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 606 simplify(SValBuilder &SVB, RangeSet::Factory &F, RangedConstraintManager &RCM, 607 ProgramStateRef State, EquivalenceClass Class); 608 609 void dumpToStream(ProgramStateRef State, raw_ostream &os) const; 610 LLVM_DUMP_METHOD void dump(ProgramStateRef State) const { 611 dumpToStream(State, llvm::errs()); 612 } 613 614 /// Check equivalence data for consistency. 615 LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED static bool 616 isClassDataConsistent(ProgramStateRef State); 617 618 LLVM_NODISCARD QualType getType() const { 619 return getRepresentativeSymbol()->getType(); 620 } 621 622 EquivalenceClass() = delete; 623 EquivalenceClass(const EquivalenceClass &) = default; 624 EquivalenceClass &operator=(const EquivalenceClass &) = delete; 625 EquivalenceClass(EquivalenceClass &&) = default; 626 EquivalenceClass &operator=(EquivalenceClass &&) = delete; 627 628 bool operator==(const EquivalenceClass &Other) const { 629 return ID == Other.ID; 630 } 631 bool operator<(const EquivalenceClass &Other) const { return ID < Other.ID; } 632 bool operator!=(const EquivalenceClass &Other) const { 633 return !operator==(Other); 634 } 635 636 static void Profile(llvm::FoldingSetNodeID &ID, uintptr_t CID) { 637 ID.AddInteger(CID); 638 } 639 640 void Profile(llvm::FoldingSetNodeID &ID) const { Profile(ID, this->ID); } 641 642 private: 643 /* implicit */ EquivalenceClass(SymbolRef Sym) 644 : ID(reinterpret_cast<uintptr_t>(Sym)) {} 645 646 /// This function is intended to be used ONLY within the class. 647 /// The fact that ID is a pointer to a symbol is an implementation detail 648 /// and should stay that way. 649 /// In the current implementation, we use it to retrieve the only member 650 /// of the trivial class. 651 SymbolRef getRepresentativeSymbol() const { 652 return reinterpret_cast<SymbolRef>(ID); 653 } 654 static inline SymbolSet::Factory &getMembersFactory(ProgramStateRef State); 655 656 inline ProgramStateRef mergeImpl(RangeSet::Factory &F, ProgramStateRef State, 657 SymbolSet Members, EquivalenceClass Other, 658 SymbolSet OtherMembers); 659 static inline bool 660 addToDisequalityInfo(DisequalityMapTy &Info, ConstraintRangeTy &Constraints, 661 RangeSet::Factory &F, ProgramStateRef State, 662 EquivalenceClass First, EquivalenceClass Second); 663 664 /// This is a unique identifier of the class. 665 uintptr_t ID; 666 }; 667 668 //===----------------------------------------------------------------------===// 669 // Constraint functions 670 //===----------------------------------------------------------------------===// 671 672 LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED bool 673 areFeasible(ConstraintRangeTy Constraints) { 674 return llvm::none_of( 675 Constraints, 676 [](const std::pair<EquivalenceClass, RangeSet> &ClassConstraint) { 677 return ClassConstraint.second.isEmpty(); 678 }); 679 } 680 681 LLVM_NODISCARD inline const RangeSet *getConstraint(ProgramStateRef State, 682 EquivalenceClass Class) { 683 return State->get<ConstraintRange>(Class); 684 } 685 686 LLVM_NODISCARD inline const RangeSet *getConstraint(ProgramStateRef State, 687 SymbolRef Sym) { 688 return getConstraint(State, EquivalenceClass::find(State, Sym)); 689 } 690 691 LLVM_NODISCARD ProgramStateRef setConstraint(ProgramStateRef State, 692 EquivalenceClass Class, 693 RangeSet Constraint) { 694 return State->set<ConstraintRange>(Class, Constraint); 695 } 696 697 LLVM_NODISCARD ProgramStateRef setConstraints(ProgramStateRef State, 698 ConstraintRangeTy Constraints) { 699 return State->set<ConstraintRange>(Constraints); 700 } 701 702 //===----------------------------------------------------------------------===// 703 // Equality/diseqiality abstraction 704 //===----------------------------------------------------------------------===// 705 706 /// A small helper function for detecting symbolic (dis)equality. 707 /// 708 /// Equality check can have different forms (like a == b or a - b) and this 709 /// class encapsulates those away if the only thing the user wants to check - 710 /// whether it's equality/diseqiality or not. 711 /// 712 /// \returns true if assuming this Sym to be true means equality of operands 713 /// false if it means disequality of operands 714 /// None otherwise 715 Optional<bool> meansEquality(const SymSymExpr *Sym) { 716 switch (Sym->getOpcode()) { 717 case BO_Sub: 718 // This case is: A - B != 0 -> disequality check. 719 return false; 720 case BO_EQ: 721 // This case is: A == B != 0 -> equality check. 722 return true; 723 case BO_NE: 724 // This case is: A != B != 0 -> diseqiality check. 725 return false; 726 default: 727 return llvm::None; 728 } 729 } 730 731 //===----------------------------------------------------------------------===// 732 // Intersection functions 733 //===----------------------------------------------------------------------===// 734 735 template <class SecondTy, class... RestTy> 736 LLVM_NODISCARD inline RangeSet intersect(RangeSet::Factory &F, RangeSet Head, 737 SecondTy Second, RestTy... Tail); 738 739 template <class... RangeTy> struct IntersectionTraits; 740 741 template <class... TailTy> struct IntersectionTraits<RangeSet, TailTy...> { 742 // Found RangeSet, no need to check any further 743 using Type = RangeSet; 744 }; 745 746 template <> struct IntersectionTraits<> { 747 // We ran out of types, and we didn't find any RangeSet, so the result should 748 // be optional. 749 using Type = Optional<RangeSet>; 750 }; 751 752 template <class OptionalOrPointer, class... TailTy> 753 struct IntersectionTraits<OptionalOrPointer, TailTy...> { 754 // If current type is Optional or a raw pointer, we should keep looking. 755 using Type = typename IntersectionTraits<TailTy...>::Type; 756 }; 757 758 template <class EndTy> 759 LLVM_NODISCARD inline EndTy intersect(RangeSet::Factory &F, EndTy End) { 760 // If the list contains only RangeSet or Optional<RangeSet>, simply return 761 // that range set. 762 return End; 763 } 764 765 LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED inline Optional<RangeSet> 766 intersect(RangeSet::Factory &F, const RangeSet *End) { 767 // This is an extraneous conversion from a raw pointer into Optional<RangeSet> 768 if (End) { 769 return *End; 770 } 771 return llvm::None; 772 } 773 774 template <class... RestTy> 775 LLVM_NODISCARD inline RangeSet intersect(RangeSet::Factory &F, RangeSet Head, 776 RangeSet Second, RestTy... Tail) { 777 // Here we call either the <RangeSet,RangeSet,...> or <RangeSet,...> version 778 // of the function and can be sure that the result is RangeSet. 779 return intersect(F, F.intersect(Head, Second), Tail...); 780 } 781 782 template <class SecondTy, class... RestTy> 783 LLVM_NODISCARD inline RangeSet intersect(RangeSet::Factory &F, RangeSet Head, 784 SecondTy Second, RestTy... Tail) { 785 if (Second) { 786 // Here we call the <RangeSet,RangeSet,...> version of the function... 787 return intersect(F, Head, *Second, Tail...); 788 } 789 // ...and here it is either <RangeSet,RangeSet,...> or <RangeSet,...>, which 790 // means that the result is definitely RangeSet. 791 return intersect(F, Head, Tail...); 792 } 793 794 /// Main generic intersect function. 795 /// It intersects all of the given range sets. If some of the given arguments 796 /// don't hold a range set (nullptr or llvm::None), the function will skip them. 797 /// 798 /// Available representations for the arguments are: 799 /// * RangeSet 800 /// * Optional<RangeSet> 801 /// * RangeSet * 802 /// Pointer to a RangeSet is automatically assumed to be nullable and will get 803 /// checked as well as the optional version. If this behaviour is undesired, 804 /// please dereference the pointer in the call. 805 /// 806 /// Return type depends on the arguments' types. If we can be sure in compile 807 /// time that there will be a range set as a result, the returning type is 808 /// simply RangeSet, in other cases we have to back off to Optional<RangeSet>. 809 /// 810 /// Please, prefer optional range sets to raw pointers. If the last argument is 811 /// a raw pointer and all previous arguments are None, it will cost one 812 /// additional check to convert RangeSet * into Optional<RangeSet>. 813 template <class HeadTy, class SecondTy, class... RestTy> 814 LLVM_NODISCARD inline 815 typename IntersectionTraits<HeadTy, SecondTy, RestTy...>::Type 816 intersect(RangeSet::Factory &F, HeadTy Head, SecondTy Second, 817 RestTy... Tail) { 818 if (Head) { 819 return intersect(F, *Head, Second, Tail...); 820 } 821 return intersect(F, Second, Tail...); 822 } 823 824 //===----------------------------------------------------------------------===// 825 // Symbolic reasoning logic 826 //===----------------------------------------------------------------------===// 827 828 /// A little component aggregating all of the reasoning we have about 829 /// the ranges of symbolic expressions. 830 /// 831 /// Even when we don't know the exact values of the operands, we still 832 /// can get a pretty good estimate of the result's range. 833 class SymbolicRangeInferrer 834 : public SymExprVisitor<SymbolicRangeInferrer, RangeSet> { 835 public: 836 template <class SourceType> 837 static RangeSet inferRange(RangeSet::Factory &F, ProgramStateRef State, 838 SourceType Origin) { 839 SymbolicRangeInferrer Inferrer(F, State); 840 return Inferrer.infer(Origin); 841 } 842 843 RangeSet VisitSymExpr(SymbolRef Sym) { 844 // If we got to this function, the actual type of the symbolic 845 // expression is not supported for advanced inference. 846 // In this case, we simply backoff to the default "let's simply 847 // infer the range from the expression's type". 848 return infer(Sym->getType()); 849 } 850 851 RangeSet VisitSymIntExpr(const SymIntExpr *Sym) { 852 return VisitBinaryOperator(Sym); 853 } 854 855 RangeSet VisitIntSymExpr(const IntSymExpr *Sym) { 856 return VisitBinaryOperator(Sym); 857 } 858 859 RangeSet VisitSymSymExpr(const SymSymExpr *Sym) { 860 return intersect( 861 RangeFactory, 862 // If Sym is (dis)equality, we might have some information 863 // on that in our equality classes data structure. 864 getRangeForEqualities(Sym), 865 // And we should always check what we can get from the operands. 866 VisitBinaryOperator(Sym)); 867 } 868 869 private: 870 SymbolicRangeInferrer(RangeSet::Factory &F, ProgramStateRef S) 871 : ValueFactory(F.getValueFactory()), RangeFactory(F), State(S) {} 872 873 /// Infer range information from the given integer constant. 874 /// 875 /// It's not a real "inference", but is here for operating with 876 /// sub-expressions in a more polymorphic manner. 877 RangeSet inferAs(const llvm::APSInt &Val, QualType) { 878 return {RangeFactory, Val}; 879 } 880 881 /// Infer range information from symbol in the context of the given type. 882 RangeSet inferAs(SymbolRef Sym, QualType DestType) { 883 QualType ActualType = Sym->getType(); 884 // Check that we can reason about the symbol at all. 885 if (ActualType->isIntegralOrEnumerationType() || 886 Loc::isLocType(ActualType)) { 887 return infer(Sym); 888 } 889 // Otherwise, let's simply infer from the destination type. 890 // We couldn't figure out nothing else about that expression. 891 return infer(DestType); 892 } 893 894 RangeSet infer(SymbolRef Sym) { 895 return intersect( 896 RangeFactory, 897 // Of course, we should take the constraint directly associated with 898 // this symbol into consideration. 899 getConstraint(State, Sym), 900 // If Sym is a difference of symbols A - B, then maybe we have range 901 // set stored for B - A. 902 // 903 // If we have range set stored for both A - B and B - A then 904 // calculate the effective range set by intersecting the range set 905 // for A - B and the negated range set of B - A. 906 getRangeForNegatedSub(Sym), 907 // If Sym is a comparison expression (except <=>), 908 // find any other comparisons with the same operands. 909 // See function description. 910 getRangeForComparisonSymbol(Sym), 911 // Apart from the Sym itself, we can infer quite a lot if we look 912 // into subexpressions of Sym. 913 Visit(Sym)); 914 } 915 916 RangeSet infer(EquivalenceClass Class) { 917 if (const RangeSet *AssociatedConstraint = getConstraint(State, Class)) 918 return *AssociatedConstraint; 919 920 return infer(Class.getType()); 921 } 922 923 /// Infer range information solely from the type. 924 RangeSet infer(QualType T) { 925 // Lazily generate a new RangeSet representing all possible values for the 926 // given symbol type. 927 RangeSet Result(RangeFactory, ValueFactory.getMinValue(T), 928 ValueFactory.getMaxValue(T)); 929 930 // References are known to be non-zero. 931 if (T->isReferenceType()) 932 return assumeNonZero(Result, T); 933 934 return Result; 935 } 936 937 template <class BinarySymExprTy> 938 RangeSet VisitBinaryOperator(const BinarySymExprTy *Sym) { 939 // TODO #1: VisitBinaryOperator implementation might not make a good 940 // use of the inferred ranges. In this case, we might be calculating 941 // everything for nothing. This being said, we should introduce some 942 // sort of laziness mechanism here. 943 // 944 // TODO #2: We didn't go into the nested expressions before, so it 945 // might cause us spending much more time doing the inference. 946 // This can be a problem for deeply nested expressions that are 947 // involved in conditions and get tested continuously. We definitely 948 // need to address this issue and introduce some sort of caching 949 // in here. 950 QualType ResultType = Sym->getType(); 951 return VisitBinaryOperator(inferAs(Sym->getLHS(), ResultType), 952 Sym->getOpcode(), 953 inferAs(Sym->getRHS(), ResultType), ResultType); 954 } 955 956 RangeSet VisitBinaryOperator(RangeSet LHS, BinaryOperator::Opcode Op, 957 RangeSet RHS, QualType T) { 958 switch (Op) { 959 case BO_Or: 960 return VisitBinaryOperator<BO_Or>(LHS, RHS, T); 961 case BO_And: 962 return VisitBinaryOperator<BO_And>(LHS, RHS, T); 963 case BO_Rem: 964 return VisitBinaryOperator<BO_Rem>(LHS, RHS, T); 965 default: 966 return infer(T); 967 } 968 } 969 970 //===----------------------------------------------------------------------===// 971 // Ranges and operators 972 //===----------------------------------------------------------------------===// 973 974 /// Return a rough approximation of the given range set. 975 /// 976 /// For the range set: 977 /// { [x_0, y_0], [x_1, y_1], ... , [x_N, y_N] } 978 /// it will return the range [x_0, y_N]. 979 static Range fillGaps(RangeSet Origin) { 980 assert(!Origin.isEmpty()); 981 return {Origin.getMinValue(), Origin.getMaxValue()}; 982 } 983 984 /// Try to convert given range into the given type. 985 /// 986 /// It will return llvm::None only when the trivial conversion is possible. 987 llvm::Optional<Range> convert(const Range &Origin, APSIntType To) { 988 if (To.testInRange(Origin.From(), false) != APSIntType::RTR_Within || 989 To.testInRange(Origin.To(), false) != APSIntType::RTR_Within) { 990 return llvm::None; 991 } 992 return Range(ValueFactory.Convert(To, Origin.From()), 993 ValueFactory.Convert(To, Origin.To())); 994 } 995 996 template <BinaryOperator::Opcode Op> 997 RangeSet VisitBinaryOperator(RangeSet LHS, RangeSet RHS, QualType T) { 998 // We should propagate information about unfeasbility of one of the 999 // operands to the resulting range. 1000 if (LHS.isEmpty() || RHS.isEmpty()) { 1001 return RangeFactory.getEmptySet(); 1002 } 1003 1004 Range CoarseLHS = fillGaps(LHS); 1005 Range CoarseRHS = fillGaps(RHS); 1006 1007 APSIntType ResultType = ValueFactory.getAPSIntType(T); 1008 1009 // We need to convert ranges to the resulting type, so we can compare values 1010 // and combine them in a meaningful (in terms of the given operation) way. 1011 auto ConvertedCoarseLHS = convert(CoarseLHS, ResultType); 1012 auto ConvertedCoarseRHS = convert(CoarseRHS, ResultType); 1013 1014 // It is hard to reason about ranges when conversion changes 1015 // borders of the ranges. 1016 if (!ConvertedCoarseLHS || !ConvertedCoarseRHS) { 1017 return infer(T); 1018 } 1019 1020 return VisitBinaryOperator<Op>(*ConvertedCoarseLHS, *ConvertedCoarseRHS, T); 1021 } 1022 1023 template <BinaryOperator::Opcode Op> 1024 RangeSet VisitBinaryOperator(Range LHS, Range RHS, QualType T) { 1025 return infer(T); 1026 } 1027 1028 /// Return a symmetrical range for the given range and type. 1029 /// 1030 /// If T is signed, return the smallest range [-x..x] that covers the original 1031 /// range, or [-min(T), max(T)] if the aforementioned symmetric range doesn't 1032 /// exist due to original range covering min(T)). 1033 /// 1034 /// If T is unsigned, return the smallest range [0..x] that covers the 1035 /// original range. 1036 Range getSymmetricalRange(Range Origin, QualType T) { 1037 APSIntType RangeType = ValueFactory.getAPSIntType(T); 1038 1039 if (RangeType.isUnsigned()) { 1040 return Range(ValueFactory.getMinValue(RangeType), Origin.To()); 1041 } 1042 1043 if (Origin.From().isMinSignedValue()) { 1044 // If mini is a minimal signed value, absolute value of it is greater 1045 // than the maximal signed value. In order to avoid these 1046 // complications, we simply return the whole range. 1047 return {ValueFactory.getMinValue(RangeType), 1048 ValueFactory.getMaxValue(RangeType)}; 1049 } 1050 1051 // At this point, we are sure that the type is signed and we can safely 1052 // use unary - operator. 1053 // 1054 // While calculating absolute maximum, we can use the following formula 1055 // because of these reasons: 1056 // * If From >= 0 then To >= From and To >= -From. 1057 // AbsMax == To == max(To, -From) 1058 // * If To <= 0 then -From >= -To and -From >= From. 1059 // AbsMax == -From == max(-From, To) 1060 // * Otherwise, From <= 0, To >= 0, and 1061 // AbsMax == max(abs(From), abs(To)) 1062 llvm::APSInt AbsMax = std::max(-Origin.From(), Origin.To()); 1063 1064 // Intersection is guaranteed to be non-empty. 1065 return {ValueFactory.getValue(-AbsMax), ValueFactory.getValue(AbsMax)}; 1066 } 1067 1068 /// Return a range set subtracting zero from \p Domain. 1069 RangeSet assumeNonZero(RangeSet Domain, QualType T) { 1070 APSIntType IntType = ValueFactory.getAPSIntType(T); 1071 return RangeFactory.deletePoint(Domain, IntType.getZeroValue()); 1072 } 1073 1074 // FIXME: Once SValBuilder supports unary minus, we should use SValBuilder to 1075 // obtain the negated symbolic expression instead of constructing the 1076 // symbol manually. This will allow us to support finding ranges of not 1077 // only negated SymSymExpr-type expressions, but also of other, simpler 1078 // expressions which we currently do not know how to negate. 1079 Optional<RangeSet> getRangeForNegatedSub(SymbolRef Sym) { 1080 if (const SymSymExpr *SSE = dyn_cast<SymSymExpr>(Sym)) { 1081 if (SSE->getOpcode() == BO_Sub) { 1082 QualType T = Sym->getType(); 1083 1084 // Do not negate unsigned ranges 1085 if (!T->isUnsignedIntegerOrEnumerationType() && 1086 !T->isSignedIntegerOrEnumerationType()) 1087 return llvm::None; 1088 1089 SymbolManager &SymMgr = State->getSymbolManager(); 1090 SymbolRef NegatedSym = 1091 SymMgr.getSymSymExpr(SSE->getRHS(), BO_Sub, SSE->getLHS(), T); 1092 1093 if (const RangeSet *NegatedRange = getConstraint(State, NegatedSym)) { 1094 return RangeFactory.negate(*NegatedRange); 1095 } 1096 } 1097 } 1098 return llvm::None; 1099 } 1100 1101 // Returns ranges only for binary comparison operators (except <=>) 1102 // when left and right operands are symbolic values. 1103 // Finds any other comparisons with the same operands. 1104 // Then do logical calculations and refuse impossible branches. 1105 // E.g. (x < y) and (x > y) at the same time are impossible. 1106 // E.g. (x >= y) and (x != y) at the same time makes (x > y) true only. 1107 // E.g. (x == y) and (y == x) are just reversed but the same. 1108 // It covers all possible combinations (see CmpOpTable description). 1109 // Note that `x` and `y` can also stand for subexpressions, 1110 // not only for actual symbols. 1111 Optional<RangeSet> getRangeForComparisonSymbol(SymbolRef Sym) { 1112 const auto *SSE = dyn_cast<SymSymExpr>(Sym); 1113 if (!SSE) 1114 return llvm::None; 1115 1116 const BinaryOperatorKind CurrentOP = SSE->getOpcode(); 1117 1118 // We currently do not support <=> (C++20). 1119 if (!BinaryOperator::isComparisonOp(CurrentOP) || (CurrentOP == BO_Cmp)) 1120 return llvm::None; 1121 1122 static const OperatorRelationsTable CmpOpTable{}; 1123 1124 const SymExpr *LHS = SSE->getLHS(); 1125 const SymExpr *RHS = SSE->getRHS(); 1126 QualType T = SSE->getType(); 1127 1128 SymbolManager &SymMgr = State->getSymbolManager(); 1129 1130 // We use this variable to store the last queried operator (`QueriedOP`) 1131 // for which the `getCmpOpState` returned with `Unknown`. If there are two 1132 // different OPs that returned `Unknown` then we have to query the special 1133 // `UnknownX2` column. We assume that `getCmpOpState(CurrentOP, CurrentOP)` 1134 // never returns `Unknown`, so `CurrentOP` is a good initial value. 1135 BinaryOperatorKind LastQueriedOpToUnknown = CurrentOP; 1136 1137 // Loop goes through all of the columns exept the last one ('UnknownX2'). 1138 // We treat `UnknownX2` column separately at the end of the loop body. 1139 for (size_t i = 0; i < CmpOpTable.getCmpOpCount(); ++i) { 1140 1141 // Let's find an expression e.g. (x < y). 1142 BinaryOperatorKind QueriedOP = OperatorRelationsTable::getOpFromIndex(i); 1143 const SymSymExpr *SymSym = SymMgr.getSymSymExpr(LHS, QueriedOP, RHS, T); 1144 const RangeSet *QueriedRangeSet = getConstraint(State, SymSym); 1145 1146 // If ranges were not previously found, 1147 // try to find a reversed expression (y > x). 1148 if (!QueriedRangeSet) { 1149 const BinaryOperatorKind ROP = 1150 BinaryOperator::reverseComparisonOp(QueriedOP); 1151 SymSym = SymMgr.getSymSymExpr(RHS, ROP, LHS, T); 1152 QueriedRangeSet = getConstraint(State, SymSym); 1153 } 1154 1155 if (!QueriedRangeSet || QueriedRangeSet->isEmpty()) 1156 continue; 1157 1158 const llvm::APSInt *ConcreteValue = QueriedRangeSet->getConcreteValue(); 1159 const bool isInFalseBranch = 1160 ConcreteValue ? (*ConcreteValue == 0) : false; 1161 1162 // If it is a false branch, we shall be guided by opposite operator, 1163 // because the table is made assuming we are in the true branch. 1164 // E.g. when (x <= y) is false, then (x > y) is true. 1165 if (isInFalseBranch) 1166 QueriedOP = BinaryOperator::negateComparisonOp(QueriedOP); 1167 1168 OperatorRelationsTable::TriStateKind BranchState = 1169 CmpOpTable.getCmpOpState(CurrentOP, QueriedOP); 1170 1171 if (BranchState == OperatorRelationsTable::Unknown) { 1172 if (LastQueriedOpToUnknown != CurrentOP && 1173 LastQueriedOpToUnknown != QueriedOP) { 1174 // If we got the Unknown state for both different operators. 1175 // if (x <= y) // assume true 1176 // if (x != y) // assume true 1177 // if (x < y) // would be also true 1178 // Get a state from `UnknownX2` column. 1179 BranchState = CmpOpTable.getCmpOpStateForUnknownX2(CurrentOP); 1180 } else { 1181 LastQueriedOpToUnknown = QueriedOP; 1182 continue; 1183 } 1184 } 1185 1186 return (BranchState == OperatorRelationsTable::True) ? getTrueRange(T) 1187 : getFalseRange(T); 1188 } 1189 1190 return llvm::None; 1191 } 1192 1193 Optional<RangeSet> getRangeForEqualities(const SymSymExpr *Sym) { 1194 Optional<bool> Equality = meansEquality(Sym); 1195 1196 if (!Equality) 1197 return llvm::None; 1198 1199 if (Optional<bool> AreEqual = 1200 EquivalenceClass::areEqual(State, Sym->getLHS(), Sym->getRHS())) { 1201 // Here we cover two cases at once: 1202 // * if Sym is equality and its operands are known to be equal -> true 1203 // * if Sym is disequality and its operands are disequal -> true 1204 if (*AreEqual == *Equality) { 1205 return getTrueRange(Sym->getType()); 1206 } 1207 // Opposite combinations result in false. 1208 return getFalseRange(Sym->getType()); 1209 } 1210 1211 return llvm::None; 1212 } 1213 1214 RangeSet getTrueRange(QualType T) { 1215 RangeSet TypeRange = infer(T); 1216 return assumeNonZero(TypeRange, T); 1217 } 1218 1219 RangeSet getFalseRange(QualType T) { 1220 const llvm::APSInt &Zero = ValueFactory.getValue(0, T); 1221 return RangeSet(RangeFactory, Zero); 1222 } 1223 1224 BasicValueFactory &ValueFactory; 1225 RangeSet::Factory &RangeFactory; 1226 ProgramStateRef State; 1227 }; 1228 1229 //===----------------------------------------------------------------------===// 1230 // Range-based reasoning about symbolic operations 1231 //===----------------------------------------------------------------------===// 1232 1233 template <> 1234 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_Or>(Range LHS, Range RHS, 1235 QualType T) { 1236 APSIntType ResultType = ValueFactory.getAPSIntType(T); 1237 llvm::APSInt Zero = ResultType.getZeroValue(); 1238 1239 bool IsLHSPositiveOrZero = LHS.From() >= Zero; 1240 bool IsRHSPositiveOrZero = RHS.From() >= Zero; 1241 1242 bool IsLHSNegative = LHS.To() < Zero; 1243 bool IsRHSNegative = RHS.To() < Zero; 1244 1245 // Check if both ranges have the same sign. 1246 if ((IsLHSPositiveOrZero && IsRHSPositiveOrZero) || 1247 (IsLHSNegative && IsRHSNegative)) { 1248 // The result is definitely greater or equal than any of the operands. 1249 const llvm::APSInt &Min = std::max(LHS.From(), RHS.From()); 1250 1251 // We estimate maximal value for positives as the maximal value for the 1252 // given type. For negatives, we estimate it with -1 (e.g. 0x11111111). 1253 // 1254 // TODO: We basically, limit the resulting range from below, but don't do 1255 // anything with the upper bound. 1256 // 1257 // For positive operands, it can be done as follows: for the upper 1258 // bound of LHS and RHS we calculate the most significant bit set. 1259 // Let's call it the N-th bit. Then we can estimate the maximal 1260 // number to be 2^(N+1)-1, i.e. the number with all the bits up to 1261 // the N-th bit set. 1262 const llvm::APSInt &Max = IsLHSNegative 1263 ? ValueFactory.getValue(--Zero) 1264 : ValueFactory.getMaxValue(ResultType); 1265 1266 return {RangeFactory, ValueFactory.getValue(Min), Max}; 1267 } 1268 1269 // Otherwise, let's check if at least one of the operands is negative. 1270 if (IsLHSNegative || IsRHSNegative) { 1271 // This means that the result is definitely negative as well. 1272 return {RangeFactory, ValueFactory.getMinValue(ResultType), 1273 ValueFactory.getValue(--Zero)}; 1274 } 1275 1276 RangeSet DefaultRange = infer(T); 1277 1278 // It is pretty hard to reason about operands with different signs 1279 // (and especially with possibly different signs). We simply check if it 1280 // can be zero. In order to conclude that the result could not be zero, 1281 // at least one of the operands should be definitely not zero itself. 1282 if (!LHS.Includes(Zero) || !RHS.Includes(Zero)) { 1283 return assumeNonZero(DefaultRange, T); 1284 } 1285 1286 // Nothing much else to do here. 1287 return DefaultRange; 1288 } 1289 1290 template <> 1291 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_And>(Range LHS, 1292 Range RHS, 1293 QualType T) { 1294 APSIntType ResultType = ValueFactory.getAPSIntType(T); 1295 llvm::APSInt Zero = ResultType.getZeroValue(); 1296 1297 bool IsLHSPositiveOrZero = LHS.From() >= Zero; 1298 bool IsRHSPositiveOrZero = RHS.From() >= Zero; 1299 1300 bool IsLHSNegative = LHS.To() < Zero; 1301 bool IsRHSNegative = RHS.To() < Zero; 1302 1303 // Check if both ranges have the same sign. 1304 if ((IsLHSPositiveOrZero && IsRHSPositiveOrZero) || 1305 (IsLHSNegative && IsRHSNegative)) { 1306 // The result is definitely less or equal than any of the operands. 1307 const llvm::APSInt &Max = std::min(LHS.To(), RHS.To()); 1308 1309 // We conservatively estimate lower bound to be the smallest positive 1310 // or negative value corresponding to the sign of the operands. 1311 const llvm::APSInt &Min = IsLHSNegative 1312 ? ValueFactory.getMinValue(ResultType) 1313 : ValueFactory.getValue(Zero); 1314 1315 return {RangeFactory, Min, Max}; 1316 } 1317 1318 // Otherwise, let's check if at least one of the operands is positive. 1319 if (IsLHSPositiveOrZero || IsRHSPositiveOrZero) { 1320 // This makes result definitely positive. 1321 // 1322 // We can also reason about a maximal value by finding the maximal 1323 // value of the positive operand. 1324 const llvm::APSInt &Max = IsLHSPositiveOrZero ? LHS.To() : RHS.To(); 1325 1326 // The minimal value on the other hand is much harder to reason about. 1327 // The only thing we know for sure is that the result is positive. 1328 return {RangeFactory, ValueFactory.getValue(Zero), 1329 ValueFactory.getValue(Max)}; 1330 } 1331 1332 // Nothing much else to do here. 1333 return infer(T); 1334 } 1335 1336 template <> 1337 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_Rem>(Range LHS, 1338 Range RHS, 1339 QualType T) { 1340 llvm::APSInt Zero = ValueFactory.getAPSIntType(T).getZeroValue(); 1341 1342 Range ConservativeRange = getSymmetricalRange(RHS, T); 1343 1344 llvm::APSInt Max = ConservativeRange.To(); 1345 llvm::APSInt Min = ConservativeRange.From(); 1346 1347 if (Max == Zero) { 1348 // It's an undefined behaviour to divide by 0 and it seems like we know 1349 // for sure that RHS is 0. Let's say that the resulting range is 1350 // simply infeasible for that matter. 1351 return RangeFactory.getEmptySet(); 1352 } 1353 1354 // At this point, our conservative range is closed. The result, however, 1355 // couldn't be greater than the RHS' maximal absolute value. Because of 1356 // this reason, we turn the range into open (or half-open in case of 1357 // unsigned integers). 1358 // 1359 // While we operate on integer values, an open interval (a, b) can be easily 1360 // represented by the closed interval [a + 1, b - 1]. And this is exactly 1361 // what we do next. 1362 // 1363 // If we are dealing with unsigned case, we shouldn't move the lower bound. 1364 if (Min.isSigned()) { 1365 ++Min; 1366 } 1367 --Max; 1368 1369 bool IsLHSPositiveOrZero = LHS.From() >= Zero; 1370 bool IsRHSPositiveOrZero = RHS.From() >= Zero; 1371 1372 // Remainder operator results with negative operands is implementation 1373 // defined. Positive cases are much easier to reason about though. 1374 if (IsLHSPositiveOrZero && IsRHSPositiveOrZero) { 1375 // If maximal value of LHS is less than maximal value of RHS, 1376 // the result won't get greater than LHS.To(). 1377 Max = std::min(LHS.To(), Max); 1378 // We want to check if it is a situation similar to the following: 1379 // 1380 // <------------|---[ LHS ]--------[ RHS ]-----> 1381 // -INF 0 +INF 1382 // 1383 // In this situation, we can conclude that (LHS / RHS) == 0 and 1384 // (LHS % RHS) == LHS. 1385 Min = LHS.To() < RHS.From() ? LHS.From() : Zero; 1386 } 1387 1388 // Nevertheless, the symmetrical range for RHS is a conservative estimate 1389 // for any sign of either LHS, or RHS. 1390 return {RangeFactory, ValueFactory.getValue(Min), ValueFactory.getValue(Max)}; 1391 } 1392 1393 //===----------------------------------------------------------------------===// 1394 // Constraint manager implementation details 1395 //===----------------------------------------------------------------------===// 1396 1397 class RangeConstraintManager : public RangedConstraintManager { 1398 public: 1399 RangeConstraintManager(ExprEngine *EE, SValBuilder &SVB) 1400 : RangedConstraintManager(EE, SVB), F(getBasicVals()) {} 1401 1402 //===------------------------------------------------------------------===// 1403 // Implementation for interface from ConstraintManager. 1404 //===------------------------------------------------------------------===// 1405 1406 bool haveEqualConstraints(ProgramStateRef S1, 1407 ProgramStateRef S2) const override { 1408 // NOTE: ClassMembers are as simple as back pointers for ClassMap, 1409 // so comparing constraint ranges and class maps should be 1410 // sufficient. 1411 return S1->get<ConstraintRange>() == S2->get<ConstraintRange>() && 1412 S1->get<ClassMap>() == S2->get<ClassMap>(); 1413 } 1414 1415 bool canReasonAbout(SVal X) const override; 1416 1417 ConditionTruthVal checkNull(ProgramStateRef State, SymbolRef Sym) override; 1418 1419 const llvm::APSInt *getSymVal(ProgramStateRef State, 1420 SymbolRef Sym) const override; 1421 1422 ProgramStateRef removeDeadBindings(ProgramStateRef State, 1423 SymbolReaper &SymReaper) override; 1424 1425 void printJson(raw_ostream &Out, ProgramStateRef State, const char *NL = "\n", 1426 unsigned int Space = 0, bool IsDot = false) const override; 1427 void printConstraints(raw_ostream &Out, ProgramStateRef State, 1428 const char *NL = "\n", unsigned int Space = 0, 1429 bool IsDot = false) const; 1430 void printEquivalenceClasses(raw_ostream &Out, ProgramStateRef State, 1431 const char *NL = "\n", unsigned int Space = 0, 1432 bool IsDot = false) const; 1433 void printDisequalities(raw_ostream &Out, ProgramStateRef State, 1434 const char *NL = "\n", unsigned int Space = 0, 1435 bool IsDot = false) const; 1436 1437 //===------------------------------------------------------------------===// 1438 // Implementation for interface from RangedConstraintManager. 1439 //===------------------------------------------------------------------===// 1440 1441 ProgramStateRef assumeSymNE(ProgramStateRef State, SymbolRef Sym, 1442 const llvm::APSInt &V, 1443 const llvm::APSInt &Adjustment) override; 1444 1445 ProgramStateRef assumeSymEQ(ProgramStateRef State, SymbolRef Sym, 1446 const llvm::APSInt &V, 1447 const llvm::APSInt &Adjustment) override; 1448 1449 ProgramStateRef assumeSymLT(ProgramStateRef State, SymbolRef Sym, 1450 const llvm::APSInt &V, 1451 const llvm::APSInt &Adjustment) override; 1452 1453 ProgramStateRef assumeSymGT(ProgramStateRef State, SymbolRef Sym, 1454 const llvm::APSInt &V, 1455 const llvm::APSInt &Adjustment) override; 1456 1457 ProgramStateRef assumeSymLE(ProgramStateRef State, SymbolRef Sym, 1458 const llvm::APSInt &V, 1459 const llvm::APSInt &Adjustment) override; 1460 1461 ProgramStateRef assumeSymGE(ProgramStateRef State, SymbolRef Sym, 1462 const llvm::APSInt &V, 1463 const llvm::APSInt &Adjustment) override; 1464 1465 ProgramStateRef assumeSymWithinInclusiveRange( 1466 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 1467 const llvm::APSInt &To, const llvm::APSInt &Adjustment) override; 1468 1469 ProgramStateRef assumeSymOutsideInclusiveRange( 1470 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 1471 const llvm::APSInt &To, const llvm::APSInt &Adjustment) override; 1472 1473 private: 1474 RangeSet::Factory F; 1475 1476 RangeSet getRange(ProgramStateRef State, SymbolRef Sym); 1477 RangeSet getRange(ProgramStateRef State, EquivalenceClass Class); 1478 ProgramStateRef setRange(ProgramStateRef State, SymbolRef Sym, 1479 RangeSet Range); 1480 ProgramStateRef setRange(ProgramStateRef State, EquivalenceClass Class, 1481 RangeSet Range); 1482 1483 RangeSet getSymLTRange(ProgramStateRef St, SymbolRef Sym, 1484 const llvm::APSInt &Int, 1485 const llvm::APSInt &Adjustment); 1486 RangeSet getSymGTRange(ProgramStateRef St, SymbolRef Sym, 1487 const llvm::APSInt &Int, 1488 const llvm::APSInt &Adjustment); 1489 RangeSet getSymLERange(ProgramStateRef St, SymbolRef Sym, 1490 const llvm::APSInt &Int, 1491 const llvm::APSInt &Adjustment); 1492 RangeSet getSymLERange(llvm::function_ref<RangeSet()> RS, 1493 const llvm::APSInt &Int, 1494 const llvm::APSInt &Adjustment); 1495 RangeSet getSymGERange(ProgramStateRef St, SymbolRef Sym, 1496 const llvm::APSInt &Int, 1497 const llvm::APSInt &Adjustment); 1498 }; 1499 1500 //===----------------------------------------------------------------------===// 1501 // Constraint assignment logic 1502 //===----------------------------------------------------------------------===// 1503 1504 /// ConstraintAssignorBase is a small utility class that unifies visitor 1505 /// for ranges with a visitor for constraints (rangeset/range/constant). 1506 /// 1507 /// It is designed to have one derived class, but generally it can have more. 1508 /// Derived class can control which types we handle by defining methods of the 1509 /// following form: 1510 /// 1511 /// bool handle${SYMBOL}To${CONSTRAINT}(const SYMBOL *Sym, 1512 /// CONSTRAINT Constraint); 1513 /// 1514 /// where SYMBOL is the type of the symbol (e.g. SymSymExpr, SymbolCast, etc.) 1515 /// CONSTRAINT is the type of constraint (RangeSet/Range/Const) 1516 /// return value signifies whether we should try other handle methods 1517 /// (i.e. false would mean to stop right after calling this method) 1518 template <class Derived> class ConstraintAssignorBase { 1519 public: 1520 using Const = const llvm::APSInt &; 1521 1522 #define DISPATCH(CLASS) return assign##CLASS##Impl(cast<CLASS>(Sym), Constraint) 1523 1524 #define ASSIGN(CLASS, TO, SYM, CONSTRAINT) \ 1525 if (!static_cast<Derived *>(this)->assign##CLASS##To##TO(SYM, CONSTRAINT)) \ 1526 return false 1527 1528 void assign(SymbolRef Sym, RangeSet Constraint) { 1529 assignImpl(Sym, Constraint); 1530 } 1531 1532 bool assignImpl(SymbolRef Sym, RangeSet Constraint) { 1533 switch (Sym->getKind()) { 1534 #define SYMBOL(Id, Parent) \ 1535 case SymExpr::Id##Kind: \ 1536 DISPATCH(Id); 1537 #include "clang/StaticAnalyzer/Core/PathSensitive/Symbols.def" 1538 } 1539 llvm_unreachable("Unknown SymExpr kind!"); 1540 } 1541 1542 #define DEFAULT_ASSIGN(Id) \ 1543 bool assign##Id##To##RangeSet(const Id *Sym, RangeSet Constraint) { \ 1544 return true; \ 1545 } \ 1546 bool assign##Id##To##Range(const Id *Sym, Range Constraint) { return true; } \ 1547 bool assign##Id##To##Const(const Id *Sym, Const Constraint) { return true; } 1548 1549 // When we dispatch for constraint types, we first try to check 1550 // if the new constraint is the constant and try the corresponding 1551 // assignor methods. If it didn't interrupt, we can proceed to the 1552 // range, and finally to the range set. 1553 #define CONSTRAINT_DISPATCH(Id) \ 1554 if (const llvm::APSInt *Const = Constraint.getConcreteValue()) { \ 1555 ASSIGN(Id, Const, Sym, *Const); \ 1556 } \ 1557 if (Constraint.size() == 1) { \ 1558 ASSIGN(Id, Range, Sym, *Constraint.begin()); \ 1559 } \ 1560 ASSIGN(Id, RangeSet, Sym, Constraint) 1561 1562 // Our internal assign method first tries to call assignor methods for all 1563 // constraint types that apply. And if not interrupted, continues with its 1564 // parent class. 1565 #define SYMBOL(Id, Parent) \ 1566 bool assign##Id##Impl(const Id *Sym, RangeSet Constraint) { \ 1567 CONSTRAINT_DISPATCH(Id); \ 1568 DISPATCH(Parent); \ 1569 } \ 1570 DEFAULT_ASSIGN(Id) 1571 #define ABSTRACT_SYMBOL(Id, Parent) SYMBOL(Id, Parent) 1572 #include "clang/StaticAnalyzer/Core/PathSensitive/Symbols.def" 1573 1574 // Default implementations for the top class that doesn't have parents. 1575 bool assignSymExprImpl(const SymExpr *Sym, RangeSet Constraint) { 1576 CONSTRAINT_DISPATCH(SymExpr); 1577 return true; 1578 } 1579 DEFAULT_ASSIGN(SymExpr); 1580 1581 #undef DISPATCH 1582 #undef CONSTRAINT_DISPATCH 1583 #undef DEFAULT_ASSIGN 1584 #undef ASSIGN 1585 }; 1586 1587 /// A little component aggregating all of the reasoning we have about 1588 /// assigning new constraints to symbols. 1589 /// 1590 /// The main purpose of this class is to associate constraints to symbols, 1591 /// and impose additional constraints on other symbols, when we can imply 1592 /// them. 1593 /// 1594 /// It has a nice symmetry with SymbolicRangeInferrer. When the latter 1595 /// can provide more precise ranges by looking into the operands of the 1596 /// expression in question, ConstraintAssignor looks into the operands 1597 /// to see if we can imply more from the new constraint. 1598 class ConstraintAssignor : public ConstraintAssignorBase<ConstraintAssignor> { 1599 public: 1600 template <class ClassOrSymbol> 1601 LLVM_NODISCARD static ProgramStateRef 1602 assign(ProgramStateRef State, RangeConstraintManager &RCM, 1603 SValBuilder &Builder, RangeSet::Factory &F, ClassOrSymbol CoS, 1604 RangeSet NewConstraint) { 1605 if (!State || NewConstraint.isEmpty()) 1606 return nullptr; 1607 1608 ConstraintAssignor Assignor{State, RCM, Builder, F}; 1609 return Assignor.assign(CoS, NewConstraint); 1610 } 1611 1612 /// Handle expressions like: a % b != 0. 1613 template <typename SymT> 1614 bool handleRemainderOp(const SymT *Sym, RangeSet Constraint) { 1615 if (Sym->getOpcode() != BO_Rem) 1616 return true; 1617 // a % b != 0 implies that a != 0. 1618 if (!Constraint.containsZero()) { 1619 SVal SymSVal = Builder.makeSymbolVal(Sym->getLHS()); 1620 if (auto NonLocSymSVal = SymSVal.getAs<nonloc::SymbolVal>()) { 1621 State = State->assume(*NonLocSymSVal, true); 1622 if (!State) 1623 return false; 1624 } 1625 } 1626 return true; 1627 } 1628 1629 inline bool assignSymExprToConst(const SymExpr *Sym, Const Constraint); 1630 inline bool assignSymIntExprToRangeSet(const SymIntExpr *Sym, 1631 RangeSet Constraint) { 1632 return handleRemainderOp(Sym, Constraint); 1633 } 1634 inline bool assignSymSymExprToRangeSet(const SymSymExpr *Sym, 1635 RangeSet Constraint); 1636 1637 private: 1638 ConstraintAssignor(ProgramStateRef State, RangeConstraintManager &RCM, 1639 SValBuilder &Builder, RangeSet::Factory &F) 1640 : State(State), RCM(RCM), Builder(Builder), RangeFactory(F) {} 1641 using Base = ConstraintAssignorBase<ConstraintAssignor>; 1642 1643 /// Base method for handling new constraints for symbols. 1644 LLVM_NODISCARD ProgramStateRef assign(SymbolRef Sym, RangeSet NewConstraint) { 1645 // All constraints are actually associated with equivalence classes, and 1646 // that's what we are going to do first. 1647 State = assign(EquivalenceClass::find(State, Sym), NewConstraint); 1648 if (!State) 1649 return nullptr; 1650 1651 // And after that we can check what other things we can get from this 1652 // constraint. 1653 Base::assign(Sym, NewConstraint); 1654 return State; 1655 } 1656 1657 /// Base method for handling new constraints for classes. 1658 LLVM_NODISCARD ProgramStateRef assign(EquivalenceClass Class, 1659 RangeSet NewConstraint) { 1660 // There is a chance that we might need to update constraints for the 1661 // classes that are known to be disequal to Class. 1662 // 1663 // In order for this to be even possible, the new constraint should 1664 // be simply a constant because we can't reason about range disequalities. 1665 if (const llvm::APSInt *Point = NewConstraint.getConcreteValue()) { 1666 1667 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1668 ConstraintRangeTy::Factory &CF = State->get_context<ConstraintRange>(); 1669 1670 // Add new constraint. 1671 Constraints = CF.add(Constraints, Class, NewConstraint); 1672 1673 for (EquivalenceClass DisequalClass : Class.getDisequalClasses(State)) { 1674 RangeSet UpdatedConstraint = SymbolicRangeInferrer::inferRange( 1675 RangeFactory, State, DisequalClass); 1676 1677 UpdatedConstraint = RangeFactory.deletePoint(UpdatedConstraint, *Point); 1678 1679 // If we end up with at least one of the disequal classes to be 1680 // constrained with an empty range-set, the state is infeasible. 1681 if (UpdatedConstraint.isEmpty()) 1682 return nullptr; 1683 1684 Constraints = CF.add(Constraints, DisequalClass, UpdatedConstraint); 1685 } 1686 assert(areFeasible(Constraints) && "Constraint manager shouldn't produce " 1687 "a state with infeasible constraints"); 1688 1689 return setConstraints(State, Constraints); 1690 } 1691 1692 return setConstraint(State, Class, NewConstraint); 1693 } 1694 1695 ProgramStateRef trackDisequality(ProgramStateRef State, SymbolRef LHS, 1696 SymbolRef RHS) { 1697 return EquivalenceClass::markDisequal(RangeFactory, State, LHS, RHS); 1698 } 1699 1700 ProgramStateRef trackEquality(ProgramStateRef State, SymbolRef LHS, 1701 SymbolRef RHS) { 1702 return EquivalenceClass::merge(RangeFactory, State, LHS, RHS); 1703 } 1704 1705 LLVM_NODISCARD Optional<bool> interpreteAsBool(RangeSet Constraint) { 1706 assert(!Constraint.isEmpty() && "Empty ranges shouldn't get here"); 1707 1708 if (Constraint.getConcreteValue()) 1709 return !Constraint.getConcreteValue()->isZero(); 1710 1711 if (!Constraint.containsZero()) 1712 return true; 1713 1714 return llvm::None; 1715 } 1716 1717 ProgramStateRef State; 1718 RangeConstraintManager &RCM; 1719 SValBuilder &Builder; 1720 RangeSet::Factory &RangeFactory; 1721 }; 1722 1723 1724 bool ConstraintAssignor::assignSymExprToConst(const SymExpr *Sym, 1725 const llvm::APSInt &Constraint) { 1726 llvm::SmallSet<EquivalenceClass, 4> SimplifiedClasses; 1727 // Iterate over all equivalence classes and try to simplify them. 1728 ClassMembersTy Members = State->get<ClassMembers>(); 1729 for (std::pair<EquivalenceClass, SymbolSet> ClassToSymbolSet : Members) { 1730 EquivalenceClass Class = ClassToSymbolSet.first; 1731 State = 1732 EquivalenceClass::simplify(Builder, RangeFactory, RCM, 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 = 1747 EquivalenceClass::simplify(Builder, RangeFactory, RCM, State, Class); 1748 if (!State) 1749 return false; 1750 } 1751 1752 return true; 1753 } 1754 1755 bool ConstraintAssignor::assignSymSymExprToRangeSet(const SymSymExpr *Sym, 1756 RangeSet Constraint) { 1757 if (!handleRemainderOp(Sym, Constraint)) 1758 return false; 1759 1760 Optional<bool> ConstraintAsBool = interpreteAsBool(Constraint); 1761 1762 if (!ConstraintAsBool) 1763 return true; 1764 1765 if (Optional<bool> Equality = meansEquality(Sym)) { 1766 // Here we cover two cases: 1767 // * if Sym is equality and the new constraint is true -> Sym's operands 1768 // should be marked as equal 1769 // * if Sym is disequality and the new constraint is false -> Sym's 1770 // operands should be also marked as equal 1771 if (*Equality == *ConstraintAsBool) { 1772 State = trackEquality(State, Sym->getLHS(), Sym->getRHS()); 1773 } else { 1774 // Other combinations leave as with disequal operands. 1775 State = trackDisequality(State, Sym->getLHS(), Sym->getRHS()); 1776 } 1777 1778 if (!State) 1779 return false; 1780 } 1781 1782 return true; 1783 } 1784 1785 } // end anonymous namespace 1786 1787 std::unique_ptr<ConstraintManager> 1788 ento::CreateRangeConstraintManager(ProgramStateManager &StMgr, 1789 ExprEngine *Eng) { 1790 return std::make_unique<RangeConstraintManager>(Eng, StMgr.getSValBuilder()); 1791 } 1792 1793 ConstraintMap ento::getConstraintMap(ProgramStateRef State) { 1794 ConstraintMap::Factory &F = State->get_context<ConstraintMap>(); 1795 ConstraintMap Result = F.getEmptyMap(); 1796 1797 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1798 for (std::pair<EquivalenceClass, RangeSet> ClassConstraint : Constraints) { 1799 EquivalenceClass Class = ClassConstraint.first; 1800 SymbolSet ClassMembers = Class.getClassMembers(State); 1801 assert(!ClassMembers.isEmpty() && 1802 "Class must always have at least one member!"); 1803 1804 SymbolRef Representative = *ClassMembers.begin(); 1805 Result = F.add(Result, Representative, ClassConstraint.second); 1806 } 1807 1808 return Result; 1809 } 1810 1811 //===----------------------------------------------------------------------===// 1812 // EqualityClass implementation details 1813 //===----------------------------------------------------------------------===// 1814 1815 LLVM_DUMP_METHOD void EquivalenceClass::dumpToStream(ProgramStateRef State, 1816 raw_ostream &os) const { 1817 SymbolSet ClassMembers = getClassMembers(State); 1818 for (const SymbolRef &MemberSym : ClassMembers) { 1819 MemberSym->dump(); 1820 os << "\n"; 1821 } 1822 } 1823 1824 inline EquivalenceClass EquivalenceClass::find(ProgramStateRef State, 1825 SymbolRef Sym) { 1826 assert(State && "State should not be null"); 1827 assert(Sym && "Symbol should not be null"); 1828 // We store far from all Symbol -> Class mappings 1829 if (const EquivalenceClass *NontrivialClass = State->get<ClassMap>(Sym)) 1830 return *NontrivialClass; 1831 1832 // This is a trivial class of Sym. 1833 return Sym; 1834 } 1835 1836 inline ProgramStateRef EquivalenceClass::merge(RangeSet::Factory &F, 1837 ProgramStateRef State, 1838 SymbolRef First, 1839 SymbolRef Second) { 1840 EquivalenceClass FirstClass = find(State, First); 1841 EquivalenceClass SecondClass = find(State, Second); 1842 1843 return FirstClass.merge(F, State, SecondClass); 1844 } 1845 1846 inline ProgramStateRef EquivalenceClass::merge(RangeSet::Factory &F, 1847 ProgramStateRef State, 1848 EquivalenceClass Other) { 1849 // It is already the same class. 1850 if (*this == Other) 1851 return State; 1852 1853 // FIXME: As of now, we support only equivalence classes of the same type. 1854 // This limitation is connected to the lack of explicit casts in 1855 // our symbolic expression model. 1856 // 1857 // That means that for `int x` and `char y` we don't distinguish 1858 // between these two very different cases: 1859 // * `x == y` 1860 // * `(char)x == y` 1861 // 1862 // The moment we introduce symbolic casts, this restriction can be 1863 // lifted. 1864 if (getType() != Other.getType()) 1865 return State; 1866 1867 SymbolSet Members = getClassMembers(State); 1868 SymbolSet OtherMembers = Other.getClassMembers(State); 1869 1870 // We estimate the size of the class by the height of tree containing 1871 // its members. Merging is not a trivial operation, so it's easier to 1872 // merge the smaller class into the bigger one. 1873 if (Members.getHeight() >= OtherMembers.getHeight()) { 1874 return mergeImpl(F, State, Members, Other, OtherMembers); 1875 } else { 1876 return Other.mergeImpl(F, State, OtherMembers, *this, Members); 1877 } 1878 } 1879 1880 inline ProgramStateRef 1881 EquivalenceClass::mergeImpl(RangeSet::Factory &RangeFactory, 1882 ProgramStateRef State, SymbolSet MyMembers, 1883 EquivalenceClass Other, SymbolSet OtherMembers) { 1884 // Essentially what we try to recreate here is some kind of union-find 1885 // data structure. It does have certain limitations due to persistence 1886 // and the need to remove elements from classes. 1887 // 1888 // In this setting, EquialityClass object is the representative of the class 1889 // or the parent element. ClassMap is a mapping of class members to their 1890 // parent. Unlike the union-find structure, they all point directly to the 1891 // class representative because we don't have an opportunity to actually do 1892 // path compression when dealing with immutability. This means that we 1893 // compress paths every time we do merges. It also means that we lose 1894 // the main amortized complexity benefit from the original data structure. 1895 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1896 ConstraintRangeTy::Factory &CRF = State->get_context<ConstraintRange>(); 1897 1898 // 1. If the merged classes have any constraints associated with them, we 1899 // need to transfer them to the class we have left. 1900 // 1901 // Intersection here makes perfect sense because both of these constraints 1902 // must hold for the whole new class. 1903 if (Optional<RangeSet> NewClassConstraint = 1904 intersect(RangeFactory, getConstraint(State, *this), 1905 getConstraint(State, Other))) { 1906 // NOTE: Essentially, NewClassConstraint should NEVER be infeasible because 1907 // range inferrer shouldn't generate ranges incompatible with 1908 // equivalence classes. However, at the moment, due to imperfections 1909 // in the solver, it is possible and the merge function can also 1910 // return infeasible states aka null states. 1911 if (NewClassConstraint->isEmpty()) 1912 // Infeasible state 1913 return nullptr; 1914 1915 // No need in tracking constraints of a now-dissolved class. 1916 Constraints = CRF.remove(Constraints, Other); 1917 // Assign new constraints for this class. 1918 Constraints = CRF.add(Constraints, *this, *NewClassConstraint); 1919 1920 assert(areFeasible(Constraints) && "Constraint manager shouldn't produce " 1921 "a state with infeasible constraints"); 1922 1923 State = State->set<ConstraintRange>(Constraints); 1924 } 1925 1926 // 2. Get ALL equivalence-related maps 1927 ClassMapTy Classes = State->get<ClassMap>(); 1928 ClassMapTy::Factory &CMF = State->get_context<ClassMap>(); 1929 1930 ClassMembersTy Members = State->get<ClassMembers>(); 1931 ClassMembersTy::Factory &MF = State->get_context<ClassMembers>(); 1932 1933 DisequalityMapTy DisequalityInfo = State->get<DisequalityMap>(); 1934 DisequalityMapTy::Factory &DF = State->get_context<DisequalityMap>(); 1935 1936 ClassSet::Factory &CF = State->get_context<ClassSet>(); 1937 SymbolSet::Factory &F = getMembersFactory(State); 1938 1939 // 2. Merge members of the Other class into the current class. 1940 SymbolSet NewClassMembers = MyMembers; 1941 for (SymbolRef Sym : OtherMembers) { 1942 NewClassMembers = F.add(NewClassMembers, Sym); 1943 // *this is now the class for all these new symbols. 1944 Classes = CMF.add(Classes, Sym, *this); 1945 } 1946 1947 // 3. Adjust member mapping. 1948 // 1949 // No need in tracking members of a now-dissolved class. 1950 Members = MF.remove(Members, Other); 1951 // Now only the current class is mapped to all the symbols. 1952 Members = MF.add(Members, *this, NewClassMembers); 1953 1954 // 4. Update disequality relations 1955 ClassSet DisequalToOther = Other.getDisequalClasses(DisequalityInfo, CF); 1956 // We are about to merge two classes but they are already known to be 1957 // non-equal. This is a contradiction. 1958 if (DisequalToOther.contains(*this)) 1959 return nullptr; 1960 1961 if (!DisequalToOther.isEmpty()) { 1962 ClassSet DisequalToThis = getDisequalClasses(DisequalityInfo, CF); 1963 DisequalityInfo = DF.remove(DisequalityInfo, Other); 1964 1965 for (EquivalenceClass DisequalClass : DisequalToOther) { 1966 DisequalToThis = CF.add(DisequalToThis, DisequalClass); 1967 1968 // Disequality is a symmetric relation meaning that if 1969 // DisequalToOther not null then the set for DisequalClass is not 1970 // empty and has at least Other. 1971 ClassSet OriginalSetLinkedToOther = 1972 *DisequalityInfo.lookup(DisequalClass); 1973 1974 // Other will be eliminated and we should replace it with the bigger 1975 // united class. 1976 ClassSet NewSet = CF.remove(OriginalSetLinkedToOther, Other); 1977 NewSet = CF.add(NewSet, *this); 1978 1979 DisequalityInfo = DF.add(DisequalityInfo, DisequalClass, NewSet); 1980 } 1981 1982 DisequalityInfo = DF.add(DisequalityInfo, *this, DisequalToThis); 1983 State = State->set<DisequalityMap>(DisequalityInfo); 1984 } 1985 1986 // 5. Update the state 1987 State = State->set<ClassMap>(Classes); 1988 State = State->set<ClassMembers>(Members); 1989 1990 return State; 1991 } 1992 1993 inline SymbolSet::Factory & 1994 EquivalenceClass::getMembersFactory(ProgramStateRef State) { 1995 return State->get_context<SymbolSet>(); 1996 } 1997 1998 SymbolSet EquivalenceClass::getClassMembers(ProgramStateRef State) const { 1999 if (const SymbolSet *Members = State->get<ClassMembers>(*this)) 2000 return *Members; 2001 2002 // This class is trivial, so we need to construct a set 2003 // with just that one symbol from the class. 2004 SymbolSet::Factory &F = getMembersFactory(State); 2005 return F.add(F.getEmptySet(), getRepresentativeSymbol()); 2006 } 2007 2008 bool EquivalenceClass::isTrivial(ProgramStateRef State) const { 2009 return State->get<ClassMembers>(*this) == nullptr; 2010 } 2011 2012 bool EquivalenceClass::isTriviallyDead(ProgramStateRef State, 2013 SymbolReaper &Reaper) const { 2014 return isTrivial(State) && Reaper.isDead(getRepresentativeSymbol()); 2015 } 2016 2017 inline ProgramStateRef EquivalenceClass::markDisequal(RangeSet::Factory &RF, 2018 ProgramStateRef State, 2019 SymbolRef First, 2020 SymbolRef Second) { 2021 return markDisequal(RF, State, find(State, First), find(State, Second)); 2022 } 2023 2024 inline ProgramStateRef EquivalenceClass::markDisequal(RangeSet::Factory &RF, 2025 ProgramStateRef State, 2026 EquivalenceClass First, 2027 EquivalenceClass Second) { 2028 return First.markDisequal(RF, State, Second); 2029 } 2030 2031 inline ProgramStateRef 2032 EquivalenceClass::markDisequal(RangeSet::Factory &RF, ProgramStateRef State, 2033 EquivalenceClass Other) const { 2034 // If we know that two classes are equal, we can only produce an infeasible 2035 // state. 2036 if (*this == Other) { 2037 return nullptr; 2038 } 2039 2040 DisequalityMapTy DisequalityInfo = State->get<DisequalityMap>(); 2041 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 2042 2043 // Disequality is a symmetric relation, so if we mark A as disequal to B, 2044 // we should also mark B as disequalt to A. 2045 if (!addToDisequalityInfo(DisequalityInfo, Constraints, RF, State, *this, 2046 Other) || 2047 !addToDisequalityInfo(DisequalityInfo, Constraints, RF, State, Other, 2048 *this)) 2049 return nullptr; 2050 2051 assert(areFeasible(Constraints) && "Constraint manager shouldn't produce " 2052 "a state with infeasible constraints"); 2053 2054 State = State->set<DisequalityMap>(DisequalityInfo); 2055 State = State->set<ConstraintRange>(Constraints); 2056 2057 return State; 2058 } 2059 2060 inline bool EquivalenceClass::addToDisequalityInfo( 2061 DisequalityMapTy &Info, ConstraintRangeTy &Constraints, 2062 RangeSet::Factory &RF, ProgramStateRef State, EquivalenceClass First, 2063 EquivalenceClass Second) { 2064 2065 // 1. Get all of the required factories. 2066 DisequalityMapTy::Factory &F = State->get_context<DisequalityMap>(); 2067 ClassSet::Factory &CF = State->get_context<ClassSet>(); 2068 ConstraintRangeTy::Factory &CRF = State->get_context<ConstraintRange>(); 2069 2070 // 2. Add Second to the set of classes disequal to First. 2071 const ClassSet *CurrentSet = Info.lookup(First); 2072 ClassSet NewSet = CurrentSet ? *CurrentSet : CF.getEmptySet(); 2073 NewSet = CF.add(NewSet, Second); 2074 2075 Info = F.add(Info, First, NewSet); 2076 2077 // 3. If Second is known to be a constant, we can delete this point 2078 // from the constraint asociated with First. 2079 // 2080 // So, if Second == 10, it means that First != 10. 2081 // At the same time, the same logic does not apply to ranges. 2082 if (const RangeSet *SecondConstraint = Constraints.lookup(Second)) 2083 if (const llvm::APSInt *Point = SecondConstraint->getConcreteValue()) { 2084 2085 RangeSet FirstConstraint = SymbolicRangeInferrer::inferRange( 2086 RF, State, First.getRepresentativeSymbol()); 2087 2088 FirstConstraint = RF.deletePoint(FirstConstraint, *Point); 2089 2090 // If the First class is about to be constrained with an empty 2091 // range-set, the state is infeasible. 2092 if (FirstConstraint.isEmpty()) 2093 return false; 2094 2095 Constraints = CRF.add(Constraints, First, FirstConstraint); 2096 } 2097 2098 return true; 2099 } 2100 2101 inline Optional<bool> EquivalenceClass::areEqual(ProgramStateRef State, 2102 SymbolRef FirstSym, 2103 SymbolRef SecondSym) { 2104 return EquivalenceClass::areEqual(State, find(State, FirstSym), 2105 find(State, SecondSym)); 2106 } 2107 2108 inline Optional<bool> EquivalenceClass::areEqual(ProgramStateRef State, 2109 EquivalenceClass First, 2110 EquivalenceClass Second) { 2111 // The same equivalence class => symbols are equal. 2112 if (First == Second) 2113 return true; 2114 2115 // Let's check if we know anything about these two classes being not equal to 2116 // each other. 2117 ClassSet DisequalToFirst = First.getDisequalClasses(State); 2118 if (DisequalToFirst.contains(Second)) 2119 return false; 2120 2121 // It is not clear. 2122 return llvm::None; 2123 } 2124 2125 // Iterate over all symbols and try to simplify them. Once a symbol is 2126 // simplified then we check if we can merge the simplified symbol's equivalence 2127 // class to this class. This way, we simplify not just the symbols but the 2128 // classes as well: we strive to keep the number of the classes to be the 2129 // absolute minimum. 2130 LLVM_NODISCARD ProgramStateRef EquivalenceClass::simplify( 2131 SValBuilder &SVB, RangeSet::Factory &F, RangedConstraintManager &RCM, 2132 ProgramStateRef State, EquivalenceClass Class) { 2133 SymbolSet ClassMembers = Class.getClassMembers(State); 2134 for (const SymbolRef &MemberSym : ClassMembers) { 2135 2136 const SVal SimplifiedMemberVal = simplifyToSVal(State, MemberSym); 2137 const SymbolRef SimplifiedMemberSym = SimplifiedMemberVal.getAsSymbol(); 2138 2139 // The symbol is collapsed to a constant, check if the current State is 2140 // still feasible. 2141 if (const auto CI = SimplifiedMemberVal.getAs<nonloc::ConcreteInt>()) { 2142 const llvm::APSInt &SV = CI->getValue(); 2143 const RangeSet *ClassConstraint = getConstraint(State, Class); 2144 // We have found a contradiction. 2145 if (ClassConstraint && !ClassConstraint->contains(SV)) 2146 return nullptr; 2147 } 2148 2149 if (SimplifiedMemberSym && MemberSym != SimplifiedMemberSym) { 2150 // The simplified symbol should be the member of the original Class, 2151 // however, it might be in another existing class at the moment. We 2152 // have to merge these classes. 2153 ProgramStateRef OldState = State; 2154 State = merge(F, State, MemberSym, SimplifiedMemberSym); 2155 if (!State) 2156 return nullptr; 2157 // No state change, no merge happened actually. 2158 if (OldState == State) 2159 continue; 2160 2161 // Initiate the reorganization of the equality information. E.g., if we 2162 // have `c + 1 == 0` then we'd like to express that `c == -1`. It makes 2163 // sense to do this only with `SymIntExpr`s. 2164 // TODO Handle `IntSymExpr` as well, once computeAdjustment can handle 2165 // them. 2166 if (const SymIntExpr *SIE = dyn_cast<SymIntExpr>(SimplifiedMemberSym)) { 2167 if (const RangeSet *ClassConstraint = getConstraint(State, Class)) { 2168 // Overestimate the individual Ranges with the RangeSet' lowest and 2169 // highest values. 2170 State = RCM.assumeSymInclusiveRange( 2171 State, SIE, ClassConstraint->getMinValue(), 2172 ClassConstraint->getMaxValue(), /*InRange=*/true); 2173 if (!State) 2174 return nullptr; 2175 } 2176 } 2177 } 2178 } 2179 return State; 2180 } 2181 2182 inline ClassSet EquivalenceClass::getDisequalClasses(ProgramStateRef State, 2183 SymbolRef Sym) { 2184 return find(State, Sym).getDisequalClasses(State); 2185 } 2186 2187 inline ClassSet 2188 EquivalenceClass::getDisequalClasses(ProgramStateRef State) const { 2189 return getDisequalClasses(State->get<DisequalityMap>(), 2190 State->get_context<ClassSet>()); 2191 } 2192 2193 inline ClassSet 2194 EquivalenceClass::getDisequalClasses(DisequalityMapTy Map, 2195 ClassSet::Factory &Factory) const { 2196 if (const ClassSet *DisequalClasses = Map.lookup(*this)) 2197 return *DisequalClasses; 2198 2199 return Factory.getEmptySet(); 2200 } 2201 2202 bool EquivalenceClass::isClassDataConsistent(ProgramStateRef State) { 2203 ClassMembersTy Members = State->get<ClassMembers>(); 2204 2205 for (std::pair<EquivalenceClass, SymbolSet> ClassMembersPair : Members) { 2206 for (SymbolRef Member : ClassMembersPair.second) { 2207 // Every member of the class should have a mapping back to the class. 2208 if (find(State, Member) == ClassMembersPair.first) { 2209 continue; 2210 } 2211 2212 return false; 2213 } 2214 } 2215 2216 DisequalityMapTy Disequalities = State->get<DisequalityMap>(); 2217 for (std::pair<EquivalenceClass, ClassSet> DisequalityInfo : Disequalities) { 2218 EquivalenceClass Class = DisequalityInfo.first; 2219 ClassSet DisequalClasses = DisequalityInfo.second; 2220 2221 // There is no use in keeping empty sets in the map. 2222 if (DisequalClasses.isEmpty()) 2223 return false; 2224 2225 // Disequality is symmetrical, i.e. for every Class A and B that A != B, 2226 // B != A should also be true. 2227 for (EquivalenceClass DisequalClass : DisequalClasses) { 2228 const ClassSet *DisequalToDisequalClasses = 2229 Disequalities.lookup(DisequalClass); 2230 2231 // It should be a set of at least one element: Class 2232 if (!DisequalToDisequalClasses || 2233 !DisequalToDisequalClasses->contains(Class)) 2234 return false; 2235 } 2236 } 2237 2238 return true; 2239 } 2240 2241 //===----------------------------------------------------------------------===// 2242 // RangeConstraintManager implementation 2243 //===----------------------------------------------------------------------===// 2244 2245 bool RangeConstraintManager::canReasonAbout(SVal X) const { 2246 Optional<nonloc::SymbolVal> SymVal = X.getAs<nonloc::SymbolVal>(); 2247 if (SymVal && SymVal->isExpression()) { 2248 const SymExpr *SE = SymVal->getSymbol(); 2249 2250 if (const SymIntExpr *SIE = dyn_cast<SymIntExpr>(SE)) { 2251 switch (SIE->getOpcode()) { 2252 // We don't reason yet about bitwise-constraints on symbolic values. 2253 case BO_And: 2254 case BO_Or: 2255 case BO_Xor: 2256 return false; 2257 // We don't reason yet about these arithmetic constraints on 2258 // symbolic values. 2259 case BO_Mul: 2260 case BO_Div: 2261 case BO_Rem: 2262 case BO_Shl: 2263 case BO_Shr: 2264 return false; 2265 // All other cases. 2266 default: 2267 return true; 2268 } 2269 } 2270 2271 if (const SymSymExpr *SSE = dyn_cast<SymSymExpr>(SE)) { 2272 // FIXME: Handle <=> here. 2273 if (BinaryOperator::isEqualityOp(SSE->getOpcode()) || 2274 BinaryOperator::isRelationalOp(SSE->getOpcode())) { 2275 // We handle Loc <> Loc comparisons, but not (yet) NonLoc <> NonLoc. 2276 // We've recently started producing Loc <> NonLoc comparisons (that 2277 // result from casts of one of the operands between eg. intptr_t and 2278 // void *), but we can't reason about them yet. 2279 if (Loc::isLocType(SSE->getLHS()->getType())) { 2280 return Loc::isLocType(SSE->getRHS()->getType()); 2281 } 2282 } 2283 } 2284 2285 return false; 2286 } 2287 2288 return true; 2289 } 2290 2291 ConditionTruthVal RangeConstraintManager::checkNull(ProgramStateRef State, 2292 SymbolRef Sym) { 2293 const RangeSet *Ranges = getConstraint(State, Sym); 2294 2295 // If we don't have any information about this symbol, it's underconstrained. 2296 if (!Ranges) 2297 return ConditionTruthVal(); 2298 2299 // If we have a concrete value, see if it's zero. 2300 if (const llvm::APSInt *Value = Ranges->getConcreteValue()) 2301 return *Value == 0; 2302 2303 BasicValueFactory &BV = getBasicVals(); 2304 APSIntType IntType = BV.getAPSIntType(Sym->getType()); 2305 llvm::APSInt Zero = IntType.getZeroValue(); 2306 2307 // Check if zero is in the set of possible values. 2308 if (!Ranges->contains(Zero)) 2309 return false; 2310 2311 // Zero is a possible value, but it is not the /only/ possible value. 2312 return ConditionTruthVal(); 2313 } 2314 2315 const llvm::APSInt *RangeConstraintManager::getSymVal(ProgramStateRef St, 2316 SymbolRef Sym) const { 2317 const RangeSet *T = getConstraint(St, Sym); 2318 return T ? T->getConcreteValue() : nullptr; 2319 } 2320 2321 //===----------------------------------------------------------------------===// 2322 // Remove dead symbols from existing constraints 2323 //===----------------------------------------------------------------------===// 2324 2325 /// Scan all symbols referenced by the constraints. If the symbol is not alive 2326 /// as marked in LSymbols, mark it as dead in DSymbols. 2327 ProgramStateRef 2328 RangeConstraintManager::removeDeadBindings(ProgramStateRef State, 2329 SymbolReaper &SymReaper) { 2330 ClassMembersTy ClassMembersMap = State->get<ClassMembers>(); 2331 ClassMembersTy NewClassMembersMap = ClassMembersMap; 2332 ClassMembersTy::Factory &EMFactory = State->get_context<ClassMembers>(); 2333 SymbolSet::Factory &SetFactory = State->get_context<SymbolSet>(); 2334 2335 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 2336 ConstraintRangeTy NewConstraints = Constraints; 2337 ConstraintRangeTy::Factory &ConstraintFactory = 2338 State->get_context<ConstraintRange>(); 2339 2340 ClassMapTy Map = State->get<ClassMap>(); 2341 ClassMapTy NewMap = Map; 2342 ClassMapTy::Factory &ClassFactory = State->get_context<ClassMap>(); 2343 2344 DisequalityMapTy Disequalities = State->get<DisequalityMap>(); 2345 DisequalityMapTy::Factory &DisequalityFactory = 2346 State->get_context<DisequalityMap>(); 2347 ClassSet::Factory &ClassSetFactory = State->get_context<ClassSet>(); 2348 2349 bool ClassMapChanged = false; 2350 bool MembersMapChanged = false; 2351 bool ConstraintMapChanged = false; 2352 bool DisequalitiesChanged = false; 2353 2354 auto removeDeadClass = [&](EquivalenceClass Class) { 2355 // Remove associated constraint ranges. 2356 Constraints = ConstraintFactory.remove(Constraints, Class); 2357 ConstraintMapChanged = true; 2358 2359 // Update disequality information to not hold any information on the 2360 // removed class. 2361 ClassSet DisequalClasses = 2362 Class.getDisequalClasses(Disequalities, ClassSetFactory); 2363 if (!DisequalClasses.isEmpty()) { 2364 for (EquivalenceClass DisequalClass : DisequalClasses) { 2365 ClassSet DisequalToDisequalSet = 2366 DisequalClass.getDisequalClasses(Disequalities, ClassSetFactory); 2367 // DisequalToDisequalSet is guaranteed to be non-empty for consistent 2368 // disequality info. 2369 assert(!DisequalToDisequalSet.isEmpty()); 2370 ClassSet NewSet = ClassSetFactory.remove(DisequalToDisequalSet, Class); 2371 2372 // No need in keeping an empty set. 2373 if (NewSet.isEmpty()) { 2374 Disequalities = 2375 DisequalityFactory.remove(Disequalities, DisequalClass); 2376 } else { 2377 Disequalities = 2378 DisequalityFactory.add(Disequalities, DisequalClass, NewSet); 2379 } 2380 } 2381 // Remove the data for the class 2382 Disequalities = DisequalityFactory.remove(Disequalities, Class); 2383 DisequalitiesChanged = true; 2384 } 2385 }; 2386 2387 // 1. Let's see if dead symbols are trivial and have associated constraints. 2388 for (std::pair<EquivalenceClass, RangeSet> ClassConstraintPair : 2389 Constraints) { 2390 EquivalenceClass Class = ClassConstraintPair.first; 2391 if (Class.isTriviallyDead(State, SymReaper)) { 2392 // If this class is trivial, we can remove its constraints right away. 2393 removeDeadClass(Class); 2394 } 2395 } 2396 2397 // 2. We don't need to track classes for dead symbols. 2398 for (std::pair<SymbolRef, EquivalenceClass> SymbolClassPair : Map) { 2399 SymbolRef Sym = SymbolClassPair.first; 2400 2401 if (SymReaper.isDead(Sym)) { 2402 ClassMapChanged = true; 2403 NewMap = ClassFactory.remove(NewMap, Sym); 2404 } 2405 } 2406 2407 // 3. Remove dead members from classes and remove dead non-trivial classes 2408 // and their constraints. 2409 for (std::pair<EquivalenceClass, SymbolSet> ClassMembersPair : 2410 ClassMembersMap) { 2411 EquivalenceClass Class = ClassMembersPair.first; 2412 SymbolSet LiveMembers = ClassMembersPair.second; 2413 bool MembersChanged = false; 2414 2415 for (SymbolRef Member : ClassMembersPair.second) { 2416 if (SymReaper.isDead(Member)) { 2417 MembersChanged = true; 2418 LiveMembers = SetFactory.remove(LiveMembers, Member); 2419 } 2420 } 2421 2422 // Check if the class changed. 2423 if (!MembersChanged) 2424 continue; 2425 2426 MembersMapChanged = true; 2427 2428 if (LiveMembers.isEmpty()) { 2429 // The class is dead now, we need to wipe it out of the members map... 2430 NewClassMembersMap = EMFactory.remove(NewClassMembersMap, Class); 2431 2432 // ...and remove all of its constraints. 2433 removeDeadClass(Class); 2434 } else { 2435 // We need to change the members associated with the class. 2436 NewClassMembersMap = 2437 EMFactory.add(NewClassMembersMap, Class, LiveMembers); 2438 } 2439 } 2440 2441 // 4. Update the state with new maps. 2442 // 2443 // Here we try to be humble and update a map only if it really changed. 2444 if (ClassMapChanged) 2445 State = State->set<ClassMap>(NewMap); 2446 2447 if (MembersMapChanged) 2448 State = State->set<ClassMembers>(NewClassMembersMap); 2449 2450 if (ConstraintMapChanged) 2451 State = State->set<ConstraintRange>(Constraints); 2452 2453 if (DisequalitiesChanged) 2454 State = State->set<DisequalityMap>(Disequalities); 2455 2456 assert(EquivalenceClass::isClassDataConsistent(State)); 2457 2458 return State; 2459 } 2460 2461 RangeSet RangeConstraintManager::getRange(ProgramStateRef State, 2462 SymbolRef Sym) { 2463 return SymbolicRangeInferrer::inferRange(F, State, Sym); 2464 } 2465 2466 ProgramStateRef RangeConstraintManager::setRange(ProgramStateRef State, 2467 SymbolRef Sym, 2468 RangeSet Range) { 2469 return ConstraintAssignor::assign(State, *this, getSValBuilder(), F, Sym, 2470 Range); 2471 } 2472 2473 //===------------------------------------------------------------------------=== 2474 // assumeSymX methods: protected interface for RangeConstraintManager. 2475 //===------------------------------------------------------------------------===/ 2476 2477 // The syntax for ranges below is mathematical, using [x, y] for closed ranges 2478 // and (x, y) for open ranges. These ranges are modular, corresponding with 2479 // a common treatment of C integer overflow. This means that these methods 2480 // do not have to worry about overflow; RangeSet::Intersect can handle such a 2481 // "wraparound" range. 2482 // As an example, the range [UINT_MAX-1, 3) contains five values: UINT_MAX-1, 2483 // UINT_MAX, 0, 1, and 2. 2484 2485 ProgramStateRef 2486 RangeConstraintManager::assumeSymNE(ProgramStateRef St, SymbolRef Sym, 2487 const llvm::APSInt &Int, 2488 const llvm::APSInt &Adjustment) { 2489 // Before we do any real work, see if the value can even show up. 2490 APSIntType AdjustmentType(Adjustment); 2491 if (AdjustmentType.testInRange(Int, true) != APSIntType::RTR_Within) 2492 return St; 2493 2494 llvm::APSInt Point = AdjustmentType.convert(Int) - Adjustment; 2495 RangeSet New = getRange(St, Sym); 2496 New = F.deletePoint(New, Point); 2497 2498 return setRange(St, Sym, New); 2499 } 2500 2501 ProgramStateRef 2502 RangeConstraintManager::assumeSymEQ(ProgramStateRef St, SymbolRef Sym, 2503 const llvm::APSInt &Int, 2504 const llvm::APSInt &Adjustment) { 2505 // Before we do any real work, see if the value can even show up. 2506 APSIntType AdjustmentType(Adjustment); 2507 if (AdjustmentType.testInRange(Int, true) != APSIntType::RTR_Within) 2508 return nullptr; 2509 2510 // [Int-Adjustment, Int-Adjustment] 2511 llvm::APSInt AdjInt = AdjustmentType.convert(Int) - Adjustment; 2512 RangeSet New = getRange(St, Sym); 2513 New = F.intersect(New, AdjInt); 2514 2515 return setRange(St, Sym, New); 2516 } 2517 2518 RangeSet RangeConstraintManager::getSymLTRange(ProgramStateRef St, 2519 SymbolRef Sym, 2520 const llvm::APSInt &Int, 2521 const llvm::APSInt &Adjustment) { 2522 // Before we do any real work, see if the value can even show up. 2523 APSIntType AdjustmentType(Adjustment); 2524 switch (AdjustmentType.testInRange(Int, true)) { 2525 case APSIntType::RTR_Below: 2526 return F.getEmptySet(); 2527 case APSIntType::RTR_Within: 2528 break; 2529 case APSIntType::RTR_Above: 2530 return getRange(St, Sym); 2531 } 2532 2533 // Special case for Int == Min. This is always false. 2534 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2535 llvm::APSInt Min = AdjustmentType.getMinValue(); 2536 if (ComparisonVal == Min) 2537 return F.getEmptySet(); 2538 2539 llvm::APSInt Lower = Min - Adjustment; 2540 llvm::APSInt Upper = ComparisonVal - Adjustment; 2541 --Upper; 2542 2543 RangeSet Result = getRange(St, Sym); 2544 return F.intersect(Result, Lower, Upper); 2545 } 2546 2547 ProgramStateRef 2548 RangeConstraintManager::assumeSymLT(ProgramStateRef St, SymbolRef Sym, 2549 const llvm::APSInt &Int, 2550 const llvm::APSInt &Adjustment) { 2551 RangeSet New = getSymLTRange(St, Sym, Int, Adjustment); 2552 return setRange(St, Sym, New); 2553 } 2554 2555 RangeSet RangeConstraintManager::getSymGTRange(ProgramStateRef St, 2556 SymbolRef Sym, 2557 const llvm::APSInt &Int, 2558 const llvm::APSInt &Adjustment) { 2559 // Before we do any real work, see if the value can even show up. 2560 APSIntType AdjustmentType(Adjustment); 2561 switch (AdjustmentType.testInRange(Int, true)) { 2562 case APSIntType::RTR_Below: 2563 return getRange(St, Sym); 2564 case APSIntType::RTR_Within: 2565 break; 2566 case APSIntType::RTR_Above: 2567 return F.getEmptySet(); 2568 } 2569 2570 // Special case for Int == Max. This is always false. 2571 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2572 llvm::APSInt Max = AdjustmentType.getMaxValue(); 2573 if (ComparisonVal == Max) 2574 return F.getEmptySet(); 2575 2576 llvm::APSInt Lower = ComparisonVal - Adjustment; 2577 llvm::APSInt Upper = Max - Adjustment; 2578 ++Lower; 2579 2580 RangeSet SymRange = getRange(St, Sym); 2581 return F.intersect(SymRange, Lower, Upper); 2582 } 2583 2584 ProgramStateRef 2585 RangeConstraintManager::assumeSymGT(ProgramStateRef St, SymbolRef Sym, 2586 const llvm::APSInt &Int, 2587 const llvm::APSInt &Adjustment) { 2588 RangeSet New = getSymGTRange(St, Sym, Int, Adjustment); 2589 return setRange(St, Sym, New); 2590 } 2591 2592 RangeSet RangeConstraintManager::getSymGERange(ProgramStateRef St, 2593 SymbolRef Sym, 2594 const llvm::APSInt &Int, 2595 const llvm::APSInt &Adjustment) { 2596 // Before we do any real work, see if the value can even show up. 2597 APSIntType AdjustmentType(Adjustment); 2598 switch (AdjustmentType.testInRange(Int, true)) { 2599 case APSIntType::RTR_Below: 2600 return getRange(St, Sym); 2601 case APSIntType::RTR_Within: 2602 break; 2603 case APSIntType::RTR_Above: 2604 return F.getEmptySet(); 2605 } 2606 2607 // Special case for Int == Min. This is always feasible. 2608 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2609 llvm::APSInt Min = AdjustmentType.getMinValue(); 2610 if (ComparisonVal == Min) 2611 return getRange(St, Sym); 2612 2613 llvm::APSInt Max = AdjustmentType.getMaxValue(); 2614 llvm::APSInt Lower = ComparisonVal - Adjustment; 2615 llvm::APSInt Upper = Max - Adjustment; 2616 2617 RangeSet SymRange = getRange(St, Sym); 2618 return F.intersect(SymRange, Lower, Upper); 2619 } 2620 2621 ProgramStateRef 2622 RangeConstraintManager::assumeSymGE(ProgramStateRef St, SymbolRef Sym, 2623 const llvm::APSInt &Int, 2624 const llvm::APSInt &Adjustment) { 2625 RangeSet New = getSymGERange(St, Sym, Int, Adjustment); 2626 return setRange(St, Sym, New); 2627 } 2628 2629 RangeSet 2630 RangeConstraintManager::getSymLERange(llvm::function_ref<RangeSet()> RS, 2631 const llvm::APSInt &Int, 2632 const llvm::APSInt &Adjustment) { 2633 // Before we do any real work, see if the value can even show up. 2634 APSIntType AdjustmentType(Adjustment); 2635 switch (AdjustmentType.testInRange(Int, true)) { 2636 case APSIntType::RTR_Below: 2637 return F.getEmptySet(); 2638 case APSIntType::RTR_Within: 2639 break; 2640 case APSIntType::RTR_Above: 2641 return RS(); 2642 } 2643 2644 // Special case for Int == Max. This is always feasible. 2645 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2646 llvm::APSInt Max = AdjustmentType.getMaxValue(); 2647 if (ComparisonVal == Max) 2648 return RS(); 2649 2650 llvm::APSInt Min = AdjustmentType.getMinValue(); 2651 llvm::APSInt Lower = Min - Adjustment; 2652 llvm::APSInt Upper = ComparisonVal - Adjustment; 2653 2654 RangeSet Default = RS(); 2655 return F.intersect(Default, Lower, Upper); 2656 } 2657 2658 RangeSet RangeConstraintManager::getSymLERange(ProgramStateRef St, 2659 SymbolRef Sym, 2660 const llvm::APSInt &Int, 2661 const llvm::APSInt &Adjustment) { 2662 return getSymLERange([&] { return getRange(St, Sym); }, Int, Adjustment); 2663 } 2664 2665 ProgramStateRef 2666 RangeConstraintManager::assumeSymLE(ProgramStateRef St, SymbolRef Sym, 2667 const llvm::APSInt &Int, 2668 const llvm::APSInt &Adjustment) { 2669 RangeSet New = getSymLERange(St, Sym, Int, Adjustment); 2670 return setRange(St, Sym, New); 2671 } 2672 2673 ProgramStateRef RangeConstraintManager::assumeSymWithinInclusiveRange( 2674 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 2675 const llvm::APSInt &To, const llvm::APSInt &Adjustment) { 2676 RangeSet New = getSymGERange(State, Sym, From, Adjustment); 2677 if (New.isEmpty()) 2678 return nullptr; 2679 RangeSet Out = getSymLERange([&] { return New; }, To, Adjustment); 2680 return setRange(State, Sym, Out); 2681 } 2682 2683 ProgramStateRef RangeConstraintManager::assumeSymOutsideInclusiveRange( 2684 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 2685 const llvm::APSInt &To, const llvm::APSInt &Adjustment) { 2686 RangeSet RangeLT = getSymLTRange(State, Sym, From, Adjustment); 2687 RangeSet RangeGT = getSymGTRange(State, Sym, To, Adjustment); 2688 RangeSet New(F.add(RangeLT, RangeGT)); 2689 return setRange(State, Sym, New); 2690 } 2691 2692 //===----------------------------------------------------------------------===// 2693 // Pretty-printing. 2694 //===----------------------------------------------------------------------===// 2695 2696 void RangeConstraintManager::printJson(raw_ostream &Out, ProgramStateRef State, 2697 const char *NL, unsigned int Space, 2698 bool IsDot) const { 2699 printConstraints(Out, State, NL, Space, IsDot); 2700 printEquivalenceClasses(Out, State, NL, Space, IsDot); 2701 printDisequalities(Out, State, NL, Space, IsDot); 2702 } 2703 2704 static std::string toString(const SymbolRef &Sym) { 2705 std::string S; 2706 llvm::raw_string_ostream O(S); 2707 Sym->dumpToStream(O); 2708 return O.str(); 2709 } 2710 2711 void RangeConstraintManager::printConstraints(raw_ostream &Out, 2712 ProgramStateRef State, 2713 const char *NL, 2714 unsigned int Space, 2715 bool IsDot) const { 2716 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 2717 2718 Indent(Out, Space, IsDot) << "\"constraints\": "; 2719 if (Constraints.isEmpty()) { 2720 Out << "null," << NL; 2721 return; 2722 } 2723 2724 std::map<std::string, RangeSet> OrderedConstraints; 2725 for (std::pair<EquivalenceClass, RangeSet> P : Constraints) { 2726 SymbolSet ClassMembers = P.first.getClassMembers(State); 2727 for (const SymbolRef &ClassMember : ClassMembers) { 2728 bool insertion_took_place; 2729 std::tie(std::ignore, insertion_took_place) = 2730 OrderedConstraints.insert({toString(ClassMember), P.second}); 2731 assert(insertion_took_place && 2732 "two symbols should not have the same dump"); 2733 } 2734 } 2735 2736 ++Space; 2737 Out << '[' << NL; 2738 bool First = true; 2739 for (std::pair<std::string, RangeSet> P : OrderedConstraints) { 2740 if (First) { 2741 First = false; 2742 } else { 2743 Out << ','; 2744 Out << NL; 2745 } 2746 Indent(Out, Space, IsDot) 2747 << "{ \"symbol\": \"" << P.first << "\", \"range\": \""; 2748 P.second.dump(Out); 2749 Out << "\" }"; 2750 } 2751 Out << NL; 2752 2753 --Space; 2754 Indent(Out, Space, IsDot) << "]," << NL; 2755 } 2756 2757 static std::string toString(ProgramStateRef State, EquivalenceClass Class) { 2758 SymbolSet ClassMembers = Class.getClassMembers(State); 2759 llvm::SmallVector<SymbolRef, 8> ClassMembersSorted(ClassMembers.begin(), 2760 ClassMembers.end()); 2761 llvm::sort(ClassMembersSorted, 2762 [](const SymbolRef &LHS, const SymbolRef &RHS) { 2763 return toString(LHS) < toString(RHS); 2764 }); 2765 2766 bool FirstMember = true; 2767 2768 std::string Str; 2769 llvm::raw_string_ostream Out(Str); 2770 Out << "[ "; 2771 for (SymbolRef ClassMember : ClassMembersSorted) { 2772 if (FirstMember) 2773 FirstMember = false; 2774 else 2775 Out << ", "; 2776 Out << "\"" << ClassMember << "\""; 2777 } 2778 Out << " ]"; 2779 return Out.str(); 2780 } 2781 2782 void RangeConstraintManager::printEquivalenceClasses(raw_ostream &Out, 2783 ProgramStateRef State, 2784 const char *NL, 2785 unsigned int Space, 2786 bool IsDot) const { 2787 ClassMembersTy Members = State->get<ClassMembers>(); 2788 2789 Indent(Out, Space, IsDot) << "\"equivalence_classes\": "; 2790 if (Members.isEmpty()) { 2791 Out << "null," << NL; 2792 return; 2793 } 2794 2795 std::set<std::string> MembersStr; 2796 for (std::pair<EquivalenceClass, SymbolSet> ClassToSymbolSet : Members) 2797 MembersStr.insert(toString(State, ClassToSymbolSet.first)); 2798 2799 ++Space; 2800 Out << '[' << NL; 2801 bool FirstClass = true; 2802 for (const std::string &Str : MembersStr) { 2803 if (FirstClass) { 2804 FirstClass = false; 2805 } else { 2806 Out << ','; 2807 Out << NL; 2808 } 2809 Indent(Out, Space, IsDot); 2810 Out << Str; 2811 } 2812 Out << NL; 2813 2814 --Space; 2815 Indent(Out, Space, IsDot) << "]," << NL; 2816 } 2817 2818 void RangeConstraintManager::printDisequalities(raw_ostream &Out, 2819 ProgramStateRef State, 2820 const char *NL, 2821 unsigned int Space, 2822 bool IsDot) const { 2823 DisequalityMapTy Disequalities = State->get<DisequalityMap>(); 2824 2825 Indent(Out, Space, IsDot) << "\"disequality_info\": "; 2826 if (Disequalities.isEmpty()) { 2827 Out << "null," << NL; 2828 return; 2829 } 2830 2831 // Transform the disequality info to an ordered map of 2832 // [string -> (ordered set of strings)] 2833 using EqClassesStrTy = std::set<std::string>; 2834 using DisequalityInfoStrTy = std::map<std::string, EqClassesStrTy>; 2835 DisequalityInfoStrTy DisequalityInfoStr; 2836 for (std::pair<EquivalenceClass, ClassSet> ClassToDisEqSet : Disequalities) { 2837 EquivalenceClass Class = ClassToDisEqSet.first; 2838 ClassSet DisequalClasses = ClassToDisEqSet.second; 2839 EqClassesStrTy MembersStr; 2840 for (EquivalenceClass DisEqClass : DisequalClasses) 2841 MembersStr.insert(toString(State, DisEqClass)); 2842 DisequalityInfoStr.insert({toString(State, Class), MembersStr}); 2843 } 2844 2845 ++Space; 2846 Out << '[' << NL; 2847 bool FirstClass = true; 2848 for (std::pair<std::string, EqClassesStrTy> ClassToDisEqSet : 2849 DisequalityInfoStr) { 2850 const std::string &Class = ClassToDisEqSet.first; 2851 if (FirstClass) { 2852 FirstClass = false; 2853 } else { 2854 Out << ','; 2855 Out << NL; 2856 } 2857 Indent(Out, Space, IsDot) << "{" << NL; 2858 unsigned int DisEqSpace = Space + 1; 2859 Indent(Out, DisEqSpace, IsDot) << "\"class\": "; 2860 Out << Class; 2861 const EqClassesStrTy &DisequalClasses = ClassToDisEqSet.second; 2862 if (!DisequalClasses.empty()) { 2863 Out << "," << NL; 2864 Indent(Out, DisEqSpace, IsDot) << "\"disequal_to\": [" << NL; 2865 unsigned int DisEqClassSpace = DisEqSpace + 1; 2866 Indent(Out, DisEqClassSpace, IsDot); 2867 bool FirstDisEqClass = true; 2868 for (const std::string &DisEqClass : DisequalClasses) { 2869 if (FirstDisEqClass) { 2870 FirstDisEqClass = false; 2871 } else { 2872 Out << ',' << NL; 2873 Indent(Out, DisEqClassSpace, IsDot); 2874 } 2875 Out << DisEqClass; 2876 } 2877 Out << "]" << NL; 2878 } 2879 Indent(Out, Space, IsDot) << "}"; 2880 } 2881 Out << NL; 2882 2883 --Space; 2884 Indent(Out, Space, IsDot) << "]," << NL; 2885 } 2886