1 //== RangeConstraintManager.cpp - Manage range constraints.------*- C++ -*--==// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file defines RangeConstraintManager, a class that tracks simple 10 // equality and inequality constraints on symbolic values of ProgramState. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "clang/Basic/JsonSupport.h" 15 #include "clang/StaticAnalyzer/Core/PathSensitive/APSIntType.h" 16 #include "clang/StaticAnalyzer/Core/PathSensitive/ProgramState.h" 17 #include "clang/StaticAnalyzer/Core/PathSensitive/ProgramStateTrait.h" 18 #include "clang/StaticAnalyzer/Core/PathSensitive/RangedConstraintManager.h" 19 #include "clang/StaticAnalyzer/Core/PathSensitive/SValVisitor.h" 20 #include "llvm/ADT/FoldingSet.h" 21 #include "llvm/ADT/ImmutableSet.h" 22 #include "llvm/ADT/STLExtras.h" 23 #include "llvm/ADT/StringExtras.h" 24 #include "llvm/ADT/SmallSet.h" 25 #include "llvm/Support/Compiler.h" 26 #include "llvm/Support/raw_ostream.h" 27 #include <algorithm> 28 #include <iterator> 29 30 using namespace clang; 31 using namespace ento; 32 33 // This class can be extended with other tables which will help to reason 34 // about ranges more precisely. 35 class OperatorRelationsTable { 36 static_assert(BO_LT < BO_GT && BO_GT < BO_LE && BO_LE < BO_GE && 37 BO_GE < BO_EQ && BO_EQ < BO_NE, 38 "This class relies on operators order. Rework it otherwise."); 39 40 public: 41 enum TriStateKind { 42 False = 0, 43 True, 44 Unknown, 45 }; 46 47 private: 48 // CmpOpTable holds states which represent the corresponding range for 49 // branching an exploded graph. We can reason about the branch if there is 50 // a previously known fact of the existence of a comparison expression with 51 // operands used in the current expression. 52 // E.g. assuming (x < y) is true that means (x != y) is surely true. 53 // if (x previous_operation y) // < | != | > 54 // if (x operation y) // != | > | < 55 // tristate // True | Unknown | False 56 // 57 // CmpOpTable represents next: 58 // __|< |> |<=|>=|==|!=|UnknownX2| 59 // < |1 |0 |* |0 |0 |* |1 | 60 // > |0 |1 |0 |* |0 |* |1 | 61 // <=|1 |0 |1 |* |1 |* |0 | 62 // >=|0 |1 |* |1 |1 |* |0 | 63 // ==|0 |0 |* |* |1 |0 |1 | 64 // !=|1 |1 |* |* |0 |1 |0 | 65 // 66 // Columns stands for a previous operator. 67 // Rows stands for a current operator. 68 // Each row has exactly two `Unknown` cases. 69 // UnknownX2 means that both `Unknown` previous operators are met in code, 70 // and there is a special column for that, for example: 71 // if (x >= y) 72 // if (x != y) 73 // if (x <= y) 74 // False only 75 static constexpr size_t CmpOpCount = BO_NE - BO_LT + 1; 76 const TriStateKind CmpOpTable[CmpOpCount][CmpOpCount + 1] = { 77 // < > <= >= == != UnknownX2 78 {True, False, Unknown, False, False, Unknown, True}, // < 79 {False, True, False, Unknown, False, Unknown, True}, // > 80 {True, False, True, Unknown, True, Unknown, False}, // <= 81 {False, True, Unknown, True, True, Unknown, False}, // >= 82 {False, False, Unknown, Unknown, True, False, True}, // == 83 {True, True, Unknown, Unknown, False, True, False}, // != 84 }; 85 86 static size_t getIndexFromOp(BinaryOperatorKind OP) { 87 return static_cast<size_t>(OP - BO_LT); 88 } 89 90 public: 91 constexpr size_t getCmpOpCount() const { return CmpOpCount; } 92 93 static BinaryOperatorKind getOpFromIndex(size_t Index) { 94 return static_cast<BinaryOperatorKind>(Index + BO_LT); 95 } 96 97 TriStateKind getCmpOpState(BinaryOperatorKind CurrentOP, 98 BinaryOperatorKind QueriedOP) const { 99 return CmpOpTable[getIndexFromOp(CurrentOP)][getIndexFromOp(QueriedOP)]; 100 } 101 102 TriStateKind getCmpOpStateForUnknownX2(BinaryOperatorKind CurrentOP) const { 103 return CmpOpTable[getIndexFromOp(CurrentOP)][CmpOpCount]; 104 } 105 }; 106 107 //===----------------------------------------------------------------------===// 108 // RangeSet implementation 109 //===----------------------------------------------------------------------===// 110 111 RangeSet::ContainerType RangeSet::Factory::EmptySet{}; 112 113 RangeSet RangeSet::Factory::add(RangeSet Original, Range Element) { 114 ContainerType Result; 115 Result.reserve(Original.size() + 1); 116 117 const_iterator Lower = llvm::lower_bound(Original, Element); 118 Result.insert(Result.end(), Original.begin(), Lower); 119 Result.push_back(Element); 120 Result.insert(Result.end(), Lower, Original.end()); 121 122 return makePersistent(std::move(Result)); 123 } 124 125 RangeSet RangeSet::Factory::add(RangeSet Original, const llvm::APSInt &Point) { 126 return add(Original, Range(Point)); 127 } 128 129 RangeSet RangeSet::Factory::getRangeSet(Range From) { 130 ContainerType Result; 131 Result.push_back(From); 132 return makePersistent(std::move(Result)); 133 } 134 135 RangeSet RangeSet::Factory::makePersistent(ContainerType &&From) { 136 llvm::FoldingSetNodeID ID; 137 void *InsertPos; 138 139 From.Profile(ID); 140 ContainerType *Result = Cache.FindNodeOrInsertPos(ID, InsertPos); 141 142 if (!Result) { 143 // It is cheaper to fully construct the resulting range on stack 144 // and move it to the freshly allocated buffer if we don't have 145 // a set like this already. 146 Result = construct(std::move(From)); 147 Cache.InsertNode(Result, InsertPos); 148 } 149 150 return Result; 151 } 152 153 RangeSet::ContainerType *RangeSet::Factory::construct(ContainerType &&From) { 154 void *Buffer = Arena.Allocate(); 155 return new (Buffer) ContainerType(std::move(From)); 156 } 157 158 RangeSet RangeSet::Factory::add(RangeSet LHS, RangeSet RHS) { 159 ContainerType Result; 160 std::merge(LHS.begin(), LHS.end(), RHS.begin(), RHS.end(), 161 std::back_inserter(Result)); 162 return makePersistent(std::move(Result)); 163 } 164 165 const llvm::APSInt &RangeSet::getMinValue() const { 166 assert(!isEmpty()); 167 return begin()->From(); 168 } 169 170 const llvm::APSInt &RangeSet::getMaxValue() const { 171 assert(!isEmpty()); 172 return std::prev(end())->To(); 173 } 174 175 bool RangeSet::containsImpl(llvm::APSInt &Point) const { 176 if (isEmpty() || !pin(Point)) 177 return false; 178 179 Range Dummy(Point); 180 const_iterator It = llvm::upper_bound(*this, Dummy); 181 if (It == begin()) 182 return false; 183 184 return std::prev(It)->Includes(Point); 185 } 186 187 bool RangeSet::pin(llvm::APSInt &Point) const { 188 APSIntType Type(getMinValue()); 189 if (Type.testInRange(Point, true) != APSIntType::RTR_Within) 190 return false; 191 192 Type.apply(Point); 193 return true; 194 } 195 196 bool RangeSet::pin(llvm::APSInt &Lower, llvm::APSInt &Upper) const { 197 // This function has nine cases, the cartesian product of range-testing 198 // both the upper and lower bounds against the symbol's type. 199 // Each case requires a different pinning operation. 200 // The function returns false if the described range is entirely outside 201 // the range of values for the associated symbol. 202 APSIntType Type(getMinValue()); 203 APSIntType::RangeTestResultKind LowerTest = Type.testInRange(Lower, true); 204 APSIntType::RangeTestResultKind UpperTest = Type.testInRange(Upper, true); 205 206 switch (LowerTest) { 207 case APSIntType::RTR_Below: 208 switch (UpperTest) { 209 case APSIntType::RTR_Below: 210 // The entire range is outside the symbol's set of possible values. 211 // If this is a conventionally-ordered range, the state is infeasible. 212 if (Lower <= Upper) 213 return false; 214 215 // However, if the range wraps around, it spans all possible values. 216 Lower = Type.getMinValue(); 217 Upper = Type.getMaxValue(); 218 break; 219 case APSIntType::RTR_Within: 220 // The range starts below what's possible but ends within it. Pin. 221 Lower = Type.getMinValue(); 222 Type.apply(Upper); 223 break; 224 case APSIntType::RTR_Above: 225 // The range spans all possible values for the symbol. Pin. 226 Lower = Type.getMinValue(); 227 Upper = Type.getMaxValue(); 228 break; 229 } 230 break; 231 case APSIntType::RTR_Within: 232 switch (UpperTest) { 233 case APSIntType::RTR_Below: 234 // The range wraps around, but all lower values are not possible. 235 Type.apply(Lower); 236 Upper = Type.getMaxValue(); 237 break; 238 case APSIntType::RTR_Within: 239 // The range may or may not wrap around, but both limits are valid. 240 Type.apply(Lower); 241 Type.apply(Upper); 242 break; 243 case APSIntType::RTR_Above: 244 // The range starts within what's possible but ends above it. Pin. 245 Type.apply(Lower); 246 Upper = Type.getMaxValue(); 247 break; 248 } 249 break; 250 case APSIntType::RTR_Above: 251 switch (UpperTest) { 252 case APSIntType::RTR_Below: 253 // The range wraps but is outside the symbol's set of possible values. 254 return false; 255 case APSIntType::RTR_Within: 256 // The range starts above what's possible but ends within it (wrap). 257 Lower = Type.getMinValue(); 258 Type.apply(Upper); 259 break; 260 case APSIntType::RTR_Above: 261 // The entire range is outside the symbol's set of possible values. 262 // If this is a conventionally-ordered range, the state is infeasible. 263 if (Lower <= Upper) 264 return false; 265 266 // However, if the range wraps around, it spans all possible values. 267 Lower = Type.getMinValue(); 268 Upper = Type.getMaxValue(); 269 break; 270 } 271 break; 272 } 273 274 return true; 275 } 276 277 RangeSet RangeSet::Factory::intersect(RangeSet What, llvm::APSInt Lower, 278 llvm::APSInt Upper) { 279 if (What.isEmpty() || !What.pin(Lower, Upper)) 280 return getEmptySet(); 281 282 ContainerType DummyContainer; 283 284 if (Lower <= Upper) { 285 // [Lower, Upper] is a regular range. 286 // 287 // Shortcut: check that there is even a possibility of the intersection 288 // by checking the two following situations: 289 // 290 // <---[ What ]---[------]------> 291 // Lower Upper 292 // -or- 293 // <----[------]----[ What ]----> 294 // Lower Upper 295 if (What.getMaxValue() < Lower || Upper < What.getMinValue()) 296 return getEmptySet(); 297 298 DummyContainer.push_back( 299 Range(ValueFactory.getValue(Lower), ValueFactory.getValue(Upper))); 300 } else { 301 // [Lower, Upper] is an inverted range, i.e. [MIN, Upper] U [Lower, MAX] 302 // 303 // Shortcut: check that there is even a possibility of the intersection 304 // by checking the following situation: 305 // 306 // <------]---[ What ]---[------> 307 // Upper Lower 308 if (What.getMaxValue() < Lower && Upper < What.getMinValue()) 309 return getEmptySet(); 310 311 DummyContainer.push_back( 312 Range(ValueFactory.getMinValue(Upper), ValueFactory.getValue(Upper))); 313 DummyContainer.push_back( 314 Range(ValueFactory.getValue(Lower), ValueFactory.getMaxValue(Lower))); 315 } 316 317 return intersect(*What.Impl, DummyContainer); 318 } 319 320 RangeSet RangeSet::Factory::intersect(const RangeSet::ContainerType &LHS, 321 const RangeSet::ContainerType &RHS) { 322 ContainerType Result; 323 Result.reserve(std::max(LHS.size(), RHS.size())); 324 325 const_iterator First = LHS.begin(), Second = RHS.begin(), 326 FirstEnd = LHS.end(), SecondEnd = RHS.end(); 327 328 const auto SwapIterators = [&First, &FirstEnd, &Second, &SecondEnd]() { 329 std::swap(First, Second); 330 std::swap(FirstEnd, SecondEnd); 331 }; 332 333 // If we ran out of ranges in one set, but not in the other, 334 // it means that those elements are definitely not in the 335 // intersection. 336 while (First != FirstEnd && Second != SecondEnd) { 337 // We want to keep the following invariant at all times: 338 // 339 // ----[ First ----------------------> 340 // --------[ Second -----------------> 341 if (Second->From() < First->From()) 342 SwapIterators(); 343 344 // Loop where the invariant holds: 345 do { 346 // Check for the following situation: 347 // 348 // ----[ First ]---------------------> 349 // ---------------[ Second ]---------> 350 // 351 // which means that... 352 if (Second->From() > First->To()) { 353 // ...First is not in the intersection. 354 // 355 // We should move on to the next range after First and break out of the 356 // loop because the invariant might not be true. 357 ++First; 358 break; 359 } 360 361 // We have a guaranteed intersection at this point! 362 // And this is the current situation: 363 // 364 // ----[ First ]-----------------> 365 // -------[ Second ------------------> 366 // 367 // Additionally, it definitely starts with Second->From(). 368 const llvm::APSInt &IntersectionStart = Second->From(); 369 370 // It is important to know which of the two ranges' ends 371 // is greater. That "longer" range might have some other 372 // intersections, while the "shorter" range might not. 373 if (Second->To() > First->To()) { 374 // Here we make a decision to keep First as the "longer" 375 // range. 376 SwapIterators(); 377 } 378 379 // At this point, we have the following situation: 380 // 381 // ---- First ]--------------------> 382 // ---- Second ]--[ Second+1 ----------> 383 // 384 // We don't know the relationship between First->From and 385 // Second->From and we don't know whether Second+1 intersects 386 // with First. 387 // 388 // However, we know that [IntersectionStart, Second->To] is 389 // a part of the intersection... 390 Result.push_back(Range(IntersectionStart, Second->To())); 391 ++Second; 392 // ...and that the invariant will hold for a valid Second+1 393 // because First->From <= Second->To < (Second+1)->From. 394 } while (Second != SecondEnd); 395 } 396 397 if (Result.empty()) 398 return getEmptySet(); 399 400 return makePersistent(std::move(Result)); 401 } 402 403 RangeSet RangeSet::Factory::intersect(RangeSet LHS, RangeSet RHS) { 404 // Shortcut: let's see if the intersection is even possible. 405 if (LHS.isEmpty() || RHS.isEmpty() || LHS.getMaxValue() < RHS.getMinValue() || 406 RHS.getMaxValue() < LHS.getMinValue()) 407 return getEmptySet(); 408 409 return intersect(*LHS.Impl, *RHS.Impl); 410 } 411 412 RangeSet RangeSet::Factory::intersect(RangeSet LHS, llvm::APSInt Point) { 413 if (LHS.containsImpl(Point)) 414 return getRangeSet(ValueFactory.getValue(Point)); 415 416 return getEmptySet(); 417 } 418 419 RangeSet RangeSet::Factory::negate(RangeSet What) { 420 if (What.isEmpty()) 421 return getEmptySet(); 422 423 const llvm::APSInt SampleValue = What.getMinValue(); 424 const llvm::APSInt &MIN = ValueFactory.getMinValue(SampleValue); 425 const llvm::APSInt &MAX = ValueFactory.getMaxValue(SampleValue); 426 427 ContainerType Result; 428 Result.reserve(What.size() + (SampleValue == MIN)); 429 430 // Handle a special case for MIN value. 431 const_iterator It = What.begin(); 432 const_iterator End = What.end(); 433 434 const llvm::APSInt &From = It->From(); 435 const llvm::APSInt &To = It->To(); 436 437 if (From == MIN) { 438 // If the range [From, To] is [MIN, MAX], then result is also [MIN, MAX]. 439 if (To == MAX) { 440 return What; 441 } 442 443 const_iterator Last = std::prev(End); 444 445 // Try to find and unite the following ranges: 446 // [MIN, MIN] & [MIN + 1, N] => [MIN, N]. 447 if (Last->To() == MAX) { 448 // It means that in the original range we have ranges 449 // [MIN, A], ... , [B, MAX] 450 // And the result should be [MIN, -B], ..., [-A, MAX] 451 Result.emplace_back(MIN, ValueFactory.getValue(-Last->From())); 452 // We already negated Last, so we can skip it. 453 End = Last; 454 } else { 455 // Add a separate range for the lowest value. 456 Result.emplace_back(MIN, MIN); 457 } 458 459 // Skip adding the second range in case when [From, To] are [MIN, MIN]. 460 if (To != MIN) { 461 Result.emplace_back(ValueFactory.getValue(-To), MAX); 462 } 463 464 // Skip the first range in the loop. 465 ++It; 466 } 467 468 // Negate all other ranges. 469 for (; It != End; ++It) { 470 // Negate int values. 471 const llvm::APSInt &NewFrom = ValueFactory.getValue(-It->To()); 472 const llvm::APSInt &NewTo = ValueFactory.getValue(-It->From()); 473 474 // Add a negated range. 475 Result.emplace_back(NewFrom, NewTo); 476 } 477 478 llvm::sort(Result); 479 return makePersistent(std::move(Result)); 480 } 481 482 RangeSet RangeSet::Factory::deletePoint(RangeSet From, 483 const llvm::APSInt &Point) { 484 if (!From.contains(Point)) 485 return From; 486 487 llvm::APSInt Upper = Point; 488 llvm::APSInt Lower = Point; 489 490 ++Upper; 491 --Lower; 492 493 // Notice that the lower bound is greater than the upper bound. 494 return intersect(From, Upper, Lower); 495 } 496 497 void Range::dump(raw_ostream &OS) const { 498 OS << '[' << toString(From(), 10) << ", " << toString(To(), 10) << ']'; 499 } 500 501 void RangeSet::dump(raw_ostream &OS) const { 502 OS << "{ "; 503 llvm::interleaveComma(*this, OS, [&OS](const Range &R) { R.dump(OS); }); 504 OS << " }"; 505 } 506 507 REGISTER_SET_FACTORY_WITH_PROGRAMSTATE(SymbolSet, SymbolRef) 508 509 namespace { 510 class EquivalenceClass; 511 } // end anonymous namespace 512 513 REGISTER_MAP_WITH_PROGRAMSTATE(ClassMap, SymbolRef, EquivalenceClass) 514 REGISTER_MAP_WITH_PROGRAMSTATE(ClassMembers, EquivalenceClass, SymbolSet) 515 REGISTER_MAP_WITH_PROGRAMSTATE(ConstraintRange, EquivalenceClass, RangeSet) 516 517 REGISTER_SET_FACTORY_WITH_PROGRAMSTATE(ClassSet, EquivalenceClass) 518 REGISTER_MAP_WITH_PROGRAMSTATE(DisequalityMap, EquivalenceClass, ClassSet) 519 520 namespace { 521 /// This class encapsulates a set of symbols equal to each other. 522 /// 523 /// The main idea of the approach requiring such classes is in narrowing 524 /// and sharing constraints between symbols within the class. Also we can 525 /// conclude that there is no practical need in storing constraints for 526 /// every member of the class separately. 527 /// 528 /// Main terminology: 529 /// 530 /// * "Equivalence class" is an object of this class, which can be efficiently 531 /// compared to other classes. It represents the whole class without 532 /// storing the actual in it. The members of the class however can be 533 /// retrieved from the state. 534 /// 535 /// * "Class members" are the symbols corresponding to the class. This means 536 /// that A == B for every member symbols A and B from the class. Members of 537 /// each class are stored in the state. 538 /// 539 /// * "Trivial class" is a class that has and ever had only one same symbol. 540 /// 541 /// * "Merge operation" merges two classes into one. It is the main operation 542 /// to produce non-trivial classes. 543 /// If, at some point, we can assume that two symbols from two distinct 544 /// classes are equal, we can merge these classes. 545 class EquivalenceClass : public llvm::FoldingSetNode { 546 public: 547 /// Find equivalence class for the given symbol in the given state. 548 LLVM_NODISCARD static inline EquivalenceClass find(ProgramStateRef State, 549 SymbolRef Sym); 550 551 /// Merge classes for the given symbols and return a new state. 552 LLVM_NODISCARD static inline ProgramStateRef merge(RangeSet::Factory &F, 553 ProgramStateRef State, 554 SymbolRef First, 555 SymbolRef Second); 556 // Merge this class with the given class and return a new state. 557 LLVM_NODISCARD inline ProgramStateRef 558 merge(RangeSet::Factory &F, ProgramStateRef State, EquivalenceClass Other); 559 560 /// Return a set of class members for the given state. 561 LLVM_NODISCARD inline SymbolSet getClassMembers(ProgramStateRef State) const; 562 563 /// Return true if the current class is trivial in the given state. 564 /// A class is trivial if and only if there is not any member relations stored 565 /// to it in State/ClassMembers. 566 /// An equivalence class with one member might seem as it does not hold any 567 /// meaningful information, i.e. that is a tautology. However, during the 568 /// removal of dead symbols we do not remove classes with one member for 569 /// resource and performance reasons. Consequently, a class with one member is 570 /// not necessarily trivial. It could happen that we have a class with two 571 /// members and then during the removal of dead symbols we remove one of its 572 /// members. In this case, the class is still non-trivial (it still has the 573 /// mappings in ClassMembers), even though it has only one member. 574 LLVM_NODISCARD inline bool isTrivial(ProgramStateRef State) const; 575 576 /// Return true if the current class is trivial and its only member is dead. 577 LLVM_NODISCARD inline bool isTriviallyDead(ProgramStateRef State, 578 SymbolReaper &Reaper) const; 579 580 LLVM_NODISCARD static inline ProgramStateRef 581 markDisequal(RangeSet::Factory &F, ProgramStateRef State, SymbolRef First, 582 SymbolRef Second); 583 LLVM_NODISCARD static inline ProgramStateRef 584 markDisequal(RangeSet::Factory &F, ProgramStateRef State, 585 EquivalenceClass First, EquivalenceClass Second); 586 LLVM_NODISCARD inline ProgramStateRef 587 markDisequal(RangeSet::Factory &F, ProgramStateRef State, 588 EquivalenceClass Other) const; 589 LLVM_NODISCARD static inline ClassSet 590 getDisequalClasses(ProgramStateRef State, SymbolRef Sym); 591 LLVM_NODISCARD inline ClassSet 592 getDisequalClasses(ProgramStateRef State) const; 593 LLVM_NODISCARD inline ClassSet 594 getDisequalClasses(DisequalityMapTy Map, ClassSet::Factory &Factory) const; 595 596 LLVM_NODISCARD static inline Optional<bool> areEqual(ProgramStateRef State, 597 EquivalenceClass First, 598 EquivalenceClass Second); 599 LLVM_NODISCARD static inline Optional<bool> 600 areEqual(ProgramStateRef State, SymbolRef First, SymbolRef Second); 601 602 /// Iterate over all symbols and try to simplify them. 603 LLVM_NODISCARD static inline ProgramStateRef simplify(SValBuilder &SVB, 604 RangeSet::Factory &F, 605 ProgramStateRef State, 606 EquivalenceClass Class); 607 608 void dumpToStream(ProgramStateRef State, raw_ostream &os) const; 609 LLVM_DUMP_METHOD void dump(ProgramStateRef State) const { 610 dumpToStream(State, llvm::errs()); 611 } 612 613 /// Check equivalence data for consistency. 614 LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED static bool 615 isClassDataConsistent(ProgramStateRef State); 616 617 LLVM_NODISCARD QualType getType() const { 618 return getRepresentativeSymbol()->getType(); 619 } 620 621 EquivalenceClass() = delete; 622 EquivalenceClass(const EquivalenceClass &) = default; 623 EquivalenceClass &operator=(const EquivalenceClass &) = delete; 624 EquivalenceClass(EquivalenceClass &&) = default; 625 EquivalenceClass &operator=(EquivalenceClass &&) = delete; 626 627 bool operator==(const EquivalenceClass &Other) const { 628 return ID == Other.ID; 629 } 630 bool operator<(const EquivalenceClass &Other) const { return ID < Other.ID; } 631 bool operator!=(const EquivalenceClass &Other) const { 632 return !operator==(Other); 633 } 634 635 static void Profile(llvm::FoldingSetNodeID &ID, uintptr_t CID) { 636 ID.AddInteger(CID); 637 } 638 639 void Profile(llvm::FoldingSetNodeID &ID) const { Profile(ID, this->ID); } 640 641 private: 642 /* implicit */ EquivalenceClass(SymbolRef Sym) 643 : ID(reinterpret_cast<uintptr_t>(Sym)) {} 644 645 /// This function is intended to be used ONLY within the class. 646 /// The fact that ID is a pointer to a symbol is an implementation detail 647 /// and should stay that way. 648 /// In the current implementation, we use it to retrieve the only member 649 /// of the trivial class. 650 SymbolRef getRepresentativeSymbol() const { 651 return reinterpret_cast<SymbolRef>(ID); 652 } 653 static inline SymbolSet::Factory &getMembersFactory(ProgramStateRef State); 654 655 inline ProgramStateRef mergeImpl(RangeSet::Factory &F, ProgramStateRef State, 656 SymbolSet Members, EquivalenceClass Other, 657 SymbolSet OtherMembers); 658 static inline bool 659 addToDisequalityInfo(DisequalityMapTy &Info, ConstraintRangeTy &Constraints, 660 RangeSet::Factory &F, ProgramStateRef State, 661 EquivalenceClass First, EquivalenceClass Second); 662 663 /// This is a unique identifier of the class. 664 uintptr_t ID; 665 }; 666 667 //===----------------------------------------------------------------------===// 668 // Constraint functions 669 //===----------------------------------------------------------------------===// 670 671 LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED bool 672 areFeasible(ConstraintRangeTy Constraints) { 673 return llvm::none_of( 674 Constraints, 675 [](const std::pair<EquivalenceClass, RangeSet> &ClassConstraint) { 676 return ClassConstraint.second.isEmpty(); 677 }); 678 } 679 680 LLVM_NODISCARD inline const RangeSet *getConstraint(ProgramStateRef State, 681 EquivalenceClass Class) { 682 return State->get<ConstraintRange>(Class); 683 } 684 685 LLVM_NODISCARD inline const RangeSet *getConstraint(ProgramStateRef State, 686 SymbolRef Sym) { 687 return getConstraint(State, EquivalenceClass::find(State, Sym)); 688 } 689 690 LLVM_NODISCARD ProgramStateRef setConstraint(ProgramStateRef State, 691 EquivalenceClass Class, 692 RangeSet Constraint) { 693 return State->set<ConstraintRange>(Class, Constraint); 694 } 695 696 LLVM_NODISCARD ProgramStateRef setConstraints(ProgramStateRef State, 697 ConstraintRangeTy Constraints) { 698 return State->set<ConstraintRange>(Constraints); 699 } 700 701 //===----------------------------------------------------------------------===// 702 // Equality/diseqiality abstraction 703 //===----------------------------------------------------------------------===// 704 705 /// A small helper function for detecting symbolic (dis)equality. 706 /// 707 /// Equality check can have different forms (like a == b or a - b) and this 708 /// class encapsulates those away if the only thing the user wants to check - 709 /// whether it's equality/diseqiality or not. 710 /// 711 /// \returns true if assuming this Sym to be true means equality of operands 712 /// false if it means disequality of operands 713 /// None otherwise 714 Optional<bool> meansEquality(const SymSymExpr *Sym) { 715 switch (Sym->getOpcode()) { 716 case BO_Sub: 717 // This case is: A - B != 0 -> disequality check. 718 return false; 719 case BO_EQ: 720 // This case is: A == B != 0 -> equality check. 721 return true; 722 case BO_NE: 723 // This case is: A != B != 0 -> diseqiality check. 724 return false; 725 default: 726 return llvm::None; 727 } 728 } 729 730 //===----------------------------------------------------------------------===// 731 // Intersection functions 732 //===----------------------------------------------------------------------===// 733 734 template <class SecondTy, class... RestTy> 735 LLVM_NODISCARD inline RangeSet intersect(RangeSet::Factory &F, RangeSet Head, 736 SecondTy Second, RestTy... Tail); 737 738 template <class... RangeTy> struct IntersectionTraits; 739 740 template <class... TailTy> struct IntersectionTraits<RangeSet, TailTy...> { 741 // Found RangeSet, no need to check any further 742 using Type = RangeSet; 743 }; 744 745 template <> struct IntersectionTraits<> { 746 // We ran out of types, and we didn't find any RangeSet, so the result should 747 // be optional. 748 using Type = Optional<RangeSet>; 749 }; 750 751 template <class OptionalOrPointer, class... TailTy> 752 struct IntersectionTraits<OptionalOrPointer, TailTy...> { 753 // If current type is Optional or a raw pointer, we should keep looking. 754 using Type = typename IntersectionTraits<TailTy...>::Type; 755 }; 756 757 template <class EndTy> 758 LLVM_NODISCARD inline EndTy intersect(RangeSet::Factory &F, EndTy End) { 759 // If the list contains only RangeSet or Optional<RangeSet>, simply return 760 // that range set. 761 return End; 762 } 763 764 LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED inline Optional<RangeSet> 765 intersect(RangeSet::Factory &F, const RangeSet *End) { 766 // This is an extraneous conversion from a raw pointer into Optional<RangeSet> 767 if (End) { 768 return *End; 769 } 770 return llvm::None; 771 } 772 773 template <class... RestTy> 774 LLVM_NODISCARD inline RangeSet intersect(RangeSet::Factory &F, RangeSet Head, 775 RangeSet Second, RestTy... Tail) { 776 // Here we call either the <RangeSet,RangeSet,...> or <RangeSet,...> version 777 // of the function and can be sure that the result is RangeSet. 778 return intersect(F, F.intersect(Head, Second), Tail...); 779 } 780 781 template <class SecondTy, class... RestTy> 782 LLVM_NODISCARD inline RangeSet intersect(RangeSet::Factory &F, RangeSet Head, 783 SecondTy Second, RestTy... Tail) { 784 if (Second) { 785 // Here we call the <RangeSet,RangeSet,...> version of the function... 786 return intersect(F, Head, *Second, Tail...); 787 } 788 // ...and here it is either <RangeSet,RangeSet,...> or <RangeSet,...>, which 789 // means that the result is definitely RangeSet. 790 return intersect(F, Head, Tail...); 791 } 792 793 /// Main generic intersect function. 794 /// It intersects all of the given range sets. If some of the given arguments 795 /// don't hold a range set (nullptr or llvm::None), the function will skip them. 796 /// 797 /// Available representations for the arguments are: 798 /// * RangeSet 799 /// * Optional<RangeSet> 800 /// * RangeSet * 801 /// Pointer to a RangeSet is automatically assumed to be nullable and will get 802 /// checked as well as the optional version. If this behaviour is undesired, 803 /// please dereference the pointer in the call. 804 /// 805 /// Return type depends on the arguments' types. If we can be sure in compile 806 /// time that there will be a range set as a result, the returning type is 807 /// simply RangeSet, in other cases we have to back off to Optional<RangeSet>. 808 /// 809 /// Please, prefer optional range sets to raw pointers. If the last argument is 810 /// a raw pointer and all previous arguments are None, it will cost one 811 /// additional check to convert RangeSet * into Optional<RangeSet>. 812 template <class HeadTy, class SecondTy, class... RestTy> 813 LLVM_NODISCARD inline 814 typename IntersectionTraits<HeadTy, SecondTy, RestTy...>::Type 815 intersect(RangeSet::Factory &F, HeadTy Head, SecondTy Second, 816 RestTy... Tail) { 817 if (Head) { 818 return intersect(F, *Head, Second, Tail...); 819 } 820 return intersect(F, Second, Tail...); 821 } 822 823 //===----------------------------------------------------------------------===// 824 // Symbolic reasoning logic 825 //===----------------------------------------------------------------------===// 826 827 /// A little component aggregating all of the reasoning we have about 828 /// the ranges of symbolic expressions. 829 /// 830 /// Even when we don't know the exact values of the operands, we still 831 /// can get a pretty good estimate of the result's range. 832 class SymbolicRangeInferrer 833 : public SymExprVisitor<SymbolicRangeInferrer, RangeSet> { 834 public: 835 template <class SourceType> 836 static RangeSet inferRange(RangeSet::Factory &F, ProgramStateRef State, 837 SourceType Origin) { 838 SymbolicRangeInferrer Inferrer(F, State); 839 return Inferrer.infer(Origin); 840 } 841 842 RangeSet VisitSymExpr(SymbolRef Sym) { 843 // If we got to this function, the actual type of the symbolic 844 // expression is not supported for advanced inference. 845 // In this case, we simply backoff to the default "let's simply 846 // infer the range from the expression's type". 847 return infer(Sym->getType()); 848 } 849 850 RangeSet VisitSymIntExpr(const SymIntExpr *Sym) { 851 return VisitBinaryOperator(Sym); 852 } 853 854 RangeSet VisitIntSymExpr(const IntSymExpr *Sym) { 855 return VisitBinaryOperator(Sym); 856 } 857 858 RangeSet VisitSymSymExpr(const SymSymExpr *Sym) { 859 return intersect( 860 RangeFactory, 861 // If Sym is (dis)equality, we might have some information 862 // on that in our equality classes data structure. 863 getRangeForEqualities(Sym), 864 // And we should always check what we can get from the operands. 865 VisitBinaryOperator(Sym)); 866 } 867 868 private: 869 SymbolicRangeInferrer(RangeSet::Factory &F, ProgramStateRef S) 870 : ValueFactory(F.getValueFactory()), RangeFactory(F), State(S) {} 871 872 /// Infer range information from the given integer constant. 873 /// 874 /// It's not a real "inference", but is here for operating with 875 /// sub-expressions in a more polymorphic manner. 876 RangeSet inferAs(const llvm::APSInt &Val, QualType) { 877 return {RangeFactory, Val}; 878 } 879 880 /// Infer range information from symbol in the context of the given type. 881 RangeSet inferAs(SymbolRef Sym, QualType DestType) { 882 QualType ActualType = Sym->getType(); 883 // Check that we can reason about the symbol at all. 884 if (ActualType->isIntegralOrEnumerationType() || 885 Loc::isLocType(ActualType)) { 886 return infer(Sym); 887 } 888 // Otherwise, let's simply infer from the destination type. 889 // We couldn't figure out nothing else about that expression. 890 return infer(DestType); 891 } 892 893 RangeSet infer(SymbolRef Sym) { 894 return intersect( 895 RangeFactory, 896 // Of course, we should take the constraint directly associated with 897 // this symbol into consideration. 898 getConstraint(State, Sym), 899 // If Sym is a difference of symbols A - B, then maybe we have range 900 // set stored for B - A. 901 // 902 // If we have range set stored for both A - B and B - A then 903 // calculate the effective range set by intersecting the range set 904 // for A - B and the negated range set of B - A. 905 getRangeForNegatedSub(Sym), 906 // If Sym is a comparison expression (except <=>), 907 // find any other comparisons with the same operands. 908 // See function description. 909 getRangeForComparisonSymbol(Sym), 910 // Apart from the Sym itself, we can infer quite a lot if we look 911 // into subexpressions of Sym. 912 Visit(Sym)); 913 } 914 915 RangeSet infer(EquivalenceClass Class) { 916 if (const RangeSet *AssociatedConstraint = getConstraint(State, Class)) 917 return *AssociatedConstraint; 918 919 return infer(Class.getType()); 920 } 921 922 /// Infer range information solely from the type. 923 RangeSet infer(QualType T) { 924 // Lazily generate a new RangeSet representing all possible values for the 925 // given symbol type. 926 RangeSet Result(RangeFactory, ValueFactory.getMinValue(T), 927 ValueFactory.getMaxValue(T)); 928 929 // References are known to be non-zero. 930 if (T->isReferenceType()) 931 return assumeNonZero(Result, T); 932 933 return Result; 934 } 935 936 template <class BinarySymExprTy> 937 RangeSet VisitBinaryOperator(const BinarySymExprTy *Sym) { 938 // TODO #1: VisitBinaryOperator implementation might not make a good 939 // use of the inferred ranges. In this case, we might be calculating 940 // everything for nothing. This being said, we should introduce some 941 // sort of laziness mechanism here. 942 // 943 // TODO #2: We didn't go into the nested expressions before, so it 944 // might cause us spending much more time doing the inference. 945 // This can be a problem for deeply nested expressions that are 946 // involved in conditions and get tested continuously. We definitely 947 // need to address this issue and introduce some sort of caching 948 // in here. 949 QualType ResultType = Sym->getType(); 950 return VisitBinaryOperator(inferAs(Sym->getLHS(), ResultType), 951 Sym->getOpcode(), 952 inferAs(Sym->getRHS(), ResultType), ResultType); 953 } 954 955 RangeSet VisitBinaryOperator(RangeSet LHS, BinaryOperator::Opcode Op, 956 RangeSet RHS, QualType T) { 957 switch (Op) { 958 case BO_Or: 959 return VisitBinaryOperator<BO_Or>(LHS, RHS, T); 960 case BO_And: 961 return VisitBinaryOperator<BO_And>(LHS, RHS, T); 962 case BO_Rem: 963 return VisitBinaryOperator<BO_Rem>(LHS, RHS, T); 964 default: 965 return infer(T); 966 } 967 } 968 969 //===----------------------------------------------------------------------===// 970 // Ranges and operators 971 //===----------------------------------------------------------------------===// 972 973 /// Return a rough approximation of the given range set. 974 /// 975 /// For the range set: 976 /// { [x_0, y_0], [x_1, y_1], ... , [x_N, y_N] } 977 /// it will return the range [x_0, y_N]. 978 static Range fillGaps(RangeSet Origin) { 979 assert(!Origin.isEmpty()); 980 return {Origin.getMinValue(), Origin.getMaxValue()}; 981 } 982 983 /// Try to convert given range into the given type. 984 /// 985 /// It will return llvm::None only when the trivial conversion is possible. 986 llvm::Optional<Range> convert(const Range &Origin, APSIntType To) { 987 if (To.testInRange(Origin.From(), false) != APSIntType::RTR_Within || 988 To.testInRange(Origin.To(), false) != APSIntType::RTR_Within) { 989 return llvm::None; 990 } 991 return Range(ValueFactory.Convert(To, Origin.From()), 992 ValueFactory.Convert(To, Origin.To())); 993 } 994 995 template <BinaryOperator::Opcode Op> 996 RangeSet VisitBinaryOperator(RangeSet LHS, RangeSet RHS, QualType T) { 997 // We should propagate information about unfeasbility of one of the 998 // operands to the resulting range. 999 if (LHS.isEmpty() || RHS.isEmpty()) { 1000 return RangeFactory.getEmptySet(); 1001 } 1002 1003 Range CoarseLHS = fillGaps(LHS); 1004 Range CoarseRHS = fillGaps(RHS); 1005 1006 APSIntType ResultType = ValueFactory.getAPSIntType(T); 1007 1008 // We need to convert ranges to the resulting type, so we can compare values 1009 // and combine them in a meaningful (in terms of the given operation) way. 1010 auto ConvertedCoarseLHS = convert(CoarseLHS, ResultType); 1011 auto ConvertedCoarseRHS = convert(CoarseRHS, ResultType); 1012 1013 // It is hard to reason about ranges when conversion changes 1014 // borders of the ranges. 1015 if (!ConvertedCoarseLHS || !ConvertedCoarseRHS) { 1016 return infer(T); 1017 } 1018 1019 return VisitBinaryOperator<Op>(*ConvertedCoarseLHS, *ConvertedCoarseRHS, T); 1020 } 1021 1022 template <BinaryOperator::Opcode Op> 1023 RangeSet VisitBinaryOperator(Range LHS, Range RHS, QualType T) { 1024 return infer(T); 1025 } 1026 1027 /// Return a symmetrical range for the given range and type. 1028 /// 1029 /// If T is signed, return the smallest range [-x..x] that covers the original 1030 /// range, or [-min(T), max(T)] if the aforementioned symmetric range doesn't 1031 /// exist due to original range covering min(T)). 1032 /// 1033 /// If T is unsigned, return the smallest range [0..x] that covers the 1034 /// original range. 1035 Range getSymmetricalRange(Range Origin, QualType T) { 1036 APSIntType RangeType = ValueFactory.getAPSIntType(T); 1037 1038 if (RangeType.isUnsigned()) { 1039 return Range(ValueFactory.getMinValue(RangeType), Origin.To()); 1040 } 1041 1042 if (Origin.From().isMinSignedValue()) { 1043 // If mini is a minimal signed value, absolute value of it is greater 1044 // than the maximal signed value. In order to avoid these 1045 // complications, we simply return the whole range. 1046 return {ValueFactory.getMinValue(RangeType), 1047 ValueFactory.getMaxValue(RangeType)}; 1048 } 1049 1050 // At this point, we are sure that the type is signed and we can safely 1051 // use unary - operator. 1052 // 1053 // While calculating absolute maximum, we can use the following formula 1054 // because of these reasons: 1055 // * If From >= 0 then To >= From and To >= -From. 1056 // AbsMax == To == max(To, -From) 1057 // * If To <= 0 then -From >= -To and -From >= From. 1058 // AbsMax == -From == max(-From, To) 1059 // * Otherwise, From <= 0, To >= 0, and 1060 // AbsMax == max(abs(From), abs(To)) 1061 llvm::APSInt AbsMax = std::max(-Origin.From(), Origin.To()); 1062 1063 // Intersection is guaranteed to be non-empty. 1064 return {ValueFactory.getValue(-AbsMax), ValueFactory.getValue(AbsMax)}; 1065 } 1066 1067 /// Return a range set subtracting zero from \p Domain. 1068 RangeSet assumeNonZero(RangeSet Domain, QualType T) { 1069 APSIntType IntType = ValueFactory.getAPSIntType(T); 1070 return RangeFactory.deletePoint(Domain, IntType.getZeroValue()); 1071 } 1072 1073 // FIXME: Once SValBuilder supports unary minus, we should use SValBuilder to 1074 // obtain the negated symbolic expression instead of constructing the 1075 // symbol manually. This will allow us to support finding ranges of not 1076 // only negated SymSymExpr-type expressions, but also of other, simpler 1077 // expressions which we currently do not know how to negate. 1078 Optional<RangeSet> getRangeForNegatedSub(SymbolRef Sym) { 1079 if (const SymSymExpr *SSE = dyn_cast<SymSymExpr>(Sym)) { 1080 if (SSE->getOpcode() == BO_Sub) { 1081 QualType T = Sym->getType(); 1082 1083 // Do not negate unsigned ranges 1084 if (!T->isUnsignedIntegerOrEnumerationType() && 1085 !T->isSignedIntegerOrEnumerationType()) 1086 return llvm::None; 1087 1088 SymbolManager &SymMgr = State->getSymbolManager(); 1089 SymbolRef NegatedSym = 1090 SymMgr.getSymSymExpr(SSE->getRHS(), BO_Sub, SSE->getLHS(), T); 1091 1092 if (const RangeSet *NegatedRange = getConstraint(State, NegatedSym)) { 1093 return RangeFactory.negate(*NegatedRange); 1094 } 1095 } 1096 } 1097 return llvm::None; 1098 } 1099 1100 // Returns ranges only for binary comparison operators (except <=>) 1101 // when left and right operands are symbolic values. 1102 // Finds any other comparisons with the same operands. 1103 // Then do logical calculations and refuse impossible branches. 1104 // E.g. (x < y) and (x > y) at the same time are impossible. 1105 // E.g. (x >= y) and (x != y) at the same time makes (x > y) true only. 1106 // E.g. (x == y) and (y == x) are just reversed but the same. 1107 // It covers all possible combinations (see CmpOpTable description). 1108 // Note that `x` and `y` can also stand for subexpressions, 1109 // not only for actual symbols. 1110 Optional<RangeSet> getRangeForComparisonSymbol(SymbolRef Sym) { 1111 const auto *SSE = dyn_cast<SymSymExpr>(Sym); 1112 if (!SSE) 1113 return llvm::None; 1114 1115 const BinaryOperatorKind CurrentOP = SSE->getOpcode(); 1116 1117 // We currently do not support <=> (C++20). 1118 if (!BinaryOperator::isComparisonOp(CurrentOP) || (CurrentOP == BO_Cmp)) 1119 return llvm::None; 1120 1121 static const OperatorRelationsTable CmpOpTable{}; 1122 1123 const SymExpr *LHS = SSE->getLHS(); 1124 const SymExpr *RHS = SSE->getRHS(); 1125 QualType T = SSE->getType(); 1126 1127 SymbolManager &SymMgr = State->getSymbolManager(); 1128 1129 // We use this variable to store the last queried operator (`QueriedOP`) 1130 // for which the `getCmpOpState` returned with `Unknown`. If there are two 1131 // different OPs that returned `Unknown` then we have to query the special 1132 // `UnknownX2` column. We assume that `getCmpOpState(CurrentOP, CurrentOP)` 1133 // never returns `Unknown`, so `CurrentOP` is a good initial value. 1134 BinaryOperatorKind LastQueriedOpToUnknown = CurrentOP; 1135 1136 // Loop goes through all of the columns exept the last one ('UnknownX2'). 1137 // We treat `UnknownX2` column separately at the end of the loop body. 1138 for (size_t i = 0; i < CmpOpTable.getCmpOpCount(); ++i) { 1139 1140 // Let's find an expression e.g. (x < y). 1141 BinaryOperatorKind QueriedOP = OperatorRelationsTable::getOpFromIndex(i); 1142 const SymSymExpr *SymSym = SymMgr.getSymSymExpr(LHS, QueriedOP, RHS, T); 1143 const RangeSet *QueriedRangeSet = getConstraint(State, SymSym); 1144 1145 // If ranges were not previously found, 1146 // try to find a reversed expression (y > x). 1147 if (!QueriedRangeSet) { 1148 const BinaryOperatorKind ROP = 1149 BinaryOperator::reverseComparisonOp(QueriedOP); 1150 SymSym = SymMgr.getSymSymExpr(RHS, ROP, LHS, T); 1151 QueriedRangeSet = getConstraint(State, SymSym); 1152 } 1153 1154 if (!QueriedRangeSet || QueriedRangeSet->isEmpty()) 1155 continue; 1156 1157 const llvm::APSInt *ConcreteValue = QueriedRangeSet->getConcreteValue(); 1158 const bool isInFalseBranch = 1159 ConcreteValue ? (*ConcreteValue == 0) : false; 1160 1161 // If it is a false branch, we shall be guided by opposite operator, 1162 // because the table is made assuming we are in the true branch. 1163 // E.g. when (x <= y) is false, then (x > y) is true. 1164 if (isInFalseBranch) 1165 QueriedOP = BinaryOperator::negateComparisonOp(QueriedOP); 1166 1167 OperatorRelationsTable::TriStateKind BranchState = 1168 CmpOpTable.getCmpOpState(CurrentOP, QueriedOP); 1169 1170 if (BranchState == OperatorRelationsTable::Unknown) { 1171 if (LastQueriedOpToUnknown != CurrentOP && 1172 LastQueriedOpToUnknown != QueriedOP) { 1173 // If we got the Unknown state for both different operators. 1174 // if (x <= y) // assume true 1175 // if (x != y) // assume true 1176 // if (x < y) // would be also true 1177 // Get a state from `UnknownX2` column. 1178 BranchState = CmpOpTable.getCmpOpStateForUnknownX2(CurrentOP); 1179 } else { 1180 LastQueriedOpToUnknown = QueriedOP; 1181 continue; 1182 } 1183 } 1184 1185 return (BranchState == OperatorRelationsTable::True) ? getTrueRange(T) 1186 : getFalseRange(T); 1187 } 1188 1189 return llvm::None; 1190 } 1191 1192 Optional<RangeSet> getRangeForEqualities(const SymSymExpr *Sym) { 1193 Optional<bool> Equality = meansEquality(Sym); 1194 1195 if (!Equality) 1196 return llvm::None; 1197 1198 if (Optional<bool> AreEqual = 1199 EquivalenceClass::areEqual(State, Sym->getLHS(), Sym->getRHS())) { 1200 // Here we cover two cases at once: 1201 // * if Sym is equality and its operands are known to be equal -> true 1202 // * if Sym is disequality and its operands are disequal -> true 1203 if (*AreEqual == *Equality) { 1204 return getTrueRange(Sym->getType()); 1205 } 1206 // Opposite combinations result in false. 1207 return getFalseRange(Sym->getType()); 1208 } 1209 1210 return llvm::None; 1211 } 1212 1213 RangeSet getTrueRange(QualType T) { 1214 RangeSet TypeRange = infer(T); 1215 return assumeNonZero(TypeRange, T); 1216 } 1217 1218 RangeSet getFalseRange(QualType T) { 1219 const llvm::APSInt &Zero = ValueFactory.getValue(0, T); 1220 return RangeSet(RangeFactory, Zero); 1221 } 1222 1223 BasicValueFactory &ValueFactory; 1224 RangeSet::Factory &RangeFactory; 1225 ProgramStateRef State; 1226 }; 1227 1228 //===----------------------------------------------------------------------===// 1229 // Range-based reasoning about symbolic operations 1230 //===----------------------------------------------------------------------===// 1231 1232 template <> 1233 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_Or>(Range LHS, Range RHS, 1234 QualType T) { 1235 APSIntType ResultType = ValueFactory.getAPSIntType(T); 1236 llvm::APSInt Zero = ResultType.getZeroValue(); 1237 1238 bool IsLHSPositiveOrZero = LHS.From() >= Zero; 1239 bool IsRHSPositiveOrZero = RHS.From() >= Zero; 1240 1241 bool IsLHSNegative = LHS.To() < Zero; 1242 bool IsRHSNegative = RHS.To() < Zero; 1243 1244 // Check if both ranges have the same sign. 1245 if ((IsLHSPositiveOrZero && IsRHSPositiveOrZero) || 1246 (IsLHSNegative && IsRHSNegative)) { 1247 // The result is definitely greater or equal than any of the operands. 1248 const llvm::APSInt &Min = std::max(LHS.From(), RHS.From()); 1249 1250 // We estimate maximal value for positives as the maximal value for the 1251 // given type. For negatives, we estimate it with -1 (e.g. 0x11111111). 1252 // 1253 // TODO: We basically, limit the resulting range from below, but don't do 1254 // anything with the upper bound. 1255 // 1256 // For positive operands, it can be done as follows: for the upper 1257 // bound of LHS and RHS we calculate the most significant bit set. 1258 // Let's call it the N-th bit. Then we can estimate the maximal 1259 // number to be 2^(N+1)-1, i.e. the number with all the bits up to 1260 // the N-th bit set. 1261 const llvm::APSInt &Max = IsLHSNegative 1262 ? ValueFactory.getValue(--Zero) 1263 : ValueFactory.getMaxValue(ResultType); 1264 1265 return {RangeFactory, ValueFactory.getValue(Min), Max}; 1266 } 1267 1268 // Otherwise, let's check if at least one of the operands is negative. 1269 if (IsLHSNegative || IsRHSNegative) { 1270 // This means that the result is definitely negative as well. 1271 return {RangeFactory, ValueFactory.getMinValue(ResultType), 1272 ValueFactory.getValue(--Zero)}; 1273 } 1274 1275 RangeSet DefaultRange = infer(T); 1276 1277 // It is pretty hard to reason about operands with different signs 1278 // (and especially with possibly different signs). We simply check if it 1279 // can be zero. In order to conclude that the result could not be zero, 1280 // at least one of the operands should be definitely not zero itself. 1281 if (!LHS.Includes(Zero) || !RHS.Includes(Zero)) { 1282 return assumeNonZero(DefaultRange, T); 1283 } 1284 1285 // Nothing much else to do here. 1286 return DefaultRange; 1287 } 1288 1289 template <> 1290 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_And>(Range LHS, 1291 Range RHS, 1292 QualType T) { 1293 APSIntType ResultType = ValueFactory.getAPSIntType(T); 1294 llvm::APSInt Zero = ResultType.getZeroValue(); 1295 1296 bool IsLHSPositiveOrZero = LHS.From() >= Zero; 1297 bool IsRHSPositiveOrZero = RHS.From() >= Zero; 1298 1299 bool IsLHSNegative = LHS.To() < Zero; 1300 bool IsRHSNegative = RHS.To() < Zero; 1301 1302 // Check if both ranges have the same sign. 1303 if ((IsLHSPositiveOrZero && IsRHSPositiveOrZero) || 1304 (IsLHSNegative && IsRHSNegative)) { 1305 // The result is definitely less or equal than any of the operands. 1306 const llvm::APSInt &Max = std::min(LHS.To(), RHS.To()); 1307 1308 // We conservatively estimate lower bound to be the smallest positive 1309 // or negative value corresponding to the sign of the operands. 1310 const llvm::APSInt &Min = IsLHSNegative 1311 ? ValueFactory.getMinValue(ResultType) 1312 : ValueFactory.getValue(Zero); 1313 1314 return {RangeFactory, Min, Max}; 1315 } 1316 1317 // Otherwise, let's check if at least one of the operands is positive. 1318 if (IsLHSPositiveOrZero || IsRHSPositiveOrZero) { 1319 // This makes result definitely positive. 1320 // 1321 // We can also reason about a maximal value by finding the maximal 1322 // value of the positive operand. 1323 const llvm::APSInt &Max = IsLHSPositiveOrZero ? LHS.To() : RHS.To(); 1324 1325 // The minimal value on the other hand is much harder to reason about. 1326 // The only thing we know for sure is that the result is positive. 1327 return {RangeFactory, ValueFactory.getValue(Zero), 1328 ValueFactory.getValue(Max)}; 1329 } 1330 1331 // Nothing much else to do here. 1332 return infer(T); 1333 } 1334 1335 template <> 1336 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_Rem>(Range LHS, 1337 Range RHS, 1338 QualType T) { 1339 llvm::APSInt Zero = ValueFactory.getAPSIntType(T).getZeroValue(); 1340 1341 Range ConservativeRange = getSymmetricalRange(RHS, T); 1342 1343 llvm::APSInt Max = ConservativeRange.To(); 1344 llvm::APSInt Min = ConservativeRange.From(); 1345 1346 if (Max == Zero) { 1347 // It's an undefined behaviour to divide by 0 and it seems like we know 1348 // for sure that RHS is 0. Let's say that the resulting range is 1349 // simply infeasible for that matter. 1350 return RangeFactory.getEmptySet(); 1351 } 1352 1353 // At this point, our conservative range is closed. The result, however, 1354 // couldn't be greater than the RHS' maximal absolute value. Because of 1355 // this reason, we turn the range into open (or half-open in case of 1356 // unsigned integers). 1357 // 1358 // While we operate on integer values, an open interval (a, b) can be easily 1359 // represented by the closed interval [a + 1, b - 1]. And this is exactly 1360 // what we do next. 1361 // 1362 // If we are dealing with unsigned case, we shouldn't move the lower bound. 1363 if (Min.isSigned()) { 1364 ++Min; 1365 } 1366 --Max; 1367 1368 bool IsLHSPositiveOrZero = LHS.From() >= Zero; 1369 bool IsRHSPositiveOrZero = RHS.From() >= Zero; 1370 1371 // Remainder operator results with negative operands is implementation 1372 // defined. Positive cases are much easier to reason about though. 1373 if (IsLHSPositiveOrZero && IsRHSPositiveOrZero) { 1374 // If maximal value of LHS is less than maximal value of RHS, 1375 // the result won't get greater than LHS.To(). 1376 Max = std::min(LHS.To(), Max); 1377 // We want to check if it is a situation similar to the following: 1378 // 1379 // <------------|---[ LHS ]--------[ RHS ]-----> 1380 // -INF 0 +INF 1381 // 1382 // In this situation, we can conclude that (LHS / RHS) == 0 and 1383 // (LHS % RHS) == LHS. 1384 Min = LHS.To() < RHS.From() ? LHS.From() : Zero; 1385 } 1386 1387 // Nevertheless, the symmetrical range for RHS is a conservative estimate 1388 // for any sign of either LHS, or RHS. 1389 return {RangeFactory, ValueFactory.getValue(Min), ValueFactory.getValue(Max)}; 1390 } 1391 1392 //===----------------------------------------------------------------------===// 1393 // Constraint assignment logic 1394 //===----------------------------------------------------------------------===// 1395 1396 /// ConstraintAssignorBase is a small utility class that unifies visitor 1397 /// for ranges with a visitor for constraints (rangeset/range/constant). 1398 /// 1399 /// It is designed to have one derived class, but generally it can have more. 1400 /// Derived class can control which types we handle by defining methods of the 1401 /// following form: 1402 /// 1403 /// bool handle${SYMBOL}To${CONSTRAINT}(const SYMBOL *Sym, 1404 /// CONSTRAINT Constraint); 1405 /// 1406 /// where SYMBOL is the type of the symbol (e.g. SymSymExpr, SymbolCast, etc.) 1407 /// CONSTRAINT is the type of constraint (RangeSet/Range/Const) 1408 /// return value signifies whether we should try other handle methods 1409 /// (i.e. false would mean to stop right after calling this method) 1410 template <class Derived> class ConstraintAssignorBase { 1411 public: 1412 using Const = const llvm::APSInt &; 1413 1414 #define DISPATCH(CLASS) return assign##CLASS##Impl(cast<CLASS>(Sym), Constraint) 1415 1416 #define ASSIGN(CLASS, TO, SYM, CONSTRAINT) \ 1417 if (!static_cast<Derived *>(this)->assign##CLASS##To##TO(SYM, CONSTRAINT)) \ 1418 return false 1419 1420 void assign(SymbolRef Sym, RangeSet Constraint) { 1421 assignImpl(Sym, Constraint); 1422 } 1423 1424 bool assignImpl(SymbolRef Sym, RangeSet Constraint) { 1425 switch (Sym->getKind()) { 1426 #define SYMBOL(Id, Parent) \ 1427 case SymExpr::Id##Kind: \ 1428 DISPATCH(Id); 1429 #include "clang/StaticAnalyzer/Core/PathSensitive/Symbols.def" 1430 } 1431 llvm_unreachable("Unknown SymExpr kind!"); 1432 } 1433 1434 #define DEFAULT_ASSIGN(Id) \ 1435 bool assign##Id##To##RangeSet(const Id *Sym, RangeSet Constraint) { \ 1436 return true; \ 1437 } \ 1438 bool assign##Id##To##Range(const Id *Sym, Range Constraint) { return true; } \ 1439 bool assign##Id##To##Const(const Id *Sym, Const Constraint) { return true; } 1440 1441 // When we dispatch for constraint types, we first try to check 1442 // if the new constraint is the constant and try the corresponding 1443 // assignor methods. If it didn't interrupt, we can proceed to the 1444 // range, and finally to the range set. 1445 #define CONSTRAINT_DISPATCH(Id) \ 1446 if (const llvm::APSInt *Const = Constraint.getConcreteValue()) { \ 1447 ASSIGN(Id, Const, Sym, *Const); \ 1448 } \ 1449 if (Constraint.size() == 1) { \ 1450 ASSIGN(Id, Range, Sym, *Constraint.begin()); \ 1451 } \ 1452 ASSIGN(Id, RangeSet, Sym, Constraint) 1453 1454 // Our internal assign method first tries to call assignor methods for all 1455 // constraint types that apply. And if not interrupted, continues with its 1456 // parent class. 1457 #define SYMBOL(Id, Parent) \ 1458 bool assign##Id##Impl(const Id *Sym, RangeSet Constraint) { \ 1459 CONSTRAINT_DISPATCH(Id); \ 1460 DISPATCH(Parent); \ 1461 } \ 1462 DEFAULT_ASSIGN(Id) 1463 #define ABSTRACT_SYMBOL(Id, Parent) SYMBOL(Id, Parent) 1464 #include "clang/StaticAnalyzer/Core/PathSensitive/Symbols.def" 1465 1466 // Default implementations for the top class that doesn't have parents. 1467 bool assignSymExprImpl(const SymExpr *Sym, RangeSet Constraint) { 1468 CONSTRAINT_DISPATCH(SymExpr); 1469 return true; 1470 } 1471 DEFAULT_ASSIGN(SymExpr); 1472 1473 #undef DISPATCH 1474 #undef CONSTRAINT_DISPATCH 1475 #undef DEFAULT_ASSIGN 1476 #undef ASSIGN 1477 }; 1478 1479 /// A little component aggregating all of the reasoning we have about 1480 /// assigning new constraints to symbols. 1481 /// 1482 /// The main purpose of this class is to associate constraints to symbols, 1483 /// and impose additional constraints on other symbols, when we can imply 1484 /// them. 1485 /// 1486 /// It has a nice symmetry with SymbolicRangeInferrer. When the latter 1487 /// can provide more precise ranges by looking into the operands of the 1488 /// expression in question, ConstraintAssignor looks into the operands 1489 /// to see if we can imply more from the new constraint. 1490 class ConstraintAssignor : public ConstraintAssignorBase<ConstraintAssignor> { 1491 public: 1492 template <class ClassOrSymbol> 1493 LLVM_NODISCARD static ProgramStateRef 1494 assign(ProgramStateRef State, SValBuilder &Builder, RangeSet::Factory &F, 1495 ClassOrSymbol CoS, RangeSet NewConstraint) { 1496 if (!State || NewConstraint.isEmpty()) 1497 return nullptr; 1498 1499 ConstraintAssignor Assignor{State, Builder, F}; 1500 return Assignor.assign(CoS, NewConstraint); 1501 } 1502 1503 inline bool assignSymExprToConst(const SymExpr *Sym, Const Constraint); 1504 inline bool assignSymSymExprToRangeSet(const SymSymExpr *Sym, 1505 RangeSet Constraint); 1506 1507 private: 1508 ConstraintAssignor(ProgramStateRef State, SValBuilder &Builder, 1509 RangeSet::Factory &F) 1510 : State(State), Builder(Builder), RangeFactory(F) {} 1511 using Base = ConstraintAssignorBase<ConstraintAssignor>; 1512 1513 /// Base method for handling new constraints for symbols. 1514 LLVM_NODISCARD ProgramStateRef assign(SymbolRef Sym, RangeSet NewConstraint) { 1515 // All constraints are actually associated with equivalence classes, and 1516 // that's what we are going to do first. 1517 State = assign(EquivalenceClass::find(State, Sym), NewConstraint); 1518 if (!State) 1519 return nullptr; 1520 1521 // And after that we can check what other things we can get from this 1522 // constraint. 1523 Base::assign(Sym, NewConstraint); 1524 return State; 1525 } 1526 1527 /// Base method for handling new constraints for classes. 1528 LLVM_NODISCARD ProgramStateRef assign(EquivalenceClass Class, 1529 RangeSet NewConstraint) { 1530 // There is a chance that we might need to update constraints for the 1531 // classes that are known to be disequal to Class. 1532 // 1533 // In order for this to be even possible, the new constraint should 1534 // be simply a constant because we can't reason about range disequalities. 1535 if (const llvm::APSInt *Point = NewConstraint.getConcreteValue()) { 1536 1537 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1538 ConstraintRangeTy::Factory &CF = State->get_context<ConstraintRange>(); 1539 1540 // Add new constraint. 1541 Constraints = CF.add(Constraints, Class, NewConstraint); 1542 1543 for (EquivalenceClass DisequalClass : Class.getDisequalClasses(State)) { 1544 RangeSet UpdatedConstraint = SymbolicRangeInferrer::inferRange( 1545 RangeFactory, State, DisequalClass); 1546 1547 UpdatedConstraint = RangeFactory.deletePoint(UpdatedConstraint, *Point); 1548 1549 // If we end up with at least one of the disequal classes to be 1550 // constrained with an empty range-set, the state is infeasible. 1551 if (UpdatedConstraint.isEmpty()) 1552 return nullptr; 1553 1554 Constraints = CF.add(Constraints, DisequalClass, UpdatedConstraint); 1555 } 1556 assert(areFeasible(Constraints) && "Constraint manager shouldn't produce " 1557 "a state with infeasible constraints"); 1558 1559 return setConstraints(State, Constraints); 1560 } 1561 1562 return setConstraint(State, Class, NewConstraint); 1563 } 1564 1565 ProgramStateRef trackDisequality(ProgramStateRef State, SymbolRef LHS, 1566 SymbolRef RHS) { 1567 return EquivalenceClass::markDisequal(RangeFactory, State, LHS, RHS); 1568 } 1569 1570 ProgramStateRef trackEquality(ProgramStateRef State, SymbolRef LHS, 1571 SymbolRef RHS) { 1572 return EquivalenceClass::merge(RangeFactory, State, LHS, RHS); 1573 } 1574 1575 LLVM_NODISCARD Optional<bool> interpreteAsBool(RangeSet Constraint) { 1576 assert(!Constraint.isEmpty() && "Empty ranges shouldn't get here"); 1577 1578 if (Constraint.getConcreteValue()) 1579 return !Constraint.getConcreteValue()->isZero(); 1580 1581 APSIntType T{Constraint.getMinValue()}; 1582 Const Zero = T.getZeroValue(); 1583 if (!Constraint.contains(Zero)) 1584 return true; 1585 1586 return llvm::None; 1587 } 1588 1589 ProgramStateRef State; 1590 SValBuilder &Builder; 1591 RangeSet::Factory &RangeFactory; 1592 }; 1593 1594 //===----------------------------------------------------------------------===// 1595 // Constraint manager implementation details 1596 //===----------------------------------------------------------------------===// 1597 1598 class RangeConstraintManager : public RangedConstraintManager { 1599 public: 1600 RangeConstraintManager(ExprEngine *EE, SValBuilder &SVB) 1601 : RangedConstraintManager(EE, SVB), F(getBasicVals()) {} 1602 1603 //===------------------------------------------------------------------===// 1604 // Implementation for interface from ConstraintManager. 1605 //===------------------------------------------------------------------===// 1606 1607 bool haveEqualConstraints(ProgramStateRef S1, 1608 ProgramStateRef S2) const override { 1609 // NOTE: ClassMembers are as simple as back pointers for ClassMap, 1610 // so comparing constraint ranges and class maps should be 1611 // sufficient. 1612 return S1->get<ConstraintRange>() == S2->get<ConstraintRange>() && 1613 S1->get<ClassMap>() == S2->get<ClassMap>(); 1614 } 1615 1616 bool canReasonAbout(SVal X) const override; 1617 1618 ConditionTruthVal checkNull(ProgramStateRef State, SymbolRef Sym) override; 1619 1620 const llvm::APSInt *getSymVal(ProgramStateRef State, 1621 SymbolRef Sym) const override; 1622 1623 ProgramStateRef removeDeadBindings(ProgramStateRef State, 1624 SymbolReaper &SymReaper) override; 1625 1626 void printJson(raw_ostream &Out, ProgramStateRef State, const char *NL = "\n", 1627 unsigned int Space = 0, bool IsDot = false) const override; 1628 void printConstraints(raw_ostream &Out, ProgramStateRef State, 1629 const char *NL = "\n", unsigned int Space = 0, 1630 bool IsDot = false) const; 1631 void printEquivalenceClasses(raw_ostream &Out, ProgramStateRef State, 1632 const char *NL = "\n", unsigned int Space = 0, 1633 bool IsDot = false) const; 1634 void printDisequalities(raw_ostream &Out, ProgramStateRef State, 1635 const char *NL = "\n", unsigned int Space = 0, 1636 bool IsDot = false) const; 1637 1638 //===------------------------------------------------------------------===// 1639 // Implementation for interface from RangedConstraintManager. 1640 //===------------------------------------------------------------------===// 1641 1642 ProgramStateRef assumeSymNE(ProgramStateRef State, SymbolRef Sym, 1643 const llvm::APSInt &V, 1644 const llvm::APSInt &Adjustment) override; 1645 1646 ProgramStateRef assumeSymEQ(ProgramStateRef State, SymbolRef Sym, 1647 const llvm::APSInt &V, 1648 const llvm::APSInt &Adjustment) override; 1649 1650 ProgramStateRef assumeSymLT(ProgramStateRef State, SymbolRef Sym, 1651 const llvm::APSInt &V, 1652 const llvm::APSInt &Adjustment) override; 1653 1654 ProgramStateRef assumeSymGT(ProgramStateRef State, SymbolRef Sym, 1655 const llvm::APSInt &V, 1656 const llvm::APSInt &Adjustment) override; 1657 1658 ProgramStateRef assumeSymLE(ProgramStateRef State, SymbolRef Sym, 1659 const llvm::APSInt &V, 1660 const llvm::APSInt &Adjustment) override; 1661 1662 ProgramStateRef assumeSymGE(ProgramStateRef State, SymbolRef Sym, 1663 const llvm::APSInt &V, 1664 const llvm::APSInt &Adjustment) override; 1665 1666 ProgramStateRef assumeSymWithinInclusiveRange( 1667 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 1668 const llvm::APSInt &To, const llvm::APSInt &Adjustment) override; 1669 1670 ProgramStateRef assumeSymOutsideInclusiveRange( 1671 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 1672 const llvm::APSInt &To, const llvm::APSInt &Adjustment) override; 1673 1674 private: 1675 RangeSet::Factory F; 1676 1677 RangeSet getRange(ProgramStateRef State, SymbolRef Sym); 1678 RangeSet getRange(ProgramStateRef State, EquivalenceClass Class); 1679 ProgramStateRef setRange(ProgramStateRef State, SymbolRef Sym, 1680 RangeSet Range); 1681 ProgramStateRef setRange(ProgramStateRef State, EquivalenceClass Class, 1682 RangeSet Range); 1683 1684 RangeSet getSymLTRange(ProgramStateRef St, SymbolRef Sym, 1685 const llvm::APSInt &Int, 1686 const llvm::APSInt &Adjustment); 1687 RangeSet getSymGTRange(ProgramStateRef St, SymbolRef Sym, 1688 const llvm::APSInt &Int, 1689 const llvm::APSInt &Adjustment); 1690 RangeSet getSymLERange(ProgramStateRef St, SymbolRef Sym, 1691 const llvm::APSInt &Int, 1692 const llvm::APSInt &Adjustment); 1693 RangeSet getSymLERange(llvm::function_ref<RangeSet()> RS, 1694 const llvm::APSInt &Int, 1695 const llvm::APSInt &Adjustment); 1696 RangeSet getSymGERange(ProgramStateRef St, SymbolRef Sym, 1697 const llvm::APSInt &Int, 1698 const llvm::APSInt &Adjustment); 1699 }; 1700 1701 bool ConstraintAssignor::assignSymExprToConst(const SymExpr *Sym, 1702 const llvm::APSInt &Constraint) { 1703 llvm::SmallSet<EquivalenceClass, 4> SimplifiedClasses; 1704 // Iterate over all equivalence classes and try to simplify them. 1705 ClassMembersTy Members = State->get<ClassMembers>(); 1706 for (std::pair<EquivalenceClass, SymbolSet> ClassToSymbolSet : Members) { 1707 EquivalenceClass Class = ClassToSymbolSet.first; 1708 State = EquivalenceClass::simplify(Builder, RangeFactory, State, Class); 1709 if (!State) 1710 return false; 1711 SimplifiedClasses.insert(Class); 1712 } 1713 1714 // Trivial equivalence classes (those that have only one symbol member) are 1715 // not stored in the State. Thus, we must skim through the constraints as 1716 // well. And we try to simplify symbols in the constraints. 1717 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1718 for (std::pair<EquivalenceClass, RangeSet> ClassConstraint : Constraints) { 1719 EquivalenceClass Class = ClassConstraint.first; 1720 if (SimplifiedClasses.count(Class)) // Already simplified. 1721 continue; 1722 State = EquivalenceClass::simplify(Builder, RangeFactory, State, Class); 1723 if (!State) 1724 return false; 1725 } 1726 1727 return true; 1728 } 1729 1730 bool ConstraintAssignor::assignSymSymExprToRangeSet(const SymSymExpr *Sym, 1731 RangeSet Constraint) { 1732 Optional<bool> ConstraintAsBool = interpreteAsBool(Constraint); 1733 1734 if (!ConstraintAsBool) 1735 return true; 1736 1737 if (Optional<bool> Equality = meansEquality(Sym)) { 1738 // Here we cover two cases: 1739 // * if Sym is equality and the new constraint is true -> Sym's operands 1740 // should be marked as equal 1741 // * if Sym is disequality and the new constraint is false -> Sym's 1742 // operands should be also marked as equal 1743 if (*Equality == *ConstraintAsBool) { 1744 State = trackEquality(State, Sym->getLHS(), Sym->getRHS()); 1745 } else { 1746 // Other combinations leave as with disequal operands. 1747 State = trackDisequality(State, Sym->getLHS(), Sym->getRHS()); 1748 } 1749 1750 if (!State) 1751 return false; 1752 } 1753 1754 return true; 1755 } 1756 1757 } // end anonymous namespace 1758 1759 std::unique_ptr<ConstraintManager> 1760 ento::CreateRangeConstraintManager(ProgramStateManager &StMgr, 1761 ExprEngine *Eng) { 1762 return std::make_unique<RangeConstraintManager>(Eng, StMgr.getSValBuilder()); 1763 } 1764 1765 ConstraintMap ento::getConstraintMap(ProgramStateRef State) { 1766 ConstraintMap::Factory &F = State->get_context<ConstraintMap>(); 1767 ConstraintMap Result = F.getEmptyMap(); 1768 1769 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1770 for (std::pair<EquivalenceClass, RangeSet> ClassConstraint : Constraints) { 1771 EquivalenceClass Class = ClassConstraint.first; 1772 SymbolSet ClassMembers = Class.getClassMembers(State); 1773 assert(!ClassMembers.isEmpty() && 1774 "Class must always have at least one member!"); 1775 1776 SymbolRef Representative = *ClassMembers.begin(); 1777 Result = F.add(Result, Representative, ClassConstraint.second); 1778 } 1779 1780 return Result; 1781 } 1782 1783 //===----------------------------------------------------------------------===// 1784 // EqualityClass implementation details 1785 //===----------------------------------------------------------------------===// 1786 1787 LLVM_DUMP_METHOD void EquivalenceClass::dumpToStream(ProgramStateRef State, 1788 raw_ostream &os) const { 1789 SymbolSet ClassMembers = getClassMembers(State); 1790 for (const SymbolRef &MemberSym : ClassMembers) { 1791 MemberSym->dump(); 1792 os << "\n"; 1793 } 1794 } 1795 1796 inline EquivalenceClass EquivalenceClass::find(ProgramStateRef State, 1797 SymbolRef Sym) { 1798 assert(State && "State should not be null"); 1799 assert(Sym && "Symbol should not be null"); 1800 // We store far from all Symbol -> Class mappings 1801 if (const EquivalenceClass *NontrivialClass = State->get<ClassMap>(Sym)) 1802 return *NontrivialClass; 1803 1804 // This is a trivial class of Sym. 1805 return Sym; 1806 } 1807 1808 inline ProgramStateRef EquivalenceClass::merge(RangeSet::Factory &F, 1809 ProgramStateRef State, 1810 SymbolRef First, 1811 SymbolRef Second) { 1812 EquivalenceClass FirstClass = find(State, First); 1813 EquivalenceClass SecondClass = find(State, Second); 1814 1815 return FirstClass.merge(F, State, SecondClass); 1816 } 1817 1818 inline ProgramStateRef EquivalenceClass::merge(RangeSet::Factory &F, 1819 ProgramStateRef State, 1820 EquivalenceClass Other) { 1821 // It is already the same class. 1822 if (*this == Other) 1823 return State; 1824 1825 // FIXME: As of now, we support only equivalence classes of the same type. 1826 // This limitation is connected to the lack of explicit casts in 1827 // our symbolic expression model. 1828 // 1829 // That means that for `int x` and `char y` we don't distinguish 1830 // between these two very different cases: 1831 // * `x == y` 1832 // * `(char)x == y` 1833 // 1834 // The moment we introduce symbolic casts, this restriction can be 1835 // lifted. 1836 if (getType() != Other.getType()) 1837 return State; 1838 1839 SymbolSet Members = getClassMembers(State); 1840 SymbolSet OtherMembers = Other.getClassMembers(State); 1841 1842 // We estimate the size of the class by the height of tree containing 1843 // its members. Merging is not a trivial operation, so it's easier to 1844 // merge the smaller class into the bigger one. 1845 if (Members.getHeight() >= OtherMembers.getHeight()) { 1846 return mergeImpl(F, State, Members, Other, OtherMembers); 1847 } else { 1848 return Other.mergeImpl(F, State, OtherMembers, *this, Members); 1849 } 1850 } 1851 1852 inline ProgramStateRef 1853 EquivalenceClass::mergeImpl(RangeSet::Factory &RangeFactory, 1854 ProgramStateRef State, SymbolSet MyMembers, 1855 EquivalenceClass Other, SymbolSet OtherMembers) { 1856 // Essentially what we try to recreate here is some kind of union-find 1857 // data structure. It does have certain limitations due to persistence 1858 // and the need to remove elements from classes. 1859 // 1860 // In this setting, EquialityClass object is the representative of the class 1861 // or the parent element. ClassMap is a mapping of class members to their 1862 // parent. Unlike the union-find structure, they all point directly to the 1863 // class representative because we don't have an opportunity to actually do 1864 // path compression when dealing with immutability. This means that we 1865 // compress paths every time we do merges. It also means that we lose 1866 // the main amortized complexity benefit from the original data structure. 1867 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1868 ConstraintRangeTy::Factory &CRF = State->get_context<ConstraintRange>(); 1869 1870 // 1. If the merged classes have any constraints associated with them, we 1871 // need to transfer them to the class we have left. 1872 // 1873 // Intersection here makes perfect sense because both of these constraints 1874 // must hold for the whole new class. 1875 if (Optional<RangeSet> NewClassConstraint = 1876 intersect(RangeFactory, getConstraint(State, *this), 1877 getConstraint(State, Other))) { 1878 // NOTE: Essentially, NewClassConstraint should NEVER be infeasible because 1879 // range inferrer shouldn't generate ranges incompatible with 1880 // equivalence classes. However, at the moment, due to imperfections 1881 // in the solver, it is possible and the merge function can also 1882 // return infeasible states aka null states. 1883 if (NewClassConstraint->isEmpty()) 1884 // Infeasible state 1885 return nullptr; 1886 1887 // No need in tracking constraints of a now-dissolved class. 1888 Constraints = CRF.remove(Constraints, Other); 1889 // Assign new constraints for this class. 1890 Constraints = CRF.add(Constraints, *this, *NewClassConstraint); 1891 1892 assert(areFeasible(Constraints) && "Constraint manager shouldn't produce " 1893 "a state with infeasible constraints"); 1894 1895 State = State->set<ConstraintRange>(Constraints); 1896 } 1897 1898 // 2. Get ALL equivalence-related maps 1899 ClassMapTy Classes = State->get<ClassMap>(); 1900 ClassMapTy::Factory &CMF = State->get_context<ClassMap>(); 1901 1902 ClassMembersTy Members = State->get<ClassMembers>(); 1903 ClassMembersTy::Factory &MF = State->get_context<ClassMembers>(); 1904 1905 DisequalityMapTy DisequalityInfo = State->get<DisequalityMap>(); 1906 DisequalityMapTy::Factory &DF = State->get_context<DisequalityMap>(); 1907 1908 ClassSet::Factory &CF = State->get_context<ClassSet>(); 1909 SymbolSet::Factory &F = getMembersFactory(State); 1910 1911 // 2. Merge members of the Other class into the current class. 1912 SymbolSet NewClassMembers = MyMembers; 1913 for (SymbolRef Sym : OtherMembers) { 1914 NewClassMembers = F.add(NewClassMembers, Sym); 1915 // *this is now the class for all these new symbols. 1916 Classes = CMF.add(Classes, Sym, *this); 1917 } 1918 1919 // 3. Adjust member mapping. 1920 // 1921 // No need in tracking members of a now-dissolved class. 1922 Members = MF.remove(Members, Other); 1923 // Now only the current class is mapped to all the symbols. 1924 Members = MF.add(Members, *this, NewClassMembers); 1925 1926 // 4. Update disequality relations 1927 ClassSet DisequalToOther = Other.getDisequalClasses(DisequalityInfo, CF); 1928 // We are about to merge two classes but they are already known to be 1929 // non-equal. This is a contradiction. 1930 if (DisequalToOther.contains(*this)) 1931 return nullptr; 1932 1933 if (!DisequalToOther.isEmpty()) { 1934 ClassSet DisequalToThis = getDisequalClasses(DisequalityInfo, CF); 1935 DisequalityInfo = DF.remove(DisequalityInfo, Other); 1936 1937 for (EquivalenceClass DisequalClass : DisequalToOther) { 1938 DisequalToThis = CF.add(DisequalToThis, DisequalClass); 1939 1940 // Disequality is a symmetric relation meaning that if 1941 // DisequalToOther not null then the set for DisequalClass is not 1942 // empty and has at least Other. 1943 ClassSet OriginalSetLinkedToOther = 1944 *DisequalityInfo.lookup(DisequalClass); 1945 1946 // Other will be eliminated and we should replace it with the bigger 1947 // united class. 1948 ClassSet NewSet = CF.remove(OriginalSetLinkedToOther, Other); 1949 NewSet = CF.add(NewSet, *this); 1950 1951 DisequalityInfo = DF.add(DisequalityInfo, DisequalClass, NewSet); 1952 } 1953 1954 DisequalityInfo = DF.add(DisequalityInfo, *this, DisequalToThis); 1955 State = State->set<DisequalityMap>(DisequalityInfo); 1956 } 1957 1958 // 5. Update the state 1959 State = State->set<ClassMap>(Classes); 1960 State = State->set<ClassMembers>(Members); 1961 1962 return State; 1963 } 1964 1965 inline SymbolSet::Factory & 1966 EquivalenceClass::getMembersFactory(ProgramStateRef State) { 1967 return State->get_context<SymbolSet>(); 1968 } 1969 1970 SymbolSet EquivalenceClass::getClassMembers(ProgramStateRef State) const { 1971 if (const SymbolSet *Members = State->get<ClassMembers>(*this)) 1972 return *Members; 1973 1974 // This class is trivial, so we need to construct a set 1975 // with just that one symbol from the class. 1976 SymbolSet::Factory &F = getMembersFactory(State); 1977 return F.add(F.getEmptySet(), getRepresentativeSymbol()); 1978 } 1979 1980 bool EquivalenceClass::isTrivial(ProgramStateRef State) const { 1981 return State->get<ClassMembers>(*this) == nullptr; 1982 } 1983 1984 bool EquivalenceClass::isTriviallyDead(ProgramStateRef State, 1985 SymbolReaper &Reaper) const { 1986 return isTrivial(State) && Reaper.isDead(getRepresentativeSymbol()); 1987 } 1988 1989 inline ProgramStateRef EquivalenceClass::markDisequal(RangeSet::Factory &RF, 1990 ProgramStateRef State, 1991 SymbolRef First, 1992 SymbolRef Second) { 1993 return markDisequal(RF, State, find(State, First), find(State, Second)); 1994 } 1995 1996 inline ProgramStateRef EquivalenceClass::markDisequal(RangeSet::Factory &RF, 1997 ProgramStateRef State, 1998 EquivalenceClass First, 1999 EquivalenceClass Second) { 2000 return First.markDisequal(RF, State, Second); 2001 } 2002 2003 inline ProgramStateRef 2004 EquivalenceClass::markDisequal(RangeSet::Factory &RF, ProgramStateRef State, 2005 EquivalenceClass Other) const { 2006 // If we know that two classes are equal, we can only produce an infeasible 2007 // state. 2008 if (*this == Other) { 2009 return nullptr; 2010 } 2011 2012 DisequalityMapTy DisequalityInfo = State->get<DisequalityMap>(); 2013 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 2014 2015 // Disequality is a symmetric relation, so if we mark A as disequal to B, 2016 // we should also mark B as disequalt to A. 2017 if (!addToDisequalityInfo(DisequalityInfo, Constraints, RF, State, *this, 2018 Other) || 2019 !addToDisequalityInfo(DisequalityInfo, Constraints, RF, State, Other, 2020 *this)) 2021 return nullptr; 2022 2023 assert(areFeasible(Constraints) && "Constraint manager shouldn't produce " 2024 "a state with infeasible constraints"); 2025 2026 State = State->set<DisequalityMap>(DisequalityInfo); 2027 State = State->set<ConstraintRange>(Constraints); 2028 2029 return State; 2030 } 2031 2032 inline bool EquivalenceClass::addToDisequalityInfo( 2033 DisequalityMapTy &Info, ConstraintRangeTy &Constraints, 2034 RangeSet::Factory &RF, ProgramStateRef State, EquivalenceClass First, 2035 EquivalenceClass Second) { 2036 2037 // 1. Get all of the required factories. 2038 DisequalityMapTy::Factory &F = State->get_context<DisequalityMap>(); 2039 ClassSet::Factory &CF = State->get_context<ClassSet>(); 2040 ConstraintRangeTy::Factory &CRF = State->get_context<ConstraintRange>(); 2041 2042 // 2. Add Second to the set of classes disequal to First. 2043 const ClassSet *CurrentSet = Info.lookup(First); 2044 ClassSet NewSet = CurrentSet ? *CurrentSet : CF.getEmptySet(); 2045 NewSet = CF.add(NewSet, Second); 2046 2047 Info = F.add(Info, First, NewSet); 2048 2049 // 3. If Second is known to be a constant, we can delete this point 2050 // from the constraint asociated with First. 2051 // 2052 // So, if Second == 10, it means that First != 10. 2053 // At the same time, the same logic does not apply to ranges. 2054 if (const RangeSet *SecondConstraint = Constraints.lookup(Second)) 2055 if (const llvm::APSInt *Point = SecondConstraint->getConcreteValue()) { 2056 2057 RangeSet FirstConstraint = SymbolicRangeInferrer::inferRange( 2058 RF, State, First.getRepresentativeSymbol()); 2059 2060 FirstConstraint = RF.deletePoint(FirstConstraint, *Point); 2061 2062 // If the First class is about to be constrained with an empty 2063 // range-set, the state is infeasible. 2064 if (FirstConstraint.isEmpty()) 2065 return false; 2066 2067 Constraints = CRF.add(Constraints, First, FirstConstraint); 2068 } 2069 2070 return true; 2071 } 2072 2073 inline Optional<bool> EquivalenceClass::areEqual(ProgramStateRef State, 2074 SymbolRef FirstSym, 2075 SymbolRef SecondSym) { 2076 return EquivalenceClass::areEqual(State, find(State, FirstSym), 2077 find(State, SecondSym)); 2078 } 2079 2080 inline Optional<bool> EquivalenceClass::areEqual(ProgramStateRef State, 2081 EquivalenceClass First, 2082 EquivalenceClass Second) { 2083 // The same equivalence class => symbols are equal. 2084 if (First == Second) 2085 return true; 2086 2087 // Let's check if we know anything about these two classes being not equal to 2088 // each other. 2089 ClassSet DisequalToFirst = First.getDisequalClasses(State); 2090 if (DisequalToFirst.contains(Second)) 2091 return false; 2092 2093 // It is not clear. 2094 return llvm::None; 2095 } 2096 2097 // Iterate over all symbols and try to simplify them. Once a symbol is 2098 // simplified then we check if we can merge the simplified symbol's equivalence 2099 // class to this class. This way, we simplify not just the symbols but the 2100 // classes as well: we strive to keep the number of the classes to be the 2101 // absolute minimum. 2102 LLVM_NODISCARD ProgramStateRef 2103 EquivalenceClass::simplify(SValBuilder &SVB, RangeSet::Factory &F, 2104 ProgramStateRef State, EquivalenceClass Class) { 2105 SymbolSet ClassMembers = Class.getClassMembers(State); 2106 for (const SymbolRef &MemberSym : ClassMembers) { 2107 SymbolRef SimplifiedMemberSym = ento::simplify(State, MemberSym); 2108 if (SimplifiedMemberSym && MemberSym != SimplifiedMemberSym) { 2109 // The simplified symbol should be the member of the original Class, 2110 // however, it might be in another existing class at the moment. We 2111 // have to merge these classes. 2112 State = merge(F, State, MemberSym, SimplifiedMemberSym); 2113 if (!State) 2114 return nullptr; 2115 } 2116 } 2117 return State; 2118 } 2119 2120 inline ClassSet EquivalenceClass::getDisequalClasses(ProgramStateRef State, 2121 SymbolRef Sym) { 2122 return find(State, Sym).getDisequalClasses(State); 2123 } 2124 2125 inline ClassSet 2126 EquivalenceClass::getDisequalClasses(ProgramStateRef State) const { 2127 return getDisequalClasses(State->get<DisequalityMap>(), 2128 State->get_context<ClassSet>()); 2129 } 2130 2131 inline ClassSet 2132 EquivalenceClass::getDisequalClasses(DisequalityMapTy Map, 2133 ClassSet::Factory &Factory) const { 2134 if (const ClassSet *DisequalClasses = Map.lookup(*this)) 2135 return *DisequalClasses; 2136 2137 return Factory.getEmptySet(); 2138 } 2139 2140 bool EquivalenceClass::isClassDataConsistent(ProgramStateRef State) { 2141 ClassMembersTy Members = State->get<ClassMembers>(); 2142 2143 for (std::pair<EquivalenceClass, SymbolSet> ClassMembersPair : Members) { 2144 for (SymbolRef Member : ClassMembersPair.second) { 2145 // Every member of the class should have a mapping back to the class. 2146 if (find(State, Member) == ClassMembersPair.first) { 2147 continue; 2148 } 2149 2150 return false; 2151 } 2152 } 2153 2154 DisequalityMapTy Disequalities = State->get<DisequalityMap>(); 2155 for (std::pair<EquivalenceClass, ClassSet> DisequalityInfo : Disequalities) { 2156 EquivalenceClass Class = DisequalityInfo.first; 2157 ClassSet DisequalClasses = DisequalityInfo.second; 2158 2159 // There is no use in keeping empty sets in the map. 2160 if (DisequalClasses.isEmpty()) 2161 return false; 2162 2163 // Disequality is symmetrical, i.e. for every Class A and B that A != B, 2164 // B != A should also be true. 2165 for (EquivalenceClass DisequalClass : DisequalClasses) { 2166 const ClassSet *DisequalToDisequalClasses = 2167 Disequalities.lookup(DisequalClass); 2168 2169 // It should be a set of at least one element: Class 2170 if (!DisequalToDisequalClasses || 2171 !DisequalToDisequalClasses->contains(Class)) 2172 return false; 2173 } 2174 } 2175 2176 return true; 2177 } 2178 2179 //===----------------------------------------------------------------------===// 2180 // RangeConstraintManager implementation 2181 //===----------------------------------------------------------------------===// 2182 2183 bool RangeConstraintManager::canReasonAbout(SVal X) const { 2184 Optional<nonloc::SymbolVal> SymVal = X.getAs<nonloc::SymbolVal>(); 2185 if (SymVal && SymVal->isExpression()) { 2186 const SymExpr *SE = SymVal->getSymbol(); 2187 2188 if (const SymIntExpr *SIE = dyn_cast<SymIntExpr>(SE)) { 2189 switch (SIE->getOpcode()) { 2190 // We don't reason yet about bitwise-constraints on symbolic values. 2191 case BO_And: 2192 case BO_Or: 2193 case BO_Xor: 2194 return false; 2195 // We don't reason yet about these arithmetic constraints on 2196 // symbolic values. 2197 case BO_Mul: 2198 case BO_Div: 2199 case BO_Rem: 2200 case BO_Shl: 2201 case BO_Shr: 2202 return false; 2203 // All other cases. 2204 default: 2205 return true; 2206 } 2207 } 2208 2209 if (const SymSymExpr *SSE = dyn_cast<SymSymExpr>(SE)) { 2210 // FIXME: Handle <=> here. 2211 if (BinaryOperator::isEqualityOp(SSE->getOpcode()) || 2212 BinaryOperator::isRelationalOp(SSE->getOpcode())) { 2213 // We handle Loc <> Loc comparisons, but not (yet) NonLoc <> NonLoc. 2214 // We've recently started producing Loc <> NonLoc comparisons (that 2215 // result from casts of one of the operands between eg. intptr_t and 2216 // void *), but we can't reason about them yet. 2217 if (Loc::isLocType(SSE->getLHS()->getType())) { 2218 return Loc::isLocType(SSE->getRHS()->getType()); 2219 } 2220 } 2221 } 2222 2223 return false; 2224 } 2225 2226 return true; 2227 } 2228 2229 ConditionTruthVal RangeConstraintManager::checkNull(ProgramStateRef State, 2230 SymbolRef Sym) { 2231 const RangeSet *Ranges = getConstraint(State, Sym); 2232 2233 // If we don't have any information about this symbol, it's underconstrained. 2234 if (!Ranges) 2235 return ConditionTruthVal(); 2236 2237 // If we have a concrete value, see if it's zero. 2238 if (const llvm::APSInt *Value = Ranges->getConcreteValue()) 2239 return *Value == 0; 2240 2241 BasicValueFactory &BV = getBasicVals(); 2242 APSIntType IntType = BV.getAPSIntType(Sym->getType()); 2243 llvm::APSInt Zero = IntType.getZeroValue(); 2244 2245 // Check if zero is in the set of possible values. 2246 if (!Ranges->contains(Zero)) 2247 return false; 2248 2249 // Zero is a possible value, but it is not the /only/ possible value. 2250 return ConditionTruthVal(); 2251 } 2252 2253 const llvm::APSInt *RangeConstraintManager::getSymVal(ProgramStateRef St, 2254 SymbolRef Sym) const { 2255 const RangeSet *T = getConstraint(St, Sym); 2256 return T ? T->getConcreteValue() : nullptr; 2257 } 2258 2259 //===----------------------------------------------------------------------===// 2260 // Remove dead symbols from existing constraints 2261 //===----------------------------------------------------------------------===// 2262 2263 /// Scan all symbols referenced by the constraints. If the symbol is not alive 2264 /// as marked in LSymbols, mark it as dead in DSymbols. 2265 ProgramStateRef 2266 RangeConstraintManager::removeDeadBindings(ProgramStateRef State, 2267 SymbolReaper &SymReaper) { 2268 ClassMembersTy ClassMembersMap = State->get<ClassMembers>(); 2269 ClassMembersTy NewClassMembersMap = ClassMembersMap; 2270 ClassMembersTy::Factory &EMFactory = State->get_context<ClassMembers>(); 2271 SymbolSet::Factory &SetFactory = State->get_context<SymbolSet>(); 2272 2273 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 2274 ConstraintRangeTy NewConstraints = Constraints; 2275 ConstraintRangeTy::Factory &ConstraintFactory = 2276 State->get_context<ConstraintRange>(); 2277 2278 ClassMapTy Map = State->get<ClassMap>(); 2279 ClassMapTy NewMap = Map; 2280 ClassMapTy::Factory &ClassFactory = State->get_context<ClassMap>(); 2281 2282 DisequalityMapTy Disequalities = State->get<DisequalityMap>(); 2283 DisequalityMapTy::Factory &DisequalityFactory = 2284 State->get_context<DisequalityMap>(); 2285 ClassSet::Factory &ClassSetFactory = State->get_context<ClassSet>(); 2286 2287 bool ClassMapChanged = false; 2288 bool MembersMapChanged = false; 2289 bool ConstraintMapChanged = false; 2290 bool DisequalitiesChanged = false; 2291 2292 auto removeDeadClass = [&](EquivalenceClass Class) { 2293 // Remove associated constraint ranges. 2294 Constraints = ConstraintFactory.remove(Constraints, Class); 2295 ConstraintMapChanged = true; 2296 2297 // Update disequality information to not hold any information on the 2298 // removed class. 2299 ClassSet DisequalClasses = 2300 Class.getDisequalClasses(Disequalities, ClassSetFactory); 2301 if (!DisequalClasses.isEmpty()) { 2302 for (EquivalenceClass DisequalClass : DisequalClasses) { 2303 ClassSet DisequalToDisequalSet = 2304 DisequalClass.getDisequalClasses(Disequalities, ClassSetFactory); 2305 // DisequalToDisequalSet is guaranteed to be non-empty for consistent 2306 // disequality info. 2307 assert(!DisequalToDisequalSet.isEmpty()); 2308 ClassSet NewSet = ClassSetFactory.remove(DisequalToDisequalSet, Class); 2309 2310 // No need in keeping an empty set. 2311 if (NewSet.isEmpty()) { 2312 Disequalities = 2313 DisequalityFactory.remove(Disequalities, DisequalClass); 2314 } else { 2315 Disequalities = 2316 DisequalityFactory.add(Disequalities, DisequalClass, NewSet); 2317 } 2318 } 2319 // Remove the data for the class 2320 Disequalities = DisequalityFactory.remove(Disequalities, Class); 2321 DisequalitiesChanged = true; 2322 } 2323 }; 2324 2325 // 1. Let's see if dead symbols are trivial and have associated constraints. 2326 for (std::pair<EquivalenceClass, RangeSet> ClassConstraintPair : 2327 Constraints) { 2328 EquivalenceClass Class = ClassConstraintPair.first; 2329 if (Class.isTriviallyDead(State, SymReaper)) { 2330 // If this class is trivial, we can remove its constraints right away. 2331 removeDeadClass(Class); 2332 } 2333 } 2334 2335 // 2. We don't need to track classes for dead symbols. 2336 for (std::pair<SymbolRef, EquivalenceClass> SymbolClassPair : Map) { 2337 SymbolRef Sym = SymbolClassPair.first; 2338 2339 if (SymReaper.isDead(Sym)) { 2340 ClassMapChanged = true; 2341 NewMap = ClassFactory.remove(NewMap, Sym); 2342 } 2343 } 2344 2345 // 3. Remove dead members from classes and remove dead non-trivial classes 2346 // and their constraints. 2347 for (std::pair<EquivalenceClass, SymbolSet> ClassMembersPair : 2348 ClassMembersMap) { 2349 EquivalenceClass Class = ClassMembersPair.first; 2350 SymbolSet LiveMembers = ClassMembersPair.second; 2351 bool MembersChanged = false; 2352 2353 for (SymbolRef Member : ClassMembersPair.second) { 2354 if (SymReaper.isDead(Member)) { 2355 MembersChanged = true; 2356 LiveMembers = SetFactory.remove(LiveMembers, Member); 2357 } 2358 } 2359 2360 // Check if the class changed. 2361 if (!MembersChanged) 2362 continue; 2363 2364 MembersMapChanged = true; 2365 2366 if (LiveMembers.isEmpty()) { 2367 // The class is dead now, we need to wipe it out of the members map... 2368 NewClassMembersMap = EMFactory.remove(NewClassMembersMap, Class); 2369 2370 // ...and remove all of its constraints. 2371 removeDeadClass(Class); 2372 } else { 2373 // We need to change the members associated with the class. 2374 NewClassMembersMap = 2375 EMFactory.add(NewClassMembersMap, Class, LiveMembers); 2376 } 2377 } 2378 2379 // 4. Update the state with new maps. 2380 // 2381 // Here we try to be humble and update a map only if it really changed. 2382 if (ClassMapChanged) 2383 State = State->set<ClassMap>(NewMap); 2384 2385 if (MembersMapChanged) 2386 State = State->set<ClassMembers>(NewClassMembersMap); 2387 2388 if (ConstraintMapChanged) 2389 State = State->set<ConstraintRange>(Constraints); 2390 2391 if (DisequalitiesChanged) 2392 State = State->set<DisequalityMap>(Disequalities); 2393 2394 assert(EquivalenceClass::isClassDataConsistent(State)); 2395 2396 return State; 2397 } 2398 2399 RangeSet RangeConstraintManager::getRange(ProgramStateRef State, 2400 SymbolRef Sym) { 2401 return SymbolicRangeInferrer::inferRange(F, State, Sym); 2402 } 2403 2404 ProgramStateRef RangeConstraintManager::setRange(ProgramStateRef State, 2405 SymbolRef Sym, 2406 RangeSet Range) { 2407 return ConstraintAssignor::assign(State, getSValBuilder(), F, Sym, Range); 2408 } 2409 2410 //===------------------------------------------------------------------------=== 2411 // assumeSymX methods: protected interface for RangeConstraintManager. 2412 //===------------------------------------------------------------------------===/ 2413 2414 // The syntax for ranges below is mathematical, using [x, y] for closed ranges 2415 // and (x, y) for open ranges. These ranges are modular, corresponding with 2416 // a common treatment of C integer overflow. This means that these methods 2417 // do not have to worry about overflow; RangeSet::Intersect can handle such a 2418 // "wraparound" range. 2419 // As an example, the range [UINT_MAX-1, 3) contains five values: UINT_MAX-1, 2420 // UINT_MAX, 0, 1, and 2. 2421 2422 ProgramStateRef 2423 RangeConstraintManager::assumeSymNE(ProgramStateRef St, SymbolRef Sym, 2424 const llvm::APSInt &Int, 2425 const llvm::APSInt &Adjustment) { 2426 // Before we do any real work, see if the value can even show up. 2427 APSIntType AdjustmentType(Adjustment); 2428 if (AdjustmentType.testInRange(Int, true) != APSIntType::RTR_Within) 2429 return St; 2430 2431 llvm::APSInt Point = AdjustmentType.convert(Int) - Adjustment; 2432 RangeSet New = getRange(St, Sym); 2433 New = F.deletePoint(New, Point); 2434 2435 return setRange(St, Sym, New); 2436 } 2437 2438 ProgramStateRef 2439 RangeConstraintManager::assumeSymEQ(ProgramStateRef St, SymbolRef Sym, 2440 const llvm::APSInt &Int, 2441 const llvm::APSInt &Adjustment) { 2442 // Before we do any real work, see if the value can even show up. 2443 APSIntType AdjustmentType(Adjustment); 2444 if (AdjustmentType.testInRange(Int, true) != APSIntType::RTR_Within) 2445 return nullptr; 2446 2447 // [Int-Adjustment, Int-Adjustment] 2448 llvm::APSInt AdjInt = AdjustmentType.convert(Int) - Adjustment; 2449 RangeSet New = getRange(St, Sym); 2450 New = F.intersect(New, AdjInt); 2451 2452 return setRange(St, Sym, New); 2453 } 2454 2455 RangeSet RangeConstraintManager::getSymLTRange(ProgramStateRef St, 2456 SymbolRef Sym, 2457 const llvm::APSInt &Int, 2458 const llvm::APSInt &Adjustment) { 2459 // Before we do any real work, see if the value can even show up. 2460 APSIntType AdjustmentType(Adjustment); 2461 switch (AdjustmentType.testInRange(Int, true)) { 2462 case APSIntType::RTR_Below: 2463 return F.getEmptySet(); 2464 case APSIntType::RTR_Within: 2465 break; 2466 case APSIntType::RTR_Above: 2467 return getRange(St, Sym); 2468 } 2469 2470 // Special case for Int == Min. This is always false. 2471 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2472 llvm::APSInt Min = AdjustmentType.getMinValue(); 2473 if (ComparisonVal == Min) 2474 return F.getEmptySet(); 2475 2476 llvm::APSInt Lower = Min - Adjustment; 2477 llvm::APSInt Upper = ComparisonVal - Adjustment; 2478 --Upper; 2479 2480 RangeSet Result = getRange(St, Sym); 2481 return F.intersect(Result, Lower, Upper); 2482 } 2483 2484 ProgramStateRef 2485 RangeConstraintManager::assumeSymLT(ProgramStateRef St, SymbolRef Sym, 2486 const llvm::APSInt &Int, 2487 const llvm::APSInt &Adjustment) { 2488 RangeSet New = getSymLTRange(St, Sym, Int, Adjustment); 2489 return setRange(St, Sym, New); 2490 } 2491 2492 RangeSet RangeConstraintManager::getSymGTRange(ProgramStateRef St, 2493 SymbolRef Sym, 2494 const llvm::APSInt &Int, 2495 const llvm::APSInt &Adjustment) { 2496 // Before we do any real work, see if the value can even show up. 2497 APSIntType AdjustmentType(Adjustment); 2498 switch (AdjustmentType.testInRange(Int, true)) { 2499 case APSIntType::RTR_Below: 2500 return getRange(St, Sym); 2501 case APSIntType::RTR_Within: 2502 break; 2503 case APSIntType::RTR_Above: 2504 return F.getEmptySet(); 2505 } 2506 2507 // Special case for Int == Max. This is always false. 2508 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2509 llvm::APSInt Max = AdjustmentType.getMaxValue(); 2510 if (ComparisonVal == Max) 2511 return F.getEmptySet(); 2512 2513 llvm::APSInt Lower = ComparisonVal - Adjustment; 2514 llvm::APSInt Upper = Max - Adjustment; 2515 ++Lower; 2516 2517 RangeSet SymRange = getRange(St, Sym); 2518 return F.intersect(SymRange, Lower, Upper); 2519 } 2520 2521 ProgramStateRef 2522 RangeConstraintManager::assumeSymGT(ProgramStateRef St, SymbolRef Sym, 2523 const llvm::APSInt &Int, 2524 const llvm::APSInt &Adjustment) { 2525 RangeSet New = getSymGTRange(St, Sym, Int, Adjustment); 2526 return setRange(St, Sym, New); 2527 } 2528 2529 RangeSet RangeConstraintManager::getSymGERange(ProgramStateRef St, 2530 SymbolRef Sym, 2531 const llvm::APSInt &Int, 2532 const llvm::APSInt &Adjustment) { 2533 // Before we do any real work, see if the value can even show up. 2534 APSIntType AdjustmentType(Adjustment); 2535 switch (AdjustmentType.testInRange(Int, true)) { 2536 case APSIntType::RTR_Below: 2537 return getRange(St, Sym); 2538 case APSIntType::RTR_Within: 2539 break; 2540 case APSIntType::RTR_Above: 2541 return F.getEmptySet(); 2542 } 2543 2544 // Special case for Int == Min. This is always feasible. 2545 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2546 llvm::APSInt Min = AdjustmentType.getMinValue(); 2547 if (ComparisonVal == Min) 2548 return getRange(St, Sym); 2549 2550 llvm::APSInt Max = AdjustmentType.getMaxValue(); 2551 llvm::APSInt Lower = ComparisonVal - Adjustment; 2552 llvm::APSInt Upper = Max - Adjustment; 2553 2554 RangeSet SymRange = getRange(St, Sym); 2555 return F.intersect(SymRange, Lower, Upper); 2556 } 2557 2558 ProgramStateRef 2559 RangeConstraintManager::assumeSymGE(ProgramStateRef St, SymbolRef Sym, 2560 const llvm::APSInt &Int, 2561 const llvm::APSInt &Adjustment) { 2562 RangeSet New = getSymGERange(St, Sym, Int, Adjustment); 2563 return setRange(St, Sym, New); 2564 } 2565 2566 RangeSet 2567 RangeConstraintManager::getSymLERange(llvm::function_ref<RangeSet()> RS, 2568 const llvm::APSInt &Int, 2569 const llvm::APSInt &Adjustment) { 2570 // Before we do any real work, see if the value can even show up. 2571 APSIntType AdjustmentType(Adjustment); 2572 switch (AdjustmentType.testInRange(Int, true)) { 2573 case APSIntType::RTR_Below: 2574 return F.getEmptySet(); 2575 case APSIntType::RTR_Within: 2576 break; 2577 case APSIntType::RTR_Above: 2578 return RS(); 2579 } 2580 2581 // Special case for Int == Max. This is always feasible. 2582 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2583 llvm::APSInt Max = AdjustmentType.getMaxValue(); 2584 if (ComparisonVal == Max) 2585 return RS(); 2586 2587 llvm::APSInt Min = AdjustmentType.getMinValue(); 2588 llvm::APSInt Lower = Min - Adjustment; 2589 llvm::APSInt Upper = ComparisonVal - Adjustment; 2590 2591 RangeSet Default = RS(); 2592 return F.intersect(Default, Lower, Upper); 2593 } 2594 2595 RangeSet RangeConstraintManager::getSymLERange(ProgramStateRef St, 2596 SymbolRef Sym, 2597 const llvm::APSInt &Int, 2598 const llvm::APSInt &Adjustment) { 2599 return getSymLERange([&] { return getRange(St, Sym); }, Int, Adjustment); 2600 } 2601 2602 ProgramStateRef 2603 RangeConstraintManager::assumeSymLE(ProgramStateRef St, SymbolRef Sym, 2604 const llvm::APSInt &Int, 2605 const llvm::APSInt &Adjustment) { 2606 RangeSet New = getSymLERange(St, Sym, Int, Adjustment); 2607 return setRange(St, Sym, New); 2608 } 2609 2610 ProgramStateRef RangeConstraintManager::assumeSymWithinInclusiveRange( 2611 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 2612 const llvm::APSInt &To, const llvm::APSInt &Adjustment) { 2613 RangeSet New = getSymGERange(State, Sym, From, Adjustment); 2614 if (New.isEmpty()) 2615 return nullptr; 2616 RangeSet Out = getSymLERange([&] { return New; }, To, Adjustment); 2617 return setRange(State, Sym, Out); 2618 } 2619 2620 ProgramStateRef RangeConstraintManager::assumeSymOutsideInclusiveRange( 2621 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 2622 const llvm::APSInt &To, const llvm::APSInt &Adjustment) { 2623 RangeSet RangeLT = getSymLTRange(State, Sym, From, Adjustment); 2624 RangeSet RangeGT = getSymGTRange(State, Sym, To, Adjustment); 2625 RangeSet New(F.add(RangeLT, RangeGT)); 2626 return setRange(State, Sym, New); 2627 } 2628 2629 //===----------------------------------------------------------------------===// 2630 // Pretty-printing. 2631 //===----------------------------------------------------------------------===// 2632 2633 void RangeConstraintManager::printJson(raw_ostream &Out, ProgramStateRef State, 2634 const char *NL, unsigned int Space, 2635 bool IsDot) const { 2636 printConstraints(Out, State, NL, Space, IsDot); 2637 printEquivalenceClasses(Out, State, NL, Space, IsDot); 2638 printDisequalities(Out, State, NL, Space, IsDot); 2639 } 2640 2641 static std::string toString(const SymbolRef &Sym) { 2642 std::string S; 2643 llvm::raw_string_ostream O(S); 2644 Sym->dumpToStream(O); 2645 return O.str(); 2646 } 2647 2648 void RangeConstraintManager::printConstraints(raw_ostream &Out, 2649 ProgramStateRef State, 2650 const char *NL, 2651 unsigned int Space, 2652 bool IsDot) const { 2653 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 2654 2655 Indent(Out, Space, IsDot) << "\"constraints\": "; 2656 if (Constraints.isEmpty()) { 2657 Out << "null," << NL; 2658 return; 2659 } 2660 2661 std::map<std::string, RangeSet> OrderedConstraints; 2662 for (std::pair<EquivalenceClass, RangeSet> P : Constraints) { 2663 SymbolSet ClassMembers = P.first.getClassMembers(State); 2664 for (const SymbolRef &ClassMember : ClassMembers) { 2665 bool insertion_took_place; 2666 std::tie(std::ignore, insertion_took_place) = 2667 OrderedConstraints.insert({toString(ClassMember), P.second}); 2668 assert(insertion_took_place && 2669 "two symbols should not have the same dump"); 2670 } 2671 } 2672 2673 ++Space; 2674 Out << '[' << NL; 2675 bool First = true; 2676 for (std::pair<std::string, RangeSet> P : OrderedConstraints) { 2677 if (First) { 2678 First = false; 2679 } else { 2680 Out << ','; 2681 Out << NL; 2682 } 2683 Indent(Out, Space, IsDot) 2684 << "{ \"symbol\": \"" << P.first << "\", \"range\": \""; 2685 P.second.dump(Out); 2686 Out << "\" }"; 2687 } 2688 Out << NL; 2689 2690 --Space; 2691 Indent(Out, Space, IsDot) << "]," << NL; 2692 } 2693 2694 static std::string toString(ProgramStateRef State, EquivalenceClass Class) { 2695 SymbolSet ClassMembers = Class.getClassMembers(State); 2696 llvm::SmallVector<SymbolRef, 8> ClassMembersSorted(ClassMembers.begin(), 2697 ClassMembers.end()); 2698 llvm::sort(ClassMembersSorted, 2699 [](const SymbolRef &LHS, const SymbolRef &RHS) { 2700 return toString(LHS) < toString(RHS); 2701 }); 2702 2703 bool FirstMember = true; 2704 2705 std::string Str; 2706 llvm::raw_string_ostream Out(Str); 2707 Out << "[ "; 2708 for (SymbolRef ClassMember : ClassMembersSorted) { 2709 if (FirstMember) 2710 FirstMember = false; 2711 else 2712 Out << ", "; 2713 Out << "\"" << ClassMember << "\""; 2714 } 2715 Out << " ]"; 2716 return Out.str(); 2717 } 2718 2719 void RangeConstraintManager::printEquivalenceClasses(raw_ostream &Out, 2720 ProgramStateRef State, 2721 const char *NL, 2722 unsigned int Space, 2723 bool IsDot) const { 2724 ClassMembersTy Members = State->get<ClassMembers>(); 2725 2726 Indent(Out, Space, IsDot) << "\"equivalence_classes\": "; 2727 if (Members.isEmpty()) { 2728 Out << "null," << NL; 2729 return; 2730 } 2731 2732 std::set<std::string> MembersStr; 2733 for (std::pair<EquivalenceClass, SymbolSet> ClassToSymbolSet : Members) 2734 MembersStr.insert(toString(State, ClassToSymbolSet.first)); 2735 2736 ++Space; 2737 Out << '[' << NL; 2738 bool FirstClass = true; 2739 for (const std::string &Str : MembersStr) { 2740 if (FirstClass) { 2741 FirstClass = false; 2742 } else { 2743 Out << ','; 2744 Out << NL; 2745 } 2746 Indent(Out, Space, IsDot); 2747 Out << Str; 2748 } 2749 Out << NL; 2750 2751 --Space; 2752 Indent(Out, Space, IsDot) << "]," << NL; 2753 } 2754 2755 void RangeConstraintManager::printDisequalities(raw_ostream &Out, 2756 ProgramStateRef State, 2757 const char *NL, 2758 unsigned int Space, 2759 bool IsDot) const { 2760 DisequalityMapTy Disequalities = State->get<DisequalityMap>(); 2761 2762 Indent(Out, Space, IsDot) << "\"disequality_info\": "; 2763 if (Disequalities.isEmpty()) { 2764 Out << "null," << NL; 2765 return; 2766 } 2767 2768 // Transform the disequality info to an ordered map of 2769 // [string -> (ordered set of strings)] 2770 using EqClassesStrTy = std::set<std::string>; 2771 using DisequalityInfoStrTy = std::map<std::string, EqClassesStrTy>; 2772 DisequalityInfoStrTy DisequalityInfoStr; 2773 for (std::pair<EquivalenceClass, ClassSet> ClassToDisEqSet : Disequalities) { 2774 EquivalenceClass Class = ClassToDisEqSet.first; 2775 ClassSet DisequalClasses = ClassToDisEqSet.second; 2776 EqClassesStrTy MembersStr; 2777 for (EquivalenceClass DisEqClass : DisequalClasses) 2778 MembersStr.insert(toString(State, DisEqClass)); 2779 DisequalityInfoStr.insert({toString(State, Class), MembersStr}); 2780 } 2781 2782 ++Space; 2783 Out << '[' << NL; 2784 bool FirstClass = true; 2785 for (std::pair<std::string, EqClassesStrTy> ClassToDisEqSet : 2786 DisequalityInfoStr) { 2787 const std::string &Class = ClassToDisEqSet.first; 2788 if (FirstClass) { 2789 FirstClass = false; 2790 } else { 2791 Out << ','; 2792 Out << NL; 2793 } 2794 Indent(Out, Space, IsDot) << "{" << NL; 2795 unsigned int DisEqSpace = Space + 1; 2796 Indent(Out, DisEqSpace, IsDot) << "\"class\": "; 2797 Out << Class; 2798 const EqClassesStrTy &DisequalClasses = ClassToDisEqSet.second; 2799 if (!DisequalClasses.empty()) { 2800 Out << "," << NL; 2801 Indent(Out, DisEqSpace, IsDot) << "\"disequal_to\": [" << NL; 2802 unsigned int DisEqClassSpace = DisEqSpace + 1; 2803 Indent(Out, DisEqClassSpace, IsDot); 2804 bool FirstDisEqClass = true; 2805 for (const std::string &DisEqClass : DisequalClasses) { 2806 if (FirstDisEqClass) { 2807 FirstDisEqClass = false; 2808 } else { 2809 Out << ',' << NL; 2810 Indent(Out, DisEqClassSpace, IsDot); 2811 } 2812 Out << DisEqClass; 2813 } 2814 Out << "]" << NL; 2815 } 2816 Indent(Out, Space, IsDot) << "}"; 2817 } 2818 Out << NL; 2819 2820 --Space; 2821 Indent(Out, Space, IsDot) << "]," << NL; 2822 } 2823