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 void Range::dump(raw_ostream &OS) const { 498 OS << '[' << toString(From(), 10) << ", " << toString(To(), 10) << ']'; 499 } 500 501 void RangeSet::dump(raw_ostream &OS) const { 502 OS << "{ "; 503 llvm::interleaveComma(*this, OS, [&OS](const Range &R) { R.dump(OS); }); 504 OS << " }"; 505 } 506 507 REGISTER_SET_FACTORY_WITH_PROGRAMSTATE(SymbolSet, SymbolRef) 508 509 namespace { 510 class EquivalenceClass; 511 } // end anonymous namespace 512 513 REGISTER_MAP_WITH_PROGRAMSTATE(ClassMap, SymbolRef, EquivalenceClass) 514 REGISTER_MAP_WITH_PROGRAMSTATE(ClassMembers, EquivalenceClass, SymbolSet) 515 REGISTER_MAP_WITH_PROGRAMSTATE(ConstraintRange, EquivalenceClass, RangeSet) 516 517 REGISTER_SET_FACTORY_WITH_PROGRAMSTATE(ClassSet, EquivalenceClass) 518 REGISTER_MAP_WITH_PROGRAMSTATE(DisequalityMap, EquivalenceClass, ClassSet) 519 520 namespace { 521 /// This class encapsulates a set of symbols equal to each other. 522 /// 523 /// The main idea of the approach requiring such classes is in narrowing 524 /// and sharing constraints between symbols within the class. Also we can 525 /// conclude that there is no practical need in storing constraints for 526 /// every member of the class separately. 527 /// 528 /// Main terminology: 529 /// 530 /// * "Equivalence class" is an object of this class, which can be efficiently 531 /// compared to other classes. It represents the whole class without 532 /// storing the actual in it. The members of the class however can be 533 /// retrieved from the state. 534 /// 535 /// * "Class members" are the symbols corresponding to the class. This means 536 /// that A == B for every member symbols A and B from the class. Members of 537 /// each class are stored in the state. 538 /// 539 /// * "Trivial class" is a class that has and ever had only one same symbol. 540 /// 541 /// * "Merge operation" merges two classes into one. It is the main operation 542 /// to produce non-trivial classes. 543 /// If, at some point, we can assume that two symbols from two distinct 544 /// classes are equal, we can merge these classes. 545 class EquivalenceClass : public llvm::FoldingSetNode { 546 public: 547 /// Find equivalence class for the given symbol in the given state. 548 LLVM_NODISCARD static inline EquivalenceClass find(ProgramStateRef State, 549 SymbolRef Sym); 550 551 /// Merge classes for the given symbols and return a new state. 552 LLVM_NODISCARD static inline ProgramStateRef merge(RangeSet::Factory &F, 553 ProgramStateRef State, 554 SymbolRef First, 555 SymbolRef Second); 556 // Merge this class with the given class and return a new state. 557 LLVM_NODISCARD inline ProgramStateRef 558 merge(RangeSet::Factory &F, ProgramStateRef State, EquivalenceClass Other); 559 560 /// Return a set of class members for the given state. 561 LLVM_NODISCARD inline SymbolSet getClassMembers(ProgramStateRef State) const; 562 563 /// Return true if the current class is trivial in the given state. 564 /// A class is trivial if and only if there is not any member relations stored 565 /// to it in State/ClassMembers. 566 /// An equivalence class with one member might seem as it does not hold any 567 /// meaningful information, i.e. that is a tautology. However, during the 568 /// removal of dead symbols we do not remove classes with one member for 569 /// resource and performance reasons. Consequently, a class with one member is 570 /// not necessarily trivial. It could happen that we have a class with two 571 /// members and then during the removal of dead symbols we remove one of its 572 /// members. In this case, the class is still non-trivial (it still has the 573 /// mappings in ClassMembers), even though it has only one member. 574 LLVM_NODISCARD inline bool isTrivial(ProgramStateRef State) const; 575 576 /// Return true if the current class is trivial and its only member is dead. 577 LLVM_NODISCARD inline bool isTriviallyDead(ProgramStateRef State, 578 SymbolReaper &Reaper) const; 579 580 LLVM_NODISCARD static inline ProgramStateRef 581 markDisequal(RangeSet::Factory &F, ProgramStateRef State, SymbolRef First, 582 SymbolRef Second); 583 LLVM_NODISCARD static inline ProgramStateRef 584 markDisequal(RangeSet::Factory &F, ProgramStateRef State, 585 EquivalenceClass First, EquivalenceClass Second); 586 LLVM_NODISCARD inline ProgramStateRef 587 markDisequal(RangeSet::Factory &F, ProgramStateRef State, 588 EquivalenceClass Other) const; 589 LLVM_NODISCARD static inline ClassSet 590 getDisequalClasses(ProgramStateRef State, SymbolRef Sym); 591 LLVM_NODISCARD inline ClassSet 592 getDisequalClasses(ProgramStateRef State) const; 593 LLVM_NODISCARD inline ClassSet 594 getDisequalClasses(DisequalityMapTy Map, ClassSet::Factory &Factory) const; 595 596 LLVM_NODISCARD static inline Optional<bool> areEqual(ProgramStateRef State, 597 EquivalenceClass First, 598 EquivalenceClass Second); 599 LLVM_NODISCARD static inline Optional<bool> 600 areEqual(ProgramStateRef State, SymbolRef First, SymbolRef Second); 601 602 /// Iterate over all symbols and try to simplify them. 603 LLVM_NODISCARD ProgramStateRef simplify(SValBuilder &SVB, 604 RangeSet::Factory &F, 605 ProgramStateRef State); 606 607 void dumpToStream(ProgramStateRef State, raw_ostream &os) const; 608 LLVM_DUMP_METHOD void dump(ProgramStateRef State) const { 609 dumpToStream(State, llvm::errs()); 610 } 611 612 /// Check equivalence data for consistency. 613 LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED static bool 614 isClassDataConsistent(ProgramStateRef State); 615 616 LLVM_NODISCARD QualType getType() const { 617 return getRepresentativeSymbol()->getType(); 618 } 619 620 EquivalenceClass() = delete; 621 EquivalenceClass(const EquivalenceClass &) = default; 622 EquivalenceClass &operator=(const EquivalenceClass &) = delete; 623 EquivalenceClass(EquivalenceClass &&) = default; 624 EquivalenceClass &operator=(EquivalenceClass &&) = delete; 625 626 bool operator==(const EquivalenceClass &Other) const { 627 return ID == Other.ID; 628 } 629 bool operator<(const EquivalenceClass &Other) const { return ID < Other.ID; } 630 bool operator!=(const EquivalenceClass &Other) const { 631 return !operator==(Other); 632 } 633 634 static void Profile(llvm::FoldingSetNodeID &ID, uintptr_t CID) { 635 ID.AddInteger(CID); 636 } 637 638 void Profile(llvm::FoldingSetNodeID &ID) const { Profile(ID, this->ID); } 639 640 private: 641 /* implicit */ EquivalenceClass(SymbolRef Sym) 642 : ID(reinterpret_cast<uintptr_t>(Sym)) {} 643 644 /// This function is intended to be used ONLY within the class. 645 /// The fact that ID is a pointer to a symbol is an implementation detail 646 /// and should stay that way. 647 /// In the current implementation, we use it to retrieve the only member 648 /// of the trivial class. 649 SymbolRef getRepresentativeSymbol() const { 650 return reinterpret_cast<SymbolRef>(ID); 651 } 652 static inline SymbolSet::Factory &getMembersFactory(ProgramStateRef State); 653 654 inline ProgramStateRef mergeImpl(RangeSet::Factory &F, ProgramStateRef State, 655 SymbolSet Members, EquivalenceClass Other, 656 SymbolSet OtherMembers); 657 static inline bool 658 addToDisequalityInfo(DisequalityMapTy &Info, ConstraintRangeTy &Constraints, 659 RangeSet::Factory &F, ProgramStateRef State, 660 EquivalenceClass First, EquivalenceClass Second); 661 662 /// This is a unique identifier of the class. 663 uintptr_t ID; 664 }; 665 666 //===----------------------------------------------------------------------===// 667 // Constraint functions 668 //===----------------------------------------------------------------------===// 669 670 LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED bool 671 areFeasible(ConstraintRangeTy Constraints) { 672 return llvm::none_of( 673 Constraints, 674 [](const std::pair<EquivalenceClass, RangeSet> &ClassConstraint) { 675 return ClassConstraint.second.isEmpty(); 676 }); 677 } 678 679 LLVM_NODISCARD inline const RangeSet *getConstraint(ProgramStateRef State, 680 EquivalenceClass Class) { 681 return State->get<ConstraintRange>(Class); 682 } 683 684 LLVM_NODISCARD inline const RangeSet *getConstraint(ProgramStateRef State, 685 SymbolRef Sym) { 686 return getConstraint(State, EquivalenceClass::find(State, Sym)); 687 } 688 689 LLVM_NODISCARD ProgramStateRef setConstraint(ProgramStateRef State, 690 EquivalenceClass Class, 691 RangeSet Constraint) { 692 return State->set<ConstraintRange>(Class, Constraint); 693 } 694 695 LLVM_NODISCARD ProgramStateRef setConstraints(ProgramStateRef State, 696 ConstraintRangeTy Constraints) { 697 return State->set<ConstraintRange>(Constraints); 698 } 699 700 //===----------------------------------------------------------------------===// 701 // Equality/diseqiality abstraction 702 //===----------------------------------------------------------------------===// 703 704 /// A small helper function for detecting symbolic (dis)equality. 705 /// 706 /// Equality check can have different forms (like a == b or a - b) and this 707 /// class encapsulates those away if the only thing the user wants to check - 708 /// whether it's equality/diseqiality or not. 709 /// 710 /// \returns true if assuming this Sym to be true means equality of operands 711 /// false if it means disequality of operands 712 /// None otherwise 713 Optional<bool> meansEquality(const SymSymExpr *Sym) { 714 switch (Sym->getOpcode()) { 715 case BO_Sub: 716 // This case is: A - B != 0 -> disequality check. 717 return false; 718 case BO_EQ: 719 // This case is: A == B != 0 -> equality check. 720 return true; 721 case BO_NE: 722 // This case is: A != B != 0 -> diseqiality check. 723 return false; 724 default: 725 return llvm::None; 726 } 727 } 728 729 //===----------------------------------------------------------------------===// 730 // Intersection functions 731 //===----------------------------------------------------------------------===// 732 733 template <class SecondTy, class... RestTy> 734 LLVM_NODISCARD inline RangeSet intersect(RangeSet::Factory &F, RangeSet Head, 735 SecondTy Second, RestTy... Tail); 736 737 template <class... RangeTy> struct IntersectionTraits; 738 739 template <class... TailTy> struct IntersectionTraits<RangeSet, TailTy...> { 740 // Found RangeSet, no need to check any further 741 using Type = RangeSet; 742 }; 743 744 template <> struct IntersectionTraits<> { 745 // We ran out of types, and we didn't find any RangeSet, so the result should 746 // be optional. 747 using Type = Optional<RangeSet>; 748 }; 749 750 template <class OptionalOrPointer, class... TailTy> 751 struct IntersectionTraits<OptionalOrPointer, TailTy...> { 752 // If current type is Optional or a raw pointer, we should keep looking. 753 using Type = typename IntersectionTraits<TailTy...>::Type; 754 }; 755 756 template <class EndTy> 757 LLVM_NODISCARD inline EndTy intersect(RangeSet::Factory &F, EndTy End) { 758 // If the list contains only RangeSet or Optional<RangeSet>, simply return 759 // that range set. 760 return End; 761 } 762 763 LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED inline Optional<RangeSet> 764 intersect(RangeSet::Factory &F, const RangeSet *End) { 765 // This is an extraneous conversion from a raw pointer into Optional<RangeSet> 766 if (End) { 767 return *End; 768 } 769 return llvm::None; 770 } 771 772 template <class... RestTy> 773 LLVM_NODISCARD inline RangeSet intersect(RangeSet::Factory &F, RangeSet Head, 774 RangeSet Second, RestTy... Tail) { 775 // Here we call either the <RangeSet,RangeSet,...> or <RangeSet,...> version 776 // of the function and can be sure that the result is RangeSet. 777 return intersect(F, F.intersect(Head, Second), Tail...); 778 } 779 780 template <class SecondTy, class... RestTy> 781 LLVM_NODISCARD inline RangeSet intersect(RangeSet::Factory &F, RangeSet Head, 782 SecondTy Second, RestTy... Tail) { 783 if (Second) { 784 // Here we call the <RangeSet,RangeSet,...> version of the function... 785 return intersect(F, Head, *Second, Tail...); 786 } 787 // ...and here it is either <RangeSet,RangeSet,...> or <RangeSet,...>, which 788 // means that the result is definitely RangeSet. 789 return intersect(F, Head, Tail...); 790 } 791 792 /// Main generic intersect function. 793 /// It intersects all of the given range sets. If some of the given arguments 794 /// don't hold a range set (nullptr or llvm::None), the function will skip them. 795 /// 796 /// Available representations for the arguments are: 797 /// * RangeSet 798 /// * Optional<RangeSet> 799 /// * RangeSet * 800 /// Pointer to a RangeSet is automatically assumed to be nullable and will get 801 /// checked as well as the optional version. If this behaviour is undesired, 802 /// please dereference the pointer in the call. 803 /// 804 /// Return type depends on the arguments' types. If we can be sure in compile 805 /// time that there will be a range set as a result, the returning type is 806 /// simply RangeSet, in other cases we have to back off to Optional<RangeSet>. 807 /// 808 /// Please, prefer optional range sets to raw pointers. If the last argument is 809 /// a raw pointer and all previous arguments are None, it will cost one 810 /// additional check to convert RangeSet * into Optional<RangeSet>. 811 template <class HeadTy, class SecondTy, class... RestTy> 812 LLVM_NODISCARD inline 813 typename IntersectionTraits<HeadTy, SecondTy, RestTy...>::Type 814 intersect(RangeSet::Factory &F, HeadTy Head, SecondTy Second, 815 RestTy... Tail) { 816 if (Head) { 817 return intersect(F, *Head, Second, Tail...); 818 } 819 return intersect(F, Second, Tail...); 820 } 821 822 //===----------------------------------------------------------------------===// 823 // Symbolic reasoning logic 824 //===----------------------------------------------------------------------===// 825 826 /// A little component aggregating all of the reasoning we have about 827 /// the ranges of symbolic expressions. 828 /// 829 /// Even when we don't know the exact values of the operands, we still 830 /// can get a pretty good estimate of the result's range. 831 class SymbolicRangeInferrer 832 : public SymExprVisitor<SymbolicRangeInferrer, RangeSet> { 833 public: 834 template <class SourceType> 835 static RangeSet inferRange(RangeSet::Factory &F, ProgramStateRef State, 836 SourceType Origin) { 837 SymbolicRangeInferrer Inferrer(F, State); 838 return Inferrer.infer(Origin); 839 } 840 841 RangeSet VisitSymExpr(SymbolRef Sym) { 842 // If we got to this function, the actual type of the symbolic 843 // expression is not supported for advanced inference. 844 // In this case, we simply backoff to the default "let's simply 845 // infer the range from the expression's type". 846 return infer(Sym->getType()); 847 } 848 849 RangeSet VisitSymIntExpr(const SymIntExpr *Sym) { 850 return VisitBinaryOperator(Sym); 851 } 852 853 RangeSet VisitIntSymExpr(const IntSymExpr *Sym) { 854 return VisitBinaryOperator(Sym); 855 } 856 857 RangeSet VisitSymSymExpr(const SymSymExpr *Sym) { 858 return intersect( 859 RangeFactory, 860 // If Sym is (dis)equality, we might have some information 861 // on that in our equality classes data structure. 862 getRangeForEqualities(Sym), 863 // And we should always check what we can get from the operands. 864 VisitBinaryOperator(Sym)); 865 } 866 867 private: 868 SymbolicRangeInferrer(RangeSet::Factory &F, ProgramStateRef S) 869 : ValueFactory(F.getValueFactory()), RangeFactory(F), State(S) {} 870 871 /// Infer range information from the given integer constant. 872 /// 873 /// It's not a real "inference", but is here for operating with 874 /// sub-expressions in a more polymorphic manner. 875 RangeSet inferAs(const llvm::APSInt &Val, QualType) { 876 return {RangeFactory, Val}; 877 } 878 879 /// Infer range information from symbol in the context of the given type. 880 RangeSet inferAs(SymbolRef Sym, QualType DestType) { 881 QualType ActualType = Sym->getType(); 882 // Check that we can reason about the symbol at all. 883 if (ActualType->isIntegralOrEnumerationType() || 884 Loc::isLocType(ActualType)) { 885 return infer(Sym); 886 } 887 // Otherwise, let's simply infer from the destination type. 888 // We couldn't figure out nothing else about that expression. 889 return infer(DestType); 890 } 891 892 RangeSet infer(SymbolRef Sym) { 893 return intersect( 894 RangeFactory, 895 // Of course, we should take the constraint directly associated with 896 // this symbol into consideration. 897 getConstraint(State, Sym), 898 // If Sym is a difference of symbols A - B, then maybe we have range 899 // set stored for B - A. 900 // 901 // If we have range set stored for both A - B and B - A then 902 // calculate the effective range set by intersecting the range set 903 // for A - B and the negated range set of B - A. 904 getRangeForNegatedSub(Sym), 905 // If Sym is a comparison expression (except <=>), 906 // find any other comparisons with the same operands. 907 // See function description. 908 getRangeForComparisonSymbol(Sym), 909 // Apart from the Sym itself, we can infer quite a lot if we look 910 // into subexpressions of Sym. 911 Visit(Sym)); 912 } 913 914 RangeSet infer(EquivalenceClass Class) { 915 if (const RangeSet *AssociatedConstraint = getConstraint(State, Class)) 916 return *AssociatedConstraint; 917 918 return infer(Class.getType()); 919 } 920 921 /// Infer range information solely from the type. 922 RangeSet infer(QualType T) { 923 // Lazily generate a new RangeSet representing all possible values for the 924 // given symbol type. 925 RangeSet Result(RangeFactory, ValueFactory.getMinValue(T), 926 ValueFactory.getMaxValue(T)); 927 928 // References are known to be non-zero. 929 if (T->isReferenceType()) 930 return assumeNonZero(Result, T); 931 932 return Result; 933 } 934 935 template <class BinarySymExprTy> 936 RangeSet VisitBinaryOperator(const BinarySymExprTy *Sym) { 937 // TODO #1: VisitBinaryOperator implementation might not make a good 938 // use of the inferred ranges. In this case, we might be calculating 939 // everything for nothing. This being said, we should introduce some 940 // sort of laziness mechanism here. 941 // 942 // TODO #2: We didn't go into the nested expressions before, so it 943 // might cause us spending much more time doing the inference. 944 // This can be a problem for deeply nested expressions that are 945 // involved in conditions and get tested continuously. We definitely 946 // need to address this issue and introduce some sort of caching 947 // in here. 948 QualType ResultType = Sym->getType(); 949 return VisitBinaryOperator(inferAs(Sym->getLHS(), ResultType), 950 Sym->getOpcode(), 951 inferAs(Sym->getRHS(), ResultType), ResultType); 952 } 953 954 RangeSet VisitBinaryOperator(RangeSet LHS, BinaryOperator::Opcode Op, 955 RangeSet RHS, QualType T) { 956 switch (Op) { 957 case BO_Or: 958 return VisitBinaryOperator<BO_Or>(LHS, RHS, T); 959 case BO_And: 960 return VisitBinaryOperator<BO_And>(LHS, RHS, T); 961 case BO_Rem: 962 return VisitBinaryOperator<BO_Rem>(LHS, RHS, T); 963 default: 964 return infer(T); 965 } 966 } 967 968 //===----------------------------------------------------------------------===// 969 // Ranges and operators 970 //===----------------------------------------------------------------------===// 971 972 /// Return a rough approximation of the given range set. 973 /// 974 /// For the range set: 975 /// { [x_0, y_0], [x_1, y_1], ... , [x_N, y_N] } 976 /// it will return the range [x_0, y_N]. 977 static Range fillGaps(RangeSet Origin) { 978 assert(!Origin.isEmpty()); 979 return {Origin.getMinValue(), Origin.getMaxValue()}; 980 } 981 982 /// Try to convert given range into the given type. 983 /// 984 /// It will return llvm::None only when the trivial conversion is possible. 985 llvm::Optional<Range> convert(const Range &Origin, APSIntType To) { 986 if (To.testInRange(Origin.From(), false) != APSIntType::RTR_Within || 987 To.testInRange(Origin.To(), false) != APSIntType::RTR_Within) { 988 return llvm::None; 989 } 990 return Range(ValueFactory.Convert(To, Origin.From()), 991 ValueFactory.Convert(To, Origin.To())); 992 } 993 994 template <BinaryOperator::Opcode Op> 995 RangeSet VisitBinaryOperator(RangeSet LHS, RangeSet RHS, QualType T) { 996 // We should propagate information about unfeasbility of one of the 997 // operands to the resulting range. 998 if (LHS.isEmpty() || RHS.isEmpty()) { 999 return RangeFactory.getEmptySet(); 1000 } 1001 1002 Range CoarseLHS = fillGaps(LHS); 1003 Range CoarseRHS = fillGaps(RHS); 1004 1005 APSIntType ResultType = ValueFactory.getAPSIntType(T); 1006 1007 // We need to convert ranges to the resulting type, so we can compare values 1008 // and combine them in a meaningful (in terms of the given operation) way. 1009 auto ConvertedCoarseLHS = convert(CoarseLHS, ResultType); 1010 auto ConvertedCoarseRHS = convert(CoarseRHS, ResultType); 1011 1012 // It is hard to reason about ranges when conversion changes 1013 // borders of the ranges. 1014 if (!ConvertedCoarseLHS || !ConvertedCoarseRHS) { 1015 return infer(T); 1016 } 1017 1018 return VisitBinaryOperator<Op>(*ConvertedCoarseLHS, *ConvertedCoarseRHS, T); 1019 } 1020 1021 template <BinaryOperator::Opcode Op> 1022 RangeSet VisitBinaryOperator(Range LHS, Range RHS, QualType T) { 1023 return infer(T); 1024 } 1025 1026 /// Return a symmetrical range for the given range and type. 1027 /// 1028 /// If T is signed, return the smallest range [-x..x] that covers the original 1029 /// range, or [-min(T), max(T)] if the aforementioned symmetric range doesn't 1030 /// exist due to original range covering min(T)). 1031 /// 1032 /// If T is unsigned, return the smallest range [0..x] that covers the 1033 /// original range. 1034 Range getSymmetricalRange(Range Origin, QualType T) { 1035 APSIntType RangeType = ValueFactory.getAPSIntType(T); 1036 1037 if (RangeType.isUnsigned()) { 1038 return Range(ValueFactory.getMinValue(RangeType), Origin.To()); 1039 } 1040 1041 if (Origin.From().isMinSignedValue()) { 1042 // If mini is a minimal signed value, absolute value of it is greater 1043 // than the maximal signed value. In order to avoid these 1044 // complications, we simply return the whole range. 1045 return {ValueFactory.getMinValue(RangeType), 1046 ValueFactory.getMaxValue(RangeType)}; 1047 } 1048 1049 // At this point, we are sure that the type is signed and we can safely 1050 // use unary - operator. 1051 // 1052 // While calculating absolute maximum, we can use the following formula 1053 // because of these reasons: 1054 // * If From >= 0 then To >= From and To >= -From. 1055 // AbsMax == To == max(To, -From) 1056 // * If To <= 0 then -From >= -To and -From >= From. 1057 // AbsMax == -From == max(-From, To) 1058 // * Otherwise, From <= 0, To >= 0, and 1059 // AbsMax == max(abs(From), abs(To)) 1060 llvm::APSInt AbsMax = std::max(-Origin.From(), Origin.To()); 1061 1062 // Intersection is guaranteed to be non-empty. 1063 return {ValueFactory.getValue(-AbsMax), ValueFactory.getValue(AbsMax)}; 1064 } 1065 1066 /// Return a range set subtracting zero from \p Domain. 1067 RangeSet assumeNonZero(RangeSet Domain, QualType T) { 1068 APSIntType IntType = ValueFactory.getAPSIntType(T); 1069 return RangeFactory.deletePoint(Domain, IntType.getZeroValue()); 1070 } 1071 1072 // FIXME: Once SValBuilder supports unary minus, we should use SValBuilder to 1073 // obtain the negated symbolic expression instead of constructing the 1074 // symbol manually. This will allow us to support finding ranges of not 1075 // only negated SymSymExpr-type expressions, but also of other, simpler 1076 // expressions which we currently do not know how to negate. 1077 Optional<RangeSet> getRangeForNegatedSub(SymbolRef Sym) { 1078 if (const SymSymExpr *SSE = dyn_cast<SymSymExpr>(Sym)) { 1079 if (SSE->getOpcode() == BO_Sub) { 1080 QualType T = Sym->getType(); 1081 1082 // Do not negate unsigned ranges 1083 if (!T->isUnsignedIntegerOrEnumerationType() && 1084 !T->isSignedIntegerOrEnumerationType()) 1085 return llvm::None; 1086 1087 SymbolManager &SymMgr = State->getSymbolManager(); 1088 SymbolRef NegatedSym = 1089 SymMgr.getSymSymExpr(SSE->getRHS(), BO_Sub, SSE->getLHS(), T); 1090 1091 if (const RangeSet *NegatedRange = getConstraint(State, NegatedSym)) { 1092 return RangeFactory.negate(*NegatedRange); 1093 } 1094 } 1095 } 1096 return llvm::None; 1097 } 1098 1099 // Returns ranges only for binary comparison operators (except <=>) 1100 // when left and right operands are symbolic values. 1101 // Finds any other comparisons with the same operands. 1102 // Then do logical calculations and refuse impossible branches. 1103 // E.g. (x < y) and (x > y) at the same time are impossible. 1104 // E.g. (x >= y) and (x != y) at the same time makes (x > y) true only. 1105 // E.g. (x == y) and (y == x) are just reversed but the same. 1106 // It covers all possible combinations (see CmpOpTable description). 1107 // Note that `x` and `y` can also stand for subexpressions, 1108 // not only for actual symbols. 1109 Optional<RangeSet> getRangeForComparisonSymbol(SymbolRef Sym) { 1110 const auto *SSE = dyn_cast<SymSymExpr>(Sym); 1111 if (!SSE) 1112 return llvm::None; 1113 1114 BinaryOperatorKind CurrentOP = SSE->getOpcode(); 1115 1116 // We currently do not support <=> (C++20). 1117 if (!BinaryOperator::isComparisonOp(CurrentOP) || (CurrentOP == BO_Cmp)) 1118 return llvm::None; 1119 1120 static const OperatorRelationsTable CmpOpTable{}; 1121 1122 const SymExpr *LHS = SSE->getLHS(); 1123 const SymExpr *RHS = SSE->getRHS(); 1124 QualType T = SSE->getType(); 1125 1126 SymbolManager &SymMgr = State->getSymbolManager(); 1127 1128 int UnknownStates = 0; 1129 1130 // Loop goes through all of the columns exept the last one ('UnknownX2'). 1131 // We treat `UnknownX2` column separately at the end of the loop body. 1132 for (size_t i = 0; i < CmpOpTable.getCmpOpCount(); ++i) { 1133 1134 // Let's find an expression e.g. (x < y). 1135 BinaryOperatorKind QueriedOP = OperatorRelationsTable::getOpFromIndex(i); 1136 const SymSymExpr *SymSym = SymMgr.getSymSymExpr(LHS, QueriedOP, RHS, T); 1137 const RangeSet *QueriedRangeSet = getConstraint(State, SymSym); 1138 1139 // If ranges were not previously found, 1140 // try to find a reversed expression (y > x). 1141 if (!QueriedRangeSet) { 1142 const BinaryOperatorKind ROP = 1143 BinaryOperator::reverseComparisonOp(QueriedOP); 1144 SymSym = SymMgr.getSymSymExpr(RHS, ROP, LHS, T); 1145 QueriedRangeSet = getConstraint(State, SymSym); 1146 } 1147 1148 if (!QueriedRangeSet || QueriedRangeSet->isEmpty()) 1149 continue; 1150 1151 const llvm::APSInt *ConcreteValue = QueriedRangeSet->getConcreteValue(); 1152 const bool isInFalseBranch = 1153 ConcreteValue ? (*ConcreteValue == 0) : false; 1154 1155 // If it is a false branch, we shall be guided by opposite operator, 1156 // because the table is made assuming we are in the true branch. 1157 // E.g. when (x <= y) is false, then (x > y) is true. 1158 if (isInFalseBranch) 1159 QueriedOP = BinaryOperator::negateComparisonOp(QueriedOP); 1160 1161 OperatorRelationsTable::TriStateKind BranchState = 1162 CmpOpTable.getCmpOpState(CurrentOP, QueriedOP); 1163 1164 if (BranchState == OperatorRelationsTable::Unknown) { 1165 if (++UnknownStates == 2) 1166 // If we met both Unknown states. 1167 // if (x <= y) // assume true 1168 // if (x != y) // assume true 1169 // if (x < y) // would be also true 1170 // Get a state from `UnknownX2` column. 1171 BranchState = CmpOpTable.getCmpOpStateForUnknownX2(CurrentOP); 1172 else 1173 continue; 1174 } 1175 1176 return (BranchState == OperatorRelationsTable::True) ? getTrueRange(T) 1177 : getFalseRange(T); 1178 } 1179 1180 return llvm::None; 1181 } 1182 1183 Optional<RangeSet> getRangeForEqualities(const SymSymExpr *Sym) { 1184 Optional<bool> Equality = meansEquality(Sym); 1185 1186 if (!Equality) 1187 return llvm::None; 1188 1189 if (Optional<bool> AreEqual = 1190 EquivalenceClass::areEqual(State, Sym->getLHS(), Sym->getRHS())) { 1191 // Here we cover two cases at once: 1192 // * if Sym is equality and its operands are known to be equal -> true 1193 // * if Sym is disequality and its operands are disequal -> true 1194 if (*AreEqual == *Equality) { 1195 return getTrueRange(Sym->getType()); 1196 } 1197 // Opposite combinations result in false. 1198 return getFalseRange(Sym->getType()); 1199 } 1200 1201 return llvm::None; 1202 } 1203 1204 RangeSet getTrueRange(QualType T) { 1205 RangeSet TypeRange = infer(T); 1206 return assumeNonZero(TypeRange, T); 1207 } 1208 1209 RangeSet getFalseRange(QualType T) { 1210 const llvm::APSInt &Zero = ValueFactory.getValue(0, T); 1211 return RangeSet(RangeFactory, Zero); 1212 } 1213 1214 BasicValueFactory &ValueFactory; 1215 RangeSet::Factory &RangeFactory; 1216 ProgramStateRef State; 1217 }; 1218 1219 //===----------------------------------------------------------------------===// 1220 // Range-based reasoning about symbolic operations 1221 //===----------------------------------------------------------------------===// 1222 1223 template <> 1224 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_Or>(Range LHS, Range RHS, 1225 QualType T) { 1226 APSIntType ResultType = ValueFactory.getAPSIntType(T); 1227 llvm::APSInt Zero = ResultType.getZeroValue(); 1228 1229 bool IsLHSPositiveOrZero = LHS.From() >= Zero; 1230 bool IsRHSPositiveOrZero = RHS.From() >= Zero; 1231 1232 bool IsLHSNegative = LHS.To() < Zero; 1233 bool IsRHSNegative = RHS.To() < Zero; 1234 1235 // Check if both ranges have the same sign. 1236 if ((IsLHSPositiveOrZero && IsRHSPositiveOrZero) || 1237 (IsLHSNegative && IsRHSNegative)) { 1238 // The result is definitely greater or equal than any of the operands. 1239 const llvm::APSInt &Min = std::max(LHS.From(), RHS.From()); 1240 1241 // We estimate maximal value for positives as the maximal value for the 1242 // given type. For negatives, we estimate it with -1 (e.g. 0x11111111). 1243 // 1244 // TODO: We basically, limit the resulting range from below, but don't do 1245 // anything with the upper bound. 1246 // 1247 // For positive operands, it can be done as follows: for the upper 1248 // bound of LHS and RHS we calculate the most significant bit set. 1249 // Let's call it the N-th bit. Then we can estimate the maximal 1250 // number to be 2^(N+1)-1, i.e. the number with all the bits up to 1251 // the N-th bit set. 1252 const llvm::APSInt &Max = IsLHSNegative 1253 ? ValueFactory.getValue(--Zero) 1254 : ValueFactory.getMaxValue(ResultType); 1255 1256 return {RangeFactory, ValueFactory.getValue(Min), Max}; 1257 } 1258 1259 // Otherwise, let's check if at least one of the operands is negative. 1260 if (IsLHSNegative || IsRHSNegative) { 1261 // This means that the result is definitely negative as well. 1262 return {RangeFactory, ValueFactory.getMinValue(ResultType), 1263 ValueFactory.getValue(--Zero)}; 1264 } 1265 1266 RangeSet DefaultRange = infer(T); 1267 1268 // It is pretty hard to reason about operands with different signs 1269 // (and especially with possibly different signs). We simply check if it 1270 // can be zero. In order to conclude that the result could not be zero, 1271 // at least one of the operands should be definitely not zero itself. 1272 if (!LHS.Includes(Zero) || !RHS.Includes(Zero)) { 1273 return assumeNonZero(DefaultRange, T); 1274 } 1275 1276 // Nothing much else to do here. 1277 return DefaultRange; 1278 } 1279 1280 template <> 1281 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_And>(Range LHS, 1282 Range RHS, 1283 QualType T) { 1284 APSIntType ResultType = ValueFactory.getAPSIntType(T); 1285 llvm::APSInt Zero = ResultType.getZeroValue(); 1286 1287 bool IsLHSPositiveOrZero = LHS.From() >= Zero; 1288 bool IsRHSPositiveOrZero = RHS.From() >= Zero; 1289 1290 bool IsLHSNegative = LHS.To() < Zero; 1291 bool IsRHSNegative = RHS.To() < Zero; 1292 1293 // Check if both ranges have the same sign. 1294 if ((IsLHSPositiveOrZero && IsRHSPositiveOrZero) || 1295 (IsLHSNegative && IsRHSNegative)) { 1296 // The result is definitely less or equal than any of the operands. 1297 const llvm::APSInt &Max = std::min(LHS.To(), RHS.To()); 1298 1299 // We conservatively estimate lower bound to be the smallest positive 1300 // or negative value corresponding to the sign of the operands. 1301 const llvm::APSInt &Min = IsLHSNegative 1302 ? ValueFactory.getMinValue(ResultType) 1303 : ValueFactory.getValue(Zero); 1304 1305 return {RangeFactory, Min, Max}; 1306 } 1307 1308 // Otherwise, let's check if at least one of the operands is positive. 1309 if (IsLHSPositiveOrZero || IsRHSPositiveOrZero) { 1310 // This makes result definitely positive. 1311 // 1312 // We can also reason about a maximal value by finding the maximal 1313 // value of the positive operand. 1314 const llvm::APSInt &Max = IsLHSPositiveOrZero ? LHS.To() : RHS.To(); 1315 1316 // The minimal value on the other hand is much harder to reason about. 1317 // The only thing we know for sure is that the result is positive. 1318 return {RangeFactory, ValueFactory.getValue(Zero), 1319 ValueFactory.getValue(Max)}; 1320 } 1321 1322 // Nothing much else to do here. 1323 return infer(T); 1324 } 1325 1326 template <> 1327 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_Rem>(Range LHS, 1328 Range RHS, 1329 QualType T) { 1330 llvm::APSInt Zero = ValueFactory.getAPSIntType(T).getZeroValue(); 1331 1332 Range ConservativeRange = getSymmetricalRange(RHS, T); 1333 1334 llvm::APSInt Max = ConservativeRange.To(); 1335 llvm::APSInt Min = ConservativeRange.From(); 1336 1337 if (Max == Zero) { 1338 // It's an undefined behaviour to divide by 0 and it seems like we know 1339 // for sure that RHS is 0. Let's say that the resulting range is 1340 // simply infeasible for that matter. 1341 return RangeFactory.getEmptySet(); 1342 } 1343 1344 // At this point, our conservative range is closed. The result, however, 1345 // couldn't be greater than the RHS' maximal absolute value. Because of 1346 // this reason, we turn the range into open (or half-open in case of 1347 // unsigned integers). 1348 // 1349 // While we operate on integer values, an open interval (a, b) can be easily 1350 // represented by the closed interval [a + 1, b - 1]. And this is exactly 1351 // what we do next. 1352 // 1353 // If we are dealing with unsigned case, we shouldn't move the lower bound. 1354 if (Min.isSigned()) { 1355 ++Min; 1356 } 1357 --Max; 1358 1359 bool IsLHSPositiveOrZero = LHS.From() >= Zero; 1360 bool IsRHSPositiveOrZero = RHS.From() >= Zero; 1361 1362 // Remainder operator results with negative operands is implementation 1363 // defined. Positive cases are much easier to reason about though. 1364 if (IsLHSPositiveOrZero && IsRHSPositiveOrZero) { 1365 // If maximal value of LHS is less than maximal value of RHS, 1366 // the result won't get greater than LHS.To(). 1367 Max = std::min(LHS.To(), Max); 1368 // We want to check if it is a situation similar to the following: 1369 // 1370 // <------------|---[ LHS ]--------[ RHS ]-----> 1371 // -INF 0 +INF 1372 // 1373 // In this situation, we can conclude that (LHS / RHS) == 0 and 1374 // (LHS % RHS) == LHS. 1375 Min = LHS.To() < RHS.From() ? LHS.From() : Zero; 1376 } 1377 1378 // Nevertheless, the symmetrical range for RHS is a conservative estimate 1379 // for any sign of either LHS, or RHS. 1380 return {RangeFactory, ValueFactory.getValue(Min), ValueFactory.getValue(Max)}; 1381 } 1382 1383 //===----------------------------------------------------------------------===// 1384 // Constraint assignment logic 1385 //===----------------------------------------------------------------------===// 1386 1387 /// ConstraintAssignorBase is a small utility class that unifies visitor 1388 /// for ranges with a visitor for constraints (rangeset/range/constant). 1389 /// 1390 /// It is designed to have one derived class, but generally it can have more. 1391 /// Derived class can control which types we handle by defining methods of the 1392 /// following form: 1393 /// 1394 /// bool handle${SYMBOL}To${CONSTRAINT}(const SYMBOL *Sym, 1395 /// CONSTRAINT Constraint); 1396 /// 1397 /// where SYMBOL is the type of the symbol (e.g. SymSymExpr, SymbolCast, etc.) 1398 /// CONSTRAINT is the type of constraint (RangeSet/Range/Const) 1399 /// return value signifies whether we should try other handle methods 1400 /// (i.e. false would mean to stop right after calling this method) 1401 template <class Derived> class ConstraintAssignorBase { 1402 public: 1403 using Const = const llvm::APSInt &; 1404 1405 #define DISPATCH(CLASS) return assign##CLASS##Impl(cast<CLASS>(Sym), Constraint) 1406 1407 #define ASSIGN(CLASS, TO, SYM, CONSTRAINT) \ 1408 if (!static_cast<Derived *>(this)->assign##CLASS##To##TO(SYM, CONSTRAINT)) \ 1409 return false 1410 1411 void assign(SymbolRef Sym, RangeSet Constraint) { 1412 assignImpl(Sym, Constraint); 1413 } 1414 1415 bool assignImpl(SymbolRef Sym, RangeSet Constraint) { 1416 switch (Sym->getKind()) { 1417 #define SYMBOL(Id, Parent) \ 1418 case SymExpr::Id##Kind: \ 1419 DISPATCH(Id); 1420 #include "clang/StaticAnalyzer/Core/PathSensitive/Symbols.def" 1421 } 1422 llvm_unreachable("Unknown SymExpr kind!"); 1423 } 1424 1425 #define DEFAULT_ASSIGN(Id) \ 1426 bool assign##Id##To##RangeSet(const Id *Sym, RangeSet Constraint) { \ 1427 return true; \ 1428 } \ 1429 bool assign##Id##To##Range(const Id *Sym, Range Constraint) { return true; } \ 1430 bool assign##Id##To##Const(const Id *Sym, Const Constraint) { return true; } 1431 1432 // When we dispatch for constraint types, we first try to check 1433 // if the new constraint is the constant and try the corresponding 1434 // assignor methods. If it didn't interrupt, we can proceed to the 1435 // range, and finally to the range set. 1436 #define CONSTRAINT_DISPATCH(Id) \ 1437 if (const llvm::APSInt *Const = Constraint.getConcreteValue()) { \ 1438 ASSIGN(Id, Const, Sym, *Const); \ 1439 } \ 1440 if (Constraint.size() == 1) { \ 1441 ASSIGN(Id, Range, Sym, *Constraint.begin()); \ 1442 } \ 1443 ASSIGN(Id, RangeSet, Sym, Constraint) 1444 1445 // Our internal assign method first tries to call assignor methods for all 1446 // constraint types that apply. And if not interrupted, continues with its 1447 // parent class. 1448 #define SYMBOL(Id, Parent) \ 1449 bool assign##Id##Impl(const Id *Sym, RangeSet Constraint) { \ 1450 CONSTRAINT_DISPATCH(Id); \ 1451 DISPATCH(Parent); \ 1452 } \ 1453 DEFAULT_ASSIGN(Id) 1454 #define ABSTRACT_SYMBOL(Id, Parent) SYMBOL(Id, Parent) 1455 #include "clang/StaticAnalyzer/Core/PathSensitive/Symbols.def" 1456 1457 // Default implementations for the top class that doesn't have parents. 1458 bool assignSymExprImpl(const SymExpr *Sym, RangeSet Constraint) { 1459 CONSTRAINT_DISPATCH(SymExpr); 1460 return true; 1461 } 1462 DEFAULT_ASSIGN(SymExpr); 1463 1464 #undef DISPATCH 1465 #undef CONSTRAINT_DISPATCH 1466 #undef DEFAULT_ASSIGN 1467 #undef ASSIGN 1468 }; 1469 1470 /// A little component aggregating all of the reasoning we have about 1471 /// assigning new constraints to symbols. 1472 /// 1473 /// The main purpose of this class is to associate constraints to symbols, 1474 /// and impose additional constraints on other symbols, when we can imply 1475 /// them. 1476 /// 1477 /// It has a nice symmetry with SymbolicRangeInferrer. When the latter 1478 /// can provide more precise ranges by looking into the operands of the 1479 /// expression in question, ConstraintAssignor looks into the operands 1480 /// to see if we can imply more from the new constraint. 1481 class ConstraintAssignor : public ConstraintAssignorBase<ConstraintAssignor> { 1482 public: 1483 template <class ClassOrSymbol> 1484 LLVM_NODISCARD static ProgramStateRef 1485 assign(ProgramStateRef State, SValBuilder &Builder, RangeSet::Factory &F, 1486 ClassOrSymbol CoS, RangeSet NewConstraint) { 1487 if (!State || NewConstraint.isEmpty()) 1488 return nullptr; 1489 1490 ConstraintAssignor Assignor{State, Builder, F}; 1491 return Assignor.assign(CoS, NewConstraint); 1492 } 1493 1494 inline bool assignSymExprToConst(const SymExpr *Sym, Const Constraint); 1495 inline bool assignSymSymExprToRangeSet(const SymSymExpr *Sym, 1496 RangeSet Constraint); 1497 1498 private: 1499 ConstraintAssignor(ProgramStateRef State, SValBuilder &Builder, 1500 RangeSet::Factory &F) 1501 : State(State), Builder(Builder), RangeFactory(F) {} 1502 using Base = ConstraintAssignorBase<ConstraintAssignor>; 1503 1504 /// Base method for handling new constraints for symbols. 1505 LLVM_NODISCARD ProgramStateRef assign(SymbolRef Sym, RangeSet NewConstraint) { 1506 // All constraints are actually associated with equivalence classes, and 1507 // that's what we are going to do first. 1508 State = assign(EquivalenceClass::find(State, Sym), NewConstraint); 1509 if (!State) 1510 return nullptr; 1511 1512 // And after that we can check what other things we can get from this 1513 // constraint. 1514 Base::assign(Sym, NewConstraint); 1515 return State; 1516 } 1517 1518 /// Base method for handling new constraints for classes. 1519 LLVM_NODISCARD ProgramStateRef assign(EquivalenceClass Class, 1520 RangeSet NewConstraint) { 1521 // There is a chance that we might need to update constraints for the 1522 // classes that are known to be disequal to Class. 1523 // 1524 // In order for this to be even possible, the new constraint should 1525 // be simply a constant because we can't reason about range disequalities. 1526 if (const llvm::APSInt *Point = NewConstraint.getConcreteValue()) { 1527 1528 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1529 ConstraintRangeTy::Factory &CF = State->get_context<ConstraintRange>(); 1530 1531 // Add new constraint. 1532 Constraints = CF.add(Constraints, Class, NewConstraint); 1533 1534 for (EquivalenceClass DisequalClass : Class.getDisequalClasses(State)) { 1535 RangeSet UpdatedConstraint = SymbolicRangeInferrer::inferRange( 1536 RangeFactory, State, DisequalClass); 1537 1538 UpdatedConstraint = RangeFactory.deletePoint(UpdatedConstraint, *Point); 1539 1540 // If we end up with at least one of the disequal classes to be 1541 // constrained with an empty range-set, the state is infeasible. 1542 if (UpdatedConstraint.isEmpty()) 1543 return nullptr; 1544 1545 Constraints = CF.add(Constraints, DisequalClass, UpdatedConstraint); 1546 } 1547 assert(areFeasible(Constraints) && "Constraint manager shouldn't produce " 1548 "a state with infeasible constraints"); 1549 1550 return setConstraints(State, Constraints); 1551 } 1552 1553 return setConstraint(State, Class, NewConstraint); 1554 } 1555 1556 ProgramStateRef trackDisequality(ProgramStateRef State, SymbolRef LHS, 1557 SymbolRef RHS) { 1558 return EquivalenceClass::markDisequal(RangeFactory, State, LHS, RHS); 1559 } 1560 1561 ProgramStateRef trackEquality(ProgramStateRef State, SymbolRef LHS, 1562 SymbolRef RHS) { 1563 return EquivalenceClass::merge(RangeFactory, State, LHS, RHS); 1564 } 1565 1566 LLVM_NODISCARD Optional<bool> interpreteAsBool(RangeSet Constraint) { 1567 assert(!Constraint.isEmpty() && "Empty ranges shouldn't get here"); 1568 1569 if (Constraint.getConcreteValue()) 1570 return !Constraint.getConcreteValue()->isNullValue(); 1571 1572 APSIntType T{Constraint.getMinValue()}; 1573 Const Zero = T.getZeroValue(); 1574 if (!Constraint.contains(Zero)) 1575 return true; 1576 1577 return llvm::None; 1578 } 1579 1580 ProgramStateRef State; 1581 SValBuilder &Builder; 1582 RangeSet::Factory &RangeFactory; 1583 }; 1584 1585 //===----------------------------------------------------------------------===// 1586 // Constraint manager implementation details 1587 //===----------------------------------------------------------------------===// 1588 1589 class RangeConstraintManager : public RangedConstraintManager { 1590 public: 1591 RangeConstraintManager(ExprEngine *EE, SValBuilder &SVB) 1592 : RangedConstraintManager(EE, SVB), F(getBasicVals()) {} 1593 1594 //===------------------------------------------------------------------===// 1595 // Implementation for interface from ConstraintManager. 1596 //===------------------------------------------------------------------===// 1597 1598 bool haveEqualConstraints(ProgramStateRef S1, 1599 ProgramStateRef S2) const override { 1600 // NOTE: ClassMembers are as simple as back pointers for ClassMap, 1601 // so comparing constraint ranges and class maps should be 1602 // sufficient. 1603 return S1->get<ConstraintRange>() == S2->get<ConstraintRange>() && 1604 S1->get<ClassMap>() == S2->get<ClassMap>(); 1605 } 1606 1607 bool canReasonAbout(SVal X) const override; 1608 1609 ConditionTruthVal checkNull(ProgramStateRef State, SymbolRef Sym) override; 1610 1611 const llvm::APSInt *getSymVal(ProgramStateRef State, 1612 SymbolRef Sym) const override; 1613 1614 ProgramStateRef removeDeadBindings(ProgramStateRef State, 1615 SymbolReaper &SymReaper) override; 1616 1617 void printJson(raw_ostream &Out, ProgramStateRef State, const char *NL = "\n", 1618 unsigned int Space = 0, bool IsDot = false) const override; 1619 void printConstraints(raw_ostream &Out, ProgramStateRef State, 1620 const char *NL = "\n", unsigned int Space = 0, 1621 bool IsDot = false) const; 1622 void printEquivalenceClasses(raw_ostream &Out, ProgramStateRef State, 1623 const char *NL = "\n", unsigned int Space = 0, 1624 bool IsDot = false) const; 1625 void printDisequalities(raw_ostream &Out, ProgramStateRef State, 1626 const char *NL = "\n", unsigned int Space = 0, 1627 bool IsDot = false) const; 1628 1629 //===------------------------------------------------------------------===// 1630 // Implementation for interface from RangedConstraintManager. 1631 //===------------------------------------------------------------------===// 1632 1633 ProgramStateRef assumeSymNE(ProgramStateRef State, SymbolRef Sym, 1634 const llvm::APSInt &V, 1635 const llvm::APSInt &Adjustment) override; 1636 1637 ProgramStateRef assumeSymEQ(ProgramStateRef State, SymbolRef Sym, 1638 const llvm::APSInt &V, 1639 const llvm::APSInt &Adjustment) override; 1640 1641 ProgramStateRef assumeSymLT(ProgramStateRef State, SymbolRef Sym, 1642 const llvm::APSInt &V, 1643 const llvm::APSInt &Adjustment) override; 1644 1645 ProgramStateRef assumeSymGT(ProgramStateRef State, SymbolRef Sym, 1646 const llvm::APSInt &V, 1647 const llvm::APSInt &Adjustment) override; 1648 1649 ProgramStateRef assumeSymLE(ProgramStateRef State, SymbolRef Sym, 1650 const llvm::APSInt &V, 1651 const llvm::APSInt &Adjustment) override; 1652 1653 ProgramStateRef assumeSymGE(ProgramStateRef State, SymbolRef Sym, 1654 const llvm::APSInt &V, 1655 const llvm::APSInt &Adjustment) override; 1656 1657 ProgramStateRef assumeSymWithinInclusiveRange( 1658 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 1659 const llvm::APSInt &To, const llvm::APSInt &Adjustment) override; 1660 1661 ProgramStateRef assumeSymOutsideInclusiveRange( 1662 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 1663 const llvm::APSInt &To, const llvm::APSInt &Adjustment) override; 1664 1665 private: 1666 RangeSet::Factory F; 1667 1668 RangeSet getRange(ProgramStateRef State, SymbolRef Sym); 1669 RangeSet getRange(ProgramStateRef State, EquivalenceClass Class); 1670 ProgramStateRef setRange(ProgramStateRef State, SymbolRef Sym, 1671 RangeSet Range); 1672 ProgramStateRef setRange(ProgramStateRef State, EquivalenceClass Class, 1673 RangeSet Range); 1674 1675 RangeSet getSymLTRange(ProgramStateRef St, SymbolRef Sym, 1676 const llvm::APSInt &Int, 1677 const llvm::APSInt &Adjustment); 1678 RangeSet getSymGTRange(ProgramStateRef St, SymbolRef Sym, 1679 const llvm::APSInt &Int, 1680 const llvm::APSInt &Adjustment); 1681 RangeSet getSymLERange(ProgramStateRef St, SymbolRef Sym, 1682 const llvm::APSInt &Int, 1683 const llvm::APSInt &Adjustment); 1684 RangeSet getSymLERange(llvm::function_ref<RangeSet()> RS, 1685 const llvm::APSInt &Int, 1686 const llvm::APSInt &Adjustment); 1687 RangeSet getSymGERange(ProgramStateRef St, SymbolRef Sym, 1688 const llvm::APSInt &Int, 1689 const llvm::APSInt &Adjustment); 1690 }; 1691 1692 bool ConstraintAssignor::assignSymExprToConst(const SymExpr *Sym, 1693 const llvm::APSInt &Constraint) { 1694 llvm::SmallSet<EquivalenceClass, 4> SimplifiedClasses; 1695 // Iterate over all equivalence classes and try to simplify them. 1696 ClassMembersTy Members = State->get<ClassMembers>(); 1697 for (std::pair<EquivalenceClass, SymbolSet> ClassToSymbolSet : Members) { 1698 EquivalenceClass Class = ClassToSymbolSet.first; 1699 State = Class.simplify(Builder, RangeFactory, State); 1700 if (!State) 1701 return false; 1702 SimplifiedClasses.insert(Class); 1703 } 1704 1705 // Trivial equivalence classes (those that have only one symbol member) are 1706 // not stored in the State. Thus, we must skim through the constraints as 1707 // well. And we try to simplify symbols in the constraints. 1708 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1709 for (std::pair<EquivalenceClass, RangeSet> ClassConstraint : Constraints) { 1710 EquivalenceClass Class = ClassConstraint.first; 1711 if (SimplifiedClasses.count(Class)) // Already simplified. 1712 continue; 1713 State = Class.simplify(Builder, RangeFactory, State); 1714 if (!State) 1715 return false; 1716 } 1717 1718 return true; 1719 } 1720 1721 bool ConstraintAssignor::assignSymSymExprToRangeSet(const SymSymExpr *Sym, 1722 RangeSet Constraint) { 1723 Optional<bool> ConstraintAsBool = interpreteAsBool(Constraint); 1724 1725 if (!ConstraintAsBool) 1726 return true; 1727 1728 if (Optional<bool> Equality = meansEquality(Sym)) { 1729 // Here we cover two cases: 1730 // * if Sym is equality and the new constraint is true -> Sym's operands 1731 // should be marked as equal 1732 // * if Sym is disequality and the new constraint is false -> Sym's 1733 // operands should be also marked as equal 1734 if (*Equality == *ConstraintAsBool) { 1735 State = trackEquality(State, Sym->getLHS(), Sym->getRHS()); 1736 } else { 1737 // Other combinations leave as with disequal operands. 1738 State = trackDisequality(State, Sym->getLHS(), Sym->getRHS()); 1739 } 1740 1741 if (!State) 1742 return false; 1743 } 1744 1745 return true; 1746 } 1747 1748 } // end anonymous namespace 1749 1750 std::unique_ptr<ConstraintManager> 1751 ento::CreateRangeConstraintManager(ProgramStateManager &StMgr, 1752 ExprEngine *Eng) { 1753 return std::make_unique<RangeConstraintManager>(Eng, StMgr.getSValBuilder()); 1754 } 1755 1756 ConstraintMap ento::getConstraintMap(ProgramStateRef State) { 1757 ConstraintMap::Factory &F = State->get_context<ConstraintMap>(); 1758 ConstraintMap Result = F.getEmptyMap(); 1759 1760 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1761 for (std::pair<EquivalenceClass, RangeSet> ClassConstraint : Constraints) { 1762 EquivalenceClass Class = ClassConstraint.first; 1763 SymbolSet ClassMembers = Class.getClassMembers(State); 1764 assert(!ClassMembers.isEmpty() && 1765 "Class must always have at least one member!"); 1766 1767 SymbolRef Representative = *ClassMembers.begin(); 1768 Result = F.add(Result, Representative, ClassConstraint.second); 1769 } 1770 1771 return Result; 1772 } 1773 1774 //===----------------------------------------------------------------------===// 1775 // EqualityClass implementation details 1776 //===----------------------------------------------------------------------===// 1777 1778 LLVM_DUMP_METHOD void EquivalenceClass::dumpToStream(ProgramStateRef State, 1779 raw_ostream &os) const { 1780 SymbolSet ClassMembers = getClassMembers(State); 1781 for (const SymbolRef &MemberSym : ClassMembers) { 1782 MemberSym->dump(); 1783 os << "\n"; 1784 } 1785 } 1786 1787 inline EquivalenceClass EquivalenceClass::find(ProgramStateRef State, 1788 SymbolRef Sym) { 1789 assert(State && "State should not be null"); 1790 assert(Sym && "Symbol should not be null"); 1791 // We store far from all Symbol -> Class mappings 1792 if (const EquivalenceClass *NontrivialClass = State->get<ClassMap>(Sym)) 1793 return *NontrivialClass; 1794 1795 // This is a trivial class of Sym. 1796 return Sym; 1797 } 1798 1799 inline ProgramStateRef EquivalenceClass::merge(RangeSet::Factory &F, 1800 ProgramStateRef State, 1801 SymbolRef First, 1802 SymbolRef Second) { 1803 EquivalenceClass FirstClass = find(State, First); 1804 EquivalenceClass SecondClass = find(State, Second); 1805 1806 return FirstClass.merge(F, State, SecondClass); 1807 } 1808 1809 inline ProgramStateRef EquivalenceClass::merge(RangeSet::Factory &F, 1810 ProgramStateRef State, 1811 EquivalenceClass Other) { 1812 // It is already the same class. 1813 if (*this == Other) 1814 return State; 1815 1816 // FIXME: As of now, we support only equivalence classes of the same type. 1817 // This limitation is connected to the lack of explicit casts in 1818 // our symbolic expression model. 1819 // 1820 // That means that for `int x` and `char y` we don't distinguish 1821 // between these two very different cases: 1822 // * `x == y` 1823 // * `(char)x == y` 1824 // 1825 // The moment we introduce symbolic casts, this restriction can be 1826 // lifted. 1827 if (getType() != Other.getType()) 1828 return State; 1829 1830 SymbolSet Members = getClassMembers(State); 1831 SymbolSet OtherMembers = Other.getClassMembers(State); 1832 1833 // We estimate the size of the class by the height of tree containing 1834 // its members. Merging is not a trivial operation, so it's easier to 1835 // merge the smaller class into the bigger one. 1836 if (Members.getHeight() >= OtherMembers.getHeight()) { 1837 return mergeImpl(F, State, Members, Other, OtherMembers); 1838 } else { 1839 return Other.mergeImpl(F, State, OtherMembers, *this, Members); 1840 } 1841 } 1842 1843 inline ProgramStateRef 1844 EquivalenceClass::mergeImpl(RangeSet::Factory &RangeFactory, 1845 ProgramStateRef State, SymbolSet MyMembers, 1846 EquivalenceClass Other, SymbolSet OtherMembers) { 1847 // Essentially what we try to recreate here is some kind of union-find 1848 // data structure. It does have certain limitations due to persistence 1849 // and the need to remove elements from classes. 1850 // 1851 // In this setting, EquialityClass object is the representative of the class 1852 // or the parent element. ClassMap is a mapping of class members to their 1853 // parent. Unlike the union-find structure, they all point directly to the 1854 // class representative because we don't have an opportunity to actually do 1855 // path compression when dealing with immutability. This means that we 1856 // compress paths every time we do merges. It also means that we lose 1857 // the main amortized complexity benefit from the original data structure. 1858 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1859 ConstraintRangeTy::Factory &CRF = State->get_context<ConstraintRange>(); 1860 1861 // 1. If the merged classes have any constraints associated with them, we 1862 // need to transfer them to the class we have left. 1863 // 1864 // Intersection here makes perfect sense because both of these constraints 1865 // must hold for the whole new class. 1866 if (Optional<RangeSet> NewClassConstraint = 1867 intersect(RangeFactory, getConstraint(State, *this), 1868 getConstraint(State, Other))) { 1869 // NOTE: Essentially, NewClassConstraint should NEVER be infeasible because 1870 // range inferrer shouldn't generate ranges incompatible with 1871 // equivalence classes. However, at the moment, due to imperfections 1872 // in the solver, it is possible and the merge function can also 1873 // return infeasible states aka null states. 1874 if (NewClassConstraint->isEmpty()) 1875 // Infeasible state 1876 return nullptr; 1877 1878 // No need in tracking constraints of a now-dissolved class. 1879 Constraints = CRF.remove(Constraints, Other); 1880 // Assign new constraints for this class. 1881 Constraints = CRF.add(Constraints, *this, *NewClassConstraint); 1882 1883 assert(areFeasible(Constraints) && "Constraint manager shouldn't produce " 1884 "a state with infeasible constraints"); 1885 1886 State = State->set<ConstraintRange>(Constraints); 1887 } 1888 1889 // 2. Get ALL equivalence-related maps 1890 ClassMapTy Classes = State->get<ClassMap>(); 1891 ClassMapTy::Factory &CMF = State->get_context<ClassMap>(); 1892 1893 ClassMembersTy Members = State->get<ClassMembers>(); 1894 ClassMembersTy::Factory &MF = State->get_context<ClassMembers>(); 1895 1896 DisequalityMapTy DisequalityInfo = State->get<DisequalityMap>(); 1897 DisequalityMapTy::Factory &DF = State->get_context<DisequalityMap>(); 1898 1899 ClassSet::Factory &CF = State->get_context<ClassSet>(); 1900 SymbolSet::Factory &F = getMembersFactory(State); 1901 1902 // 2. Merge members of the Other class into the current class. 1903 SymbolSet NewClassMembers = MyMembers; 1904 for (SymbolRef Sym : OtherMembers) { 1905 NewClassMembers = F.add(NewClassMembers, Sym); 1906 // *this is now the class for all these new symbols. 1907 Classes = CMF.add(Classes, Sym, *this); 1908 } 1909 1910 // 3. Adjust member mapping. 1911 // 1912 // No need in tracking members of a now-dissolved class. 1913 Members = MF.remove(Members, Other); 1914 // Now only the current class is mapped to all the symbols. 1915 Members = MF.add(Members, *this, NewClassMembers); 1916 1917 // 4. Update disequality relations 1918 ClassSet DisequalToOther = Other.getDisequalClasses(DisequalityInfo, CF); 1919 // We are about to merge two classes but they are already known to be 1920 // non-equal. This is a contradiction. 1921 if (DisequalToOther.contains(*this)) 1922 return nullptr; 1923 1924 if (!DisequalToOther.isEmpty()) { 1925 ClassSet DisequalToThis = getDisequalClasses(DisequalityInfo, CF); 1926 DisequalityInfo = DF.remove(DisequalityInfo, Other); 1927 1928 for (EquivalenceClass DisequalClass : DisequalToOther) { 1929 DisequalToThis = CF.add(DisequalToThis, DisequalClass); 1930 1931 // Disequality is a symmetric relation meaning that if 1932 // DisequalToOther not null then the set for DisequalClass is not 1933 // empty and has at least Other. 1934 ClassSet OriginalSetLinkedToOther = 1935 *DisequalityInfo.lookup(DisequalClass); 1936 1937 // Other will be eliminated and we should replace it with the bigger 1938 // united class. 1939 ClassSet NewSet = CF.remove(OriginalSetLinkedToOther, Other); 1940 NewSet = CF.add(NewSet, *this); 1941 1942 DisequalityInfo = DF.add(DisequalityInfo, DisequalClass, NewSet); 1943 } 1944 1945 DisequalityInfo = DF.add(DisequalityInfo, *this, DisequalToThis); 1946 State = State->set<DisequalityMap>(DisequalityInfo); 1947 } 1948 1949 // 5. Update the state 1950 State = State->set<ClassMap>(Classes); 1951 State = State->set<ClassMembers>(Members); 1952 1953 return State; 1954 } 1955 1956 inline SymbolSet::Factory & 1957 EquivalenceClass::getMembersFactory(ProgramStateRef State) { 1958 return State->get_context<SymbolSet>(); 1959 } 1960 1961 SymbolSet EquivalenceClass::getClassMembers(ProgramStateRef State) const { 1962 if (const SymbolSet *Members = State->get<ClassMembers>(*this)) 1963 return *Members; 1964 1965 // This class is trivial, so we need to construct a set 1966 // with just that one symbol from the class. 1967 SymbolSet::Factory &F = getMembersFactory(State); 1968 return F.add(F.getEmptySet(), getRepresentativeSymbol()); 1969 } 1970 1971 bool EquivalenceClass::isTrivial(ProgramStateRef State) const { 1972 return State->get<ClassMembers>(*this) == nullptr; 1973 } 1974 1975 bool EquivalenceClass::isTriviallyDead(ProgramStateRef State, 1976 SymbolReaper &Reaper) const { 1977 return isTrivial(State) && Reaper.isDead(getRepresentativeSymbol()); 1978 } 1979 1980 inline ProgramStateRef EquivalenceClass::markDisequal(RangeSet::Factory &RF, 1981 ProgramStateRef State, 1982 SymbolRef First, 1983 SymbolRef Second) { 1984 return markDisequal(RF, State, find(State, First), find(State, Second)); 1985 } 1986 1987 inline ProgramStateRef EquivalenceClass::markDisequal(RangeSet::Factory &RF, 1988 ProgramStateRef State, 1989 EquivalenceClass First, 1990 EquivalenceClass Second) { 1991 return First.markDisequal(RF, State, Second); 1992 } 1993 1994 inline ProgramStateRef 1995 EquivalenceClass::markDisequal(RangeSet::Factory &RF, ProgramStateRef State, 1996 EquivalenceClass Other) const { 1997 // If we know that two classes are equal, we can only produce an infeasible 1998 // state. 1999 if (*this == Other) { 2000 return nullptr; 2001 } 2002 2003 DisequalityMapTy DisequalityInfo = State->get<DisequalityMap>(); 2004 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 2005 2006 // Disequality is a symmetric relation, so if we mark A as disequal to B, 2007 // we should also mark B as disequalt to A. 2008 if (!addToDisequalityInfo(DisequalityInfo, Constraints, RF, State, *this, 2009 Other) || 2010 !addToDisequalityInfo(DisequalityInfo, Constraints, RF, State, Other, 2011 *this)) 2012 return nullptr; 2013 2014 assert(areFeasible(Constraints) && "Constraint manager shouldn't produce " 2015 "a state with infeasible constraints"); 2016 2017 State = State->set<DisequalityMap>(DisequalityInfo); 2018 State = State->set<ConstraintRange>(Constraints); 2019 2020 return State; 2021 } 2022 2023 inline bool EquivalenceClass::addToDisequalityInfo( 2024 DisequalityMapTy &Info, ConstraintRangeTy &Constraints, 2025 RangeSet::Factory &RF, ProgramStateRef State, EquivalenceClass First, 2026 EquivalenceClass Second) { 2027 2028 // 1. Get all of the required factories. 2029 DisequalityMapTy::Factory &F = State->get_context<DisequalityMap>(); 2030 ClassSet::Factory &CF = State->get_context<ClassSet>(); 2031 ConstraintRangeTy::Factory &CRF = State->get_context<ConstraintRange>(); 2032 2033 // 2. Add Second to the set of classes disequal to First. 2034 const ClassSet *CurrentSet = Info.lookup(First); 2035 ClassSet NewSet = CurrentSet ? *CurrentSet : CF.getEmptySet(); 2036 NewSet = CF.add(NewSet, Second); 2037 2038 Info = F.add(Info, First, NewSet); 2039 2040 // 3. If Second is known to be a constant, we can delete this point 2041 // from the constraint asociated with First. 2042 // 2043 // So, if Second == 10, it means that First != 10. 2044 // At the same time, the same logic does not apply to ranges. 2045 if (const RangeSet *SecondConstraint = Constraints.lookup(Second)) 2046 if (const llvm::APSInt *Point = SecondConstraint->getConcreteValue()) { 2047 2048 RangeSet FirstConstraint = SymbolicRangeInferrer::inferRange( 2049 RF, State, First.getRepresentativeSymbol()); 2050 2051 FirstConstraint = RF.deletePoint(FirstConstraint, *Point); 2052 2053 // If the First class is about to be constrained with an empty 2054 // range-set, the state is infeasible. 2055 if (FirstConstraint.isEmpty()) 2056 return false; 2057 2058 Constraints = CRF.add(Constraints, First, FirstConstraint); 2059 } 2060 2061 return true; 2062 } 2063 2064 inline Optional<bool> EquivalenceClass::areEqual(ProgramStateRef State, 2065 SymbolRef FirstSym, 2066 SymbolRef SecondSym) { 2067 return EquivalenceClass::areEqual(State, find(State, FirstSym), 2068 find(State, SecondSym)); 2069 } 2070 2071 inline Optional<bool> EquivalenceClass::areEqual(ProgramStateRef State, 2072 EquivalenceClass First, 2073 EquivalenceClass Second) { 2074 // The same equivalence class => symbols are equal. 2075 if (First == Second) 2076 return true; 2077 2078 // Let's check if we know anything about these two classes being not equal to 2079 // each other. 2080 ClassSet DisequalToFirst = First.getDisequalClasses(State); 2081 if (DisequalToFirst.contains(Second)) 2082 return false; 2083 2084 // It is not clear. 2085 return llvm::None; 2086 } 2087 2088 // Iterate over all symbols and try to simplify them. Once a symbol is 2089 // simplified then we check if we can merge the simplified symbol's equivalence 2090 // class to this class. This way, we simplify not just the symbols but the 2091 // classes as well: we strive to keep the number of the classes to be the 2092 // absolute minimum. 2093 LLVM_NODISCARD ProgramStateRef EquivalenceClass::simplify( 2094 SValBuilder &SVB, RangeSet::Factory &F, ProgramStateRef State) { 2095 SymbolSet ClassMembers = getClassMembers(State); 2096 for (const SymbolRef &MemberSym : ClassMembers) { 2097 SymbolRef SimplifiedMemberSym = ento::simplify(State, MemberSym); 2098 if (SimplifiedMemberSym && MemberSym != SimplifiedMemberSym) { 2099 EquivalenceClass ClassOfSimplifiedSym = 2100 EquivalenceClass::find(State, SimplifiedMemberSym); 2101 // The simplified symbol should be the member of the original Class, 2102 // however, it might be in another existing class at the moment. We 2103 // have to merge these classes. 2104 State = merge(F, State, ClassOfSimplifiedSym); 2105 if (!State) 2106 return nullptr; 2107 } 2108 } 2109 return State; 2110 } 2111 2112 inline ClassSet EquivalenceClass::getDisequalClasses(ProgramStateRef State, 2113 SymbolRef Sym) { 2114 return find(State, Sym).getDisequalClasses(State); 2115 } 2116 2117 inline ClassSet 2118 EquivalenceClass::getDisequalClasses(ProgramStateRef State) const { 2119 return getDisequalClasses(State->get<DisequalityMap>(), 2120 State->get_context<ClassSet>()); 2121 } 2122 2123 inline ClassSet 2124 EquivalenceClass::getDisequalClasses(DisequalityMapTy Map, 2125 ClassSet::Factory &Factory) const { 2126 if (const ClassSet *DisequalClasses = Map.lookup(*this)) 2127 return *DisequalClasses; 2128 2129 return Factory.getEmptySet(); 2130 } 2131 2132 bool EquivalenceClass::isClassDataConsistent(ProgramStateRef State) { 2133 ClassMembersTy Members = State->get<ClassMembers>(); 2134 2135 for (std::pair<EquivalenceClass, SymbolSet> ClassMembersPair : Members) { 2136 for (SymbolRef Member : ClassMembersPair.second) { 2137 // Every member of the class should have a mapping back to the class. 2138 if (find(State, Member) == ClassMembersPair.first) { 2139 continue; 2140 } 2141 2142 return false; 2143 } 2144 } 2145 2146 DisequalityMapTy Disequalities = State->get<DisequalityMap>(); 2147 for (std::pair<EquivalenceClass, ClassSet> DisequalityInfo : Disequalities) { 2148 EquivalenceClass Class = DisequalityInfo.first; 2149 ClassSet DisequalClasses = DisequalityInfo.second; 2150 2151 // There is no use in keeping empty sets in the map. 2152 if (DisequalClasses.isEmpty()) 2153 return false; 2154 2155 // Disequality is symmetrical, i.e. for every Class A and B that A != B, 2156 // B != A should also be true. 2157 for (EquivalenceClass DisequalClass : DisequalClasses) { 2158 const ClassSet *DisequalToDisequalClasses = 2159 Disequalities.lookup(DisequalClass); 2160 2161 // It should be a set of at least one element: Class 2162 if (!DisequalToDisequalClasses || 2163 !DisequalToDisequalClasses->contains(Class)) 2164 return false; 2165 } 2166 } 2167 2168 return true; 2169 } 2170 2171 //===----------------------------------------------------------------------===// 2172 // RangeConstraintManager implementation 2173 //===----------------------------------------------------------------------===// 2174 2175 bool RangeConstraintManager::canReasonAbout(SVal X) const { 2176 Optional<nonloc::SymbolVal> SymVal = X.getAs<nonloc::SymbolVal>(); 2177 if (SymVal && SymVal->isExpression()) { 2178 const SymExpr *SE = SymVal->getSymbol(); 2179 2180 if (const SymIntExpr *SIE = dyn_cast<SymIntExpr>(SE)) { 2181 switch (SIE->getOpcode()) { 2182 // We don't reason yet about bitwise-constraints on symbolic values. 2183 case BO_And: 2184 case BO_Or: 2185 case BO_Xor: 2186 return false; 2187 // We don't reason yet about these arithmetic constraints on 2188 // symbolic values. 2189 case BO_Mul: 2190 case BO_Div: 2191 case BO_Rem: 2192 case BO_Shl: 2193 case BO_Shr: 2194 return false; 2195 // All other cases. 2196 default: 2197 return true; 2198 } 2199 } 2200 2201 if (const SymSymExpr *SSE = dyn_cast<SymSymExpr>(SE)) { 2202 // FIXME: Handle <=> here. 2203 if (BinaryOperator::isEqualityOp(SSE->getOpcode()) || 2204 BinaryOperator::isRelationalOp(SSE->getOpcode())) { 2205 // We handle Loc <> Loc comparisons, but not (yet) NonLoc <> NonLoc. 2206 // We've recently started producing Loc <> NonLoc comparisons (that 2207 // result from casts of one of the operands between eg. intptr_t and 2208 // void *), but we can't reason about them yet. 2209 if (Loc::isLocType(SSE->getLHS()->getType())) { 2210 return Loc::isLocType(SSE->getRHS()->getType()); 2211 } 2212 } 2213 } 2214 2215 return false; 2216 } 2217 2218 return true; 2219 } 2220 2221 ConditionTruthVal RangeConstraintManager::checkNull(ProgramStateRef State, 2222 SymbolRef Sym) { 2223 const RangeSet *Ranges = getConstraint(State, Sym); 2224 2225 // If we don't have any information about this symbol, it's underconstrained. 2226 if (!Ranges) 2227 return ConditionTruthVal(); 2228 2229 // If we have a concrete value, see if it's zero. 2230 if (const llvm::APSInt *Value = Ranges->getConcreteValue()) 2231 return *Value == 0; 2232 2233 BasicValueFactory &BV = getBasicVals(); 2234 APSIntType IntType = BV.getAPSIntType(Sym->getType()); 2235 llvm::APSInt Zero = IntType.getZeroValue(); 2236 2237 // Check if zero is in the set of possible values. 2238 if (!Ranges->contains(Zero)) 2239 return false; 2240 2241 // Zero is a possible value, but it is not the /only/ possible value. 2242 return ConditionTruthVal(); 2243 } 2244 2245 const llvm::APSInt *RangeConstraintManager::getSymVal(ProgramStateRef St, 2246 SymbolRef Sym) const { 2247 const RangeSet *T = getConstraint(St, Sym); 2248 return T ? T->getConcreteValue() : nullptr; 2249 } 2250 2251 //===----------------------------------------------------------------------===// 2252 // Remove dead symbols from existing constraints 2253 //===----------------------------------------------------------------------===// 2254 2255 /// Scan all symbols referenced by the constraints. If the symbol is not alive 2256 /// as marked in LSymbols, mark it as dead in DSymbols. 2257 ProgramStateRef 2258 RangeConstraintManager::removeDeadBindings(ProgramStateRef State, 2259 SymbolReaper &SymReaper) { 2260 ClassMembersTy ClassMembersMap = State->get<ClassMembers>(); 2261 ClassMembersTy NewClassMembersMap = ClassMembersMap; 2262 ClassMembersTy::Factory &EMFactory = State->get_context<ClassMembers>(); 2263 SymbolSet::Factory &SetFactory = State->get_context<SymbolSet>(); 2264 2265 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 2266 ConstraintRangeTy NewConstraints = Constraints; 2267 ConstraintRangeTy::Factory &ConstraintFactory = 2268 State->get_context<ConstraintRange>(); 2269 2270 ClassMapTy Map = State->get<ClassMap>(); 2271 ClassMapTy NewMap = Map; 2272 ClassMapTy::Factory &ClassFactory = State->get_context<ClassMap>(); 2273 2274 DisequalityMapTy Disequalities = State->get<DisequalityMap>(); 2275 DisequalityMapTy::Factory &DisequalityFactory = 2276 State->get_context<DisequalityMap>(); 2277 ClassSet::Factory &ClassSetFactory = State->get_context<ClassSet>(); 2278 2279 bool ClassMapChanged = false; 2280 bool MembersMapChanged = false; 2281 bool ConstraintMapChanged = false; 2282 bool DisequalitiesChanged = false; 2283 2284 auto removeDeadClass = [&](EquivalenceClass Class) { 2285 // Remove associated constraint ranges. 2286 Constraints = ConstraintFactory.remove(Constraints, Class); 2287 ConstraintMapChanged = true; 2288 2289 // Update disequality information to not hold any information on the 2290 // removed class. 2291 ClassSet DisequalClasses = 2292 Class.getDisequalClasses(Disequalities, ClassSetFactory); 2293 if (!DisequalClasses.isEmpty()) { 2294 for (EquivalenceClass DisequalClass : DisequalClasses) { 2295 ClassSet DisequalToDisequalSet = 2296 DisequalClass.getDisequalClasses(Disequalities, ClassSetFactory); 2297 // DisequalToDisequalSet is guaranteed to be non-empty for consistent 2298 // disequality info. 2299 assert(!DisequalToDisequalSet.isEmpty()); 2300 ClassSet NewSet = ClassSetFactory.remove(DisequalToDisequalSet, Class); 2301 2302 // No need in keeping an empty set. 2303 if (NewSet.isEmpty()) { 2304 Disequalities = 2305 DisequalityFactory.remove(Disequalities, DisequalClass); 2306 } else { 2307 Disequalities = 2308 DisequalityFactory.add(Disequalities, DisequalClass, NewSet); 2309 } 2310 } 2311 // Remove the data for the class 2312 Disequalities = DisequalityFactory.remove(Disequalities, Class); 2313 DisequalitiesChanged = true; 2314 } 2315 }; 2316 2317 // 1. Let's see if dead symbols are trivial and have associated constraints. 2318 for (std::pair<EquivalenceClass, RangeSet> ClassConstraintPair : 2319 Constraints) { 2320 EquivalenceClass Class = ClassConstraintPair.first; 2321 if (Class.isTriviallyDead(State, SymReaper)) { 2322 // If this class is trivial, we can remove its constraints right away. 2323 removeDeadClass(Class); 2324 } 2325 } 2326 2327 // 2. We don't need to track classes for dead symbols. 2328 for (std::pair<SymbolRef, EquivalenceClass> SymbolClassPair : Map) { 2329 SymbolRef Sym = SymbolClassPair.first; 2330 2331 if (SymReaper.isDead(Sym)) { 2332 ClassMapChanged = true; 2333 NewMap = ClassFactory.remove(NewMap, Sym); 2334 } 2335 } 2336 2337 // 3. Remove dead members from classes and remove dead non-trivial classes 2338 // and their constraints. 2339 for (std::pair<EquivalenceClass, SymbolSet> ClassMembersPair : 2340 ClassMembersMap) { 2341 EquivalenceClass Class = ClassMembersPair.first; 2342 SymbolSet LiveMembers = ClassMembersPair.second; 2343 bool MembersChanged = false; 2344 2345 for (SymbolRef Member : ClassMembersPair.second) { 2346 if (SymReaper.isDead(Member)) { 2347 MembersChanged = true; 2348 LiveMembers = SetFactory.remove(LiveMembers, Member); 2349 } 2350 } 2351 2352 // Check if the class changed. 2353 if (!MembersChanged) 2354 continue; 2355 2356 MembersMapChanged = true; 2357 2358 if (LiveMembers.isEmpty()) { 2359 // The class is dead now, we need to wipe it out of the members map... 2360 NewClassMembersMap = EMFactory.remove(NewClassMembersMap, Class); 2361 2362 // ...and remove all of its constraints. 2363 removeDeadClass(Class); 2364 } else { 2365 // We need to change the members associated with the class. 2366 NewClassMembersMap = 2367 EMFactory.add(NewClassMembersMap, Class, LiveMembers); 2368 } 2369 } 2370 2371 // 4. Update the state with new maps. 2372 // 2373 // Here we try to be humble and update a map only if it really changed. 2374 if (ClassMapChanged) 2375 State = State->set<ClassMap>(NewMap); 2376 2377 if (MembersMapChanged) 2378 State = State->set<ClassMembers>(NewClassMembersMap); 2379 2380 if (ConstraintMapChanged) 2381 State = State->set<ConstraintRange>(Constraints); 2382 2383 if (DisequalitiesChanged) 2384 State = State->set<DisequalityMap>(Disequalities); 2385 2386 assert(EquivalenceClass::isClassDataConsistent(State)); 2387 2388 return State; 2389 } 2390 2391 RangeSet RangeConstraintManager::getRange(ProgramStateRef State, 2392 SymbolRef Sym) { 2393 return SymbolicRangeInferrer::inferRange(F, State, Sym); 2394 } 2395 2396 ProgramStateRef RangeConstraintManager::setRange(ProgramStateRef State, 2397 SymbolRef Sym, 2398 RangeSet Range) { 2399 return ConstraintAssignor::assign(State, getSValBuilder(), F, Sym, Range); 2400 } 2401 2402 //===------------------------------------------------------------------------=== 2403 // assumeSymX methods: protected interface for RangeConstraintManager. 2404 //===------------------------------------------------------------------------===/ 2405 2406 // The syntax for ranges below is mathematical, using [x, y] for closed ranges 2407 // and (x, y) for open ranges. These ranges are modular, corresponding with 2408 // a common treatment of C integer overflow. This means that these methods 2409 // do not have to worry about overflow; RangeSet::Intersect can handle such a 2410 // "wraparound" range. 2411 // As an example, the range [UINT_MAX-1, 3) contains five values: UINT_MAX-1, 2412 // UINT_MAX, 0, 1, and 2. 2413 2414 ProgramStateRef 2415 RangeConstraintManager::assumeSymNE(ProgramStateRef St, SymbolRef Sym, 2416 const llvm::APSInt &Int, 2417 const llvm::APSInt &Adjustment) { 2418 // Before we do any real work, see if the value can even show up. 2419 APSIntType AdjustmentType(Adjustment); 2420 if (AdjustmentType.testInRange(Int, true) != APSIntType::RTR_Within) 2421 return St; 2422 2423 llvm::APSInt Point = AdjustmentType.convert(Int) - Adjustment; 2424 RangeSet New = getRange(St, Sym); 2425 New = F.deletePoint(New, Point); 2426 2427 return setRange(St, Sym, New); 2428 } 2429 2430 ProgramStateRef 2431 RangeConstraintManager::assumeSymEQ(ProgramStateRef St, SymbolRef Sym, 2432 const llvm::APSInt &Int, 2433 const llvm::APSInt &Adjustment) { 2434 // Before we do any real work, see if the value can even show up. 2435 APSIntType AdjustmentType(Adjustment); 2436 if (AdjustmentType.testInRange(Int, true) != APSIntType::RTR_Within) 2437 return nullptr; 2438 2439 // [Int-Adjustment, Int-Adjustment] 2440 llvm::APSInt AdjInt = AdjustmentType.convert(Int) - Adjustment; 2441 RangeSet New = getRange(St, Sym); 2442 New = F.intersect(New, AdjInt); 2443 2444 return setRange(St, Sym, New); 2445 } 2446 2447 RangeSet RangeConstraintManager::getSymLTRange(ProgramStateRef St, 2448 SymbolRef Sym, 2449 const llvm::APSInt &Int, 2450 const llvm::APSInt &Adjustment) { 2451 // Before we do any real work, see if the value can even show up. 2452 APSIntType AdjustmentType(Adjustment); 2453 switch (AdjustmentType.testInRange(Int, true)) { 2454 case APSIntType::RTR_Below: 2455 return F.getEmptySet(); 2456 case APSIntType::RTR_Within: 2457 break; 2458 case APSIntType::RTR_Above: 2459 return getRange(St, Sym); 2460 } 2461 2462 // Special case for Int == Min. This is always false. 2463 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2464 llvm::APSInt Min = AdjustmentType.getMinValue(); 2465 if (ComparisonVal == Min) 2466 return F.getEmptySet(); 2467 2468 llvm::APSInt Lower = Min - Adjustment; 2469 llvm::APSInt Upper = ComparisonVal - Adjustment; 2470 --Upper; 2471 2472 RangeSet Result = getRange(St, Sym); 2473 return F.intersect(Result, Lower, Upper); 2474 } 2475 2476 ProgramStateRef 2477 RangeConstraintManager::assumeSymLT(ProgramStateRef St, SymbolRef Sym, 2478 const llvm::APSInt &Int, 2479 const llvm::APSInt &Adjustment) { 2480 RangeSet New = getSymLTRange(St, Sym, Int, Adjustment); 2481 return setRange(St, Sym, New); 2482 } 2483 2484 RangeSet RangeConstraintManager::getSymGTRange(ProgramStateRef St, 2485 SymbolRef Sym, 2486 const llvm::APSInt &Int, 2487 const llvm::APSInt &Adjustment) { 2488 // Before we do any real work, see if the value can even show up. 2489 APSIntType AdjustmentType(Adjustment); 2490 switch (AdjustmentType.testInRange(Int, true)) { 2491 case APSIntType::RTR_Below: 2492 return getRange(St, Sym); 2493 case APSIntType::RTR_Within: 2494 break; 2495 case APSIntType::RTR_Above: 2496 return F.getEmptySet(); 2497 } 2498 2499 // Special case for Int == Max. This is always false. 2500 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2501 llvm::APSInt Max = AdjustmentType.getMaxValue(); 2502 if (ComparisonVal == Max) 2503 return F.getEmptySet(); 2504 2505 llvm::APSInt Lower = ComparisonVal - Adjustment; 2506 llvm::APSInt Upper = Max - Adjustment; 2507 ++Lower; 2508 2509 RangeSet SymRange = getRange(St, Sym); 2510 return F.intersect(SymRange, Lower, Upper); 2511 } 2512 2513 ProgramStateRef 2514 RangeConstraintManager::assumeSymGT(ProgramStateRef St, SymbolRef Sym, 2515 const llvm::APSInt &Int, 2516 const llvm::APSInt &Adjustment) { 2517 RangeSet New = getSymGTRange(St, Sym, Int, Adjustment); 2518 return setRange(St, Sym, New); 2519 } 2520 2521 RangeSet RangeConstraintManager::getSymGERange(ProgramStateRef St, 2522 SymbolRef Sym, 2523 const llvm::APSInt &Int, 2524 const llvm::APSInt &Adjustment) { 2525 // Before we do any real work, see if the value can even show up. 2526 APSIntType AdjustmentType(Adjustment); 2527 switch (AdjustmentType.testInRange(Int, true)) { 2528 case APSIntType::RTR_Below: 2529 return getRange(St, Sym); 2530 case APSIntType::RTR_Within: 2531 break; 2532 case APSIntType::RTR_Above: 2533 return F.getEmptySet(); 2534 } 2535 2536 // Special case for Int == Min. This is always feasible. 2537 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2538 llvm::APSInt Min = AdjustmentType.getMinValue(); 2539 if (ComparisonVal == Min) 2540 return getRange(St, Sym); 2541 2542 llvm::APSInt Max = AdjustmentType.getMaxValue(); 2543 llvm::APSInt Lower = ComparisonVal - Adjustment; 2544 llvm::APSInt Upper = Max - Adjustment; 2545 2546 RangeSet SymRange = getRange(St, Sym); 2547 return F.intersect(SymRange, Lower, Upper); 2548 } 2549 2550 ProgramStateRef 2551 RangeConstraintManager::assumeSymGE(ProgramStateRef St, SymbolRef Sym, 2552 const llvm::APSInt &Int, 2553 const llvm::APSInt &Adjustment) { 2554 RangeSet New = getSymGERange(St, Sym, Int, Adjustment); 2555 return setRange(St, Sym, New); 2556 } 2557 2558 RangeSet 2559 RangeConstraintManager::getSymLERange(llvm::function_ref<RangeSet()> RS, 2560 const llvm::APSInt &Int, 2561 const llvm::APSInt &Adjustment) { 2562 // Before we do any real work, see if the value can even show up. 2563 APSIntType AdjustmentType(Adjustment); 2564 switch (AdjustmentType.testInRange(Int, true)) { 2565 case APSIntType::RTR_Below: 2566 return F.getEmptySet(); 2567 case APSIntType::RTR_Within: 2568 break; 2569 case APSIntType::RTR_Above: 2570 return RS(); 2571 } 2572 2573 // Special case for Int == Max. This is always feasible. 2574 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2575 llvm::APSInt Max = AdjustmentType.getMaxValue(); 2576 if (ComparisonVal == Max) 2577 return RS(); 2578 2579 llvm::APSInt Min = AdjustmentType.getMinValue(); 2580 llvm::APSInt Lower = Min - Adjustment; 2581 llvm::APSInt Upper = ComparisonVal - Adjustment; 2582 2583 RangeSet Default = RS(); 2584 return F.intersect(Default, Lower, Upper); 2585 } 2586 2587 RangeSet RangeConstraintManager::getSymLERange(ProgramStateRef St, 2588 SymbolRef Sym, 2589 const llvm::APSInt &Int, 2590 const llvm::APSInt &Adjustment) { 2591 return getSymLERange([&] { return getRange(St, Sym); }, Int, Adjustment); 2592 } 2593 2594 ProgramStateRef 2595 RangeConstraintManager::assumeSymLE(ProgramStateRef St, SymbolRef Sym, 2596 const llvm::APSInt &Int, 2597 const llvm::APSInt &Adjustment) { 2598 RangeSet New = getSymLERange(St, Sym, Int, Adjustment); 2599 return setRange(St, Sym, New); 2600 } 2601 2602 ProgramStateRef RangeConstraintManager::assumeSymWithinInclusiveRange( 2603 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 2604 const llvm::APSInt &To, const llvm::APSInt &Adjustment) { 2605 RangeSet New = getSymGERange(State, Sym, From, Adjustment); 2606 if (New.isEmpty()) 2607 return nullptr; 2608 RangeSet Out = getSymLERange([&] { return New; }, To, Adjustment); 2609 return setRange(State, Sym, Out); 2610 } 2611 2612 ProgramStateRef RangeConstraintManager::assumeSymOutsideInclusiveRange( 2613 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 2614 const llvm::APSInt &To, const llvm::APSInt &Adjustment) { 2615 RangeSet RangeLT = getSymLTRange(State, Sym, From, Adjustment); 2616 RangeSet RangeGT = getSymGTRange(State, Sym, To, Adjustment); 2617 RangeSet New(F.add(RangeLT, RangeGT)); 2618 return setRange(State, Sym, New); 2619 } 2620 2621 //===----------------------------------------------------------------------===// 2622 // Pretty-printing. 2623 //===----------------------------------------------------------------------===// 2624 2625 void RangeConstraintManager::printJson(raw_ostream &Out, ProgramStateRef State, 2626 const char *NL, unsigned int Space, 2627 bool IsDot) const { 2628 printConstraints(Out, State, NL, Space, IsDot); 2629 printEquivalenceClasses(Out, State, NL, Space, IsDot); 2630 printDisequalities(Out, State, NL, Space, IsDot); 2631 } 2632 2633 void RangeConstraintManager::printConstraints(raw_ostream &Out, 2634 ProgramStateRef State, 2635 const char *NL, 2636 unsigned int Space, 2637 bool IsDot) const { 2638 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 2639 2640 Indent(Out, Space, IsDot) << "\"constraints\": "; 2641 if (Constraints.isEmpty()) { 2642 Out << "null," << NL; 2643 return; 2644 } 2645 2646 ++Space; 2647 Out << '[' << NL; 2648 bool First = true; 2649 for (std::pair<EquivalenceClass, RangeSet> P : Constraints) { 2650 SymbolSet ClassMembers = P.first.getClassMembers(State); 2651 2652 // We can print the same constraint for every class member. 2653 for (SymbolRef ClassMember : ClassMembers) { 2654 if (First) { 2655 First = false; 2656 } else { 2657 Out << ','; 2658 Out << NL; 2659 } 2660 Indent(Out, Space, IsDot) 2661 << "{ \"symbol\": \"" << ClassMember << "\", \"range\": \""; 2662 P.second.dump(Out); 2663 Out << "\" }"; 2664 } 2665 } 2666 Out << NL; 2667 2668 --Space; 2669 Indent(Out, Space, IsDot) << "]," << NL; 2670 } 2671 2672 static std::string toString(const SymbolRef &Sym) { 2673 std::string S; 2674 llvm::raw_string_ostream O(S); 2675 Sym->dumpToStream(O); 2676 return O.str(); 2677 } 2678 2679 static std::string toString(ProgramStateRef State, EquivalenceClass Class) { 2680 SymbolSet ClassMembers = Class.getClassMembers(State); 2681 llvm::SmallVector<SymbolRef, 8> ClassMembersSorted(ClassMembers.begin(), 2682 ClassMembers.end()); 2683 llvm::sort(ClassMembersSorted, 2684 [](const SymbolRef &LHS, const SymbolRef &RHS) { 2685 return toString(LHS) < toString(RHS); 2686 }); 2687 2688 bool FirstMember = true; 2689 2690 std::string Str; 2691 llvm::raw_string_ostream Out(Str); 2692 Out << "[ "; 2693 for (SymbolRef ClassMember : ClassMembersSorted) { 2694 if (FirstMember) 2695 FirstMember = false; 2696 else 2697 Out << ", "; 2698 Out << "\"" << ClassMember << "\""; 2699 } 2700 Out << " ]"; 2701 return Out.str(); 2702 } 2703 2704 void RangeConstraintManager::printEquivalenceClasses(raw_ostream &Out, 2705 ProgramStateRef State, 2706 const char *NL, 2707 unsigned int Space, 2708 bool IsDot) const { 2709 ClassMembersTy Members = State->get<ClassMembers>(); 2710 2711 Indent(Out, Space, IsDot) << "\"equivalence_classes\": "; 2712 if (Members.isEmpty()) { 2713 Out << "null," << NL; 2714 return; 2715 } 2716 2717 std::set<std::string> MembersStr; 2718 for (std::pair<EquivalenceClass, SymbolSet> ClassToSymbolSet : Members) 2719 MembersStr.insert(toString(State, ClassToSymbolSet.first)); 2720 2721 ++Space; 2722 Out << '[' << NL; 2723 bool FirstClass = true; 2724 for (const std::string &Str : MembersStr) { 2725 if (FirstClass) { 2726 FirstClass = false; 2727 } else { 2728 Out << ','; 2729 Out << NL; 2730 } 2731 Indent(Out, Space, IsDot); 2732 Out << Str; 2733 } 2734 Out << NL; 2735 2736 --Space; 2737 Indent(Out, Space, IsDot) << "]," << NL; 2738 } 2739 2740 void RangeConstraintManager::printDisequalities(raw_ostream &Out, 2741 ProgramStateRef State, 2742 const char *NL, 2743 unsigned int Space, 2744 bool IsDot) const { 2745 DisequalityMapTy Disequalities = State->get<DisequalityMap>(); 2746 2747 Indent(Out, Space, IsDot) << "\"disequality_info\": "; 2748 if (Disequalities.isEmpty()) { 2749 Out << "null," << NL; 2750 return; 2751 } 2752 2753 // Transform the disequality info to an ordered map of 2754 // [string -> (ordered set of strings)] 2755 using EqClassesStrTy = std::set<std::string>; 2756 using DisequalityInfoStrTy = std::map<std::string, EqClassesStrTy>; 2757 DisequalityInfoStrTy DisequalityInfoStr; 2758 for (std::pair<EquivalenceClass, ClassSet> ClassToDisEqSet : Disequalities) { 2759 EquivalenceClass Class = ClassToDisEqSet.first; 2760 ClassSet DisequalClasses = ClassToDisEqSet.second; 2761 EqClassesStrTy MembersStr; 2762 for (EquivalenceClass DisEqClass : DisequalClasses) 2763 MembersStr.insert(toString(State, DisEqClass)); 2764 DisequalityInfoStr.insert({toString(State, Class), MembersStr}); 2765 } 2766 2767 ++Space; 2768 Out << '[' << NL; 2769 bool FirstClass = true; 2770 for (std::pair<std::string, EqClassesStrTy> ClassToDisEqSet : 2771 DisequalityInfoStr) { 2772 const std::string &Class = ClassToDisEqSet.first; 2773 if (FirstClass) { 2774 FirstClass = false; 2775 } else { 2776 Out << ','; 2777 Out << NL; 2778 } 2779 Indent(Out, Space, IsDot) << "{" << NL; 2780 unsigned int DisEqSpace = Space + 1; 2781 Indent(Out, DisEqSpace, IsDot) << "\"class\": "; 2782 Out << Class; 2783 const EqClassesStrTy &DisequalClasses = ClassToDisEqSet.second; 2784 if (!DisequalClasses.empty()) { 2785 Out << "," << NL; 2786 Indent(Out, DisEqSpace, IsDot) << "\"disequal_to\": [" << NL; 2787 unsigned int DisEqClassSpace = DisEqSpace + 1; 2788 Indent(Out, DisEqClassSpace, IsDot); 2789 bool FirstDisEqClass = true; 2790 for (const std::string &DisEqClass : DisequalClasses) { 2791 if (FirstDisEqClass) { 2792 FirstDisEqClass = false; 2793 } else { 2794 Out << ',' << NL; 2795 Indent(Out, DisEqClassSpace, IsDot); 2796 } 2797 Out << DisEqClass; 2798 } 2799 Out << "]" << NL; 2800 } 2801 Indent(Out, Space, IsDot) << "}"; 2802 } 2803 Out << NL; 2804 2805 --Space; 2806 Indent(Out, Space, IsDot) << "]," << NL; 2807 } 2808