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