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