isTriviallyDead(ProgramStateRef State,
1987                                        SymbolReaper &Reaper) const {
1988   return isTrivial(State) && Reaper.isDead(getRepresentativeSymbol());
1989 }
1990 
1991 inline ProgramStateRef EquivalenceClass::markDisequal(RangeSet::Factory &RF,
1992                                                       ProgramStateRef State,
1993                                                       SymbolRef First,
1994                                                       SymbolRef Second) {
1995   return markDisequal(RF, State, find(State, First), find(State, Second));
1996 }
1997 
1998 inline ProgramStateRef EquivalenceClass::markDisequal(RangeSet::Factory &RF,
1999                                                       ProgramStateRef State,
2000                                                       EquivalenceClass First,
2001                                                       EquivalenceClass Second) {
2002   return First.markDisequal(RF, State, Second);
2003 }
2004 
2005 inline ProgramStateRef
2006 EquivalenceClass::markDisequal(RangeSet::Factory &RF, ProgramStateRef State,
2007                                EquivalenceClass Other) const {
2008   // If we know that two classes are equal, we can only produce an infeasible
2009   // state.
2010   if (*this == Other) {
2011     return nullptr;
2012   }
2013 
2014   DisequalityMapTy DisequalityInfo = State->get<DisequalityMap>();
2015   ConstraintRangeTy Constraints = State->get<ConstraintRange>();
2016 
2017   // Disequality is a symmetric relation, so if we mark A as disequal to B,
2018   // we should also mark B as disequalt to A.
2019   if (!addToDisequalityInfo(DisequalityInfo, Constraints, RF, State, *this,
2020                             Other) ||
2021       !addToDisequalityInfo(DisequalityInfo, Constraints, RF, State, Other,
2022                             *this))
2023     return nullptr;
2024 
2025   assert(areFeasible(Constraints) && "Constraint manager shouldn't produce "
2026                                      "a state with infeasible constraints");
2027 
2028   State = State->set<DisequalityMap>(DisequalityInfo);
2029   State = State->set<ConstraintRange>(Constraints);
2030 
2031   return State;
2032 }
2033 
2034 inline bool EquivalenceClass::addToDisequalityInfo(
2035     DisequalityMapTy &Info, ConstraintRangeTy &Constraints,
2036     RangeSet::Factory &RF, ProgramStateRef State, EquivalenceClass First,
2037     EquivalenceClass Second) {
2038 
2039   // 1. Get all of the required factories.
2040   DisequalityMapTy::Factory &F = State->get_context<DisequalityMap>();
2041   ClassSet::Factory &CF = State->get_context<ClassSet>();
2042   ConstraintRangeTy::Factory &CRF = State->get_context<ConstraintRange>();
2043 
2044   // 2. Add Second to the set of classes disequal to First.
2045   const ClassSet *CurrentSet = Info.lookup(First);
2046   ClassSet NewSet = CurrentSet ? *CurrentSet : CF.getEmptySet();
2047   NewSet = CF.add(NewSet, Second);
2048 
2049   Info = F.add(Info, First, NewSet);
2050 
2051   // 3. If Second is known to be a constant, we can delete this point
2052   //    from the constraint asociated with First.
2053   //
2054   //    So, if Second == 10, it means that First != 10.
2055   //    At the same time, the same logic does not apply to ranges.
2056   if (const RangeSet *SecondConstraint = Constraints.lookup(Second))
2057     if (const llvm::APSInt *Point = SecondConstraint->getConcreteValue()) {
2058 
2059       RangeSet FirstConstraint = SymbolicRangeInferrer::inferRange(
2060           RF, State, First.getRepresentativeSymbol());
2061 
2062       FirstConstraint = RF.deletePoint(FirstConstraint, *Point);
2063 
2064       // If the First class is about to be constrained with an empty
2065       // range-set, the state is infeasible.
2066       if (FirstConstraint.isEmpty())
2067         return false;
2068 
2069       Constraints = CRF.add(Constraints, First, FirstConstraint);
2070     }
2071 
2072   return true;
2073 }
2074 
2075 inline Optional<bool> EquivalenceClass::areEqual(ProgramStateRef State,
2076                                                  SymbolRef FirstSym,
2077                                                  SymbolRef SecondSym) {
2078   return EquivalenceClass::areEqual(State, find(State, FirstSym),
2079                                     find(State, SecondSym));
2080 }
2081 
2082 inline Optional<bool> EquivalenceClass::areEqual(ProgramStateRef State,
2083                                                  EquivalenceClass First,
2084                                                  EquivalenceClass Second) {
2085   // The same equivalence class => symbols are equal.
2086   if (First == Second)
2087     return true;
2088 
2089   // Let's check if we know anything about these two classes being not equal to
2090   // each other.
2091   ClassSet DisequalToFirst = First.getDisequalClasses(State);
2092   if (DisequalToFirst.contains(Second))
2093     return false;
2094 
2095   // It is not clear.
2096   return llvm::None;
2097 }
2098 
2099 // Iterate over all symbols and try to simplify them. Once a symbol is
2100 // simplified then we check if we can merge the simplified symbol's equivalence
2101 // class to this class. This way, we simplify not just the symbols but the
2102 // classes as well: we strive to keep the number of the classes to be the
2103 // absolute minimum.
2104 LLVM_NODISCARD ProgramStateRef
2105 EquivalenceClass::simplify(SValBuilder &SVB, RangeSet::Factory &F,
2106                            ProgramStateRef State, EquivalenceClass Class) {
2107   SymbolSet ClassMembers = Class.getClassMembers(State);
2108   for (const SymbolRef &MemberSym : ClassMembers) {
2109     SymbolRef SimplifiedMemberSym = ento::simplify(State, MemberSym);
2110     if (SimplifiedMemberSym && MemberSym != SimplifiedMemberSym) {
2111       // The simplified symbol should be the member of the original Class,
2112       // however, it might be in another existing class at the moment. We
2113       // have to merge these classes.
2114       State = merge(F, State, MemberSym, SimplifiedMemberSym);
2115       if (!State)
2116         return nullptr;
2117     }
2118   }
2119   return State;
2120 }
2121 
2122 inline ClassSet EquivalenceClass::getDisequalClasses(ProgramStateRef State,
2123                                                      SymbolRef Sym) {
2124   return find(State, Sym).getDisequalClasses(State);
2125 }
2126 
2127 inline ClassSet
2128 EquivalenceClass::getDisequalClasses(ProgramStateRef State) const {
2129   return getDisequalClasses(State->get<DisequalityMap>(),
2130                             State->get_context<ClassSet>());
2131 }
2132 
2133 inline ClassSet
2134 EquivalenceClass::getDisequalClasses(DisequalityMapTy Map,
2135                                      ClassSet::Factory &Factory) const {
2136   if (const ClassSet *DisequalClasses = Map.lookup(*this))
2137     return *DisequalClasses;
2138 
2139   return Factory.getEmptySet();
2140 }
2141 
2142 bool EquivalenceClass::isClassDataConsistent(ProgramStateRef State) {
2143   ClassMembersTy Members = State->get<ClassMembers>();
2144 
2145   for (std::pair<EquivalenceClass, SymbolSet> ClassMembersPair : Members) {
2146     for (SymbolRef Member : ClassMembersPair.second) {
2147       // Every member of the class should have a mapping back to the class.
2148       if (find(State, Member) == ClassMembersPair.first) {
2149         continue;
2150       }
2151 
2152       return false;
2153     }
2154   }
2155 
2156   DisequalityMapTy Disequalities = State->get<DisequalityMap>();
2157   for (std::pair<EquivalenceClass, ClassSet> DisequalityInfo : Disequalities) {
2158     EquivalenceClass Class = DisequalityInfo.first;
2159     ClassSet DisequalClasses = DisequalityInfo.second;
2160 
2161     // There is no use in keeping empty sets in the map.
2162     if (DisequalClasses.isEmpty())
2163       return false;
2164 
2165     // Disequality is symmetrical, i.e. for every Class A and B that A != B,
2166     // B != A should also be true.
2167     for (EquivalenceClass DisequalClass : DisequalClasses) {
2168       const ClassSet *DisequalToDisequalClasses =
2169           Disequalities.lookup(DisequalClass);
2170 
2171       // It should be a set of at least one element: Class
2172       if (!DisequalToDisequalClasses ||
2173           !DisequalToDisequalClasses->contains(Class))
2174         return false;
2175     }
2176   }
2177 
2178   return true;
2179 }
2180 
2181 //===----------------------------------------------------------------------===//
2182 //                    RangeConstraintManager implementation
2183 //===----------------------------------------------------------------------===//
2184 
2185 bool RangeConstraintManager::canReasonAbout(SVal X) const {
2186   Optional<nonloc::SymbolVal> SymVal = X.getAs<nonloc::SymbolVal>();
2187   if (SymVal && SymVal->isExpression()) {
2188     const SymExpr *SE = SymVal->getSymbol();
2189 
2190     if (const SymIntExpr *SIE = dyn_cast<SymIntExpr>(SE)) {
2191       switch (SIE->getOpcode()) {
2192       // We don't reason yet about bitwise-constraints on symbolic values.
2193       case BO_And:
2194       case BO_Or:
2195       case BO_Xor:
2196         return false;
2197       // We don't reason yet about these arithmetic constraints on
2198       // symbolic values.
2199       case BO_Mul:
2200       case BO_Div:
2201       case BO_Rem:
2202       case BO_Shl:
2203       case BO_Shr:
2204         return false;
2205       // All other cases.
2206       default:
2207         return true;
2208       }
2209     }
2210 
2211     if (const SymSymExpr *SSE = dyn_cast<SymSymExpr>(SE)) {
2212       // FIXME: Handle <=> here.
2213       if (BinaryOperator::isEqualityOp(SSE->getOpcode()) ||
2214           BinaryOperator::isRelationalOp(SSE->getOpcode())) {
2215         // We handle Loc <> Loc comparisons, but not (yet) NonLoc <> NonLoc.
2216         // We've recently started producing Loc <> NonLoc comparisons (that
2217         // result from casts of one of the operands between eg. intptr_t and
2218         // void *), but we can't reason about them yet.
2219         if (Loc::isLocType(SSE->getLHS()->getType())) {
2220           return Loc::isLocType(SSE->getRHS()->getType());
2221         }
2222       }
2223     }
2224 
2225     return false;
2226   }
2227 
2228   return true;
2229 }
2230 
2231 ConditionTruthVal RangeConstraintManager::checkNull(ProgramStateRef State,
2232                                                     SymbolRef Sym) {
2233   const RangeSet *Ranges = getConstraint(State, Sym);
2234 
2235   // If we don't have any information about this symbol, it's underconstrained.
2236   if (!Ranges)
2237     return ConditionTruthVal();
2238 
2239   // If we have a concrete value, see if it's zero.
2240   if (const llvm::APSInt *Value = Ranges->getConcreteValue())
2241     return *Value == 0;
2242 
2243   BasicValueFactory &BV = getBasicVals();
2244   APSIntType IntType = BV.getAPSIntType(Sym->getType());
2245   llvm::APSInt Zero = IntType.getZeroValue();
2246 
2247   // Check if zero is in the set of possible values.
2248   if (!Ranges->contains(Zero))
2249     return false;
2250 
2251   // Zero is a possible value, but it is not the /only/ possible value.
2252   return ConditionTruthVal();
2253 }
2254 
2255 const llvm::APSInt *RangeConstraintManager::getSymVal(ProgramStateRef St,
2256                                                       SymbolRef Sym) const {
2257   const RangeSet *T = getConstraint(St, Sym);
2258   return T ? T->getConcreteValue() : nullptr;
2259 }
2260 
2261 //===----------------------------------------------------------------------===//
2262 //                Remove dead symbols from existing constraints
2263 //===----------------------------------------------------------------------===//
2264 
2265 /// Scan all symbols referenced by the constraints. If the symbol is not alive
2266 /// as marked in LSymbols, mark it as dead in DSymbols.
2267 ProgramStateRef
2268 RangeConstraintManager::removeDeadBindings(ProgramStateRef State,
2269                                            SymbolReaper &SymReaper) {
2270   ClassMembersTy ClassMembersMap = State->get<ClassMembers>();
2271   ClassMembersTy NewClassMembersMap = ClassMembersMap;
2272   ClassMembersTy::Factory &EMFactory = State->get_context<ClassMembers>();
2273   SymbolSet::Factory &SetFactory = State->get_context<SymbolSet>();
2274 
2275   ConstraintRangeTy Constraints = State->get<ConstraintRange>();
2276   ConstraintRangeTy NewConstraints = Constraints;
2277   ConstraintRangeTy::Factory &ConstraintFactory =
2278       State->get_context<ConstraintRange>();
2279 
2280   ClassMapTy Map = State->get<ClassMap>();
2281   ClassMapTy NewMap = Map;
2282   ClassMapTy::Factory &ClassFactory = State->get_context<ClassMap>();
2283 
2284   DisequalityMapTy Disequalities = State->get<DisequalityMap>();
2285   DisequalityMapTy::Factory &DisequalityFactory =
2286       State->get_context<DisequalityMap>();
2287   ClassSet::Factory &ClassSetFactory = State->get_context<ClassSet>();
2288 
2289   bool ClassMapChanged = false;
2290   bool MembersMapChanged = false;
2291   bool ConstraintMapChanged = false;
2292   bool DisequalitiesChanged = false;
2293 
2294   auto removeDeadClass = [&](EquivalenceClass Class) {
2295     // Remove associated constraint ranges.
2296     Constraints = ConstraintFactory.remove(Constraints, Class);
2297     ConstraintMapChanged = true;
2298 
2299     // Update disequality information to not hold any information on the
2300     // removed class.
2301     ClassSet DisequalClasses =
2302         Class.getDisequalClasses(Disequalities, ClassSetFactory);
2303     if (!DisequalClasses.isEmpty()) {
2304       for (EquivalenceClass DisequalClass : DisequalClasses) {
2305         ClassSet DisequalToDisequalSet =
2306             DisequalClass.getDisequalClasses(Disequalities, ClassSetFactory);
2307         // DisequalToDisequalSet is guaranteed to be non-empty for consistent
2308         // disequality info.
2309         assert(!DisequalToDisequalSet.isEmpty());
2310         ClassSet NewSet = ClassSetFactory.remove(DisequalToDisequalSet, Class);
2311 
2312         // No need in keeping an empty set.
2313         if (NewSet.isEmpty()) {
2314           Disequalities =
2315               DisequalityFactory.remove(Disequalities, DisequalClass);
2316         } else {
2317           Disequalities =
2318               DisequalityFactory.add(Disequalities, DisequalClass, NewSet);
2319         }
2320       }
2321       // Remove the data for the class
2322       Disequalities = DisequalityFactory.remove(Disequalities, Class);
2323       DisequalitiesChanged = true;
2324     }
2325   };
2326 
2327   // 1. Let's see if dead symbols are trivial and have associated constraints.
2328   for (std::pair<EquivalenceClass, RangeSet> ClassConstraintPair :
2329        Constraints) {
2330     EquivalenceClass Class = ClassConstraintPair.first;
2331     if (Class.isTriviallyDead(State, SymReaper)) {
2332       // If this class is trivial, we can remove its constraints right away.
2333       removeDeadClass(Class);
2334     }
2335   }
2336 
2337   // 2. We don't need to track classes for dead symbols.
2338   for (std::pair<SymbolRef, EquivalenceClass> SymbolClassPair : Map) {
2339     SymbolRef Sym = SymbolClassPair.first;
2340 
2341     if (SymReaper.isDead(Sym)) {
2342       ClassMapChanged = true;
2343       NewMap = ClassFactory.remove(NewMap, Sym);
2344     }
2345   }
2346 
2347   // 3. Remove dead members from classes and remove dead non-trivial classes
2348   //    and their constraints.
2349   for (std::pair<EquivalenceClass, SymbolSet> ClassMembersPair :
2350        ClassMembersMap) {
2351     EquivalenceClass Class = ClassMembersPair.first;
2352     SymbolSet LiveMembers = ClassMembersPair.second;
2353     bool MembersChanged = false;
2354 
2355     for (SymbolRef Member : ClassMembersPair.second) {
2356       if (SymReaper.isDead(Member)) {
2357         MembersChanged = true;
2358         LiveMembers = SetFactory.remove(LiveMembers, Member);
2359       }
2360     }
2361 
2362     // Check if the class changed.
2363     if (!MembersChanged)
2364       continue;
2365 
2366     MembersMapChanged = true;
2367 
2368     if (LiveMembers.isEmpty()) {
2369       // The class is dead now, we need to wipe it out of the members map...
2370       NewClassMembersMap = EMFactory.remove(NewClassMembersMap, Class);
2371 
2372       // ...and remove all of its constraints.
2373       removeDeadClass(Class);
2374     } else {
2375       // We need to change the members associated with the class.
2376       NewClassMembersMap =
2377           EMFactory.add(NewClassMembersMap, Class, LiveMembers);
2378     }
2379   }
2380 
2381   // 4. Update the state with new maps.
2382   //
2383   // Here we try to be humble and update a map only if it really changed.
2384   if (ClassMapChanged)
2385     State = State->set<ClassMap>(NewMap);
2386 
2387   if (MembersMapChanged)
2388     State = State->set<ClassMembers>(NewClassMembersMap);
2389 
2390   if (ConstraintMapChanged)
2391     State = State->set<ConstraintRange>(Constraints);
2392 
2393   if (DisequalitiesChanged)
2394     State = State->set<DisequalityMap>(Disequalities);
2395 
2396   assert(EquivalenceClass::isClassDataConsistent(State));
2397 
2398   return State;
2399 }
2400 
2401 RangeSet RangeConstraintManager::getRange(ProgramStateRef State,
2402                                           SymbolRef Sym) {
2403   return SymbolicRangeInferrer::inferRange(F, State, Sym);
2404 }
2405 
2406 ProgramStateRef RangeConstraintManager::setRange(ProgramStateRef State,
2407                                                  SymbolRef Sym,
2408                                                  RangeSet Range) {
2409   return ConstraintAssignor::assign(State, getSValBuilder(), F, Sym, Range);
2410 }
2411 
2412 //===------------------------------------------------------------------------===
2413 // assumeSymX methods: protected interface for RangeConstraintManager.
2414 //===------------------------------------------------------------------------===/
2415 
2416 // The syntax for ranges below is mathematical, using [x, y] for closed ranges
2417 // and (x, y) for open ranges. These ranges are modular, corresponding with
2418 // a common treatment of C integer overflow. This means that these methods
2419 // do not have to worry about overflow; RangeSet::Intersect can handle such a
2420 // "wraparound" range.
2421 // As an example, the range [UINT_MAX-1, 3) contains five values: UINT_MAX-1,
2422 // UINT_MAX, 0, 1, and 2.
2423 
2424 ProgramStateRef
2425 RangeConstraintManager::assumeSymNE(ProgramStateRef St, SymbolRef Sym,
2426                                     const llvm::APSInt &Int,
2427                                     const llvm::APSInt &Adjustment) {
2428   // Before we do any real work, see if the value can even show up.
2429   APSIntType AdjustmentType(Adjustment);
2430   if (AdjustmentType.testInRange(Int, true) != APSIntType::RTR_Within)
2431     return St;
2432 
2433   llvm::APSInt Point = AdjustmentType.convert(Int) - Adjustment;
2434   RangeSet New = getRange(St, Sym);
2435   New = F.deletePoint(New, Point);
2436 
2437   return setRange(St, Sym, New);
2438 }
2439 
2440 ProgramStateRef
2441 RangeConstraintManager::assumeSymEQ(ProgramStateRef St, SymbolRef Sym,
2442                                     const llvm::APSInt &Int,
2443                                     const llvm::APSInt &Adjustment) {
2444   // Before we do any real work, see if the value can even show up.
2445   APSIntType AdjustmentType(Adjustment);
2446   if (AdjustmentType.testInRange(Int, true) != APSIntType::RTR_Within)
2447     return nullptr;
2448 
2449   // [Int-Adjustment, Int-Adjustment]
2450   llvm::APSInt AdjInt = AdjustmentType.convert(Int) - Adjustment;
2451   RangeSet New = getRange(St, Sym);
2452   New = F.intersect(New, AdjInt);
2453 
2454   return setRange(St, Sym, New);
2455 }
2456 
2457 RangeSet RangeConstraintManager::getSymLTRange(ProgramStateRef St,
2458                                                SymbolRef Sym,
2459                                                const llvm::APSInt &Int,
2460                                                const llvm::APSInt &Adjustment) {
2461   // Before we do any real work, see if the value can even show up.
2462   APSIntType AdjustmentType(Adjustment);
2463   switch (AdjustmentType.testInRange(Int, true)) {
2464   case APSIntType::RTR_Below:
2465     return F.getEmptySet();
2466   case APSIntType::RTR_Within:
2467     break;
2468   case APSIntType::RTR_Above:
2469     return getRange(St, Sym);
2470   }
2471 
2472   // Special case for Int == Min. This is always false.
2473   llvm::APSInt ComparisonVal = AdjustmentType.convert(Int);
2474   llvm::APSInt Min = AdjustmentType.getMinValue();
2475   if (ComparisonVal == Min)
2476     return F.getEmptySet();
2477 
2478   llvm::APSInt Lower = Min - Adjustment;
2479   llvm::APSInt Upper = ComparisonVal - Adjustment;
2480   --Upper;
2481 
2482   RangeSet Result = getRange(St, Sym);
2483   return F.intersect(Result, Lower, Upper);
2484 }
2485 
2486 ProgramStateRef
2487 RangeConstraintManager::assumeSymLT(ProgramStateRef St, SymbolRef Sym,
2488                                     const llvm::APSInt &Int,
2489                                     const llvm::APSInt &Adjustment) {
2490   RangeSet New = getSymLTRange(St, Sym, Int, Adjustment);
2491   return setRange(St, Sym, New);
2492 }
2493 
2494 RangeSet RangeConstraintManager::getSymGTRange(ProgramStateRef St,
2495                                                SymbolRef Sym,
2496                                                const llvm::APSInt &Int,
2497                                                const llvm::APSInt &Adjustment) {
2498   // Before we do any real work, see if the value can even show up.
2499   APSIntType AdjustmentType(Adjustment);
2500   switch (AdjustmentType.testInRange(Int, true)) {
2501   case APSIntType::RTR_Below:
2502     return getRange(St, Sym);
2503   case APSIntType::RTR_Within:
2504     break;
2505   case APSIntType::RTR_Above:
2506     return F.getEmptySet();
2507   }
2508 
2509   // Special case for Int == Max. This is always false.
2510   llvm::APSInt ComparisonVal = AdjustmentType.convert(Int);
2511   llvm::APSInt Max = AdjustmentType.getMaxValue();
2512   if (ComparisonVal == Max)
2513     return F.getEmptySet();
2514 
2515   llvm::APSInt Lower = ComparisonVal - Adjustment;
2516   llvm::APSInt Upper = Max - Adjustment;
2517   ++Lower;
2518 
2519   RangeSet SymRange = getRange(St, Sym);
2520   return F.intersect(SymRange, Lower, Upper);
2521 }
2522 
2523 ProgramStateRef
2524 RangeConstraintManager::assumeSymGT(ProgramStateRef St, SymbolRef Sym,
2525                                     const llvm::APSInt &Int,
2526                                     const llvm::APSInt &Adjustment) {
2527   RangeSet New = getSymGTRange(St, Sym, Int, Adjustment);
2528   return setRange(St, Sym, New);
2529 }
2530 
2531 RangeSet RangeConstraintManager::getSymGERange(ProgramStateRef St,
2532                                                SymbolRef Sym,
2533                                                const llvm::APSInt &Int,
2534                                                const llvm::APSInt &Adjustment) {
2535   // Before we do any real work, see if the value can even show up.
2536   APSIntType AdjustmentType(Adjustment);
2537   switch (AdjustmentType.testInRange(Int, true)) {
2538   case APSIntType::RTR_Below:
2539     return getRange(St, Sym);
2540   case APSIntType::RTR_Within:
2541     break;
2542   case APSIntType::RTR_Above:
2543     return F.getEmptySet();
2544   }
2545 
2546   // Special case for Int == Min. This is always feasible.
2547   llvm::APSInt ComparisonVal = AdjustmentType.convert(Int);
2548   llvm::APSInt Min = AdjustmentType.getMinValue();
2549   if (ComparisonVal == Min)
2550     return getRange(St, Sym);
2551 
2552   llvm::APSInt Max = AdjustmentType.getMaxValue();
2553   llvm::APSInt Lower = ComparisonVal - Adjustment;
2554   llvm::APSInt Upper = Max - Adjustment;
2555 
2556   RangeSet SymRange = getRange(St, Sym);
2557   return F.intersect(SymRange, Lower, Upper);
2558 }
2559 
2560 ProgramStateRef
2561 RangeConstraintManager::assumeSymGE(ProgramStateRef St, SymbolRef Sym,
2562                                     const llvm::APSInt &Int,
2563                                     const llvm::APSInt &Adjustment) {
2564   RangeSet New = getSymGERange(St, Sym, Int, Adjustment);
2565   return setRange(St, Sym, New);
2566 }
2567 
2568 RangeSet
2569 RangeConstraintManager::getSymLERange(llvm::function_ref<RangeSet()> RS,
2570                                       const llvm::APSInt &Int,
2571                                       const llvm::APSInt &Adjustment) {
2572   // Before we do any real work, see if the value can even show up.
2573   APSIntType AdjustmentType(Adjustment);
2574   switch (AdjustmentType.testInRange(Int, true)) {
2575   case APSIntType::RTR_Below:
2576     return F.getEmptySet();
2577   case APSIntType::RTR_Within:
2578     break;
2579   case APSIntType::RTR_Above:
2580     return RS();
2581   }
2582 
2583   // Special case for Int == Max. This is always feasible.
2584   llvm::APSInt ComparisonVal = AdjustmentType.convert(Int);
2585   llvm::APSInt Max = AdjustmentType.getMaxValue();
2586   if (ComparisonVal == Max)
2587     return RS();
2588 
2589   llvm::APSInt Min = AdjustmentType.getMinValue();
2590   llvm::APSInt Lower = Min - Adjustment;
2591   llvm::APSInt Upper = ComparisonVal - Adjustment;
2592 
2593   RangeSet Default = RS();
2594   return F.intersect(Default, Lower, Upper);
2595 }
2596 
2597 RangeSet RangeConstraintManager::getSymLERange(ProgramStateRef St,
2598                                                SymbolRef Sym,
2599                                                const llvm::APSInt &Int,
2600                                                const llvm::APSInt &Adjustment) {
2601   return getSymLERange([&] { return getRange(St, Sym); }, Int, Adjustment);
2602 }
2603 
2604 ProgramStateRef
2605 RangeConstraintManager::assumeSymLE(ProgramStateRef St, SymbolRef Sym,
2606                                     const llvm::APSInt &Int,
2607                                     const llvm::APSInt &Adjustment) {
2608   RangeSet New = getSymLERange(St, Sym, Int, Adjustment);
2609   return setRange(St, Sym, New);
2610 }
2611 
2612 ProgramStateRef RangeConstraintManager::assumeSymWithinInclusiveRange(
2613     ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From,
2614     const llvm::APSInt &To, const llvm::APSInt &Adjustment) {
2615   RangeSet New = getSymGERange(State, Sym, From, Adjustment);
2616   if (New.isEmpty())
2617     return nullptr;
2618   RangeSet Out = getSymLERange([&] { return New; }, To, Adjustment);
2619   return setRange(State, Sym, Out);
2620 }
2621 
2622 ProgramStateRef RangeConstraintManager::assumeSymOutsideInclusiveRange(
2623     ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From,
2624     const llvm::APSInt &To, const llvm::APSInt &Adjustment) {
2625   RangeSet RangeLT = getSymLTRange(State, Sym, From, Adjustment);
2626   RangeSet RangeGT = getSymGTRange(State, Sym, To, Adjustment);
2627   RangeSet New(F.add(RangeLT, RangeGT));
2628   return setRange(State, Sym, New);
2629 }
2630 
2631 //===----------------------------------------------------------------------===//
2632 // Pretty-printing.
2633 //===----------------------------------------------------------------------===//
2634 
2635 void RangeConstraintManager::printJson(raw_ostream &Out, ProgramStateRef State,
2636                                        const char *NL, unsigned int Space,
2637                                        bool IsDot) const {
2638   printConstraints(Out, State, NL, Space, IsDot);
2639   printEquivalenceClasses(Out, State, NL, Space, IsDot);
2640   printDisequalities(Out, State, NL, Space, IsDot);
2641 }
2642 
2643 static std::string toString(const SymbolRef &Sym) {
2644   std::string S;
2645   llvm::raw_string_ostream O(S);
2646   Sym->dumpToStream(O);
2647   return O.str();
2648 }
2649 
2650 void RangeConstraintManager::printConstraints(raw_ostream &Out,
2651                                               ProgramStateRef State,
2652                                               const char *NL,
2653                                               unsigned int Space,
2654                                               bool IsDot) const {
2655   ConstraintRangeTy Constraints = State->get<ConstraintRange>();
2656 
2657   Indent(Out, Space, IsDot) << "\"constraints\": ";
2658   if (Constraints.isEmpty()) {
2659     Out << "null," << NL;
2660     return;
2661   }
2662 
2663   std::map<std::string, RangeSet> OrderedConstraints;
2664   for (std::pair<EquivalenceClass, RangeSet> P : Constraints) {
2665     SymbolSet ClassMembers = P.first.getClassMembers(State);
2666     for (const SymbolRef &ClassMember : ClassMembers) {
2667       bool insertion_took_place;
2668       std::tie(std::ignore, insertion_took_place) =
2669           OrderedConstraints.insert({toString(ClassMember), P.second});
2670       assert(insertion_took_place &&
2671              "two symbols should not have the same dump");
2672     }
2673   }
2674 
2675   ++Space;
2676   Out << '[' << NL;
2677   bool First = true;
2678   for (std::pair<std::string, RangeSet> P : OrderedConstraints) {
2679     if (First) {
2680       First = false;
2681     } else {
2682       Out << ',';
2683       Out << NL;
2684     }
2685     Indent(Out, Space, IsDot)
2686         << "{ \"symbol\": \"" << P.first << "\", \"range\": \"";
2687     P.second.dump(Out);
2688     Out << "\" }";
2689   }
2690   Out << NL;
2691 
2692   --Space;
2693   Indent(Out, Space, IsDot) << "]," << NL;
2694 }
2695 
2696 static std::string toString(ProgramStateRef State, EquivalenceClass Class) {
2697   SymbolSet ClassMembers = Class.getClassMembers(State);
2698   llvm::SmallVector<SymbolRef, 8> ClassMembersSorted(ClassMembers.begin(),
2699                                                      ClassMembers.end());
2700   llvm::sort(ClassMembersSorted,
2701              [](const SymbolRef &LHS, const SymbolRef &RHS) {
2702                return toString(LHS) < toString(RHS);
2703              });
2704 
2705   bool FirstMember = true;
2706 
2707   std::string Str;
2708   llvm::raw_string_ostream Out(Str);
2709   Out << "[ ";
2710   for (SymbolRef ClassMember : ClassMembersSorted) {
2711     if (FirstMember)
2712       FirstMember = false;
2713     else
2714       Out << ", ";
2715     Out << "\"" << ClassMember << "\"";
2716   }
2717   Out << " ]";
2718   return Out.str();
2719 }
2720 
2721 void RangeConstraintManager::printEquivalenceClasses(raw_ostream &Out,
2722                                                      ProgramStateRef State,
2723                                                      const char *NL,
2724                                                      unsigned int Space,
2725                                                      bool IsDot) const {
2726   ClassMembersTy Members = State->get<ClassMembers>();
2727 
2728   Indent(Out, Space, IsDot) << "\"equivalence_classes\": ";
2729   if (Members.isEmpty()) {
2730     Out << "null," << NL;
2731     return;
2732   }
2733 
2734   std::set<std::string> MembersStr;
2735   for (std::pair<EquivalenceClass, SymbolSet> ClassToSymbolSet : Members)
2736     MembersStr.insert(toString(State, ClassToSymbolSet.first));
2737 
2738   ++Space;
2739   Out << '[' << NL;
2740   bool FirstClass = true;
2741   for (const std::string &Str : MembersStr) {
2742     if (FirstClass) {
2743       FirstClass = false;
2744     } else {
2745       Out << ',';
2746       Out << NL;
2747     }
2748     Indent(Out, Space, IsDot);
2749     Out << Str;
2750   }
2751   Out << NL;
2752 
2753   --Space;
2754   Indent(Out, Space, IsDot) << "]," << NL;
2755 }
2756 
2757 void RangeConstraintManager::printDisequalities(raw_ostream &Out,
2758                                                 ProgramStateRef State,
2759                                                 const char *NL,
2760                                                 unsigned int Space,
2761                                                 bool IsDot) const {
2762   DisequalityMapTy Disequalities = State->get<DisequalityMap>();
2763 
2764   Indent(Out, Space, IsDot) << "\"disequality_info\": ";
2765   if (Disequalities.isEmpty()) {
2766     Out << "null," << NL;
2767     return;
2768   }
2769 
2770   // Transform the disequality info to an ordered map of
2771   // [string -> (ordered set of strings)]
2772   using EqClassesStrTy = std::set<std::string>;
2773   using DisequalityInfoStrTy = std::map<std::string, EqClassesStrTy>;
2774   DisequalityInfoStrTy DisequalityInfoStr;
2775   for (std::pair<EquivalenceClass, ClassSet> ClassToDisEqSet : Disequalities) {
2776     EquivalenceClass Class = ClassToDisEqSet.first;
2777     ClassSet DisequalClasses = ClassToDisEqSet.second;
2778     EqClassesStrTy MembersStr;
2779     for (EquivalenceClass DisEqClass : DisequalClasses)
2780       MembersStr.insert(toString(State, DisEqClass));
2781     DisequalityInfoStr.insert({toString(State, Class), MembersStr});
2782   }
2783 
2784   ++Space;
2785   Out << '[' << NL;
2786   bool FirstClass = true;
2787   for (std::pair<std::string, EqClassesStrTy> ClassToDisEqSet :
2788        DisequalityInfoStr) {
2789     const std::string &Class = ClassToDisEqSet.first;
2790     if (FirstClass) {
2791       FirstClass = false;
2792     } else {
2793       Out << ',';
2794       Out << NL;
2795     }
2796     Indent(Out, Space, IsDot) << "{" << NL;
2797     unsigned int DisEqSpace = Space + 1;
2798     Indent(Out, DisEqSpace, IsDot) << "\"class\": ";
2799     Out << Class;
2800     const EqClassesStrTy &DisequalClasses = ClassToDisEqSet.second;
2801     if (!DisequalClasses.empty()) {
2802       Out << "," << NL;
2803       Indent(Out, DisEqSpace, IsDot) << "\"disequal_to\": [" << NL;
2804       unsigned int DisEqClassSpace = DisEqSpace + 1;
2805       Indent(Out, DisEqClassSpace, IsDot);
2806       bool FirstDisEqClass = true;
2807       for (const std::string &DisEqClass : DisequalClasses) {
2808         if (FirstDisEqClass) {
2809           FirstDisEqClass = false;
2810         } else {
2811           Out << ',' << NL;
2812           Indent(Out, DisEqClassSpace, IsDot);
2813         }
2814         Out << DisEqClass;
2815       }
2816       Out << "]" << NL;
2817     }
2818     Indent(Out, Space, IsDot) << "}";
2819   }
2820   Out << NL;
2821 
2822   --Space;
2823   Indent(Out, Space, IsDot) << "]," << NL;
2824 }
2825