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 /// Return true if the current class is trivial in the given state. 563 LLVM_NODISCARD inline bool isTrivial(ProgramStateRef State) const; 564 /// Return true if the current class is trivial and its only member is dead. 565 LLVM_NODISCARD inline bool isTriviallyDead(ProgramStateRef State, 566 SymbolReaper &Reaper) const; 567 568 LLVM_NODISCARD static inline ProgramStateRef 569 markDisequal(RangeSet::Factory &F, ProgramStateRef State, SymbolRef First, 570 SymbolRef Second); 571 LLVM_NODISCARD static inline ProgramStateRef 572 markDisequal(RangeSet::Factory &F, ProgramStateRef State, 573 EquivalenceClass First, EquivalenceClass Second); 574 LLVM_NODISCARD inline ProgramStateRef 575 markDisequal(RangeSet::Factory &F, ProgramStateRef State, 576 EquivalenceClass Other) const; 577 LLVM_NODISCARD static inline ClassSet 578 getDisequalClasses(ProgramStateRef State, SymbolRef Sym); 579 LLVM_NODISCARD inline ClassSet 580 getDisequalClasses(ProgramStateRef State) const; 581 LLVM_NODISCARD inline ClassSet 582 getDisequalClasses(DisequalityMapTy Map, ClassSet::Factory &Factory) const; 583 584 LLVM_NODISCARD static inline Optional<bool> areEqual(ProgramStateRef State, 585 EquivalenceClass First, 586 EquivalenceClass Second); 587 LLVM_NODISCARD static inline Optional<bool> 588 areEqual(ProgramStateRef State, SymbolRef First, SymbolRef Second); 589 590 /// Iterate over all symbols and try to simplify them. 591 LLVM_NODISCARD ProgramStateRef simplify(SValBuilder &SVB, 592 RangeSet::Factory &F, 593 ProgramStateRef State); 594 595 /// Check equivalence data for consistency. 596 LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED static bool 597 isClassDataConsistent(ProgramStateRef State); 598 599 LLVM_NODISCARD QualType getType() const { 600 return getRepresentativeSymbol()->getType(); 601 } 602 603 EquivalenceClass() = delete; 604 EquivalenceClass(const EquivalenceClass &) = default; 605 EquivalenceClass &operator=(const EquivalenceClass &) = delete; 606 EquivalenceClass(EquivalenceClass &&) = default; 607 EquivalenceClass &operator=(EquivalenceClass &&) = delete; 608 609 bool operator==(const EquivalenceClass &Other) const { 610 return ID == Other.ID; 611 } 612 bool operator<(const EquivalenceClass &Other) const { return ID < Other.ID; } 613 bool operator!=(const EquivalenceClass &Other) const { 614 return !operator==(Other); 615 } 616 617 static void Profile(llvm::FoldingSetNodeID &ID, uintptr_t CID) { 618 ID.AddInteger(CID); 619 } 620 621 void Profile(llvm::FoldingSetNodeID &ID) const { Profile(ID, this->ID); } 622 623 private: 624 /* implicit */ EquivalenceClass(SymbolRef Sym) 625 : ID(reinterpret_cast<uintptr_t>(Sym)) {} 626 627 /// This function is intended to be used ONLY within the class. 628 /// The fact that ID is a pointer to a symbol is an implementation detail 629 /// and should stay that way. 630 /// In the current implementation, we use it to retrieve the only member 631 /// of the trivial class. 632 SymbolRef getRepresentativeSymbol() const { 633 return reinterpret_cast<SymbolRef>(ID); 634 } 635 static inline SymbolSet::Factory &getMembersFactory(ProgramStateRef State); 636 637 inline ProgramStateRef mergeImpl(RangeSet::Factory &F, ProgramStateRef State, 638 SymbolSet Members, EquivalenceClass Other, 639 SymbolSet OtherMembers); 640 static inline bool 641 addToDisequalityInfo(DisequalityMapTy &Info, ConstraintRangeTy &Constraints, 642 RangeSet::Factory &F, ProgramStateRef State, 643 EquivalenceClass First, EquivalenceClass Second); 644 645 /// This is a unique identifier of the class. 646 uintptr_t ID; 647 }; 648 649 //===----------------------------------------------------------------------===// 650 // Constraint functions 651 //===----------------------------------------------------------------------===// 652 653 LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED bool 654 areFeasible(ConstraintRangeTy Constraints) { 655 return llvm::none_of( 656 Constraints, 657 [](const std::pair<EquivalenceClass, RangeSet> &ClassConstraint) { 658 return ClassConstraint.second.isEmpty(); 659 }); 660 } 661 662 LLVM_NODISCARD inline const RangeSet *getConstraint(ProgramStateRef State, 663 EquivalenceClass Class) { 664 return State->get<ConstraintRange>(Class); 665 } 666 667 LLVM_NODISCARD inline const RangeSet *getConstraint(ProgramStateRef State, 668 SymbolRef Sym) { 669 return getConstraint(State, EquivalenceClass::find(State, Sym)); 670 } 671 672 //===----------------------------------------------------------------------===// 673 // Equality/diseqiality abstraction 674 //===----------------------------------------------------------------------===// 675 676 /// A small helper structure representing symbolic equality. 677 /// 678 /// Equality check can have different forms (like a == b or a - b) and this 679 /// class encapsulates those away if the only thing the user wants to check - 680 /// whether it's equality/diseqiality or not and have an easy access to the 681 /// compared symbols. 682 struct EqualityInfo { 683 public: 684 SymbolRef Left, Right; 685 // true for equality and false for disequality. 686 bool IsEquality = true; 687 688 void invert() { IsEquality = !IsEquality; } 689 /// Extract equality information from the given symbol and the constants. 690 /// 691 /// This function assumes the following expression Sym + Adjustment != Int. 692 /// It is a default because the most widespread case of the equality check 693 /// is (A == B) + 0 != 0. 694 static Optional<EqualityInfo> extract(SymbolRef Sym, const llvm::APSInt &Int, 695 const llvm::APSInt &Adjustment) { 696 // As of now, the only equality form supported is Sym + 0 != 0. 697 if (!Int.isNullValue() || !Adjustment.isNullValue()) 698 return llvm::None; 699 700 return extract(Sym); 701 } 702 /// Extract equality information from the given symbol. 703 static Optional<EqualityInfo> extract(SymbolRef Sym) { 704 return EqualityExtractor().Visit(Sym); 705 } 706 707 private: 708 class EqualityExtractor 709 : public SymExprVisitor<EqualityExtractor, Optional<EqualityInfo>> { 710 public: 711 Optional<EqualityInfo> VisitSymSymExpr(const SymSymExpr *Sym) const { 712 switch (Sym->getOpcode()) { 713 case BO_Sub: 714 // This case is: A - B != 0 -> disequality check. 715 return EqualityInfo{Sym->getLHS(), Sym->getRHS(), false}; 716 case BO_EQ: 717 // This case is: A == B != 0 -> equality check. 718 return EqualityInfo{Sym->getLHS(), Sym->getRHS(), true}; 719 case BO_NE: 720 // This case is: A != B != 0 -> diseqiality check. 721 return EqualityInfo{Sym->getLHS(), Sym->getRHS(), false}; 722 default: 723 return llvm::None; 724 } 725 } 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 VisitBinaryOperator(Sym); 859 } 860 861 private: 862 SymbolicRangeInferrer(RangeSet::Factory &F, ProgramStateRef S) 863 : ValueFactory(F.getValueFactory()), RangeFactory(F), State(S) {} 864 865 /// Infer range information from the given integer constant. 866 /// 867 /// It's not a real "inference", but is here for operating with 868 /// sub-expressions in a more polymorphic manner. 869 RangeSet inferAs(const llvm::APSInt &Val, QualType) { 870 return {RangeFactory, Val}; 871 } 872 873 /// Infer range information from symbol in the context of the given type. 874 RangeSet inferAs(SymbolRef Sym, QualType DestType) { 875 QualType ActualType = Sym->getType(); 876 // Check that we can reason about the symbol at all. 877 if (ActualType->isIntegralOrEnumerationType() || 878 Loc::isLocType(ActualType)) { 879 return infer(Sym); 880 } 881 // Otherwise, let's simply infer from the destination type. 882 // We couldn't figure out nothing else about that expression. 883 return infer(DestType); 884 } 885 886 RangeSet infer(SymbolRef Sym) { 887 return intersect( 888 RangeFactory, 889 // Of course, we should take the constraint directly associated with 890 // this symbol into consideration. 891 getConstraint(State, Sym), 892 // If Sym is a difference of symbols A - B, then maybe we have range 893 // set stored for B - A. 894 // 895 // If we have range set stored for both A - B and B - A then 896 // calculate the effective range set by intersecting the range set 897 // for A - B and the negated range set of B - A. 898 getRangeForNegatedSub(Sym), 899 // If Sym is (dis)equality, we might have some information on that 900 // in our equality classes data structure. 901 getRangeForEqualities(Sym), 902 // If Sym is a comparison expression (except <=>), 903 // find any other comparisons with the same operands. 904 // See function description. 905 getRangeForComparisonSymbol(Sym), 906 // Apart from the Sym itself, we can infer quite a lot if we look 907 // into subexpressions of Sym. 908 Visit(Sym)); 909 } 910 911 RangeSet infer(EquivalenceClass Class) { 912 if (const RangeSet *AssociatedConstraint = getConstraint(State, Class)) 913 return *AssociatedConstraint; 914 915 return infer(Class.getType()); 916 } 917 918 /// Infer range information solely from the type. 919 RangeSet infer(QualType T) { 920 // Lazily generate a new RangeSet representing all possible values for the 921 // given symbol type. 922 RangeSet Result(RangeFactory, ValueFactory.getMinValue(T), 923 ValueFactory.getMaxValue(T)); 924 925 // References are known to be non-zero. 926 if (T->isReferenceType()) 927 return assumeNonZero(Result, T); 928 929 return Result; 930 } 931 932 template <class BinarySymExprTy> 933 RangeSet VisitBinaryOperator(const BinarySymExprTy *Sym) { 934 // TODO #1: VisitBinaryOperator implementation might not make a good 935 // use of the inferred ranges. In this case, we might be calculating 936 // everything for nothing. This being said, we should introduce some 937 // sort of laziness mechanism here. 938 // 939 // TODO #2: We didn't go into the nested expressions before, so it 940 // might cause us spending much more time doing the inference. 941 // This can be a problem for deeply nested expressions that are 942 // involved in conditions and get tested continuously. We definitely 943 // need to address this issue and introduce some sort of caching 944 // in here. 945 QualType ResultType = Sym->getType(); 946 return VisitBinaryOperator(inferAs(Sym->getLHS(), ResultType), 947 Sym->getOpcode(), 948 inferAs(Sym->getRHS(), ResultType), ResultType); 949 } 950 951 RangeSet VisitBinaryOperator(RangeSet LHS, BinaryOperator::Opcode Op, 952 RangeSet RHS, QualType T) { 953 switch (Op) { 954 case BO_Or: 955 return VisitBinaryOperator<BO_Or>(LHS, RHS, T); 956 case BO_And: 957 return VisitBinaryOperator<BO_And>(LHS, RHS, T); 958 case BO_Rem: 959 return VisitBinaryOperator<BO_Rem>(LHS, RHS, T); 960 default: 961 return infer(T); 962 } 963 } 964 965 //===----------------------------------------------------------------------===// 966 // Ranges and operators 967 //===----------------------------------------------------------------------===// 968 969 /// Return a rough approximation of the given range set. 970 /// 971 /// For the range set: 972 /// { [x_0, y_0], [x_1, y_1], ... , [x_N, y_N] } 973 /// it will return the range [x_0, y_N]. 974 static Range fillGaps(RangeSet Origin) { 975 assert(!Origin.isEmpty()); 976 return {Origin.getMinValue(), Origin.getMaxValue()}; 977 } 978 979 /// Try to convert given range into the given type. 980 /// 981 /// It will return llvm::None only when the trivial conversion is possible. 982 llvm::Optional<Range> convert(const Range &Origin, APSIntType To) { 983 if (To.testInRange(Origin.From(), false) != APSIntType::RTR_Within || 984 To.testInRange(Origin.To(), false) != APSIntType::RTR_Within) { 985 return llvm::None; 986 } 987 return Range(ValueFactory.Convert(To, Origin.From()), 988 ValueFactory.Convert(To, Origin.To())); 989 } 990 991 template <BinaryOperator::Opcode Op> 992 RangeSet VisitBinaryOperator(RangeSet LHS, RangeSet RHS, QualType T) { 993 // We should propagate information about unfeasbility of one of the 994 // operands to the resulting range. 995 if (LHS.isEmpty() || RHS.isEmpty()) { 996 return RangeFactory.getEmptySet(); 997 } 998 999 Range CoarseLHS = fillGaps(LHS); 1000 Range CoarseRHS = fillGaps(RHS); 1001 1002 APSIntType ResultType = ValueFactory.getAPSIntType(T); 1003 1004 // We need to convert ranges to the resulting type, so we can compare values 1005 // and combine them in a meaningful (in terms of the given operation) way. 1006 auto ConvertedCoarseLHS = convert(CoarseLHS, ResultType); 1007 auto ConvertedCoarseRHS = convert(CoarseRHS, ResultType); 1008 1009 // It is hard to reason about ranges when conversion changes 1010 // borders of the ranges. 1011 if (!ConvertedCoarseLHS || !ConvertedCoarseRHS) { 1012 return infer(T); 1013 } 1014 1015 return VisitBinaryOperator<Op>(*ConvertedCoarseLHS, *ConvertedCoarseRHS, T); 1016 } 1017 1018 template <BinaryOperator::Opcode Op> 1019 RangeSet VisitBinaryOperator(Range LHS, Range RHS, QualType T) { 1020 return infer(T); 1021 } 1022 1023 /// Return a symmetrical range for the given range and type. 1024 /// 1025 /// If T is signed, return the smallest range [-x..x] that covers the original 1026 /// range, or [-min(T), max(T)] if the aforementioned symmetric range doesn't 1027 /// exist due to original range covering min(T)). 1028 /// 1029 /// If T is unsigned, return the smallest range [0..x] that covers the 1030 /// original range. 1031 Range getSymmetricalRange(Range Origin, QualType T) { 1032 APSIntType RangeType = ValueFactory.getAPSIntType(T); 1033 1034 if (RangeType.isUnsigned()) { 1035 return Range(ValueFactory.getMinValue(RangeType), Origin.To()); 1036 } 1037 1038 if (Origin.From().isMinSignedValue()) { 1039 // If mini is a minimal signed value, absolute value of it is greater 1040 // than the maximal signed value. In order to avoid these 1041 // complications, we simply return the whole range. 1042 return {ValueFactory.getMinValue(RangeType), 1043 ValueFactory.getMaxValue(RangeType)}; 1044 } 1045 1046 // At this point, we are sure that the type is signed and we can safely 1047 // use unary - operator. 1048 // 1049 // While calculating absolute maximum, we can use the following formula 1050 // because of these reasons: 1051 // * If From >= 0 then To >= From and To >= -From. 1052 // AbsMax == To == max(To, -From) 1053 // * If To <= 0 then -From >= -To and -From >= From. 1054 // AbsMax == -From == max(-From, To) 1055 // * Otherwise, From <= 0, To >= 0, and 1056 // AbsMax == max(abs(From), abs(To)) 1057 llvm::APSInt AbsMax = std::max(-Origin.From(), Origin.To()); 1058 1059 // Intersection is guaranteed to be non-empty. 1060 return {ValueFactory.getValue(-AbsMax), ValueFactory.getValue(AbsMax)}; 1061 } 1062 1063 /// Return a range set subtracting zero from \p Domain. 1064 RangeSet assumeNonZero(RangeSet Domain, QualType T) { 1065 APSIntType IntType = ValueFactory.getAPSIntType(T); 1066 return RangeFactory.deletePoint(Domain, IntType.getZeroValue()); 1067 } 1068 1069 // FIXME: Once SValBuilder supports unary minus, we should use SValBuilder to 1070 // obtain the negated symbolic expression instead of constructing the 1071 // symbol manually. This will allow us to support finding ranges of not 1072 // only negated SymSymExpr-type expressions, but also of other, simpler 1073 // expressions which we currently do not know how to negate. 1074 Optional<RangeSet> getRangeForNegatedSub(SymbolRef Sym) { 1075 if (const SymSymExpr *SSE = dyn_cast<SymSymExpr>(Sym)) { 1076 if (SSE->getOpcode() == BO_Sub) { 1077 QualType T = Sym->getType(); 1078 1079 // Do not negate unsigned ranges 1080 if (!T->isUnsignedIntegerOrEnumerationType() && 1081 !T->isSignedIntegerOrEnumerationType()) 1082 return llvm::None; 1083 1084 SymbolManager &SymMgr = State->getSymbolManager(); 1085 SymbolRef NegatedSym = 1086 SymMgr.getSymSymExpr(SSE->getRHS(), BO_Sub, SSE->getLHS(), T); 1087 1088 if (const RangeSet *NegatedRange = getConstraint(State, NegatedSym)) { 1089 return RangeFactory.negate(*NegatedRange); 1090 } 1091 } 1092 } 1093 return llvm::None; 1094 } 1095 1096 // Returns ranges only for binary comparison operators (except <=>) 1097 // when left and right operands are symbolic values. 1098 // Finds any other comparisons with the same operands. 1099 // Then do logical calculations and refuse impossible branches. 1100 // E.g. (x < y) and (x > y) at the same time are impossible. 1101 // E.g. (x >= y) and (x != y) at the same time makes (x > y) true only. 1102 // E.g. (x == y) and (y == x) are just reversed but the same. 1103 // It covers all possible combinations (see CmpOpTable description). 1104 // Note that `x` and `y` can also stand for subexpressions, 1105 // not only for actual symbols. 1106 Optional<RangeSet> getRangeForComparisonSymbol(SymbolRef Sym) { 1107 const auto *SSE = dyn_cast<SymSymExpr>(Sym); 1108 if (!SSE) 1109 return llvm::None; 1110 1111 BinaryOperatorKind CurrentOP = SSE->getOpcode(); 1112 1113 // We currently do not support <=> (C++20). 1114 if (!BinaryOperator::isComparisonOp(CurrentOP) || (CurrentOP == BO_Cmp)) 1115 return llvm::None; 1116 1117 static const OperatorRelationsTable CmpOpTable{}; 1118 1119 const SymExpr *LHS = SSE->getLHS(); 1120 const SymExpr *RHS = SSE->getRHS(); 1121 QualType T = SSE->getType(); 1122 1123 SymbolManager &SymMgr = State->getSymbolManager(); 1124 1125 int UnknownStates = 0; 1126 1127 // Loop goes through all of the columns exept the last one ('UnknownX2'). 1128 // We treat `UnknownX2` column separately at the end of the loop body. 1129 for (size_t i = 0; i < CmpOpTable.getCmpOpCount(); ++i) { 1130 1131 // Let's find an expression e.g. (x < y). 1132 BinaryOperatorKind QueriedOP = OperatorRelationsTable::getOpFromIndex(i); 1133 const SymSymExpr *SymSym = SymMgr.getSymSymExpr(LHS, QueriedOP, RHS, T); 1134 const RangeSet *QueriedRangeSet = getConstraint(State, SymSym); 1135 1136 // If ranges were not previously found, 1137 // try to find a reversed expression (y > x). 1138 if (!QueriedRangeSet) { 1139 const BinaryOperatorKind ROP = 1140 BinaryOperator::reverseComparisonOp(QueriedOP); 1141 SymSym = SymMgr.getSymSymExpr(RHS, ROP, LHS, T); 1142 QueriedRangeSet = getConstraint(State, SymSym); 1143 } 1144 1145 if (!QueriedRangeSet || QueriedRangeSet->isEmpty()) 1146 continue; 1147 1148 const llvm::APSInt *ConcreteValue = QueriedRangeSet->getConcreteValue(); 1149 const bool isInFalseBranch = 1150 ConcreteValue ? (*ConcreteValue == 0) : false; 1151 1152 // If it is a false branch, we shall be guided by opposite operator, 1153 // because the table is made assuming we are in the true branch. 1154 // E.g. when (x <= y) is false, then (x > y) is true. 1155 if (isInFalseBranch) 1156 QueriedOP = BinaryOperator::negateComparisonOp(QueriedOP); 1157 1158 OperatorRelationsTable::TriStateKind BranchState = 1159 CmpOpTable.getCmpOpState(CurrentOP, QueriedOP); 1160 1161 if (BranchState == OperatorRelationsTable::Unknown) { 1162 if (++UnknownStates == 2) 1163 // If we met both Unknown states. 1164 // if (x <= y) // assume true 1165 // if (x != y) // assume true 1166 // if (x < y) // would be also true 1167 // Get a state from `UnknownX2` column. 1168 BranchState = CmpOpTable.getCmpOpStateForUnknownX2(CurrentOP); 1169 else 1170 continue; 1171 } 1172 1173 return (BranchState == OperatorRelationsTable::True) ? getTrueRange(T) 1174 : getFalseRange(T); 1175 } 1176 1177 return llvm::None; 1178 } 1179 1180 Optional<RangeSet> getRangeForEqualities(SymbolRef Sym) { 1181 Optional<EqualityInfo> Equality = EqualityInfo::extract(Sym); 1182 1183 if (!Equality) 1184 return llvm::None; 1185 1186 if (Optional<bool> AreEqual = EquivalenceClass::areEqual( 1187 State, Equality->Left, Equality->Right)) { 1188 if (*AreEqual == Equality->IsEquality) { 1189 return getTrueRange(Sym->getType()); 1190 } 1191 return getFalseRange(Sym->getType()); 1192 } 1193 1194 return llvm::None; 1195 } 1196 1197 RangeSet getTrueRange(QualType T) { 1198 RangeSet TypeRange = infer(T); 1199 return assumeNonZero(TypeRange, T); 1200 } 1201 1202 RangeSet getFalseRange(QualType T) { 1203 const llvm::APSInt &Zero = ValueFactory.getValue(0, T); 1204 return RangeSet(RangeFactory, Zero); 1205 } 1206 1207 BasicValueFactory &ValueFactory; 1208 RangeSet::Factory &RangeFactory; 1209 ProgramStateRef State; 1210 }; 1211 1212 //===----------------------------------------------------------------------===// 1213 // Range-based reasoning about symbolic operations 1214 //===----------------------------------------------------------------------===// 1215 1216 template <> 1217 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_Or>(Range LHS, Range RHS, 1218 QualType T) { 1219 APSIntType ResultType = ValueFactory.getAPSIntType(T); 1220 llvm::APSInt Zero = ResultType.getZeroValue(); 1221 1222 bool IsLHSPositiveOrZero = LHS.From() >= Zero; 1223 bool IsRHSPositiveOrZero = RHS.From() >= Zero; 1224 1225 bool IsLHSNegative = LHS.To() < Zero; 1226 bool IsRHSNegative = RHS.To() < Zero; 1227 1228 // Check if both ranges have the same sign. 1229 if ((IsLHSPositiveOrZero && IsRHSPositiveOrZero) || 1230 (IsLHSNegative && IsRHSNegative)) { 1231 // The result is definitely greater or equal than any of the operands. 1232 const llvm::APSInt &Min = std::max(LHS.From(), RHS.From()); 1233 1234 // We estimate maximal value for positives as the maximal value for the 1235 // given type. For negatives, we estimate it with -1 (e.g. 0x11111111). 1236 // 1237 // TODO: We basically, limit the resulting range from below, but don't do 1238 // anything with the upper bound. 1239 // 1240 // For positive operands, it can be done as follows: for the upper 1241 // bound of LHS and RHS we calculate the most significant bit set. 1242 // Let's call it the N-th bit. Then we can estimate the maximal 1243 // number to be 2^(N+1)-1, i.e. the number with all the bits up to 1244 // the N-th bit set. 1245 const llvm::APSInt &Max = IsLHSNegative 1246 ? ValueFactory.getValue(--Zero) 1247 : ValueFactory.getMaxValue(ResultType); 1248 1249 return {RangeFactory, ValueFactory.getValue(Min), Max}; 1250 } 1251 1252 // Otherwise, let's check if at least one of the operands is negative. 1253 if (IsLHSNegative || IsRHSNegative) { 1254 // This means that the result is definitely negative as well. 1255 return {RangeFactory, ValueFactory.getMinValue(ResultType), 1256 ValueFactory.getValue(--Zero)}; 1257 } 1258 1259 RangeSet DefaultRange = infer(T); 1260 1261 // It is pretty hard to reason about operands with different signs 1262 // (and especially with possibly different signs). We simply check if it 1263 // can be zero. In order to conclude that the result could not be zero, 1264 // at least one of the operands should be definitely not zero itself. 1265 if (!LHS.Includes(Zero) || !RHS.Includes(Zero)) { 1266 return assumeNonZero(DefaultRange, T); 1267 } 1268 1269 // Nothing much else to do here. 1270 return DefaultRange; 1271 } 1272 1273 template <> 1274 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_And>(Range LHS, 1275 Range RHS, 1276 QualType T) { 1277 APSIntType ResultType = ValueFactory.getAPSIntType(T); 1278 llvm::APSInt Zero = ResultType.getZeroValue(); 1279 1280 bool IsLHSPositiveOrZero = LHS.From() >= Zero; 1281 bool IsRHSPositiveOrZero = RHS.From() >= Zero; 1282 1283 bool IsLHSNegative = LHS.To() < Zero; 1284 bool IsRHSNegative = RHS.To() < Zero; 1285 1286 // Check if both ranges have the same sign. 1287 if ((IsLHSPositiveOrZero && IsRHSPositiveOrZero) || 1288 (IsLHSNegative && IsRHSNegative)) { 1289 // The result is definitely less or equal than any of the operands. 1290 const llvm::APSInt &Max = std::min(LHS.To(), RHS.To()); 1291 1292 // We conservatively estimate lower bound to be the smallest positive 1293 // or negative value corresponding to the sign of the operands. 1294 const llvm::APSInt &Min = IsLHSNegative 1295 ? ValueFactory.getMinValue(ResultType) 1296 : ValueFactory.getValue(Zero); 1297 1298 return {RangeFactory, Min, Max}; 1299 } 1300 1301 // Otherwise, let's check if at least one of the operands is positive. 1302 if (IsLHSPositiveOrZero || IsRHSPositiveOrZero) { 1303 // This makes result definitely positive. 1304 // 1305 // We can also reason about a maximal value by finding the maximal 1306 // value of the positive operand. 1307 const llvm::APSInt &Max = IsLHSPositiveOrZero ? LHS.To() : RHS.To(); 1308 1309 // The minimal value on the other hand is much harder to reason about. 1310 // The only thing we know for sure is that the result is positive. 1311 return {RangeFactory, ValueFactory.getValue(Zero), 1312 ValueFactory.getValue(Max)}; 1313 } 1314 1315 // Nothing much else to do here. 1316 return infer(T); 1317 } 1318 1319 template <> 1320 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_Rem>(Range LHS, 1321 Range RHS, 1322 QualType T) { 1323 llvm::APSInt Zero = ValueFactory.getAPSIntType(T).getZeroValue(); 1324 1325 Range ConservativeRange = getSymmetricalRange(RHS, T); 1326 1327 llvm::APSInt Max = ConservativeRange.To(); 1328 llvm::APSInt Min = ConservativeRange.From(); 1329 1330 if (Max == Zero) { 1331 // It's an undefined behaviour to divide by 0 and it seems like we know 1332 // for sure that RHS is 0. Let's say that the resulting range is 1333 // simply infeasible for that matter. 1334 return RangeFactory.getEmptySet(); 1335 } 1336 1337 // At this point, our conservative range is closed. The result, however, 1338 // couldn't be greater than the RHS' maximal absolute value. Because of 1339 // this reason, we turn the range into open (or half-open in case of 1340 // unsigned integers). 1341 // 1342 // While we operate on integer values, an open interval (a, b) can be easily 1343 // represented by the closed interval [a + 1, b - 1]. And this is exactly 1344 // what we do next. 1345 // 1346 // If we are dealing with unsigned case, we shouldn't move the lower bound. 1347 if (Min.isSigned()) { 1348 ++Min; 1349 } 1350 --Max; 1351 1352 bool IsLHSPositiveOrZero = LHS.From() >= Zero; 1353 bool IsRHSPositiveOrZero = RHS.From() >= Zero; 1354 1355 // Remainder operator results with negative operands is implementation 1356 // defined. Positive cases are much easier to reason about though. 1357 if (IsLHSPositiveOrZero && IsRHSPositiveOrZero) { 1358 // If maximal value of LHS is less than maximal value of RHS, 1359 // the result won't get greater than LHS.To(). 1360 Max = std::min(LHS.To(), Max); 1361 // We want to check if it is a situation similar to the following: 1362 // 1363 // <------------|---[ LHS ]--------[ RHS ]-----> 1364 // -INF 0 +INF 1365 // 1366 // In this situation, we can conclude that (LHS / RHS) == 0 and 1367 // (LHS % RHS) == LHS. 1368 Min = LHS.To() < RHS.From() ? LHS.From() : Zero; 1369 } 1370 1371 // Nevertheless, the symmetrical range for RHS is a conservative estimate 1372 // for any sign of either LHS, or RHS. 1373 return {RangeFactory, ValueFactory.getValue(Min), ValueFactory.getValue(Max)}; 1374 } 1375 1376 //===----------------------------------------------------------------------===// 1377 // Constraint manager implementation details 1378 //===----------------------------------------------------------------------===// 1379 1380 class RangeConstraintManager : public RangedConstraintManager { 1381 public: 1382 RangeConstraintManager(ExprEngine *EE, SValBuilder &SVB) 1383 : RangedConstraintManager(EE, SVB), F(getBasicVals()) {} 1384 1385 //===------------------------------------------------------------------===// 1386 // Implementation for interface from ConstraintManager. 1387 //===------------------------------------------------------------------===// 1388 1389 bool haveEqualConstraints(ProgramStateRef S1, 1390 ProgramStateRef S2) const override { 1391 // NOTE: ClassMembers are as simple as back pointers for ClassMap, 1392 // so comparing constraint ranges and class maps should be 1393 // sufficient. 1394 return S1->get<ConstraintRange>() == S2->get<ConstraintRange>() && 1395 S1->get<ClassMap>() == S2->get<ClassMap>(); 1396 } 1397 1398 bool canReasonAbout(SVal X) const override; 1399 1400 ConditionTruthVal checkNull(ProgramStateRef State, SymbolRef Sym) override; 1401 1402 const llvm::APSInt *getSymVal(ProgramStateRef State, 1403 SymbolRef Sym) const override; 1404 1405 ProgramStateRef removeDeadBindings(ProgramStateRef State, 1406 SymbolReaper &SymReaper) override; 1407 1408 void printJson(raw_ostream &Out, ProgramStateRef State, const char *NL = "\n", 1409 unsigned int Space = 0, bool IsDot = false) const override; 1410 1411 //===------------------------------------------------------------------===// 1412 // Implementation for interface from RangedConstraintManager. 1413 //===------------------------------------------------------------------===// 1414 1415 ProgramStateRef assumeSymNE(ProgramStateRef State, SymbolRef Sym, 1416 const llvm::APSInt &V, 1417 const llvm::APSInt &Adjustment) override; 1418 1419 ProgramStateRef assumeSymEQ(ProgramStateRef State, SymbolRef Sym, 1420 const llvm::APSInt &V, 1421 const llvm::APSInt &Adjustment) override; 1422 1423 ProgramStateRef assumeSymLT(ProgramStateRef State, SymbolRef Sym, 1424 const llvm::APSInt &V, 1425 const llvm::APSInt &Adjustment) override; 1426 1427 ProgramStateRef assumeSymGT(ProgramStateRef State, SymbolRef Sym, 1428 const llvm::APSInt &V, 1429 const llvm::APSInt &Adjustment) override; 1430 1431 ProgramStateRef assumeSymLE(ProgramStateRef State, SymbolRef Sym, 1432 const llvm::APSInt &V, 1433 const llvm::APSInt &Adjustment) override; 1434 1435 ProgramStateRef assumeSymGE(ProgramStateRef State, SymbolRef Sym, 1436 const llvm::APSInt &V, 1437 const llvm::APSInt &Adjustment) override; 1438 1439 ProgramStateRef assumeSymWithinInclusiveRange( 1440 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 1441 const llvm::APSInt &To, const llvm::APSInt &Adjustment) override; 1442 1443 ProgramStateRef assumeSymOutsideInclusiveRange( 1444 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 1445 const llvm::APSInt &To, const llvm::APSInt &Adjustment) override; 1446 1447 private: 1448 RangeSet::Factory F; 1449 1450 RangeSet getRange(ProgramStateRef State, SymbolRef Sym); 1451 RangeSet getRange(ProgramStateRef State, EquivalenceClass Class); 1452 1453 RangeSet getSymLTRange(ProgramStateRef St, SymbolRef Sym, 1454 const llvm::APSInt &Int, 1455 const llvm::APSInt &Adjustment); 1456 RangeSet getSymGTRange(ProgramStateRef St, SymbolRef Sym, 1457 const llvm::APSInt &Int, 1458 const llvm::APSInt &Adjustment); 1459 RangeSet getSymLERange(ProgramStateRef St, SymbolRef Sym, 1460 const llvm::APSInt &Int, 1461 const llvm::APSInt &Adjustment); 1462 RangeSet getSymLERange(llvm::function_ref<RangeSet()> RS, 1463 const llvm::APSInt &Int, 1464 const llvm::APSInt &Adjustment); 1465 RangeSet getSymGERange(ProgramStateRef St, SymbolRef Sym, 1466 const llvm::APSInt &Int, 1467 const llvm::APSInt &Adjustment); 1468 1469 //===------------------------------------------------------------------===// 1470 // Equality tracking implementation 1471 //===------------------------------------------------------------------===// 1472 1473 ProgramStateRef trackEQ(RangeSet NewConstraint, ProgramStateRef State, 1474 SymbolRef Sym, const llvm::APSInt &Int, 1475 const llvm::APSInt &Adjustment) { 1476 return track<true>(NewConstraint, State, Sym, Int, Adjustment); 1477 } 1478 1479 ProgramStateRef trackNE(RangeSet NewConstraint, ProgramStateRef State, 1480 SymbolRef Sym, const llvm::APSInt &Int, 1481 const llvm::APSInt &Adjustment) { 1482 return track<false>(NewConstraint, State, Sym, Int, Adjustment); 1483 } 1484 1485 template <bool EQ> 1486 ProgramStateRef track(RangeSet NewConstraint, ProgramStateRef State, 1487 SymbolRef Sym, const llvm::APSInt &Int, 1488 const llvm::APSInt &Adjustment) { 1489 if (NewConstraint.isEmpty()) 1490 // This is an infeasible assumption. 1491 return nullptr; 1492 1493 if (ProgramStateRef NewState = setConstraint(State, Sym, NewConstraint)) { 1494 if (auto Equality = EqualityInfo::extract(Sym, Int, Adjustment)) { 1495 // If the original assumption is not Sym + Adjustment !=/</> Int, 1496 // we should invert IsEquality flag. 1497 Equality->IsEquality = Equality->IsEquality != EQ; 1498 return track(NewState, *Equality); 1499 } 1500 1501 return NewState; 1502 } 1503 1504 return nullptr; 1505 } 1506 1507 ProgramStateRef track(ProgramStateRef State, EqualityInfo ToTrack) { 1508 if (ToTrack.IsEquality) { 1509 return trackEquality(State, ToTrack.Left, ToTrack.Right); 1510 } 1511 return trackDisequality(State, ToTrack.Left, ToTrack.Right); 1512 } 1513 1514 ProgramStateRef trackDisequality(ProgramStateRef State, SymbolRef LHS, 1515 SymbolRef RHS) { 1516 return EquivalenceClass::markDisequal(F, State, LHS, RHS); 1517 } 1518 1519 ProgramStateRef trackEquality(ProgramStateRef State, SymbolRef LHS, 1520 SymbolRef RHS) { 1521 return EquivalenceClass::merge(F, State, LHS, RHS); 1522 } 1523 1524 LLVM_NODISCARD ProgramStateRef setConstraint(ProgramStateRef State, 1525 EquivalenceClass Class, 1526 RangeSet Constraint) { 1527 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1528 ConstraintRangeTy::Factory &CF = State->get_context<ConstraintRange>(); 1529 1530 assert(!Constraint.isEmpty() && "New constraint should not be empty"); 1531 1532 // Add new constraint. 1533 Constraints = CF.add(Constraints, Class, Constraint); 1534 1535 // There is a chance that we might need to update constraints for the 1536 // classes that are known to be disequal to Class. 1537 // 1538 // In order for this to be even possible, the new constraint should 1539 // be simply a constant because we can't reason about range disequalities. 1540 if (const llvm::APSInt *Point = Constraint.getConcreteValue()) 1541 for (EquivalenceClass DisequalClass : Class.getDisequalClasses(State)) { 1542 RangeSet UpdatedConstraint = getRange(State, DisequalClass); 1543 UpdatedConstraint = F.deletePoint(UpdatedConstraint, *Point); 1544 1545 // If we end up with at least one of the disequal classes to be 1546 // constrained with an empty range-set, the state is infeasible. 1547 if (UpdatedConstraint.isEmpty()) 1548 return nullptr; 1549 1550 Constraints = CF.add(Constraints, DisequalClass, UpdatedConstraint); 1551 } 1552 1553 assert(areFeasible(Constraints) && "Constraint manager shouldn't produce " 1554 "a state with infeasible constraints"); 1555 1556 return State->set<ConstraintRange>(Constraints); 1557 } 1558 1559 // Associate a constraint to a symbolic expression. First, we set the 1560 // constraint in the State, then we try to simplify existing symbolic 1561 // expressions based on the newly set constraint. 1562 LLVM_NODISCARD inline ProgramStateRef 1563 setConstraint(ProgramStateRef State, SymbolRef Sym, RangeSet Constraint) { 1564 assert(State); 1565 1566 State = setConstraint(State, EquivalenceClass::find(State, Sym), Constraint); 1567 if (!State) 1568 return nullptr; 1569 1570 // We have a chance to simplify existing symbolic values if the new 1571 // constraint is a constant. 1572 if (!Constraint.getConcreteValue()) 1573 return State; 1574 1575 llvm::SmallSet<EquivalenceClass, 4> SimplifiedClasses; 1576 // Iterate over all equivalence classes and try to simplify them. 1577 ClassMembersTy Members = State->get<ClassMembers>(); 1578 for (std::pair<EquivalenceClass, SymbolSet> ClassToSymbolSet : Members) { 1579 EquivalenceClass Class = ClassToSymbolSet.first; 1580 State = Class.simplify(getSValBuilder(), F, State); 1581 if (!State) 1582 return nullptr; 1583 SimplifiedClasses.insert(Class); 1584 } 1585 1586 // Trivial equivalence classes (those that have only one symbol member) are 1587 // not stored in the State. Thus, we must skim through the constraints as 1588 // well. And we try to simplify symbols in the constraints. 1589 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1590 for (std::pair<EquivalenceClass, RangeSet> ClassConstraint : Constraints) { 1591 EquivalenceClass Class = ClassConstraint.first; 1592 if (SimplifiedClasses.count(Class)) // Already simplified. 1593 continue; 1594 State = Class.simplify(getSValBuilder(), F, State); 1595 if (!State) 1596 return nullptr; 1597 } 1598 1599 return State; 1600 } 1601 }; 1602 1603 } // end anonymous namespace 1604 1605 std::unique_ptr<ConstraintManager> 1606 ento::CreateRangeConstraintManager(ProgramStateManager &StMgr, 1607 ExprEngine *Eng) { 1608 return std::make_unique<RangeConstraintManager>(Eng, StMgr.getSValBuilder()); 1609 } 1610 1611 ConstraintMap ento::getConstraintMap(ProgramStateRef State) { 1612 ConstraintMap::Factory &F = State->get_context<ConstraintMap>(); 1613 ConstraintMap Result = F.getEmptyMap(); 1614 1615 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1616 for (std::pair<EquivalenceClass, RangeSet> ClassConstraint : Constraints) { 1617 EquivalenceClass Class = ClassConstraint.first; 1618 SymbolSet ClassMembers = Class.getClassMembers(State); 1619 assert(!ClassMembers.isEmpty() && 1620 "Class must always have at least one member!"); 1621 1622 SymbolRef Representative = *ClassMembers.begin(); 1623 Result = F.add(Result, Representative, ClassConstraint.second); 1624 } 1625 1626 return Result; 1627 } 1628 1629 //===----------------------------------------------------------------------===// 1630 // EqualityClass implementation details 1631 //===----------------------------------------------------------------------===// 1632 1633 inline EquivalenceClass EquivalenceClass::find(ProgramStateRef State, 1634 SymbolRef Sym) { 1635 assert(State && "State should not be null"); 1636 assert(Sym && "Symbol should not be null"); 1637 // We store far from all Symbol -> Class mappings 1638 if (const EquivalenceClass *NontrivialClass = State->get<ClassMap>(Sym)) 1639 return *NontrivialClass; 1640 1641 // This is a trivial class of Sym. 1642 return Sym; 1643 } 1644 1645 inline ProgramStateRef EquivalenceClass::merge(RangeSet::Factory &F, 1646 ProgramStateRef State, 1647 SymbolRef First, 1648 SymbolRef Second) { 1649 EquivalenceClass FirstClass = find(State, First); 1650 EquivalenceClass SecondClass = find(State, Second); 1651 1652 return FirstClass.merge(F, State, SecondClass); 1653 } 1654 1655 inline ProgramStateRef EquivalenceClass::merge(RangeSet::Factory &F, 1656 ProgramStateRef State, 1657 EquivalenceClass Other) { 1658 // It is already the same class. 1659 if (*this == Other) 1660 return State; 1661 1662 // FIXME: As of now, we support only equivalence classes of the same type. 1663 // This limitation is connected to the lack of explicit casts in 1664 // our symbolic expression model. 1665 // 1666 // That means that for `int x` and `char y` we don't distinguish 1667 // between these two very different cases: 1668 // * `x == y` 1669 // * `(char)x == y` 1670 // 1671 // The moment we introduce symbolic casts, this restriction can be 1672 // lifted. 1673 if (getType() != Other.getType()) 1674 return State; 1675 1676 SymbolSet Members = getClassMembers(State); 1677 SymbolSet OtherMembers = Other.getClassMembers(State); 1678 1679 // We estimate the size of the class by the height of tree containing 1680 // its members. Merging is not a trivial operation, so it's easier to 1681 // merge the smaller class into the bigger one. 1682 if (Members.getHeight() >= OtherMembers.getHeight()) { 1683 return mergeImpl(F, State, Members, Other, OtherMembers); 1684 } else { 1685 return Other.mergeImpl(F, State, OtherMembers, *this, Members); 1686 } 1687 } 1688 1689 inline ProgramStateRef 1690 EquivalenceClass::mergeImpl(RangeSet::Factory &RangeFactory, 1691 ProgramStateRef State, SymbolSet MyMembers, 1692 EquivalenceClass Other, SymbolSet OtherMembers) { 1693 // Essentially what we try to recreate here is some kind of union-find 1694 // data structure. It does have certain limitations due to persistence 1695 // and the need to remove elements from classes. 1696 // 1697 // In this setting, EquialityClass object is the representative of the class 1698 // or the parent element. ClassMap is a mapping of class members to their 1699 // parent. Unlike the union-find structure, they all point directly to the 1700 // class representative because we don't have an opportunity to actually do 1701 // path compression when dealing with immutability. This means that we 1702 // compress paths every time we do merges. It also means that we lose 1703 // the main amortized complexity benefit from the original data structure. 1704 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1705 ConstraintRangeTy::Factory &CRF = State->get_context<ConstraintRange>(); 1706 1707 // 1. If the merged classes have any constraints associated with them, we 1708 // need to transfer them to the class we have left. 1709 // 1710 // Intersection here makes perfect sense because both of these constraints 1711 // must hold for the whole new class. 1712 if (Optional<RangeSet> NewClassConstraint = 1713 intersect(RangeFactory, getConstraint(State, *this), 1714 getConstraint(State, Other))) { 1715 // NOTE: Essentially, NewClassConstraint should NEVER be infeasible because 1716 // range inferrer shouldn't generate ranges incompatible with 1717 // equivalence classes. However, at the moment, due to imperfections 1718 // in the solver, it is possible and the merge function can also 1719 // return infeasible states aka null states. 1720 if (NewClassConstraint->isEmpty()) 1721 // Infeasible state 1722 return nullptr; 1723 1724 // No need in tracking constraints of a now-dissolved class. 1725 Constraints = CRF.remove(Constraints, Other); 1726 // Assign new constraints for this class. 1727 Constraints = CRF.add(Constraints, *this, *NewClassConstraint); 1728 1729 assert(areFeasible(Constraints) && "Constraint manager shouldn't produce " 1730 "a state with infeasible constraints"); 1731 1732 State = State->set<ConstraintRange>(Constraints); 1733 } 1734 1735 // 2. Get ALL equivalence-related maps 1736 ClassMapTy Classes = State->get<ClassMap>(); 1737 ClassMapTy::Factory &CMF = State->get_context<ClassMap>(); 1738 1739 ClassMembersTy Members = State->get<ClassMembers>(); 1740 ClassMembersTy::Factory &MF = State->get_context<ClassMembers>(); 1741 1742 DisequalityMapTy DisequalityInfo = State->get<DisequalityMap>(); 1743 DisequalityMapTy::Factory &DF = State->get_context<DisequalityMap>(); 1744 1745 ClassSet::Factory &CF = State->get_context<ClassSet>(); 1746 SymbolSet::Factory &F = getMembersFactory(State); 1747 1748 // 2. Merge members of the Other class into the current class. 1749 SymbolSet NewClassMembers = MyMembers; 1750 for (SymbolRef Sym : OtherMembers) { 1751 NewClassMembers = F.add(NewClassMembers, Sym); 1752 // *this is now the class for all these new symbols. 1753 Classes = CMF.add(Classes, Sym, *this); 1754 } 1755 1756 // 3. Adjust member mapping. 1757 // 1758 // No need in tracking members of a now-dissolved class. 1759 Members = MF.remove(Members, Other); 1760 // Now only the current class is mapped to all the symbols. 1761 Members = MF.add(Members, *this, NewClassMembers); 1762 1763 // 4. Update disequality relations 1764 ClassSet DisequalToOther = Other.getDisequalClasses(DisequalityInfo, CF); 1765 // We are about to merge two classes but they are already known to be 1766 // non-equal. This is a contradiction. 1767 if (DisequalToOther.contains(*this)) 1768 return nullptr; 1769 1770 if (!DisequalToOther.isEmpty()) { 1771 ClassSet DisequalToThis = getDisequalClasses(DisequalityInfo, CF); 1772 DisequalityInfo = DF.remove(DisequalityInfo, Other); 1773 1774 for (EquivalenceClass DisequalClass : DisequalToOther) { 1775 DisequalToThis = CF.add(DisequalToThis, DisequalClass); 1776 1777 // Disequality is a symmetric relation meaning that if 1778 // DisequalToOther not null then the set for DisequalClass is not 1779 // empty and has at least Other. 1780 ClassSet OriginalSetLinkedToOther = 1781 *DisequalityInfo.lookup(DisequalClass); 1782 1783 // Other will be eliminated and we should replace it with the bigger 1784 // united class. 1785 ClassSet NewSet = CF.remove(OriginalSetLinkedToOther, Other); 1786 NewSet = CF.add(NewSet, *this); 1787 1788 DisequalityInfo = DF.add(DisequalityInfo, DisequalClass, NewSet); 1789 } 1790 1791 DisequalityInfo = DF.add(DisequalityInfo, *this, DisequalToThis); 1792 State = State->set<DisequalityMap>(DisequalityInfo); 1793 } 1794 1795 // 5. Update the state 1796 State = State->set<ClassMap>(Classes); 1797 State = State->set<ClassMembers>(Members); 1798 1799 return State; 1800 } 1801 1802 inline SymbolSet::Factory & 1803 EquivalenceClass::getMembersFactory(ProgramStateRef State) { 1804 return State->get_context<SymbolSet>(); 1805 } 1806 1807 SymbolSet EquivalenceClass::getClassMembers(ProgramStateRef State) const { 1808 if (const SymbolSet *Members = State->get<ClassMembers>(*this)) 1809 return *Members; 1810 1811 // This class is trivial, so we need to construct a set 1812 // with just that one symbol from the class. 1813 SymbolSet::Factory &F = getMembersFactory(State); 1814 return F.add(F.getEmptySet(), getRepresentativeSymbol()); 1815 } 1816 1817 bool EquivalenceClass::isTrivial(ProgramStateRef State) const { 1818 return State->get<ClassMembers>(*this) == nullptr; 1819 } 1820 1821 bool EquivalenceClass::isTriviallyDead(ProgramStateRef State, 1822 SymbolReaper &Reaper) const { 1823 return isTrivial(State) && Reaper.isDead(getRepresentativeSymbol()); 1824 } 1825 1826 inline ProgramStateRef EquivalenceClass::markDisequal(RangeSet::Factory &RF, 1827 ProgramStateRef State, 1828 SymbolRef First, 1829 SymbolRef Second) { 1830 return markDisequal(RF, State, find(State, First), find(State, Second)); 1831 } 1832 1833 inline ProgramStateRef EquivalenceClass::markDisequal(RangeSet::Factory &RF, 1834 ProgramStateRef State, 1835 EquivalenceClass First, 1836 EquivalenceClass Second) { 1837 return First.markDisequal(RF, State, Second); 1838 } 1839 1840 inline ProgramStateRef 1841 EquivalenceClass::markDisequal(RangeSet::Factory &RF, ProgramStateRef State, 1842 EquivalenceClass Other) const { 1843 // If we know that two classes are equal, we can only produce an infeasible 1844 // state. 1845 if (*this == Other) { 1846 return nullptr; 1847 } 1848 1849 DisequalityMapTy DisequalityInfo = State->get<DisequalityMap>(); 1850 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1851 1852 // Disequality is a symmetric relation, so if we mark A as disequal to B, 1853 // we should also mark B as disequalt to A. 1854 if (!addToDisequalityInfo(DisequalityInfo, Constraints, RF, State, *this, 1855 Other) || 1856 !addToDisequalityInfo(DisequalityInfo, Constraints, RF, State, Other, 1857 *this)) 1858 return nullptr; 1859 1860 assert(areFeasible(Constraints) && "Constraint manager shouldn't produce " 1861 "a state with infeasible constraints"); 1862 1863 State = State->set<DisequalityMap>(DisequalityInfo); 1864 State = State->set<ConstraintRange>(Constraints); 1865 1866 return State; 1867 } 1868 1869 inline bool EquivalenceClass::addToDisequalityInfo( 1870 DisequalityMapTy &Info, ConstraintRangeTy &Constraints, 1871 RangeSet::Factory &RF, ProgramStateRef State, EquivalenceClass First, 1872 EquivalenceClass Second) { 1873 1874 // 1. Get all of the required factories. 1875 DisequalityMapTy::Factory &F = State->get_context<DisequalityMap>(); 1876 ClassSet::Factory &CF = State->get_context<ClassSet>(); 1877 ConstraintRangeTy::Factory &CRF = State->get_context<ConstraintRange>(); 1878 1879 // 2. Add Second to the set of classes disequal to First. 1880 const ClassSet *CurrentSet = Info.lookup(First); 1881 ClassSet NewSet = CurrentSet ? *CurrentSet : CF.getEmptySet(); 1882 NewSet = CF.add(NewSet, Second); 1883 1884 Info = F.add(Info, First, NewSet); 1885 1886 // 3. If Second is known to be a constant, we can delete this point 1887 // from the constraint asociated with First. 1888 // 1889 // So, if Second == 10, it means that First != 10. 1890 // At the same time, the same logic does not apply to ranges. 1891 if (const RangeSet *SecondConstraint = Constraints.lookup(Second)) 1892 if (const llvm::APSInt *Point = SecondConstraint->getConcreteValue()) { 1893 1894 RangeSet FirstConstraint = SymbolicRangeInferrer::inferRange( 1895 RF, State, First.getRepresentativeSymbol()); 1896 1897 FirstConstraint = RF.deletePoint(FirstConstraint, *Point); 1898 1899 // If the First class is about to be constrained with an empty 1900 // range-set, the state is infeasible. 1901 if (FirstConstraint.isEmpty()) 1902 return false; 1903 1904 Constraints = CRF.add(Constraints, First, FirstConstraint); 1905 } 1906 1907 return true; 1908 } 1909 1910 inline Optional<bool> EquivalenceClass::areEqual(ProgramStateRef State, 1911 SymbolRef FirstSym, 1912 SymbolRef SecondSym) { 1913 return EquivalenceClass::areEqual(State, find(State, FirstSym), 1914 find(State, SecondSym)); 1915 } 1916 1917 inline Optional<bool> EquivalenceClass::areEqual(ProgramStateRef State, 1918 EquivalenceClass First, 1919 EquivalenceClass Second) { 1920 // The same equivalence class => symbols are equal. 1921 if (First == Second) 1922 return true; 1923 1924 // Let's check if we know anything about these two classes being not equal to 1925 // each other. 1926 ClassSet DisequalToFirst = First.getDisequalClasses(State); 1927 if (DisequalToFirst.contains(Second)) 1928 return false; 1929 1930 // It is not clear. 1931 return llvm::None; 1932 } 1933 1934 // Iterate over all symbols and try to simplify them. Once a symbol is 1935 // simplified then we check if we can merge the simplified symbol's equivalence 1936 // class to this class. This way, we simplify not just the symbols but the 1937 // classes as well: we strive to keep the number of the classes to be the 1938 // absolute minimum. 1939 LLVM_NODISCARD ProgramStateRef EquivalenceClass::simplify( 1940 SValBuilder &SVB, RangeSet::Factory &F, ProgramStateRef State) { 1941 SymbolSet ClassMembers = getClassMembers(State); 1942 for (const SymbolRef &MemberSym : ClassMembers) { 1943 SymbolRef SimplifiedMemberSym = ento::simplify(State, MemberSym); 1944 if (SimplifiedMemberSym && MemberSym != SimplifiedMemberSym) { 1945 EquivalenceClass ClassOfSimplifiedSym = 1946 EquivalenceClass::find(State, SimplifiedMemberSym); 1947 // The simplified symbol should be the member of the original Class, 1948 // however, it might be in another existing class at the moment. We 1949 // have to merge these classes. 1950 State = merge(F, State, ClassOfSimplifiedSym); 1951 if (!State) 1952 return nullptr; 1953 } 1954 } 1955 return State; 1956 } 1957 1958 inline ClassSet EquivalenceClass::getDisequalClasses(ProgramStateRef State, 1959 SymbolRef Sym) { 1960 return find(State, Sym).getDisequalClasses(State); 1961 } 1962 1963 inline ClassSet 1964 EquivalenceClass::getDisequalClasses(ProgramStateRef State) const { 1965 return getDisequalClasses(State->get<DisequalityMap>(), 1966 State->get_context<ClassSet>()); 1967 } 1968 1969 inline ClassSet 1970 EquivalenceClass::getDisequalClasses(DisequalityMapTy Map, 1971 ClassSet::Factory &Factory) const { 1972 if (const ClassSet *DisequalClasses = Map.lookup(*this)) 1973 return *DisequalClasses; 1974 1975 return Factory.getEmptySet(); 1976 } 1977 1978 bool EquivalenceClass::isClassDataConsistent(ProgramStateRef State) { 1979 ClassMembersTy Members = State->get<ClassMembers>(); 1980 1981 for (std::pair<EquivalenceClass, SymbolSet> ClassMembersPair : Members) { 1982 for (SymbolRef Member : ClassMembersPair.second) { 1983 // Every member of the class should have a mapping back to the class. 1984 if (find(State, Member) == ClassMembersPair.first) { 1985 continue; 1986 } 1987 1988 return false; 1989 } 1990 } 1991 1992 DisequalityMapTy Disequalities = State->get<DisequalityMap>(); 1993 for (std::pair<EquivalenceClass, ClassSet> DisequalityInfo : Disequalities) { 1994 EquivalenceClass Class = DisequalityInfo.first; 1995 ClassSet DisequalClasses = DisequalityInfo.second; 1996 1997 // There is no use in keeping empty sets in the map. 1998 if (DisequalClasses.isEmpty()) 1999 return false; 2000 2001 // Disequality is symmetrical, i.e. for every Class A and B that A != B, 2002 // B != A should also be true. 2003 for (EquivalenceClass DisequalClass : DisequalClasses) { 2004 const ClassSet *DisequalToDisequalClasses = 2005 Disequalities.lookup(DisequalClass); 2006 2007 // It should be a set of at least one element: Class 2008 if (!DisequalToDisequalClasses || 2009 !DisequalToDisequalClasses->contains(Class)) 2010 return false; 2011 } 2012 } 2013 2014 return true; 2015 } 2016 2017 //===----------------------------------------------------------------------===// 2018 // RangeConstraintManager implementation 2019 //===----------------------------------------------------------------------===// 2020 2021 bool RangeConstraintManager::canReasonAbout(SVal X) const { 2022 Optional<nonloc::SymbolVal> SymVal = X.getAs<nonloc::SymbolVal>(); 2023 if (SymVal && SymVal->isExpression()) { 2024 const SymExpr *SE = SymVal->getSymbol(); 2025 2026 if (const SymIntExpr *SIE = dyn_cast<SymIntExpr>(SE)) { 2027 switch (SIE->getOpcode()) { 2028 // We don't reason yet about bitwise-constraints on symbolic values. 2029 case BO_And: 2030 case BO_Or: 2031 case BO_Xor: 2032 return false; 2033 // We don't reason yet about these arithmetic constraints on 2034 // symbolic values. 2035 case BO_Mul: 2036 case BO_Div: 2037 case BO_Rem: 2038 case BO_Shl: 2039 case BO_Shr: 2040 return false; 2041 // All other cases. 2042 default: 2043 return true; 2044 } 2045 } 2046 2047 if (const SymSymExpr *SSE = dyn_cast<SymSymExpr>(SE)) { 2048 // FIXME: Handle <=> here. 2049 if (BinaryOperator::isEqualityOp(SSE->getOpcode()) || 2050 BinaryOperator::isRelationalOp(SSE->getOpcode())) { 2051 // We handle Loc <> Loc comparisons, but not (yet) NonLoc <> NonLoc. 2052 // We've recently started producing Loc <> NonLoc comparisons (that 2053 // result from casts of one of the operands between eg. intptr_t and 2054 // void *), but we can't reason about them yet. 2055 if (Loc::isLocType(SSE->getLHS()->getType())) { 2056 return Loc::isLocType(SSE->getRHS()->getType()); 2057 } 2058 } 2059 } 2060 2061 return false; 2062 } 2063 2064 return true; 2065 } 2066 2067 ConditionTruthVal RangeConstraintManager::checkNull(ProgramStateRef State, 2068 SymbolRef Sym) { 2069 const RangeSet *Ranges = getConstraint(State, Sym); 2070 2071 // If we don't have any information about this symbol, it's underconstrained. 2072 if (!Ranges) 2073 return ConditionTruthVal(); 2074 2075 // If we have a concrete value, see if it's zero. 2076 if (const llvm::APSInt *Value = Ranges->getConcreteValue()) 2077 return *Value == 0; 2078 2079 BasicValueFactory &BV = getBasicVals(); 2080 APSIntType IntType = BV.getAPSIntType(Sym->getType()); 2081 llvm::APSInt Zero = IntType.getZeroValue(); 2082 2083 // Check if zero is in the set of possible values. 2084 if (!Ranges->contains(Zero)) 2085 return false; 2086 2087 // Zero is a possible value, but it is not the /only/ possible value. 2088 return ConditionTruthVal(); 2089 } 2090 2091 const llvm::APSInt *RangeConstraintManager::getSymVal(ProgramStateRef St, 2092 SymbolRef Sym) const { 2093 const RangeSet *T = getConstraint(St, Sym); 2094 return T ? T->getConcreteValue() : nullptr; 2095 } 2096 2097 //===----------------------------------------------------------------------===// 2098 // Remove dead symbols from existing constraints 2099 //===----------------------------------------------------------------------===// 2100 2101 /// Scan all symbols referenced by the constraints. If the symbol is not alive 2102 /// as marked in LSymbols, mark it as dead in DSymbols. 2103 ProgramStateRef 2104 RangeConstraintManager::removeDeadBindings(ProgramStateRef State, 2105 SymbolReaper &SymReaper) { 2106 ClassMembersTy ClassMembersMap = State->get<ClassMembers>(); 2107 ClassMembersTy NewClassMembersMap = ClassMembersMap; 2108 ClassMembersTy::Factory &EMFactory = State->get_context<ClassMembers>(); 2109 SymbolSet::Factory &SetFactory = State->get_context<SymbolSet>(); 2110 2111 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 2112 ConstraintRangeTy NewConstraints = Constraints; 2113 ConstraintRangeTy::Factory &ConstraintFactory = 2114 State->get_context<ConstraintRange>(); 2115 2116 ClassMapTy Map = State->get<ClassMap>(); 2117 ClassMapTy NewMap = Map; 2118 ClassMapTy::Factory &ClassFactory = State->get_context<ClassMap>(); 2119 2120 DisequalityMapTy Disequalities = State->get<DisequalityMap>(); 2121 DisequalityMapTy::Factory &DisequalityFactory = 2122 State->get_context<DisequalityMap>(); 2123 ClassSet::Factory &ClassSetFactory = State->get_context<ClassSet>(); 2124 2125 bool ClassMapChanged = false; 2126 bool MembersMapChanged = false; 2127 bool ConstraintMapChanged = false; 2128 bool DisequalitiesChanged = false; 2129 2130 auto removeDeadClass = [&](EquivalenceClass Class) { 2131 // Remove associated constraint ranges. 2132 Constraints = ConstraintFactory.remove(Constraints, Class); 2133 ConstraintMapChanged = true; 2134 2135 // Update disequality information to not hold any information on the 2136 // removed class. 2137 ClassSet DisequalClasses = 2138 Class.getDisequalClasses(Disequalities, ClassSetFactory); 2139 if (!DisequalClasses.isEmpty()) { 2140 for (EquivalenceClass DisequalClass : DisequalClasses) { 2141 ClassSet DisequalToDisequalSet = 2142 DisequalClass.getDisequalClasses(Disequalities, ClassSetFactory); 2143 // DisequalToDisequalSet is guaranteed to be non-empty for consistent 2144 // disequality info. 2145 assert(!DisequalToDisequalSet.isEmpty()); 2146 ClassSet NewSet = ClassSetFactory.remove(DisequalToDisequalSet, Class); 2147 2148 // No need in keeping an empty set. 2149 if (NewSet.isEmpty()) { 2150 Disequalities = 2151 DisequalityFactory.remove(Disequalities, DisequalClass); 2152 } else { 2153 Disequalities = 2154 DisequalityFactory.add(Disequalities, DisequalClass, NewSet); 2155 } 2156 } 2157 // Remove the data for the class 2158 Disequalities = DisequalityFactory.remove(Disequalities, Class); 2159 DisequalitiesChanged = true; 2160 } 2161 }; 2162 2163 // 1. Let's see if dead symbols are trivial and have associated constraints. 2164 for (std::pair<EquivalenceClass, RangeSet> ClassConstraintPair : 2165 Constraints) { 2166 EquivalenceClass Class = ClassConstraintPair.first; 2167 if (Class.isTriviallyDead(State, SymReaper)) { 2168 // If this class is trivial, we can remove its constraints right away. 2169 removeDeadClass(Class); 2170 } 2171 } 2172 2173 // 2. We don't need to track classes for dead symbols. 2174 for (std::pair<SymbolRef, EquivalenceClass> SymbolClassPair : Map) { 2175 SymbolRef Sym = SymbolClassPair.first; 2176 2177 if (SymReaper.isDead(Sym)) { 2178 ClassMapChanged = true; 2179 NewMap = ClassFactory.remove(NewMap, Sym); 2180 } 2181 } 2182 2183 // 3. Remove dead members from classes and remove dead non-trivial classes 2184 // and their constraints. 2185 for (std::pair<EquivalenceClass, SymbolSet> ClassMembersPair : 2186 ClassMembersMap) { 2187 EquivalenceClass Class = ClassMembersPair.first; 2188 SymbolSet LiveMembers = ClassMembersPair.second; 2189 bool MembersChanged = false; 2190 2191 for (SymbolRef Member : ClassMembersPair.second) { 2192 if (SymReaper.isDead(Member)) { 2193 MembersChanged = true; 2194 LiveMembers = SetFactory.remove(LiveMembers, Member); 2195 } 2196 } 2197 2198 // Check if the class changed. 2199 if (!MembersChanged) 2200 continue; 2201 2202 MembersMapChanged = true; 2203 2204 if (LiveMembers.isEmpty()) { 2205 // The class is dead now, we need to wipe it out of the members map... 2206 NewClassMembersMap = EMFactory.remove(NewClassMembersMap, Class); 2207 2208 // ...and remove all of its constraints. 2209 removeDeadClass(Class); 2210 } else { 2211 // We need to change the members associated with the class. 2212 NewClassMembersMap = 2213 EMFactory.add(NewClassMembersMap, Class, LiveMembers); 2214 } 2215 } 2216 2217 // 4. Update the state with new maps. 2218 // 2219 // Here we try to be humble and update a map only if it really changed. 2220 if (ClassMapChanged) 2221 State = State->set<ClassMap>(NewMap); 2222 2223 if (MembersMapChanged) 2224 State = State->set<ClassMembers>(NewClassMembersMap); 2225 2226 if (ConstraintMapChanged) 2227 State = State->set<ConstraintRange>(Constraints); 2228 2229 if (DisequalitiesChanged) 2230 State = State->set<DisequalityMap>(Disequalities); 2231 2232 assert(EquivalenceClass::isClassDataConsistent(State)); 2233 2234 return State; 2235 } 2236 2237 RangeSet RangeConstraintManager::getRange(ProgramStateRef State, 2238 SymbolRef Sym) { 2239 return SymbolicRangeInferrer::inferRange(F, State, Sym); 2240 } 2241 2242 RangeSet RangeConstraintManager::getRange(ProgramStateRef State, 2243 EquivalenceClass Class) { 2244 return SymbolicRangeInferrer::inferRange(F, State, Class); 2245 } 2246 2247 //===------------------------------------------------------------------------=== 2248 // assumeSymX methods: protected interface for RangeConstraintManager. 2249 //===------------------------------------------------------------------------===/ 2250 2251 // The syntax for ranges below is mathematical, using [x, y] for closed ranges 2252 // and (x, y) for open ranges. These ranges are modular, corresponding with 2253 // a common treatment of C integer overflow. This means that these methods 2254 // do not have to worry about overflow; RangeSet::Intersect can handle such a 2255 // "wraparound" range. 2256 // As an example, the range [UINT_MAX-1, 3) contains five values: UINT_MAX-1, 2257 // UINT_MAX, 0, 1, and 2. 2258 2259 ProgramStateRef 2260 RangeConstraintManager::assumeSymNE(ProgramStateRef St, SymbolRef Sym, 2261 const llvm::APSInt &Int, 2262 const llvm::APSInt &Adjustment) { 2263 // Before we do any real work, see if the value can even show up. 2264 APSIntType AdjustmentType(Adjustment); 2265 if (AdjustmentType.testInRange(Int, true) != APSIntType::RTR_Within) 2266 return St; 2267 2268 llvm::APSInt Point = AdjustmentType.convert(Int) - Adjustment; 2269 RangeSet New = getRange(St, Sym); 2270 New = F.deletePoint(New, Point); 2271 2272 return trackNE(New, St, Sym, Int, Adjustment); 2273 } 2274 2275 ProgramStateRef 2276 RangeConstraintManager::assumeSymEQ(ProgramStateRef St, SymbolRef Sym, 2277 const llvm::APSInt &Int, 2278 const llvm::APSInt &Adjustment) { 2279 // Before we do any real work, see if the value can even show up. 2280 APSIntType AdjustmentType(Adjustment); 2281 if (AdjustmentType.testInRange(Int, true) != APSIntType::RTR_Within) 2282 return nullptr; 2283 2284 // [Int-Adjustment, Int-Adjustment] 2285 llvm::APSInt AdjInt = AdjustmentType.convert(Int) - Adjustment; 2286 RangeSet New = getRange(St, Sym); 2287 New = F.intersect(New, AdjInt); 2288 2289 return trackEQ(New, St, Sym, Int, Adjustment); 2290 } 2291 2292 RangeSet RangeConstraintManager::getSymLTRange(ProgramStateRef St, 2293 SymbolRef Sym, 2294 const llvm::APSInt &Int, 2295 const llvm::APSInt &Adjustment) { 2296 // Before we do any real work, see if the value can even show up. 2297 APSIntType AdjustmentType(Adjustment); 2298 switch (AdjustmentType.testInRange(Int, true)) { 2299 case APSIntType::RTR_Below: 2300 return F.getEmptySet(); 2301 case APSIntType::RTR_Within: 2302 break; 2303 case APSIntType::RTR_Above: 2304 return getRange(St, Sym); 2305 } 2306 2307 // Special case for Int == Min. This is always false. 2308 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2309 llvm::APSInt Min = AdjustmentType.getMinValue(); 2310 if (ComparisonVal == Min) 2311 return F.getEmptySet(); 2312 2313 llvm::APSInt Lower = Min - Adjustment; 2314 llvm::APSInt Upper = ComparisonVal - Adjustment; 2315 --Upper; 2316 2317 RangeSet Result = getRange(St, Sym); 2318 return F.intersect(Result, Lower, Upper); 2319 } 2320 2321 ProgramStateRef 2322 RangeConstraintManager::assumeSymLT(ProgramStateRef St, SymbolRef Sym, 2323 const llvm::APSInt &Int, 2324 const llvm::APSInt &Adjustment) { 2325 RangeSet New = getSymLTRange(St, Sym, Int, Adjustment); 2326 return trackNE(New, St, Sym, Int, Adjustment); 2327 } 2328 2329 RangeSet RangeConstraintManager::getSymGTRange(ProgramStateRef St, 2330 SymbolRef Sym, 2331 const llvm::APSInt &Int, 2332 const llvm::APSInt &Adjustment) { 2333 // Before we do any real work, see if the value can even show up. 2334 APSIntType AdjustmentType(Adjustment); 2335 switch (AdjustmentType.testInRange(Int, true)) { 2336 case APSIntType::RTR_Below: 2337 return getRange(St, Sym); 2338 case APSIntType::RTR_Within: 2339 break; 2340 case APSIntType::RTR_Above: 2341 return F.getEmptySet(); 2342 } 2343 2344 // Special case for Int == Max. This is always false. 2345 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2346 llvm::APSInt Max = AdjustmentType.getMaxValue(); 2347 if (ComparisonVal == Max) 2348 return F.getEmptySet(); 2349 2350 llvm::APSInt Lower = ComparisonVal - Adjustment; 2351 llvm::APSInt Upper = Max - Adjustment; 2352 ++Lower; 2353 2354 RangeSet SymRange = getRange(St, Sym); 2355 return F.intersect(SymRange, Lower, Upper); 2356 } 2357 2358 ProgramStateRef 2359 RangeConstraintManager::assumeSymGT(ProgramStateRef St, SymbolRef Sym, 2360 const llvm::APSInt &Int, 2361 const llvm::APSInt &Adjustment) { 2362 RangeSet New = getSymGTRange(St, Sym, Int, Adjustment); 2363 return trackNE(New, St, Sym, Int, Adjustment); 2364 } 2365 2366 RangeSet RangeConstraintManager::getSymGERange(ProgramStateRef St, 2367 SymbolRef Sym, 2368 const llvm::APSInt &Int, 2369 const llvm::APSInt &Adjustment) { 2370 // Before we do any real work, see if the value can even show up. 2371 APSIntType AdjustmentType(Adjustment); 2372 switch (AdjustmentType.testInRange(Int, true)) { 2373 case APSIntType::RTR_Below: 2374 return getRange(St, Sym); 2375 case APSIntType::RTR_Within: 2376 break; 2377 case APSIntType::RTR_Above: 2378 return F.getEmptySet(); 2379 } 2380 2381 // Special case for Int == Min. This is always feasible. 2382 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2383 llvm::APSInt Min = AdjustmentType.getMinValue(); 2384 if (ComparisonVal == Min) 2385 return getRange(St, Sym); 2386 2387 llvm::APSInt Max = AdjustmentType.getMaxValue(); 2388 llvm::APSInt Lower = ComparisonVal - Adjustment; 2389 llvm::APSInt Upper = Max - Adjustment; 2390 2391 RangeSet SymRange = getRange(St, Sym); 2392 return F.intersect(SymRange, Lower, Upper); 2393 } 2394 2395 ProgramStateRef 2396 RangeConstraintManager::assumeSymGE(ProgramStateRef St, SymbolRef Sym, 2397 const llvm::APSInt &Int, 2398 const llvm::APSInt &Adjustment) { 2399 RangeSet New = getSymGERange(St, Sym, Int, Adjustment); 2400 return New.isEmpty() ? nullptr : setConstraint(St, Sym, New); 2401 } 2402 2403 RangeSet 2404 RangeConstraintManager::getSymLERange(llvm::function_ref<RangeSet()> RS, 2405 const llvm::APSInt &Int, 2406 const llvm::APSInt &Adjustment) { 2407 // Before we do any real work, see if the value can even show up. 2408 APSIntType AdjustmentType(Adjustment); 2409 switch (AdjustmentType.testInRange(Int, true)) { 2410 case APSIntType::RTR_Below: 2411 return F.getEmptySet(); 2412 case APSIntType::RTR_Within: 2413 break; 2414 case APSIntType::RTR_Above: 2415 return RS(); 2416 } 2417 2418 // Special case for Int == Max. This is always feasible. 2419 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2420 llvm::APSInt Max = AdjustmentType.getMaxValue(); 2421 if (ComparisonVal == Max) 2422 return RS(); 2423 2424 llvm::APSInt Min = AdjustmentType.getMinValue(); 2425 llvm::APSInt Lower = Min - Adjustment; 2426 llvm::APSInt Upper = ComparisonVal - Adjustment; 2427 2428 RangeSet Default = RS(); 2429 return F.intersect(Default, Lower, Upper); 2430 } 2431 2432 RangeSet RangeConstraintManager::getSymLERange(ProgramStateRef St, 2433 SymbolRef Sym, 2434 const llvm::APSInt &Int, 2435 const llvm::APSInt &Adjustment) { 2436 return getSymLERange([&] { return getRange(St, Sym); }, Int, Adjustment); 2437 } 2438 2439 ProgramStateRef 2440 RangeConstraintManager::assumeSymLE(ProgramStateRef St, SymbolRef Sym, 2441 const llvm::APSInt &Int, 2442 const llvm::APSInt &Adjustment) { 2443 RangeSet New = getSymLERange(St, Sym, Int, Adjustment); 2444 return New.isEmpty() ? nullptr : setConstraint(St, Sym, New); 2445 } 2446 2447 ProgramStateRef RangeConstraintManager::assumeSymWithinInclusiveRange( 2448 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 2449 const llvm::APSInt &To, const llvm::APSInt &Adjustment) { 2450 RangeSet New = getSymGERange(State, Sym, From, Adjustment); 2451 if (New.isEmpty()) 2452 return nullptr; 2453 RangeSet Out = getSymLERange([&] { return New; }, To, Adjustment); 2454 return Out.isEmpty() ? nullptr : setConstraint(State, Sym, Out); 2455 } 2456 2457 ProgramStateRef RangeConstraintManager::assumeSymOutsideInclusiveRange( 2458 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 2459 const llvm::APSInt &To, const llvm::APSInt &Adjustment) { 2460 RangeSet RangeLT = getSymLTRange(State, Sym, From, Adjustment); 2461 RangeSet RangeGT = getSymGTRange(State, Sym, To, Adjustment); 2462 RangeSet New(F.add(RangeLT, RangeGT)); 2463 return New.isEmpty() ? nullptr : setConstraint(State, Sym, New); 2464 } 2465 2466 //===----------------------------------------------------------------------===// 2467 // Pretty-printing. 2468 //===----------------------------------------------------------------------===// 2469 2470 void RangeConstraintManager::printJson(raw_ostream &Out, ProgramStateRef State, 2471 const char *NL, unsigned int Space, 2472 bool IsDot) const { 2473 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 2474 2475 Indent(Out, Space, IsDot) << "\"constraints\": "; 2476 if (Constraints.isEmpty()) { 2477 Out << "null," << NL; 2478 return; 2479 } 2480 2481 ++Space; 2482 Out << '[' << NL; 2483 bool First = true; 2484 for (std::pair<EquivalenceClass, RangeSet> P : Constraints) { 2485 SymbolSet ClassMembers = P.first.getClassMembers(State); 2486 2487 // We can print the same constraint for every class member. 2488 for (SymbolRef ClassMember : ClassMembers) { 2489 if (First) { 2490 First = false; 2491 } else { 2492 Out << ','; 2493 Out << NL; 2494 } 2495 Indent(Out, Space, IsDot) 2496 << "{ \"symbol\": \"" << ClassMember << "\", \"range\": \""; 2497 P.second.dump(Out); 2498 Out << "\" }"; 2499 } 2500 } 2501 Out << NL; 2502 2503 --Space; 2504 Indent(Out, Space, IsDot) << "]," << NL; 2505 } 2506