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