1 //===- ThreadSafety.cpp ----------------------------------------*- 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 // A intra-procedural analysis for thread safety (e.g. deadlocks and race 11 // conditions), based off of an annotation system. 12 // 13 // See http://clang.llvm.org/docs/ThreadSafetyAnalysis.html 14 // for more information. 15 // 16 //===----------------------------------------------------------------------===// 17 18 #include "clang/AST/Attr.h" 19 #include "clang/AST/DeclCXX.h" 20 #include "clang/AST/ExprCXX.h" 21 #include "clang/AST/StmtCXX.h" 22 #include "clang/AST/StmtVisitor.h" 23 #include "clang/Analysis/Analyses/PostOrderCFGView.h" 24 #include "clang/Analysis/Analyses/ThreadSafety.h" 25 #include "clang/Analysis/Analyses/ThreadSafetyTIL.h" 26 #include "clang/Analysis/Analyses/ThreadSafetyCommon.h" 27 #include "clang/Analysis/AnalysisContext.h" 28 #include "clang/Analysis/CFG.h" 29 #include "clang/Analysis/CFGStmtMap.h" 30 #include "clang/Basic/OperatorKinds.h" 31 #include "clang/Basic/SourceLocation.h" 32 #include "clang/Basic/SourceManager.h" 33 #include "llvm/ADT/BitVector.h" 34 #include "llvm/ADT/FoldingSet.h" 35 #include "llvm/ADT/ImmutableMap.h" 36 #include "llvm/ADT/PostOrderIterator.h" 37 #include "llvm/ADT/SmallVector.h" 38 #include "llvm/ADT/StringRef.h" 39 #include "llvm/Support/raw_ostream.h" 40 #include <algorithm> 41 #include <utility> 42 #include <vector> 43 44 using namespace clang; 45 using namespace thread_safety; 46 47 // Key method definition 48 ThreadSafetyHandler::~ThreadSafetyHandler() {} 49 50 namespace { 51 52 /// SExpr implements a simple expression language that is used to store, 53 /// compare, and pretty-print C++ expressions. Unlike a clang Expr, a SExpr 54 /// does not capture surface syntax, and it does not distinguish between 55 /// C++ concepts, like pointers and references, that have no real semantic 56 /// differences. This simplicity allows SExprs to be meaningfully compared, 57 /// e.g. 58 /// (x) = x 59 /// (*this).foo = this->foo 60 /// *&a = a 61 /// 62 /// Thread-safety analysis works by comparing lock expressions. Within the 63 /// body of a function, an expression such as "x->foo->bar.mu" will resolve to 64 /// a particular mutex object at run-time. Subsequent occurrences of the same 65 /// expression (where "same" means syntactic equality) will refer to the same 66 /// run-time object if three conditions hold: 67 /// (1) Local variables in the expression, such as "x" have not changed. 68 /// (2) Values on the heap that affect the expression have not changed. 69 /// (3) The expression involves only pure function calls. 70 /// 71 /// The current implementation assumes, but does not verify, that multiple uses 72 /// of the same lock expression satisfies these criteria. 73 class SExpr { 74 private: 75 enum ExprOp { 76 EOP_Nop, ///< No-op 77 EOP_Wildcard, ///< Matches anything. 78 EOP_Universal, ///< Universal lock. 79 EOP_This, ///< This keyword. 80 EOP_NVar, ///< Named variable. 81 EOP_LVar, ///< Local variable. 82 EOP_Dot, ///< Field access 83 EOP_Call, ///< Function call 84 EOP_MCall, ///< Method call 85 EOP_Index, ///< Array index 86 EOP_Unary, ///< Unary operation 87 EOP_Binary, ///< Binary operation 88 EOP_Unknown ///< Catchall for everything else 89 }; 90 91 92 class SExprNode { 93 private: 94 unsigned char Op; ///< Opcode of the root node 95 unsigned char Flags; ///< Additional opcode-specific data 96 unsigned short Sz; ///< Number of child nodes 97 const void* Data; ///< Additional opcode-specific data 98 99 public: 100 SExprNode(ExprOp O, unsigned F, const void* D) 101 : Op(static_cast<unsigned char>(O)), 102 Flags(static_cast<unsigned char>(F)), Sz(1), Data(D) 103 { } 104 105 unsigned size() const { return Sz; } 106 void setSize(unsigned S) { Sz = S; } 107 108 ExprOp kind() const { return static_cast<ExprOp>(Op); } 109 110 const NamedDecl* getNamedDecl() const { 111 assert(Op == EOP_NVar || Op == EOP_LVar || Op == EOP_Dot); 112 return reinterpret_cast<const NamedDecl*>(Data); 113 } 114 115 const NamedDecl* getFunctionDecl() const { 116 assert(Op == EOP_Call || Op == EOP_MCall); 117 return reinterpret_cast<const NamedDecl*>(Data); 118 } 119 120 bool isArrow() const { return Op == EOP_Dot && Flags == 1; } 121 void setArrow(bool A) { Flags = A ? 1 : 0; } 122 123 unsigned arity() const { 124 switch (Op) { 125 case EOP_Nop: return 0; 126 case EOP_Wildcard: return 0; 127 case EOP_Universal: return 0; 128 case EOP_NVar: return 0; 129 case EOP_LVar: return 0; 130 case EOP_This: return 0; 131 case EOP_Dot: return 1; 132 case EOP_Call: return Flags+1; // First arg is function. 133 case EOP_MCall: return Flags+1; // First arg is implicit obj. 134 case EOP_Index: return 2; 135 case EOP_Unary: return 1; 136 case EOP_Binary: return 2; 137 case EOP_Unknown: return Flags; 138 } 139 return 0; 140 } 141 142 bool operator==(const SExprNode& Other) const { 143 // Ignore flags and size -- they don't matter. 144 return (Op == Other.Op && 145 Data == Other.Data); 146 } 147 148 bool operator!=(const SExprNode& Other) const { 149 return !(*this == Other); 150 } 151 152 bool matches(const SExprNode& Other) const { 153 return (*this == Other) || 154 (Op == EOP_Wildcard) || 155 (Other.Op == EOP_Wildcard); 156 } 157 }; 158 159 160 /// \brief Encapsulates the lexical context of a function call. The lexical 161 /// context includes the arguments to the call, including the implicit object 162 /// argument. When an attribute containing a mutex expression is attached to 163 /// a method, the expression may refer to formal parameters of the method. 164 /// Actual arguments must be substituted for formal parameters to derive 165 /// the appropriate mutex expression in the lexical context where the function 166 /// is called. PrevCtx holds the context in which the arguments themselves 167 /// should be evaluated; multiple calling contexts can be chained together 168 /// by the lock_returned attribute. 169 struct CallingContext { 170 const NamedDecl* AttrDecl; // The decl to which the attribute is attached. 171 const Expr* SelfArg; // Implicit object argument -- e.g. 'this' 172 bool SelfArrow; // is Self referred to with -> or .? 173 unsigned NumArgs; // Number of funArgs 174 const Expr* const* FunArgs; // Function arguments 175 CallingContext* PrevCtx; // The previous context; or 0 if none. 176 177 CallingContext(const NamedDecl *D) 178 : AttrDecl(D), SelfArg(0), SelfArrow(false), NumArgs(0), FunArgs(0), 179 PrevCtx(0) {} 180 }; 181 182 typedef SmallVector<SExprNode, 4> NodeVector; 183 184 private: 185 // A SExpr is a list of SExprNodes in prefix order. The Size field allows 186 // the list to be traversed as a tree. 187 NodeVector NodeVec; 188 189 private: 190 unsigned make(ExprOp O, unsigned F = 0, const void *D = 0) { 191 NodeVec.push_back(SExprNode(O, F, D)); 192 return NodeVec.size() - 1; 193 } 194 195 unsigned makeNop() { 196 return make(EOP_Nop); 197 } 198 199 unsigned makeWildcard() { 200 return make(EOP_Wildcard); 201 } 202 203 unsigned makeUniversal() { 204 return make(EOP_Universal); 205 } 206 207 unsigned makeNamedVar(const NamedDecl *D) { 208 return make(EOP_NVar, 0, D); 209 } 210 211 unsigned makeLocalVar(const NamedDecl *D) { 212 return make(EOP_LVar, 0, D); 213 } 214 215 unsigned makeThis() { 216 return make(EOP_This); 217 } 218 219 unsigned makeDot(const NamedDecl *D, bool Arrow) { 220 return make(EOP_Dot, Arrow ? 1 : 0, D); 221 } 222 223 unsigned makeCall(unsigned NumArgs, const NamedDecl *D) { 224 return make(EOP_Call, NumArgs, D); 225 } 226 227 // Grab the very first declaration of virtual method D 228 const CXXMethodDecl* getFirstVirtualDecl(const CXXMethodDecl *D) { 229 while (true) { 230 D = D->getCanonicalDecl(); 231 CXXMethodDecl::method_iterator I = D->begin_overridden_methods(), 232 E = D->end_overridden_methods(); 233 if (I == E) 234 return D; // Method does not override anything 235 D = *I; // FIXME: this does not work with multiple inheritance. 236 } 237 return 0; 238 } 239 240 unsigned makeMCall(unsigned NumArgs, const CXXMethodDecl *D) { 241 return make(EOP_MCall, NumArgs, getFirstVirtualDecl(D)); 242 } 243 244 unsigned makeIndex() { 245 return make(EOP_Index); 246 } 247 248 unsigned makeUnary() { 249 return make(EOP_Unary); 250 } 251 252 unsigned makeBinary() { 253 return make(EOP_Binary); 254 } 255 256 unsigned makeUnknown(unsigned Arity) { 257 return make(EOP_Unknown, Arity); 258 } 259 260 inline bool isCalleeArrow(const Expr *E) { 261 const MemberExpr *ME = dyn_cast<MemberExpr>(E->IgnoreParenCasts()); 262 return ME ? ME->isArrow() : false; 263 } 264 265 /// Build an SExpr from the given C++ expression. 266 /// Recursive function that terminates on DeclRefExpr. 267 /// Note: this function merely creates a SExpr; it does not check to 268 /// ensure that the original expression is a valid mutex expression. 269 /// 270 /// NDeref returns the number of Derefence and AddressOf operations 271 /// preceding the Expr; this is used to decide whether to pretty-print 272 /// SExprs with . or ->. 273 unsigned buildSExpr(const Expr *Exp, CallingContext* CallCtx, 274 int* NDeref = 0) { 275 if (!Exp) 276 return 0; 277 278 if (const DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(Exp)) { 279 const NamedDecl *ND = cast<NamedDecl>(DRE->getDecl()->getCanonicalDecl()); 280 const ParmVarDecl *PV = dyn_cast_or_null<ParmVarDecl>(ND); 281 if (PV) { 282 const FunctionDecl *FD = 283 cast<FunctionDecl>(PV->getDeclContext())->getCanonicalDecl(); 284 unsigned i = PV->getFunctionScopeIndex(); 285 286 if (CallCtx && CallCtx->FunArgs && 287 FD == CallCtx->AttrDecl->getCanonicalDecl()) { 288 // Substitute call arguments for references to function parameters 289 assert(i < CallCtx->NumArgs); 290 return buildSExpr(CallCtx->FunArgs[i], CallCtx->PrevCtx, NDeref); 291 } 292 // Map the param back to the param of the original function declaration. 293 makeNamedVar(FD->getParamDecl(i)); 294 return 1; 295 } 296 // Not a function parameter -- just store the reference. 297 makeNamedVar(ND); 298 return 1; 299 } else if (isa<CXXThisExpr>(Exp)) { 300 // Substitute parent for 'this' 301 if (CallCtx && CallCtx->SelfArg) { 302 if (!CallCtx->SelfArrow && NDeref) 303 // 'this' is a pointer, but self is not, so need to take address. 304 --(*NDeref); 305 return buildSExpr(CallCtx->SelfArg, CallCtx->PrevCtx, NDeref); 306 } 307 else { 308 makeThis(); 309 return 1; 310 } 311 } else if (const MemberExpr *ME = dyn_cast<MemberExpr>(Exp)) { 312 const NamedDecl *ND = ME->getMemberDecl(); 313 int ImplicitDeref = ME->isArrow() ? 1 : 0; 314 unsigned Root = makeDot(ND, false); 315 unsigned Sz = buildSExpr(ME->getBase(), CallCtx, &ImplicitDeref); 316 NodeVec[Root].setArrow(ImplicitDeref > 0); 317 NodeVec[Root].setSize(Sz + 1); 318 return Sz + 1; 319 } else if (const CXXMemberCallExpr *CMCE = dyn_cast<CXXMemberCallExpr>(Exp)) { 320 // When calling a function with a lock_returned attribute, replace 321 // the function call with the expression in lock_returned. 322 const CXXMethodDecl *MD = CMCE->getMethodDecl()->getMostRecentDecl(); 323 if (LockReturnedAttr* At = MD->getAttr<LockReturnedAttr>()) { 324 CallingContext LRCallCtx(CMCE->getMethodDecl()); 325 LRCallCtx.SelfArg = CMCE->getImplicitObjectArgument(); 326 LRCallCtx.SelfArrow = isCalleeArrow(CMCE->getCallee()); 327 LRCallCtx.NumArgs = CMCE->getNumArgs(); 328 LRCallCtx.FunArgs = CMCE->getArgs(); 329 LRCallCtx.PrevCtx = CallCtx; 330 return buildSExpr(At->getArg(), &LRCallCtx); 331 } 332 // Hack to treat smart pointers and iterators as pointers; 333 // ignore any method named get(). 334 if (CMCE->getMethodDecl()->getNameAsString() == "get" && 335 CMCE->getNumArgs() == 0) { 336 if (NDeref && isCalleeArrow(CMCE->getCallee())) 337 ++(*NDeref); 338 return buildSExpr(CMCE->getImplicitObjectArgument(), CallCtx, NDeref); 339 } 340 unsigned NumCallArgs = CMCE->getNumArgs(); 341 unsigned Root = makeMCall(NumCallArgs, CMCE->getMethodDecl()); 342 unsigned Sz = buildSExpr(CMCE->getImplicitObjectArgument(), CallCtx); 343 const Expr* const* CallArgs = CMCE->getArgs(); 344 for (unsigned i = 0; i < NumCallArgs; ++i) { 345 Sz += buildSExpr(CallArgs[i], CallCtx); 346 } 347 NodeVec[Root].setSize(Sz + 1); 348 return Sz + 1; 349 } else if (const CallExpr *CE = dyn_cast<CallExpr>(Exp)) { 350 const FunctionDecl *FD = CE->getDirectCallee()->getMostRecentDecl(); 351 if (LockReturnedAttr* At = FD->getAttr<LockReturnedAttr>()) { 352 CallingContext LRCallCtx(CE->getDirectCallee()); 353 LRCallCtx.NumArgs = CE->getNumArgs(); 354 LRCallCtx.FunArgs = CE->getArgs(); 355 LRCallCtx.PrevCtx = CallCtx; 356 return buildSExpr(At->getArg(), &LRCallCtx); 357 } 358 // Treat smart pointers and iterators as pointers; 359 // ignore the * and -> operators. 360 if (const CXXOperatorCallExpr *OE = dyn_cast<CXXOperatorCallExpr>(CE)) { 361 OverloadedOperatorKind k = OE->getOperator(); 362 if (k == OO_Star) { 363 if (NDeref) ++(*NDeref); 364 return buildSExpr(OE->getArg(0), CallCtx, NDeref); 365 } 366 else if (k == OO_Arrow) { 367 return buildSExpr(OE->getArg(0), CallCtx, NDeref); 368 } 369 } 370 unsigned NumCallArgs = CE->getNumArgs(); 371 unsigned Root = makeCall(NumCallArgs, 0); 372 unsigned Sz = buildSExpr(CE->getCallee(), CallCtx); 373 const Expr* const* CallArgs = CE->getArgs(); 374 for (unsigned i = 0; i < NumCallArgs; ++i) { 375 Sz += buildSExpr(CallArgs[i], CallCtx); 376 } 377 NodeVec[Root].setSize(Sz+1); 378 return Sz+1; 379 } else if (const BinaryOperator *BOE = dyn_cast<BinaryOperator>(Exp)) { 380 unsigned Root = makeBinary(); 381 unsigned Sz = buildSExpr(BOE->getLHS(), CallCtx); 382 Sz += buildSExpr(BOE->getRHS(), CallCtx); 383 NodeVec[Root].setSize(Sz); 384 return Sz; 385 } else if (const UnaryOperator *UOE = dyn_cast<UnaryOperator>(Exp)) { 386 // Ignore & and * operators -- they're no-ops. 387 // However, we try to figure out whether the expression is a pointer, 388 // so we can use . and -> appropriately in error messages. 389 if (UOE->getOpcode() == UO_Deref) { 390 if (NDeref) ++(*NDeref); 391 return buildSExpr(UOE->getSubExpr(), CallCtx, NDeref); 392 } 393 if (UOE->getOpcode() == UO_AddrOf) { 394 if (DeclRefExpr* DRE = dyn_cast<DeclRefExpr>(UOE->getSubExpr())) { 395 if (DRE->getDecl()->isCXXInstanceMember()) { 396 // This is a pointer-to-member expression, e.g. &MyClass::mu_. 397 // We interpret this syntax specially, as a wildcard. 398 unsigned Root = makeDot(DRE->getDecl(), false); 399 makeWildcard(); 400 NodeVec[Root].setSize(2); 401 return 2; 402 } 403 } 404 if (NDeref) --(*NDeref); 405 return buildSExpr(UOE->getSubExpr(), CallCtx, NDeref); 406 } 407 unsigned Root = makeUnary(); 408 unsigned Sz = buildSExpr(UOE->getSubExpr(), CallCtx); 409 NodeVec[Root].setSize(Sz); 410 return Sz; 411 } else if (const ArraySubscriptExpr *ASE = 412 dyn_cast<ArraySubscriptExpr>(Exp)) { 413 unsigned Root = makeIndex(); 414 unsigned Sz = buildSExpr(ASE->getBase(), CallCtx); 415 Sz += buildSExpr(ASE->getIdx(), CallCtx); 416 NodeVec[Root].setSize(Sz); 417 return Sz; 418 } else if (const AbstractConditionalOperator *CE = 419 dyn_cast<AbstractConditionalOperator>(Exp)) { 420 unsigned Root = makeUnknown(3); 421 unsigned Sz = buildSExpr(CE->getCond(), CallCtx); 422 Sz += buildSExpr(CE->getTrueExpr(), CallCtx); 423 Sz += buildSExpr(CE->getFalseExpr(), CallCtx); 424 NodeVec[Root].setSize(Sz); 425 return Sz; 426 } else if (const ChooseExpr *CE = dyn_cast<ChooseExpr>(Exp)) { 427 unsigned Root = makeUnknown(3); 428 unsigned Sz = buildSExpr(CE->getCond(), CallCtx); 429 Sz += buildSExpr(CE->getLHS(), CallCtx); 430 Sz += buildSExpr(CE->getRHS(), CallCtx); 431 NodeVec[Root].setSize(Sz); 432 return Sz; 433 } else if (const CastExpr *CE = dyn_cast<CastExpr>(Exp)) { 434 return buildSExpr(CE->getSubExpr(), CallCtx, NDeref); 435 } else if (const ParenExpr *PE = dyn_cast<ParenExpr>(Exp)) { 436 return buildSExpr(PE->getSubExpr(), CallCtx, NDeref); 437 } else if (const ExprWithCleanups *EWC = dyn_cast<ExprWithCleanups>(Exp)) { 438 return buildSExpr(EWC->getSubExpr(), CallCtx, NDeref); 439 } else if (const CXXBindTemporaryExpr *E = dyn_cast<CXXBindTemporaryExpr>(Exp)) { 440 return buildSExpr(E->getSubExpr(), CallCtx, NDeref); 441 } else if (isa<CharacterLiteral>(Exp) || 442 isa<CXXNullPtrLiteralExpr>(Exp) || 443 isa<GNUNullExpr>(Exp) || 444 isa<CXXBoolLiteralExpr>(Exp) || 445 isa<FloatingLiteral>(Exp) || 446 isa<ImaginaryLiteral>(Exp) || 447 isa<IntegerLiteral>(Exp) || 448 isa<StringLiteral>(Exp) || 449 isa<ObjCStringLiteral>(Exp)) { 450 makeNop(); 451 return 1; // FIXME: Ignore literals for now 452 } else { 453 makeNop(); 454 return 1; // Ignore. FIXME: mark as invalid expression? 455 } 456 } 457 458 /// \brief Construct a SExpr from an expression. 459 /// \param MutexExp The original mutex expression within an attribute 460 /// \param DeclExp An expression involving the Decl on which the attribute 461 /// occurs. 462 /// \param D The declaration to which the lock/unlock attribute is attached. 463 void buildSExprFromExpr(const Expr *MutexExp, const Expr *DeclExp, 464 const NamedDecl *D, VarDecl *SelfDecl = 0) { 465 CallingContext CallCtx(D); 466 467 if (MutexExp) { 468 if (const StringLiteral* SLit = dyn_cast<StringLiteral>(MutexExp)) { 469 if (SLit->getString() == StringRef("*")) 470 // The "*" expr is a universal lock, which essentially turns off 471 // checks until it is removed from the lockset. 472 makeUniversal(); 473 else 474 // Ignore other string literals for now. 475 makeNop(); 476 return; 477 } 478 } 479 480 // If we are processing a raw attribute expression, with no substitutions. 481 if (DeclExp == 0) { 482 buildSExpr(MutexExp, 0); 483 return; 484 } 485 486 // Examine DeclExp to find SelfArg and FunArgs, which are used to substitute 487 // for formal parameters when we call buildMutexID later. 488 if (const MemberExpr *ME = dyn_cast<MemberExpr>(DeclExp)) { 489 CallCtx.SelfArg = ME->getBase(); 490 CallCtx.SelfArrow = ME->isArrow(); 491 } else if (const CXXMemberCallExpr *CE = 492 dyn_cast<CXXMemberCallExpr>(DeclExp)) { 493 CallCtx.SelfArg = CE->getImplicitObjectArgument(); 494 CallCtx.SelfArrow = isCalleeArrow(CE->getCallee()); 495 CallCtx.NumArgs = CE->getNumArgs(); 496 CallCtx.FunArgs = CE->getArgs(); 497 } else if (const CallExpr *CE = dyn_cast<CallExpr>(DeclExp)) { 498 CallCtx.NumArgs = CE->getNumArgs(); 499 CallCtx.FunArgs = CE->getArgs(); 500 } else if (const CXXConstructExpr *CE = 501 dyn_cast<CXXConstructExpr>(DeclExp)) { 502 CallCtx.SelfArg = 0; // Will be set below 503 CallCtx.NumArgs = CE->getNumArgs(); 504 CallCtx.FunArgs = CE->getArgs(); 505 } else if (D && isa<CXXDestructorDecl>(D)) { 506 // There's no such thing as a "destructor call" in the AST. 507 CallCtx.SelfArg = DeclExp; 508 } 509 510 // Hack to handle constructors, where self cannot be recovered from 511 // the expression. 512 if (SelfDecl && !CallCtx.SelfArg) { 513 DeclRefExpr SelfDRE(SelfDecl, false, SelfDecl->getType(), VK_LValue, 514 SelfDecl->getLocation()); 515 CallCtx.SelfArg = &SelfDRE; 516 517 // If the attribute has no arguments, then assume the argument is "this". 518 if (MutexExp == 0) 519 buildSExpr(CallCtx.SelfArg, 0); 520 else // For most attributes. 521 buildSExpr(MutexExp, &CallCtx); 522 return; 523 } 524 525 // If the attribute has no arguments, then assume the argument is "this". 526 if (MutexExp == 0) 527 buildSExpr(CallCtx.SelfArg, 0); 528 else // For most attributes. 529 buildSExpr(MutexExp, &CallCtx); 530 } 531 532 /// \brief Get index of next sibling of node i. 533 unsigned getNextSibling(unsigned i) const { 534 return i + NodeVec[i].size(); 535 } 536 537 public: 538 explicit SExpr(clang::Decl::EmptyShell e) { NodeVec.clear(); } 539 540 /// \param MutexExp The original mutex expression within an attribute 541 /// \param DeclExp An expression involving the Decl on which the attribute 542 /// occurs. 543 /// \param D The declaration to which the lock/unlock attribute is attached. 544 /// Caller must check isValid() after construction. 545 SExpr(const Expr* MutexExp, const Expr *DeclExp, const NamedDecl* D, 546 VarDecl *SelfDecl=0) { 547 buildSExprFromExpr(MutexExp, DeclExp, D, SelfDecl); 548 } 549 550 /// Return true if this is a valid decl sequence. 551 /// Caller must call this by hand after construction to handle errors. 552 bool isValid() const { 553 return !NodeVec.empty(); 554 } 555 556 bool shouldIgnore() const { 557 // Nop is a mutex that we have decided to deliberately ignore. 558 assert(NodeVec.size() > 0 && "Invalid Mutex"); 559 return NodeVec[0].kind() == EOP_Nop; 560 } 561 562 bool isUniversal() const { 563 assert(NodeVec.size() > 0 && "Invalid Mutex"); 564 return NodeVec[0].kind() == EOP_Universal; 565 } 566 567 /// Issue a warning about an invalid lock expression 568 static void warnInvalidLock(ThreadSafetyHandler &Handler, 569 const Expr *MutexExp, const Expr *DeclExp, 570 const NamedDecl *D, StringRef Kind) { 571 SourceLocation Loc; 572 if (DeclExp) 573 Loc = DeclExp->getExprLoc(); 574 575 // FIXME: add a note about the attribute location in MutexExp or D 576 if (Loc.isValid()) 577 Handler.handleInvalidLockExp(Kind, Loc); 578 } 579 580 bool operator==(const SExpr &other) const { 581 return NodeVec == other.NodeVec; 582 } 583 584 bool operator!=(const SExpr &other) const { 585 return !(*this == other); 586 } 587 588 bool matches(const SExpr &Other, unsigned i = 0, unsigned j = 0) const { 589 if (NodeVec[i].matches(Other.NodeVec[j])) { 590 unsigned ni = NodeVec[i].arity(); 591 unsigned nj = Other.NodeVec[j].arity(); 592 unsigned n = (ni < nj) ? ni : nj; 593 bool Result = true; 594 unsigned ci = i+1; // first child of i 595 unsigned cj = j+1; // first child of j 596 for (unsigned k = 0; k < n; 597 ++k, ci=getNextSibling(ci), cj = Other.getNextSibling(cj)) { 598 Result = Result && matches(Other, ci, cj); 599 } 600 return Result; 601 } 602 return false; 603 } 604 605 // A partial match between a.mu and b.mu returns true a and b have the same 606 // type (and thus mu refers to the same mutex declaration), regardless of 607 // whether a and b are different objects or not. 608 bool partiallyMatches(const SExpr &Other) const { 609 if (NodeVec[0].kind() == EOP_Dot) 610 return NodeVec[0].matches(Other.NodeVec[0]); 611 return false; 612 } 613 614 /// \brief Pretty print a lock expression for use in error messages. 615 std::string toString(unsigned i = 0) const { 616 assert(isValid()); 617 if (i >= NodeVec.size()) 618 return ""; 619 620 const SExprNode* N = &NodeVec[i]; 621 switch (N->kind()) { 622 case EOP_Nop: 623 return "_"; 624 case EOP_Wildcard: 625 return "(?)"; 626 case EOP_Universal: 627 return "*"; 628 case EOP_This: 629 return "this"; 630 case EOP_NVar: 631 case EOP_LVar: { 632 return N->getNamedDecl()->getNameAsString(); 633 } 634 case EOP_Dot: { 635 if (NodeVec[i+1].kind() == EOP_Wildcard) { 636 std::string S = "&"; 637 S += N->getNamedDecl()->getQualifiedNameAsString(); 638 return S; 639 } 640 std::string FieldName = N->getNamedDecl()->getNameAsString(); 641 if (NodeVec[i+1].kind() == EOP_This) 642 return FieldName; 643 644 std::string S = toString(i+1); 645 if (N->isArrow()) 646 return S + "->" + FieldName; 647 else 648 return S + "." + FieldName; 649 } 650 case EOP_Call: { 651 std::string S = toString(i+1) + "("; 652 unsigned NumArgs = N->arity()-1; 653 unsigned ci = getNextSibling(i+1); 654 for (unsigned k=0; k<NumArgs; ++k, ci = getNextSibling(ci)) { 655 S += toString(ci); 656 if (k+1 < NumArgs) S += ","; 657 } 658 S += ")"; 659 return S; 660 } 661 case EOP_MCall: { 662 std::string S = ""; 663 if (NodeVec[i+1].kind() != EOP_This) 664 S = toString(i+1) + "."; 665 if (const NamedDecl *D = N->getFunctionDecl()) 666 S += D->getNameAsString() + "("; 667 else 668 S += "#("; 669 unsigned NumArgs = N->arity()-1; 670 unsigned ci = getNextSibling(i+1); 671 for (unsigned k=0; k<NumArgs; ++k, ci = getNextSibling(ci)) { 672 S += toString(ci); 673 if (k+1 < NumArgs) S += ","; 674 } 675 S += ")"; 676 return S; 677 } 678 case EOP_Index: { 679 std::string S1 = toString(i+1); 680 std::string S2 = toString(i+1 + NodeVec[i+1].size()); 681 return S1 + "[" + S2 + "]"; 682 } 683 case EOP_Unary: { 684 std::string S = toString(i+1); 685 return "#" + S; 686 } 687 case EOP_Binary: { 688 std::string S1 = toString(i+1); 689 std::string S2 = toString(i+1 + NodeVec[i+1].size()); 690 return "(" + S1 + "#" + S2 + ")"; 691 } 692 case EOP_Unknown: { 693 unsigned NumChildren = N->arity(); 694 if (NumChildren == 0) 695 return "(...)"; 696 std::string S = "("; 697 unsigned ci = i+1; 698 for (unsigned j = 0; j < NumChildren; ++j, ci = getNextSibling(ci)) { 699 S += toString(ci); 700 if (j+1 < NumChildren) S += "#"; 701 } 702 S += ")"; 703 return S; 704 } 705 } 706 return ""; 707 } 708 }; 709 710 /// \brief A short list of SExprs 711 class MutexIDList : public SmallVector<SExpr, 3> { 712 public: 713 /// \brief Push M onto list, but discard duplicates. 714 void push_back_nodup(const SExpr& M) { 715 if (end() == std::find(begin(), end(), M)) 716 push_back(M); 717 } 718 }; 719 720 /// \brief This is a helper class that stores info about the most recent 721 /// accquire of a Lock. 722 /// 723 /// The main body of the analysis maps MutexIDs to LockDatas. 724 struct LockData { 725 SourceLocation AcquireLoc; 726 727 /// \brief LKind stores whether a lock is held shared or exclusively. 728 /// Note that this analysis does not currently support either re-entrant 729 /// locking or lock "upgrading" and "downgrading" between exclusive and 730 /// shared. 731 /// 732 /// FIXME: add support for re-entrant locking and lock up/downgrading 733 LockKind LKind; 734 bool Asserted; // for asserted locks 735 bool Managed; // for ScopedLockable objects 736 SExpr UnderlyingMutex; // for ScopedLockable objects 737 738 LockData(SourceLocation AcquireLoc, LockKind LKind, bool M=false, 739 bool Asrt=false) 740 : AcquireLoc(AcquireLoc), LKind(LKind), Asserted(Asrt), Managed(M), 741 UnderlyingMutex(Decl::EmptyShell()) 742 {} 743 744 LockData(SourceLocation AcquireLoc, LockKind LKind, const SExpr &Mu) 745 : AcquireLoc(AcquireLoc), LKind(LKind), Asserted(false), Managed(false), 746 UnderlyingMutex(Mu) 747 {} 748 749 bool operator==(const LockData &other) const { 750 return AcquireLoc == other.AcquireLoc && LKind == other.LKind; 751 } 752 753 bool operator!=(const LockData &other) const { 754 return !(*this == other); 755 } 756 757 void Profile(llvm::FoldingSetNodeID &ID) const { 758 ID.AddInteger(AcquireLoc.getRawEncoding()); 759 ID.AddInteger(LKind); 760 } 761 762 bool isAtLeast(LockKind LK) { 763 return (LK == LK_Shared) || (LKind == LK_Exclusive); 764 } 765 }; 766 767 768 /// \brief A FactEntry stores a single fact that is known at a particular point 769 /// in the program execution. Currently, this is information regarding a lock 770 /// that is held at that point. 771 struct FactEntry { 772 SExpr MutID; 773 LockData LDat; 774 775 FactEntry(const SExpr& M, const LockData& L) 776 : MutID(M), LDat(L) 777 { } 778 }; 779 780 781 typedef unsigned short FactID; 782 783 /// \brief FactManager manages the memory for all facts that are created during 784 /// the analysis of a single routine. 785 class FactManager { 786 private: 787 std::vector<FactEntry> Facts; 788 789 public: 790 FactID newLock(const SExpr& M, const LockData& L) { 791 Facts.push_back(FactEntry(M,L)); 792 return static_cast<unsigned short>(Facts.size() - 1); 793 } 794 795 const FactEntry& operator[](FactID F) const { return Facts[F]; } 796 FactEntry& operator[](FactID F) { return Facts[F]; } 797 }; 798 799 800 /// \brief A FactSet is the set of facts that are known to be true at a 801 /// particular program point. FactSets must be small, because they are 802 /// frequently copied, and are thus implemented as a set of indices into a 803 /// table maintained by a FactManager. A typical FactSet only holds 1 or 2 804 /// locks, so we can get away with doing a linear search for lookup. Note 805 /// that a hashtable or map is inappropriate in this case, because lookups 806 /// may involve partial pattern matches, rather than exact matches. 807 class FactSet { 808 private: 809 typedef SmallVector<FactID, 4> FactVec; 810 811 FactVec FactIDs; 812 813 public: 814 typedef FactVec::iterator iterator; 815 typedef FactVec::const_iterator const_iterator; 816 817 iterator begin() { return FactIDs.begin(); } 818 const_iterator begin() const { return FactIDs.begin(); } 819 820 iterator end() { return FactIDs.end(); } 821 const_iterator end() const { return FactIDs.end(); } 822 823 bool isEmpty() const { return FactIDs.size() == 0; } 824 825 FactID addLock(FactManager& FM, const SExpr& M, const LockData& L) { 826 FactID F = FM.newLock(M, L); 827 FactIDs.push_back(F); 828 return F; 829 } 830 831 bool removeLock(FactManager& FM, const SExpr& M) { 832 unsigned n = FactIDs.size(); 833 if (n == 0) 834 return false; 835 836 for (unsigned i = 0; i < n-1; ++i) { 837 if (FM[FactIDs[i]].MutID.matches(M)) { 838 FactIDs[i] = FactIDs[n-1]; 839 FactIDs.pop_back(); 840 return true; 841 } 842 } 843 if (FM[FactIDs[n-1]].MutID.matches(M)) { 844 FactIDs.pop_back(); 845 return true; 846 } 847 return false; 848 } 849 850 // Returns an iterator 851 iterator findLockIter(FactManager &FM, const SExpr &M) { 852 for (iterator I = begin(), E = end(); I != E; ++I) { 853 const SExpr &Exp = FM[*I].MutID; 854 if (Exp.matches(M)) 855 return I; 856 } 857 return end(); 858 } 859 860 LockData* findLock(FactManager &FM, const SExpr &M) const { 861 for (const_iterator I = begin(), E = end(); I != E; ++I) { 862 const SExpr &Exp = FM[*I].MutID; 863 if (Exp.matches(M)) 864 return &FM[*I].LDat; 865 } 866 return 0; 867 } 868 869 LockData* findLockUniv(FactManager &FM, const SExpr &M) const { 870 for (const_iterator I = begin(), E = end(); I != E; ++I) { 871 const SExpr &Exp = FM[*I].MutID; 872 if (Exp.matches(M) || Exp.isUniversal()) 873 return &FM[*I].LDat; 874 } 875 return 0; 876 } 877 878 FactEntry* findPartialMatch(FactManager &FM, const SExpr &M) const { 879 for (const_iterator I=begin(), E=end(); I != E; ++I) { 880 const SExpr& Exp = FM[*I].MutID; 881 if (Exp.partiallyMatches(M)) return &FM[*I]; 882 } 883 return 0; 884 } 885 }; 886 887 888 889 /// A Lockset maps each SExpr (defined above) to information about how it has 890 /// been locked. 891 typedef llvm::ImmutableMap<SExpr, LockData> Lockset; 892 typedef llvm::ImmutableMap<const NamedDecl*, unsigned> LocalVarContext; 893 894 class LocalVariableMap; 895 896 /// A side (entry or exit) of a CFG node. 897 enum CFGBlockSide { CBS_Entry, CBS_Exit }; 898 899 /// CFGBlockInfo is a struct which contains all the information that is 900 /// maintained for each block in the CFG. See LocalVariableMap for more 901 /// information about the contexts. 902 struct CFGBlockInfo { 903 FactSet EntrySet; // Lockset held at entry to block 904 FactSet ExitSet; // Lockset held at exit from block 905 LocalVarContext EntryContext; // Context held at entry to block 906 LocalVarContext ExitContext; // Context held at exit from block 907 SourceLocation EntryLoc; // Location of first statement in block 908 SourceLocation ExitLoc; // Location of last statement in block. 909 unsigned EntryIndex; // Used to replay contexts later 910 bool Reachable; // Is this block reachable? 911 912 const FactSet &getSet(CFGBlockSide Side) const { 913 return Side == CBS_Entry ? EntrySet : ExitSet; 914 } 915 SourceLocation getLocation(CFGBlockSide Side) const { 916 return Side == CBS_Entry ? EntryLoc : ExitLoc; 917 } 918 919 private: 920 CFGBlockInfo(LocalVarContext EmptyCtx) 921 : EntryContext(EmptyCtx), ExitContext(EmptyCtx), Reachable(false) 922 { } 923 924 public: 925 static CFGBlockInfo getEmptyBlockInfo(LocalVariableMap &M); 926 }; 927 928 929 930 // A LocalVariableMap maintains a map from local variables to their currently 931 // valid definitions. It provides SSA-like functionality when traversing the 932 // CFG. Like SSA, each definition or assignment to a variable is assigned a 933 // unique name (an integer), which acts as the SSA name for that definition. 934 // The total set of names is shared among all CFG basic blocks. 935 // Unlike SSA, we do not rewrite expressions to replace local variables declrefs 936 // with their SSA-names. Instead, we compute a Context for each point in the 937 // code, which maps local variables to the appropriate SSA-name. This map 938 // changes with each assignment. 939 // 940 // The map is computed in a single pass over the CFG. Subsequent analyses can 941 // then query the map to find the appropriate Context for a statement, and use 942 // that Context to look up the definitions of variables. 943 class LocalVariableMap { 944 public: 945 typedef LocalVarContext Context; 946 947 /// A VarDefinition consists of an expression, representing the value of the 948 /// variable, along with the context in which that expression should be 949 /// interpreted. A reference VarDefinition does not itself contain this 950 /// information, but instead contains a pointer to a previous VarDefinition. 951 struct VarDefinition { 952 public: 953 friend class LocalVariableMap; 954 955 const NamedDecl *Dec; // The original declaration for this variable. 956 const Expr *Exp; // The expression for this variable, OR 957 unsigned Ref; // Reference to another VarDefinition 958 Context Ctx; // The map with which Exp should be interpreted. 959 960 bool isReference() { return !Exp; } 961 962 private: 963 // Create ordinary variable definition 964 VarDefinition(const NamedDecl *D, const Expr *E, Context C) 965 : Dec(D), Exp(E), Ref(0), Ctx(C) 966 { } 967 968 // Create reference to previous definition 969 VarDefinition(const NamedDecl *D, unsigned R, Context C) 970 : Dec(D), Exp(0), Ref(R), Ctx(C) 971 { } 972 }; 973 974 private: 975 Context::Factory ContextFactory; 976 std::vector<VarDefinition> VarDefinitions; 977 std::vector<unsigned> CtxIndices; 978 std::vector<std::pair<Stmt*, Context> > SavedContexts; 979 980 public: 981 LocalVariableMap() { 982 // index 0 is a placeholder for undefined variables (aka phi-nodes). 983 VarDefinitions.push_back(VarDefinition(0, 0u, getEmptyContext())); 984 } 985 986 /// Look up a definition, within the given context. 987 const VarDefinition* lookup(const NamedDecl *D, Context Ctx) { 988 const unsigned *i = Ctx.lookup(D); 989 if (!i) 990 return 0; 991 assert(*i < VarDefinitions.size()); 992 return &VarDefinitions[*i]; 993 } 994 995 /// Look up the definition for D within the given context. Returns 996 /// NULL if the expression is not statically known. If successful, also 997 /// modifies Ctx to hold the context of the return Expr. 998 const Expr* lookupExpr(const NamedDecl *D, Context &Ctx) { 999 const unsigned *P = Ctx.lookup(D); 1000 if (!P) 1001 return 0; 1002 1003 unsigned i = *P; 1004 while (i > 0) { 1005 if (VarDefinitions[i].Exp) { 1006 Ctx = VarDefinitions[i].Ctx; 1007 return VarDefinitions[i].Exp; 1008 } 1009 i = VarDefinitions[i].Ref; 1010 } 1011 return 0; 1012 } 1013 1014 Context getEmptyContext() { return ContextFactory.getEmptyMap(); } 1015 1016 /// Return the next context after processing S. This function is used by 1017 /// clients of the class to get the appropriate context when traversing the 1018 /// CFG. It must be called for every assignment or DeclStmt. 1019 Context getNextContext(unsigned &CtxIndex, Stmt *S, Context C) { 1020 if (SavedContexts[CtxIndex+1].first == S) { 1021 CtxIndex++; 1022 Context Result = SavedContexts[CtxIndex].second; 1023 return Result; 1024 } 1025 return C; 1026 } 1027 1028 void dumpVarDefinitionName(unsigned i) { 1029 if (i == 0) { 1030 llvm::errs() << "Undefined"; 1031 return; 1032 } 1033 const NamedDecl *Dec = VarDefinitions[i].Dec; 1034 if (!Dec) { 1035 llvm::errs() << "<<NULL>>"; 1036 return; 1037 } 1038 Dec->printName(llvm::errs()); 1039 llvm::errs() << "." << i << " " << ((const void*) Dec); 1040 } 1041 1042 /// Dumps an ASCII representation of the variable map to llvm::errs() 1043 void dump() { 1044 for (unsigned i = 1, e = VarDefinitions.size(); i < e; ++i) { 1045 const Expr *Exp = VarDefinitions[i].Exp; 1046 unsigned Ref = VarDefinitions[i].Ref; 1047 1048 dumpVarDefinitionName(i); 1049 llvm::errs() << " = "; 1050 if (Exp) Exp->dump(); 1051 else { 1052 dumpVarDefinitionName(Ref); 1053 llvm::errs() << "\n"; 1054 } 1055 } 1056 } 1057 1058 /// Dumps an ASCII representation of a Context to llvm::errs() 1059 void dumpContext(Context C) { 1060 for (Context::iterator I = C.begin(), E = C.end(); I != E; ++I) { 1061 const NamedDecl *D = I.getKey(); 1062 D->printName(llvm::errs()); 1063 const unsigned *i = C.lookup(D); 1064 llvm::errs() << " -> "; 1065 dumpVarDefinitionName(*i); 1066 llvm::errs() << "\n"; 1067 } 1068 } 1069 1070 /// Builds the variable map. 1071 void traverseCFG(CFG *CFGraph, PostOrderCFGView *SortedGraph, 1072 std::vector<CFGBlockInfo> &BlockInfo); 1073 1074 protected: 1075 // Get the current context index 1076 unsigned getContextIndex() { return SavedContexts.size()-1; } 1077 1078 // Save the current context for later replay 1079 void saveContext(Stmt *S, Context C) { 1080 SavedContexts.push_back(std::make_pair(S,C)); 1081 } 1082 1083 // Adds a new definition to the given context, and returns a new context. 1084 // This method should be called when declaring a new variable. 1085 Context addDefinition(const NamedDecl *D, Expr *Exp, Context Ctx) { 1086 assert(!Ctx.contains(D)); 1087 unsigned newID = VarDefinitions.size(); 1088 Context NewCtx = ContextFactory.add(Ctx, D, newID); 1089 VarDefinitions.push_back(VarDefinition(D, Exp, Ctx)); 1090 return NewCtx; 1091 } 1092 1093 // Add a new reference to an existing definition. 1094 Context addReference(const NamedDecl *D, unsigned i, Context Ctx) { 1095 unsigned newID = VarDefinitions.size(); 1096 Context NewCtx = ContextFactory.add(Ctx, D, newID); 1097 VarDefinitions.push_back(VarDefinition(D, i, Ctx)); 1098 return NewCtx; 1099 } 1100 1101 // Updates a definition only if that definition is already in the map. 1102 // This method should be called when assigning to an existing variable. 1103 Context updateDefinition(const NamedDecl *D, Expr *Exp, Context Ctx) { 1104 if (Ctx.contains(D)) { 1105 unsigned newID = VarDefinitions.size(); 1106 Context NewCtx = ContextFactory.remove(Ctx, D); 1107 NewCtx = ContextFactory.add(NewCtx, D, newID); 1108 VarDefinitions.push_back(VarDefinition(D, Exp, Ctx)); 1109 return NewCtx; 1110 } 1111 return Ctx; 1112 } 1113 1114 // Removes a definition from the context, but keeps the variable name 1115 // as a valid variable. The index 0 is a placeholder for cleared definitions. 1116 Context clearDefinition(const NamedDecl *D, Context Ctx) { 1117 Context NewCtx = Ctx; 1118 if (NewCtx.contains(D)) { 1119 NewCtx = ContextFactory.remove(NewCtx, D); 1120 NewCtx = ContextFactory.add(NewCtx, D, 0); 1121 } 1122 return NewCtx; 1123 } 1124 1125 // Remove a definition entirely frmo the context. 1126 Context removeDefinition(const NamedDecl *D, Context Ctx) { 1127 Context NewCtx = Ctx; 1128 if (NewCtx.contains(D)) { 1129 NewCtx = ContextFactory.remove(NewCtx, D); 1130 } 1131 return NewCtx; 1132 } 1133 1134 Context intersectContexts(Context C1, Context C2); 1135 Context createReferenceContext(Context C); 1136 void intersectBackEdge(Context C1, Context C2); 1137 1138 friend class VarMapBuilder; 1139 }; 1140 1141 1142 // This has to be defined after LocalVariableMap. 1143 CFGBlockInfo CFGBlockInfo::getEmptyBlockInfo(LocalVariableMap &M) { 1144 return CFGBlockInfo(M.getEmptyContext()); 1145 } 1146 1147 1148 /// Visitor which builds a LocalVariableMap 1149 class VarMapBuilder : public StmtVisitor<VarMapBuilder> { 1150 public: 1151 LocalVariableMap* VMap; 1152 LocalVariableMap::Context Ctx; 1153 1154 VarMapBuilder(LocalVariableMap *VM, LocalVariableMap::Context C) 1155 : VMap(VM), Ctx(C) {} 1156 1157 void VisitDeclStmt(DeclStmt *S); 1158 void VisitBinaryOperator(BinaryOperator *BO); 1159 }; 1160 1161 1162 // Add new local variables to the variable map 1163 void VarMapBuilder::VisitDeclStmt(DeclStmt *S) { 1164 bool modifiedCtx = false; 1165 DeclGroupRef DGrp = S->getDeclGroup(); 1166 for (DeclGroupRef::iterator I = DGrp.begin(), E = DGrp.end(); I != E; ++I) { 1167 if (VarDecl *VD = dyn_cast_or_null<VarDecl>(*I)) { 1168 Expr *E = VD->getInit(); 1169 1170 // Add local variables with trivial type to the variable map 1171 QualType T = VD->getType(); 1172 if (T.isTrivialType(VD->getASTContext())) { 1173 Ctx = VMap->addDefinition(VD, E, Ctx); 1174 modifiedCtx = true; 1175 } 1176 } 1177 } 1178 if (modifiedCtx) 1179 VMap->saveContext(S, Ctx); 1180 } 1181 1182 // Update local variable definitions in variable map 1183 void VarMapBuilder::VisitBinaryOperator(BinaryOperator *BO) { 1184 if (!BO->isAssignmentOp()) 1185 return; 1186 1187 Expr *LHSExp = BO->getLHS()->IgnoreParenCasts(); 1188 1189 // Update the variable map and current context. 1190 if (DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(LHSExp)) { 1191 ValueDecl *VDec = DRE->getDecl(); 1192 if (Ctx.lookup(VDec)) { 1193 if (BO->getOpcode() == BO_Assign) 1194 Ctx = VMap->updateDefinition(VDec, BO->getRHS(), Ctx); 1195 else 1196 // FIXME -- handle compound assignment operators 1197 Ctx = VMap->clearDefinition(VDec, Ctx); 1198 VMap->saveContext(BO, Ctx); 1199 } 1200 } 1201 } 1202 1203 1204 // Computes the intersection of two contexts. The intersection is the 1205 // set of variables which have the same definition in both contexts; 1206 // variables with different definitions are discarded. 1207 LocalVariableMap::Context 1208 LocalVariableMap::intersectContexts(Context C1, Context C2) { 1209 Context Result = C1; 1210 for (Context::iterator I = C1.begin(), E = C1.end(); I != E; ++I) { 1211 const NamedDecl *Dec = I.getKey(); 1212 unsigned i1 = I.getData(); 1213 const unsigned *i2 = C2.lookup(Dec); 1214 if (!i2) // variable doesn't exist on second path 1215 Result = removeDefinition(Dec, Result); 1216 else if (*i2 != i1) // variable exists, but has different definition 1217 Result = clearDefinition(Dec, Result); 1218 } 1219 return Result; 1220 } 1221 1222 // For every variable in C, create a new variable that refers to the 1223 // definition in C. Return a new context that contains these new variables. 1224 // (We use this for a naive implementation of SSA on loop back-edges.) 1225 LocalVariableMap::Context LocalVariableMap::createReferenceContext(Context C) { 1226 Context Result = getEmptyContext(); 1227 for (Context::iterator I = C.begin(), E = C.end(); I != E; ++I) { 1228 const NamedDecl *Dec = I.getKey(); 1229 unsigned i = I.getData(); 1230 Result = addReference(Dec, i, Result); 1231 } 1232 return Result; 1233 } 1234 1235 // This routine also takes the intersection of C1 and C2, but it does so by 1236 // altering the VarDefinitions. C1 must be the result of an earlier call to 1237 // createReferenceContext. 1238 void LocalVariableMap::intersectBackEdge(Context C1, Context C2) { 1239 for (Context::iterator I = C1.begin(), E = C1.end(); I != E; ++I) { 1240 const NamedDecl *Dec = I.getKey(); 1241 unsigned i1 = I.getData(); 1242 VarDefinition *VDef = &VarDefinitions[i1]; 1243 assert(VDef->isReference()); 1244 1245 const unsigned *i2 = C2.lookup(Dec); 1246 if (!i2 || (*i2 != i1)) 1247 VDef->Ref = 0; // Mark this variable as undefined 1248 } 1249 } 1250 1251 1252 // Traverse the CFG in topological order, so all predecessors of a block 1253 // (excluding back-edges) are visited before the block itself. At 1254 // each point in the code, we calculate a Context, which holds the set of 1255 // variable definitions which are visible at that point in execution. 1256 // Visible variables are mapped to their definitions using an array that 1257 // contains all definitions. 1258 // 1259 // At join points in the CFG, the set is computed as the intersection of 1260 // the incoming sets along each edge, E.g. 1261 // 1262 // { Context | VarDefinitions } 1263 // int x = 0; { x -> x1 | x1 = 0 } 1264 // int y = 0; { x -> x1, y -> y1 | y1 = 0, x1 = 0 } 1265 // if (b) x = 1; { x -> x2, y -> y1 | x2 = 1, y1 = 0, ... } 1266 // else x = 2; { x -> x3, y -> y1 | x3 = 2, x2 = 1, ... } 1267 // ... { y -> y1 (x is unknown) | x3 = 2, x2 = 1, ... } 1268 // 1269 // This is essentially a simpler and more naive version of the standard SSA 1270 // algorithm. Those definitions that remain in the intersection are from blocks 1271 // that strictly dominate the current block. We do not bother to insert proper 1272 // phi nodes, because they are not used in our analysis; instead, wherever 1273 // a phi node would be required, we simply remove that definition from the 1274 // context (E.g. x above). 1275 // 1276 // The initial traversal does not capture back-edges, so those need to be 1277 // handled on a separate pass. Whenever the first pass encounters an 1278 // incoming back edge, it duplicates the context, creating new definitions 1279 // that refer back to the originals. (These correspond to places where SSA 1280 // might have to insert a phi node.) On the second pass, these definitions are 1281 // set to NULL if the variable has changed on the back-edge (i.e. a phi 1282 // node was actually required.) E.g. 1283 // 1284 // { Context | VarDefinitions } 1285 // int x = 0, y = 0; { x -> x1, y -> y1 | y1 = 0, x1 = 0 } 1286 // while (b) { x -> x2, y -> y1 | [1st:] x2=x1; [2nd:] x2=NULL; } 1287 // x = x+1; { x -> x3, y -> y1 | x3 = x2 + 1, ... } 1288 // ... { y -> y1 | x3 = 2, x2 = 1, ... } 1289 // 1290 void LocalVariableMap::traverseCFG(CFG *CFGraph, 1291 PostOrderCFGView *SortedGraph, 1292 std::vector<CFGBlockInfo> &BlockInfo) { 1293 PostOrderCFGView::CFGBlockSet VisitedBlocks(CFGraph); 1294 1295 CtxIndices.resize(CFGraph->getNumBlockIDs()); 1296 1297 for (PostOrderCFGView::iterator I = SortedGraph->begin(), 1298 E = SortedGraph->end(); I!= E; ++I) { 1299 const CFGBlock *CurrBlock = *I; 1300 int CurrBlockID = CurrBlock->getBlockID(); 1301 CFGBlockInfo *CurrBlockInfo = &BlockInfo[CurrBlockID]; 1302 1303 VisitedBlocks.insert(CurrBlock); 1304 1305 // Calculate the entry context for the current block 1306 bool HasBackEdges = false; 1307 bool CtxInit = true; 1308 for (CFGBlock::const_pred_iterator PI = CurrBlock->pred_begin(), 1309 PE = CurrBlock->pred_end(); PI != PE; ++PI) { 1310 // if *PI -> CurrBlock is a back edge, so skip it 1311 if (*PI == 0 || !VisitedBlocks.alreadySet(*PI)) { 1312 HasBackEdges = true; 1313 continue; 1314 } 1315 1316 int PrevBlockID = (*PI)->getBlockID(); 1317 CFGBlockInfo *PrevBlockInfo = &BlockInfo[PrevBlockID]; 1318 1319 if (CtxInit) { 1320 CurrBlockInfo->EntryContext = PrevBlockInfo->ExitContext; 1321 CtxInit = false; 1322 } 1323 else { 1324 CurrBlockInfo->EntryContext = 1325 intersectContexts(CurrBlockInfo->EntryContext, 1326 PrevBlockInfo->ExitContext); 1327 } 1328 } 1329 1330 // Duplicate the context if we have back-edges, so we can call 1331 // intersectBackEdges later. 1332 if (HasBackEdges) 1333 CurrBlockInfo->EntryContext = 1334 createReferenceContext(CurrBlockInfo->EntryContext); 1335 1336 // Create a starting context index for the current block 1337 saveContext(0, CurrBlockInfo->EntryContext); 1338 CurrBlockInfo->EntryIndex = getContextIndex(); 1339 1340 // Visit all the statements in the basic block. 1341 VarMapBuilder VMapBuilder(this, CurrBlockInfo->EntryContext); 1342 for (CFGBlock::const_iterator BI = CurrBlock->begin(), 1343 BE = CurrBlock->end(); BI != BE; ++BI) { 1344 switch (BI->getKind()) { 1345 case CFGElement::Statement: { 1346 CFGStmt CS = BI->castAs<CFGStmt>(); 1347 VMapBuilder.Visit(const_cast<Stmt*>(CS.getStmt())); 1348 break; 1349 } 1350 default: 1351 break; 1352 } 1353 } 1354 CurrBlockInfo->ExitContext = VMapBuilder.Ctx; 1355 1356 // Mark variables on back edges as "unknown" if they've been changed. 1357 for (CFGBlock::const_succ_iterator SI = CurrBlock->succ_begin(), 1358 SE = CurrBlock->succ_end(); SI != SE; ++SI) { 1359 // if CurrBlock -> *SI is *not* a back edge 1360 if (*SI == 0 || !VisitedBlocks.alreadySet(*SI)) 1361 continue; 1362 1363 CFGBlock *FirstLoopBlock = *SI; 1364 Context LoopBegin = BlockInfo[FirstLoopBlock->getBlockID()].EntryContext; 1365 Context LoopEnd = CurrBlockInfo->ExitContext; 1366 intersectBackEdge(LoopBegin, LoopEnd); 1367 } 1368 } 1369 1370 // Put an extra entry at the end of the indexed context array 1371 unsigned exitID = CFGraph->getExit().getBlockID(); 1372 saveContext(0, BlockInfo[exitID].ExitContext); 1373 } 1374 1375 /// Find the appropriate source locations to use when producing diagnostics for 1376 /// each block in the CFG. 1377 static void findBlockLocations(CFG *CFGraph, 1378 PostOrderCFGView *SortedGraph, 1379 std::vector<CFGBlockInfo> &BlockInfo) { 1380 for (PostOrderCFGView::iterator I = SortedGraph->begin(), 1381 E = SortedGraph->end(); I!= E; ++I) { 1382 const CFGBlock *CurrBlock = *I; 1383 CFGBlockInfo *CurrBlockInfo = &BlockInfo[CurrBlock->getBlockID()]; 1384 1385 // Find the source location of the last statement in the block, if the 1386 // block is not empty. 1387 if (const Stmt *S = CurrBlock->getTerminator()) { 1388 CurrBlockInfo->EntryLoc = CurrBlockInfo->ExitLoc = S->getLocStart(); 1389 } else { 1390 for (CFGBlock::const_reverse_iterator BI = CurrBlock->rbegin(), 1391 BE = CurrBlock->rend(); BI != BE; ++BI) { 1392 // FIXME: Handle other CFGElement kinds. 1393 if (Optional<CFGStmt> CS = BI->getAs<CFGStmt>()) { 1394 CurrBlockInfo->ExitLoc = CS->getStmt()->getLocStart(); 1395 break; 1396 } 1397 } 1398 } 1399 1400 if (!CurrBlockInfo->ExitLoc.isInvalid()) { 1401 // This block contains at least one statement. Find the source location 1402 // of the first statement in the block. 1403 for (CFGBlock::const_iterator BI = CurrBlock->begin(), 1404 BE = CurrBlock->end(); BI != BE; ++BI) { 1405 // FIXME: Handle other CFGElement kinds. 1406 if (Optional<CFGStmt> CS = BI->getAs<CFGStmt>()) { 1407 CurrBlockInfo->EntryLoc = CS->getStmt()->getLocStart(); 1408 break; 1409 } 1410 } 1411 } else if (CurrBlock->pred_size() == 1 && *CurrBlock->pred_begin() && 1412 CurrBlock != &CFGraph->getExit()) { 1413 // The block is empty, and has a single predecessor. Use its exit 1414 // location. 1415 CurrBlockInfo->EntryLoc = CurrBlockInfo->ExitLoc = 1416 BlockInfo[(*CurrBlock->pred_begin())->getBlockID()].ExitLoc; 1417 } 1418 } 1419 } 1420 1421 /// \brief Class which implements the core thread safety analysis routines. 1422 class ThreadSafetyAnalyzer { 1423 friend class BuildLockset; 1424 1425 ThreadSafetyHandler &Handler; 1426 LocalVariableMap LocalVarMap; 1427 FactManager FactMan; 1428 std::vector<CFGBlockInfo> BlockInfo; 1429 1430 public: 1431 ThreadSafetyAnalyzer(ThreadSafetyHandler &H) : Handler(H) {} 1432 1433 void addLock(FactSet &FSet, const SExpr &Mutex, const LockData &LDat, 1434 StringRef DiagKind); 1435 void removeLock(FactSet &FSet, const SExpr &Mutex, SourceLocation UnlockLoc, 1436 bool FullyRemove, LockKind Kind, StringRef DiagKind); 1437 1438 template <typename AttrType> 1439 void getMutexIDs(MutexIDList &Mtxs, AttrType *Attr, Expr *Exp, 1440 const NamedDecl *D, VarDecl *SelfDecl=0); 1441 1442 template <class AttrType> 1443 void getMutexIDs(MutexIDList &Mtxs, AttrType *Attr, Expr *Exp, 1444 const NamedDecl *D, 1445 const CFGBlock *PredBlock, const CFGBlock *CurrBlock, 1446 Expr *BrE, bool Neg); 1447 1448 const CallExpr* getTrylockCallExpr(const Stmt *Cond, LocalVarContext C, 1449 bool &Negate); 1450 1451 void getEdgeLockset(FactSet &Result, const FactSet &ExitSet, 1452 const CFGBlock* PredBlock, 1453 const CFGBlock *CurrBlock); 1454 1455 void intersectAndWarn(FactSet &FSet1, const FactSet &FSet2, 1456 SourceLocation JoinLoc, 1457 LockErrorKind LEK1, LockErrorKind LEK2, 1458 bool Modify=true); 1459 1460 void intersectAndWarn(FactSet &FSet1, const FactSet &FSet2, 1461 SourceLocation JoinLoc, LockErrorKind LEK1, 1462 bool Modify=true) { 1463 intersectAndWarn(FSet1, FSet2, JoinLoc, LEK1, LEK1, Modify); 1464 } 1465 1466 void runAnalysis(AnalysisDeclContext &AC); 1467 }; 1468 1469 /// \brief Gets the value decl pointer from DeclRefExprs or MemberExprs. 1470 static const ValueDecl *getValueDecl(const Expr *Exp) { 1471 if (const auto *CE = dyn_cast<ImplicitCastExpr>(Exp)) 1472 return getValueDecl(CE->getSubExpr()); 1473 1474 if (const auto *DR = dyn_cast<DeclRefExpr>(Exp)) 1475 return DR->getDecl(); 1476 1477 if (const auto *ME = dyn_cast<MemberExpr>(Exp)) 1478 return ME->getMemberDecl(); 1479 1480 return nullptr; 1481 } 1482 1483 template <typename Ty> 1484 class has_arg_iterator { 1485 typedef char yes[1]; 1486 typedef char no[2]; 1487 1488 template <typename Inner> 1489 static yes& test(Inner *I, decltype(I->args_begin()) * = nullptr); 1490 1491 template <typename> 1492 static no& test(...); 1493 1494 public: 1495 static const bool value = sizeof(test<Ty>(nullptr)) == sizeof(yes); 1496 }; 1497 1498 static StringRef ClassifyDiagnostic(const CapabilityAttr *A) { 1499 return A->getName(); 1500 } 1501 1502 static StringRef ClassifyDiagnostic(QualType VDT) { 1503 // We need to look at the declaration of the type of the value to determine 1504 // which it is. The type should either be a record or a typedef, or a pointer 1505 // or reference thereof. 1506 if (const auto *RT = VDT->getAs<RecordType>()) { 1507 if (const auto *RD = RT->getDecl()) 1508 if (const auto *CA = RD->getAttr<CapabilityAttr>()) 1509 return ClassifyDiagnostic(CA); 1510 } else if (const auto *TT = VDT->getAs<TypedefType>()) { 1511 if (const auto *TD = TT->getDecl()) 1512 if (const auto *CA = TD->getAttr<CapabilityAttr>()) 1513 return ClassifyDiagnostic(CA); 1514 } else if (VDT->isPointerType() || VDT->isReferenceType()) 1515 return ClassifyDiagnostic(VDT->getPointeeType()); 1516 1517 return "mutex"; 1518 } 1519 1520 static StringRef ClassifyDiagnostic(const ValueDecl *VD) { 1521 assert(VD && "No ValueDecl passed"); 1522 1523 // The ValueDecl is the declaration of a mutex or role (hopefully). 1524 return ClassifyDiagnostic(VD->getType()); 1525 } 1526 1527 template <typename AttrTy> 1528 static typename std::enable_if<!has_arg_iterator<AttrTy>::value, 1529 StringRef>::type 1530 ClassifyDiagnostic(const AttrTy *A) { 1531 if (const ValueDecl *VD = getValueDecl(A->getArg())) 1532 return ClassifyDiagnostic(VD); 1533 return "mutex"; 1534 } 1535 1536 template <typename AttrTy> 1537 static typename std::enable_if<has_arg_iterator<AttrTy>::value, 1538 StringRef>::type 1539 ClassifyDiagnostic(const AttrTy *A) { 1540 for (auto I = A->args_begin(), E = A->args_end(); I != E; ++I) { 1541 if (const ValueDecl *VD = getValueDecl(*I)) 1542 return ClassifyDiagnostic(VD); 1543 } 1544 return "mutex"; 1545 } 1546 1547 /// \brief Add a new lock to the lockset, warning if the lock is already there. 1548 /// \param Mutex -- the Mutex expression for the lock 1549 /// \param LDat -- the LockData for the lock 1550 void ThreadSafetyAnalyzer::addLock(FactSet &FSet, const SExpr &Mutex, 1551 const LockData &LDat, StringRef DiagKind) { 1552 // FIXME: deal with acquired before/after annotations. 1553 // FIXME: Don't always warn when we have support for reentrant locks. 1554 if (Mutex.shouldIgnore()) 1555 return; 1556 1557 if (FSet.findLock(FactMan, Mutex)) { 1558 if (!LDat.Asserted) 1559 Handler.handleDoubleLock(DiagKind, Mutex.toString(), LDat.AcquireLoc); 1560 } else { 1561 FSet.addLock(FactMan, Mutex, LDat); 1562 } 1563 } 1564 1565 1566 /// \brief Remove a lock from the lockset, warning if the lock is not there. 1567 /// \param Mutex The lock expression corresponding to the lock to be removed 1568 /// \param UnlockLoc The source location of the unlock (only used in error msg) 1569 void ThreadSafetyAnalyzer::removeLock(FactSet &FSet, const SExpr &Mutex, 1570 SourceLocation UnlockLoc, 1571 bool FullyRemove, LockKind ReceivedKind, 1572 StringRef DiagKind) { 1573 if (Mutex.shouldIgnore()) 1574 return; 1575 1576 const LockData *LDat = FSet.findLock(FactMan, Mutex); 1577 if (!LDat) { 1578 Handler.handleUnmatchedUnlock(DiagKind, Mutex.toString(), UnlockLoc); 1579 return; 1580 } 1581 1582 // Generic lock removal doesn't care about lock kind mismatches, but 1583 // otherwise diagnose when the lock kinds are mismatched. 1584 if (ReceivedKind != LK_Generic && LDat->LKind != ReceivedKind) { 1585 Handler.handleIncorrectUnlockKind(DiagKind, Mutex.toString(), LDat->LKind, 1586 ReceivedKind, UnlockLoc); 1587 return; 1588 } 1589 1590 if (LDat->UnderlyingMutex.isValid()) { 1591 // This is scoped lockable object, which manages the real mutex. 1592 if (FullyRemove) { 1593 // We're destroying the managing object. 1594 // Remove the underlying mutex if it exists; but don't warn. 1595 if (FSet.findLock(FactMan, LDat->UnderlyingMutex)) 1596 FSet.removeLock(FactMan, LDat->UnderlyingMutex); 1597 } else { 1598 // We're releasing the underlying mutex, but not destroying the 1599 // managing object. Warn on dual release. 1600 if (!FSet.findLock(FactMan, LDat->UnderlyingMutex)) { 1601 Handler.handleUnmatchedUnlock( 1602 DiagKind, LDat->UnderlyingMutex.toString(), UnlockLoc); 1603 } 1604 FSet.removeLock(FactMan, LDat->UnderlyingMutex); 1605 return; 1606 } 1607 } 1608 FSet.removeLock(FactMan, Mutex); 1609 } 1610 1611 1612 /// \brief Extract the list of mutexIDs from the attribute on an expression, 1613 /// and push them onto Mtxs, discarding any duplicates. 1614 template <typename AttrType> 1615 void ThreadSafetyAnalyzer::getMutexIDs(MutexIDList &Mtxs, AttrType *Attr, 1616 Expr *Exp, const NamedDecl *D, 1617 VarDecl *SelfDecl) { 1618 typedef typename AttrType::args_iterator iterator_type; 1619 1620 if (Attr->args_size() == 0) { 1621 // The mutex held is the "this" object. 1622 SExpr Mu(0, Exp, D, SelfDecl); 1623 if (!Mu.isValid()) 1624 SExpr::warnInvalidLock(Handler, 0, Exp, D, ClassifyDiagnostic(Attr)); 1625 else 1626 Mtxs.push_back_nodup(Mu); 1627 return; 1628 } 1629 1630 for (iterator_type I=Attr->args_begin(), E=Attr->args_end(); I != E; ++I) { 1631 SExpr Mu(*I, Exp, D, SelfDecl); 1632 if (!Mu.isValid()) 1633 SExpr::warnInvalidLock(Handler, *I, Exp, D, ClassifyDiagnostic(Attr)); 1634 else 1635 Mtxs.push_back_nodup(Mu); 1636 } 1637 } 1638 1639 1640 /// \brief Extract the list of mutexIDs from a trylock attribute. If the 1641 /// trylock applies to the given edge, then push them onto Mtxs, discarding 1642 /// any duplicates. 1643 template <class AttrType> 1644 void ThreadSafetyAnalyzer::getMutexIDs(MutexIDList &Mtxs, AttrType *Attr, 1645 Expr *Exp, const NamedDecl *D, 1646 const CFGBlock *PredBlock, 1647 const CFGBlock *CurrBlock, 1648 Expr *BrE, bool Neg) { 1649 // Find out which branch has the lock 1650 bool branch = 0; 1651 if (CXXBoolLiteralExpr *BLE = dyn_cast_or_null<CXXBoolLiteralExpr>(BrE)) { 1652 branch = BLE->getValue(); 1653 } 1654 else if (IntegerLiteral *ILE = dyn_cast_or_null<IntegerLiteral>(BrE)) { 1655 branch = ILE->getValue().getBoolValue(); 1656 } 1657 int branchnum = branch ? 0 : 1; 1658 if (Neg) branchnum = !branchnum; 1659 1660 // If we've taken the trylock branch, then add the lock 1661 int i = 0; 1662 for (CFGBlock::const_succ_iterator SI = PredBlock->succ_begin(), 1663 SE = PredBlock->succ_end(); SI != SE && i < 2; ++SI, ++i) { 1664 if (*SI == CurrBlock && i == branchnum) { 1665 getMutexIDs(Mtxs, Attr, Exp, D); 1666 } 1667 } 1668 } 1669 1670 1671 bool getStaticBooleanValue(Expr* E, bool& TCond) { 1672 if (isa<CXXNullPtrLiteralExpr>(E) || isa<GNUNullExpr>(E)) { 1673 TCond = false; 1674 return true; 1675 } else if (CXXBoolLiteralExpr *BLE = dyn_cast<CXXBoolLiteralExpr>(E)) { 1676 TCond = BLE->getValue(); 1677 return true; 1678 } else if (IntegerLiteral *ILE = dyn_cast<IntegerLiteral>(E)) { 1679 TCond = ILE->getValue().getBoolValue(); 1680 return true; 1681 } else if (ImplicitCastExpr *CE = dyn_cast<ImplicitCastExpr>(E)) { 1682 return getStaticBooleanValue(CE->getSubExpr(), TCond); 1683 } 1684 return false; 1685 } 1686 1687 1688 // If Cond can be traced back to a function call, return the call expression. 1689 // The negate variable should be called with false, and will be set to true 1690 // if the function call is negated, e.g. if (!mu.tryLock(...)) 1691 const CallExpr* ThreadSafetyAnalyzer::getTrylockCallExpr(const Stmt *Cond, 1692 LocalVarContext C, 1693 bool &Negate) { 1694 if (!Cond) 1695 return 0; 1696 1697 if (const CallExpr *CallExp = dyn_cast<CallExpr>(Cond)) { 1698 return CallExp; 1699 } 1700 else if (const ParenExpr *PE = dyn_cast<ParenExpr>(Cond)) { 1701 return getTrylockCallExpr(PE->getSubExpr(), C, Negate); 1702 } 1703 else if (const ImplicitCastExpr *CE = dyn_cast<ImplicitCastExpr>(Cond)) { 1704 return getTrylockCallExpr(CE->getSubExpr(), C, Negate); 1705 } 1706 else if (const ExprWithCleanups* EWC = dyn_cast<ExprWithCleanups>(Cond)) { 1707 return getTrylockCallExpr(EWC->getSubExpr(), C, Negate); 1708 } 1709 else if (const DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(Cond)) { 1710 const Expr *E = LocalVarMap.lookupExpr(DRE->getDecl(), C); 1711 return getTrylockCallExpr(E, C, Negate); 1712 } 1713 else if (const UnaryOperator *UOP = dyn_cast<UnaryOperator>(Cond)) { 1714 if (UOP->getOpcode() == UO_LNot) { 1715 Negate = !Negate; 1716 return getTrylockCallExpr(UOP->getSubExpr(), C, Negate); 1717 } 1718 return 0; 1719 } 1720 else if (const BinaryOperator *BOP = dyn_cast<BinaryOperator>(Cond)) { 1721 if (BOP->getOpcode() == BO_EQ || BOP->getOpcode() == BO_NE) { 1722 if (BOP->getOpcode() == BO_NE) 1723 Negate = !Negate; 1724 1725 bool TCond = false; 1726 if (getStaticBooleanValue(BOP->getRHS(), TCond)) { 1727 if (!TCond) Negate = !Negate; 1728 return getTrylockCallExpr(BOP->getLHS(), C, Negate); 1729 } 1730 TCond = false; 1731 if (getStaticBooleanValue(BOP->getLHS(), TCond)) { 1732 if (!TCond) Negate = !Negate; 1733 return getTrylockCallExpr(BOP->getRHS(), C, Negate); 1734 } 1735 return 0; 1736 } 1737 if (BOP->getOpcode() == BO_LAnd) { 1738 // LHS must have been evaluated in a different block. 1739 return getTrylockCallExpr(BOP->getRHS(), C, Negate); 1740 } 1741 if (BOP->getOpcode() == BO_LOr) { 1742 return getTrylockCallExpr(BOP->getRHS(), C, Negate); 1743 } 1744 return 0; 1745 } 1746 return 0; 1747 } 1748 1749 1750 /// \brief Find the lockset that holds on the edge between PredBlock 1751 /// and CurrBlock. The edge set is the exit set of PredBlock (passed 1752 /// as the ExitSet parameter) plus any trylocks, which are conditionally held. 1753 void ThreadSafetyAnalyzer::getEdgeLockset(FactSet& Result, 1754 const FactSet &ExitSet, 1755 const CFGBlock *PredBlock, 1756 const CFGBlock *CurrBlock) { 1757 Result = ExitSet; 1758 1759 const Stmt *Cond = PredBlock->getTerminatorCondition(); 1760 if (!Cond) 1761 return; 1762 1763 bool Negate = false; 1764 const CFGBlockInfo *PredBlockInfo = &BlockInfo[PredBlock->getBlockID()]; 1765 const LocalVarContext &LVarCtx = PredBlockInfo->ExitContext; 1766 StringRef CapDiagKind = "mutex"; 1767 1768 CallExpr *Exp = 1769 const_cast<CallExpr*>(getTrylockCallExpr(Cond, LVarCtx, Negate)); 1770 if (!Exp) 1771 return; 1772 1773 NamedDecl *FunDecl = dyn_cast_or_null<NamedDecl>(Exp->getCalleeDecl()); 1774 if(!FunDecl || !FunDecl->hasAttrs()) 1775 return; 1776 1777 MutexIDList ExclusiveLocksToAdd; 1778 MutexIDList SharedLocksToAdd; 1779 1780 // If the condition is a call to a Trylock function, then grab the attributes 1781 AttrVec &ArgAttrs = FunDecl->getAttrs(); 1782 for (unsigned i = 0; i < ArgAttrs.size(); ++i) { 1783 Attr *Attr = ArgAttrs[i]; 1784 switch (Attr->getKind()) { 1785 case attr::ExclusiveTrylockFunction: { 1786 ExclusiveTrylockFunctionAttr *A = 1787 cast<ExclusiveTrylockFunctionAttr>(Attr); 1788 getMutexIDs(ExclusiveLocksToAdd, A, Exp, FunDecl, 1789 PredBlock, CurrBlock, A->getSuccessValue(), Negate); 1790 CapDiagKind = ClassifyDiagnostic(A); 1791 break; 1792 } 1793 case attr::SharedTrylockFunction: { 1794 SharedTrylockFunctionAttr *A = 1795 cast<SharedTrylockFunctionAttr>(Attr); 1796 getMutexIDs(SharedLocksToAdd, A, Exp, FunDecl, 1797 PredBlock, CurrBlock, A->getSuccessValue(), Negate); 1798 CapDiagKind = ClassifyDiagnostic(A); 1799 break; 1800 } 1801 default: 1802 break; 1803 } 1804 } 1805 1806 // Add and remove locks. 1807 SourceLocation Loc = Exp->getExprLoc(); 1808 for (const auto &ExclusiveLockToAdd : ExclusiveLocksToAdd) 1809 addLock(Result, ExclusiveLockToAdd, LockData(Loc, LK_Exclusive), 1810 CapDiagKind); 1811 for (const auto &SharedLockToAdd : SharedLocksToAdd) 1812 addLock(Result, SharedLockToAdd, LockData(Loc, LK_Shared), CapDiagKind); 1813 } 1814 1815 /// \brief We use this class to visit different types of expressions in 1816 /// CFGBlocks, and build up the lockset. 1817 /// An expression may cause us to add or remove locks from the lockset, or else 1818 /// output error messages related to missing locks. 1819 /// FIXME: In future, we may be able to not inherit from a visitor. 1820 class BuildLockset : public StmtVisitor<BuildLockset> { 1821 friend class ThreadSafetyAnalyzer; 1822 1823 ThreadSafetyAnalyzer *Analyzer; 1824 FactSet FSet; 1825 LocalVariableMap::Context LVarCtx; 1826 unsigned CtxIndex; 1827 1828 // Helper functions 1829 1830 void warnIfMutexNotHeld(const NamedDecl *D, const Expr *Exp, AccessKind AK, 1831 Expr *MutexExp, ProtectedOperationKind POK, 1832 StringRef DiagKind); 1833 void warnIfMutexHeld(const NamedDecl *D, const Expr *Exp, Expr *MutexExp, 1834 StringRef DiagKind); 1835 1836 void checkAccess(const Expr *Exp, AccessKind AK); 1837 void checkPtAccess(const Expr *Exp, AccessKind AK); 1838 1839 void handleCall(Expr *Exp, const NamedDecl *D, VarDecl *VD = 0); 1840 1841 public: 1842 BuildLockset(ThreadSafetyAnalyzer *Anlzr, CFGBlockInfo &Info) 1843 : StmtVisitor<BuildLockset>(), 1844 Analyzer(Anlzr), 1845 FSet(Info.EntrySet), 1846 LVarCtx(Info.EntryContext), 1847 CtxIndex(Info.EntryIndex) 1848 {} 1849 1850 void VisitUnaryOperator(UnaryOperator *UO); 1851 void VisitBinaryOperator(BinaryOperator *BO); 1852 void VisitCastExpr(CastExpr *CE); 1853 void VisitCallExpr(CallExpr *Exp); 1854 void VisitCXXConstructExpr(CXXConstructExpr *Exp); 1855 void VisitDeclStmt(DeclStmt *S); 1856 }; 1857 1858 /// \brief Warn if the LSet does not contain a lock sufficient to protect access 1859 /// of at least the passed in AccessKind. 1860 void BuildLockset::warnIfMutexNotHeld(const NamedDecl *D, const Expr *Exp, 1861 AccessKind AK, Expr *MutexExp, 1862 ProtectedOperationKind POK, 1863 StringRef DiagKind) { 1864 LockKind LK = getLockKindFromAccessKind(AK); 1865 1866 SExpr Mutex(MutexExp, Exp, D); 1867 if (!Mutex.isValid()) { 1868 SExpr::warnInvalidLock(Analyzer->Handler, MutexExp, Exp, D, DiagKind); 1869 return; 1870 } else if (Mutex.shouldIgnore()) { 1871 return; 1872 } 1873 1874 LockData* LDat = FSet.findLockUniv(Analyzer->FactMan, Mutex); 1875 bool NoError = true; 1876 if (!LDat) { 1877 // No exact match found. Look for a partial match. 1878 FactEntry* FEntry = FSet.findPartialMatch(Analyzer->FactMan, Mutex); 1879 if (FEntry) { 1880 // Warn that there's no precise match. 1881 LDat = &FEntry->LDat; 1882 std::string PartMatchStr = FEntry->MutID.toString(); 1883 StringRef PartMatchName(PartMatchStr); 1884 Analyzer->Handler.handleMutexNotHeld(DiagKind, D, POK, Mutex.toString(), 1885 LK, Exp->getExprLoc(), 1886 &PartMatchName); 1887 } else { 1888 // Warn that there's no match at all. 1889 Analyzer->Handler.handleMutexNotHeld(DiagKind, D, POK, Mutex.toString(), 1890 LK, Exp->getExprLoc()); 1891 } 1892 NoError = false; 1893 } 1894 // Make sure the mutex we found is the right kind. 1895 if (NoError && LDat && !LDat->isAtLeast(LK)) 1896 Analyzer->Handler.handleMutexNotHeld(DiagKind, D, POK, Mutex.toString(), LK, 1897 Exp->getExprLoc()); 1898 } 1899 1900 /// \brief Warn if the LSet contains the given lock. 1901 void BuildLockset::warnIfMutexHeld(const NamedDecl *D, const Expr *Exp, 1902 Expr *MutexExp, 1903 StringRef DiagKind) { 1904 SExpr Mutex(MutexExp, Exp, D); 1905 if (!Mutex.isValid()) { 1906 SExpr::warnInvalidLock(Analyzer->Handler, MutexExp, Exp, D, DiagKind); 1907 return; 1908 } 1909 1910 LockData* LDat = FSet.findLock(Analyzer->FactMan, Mutex); 1911 if (LDat) 1912 Analyzer->Handler.handleFunExcludesLock( 1913 DiagKind, D->getNameAsString(), Mutex.toString(), Exp->getExprLoc()); 1914 } 1915 1916 /// \brief Checks guarded_by and pt_guarded_by attributes. 1917 /// Whenever we identify an access (read or write) to a DeclRefExpr that is 1918 /// marked with guarded_by, we must ensure the appropriate mutexes are held. 1919 /// Similarly, we check if the access is to an expression that dereferences 1920 /// a pointer marked with pt_guarded_by. 1921 void BuildLockset::checkAccess(const Expr *Exp, AccessKind AK) { 1922 Exp = Exp->IgnoreParenCasts(); 1923 1924 if (const UnaryOperator *UO = dyn_cast<UnaryOperator>(Exp)) { 1925 // For dereferences 1926 if (UO->getOpcode() == clang::UO_Deref) 1927 checkPtAccess(UO->getSubExpr(), AK); 1928 return; 1929 } 1930 1931 if (const ArraySubscriptExpr *AE = dyn_cast<ArraySubscriptExpr>(Exp)) { 1932 checkPtAccess(AE->getLHS(), AK); 1933 return; 1934 } 1935 1936 if (const MemberExpr *ME = dyn_cast<MemberExpr>(Exp)) { 1937 if (ME->isArrow()) 1938 checkPtAccess(ME->getBase(), AK); 1939 else 1940 checkAccess(ME->getBase(), AK); 1941 } 1942 1943 const ValueDecl *D = getValueDecl(Exp); 1944 if (!D || !D->hasAttrs()) 1945 return; 1946 1947 if (D->hasAttr<GuardedVarAttr>() && FSet.isEmpty()) 1948 Analyzer->Handler.handleNoMutexHeld("mutex", D, POK_VarAccess, AK, 1949 Exp->getExprLoc()); 1950 1951 for (const auto *I : D->specific_attrs<GuardedByAttr>()) 1952 warnIfMutexNotHeld(D, Exp, AK, I->getArg(), POK_VarAccess, 1953 ClassifyDiagnostic(I)); 1954 } 1955 1956 /// \brief Checks pt_guarded_by and pt_guarded_var attributes. 1957 void BuildLockset::checkPtAccess(const Expr *Exp, AccessKind AK) { 1958 while (true) { 1959 if (const ParenExpr *PE = dyn_cast<ParenExpr>(Exp)) { 1960 Exp = PE->getSubExpr(); 1961 continue; 1962 } 1963 if (const CastExpr *CE = dyn_cast<CastExpr>(Exp)) { 1964 if (CE->getCastKind() == CK_ArrayToPointerDecay) { 1965 // If it's an actual array, and not a pointer, then it's elements 1966 // are protected by GUARDED_BY, not PT_GUARDED_BY; 1967 checkAccess(CE->getSubExpr(), AK); 1968 return; 1969 } 1970 Exp = CE->getSubExpr(); 1971 continue; 1972 } 1973 break; 1974 } 1975 1976 const ValueDecl *D = getValueDecl(Exp); 1977 if (!D || !D->hasAttrs()) 1978 return; 1979 1980 if (D->hasAttr<PtGuardedVarAttr>() && FSet.isEmpty()) 1981 Analyzer->Handler.handleNoMutexHeld("mutex", D, POK_VarDereference, AK, 1982 Exp->getExprLoc()); 1983 1984 for (auto const *I : D->specific_attrs<PtGuardedByAttr>()) 1985 warnIfMutexNotHeld(D, Exp, AK, I->getArg(), POK_VarDereference, 1986 ClassifyDiagnostic(I)); 1987 } 1988 1989 /// \brief Process a function call, method call, constructor call, 1990 /// or destructor call. This involves looking at the attributes on the 1991 /// corresponding function/method/constructor/destructor, issuing warnings, 1992 /// and updating the locksets accordingly. 1993 /// 1994 /// FIXME: For classes annotated with one of the guarded annotations, we need 1995 /// to treat const method calls as reads and non-const method calls as writes, 1996 /// and check that the appropriate locks are held. Non-const method calls with 1997 /// the same signature as const method calls can be also treated as reads. 1998 /// 1999 void BuildLockset::handleCall(Expr *Exp, const NamedDecl *D, VarDecl *VD) { 2000 SourceLocation Loc = Exp->getExprLoc(); 2001 const AttrVec &ArgAttrs = D->getAttrs(); 2002 MutexIDList ExclusiveLocksToAdd, SharedLocksToAdd; 2003 MutexIDList ExclusiveLocksToRemove, SharedLocksToRemove, GenericLocksToRemove; 2004 StringRef CapDiagKind = "mutex"; 2005 2006 for(unsigned i = 0; i < ArgAttrs.size(); ++i) { 2007 Attr *At = const_cast<Attr*>(ArgAttrs[i]); 2008 switch (At->getKind()) { 2009 // When we encounter a lock function, we need to add the lock to our 2010 // lockset. 2011 case attr::AcquireCapability: { 2012 auto *A = cast<AcquireCapabilityAttr>(At); 2013 Analyzer->getMutexIDs(A->isShared() ? SharedLocksToAdd 2014 : ExclusiveLocksToAdd, 2015 A, Exp, D, VD); 2016 2017 CapDiagKind = ClassifyDiagnostic(A); 2018 break; 2019 } 2020 2021 // An assert will add a lock to the lockset, but will not generate 2022 // a warning if it is already there, and will not generate a warning 2023 // if it is not removed. 2024 case attr::AssertExclusiveLock: { 2025 AssertExclusiveLockAttr *A = cast<AssertExclusiveLockAttr>(At); 2026 2027 MutexIDList AssertLocks; 2028 Analyzer->getMutexIDs(AssertLocks, A, Exp, D, VD); 2029 for (const auto &AssertLock : AssertLocks) 2030 Analyzer->addLock(FSet, AssertLock, 2031 LockData(Loc, LK_Exclusive, false, true), 2032 ClassifyDiagnostic(A)); 2033 break; 2034 } 2035 case attr::AssertSharedLock: { 2036 AssertSharedLockAttr *A = cast<AssertSharedLockAttr>(At); 2037 2038 MutexIDList AssertLocks; 2039 Analyzer->getMutexIDs(AssertLocks, A, Exp, D, VD); 2040 for (const auto &AssertLock : AssertLocks) 2041 Analyzer->addLock(FSet, AssertLock, 2042 LockData(Loc, LK_Shared, false, true), 2043 ClassifyDiagnostic(A)); 2044 break; 2045 } 2046 2047 // When we encounter an unlock function, we need to remove unlocked 2048 // mutexes from the lockset, and flag a warning if they are not there. 2049 case attr::ReleaseCapability: { 2050 auto *A = cast<ReleaseCapabilityAttr>(At); 2051 if (A->isGeneric()) 2052 Analyzer->getMutexIDs(GenericLocksToRemove, A, Exp, D, VD); 2053 else if (A->isShared()) 2054 Analyzer->getMutexIDs(SharedLocksToRemove, A, Exp, D, VD); 2055 else 2056 Analyzer->getMutexIDs(ExclusiveLocksToRemove, A, Exp, D, VD); 2057 2058 CapDiagKind = ClassifyDiagnostic(A); 2059 break; 2060 } 2061 2062 case attr::RequiresCapability: { 2063 RequiresCapabilityAttr *A = cast<RequiresCapabilityAttr>(At); 2064 2065 for (RequiresCapabilityAttr::args_iterator I = A->args_begin(), 2066 E = A->args_end(); I != E; ++I) 2067 warnIfMutexNotHeld(D, Exp, A->isShared() ? AK_Read : AK_Written, *I, 2068 POK_FunctionCall, ClassifyDiagnostic(A)); 2069 break; 2070 } 2071 2072 case attr::LocksExcluded: { 2073 LocksExcludedAttr *A = cast<LocksExcludedAttr>(At); 2074 2075 for (LocksExcludedAttr::args_iterator I = A->args_begin(), 2076 E = A->args_end(); I != E; ++I) { 2077 warnIfMutexHeld(D, Exp, *I, ClassifyDiagnostic(A)); 2078 } 2079 break; 2080 } 2081 2082 // Ignore attributes unrelated to thread-safety 2083 default: 2084 break; 2085 } 2086 } 2087 2088 // Figure out if we're calling the constructor of scoped lockable class 2089 bool isScopedVar = false; 2090 if (VD) { 2091 if (const CXXConstructorDecl *CD = dyn_cast<const CXXConstructorDecl>(D)) { 2092 const CXXRecordDecl* PD = CD->getParent(); 2093 if (PD && PD->hasAttr<ScopedLockableAttr>()) 2094 isScopedVar = true; 2095 } 2096 } 2097 2098 // Add locks. 2099 for (const auto &M : ExclusiveLocksToAdd) 2100 Analyzer->addLock(FSet, M, LockData(Loc, LK_Exclusive, isScopedVar), 2101 CapDiagKind); 2102 for (const auto &M : SharedLocksToAdd) 2103 Analyzer->addLock(FSet, M, LockData(Loc, LK_Shared, isScopedVar), 2104 CapDiagKind); 2105 2106 // Add the managing object as a dummy mutex, mapped to the underlying mutex. 2107 // FIXME -- this doesn't work if we acquire multiple locks. 2108 if (isScopedVar) { 2109 SourceLocation MLoc = VD->getLocation(); 2110 DeclRefExpr DRE(VD, false, VD->getType(), VK_LValue, VD->getLocation()); 2111 SExpr SMutex(&DRE, 0, 0); 2112 2113 for (const auto &M : ExclusiveLocksToAdd) 2114 Analyzer->addLock(FSet, SMutex, LockData(MLoc, LK_Exclusive, M), 2115 CapDiagKind); 2116 for (const auto &M : SharedLocksToAdd) 2117 Analyzer->addLock(FSet, SMutex, LockData(MLoc, LK_Shared, M), 2118 CapDiagKind); 2119 } 2120 2121 // Remove locks. 2122 // FIXME -- should only fully remove if the attribute refers to 'this'. 2123 bool Dtor = isa<CXXDestructorDecl>(D); 2124 for (const auto &M : ExclusiveLocksToRemove) 2125 Analyzer->removeLock(FSet, M, Loc, Dtor, LK_Exclusive, CapDiagKind); 2126 for (const auto &M : SharedLocksToRemove) 2127 Analyzer->removeLock(FSet, M, Loc, Dtor, LK_Shared, CapDiagKind); 2128 for (const auto &M : GenericLocksToRemove) 2129 Analyzer->removeLock(FSet, M, Loc, Dtor, LK_Generic, CapDiagKind); 2130 } 2131 2132 2133 /// \brief For unary operations which read and write a variable, we need to 2134 /// check whether we hold any required mutexes. Reads are checked in 2135 /// VisitCastExpr. 2136 void BuildLockset::VisitUnaryOperator(UnaryOperator *UO) { 2137 switch (UO->getOpcode()) { 2138 case clang::UO_PostDec: 2139 case clang::UO_PostInc: 2140 case clang::UO_PreDec: 2141 case clang::UO_PreInc: { 2142 checkAccess(UO->getSubExpr(), AK_Written); 2143 break; 2144 } 2145 default: 2146 break; 2147 } 2148 } 2149 2150 /// For binary operations which assign to a variable (writes), we need to check 2151 /// whether we hold any required mutexes. 2152 /// FIXME: Deal with non-primitive types. 2153 void BuildLockset::VisitBinaryOperator(BinaryOperator *BO) { 2154 if (!BO->isAssignmentOp()) 2155 return; 2156 2157 // adjust the context 2158 LVarCtx = Analyzer->LocalVarMap.getNextContext(CtxIndex, BO, LVarCtx); 2159 2160 checkAccess(BO->getLHS(), AK_Written); 2161 } 2162 2163 2164 /// Whenever we do an LValue to Rvalue cast, we are reading a variable and 2165 /// need to ensure we hold any required mutexes. 2166 /// FIXME: Deal with non-primitive types. 2167 void BuildLockset::VisitCastExpr(CastExpr *CE) { 2168 if (CE->getCastKind() != CK_LValueToRValue) 2169 return; 2170 checkAccess(CE->getSubExpr(), AK_Read); 2171 } 2172 2173 2174 void BuildLockset::VisitCallExpr(CallExpr *Exp) { 2175 if (CXXMemberCallExpr *CE = dyn_cast<CXXMemberCallExpr>(Exp)) { 2176 MemberExpr *ME = dyn_cast<MemberExpr>(CE->getCallee()); 2177 // ME can be null when calling a method pointer 2178 CXXMethodDecl *MD = CE->getMethodDecl(); 2179 2180 if (ME && MD) { 2181 if (ME->isArrow()) { 2182 if (MD->isConst()) { 2183 checkPtAccess(CE->getImplicitObjectArgument(), AK_Read); 2184 } else { // FIXME -- should be AK_Written 2185 checkPtAccess(CE->getImplicitObjectArgument(), AK_Read); 2186 } 2187 } else { 2188 if (MD->isConst()) 2189 checkAccess(CE->getImplicitObjectArgument(), AK_Read); 2190 else // FIXME -- should be AK_Written 2191 checkAccess(CE->getImplicitObjectArgument(), AK_Read); 2192 } 2193 } 2194 } else if (CXXOperatorCallExpr *OE = dyn_cast<CXXOperatorCallExpr>(Exp)) { 2195 switch (OE->getOperator()) { 2196 case OO_Equal: { 2197 const Expr *Target = OE->getArg(0); 2198 const Expr *Source = OE->getArg(1); 2199 checkAccess(Target, AK_Written); 2200 checkAccess(Source, AK_Read); 2201 break; 2202 } 2203 case OO_Star: 2204 case OO_Arrow: 2205 case OO_Subscript: { 2206 const Expr *Obj = OE->getArg(0); 2207 checkAccess(Obj, AK_Read); 2208 checkPtAccess(Obj, AK_Read); 2209 break; 2210 } 2211 default: { 2212 const Expr *Obj = OE->getArg(0); 2213 checkAccess(Obj, AK_Read); 2214 break; 2215 } 2216 } 2217 } 2218 NamedDecl *D = dyn_cast_or_null<NamedDecl>(Exp->getCalleeDecl()); 2219 if(!D || !D->hasAttrs()) 2220 return; 2221 handleCall(Exp, D); 2222 } 2223 2224 void BuildLockset::VisitCXXConstructExpr(CXXConstructExpr *Exp) { 2225 const CXXConstructorDecl *D = Exp->getConstructor(); 2226 if (D && D->isCopyConstructor()) { 2227 const Expr* Source = Exp->getArg(0); 2228 checkAccess(Source, AK_Read); 2229 } 2230 // FIXME -- only handles constructors in DeclStmt below. 2231 } 2232 2233 void BuildLockset::VisitDeclStmt(DeclStmt *S) { 2234 // adjust the context 2235 LVarCtx = Analyzer->LocalVarMap.getNextContext(CtxIndex, S, LVarCtx); 2236 2237 DeclGroupRef DGrp = S->getDeclGroup(); 2238 for (DeclGroupRef::iterator I = DGrp.begin(), E = DGrp.end(); I != E; ++I) { 2239 Decl *D = *I; 2240 if (VarDecl *VD = dyn_cast_or_null<VarDecl>(D)) { 2241 Expr *E = VD->getInit(); 2242 // handle constructors that involve temporaries 2243 if (ExprWithCleanups *EWC = dyn_cast_or_null<ExprWithCleanups>(E)) 2244 E = EWC->getSubExpr(); 2245 2246 if (CXXConstructExpr *CE = dyn_cast_or_null<CXXConstructExpr>(E)) { 2247 NamedDecl *CtorD = dyn_cast_or_null<NamedDecl>(CE->getConstructor()); 2248 if (!CtorD || !CtorD->hasAttrs()) 2249 return; 2250 handleCall(CE, CtorD, VD); 2251 } 2252 } 2253 } 2254 } 2255 2256 2257 2258 /// \brief Compute the intersection of two locksets and issue warnings for any 2259 /// locks in the symmetric difference. 2260 /// 2261 /// This function is used at a merge point in the CFG when comparing the lockset 2262 /// of each branch being merged. For example, given the following sequence: 2263 /// A; if () then B; else C; D; we need to check that the lockset after B and C 2264 /// are the same. In the event of a difference, we use the intersection of these 2265 /// two locksets at the start of D. 2266 /// 2267 /// \param FSet1 The first lockset. 2268 /// \param FSet2 The second lockset. 2269 /// \param JoinLoc The location of the join point for error reporting 2270 /// \param LEK1 The error message to report if a mutex is missing from LSet1 2271 /// \param LEK2 The error message to report if a mutex is missing from Lset2 2272 void ThreadSafetyAnalyzer::intersectAndWarn(FactSet &FSet1, 2273 const FactSet &FSet2, 2274 SourceLocation JoinLoc, 2275 LockErrorKind LEK1, 2276 LockErrorKind LEK2, 2277 bool Modify) { 2278 FactSet FSet1Orig = FSet1; 2279 2280 // Find locks in FSet2 that conflict or are not in FSet1, and warn. 2281 for (FactSet::const_iterator I = FSet2.begin(), E = FSet2.end(); 2282 I != E; ++I) { 2283 const SExpr &FSet2Mutex = FactMan[*I].MutID; 2284 const LockData &LDat2 = FactMan[*I].LDat; 2285 FactSet::iterator I1 = FSet1.findLockIter(FactMan, FSet2Mutex); 2286 2287 if (I1 != FSet1.end()) { 2288 const LockData* LDat1 = &FactMan[*I1].LDat; 2289 if (LDat1->LKind != LDat2.LKind) { 2290 Handler.handleExclusiveAndShared("mutex", FSet2Mutex.toString(), 2291 LDat2.AcquireLoc, LDat1->AcquireLoc); 2292 if (Modify && LDat1->LKind != LK_Exclusive) { 2293 // Take the exclusive lock, which is the one in FSet2. 2294 *I1 = *I; 2295 } 2296 } 2297 else if (LDat1->Asserted && !LDat2.Asserted) { 2298 // The non-asserted lock in FSet2 is the one we want to track. 2299 *I1 = *I; 2300 } 2301 } else { 2302 if (LDat2.UnderlyingMutex.isValid()) { 2303 if (FSet2.findLock(FactMan, LDat2.UnderlyingMutex)) { 2304 // If this is a scoped lock that manages another mutex, and if the 2305 // underlying mutex is still held, then warn about the underlying 2306 // mutex. 2307 Handler.handleMutexHeldEndOfScope("mutex", 2308 LDat2.UnderlyingMutex.toString(), 2309 LDat2.AcquireLoc, JoinLoc, LEK1); 2310 } 2311 } 2312 else if (!LDat2.Managed && !FSet2Mutex.isUniversal() && !LDat2.Asserted) 2313 Handler.handleMutexHeldEndOfScope("mutex", FSet2Mutex.toString(), 2314 LDat2.AcquireLoc, JoinLoc, LEK1); 2315 } 2316 } 2317 2318 // Find locks in FSet1 that are not in FSet2, and remove them. 2319 for (FactSet::const_iterator I = FSet1Orig.begin(), E = FSet1Orig.end(); 2320 I != E; ++I) { 2321 const SExpr &FSet1Mutex = FactMan[*I].MutID; 2322 const LockData &LDat1 = FactMan[*I].LDat; 2323 2324 if (!FSet2.findLock(FactMan, FSet1Mutex)) { 2325 if (LDat1.UnderlyingMutex.isValid()) { 2326 if (FSet1Orig.findLock(FactMan, LDat1.UnderlyingMutex)) { 2327 // If this is a scoped lock that manages another mutex, and if the 2328 // underlying mutex is still held, then warn about the underlying 2329 // mutex. 2330 Handler.handleMutexHeldEndOfScope("mutex", 2331 LDat1.UnderlyingMutex.toString(), 2332 LDat1.AcquireLoc, JoinLoc, LEK1); 2333 } 2334 } 2335 else if (!LDat1.Managed && !FSet1Mutex.isUniversal() && !LDat1.Asserted) 2336 Handler.handleMutexHeldEndOfScope("mutex", FSet1Mutex.toString(), 2337 LDat1.AcquireLoc, JoinLoc, LEK2); 2338 if (Modify) 2339 FSet1.removeLock(FactMan, FSet1Mutex); 2340 } 2341 } 2342 } 2343 2344 2345 // Return true if block B never continues to its successors. 2346 inline bool neverReturns(const CFGBlock* B) { 2347 if (B->hasNoReturnElement()) 2348 return true; 2349 if (B->empty()) 2350 return false; 2351 2352 CFGElement Last = B->back(); 2353 if (Optional<CFGStmt> S = Last.getAs<CFGStmt>()) { 2354 if (isa<CXXThrowExpr>(S->getStmt())) 2355 return true; 2356 } 2357 return false; 2358 } 2359 2360 2361 /// \brief Check a function's CFG for thread-safety violations. 2362 /// 2363 /// We traverse the blocks in the CFG, compute the set of mutexes that are held 2364 /// at the end of each block, and issue warnings for thread safety violations. 2365 /// Each block in the CFG is traversed exactly once. 2366 void ThreadSafetyAnalyzer::runAnalysis(AnalysisDeclContext &AC) { 2367 // TODO: this whole function needs be rewritten as a visitor for CFGWalker. 2368 // For now, we just use the walker to set things up. 2369 threadSafety::CFGWalker walker; 2370 if (!walker.init(AC)) 2371 return; 2372 2373 // AC.dumpCFG(true); 2374 // threadSafety::printSCFG(walker); 2375 2376 CFG *CFGraph = walker.CFGraph; 2377 const NamedDecl *D = walker.FDecl; 2378 2379 if (D->hasAttr<NoThreadSafetyAnalysisAttr>()) 2380 return; 2381 2382 // FIXME: Do something a bit more intelligent inside constructor and 2383 // destructor code. Constructors and destructors must assume unique access 2384 // to 'this', so checks on member variable access is disabled, but we should 2385 // still enable checks on other objects. 2386 if (isa<CXXConstructorDecl>(D)) 2387 return; // Don't check inside constructors. 2388 if (isa<CXXDestructorDecl>(D)) 2389 return; // Don't check inside destructors. 2390 2391 BlockInfo.resize(CFGraph->getNumBlockIDs(), 2392 CFGBlockInfo::getEmptyBlockInfo(LocalVarMap)); 2393 2394 // We need to explore the CFG via a "topological" ordering. 2395 // That way, we will be guaranteed to have information about required 2396 // predecessor locksets when exploring a new block. 2397 PostOrderCFGView *SortedGraph = walker.SortedGraph; 2398 PostOrderCFGView::CFGBlockSet VisitedBlocks(CFGraph); 2399 2400 // Mark entry block as reachable 2401 BlockInfo[CFGraph->getEntry().getBlockID()].Reachable = true; 2402 2403 // Compute SSA names for local variables 2404 LocalVarMap.traverseCFG(CFGraph, SortedGraph, BlockInfo); 2405 2406 // Fill in source locations for all CFGBlocks. 2407 findBlockLocations(CFGraph, SortedGraph, BlockInfo); 2408 2409 MutexIDList ExclusiveLocksAcquired; 2410 MutexIDList SharedLocksAcquired; 2411 MutexIDList LocksReleased; 2412 2413 // Add locks from exclusive_locks_required and shared_locks_required 2414 // to initial lockset. Also turn off checking for lock and unlock functions. 2415 // FIXME: is there a more intelligent way to check lock/unlock functions? 2416 if (!SortedGraph->empty() && D->hasAttrs()) { 2417 const CFGBlock *FirstBlock = *SortedGraph->begin(); 2418 FactSet &InitialLockset = BlockInfo[FirstBlock->getBlockID()].EntrySet; 2419 const AttrVec &ArgAttrs = D->getAttrs(); 2420 2421 MutexIDList ExclusiveLocksToAdd; 2422 MutexIDList SharedLocksToAdd; 2423 StringRef CapDiagKind = "mutex"; 2424 2425 SourceLocation Loc = D->getLocation(); 2426 for (unsigned i = 0; i < ArgAttrs.size(); ++i) { 2427 Attr *Attr = ArgAttrs[i]; 2428 Loc = Attr->getLocation(); 2429 if (RequiresCapabilityAttr *A = dyn_cast<RequiresCapabilityAttr>(Attr)) { 2430 getMutexIDs(A->isShared() ? SharedLocksToAdd : ExclusiveLocksToAdd, A, 2431 0, D); 2432 CapDiagKind = ClassifyDiagnostic(A); 2433 } else if (auto *A = dyn_cast<ReleaseCapabilityAttr>(Attr)) { 2434 // UNLOCK_FUNCTION() is used to hide the underlying lock implementation. 2435 // We must ignore such methods. 2436 if (A->args_size() == 0) 2437 return; 2438 // FIXME -- deal with exclusive vs. shared unlock functions? 2439 getMutexIDs(ExclusiveLocksToAdd, A, (Expr*) 0, D); 2440 getMutexIDs(LocksReleased, A, (Expr*) 0, D); 2441 CapDiagKind = ClassifyDiagnostic(A); 2442 } else if (auto *A = dyn_cast<AcquireCapabilityAttr>(Attr)) { 2443 if (A->args_size() == 0) 2444 return; 2445 getMutexIDs(A->isShared() ? SharedLocksAcquired 2446 : ExclusiveLocksAcquired, 2447 A, nullptr, D); 2448 CapDiagKind = ClassifyDiagnostic(A); 2449 } else if (isa<ExclusiveTrylockFunctionAttr>(Attr)) { 2450 // Don't try to check trylock functions for now 2451 return; 2452 } else if (isa<SharedTrylockFunctionAttr>(Attr)) { 2453 // Don't try to check trylock functions for now 2454 return; 2455 } 2456 } 2457 2458 // FIXME -- Loc can be wrong here. 2459 for (const auto &ExclusiveLockToAdd : ExclusiveLocksToAdd) 2460 addLock(InitialLockset, ExclusiveLockToAdd, LockData(Loc, LK_Exclusive), 2461 CapDiagKind); 2462 for (const auto &SharedLockToAdd : SharedLocksToAdd) 2463 addLock(InitialLockset, SharedLockToAdd, LockData(Loc, LK_Shared), 2464 CapDiagKind); 2465 } 2466 2467 for (PostOrderCFGView::iterator I = SortedGraph->begin(), 2468 E = SortedGraph->end(); I!= E; ++I) { 2469 const CFGBlock *CurrBlock = *I; 2470 int CurrBlockID = CurrBlock->getBlockID(); 2471 CFGBlockInfo *CurrBlockInfo = &BlockInfo[CurrBlockID]; 2472 2473 // Use the default initial lockset in case there are no predecessors. 2474 VisitedBlocks.insert(CurrBlock); 2475 2476 // Iterate through the predecessor blocks and warn if the lockset for all 2477 // predecessors is not the same. We take the entry lockset of the current 2478 // block to be the intersection of all previous locksets. 2479 // FIXME: By keeping the intersection, we may output more errors in future 2480 // for a lock which is not in the intersection, but was in the union. We 2481 // may want to also keep the union in future. As an example, let's say 2482 // the intersection contains Mutex L, and the union contains L and M. 2483 // Later we unlock M. At this point, we would output an error because we 2484 // never locked M; although the real error is probably that we forgot to 2485 // lock M on all code paths. Conversely, let's say that later we lock M. 2486 // In this case, we should compare against the intersection instead of the 2487 // union because the real error is probably that we forgot to unlock M on 2488 // all code paths. 2489 bool LocksetInitialized = false; 2490 SmallVector<CFGBlock *, 8> SpecialBlocks; 2491 for (CFGBlock::const_pred_iterator PI = CurrBlock->pred_begin(), 2492 PE = CurrBlock->pred_end(); PI != PE; ++PI) { 2493 2494 // if *PI -> CurrBlock is a back edge 2495 if (*PI == 0 || !VisitedBlocks.alreadySet(*PI)) 2496 continue; 2497 2498 int PrevBlockID = (*PI)->getBlockID(); 2499 CFGBlockInfo *PrevBlockInfo = &BlockInfo[PrevBlockID]; 2500 2501 // Ignore edges from blocks that can't return. 2502 if (neverReturns(*PI) || !PrevBlockInfo->Reachable) 2503 continue; 2504 2505 // Okay, we can reach this block from the entry. 2506 CurrBlockInfo->Reachable = true; 2507 2508 // If the previous block ended in a 'continue' or 'break' statement, then 2509 // a difference in locksets is probably due to a bug in that block, rather 2510 // than in some other predecessor. In that case, keep the other 2511 // predecessor's lockset. 2512 if (const Stmt *Terminator = (*PI)->getTerminator()) { 2513 if (isa<ContinueStmt>(Terminator) || isa<BreakStmt>(Terminator)) { 2514 SpecialBlocks.push_back(*PI); 2515 continue; 2516 } 2517 } 2518 2519 FactSet PrevLockset; 2520 getEdgeLockset(PrevLockset, PrevBlockInfo->ExitSet, *PI, CurrBlock); 2521 2522 if (!LocksetInitialized) { 2523 CurrBlockInfo->EntrySet = PrevLockset; 2524 LocksetInitialized = true; 2525 } else { 2526 intersectAndWarn(CurrBlockInfo->EntrySet, PrevLockset, 2527 CurrBlockInfo->EntryLoc, 2528 LEK_LockedSomePredecessors); 2529 } 2530 } 2531 2532 // Skip rest of block if it's not reachable. 2533 if (!CurrBlockInfo->Reachable) 2534 continue; 2535 2536 // Process continue and break blocks. Assume that the lockset for the 2537 // resulting block is unaffected by any discrepancies in them. 2538 for (unsigned SpecialI = 0, SpecialN = SpecialBlocks.size(); 2539 SpecialI < SpecialN; ++SpecialI) { 2540 CFGBlock *PrevBlock = SpecialBlocks[SpecialI]; 2541 int PrevBlockID = PrevBlock->getBlockID(); 2542 CFGBlockInfo *PrevBlockInfo = &BlockInfo[PrevBlockID]; 2543 2544 if (!LocksetInitialized) { 2545 CurrBlockInfo->EntrySet = PrevBlockInfo->ExitSet; 2546 LocksetInitialized = true; 2547 } else { 2548 // Determine whether this edge is a loop terminator for diagnostic 2549 // purposes. FIXME: A 'break' statement might be a loop terminator, but 2550 // it might also be part of a switch. Also, a subsequent destructor 2551 // might add to the lockset, in which case the real issue might be a 2552 // double lock on the other path. 2553 const Stmt *Terminator = PrevBlock->getTerminator(); 2554 bool IsLoop = Terminator && isa<ContinueStmt>(Terminator); 2555 2556 FactSet PrevLockset; 2557 getEdgeLockset(PrevLockset, PrevBlockInfo->ExitSet, 2558 PrevBlock, CurrBlock); 2559 2560 // Do not update EntrySet. 2561 intersectAndWarn(CurrBlockInfo->EntrySet, PrevLockset, 2562 PrevBlockInfo->ExitLoc, 2563 IsLoop ? LEK_LockedSomeLoopIterations 2564 : LEK_LockedSomePredecessors, 2565 false); 2566 } 2567 } 2568 2569 BuildLockset LocksetBuilder(this, *CurrBlockInfo); 2570 2571 // Visit all the statements in the basic block. 2572 for (CFGBlock::const_iterator BI = CurrBlock->begin(), 2573 BE = CurrBlock->end(); BI != BE; ++BI) { 2574 switch (BI->getKind()) { 2575 case CFGElement::Statement: { 2576 CFGStmt CS = BI->castAs<CFGStmt>(); 2577 LocksetBuilder.Visit(const_cast<Stmt*>(CS.getStmt())); 2578 break; 2579 } 2580 // Ignore BaseDtor, MemberDtor, and TemporaryDtor for now. 2581 case CFGElement::AutomaticObjectDtor: { 2582 CFGAutomaticObjDtor AD = BI->castAs<CFGAutomaticObjDtor>(); 2583 CXXDestructorDecl *DD = const_cast<CXXDestructorDecl *>( 2584 AD.getDestructorDecl(AC.getASTContext())); 2585 if (!DD->hasAttrs()) 2586 break; 2587 2588 // Create a dummy expression, 2589 VarDecl *VD = const_cast<VarDecl*>(AD.getVarDecl()); 2590 DeclRefExpr DRE(VD, false, VD->getType(), VK_LValue, 2591 AD.getTriggerStmt()->getLocEnd()); 2592 LocksetBuilder.handleCall(&DRE, DD); 2593 break; 2594 } 2595 default: 2596 break; 2597 } 2598 } 2599 CurrBlockInfo->ExitSet = LocksetBuilder.FSet; 2600 2601 // For every back edge from CurrBlock (the end of the loop) to another block 2602 // (FirstLoopBlock) we need to check that the Lockset of Block is equal to 2603 // the one held at the beginning of FirstLoopBlock. We can look up the 2604 // Lockset held at the beginning of FirstLoopBlock in the EntryLockSets map. 2605 for (CFGBlock::const_succ_iterator SI = CurrBlock->succ_begin(), 2606 SE = CurrBlock->succ_end(); SI != SE; ++SI) { 2607 2608 // if CurrBlock -> *SI is *not* a back edge 2609 if (*SI == 0 || !VisitedBlocks.alreadySet(*SI)) 2610 continue; 2611 2612 CFGBlock *FirstLoopBlock = *SI; 2613 CFGBlockInfo *PreLoop = &BlockInfo[FirstLoopBlock->getBlockID()]; 2614 CFGBlockInfo *LoopEnd = &BlockInfo[CurrBlockID]; 2615 intersectAndWarn(LoopEnd->ExitSet, PreLoop->EntrySet, 2616 PreLoop->EntryLoc, 2617 LEK_LockedSomeLoopIterations, 2618 false); 2619 } 2620 } 2621 2622 CFGBlockInfo *Initial = &BlockInfo[CFGraph->getEntry().getBlockID()]; 2623 CFGBlockInfo *Final = &BlockInfo[CFGraph->getExit().getBlockID()]; 2624 2625 // Skip the final check if the exit block is unreachable. 2626 if (!Final->Reachable) 2627 return; 2628 2629 // By default, we expect all locks held on entry to be held on exit. 2630 FactSet ExpectedExitSet = Initial->EntrySet; 2631 2632 // Adjust the expected exit set by adding or removing locks, as declared 2633 // by *-LOCK_FUNCTION and UNLOCK_FUNCTION. The intersect below will then 2634 // issue the appropriate warning. 2635 // FIXME: the location here is not quite right. 2636 for (unsigned i=0,n=ExclusiveLocksAcquired.size(); i<n; ++i) { 2637 ExpectedExitSet.addLock(FactMan, ExclusiveLocksAcquired[i], 2638 LockData(D->getLocation(), LK_Exclusive)); 2639 } 2640 for (unsigned i=0,n=SharedLocksAcquired.size(); i<n; ++i) { 2641 ExpectedExitSet.addLock(FactMan, SharedLocksAcquired[i], 2642 LockData(D->getLocation(), LK_Shared)); 2643 } 2644 for (unsigned i=0,n=LocksReleased.size(); i<n; ++i) { 2645 ExpectedExitSet.removeLock(FactMan, LocksReleased[i]); 2646 } 2647 2648 // FIXME: Should we call this function for all blocks which exit the function? 2649 intersectAndWarn(ExpectedExitSet, Final->ExitSet, 2650 Final->ExitLoc, 2651 LEK_LockedAtEndOfFunction, 2652 LEK_NotLockedAtEndOfFunction, 2653 false); 2654 } 2655 2656 } // end anonymous namespace 2657 2658 2659 namespace clang { 2660 namespace thread_safety { 2661 2662 /// \brief Check a function's CFG for thread-safety violations. 2663 /// 2664 /// We traverse the blocks in the CFG, compute the set of mutexes that are held 2665 /// at the end of each block, and issue warnings for thread safety violations. 2666 /// Each block in the CFG is traversed exactly once. 2667 void runThreadSafetyAnalysis(AnalysisDeclContext &AC, 2668 ThreadSafetyHandler &Handler) { 2669 ThreadSafetyAnalyzer Analyzer(Handler); 2670 Analyzer.runAnalysis(AC); 2671 } 2672 2673 /// \brief Helper function that returns a LockKind required for the given level 2674 /// of access. 2675 LockKind getLockKindFromAccessKind(AccessKind AK) { 2676 switch (AK) { 2677 case AK_Read : 2678 return LK_Shared; 2679 case AK_Written : 2680 return LK_Exclusive; 2681 } 2682 llvm_unreachable("Unknown AccessKind"); 2683 } 2684 2685 }} // end namespace clang::thread_safety 2686