1 // SimpleSValBuilder.cpp - A basic SValBuilder -----------------------*- C++ -*- 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file defines SimpleSValBuilder, a basic implementation of SValBuilder. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "clang/StaticAnalyzer/Core/PathSensitive/APSIntType.h" 15 #include "clang/StaticAnalyzer/Core/PathSensitive/SValBuilder.h" 16 #include "clang/StaticAnalyzer/Core/PathSensitive/ProgramState.h" 17 18 using namespace clang; 19 using namespace ento; 20 21 namespace { 22 class SimpleSValBuilder : public SValBuilder { 23 protected: 24 virtual SVal dispatchCast(SVal val, QualType castTy); 25 virtual SVal evalCastFromNonLoc(NonLoc val, QualType castTy); 26 virtual SVal evalCastFromLoc(Loc val, QualType castTy); 27 28 public: 29 SimpleSValBuilder(llvm::BumpPtrAllocator &alloc, ASTContext &context, 30 ProgramStateManager &stateMgr) 31 : SValBuilder(alloc, context, stateMgr) {} 32 virtual ~SimpleSValBuilder() {} 33 34 virtual SVal evalMinus(NonLoc val); 35 virtual SVal evalComplement(NonLoc val); 36 virtual SVal evalBinOpNN(ProgramStateRef state, BinaryOperator::Opcode op, 37 NonLoc lhs, NonLoc rhs, QualType resultTy); 38 virtual SVal evalBinOpLL(ProgramStateRef state, BinaryOperator::Opcode op, 39 Loc lhs, Loc rhs, QualType resultTy); 40 virtual SVal evalBinOpLN(ProgramStateRef state, BinaryOperator::Opcode op, 41 Loc lhs, NonLoc rhs, QualType resultTy); 42 43 /// getKnownValue - evaluates a given SVal. If the SVal has only one possible 44 /// (integer) value, that value is returned. Otherwise, returns NULL. 45 virtual const llvm::APSInt *getKnownValue(ProgramStateRef state, SVal V); 46 47 SVal MakeSymIntVal(const SymExpr *LHS, BinaryOperator::Opcode op, 48 const llvm::APSInt &RHS, QualType resultTy); 49 }; 50 } // end anonymous namespace 51 52 SValBuilder *ento::createSimpleSValBuilder(llvm::BumpPtrAllocator &alloc, 53 ASTContext &context, 54 ProgramStateManager &stateMgr) { 55 return new SimpleSValBuilder(alloc, context, stateMgr); 56 } 57 58 //===----------------------------------------------------------------------===// 59 // Transfer function for Casts. 60 //===----------------------------------------------------------------------===// 61 62 SVal SimpleSValBuilder::dispatchCast(SVal Val, QualType CastTy) { 63 assert(isa<Loc>(&Val) || isa<NonLoc>(&Val)); 64 return isa<Loc>(Val) ? evalCastFromLoc(cast<Loc>(Val), CastTy) 65 : evalCastFromNonLoc(cast<NonLoc>(Val), CastTy); 66 } 67 68 SVal SimpleSValBuilder::evalCastFromNonLoc(NonLoc val, QualType castTy) { 69 70 bool isLocType = Loc::isLocType(castTy); 71 72 if (nonloc::LocAsInteger *LI = dyn_cast<nonloc::LocAsInteger>(&val)) { 73 if (isLocType) 74 return LI->getLoc(); 75 76 // FIXME: Correctly support promotions/truncations. 77 unsigned castSize = Context.getTypeSize(castTy); 78 if (castSize == LI->getNumBits()) 79 return val; 80 return makeLocAsInteger(LI->getLoc(), castSize); 81 } 82 83 if (const SymExpr *se = val.getAsSymbolicExpression()) { 84 QualType T = Context.getCanonicalType(se->getType(Context)); 85 // If types are the same or both are integers, ignore the cast. 86 // FIXME: Remove this hack when we support symbolic truncation/extension. 87 // HACK: If both castTy and T are integers, ignore the cast. This is 88 // not a permanent solution. Eventually we want to precisely handle 89 // extension/truncation of symbolic integers. This prevents us from losing 90 // precision when we assign 'x = y' and 'y' is symbolic and x and y are 91 // different integer types. 92 if (haveSameType(T, castTy)) 93 return val; 94 95 if (!isLocType) 96 return makeNonLoc(se, T, castTy); 97 return UnknownVal(); 98 } 99 100 // If value is a non integer constant, produce unknown. 101 if (!isa<nonloc::ConcreteInt>(val)) 102 return UnknownVal(); 103 104 // Only handle casts from integers to integers - if val is an integer constant 105 // being cast to a non integer type, produce unknown. 106 if (!isLocType && !castTy->isIntegerType()) 107 return UnknownVal(); 108 109 llvm::APSInt i = cast<nonloc::ConcreteInt>(val).getValue(); 110 BasicVals.getAPSIntType(castTy).apply(i); 111 112 if (isLocType) 113 return makeIntLocVal(i); 114 else 115 return makeIntVal(i); 116 } 117 118 SVal SimpleSValBuilder::evalCastFromLoc(Loc val, QualType castTy) { 119 120 // Casts from pointers -> pointers, just return the lval. 121 // 122 // Casts from pointers -> references, just return the lval. These 123 // can be introduced by the frontend for corner cases, e.g 124 // casting from va_list* to __builtin_va_list&. 125 // 126 if (Loc::isLocType(castTy) || castTy->isReferenceType()) 127 return val; 128 129 // FIXME: Handle transparent unions where a value can be "transparently" 130 // lifted into a union type. 131 if (castTy->isUnionType()) 132 return UnknownVal(); 133 134 if (castTy->isIntegerType()) { 135 unsigned BitWidth = Context.getTypeSize(castTy); 136 137 if (!isa<loc::ConcreteInt>(val)) 138 return makeLocAsInteger(val, BitWidth); 139 140 llvm::APSInt i = cast<loc::ConcreteInt>(val).getValue(); 141 BasicVals.getAPSIntType(castTy).apply(i); 142 return makeIntVal(i); 143 } 144 145 // All other cases: return 'UnknownVal'. This includes casting pointers 146 // to floats, which is probably badness it itself, but this is a good 147 // intermediate solution until we do something better. 148 return UnknownVal(); 149 } 150 151 //===----------------------------------------------------------------------===// 152 // Transfer function for unary operators. 153 //===----------------------------------------------------------------------===// 154 155 SVal SimpleSValBuilder::evalMinus(NonLoc val) { 156 switch (val.getSubKind()) { 157 case nonloc::ConcreteIntKind: 158 return cast<nonloc::ConcreteInt>(val).evalMinus(*this); 159 default: 160 return UnknownVal(); 161 } 162 } 163 164 SVal SimpleSValBuilder::evalComplement(NonLoc X) { 165 switch (X.getSubKind()) { 166 case nonloc::ConcreteIntKind: 167 return cast<nonloc::ConcreteInt>(X).evalComplement(*this); 168 default: 169 return UnknownVal(); 170 } 171 } 172 173 //===----------------------------------------------------------------------===// 174 // Transfer function for binary operators. 175 //===----------------------------------------------------------------------===// 176 177 static BinaryOperator::Opcode NegateComparison(BinaryOperator::Opcode op) { 178 switch (op) { 179 default: 180 llvm_unreachable("Invalid opcode."); 181 case BO_LT: return BO_GE; 182 case BO_GT: return BO_LE; 183 case BO_LE: return BO_GT; 184 case BO_GE: return BO_LT; 185 case BO_EQ: return BO_NE; 186 case BO_NE: return BO_EQ; 187 } 188 } 189 190 static BinaryOperator::Opcode ReverseComparison(BinaryOperator::Opcode op) { 191 switch (op) { 192 default: 193 llvm_unreachable("Invalid opcode."); 194 case BO_LT: return BO_GT; 195 case BO_GT: return BO_LT; 196 case BO_LE: return BO_GE; 197 case BO_GE: return BO_LE; 198 case BO_EQ: 199 case BO_NE: 200 return op; 201 } 202 } 203 204 SVal SimpleSValBuilder::MakeSymIntVal(const SymExpr *LHS, 205 BinaryOperator::Opcode op, 206 const llvm::APSInt &RHS, 207 QualType resultTy) { 208 bool isIdempotent = false; 209 210 // Check for a few special cases with known reductions first. 211 switch (op) { 212 default: 213 // We can't reduce this case; just treat it normally. 214 break; 215 case BO_Mul: 216 // a*0 and a*1 217 if (RHS == 0) 218 return makeIntVal(0, resultTy); 219 else if (RHS == 1) 220 isIdempotent = true; 221 break; 222 case BO_Div: 223 // a/0 and a/1 224 if (RHS == 0) 225 // This is also handled elsewhere. 226 return UndefinedVal(); 227 else if (RHS == 1) 228 isIdempotent = true; 229 break; 230 case BO_Rem: 231 // a%0 and a%1 232 if (RHS == 0) 233 // This is also handled elsewhere. 234 return UndefinedVal(); 235 else if (RHS == 1) 236 return makeIntVal(0, resultTy); 237 break; 238 case BO_Add: 239 case BO_Sub: 240 case BO_Shl: 241 case BO_Shr: 242 case BO_Xor: 243 // a+0, a-0, a<<0, a>>0, a^0 244 if (RHS == 0) 245 isIdempotent = true; 246 break; 247 case BO_And: 248 // a&0 and a&(~0) 249 if (RHS == 0) 250 return makeIntVal(0, resultTy); 251 else if (RHS.isAllOnesValue()) 252 isIdempotent = true; 253 break; 254 case BO_Or: 255 // a|0 and a|(~0) 256 if (RHS == 0) 257 isIdempotent = true; 258 else if (RHS.isAllOnesValue()) { 259 const llvm::APSInt &Result = BasicVals.Convert(resultTy, RHS); 260 return nonloc::ConcreteInt(Result); 261 } 262 break; 263 } 264 265 // Idempotent ops (like a*1) can still change the type of an expression. 266 // Wrap the LHS up in a NonLoc again and let evalCastFromNonLoc do the 267 // dirty work. 268 if (isIdempotent) 269 return evalCastFromNonLoc(nonloc::SymbolVal(LHS), resultTy); 270 271 // If we reach this point, the expression cannot be simplified. 272 // Make a SymbolVal for the entire expression, after converting the RHS. 273 const llvm::APSInt *ConvertedRHS = &RHS; 274 if (BinaryOperator::isComparisonOp(op)) { 275 // We're looking for a type big enough to compare the symbolic value 276 // with the given constant. 277 // FIXME: This is an approximation of Sema::UsualArithmeticConversions. 278 ASTContext &Ctx = getContext(); 279 QualType SymbolType = LHS->getType(Ctx); 280 uint64_t ValWidth = RHS.getBitWidth(); 281 uint64_t TypeWidth = Ctx.getTypeSize(SymbolType); 282 283 if (ValWidth < TypeWidth) { 284 // If the value is too small, extend it. 285 ConvertedRHS = &BasicVals.Convert(SymbolType, RHS); 286 } else if (ValWidth == TypeWidth) { 287 // If the value is signed but the symbol is unsigned, do the comparison 288 // in unsigned space. [C99 6.3.1.8] 289 // (For the opposite case, the value is already unsigned.) 290 if (RHS.isSigned() && !SymbolType->isSignedIntegerOrEnumerationType()) 291 ConvertedRHS = &BasicVals.Convert(SymbolType, RHS); 292 } 293 } else 294 ConvertedRHS = &BasicVals.Convert(resultTy, RHS); 295 296 return makeNonLoc(LHS, op, *ConvertedRHS, resultTy); 297 } 298 299 SVal SimpleSValBuilder::evalBinOpNN(ProgramStateRef state, 300 BinaryOperator::Opcode op, 301 NonLoc lhs, NonLoc rhs, 302 QualType resultTy) { 303 NonLoc InputLHS = lhs; 304 NonLoc InputRHS = rhs; 305 306 // Handle trivial case where left-side and right-side are the same. 307 if (lhs == rhs) 308 switch (op) { 309 default: 310 break; 311 case BO_EQ: 312 case BO_LE: 313 case BO_GE: 314 return makeTruthVal(true, resultTy); 315 case BO_LT: 316 case BO_GT: 317 case BO_NE: 318 return makeTruthVal(false, resultTy); 319 case BO_Xor: 320 case BO_Sub: 321 if (resultTy->isIntegralOrEnumerationType()) 322 return makeIntVal(0, resultTy); 323 return evalCastFromNonLoc(makeIntVal(0, /*Unsigned=*/false), resultTy); 324 case BO_Or: 325 case BO_And: 326 return evalCastFromNonLoc(lhs, resultTy); 327 } 328 329 while (1) { 330 switch (lhs.getSubKind()) { 331 default: 332 return makeSymExprValNN(state, op, lhs, rhs, resultTy); 333 case nonloc::LocAsIntegerKind: { 334 Loc lhsL = cast<nonloc::LocAsInteger>(lhs).getLoc(); 335 switch (rhs.getSubKind()) { 336 case nonloc::LocAsIntegerKind: 337 return evalBinOpLL(state, op, lhsL, 338 cast<nonloc::LocAsInteger>(rhs).getLoc(), 339 resultTy); 340 case nonloc::ConcreteIntKind: { 341 // Transform the integer into a location and compare. 342 llvm::APSInt i = cast<nonloc::ConcreteInt>(rhs).getValue(); 343 BasicVals.getAPSIntType(Context.VoidPtrTy).apply(i); 344 return evalBinOpLL(state, op, lhsL, makeLoc(i), resultTy); 345 } 346 default: 347 switch (op) { 348 case BO_EQ: 349 return makeTruthVal(false, resultTy); 350 case BO_NE: 351 return makeTruthVal(true, resultTy); 352 default: 353 // This case also handles pointer arithmetic. 354 return makeSymExprValNN(state, op, InputLHS, InputRHS, resultTy); 355 } 356 } 357 } 358 case nonloc::ConcreteIntKind: { 359 llvm::APSInt LHSValue = cast<nonloc::ConcreteInt>(lhs).getValue(); 360 361 // If we're dealing with two known constants, just perform the operation. 362 if (const llvm::APSInt *KnownRHSValue = getKnownValue(state, rhs)) { 363 llvm::APSInt RHSValue = *KnownRHSValue; 364 if (BinaryOperator::isComparisonOp(op)) { 365 // We're looking for a type big enough to compare the two values. 366 // FIXME: This is not correct. char + short will result in a promotion 367 // to int. Unfortunately we have lost types by this point. 368 APSIntType CompareType = std::max(APSIntType(LHSValue), 369 APSIntType(RHSValue)); 370 CompareType.apply(LHSValue); 371 CompareType.apply(RHSValue); 372 } else if (!BinaryOperator::isShiftOp(op)) { 373 APSIntType IntType = BasicVals.getAPSIntType(resultTy); 374 IntType.apply(LHSValue); 375 IntType.apply(RHSValue); 376 } 377 378 const llvm::APSInt *Result = 379 BasicVals.evalAPSInt(op, LHSValue, RHSValue); 380 if (!Result) 381 return UndefinedVal(); 382 383 return nonloc::ConcreteInt(*Result); 384 } 385 386 // Swap the left and right sides and flip the operator if doing so 387 // allows us to better reason about the expression (this is a form 388 // of expression canonicalization). 389 // While we're at it, catch some special cases for non-commutative ops. 390 switch (op) { 391 case BO_LT: 392 case BO_GT: 393 case BO_LE: 394 case BO_GE: 395 op = ReverseComparison(op); 396 // FALL-THROUGH 397 case BO_EQ: 398 case BO_NE: 399 case BO_Add: 400 case BO_Mul: 401 case BO_And: 402 case BO_Xor: 403 case BO_Or: 404 std::swap(lhs, rhs); 405 continue; 406 case BO_Shr: 407 // (~0)>>a 408 if (LHSValue.isAllOnesValue() && LHSValue.isSigned()) 409 return evalCastFromNonLoc(lhs, resultTy); 410 // FALL-THROUGH 411 case BO_Shl: 412 // 0<<a and 0>>a 413 if (LHSValue == 0) 414 return evalCastFromNonLoc(lhs, resultTy); 415 return makeSymExprValNN(state, op, InputLHS, InputRHS, resultTy); 416 default: 417 return makeSymExprValNN(state, op, InputLHS, InputRHS, resultTy); 418 } 419 } 420 case nonloc::SymbolValKind: { 421 // We only handle LHS as simple symbols or SymIntExprs. 422 SymbolRef Sym = cast<nonloc::SymbolVal>(lhs).getSymbol(); 423 424 // LHS is a symbolic expression. 425 if (const SymIntExpr *symIntExpr = dyn_cast<SymIntExpr>(Sym)) { 426 427 // Is this a logical not? (!x is represented as x == 0.) 428 if (op == BO_EQ && rhs.isZeroConstant()) { 429 // We know how to negate certain expressions. Simplify them here. 430 431 BinaryOperator::Opcode opc = symIntExpr->getOpcode(); 432 switch (opc) { 433 default: 434 // We don't know how to negate this operation. 435 // Just handle it as if it were a normal comparison to 0. 436 break; 437 case BO_LAnd: 438 case BO_LOr: 439 llvm_unreachable("Logical operators handled by branching logic."); 440 case BO_Assign: 441 case BO_MulAssign: 442 case BO_DivAssign: 443 case BO_RemAssign: 444 case BO_AddAssign: 445 case BO_SubAssign: 446 case BO_ShlAssign: 447 case BO_ShrAssign: 448 case BO_AndAssign: 449 case BO_XorAssign: 450 case BO_OrAssign: 451 case BO_Comma: 452 llvm_unreachable("'=' and ',' operators handled by ExprEngine."); 453 case BO_PtrMemD: 454 case BO_PtrMemI: 455 llvm_unreachable("Pointer arithmetic not handled here."); 456 case BO_LT: 457 case BO_GT: 458 case BO_LE: 459 case BO_GE: 460 case BO_EQ: 461 case BO_NE: 462 // Negate the comparison and make a value. 463 opc = NegateComparison(opc); 464 assert(symIntExpr->getType(Context) == resultTy); 465 return makeNonLoc(symIntExpr->getLHS(), opc, 466 symIntExpr->getRHS(), resultTy); 467 } 468 } 469 470 // For now, only handle expressions whose RHS is a constant. 471 if (const llvm::APSInt *RHSValue = getKnownValue(state, rhs)) { 472 // If both the LHS and the current expression are additive, 473 // fold their constants and try again. 474 if (BinaryOperator::isAdditiveOp(op)) { 475 BinaryOperator::Opcode lop = symIntExpr->getOpcode(); 476 if (BinaryOperator::isAdditiveOp(lop)) { 477 // Convert the two constants to a common type, then combine them. 478 479 // resultTy may not be the best type to convert to, but it's 480 // probably the best choice in expressions with mixed type 481 // (such as x+1U+2LL). The rules for implicit conversions should 482 // choose a reasonable type to preserve the expression, and will 483 // at least match how the value is going to be used. 484 APSIntType IntType = BasicVals.getAPSIntType(resultTy); 485 const llvm::APSInt &first = IntType.convert(symIntExpr->getRHS()); 486 const llvm::APSInt &second = IntType.convert(*RHSValue); 487 488 const llvm::APSInt *newRHS; 489 if (lop == op) 490 newRHS = BasicVals.evalAPSInt(BO_Add, first, second); 491 else 492 newRHS = BasicVals.evalAPSInt(BO_Sub, first, second); 493 494 assert(newRHS && "Invalid operation despite common type!"); 495 rhs = nonloc::ConcreteInt(*newRHS); 496 lhs = nonloc::SymbolVal(symIntExpr->getLHS()); 497 op = lop; 498 continue; 499 } 500 } 501 502 // Otherwise, make a SymIntExpr out of the expression. 503 return MakeSymIntVal(symIntExpr, op, *RHSValue, resultTy); 504 } 505 506 507 } else if (isa<SymbolData>(Sym)) { 508 // Does the symbol simplify to a constant? If so, "fold" the constant 509 // by setting 'lhs' to a ConcreteInt and try again. 510 if (const llvm::APSInt *Constant = state->getConstraintManager() 511 .getSymVal(state, Sym)) { 512 lhs = nonloc::ConcreteInt(*Constant); 513 continue; 514 } 515 516 // Is the RHS a constant? 517 if (const llvm::APSInt *RHSValue = getKnownValue(state, rhs)) 518 return MakeSymIntVal(Sym, op, *RHSValue, resultTy); 519 } 520 521 // Give up -- this is not a symbolic expression we can handle. 522 return makeSymExprValNN(state, op, InputLHS, InputRHS, resultTy); 523 } 524 } 525 } 526 } 527 528 // FIXME: all this logic will change if/when we have MemRegion::getLocation(). 529 SVal SimpleSValBuilder::evalBinOpLL(ProgramStateRef state, 530 BinaryOperator::Opcode op, 531 Loc lhs, Loc rhs, 532 QualType resultTy) { 533 // Only comparisons and subtractions are valid operations on two pointers. 534 // See [C99 6.5.5 through 6.5.14] or [C++0x 5.6 through 5.15]. 535 // However, if a pointer is casted to an integer, evalBinOpNN may end up 536 // calling this function with another operation (PR7527). We don't attempt to 537 // model this for now, but it could be useful, particularly when the 538 // "location" is actually an integer value that's been passed through a void*. 539 if (!(BinaryOperator::isComparisonOp(op) || op == BO_Sub)) 540 return UnknownVal(); 541 542 // Special cases for when both sides are identical. 543 if (lhs == rhs) { 544 switch (op) { 545 default: 546 llvm_unreachable("Unimplemented operation for two identical values"); 547 case BO_Sub: 548 return makeZeroVal(resultTy); 549 case BO_EQ: 550 case BO_LE: 551 case BO_GE: 552 return makeTruthVal(true, resultTy); 553 case BO_NE: 554 case BO_LT: 555 case BO_GT: 556 return makeTruthVal(false, resultTy); 557 } 558 } 559 560 switch (lhs.getSubKind()) { 561 default: 562 llvm_unreachable("Ordering not implemented for this Loc."); 563 564 case loc::GotoLabelKind: 565 // The only thing we know about labels is that they're non-null. 566 if (rhs.isZeroConstant()) { 567 switch (op) { 568 default: 569 break; 570 case BO_Sub: 571 return evalCastFromLoc(lhs, resultTy); 572 case BO_EQ: 573 case BO_LE: 574 case BO_LT: 575 return makeTruthVal(false, resultTy); 576 case BO_NE: 577 case BO_GT: 578 case BO_GE: 579 return makeTruthVal(true, resultTy); 580 } 581 } 582 // There may be two labels for the same location, and a function region may 583 // have the same address as a label at the start of the function (depending 584 // on the ABI). 585 // FIXME: we can probably do a comparison against other MemRegions, though. 586 // FIXME: is there a way to tell if two labels refer to the same location? 587 return UnknownVal(); 588 589 case loc::ConcreteIntKind: { 590 // If one of the operands is a symbol and the other is a constant, 591 // build an expression for use by the constraint manager. 592 if (SymbolRef rSym = rhs.getAsLocSymbol()) { 593 // We can only build expressions with symbols on the left, 594 // so we need a reversible operator. 595 if (!BinaryOperator::isComparisonOp(op)) 596 return UnknownVal(); 597 598 const llvm::APSInt &lVal = cast<loc::ConcreteInt>(lhs).getValue(); 599 return makeNonLoc(rSym, ReverseComparison(op), lVal, resultTy); 600 } 601 602 // If both operands are constants, just perform the operation. 603 if (loc::ConcreteInt *rInt = dyn_cast<loc::ConcreteInt>(&rhs)) { 604 SVal ResultVal = cast<loc::ConcreteInt>(lhs).evalBinOp(BasicVals, op, 605 *rInt); 606 if (Loc *Result = dyn_cast<Loc>(&ResultVal)) 607 return evalCastFromLoc(*Result, resultTy); 608 else 609 return UnknownVal(); 610 } 611 612 // Special case comparisons against NULL. 613 // This must come after the test if the RHS is a symbol, which is used to 614 // build constraints. The address of any non-symbolic region is guaranteed 615 // to be non-NULL, as is any label. 616 assert(isa<loc::MemRegionVal>(rhs) || isa<loc::GotoLabel>(rhs)); 617 if (lhs.isZeroConstant()) { 618 switch (op) { 619 default: 620 break; 621 case BO_EQ: 622 case BO_GT: 623 case BO_GE: 624 return makeTruthVal(false, resultTy); 625 case BO_NE: 626 case BO_LT: 627 case BO_LE: 628 return makeTruthVal(true, resultTy); 629 } 630 } 631 632 // Comparing an arbitrary integer to a region or label address is 633 // completely unknowable. 634 return UnknownVal(); 635 } 636 case loc::MemRegionKind: { 637 if (loc::ConcreteInt *rInt = dyn_cast<loc::ConcreteInt>(&rhs)) { 638 // If one of the operands is a symbol and the other is a constant, 639 // build an expression for use by the constraint manager. 640 if (SymbolRef lSym = lhs.getAsLocSymbol()) 641 return MakeSymIntVal(lSym, op, rInt->getValue(), resultTy); 642 643 // Special case comparisons to NULL. 644 // This must come after the test if the LHS is a symbol, which is used to 645 // build constraints. The address of any non-symbolic region is guaranteed 646 // to be non-NULL. 647 if (rInt->isZeroConstant()) { 648 switch (op) { 649 default: 650 break; 651 case BO_Sub: 652 return evalCastFromLoc(lhs, resultTy); 653 case BO_EQ: 654 case BO_LT: 655 case BO_LE: 656 return makeTruthVal(false, resultTy); 657 case BO_NE: 658 case BO_GT: 659 case BO_GE: 660 return makeTruthVal(true, resultTy); 661 } 662 } 663 664 // Comparing a region to an arbitrary integer is completely unknowable. 665 return UnknownVal(); 666 } 667 668 // Get both values as regions, if possible. 669 const MemRegion *LeftMR = lhs.getAsRegion(); 670 assert(LeftMR && "MemRegionKind SVal doesn't have a region!"); 671 672 const MemRegion *RightMR = rhs.getAsRegion(); 673 if (!RightMR) 674 // The RHS is probably a label, which in theory could address a region. 675 // FIXME: we can probably make a more useful statement about non-code 676 // regions, though. 677 return UnknownVal(); 678 679 const MemSpaceRegion *LeftMS = LeftMR->getMemorySpace(); 680 const MemSpaceRegion *RightMS = RightMR->getMemorySpace(); 681 const MemSpaceRegion *UnknownMS = MemMgr.getUnknownRegion(); 682 const MemRegion *LeftBase = LeftMR->getBaseRegion(); 683 const MemRegion *RightBase = RightMR->getBaseRegion(); 684 685 // If the two regions are from different known memory spaces they cannot be 686 // equal. Also, assume that no symbolic region (whose memory space is 687 // unknown) is on the stack. 688 if (LeftMS != RightMS && 689 ((LeftMS != UnknownMS && RightMS != UnknownMS) || 690 (isa<StackSpaceRegion>(LeftMS) || isa<StackSpaceRegion>(RightMS)))) { 691 switch (op) { 692 default: 693 return UnknownVal(); 694 case BO_EQ: 695 return makeTruthVal(false, resultTy); 696 case BO_NE: 697 return makeTruthVal(true, resultTy); 698 } 699 } 700 701 // If both values wrap regions, see if they're from different base regions. 702 // Note, heap base symbolic regions are assumed to not alias with 703 // each other; for example, we assume that malloc returns different address 704 // on each invocation. 705 if (LeftBase != RightBase && 706 ((!isa<SymbolicRegion>(LeftBase) && !isa<SymbolicRegion>(RightBase)) || 707 (isa<HeapSpaceRegion>(LeftMS) || isa<HeapSpaceRegion>(RightMS))) ){ 708 switch (op) { 709 default: 710 return UnknownVal(); 711 case BO_EQ: 712 return makeTruthVal(false, resultTy); 713 case BO_NE: 714 return makeTruthVal(true, resultTy); 715 } 716 } 717 718 // FIXME: If/when there is a getAsRawOffset() for FieldRegions, this 719 // ElementRegion path and the FieldRegion path below should be unified. 720 if (const ElementRegion *LeftER = dyn_cast<ElementRegion>(LeftMR)) { 721 // First see if the right region is also an ElementRegion. 722 const ElementRegion *RightER = dyn_cast<ElementRegion>(RightMR); 723 if (!RightER) 724 return UnknownVal(); 725 726 // Next, see if the two ERs have the same super-region and matching types. 727 // FIXME: This should do something useful even if the types don't match, 728 // though if both indexes are constant the RegionRawOffset path will 729 // give the correct answer. 730 if (LeftER->getSuperRegion() == RightER->getSuperRegion() && 731 LeftER->getElementType() == RightER->getElementType()) { 732 // Get the left index and cast it to the correct type. 733 // If the index is unknown or undefined, bail out here. 734 SVal LeftIndexVal = LeftER->getIndex(); 735 NonLoc *LeftIndex = dyn_cast<NonLoc>(&LeftIndexVal); 736 if (!LeftIndex) 737 return UnknownVal(); 738 LeftIndexVal = evalCastFromNonLoc(*LeftIndex, resultTy); 739 LeftIndex = dyn_cast<NonLoc>(&LeftIndexVal); 740 if (!LeftIndex) 741 return UnknownVal(); 742 743 // Do the same for the right index. 744 SVal RightIndexVal = RightER->getIndex(); 745 NonLoc *RightIndex = dyn_cast<NonLoc>(&RightIndexVal); 746 if (!RightIndex) 747 return UnknownVal(); 748 RightIndexVal = evalCastFromNonLoc(*RightIndex, resultTy); 749 RightIndex = dyn_cast<NonLoc>(&RightIndexVal); 750 if (!RightIndex) 751 return UnknownVal(); 752 753 // Actually perform the operation. 754 // evalBinOpNN expects the two indexes to already be the right type. 755 return evalBinOpNN(state, op, *LeftIndex, *RightIndex, resultTy); 756 } 757 758 // If the element indexes aren't comparable, see if the raw offsets are. 759 RegionRawOffset LeftOffset = LeftER->getAsArrayOffset(); 760 RegionRawOffset RightOffset = RightER->getAsArrayOffset(); 761 762 if (LeftOffset.getRegion() != NULL && 763 LeftOffset.getRegion() == RightOffset.getRegion()) { 764 CharUnits left = LeftOffset.getOffset(); 765 CharUnits right = RightOffset.getOffset(); 766 767 switch (op) { 768 default: 769 return UnknownVal(); 770 case BO_LT: 771 return makeTruthVal(left < right, resultTy); 772 case BO_GT: 773 return makeTruthVal(left > right, resultTy); 774 case BO_LE: 775 return makeTruthVal(left <= right, resultTy); 776 case BO_GE: 777 return makeTruthVal(left >= right, resultTy); 778 case BO_EQ: 779 return makeTruthVal(left == right, resultTy); 780 case BO_NE: 781 return makeTruthVal(left != right, resultTy); 782 } 783 } 784 785 // If we get here, we have no way of comparing the ElementRegions. 786 return UnknownVal(); 787 } 788 789 // See if both regions are fields of the same structure. 790 // FIXME: This doesn't handle nesting, inheritance, or Objective-C ivars. 791 if (const FieldRegion *LeftFR = dyn_cast<FieldRegion>(LeftMR)) { 792 // Only comparisons are meaningful here! 793 if (!BinaryOperator::isComparisonOp(op)) 794 return UnknownVal(); 795 796 // First see if the right region is also a FieldRegion. 797 const FieldRegion *RightFR = dyn_cast<FieldRegion>(RightMR); 798 if (!RightFR) 799 return UnknownVal(); 800 801 // Next, see if the two FRs have the same super-region. 802 // FIXME: This doesn't handle casts yet, and simply stripping the casts 803 // doesn't help. 804 if (LeftFR->getSuperRegion() != RightFR->getSuperRegion()) 805 return UnknownVal(); 806 807 const FieldDecl *LeftFD = LeftFR->getDecl(); 808 const FieldDecl *RightFD = RightFR->getDecl(); 809 const RecordDecl *RD = LeftFD->getParent(); 810 811 // Make sure the two FRs are from the same kind of record. Just in case! 812 // FIXME: This is probably where inheritance would be a problem. 813 if (RD != RightFD->getParent()) 814 return UnknownVal(); 815 816 // We know for sure that the two fields are not the same, since that 817 // would have given us the same SVal. 818 if (op == BO_EQ) 819 return makeTruthVal(false, resultTy); 820 if (op == BO_NE) 821 return makeTruthVal(true, resultTy); 822 823 // Iterate through the fields and see which one comes first. 824 // [C99 6.7.2.1.13] "Within a structure object, the non-bit-field 825 // members and the units in which bit-fields reside have addresses that 826 // increase in the order in which they are declared." 827 bool leftFirst = (op == BO_LT || op == BO_LE); 828 for (RecordDecl::field_iterator I = RD->field_begin(), 829 E = RD->field_end(); I!=E; ++I) { 830 if (*I == LeftFD) 831 return makeTruthVal(leftFirst, resultTy); 832 if (*I == RightFD) 833 return makeTruthVal(!leftFirst, resultTy); 834 } 835 836 llvm_unreachable("Fields not found in parent record's definition"); 837 } 838 839 // If we get here, we have no way of comparing the regions. 840 return UnknownVal(); 841 } 842 } 843 } 844 845 SVal SimpleSValBuilder::evalBinOpLN(ProgramStateRef state, 846 BinaryOperator::Opcode op, 847 Loc lhs, NonLoc rhs, QualType resultTy) { 848 849 // Special case: rhs is a zero constant. 850 if (rhs.isZeroConstant()) 851 return lhs; 852 853 // Special case: 'rhs' is an integer that has the same width as a pointer and 854 // we are using the integer location in a comparison. Normally this cannot be 855 // triggered, but transfer functions like those for OSCommpareAndSwapBarrier32 856 // can generate comparisons that trigger this code. 857 // FIXME: Are all locations guaranteed to have pointer width? 858 if (BinaryOperator::isComparisonOp(op)) { 859 if (nonloc::ConcreteInt *rhsInt = dyn_cast<nonloc::ConcreteInt>(&rhs)) { 860 const llvm::APSInt *x = &rhsInt->getValue(); 861 ASTContext &ctx = Context; 862 if (ctx.getTypeSize(ctx.VoidPtrTy) == x->getBitWidth()) { 863 // Convert the signedness of the integer (if necessary). 864 if (x->isSigned()) 865 x = &getBasicValueFactory().getValue(*x, true); 866 867 return evalBinOpLL(state, op, lhs, loc::ConcreteInt(*x), resultTy); 868 } 869 } 870 return UnknownVal(); 871 } 872 873 // We are dealing with pointer arithmetic. 874 875 // Handle pointer arithmetic on constant values. 876 if (nonloc::ConcreteInt *rhsInt = dyn_cast<nonloc::ConcreteInt>(&rhs)) { 877 if (loc::ConcreteInt *lhsInt = dyn_cast<loc::ConcreteInt>(&lhs)) { 878 const llvm::APSInt &leftI = lhsInt->getValue(); 879 assert(leftI.isUnsigned()); 880 llvm::APSInt rightI(rhsInt->getValue(), /* isUnsigned */ true); 881 882 // Convert the bitwidth of rightI. This should deal with overflow 883 // since we are dealing with concrete values. 884 rightI = rightI.extOrTrunc(leftI.getBitWidth()); 885 886 // Offset the increment by the pointer size. 887 llvm::APSInt Multiplicand(rightI.getBitWidth(), /* isUnsigned */ true); 888 rightI *= Multiplicand; 889 890 // Compute the adjusted pointer. 891 switch (op) { 892 case BO_Add: 893 rightI = leftI + rightI; 894 break; 895 case BO_Sub: 896 rightI = leftI - rightI; 897 break; 898 default: 899 llvm_unreachable("Invalid pointer arithmetic operation"); 900 } 901 return loc::ConcreteInt(getBasicValueFactory().getValue(rightI)); 902 } 903 } 904 905 // Handle cases where 'lhs' is a region. 906 if (const MemRegion *region = lhs.getAsRegion()) { 907 rhs = cast<NonLoc>(convertToArrayIndex(rhs)); 908 SVal index = UnknownVal(); 909 const MemRegion *superR = 0; 910 QualType elementType; 911 912 if (const ElementRegion *elemReg = dyn_cast<ElementRegion>(region)) { 913 assert(op == BO_Add || op == BO_Sub); 914 index = evalBinOpNN(state, op, elemReg->getIndex(), rhs, 915 getArrayIndexType()); 916 superR = elemReg->getSuperRegion(); 917 elementType = elemReg->getElementType(); 918 } 919 else if (isa<SubRegion>(region)) { 920 superR = region; 921 index = rhs; 922 if (resultTy->isAnyPointerType()) 923 elementType = resultTy->getPointeeType(); 924 } 925 926 if (NonLoc *indexV = dyn_cast<NonLoc>(&index)) { 927 return loc::MemRegionVal(MemMgr.getElementRegion(elementType, *indexV, 928 superR, getContext())); 929 } 930 } 931 return UnknownVal(); 932 } 933 934 const llvm::APSInt *SimpleSValBuilder::getKnownValue(ProgramStateRef state, 935 SVal V) { 936 if (V.isUnknownOrUndef()) 937 return NULL; 938 939 if (loc::ConcreteInt* X = dyn_cast<loc::ConcreteInt>(&V)) 940 return &X->getValue(); 941 942 if (nonloc::ConcreteInt* X = dyn_cast<nonloc::ConcreteInt>(&V)) 943 return &X->getValue(); 944 945 if (SymbolRef Sym = V.getAsSymbol()) 946 return state->getConstraintManager().getSymVal(state, Sym); 947 948 // FIXME: Add support for SymExprs. 949 return NULL; 950 } 951