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 LLVM_DUMP_METHOD void Range::dump(raw_ostream &OS) const { 498 OS << '[' << toString(From(), 10) << ", " << toString(To(), 10) << ']'; 499 } 500 LLVM_DUMP_METHOD void Range::dump() const { dump(llvm::errs()); } 501 502 LLVM_DUMP_METHOD void RangeSet::dump(raw_ostream &OS) const { 503 OS << "{ "; 504 llvm::interleaveComma(*this, OS, [&OS](const Range &R) { R.dump(OS); }); 505 OS << " }"; 506 } 507 LLVM_DUMP_METHOD void RangeSet::dump() const { dump(llvm::errs()); } 508 509 REGISTER_SET_FACTORY_WITH_PROGRAMSTATE(SymbolSet, SymbolRef) 510 511 namespace { 512 class EquivalenceClass; 513 } // end anonymous namespace 514 515 REGISTER_MAP_WITH_PROGRAMSTATE(ClassMap, SymbolRef, EquivalenceClass) 516 REGISTER_MAP_WITH_PROGRAMSTATE(ClassMembers, EquivalenceClass, SymbolSet) 517 REGISTER_MAP_WITH_PROGRAMSTATE(ConstraintRange, EquivalenceClass, RangeSet) 518 519 REGISTER_SET_FACTORY_WITH_PROGRAMSTATE(ClassSet, EquivalenceClass) 520 REGISTER_MAP_WITH_PROGRAMSTATE(DisequalityMap, EquivalenceClass, ClassSet) 521 522 namespace { 523 /// This class encapsulates a set of symbols equal to each other. 524 /// 525 /// The main idea of the approach requiring such classes is in narrowing 526 /// and sharing constraints between symbols within the class. Also we can 527 /// conclude that there is no practical need in storing constraints for 528 /// every member of the class separately. 529 /// 530 /// Main terminology: 531 /// 532 /// * "Equivalence class" is an object of this class, which can be efficiently 533 /// compared to other classes. It represents the whole class without 534 /// storing the actual in it. The members of the class however can be 535 /// retrieved from the state. 536 /// 537 /// * "Class members" are the symbols corresponding to the class. This means 538 /// that A == B for every member symbols A and B from the class. Members of 539 /// each class are stored in the state. 540 /// 541 /// * "Trivial class" is a class that has and ever had only one same symbol. 542 /// 543 /// * "Merge operation" merges two classes into one. It is the main operation 544 /// to produce non-trivial classes. 545 /// If, at some point, we can assume that two symbols from two distinct 546 /// classes are equal, we can merge these classes. 547 class EquivalenceClass : public llvm::FoldingSetNode { 548 public: 549 /// Find equivalence class for the given symbol in the given state. 550 LLVM_NODISCARD static inline EquivalenceClass find(ProgramStateRef State, 551 SymbolRef Sym); 552 553 /// Merge classes for the given symbols and return a new state. 554 LLVM_NODISCARD static inline ProgramStateRef merge(RangeSet::Factory &F, 555 ProgramStateRef State, 556 SymbolRef First, 557 SymbolRef Second); 558 // Merge this class with the given class and return a new state. 559 LLVM_NODISCARD inline ProgramStateRef 560 merge(RangeSet::Factory &F, ProgramStateRef State, EquivalenceClass Other); 561 562 /// Return a set of class members for the given state. 563 LLVM_NODISCARD inline SymbolSet getClassMembers(ProgramStateRef State) const; 564 565 /// Return true if the current class is trivial in the given state. 566 /// A class is trivial if and only if there is not any member relations stored 567 /// to it in State/ClassMembers. 568 /// An equivalence class with one member might seem as it does not hold any 569 /// meaningful information, i.e. that is a tautology. However, during the 570 /// removal of dead symbols we do not remove classes with one member for 571 /// resource and performance reasons. Consequently, a class with one member is 572 /// not necessarily trivial. It could happen that we have a class with two 573 /// members and then during the removal of dead symbols we remove one of its 574 /// members. In this case, the class is still non-trivial (it still has the 575 /// mappings in ClassMembers), even though it has only one member. 576 LLVM_NODISCARD inline bool isTrivial(ProgramStateRef State) const; 577 578 /// Return true if the current class is trivial and its only member is dead. 579 LLVM_NODISCARD inline bool isTriviallyDead(ProgramStateRef State, 580 SymbolReaper &Reaper) const; 581 582 LLVM_NODISCARD static inline ProgramStateRef 583 markDisequal(RangeSet::Factory &F, ProgramStateRef State, SymbolRef First, 584 SymbolRef Second); 585 LLVM_NODISCARD static inline ProgramStateRef 586 markDisequal(RangeSet::Factory &F, ProgramStateRef State, 587 EquivalenceClass First, EquivalenceClass Second); 588 LLVM_NODISCARD inline ProgramStateRef 589 markDisequal(RangeSet::Factory &F, ProgramStateRef State, 590 EquivalenceClass Other) const; 591 LLVM_NODISCARD static inline ClassSet 592 getDisequalClasses(ProgramStateRef State, SymbolRef Sym); 593 LLVM_NODISCARD inline ClassSet 594 getDisequalClasses(ProgramStateRef State) const; 595 LLVM_NODISCARD inline ClassSet 596 getDisequalClasses(DisequalityMapTy Map, ClassSet::Factory &Factory) const; 597 598 LLVM_NODISCARD static inline Optional<bool> areEqual(ProgramStateRef State, 599 EquivalenceClass First, 600 EquivalenceClass Second); 601 LLVM_NODISCARD static inline Optional<bool> 602 areEqual(ProgramStateRef State, SymbolRef First, SymbolRef Second); 603 604 /// Iterate over all symbols and try to simplify them. 605 LLVM_NODISCARD static inline ProgramStateRef simplify(SValBuilder &SVB, 606 RangeSet::Factory &F, 607 ProgramStateRef State, 608 EquivalenceClass Class); 609 610 void dumpToStream(ProgramStateRef State, raw_ostream &os) const; 611 LLVM_DUMP_METHOD void dump(ProgramStateRef State) const { 612 dumpToStream(State, llvm::errs()); 613 } 614 615 /// Check equivalence data for consistency. 616 LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED static bool 617 isClassDataConsistent(ProgramStateRef State); 618 619 LLVM_NODISCARD QualType getType() const { 620 return getRepresentativeSymbol()->getType(); 621 } 622 623 EquivalenceClass() = delete; 624 EquivalenceClass(const EquivalenceClass &) = default; 625 EquivalenceClass &operator=(const EquivalenceClass &) = delete; 626 EquivalenceClass(EquivalenceClass &&) = default; 627 EquivalenceClass &operator=(EquivalenceClass &&) = delete; 628 629 bool operator==(const EquivalenceClass &Other) const { 630 return ID == Other.ID; 631 } 632 bool operator<(const EquivalenceClass &Other) const { return ID < Other.ID; } 633 bool operator!=(const EquivalenceClass &Other) const { 634 return !operator==(Other); 635 } 636 637 static void Profile(llvm::FoldingSetNodeID &ID, uintptr_t CID) { 638 ID.AddInteger(CID); 639 } 640 641 void Profile(llvm::FoldingSetNodeID &ID) const { Profile(ID, this->ID); } 642 643 private: 644 /* implicit */ EquivalenceClass(SymbolRef Sym) 645 : ID(reinterpret_cast<uintptr_t>(Sym)) {} 646 647 /// This function is intended to be used ONLY within the class. 648 /// The fact that ID is a pointer to a symbol is an implementation detail 649 /// and should stay that way. 650 /// In the current implementation, we use it to retrieve the only member 651 /// of the trivial class. 652 SymbolRef getRepresentativeSymbol() const { 653 return reinterpret_cast<SymbolRef>(ID); 654 } 655 static inline SymbolSet::Factory &getMembersFactory(ProgramStateRef State); 656 657 inline ProgramStateRef mergeImpl(RangeSet::Factory &F, ProgramStateRef State, 658 SymbolSet Members, EquivalenceClass Other, 659 SymbolSet OtherMembers); 660 static inline bool 661 addToDisequalityInfo(DisequalityMapTy &Info, ConstraintRangeTy &Constraints, 662 RangeSet::Factory &F, ProgramStateRef State, 663 EquivalenceClass First, EquivalenceClass Second); 664 665 /// This is a unique identifier of the class. 666 uintptr_t ID; 667 }; 668 669 //===----------------------------------------------------------------------===// 670 // Constraint functions 671 //===----------------------------------------------------------------------===// 672 673 LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED bool 674 areFeasible(ConstraintRangeTy Constraints) { 675 return llvm::none_of( 676 Constraints, 677 [](const std::pair<EquivalenceClass, RangeSet> &ClassConstraint) { 678 return ClassConstraint.second.isEmpty(); 679 }); 680 } 681 682 LLVM_NODISCARD inline const RangeSet *getConstraint(ProgramStateRef State, 683 EquivalenceClass Class) { 684 return State->get<ConstraintRange>(Class); 685 } 686 687 LLVM_NODISCARD inline const RangeSet *getConstraint(ProgramStateRef State, 688 SymbolRef Sym) { 689 return getConstraint(State, EquivalenceClass::find(State, Sym)); 690 } 691 692 LLVM_NODISCARD ProgramStateRef setConstraint(ProgramStateRef State, 693 EquivalenceClass Class, 694 RangeSet Constraint) { 695 return State->set<ConstraintRange>(Class, Constraint); 696 } 697 698 LLVM_NODISCARD ProgramStateRef setConstraints(ProgramStateRef State, 699 ConstraintRangeTy Constraints) { 700 return State->set<ConstraintRange>(Constraints); 701 } 702 703 //===----------------------------------------------------------------------===// 704 // Equality/diseqiality abstraction 705 //===----------------------------------------------------------------------===// 706 707 /// A small helper function for detecting symbolic (dis)equality. 708 /// 709 /// Equality check can have different forms (like a == b or a - b) and this 710 /// class encapsulates those away if the only thing the user wants to check - 711 /// whether it's equality/diseqiality or not. 712 /// 713 /// \returns true if assuming this Sym to be true means equality of operands 714 /// false if it means disequality of operands 715 /// None otherwise 716 Optional<bool> meansEquality(const SymSymExpr *Sym) { 717 switch (Sym->getOpcode()) { 718 case BO_Sub: 719 // This case is: A - B != 0 -> disequality check. 720 return false; 721 case BO_EQ: 722 // This case is: A == B != 0 -> equality check. 723 return true; 724 case BO_NE: 725 // This case is: A != B != 0 -> diseqiality check. 726 return false; 727 default: 728 return llvm::None; 729 } 730 } 731 732 //===----------------------------------------------------------------------===// 733 // Intersection functions 734 //===----------------------------------------------------------------------===// 735 736 template <class SecondTy, class... RestTy> 737 LLVM_NODISCARD inline RangeSet intersect(RangeSet::Factory &F, RangeSet Head, 738 SecondTy Second, RestTy... Tail); 739 740 template <class... RangeTy> struct IntersectionTraits; 741 742 template <class... TailTy> struct IntersectionTraits<RangeSet, TailTy...> { 743 // Found RangeSet, no need to check any further 744 using Type = RangeSet; 745 }; 746 747 template <> struct IntersectionTraits<> { 748 // We ran out of types, and we didn't find any RangeSet, so the result should 749 // be optional. 750 using Type = Optional<RangeSet>; 751 }; 752 753 template <class OptionalOrPointer, class... TailTy> 754 struct IntersectionTraits<OptionalOrPointer, TailTy...> { 755 // If current type is Optional or a raw pointer, we should keep looking. 756 using Type = typename IntersectionTraits<TailTy...>::Type; 757 }; 758 759 template <class EndTy> 760 LLVM_NODISCARD inline EndTy intersect(RangeSet::Factory &F, EndTy End) { 761 // If the list contains only RangeSet or Optional<RangeSet>, simply return 762 // that range set. 763 return End; 764 } 765 766 LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED inline Optional<RangeSet> 767 intersect(RangeSet::Factory &F, const RangeSet *End) { 768 // This is an extraneous conversion from a raw pointer into Optional<RangeSet> 769 if (End) { 770 return *End; 771 } 772 return llvm::None; 773 } 774 775 template <class... RestTy> 776 LLVM_NODISCARD inline RangeSet intersect(RangeSet::Factory &F, RangeSet Head, 777 RangeSet Second, RestTy... Tail) { 778 // Here we call either the <RangeSet,RangeSet,...> or <RangeSet,...> version 779 // of the function and can be sure that the result is RangeSet. 780 return intersect(F, F.intersect(Head, Second), Tail...); 781 } 782 783 template <class SecondTy, class... RestTy> 784 LLVM_NODISCARD inline RangeSet intersect(RangeSet::Factory &F, RangeSet Head, 785 SecondTy Second, RestTy... Tail) { 786 if (Second) { 787 // Here we call the <RangeSet,RangeSet,...> version of the function... 788 return intersect(F, Head, *Second, Tail...); 789 } 790 // ...and here it is either <RangeSet,RangeSet,...> or <RangeSet,...>, which 791 // means that the result is definitely RangeSet. 792 return intersect(F, Head, Tail...); 793 } 794 795 /// Main generic intersect function. 796 /// It intersects all of the given range sets. If some of the given arguments 797 /// don't hold a range set (nullptr or llvm::None), the function will skip them. 798 /// 799 /// Available representations for the arguments are: 800 /// * RangeSet 801 /// * Optional<RangeSet> 802 /// * RangeSet * 803 /// Pointer to a RangeSet is automatically assumed to be nullable and will get 804 /// checked as well as the optional version. If this behaviour is undesired, 805 /// please dereference the pointer in the call. 806 /// 807 /// Return type depends on the arguments' types. If we can be sure in compile 808 /// time that there will be a range set as a result, the returning type is 809 /// simply RangeSet, in other cases we have to back off to Optional<RangeSet>. 810 /// 811 /// Please, prefer optional range sets to raw pointers. If the last argument is 812 /// a raw pointer and all previous arguments are None, it will cost one 813 /// additional check to convert RangeSet * into Optional<RangeSet>. 814 template <class HeadTy, class SecondTy, class... RestTy> 815 LLVM_NODISCARD inline 816 typename IntersectionTraits<HeadTy, SecondTy, RestTy...>::Type 817 intersect(RangeSet::Factory &F, HeadTy Head, SecondTy Second, 818 RestTy... Tail) { 819 if (Head) { 820 return intersect(F, *Head, Second, Tail...); 821 } 822 return intersect(F, Second, Tail...); 823 } 824 825 //===----------------------------------------------------------------------===// 826 // Symbolic reasoning logic 827 //===----------------------------------------------------------------------===// 828 829 /// A little component aggregating all of the reasoning we have about 830 /// the ranges of symbolic expressions. 831 /// 832 /// Even when we don't know the exact values of the operands, we still 833 /// can get a pretty good estimate of the result's range. 834 class SymbolicRangeInferrer 835 : public SymExprVisitor<SymbolicRangeInferrer, RangeSet> { 836 public: 837 template <class SourceType> 838 static RangeSet inferRange(RangeSet::Factory &F, ProgramStateRef State, 839 SourceType Origin) { 840 SymbolicRangeInferrer Inferrer(F, State); 841 return Inferrer.infer(Origin); 842 } 843 844 RangeSet VisitSymExpr(SymbolRef Sym) { 845 // If we got to this function, the actual type of the symbolic 846 // expression is not supported for advanced inference. 847 // In this case, we simply backoff to the default "let's simply 848 // infer the range from the expression's type". 849 return infer(Sym->getType()); 850 } 851 852 RangeSet VisitSymIntExpr(const SymIntExpr *Sym) { 853 return VisitBinaryOperator(Sym); 854 } 855 856 RangeSet VisitIntSymExpr(const IntSymExpr *Sym) { 857 return VisitBinaryOperator(Sym); 858 } 859 860 RangeSet VisitSymSymExpr(const SymSymExpr *Sym) { 861 return intersect( 862 RangeFactory, 863 // If Sym is (dis)equality, we might have some information 864 // on that in our equality classes data structure. 865 getRangeForEqualities(Sym), 866 // And we should always check what we can get from the operands. 867 VisitBinaryOperator(Sym)); 868 } 869 870 private: 871 SymbolicRangeInferrer(RangeSet::Factory &F, ProgramStateRef S) 872 : ValueFactory(F.getValueFactory()), RangeFactory(F), State(S) {} 873 874 /// Infer range information from the given integer constant. 875 /// 876 /// It's not a real "inference", but is here for operating with 877 /// sub-expressions in a more polymorphic manner. 878 RangeSet inferAs(const llvm::APSInt &Val, QualType) { 879 return {RangeFactory, Val}; 880 } 881 882 /// Infer range information from symbol in the context of the given type. 883 RangeSet inferAs(SymbolRef Sym, QualType DestType) { 884 QualType ActualType = Sym->getType(); 885 // Check that we can reason about the symbol at all. 886 if (ActualType->isIntegralOrEnumerationType() || 887 Loc::isLocType(ActualType)) { 888 return infer(Sym); 889 } 890 // Otherwise, let's simply infer from the destination type. 891 // We couldn't figure out nothing else about that expression. 892 return infer(DestType); 893 } 894 895 RangeSet infer(SymbolRef Sym) { 896 return intersect( 897 RangeFactory, 898 // Of course, we should take the constraint directly associated with 899 // this symbol into consideration. 900 getConstraint(State, Sym), 901 // If Sym is a difference of symbols A - B, then maybe we have range 902 // set stored for B - A. 903 // 904 // If we have range set stored for both A - B and B - A then 905 // calculate the effective range set by intersecting the range set 906 // for A - B and the negated range set of B - A. 907 getRangeForNegatedSub(Sym), 908 // If Sym is a comparison expression (except <=>), 909 // find any other comparisons with the same operands. 910 // See function description. 911 getRangeForComparisonSymbol(Sym), 912 // Apart from the Sym itself, we can infer quite a lot if we look 913 // into subexpressions of Sym. 914 Visit(Sym)); 915 } 916 917 RangeSet infer(EquivalenceClass Class) { 918 if (const RangeSet *AssociatedConstraint = getConstraint(State, Class)) 919 return *AssociatedConstraint; 920 921 return infer(Class.getType()); 922 } 923 924 /// Infer range information solely from the type. 925 RangeSet infer(QualType T) { 926 // Lazily generate a new RangeSet representing all possible values for the 927 // given symbol type. 928 RangeSet Result(RangeFactory, ValueFactory.getMinValue(T), 929 ValueFactory.getMaxValue(T)); 930 931 // References are known to be non-zero. 932 if (T->isReferenceType()) 933 return assumeNonZero(Result, T); 934 935 return Result; 936 } 937 938 template <class BinarySymExprTy> 939 RangeSet VisitBinaryOperator(const BinarySymExprTy *Sym) { 940 // TODO #1: VisitBinaryOperator implementation might not make a good 941 // use of the inferred ranges. In this case, we might be calculating 942 // everything for nothing. This being said, we should introduce some 943 // sort of laziness mechanism here. 944 // 945 // TODO #2: We didn't go into the nested expressions before, so it 946 // might cause us spending much more time doing the inference. 947 // This can be a problem for deeply nested expressions that are 948 // involved in conditions and get tested continuously. We definitely 949 // need to address this issue and introduce some sort of caching 950 // in here. 951 QualType ResultType = Sym->getType(); 952 return VisitBinaryOperator(inferAs(Sym->getLHS(), ResultType), 953 Sym->getOpcode(), 954 inferAs(Sym->getRHS(), ResultType), ResultType); 955 } 956 957 RangeSet VisitBinaryOperator(RangeSet LHS, BinaryOperator::Opcode Op, 958 RangeSet RHS, QualType T) { 959 switch (Op) { 960 case BO_Or: 961 return VisitBinaryOperator<BO_Or>(LHS, RHS, T); 962 case BO_And: 963 return VisitBinaryOperator<BO_And>(LHS, RHS, T); 964 case BO_Rem: 965 return VisitBinaryOperator<BO_Rem>(LHS, RHS, T); 966 default: 967 return infer(T); 968 } 969 } 970 971 //===----------------------------------------------------------------------===// 972 // Ranges and operators 973 //===----------------------------------------------------------------------===// 974 975 /// Return a rough approximation of the given range set. 976 /// 977 /// For the range set: 978 /// { [x_0, y_0], [x_1, y_1], ... , [x_N, y_N] } 979 /// it will return the range [x_0, y_N]. 980 static Range fillGaps(RangeSet Origin) { 981 assert(!Origin.isEmpty()); 982 return {Origin.getMinValue(), Origin.getMaxValue()}; 983 } 984 985 /// Try to convert given range into the given type. 986 /// 987 /// It will return llvm::None only when the trivial conversion is possible. 988 llvm::Optional<Range> convert(const Range &Origin, APSIntType To) { 989 if (To.testInRange(Origin.From(), false) != APSIntType::RTR_Within || 990 To.testInRange(Origin.To(), false) != APSIntType::RTR_Within) { 991 return llvm::None; 992 } 993 return Range(ValueFactory.Convert(To, Origin.From()), 994 ValueFactory.Convert(To, Origin.To())); 995 } 996 997 template <BinaryOperator::Opcode Op> 998 RangeSet VisitBinaryOperator(RangeSet LHS, RangeSet RHS, QualType T) { 999 // We should propagate information about unfeasbility of one of the 1000 // operands to the resulting range. 1001 if (LHS.isEmpty() || RHS.isEmpty()) { 1002 return RangeFactory.getEmptySet(); 1003 } 1004 1005 Range CoarseLHS = fillGaps(LHS); 1006 Range CoarseRHS = fillGaps(RHS); 1007 1008 APSIntType ResultType = ValueFactory.getAPSIntType(T); 1009 1010 // We need to convert ranges to the resulting type, so we can compare values 1011 // and combine them in a meaningful (in terms of the given operation) way. 1012 auto ConvertedCoarseLHS = convert(CoarseLHS, ResultType); 1013 auto ConvertedCoarseRHS = convert(CoarseRHS, ResultType); 1014 1015 // It is hard to reason about ranges when conversion changes 1016 // borders of the ranges. 1017 if (!ConvertedCoarseLHS || !ConvertedCoarseRHS) { 1018 return infer(T); 1019 } 1020 1021 return VisitBinaryOperator<Op>(*ConvertedCoarseLHS, *ConvertedCoarseRHS, T); 1022 } 1023 1024 template <BinaryOperator::Opcode Op> 1025 RangeSet VisitBinaryOperator(Range LHS, Range RHS, QualType T) { 1026 return infer(T); 1027 } 1028 1029 /// Return a symmetrical range for the given range and type. 1030 /// 1031 /// If T is signed, return the smallest range [-x..x] that covers the original 1032 /// range, or [-min(T), max(T)] if the aforementioned symmetric range doesn't 1033 /// exist due to original range covering min(T)). 1034 /// 1035 /// If T is unsigned, return the smallest range [0..x] that covers the 1036 /// original range. 1037 Range getSymmetricalRange(Range Origin, QualType T) { 1038 APSIntType RangeType = ValueFactory.getAPSIntType(T); 1039 1040 if (RangeType.isUnsigned()) { 1041 return Range(ValueFactory.getMinValue(RangeType), Origin.To()); 1042 } 1043 1044 if (Origin.From().isMinSignedValue()) { 1045 // If mini is a minimal signed value, absolute value of it is greater 1046 // than the maximal signed value. In order to avoid these 1047 // complications, we simply return the whole range. 1048 return {ValueFactory.getMinValue(RangeType), 1049 ValueFactory.getMaxValue(RangeType)}; 1050 } 1051 1052 // At this point, we are sure that the type is signed and we can safely 1053 // use unary - operator. 1054 // 1055 // While calculating absolute maximum, we can use the following formula 1056 // because of these reasons: 1057 // * If From >= 0 then To >= From and To >= -From. 1058 // AbsMax == To == max(To, -From) 1059 // * If To <= 0 then -From >= -To and -From >= From. 1060 // AbsMax == -From == max(-From, To) 1061 // * Otherwise, From <= 0, To >= 0, and 1062 // AbsMax == max(abs(From), abs(To)) 1063 llvm::APSInt AbsMax = std::max(-Origin.From(), Origin.To()); 1064 1065 // Intersection is guaranteed to be non-empty. 1066 return {ValueFactory.getValue(-AbsMax), ValueFactory.getValue(AbsMax)}; 1067 } 1068 1069 /// Return a range set subtracting zero from \p Domain. 1070 RangeSet assumeNonZero(RangeSet Domain, QualType T) { 1071 APSIntType IntType = ValueFactory.getAPSIntType(T); 1072 return RangeFactory.deletePoint(Domain, IntType.getZeroValue()); 1073 } 1074 1075 // FIXME: Once SValBuilder supports unary minus, we should use SValBuilder to 1076 // obtain the negated symbolic expression instead of constructing the 1077 // symbol manually. This will allow us to support finding ranges of not 1078 // only negated SymSymExpr-type expressions, but also of other, simpler 1079 // expressions which we currently do not know how to negate. 1080 Optional<RangeSet> getRangeForNegatedSub(SymbolRef Sym) { 1081 if (const SymSymExpr *SSE = dyn_cast<SymSymExpr>(Sym)) { 1082 if (SSE->getOpcode() == BO_Sub) { 1083 QualType T = Sym->getType(); 1084 1085 // Do not negate unsigned ranges 1086 if (!T->isUnsignedIntegerOrEnumerationType() && 1087 !T->isSignedIntegerOrEnumerationType()) 1088 return llvm::None; 1089 1090 SymbolManager &SymMgr = State->getSymbolManager(); 1091 SymbolRef NegatedSym = 1092 SymMgr.getSymSymExpr(SSE->getRHS(), BO_Sub, SSE->getLHS(), T); 1093 1094 if (const RangeSet *NegatedRange = getConstraint(State, NegatedSym)) { 1095 return RangeFactory.negate(*NegatedRange); 1096 } 1097 } 1098 } 1099 return llvm::None; 1100 } 1101 1102 // Returns ranges only for binary comparison operators (except <=>) 1103 // when left and right operands are symbolic values. 1104 // Finds any other comparisons with the same operands. 1105 // Then do logical calculations and refuse impossible branches. 1106 // E.g. (x < y) and (x > y) at the same time are impossible. 1107 // E.g. (x >= y) and (x != y) at the same time makes (x > y) true only. 1108 // E.g. (x == y) and (y == x) are just reversed but the same. 1109 // It covers all possible combinations (see CmpOpTable description). 1110 // Note that `x` and `y` can also stand for subexpressions, 1111 // not only for actual symbols. 1112 Optional<RangeSet> getRangeForComparisonSymbol(SymbolRef Sym) { 1113 const auto *SSE = dyn_cast<SymSymExpr>(Sym); 1114 if (!SSE) 1115 return llvm::None; 1116 1117 const BinaryOperatorKind CurrentOP = SSE->getOpcode(); 1118 1119 // We currently do not support <=> (C++20). 1120 if (!BinaryOperator::isComparisonOp(CurrentOP) || (CurrentOP == BO_Cmp)) 1121 return llvm::None; 1122 1123 static const OperatorRelationsTable CmpOpTable{}; 1124 1125 const SymExpr *LHS = SSE->getLHS(); 1126 const SymExpr *RHS = SSE->getRHS(); 1127 QualType T = SSE->getType(); 1128 1129 SymbolManager &SymMgr = State->getSymbolManager(); 1130 1131 // We use this variable to store the last queried operator (`QueriedOP`) 1132 // for which the `getCmpOpState` returned with `Unknown`. If there are two 1133 // different OPs that returned `Unknown` then we have to query the special 1134 // `UnknownX2` column. We assume that `getCmpOpState(CurrentOP, CurrentOP)` 1135 // never returns `Unknown`, so `CurrentOP` is a good initial value. 1136 BinaryOperatorKind LastQueriedOpToUnknown = CurrentOP; 1137 1138 // Loop goes through all of the columns exept the last one ('UnknownX2'). 1139 // We treat `UnknownX2` column separately at the end of the loop body. 1140 for (size_t i = 0; i < CmpOpTable.getCmpOpCount(); ++i) { 1141 1142 // Let's find an expression e.g. (x < y). 1143 BinaryOperatorKind QueriedOP = OperatorRelationsTable::getOpFromIndex(i); 1144 const SymSymExpr *SymSym = SymMgr.getSymSymExpr(LHS, QueriedOP, RHS, T); 1145 const RangeSet *QueriedRangeSet = getConstraint(State, SymSym); 1146 1147 // If ranges were not previously found, 1148 // try to find a reversed expression (y > x). 1149 if (!QueriedRangeSet) { 1150 const BinaryOperatorKind ROP = 1151 BinaryOperator::reverseComparisonOp(QueriedOP); 1152 SymSym = SymMgr.getSymSymExpr(RHS, ROP, LHS, T); 1153 QueriedRangeSet = getConstraint(State, SymSym); 1154 } 1155 1156 if (!QueriedRangeSet || QueriedRangeSet->isEmpty()) 1157 continue; 1158 1159 const llvm::APSInt *ConcreteValue = QueriedRangeSet->getConcreteValue(); 1160 const bool isInFalseBranch = 1161 ConcreteValue ? (*ConcreteValue == 0) : false; 1162 1163 // If it is a false branch, we shall be guided by opposite operator, 1164 // because the table is made assuming we are in the true branch. 1165 // E.g. when (x <= y) is false, then (x > y) is true. 1166 if (isInFalseBranch) 1167 QueriedOP = BinaryOperator::negateComparisonOp(QueriedOP); 1168 1169 OperatorRelationsTable::TriStateKind BranchState = 1170 CmpOpTable.getCmpOpState(CurrentOP, QueriedOP); 1171 1172 if (BranchState == OperatorRelationsTable::Unknown) { 1173 if (LastQueriedOpToUnknown != CurrentOP && 1174 LastQueriedOpToUnknown != QueriedOP) { 1175 // If we got the Unknown state for both different operators. 1176 // if (x <= y) // assume true 1177 // if (x != y) // assume true 1178 // if (x < y) // would be also true 1179 // Get a state from `UnknownX2` column. 1180 BranchState = CmpOpTable.getCmpOpStateForUnknownX2(CurrentOP); 1181 } else { 1182 LastQueriedOpToUnknown = QueriedOP; 1183 continue; 1184 } 1185 } 1186 1187 return (BranchState == OperatorRelationsTable::True) ? getTrueRange(T) 1188 : getFalseRange(T); 1189 } 1190 1191 return llvm::None; 1192 } 1193 1194 Optional<RangeSet> getRangeForEqualities(const SymSymExpr *Sym) { 1195 Optional<bool> Equality = meansEquality(Sym); 1196 1197 if (!Equality) 1198 return llvm::None; 1199 1200 if (Optional<bool> AreEqual = 1201 EquivalenceClass::areEqual(State, Sym->getLHS(), Sym->getRHS())) { 1202 // Here we cover two cases at once: 1203 // * if Sym is equality and its operands are known to be equal -> true 1204 // * if Sym is disequality and its operands are disequal -> true 1205 if (*AreEqual == *Equality) { 1206 return getTrueRange(Sym->getType()); 1207 } 1208 // Opposite combinations result in false. 1209 return getFalseRange(Sym->getType()); 1210 } 1211 1212 return llvm::None; 1213 } 1214 1215 RangeSet getTrueRange(QualType T) { 1216 RangeSet TypeRange = infer(T); 1217 return assumeNonZero(TypeRange, T); 1218 } 1219 1220 RangeSet getFalseRange(QualType T) { 1221 const llvm::APSInt &Zero = ValueFactory.getValue(0, T); 1222 return RangeSet(RangeFactory, Zero); 1223 } 1224 1225 BasicValueFactory &ValueFactory; 1226 RangeSet::Factory &RangeFactory; 1227 ProgramStateRef State; 1228 }; 1229 1230 //===----------------------------------------------------------------------===// 1231 // Range-based reasoning about symbolic operations 1232 //===----------------------------------------------------------------------===// 1233 1234 template <> 1235 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_Or>(Range LHS, Range RHS, 1236 QualType T) { 1237 APSIntType ResultType = ValueFactory.getAPSIntType(T); 1238 llvm::APSInt Zero = ResultType.getZeroValue(); 1239 1240 bool IsLHSPositiveOrZero = LHS.From() >= Zero; 1241 bool IsRHSPositiveOrZero = RHS.From() >= Zero; 1242 1243 bool IsLHSNegative = LHS.To() < Zero; 1244 bool IsRHSNegative = RHS.To() < Zero; 1245 1246 // Check if both ranges have the same sign. 1247 if ((IsLHSPositiveOrZero && IsRHSPositiveOrZero) || 1248 (IsLHSNegative && IsRHSNegative)) { 1249 // The result is definitely greater or equal than any of the operands. 1250 const llvm::APSInt &Min = std::max(LHS.From(), RHS.From()); 1251 1252 // We estimate maximal value for positives as the maximal value for the 1253 // given type. For negatives, we estimate it with -1 (e.g. 0x11111111). 1254 // 1255 // TODO: We basically, limit the resulting range from below, but don't do 1256 // anything with the upper bound. 1257 // 1258 // For positive operands, it can be done as follows: for the upper 1259 // bound of LHS and RHS we calculate the most significant bit set. 1260 // Let's call it the N-th bit. Then we can estimate the maximal 1261 // number to be 2^(N+1)-1, i.e. the number with all the bits up to 1262 // the N-th bit set. 1263 const llvm::APSInt &Max = IsLHSNegative 1264 ? ValueFactory.getValue(--Zero) 1265 : ValueFactory.getMaxValue(ResultType); 1266 1267 return {RangeFactory, ValueFactory.getValue(Min), Max}; 1268 } 1269 1270 // Otherwise, let's check if at least one of the operands is negative. 1271 if (IsLHSNegative || IsRHSNegative) { 1272 // This means that the result is definitely negative as well. 1273 return {RangeFactory, ValueFactory.getMinValue(ResultType), 1274 ValueFactory.getValue(--Zero)}; 1275 } 1276 1277 RangeSet DefaultRange = infer(T); 1278 1279 // It is pretty hard to reason about operands with different signs 1280 // (and especially with possibly different signs). We simply check if it 1281 // can be zero. In order to conclude that the result could not be zero, 1282 // at least one of the operands should be definitely not zero itself. 1283 if (!LHS.Includes(Zero) || !RHS.Includes(Zero)) { 1284 return assumeNonZero(DefaultRange, T); 1285 } 1286 1287 // Nothing much else to do here. 1288 return DefaultRange; 1289 } 1290 1291 template <> 1292 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_And>(Range LHS, 1293 Range RHS, 1294 QualType T) { 1295 APSIntType ResultType = ValueFactory.getAPSIntType(T); 1296 llvm::APSInt Zero = ResultType.getZeroValue(); 1297 1298 bool IsLHSPositiveOrZero = LHS.From() >= Zero; 1299 bool IsRHSPositiveOrZero = RHS.From() >= Zero; 1300 1301 bool IsLHSNegative = LHS.To() < Zero; 1302 bool IsRHSNegative = RHS.To() < Zero; 1303 1304 // Check if both ranges have the same sign. 1305 if ((IsLHSPositiveOrZero && IsRHSPositiveOrZero) || 1306 (IsLHSNegative && IsRHSNegative)) { 1307 // The result is definitely less or equal than any of the operands. 1308 const llvm::APSInt &Max = std::min(LHS.To(), RHS.To()); 1309 1310 // We conservatively estimate lower bound to be the smallest positive 1311 // or negative value corresponding to the sign of the operands. 1312 const llvm::APSInt &Min = IsLHSNegative 1313 ? ValueFactory.getMinValue(ResultType) 1314 : ValueFactory.getValue(Zero); 1315 1316 return {RangeFactory, Min, Max}; 1317 } 1318 1319 // Otherwise, let's check if at least one of the operands is positive. 1320 if (IsLHSPositiveOrZero || IsRHSPositiveOrZero) { 1321 // This makes result definitely positive. 1322 // 1323 // We can also reason about a maximal value by finding the maximal 1324 // value of the positive operand. 1325 const llvm::APSInt &Max = IsLHSPositiveOrZero ? LHS.To() : RHS.To(); 1326 1327 // The minimal value on the other hand is much harder to reason about. 1328 // The only thing we know for sure is that the result is positive. 1329 return {RangeFactory, ValueFactory.getValue(Zero), 1330 ValueFactory.getValue(Max)}; 1331 } 1332 1333 // Nothing much else to do here. 1334 return infer(T); 1335 } 1336 1337 template <> 1338 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_Rem>(Range LHS, 1339 Range RHS, 1340 QualType T) { 1341 llvm::APSInt Zero = ValueFactory.getAPSIntType(T).getZeroValue(); 1342 1343 Range ConservativeRange = getSymmetricalRange(RHS, T); 1344 1345 llvm::APSInt Max = ConservativeRange.To(); 1346 llvm::APSInt Min = ConservativeRange.From(); 1347 1348 if (Max == Zero) { 1349 // It's an undefined behaviour to divide by 0 and it seems like we know 1350 // for sure that RHS is 0. Let's say that the resulting range is 1351 // simply infeasible for that matter. 1352 return RangeFactory.getEmptySet(); 1353 } 1354 1355 // At this point, our conservative range is closed. The result, however, 1356 // couldn't be greater than the RHS' maximal absolute value. Because of 1357 // this reason, we turn the range into open (or half-open in case of 1358 // unsigned integers). 1359 // 1360 // While we operate on integer values, an open interval (a, b) can be easily 1361 // represented by the closed interval [a + 1, b - 1]. And this is exactly 1362 // what we do next. 1363 // 1364 // If we are dealing with unsigned case, we shouldn't move the lower bound. 1365 if (Min.isSigned()) { 1366 ++Min; 1367 } 1368 --Max; 1369 1370 bool IsLHSPositiveOrZero = LHS.From() >= Zero; 1371 bool IsRHSPositiveOrZero = RHS.From() >= Zero; 1372 1373 // Remainder operator results with negative operands is implementation 1374 // defined. Positive cases are much easier to reason about though. 1375 if (IsLHSPositiveOrZero && IsRHSPositiveOrZero) { 1376 // If maximal value of LHS is less than maximal value of RHS, 1377 // the result won't get greater than LHS.To(). 1378 Max = std::min(LHS.To(), Max); 1379 // We want to check if it is a situation similar to the following: 1380 // 1381 // <------------|---[ LHS ]--------[ RHS ]-----> 1382 // -INF 0 +INF 1383 // 1384 // In this situation, we can conclude that (LHS / RHS) == 0 and 1385 // (LHS % RHS) == LHS. 1386 Min = LHS.To() < RHS.From() ? LHS.From() : Zero; 1387 } 1388 1389 // Nevertheless, the symmetrical range for RHS is a conservative estimate 1390 // for any sign of either LHS, or RHS. 1391 return {RangeFactory, ValueFactory.getValue(Min), ValueFactory.getValue(Max)}; 1392 } 1393 1394 //===----------------------------------------------------------------------===// 1395 // Constraint manager implementation details 1396 //===----------------------------------------------------------------------===// 1397 1398 class RangeConstraintManager : public RangedConstraintManager { 1399 public: 1400 RangeConstraintManager(ExprEngine *EE, SValBuilder &SVB) 1401 : RangedConstraintManager(EE, SVB), F(getBasicVals()) {} 1402 1403 //===------------------------------------------------------------------===// 1404 // Implementation for interface from ConstraintManager. 1405 //===------------------------------------------------------------------===// 1406 1407 bool haveEqualConstraints(ProgramStateRef S1, 1408 ProgramStateRef S2) const override { 1409 // NOTE: ClassMembers are as simple as back pointers for ClassMap, 1410 // so comparing constraint ranges and class maps should be 1411 // sufficient. 1412 return S1->get<ConstraintRange>() == S2->get<ConstraintRange>() && 1413 S1->get<ClassMap>() == S2->get<ClassMap>(); 1414 } 1415 1416 bool canReasonAbout(SVal X) const override; 1417 1418 ConditionTruthVal checkNull(ProgramStateRef State, SymbolRef Sym) override; 1419 1420 const llvm::APSInt *getSymVal(ProgramStateRef State, 1421 SymbolRef Sym) const override; 1422 1423 ProgramStateRef removeDeadBindings(ProgramStateRef State, 1424 SymbolReaper &SymReaper) override; 1425 1426 void printJson(raw_ostream &Out, ProgramStateRef State, const char *NL = "\n", 1427 unsigned int Space = 0, bool IsDot = false) const override; 1428 void printConstraints(raw_ostream &Out, ProgramStateRef State, 1429 const char *NL = "\n", unsigned int Space = 0, 1430 bool IsDot = false) const; 1431 void printEquivalenceClasses(raw_ostream &Out, ProgramStateRef State, 1432 const char *NL = "\n", unsigned int Space = 0, 1433 bool IsDot = false) const; 1434 void printDisequalities(raw_ostream &Out, ProgramStateRef State, 1435 const char *NL = "\n", unsigned int Space = 0, 1436 bool IsDot = false) const; 1437 1438 //===------------------------------------------------------------------===// 1439 // Implementation for interface from RangedConstraintManager. 1440 //===------------------------------------------------------------------===// 1441 1442 ProgramStateRef assumeSymNE(ProgramStateRef State, SymbolRef Sym, 1443 const llvm::APSInt &V, 1444 const llvm::APSInt &Adjustment) override; 1445 1446 ProgramStateRef assumeSymEQ(ProgramStateRef State, SymbolRef Sym, 1447 const llvm::APSInt &V, 1448 const llvm::APSInt &Adjustment) override; 1449 1450 ProgramStateRef assumeSymLT(ProgramStateRef State, SymbolRef Sym, 1451 const llvm::APSInt &V, 1452 const llvm::APSInt &Adjustment) override; 1453 1454 ProgramStateRef assumeSymGT(ProgramStateRef State, SymbolRef Sym, 1455 const llvm::APSInt &V, 1456 const llvm::APSInt &Adjustment) override; 1457 1458 ProgramStateRef assumeSymLE(ProgramStateRef State, SymbolRef Sym, 1459 const llvm::APSInt &V, 1460 const llvm::APSInt &Adjustment) override; 1461 1462 ProgramStateRef assumeSymGE(ProgramStateRef State, SymbolRef Sym, 1463 const llvm::APSInt &V, 1464 const llvm::APSInt &Adjustment) override; 1465 1466 ProgramStateRef assumeSymWithinInclusiveRange( 1467 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 1468 const llvm::APSInt &To, const llvm::APSInt &Adjustment) override; 1469 1470 ProgramStateRef assumeSymOutsideInclusiveRange( 1471 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 1472 const llvm::APSInt &To, const llvm::APSInt &Adjustment) override; 1473 1474 private: 1475 RangeSet::Factory F; 1476 1477 RangeSet getRange(ProgramStateRef State, SymbolRef Sym); 1478 RangeSet getRange(ProgramStateRef State, EquivalenceClass Class); 1479 ProgramStateRef setRange(ProgramStateRef State, SymbolRef Sym, 1480 RangeSet Range); 1481 ProgramStateRef setRange(ProgramStateRef State, EquivalenceClass Class, 1482 RangeSet Range); 1483 1484 RangeSet getSymLTRange(ProgramStateRef St, SymbolRef Sym, 1485 const llvm::APSInt &Int, 1486 const llvm::APSInt &Adjustment); 1487 RangeSet getSymGTRange(ProgramStateRef St, SymbolRef Sym, 1488 const llvm::APSInt &Int, 1489 const llvm::APSInt &Adjustment); 1490 RangeSet getSymLERange(ProgramStateRef St, SymbolRef Sym, 1491 const llvm::APSInt &Int, 1492 const llvm::APSInt &Adjustment); 1493 RangeSet getSymLERange(llvm::function_ref<RangeSet()> RS, 1494 const llvm::APSInt &Int, 1495 const llvm::APSInt &Adjustment); 1496 RangeSet getSymGERange(ProgramStateRef St, SymbolRef Sym, 1497 const llvm::APSInt &Int, 1498 const llvm::APSInt &Adjustment); 1499 }; 1500 1501 //===----------------------------------------------------------------------===// 1502 // Constraint assignment logic 1503 //===----------------------------------------------------------------------===// 1504 1505 /// ConstraintAssignorBase is a small utility class that unifies visitor 1506 /// for ranges with a visitor for constraints (rangeset/range/constant). 1507 /// 1508 /// It is designed to have one derived class, but generally it can have more. 1509 /// Derived class can control which types we handle by defining methods of the 1510 /// following form: 1511 /// 1512 /// bool handle${SYMBOL}To${CONSTRAINT}(const SYMBOL *Sym, 1513 /// CONSTRAINT Constraint); 1514 /// 1515 /// where SYMBOL is the type of the symbol (e.g. SymSymExpr, SymbolCast, etc.) 1516 /// CONSTRAINT is the type of constraint (RangeSet/Range/Const) 1517 /// return value signifies whether we should try other handle methods 1518 /// (i.e. false would mean to stop right after calling this method) 1519 template <class Derived> class ConstraintAssignorBase { 1520 public: 1521 using Const = const llvm::APSInt &; 1522 1523 #define DISPATCH(CLASS) return assign##CLASS##Impl(cast<CLASS>(Sym), Constraint) 1524 1525 #define ASSIGN(CLASS, TO, SYM, CONSTRAINT) \ 1526 if (!static_cast<Derived *>(this)->assign##CLASS##To##TO(SYM, CONSTRAINT)) \ 1527 return false 1528 1529 void assign(SymbolRef Sym, RangeSet Constraint) { 1530 assignImpl(Sym, Constraint); 1531 } 1532 1533 bool assignImpl(SymbolRef Sym, RangeSet Constraint) { 1534 switch (Sym->getKind()) { 1535 #define SYMBOL(Id, Parent) \ 1536 case SymExpr::Id##Kind: \ 1537 DISPATCH(Id); 1538 #include "clang/StaticAnalyzer/Core/PathSensitive/Symbols.def" 1539 } 1540 llvm_unreachable("Unknown SymExpr kind!"); 1541 } 1542 1543 #define DEFAULT_ASSIGN(Id) \ 1544 bool assign##Id##To##RangeSet(const Id *Sym, RangeSet Constraint) { \ 1545 return true; \ 1546 } \ 1547 bool assign##Id##To##Range(const Id *Sym, Range Constraint) { return true; } \ 1548 bool assign##Id##To##Const(const Id *Sym, Const Constraint) { return true; } 1549 1550 // When we dispatch for constraint types, we first try to check 1551 // if the new constraint is the constant and try the corresponding 1552 // assignor methods. If it didn't interrupt, we can proceed to the 1553 // range, and finally to the range set. 1554 #define CONSTRAINT_DISPATCH(Id) \ 1555 if (const llvm::APSInt *Const = Constraint.getConcreteValue()) { \ 1556 ASSIGN(Id, Const, Sym, *Const); \ 1557 } \ 1558 if (Constraint.size() == 1) { \ 1559 ASSIGN(Id, Range, Sym, *Constraint.begin()); \ 1560 } \ 1561 ASSIGN(Id, RangeSet, Sym, Constraint) 1562 1563 // Our internal assign method first tries to call assignor methods for all 1564 // constraint types that apply. And if not interrupted, continues with its 1565 // parent class. 1566 #define SYMBOL(Id, Parent) \ 1567 bool assign##Id##Impl(const Id *Sym, RangeSet Constraint) { \ 1568 CONSTRAINT_DISPATCH(Id); \ 1569 DISPATCH(Parent); \ 1570 } \ 1571 DEFAULT_ASSIGN(Id) 1572 #define ABSTRACT_SYMBOL(Id, Parent) SYMBOL(Id, Parent) 1573 #include "clang/StaticAnalyzer/Core/PathSensitive/Symbols.def" 1574 1575 // Default implementations for the top class that doesn't have parents. 1576 bool assignSymExprImpl(const SymExpr *Sym, RangeSet Constraint) { 1577 CONSTRAINT_DISPATCH(SymExpr); 1578 return true; 1579 } 1580 DEFAULT_ASSIGN(SymExpr); 1581 1582 #undef DISPATCH 1583 #undef CONSTRAINT_DISPATCH 1584 #undef DEFAULT_ASSIGN 1585 #undef ASSIGN 1586 }; 1587 1588 /// A little component aggregating all of the reasoning we have about 1589 /// assigning new constraints to symbols. 1590 /// 1591 /// The main purpose of this class is to associate constraints to symbols, 1592 /// and impose additional constraints on other symbols, when we can imply 1593 /// them. 1594 /// 1595 /// It has a nice symmetry with SymbolicRangeInferrer. When the latter 1596 /// can provide more precise ranges by looking into the operands of the 1597 /// expression in question, ConstraintAssignor looks into the operands 1598 /// to see if we can imply more from the new constraint. 1599 class ConstraintAssignor : public ConstraintAssignorBase<ConstraintAssignor> { 1600 public: 1601 template <class ClassOrSymbol> 1602 LLVM_NODISCARD static ProgramStateRef 1603 assign(ProgramStateRef State, RangeConstraintManager &RCM, 1604 SValBuilder &Builder, RangeSet::Factory &F, ClassOrSymbol CoS, 1605 RangeSet NewConstraint) { 1606 if (!State || NewConstraint.isEmpty()) 1607 return nullptr; 1608 1609 ConstraintAssignor Assignor{State, RCM, Builder, F}; 1610 return Assignor.assign(CoS, NewConstraint); 1611 } 1612 1613 /// Handle expressions like: a % b != 0. 1614 template <typename SymT> 1615 bool handleRemainderOp(const SymT *Sym, RangeSet Constraint) { 1616 if (Sym->getOpcode() != BO_Rem) 1617 return true; 1618 const SymbolRef LHS = Sym->getLHS(); 1619 const llvm::APSInt &Zero = 1620 Builder.getBasicValueFactory().getValue(0, Sym->getType()); 1621 // a % b != 0 implies that a != 0. 1622 if (!Constraint.containsZero()) { 1623 State = RCM.assumeSymNE(State, LHS, Zero, Zero); 1624 if (!State) 1625 return false; 1626 } 1627 return true; 1628 } 1629 1630 inline bool assignSymExprToConst(const SymExpr *Sym, Const Constraint); 1631 inline bool assignSymIntExprToRangeSet(const SymIntExpr *Sym, 1632 RangeSet Constraint) { 1633 return handleRemainderOp(Sym, Constraint); 1634 } 1635 inline bool assignSymSymExprToRangeSet(const SymSymExpr *Sym, 1636 RangeSet Constraint); 1637 1638 private: 1639 ConstraintAssignor(ProgramStateRef State, RangeConstraintManager &RCM, 1640 SValBuilder &Builder, RangeSet::Factory &F) 1641 : State(State), RCM(RCM), Builder(Builder), RangeFactory(F) {} 1642 using Base = ConstraintAssignorBase<ConstraintAssignor>; 1643 1644 /// Base method for handling new constraints for symbols. 1645 LLVM_NODISCARD ProgramStateRef assign(SymbolRef Sym, RangeSet NewConstraint) { 1646 // All constraints are actually associated with equivalence classes, and 1647 // that's what we are going to do first. 1648 State = assign(EquivalenceClass::find(State, Sym), NewConstraint); 1649 if (!State) 1650 return nullptr; 1651 1652 // And after that we can check what other things we can get from this 1653 // constraint. 1654 Base::assign(Sym, NewConstraint); 1655 return State; 1656 } 1657 1658 /// Base method for handling new constraints for classes. 1659 LLVM_NODISCARD ProgramStateRef assign(EquivalenceClass Class, 1660 RangeSet NewConstraint) { 1661 // There is a chance that we might need to update constraints for the 1662 // classes that are known to be disequal to Class. 1663 // 1664 // In order for this to be even possible, the new constraint should 1665 // be simply a constant because we can't reason about range disequalities. 1666 if (const llvm::APSInt *Point = NewConstraint.getConcreteValue()) { 1667 1668 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1669 ConstraintRangeTy::Factory &CF = State->get_context<ConstraintRange>(); 1670 1671 // Add new constraint. 1672 Constraints = CF.add(Constraints, Class, NewConstraint); 1673 1674 for (EquivalenceClass DisequalClass : Class.getDisequalClasses(State)) { 1675 RangeSet UpdatedConstraint = SymbolicRangeInferrer::inferRange( 1676 RangeFactory, State, DisequalClass); 1677 1678 UpdatedConstraint = RangeFactory.deletePoint(UpdatedConstraint, *Point); 1679 1680 // If we end up with at least one of the disequal classes to be 1681 // constrained with an empty range-set, the state is infeasible. 1682 if (UpdatedConstraint.isEmpty()) 1683 return nullptr; 1684 1685 Constraints = CF.add(Constraints, DisequalClass, UpdatedConstraint); 1686 } 1687 assert(areFeasible(Constraints) && "Constraint manager shouldn't produce " 1688 "a state with infeasible constraints"); 1689 1690 return setConstraints(State, Constraints); 1691 } 1692 1693 return setConstraint(State, Class, NewConstraint); 1694 } 1695 1696 ProgramStateRef trackDisequality(ProgramStateRef State, SymbolRef LHS, 1697 SymbolRef RHS) { 1698 return EquivalenceClass::markDisequal(RangeFactory, State, LHS, RHS); 1699 } 1700 1701 ProgramStateRef trackEquality(ProgramStateRef State, SymbolRef LHS, 1702 SymbolRef RHS) { 1703 return EquivalenceClass::merge(RangeFactory, State, LHS, RHS); 1704 } 1705 1706 LLVM_NODISCARD Optional<bool> interpreteAsBool(RangeSet Constraint) { 1707 assert(!Constraint.isEmpty() && "Empty ranges shouldn't get here"); 1708 1709 if (Constraint.getConcreteValue()) 1710 return !Constraint.getConcreteValue()->isZero(); 1711 1712 if (!Constraint.containsZero()) 1713 return true; 1714 1715 return llvm::None; 1716 } 1717 1718 ProgramStateRef State; 1719 RangeConstraintManager &RCM; 1720 SValBuilder &Builder; 1721 RangeSet::Factory &RangeFactory; 1722 }; 1723 1724 1725 bool ConstraintAssignor::assignSymExprToConst(const SymExpr *Sym, 1726 const llvm::APSInt &Constraint) { 1727 llvm::SmallSet<EquivalenceClass, 4> SimplifiedClasses; 1728 // Iterate over all equivalence classes and try to simplify them. 1729 ClassMembersTy Members = State->get<ClassMembers>(); 1730 for (std::pair<EquivalenceClass, SymbolSet> ClassToSymbolSet : Members) { 1731 EquivalenceClass Class = ClassToSymbolSet.first; 1732 State = EquivalenceClass::simplify(Builder, RangeFactory, State, Class); 1733 if (!State) 1734 return false; 1735 SimplifiedClasses.insert(Class); 1736 } 1737 1738 // Trivial equivalence classes (those that have only one symbol member) are 1739 // not stored in the State. Thus, we must skim through the constraints as 1740 // well. And we try to simplify symbols in the constraints. 1741 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1742 for (std::pair<EquivalenceClass, RangeSet> ClassConstraint : Constraints) { 1743 EquivalenceClass Class = ClassConstraint.first; 1744 if (SimplifiedClasses.count(Class)) // Already simplified. 1745 continue; 1746 State = EquivalenceClass::simplify(Builder, RangeFactory, State, Class); 1747 if (!State) 1748 return false; 1749 } 1750 1751 return true; 1752 } 1753 1754 bool ConstraintAssignor::assignSymSymExprToRangeSet(const SymSymExpr *Sym, 1755 RangeSet Constraint) { 1756 if (!handleRemainderOp(Sym, Constraint)) 1757 return false; 1758 1759 Optional<bool> ConstraintAsBool = interpreteAsBool(Constraint); 1760 1761 if (!ConstraintAsBool) 1762 return true; 1763 1764 if (Optional<bool> Equality = meansEquality(Sym)) { 1765 // Here we cover two cases: 1766 // * if Sym is equality and the new constraint is true -> Sym's operands 1767 // should be marked as equal 1768 // * if Sym is disequality and the new constraint is false -> Sym's 1769 // operands should be also marked as equal 1770 if (*Equality == *ConstraintAsBool) { 1771 State = trackEquality(State, Sym->getLHS(), Sym->getRHS()); 1772 } else { 1773 // Other combinations leave as with disequal operands. 1774 State = trackDisequality(State, Sym->getLHS(), Sym->getRHS()); 1775 } 1776 1777 if (!State) 1778 return false; 1779 } 1780 1781 return true; 1782 } 1783 1784 } // end anonymous namespace 1785 1786 std::unique_ptr<ConstraintManager> 1787 ento::CreateRangeConstraintManager(ProgramStateManager &StMgr, 1788 ExprEngine *Eng) { 1789 return std::make_unique<RangeConstraintManager>(Eng, StMgr.getSValBuilder()); 1790 } 1791 1792 ConstraintMap ento::getConstraintMap(ProgramStateRef State) { 1793 ConstraintMap::Factory &F = State->get_context<ConstraintMap>(); 1794 ConstraintMap Result = F.getEmptyMap(); 1795 1796 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1797 for (std::pair<EquivalenceClass, RangeSet> ClassConstraint : Constraints) { 1798 EquivalenceClass Class = ClassConstraint.first; 1799 SymbolSet ClassMembers = Class.getClassMembers(State); 1800 assert(!ClassMembers.isEmpty() && 1801 "Class must always have at least one member!"); 1802 1803 SymbolRef Representative = *ClassMembers.begin(); 1804 Result = F.add(Result, Representative, ClassConstraint.second); 1805 } 1806 1807 return Result; 1808 } 1809 1810 //===----------------------------------------------------------------------===// 1811 // EqualityClass implementation details 1812 //===----------------------------------------------------------------------===// 1813 1814 LLVM_DUMP_METHOD void EquivalenceClass::dumpToStream(ProgramStateRef State, 1815 raw_ostream &os) const { 1816 SymbolSet ClassMembers = getClassMembers(State); 1817 for (const SymbolRef &MemberSym : ClassMembers) { 1818 MemberSym->dump(); 1819 os << "\n"; 1820 } 1821 } 1822 1823 inline EquivalenceClass EquivalenceClass::find(ProgramStateRef State, 1824 SymbolRef Sym) { 1825 assert(State && "State should not be null"); 1826 assert(Sym && "Symbol should not be null"); 1827 // We store far from all Symbol -> Class mappings 1828 if (const EquivalenceClass *NontrivialClass = State->get<ClassMap>(Sym)) 1829 return *NontrivialClass; 1830 1831 // This is a trivial class of Sym. 1832 return Sym; 1833 } 1834 1835 inline ProgramStateRef EquivalenceClass::merge(RangeSet::Factory &F, 1836 ProgramStateRef State, 1837 SymbolRef First, 1838 SymbolRef Second) { 1839 EquivalenceClass FirstClass = find(State, First); 1840 EquivalenceClass SecondClass = find(State, Second); 1841 1842 return FirstClass.merge(F, State, SecondClass); 1843 } 1844 1845 inline ProgramStateRef EquivalenceClass::merge(RangeSet::Factory &F, 1846 ProgramStateRef State, 1847 EquivalenceClass Other) { 1848 // It is already the same class. 1849 if (*this == Other) 1850 return State; 1851 1852 // FIXME: As of now, we support only equivalence classes of the same type. 1853 // This limitation is connected to the lack of explicit casts in 1854 // our symbolic expression model. 1855 // 1856 // That means that for `int x` and `char y` we don't distinguish 1857 // between these two very different cases: 1858 // * `x == y` 1859 // * `(char)x == y` 1860 // 1861 // The moment we introduce symbolic casts, this restriction can be 1862 // lifted. 1863 if (getType() != Other.getType()) 1864 return State; 1865 1866 SymbolSet Members = getClassMembers(State); 1867 SymbolSet OtherMembers = Other.getClassMembers(State); 1868 1869 // We estimate the size of the class by the height of tree containing 1870 // its members. Merging is not a trivial operation, so it's easier to 1871 // merge the smaller class into the bigger one. 1872 if (Members.getHeight() >= OtherMembers.getHeight()) { 1873 return mergeImpl(F, State, Members, Other, OtherMembers); 1874 } else { 1875 return Other.mergeImpl(F, State, OtherMembers, *this, Members); 1876 } 1877 } 1878 1879 inline ProgramStateRef 1880 EquivalenceClass::mergeImpl(RangeSet::Factory &RangeFactory, 1881 ProgramStateRef State, SymbolSet MyMembers, 1882 EquivalenceClass Other, SymbolSet OtherMembers) { 1883 // Essentially what we try to recreate here is some kind of union-find 1884 // data structure. It does have certain limitations due to persistence 1885 // and the need to remove elements from classes. 1886 // 1887 // In this setting, EquialityClass object is the representative of the class 1888 // or the parent element. ClassMap is a mapping of class members to their 1889 // parent. Unlike the union-find structure, they all point directly to the 1890 // class representative because we don't have an opportunity to actually do 1891 // path compression when dealing with immutability. This means that we 1892 // compress paths every time we do merges. It also means that we lose 1893 // the main amortized complexity benefit from the original data structure. 1894 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1895 ConstraintRangeTy::Factory &CRF = State->get_context<ConstraintRange>(); 1896 1897 // 1. If the merged classes have any constraints associated with them, we 1898 // need to transfer them to the class we have left. 1899 // 1900 // Intersection here makes perfect sense because both of these constraints 1901 // must hold for the whole new class. 1902 if (Optional<RangeSet> NewClassConstraint = 1903 intersect(RangeFactory, getConstraint(State, *this), 1904 getConstraint(State, Other))) { 1905 // NOTE: Essentially, NewClassConstraint should NEVER be infeasible because 1906 // range inferrer shouldn't generate ranges incompatible with 1907 // equivalence classes. However, at the moment, due to imperfections 1908 // in the solver, it is possible and the merge function can also 1909 // return infeasible states aka null states. 1910 if (NewClassConstraint->isEmpty()) 1911 // Infeasible state 1912 return nullptr; 1913 1914 // No need in tracking constraints of a now-dissolved class. 1915 Constraints = CRF.remove(Constraints, Other); 1916 // Assign new constraints for this class. 1917 Constraints = CRF.add(Constraints, *this, *NewClassConstraint); 1918 1919 assert(areFeasible(Constraints) && "Constraint manager shouldn't produce " 1920 "a state with infeasible constraints"); 1921 1922 State = State->set<ConstraintRange>(Constraints); 1923 } 1924 1925 // 2. Get ALL equivalence-related maps 1926 ClassMapTy Classes = State->get<ClassMap>(); 1927 ClassMapTy::Factory &CMF = State->get_context<ClassMap>(); 1928 1929 ClassMembersTy Members = State->get<ClassMembers>(); 1930 ClassMembersTy::Factory &MF = State->get_context<ClassMembers>(); 1931 1932 DisequalityMapTy DisequalityInfo = State->get<DisequalityMap>(); 1933 DisequalityMapTy::Factory &DF = State->get_context<DisequalityMap>(); 1934 1935 ClassSet::Factory &CF = State->get_context<ClassSet>(); 1936 SymbolSet::Factory &F = getMembersFactory(State); 1937 1938 // 2. Merge members of the Other class into the current class. 1939 SymbolSet NewClassMembers = MyMembers; 1940 for (SymbolRef Sym : OtherMembers) { 1941 NewClassMembers = F.add(NewClassMembers, Sym); 1942 // *this is now the class for all these new symbols. 1943 Classes = CMF.add(Classes, Sym, *this); 1944 } 1945 1946 // 3. Adjust member mapping. 1947 // 1948 // No need in tracking members of a now-dissolved class. 1949 Members = MF.remove(Members, Other); 1950 // Now only the current class is mapped to all the symbols. 1951 Members = MF.add(Members, *this, NewClassMembers); 1952 1953 // 4. Update disequality relations 1954 ClassSet DisequalToOther = Other.getDisequalClasses(DisequalityInfo, CF); 1955 // We are about to merge two classes but they are already known to be 1956 // non-equal. This is a contradiction. 1957 if (DisequalToOther.contains(*this)) 1958 return nullptr; 1959 1960 if (!DisequalToOther.isEmpty()) { 1961 ClassSet DisequalToThis = getDisequalClasses(DisequalityInfo, CF); 1962 DisequalityInfo = DF.remove(DisequalityInfo, Other); 1963 1964 for (EquivalenceClass DisequalClass : DisequalToOther) { 1965 DisequalToThis = CF.add(DisequalToThis, DisequalClass); 1966 1967 // Disequality is a symmetric relation meaning that if 1968 // DisequalToOther not null then the set for DisequalClass is not 1969 // empty and has at least Other. 1970 ClassSet OriginalSetLinkedToOther = 1971 *DisequalityInfo.lookup(DisequalClass); 1972 1973 // Other will be eliminated and we should replace it with the bigger 1974 // united class. 1975 ClassSet NewSet = CF.remove(OriginalSetLinkedToOther, Other); 1976 NewSet = CF.add(NewSet, *this); 1977 1978 DisequalityInfo = DF.add(DisequalityInfo, DisequalClass, NewSet); 1979 } 1980 1981 DisequalityInfo = DF.add(DisequalityInfo, *this, DisequalToThis); 1982 State = State->set<DisequalityMap>(DisequalityInfo); 1983 } 1984 1985 // 5. Update the state 1986 State = State->set<ClassMap>(Classes); 1987 State = State->set<ClassMembers>(Members); 1988 1989 return State; 1990 } 1991 1992 inline SymbolSet::Factory & 1993 EquivalenceClass::getMembersFactory(ProgramStateRef State) { 1994 return State->get_context<SymbolSet>(); 1995 } 1996 1997 SymbolSet EquivalenceClass::getClassMembers(ProgramStateRef State) const { 1998 if (const SymbolSet *Members = State->get<ClassMembers>(*this)) 1999 return *Members; 2000 2001 // This class is trivial, so we need to construct a set 2002 // with just that one symbol from the class. 2003 SymbolSet::Factory &F = getMembersFactory(State); 2004 return F.add(F.getEmptySet(), getRepresentativeSymbol()); 2005 } 2006 2007 bool EquivalenceClass::isTrivial(ProgramStateRef State) const { 2008 return State->get<ClassMembers>(*this) == nullptr; 2009 } 2010 2011 bool EquivalenceClass::isTriviallyDead(ProgramStateRef State, 2012 SymbolReaper &Reaper) const { 2013 return isTrivial(State) && Reaper.isDead(getRepresentativeSymbol()); 2014 } 2015 2016 inline ProgramStateRef EquivalenceClass::markDisequal(RangeSet::Factory &RF, 2017 ProgramStateRef State, 2018 SymbolRef First, 2019 SymbolRef Second) { 2020 return markDisequal(RF, State, find(State, First), find(State, Second)); 2021 } 2022 2023 inline ProgramStateRef EquivalenceClass::markDisequal(RangeSet::Factory &RF, 2024 ProgramStateRef State, 2025 EquivalenceClass First, 2026 EquivalenceClass Second) { 2027 return First.markDisequal(RF, State, Second); 2028 } 2029 2030 inline ProgramStateRef 2031 EquivalenceClass::markDisequal(RangeSet::Factory &RF, ProgramStateRef State, 2032 EquivalenceClass Other) const { 2033 // If we know that two classes are equal, we can only produce an infeasible 2034 // state. 2035 if (*this == Other) { 2036 return nullptr; 2037 } 2038 2039 DisequalityMapTy DisequalityInfo = State->get<DisequalityMap>(); 2040 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 2041 2042 // Disequality is a symmetric relation, so if we mark A as disequal to B, 2043 // we should also mark B as disequalt to A. 2044 if (!addToDisequalityInfo(DisequalityInfo, Constraints, RF, State, *this, 2045 Other) || 2046 !addToDisequalityInfo(DisequalityInfo, Constraints, RF, State, Other, 2047 *this)) 2048 return nullptr; 2049 2050 assert(areFeasible(Constraints) && "Constraint manager shouldn't produce " 2051 "a state with infeasible constraints"); 2052 2053 State = State->set<DisequalityMap>(DisequalityInfo); 2054 State = State->set<ConstraintRange>(Constraints); 2055 2056 return State; 2057 } 2058 2059 inline bool EquivalenceClass::addToDisequalityInfo( 2060 DisequalityMapTy &Info, ConstraintRangeTy &Constraints, 2061 RangeSet::Factory &RF, ProgramStateRef State, EquivalenceClass First, 2062 EquivalenceClass Second) { 2063 2064 // 1. Get all of the required factories. 2065 DisequalityMapTy::Factory &F = State->get_context<DisequalityMap>(); 2066 ClassSet::Factory &CF = State->get_context<ClassSet>(); 2067 ConstraintRangeTy::Factory &CRF = State->get_context<ConstraintRange>(); 2068 2069 // 2. Add Second to the set of classes disequal to First. 2070 const ClassSet *CurrentSet = Info.lookup(First); 2071 ClassSet NewSet = CurrentSet ? *CurrentSet : CF.getEmptySet(); 2072 NewSet = CF.add(NewSet, Second); 2073 2074 Info = F.add(Info, First, NewSet); 2075 2076 // 3. If Second is known to be a constant, we can delete this point 2077 // from the constraint asociated with First. 2078 // 2079 // So, if Second == 10, it means that First != 10. 2080 // At the same time, the same logic does not apply to ranges. 2081 if (const RangeSet *SecondConstraint = Constraints.lookup(Second)) 2082 if (const llvm::APSInt *Point = SecondConstraint->getConcreteValue()) { 2083 2084 RangeSet FirstConstraint = SymbolicRangeInferrer::inferRange( 2085 RF, State, First.getRepresentativeSymbol()); 2086 2087 FirstConstraint = RF.deletePoint(FirstConstraint, *Point); 2088 2089 // If the First class is about to be constrained with an empty 2090 // range-set, the state is infeasible. 2091 if (FirstConstraint.isEmpty()) 2092 return false; 2093 2094 Constraints = CRF.add(Constraints, First, FirstConstraint); 2095 } 2096 2097 return true; 2098 } 2099 2100 inline Optional<bool> EquivalenceClass::areEqual(ProgramStateRef State, 2101 SymbolRef FirstSym, 2102 SymbolRef SecondSym) { 2103 return EquivalenceClass::areEqual(State, find(State, FirstSym), 2104 find(State, SecondSym)); 2105 } 2106 2107 inline Optional<bool> EquivalenceClass::areEqual(ProgramStateRef State, 2108 EquivalenceClass First, 2109 EquivalenceClass Second) { 2110 // The same equivalence class => symbols are equal. 2111 if (First == Second) 2112 return true; 2113 2114 // Let's check if we know anything about these two classes being not equal to 2115 // each other. 2116 ClassSet DisequalToFirst = First.getDisequalClasses(State); 2117 if (DisequalToFirst.contains(Second)) 2118 return false; 2119 2120 // It is not clear. 2121 return llvm::None; 2122 } 2123 2124 // Iterate over all symbols and try to simplify them. Once a symbol is 2125 // simplified then we check if we can merge the simplified symbol's equivalence 2126 // class to this class. This way, we simplify not just the symbols but the 2127 // classes as well: we strive to keep the number of the classes to be the 2128 // absolute minimum. 2129 LLVM_NODISCARD ProgramStateRef 2130 EquivalenceClass::simplify(SValBuilder &SVB, RangeSet::Factory &F, 2131 ProgramStateRef State, EquivalenceClass Class) { 2132 SymbolSet ClassMembers = Class.getClassMembers(State); 2133 for (const SymbolRef &MemberSym : ClassMembers) { 2134 2135 const SVal SimplifiedMemberVal = simplifyToSVal(State, MemberSym); 2136 const SymbolRef SimplifiedMemberSym = SimplifiedMemberVal.getAsSymbol(); 2137 2138 // The symbol is collapsed to a constant, check if the current State is 2139 // still feasible. 2140 if (const auto CI = SimplifiedMemberVal.getAs<nonloc::ConcreteInt>()) { 2141 const llvm::APSInt &SV = CI->getValue(); 2142 const RangeSet *ClassConstraint = getConstraint(State, Class); 2143 // We have found a contradiction. 2144 if (ClassConstraint && !ClassConstraint->contains(SV)) 2145 return nullptr; 2146 } 2147 2148 if (SimplifiedMemberSym && MemberSym != SimplifiedMemberSym) { 2149 // The simplified symbol should be the member of the original Class, 2150 // however, it might be in another existing class at the moment. We 2151 // have to merge these classes. 2152 State = merge(F, State, MemberSym, SimplifiedMemberSym); 2153 if (!State) 2154 return nullptr; 2155 } 2156 } 2157 return State; 2158 } 2159 2160 inline ClassSet EquivalenceClass::getDisequalClasses(ProgramStateRef State, 2161 SymbolRef Sym) { 2162 return find(State, Sym).getDisequalClasses(State); 2163 } 2164 2165 inline ClassSet 2166 EquivalenceClass::getDisequalClasses(ProgramStateRef State) const { 2167 return getDisequalClasses(State->get<DisequalityMap>(), 2168 State->get_context<ClassSet>()); 2169 } 2170 2171 inline ClassSet 2172 EquivalenceClass::getDisequalClasses(DisequalityMapTy Map, 2173 ClassSet::Factory &Factory) const { 2174 if (const ClassSet *DisequalClasses = Map.lookup(*this)) 2175 return *DisequalClasses; 2176 2177 return Factory.getEmptySet(); 2178 } 2179 2180 bool EquivalenceClass::isClassDataConsistent(ProgramStateRef State) { 2181 ClassMembersTy Members = State->get<ClassMembers>(); 2182 2183 for (std::pair<EquivalenceClass, SymbolSet> ClassMembersPair : Members) { 2184 for (SymbolRef Member : ClassMembersPair.second) { 2185 // Every member of the class should have a mapping back to the class. 2186 if (find(State, Member) == ClassMembersPair.first) { 2187 continue; 2188 } 2189 2190 return false; 2191 } 2192 } 2193 2194 DisequalityMapTy Disequalities = State->get<DisequalityMap>(); 2195 for (std::pair<EquivalenceClass, ClassSet> DisequalityInfo : Disequalities) { 2196 EquivalenceClass Class = DisequalityInfo.first; 2197 ClassSet DisequalClasses = DisequalityInfo.second; 2198 2199 // There is no use in keeping empty sets in the map. 2200 if (DisequalClasses.isEmpty()) 2201 return false; 2202 2203 // Disequality is symmetrical, i.e. for every Class A and B that A != B, 2204 // B != A should also be true. 2205 for (EquivalenceClass DisequalClass : DisequalClasses) { 2206 const ClassSet *DisequalToDisequalClasses = 2207 Disequalities.lookup(DisequalClass); 2208 2209 // It should be a set of at least one element: Class 2210 if (!DisequalToDisequalClasses || 2211 !DisequalToDisequalClasses->contains(Class)) 2212 return false; 2213 } 2214 } 2215 2216 return true; 2217 } 2218 2219 //===----------------------------------------------------------------------===// 2220 // RangeConstraintManager implementation 2221 //===----------------------------------------------------------------------===// 2222 2223 bool RangeConstraintManager::canReasonAbout(SVal X) const { 2224 Optional<nonloc::SymbolVal> SymVal = X.getAs<nonloc::SymbolVal>(); 2225 if (SymVal && SymVal->isExpression()) { 2226 const SymExpr *SE = SymVal->getSymbol(); 2227 2228 if (const SymIntExpr *SIE = dyn_cast<SymIntExpr>(SE)) { 2229 switch (SIE->getOpcode()) { 2230 // We don't reason yet about bitwise-constraints on symbolic values. 2231 case BO_And: 2232 case BO_Or: 2233 case BO_Xor: 2234 return false; 2235 // We don't reason yet about these arithmetic constraints on 2236 // symbolic values. 2237 case BO_Mul: 2238 case BO_Div: 2239 case BO_Rem: 2240 case BO_Shl: 2241 case BO_Shr: 2242 return false; 2243 // All other cases. 2244 default: 2245 return true; 2246 } 2247 } 2248 2249 if (const SymSymExpr *SSE = dyn_cast<SymSymExpr>(SE)) { 2250 // FIXME: Handle <=> here. 2251 if (BinaryOperator::isEqualityOp(SSE->getOpcode()) || 2252 BinaryOperator::isRelationalOp(SSE->getOpcode())) { 2253 // We handle Loc <> Loc comparisons, but not (yet) NonLoc <> NonLoc. 2254 // We've recently started producing Loc <> NonLoc comparisons (that 2255 // result from casts of one of the operands between eg. intptr_t and 2256 // void *), but we can't reason about them yet. 2257 if (Loc::isLocType(SSE->getLHS()->getType())) { 2258 return Loc::isLocType(SSE->getRHS()->getType()); 2259 } 2260 } 2261 } 2262 2263 return false; 2264 } 2265 2266 return true; 2267 } 2268 2269 ConditionTruthVal RangeConstraintManager::checkNull(ProgramStateRef State, 2270 SymbolRef Sym) { 2271 const RangeSet *Ranges = getConstraint(State, Sym); 2272 2273 // If we don't have any information about this symbol, it's underconstrained. 2274 if (!Ranges) 2275 return ConditionTruthVal(); 2276 2277 // If we have a concrete value, see if it's zero. 2278 if (const llvm::APSInt *Value = Ranges->getConcreteValue()) 2279 return *Value == 0; 2280 2281 BasicValueFactory &BV = getBasicVals(); 2282 APSIntType IntType = BV.getAPSIntType(Sym->getType()); 2283 llvm::APSInt Zero = IntType.getZeroValue(); 2284 2285 // Check if zero is in the set of possible values. 2286 if (!Ranges->contains(Zero)) 2287 return false; 2288 2289 // Zero is a possible value, but it is not the /only/ possible value. 2290 return ConditionTruthVal(); 2291 } 2292 2293 const llvm::APSInt *RangeConstraintManager::getSymVal(ProgramStateRef St, 2294 SymbolRef Sym) const { 2295 const RangeSet *T = getConstraint(St, Sym); 2296 return T ? T->getConcreteValue() : nullptr; 2297 } 2298 2299 //===----------------------------------------------------------------------===// 2300 // Remove dead symbols from existing constraints 2301 //===----------------------------------------------------------------------===// 2302 2303 /// Scan all symbols referenced by the constraints. If the symbol is not alive 2304 /// as marked in LSymbols, mark it as dead in DSymbols. 2305 ProgramStateRef 2306 RangeConstraintManager::removeDeadBindings(ProgramStateRef State, 2307 SymbolReaper &SymReaper) { 2308 ClassMembersTy ClassMembersMap = State->get<ClassMembers>(); 2309 ClassMembersTy NewClassMembersMap = ClassMembersMap; 2310 ClassMembersTy::Factory &EMFactory = State->get_context<ClassMembers>(); 2311 SymbolSet::Factory &SetFactory = State->get_context<SymbolSet>(); 2312 2313 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 2314 ConstraintRangeTy NewConstraints = Constraints; 2315 ConstraintRangeTy::Factory &ConstraintFactory = 2316 State->get_context<ConstraintRange>(); 2317 2318 ClassMapTy Map = State->get<ClassMap>(); 2319 ClassMapTy NewMap = Map; 2320 ClassMapTy::Factory &ClassFactory = State->get_context<ClassMap>(); 2321 2322 DisequalityMapTy Disequalities = State->get<DisequalityMap>(); 2323 DisequalityMapTy::Factory &DisequalityFactory = 2324 State->get_context<DisequalityMap>(); 2325 ClassSet::Factory &ClassSetFactory = State->get_context<ClassSet>(); 2326 2327 bool ClassMapChanged = false; 2328 bool MembersMapChanged = false; 2329 bool ConstraintMapChanged = false; 2330 bool DisequalitiesChanged = false; 2331 2332 auto removeDeadClass = [&](EquivalenceClass Class) { 2333 // Remove associated constraint ranges. 2334 Constraints = ConstraintFactory.remove(Constraints, Class); 2335 ConstraintMapChanged = true; 2336 2337 // Update disequality information to not hold any information on the 2338 // removed class. 2339 ClassSet DisequalClasses = 2340 Class.getDisequalClasses(Disequalities, ClassSetFactory); 2341 if (!DisequalClasses.isEmpty()) { 2342 for (EquivalenceClass DisequalClass : DisequalClasses) { 2343 ClassSet DisequalToDisequalSet = 2344 DisequalClass.getDisequalClasses(Disequalities, ClassSetFactory); 2345 // DisequalToDisequalSet is guaranteed to be non-empty for consistent 2346 // disequality info. 2347 assert(!DisequalToDisequalSet.isEmpty()); 2348 ClassSet NewSet = ClassSetFactory.remove(DisequalToDisequalSet, Class); 2349 2350 // No need in keeping an empty set. 2351 if (NewSet.isEmpty()) { 2352 Disequalities = 2353 DisequalityFactory.remove(Disequalities, DisequalClass); 2354 } else { 2355 Disequalities = 2356 DisequalityFactory.add(Disequalities, DisequalClass, NewSet); 2357 } 2358 } 2359 // Remove the data for the class 2360 Disequalities = DisequalityFactory.remove(Disequalities, Class); 2361 DisequalitiesChanged = true; 2362 } 2363 }; 2364 2365 // 1. Let's see if dead symbols are trivial and have associated constraints. 2366 for (std::pair<EquivalenceClass, RangeSet> ClassConstraintPair : 2367 Constraints) { 2368 EquivalenceClass Class = ClassConstraintPair.first; 2369 if (Class.isTriviallyDead(State, SymReaper)) { 2370 // If this class is trivial, we can remove its constraints right away. 2371 removeDeadClass(Class); 2372 } 2373 } 2374 2375 // 2. We don't need to track classes for dead symbols. 2376 for (std::pair<SymbolRef, EquivalenceClass> SymbolClassPair : Map) { 2377 SymbolRef Sym = SymbolClassPair.first; 2378 2379 if (SymReaper.isDead(Sym)) { 2380 ClassMapChanged = true; 2381 NewMap = ClassFactory.remove(NewMap, Sym); 2382 } 2383 } 2384 2385 // 3. Remove dead members from classes and remove dead non-trivial classes 2386 // and their constraints. 2387 for (std::pair<EquivalenceClass, SymbolSet> ClassMembersPair : 2388 ClassMembersMap) { 2389 EquivalenceClass Class = ClassMembersPair.first; 2390 SymbolSet LiveMembers = ClassMembersPair.second; 2391 bool MembersChanged = false; 2392 2393 for (SymbolRef Member : ClassMembersPair.second) { 2394 if (SymReaper.isDead(Member)) { 2395 MembersChanged = true; 2396 LiveMembers = SetFactory.remove(LiveMembers, Member); 2397 } 2398 } 2399 2400 // Check if the class changed. 2401 if (!MembersChanged) 2402 continue; 2403 2404 MembersMapChanged = true; 2405 2406 if (LiveMembers.isEmpty()) { 2407 // The class is dead now, we need to wipe it out of the members map... 2408 NewClassMembersMap = EMFactory.remove(NewClassMembersMap, Class); 2409 2410 // ...and remove all of its constraints. 2411 removeDeadClass(Class); 2412 } else { 2413 // We need to change the members associated with the class. 2414 NewClassMembersMap = 2415 EMFactory.add(NewClassMembersMap, Class, LiveMembers); 2416 } 2417 } 2418 2419 // 4. Update the state with new maps. 2420 // 2421 // Here we try to be humble and update a map only if it really changed. 2422 if (ClassMapChanged) 2423 State = State->set<ClassMap>(NewMap); 2424 2425 if (MembersMapChanged) 2426 State = State->set<ClassMembers>(NewClassMembersMap); 2427 2428 if (ConstraintMapChanged) 2429 State = State->set<ConstraintRange>(Constraints); 2430 2431 if (DisequalitiesChanged) 2432 State = State->set<DisequalityMap>(Disequalities); 2433 2434 assert(EquivalenceClass::isClassDataConsistent(State)); 2435 2436 return State; 2437 } 2438 2439 RangeSet RangeConstraintManager::getRange(ProgramStateRef State, 2440 SymbolRef Sym) { 2441 return SymbolicRangeInferrer::inferRange(F, State, Sym); 2442 } 2443 2444 ProgramStateRef RangeConstraintManager::setRange(ProgramStateRef State, 2445 SymbolRef Sym, 2446 RangeSet Range) { 2447 return ConstraintAssignor::assign(State, *this, getSValBuilder(), F, Sym, 2448 Range); 2449 } 2450 2451 //===------------------------------------------------------------------------=== 2452 // assumeSymX methods: protected interface for RangeConstraintManager. 2453 //===------------------------------------------------------------------------===/ 2454 2455 // The syntax for ranges below is mathematical, using [x, y] for closed ranges 2456 // and (x, y) for open ranges. These ranges are modular, corresponding with 2457 // a common treatment of C integer overflow. This means that these methods 2458 // do not have to worry about overflow; RangeSet::Intersect can handle such a 2459 // "wraparound" range. 2460 // As an example, the range [UINT_MAX-1, 3) contains five values: UINT_MAX-1, 2461 // UINT_MAX, 0, 1, and 2. 2462 2463 ProgramStateRef 2464 RangeConstraintManager::assumeSymNE(ProgramStateRef St, SymbolRef Sym, 2465 const llvm::APSInt &Int, 2466 const llvm::APSInt &Adjustment) { 2467 // Before we do any real work, see if the value can even show up. 2468 APSIntType AdjustmentType(Adjustment); 2469 if (AdjustmentType.testInRange(Int, true) != APSIntType::RTR_Within) 2470 return St; 2471 2472 llvm::APSInt Point = AdjustmentType.convert(Int) - Adjustment; 2473 RangeSet New = getRange(St, Sym); 2474 New = F.deletePoint(New, Point); 2475 2476 return setRange(St, Sym, New); 2477 } 2478 2479 ProgramStateRef 2480 RangeConstraintManager::assumeSymEQ(ProgramStateRef St, SymbolRef Sym, 2481 const llvm::APSInt &Int, 2482 const llvm::APSInt &Adjustment) { 2483 // Before we do any real work, see if the value can even show up. 2484 APSIntType AdjustmentType(Adjustment); 2485 if (AdjustmentType.testInRange(Int, true) != APSIntType::RTR_Within) 2486 return nullptr; 2487 2488 // [Int-Adjustment, Int-Adjustment] 2489 llvm::APSInt AdjInt = AdjustmentType.convert(Int) - Adjustment; 2490 RangeSet New = getRange(St, Sym); 2491 New = F.intersect(New, AdjInt); 2492 2493 return setRange(St, Sym, New); 2494 } 2495 2496 RangeSet RangeConstraintManager::getSymLTRange(ProgramStateRef St, 2497 SymbolRef Sym, 2498 const llvm::APSInt &Int, 2499 const llvm::APSInt &Adjustment) { 2500 // Before we do any real work, see if the value can even show up. 2501 APSIntType AdjustmentType(Adjustment); 2502 switch (AdjustmentType.testInRange(Int, true)) { 2503 case APSIntType::RTR_Below: 2504 return F.getEmptySet(); 2505 case APSIntType::RTR_Within: 2506 break; 2507 case APSIntType::RTR_Above: 2508 return getRange(St, Sym); 2509 } 2510 2511 // Special case for Int == Min. This is always false. 2512 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2513 llvm::APSInt Min = AdjustmentType.getMinValue(); 2514 if (ComparisonVal == Min) 2515 return F.getEmptySet(); 2516 2517 llvm::APSInt Lower = Min - Adjustment; 2518 llvm::APSInt Upper = ComparisonVal - Adjustment; 2519 --Upper; 2520 2521 RangeSet Result = getRange(St, Sym); 2522 return F.intersect(Result, Lower, Upper); 2523 } 2524 2525 ProgramStateRef 2526 RangeConstraintManager::assumeSymLT(ProgramStateRef St, SymbolRef Sym, 2527 const llvm::APSInt &Int, 2528 const llvm::APSInt &Adjustment) { 2529 RangeSet New = getSymLTRange(St, Sym, Int, Adjustment); 2530 return setRange(St, Sym, New); 2531 } 2532 2533 RangeSet RangeConstraintManager::getSymGTRange(ProgramStateRef St, 2534 SymbolRef Sym, 2535 const llvm::APSInt &Int, 2536 const llvm::APSInt &Adjustment) { 2537 // Before we do any real work, see if the value can even show up. 2538 APSIntType AdjustmentType(Adjustment); 2539 switch (AdjustmentType.testInRange(Int, true)) { 2540 case APSIntType::RTR_Below: 2541 return getRange(St, Sym); 2542 case APSIntType::RTR_Within: 2543 break; 2544 case APSIntType::RTR_Above: 2545 return F.getEmptySet(); 2546 } 2547 2548 // Special case for Int == Max. This is always false. 2549 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2550 llvm::APSInt Max = AdjustmentType.getMaxValue(); 2551 if (ComparisonVal == Max) 2552 return F.getEmptySet(); 2553 2554 llvm::APSInt Lower = ComparisonVal - Adjustment; 2555 llvm::APSInt Upper = Max - Adjustment; 2556 ++Lower; 2557 2558 RangeSet SymRange = getRange(St, Sym); 2559 return F.intersect(SymRange, Lower, Upper); 2560 } 2561 2562 ProgramStateRef 2563 RangeConstraintManager::assumeSymGT(ProgramStateRef St, SymbolRef Sym, 2564 const llvm::APSInt &Int, 2565 const llvm::APSInt &Adjustment) { 2566 RangeSet New = getSymGTRange(St, Sym, Int, Adjustment); 2567 return setRange(St, Sym, New); 2568 } 2569 2570 RangeSet RangeConstraintManager::getSymGERange(ProgramStateRef St, 2571 SymbolRef Sym, 2572 const llvm::APSInt &Int, 2573 const llvm::APSInt &Adjustment) { 2574 // Before we do any real work, see if the value can even show up. 2575 APSIntType AdjustmentType(Adjustment); 2576 switch (AdjustmentType.testInRange(Int, true)) { 2577 case APSIntType::RTR_Below: 2578 return getRange(St, Sym); 2579 case APSIntType::RTR_Within: 2580 break; 2581 case APSIntType::RTR_Above: 2582 return F.getEmptySet(); 2583 } 2584 2585 // Special case for Int == Min. This is always feasible. 2586 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2587 llvm::APSInt Min = AdjustmentType.getMinValue(); 2588 if (ComparisonVal == Min) 2589 return getRange(St, Sym); 2590 2591 llvm::APSInt Max = AdjustmentType.getMaxValue(); 2592 llvm::APSInt Lower = ComparisonVal - Adjustment; 2593 llvm::APSInt Upper = Max - Adjustment; 2594 2595 RangeSet SymRange = getRange(St, Sym); 2596 return F.intersect(SymRange, Lower, Upper); 2597 } 2598 2599 ProgramStateRef 2600 RangeConstraintManager::assumeSymGE(ProgramStateRef St, SymbolRef Sym, 2601 const llvm::APSInt &Int, 2602 const llvm::APSInt &Adjustment) { 2603 RangeSet New = getSymGERange(St, Sym, Int, Adjustment); 2604 return setRange(St, Sym, New); 2605 } 2606 2607 RangeSet 2608 RangeConstraintManager::getSymLERange(llvm::function_ref<RangeSet()> RS, 2609 const llvm::APSInt &Int, 2610 const llvm::APSInt &Adjustment) { 2611 // Before we do any real work, see if the value can even show up. 2612 APSIntType AdjustmentType(Adjustment); 2613 switch (AdjustmentType.testInRange(Int, true)) { 2614 case APSIntType::RTR_Below: 2615 return F.getEmptySet(); 2616 case APSIntType::RTR_Within: 2617 break; 2618 case APSIntType::RTR_Above: 2619 return RS(); 2620 } 2621 2622 // Special case for Int == Max. This is always feasible. 2623 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2624 llvm::APSInt Max = AdjustmentType.getMaxValue(); 2625 if (ComparisonVal == Max) 2626 return RS(); 2627 2628 llvm::APSInt Min = AdjustmentType.getMinValue(); 2629 llvm::APSInt Lower = Min - Adjustment; 2630 llvm::APSInt Upper = ComparisonVal - Adjustment; 2631 2632 RangeSet Default = RS(); 2633 return F.intersect(Default, Lower, Upper); 2634 } 2635 2636 RangeSet RangeConstraintManager::getSymLERange(ProgramStateRef St, 2637 SymbolRef Sym, 2638 const llvm::APSInt &Int, 2639 const llvm::APSInt &Adjustment) { 2640 return getSymLERange([&] { return getRange(St, Sym); }, Int, Adjustment); 2641 } 2642 2643 ProgramStateRef 2644 RangeConstraintManager::assumeSymLE(ProgramStateRef St, SymbolRef Sym, 2645 const llvm::APSInt &Int, 2646 const llvm::APSInt &Adjustment) { 2647 RangeSet New = getSymLERange(St, Sym, Int, Adjustment); 2648 return setRange(St, Sym, New); 2649 } 2650 2651 ProgramStateRef RangeConstraintManager::assumeSymWithinInclusiveRange( 2652 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 2653 const llvm::APSInt &To, const llvm::APSInt &Adjustment) { 2654 RangeSet New = getSymGERange(State, Sym, From, Adjustment); 2655 if (New.isEmpty()) 2656 return nullptr; 2657 RangeSet Out = getSymLERange([&] { return New; }, To, Adjustment); 2658 return setRange(State, Sym, Out); 2659 } 2660 2661 ProgramStateRef RangeConstraintManager::assumeSymOutsideInclusiveRange( 2662 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 2663 const llvm::APSInt &To, const llvm::APSInt &Adjustment) { 2664 RangeSet RangeLT = getSymLTRange(State, Sym, From, Adjustment); 2665 RangeSet RangeGT = getSymGTRange(State, Sym, To, Adjustment); 2666 RangeSet New(F.add(RangeLT, RangeGT)); 2667 return setRange(State, Sym, New); 2668 } 2669 2670 //===----------------------------------------------------------------------===// 2671 // Pretty-printing. 2672 //===----------------------------------------------------------------------===// 2673 2674 void RangeConstraintManager::printJson(raw_ostream &Out, ProgramStateRef State, 2675 const char *NL, unsigned int Space, 2676 bool IsDot) const { 2677 printConstraints(Out, State, NL, Space, IsDot); 2678 printEquivalenceClasses(Out, State, NL, Space, IsDot); 2679 printDisequalities(Out, State, NL, Space, IsDot); 2680 } 2681 2682 static std::string toString(const SymbolRef &Sym) { 2683 std::string S; 2684 llvm::raw_string_ostream O(S); 2685 Sym->dumpToStream(O); 2686 return O.str(); 2687 } 2688 2689 void RangeConstraintManager::printConstraints(raw_ostream &Out, 2690 ProgramStateRef State, 2691 const char *NL, 2692 unsigned int Space, 2693 bool IsDot) const { 2694 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 2695 2696 Indent(Out, Space, IsDot) << "\"constraints\": "; 2697 if (Constraints.isEmpty()) { 2698 Out << "null," << NL; 2699 return; 2700 } 2701 2702 std::map<std::string, RangeSet> OrderedConstraints; 2703 for (std::pair<EquivalenceClass, RangeSet> P : Constraints) { 2704 SymbolSet ClassMembers = P.first.getClassMembers(State); 2705 for (const SymbolRef &ClassMember : ClassMembers) { 2706 bool insertion_took_place; 2707 std::tie(std::ignore, insertion_took_place) = 2708 OrderedConstraints.insert({toString(ClassMember), P.second}); 2709 assert(insertion_took_place && 2710 "two symbols should not have the same dump"); 2711 } 2712 } 2713 2714 ++Space; 2715 Out << '[' << NL; 2716 bool First = true; 2717 for (std::pair<std::string, RangeSet> P : OrderedConstraints) { 2718 if (First) { 2719 First = false; 2720 } else { 2721 Out << ','; 2722 Out << NL; 2723 } 2724 Indent(Out, Space, IsDot) 2725 << "{ \"symbol\": \"" << P.first << "\", \"range\": \""; 2726 P.second.dump(Out); 2727 Out << "\" }"; 2728 } 2729 Out << NL; 2730 2731 --Space; 2732 Indent(Out, Space, IsDot) << "]," << NL; 2733 } 2734 2735 static std::string toString(ProgramStateRef State, EquivalenceClass Class) { 2736 SymbolSet ClassMembers = Class.getClassMembers(State); 2737 llvm::SmallVector<SymbolRef, 8> ClassMembersSorted(ClassMembers.begin(), 2738 ClassMembers.end()); 2739 llvm::sort(ClassMembersSorted, 2740 [](const SymbolRef &LHS, const SymbolRef &RHS) { 2741 return toString(LHS) < toString(RHS); 2742 }); 2743 2744 bool FirstMember = true; 2745 2746 std::string Str; 2747 llvm::raw_string_ostream Out(Str); 2748 Out << "[ "; 2749 for (SymbolRef ClassMember : ClassMembersSorted) { 2750 if (FirstMember) 2751 FirstMember = false; 2752 else 2753 Out << ", "; 2754 Out << "\"" << ClassMember << "\""; 2755 } 2756 Out << " ]"; 2757 return Out.str(); 2758 } 2759 2760 void RangeConstraintManager::printEquivalenceClasses(raw_ostream &Out, 2761 ProgramStateRef State, 2762 const char *NL, 2763 unsigned int Space, 2764 bool IsDot) const { 2765 ClassMembersTy Members = State->get<ClassMembers>(); 2766 2767 Indent(Out, Space, IsDot) << "\"equivalence_classes\": "; 2768 if (Members.isEmpty()) { 2769 Out << "null," << NL; 2770 return; 2771 } 2772 2773 std::set<std::string> MembersStr; 2774 for (std::pair<EquivalenceClass, SymbolSet> ClassToSymbolSet : Members) 2775 MembersStr.insert(toString(State, ClassToSymbolSet.first)); 2776 2777 ++Space; 2778 Out << '[' << NL; 2779 bool FirstClass = true; 2780 for (const std::string &Str : MembersStr) { 2781 if (FirstClass) { 2782 FirstClass = false; 2783 } else { 2784 Out << ','; 2785 Out << NL; 2786 } 2787 Indent(Out, Space, IsDot); 2788 Out << Str; 2789 } 2790 Out << NL; 2791 2792 --Space; 2793 Indent(Out, Space, IsDot) << "]," << NL; 2794 } 2795 2796 void RangeConstraintManager::printDisequalities(raw_ostream &Out, 2797 ProgramStateRef State, 2798 const char *NL, 2799 unsigned int Space, 2800 bool IsDot) const { 2801 DisequalityMapTy Disequalities = State->get<DisequalityMap>(); 2802 2803 Indent(Out, Space, IsDot) << "\"disequality_info\": "; 2804 if (Disequalities.isEmpty()) { 2805 Out << "null," << NL; 2806 return; 2807 } 2808 2809 // Transform the disequality info to an ordered map of 2810 // [string -> (ordered set of strings)] 2811 using EqClassesStrTy = std::set<std::string>; 2812 using DisequalityInfoStrTy = std::map<std::string, EqClassesStrTy>; 2813 DisequalityInfoStrTy DisequalityInfoStr; 2814 for (std::pair<EquivalenceClass, ClassSet> ClassToDisEqSet : Disequalities) { 2815 EquivalenceClass Class = ClassToDisEqSet.first; 2816 ClassSet DisequalClasses = ClassToDisEqSet.second; 2817 EqClassesStrTy MembersStr; 2818 for (EquivalenceClass DisEqClass : DisequalClasses) 2819 MembersStr.insert(toString(State, DisEqClass)); 2820 DisequalityInfoStr.insert({toString(State, Class), MembersStr}); 2821 } 2822 2823 ++Space; 2824 Out << '[' << NL; 2825 bool FirstClass = true; 2826 for (std::pair<std::string, EqClassesStrTy> ClassToDisEqSet : 2827 DisequalityInfoStr) { 2828 const std::string &Class = ClassToDisEqSet.first; 2829 if (FirstClass) { 2830 FirstClass = false; 2831 } else { 2832 Out << ','; 2833 Out << NL; 2834 } 2835 Indent(Out, Space, IsDot) << "{" << NL; 2836 unsigned int DisEqSpace = Space + 1; 2837 Indent(Out, DisEqSpace, IsDot) << "\"class\": "; 2838 Out << Class; 2839 const EqClassesStrTy &DisequalClasses = ClassToDisEqSet.second; 2840 if (!DisequalClasses.empty()) { 2841 Out << "," << NL; 2842 Indent(Out, DisEqSpace, IsDot) << "\"disequal_to\": [" << NL; 2843 unsigned int DisEqClassSpace = DisEqSpace + 1; 2844 Indent(Out, DisEqClassSpace, IsDot); 2845 bool FirstDisEqClass = true; 2846 for (const std::string &DisEqClass : DisequalClasses) { 2847 if (FirstDisEqClass) { 2848 FirstDisEqClass = false; 2849 } else { 2850 Out << ',' << NL; 2851 Indent(Out, DisEqClassSpace, IsDot); 2852 } 2853 Out << DisEqClass; 2854 } 2855 Out << "]" << NL; 2856 } 2857 Indent(Out, Space, IsDot) << "}"; 2858 } 2859 Out << NL; 2860 2861 --Space; 2862 Indent(Out, Space, IsDot) << "]," << NL; 2863 } 2864