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 LHS, RangeSet RHS) { 114 ContainerType Result; 115 Result.reserve(LHS.size() + RHS.size()); 116 std::merge(LHS.begin(), LHS.end(), RHS.begin(), RHS.end(), 117 std::back_inserter(Result)); 118 return makePersistent(std::move(Result)); 119 } 120 121 RangeSet RangeSet::Factory::add(RangeSet Original, Range Element) { 122 ContainerType Result; 123 Result.reserve(Original.size() + 1); 124 125 const_iterator Lower = llvm::lower_bound(Original, Element); 126 Result.insert(Result.end(), Original.begin(), Lower); 127 Result.push_back(Element); 128 Result.insert(Result.end(), Lower, Original.end()); 129 130 return makePersistent(std::move(Result)); 131 } 132 133 RangeSet RangeSet::Factory::add(RangeSet Original, const llvm::APSInt &Point) { 134 return add(Original, Range(Point)); 135 } 136 137 RangeSet RangeSet::Factory::unite(RangeSet LHS, RangeSet RHS) { 138 ContainerType Result = unite(*LHS.Impl, *RHS.Impl); 139 return makePersistent(std::move(Result)); 140 } 141 142 RangeSet RangeSet::Factory::unite(RangeSet Original, Range R) { 143 ContainerType Result; 144 Result.push_back(R); 145 Result = unite(*Original.Impl, Result); 146 return makePersistent(std::move(Result)); 147 } 148 149 RangeSet RangeSet::Factory::unite(RangeSet Original, llvm::APSInt Point) { 150 return unite(Original, Range(ValueFactory.getValue(Point))); 151 } 152 153 RangeSet RangeSet::Factory::unite(RangeSet Original, llvm::APSInt From, 154 llvm::APSInt To) { 155 return unite(Original, 156 Range(ValueFactory.getValue(From), ValueFactory.getValue(To))); 157 } 158 159 template <typename T> 160 void swapIterators(T &First, T &FirstEnd, T &Second, T &SecondEnd) { 161 std::swap(First, Second); 162 std::swap(FirstEnd, SecondEnd); 163 } 164 165 RangeSet::ContainerType RangeSet::Factory::unite(const ContainerType &LHS, 166 const ContainerType &RHS) { 167 if (LHS.empty()) 168 return RHS; 169 if (RHS.empty()) 170 return LHS; 171 172 using llvm::APSInt; 173 using iterator = ContainerType::const_iterator; 174 175 iterator First = LHS.begin(); 176 iterator FirstEnd = LHS.end(); 177 iterator Second = RHS.begin(); 178 iterator SecondEnd = RHS.end(); 179 APSIntType Ty = APSIntType(First->From()); 180 const APSInt Min = Ty.getMinValue(); 181 182 // Handle a corner case first when both range sets start from MIN. 183 // This helps to avoid complicated conditions below. Specifically, this 184 // particular check for `MIN` is not needed in the loop below every time 185 // when we do `Second->From() - One` operation. 186 if (Min == First->From() && Min == Second->From()) { 187 if (First->To() > Second->To()) { 188 // [ First ]---> 189 // [ Second ]-----> 190 // MIN^ 191 // The Second range is entirely inside the First one. 192 193 // Check if Second is the last in its RangeSet. 194 if (++Second == SecondEnd) 195 // [ First ]--[ First + 1 ]---> 196 // [ Second ]---------------------> 197 // MIN^ 198 // The Union is equal to First's RangeSet. 199 return LHS; 200 } else { 201 // case 1: [ First ]-----> 202 // case 2: [ First ]---> 203 // [ Second ]---> 204 // MIN^ 205 // The First range is entirely inside or equal to the Second one. 206 207 // Check if First is the last in its RangeSet. 208 if (++First == FirstEnd) 209 // [ First ]-----------------------> 210 // [ Second ]--[ Second + 1 ]----> 211 // MIN^ 212 // The Union is equal to Second's RangeSet. 213 return RHS; 214 } 215 } 216 217 const APSInt One = Ty.getValue(1); 218 ContainerType Result; 219 220 // This is called when there are no ranges left in one of the ranges. 221 // Append the rest of the ranges from another range set to the Result 222 // and return with that. 223 const auto AppendTheRest = [&Result](iterator I, iterator E) { 224 Result.append(I, E); 225 return Result; 226 }; 227 228 while (true) { 229 // We want to keep the following invariant at all times: 230 // ---[ First ------> 231 // -----[ Second ---> 232 if (First->From() > Second->From()) 233 swapIterators(First, FirstEnd, Second, SecondEnd); 234 235 // The Union definitely starts with First->From(). 236 // ----------[ First ------> 237 // ------------[ Second ---> 238 // ----------[ Union ------> 239 // UnionStart^ 240 const llvm::APSInt &UnionStart = First->From(); 241 242 // Loop where the invariant holds. 243 while (true) { 244 // Skip all enclosed ranges. 245 // ---[ First ]---> 246 // -----[ Second ]--[ Second + 1 ]--[ Second + N ]-----> 247 while (First->To() >= Second->To()) { 248 // Check if Second is the last in its RangeSet. 249 if (++Second == SecondEnd) { 250 // Append the Union. 251 // ---[ Union ]---> 252 // -----[ Second ]-----> 253 // --------[ First ]---> 254 // UnionEnd^ 255 Result.emplace_back(UnionStart, First->To()); 256 // ---[ Union ]-----------------> 257 // --------------[ First + 1]---> 258 // Append all remaining ranges from the First's RangeSet. 259 return AppendTheRest(++First, FirstEnd); 260 } 261 } 262 263 // Check if First and Second are disjoint. It means that we find 264 // the end of the Union. Exit the loop and append the Union. 265 // ---[ First ]=-------------> 266 // ------------=[ Second ]---> 267 // ----MinusOne^ 268 if (First->To() < Second->From() - One) 269 break; 270 271 // First is entirely inside the Union. Go next. 272 // ---[ Union -----------> 273 // ---- [ First ]--------> 274 // -------[ Second ]-----> 275 // Check if First is the last in its RangeSet. 276 if (++First == FirstEnd) { 277 // Append the Union. 278 // ---[ Union ]---> 279 // -----[ First ]-------> 280 // --------[ Second ]---> 281 // UnionEnd^ 282 Result.emplace_back(UnionStart, Second->To()); 283 // ---[ Union ]------------------> 284 // --------------[ Second + 1]---> 285 // Append all remaining ranges from the Second's RangeSet. 286 return AppendTheRest(++Second, SecondEnd); 287 } 288 289 // We know that we are at one of the two cases: 290 // case 1: --[ First ]---------> 291 // case 2: ----[ First ]-------> 292 // --------[ Second ]----------> 293 // In both cases First starts after Second->From(). 294 // Make sure that the loop invariant holds. 295 swapIterators(First, FirstEnd, Second, SecondEnd); 296 } 297 298 // Here First and Second are disjoint. 299 // Append the Union. 300 // ---[ Union ]---------------> 301 // -----------------[ Second ]---> 302 // ------[ First ]---------------> 303 // UnionEnd^ 304 Result.emplace_back(UnionStart, First->To()); 305 306 // Check if First is the last in its RangeSet. 307 if (++First == FirstEnd) 308 // ---[ Union ]---------------> 309 // --------------[ Second ]---> 310 // Append all remaining ranges from the Second's RangeSet. 311 return AppendTheRest(Second, SecondEnd); 312 } 313 314 llvm_unreachable("Normally, we should not reach here"); 315 } 316 317 RangeSet RangeSet::Factory::getRangeSet(Range From) { 318 ContainerType Result; 319 Result.push_back(From); 320 return makePersistent(std::move(Result)); 321 } 322 323 RangeSet RangeSet::Factory::makePersistent(ContainerType &&From) { 324 llvm::FoldingSetNodeID ID; 325 void *InsertPos; 326 327 From.Profile(ID); 328 ContainerType *Result = Cache.FindNodeOrInsertPos(ID, InsertPos); 329 330 if (!Result) { 331 // It is cheaper to fully construct the resulting range on stack 332 // and move it to the freshly allocated buffer if we don't have 333 // a set like this already. 334 Result = construct(std::move(From)); 335 Cache.InsertNode(Result, InsertPos); 336 } 337 338 return Result; 339 } 340 341 RangeSet::ContainerType *RangeSet::Factory::construct(ContainerType &&From) { 342 void *Buffer = Arena.Allocate(); 343 return new (Buffer) ContainerType(std::move(From)); 344 } 345 346 const llvm::APSInt &RangeSet::getMinValue() const { 347 assert(!isEmpty()); 348 return begin()->From(); 349 } 350 351 const llvm::APSInt &RangeSet::getMaxValue() const { 352 assert(!isEmpty()); 353 return std::prev(end())->To(); 354 } 355 356 bool RangeSet::containsImpl(llvm::APSInt &Point) const { 357 if (isEmpty() || !pin(Point)) 358 return false; 359 360 Range Dummy(Point); 361 const_iterator It = llvm::upper_bound(*this, Dummy); 362 if (It == begin()) 363 return false; 364 365 return std::prev(It)->Includes(Point); 366 } 367 368 bool RangeSet::pin(llvm::APSInt &Point) const { 369 APSIntType Type(getMinValue()); 370 if (Type.testInRange(Point, true) != APSIntType::RTR_Within) 371 return false; 372 373 Type.apply(Point); 374 return true; 375 } 376 377 bool RangeSet::pin(llvm::APSInt &Lower, llvm::APSInt &Upper) const { 378 // This function has nine cases, the cartesian product of range-testing 379 // both the upper and lower bounds against the symbol's type. 380 // Each case requires a different pinning operation. 381 // The function returns false if the described range is entirely outside 382 // the range of values for the associated symbol. 383 APSIntType Type(getMinValue()); 384 APSIntType::RangeTestResultKind LowerTest = Type.testInRange(Lower, true); 385 APSIntType::RangeTestResultKind UpperTest = Type.testInRange(Upper, true); 386 387 switch (LowerTest) { 388 case APSIntType::RTR_Below: 389 switch (UpperTest) { 390 case APSIntType::RTR_Below: 391 // The entire range is outside the symbol's set of possible values. 392 // If this is a conventionally-ordered range, the state is infeasible. 393 if (Lower <= Upper) 394 return false; 395 396 // However, if the range wraps around, it spans all possible values. 397 Lower = Type.getMinValue(); 398 Upper = Type.getMaxValue(); 399 break; 400 case APSIntType::RTR_Within: 401 // The range starts below what's possible but ends within it. Pin. 402 Lower = Type.getMinValue(); 403 Type.apply(Upper); 404 break; 405 case APSIntType::RTR_Above: 406 // The range spans all possible values for the symbol. Pin. 407 Lower = Type.getMinValue(); 408 Upper = Type.getMaxValue(); 409 break; 410 } 411 break; 412 case APSIntType::RTR_Within: 413 switch (UpperTest) { 414 case APSIntType::RTR_Below: 415 // The range wraps around, but all lower values are not possible. 416 Type.apply(Lower); 417 Upper = Type.getMaxValue(); 418 break; 419 case APSIntType::RTR_Within: 420 // The range may or may not wrap around, but both limits are valid. 421 Type.apply(Lower); 422 Type.apply(Upper); 423 break; 424 case APSIntType::RTR_Above: 425 // The range starts within what's possible but ends above it. Pin. 426 Type.apply(Lower); 427 Upper = Type.getMaxValue(); 428 break; 429 } 430 break; 431 case APSIntType::RTR_Above: 432 switch (UpperTest) { 433 case APSIntType::RTR_Below: 434 // The range wraps but is outside the symbol's set of possible values. 435 return false; 436 case APSIntType::RTR_Within: 437 // The range starts above what's possible but ends within it (wrap). 438 Lower = Type.getMinValue(); 439 Type.apply(Upper); 440 break; 441 case APSIntType::RTR_Above: 442 // The entire range is outside the symbol's set of possible values. 443 // If this is a conventionally-ordered range, the state is infeasible. 444 if (Lower <= Upper) 445 return false; 446 447 // However, if the range wraps around, it spans all possible values. 448 Lower = Type.getMinValue(); 449 Upper = Type.getMaxValue(); 450 break; 451 } 452 break; 453 } 454 455 return true; 456 } 457 458 RangeSet RangeSet::Factory::intersect(RangeSet What, llvm::APSInt Lower, 459 llvm::APSInt Upper) { 460 if (What.isEmpty() || !What.pin(Lower, Upper)) 461 return getEmptySet(); 462 463 ContainerType DummyContainer; 464 465 if (Lower <= Upper) { 466 // [Lower, Upper] is a regular range. 467 // 468 // Shortcut: check that there is even a possibility of the intersection 469 // by checking the two following situations: 470 // 471 // <---[ What ]---[------]------> 472 // Lower Upper 473 // -or- 474 // <----[------]----[ What ]----> 475 // Lower Upper 476 if (What.getMaxValue() < Lower || Upper < What.getMinValue()) 477 return getEmptySet(); 478 479 DummyContainer.push_back( 480 Range(ValueFactory.getValue(Lower), ValueFactory.getValue(Upper))); 481 } else { 482 // [Lower, Upper] is an inverted range, i.e. [MIN, Upper] U [Lower, MAX] 483 // 484 // Shortcut: check that there is even a possibility of the intersection 485 // by checking the following situation: 486 // 487 // <------]---[ What ]---[------> 488 // Upper Lower 489 if (What.getMaxValue() < Lower && Upper < What.getMinValue()) 490 return getEmptySet(); 491 492 DummyContainer.push_back( 493 Range(ValueFactory.getMinValue(Upper), ValueFactory.getValue(Upper))); 494 DummyContainer.push_back( 495 Range(ValueFactory.getValue(Lower), ValueFactory.getMaxValue(Lower))); 496 } 497 498 return intersect(*What.Impl, DummyContainer); 499 } 500 501 RangeSet RangeSet::Factory::intersect(const RangeSet::ContainerType &LHS, 502 const RangeSet::ContainerType &RHS) { 503 ContainerType Result; 504 Result.reserve(std::max(LHS.size(), RHS.size())); 505 506 const_iterator First = LHS.begin(), Second = RHS.begin(), 507 FirstEnd = LHS.end(), SecondEnd = RHS.end(); 508 509 // If we ran out of ranges in one set, but not in the other, 510 // it means that those elements are definitely not in the 511 // intersection. 512 while (First != FirstEnd && Second != SecondEnd) { 513 // We want to keep the following invariant at all times: 514 // 515 // ----[ First ----------------------> 516 // --------[ Second -----------------> 517 if (Second->From() < First->From()) 518 swapIterators(First, FirstEnd, Second, SecondEnd); 519 520 // Loop where the invariant holds: 521 do { 522 // Check for the following situation: 523 // 524 // ----[ First ]---------------------> 525 // ---------------[ Second ]---------> 526 // 527 // which means that... 528 if (Second->From() > First->To()) { 529 // ...First is not in the intersection. 530 // 531 // We should move on to the next range after First and break out of the 532 // loop because the invariant might not be true. 533 ++First; 534 break; 535 } 536 537 // We have a guaranteed intersection at this point! 538 // And this is the current situation: 539 // 540 // ----[ First ]-----------------> 541 // -------[ Second ------------------> 542 // 543 // Additionally, it definitely starts with Second->From(). 544 const llvm::APSInt &IntersectionStart = Second->From(); 545 546 // It is important to know which of the two ranges' ends 547 // is greater. That "longer" range might have some other 548 // intersections, while the "shorter" range might not. 549 if (Second->To() > First->To()) { 550 // Here we make a decision to keep First as the "longer" 551 // range. 552 swapIterators(First, FirstEnd, Second, SecondEnd); 553 } 554 555 // At this point, we have the following situation: 556 // 557 // ---- First ]--------------------> 558 // ---- Second ]--[ Second+1 ----------> 559 // 560 // We don't know the relationship between First->From and 561 // Second->From and we don't know whether Second+1 intersects 562 // with First. 563 // 564 // However, we know that [IntersectionStart, Second->To] is 565 // a part of the intersection... 566 Result.push_back(Range(IntersectionStart, Second->To())); 567 ++Second; 568 // ...and that the invariant will hold for a valid Second+1 569 // because First->From <= Second->To < (Second+1)->From. 570 } while (Second != SecondEnd); 571 } 572 573 if (Result.empty()) 574 return getEmptySet(); 575 576 return makePersistent(std::move(Result)); 577 } 578 579 RangeSet RangeSet::Factory::intersect(RangeSet LHS, RangeSet RHS) { 580 // Shortcut: let's see if the intersection is even possible. 581 if (LHS.isEmpty() || RHS.isEmpty() || LHS.getMaxValue() < RHS.getMinValue() || 582 RHS.getMaxValue() < LHS.getMinValue()) 583 return getEmptySet(); 584 585 return intersect(*LHS.Impl, *RHS.Impl); 586 } 587 588 RangeSet RangeSet::Factory::intersect(RangeSet LHS, llvm::APSInt Point) { 589 if (LHS.containsImpl(Point)) 590 return getRangeSet(ValueFactory.getValue(Point)); 591 592 return getEmptySet(); 593 } 594 595 RangeSet RangeSet::Factory::negate(RangeSet What) { 596 if (What.isEmpty()) 597 return getEmptySet(); 598 599 const llvm::APSInt SampleValue = What.getMinValue(); 600 const llvm::APSInt &MIN = ValueFactory.getMinValue(SampleValue); 601 const llvm::APSInt &MAX = ValueFactory.getMaxValue(SampleValue); 602 603 ContainerType Result; 604 Result.reserve(What.size() + (SampleValue == MIN)); 605 606 // Handle a special case for MIN value. 607 const_iterator It = What.begin(); 608 const_iterator End = What.end(); 609 610 const llvm::APSInt &From = It->From(); 611 const llvm::APSInt &To = It->To(); 612 613 if (From == MIN) { 614 // If the range [From, To] is [MIN, MAX], then result is also [MIN, MAX]. 615 if (To == MAX) { 616 return What; 617 } 618 619 const_iterator Last = std::prev(End); 620 621 // Try to find and unite the following ranges: 622 // [MIN, MIN] & [MIN + 1, N] => [MIN, N]. 623 if (Last->To() == MAX) { 624 // It means that in the original range we have ranges 625 // [MIN, A], ... , [B, MAX] 626 // And the result should be [MIN, -B], ..., [-A, MAX] 627 Result.emplace_back(MIN, ValueFactory.getValue(-Last->From())); 628 // We already negated Last, so we can skip it. 629 End = Last; 630 } else { 631 // Add a separate range for the lowest value. 632 Result.emplace_back(MIN, MIN); 633 } 634 635 // Skip adding the second range in case when [From, To] are [MIN, MIN]. 636 if (To != MIN) { 637 Result.emplace_back(ValueFactory.getValue(-To), MAX); 638 } 639 640 // Skip the first range in the loop. 641 ++It; 642 } 643 644 // Negate all other ranges. 645 for (; It != End; ++It) { 646 // Negate int values. 647 const llvm::APSInt &NewFrom = ValueFactory.getValue(-It->To()); 648 const llvm::APSInt &NewTo = ValueFactory.getValue(-It->From()); 649 650 // Add a negated range. 651 Result.emplace_back(NewFrom, NewTo); 652 } 653 654 llvm::sort(Result); 655 return makePersistent(std::move(Result)); 656 } 657 658 RangeSet RangeSet::Factory::deletePoint(RangeSet From, 659 const llvm::APSInt &Point) { 660 if (!From.contains(Point)) 661 return From; 662 663 llvm::APSInt Upper = Point; 664 llvm::APSInt Lower = Point; 665 666 ++Upper; 667 --Lower; 668 669 // Notice that the lower bound is greater than the upper bound. 670 return intersect(From, Upper, Lower); 671 } 672 673 LLVM_DUMP_METHOD void Range::dump(raw_ostream &OS) const { 674 OS << '[' << toString(From(), 10) << ", " << toString(To(), 10) << ']'; 675 } 676 LLVM_DUMP_METHOD void Range::dump() const { dump(llvm::errs()); } 677 678 LLVM_DUMP_METHOD void RangeSet::dump(raw_ostream &OS) const { 679 OS << "{ "; 680 llvm::interleaveComma(*this, OS, [&OS](const Range &R) { R.dump(OS); }); 681 OS << " }"; 682 } 683 LLVM_DUMP_METHOD void RangeSet::dump() const { dump(llvm::errs()); } 684 685 REGISTER_SET_FACTORY_WITH_PROGRAMSTATE(SymbolSet, SymbolRef) 686 687 namespace { 688 class EquivalenceClass; 689 } // end anonymous namespace 690 691 REGISTER_MAP_WITH_PROGRAMSTATE(ClassMap, SymbolRef, EquivalenceClass) 692 REGISTER_MAP_WITH_PROGRAMSTATE(ClassMembers, EquivalenceClass, SymbolSet) 693 REGISTER_MAP_WITH_PROGRAMSTATE(ConstraintRange, EquivalenceClass, RangeSet) 694 695 REGISTER_SET_FACTORY_WITH_PROGRAMSTATE(ClassSet, EquivalenceClass) 696 REGISTER_MAP_WITH_PROGRAMSTATE(DisequalityMap, EquivalenceClass, ClassSet) 697 698 namespace { 699 /// This class encapsulates a set of symbols equal to each other. 700 /// 701 /// The main idea of the approach requiring such classes is in narrowing 702 /// and sharing constraints between symbols within the class. Also we can 703 /// conclude that there is no practical need in storing constraints for 704 /// every member of the class separately. 705 /// 706 /// Main terminology: 707 /// 708 /// * "Equivalence class" is an object of this class, which can be efficiently 709 /// compared to other classes. It represents the whole class without 710 /// storing the actual in it. The members of the class however can be 711 /// retrieved from the state. 712 /// 713 /// * "Class members" are the symbols corresponding to the class. This means 714 /// that A == B for every member symbols A and B from the class. Members of 715 /// each class are stored in the state. 716 /// 717 /// * "Trivial class" is a class that has and ever had only one same symbol. 718 /// 719 /// * "Merge operation" merges two classes into one. It is the main operation 720 /// to produce non-trivial classes. 721 /// If, at some point, we can assume that two symbols from two distinct 722 /// classes are equal, we can merge these classes. 723 class EquivalenceClass : public llvm::FoldingSetNode { 724 public: 725 /// Find equivalence class for the given symbol in the given state. 726 LLVM_NODISCARD static inline EquivalenceClass find(ProgramStateRef State, 727 SymbolRef Sym); 728 729 /// Merge classes for the given symbols and return a new state. 730 LLVM_NODISCARD static inline ProgramStateRef merge(RangeSet::Factory &F, 731 ProgramStateRef State, 732 SymbolRef First, 733 SymbolRef Second); 734 // Merge this class with the given class and return a new state. 735 LLVM_NODISCARD inline ProgramStateRef 736 merge(RangeSet::Factory &F, ProgramStateRef State, EquivalenceClass Other); 737 738 /// Return a set of class members for the given state. 739 LLVM_NODISCARD inline SymbolSet getClassMembers(ProgramStateRef State) const; 740 741 /// Return true if the current class is trivial in the given state. 742 /// A class is trivial if and only if there is not any member relations stored 743 /// to it in State/ClassMembers. 744 /// An equivalence class with one member might seem as it does not hold any 745 /// meaningful information, i.e. that is a tautology. However, during the 746 /// removal of dead symbols we do not remove classes with one member for 747 /// resource and performance reasons. Consequently, a class with one member is 748 /// not necessarily trivial. It could happen that we have a class with two 749 /// members and then during the removal of dead symbols we remove one of its 750 /// members. In this case, the class is still non-trivial (it still has the 751 /// mappings in ClassMembers), even though it has only one member. 752 LLVM_NODISCARD inline bool isTrivial(ProgramStateRef State) const; 753 754 /// Return true if the current class is trivial and its only member is dead. 755 LLVM_NODISCARD inline bool isTriviallyDead(ProgramStateRef State, 756 SymbolReaper &Reaper) const; 757 758 LLVM_NODISCARD static inline ProgramStateRef 759 markDisequal(RangeSet::Factory &F, ProgramStateRef State, SymbolRef First, 760 SymbolRef Second); 761 LLVM_NODISCARD static inline ProgramStateRef 762 markDisequal(RangeSet::Factory &F, ProgramStateRef State, 763 EquivalenceClass First, EquivalenceClass Second); 764 LLVM_NODISCARD inline ProgramStateRef 765 markDisequal(RangeSet::Factory &F, ProgramStateRef State, 766 EquivalenceClass Other) const; 767 LLVM_NODISCARD static inline ClassSet 768 getDisequalClasses(ProgramStateRef State, SymbolRef Sym); 769 LLVM_NODISCARD inline ClassSet 770 getDisequalClasses(ProgramStateRef State) const; 771 LLVM_NODISCARD inline ClassSet 772 getDisequalClasses(DisequalityMapTy Map, ClassSet::Factory &Factory) const; 773 774 LLVM_NODISCARD static inline Optional<bool> areEqual(ProgramStateRef State, 775 EquivalenceClass First, 776 EquivalenceClass Second); 777 LLVM_NODISCARD static inline Optional<bool> 778 areEqual(ProgramStateRef State, SymbolRef First, SymbolRef Second); 779 780 /// Remove one member from the class. 781 LLVM_NODISCARD ProgramStateRef removeMember(ProgramStateRef State, 782 const SymbolRef Old); 783 784 /// Iterate over all symbols and try to simplify them. 785 LLVM_NODISCARD static inline ProgramStateRef simplify(SValBuilder &SVB, 786 RangeSet::Factory &F, 787 ProgramStateRef State, 788 EquivalenceClass Class); 789 790 void dumpToStream(ProgramStateRef State, raw_ostream &os) const; 791 LLVM_DUMP_METHOD void dump(ProgramStateRef State) const { 792 dumpToStream(State, llvm::errs()); 793 } 794 795 /// Check equivalence data for consistency. 796 LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED static bool 797 isClassDataConsistent(ProgramStateRef State); 798 799 LLVM_NODISCARD QualType getType() const { 800 return getRepresentativeSymbol()->getType(); 801 } 802 803 EquivalenceClass() = delete; 804 EquivalenceClass(const EquivalenceClass &) = default; 805 EquivalenceClass &operator=(const EquivalenceClass &) = delete; 806 EquivalenceClass(EquivalenceClass &&) = default; 807 EquivalenceClass &operator=(EquivalenceClass &&) = delete; 808 809 bool operator==(const EquivalenceClass &Other) const { 810 return ID == Other.ID; 811 } 812 bool operator<(const EquivalenceClass &Other) const { return ID < Other.ID; } 813 bool operator!=(const EquivalenceClass &Other) const { 814 return !operator==(Other); 815 } 816 817 static void Profile(llvm::FoldingSetNodeID &ID, uintptr_t CID) { 818 ID.AddInteger(CID); 819 } 820 821 void Profile(llvm::FoldingSetNodeID &ID) const { Profile(ID, this->ID); } 822 823 private: 824 /* implicit */ EquivalenceClass(SymbolRef Sym) 825 : ID(reinterpret_cast<uintptr_t>(Sym)) {} 826 827 /// This function is intended to be used ONLY within the class. 828 /// The fact that ID is a pointer to a symbol is an implementation detail 829 /// and should stay that way. 830 /// In the current implementation, we use it to retrieve the only member 831 /// of the trivial class. 832 SymbolRef getRepresentativeSymbol() const { 833 return reinterpret_cast<SymbolRef>(ID); 834 } 835 static inline SymbolSet::Factory &getMembersFactory(ProgramStateRef State); 836 837 inline ProgramStateRef mergeImpl(RangeSet::Factory &F, ProgramStateRef State, 838 SymbolSet Members, EquivalenceClass Other, 839 SymbolSet OtherMembers); 840 841 static inline bool 842 addToDisequalityInfo(DisequalityMapTy &Info, ConstraintRangeTy &Constraints, 843 RangeSet::Factory &F, ProgramStateRef State, 844 EquivalenceClass First, EquivalenceClass Second); 845 846 /// This is a unique identifier of the class. 847 uintptr_t ID; 848 }; 849 850 //===----------------------------------------------------------------------===// 851 // Constraint functions 852 //===----------------------------------------------------------------------===// 853 854 LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED bool 855 areFeasible(ConstraintRangeTy Constraints) { 856 return llvm::none_of( 857 Constraints, 858 [](const std::pair<EquivalenceClass, RangeSet> &ClassConstraint) { 859 return ClassConstraint.second.isEmpty(); 860 }); 861 } 862 863 LLVM_NODISCARD inline const RangeSet *getConstraint(ProgramStateRef State, 864 EquivalenceClass Class) { 865 return State->get<ConstraintRange>(Class); 866 } 867 868 LLVM_NODISCARD inline const RangeSet *getConstraint(ProgramStateRef State, 869 SymbolRef Sym) { 870 return getConstraint(State, EquivalenceClass::find(State, Sym)); 871 } 872 873 LLVM_NODISCARD ProgramStateRef setConstraint(ProgramStateRef State, 874 EquivalenceClass Class, 875 RangeSet Constraint) { 876 return State->set<ConstraintRange>(Class, Constraint); 877 } 878 879 LLVM_NODISCARD ProgramStateRef setConstraints(ProgramStateRef State, 880 ConstraintRangeTy Constraints) { 881 return State->set<ConstraintRange>(Constraints); 882 } 883 884 //===----------------------------------------------------------------------===// 885 // Equality/diseqiality abstraction 886 //===----------------------------------------------------------------------===// 887 888 /// A small helper function for detecting symbolic (dis)equality. 889 /// 890 /// Equality check can have different forms (like a == b or a - b) and this 891 /// class encapsulates those away if the only thing the user wants to check - 892 /// whether it's equality/diseqiality or not. 893 /// 894 /// \returns true if assuming this Sym to be true means equality of operands 895 /// false if it means disequality of operands 896 /// None otherwise 897 Optional<bool> meansEquality(const SymSymExpr *Sym) { 898 switch (Sym->getOpcode()) { 899 case BO_Sub: 900 // This case is: A - B != 0 -> disequality check. 901 return false; 902 case BO_EQ: 903 // This case is: A == B != 0 -> equality check. 904 return true; 905 case BO_NE: 906 // This case is: A != B != 0 -> diseqiality check. 907 return false; 908 default: 909 return llvm::None; 910 } 911 } 912 913 //===----------------------------------------------------------------------===// 914 // Intersection functions 915 //===----------------------------------------------------------------------===// 916 917 template <class SecondTy, class... RestTy> 918 LLVM_NODISCARD inline RangeSet intersect(RangeSet::Factory &F, RangeSet Head, 919 SecondTy Second, RestTy... Tail); 920 921 template <class... RangeTy> struct IntersectionTraits; 922 923 template <class... TailTy> struct IntersectionTraits<RangeSet, TailTy...> { 924 // Found RangeSet, no need to check any further 925 using Type = RangeSet; 926 }; 927 928 template <> struct IntersectionTraits<> { 929 // We ran out of types, and we didn't find any RangeSet, so the result should 930 // be optional. 931 using Type = Optional<RangeSet>; 932 }; 933 934 template <class OptionalOrPointer, class... TailTy> 935 struct IntersectionTraits<OptionalOrPointer, TailTy...> { 936 // If current type is Optional or a raw pointer, we should keep looking. 937 using Type = typename IntersectionTraits<TailTy...>::Type; 938 }; 939 940 template <class EndTy> 941 LLVM_NODISCARD inline EndTy intersect(RangeSet::Factory &F, EndTy End) { 942 // If the list contains only RangeSet or Optional<RangeSet>, simply return 943 // that range set. 944 return End; 945 } 946 947 LLVM_NODISCARD LLVM_ATTRIBUTE_UNUSED inline Optional<RangeSet> 948 intersect(RangeSet::Factory &F, const RangeSet *End) { 949 // This is an extraneous conversion from a raw pointer into Optional<RangeSet> 950 if (End) { 951 return *End; 952 } 953 return llvm::None; 954 } 955 956 template <class... RestTy> 957 LLVM_NODISCARD inline RangeSet intersect(RangeSet::Factory &F, RangeSet Head, 958 RangeSet Second, RestTy... Tail) { 959 // Here we call either the <RangeSet,RangeSet,...> or <RangeSet,...> version 960 // of the function and can be sure that the result is RangeSet. 961 return intersect(F, F.intersect(Head, Second), Tail...); 962 } 963 964 template <class SecondTy, class... RestTy> 965 LLVM_NODISCARD inline RangeSet intersect(RangeSet::Factory &F, RangeSet Head, 966 SecondTy Second, RestTy... Tail) { 967 if (Second) { 968 // Here we call the <RangeSet,RangeSet,...> version of the function... 969 return intersect(F, Head, *Second, Tail...); 970 } 971 // ...and here it is either <RangeSet,RangeSet,...> or <RangeSet,...>, which 972 // means that the result is definitely RangeSet. 973 return intersect(F, Head, Tail...); 974 } 975 976 /// Main generic intersect function. 977 /// It intersects all of the given range sets. If some of the given arguments 978 /// don't hold a range set (nullptr or llvm::None), the function will skip them. 979 /// 980 /// Available representations for the arguments are: 981 /// * RangeSet 982 /// * Optional<RangeSet> 983 /// * RangeSet * 984 /// Pointer to a RangeSet is automatically assumed to be nullable and will get 985 /// checked as well as the optional version. If this behaviour is undesired, 986 /// please dereference the pointer in the call. 987 /// 988 /// Return type depends on the arguments' types. If we can be sure in compile 989 /// time that there will be a range set as a result, the returning type is 990 /// simply RangeSet, in other cases we have to back off to Optional<RangeSet>. 991 /// 992 /// Please, prefer optional range sets to raw pointers. If the last argument is 993 /// a raw pointer and all previous arguments are None, it will cost one 994 /// additional check to convert RangeSet * into Optional<RangeSet>. 995 template <class HeadTy, class SecondTy, class... RestTy> 996 LLVM_NODISCARD inline 997 typename IntersectionTraits<HeadTy, SecondTy, RestTy...>::Type 998 intersect(RangeSet::Factory &F, HeadTy Head, SecondTy Second, 999 RestTy... Tail) { 1000 if (Head) { 1001 return intersect(F, *Head, Second, Tail...); 1002 } 1003 return intersect(F, Second, Tail...); 1004 } 1005 1006 //===----------------------------------------------------------------------===// 1007 // Symbolic reasoning logic 1008 //===----------------------------------------------------------------------===// 1009 1010 /// A little component aggregating all of the reasoning we have about 1011 /// the ranges of symbolic expressions. 1012 /// 1013 /// Even when we don't know the exact values of the operands, we still 1014 /// can get a pretty good estimate of the result's range. 1015 class SymbolicRangeInferrer 1016 : public SymExprVisitor<SymbolicRangeInferrer, RangeSet> { 1017 public: 1018 template <class SourceType> 1019 static RangeSet inferRange(RangeSet::Factory &F, ProgramStateRef State, 1020 SourceType Origin) { 1021 SymbolicRangeInferrer Inferrer(F, State); 1022 return Inferrer.infer(Origin); 1023 } 1024 1025 RangeSet VisitSymExpr(SymbolRef Sym) { 1026 // If we got to this function, the actual type of the symbolic 1027 // expression is not supported for advanced inference. 1028 // In this case, we simply backoff to the default "let's simply 1029 // infer the range from the expression's type". 1030 return infer(Sym->getType()); 1031 } 1032 1033 RangeSet VisitSymIntExpr(const SymIntExpr *Sym) { 1034 return VisitBinaryOperator(Sym); 1035 } 1036 1037 RangeSet VisitIntSymExpr(const IntSymExpr *Sym) { 1038 return VisitBinaryOperator(Sym); 1039 } 1040 1041 RangeSet VisitSymSymExpr(const SymSymExpr *Sym) { 1042 return intersect( 1043 RangeFactory, 1044 // If Sym is (dis)equality, we might have some information 1045 // on that in our equality classes data structure. 1046 getRangeForEqualities(Sym), 1047 // And we should always check what we can get from the operands. 1048 VisitBinaryOperator(Sym)); 1049 } 1050 1051 private: 1052 SymbolicRangeInferrer(RangeSet::Factory &F, ProgramStateRef S) 1053 : ValueFactory(F.getValueFactory()), RangeFactory(F), State(S) {} 1054 1055 /// Infer range information from the given integer constant. 1056 /// 1057 /// It's not a real "inference", but is here for operating with 1058 /// sub-expressions in a more polymorphic manner. 1059 RangeSet inferAs(const llvm::APSInt &Val, QualType) { 1060 return {RangeFactory, Val}; 1061 } 1062 1063 /// Infer range information from symbol in the context of the given type. 1064 RangeSet inferAs(SymbolRef Sym, QualType DestType) { 1065 QualType ActualType = Sym->getType(); 1066 // Check that we can reason about the symbol at all. 1067 if (ActualType->isIntegralOrEnumerationType() || 1068 Loc::isLocType(ActualType)) { 1069 return infer(Sym); 1070 } 1071 // Otherwise, let's simply infer from the destination type. 1072 // We couldn't figure out nothing else about that expression. 1073 return infer(DestType); 1074 } 1075 1076 RangeSet infer(SymbolRef Sym) { 1077 return intersect( 1078 RangeFactory, 1079 // Of course, we should take the constraint directly associated with 1080 // this symbol into consideration. 1081 getConstraint(State, Sym), 1082 // If Sym is a difference of symbols A - B, then maybe we have range 1083 // set stored for B - A. 1084 // 1085 // If we have range set stored for both A - B and B - A then 1086 // calculate the effective range set by intersecting the range set 1087 // for A - B and the negated range set of B - A. 1088 getRangeForNegatedSub(Sym), 1089 // If Sym is a comparison expression (except <=>), 1090 // find any other comparisons with the same operands. 1091 // See function description. 1092 getRangeForComparisonSymbol(Sym), 1093 // Apart from the Sym itself, we can infer quite a lot if we look 1094 // into subexpressions of Sym. 1095 Visit(Sym)); 1096 } 1097 1098 RangeSet infer(EquivalenceClass Class) { 1099 if (const RangeSet *AssociatedConstraint = getConstraint(State, Class)) 1100 return *AssociatedConstraint; 1101 1102 return infer(Class.getType()); 1103 } 1104 1105 /// Infer range information solely from the type. 1106 RangeSet infer(QualType T) { 1107 // Lazily generate a new RangeSet representing all possible values for the 1108 // given symbol type. 1109 RangeSet Result(RangeFactory, ValueFactory.getMinValue(T), 1110 ValueFactory.getMaxValue(T)); 1111 1112 // References are known to be non-zero. 1113 if (T->isReferenceType()) 1114 return assumeNonZero(Result, T); 1115 1116 return Result; 1117 } 1118 1119 template <class BinarySymExprTy> 1120 RangeSet VisitBinaryOperator(const BinarySymExprTy *Sym) { 1121 // TODO #1: VisitBinaryOperator implementation might not make a good 1122 // use of the inferred ranges. In this case, we might be calculating 1123 // everything for nothing. This being said, we should introduce some 1124 // sort of laziness mechanism here. 1125 // 1126 // TODO #2: We didn't go into the nested expressions before, so it 1127 // might cause us spending much more time doing the inference. 1128 // This can be a problem for deeply nested expressions that are 1129 // involved in conditions and get tested continuously. We definitely 1130 // need to address this issue and introduce some sort of caching 1131 // in here. 1132 QualType ResultType = Sym->getType(); 1133 return VisitBinaryOperator(inferAs(Sym->getLHS(), ResultType), 1134 Sym->getOpcode(), 1135 inferAs(Sym->getRHS(), ResultType), ResultType); 1136 } 1137 1138 RangeSet VisitBinaryOperator(RangeSet LHS, BinaryOperator::Opcode Op, 1139 RangeSet RHS, QualType T) { 1140 switch (Op) { 1141 case BO_Or: 1142 return VisitBinaryOperator<BO_Or>(LHS, RHS, T); 1143 case BO_And: 1144 return VisitBinaryOperator<BO_And>(LHS, RHS, T); 1145 case BO_Rem: 1146 return VisitBinaryOperator<BO_Rem>(LHS, RHS, T); 1147 default: 1148 return infer(T); 1149 } 1150 } 1151 1152 //===----------------------------------------------------------------------===// 1153 // Ranges and operators 1154 //===----------------------------------------------------------------------===// 1155 1156 /// Return a rough approximation of the given range set. 1157 /// 1158 /// For the range set: 1159 /// { [x_0, y_0], [x_1, y_1], ... , [x_N, y_N] } 1160 /// it will return the range [x_0, y_N]. 1161 static Range fillGaps(RangeSet Origin) { 1162 assert(!Origin.isEmpty()); 1163 return {Origin.getMinValue(), Origin.getMaxValue()}; 1164 } 1165 1166 /// Try to convert given range into the given type. 1167 /// 1168 /// It will return llvm::None only when the trivial conversion is possible. 1169 llvm::Optional<Range> convert(const Range &Origin, APSIntType To) { 1170 if (To.testInRange(Origin.From(), false) != APSIntType::RTR_Within || 1171 To.testInRange(Origin.To(), false) != APSIntType::RTR_Within) { 1172 return llvm::None; 1173 } 1174 return Range(ValueFactory.Convert(To, Origin.From()), 1175 ValueFactory.Convert(To, Origin.To())); 1176 } 1177 1178 template <BinaryOperator::Opcode Op> 1179 RangeSet VisitBinaryOperator(RangeSet LHS, RangeSet RHS, QualType T) { 1180 // We should propagate information about unfeasbility of one of the 1181 // operands to the resulting range. 1182 if (LHS.isEmpty() || RHS.isEmpty()) { 1183 return RangeFactory.getEmptySet(); 1184 } 1185 1186 Range CoarseLHS = fillGaps(LHS); 1187 Range CoarseRHS = fillGaps(RHS); 1188 1189 APSIntType ResultType = ValueFactory.getAPSIntType(T); 1190 1191 // We need to convert ranges to the resulting type, so we can compare values 1192 // and combine them in a meaningful (in terms of the given operation) way. 1193 auto ConvertedCoarseLHS = convert(CoarseLHS, ResultType); 1194 auto ConvertedCoarseRHS = convert(CoarseRHS, ResultType); 1195 1196 // It is hard to reason about ranges when conversion changes 1197 // borders of the ranges. 1198 if (!ConvertedCoarseLHS || !ConvertedCoarseRHS) { 1199 return infer(T); 1200 } 1201 1202 return VisitBinaryOperator<Op>(*ConvertedCoarseLHS, *ConvertedCoarseRHS, T); 1203 } 1204 1205 template <BinaryOperator::Opcode Op> 1206 RangeSet VisitBinaryOperator(Range LHS, Range RHS, QualType T) { 1207 return infer(T); 1208 } 1209 1210 /// Return a symmetrical range for the given range and type. 1211 /// 1212 /// If T is signed, return the smallest range [-x..x] that covers the original 1213 /// range, or [-min(T), max(T)] if the aforementioned symmetric range doesn't 1214 /// exist due to original range covering min(T)). 1215 /// 1216 /// If T is unsigned, return the smallest range [0..x] that covers the 1217 /// original range. 1218 Range getSymmetricalRange(Range Origin, QualType T) { 1219 APSIntType RangeType = ValueFactory.getAPSIntType(T); 1220 1221 if (RangeType.isUnsigned()) { 1222 return Range(ValueFactory.getMinValue(RangeType), Origin.To()); 1223 } 1224 1225 if (Origin.From().isMinSignedValue()) { 1226 // If mini is a minimal signed value, absolute value of it is greater 1227 // than the maximal signed value. In order to avoid these 1228 // complications, we simply return the whole range. 1229 return {ValueFactory.getMinValue(RangeType), 1230 ValueFactory.getMaxValue(RangeType)}; 1231 } 1232 1233 // At this point, we are sure that the type is signed and we can safely 1234 // use unary - operator. 1235 // 1236 // While calculating absolute maximum, we can use the following formula 1237 // because of these reasons: 1238 // * If From >= 0 then To >= From and To >= -From. 1239 // AbsMax == To == max(To, -From) 1240 // * If To <= 0 then -From >= -To and -From >= From. 1241 // AbsMax == -From == max(-From, To) 1242 // * Otherwise, From <= 0, To >= 0, and 1243 // AbsMax == max(abs(From), abs(To)) 1244 llvm::APSInt AbsMax = std::max(-Origin.From(), Origin.To()); 1245 1246 // Intersection is guaranteed to be non-empty. 1247 return {ValueFactory.getValue(-AbsMax), ValueFactory.getValue(AbsMax)}; 1248 } 1249 1250 /// Return a range set subtracting zero from \p Domain. 1251 RangeSet assumeNonZero(RangeSet Domain, QualType T) { 1252 APSIntType IntType = ValueFactory.getAPSIntType(T); 1253 return RangeFactory.deletePoint(Domain, IntType.getZeroValue()); 1254 } 1255 1256 // FIXME: Once SValBuilder supports unary minus, we should use SValBuilder to 1257 // obtain the negated symbolic expression instead of constructing the 1258 // symbol manually. This will allow us to support finding ranges of not 1259 // only negated SymSymExpr-type expressions, but also of other, simpler 1260 // expressions which we currently do not know how to negate. 1261 Optional<RangeSet> getRangeForNegatedSub(SymbolRef Sym) { 1262 if (const SymSymExpr *SSE = dyn_cast<SymSymExpr>(Sym)) { 1263 if (SSE->getOpcode() == BO_Sub) { 1264 QualType T = Sym->getType(); 1265 1266 // Do not negate unsigned ranges 1267 if (!T->isUnsignedIntegerOrEnumerationType() && 1268 !T->isSignedIntegerOrEnumerationType()) 1269 return llvm::None; 1270 1271 SymbolManager &SymMgr = State->getSymbolManager(); 1272 SymbolRef NegatedSym = 1273 SymMgr.getSymSymExpr(SSE->getRHS(), BO_Sub, SSE->getLHS(), T); 1274 1275 if (const RangeSet *NegatedRange = getConstraint(State, NegatedSym)) { 1276 return RangeFactory.negate(*NegatedRange); 1277 } 1278 } 1279 } 1280 return llvm::None; 1281 } 1282 1283 // Returns ranges only for binary comparison operators (except <=>) 1284 // when left and right operands are symbolic values. 1285 // Finds any other comparisons with the same operands. 1286 // Then do logical calculations and refuse impossible branches. 1287 // E.g. (x < y) and (x > y) at the same time are impossible. 1288 // E.g. (x >= y) and (x != y) at the same time makes (x > y) true only. 1289 // E.g. (x == y) and (y == x) are just reversed but the same. 1290 // It covers all possible combinations (see CmpOpTable description). 1291 // Note that `x` and `y` can also stand for subexpressions, 1292 // not only for actual symbols. 1293 Optional<RangeSet> getRangeForComparisonSymbol(SymbolRef Sym) { 1294 const auto *SSE = dyn_cast<SymSymExpr>(Sym); 1295 if (!SSE) 1296 return llvm::None; 1297 1298 const BinaryOperatorKind CurrentOP = SSE->getOpcode(); 1299 1300 // We currently do not support <=> (C++20). 1301 if (!BinaryOperator::isComparisonOp(CurrentOP) || (CurrentOP == BO_Cmp)) 1302 return llvm::None; 1303 1304 static const OperatorRelationsTable CmpOpTable{}; 1305 1306 const SymExpr *LHS = SSE->getLHS(); 1307 const SymExpr *RHS = SSE->getRHS(); 1308 QualType T = SSE->getType(); 1309 1310 SymbolManager &SymMgr = State->getSymbolManager(); 1311 1312 // We use this variable to store the last queried operator (`QueriedOP`) 1313 // for which the `getCmpOpState` returned with `Unknown`. If there are two 1314 // different OPs that returned `Unknown` then we have to query the special 1315 // `UnknownX2` column. We assume that `getCmpOpState(CurrentOP, CurrentOP)` 1316 // never returns `Unknown`, so `CurrentOP` is a good initial value. 1317 BinaryOperatorKind LastQueriedOpToUnknown = CurrentOP; 1318 1319 // Loop goes through all of the columns exept the last one ('UnknownX2'). 1320 // We treat `UnknownX2` column separately at the end of the loop body. 1321 for (size_t i = 0; i < CmpOpTable.getCmpOpCount(); ++i) { 1322 1323 // Let's find an expression e.g. (x < y). 1324 BinaryOperatorKind QueriedOP = OperatorRelationsTable::getOpFromIndex(i); 1325 const SymSymExpr *SymSym = SymMgr.getSymSymExpr(LHS, QueriedOP, RHS, T); 1326 const RangeSet *QueriedRangeSet = getConstraint(State, SymSym); 1327 1328 // If ranges were not previously found, 1329 // try to find a reversed expression (y > x). 1330 if (!QueriedRangeSet) { 1331 const BinaryOperatorKind ROP = 1332 BinaryOperator::reverseComparisonOp(QueriedOP); 1333 SymSym = SymMgr.getSymSymExpr(RHS, ROP, LHS, T); 1334 QueriedRangeSet = getConstraint(State, SymSym); 1335 } 1336 1337 if (!QueriedRangeSet || QueriedRangeSet->isEmpty()) 1338 continue; 1339 1340 const llvm::APSInt *ConcreteValue = QueriedRangeSet->getConcreteValue(); 1341 const bool isInFalseBranch = 1342 ConcreteValue ? (*ConcreteValue == 0) : false; 1343 1344 // If it is a false branch, we shall be guided by opposite operator, 1345 // because the table is made assuming we are in the true branch. 1346 // E.g. when (x <= y) is false, then (x > y) is true. 1347 if (isInFalseBranch) 1348 QueriedOP = BinaryOperator::negateComparisonOp(QueriedOP); 1349 1350 OperatorRelationsTable::TriStateKind BranchState = 1351 CmpOpTable.getCmpOpState(CurrentOP, QueriedOP); 1352 1353 if (BranchState == OperatorRelationsTable::Unknown) { 1354 if (LastQueriedOpToUnknown != CurrentOP && 1355 LastQueriedOpToUnknown != QueriedOP) { 1356 // If we got the Unknown state for both different operators. 1357 // if (x <= y) // assume true 1358 // if (x != y) // assume true 1359 // if (x < y) // would be also true 1360 // Get a state from `UnknownX2` column. 1361 BranchState = CmpOpTable.getCmpOpStateForUnknownX2(CurrentOP); 1362 } else { 1363 LastQueriedOpToUnknown = QueriedOP; 1364 continue; 1365 } 1366 } 1367 1368 return (BranchState == OperatorRelationsTable::True) ? getTrueRange(T) 1369 : getFalseRange(T); 1370 } 1371 1372 return llvm::None; 1373 } 1374 1375 Optional<RangeSet> getRangeForEqualities(const SymSymExpr *Sym) { 1376 Optional<bool> Equality = meansEquality(Sym); 1377 1378 if (!Equality) 1379 return llvm::None; 1380 1381 if (Optional<bool> AreEqual = 1382 EquivalenceClass::areEqual(State, Sym->getLHS(), Sym->getRHS())) { 1383 // Here we cover two cases at once: 1384 // * if Sym is equality and its operands are known to be equal -> true 1385 // * if Sym is disequality and its operands are disequal -> true 1386 if (*AreEqual == *Equality) { 1387 return getTrueRange(Sym->getType()); 1388 } 1389 // Opposite combinations result in false. 1390 return getFalseRange(Sym->getType()); 1391 } 1392 1393 return llvm::None; 1394 } 1395 1396 RangeSet getTrueRange(QualType T) { 1397 RangeSet TypeRange = infer(T); 1398 return assumeNonZero(TypeRange, T); 1399 } 1400 1401 RangeSet getFalseRange(QualType T) { 1402 const llvm::APSInt &Zero = ValueFactory.getValue(0, T); 1403 return RangeSet(RangeFactory, Zero); 1404 } 1405 1406 BasicValueFactory &ValueFactory; 1407 RangeSet::Factory &RangeFactory; 1408 ProgramStateRef State; 1409 }; 1410 1411 //===----------------------------------------------------------------------===// 1412 // Range-based reasoning about symbolic operations 1413 //===----------------------------------------------------------------------===// 1414 1415 template <> 1416 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_Or>(Range LHS, Range RHS, 1417 QualType T) { 1418 APSIntType ResultType = ValueFactory.getAPSIntType(T); 1419 llvm::APSInt Zero = ResultType.getZeroValue(); 1420 1421 bool IsLHSPositiveOrZero = LHS.From() >= Zero; 1422 bool IsRHSPositiveOrZero = RHS.From() >= Zero; 1423 1424 bool IsLHSNegative = LHS.To() < Zero; 1425 bool IsRHSNegative = RHS.To() < Zero; 1426 1427 // Check if both ranges have the same sign. 1428 if ((IsLHSPositiveOrZero && IsRHSPositiveOrZero) || 1429 (IsLHSNegative && IsRHSNegative)) { 1430 // The result is definitely greater or equal than any of the operands. 1431 const llvm::APSInt &Min = std::max(LHS.From(), RHS.From()); 1432 1433 // We estimate maximal value for positives as the maximal value for the 1434 // given type. For negatives, we estimate it with -1 (e.g. 0x11111111). 1435 // 1436 // TODO: We basically, limit the resulting range from below, but don't do 1437 // anything with the upper bound. 1438 // 1439 // For positive operands, it can be done as follows: for the upper 1440 // bound of LHS and RHS we calculate the most significant bit set. 1441 // Let's call it the N-th bit. Then we can estimate the maximal 1442 // number to be 2^(N+1)-1, i.e. the number with all the bits up to 1443 // the N-th bit set. 1444 const llvm::APSInt &Max = IsLHSNegative 1445 ? ValueFactory.getValue(--Zero) 1446 : ValueFactory.getMaxValue(ResultType); 1447 1448 return {RangeFactory, ValueFactory.getValue(Min), Max}; 1449 } 1450 1451 // Otherwise, let's check if at least one of the operands is negative. 1452 if (IsLHSNegative || IsRHSNegative) { 1453 // This means that the result is definitely negative as well. 1454 return {RangeFactory, ValueFactory.getMinValue(ResultType), 1455 ValueFactory.getValue(--Zero)}; 1456 } 1457 1458 RangeSet DefaultRange = infer(T); 1459 1460 // It is pretty hard to reason about operands with different signs 1461 // (and especially with possibly different signs). We simply check if it 1462 // can be zero. In order to conclude that the result could not be zero, 1463 // at least one of the operands should be definitely not zero itself. 1464 if (!LHS.Includes(Zero) || !RHS.Includes(Zero)) { 1465 return assumeNonZero(DefaultRange, T); 1466 } 1467 1468 // Nothing much else to do here. 1469 return DefaultRange; 1470 } 1471 1472 template <> 1473 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_And>(Range LHS, 1474 Range RHS, 1475 QualType T) { 1476 APSIntType ResultType = ValueFactory.getAPSIntType(T); 1477 llvm::APSInt Zero = ResultType.getZeroValue(); 1478 1479 bool IsLHSPositiveOrZero = LHS.From() >= Zero; 1480 bool IsRHSPositiveOrZero = RHS.From() >= Zero; 1481 1482 bool IsLHSNegative = LHS.To() < Zero; 1483 bool IsRHSNegative = RHS.To() < Zero; 1484 1485 // Check if both ranges have the same sign. 1486 if ((IsLHSPositiveOrZero && IsRHSPositiveOrZero) || 1487 (IsLHSNegative && IsRHSNegative)) { 1488 // The result is definitely less or equal than any of the operands. 1489 const llvm::APSInt &Max = std::min(LHS.To(), RHS.To()); 1490 1491 // We conservatively estimate lower bound to be the smallest positive 1492 // or negative value corresponding to the sign of the operands. 1493 const llvm::APSInt &Min = IsLHSNegative 1494 ? ValueFactory.getMinValue(ResultType) 1495 : ValueFactory.getValue(Zero); 1496 1497 return {RangeFactory, Min, Max}; 1498 } 1499 1500 // Otherwise, let's check if at least one of the operands is positive. 1501 if (IsLHSPositiveOrZero || IsRHSPositiveOrZero) { 1502 // This makes result definitely positive. 1503 // 1504 // We can also reason about a maximal value by finding the maximal 1505 // value of the positive operand. 1506 const llvm::APSInt &Max = IsLHSPositiveOrZero ? LHS.To() : RHS.To(); 1507 1508 // The minimal value on the other hand is much harder to reason about. 1509 // The only thing we know for sure is that the result is positive. 1510 return {RangeFactory, ValueFactory.getValue(Zero), 1511 ValueFactory.getValue(Max)}; 1512 } 1513 1514 // Nothing much else to do here. 1515 return infer(T); 1516 } 1517 1518 template <> 1519 RangeSet SymbolicRangeInferrer::VisitBinaryOperator<BO_Rem>(Range LHS, 1520 Range RHS, 1521 QualType T) { 1522 llvm::APSInt Zero = ValueFactory.getAPSIntType(T).getZeroValue(); 1523 1524 Range ConservativeRange = getSymmetricalRange(RHS, T); 1525 1526 llvm::APSInt Max = ConservativeRange.To(); 1527 llvm::APSInt Min = ConservativeRange.From(); 1528 1529 if (Max == Zero) { 1530 // It's an undefined behaviour to divide by 0 and it seems like we know 1531 // for sure that RHS is 0. Let's say that the resulting range is 1532 // simply infeasible for that matter. 1533 return RangeFactory.getEmptySet(); 1534 } 1535 1536 // At this point, our conservative range is closed. The result, however, 1537 // couldn't be greater than the RHS' maximal absolute value. Because of 1538 // this reason, we turn the range into open (or half-open in case of 1539 // unsigned integers). 1540 // 1541 // While we operate on integer values, an open interval (a, b) can be easily 1542 // represented by the closed interval [a + 1, b - 1]. And this is exactly 1543 // what we do next. 1544 // 1545 // If we are dealing with unsigned case, we shouldn't move the lower bound. 1546 if (Min.isSigned()) { 1547 ++Min; 1548 } 1549 --Max; 1550 1551 bool IsLHSPositiveOrZero = LHS.From() >= Zero; 1552 bool IsRHSPositiveOrZero = RHS.From() >= Zero; 1553 1554 // Remainder operator results with negative operands is implementation 1555 // defined. Positive cases are much easier to reason about though. 1556 if (IsLHSPositiveOrZero && IsRHSPositiveOrZero) { 1557 // If maximal value of LHS is less than maximal value of RHS, 1558 // the result won't get greater than LHS.To(). 1559 Max = std::min(LHS.To(), Max); 1560 // We want to check if it is a situation similar to the following: 1561 // 1562 // <------------|---[ LHS ]--------[ RHS ]-----> 1563 // -INF 0 +INF 1564 // 1565 // In this situation, we can conclude that (LHS / RHS) == 0 and 1566 // (LHS % RHS) == LHS. 1567 Min = LHS.To() < RHS.From() ? LHS.From() : Zero; 1568 } 1569 1570 // Nevertheless, the symmetrical range for RHS is a conservative estimate 1571 // for any sign of either LHS, or RHS. 1572 return {RangeFactory, ValueFactory.getValue(Min), ValueFactory.getValue(Max)}; 1573 } 1574 1575 //===----------------------------------------------------------------------===// 1576 // Constraint manager implementation details 1577 //===----------------------------------------------------------------------===// 1578 1579 class RangeConstraintManager : public RangedConstraintManager { 1580 public: 1581 RangeConstraintManager(ExprEngine *EE, SValBuilder &SVB) 1582 : RangedConstraintManager(EE, SVB), F(getBasicVals()) {} 1583 1584 //===------------------------------------------------------------------===// 1585 // Implementation for interface from ConstraintManager. 1586 //===------------------------------------------------------------------===// 1587 1588 bool haveEqualConstraints(ProgramStateRef S1, 1589 ProgramStateRef S2) const override { 1590 // NOTE: ClassMembers are as simple as back pointers for ClassMap, 1591 // so comparing constraint ranges and class maps should be 1592 // sufficient. 1593 return S1->get<ConstraintRange>() == S2->get<ConstraintRange>() && 1594 S1->get<ClassMap>() == S2->get<ClassMap>(); 1595 } 1596 1597 bool canReasonAbout(SVal X) const override; 1598 1599 ConditionTruthVal checkNull(ProgramStateRef State, SymbolRef Sym) override; 1600 1601 const llvm::APSInt *getSymVal(ProgramStateRef State, 1602 SymbolRef Sym) const override; 1603 1604 ProgramStateRef removeDeadBindings(ProgramStateRef State, 1605 SymbolReaper &SymReaper) override; 1606 1607 void printJson(raw_ostream &Out, ProgramStateRef State, const char *NL = "\n", 1608 unsigned int Space = 0, bool IsDot = false) const override; 1609 void printConstraints(raw_ostream &Out, ProgramStateRef State, 1610 const char *NL = "\n", unsigned int Space = 0, 1611 bool IsDot = false) const; 1612 void printEquivalenceClasses(raw_ostream &Out, ProgramStateRef State, 1613 const char *NL = "\n", unsigned int Space = 0, 1614 bool IsDot = false) const; 1615 void printDisequalities(raw_ostream &Out, ProgramStateRef State, 1616 const char *NL = "\n", unsigned int Space = 0, 1617 bool IsDot = false) const; 1618 1619 //===------------------------------------------------------------------===// 1620 // Implementation for interface from RangedConstraintManager. 1621 //===------------------------------------------------------------------===// 1622 1623 ProgramStateRef assumeSymNE(ProgramStateRef State, SymbolRef Sym, 1624 const llvm::APSInt &V, 1625 const llvm::APSInt &Adjustment) override; 1626 1627 ProgramStateRef assumeSymEQ(ProgramStateRef State, SymbolRef Sym, 1628 const llvm::APSInt &V, 1629 const llvm::APSInt &Adjustment) override; 1630 1631 ProgramStateRef assumeSymLT(ProgramStateRef State, SymbolRef Sym, 1632 const llvm::APSInt &V, 1633 const llvm::APSInt &Adjustment) override; 1634 1635 ProgramStateRef assumeSymGT(ProgramStateRef State, SymbolRef Sym, 1636 const llvm::APSInt &V, 1637 const llvm::APSInt &Adjustment) override; 1638 1639 ProgramStateRef assumeSymLE(ProgramStateRef State, SymbolRef Sym, 1640 const llvm::APSInt &V, 1641 const llvm::APSInt &Adjustment) override; 1642 1643 ProgramStateRef assumeSymGE(ProgramStateRef State, SymbolRef Sym, 1644 const llvm::APSInt &V, 1645 const llvm::APSInt &Adjustment) override; 1646 1647 ProgramStateRef assumeSymWithinInclusiveRange( 1648 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 1649 const llvm::APSInt &To, const llvm::APSInt &Adjustment) override; 1650 1651 ProgramStateRef assumeSymOutsideInclusiveRange( 1652 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 1653 const llvm::APSInt &To, const llvm::APSInt &Adjustment) override; 1654 1655 private: 1656 RangeSet::Factory F; 1657 1658 RangeSet getRange(ProgramStateRef State, SymbolRef Sym); 1659 RangeSet getRange(ProgramStateRef State, EquivalenceClass Class); 1660 ProgramStateRef setRange(ProgramStateRef State, SymbolRef Sym, 1661 RangeSet Range); 1662 ProgramStateRef setRange(ProgramStateRef State, EquivalenceClass Class, 1663 RangeSet Range); 1664 1665 RangeSet getSymLTRange(ProgramStateRef St, SymbolRef Sym, 1666 const llvm::APSInt &Int, 1667 const llvm::APSInt &Adjustment); 1668 RangeSet getSymGTRange(ProgramStateRef St, SymbolRef Sym, 1669 const llvm::APSInt &Int, 1670 const llvm::APSInt &Adjustment); 1671 RangeSet getSymLERange(ProgramStateRef St, SymbolRef Sym, 1672 const llvm::APSInt &Int, 1673 const llvm::APSInt &Adjustment); 1674 RangeSet getSymLERange(llvm::function_ref<RangeSet()> RS, 1675 const llvm::APSInt &Int, 1676 const llvm::APSInt &Adjustment); 1677 RangeSet getSymGERange(ProgramStateRef St, SymbolRef Sym, 1678 const llvm::APSInt &Int, 1679 const llvm::APSInt &Adjustment); 1680 }; 1681 1682 //===----------------------------------------------------------------------===// 1683 // Constraint assignment logic 1684 //===----------------------------------------------------------------------===// 1685 1686 /// ConstraintAssignorBase is a small utility class that unifies visitor 1687 /// for ranges with a visitor for constraints (rangeset/range/constant). 1688 /// 1689 /// It is designed to have one derived class, but generally it can have more. 1690 /// Derived class can control which types we handle by defining methods of the 1691 /// following form: 1692 /// 1693 /// bool handle${SYMBOL}To${CONSTRAINT}(const SYMBOL *Sym, 1694 /// CONSTRAINT Constraint); 1695 /// 1696 /// where SYMBOL is the type of the symbol (e.g. SymSymExpr, SymbolCast, etc.) 1697 /// CONSTRAINT is the type of constraint (RangeSet/Range/Const) 1698 /// return value signifies whether we should try other handle methods 1699 /// (i.e. false would mean to stop right after calling this method) 1700 template <class Derived> class ConstraintAssignorBase { 1701 public: 1702 using Const = const llvm::APSInt &; 1703 1704 #define DISPATCH(CLASS) return assign##CLASS##Impl(cast<CLASS>(Sym), Constraint) 1705 1706 #define ASSIGN(CLASS, TO, SYM, CONSTRAINT) \ 1707 if (!static_cast<Derived *>(this)->assign##CLASS##To##TO(SYM, CONSTRAINT)) \ 1708 return false 1709 1710 void assign(SymbolRef Sym, RangeSet Constraint) { 1711 assignImpl(Sym, Constraint); 1712 } 1713 1714 bool assignImpl(SymbolRef Sym, RangeSet Constraint) { 1715 switch (Sym->getKind()) { 1716 #define SYMBOL(Id, Parent) \ 1717 case SymExpr::Id##Kind: \ 1718 DISPATCH(Id); 1719 #include "clang/StaticAnalyzer/Core/PathSensitive/Symbols.def" 1720 } 1721 llvm_unreachable("Unknown SymExpr kind!"); 1722 } 1723 1724 #define DEFAULT_ASSIGN(Id) \ 1725 bool assign##Id##To##RangeSet(const Id *Sym, RangeSet Constraint) { \ 1726 return true; \ 1727 } \ 1728 bool assign##Id##To##Range(const Id *Sym, Range Constraint) { return true; } \ 1729 bool assign##Id##To##Const(const Id *Sym, Const Constraint) { return true; } 1730 1731 // When we dispatch for constraint types, we first try to check 1732 // if the new constraint is the constant and try the corresponding 1733 // assignor methods. If it didn't interrupt, we can proceed to the 1734 // range, and finally to the range set. 1735 #define CONSTRAINT_DISPATCH(Id) \ 1736 if (const llvm::APSInt *Const = Constraint.getConcreteValue()) { \ 1737 ASSIGN(Id, Const, Sym, *Const); \ 1738 } \ 1739 if (Constraint.size() == 1) { \ 1740 ASSIGN(Id, Range, Sym, *Constraint.begin()); \ 1741 } \ 1742 ASSIGN(Id, RangeSet, Sym, Constraint) 1743 1744 // Our internal assign method first tries to call assignor methods for all 1745 // constraint types that apply. And if not interrupted, continues with its 1746 // parent class. 1747 #define SYMBOL(Id, Parent) \ 1748 bool assign##Id##Impl(const Id *Sym, RangeSet Constraint) { \ 1749 CONSTRAINT_DISPATCH(Id); \ 1750 DISPATCH(Parent); \ 1751 } \ 1752 DEFAULT_ASSIGN(Id) 1753 #define ABSTRACT_SYMBOL(Id, Parent) SYMBOL(Id, Parent) 1754 #include "clang/StaticAnalyzer/Core/PathSensitive/Symbols.def" 1755 1756 // Default implementations for the top class that doesn't have parents. 1757 bool assignSymExprImpl(const SymExpr *Sym, RangeSet Constraint) { 1758 CONSTRAINT_DISPATCH(SymExpr); 1759 return true; 1760 } 1761 DEFAULT_ASSIGN(SymExpr); 1762 1763 #undef DISPATCH 1764 #undef CONSTRAINT_DISPATCH 1765 #undef DEFAULT_ASSIGN 1766 #undef ASSIGN 1767 }; 1768 1769 /// A little component aggregating all of the reasoning we have about 1770 /// assigning new constraints to symbols. 1771 /// 1772 /// The main purpose of this class is to associate constraints to symbols, 1773 /// and impose additional constraints on other symbols, when we can imply 1774 /// them. 1775 /// 1776 /// It has a nice symmetry with SymbolicRangeInferrer. When the latter 1777 /// can provide more precise ranges by looking into the operands of the 1778 /// expression in question, ConstraintAssignor looks into the operands 1779 /// to see if we can imply more from the new constraint. 1780 class ConstraintAssignor : public ConstraintAssignorBase<ConstraintAssignor> { 1781 public: 1782 template <class ClassOrSymbol> 1783 LLVM_NODISCARD static ProgramStateRef 1784 assign(ProgramStateRef State, SValBuilder &Builder, RangeSet::Factory &F, 1785 ClassOrSymbol CoS, RangeSet NewConstraint) { 1786 if (!State || NewConstraint.isEmpty()) 1787 return nullptr; 1788 1789 ConstraintAssignor Assignor{State, Builder, F}; 1790 return Assignor.assign(CoS, NewConstraint); 1791 } 1792 1793 /// Handle expressions like: a % b != 0. 1794 template <typename SymT> 1795 bool handleRemainderOp(const SymT *Sym, RangeSet Constraint) { 1796 if (Sym->getOpcode() != BO_Rem) 1797 return true; 1798 // a % b != 0 implies that a != 0. 1799 if (!Constraint.containsZero()) { 1800 SVal SymSVal = Builder.makeSymbolVal(Sym->getLHS()); 1801 if (auto NonLocSymSVal = SymSVal.getAs<nonloc::SymbolVal>()) { 1802 State = State->assume(*NonLocSymSVal, true); 1803 if (!State) 1804 return false; 1805 } 1806 } 1807 return true; 1808 } 1809 1810 inline bool assignSymExprToConst(const SymExpr *Sym, Const Constraint); 1811 inline bool assignSymIntExprToRangeSet(const SymIntExpr *Sym, 1812 RangeSet Constraint) { 1813 return handleRemainderOp(Sym, Constraint); 1814 } 1815 inline bool assignSymSymExprToRangeSet(const SymSymExpr *Sym, 1816 RangeSet Constraint); 1817 1818 private: 1819 ConstraintAssignor(ProgramStateRef State, SValBuilder &Builder, 1820 RangeSet::Factory &F) 1821 : State(State), Builder(Builder), RangeFactory(F) {} 1822 using Base = ConstraintAssignorBase<ConstraintAssignor>; 1823 1824 /// Base method for handling new constraints for symbols. 1825 LLVM_NODISCARD ProgramStateRef assign(SymbolRef Sym, RangeSet NewConstraint) { 1826 // All constraints are actually associated with equivalence classes, and 1827 // that's what we are going to do first. 1828 State = assign(EquivalenceClass::find(State, Sym), NewConstraint); 1829 if (!State) 1830 return nullptr; 1831 1832 // And after that we can check what other things we can get from this 1833 // constraint. 1834 Base::assign(Sym, NewConstraint); 1835 return State; 1836 } 1837 1838 /// Base method for handling new constraints for classes. 1839 LLVM_NODISCARD ProgramStateRef assign(EquivalenceClass Class, 1840 RangeSet NewConstraint) { 1841 // There is a chance that we might need to update constraints for the 1842 // classes that are known to be disequal to Class. 1843 // 1844 // In order for this to be even possible, the new constraint should 1845 // be simply a constant because we can't reason about range disequalities. 1846 if (const llvm::APSInt *Point = NewConstraint.getConcreteValue()) { 1847 1848 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1849 ConstraintRangeTy::Factory &CF = State->get_context<ConstraintRange>(); 1850 1851 // Add new constraint. 1852 Constraints = CF.add(Constraints, Class, NewConstraint); 1853 1854 for (EquivalenceClass DisequalClass : Class.getDisequalClasses(State)) { 1855 RangeSet UpdatedConstraint = SymbolicRangeInferrer::inferRange( 1856 RangeFactory, State, DisequalClass); 1857 1858 UpdatedConstraint = RangeFactory.deletePoint(UpdatedConstraint, *Point); 1859 1860 // If we end up with at least one of the disequal classes to be 1861 // constrained with an empty range-set, the state is infeasible. 1862 if (UpdatedConstraint.isEmpty()) 1863 return nullptr; 1864 1865 Constraints = CF.add(Constraints, DisequalClass, UpdatedConstraint); 1866 } 1867 assert(areFeasible(Constraints) && "Constraint manager shouldn't produce " 1868 "a state with infeasible constraints"); 1869 1870 return setConstraints(State, Constraints); 1871 } 1872 1873 return setConstraint(State, Class, NewConstraint); 1874 } 1875 1876 ProgramStateRef trackDisequality(ProgramStateRef State, SymbolRef LHS, 1877 SymbolRef RHS) { 1878 return EquivalenceClass::markDisequal(RangeFactory, State, LHS, RHS); 1879 } 1880 1881 ProgramStateRef trackEquality(ProgramStateRef State, SymbolRef LHS, 1882 SymbolRef RHS) { 1883 return EquivalenceClass::merge(RangeFactory, State, LHS, RHS); 1884 } 1885 1886 LLVM_NODISCARD Optional<bool> interpreteAsBool(RangeSet Constraint) { 1887 assert(!Constraint.isEmpty() && "Empty ranges shouldn't get here"); 1888 1889 if (Constraint.getConcreteValue()) 1890 return !Constraint.getConcreteValue()->isZero(); 1891 1892 if (!Constraint.containsZero()) 1893 return true; 1894 1895 return llvm::None; 1896 } 1897 1898 ProgramStateRef State; 1899 SValBuilder &Builder; 1900 RangeSet::Factory &RangeFactory; 1901 }; 1902 1903 1904 bool ConstraintAssignor::assignSymExprToConst(const SymExpr *Sym, 1905 const llvm::APSInt &Constraint) { 1906 llvm::SmallSet<EquivalenceClass, 4> SimplifiedClasses; 1907 // Iterate over all equivalence classes and try to simplify them. 1908 ClassMembersTy Members = State->get<ClassMembers>(); 1909 for (std::pair<EquivalenceClass, SymbolSet> ClassToSymbolSet : Members) { 1910 EquivalenceClass Class = ClassToSymbolSet.first; 1911 State = EquivalenceClass::simplify(Builder, RangeFactory, State, Class); 1912 if (!State) 1913 return false; 1914 SimplifiedClasses.insert(Class); 1915 } 1916 1917 // Trivial equivalence classes (those that have only one symbol member) are 1918 // not stored in the State. Thus, we must skim through the constraints as 1919 // well. And we try to simplify symbols in the constraints. 1920 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1921 for (std::pair<EquivalenceClass, RangeSet> ClassConstraint : Constraints) { 1922 EquivalenceClass Class = ClassConstraint.first; 1923 if (SimplifiedClasses.count(Class)) // Already simplified. 1924 continue; 1925 State = EquivalenceClass::simplify(Builder, RangeFactory, State, Class); 1926 if (!State) 1927 return false; 1928 } 1929 1930 // We may have trivial equivalence classes in the disequality info as 1931 // well, and we need to simplify them. 1932 DisequalityMapTy DisequalityInfo = State->get<DisequalityMap>(); 1933 for (std::pair<EquivalenceClass, ClassSet> DisequalityEntry : 1934 DisequalityInfo) { 1935 EquivalenceClass Class = DisequalityEntry.first; 1936 ClassSet DisequalClasses = DisequalityEntry.second; 1937 State = EquivalenceClass::simplify(Builder, RangeFactory, State, Class); 1938 if (!State) 1939 return false; 1940 } 1941 1942 return true; 1943 } 1944 1945 bool ConstraintAssignor::assignSymSymExprToRangeSet(const SymSymExpr *Sym, 1946 RangeSet Constraint) { 1947 if (!handleRemainderOp(Sym, Constraint)) 1948 return false; 1949 1950 Optional<bool> ConstraintAsBool = interpreteAsBool(Constraint); 1951 1952 if (!ConstraintAsBool) 1953 return true; 1954 1955 if (Optional<bool> Equality = meansEquality(Sym)) { 1956 // Here we cover two cases: 1957 // * if Sym is equality and the new constraint is true -> Sym's operands 1958 // should be marked as equal 1959 // * if Sym is disequality and the new constraint is false -> Sym's 1960 // operands should be also marked as equal 1961 if (*Equality == *ConstraintAsBool) { 1962 State = trackEquality(State, Sym->getLHS(), Sym->getRHS()); 1963 } else { 1964 // Other combinations leave as with disequal operands. 1965 State = trackDisequality(State, Sym->getLHS(), Sym->getRHS()); 1966 } 1967 1968 if (!State) 1969 return false; 1970 } 1971 1972 return true; 1973 } 1974 1975 } // end anonymous namespace 1976 1977 std::unique_ptr<ConstraintManager> 1978 ento::CreateRangeConstraintManager(ProgramStateManager &StMgr, 1979 ExprEngine *Eng) { 1980 return std::make_unique<RangeConstraintManager>(Eng, StMgr.getSValBuilder()); 1981 } 1982 1983 ConstraintMap ento::getConstraintMap(ProgramStateRef State) { 1984 ConstraintMap::Factory &F = State->get_context<ConstraintMap>(); 1985 ConstraintMap Result = F.getEmptyMap(); 1986 1987 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 1988 for (std::pair<EquivalenceClass, RangeSet> ClassConstraint : Constraints) { 1989 EquivalenceClass Class = ClassConstraint.first; 1990 SymbolSet ClassMembers = Class.getClassMembers(State); 1991 assert(!ClassMembers.isEmpty() && 1992 "Class must always have at least one member!"); 1993 1994 SymbolRef Representative = *ClassMembers.begin(); 1995 Result = F.add(Result, Representative, ClassConstraint.second); 1996 } 1997 1998 return Result; 1999 } 2000 2001 //===----------------------------------------------------------------------===// 2002 // EqualityClass implementation details 2003 //===----------------------------------------------------------------------===// 2004 2005 LLVM_DUMP_METHOD void EquivalenceClass::dumpToStream(ProgramStateRef State, 2006 raw_ostream &os) const { 2007 SymbolSet ClassMembers = getClassMembers(State); 2008 for (const SymbolRef &MemberSym : ClassMembers) { 2009 MemberSym->dump(); 2010 os << "\n"; 2011 } 2012 } 2013 2014 inline EquivalenceClass EquivalenceClass::find(ProgramStateRef State, 2015 SymbolRef Sym) { 2016 assert(State && "State should not be null"); 2017 assert(Sym && "Symbol should not be null"); 2018 // We store far from all Symbol -> Class mappings 2019 if (const EquivalenceClass *NontrivialClass = State->get<ClassMap>(Sym)) 2020 return *NontrivialClass; 2021 2022 // This is a trivial class of Sym. 2023 return Sym; 2024 } 2025 2026 inline ProgramStateRef EquivalenceClass::merge(RangeSet::Factory &F, 2027 ProgramStateRef State, 2028 SymbolRef First, 2029 SymbolRef Second) { 2030 EquivalenceClass FirstClass = find(State, First); 2031 EquivalenceClass SecondClass = find(State, Second); 2032 2033 return FirstClass.merge(F, State, SecondClass); 2034 } 2035 2036 inline ProgramStateRef EquivalenceClass::merge(RangeSet::Factory &F, 2037 ProgramStateRef State, 2038 EquivalenceClass Other) { 2039 // It is already the same class. 2040 if (*this == Other) 2041 return State; 2042 2043 // FIXME: As of now, we support only equivalence classes of the same type. 2044 // This limitation is connected to the lack of explicit casts in 2045 // our symbolic expression model. 2046 // 2047 // That means that for `int x` and `char y` we don't distinguish 2048 // between these two very different cases: 2049 // * `x == y` 2050 // * `(char)x == y` 2051 // 2052 // The moment we introduce symbolic casts, this restriction can be 2053 // lifted. 2054 if (getType() != Other.getType()) 2055 return State; 2056 2057 SymbolSet Members = getClassMembers(State); 2058 SymbolSet OtherMembers = Other.getClassMembers(State); 2059 2060 // We estimate the size of the class by the height of tree containing 2061 // its members. Merging is not a trivial operation, so it's easier to 2062 // merge the smaller class into the bigger one. 2063 if (Members.getHeight() >= OtherMembers.getHeight()) { 2064 return mergeImpl(F, State, Members, Other, OtherMembers); 2065 } else { 2066 return Other.mergeImpl(F, State, OtherMembers, *this, Members); 2067 } 2068 } 2069 2070 inline ProgramStateRef 2071 EquivalenceClass::mergeImpl(RangeSet::Factory &RangeFactory, 2072 ProgramStateRef State, SymbolSet MyMembers, 2073 EquivalenceClass Other, SymbolSet OtherMembers) { 2074 // Essentially what we try to recreate here is some kind of union-find 2075 // data structure. It does have certain limitations due to persistence 2076 // and the need to remove elements from classes. 2077 // 2078 // In this setting, EquialityClass object is the representative of the class 2079 // or the parent element. ClassMap is a mapping of class members to their 2080 // parent. Unlike the union-find structure, they all point directly to the 2081 // class representative because we don't have an opportunity to actually do 2082 // path compression when dealing with immutability. This means that we 2083 // compress paths every time we do merges. It also means that we lose 2084 // the main amortized complexity benefit from the original data structure. 2085 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 2086 ConstraintRangeTy::Factory &CRF = State->get_context<ConstraintRange>(); 2087 2088 // 1. If the merged classes have any constraints associated with them, we 2089 // need to transfer them to the class we have left. 2090 // 2091 // Intersection here makes perfect sense because both of these constraints 2092 // must hold for the whole new class. 2093 if (Optional<RangeSet> NewClassConstraint = 2094 intersect(RangeFactory, getConstraint(State, *this), 2095 getConstraint(State, Other))) { 2096 // NOTE: Essentially, NewClassConstraint should NEVER be infeasible because 2097 // range inferrer shouldn't generate ranges incompatible with 2098 // equivalence classes. However, at the moment, due to imperfections 2099 // in the solver, it is possible and the merge function can also 2100 // return infeasible states aka null states. 2101 if (NewClassConstraint->isEmpty()) 2102 // Infeasible state 2103 return nullptr; 2104 2105 // No need in tracking constraints of a now-dissolved class. 2106 Constraints = CRF.remove(Constraints, Other); 2107 // Assign new constraints for this class. 2108 Constraints = CRF.add(Constraints, *this, *NewClassConstraint); 2109 2110 assert(areFeasible(Constraints) && "Constraint manager shouldn't produce " 2111 "a state with infeasible constraints"); 2112 2113 State = State->set<ConstraintRange>(Constraints); 2114 } 2115 2116 // 2. Get ALL equivalence-related maps 2117 ClassMapTy Classes = State->get<ClassMap>(); 2118 ClassMapTy::Factory &CMF = State->get_context<ClassMap>(); 2119 2120 ClassMembersTy Members = State->get<ClassMembers>(); 2121 ClassMembersTy::Factory &MF = State->get_context<ClassMembers>(); 2122 2123 DisequalityMapTy DisequalityInfo = State->get<DisequalityMap>(); 2124 DisequalityMapTy::Factory &DF = State->get_context<DisequalityMap>(); 2125 2126 ClassSet::Factory &CF = State->get_context<ClassSet>(); 2127 SymbolSet::Factory &F = getMembersFactory(State); 2128 2129 // 2. Merge members of the Other class into the current class. 2130 SymbolSet NewClassMembers = MyMembers; 2131 for (SymbolRef Sym : OtherMembers) { 2132 NewClassMembers = F.add(NewClassMembers, Sym); 2133 // *this is now the class for all these new symbols. 2134 Classes = CMF.add(Classes, Sym, *this); 2135 } 2136 2137 // 3. Adjust member mapping. 2138 // 2139 // No need in tracking members of a now-dissolved class. 2140 Members = MF.remove(Members, Other); 2141 // Now only the current class is mapped to all the symbols. 2142 Members = MF.add(Members, *this, NewClassMembers); 2143 2144 // 4. Update disequality relations 2145 ClassSet DisequalToOther = Other.getDisequalClasses(DisequalityInfo, CF); 2146 // We are about to merge two classes but they are already known to be 2147 // non-equal. This is a contradiction. 2148 if (DisequalToOther.contains(*this)) 2149 return nullptr; 2150 2151 if (!DisequalToOther.isEmpty()) { 2152 ClassSet DisequalToThis = getDisequalClasses(DisequalityInfo, CF); 2153 DisequalityInfo = DF.remove(DisequalityInfo, Other); 2154 2155 for (EquivalenceClass DisequalClass : DisequalToOther) { 2156 DisequalToThis = CF.add(DisequalToThis, DisequalClass); 2157 2158 // Disequality is a symmetric relation meaning that if 2159 // DisequalToOther not null then the set for DisequalClass is not 2160 // empty and has at least Other. 2161 ClassSet OriginalSetLinkedToOther = 2162 *DisequalityInfo.lookup(DisequalClass); 2163 2164 // Other will be eliminated and we should replace it with the bigger 2165 // united class. 2166 ClassSet NewSet = CF.remove(OriginalSetLinkedToOther, Other); 2167 NewSet = CF.add(NewSet, *this); 2168 2169 DisequalityInfo = DF.add(DisequalityInfo, DisequalClass, NewSet); 2170 } 2171 2172 DisequalityInfo = DF.add(DisequalityInfo, *this, DisequalToThis); 2173 State = State->set<DisequalityMap>(DisequalityInfo); 2174 } 2175 2176 // 5. Update the state 2177 State = State->set<ClassMap>(Classes); 2178 State = State->set<ClassMembers>(Members); 2179 2180 return State; 2181 } 2182 2183 inline SymbolSet::Factory & 2184 EquivalenceClass::getMembersFactory(ProgramStateRef State) { 2185 return State->get_context<SymbolSet>(); 2186 } 2187 2188 SymbolSet EquivalenceClass::getClassMembers(ProgramStateRef State) const { 2189 if (const SymbolSet *Members = State->get<ClassMembers>(*this)) 2190 return *Members; 2191 2192 // This class is trivial, so we need to construct a set 2193 // with just that one symbol from the class. 2194 SymbolSet::Factory &F = getMembersFactory(State); 2195 return F.add(F.getEmptySet(), getRepresentativeSymbol()); 2196 } 2197 2198 bool EquivalenceClass::isTrivial(ProgramStateRef State) const { 2199 return State->get<ClassMembers>(*this) == nullptr; 2200 } 2201 2202 bool EquivalenceClass::isTriviallyDead(ProgramStateRef State, 2203 SymbolReaper &Reaper) const { 2204 return isTrivial(State) && Reaper.isDead(getRepresentativeSymbol()); 2205 } 2206 2207 inline ProgramStateRef EquivalenceClass::markDisequal(RangeSet::Factory &RF, 2208 ProgramStateRef State, 2209 SymbolRef First, 2210 SymbolRef Second) { 2211 return markDisequal(RF, State, find(State, First), find(State, Second)); 2212 } 2213 2214 inline ProgramStateRef EquivalenceClass::markDisequal(RangeSet::Factory &RF, 2215 ProgramStateRef State, 2216 EquivalenceClass First, 2217 EquivalenceClass Second) { 2218 return First.markDisequal(RF, State, Second); 2219 } 2220 2221 inline ProgramStateRef 2222 EquivalenceClass::markDisequal(RangeSet::Factory &RF, ProgramStateRef State, 2223 EquivalenceClass Other) const { 2224 // If we know that two classes are equal, we can only produce an infeasible 2225 // state. 2226 if (*this == Other) { 2227 return nullptr; 2228 } 2229 2230 DisequalityMapTy DisequalityInfo = State->get<DisequalityMap>(); 2231 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 2232 2233 // Disequality is a symmetric relation, so if we mark A as disequal to B, 2234 // we should also mark B as disequalt to A. 2235 if (!addToDisequalityInfo(DisequalityInfo, Constraints, RF, State, *this, 2236 Other) || 2237 !addToDisequalityInfo(DisequalityInfo, Constraints, RF, State, Other, 2238 *this)) 2239 return nullptr; 2240 2241 assert(areFeasible(Constraints) && "Constraint manager shouldn't produce " 2242 "a state with infeasible constraints"); 2243 2244 State = State->set<DisequalityMap>(DisequalityInfo); 2245 State = State->set<ConstraintRange>(Constraints); 2246 2247 return State; 2248 } 2249 2250 inline bool EquivalenceClass::addToDisequalityInfo( 2251 DisequalityMapTy &Info, ConstraintRangeTy &Constraints, 2252 RangeSet::Factory &RF, ProgramStateRef State, EquivalenceClass First, 2253 EquivalenceClass Second) { 2254 2255 // 1. Get all of the required factories. 2256 DisequalityMapTy::Factory &F = State->get_context<DisequalityMap>(); 2257 ClassSet::Factory &CF = State->get_context<ClassSet>(); 2258 ConstraintRangeTy::Factory &CRF = State->get_context<ConstraintRange>(); 2259 2260 // 2. Add Second to the set of classes disequal to First. 2261 const ClassSet *CurrentSet = Info.lookup(First); 2262 ClassSet NewSet = CurrentSet ? *CurrentSet : CF.getEmptySet(); 2263 NewSet = CF.add(NewSet, Second); 2264 2265 Info = F.add(Info, First, NewSet); 2266 2267 // 3. If Second is known to be a constant, we can delete this point 2268 // from the constraint asociated with First. 2269 // 2270 // So, if Second == 10, it means that First != 10. 2271 // At the same time, the same logic does not apply to ranges. 2272 if (const RangeSet *SecondConstraint = Constraints.lookup(Second)) 2273 if (const llvm::APSInt *Point = SecondConstraint->getConcreteValue()) { 2274 2275 RangeSet FirstConstraint = SymbolicRangeInferrer::inferRange( 2276 RF, State, First.getRepresentativeSymbol()); 2277 2278 FirstConstraint = RF.deletePoint(FirstConstraint, *Point); 2279 2280 // If the First class is about to be constrained with an empty 2281 // range-set, the state is infeasible. 2282 if (FirstConstraint.isEmpty()) 2283 return false; 2284 2285 Constraints = CRF.add(Constraints, First, FirstConstraint); 2286 } 2287 2288 return true; 2289 } 2290 2291 inline Optional<bool> EquivalenceClass::areEqual(ProgramStateRef State, 2292 SymbolRef FirstSym, 2293 SymbolRef SecondSym) { 2294 return EquivalenceClass::areEqual(State, find(State, FirstSym), 2295 find(State, SecondSym)); 2296 } 2297 2298 inline Optional<bool> EquivalenceClass::areEqual(ProgramStateRef State, 2299 EquivalenceClass First, 2300 EquivalenceClass Second) { 2301 // The same equivalence class => symbols are equal. 2302 if (First == Second) 2303 return true; 2304 2305 // Let's check if we know anything about these two classes being not equal to 2306 // each other. 2307 ClassSet DisequalToFirst = First.getDisequalClasses(State); 2308 if (DisequalToFirst.contains(Second)) 2309 return false; 2310 2311 // It is not clear. 2312 return llvm::None; 2313 } 2314 2315 LLVM_NODISCARD ProgramStateRef 2316 EquivalenceClass::removeMember(ProgramStateRef State, const SymbolRef Old) { 2317 2318 SymbolSet ClsMembers = getClassMembers(State); 2319 assert(ClsMembers.contains(Old)); 2320 2321 // We don't remove `Old`'s Sym->Class relation for two reasons: 2322 // 1) This way constraints for the old symbol can still be found via it's 2323 // equivalence class that it used to be the member of. 2324 // 2) Performance and resource reasons. We can spare one removal and thus one 2325 // additional tree in the forest of `ClassMap`. 2326 2327 // Remove `Old`'s Class->Sym relation. 2328 SymbolSet::Factory &F = getMembersFactory(State); 2329 ClassMembersTy::Factory &EMFactory = State->get_context<ClassMembers>(); 2330 ClsMembers = F.remove(ClsMembers, Old); 2331 // Ensure another precondition of the removeMember function (we can check 2332 // this only with isEmpty, thus we have to do the remove first). 2333 assert(!ClsMembers.isEmpty() && 2334 "Class should have had at least two members before member removal"); 2335 // Overwrite the existing members assigned to this class. 2336 ClassMembersTy ClassMembersMap = State->get<ClassMembers>(); 2337 ClassMembersMap = EMFactory.add(ClassMembersMap, *this, ClsMembers); 2338 State = State->set<ClassMembers>(ClassMembersMap); 2339 2340 return State; 2341 } 2342 2343 // Re-evaluate an SVal with top-level `State->assume` logic. 2344 LLVM_NODISCARD ProgramStateRef reAssume(ProgramStateRef State, 2345 const RangeSet *Constraint, 2346 SVal TheValue) { 2347 if (!Constraint) 2348 return State; 2349 2350 const auto DefinedVal = TheValue.castAs<DefinedSVal>(); 2351 2352 // If the SVal is 0, we can simply interpret that as `false`. 2353 if (Constraint->encodesFalseRange()) 2354 return State->assume(DefinedVal, false); 2355 2356 // If the constraint does not encode 0 then we can interpret that as `true` 2357 // AND as a Range(Set). 2358 if (Constraint->encodesTrueRange()) { 2359 State = State->assume(DefinedVal, true); 2360 if (!State) 2361 return nullptr; 2362 // Fall through, re-assume based on the range values as well. 2363 } 2364 // Overestimate the individual Ranges with the RangeSet' lowest and 2365 // highest values. 2366 return State->assumeInclusiveRange(DefinedVal, Constraint->getMinValue(), 2367 Constraint->getMaxValue(), true); 2368 } 2369 2370 // Iterate over all symbols and try to simplify them. Once a symbol is 2371 // simplified then we check if we can merge the simplified symbol's equivalence 2372 // class to this class. This way, we simplify not just the symbols but the 2373 // classes as well: we strive to keep the number of the classes to be the 2374 // absolute minimum. 2375 LLVM_NODISCARD ProgramStateRef 2376 EquivalenceClass::simplify(SValBuilder &SVB, RangeSet::Factory &F, 2377 ProgramStateRef State, EquivalenceClass Class) { 2378 SymbolSet ClassMembers = Class.getClassMembers(State); 2379 for (const SymbolRef &MemberSym : ClassMembers) { 2380 2381 const SVal SimplifiedMemberVal = simplifyToSVal(State, MemberSym); 2382 const SymbolRef SimplifiedMemberSym = SimplifiedMemberVal.getAsSymbol(); 2383 2384 // The symbol is collapsed to a constant, check if the current State is 2385 // still feasible. 2386 if (const auto CI = SimplifiedMemberVal.getAs<nonloc::ConcreteInt>()) { 2387 const llvm::APSInt &SV = CI->getValue(); 2388 const RangeSet *ClassConstraint = getConstraint(State, Class); 2389 // We have found a contradiction. 2390 if (ClassConstraint && !ClassConstraint->contains(SV)) 2391 return nullptr; 2392 } 2393 2394 if (SimplifiedMemberSym && MemberSym != SimplifiedMemberSym) { 2395 // The simplified symbol should be the member of the original Class, 2396 // however, it might be in another existing class at the moment. We 2397 // have to merge these classes. 2398 ProgramStateRef OldState = State; 2399 State = merge(F, State, MemberSym, SimplifiedMemberSym); 2400 if (!State) 2401 return nullptr; 2402 // No state change, no merge happened actually. 2403 if (OldState == State) 2404 continue; 2405 2406 assert(find(State, MemberSym) == find(State, SimplifiedMemberSym)); 2407 // Remove the old and more complex symbol. 2408 State = find(State, MemberSym).removeMember(State, MemberSym); 2409 2410 // Query the class constraint again b/c that may have changed during the 2411 // merge above. 2412 const RangeSet *ClassConstraint = getConstraint(State, Class); 2413 2414 // Re-evaluate an SVal with top-level `State->assume`, this ignites 2415 // a RECURSIVE algorithm that will reach a FIXPOINT. 2416 // 2417 // About performance and complexity: Let us assume that in a State we 2418 // have N non-trivial equivalence classes and that all constraints and 2419 // disequality info is related to non-trivial classes. In the worst case, 2420 // we can simplify only one symbol of one class in each iteration. The 2421 // number of symbols in one class cannot grow b/c we replace the old 2422 // symbol with the simplified one. Also, the number of the equivalence 2423 // classes can decrease only, b/c the algorithm does a merge operation 2424 // optionally. We need N iterations in this case to reach the fixpoint. 2425 // Thus, the steps needed to be done in the worst case is proportional to 2426 // N*N. 2427 // 2428 // This worst case scenario can be extended to that case when we have 2429 // trivial classes in the constraints and in the disequality map. This 2430 // case can be reduced to the case with a State where there are only 2431 // non-trivial classes. This is because a merge operation on two trivial 2432 // classes results in one non-trivial class. 2433 State = reAssume(State, ClassConstraint, SimplifiedMemberVal); 2434 if (!State) 2435 return nullptr; 2436 } 2437 } 2438 return State; 2439 } 2440 2441 inline ClassSet EquivalenceClass::getDisequalClasses(ProgramStateRef State, 2442 SymbolRef Sym) { 2443 return find(State, Sym).getDisequalClasses(State); 2444 } 2445 2446 inline ClassSet 2447 EquivalenceClass::getDisequalClasses(ProgramStateRef State) const { 2448 return getDisequalClasses(State->get<DisequalityMap>(), 2449 State->get_context<ClassSet>()); 2450 } 2451 2452 inline ClassSet 2453 EquivalenceClass::getDisequalClasses(DisequalityMapTy Map, 2454 ClassSet::Factory &Factory) const { 2455 if (const ClassSet *DisequalClasses = Map.lookup(*this)) 2456 return *DisequalClasses; 2457 2458 return Factory.getEmptySet(); 2459 } 2460 2461 bool EquivalenceClass::isClassDataConsistent(ProgramStateRef State) { 2462 ClassMembersTy Members = State->get<ClassMembers>(); 2463 2464 for (std::pair<EquivalenceClass, SymbolSet> ClassMembersPair : Members) { 2465 for (SymbolRef Member : ClassMembersPair.second) { 2466 // Every member of the class should have a mapping back to the class. 2467 if (find(State, Member) == ClassMembersPair.first) { 2468 continue; 2469 } 2470 2471 return false; 2472 } 2473 } 2474 2475 DisequalityMapTy Disequalities = State->get<DisequalityMap>(); 2476 for (std::pair<EquivalenceClass, ClassSet> DisequalityInfo : Disequalities) { 2477 EquivalenceClass Class = DisequalityInfo.first; 2478 ClassSet DisequalClasses = DisequalityInfo.second; 2479 2480 // There is no use in keeping empty sets in the map. 2481 if (DisequalClasses.isEmpty()) 2482 return false; 2483 2484 // Disequality is symmetrical, i.e. for every Class A and B that A != B, 2485 // B != A should also be true. 2486 for (EquivalenceClass DisequalClass : DisequalClasses) { 2487 const ClassSet *DisequalToDisequalClasses = 2488 Disequalities.lookup(DisequalClass); 2489 2490 // It should be a set of at least one element: Class 2491 if (!DisequalToDisequalClasses || 2492 !DisequalToDisequalClasses->contains(Class)) 2493 return false; 2494 } 2495 } 2496 2497 return true; 2498 } 2499 2500 //===----------------------------------------------------------------------===// 2501 // RangeConstraintManager implementation 2502 //===----------------------------------------------------------------------===// 2503 2504 bool RangeConstraintManager::canReasonAbout(SVal X) const { 2505 Optional<nonloc::SymbolVal> SymVal = X.getAs<nonloc::SymbolVal>(); 2506 if (SymVal && SymVal->isExpression()) { 2507 const SymExpr *SE = SymVal->getSymbol(); 2508 2509 if (const SymIntExpr *SIE = dyn_cast<SymIntExpr>(SE)) { 2510 switch (SIE->getOpcode()) { 2511 // We don't reason yet about bitwise-constraints on symbolic values. 2512 case BO_And: 2513 case BO_Or: 2514 case BO_Xor: 2515 return false; 2516 // We don't reason yet about these arithmetic constraints on 2517 // symbolic values. 2518 case BO_Mul: 2519 case BO_Div: 2520 case BO_Rem: 2521 case BO_Shl: 2522 case BO_Shr: 2523 return false; 2524 // All other cases. 2525 default: 2526 return true; 2527 } 2528 } 2529 2530 if (const SymSymExpr *SSE = dyn_cast<SymSymExpr>(SE)) { 2531 // FIXME: Handle <=> here. 2532 if (BinaryOperator::isEqualityOp(SSE->getOpcode()) || 2533 BinaryOperator::isRelationalOp(SSE->getOpcode())) { 2534 // We handle Loc <> Loc comparisons, but not (yet) NonLoc <> NonLoc. 2535 // We've recently started producing Loc <> NonLoc comparisons (that 2536 // result from casts of one of the operands between eg. intptr_t and 2537 // void *), but we can't reason about them yet. 2538 if (Loc::isLocType(SSE->getLHS()->getType())) { 2539 return Loc::isLocType(SSE->getRHS()->getType()); 2540 } 2541 } 2542 } 2543 2544 return false; 2545 } 2546 2547 return true; 2548 } 2549 2550 ConditionTruthVal RangeConstraintManager::checkNull(ProgramStateRef State, 2551 SymbolRef Sym) { 2552 const RangeSet *Ranges = getConstraint(State, Sym); 2553 2554 // If we don't have any information about this symbol, it's underconstrained. 2555 if (!Ranges) 2556 return ConditionTruthVal(); 2557 2558 // If we have a concrete value, see if it's zero. 2559 if (const llvm::APSInt *Value = Ranges->getConcreteValue()) 2560 return *Value == 0; 2561 2562 BasicValueFactory &BV = getBasicVals(); 2563 APSIntType IntType = BV.getAPSIntType(Sym->getType()); 2564 llvm::APSInt Zero = IntType.getZeroValue(); 2565 2566 // Check if zero is in the set of possible values. 2567 if (!Ranges->contains(Zero)) 2568 return false; 2569 2570 // Zero is a possible value, but it is not the /only/ possible value. 2571 return ConditionTruthVal(); 2572 } 2573 2574 const llvm::APSInt *RangeConstraintManager::getSymVal(ProgramStateRef St, 2575 SymbolRef Sym) const { 2576 const RangeSet *T = getConstraint(St, Sym); 2577 return T ? T->getConcreteValue() : nullptr; 2578 } 2579 2580 //===----------------------------------------------------------------------===// 2581 // Remove dead symbols from existing constraints 2582 //===----------------------------------------------------------------------===// 2583 2584 /// Scan all symbols referenced by the constraints. If the symbol is not alive 2585 /// as marked in LSymbols, mark it as dead in DSymbols. 2586 ProgramStateRef 2587 RangeConstraintManager::removeDeadBindings(ProgramStateRef State, 2588 SymbolReaper &SymReaper) { 2589 ClassMembersTy ClassMembersMap = State->get<ClassMembers>(); 2590 ClassMembersTy NewClassMembersMap = ClassMembersMap; 2591 ClassMembersTy::Factory &EMFactory = State->get_context<ClassMembers>(); 2592 SymbolSet::Factory &SetFactory = State->get_context<SymbolSet>(); 2593 2594 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 2595 ConstraintRangeTy NewConstraints = Constraints; 2596 ConstraintRangeTy::Factory &ConstraintFactory = 2597 State->get_context<ConstraintRange>(); 2598 2599 ClassMapTy Map = State->get<ClassMap>(); 2600 ClassMapTy NewMap = Map; 2601 ClassMapTy::Factory &ClassFactory = State->get_context<ClassMap>(); 2602 2603 DisequalityMapTy Disequalities = State->get<DisequalityMap>(); 2604 DisequalityMapTy::Factory &DisequalityFactory = 2605 State->get_context<DisequalityMap>(); 2606 ClassSet::Factory &ClassSetFactory = State->get_context<ClassSet>(); 2607 2608 bool ClassMapChanged = false; 2609 bool MembersMapChanged = false; 2610 bool ConstraintMapChanged = false; 2611 bool DisequalitiesChanged = false; 2612 2613 auto removeDeadClass = [&](EquivalenceClass Class) { 2614 // Remove associated constraint ranges. 2615 Constraints = ConstraintFactory.remove(Constraints, Class); 2616 ConstraintMapChanged = true; 2617 2618 // Update disequality information to not hold any information on the 2619 // removed class. 2620 ClassSet DisequalClasses = 2621 Class.getDisequalClasses(Disequalities, ClassSetFactory); 2622 if (!DisequalClasses.isEmpty()) { 2623 for (EquivalenceClass DisequalClass : DisequalClasses) { 2624 ClassSet DisequalToDisequalSet = 2625 DisequalClass.getDisequalClasses(Disequalities, ClassSetFactory); 2626 // DisequalToDisequalSet is guaranteed to be non-empty for consistent 2627 // disequality info. 2628 assert(!DisequalToDisequalSet.isEmpty()); 2629 ClassSet NewSet = ClassSetFactory.remove(DisequalToDisequalSet, Class); 2630 2631 // No need in keeping an empty set. 2632 if (NewSet.isEmpty()) { 2633 Disequalities = 2634 DisequalityFactory.remove(Disequalities, DisequalClass); 2635 } else { 2636 Disequalities = 2637 DisequalityFactory.add(Disequalities, DisequalClass, NewSet); 2638 } 2639 } 2640 // Remove the data for the class 2641 Disequalities = DisequalityFactory.remove(Disequalities, Class); 2642 DisequalitiesChanged = true; 2643 } 2644 }; 2645 2646 // 1. Let's see if dead symbols are trivial and have associated constraints. 2647 for (std::pair<EquivalenceClass, RangeSet> ClassConstraintPair : 2648 Constraints) { 2649 EquivalenceClass Class = ClassConstraintPair.first; 2650 if (Class.isTriviallyDead(State, SymReaper)) { 2651 // If this class is trivial, we can remove its constraints right away. 2652 removeDeadClass(Class); 2653 } 2654 } 2655 2656 // 2. We don't need to track classes for dead symbols. 2657 for (std::pair<SymbolRef, EquivalenceClass> SymbolClassPair : Map) { 2658 SymbolRef Sym = SymbolClassPair.first; 2659 2660 if (SymReaper.isDead(Sym)) { 2661 ClassMapChanged = true; 2662 NewMap = ClassFactory.remove(NewMap, Sym); 2663 } 2664 } 2665 2666 // 3. Remove dead members from classes and remove dead non-trivial classes 2667 // and their constraints. 2668 for (std::pair<EquivalenceClass, SymbolSet> ClassMembersPair : 2669 ClassMembersMap) { 2670 EquivalenceClass Class = ClassMembersPair.first; 2671 SymbolSet LiveMembers = ClassMembersPair.second; 2672 bool MembersChanged = false; 2673 2674 for (SymbolRef Member : ClassMembersPair.second) { 2675 if (SymReaper.isDead(Member)) { 2676 MembersChanged = true; 2677 LiveMembers = SetFactory.remove(LiveMembers, Member); 2678 } 2679 } 2680 2681 // Check if the class changed. 2682 if (!MembersChanged) 2683 continue; 2684 2685 MembersMapChanged = true; 2686 2687 if (LiveMembers.isEmpty()) { 2688 // The class is dead now, we need to wipe it out of the members map... 2689 NewClassMembersMap = EMFactory.remove(NewClassMembersMap, Class); 2690 2691 // ...and remove all of its constraints. 2692 removeDeadClass(Class); 2693 } else { 2694 // We need to change the members associated with the class. 2695 NewClassMembersMap = 2696 EMFactory.add(NewClassMembersMap, Class, LiveMembers); 2697 } 2698 } 2699 2700 // 4. Update the state with new maps. 2701 // 2702 // Here we try to be humble and update a map only if it really changed. 2703 if (ClassMapChanged) 2704 State = State->set<ClassMap>(NewMap); 2705 2706 if (MembersMapChanged) 2707 State = State->set<ClassMembers>(NewClassMembersMap); 2708 2709 if (ConstraintMapChanged) 2710 State = State->set<ConstraintRange>(Constraints); 2711 2712 if (DisequalitiesChanged) 2713 State = State->set<DisequalityMap>(Disequalities); 2714 2715 assert(EquivalenceClass::isClassDataConsistent(State)); 2716 2717 return State; 2718 } 2719 2720 RangeSet RangeConstraintManager::getRange(ProgramStateRef State, 2721 SymbolRef Sym) { 2722 return SymbolicRangeInferrer::inferRange(F, State, Sym); 2723 } 2724 2725 ProgramStateRef RangeConstraintManager::setRange(ProgramStateRef State, 2726 SymbolRef Sym, 2727 RangeSet Range) { 2728 return ConstraintAssignor::assign(State, getSValBuilder(), F, Sym, Range); 2729 } 2730 2731 //===------------------------------------------------------------------------=== 2732 // assumeSymX methods: protected interface for RangeConstraintManager. 2733 //===------------------------------------------------------------------------===/ 2734 2735 // The syntax for ranges below is mathematical, using [x, y] for closed ranges 2736 // and (x, y) for open ranges. These ranges are modular, corresponding with 2737 // a common treatment of C integer overflow. This means that these methods 2738 // do not have to worry about overflow; RangeSet::Intersect can handle such a 2739 // "wraparound" range. 2740 // As an example, the range [UINT_MAX-1, 3) contains five values: UINT_MAX-1, 2741 // UINT_MAX, 0, 1, and 2. 2742 2743 ProgramStateRef 2744 RangeConstraintManager::assumeSymNE(ProgramStateRef St, SymbolRef Sym, 2745 const llvm::APSInt &Int, 2746 const llvm::APSInt &Adjustment) { 2747 // Before we do any real work, see if the value can even show up. 2748 APSIntType AdjustmentType(Adjustment); 2749 if (AdjustmentType.testInRange(Int, true) != APSIntType::RTR_Within) 2750 return St; 2751 2752 llvm::APSInt Point = AdjustmentType.convert(Int) - Adjustment; 2753 RangeSet New = getRange(St, Sym); 2754 New = F.deletePoint(New, Point); 2755 2756 return setRange(St, Sym, New); 2757 } 2758 2759 ProgramStateRef 2760 RangeConstraintManager::assumeSymEQ(ProgramStateRef St, SymbolRef Sym, 2761 const llvm::APSInt &Int, 2762 const llvm::APSInt &Adjustment) { 2763 // Before we do any real work, see if the value can even show up. 2764 APSIntType AdjustmentType(Adjustment); 2765 if (AdjustmentType.testInRange(Int, true) != APSIntType::RTR_Within) 2766 return nullptr; 2767 2768 // [Int-Adjustment, Int-Adjustment] 2769 llvm::APSInt AdjInt = AdjustmentType.convert(Int) - Adjustment; 2770 RangeSet New = getRange(St, Sym); 2771 New = F.intersect(New, AdjInt); 2772 2773 return setRange(St, Sym, New); 2774 } 2775 2776 RangeSet RangeConstraintManager::getSymLTRange(ProgramStateRef St, 2777 SymbolRef Sym, 2778 const llvm::APSInt &Int, 2779 const llvm::APSInt &Adjustment) { 2780 // Before we do any real work, see if the value can even show up. 2781 APSIntType AdjustmentType(Adjustment); 2782 switch (AdjustmentType.testInRange(Int, true)) { 2783 case APSIntType::RTR_Below: 2784 return F.getEmptySet(); 2785 case APSIntType::RTR_Within: 2786 break; 2787 case APSIntType::RTR_Above: 2788 return getRange(St, Sym); 2789 } 2790 2791 // Special case for Int == Min. This is always false. 2792 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2793 llvm::APSInt Min = AdjustmentType.getMinValue(); 2794 if (ComparisonVal == Min) 2795 return F.getEmptySet(); 2796 2797 llvm::APSInt Lower = Min - Adjustment; 2798 llvm::APSInt Upper = ComparisonVal - Adjustment; 2799 --Upper; 2800 2801 RangeSet Result = getRange(St, Sym); 2802 return F.intersect(Result, Lower, Upper); 2803 } 2804 2805 ProgramStateRef 2806 RangeConstraintManager::assumeSymLT(ProgramStateRef St, SymbolRef Sym, 2807 const llvm::APSInt &Int, 2808 const llvm::APSInt &Adjustment) { 2809 RangeSet New = getSymLTRange(St, Sym, Int, Adjustment); 2810 return setRange(St, Sym, New); 2811 } 2812 2813 RangeSet RangeConstraintManager::getSymGTRange(ProgramStateRef St, 2814 SymbolRef Sym, 2815 const llvm::APSInt &Int, 2816 const llvm::APSInt &Adjustment) { 2817 // Before we do any real work, see if the value can even show up. 2818 APSIntType AdjustmentType(Adjustment); 2819 switch (AdjustmentType.testInRange(Int, true)) { 2820 case APSIntType::RTR_Below: 2821 return getRange(St, Sym); 2822 case APSIntType::RTR_Within: 2823 break; 2824 case APSIntType::RTR_Above: 2825 return F.getEmptySet(); 2826 } 2827 2828 // Special case for Int == Max. This is always false. 2829 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2830 llvm::APSInt Max = AdjustmentType.getMaxValue(); 2831 if (ComparisonVal == Max) 2832 return F.getEmptySet(); 2833 2834 llvm::APSInt Lower = ComparisonVal - Adjustment; 2835 llvm::APSInt Upper = Max - Adjustment; 2836 ++Lower; 2837 2838 RangeSet SymRange = getRange(St, Sym); 2839 return F.intersect(SymRange, Lower, Upper); 2840 } 2841 2842 ProgramStateRef 2843 RangeConstraintManager::assumeSymGT(ProgramStateRef St, SymbolRef Sym, 2844 const llvm::APSInt &Int, 2845 const llvm::APSInt &Adjustment) { 2846 RangeSet New = getSymGTRange(St, Sym, Int, Adjustment); 2847 return setRange(St, Sym, New); 2848 } 2849 2850 RangeSet RangeConstraintManager::getSymGERange(ProgramStateRef St, 2851 SymbolRef Sym, 2852 const llvm::APSInt &Int, 2853 const llvm::APSInt &Adjustment) { 2854 // Before we do any real work, see if the value can even show up. 2855 APSIntType AdjustmentType(Adjustment); 2856 switch (AdjustmentType.testInRange(Int, true)) { 2857 case APSIntType::RTR_Below: 2858 return getRange(St, Sym); 2859 case APSIntType::RTR_Within: 2860 break; 2861 case APSIntType::RTR_Above: 2862 return F.getEmptySet(); 2863 } 2864 2865 // Special case for Int == Min. This is always feasible. 2866 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2867 llvm::APSInt Min = AdjustmentType.getMinValue(); 2868 if (ComparisonVal == Min) 2869 return getRange(St, Sym); 2870 2871 llvm::APSInt Max = AdjustmentType.getMaxValue(); 2872 llvm::APSInt Lower = ComparisonVal - Adjustment; 2873 llvm::APSInt Upper = Max - Adjustment; 2874 2875 RangeSet SymRange = getRange(St, Sym); 2876 return F.intersect(SymRange, Lower, Upper); 2877 } 2878 2879 ProgramStateRef 2880 RangeConstraintManager::assumeSymGE(ProgramStateRef St, SymbolRef Sym, 2881 const llvm::APSInt &Int, 2882 const llvm::APSInt &Adjustment) { 2883 RangeSet New = getSymGERange(St, Sym, Int, Adjustment); 2884 return setRange(St, Sym, New); 2885 } 2886 2887 RangeSet 2888 RangeConstraintManager::getSymLERange(llvm::function_ref<RangeSet()> RS, 2889 const llvm::APSInt &Int, 2890 const llvm::APSInt &Adjustment) { 2891 // Before we do any real work, see if the value can even show up. 2892 APSIntType AdjustmentType(Adjustment); 2893 switch (AdjustmentType.testInRange(Int, true)) { 2894 case APSIntType::RTR_Below: 2895 return F.getEmptySet(); 2896 case APSIntType::RTR_Within: 2897 break; 2898 case APSIntType::RTR_Above: 2899 return RS(); 2900 } 2901 2902 // Special case for Int == Max. This is always feasible. 2903 llvm::APSInt ComparisonVal = AdjustmentType.convert(Int); 2904 llvm::APSInt Max = AdjustmentType.getMaxValue(); 2905 if (ComparisonVal == Max) 2906 return RS(); 2907 2908 llvm::APSInt Min = AdjustmentType.getMinValue(); 2909 llvm::APSInt Lower = Min - Adjustment; 2910 llvm::APSInt Upper = ComparisonVal - Adjustment; 2911 2912 RangeSet Default = RS(); 2913 return F.intersect(Default, Lower, Upper); 2914 } 2915 2916 RangeSet RangeConstraintManager::getSymLERange(ProgramStateRef St, 2917 SymbolRef Sym, 2918 const llvm::APSInt &Int, 2919 const llvm::APSInt &Adjustment) { 2920 return getSymLERange([&] { return getRange(St, Sym); }, Int, Adjustment); 2921 } 2922 2923 ProgramStateRef 2924 RangeConstraintManager::assumeSymLE(ProgramStateRef St, SymbolRef Sym, 2925 const llvm::APSInt &Int, 2926 const llvm::APSInt &Adjustment) { 2927 RangeSet New = getSymLERange(St, Sym, Int, Adjustment); 2928 return setRange(St, Sym, New); 2929 } 2930 2931 ProgramStateRef RangeConstraintManager::assumeSymWithinInclusiveRange( 2932 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 2933 const llvm::APSInt &To, const llvm::APSInt &Adjustment) { 2934 RangeSet New = getSymGERange(State, Sym, From, Adjustment); 2935 if (New.isEmpty()) 2936 return nullptr; 2937 RangeSet Out = getSymLERange([&] { return New; }, To, Adjustment); 2938 return setRange(State, Sym, Out); 2939 } 2940 2941 ProgramStateRef RangeConstraintManager::assumeSymOutsideInclusiveRange( 2942 ProgramStateRef State, SymbolRef Sym, const llvm::APSInt &From, 2943 const llvm::APSInt &To, const llvm::APSInt &Adjustment) { 2944 RangeSet RangeLT = getSymLTRange(State, Sym, From, Adjustment); 2945 RangeSet RangeGT = getSymGTRange(State, Sym, To, Adjustment); 2946 RangeSet New(F.add(RangeLT, RangeGT)); 2947 return setRange(State, Sym, New); 2948 } 2949 2950 //===----------------------------------------------------------------------===// 2951 // Pretty-printing. 2952 //===----------------------------------------------------------------------===// 2953 2954 void RangeConstraintManager::printJson(raw_ostream &Out, ProgramStateRef State, 2955 const char *NL, unsigned int Space, 2956 bool IsDot) const { 2957 printConstraints(Out, State, NL, Space, IsDot); 2958 printEquivalenceClasses(Out, State, NL, Space, IsDot); 2959 printDisequalities(Out, State, NL, Space, IsDot); 2960 } 2961 2962 static std::string toString(const SymbolRef &Sym) { 2963 std::string S; 2964 llvm::raw_string_ostream O(S); 2965 Sym->dumpToStream(O); 2966 return O.str(); 2967 } 2968 2969 void RangeConstraintManager::printConstraints(raw_ostream &Out, 2970 ProgramStateRef State, 2971 const char *NL, 2972 unsigned int Space, 2973 bool IsDot) const { 2974 ConstraintRangeTy Constraints = State->get<ConstraintRange>(); 2975 2976 Indent(Out, Space, IsDot) << "\"constraints\": "; 2977 if (Constraints.isEmpty()) { 2978 Out << "null," << NL; 2979 return; 2980 } 2981 2982 std::map<std::string, RangeSet> OrderedConstraints; 2983 for (std::pair<EquivalenceClass, RangeSet> P : Constraints) { 2984 SymbolSet ClassMembers = P.first.getClassMembers(State); 2985 for (const SymbolRef &ClassMember : ClassMembers) { 2986 bool insertion_took_place; 2987 std::tie(std::ignore, insertion_took_place) = 2988 OrderedConstraints.insert({toString(ClassMember), P.second}); 2989 assert(insertion_took_place && 2990 "two symbols should not have the same dump"); 2991 } 2992 } 2993 2994 ++Space; 2995 Out << '[' << NL; 2996 bool First = true; 2997 for (std::pair<std::string, RangeSet> P : OrderedConstraints) { 2998 if (First) { 2999 First = false; 3000 } else { 3001 Out << ','; 3002 Out << NL; 3003 } 3004 Indent(Out, Space, IsDot) 3005 << "{ \"symbol\": \"" << P.first << "\", \"range\": \""; 3006 P.second.dump(Out); 3007 Out << "\" }"; 3008 } 3009 Out << NL; 3010 3011 --Space; 3012 Indent(Out, Space, IsDot) << "]," << NL; 3013 } 3014 3015 static std::string toString(ProgramStateRef State, EquivalenceClass Class) { 3016 SymbolSet ClassMembers = Class.getClassMembers(State); 3017 llvm::SmallVector<SymbolRef, 8> ClassMembersSorted(ClassMembers.begin(), 3018 ClassMembers.end()); 3019 llvm::sort(ClassMembersSorted, 3020 [](const SymbolRef &LHS, const SymbolRef &RHS) { 3021 return toString(LHS) < toString(RHS); 3022 }); 3023 3024 bool FirstMember = true; 3025 3026 std::string Str; 3027 llvm::raw_string_ostream Out(Str); 3028 Out << "[ "; 3029 for (SymbolRef ClassMember : ClassMembersSorted) { 3030 if (FirstMember) 3031 FirstMember = false; 3032 else 3033 Out << ", "; 3034 Out << "\"" << ClassMember << "\""; 3035 } 3036 Out << " ]"; 3037 return Out.str(); 3038 } 3039 3040 void RangeConstraintManager::printEquivalenceClasses(raw_ostream &Out, 3041 ProgramStateRef State, 3042 const char *NL, 3043 unsigned int Space, 3044 bool IsDot) const { 3045 ClassMembersTy Members = State->get<ClassMembers>(); 3046 3047 Indent(Out, Space, IsDot) << "\"equivalence_classes\": "; 3048 if (Members.isEmpty()) { 3049 Out << "null," << NL; 3050 return; 3051 } 3052 3053 std::set<std::string> MembersStr; 3054 for (std::pair<EquivalenceClass, SymbolSet> ClassToSymbolSet : Members) 3055 MembersStr.insert(toString(State, ClassToSymbolSet.first)); 3056 3057 ++Space; 3058 Out << '[' << NL; 3059 bool FirstClass = true; 3060 for (const std::string &Str : MembersStr) { 3061 if (FirstClass) { 3062 FirstClass = false; 3063 } else { 3064 Out << ','; 3065 Out << NL; 3066 } 3067 Indent(Out, Space, IsDot); 3068 Out << Str; 3069 } 3070 Out << NL; 3071 3072 --Space; 3073 Indent(Out, Space, IsDot) << "]," << NL; 3074 } 3075 3076 void RangeConstraintManager::printDisequalities(raw_ostream &Out, 3077 ProgramStateRef State, 3078 const char *NL, 3079 unsigned int Space, 3080 bool IsDot) const { 3081 DisequalityMapTy Disequalities = State->get<DisequalityMap>(); 3082 3083 Indent(Out, Space, IsDot) << "\"disequality_info\": "; 3084 if (Disequalities.isEmpty()) { 3085 Out << "null," << NL; 3086 return; 3087 } 3088 3089 // Transform the disequality info to an ordered map of 3090 // [string -> (ordered set of strings)] 3091 using EqClassesStrTy = std::set<std::string>; 3092 using DisequalityInfoStrTy = std::map<std::string, EqClassesStrTy>; 3093 DisequalityInfoStrTy DisequalityInfoStr; 3094 for (std::pair<EquivalenceClass, ClassSet> ClassToDisEqSet : Disequalities) { 3095 EquivalenceClass Class = ClassToDisEqSet.first; 3096 ClassSet DisequalClasses = ClassToDisEqSet.second; 3097 EqClassesStrTy MembersStr; 3098 for (EquivalenceClass DisEqClass : DisequalClasses) 3099 MembersStr.insert(toString(State, DisEqClass)); 3100 DisequalityInfoStr.insert({toString(State, Class), MembersStr}); 3101 } 3102 3103 ++Space; 3104 Out << '[' << NL; 3105 bool FirstClass = true; 3106 for (std::pair<std::string, EqClassesStrTy> ClassToDisEqSet : 3107 DisequalityInfoStr) { 3108 const std::string &Class = ClassToDisEqSet.first; 3109 if (FirstClass) { 3110 FirstClass = false; 3111 } else { 3112 Out << ','; 3113 Out << NL; 3114 } 3115 Indent(Out, Space, IsDot) << "{" << NL; 3116 unsigned int DisEqSpace = Space + 1; 3117 Indent(Out, DisEqSpace, IsDot) << "\"class\": "; 3118 Out << Class; 3119 const EqClassesStrTy &DisequalClasses = ClassToDisEqSet.second; 3120 if (!DisequalClasses.empty()) { 3121 Out << "," << NL; 3122 Indent(Out, DisEqSpace, IsDot) << "\"disequal_to\": [" << NL; 3123 unsigned int DisEqClassSpace = DisEqSpace + 1; 3124 Indent(Out, DisEqClassSpace, IsDot); 3125 bool FirstDisEqClass = true; 3126 for (const std::string &DisEqClass : DisequalClasses) { 3127 if (FirstDisEqClass) { 3128 FirstDisEqClass = false; 3129 } else { 3130 Out << ',' << NL; 3131 Indent(Out, DisEqClassSpace, IsDot); 3132 } 3133 Out << DisEqClass; 3134 } 3135 Out << "]" << NL; 3136 } 3137 Indent(Out, Space, IsDot) << "}"; 3138 } 3139 Out << NL; 3140 3141 --Space; 3142 Indent(Out, Space, IsDot) << "]," << NL; 3143 } 3144