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 VarDecl *SelfDecl = 0) { 468 CallingContext CallCtx(D); 469 470 if (MutexExp) { 471 if (StringLiteral* SLit = dyn_cast<StringLiteral>(MutexExp)) { 472 if (SLit->getString() == StringRef("*")) 473 // The "*" expr is a universal lock, which essentially turns off 474 // checks until it is removed from the lockset. 475 makeUniversal(); 476 else 477 // Ignore other string literals for now. 478 makeNop(); 479 return; 480 } 481 } 482 483 // If we are processing a raw attribute expression, with no substitutions. 484 if (DeclExp == 0) { 485 buildSExpr(MutexExp, 0); 486 return; 487 } 488 489 // Examine DeclExp to find SelfArg and FunArgs, which are used to substitute 490 // for formal parameters when we call buildMutexID later. 491 if (MemberExpr *ME = dyn_cast<MemberExpr>(DeclExp)) { 492 CallCtx.SelfArg = ME->getBase(); 493 CallCtx.SelfArrow = ME->isArrow(); 494 } else if (CXXMemberCallExpr *CE = dyn_cast<CXXMemberCallExpr>(DeclExp)) { 495 CallCtx.SelfArg = CE->getImplicitObjectArgument(); 496 CallCtx.SelfArrow = dyn_cast<MemberExpr>(CE->getCallee())->isArrow(); 497 CallCtx.NumArgs = CE->getNumArgs(); 498 CallCtx.FunArgs = CE->getArgs(); 499 } else if (CallExpr *CE = dyn_cast<CallExpr>(DeclExp)) { 500 CallCtx.NumArgs = CE->getNumArgs(); 501 CallCtx.FunArgs = CE->getArgs(); 502 } else if (CXXConstructExpr *CE = dyn_cast<CXXConstructExpr>(DeclExp)) { 503 CallCtx.SelfArg = 0; // Will be set below 504 CallCtx.NumArgs = CE->getNumArgs(); 505 CallCtx.FunArgs = CE->getArgs(); 506 } else if (D && isa<CXXDestructorDecl>(D)) { 507 // There's no such thing as a "destructor call" in the AST. 508 CallCtx.SelfArg = DeclExp; 509 } 510 511 // Hack to handle constructors, where self cannot be recovered from 512 // the expression. 513 if (SelfDecl && !CallCtx.SelfArg) { 514 DeclRefExpr SelfDRE(SelfDecl, false, SelfDecl->getType(), VK_LValue, 515 SelfDecl->getLocation()); 516 CallCtx.SelfArg = &SelfDRE; 517 518 // If the attribute has no arguments, then assume the argument is "this". 519 if (MutexExp == 0) 520 buildSExpr(CallCtx.SelfArg, 0); 521 else // For most attributes. 522 buildSExpr(MutexExp, &CallCtx); 523 return; 524 } 525 526 // If the attribute has no arguments, then assume the argument is "this". 527 if (MutexExp == 0) 528 buildSExpr(CallCtx.SelfArg, 0); 529 else // For most attributes. 530 buildSExpr(MutexExp, &CallCtx); 531 } 532 533 /// \brief Get index of next sibling of node i. 534 unsigned getNextSibling(unsigned i) const { 535 return i + NodeVec[i].size(); 536 } 537 538 public: 539 explicit SExpr(clang::Decl::EmptyShell e) { NodeVec.clear(); } 540 541 /// \param MutexExp The original mutex expression within an attribute 542 /// \param DeclExp An expression involving the Decl on which the attribute 543 /// occurs. 544 /// \param D The declaration to which the lock/unlock attribute is attached. 545 /// Caller must check isValid() after construction. 546 SExpr(Expr* MutexExp, Expr *DeclExp, const NamedDecl* D, 547 VarDecl *SelfDecl=0) { 548 buildSExprFromExpr(MutexExp, DeclExp, D, SelfDecl); 549 } 550 551 /// Return true if this is a valid decl sequence. 552 /// Caller must call this by hand after construction to handle errors. 553 bool isValid() const { 554 return !NodeVec.empty(); 555 } 556 557 bool shouldIgnore() const { 558 // Nop is a mutex that we have decided to deliberately ignore. 559 assert(NodeVec.size() > 0 && "Invalid Mutex"); 560 return NodeVec[0].kind() == EOP_Nop; 561 } 562 563 bool isUniversal() const { 564 assert(NodeVec.size() > 0 && "Invalid Mutex"); 565 return NodeVec[0].kind() == EOP_Universal; 566 } 567 568 /// Issue a warning about an invalid lock expression 569 static void warnInvalidLock(ThreadSafetyHandler &Handler, Expr* MutexExp, 570 Expr *DeclExp, const NamedDecl* D) { 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(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 }; 709 710 711 712 /// \brief A short list of SExprs 713 class MutexIDList : public SmallVector<SExpr, 3> { 714 public: 715 /// \brief Return true if the list contains the specified SExpr 716 /// Performs a linear search, because these lists are almost always very small. 717 bool contains(const SExpr& M) { 718 for (iterator I=begin(),E=end(); I != E; ++I) 719 if ((*I) == M) return true; 720 return false; 721 } 722 723 /// \brief Push M onto list, bud discard duplicates 724 void push_back_nodup(const SExpr& M) { 725 if (!contains(M)) push_back(M); 726 } 727 }; 728 729 730 731 /// \brief This is a helper class that stores info about the most recent 732 /// accquire of a Lock. 733 /// 734 /// The main body of the analysis maps MutexIDs to LockDatas. 735 struct LockData { 736 SourceLocation AcquireLoc; 737 738 /// \brief LKind stores whether a lock is held shared or exclusively. 739 /// Note that this analysis does not currently support either re-entrant 740 /// locking or lock "upgrading" and "downgrading" between exclusive and 741 /// shared. 742 /// 743 /// FIXME: add support for re-entrant locking and lock up/downgrading 744 LockKind LKind; 745 bool Managed; // for ScopedLockable objects 746 SExpr UnderlyingMutex; // for ScopedLockable objects 747 748 LockData(SourceLocation AcquireLoc, LockKind LKind, bool M = false) 749 : AcquireLoc(AcquireLoc), LKind(LKind), Managed(M), 750 UnderlyingMutex(Decl::EmptyShell()) 751 {} 752 753 LockData(SourceLocation AcquireLoc, LockKind LKind, const SExpr &Mu) 754 : AcquireLoc(AcquireLoc), LKind(LKind), Managed(false), 755 UnderlyingMutex(Mu) 756 {} 757 758 bool operator==(const LockData &other) const { 759 return AcquireLoc == other.AcquireLoc && LKind == other.LKind; 760 } 761 762 bool operator!=(const LockData &other) const { 763 return !(*this == other); 764 } 765 766 void Profile(llvm::FoldingSetNodeID &ID) const { 767 ID.AddInteger(AcquireLoc.getRawEncoding()); 768 ID.AddInteger(LKind); 769 } 770 771 bool isAtLeast(LockKind LK) { 772 return (LK == LK_Shared) || (LKind == LK_Exclusive); 773 } 774 }; 775 776 777 /// \brief A FactEntry stores a single fact that is known at a particular point 778 /// in the program execution. Currently, this is information regarding a lock 779 /// that is held at that point. 780 struct FactEntry { 781 SExpr MutID; 782 LockData LDat; 783 784 FactEntry(const SExpr& M, const LockData& L) 785 : MutID(M), LDat(L) 786 { } 787 }; 788 789 790 typedef unsigned short FactID; 791 792 /// \brief FactManager manages the memory for all facts that are created during 793 /// the analysis of a single routine. 794 class FactManager { 795 private: 796 std::vector<FactEntry> Facts; 797 798 public: 799 FactID newLock(const SExpr& M, const LockData& L) { 800 Facts.push_back(FactEntry(M,L)); 801 return static_cast<unsigned short>(Facts.size() - 1); 802 } 803 804 const FactEntry& operator[](FactID F) const { return Facts[F]; } 805 FactEntry& operator[](FactID F) { return Facts[F]; } 806 }; 807 808 809 /// \brief A FactSet is the set of facts that are known to be true at a 810 /// particular program point. FactSets must be small, because they are 811 /// frequently copied, and are thus implemented as a set of indices into a 812 /// table maintained by a FactManager. A typical FactSet only holds 1 or 2 813 /// locks, so we can get away with doing a linear search for lookup. Note 814 /// that a hashtable or map is inappropriate in this case, because lookups 815 /// may involve partial pattern matches, rather than exact matches. 816 class FactSet { 817 private: 818 typedef SmallVector<FactID, 4> FactVec; 819 820 FactVec FactIDs; 821 822 public: 823 typedef FactVec::iterator iterator; 824 typedef FactVec::const_iterator const_iterator; 825 826 iterator begin() { return FactIDs.begin(); } 827 const_iterator begin() const { return FactIDs.begin(); } 828 829 iterator end() { return FactIDs.end(); } 830 const_iterator end() const { return FactIDs.end(); } 831 832 bool isEmpty() const { return FactIDs.size() == 0; } 833 834 FactID addLock(FactManager& FM, const SExpr& M, const LockData& L) { 835 FactID F = FM.newLock(M, L); 836 FactIDs.push_back(F); 837 return F; 838 } 839 840 bool removeLock(FactManager& FM, const SExpr& M) { 841 unsigned n = FactIDs.size(); 842 if (n == 0) 843 return false; 844 845 for (unsigned i = 0; i < n-1; ++i) { 846 if (FM[FactIDs[i]].MutID.matches(M)) { 847 FactIDs[i] = FactIDs[n-1]; 848 FactIDs.pop_back(); 849 return true; 850 } 851 } 852 if (FM[FactIDs[n-1]].MutID.matches(M)) { 853 FactIDs.pop_back(); 854 return true; 855 } 856 return false; 857 } 858 859 LockData* findLock(FactManager &FM, const SExpr &M) const { 860 for (const_iterator I = begin(), E = end(); I != E; ++I) { 861 const SExpr &Exp = FM[*I].MutID; 862 if (Exp.matches(M)) 863 return &FM[*I].LDat; 864 } 865 return 0; 866 } 867 868 LockData* findLockUniv(FactManager &FM, const SExpr &M) const { 869 for (const_iterator I = begin(), E = end(); I != E; ++I) { 870 const SExpr &Exp = FM[*I].MutID; 871 if (Exp.matches(M) || Exp.isUniversal()) 872 return &FM[*I].LDat; 873 } 874 return 0; 875 } 876 877 FactEntry* findPartialMatch(FactManager &FM, const SExpr &M) const { 878 for (const_iterator I=begin(), E=end(); I != E; ++I) { 879 const SExpr& Exp = FM[*I].MutID; 880 if (Exp.partiallyMatches(M)) return &FM[*I]; 881 } 882 return 0; 883 } 884 }; 885 886 887 888 /// A Lockset maps each SExpr (defined above) to information about how it has 889 /// been locked. 890 typedef llvm::ImmutableMap<SExpr, LockData> Lockset; 891 typedef llvm::ImmutableMap<const NamedDecl*, unsigned> LocalVarContext; 892 893 class LocalVariableMap; 894 895 /// A side (entry or exit) of a CFG node. 896 enum CFGBlockSide { CBS_Entry, CBS_Exit }; 897 898 /// CFGBlockInfo is a struct which contains all the information that is 899 /// maintained for each block in the CFG. See LocalVariableMap for more 900 /// information about the contexts. 901 struct CFGBlockInfo { 902 FactSet EntrySet; // Lockset held at entry to block 903 FactSet ExitSet; // Lockset held at exit from block 904 LocalVarContext EntryContext; // Context held at entry to block 905 LocalVarContext ExitContext; // Context held at exit from block 906 SourceLocation EntryLoc; // Location of first statement in block 907 SourceLocation ExitLoc; // Location of last statement in block. 908 unsigned EntryIndex; // Used to replay contexts later 909 bool Reachable; // Is this block reachable? 910 911 const FactSet &getSet(CFGBlockSide Side) const { 912 return Side == CBS_Entry ? EntrySet : ExitSet; 913 } 914 SourceLocation getLocation(CFGBlockSide Side) const { 915 return Side == CBS_Entry ? EntryLoc : ExitLoc; 916 } 917 918 private: 919 CFGBlockInfo(LocalVarContext EmptyCtx) 920 : EntryContext(EmptyCtx), ExitContext(EmptyCtx), Reachable(false) 921 { } 922 923 public: 924 static CFGBlockInfo getEmptyBlockInfo(LocalVariableMap &M); 925 }; 926 927 928 929 // A LocalVariableMap maintains a map from local variables to their currently 930 // valid definitions. It provides SSA-like functionality when traversing the 931 // CFG. Like SSA, each definition or assignment to a variable is assigned a 932 // unique name (an integer), which acts as the SSA name for that definition. 933 // The total set of names is shared among all CFG basic blocks. 934 // Unlike SSA, we do not rewrite expressions to replace local variables declrefs 935 // with their SSA-names. Instead, we compute a Context for each point in the 936 // code, which maps local variables to the appropriate SSA-name. This map 937 // changes with each assignment. 938 // 939 // The map is computed in a single pass over the CFG. Subsequent analyses can 940 // then query the map to find the appropriate Context for a statement, and use 941 // that Context to look up the definitions of variables. 942 class LocalVariableMap { 943 public: 944 typedef LocalVarContext Context; 945 946 /// A VarDefinition consists of an expression, representing the value of the 947 /// variable, along with the context in which that expression should be 948 /// interpreted. A reference VarDefinition does not itself contain this 949 /// information, but instead contains a pointer to a previous VarDefinition. 950 struct VarDefinition { 951 public: 952 friend class LocalVariableMap; 953 954 const NamedDecl *Dec; // The original declaration for this variable. 955 const Expr *Exp; // The expression for this variable, OR 956 unsigned Ref; // Reference to another VarDefinition 957 Context Ctx; // The map with which Exp should be interpreted. 958 959 bool isReference() { return !Exp; } 960 961 private: 962 // Create ordinary variable definition 963 VarDefinition(const NamedDecl *D, const Expr *E, Context C) 964 : Dec(D), Exp(E), Ref(0), Ctx(C) 965 { } 966 967 // Create reference to previous definition 968 VarDefinition(const NamedDecl *D, unsigned R, Context C) 969 : Dec(D), Exp(0), Ref(R), Ctx(C) 970 { } 971 }; 972 973 private: 974 Context::Factory ContextFactory; 975 std::vector<VarDefinition> VarDefinitions; 976 std::vector<unsigned> CtxIndices; 977 std::vector<std::pair<Stmt*, Context> > SavedContexts; 978 979 public: 980 LocalVariableMap() { 981 // index 0 is a placeholder for undefined variables (aka phi-nodes). 982 VarDefinitions.push_back(VarDefinition(0, 0u, getEmptyContext())); 983 } 984 985 /// Look up a definition, within the given context. 986 const VarDefinition* lookup(const NamedDecl *D, Context Ctx) { 987 const unsigned *i = Ctx.lookup(D); 988 if (!i) 989 return 0; 990 assert(*i < VarDefinitions.size()); 991 return &VarDefinitions[*i]; 992 } 993 994 /// Look up the definition for D within the given context. Returns 995 /// NULL if the expression is not statically known. If successful, also 996 /// modifies Ctx to hold the context of the return Expr. 997 const Expr* lookupExpr(const NamedDecl *D, Context &Ctx) { 998 const unsigned *P = Ctx.lookup(D); 999 if (!P) 1000 return 0; 1001 1002 unsigned i = *P; 1003 while (i > 0) { 1004 if (VarDefinitions[i].Exp) { 1005 Ctx = VarDefinitions[i].Ctx; 1006 return VarDefinitions[i].Exp; 1007 } 1008 i = VarDefinitions[i].Ref; 1009 } 1010 return 0; 1011 } 1012 1013 Context getEmptyContext() { return ContextFactory.getEmptyMap(); } 1014 1015 /// Return the next context after processing S. This function is used by 1016 /// clients of the class to get the appropriate context when traversing the 1017 /// CFG. It must be called for every assignment or DeclStmt. 1018 Context getNextContext(unsigned &CtxIndex, Stmt *S, Context C) { 1019 if (SavedContexts[CtxIndex+1].first == S) { 1020 CtxIndex++; 1021 Context Result = SavedContexts[CtxIndex].second; 1022 return Result; 1023 } 1024 return C; 1025 } 1026 1027 void dumpVarDefinitionName(unsigned i) { 1028 if (i == 0) { 1029 llvm::errs() << "Undefined"; 1030 return; 1031 } 1032 const NamedDecl *Dec = VarDefinitions[i].Dec; 1033 if (!Dec) { 1034 llvm::errs() << "<<NULL>>"; 1035 return; 1036 } 1037 Dec->printName(llvm::errs()); 1038 llvm::errs() << "." << i << " " << ((const void*) Dec); 1039 } 1040 1041 /// Dumps an ASCII representation of the variable map to llvm::errs() 1042 void dump() { 1043 for (unsigned i = 1, e = VarDefinitions.size(); i < e; ++i) { 1044 const Expr *Exp = VarDefinitions[i].Exp; 1045 unsigned Ref = VarDefinitions[i].Ref; 1046 1047 dumpVarDefinitionName(i); 1048 llvm::errs() << " = "; 1049 if (Exp) Exp->dump(); 1050 else { 1051 dumpVarDefinitionName(Ref); 1052 llvm::errs() << "\n"; 1053 } 1054 } 1055 } 1056 1057 /// Dumps an ASCII representation of a Context to llvm::errs() 1058 void dumpContext(Context C) { 1059 for (Context::iterator I = C.begin(), E = C.end(); I != E; ++I) { 1060 const NamedDecl *D = I.getKey(); 1061 D->printName(llvm::errs()); 1062 const unsigned *i = C.lookup(D); 1063 llvm::errs() << " -> "; 1064 dumpVarDefinitionName(*i); 1065 llvm::errs() << "\n"; 1066 } 1067 } 1068 1069 /// Builds the variable map. 1070 void traverseCFG(CFG *CFGraph, PostOrderCFGView *SortedGraph, 1071 std::vector<CFGBlockInfo> &BlockInfo); 1072 1073 protected: 1074 // Get the current context index 1075 unsigned getContextIndex() { return SavedContexts.size()-1; } 1076 1077 // Save the current context for later replay 1078 void saveContext(Stmt *S, Context C) { 1079 SavedContexts.push_back(std::make_pair(S,C)); 1080 } 1081 1082 // Adds a new definition to the given context, and returns a new context. 1083 // This method should be called when declaring a new variable. 1084 Context addDefinition(const NamedDecl *D, Expr *Exp, Context Ctx) { 1085 assert(!Ctx.contains(D)); 1086 unsigned newID = VarDefinitions.size(); 1087 Context NewCtx = ContextFactory.add(Ctx, D, newID); 1088 VarDefinitions.push_back(VarDefinition(D, Exp, Ctx)); 1089 return NewCtx; 1090 } 1091 1092 // Add a new reference to an existing definition. 1093 Context addReference(const NamedDecl *D, unsigned i, Context Ctx) { 1094 unsigned newID = VarDefinitions.size(); 1095 Context NewCtx = ContextFactory.add(Ctx, D, newID); 1096 VarDefinitions.push_back(VarDefinition(D, i, Ctx)); 1097 return NewCtx; 1098 } 1099 1100 // Updates a definition only if that definition is already in the map. 1101 // This method should be called when assigning to an existing variable. 1102 Context updateDefinition(const NamedDecl *D, Expr *Exp, Context Ctx) { 1103 if (Ctx.contains(D)) { 1104 unsigned newID = VarDefinitions.size(); 1105 Context NewCtx = ContextFactory.remove(Ctx, D); 1106 NewCtx = ContextFactory.add(NewCtx, D, newID); 1107 VarDefinitions.push_back(VarDefinition(D, Exp, Ctx)); 1108 return NewCtx; 1109 } 1110 return Ctx; 1111 } 1112 1113 // Removes a definition from the context, but keeps the variable name 1114 // as a valid variable. The index 0 is a placeholder for cleared definitions. 1115 Context clearDefinition(const NamedDecl *D, Context Ctx) { 1116 Context NewCtx = Ctx; 1117 if (NewCtx.contains(D)) { 1118 NewCtx = ContextFactory.remove(NewCtx, D); 1119 NewCtx = ContextFactory.add(NewCtx, D, 0); 1120 } 1121 return NewCtx; 1122 } 1123 1124 // Remove a definition entirely frmo the context. 1125 Context removeDefinition(const NamedDecl *D, Context Ctx) { 1126 Context NewCtx = Ctx; 1127 if (NewCtx.contains(D)) { 1128 NewCtx = ContextFactory.remove(NewCtx, D); 1129 } 1130 return NewCtx; 1131 } 1132 1133 Context intersectContexts(Context C1, Context C2); 1134 Context createReferenceContext(Context C); 1135 void intersectBackEdge(Context C1, Context C2); 1136 1137 friend class VarMapBuilder; 1138 }; 1139 1140 1141 // This has to be defined after LocalVariableMap. 1142 CFGBlockInfo CFGBlockInfo::getEmptyBlockInfo(LocalVariableMap &M) { 1143 return CFGBlockInfo(M.getEmptyContext()); 1144 } 1145 1146 1147 /// Visitor which builds a LocalVariableMap 1148 class VarMapBuilder : public StmtVisitor<VarMapBuilder> { 1149 public: 1150 LocalVariableMap* VMap; 1151 LocalVariableMap::Context Ctx; 1152 1153 VarMapBuilder(LocalVariableMap *VM, LocalVariableMap::Context C) 1154 : VMap(VM), Ctx(C) {} 1155 1156 void VisitDeclStmt(DeclStmt *S); 1157 void VisitBinaryOperator(BinaryOperator *BO); 1158 }; 1159 1160 1161 // Add new local variables to the variable map 1162 void VarMapBuilder::VisitDeclStmt(DeclStmt *S) { 1163 bool modifiedCtx = false; 1164 DeclGroupRef DGrp = S->getDeclGroup(); 1165 for (DeclGroupRef::iterator I = DGrp.begin(), E = DGrp.end(); I != E; ++I) { 1166 if (VarDecl *VD = dyn_cast_or_null<VarDecl>(*I)) { 1167 Expr *E = VD->getInit(); 1168 1169 // Add local variables with trivial type to the variable map 1170 QualType T = VD->getType(); 1171 if (T.isTrivialType(VD->getASTContext())) { 1172 Ctx = VMap->addDefinition(VD, E, Ctx); 1173 modifiedCtx = true; 1174 } 1175 } 1176 } 1177 if (modifiedCtx) 1178 VMap->saveContext(S, Ctx); 1179 } 1180 1181 // Update local variable definitions in variable map 1182 void VarMapBuilder::VisitBinaryOperator(BinaryOperator *BO) { 1183 if (!BO->isAssignmentOp()) 1184 return; 1185 1186 Expr *LHSExp = BO->getLHS()->IgnoreParenCasts(); 1187 1188 // Update the variable map and current context. 1189 if (DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(LHSExp)) { 1190 ValueDecl *VDec = DRE->getDecl(); 1191 if (Ctx.lookup(VDec)) { 1192 if (BO->getOpcode() == BO_Assign) 1193 Ctx = VMap->updateDefinition(VDec, BO->getRHS(), Ctx); 1194 else 1195 // FIXME -- handle compound assignment operators 1196 Ctx = VMap->clearDefinition(VDec, Ctx); 1197 VMap->saveContext(BO, Ctx); 1198 } 1199 } 1200 } 1201 1202 1203 // Computes the intersection of two contexts. The intersection is the 1204 // set of variables which have the same definition in both contexts; 1205 // variables with different definitions are discarded. 1206 LocalVariableMap::Context 1207 LocalVariableMap::intersectContexts(Context C1, Context C2) { 1208 Context Result = C1; 1209 for (Context::iterator I = C1.begin(), E = C1.end(); I != E; ++I) { 1210 const NamedDecl *Dec = I.getKey(); 1211 unsigned i1 = I.getData(); 1212 const unsigned *i2 = C2.lookup(Dec); 1213 if (!i2) // variable doesn't exist on second path 1214 Result = removeDefinition(Dec, Result); 1215 else if (*i2 != i1) // variable exists, but has different definition 1216 Result = clearDefinition(Dec, Result); 1217 } 1218 return Result; 1219 } 1220 1221 // For every variable in C, create a new variable that refers to the 1222 // definition in C. Return a new context that contains these new variables. 1223 // (We use this for a naive implementation of SSA on loop back-edges.) 1224 LocalVariableMap::Context LocalVariableMap::createReferenceContext(Context C) { 1225 Context Result = getEmptyContext(); 1226 for (Context::iterator I = C.begin(), E = C.end(); I != E; ++I) { 1227 const NamedDecl *Dec = I.getKey(); 1228 unsigned i = I.getData(); 1229 Result = addReference(Dec, i, Result); 1230 } 1231 return Result; 1232 } 1233 1234 // This routine also takes the intersection of C1 and C2, but it does so by 1235 // altering the VarDefinitions. C1 must be the result of an earlier call to 1236 // createReferenceContext. 1237 void LocalVariableMap::intersectBackEdge(Context C1, Context C2) { 1238 for (Context::iterator I = C1.begin(), E = C1.end(); I != E; ++I) { 1239 const NamedDecl *Dec = I.getKey(); 1240 unsigned i1 = I.getData(); 1241 VarDefinition *VDef = &VarDefinitions[i1]; 1242 assert(VDef->isReference()); 1243 1244 const unsigned *i2 = C2.lookup(Dec); 1245 if (!i2 || (*i2 != i1)) 1246 VDef->Ref = 0; // Mark this variable as undefined 1247 } 1248 } 1249 1250 1251 // Traverse the CFG in topological order, so all predecessors of a block 1252 // (excluding back-edges) are visited before the block itself. At 1253 // each point in the code, we calculate a Context, which holds the set of 1254 // variable definitions which are visible at that point in execution. 1255 // Visible variables are mapped to their definitions using an array that 1256 // contains all definitions. 1257 // 1258 // At join points in the CFG, the set is computed as the intersection of 1259 // the incoming sets along each edge, E.g. 1260 // 1261 // { Context | VarDefinitions } 1262 // int x = 0; { x -> x1 | x1 = 0 } 1263 // int y = 0; { x -> x1, y -> y1 | y1 = 0, x1 = 0 } 1264 // if (b) x = 1; { x -> x2, y -> y1 | x2 = 1, y1 = 0, ... } 1265 // else x = 2; { x -> x3, y -> y1 | x3 = 2, x2 = 1, ... } 1266 // ... { y -> y1 (x is unknown) | x3 = 2, x2 = 1, ... } 1267 // 1268 // This is essentially a simpler and more naive version of the standard SSA 1269 // algorithm. Those definitions that remain in the intersection are from blocks 1270 // that strictly dominate the current block. We do not bother to insert proper 1271 // phi nodes, because they are not used in our analysis; instead, wherever 1272 // a phi node would be required, we simply remove that definition from the 1273 // context (E.g. x above). 1274 // 1275 // The initial traversal does not capture back-edges, so those need to be 1276 // handled on a separate pass. Whenever the first pass encounters an 1277 // incoming back edge, it duplicates the context, creating new definitions 1278 // that refer back to the originals. (These correspond to places where SSA 1279 // might have to insert a phi node.) On the second pass, these definitions are 1280 // set to NULL if the variable has changed on the back-edge (i.e. a phi 1281 // node was actually required.) E.g. 1282 // 1283 // { Context | VarDefinitions } 1284 // int x = 0, y = 0; { x -> x1, y -> y1 | y1 = 0, x1 = 0 } 1285 // while (b) { x -> x2, y -> y1 | [1st:] x2=x1; [2nd:] x2=NULL; } 1286 // x = x+1; { x -> x3, y -> y1 | x3 = x2 + 1, ... } 1287 // ... { y -> y1 | x3 = 2, x2 = 1, ... } 1288 // 1289 void LocalVariableMap::traverseCFG(CFG *CFGraph, 1290 PostOrderCFGView *SortedGraph, 1291 std::vector<CFGBlockInfo> &BlockInfo) { 1292 PostOrderCFGView::CFGBlockSet VisitedBlocks(CFGraph); 1293 1294 CtxIndices.resize(CFGraph->getNumBlockIDs()); 1295 1296 for (PostOrderCFGView::iterator I = SortedGraph->begin(), 1297 E = SortedGraph->end(); I!= E; ++I) { 1298 const CFGBlock *CurrBlock = *I; 1299 int CurrBlockID = CurrBlock->getBlockID(); 1300 CFGBlockInfo *CurrBlockInfo = &BlockInfo[CurrBlockID]; 1301 1302 VisitedBlocks.insert(CurrBlock); 1303 1304 // Calculate the entry context for the current block 1305 bool HasBackEdges = false; 1306 bool CtxInit = true; 1307 for (CFGBlock::const_pred_iterator PI = CurrBlock->pred_begin(), 1308 PE = CurrBlock->pred_end(); PI != PE; ++PI) { 1309 // if *PI -> CurrBlock is a back edge, so skip it 1310 if (*PI == 0 || !VisitedBlocks.alreadySet(*PI)) { 1311 HasBackEdges = true; 1312 continue; 1313 } 1314 1315 int PrevBlockID = (*PI)->getBlockID(); 1316 CFGBlockInfo *PrevBlockInfo = &BlockInfo[PrevBlockID]; 1317 1318 if (CtxInit) { 1319 CurrBlockInfo->EntryContext = PrevBlockInfo->ExitContext; 1320 CtxInit = false; 1321 } 1322 else { 1323 CurrBlockInfo->EntryContext = 1324 intersectContexts(CurrBlockInfo->EntryContext, 1325 PrevBlockInfo->ExitContext); 1326 } 1327 } 1328 1329 // Duplicate the context if we have back-edges, so we can call 1330 // intersectBackEdges later. 1331 if (HasBackEdges) 1332 CurrBlockInfo->EntryContext = 1333 createReferenceContext(CurrBlockInfo->EntryContext); 1334 1335 // Create a starting context index for the current block 1336 saveContext(0, CurrBlockInfo->EntryContext); 1337 CurrBlockInfo->EntryIndex = getContextIndex(); 1338 1339 // Visit all the statements in the basic block. 1340 VarMapBuilder VMapBuilder(this, CurrBlockInfo->EntryContext); 1341 for (CFGBlock::const_iterator BI = CurrBlock->begin(), 1342 BE = CurrBlock->end(); BI != BE; ++BI) { 1343 switch (BI->getKind()) { 1344 case CFGElement::Statement: { 1345 const CFGStmt *CS = cast<CFGStmt>(&*BI); 1346 VMapBuilder.Visit(const_cast<Stmt*>(CS->getStmt())); 1347 break; 1348 } 1349 default: 1350 break; 1351 } 1352 } 1353 CurrBlockInfo->ExitContext = VMapBuilder.Ctx; 1354 1355 // Mark variables on back edges as "unknown" if they've been changed. 1356 for (CFGBlock::const_succ_iterator SI = CurrBlock->succ_begin(), 1357 SE = CurrBlock->succ_end(); SI != SE; ++SI) { 1358 // if CurrBlock -> *SI is *not* a back edge 1359 if (*SI == 0 || !VisitedBlocks.alreadySet(*SI)) 1360 continue; 1361 1362 CFGBlock *FirstLoopBlock = *SI; 1363 Context LoopBegin = BlockInfo[FirstLoopBlock->getBlockID()].EntryContext; 1364 Context LoopEnd = CurrBlockInfo->ExitContext; 1365 intersectBackEdge(LoopBegin, LoopEnd); 1366 } 1367 } 1368 1369 // Put an extra entry at the end of the indexed context array 1370 unsigned exitID = CFGraph->getExit().getBlockID(); 1371 saveContext(0, BlockInfo[exitID].ExitContext); 1372 } 1373 1374 /// Find the appropriate source locations to use when producing diagnostics for 1375 /// each block in the CFG. 1376 static void findBlockLocations(CFG *CFGraph, 1377 PostOrderCFGView *SortedGraph, 1378 std::vector<CFGBlockInfo> &BlockInfo) { 1379 for (PostOrderCFGView::iterator I = SortedGraph->begin(), 1380 E = SortedGraph->end(); I!= E; ++I) { 1381 const CFGBlock *CurrBlock = *I; 1382 CFGBlockInfo *CurrBlockInfo = &BlockInfo[CurrBlock->getBlockID()]; 1383 1384 // Find the source location of the last statement in the block, if the 1385 // block is not empty. 1386 if (const Stmt *S = CurrBlock->getTerminator()) { 1387 CurrBlockInfo->EntryLoc = CurrBlockInfo->ExitLoc = S->getLocStart(); 1388 } else { 1389 for (CFGBlock::const_reverse_iterator BI = CurrBlock->rbegin(), 1390 BE = CurrBlock->rend(); BI != BE; ++BI) { 1391 // FIXME: Handle other CFGElement kinds. 1392 if (const CFGStmt *CS = dyn_cast<CFGStmt>(&*BI)) { 1393 CurrBlockInfo->ExitLoc = CS->getStmt()->getLocStart(); 1394 break; 1395 } 1396 } 1397 } 1398 1399 if (!CurrBlockInfo->ExitLoc.isInvalid()) { 1400 // This block contains at least one statement. Find the source location 1401 // of the first statement in the block. 1402 for (CFGBlock::const_iterator BI = CurrBlock->begin(), 1403 BE = CurrBlock->end(); BI != BE; ++BI) { 1404 // FIXME: Handle other CFGElement kinds. 1405 if (const CFGStmt *CS = dyn_cast<CFGStmt>(&*BI)) { 1406 CurrBlockInfo->EntryLoc = CS->getStmt()->getLocStart(); 1407 break; 1408 } 1409 } 1410 } else if (CurrBlock->pred_size() == 1 && *CurrBlock->pred_begin() && 1411 CurrBlock != &CFGraph->getExit()) { 1412 // The block is empty, and has a single predecessor. Use its exit 1413 // location. 1414 CurrBlockInfo->EntryLoc = CurrBlockInfo->ExitLoc = 1415 BlockInfo[(*CurrBlock->pred_begin())->getBlockID()].ExitLoc; 1416 } 1417 } 1418 } 1419 1420 /// \brief Class which implements the core thread safety analysis routines. 1421 class ThreadSafetyAnalyzer { 1422 friend class BuildLockset; 1423 1424 ThreadSafetyHandler &Handler; 1425 LocalVariableMap LocalVarMap; 1426 FactManager FactMan; 1427 std::vector<CFGBlockInfo> BlockInfo; 1428 1429 public: 1430 ThreadSafetyAnalyzer(ThreadSafetyHandler &H) : Handler(H) {} 1431 1432 void addLock(FactSet &FSet, const SExpr &Mutex, const LockData &LDat); 1433 void removeLock(FactSet &FSet, const SExpr &Mutex, 1434 SourceLocation UnlockLoc, bool FullyRemove=false); 1435 1436 template <typename AttrType> 1437 void getMutexIDs(MutexIDList &Mtxs, AttrType *Attr, Expr *Exp, 1438 const NamedDecl *D, VarDecl *SelfDecl=0); 1439 1440 template <class AttrType> 1441 void getMutexIDs(MutexIDList &Mtxs, AttrType *Attr, Expr *Exp, 1442 const NamedDecl *D, 1443 const CFGBlock *PredBlock, const CFGBlock *CurrBlock, 1444 Expr *BrE, bool Neg); 1445 1446 const CallExpr* getTrylockCallExpr(const Stmt *Cond, LocalVarContext C, 1447 bool &Negate); 1448 1449 void getEdgeLockset(FactSet &Result, const FactSet &ExitSet, 1450 const CFGBlock* PredBlock, 1451 const CFGBlock *CurrBlock); 1452 1453 void intersectAndWarn(FactSet &FSet1, const FactSet &FSet2, 1454 SourceLocation JoinLoc, 1455 LockErrorKind LEK1, LockErrorKind LEK2, 1456 bool Modify=true); 1457 1458 void intersectAndWarn(FactSet &FSet1, const FactSet &FSet2, 1459 SourceLocation JoinLoc, LockErrorKind LEK1, 1460 bool Modify=true) { 1461 intersectAndWarn(FSet1, FSet2, JoinLoc, LEK1, LEK1, Modify); 1462 } 1463 1464 void runAnalysis(AnalysisDeclContext &AC); 1465 }; 1466 1467 1468 /// \brief Add a new lock to the lockset, warning if the lock is already there. 1469 /// \param Mutex -- the Mutex expression for the lock 1470 /// \param LDat -- the LockData for the lock 1471 void ThreadSafetyAnalyzer::addLock(FactSet &FSet, const SExpr &Mutex, 1472 const LockData &LDat) { 1473 // FIXME: deal with acquired before/after annotations. 1474 // FIXME: Don't always warn when we have support for reentrant locks. 1475 if (Mutex.shouldIgnore()) 1476 return; 1477 1478 if (FSet.findLock(FactMan, Mutex)) { 1479 Handler.handleDoubleLock(Mutex.toString(), LDat.AcquireLoc); 1480 } else { 1481 FSet.addLock(FactMan, Mutex, LDat); 1482 } 1483 } 1484 1485 1486 /// \brief Remove a lock from the lockset, warning if the lock is not there. 1487 /// \param Mutex The lock expression corresponding to the lock to be removed 1488 /// \param UnlockLoc The source location of the unlock (only used in error msg) 1489 void ThreadSafetyAnalyzer::removeLock(FactSet &FSet, 1490 const SExpr &Mutex, 1491 SourceLocation UnlockLoc, 1492 bool FullyRemove) { 1493 if (Mutex.shouldIgnore()) 1494 return; 1495 1496 const LockData *LDat = FSet.findLock(FactMan, Mutex); 1497 if (!LDat) { 1498 Handler.handleUnmatchedUnlock(Mutex.toString(), UnlockLoc); 1499 return; 1500 } 1501 1502 if (LDat->UnderlyingMutex.isValid()) { 1503 // This is scoped lockable object, which manages the real mutex. 1504 if (FullyRemove) { 1505 // We're destroying the managing object. 1506 // Remove the underlying mutex if it exists; but don't warn. 1507 if (FSet.findLock(FactMan, LDat->UnderlyingMutex)) 1508 FSet.removeLock(FactMan, LDat->UnderlyingMutex); 1509 } else { 1510 // We're releasing the underlying mutex, but not destroying the 1511 // managing object. Warn on dual release. 1512 if (!FSet.findLock(FactMan, LDat->UnderlyingMutex)) { 1513 Handler.handleUnmatchedUnlock(LDat->UnderlyingMutex.toString(), 1514 UnlockLoc); 1515 } 1516 FSet.removeLock(FactMan, LDat->UnderlyingMutex); 1517 return; 1518 } 1519 } 1520 FSet.removeLock(FactMan, Mutex); 1521 } 1522 1523 1524 /// \brief Extract the list of mutexIDs from the attribute on an expression, 1525 /// and push them onto Mtxs, discarding any duplicates. 1526 template <typename AttrType> 1527 void ThreadSafetyAnalyzer::getMutexIDs(MutexIDList &Mtxs, AttrType *Attr, 1528 Expr *Exp, const NamedDecl *D, 1529 VarDecl *SelfDecl) { 1530 typedef typename AttrType::args_iterator iterator_type; 1531 1532 if (Attr->args_size() == 0) { 1533 // The mutex held is the "this" object. 1534 SExpr Mu(0, Exp, D, SelfDecl); 1535 if (!Mu.isValid()) 1536 SExpr::warnInvalidLock(Handler, 0, Exp, D); 1537 else 1538 Mtxs.push_back_nodup(Mu); 1539 return; 1540 } 1541 1542 for (iterator_type I=Attr->args_begin(), E=Attr->args_end(); I != E; ++I) { 1543 SExpr Mu(*I, Exp, D, SelfDecl); 1544 if (!Mu.isValid()) 1545 SExpr::warnInvalidLock(Handler, *I, Exp, D); 1546 else 1547 Mtxs.push_back_nodup(Mu); 1548 } 1549 } 1550 1551 1552 /// \brief Extract the list of mutexIDs from a trylock attribute. If the 1553 /// trylock applies to the given edge, then push them onto Mtxs, discarding 1554 /// any duplicates. 1555 template <class AttrType> 1556 void ThreadSafetyAnalyzer::getMutexIDs(MutexIDList &Mtxs, AttrType *Attr, 1557 Expr *Exp, const NamedDecl *D, 1558 const CFGBlock *PredBlock, 1559 const CFGBlock *CurrBlock, 1560 Expr *BrE, bool Neg) { 1561 // Find out which branch has the lock 1562 bool branch = 0; 1563 if (CXXBoolLiteralExpr *BLE = dyn_cast_or_null<CXXBoolLiteralExpr>(BrE)) { 1564 branch = BLE->getValue(); 1565 } 1566 else if (IntegerLiteral *ILE = dyn_cast_or_null<IntegerLiteral>(BrE)) { 1567 branch = ILE->getValue().getBoolValue(); 1568 } 1569 int branchnum = branch ? 0 : 1; 1570 if (Neg) branchnum = !branchnum; 1571 1572 // If we've taken the trylock branch, then add the lock 1573 int i = 0; 1574 for (CFGBlock::const_succ_iterator SI = PredBlock->succ_begin(), 1575 SE = PredBlock->succ_end(); SI != SE && i < 2; ++SI, ++i) { 1576 if (*SI == CurrBlock && i == branchnum) { 1577 getMutexIDs(Mtxs, Attr, Exp, D); 1578 } 1579 } 1580 } 1581 1582 1583 bool getStaticBooleanValue(Expr* E, bool& TCond) { 1584 if (isa<CXXNullPtrLiteralExpr>(E) || isa<GNUNullExpr>(E)) { 1585 TCond = false; 1586 return true; 1587 } else if (CXXBoolLiteralExpr *BLE = dyn_cast<CXXBoolLiteralExpr>(E)) { 1588 TCond = BLE->getValue(); 1589 return true; 1590 } else if (IntegerLiteral *ILE = dyn_cast<IntegerLiteral>(E)) { 1591 TCond = ILE->getValue().getBoolValue(); 1592 return true; 1593 } else if (ImplicitCastExpr *CE = dyn_cast<ImplicitCastExpr>(E)) { 1594 return getStaticBooleanValue(CE->getSubExpr(), TCond); 1595 } 1596 return false; 1597 } 1598 1599 1600 // If Cond can be traced back to a function call, return the call expression. 1601 // The negate variable should be called with false, and will be set to true 1602 // if the function call is negated, e.g. if (!mu.tryLock(...)) 1603 const CallExpr* ThreadSafetyAnalyzer::getTrylockCallExpr(const Stmt *Cond, 1604 LocalVarContext C, 1605 bool &Negate) { 1606 if (!Cond) 1607 return 0; 1608 1609 if (const CallExpr *CallExp = dyn_cast<CallExpr>(Cond)) { 1610 return CallExp; 1611 } 1612 else if (const ParenExpr *PE = dyn_cast<ParenExpr>(Cond)) { 1613 return getTrylockCallExpr(PE->getSubExpr(), C, Negate); 1614 } 1615 else if (const ImplicitCastExpr *CE = dyn_cast<ImplicitCastExpr>(Cond)) { 1616 return getTrylockCallExpr(CE->getSubExpr(), C, Negate); 1617 } 1618 else if (const ExprWithCleanups* EWC = dyn_cast<ExprWithCleanups>(Cond)) { 1619 return getTrylockCallExpr(EWC->getSubExpr(), C, Negate); 1620 } 1621 else if (const DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(Cond)) { 1622 const Expr *E = LocalVarMap.lookupExpr(DRE->getDecl(), C); 1623 return getTrylockCallExpr(E, C, Negate); 1624 } 1625 else if (const UnaryOperator *UOP = dyn_cast<UnaryOperator>(Cond)) { 1626 if (UOP->getOpcode() == UO_LNot) { 1627 Negate = !Negate; 1628 return getTrylockCallExpr(UOP->getSubExpr(), C, Negate); 1629 } 1630 return 0; 1631 } 1632 else if (const BinaryOperator *BOP = dyn_cast<BinaryOperator>(Cond)) { 1633 if (BOP->getOpcode() == BO_EQ || BOP->getOpcode() == BO_NE) { 1634 if (BOP->getOpcode() == BO_NE) 1635 Negate = !Negate; 1636 1637 bool TCond = false; 1638 if (getStaticBooleanValue(BOP->getRHS(), TCond)) { 1639 if (!TCond) Negate = !Negate; 1640 return getTrylockCallExpr(BOP->getLHS(), C, Negate); 1641 } 1642 else if (getStaticBooleanValue(BOP->getLHS(), TCond)) { 1643 if (!TCond) Negate = !Negate; 1644 return getTrylockCallExpr(BOP->getRHS(), C, Negate); 1645 } 1646 return 0; 1647 } 1648 return 0; 1649 } 1650 // FIXME -- handle && and || as well. 1651 return 0; 1652 } 1653 1654 1655 /// \brief Find the lockset that holds on the edge between PredBlock 1656 /// and CurrBlock. The edge set is the exit set of PredBlock (passed 1657 /// as the ExitSet parameter) plus any trylocks, which are conditionally held. 1658 void ThreadSafetyAnalyzer::getEdgeLockset(FactSet& Result, 1659 const FactSet &ExitSet, 1660 const CFGBlock *PredBlock, 1661 const CFGBlock *CurrBlock) { 1662 Result = ExitSet; 1663 1664 if (!PredBlock->getTerminatorCondition()) 1665 return; 1666 1667 bool Negate = false; 1668 const Stmt *Cond = PredBlock->getTerminatorCondition(); 1669 const CFGBlockInfo *PredBlockInfo = &BlockInfo[PredBlock->getBlockID()]; 1670 const LocalVarContext &LVarCtx = PredBlockInfo->ExitContext; 1671 1672 CallExpr *Exp = 1673 const_cast<CallExpr*>(getTrylockCallExpr(Cond, LVarCtx, Negate)); 1674 if (!Exp) 1675 return; 1676 1677 NamedDecl *FunDecl = dyn_cast_or_null<NamedDecl>(Exp->getCalleeDecl()); 1678 if(!FunDecl || !FunDecl->hasAttrs()) 1679 return; 1680 1681 1682 MutexIDList ExclusiveLocksToAdd; 1683 MutexIDList SharedLocksToAdd; 1684 1685 // If the condition is a call to a Trylock function, then grab the attributes 1686 AttrVec &ArgAttrs = FunDecl->getAttrs(); 1687 for (unsigned i = 0; i < ArgAttrs.size(); ++i) { 1688 Attr *Attr = ArgAttrs[i]; 1689 switch (Attr->getKind()) { 1690 case attr::ExclusiveTrylockFunction: { 1691 ExclusiveTrylockFunctionAttr *A = 1692 cast<ExclusiveTrylockFunctionAttr>(Attr); 1693 getMutexIDs(ExclusiveLocksToAdd, A, Exp, FunDecl, 1694 PredBlock, CurrBlock, A->getSuccessValue(), Negate); 1695 break; 1696 } 1697 case attr::SharedTrylockFunction: { 1698 SharedTrylockFunctionAttr *A = 1699 cast<SharedTrylockFunctionAttr>(Attr); 1700 getMutexIDs(SharedLocksToAdd, A, Exp, FunDecl, 1701 PredBlock, CurrBlock, A->getSuccessValue(), Negate); 1702 break; 1703 } 1704 default: 1705 break; 1706 } 1707 } 1708 1709 // Add and remove locks. 1710 SourceLocation Loc = Exp->getExprLoc(); 1711 for (unsigned i=0,n=ExclusiveLocksToAdd.size(); i<n; ++i) { 1712 addLock(Result, ExclusiveLocksToAdd[i], 1713 LockData(Loc, LK_Exclusive)); 1714 } 1715 for (unsigned i=0,n=SharedLocksToAdd.size(); i<n; ++i) { 1716 addLock(Result, SharedLocksToAdd[i], 1717 LockData(Loc, LK_Shared)); 1718 } 1719 } 1720 1721 1722 /// \brief We use this class to visit different types of expressions in 1723 /// CFGBlocks, and build up the lockset. 1724 /// An expression may cause us to add or remove locks from the lockset, or else 1725 /// output error messages related to missing locks. 1726 /// FIXME: In future, we may be able to not inherit from a visitor. 1727 class BuildLockset : public StmtVisitor<BuildLockset> { 1728 friend class ThreadSafetyAnalyzer; 1729 1730 ThreadSafetyAnalyzer *Analyzer; 1731 FactSet FSet; 1732 LocalVariableMap::Context LVarCtx; 1733 unsigned CtxIndex; 1734 1735 // Helper functions 1736 const ValueDecl *getValueDecl(Expr *Exp); 1737 1738 void warnIfMutexNotHeld(const NamedDecl *D, Expr *Exp, AccessKind AK, 1739 Expr *MutexExp, ProtectedOperationKind POK); 1740 void warnIfMutexHeld(const NamedDecl *D, Expr *Exp, Expr *MutexExp); 1741 1742 void checkAccess(Expr *Exp, AccessKind AK); 1743 void checkDereference(Expr *Exp, AccessKind AK); 1744 void handleCall(Expr *Exp, const NamedDecl *D, VarDecl *VD = 0); 1745 1746 public: 1747 BuildLockset(ThreadSafetyAnalyzer *Anlzr, CFGBlockInfo &Info) 1748 : StmtVisitor<BuildLockset>(), 1749 Analyzer(Anlzr), 1750 FSet(Info.EntrySet), 1751 LVarCtx(Info.EntryContext), 1752 CtxIndex(Info.EntryIndex) 1753 {} 1754 1755 void VisitUnaryOperator(UnaryOperator *UO); 1756 void VisitBinaryOperator(BinaryOperator *BO); 1757 void VisitCastExpr(CastExpr *CE); 1758 void VisitCallExpr(CallExpr *Exp); 1759 void VisitCXXConstructExpr(CXXConstructExpr *Exp); 1760 void VisitDeclStmt(DeclStmt *S); 1761 }; 1762 1763 1764 /// \brief Gets the value decl pointer from DeclRefExprs or MemberExprs 1765 const ValueDecl *BuildLockset::getValueDecl(Expr *Exp) { 1766 if (const DeclRefExpr *DR = dyn_cast<DeclRefExpr>(Exp)) 1767 return DR->getDecl(); 1768 1769 if (const MemberExpr *ME = dyn_cast<MemberExpr>(Exp)) 1770 return ME->getMemberDecl(); 1771 1772 return 0; 1773 } 1774 1775 /// \brief Warn if the LSet does not contain a lock sufficient to protect access 1776 /// of at least the passed in AccessKind. 1777 void BuildLockset::warnIfMutexNotHeld(const NamedDecl *D, Expr *Exp, 1778 AccessKind AK, Expr *MutexExp, 1779 ProtectedOperationKind POK) { 1780 LockKind LK = getLockKindFromAccessKind(AK); 1781 1782 SExpr Mutex(MutexExp, Exp, D); 1783 if (!Mutex.isValid()) { 1784 SExpr::warnInvalidLock(Analyzer->Handler, MutexExp, Exp, D); 1785 return; 1786 } else if (Mutex.shouldIgnore()) { 1787 return; 1788 } 1789 1790 LockData* LDat = FSet.findLockUniv(Analyzer->FactMan, Mutex); 1791 bool NoError = true; 1792 if (!LDat) { 1793 // No exact match found. Look for a partial match. 1794 FactEntry* FEntry = FSet.findPartialMatch(Analyzer->FactMan, Mutex); 1795 if (FEntry) { 1796 // Warn that there's no precise match. 1797 LDat = &FEntry->LDat; 1798 std::string PartMatchStr = FEntry->MutID.toString(); 1799 StringRef PartMatchName(PartMatchStr); 1800 Analyzer->Handler.handleMutexNotHeld(D, POK, Mutex.toString(), LK, 1801 Exp->getExprLoc(), &PartMatchName); 1802 } else { 1803 // Warn that there's no match at all. 1804 Analyzer->Handler.handleMutexNotHeld(D, POK, Mutex.toString(), LK, 1805 Exp->getExprLoc()); 1806 } 1807 NoError = false; 1808 } 1809 // Make sure the mutex we found is the right kind. 1810 if (NoError && LDat && !LDat->isAtLeast(LK)) 1811 Analyzer->Handler.handleMutexNotHeld(D, POK, Mutex.toString(), LK, 1812 Exp->getExprLoc()); 1813 } 1814 1815 /// \brief Warn if the LSet contains the given lock. 1816 void BuildLockset::warnIfMutexHeld(const NamedDecl *D, Expr* Exp, 1817 Expr *MutexExp) { 1818 SExpr Mutex(MutexExp, Exp, D); 1819 if (!Mutex.isValid()) { 1820 SExpr::warnInvalidLock(Analyzer->Handler, MutexExp, Exp, D); 1821 return; 1822 } 1823 1824 LockData* LDat = FSet.findLock(Analyzer->FactMan, Mutex); 1825 if (LDat) { 1826 std::string DeclName = D->getNameAsString(); 1827 StringRef DeclNameSR (DeclName); 1828 Analyzer->Handler.handleFunExcludesLock(DeclNameSR, Mutex.toString(), 1829 Exp->getExprLoc()); 1830 } 1831 } 1832 1833 1834 /// \brief This method identifies variable dereferences and checks pt_guarded_by 1835 /// and pt_guarded_var annotations. Note that we only check these annotations 1836 /// at the time a pointer is dereferenced. 1837 /// FIXME: We need to check for other types of pointer dereferences 1838 /// (e.g. [], ->) and deal with them here. 1839 /// \param Exp An expression that has been read or written. 1840 void BuildLockset::checkDereference(Expr *Exp, AccessKind AK) { 1841 UnaryOperator *UO = dyn_cast<UnaryOperator>(Exp); 1842 if (!UO || UO->getOpcode() != clang::UO_Deref) 1843 return; 1844 Exp = UO->getSubExpr()->IgnoreParenCasts(); 1845 1846 const ValueDecl *D = getValueDecl(Exp); 1847 if(!D || !D->hasAttrs()) 1848 return; 1849 1850 if (D->getAttr<PtGuardedVarAttr>() && FSet.isEmpty()) 1851 Analyzer->Handler.handleNoMutexHeld(D, POK_VarDereference, AK, 1852 Exp->getExprLoc()); 1853 1854 const AttrVec &ArgAttrs = D->getAttrs(); 1855 for(unsigned i = 0, Size = ArgAttrs.size(); i < Size; ++i) 1856 if (PtGuardedByAttr *PGBAttr = dyn_cast<PtGuardedByAttr>(ArgAttrs[i])) 1857 warnIfMutexNotHeld(D, Exp, AK, PGBAttr->getArg(), POK_VarDereference); 1858 } 1859 1860 /// \brief Checks guarded_by and guarded_var attributes. 1861 /// Whenever we identify an access (read or write) of a DeclRefExpr or 1862 /// MemberExpr, we need to check whether there are any guarded_by or 1863 /// guarded_var attributes, and make sure we hold the appropriate mutexes. 1864 void BuildLockset::checkAccess(Expr *Exp, AccessKind AK) { 1865 const ValueDecl *D = getValueDecl(Exp); 1866 if(!D || !D->hasAttrs()) 1867 return; 1868 1869 if (D->getAttr<GuardedVarAttr>() && FSet.isEmpty()) 1870 Analyzer->Handler.handleNoMutexHeld(D, POK_VarAccess, AK, 1871 Exp->getExprLoc()); 1872 1873 const AttrVec &ArgAttrs = D->getAttrs(); 1874 for(unsigned i = 0, Size = ArgAttrs.size(); i < Size; ++i) 1875 if (GuardedByAttr *GBAttr = dyn_cast<GuardedByAttr>(ArgAttrs[i])) 1876 warnIfMutexNotHeld(D, Exp, AK, GBAttr->getArg(), POK_VarAccess); 1877 } 1878 1879 /// \brief Process a function call, method call, constructor call, 1880 /// or destructor call. This involves looking at the attributes on the 1881 /// corresponding function/method/constructor/destructor, issuing warnings, 1882 /// and updating the locksets accordingly. 1883 /// 1884 /// FIXME: For classes annotated with one of the guarded annotations, we need 1885 /// to treat const method calls as reads and non-const method calls as writes, 1886 /// and check that the appropriate locks are held. Non-const method calls with 1887 /// the same signature as const method calls can be also treated as reads. 1888 /// 1889 void BuildLockset::handleCall(Expr *Exp, const NamedDecl *D, VarDecl *VD) { 1890 const AttrVec &ArgAttrs = D->getAttrs(); 1891 MutexIDList ExclusiveLocksToAdd; 1892 MutexIDList SharedLocksToAdd; 1893 MutexIDList LocksToRemove; 1894 1895 for(unsigned i = 0; i < ArgAttrs.size(); ++i) { 1896 Attr *At = const_cast<Attr*>(ArgAttrs[i]); 1897 switch (At->getKind()) { 1898 // When we encounter an exclusive lock function, we need to add the lock 1899 // to our lockset with kind exclusive. 1900 case attr::ExclusiveLockFunction: { 1901 ExclusiveLockFunctionAttr *A = cast<ExclusiveLockFunctionAttr>(At); 1902 Analyzer->getMutexIDs(ExclusiveLocksToAdd, A, Exp, D, VD); 1903 break; 1904 } 1905 1906 // When we encounter a shared lock function, we need to add the lock 1907 // to our lockset with kind shared. 1908 case attr::SharedLockFunction: { 1909 SharedLockFunctionAttr *A = cast<SharedLockFunctionAttr>(At); 1910 Analyzer->getMutexIDs(SharedLocksToAdd, A, Exp, D, VD); 1911 break; 1912 } 1913 1914 // When we encounter an unlock function, we need to remove unlocked 1915 // mutexes from the lockset, and flag a warning if they are not there. 1916 case attr::UnlockFunction: { 1917 UnlockFunctionAttr *A = cast<UnlockFunctionAttr>(At); 1918 Analyzer->getMutexIDs(LocksToRemove, A, Exp, D, VD); 1919 break; 1920 } 1921 1922 case attr::ExclusiveLocksRequired: { 1923 ExclusiveLocksRequiredAttr *A = cast<ExclusiveLocksRequiredAttr>(At); 1924 1925 for (ExclusiveLocksRequiredAttr::args_iterator 1926 I = A->args_begin(), E = A->args_end(); I != E; ++I) 1927 warnIfMutexNotHeld(D, Exp, AK_Written, *I, POK_FunctionCall); 1928 break; 1929 } 1930 1931 case attr::SharedLocksRequired: { 1932 SharedLocksRequiredAttr *A = cast<SharedLocksRequiredAttr>(At); 1933 1934 for (SharedLocksRequiredAttr::args_iterator I = A->args_begin(), 1935 E = A->args_end(); I != E; ++I) 1936 warnIfMutexNotHeld(D, Exp, AK_Read, *I, POK_FunctionCall); 1937 break; 1938 } 1939 1940 case attr::LocksExcluded: { 1941 LocksExcludedAttr *A = cast<LocksExcludedAttr>(At); 1942 1943 for (LocksExcludedAttr::args_iterator I = A->args_begin(), 1944 E = A->args_end(); I != E; ++I) { 1945 warnIfMutexHeld(D, Exp, *I); 1946 } 1947 break; 1948 } 1949 1950 // Ignore other (non thread-safety) attributes 1951 default: 1952 break; 1953 } 1954 } 1955 1956 // Figure out if we're calling the constructor of scoped lockable class 1957 bool isScopedVar = false; 1958 if (VD) { 1959 if (const CXXConstructorDecl *CD = dyn_cast<const CXXConstructorDecl>(D)) { 1960 const CXXRecordDecl* PD = CD->getParent(); 1961 if (PD && PD->getAttr<ScopedLockableAttr>()) 1962 isScopedVar = true; 1963 } 1964 } 1965 1966 // Add locks. 1967 SourceLocation Loc = Exp->getExprLoc(); 1968 for (unsigned i=0,n=ExclusiveLocksToAdd.size(); i<n; ++i) { 1969 Analyzer->addLock(FSet, ExclusiveLocksToAdd[i], 1970 LockData(Loc, LK_Exclusive, isScopedVar)); 1971 } 1972 for (unsigned i=0,n=SharedLocksToAdd.size(); i<n; ++i) { 1973 Analyzer->addLock(FSet, SharedLocksToAdd[i], 1974 LockData(Loc, LK_Shared, isScopedVar)); 1975 } 1976 1977 // Add the managing object as a dummy mutex, mapped to the underlying mutex. 1978 // FIXME -- this doesn't work if we acquire multiple locks. 1979 if (isScopedVar) { 1980 SourceLocation MLoc = VD->getLocation(); 1981 DeclRefExpr DRE(VD, false, VD->getType(), VK_LValue, VD->getLocation()); 1982 SExpr SMutex(&DRE, 0, 0); 1983 1984 for (unsigned i=0,n=ExclusiveLocksToAdd.size(); i<n; ++i) { 1985 Analyzer->addLock(FSet, SMutex, LockData(MLoc, LK_Exclusive, 1986 ExclusiveLocksToAdd[i])); 1987 } 1988 for (unsigned i=0,n=SharedLocksToAdd.size(); i<n; ++i) { 1989 Analyzer->addLock(FSet, SMutex, LockData(MLoc, LK_Shared, 1990 SharedLocksToAdd[i])); 1991 } 1992 } 1993 1994 // Remove locks. 1995 // FIXME -- should only fully remove if the attribute refers to 'this'. 1996 bool Dtor = isa<CXXDestructorDecl>(D); 1997 for (unsigned i=0,n=LocksToRemove.size(); i<n; ++i) { 1998 Analyzer->removeLock(FSet, LocksToRemove[i], Loc, Dtor); 1999 } 2000 } 2001 2002 2003 /// \brief For unary operations which read and write a variable, we need to 2004 /// check whether we hold any required mutexes. Reads are checked in 2005 /// VisitCastExpr. 2006 void BuildLockset::VisitUnaryOperator(UnaryOperator *UO) { 2007 switch (UO->getOpcode()) { 2008 case clang::UO_PostDec: 2009 case clang::UO_PostInc: 2010 case clang::UO_PreDec: 2011 case clang::UO_PreInc: { 2012 Expr *SubExp = UO->getSubExpr()->IgnoreParenCasts(); 2013 checkAccess(SubExp, AK_Written); 2014 checkDereference(SubExp, AK_Written); 2015 break; 2016 } 2017 default: 2018 break; 2019 } 2020 } 2021 2022 /// For binary operations which assign to a variable (writes), we need to check 2023 /// whether we hold any required mutexes. 2024 /// FIXME: Deal with non-primitive types. 2025 void BuildLockset::VisitBinaryOperator(BinaryOperator *BO) { 2026 if (!BO->isAssignmentOp()) 2027 return; 2028 2029 // adjust the context 2030 LVarCtx = Analyzer->LocalVarMap.getNextContext(CtxIndex, BO, LVarCtx); 2031 2032 Expr *LHSExp = BO->getLHS()->IgnoreParenCasts(); 2033 checkAccess(LHSExp, AK_Written); 2034 checkDereference(LHSExp, AK_Written); 2035 } 2036 2037 /// Whenever we do an LValue to Rvalue cast, we are reading a variable and 2038 /// need to ensure we hold any required mutexes. 2039 /// FIXME: Deal with non-primitive types. 2040 void BuildLockset::VisitCastExpr(CastExpr *CE) { 2041 if (CE->getCastKind() != CK_LValueToRValue) 2042 return; 2043 Expr *SubExp = CE->getSubExpr()->IgnoreParenCasts(); 2044 checkAccess(SubExp, AK_Read); 2045 checkDereference(SubExp, AK_Read); 2046 } 2047 2048 2049 void BuildLockset::VisitCallExpr(CallExpr *Exp) { 2050 NamedDecl *D = dyn_cast_or_null<NamedDecl>(Exp->getCalleeDecl()); 2051 if(!D || !D->hasAttrs()) 2052 return; 2053 handleCall(Exp, D); 2054 } 2055 2056 void BuildLockset::VisitCXXConstructExpr(CXXConstructExpr *Exp) { 2057 // FIXME -- only handles constructors in DeclStmt below. 2058 } 2059 2060 void BuildLockset::VisitDeclStmt(DeclStmt *S) { 2061 // adjust the context 2062 LVarCtx = Analyzer->LocalVarMap.getNextContext(CtxIndex, S, LVarCtx); 2063 2064 DeclGroupRef DGrp = S->getDeclGroup(); 2065 for (DeclGroupRef::iterator I = DGrp.begin(), E = DGrp.end(); I != E; ++I) { 2066 Decl *D = *I; 2067 if (VarDecl *VD = dyn_cast_or_null<VarDecl>(D)) { 2068 Expr *E = VD->getInit(); 2069 // handle constructors that involve temporaries 2070 if (ExprWithCleanups *EWC = dyn_cast_or_null<ExprWithCleanups>(E)) 2071 E = EWC->getSubExpr(); 2072 2073 if (CXXConstructExpr *CE = dyn_cast_or_null<CXXConstructExpr>(E)) { 2074 NamedDecl *CtorD = dyn_cast_or_null<NamedDecl>(CE->getConstructor()); 2075 if (!CtorD || !CtorD->hasAttrs()) 2076 return; 2077 handleCall(CE, CtorD, VD); 2078 } 2079 } 2080 } 2081 } 2082 2083 2084 2085 /// \brief Compute the intersection of two locksets and issue warnings for any 2086 /// locks in the symmetric difference. 2087 /// 2088 /// This function is used at a merge point in the CFG when comparing the lockset 2089 /// of each branch being merged. For example, given the following sequence: 2090 /// A; if () then B; else C; D; we need to check that the lockset after B and C 2091 /// are the same. In the event of a difference, we use the intersection of these 2092 /// two locksets at the start of D. 2093 /// 2094 /// \param FSet1 The first lockset. 2095 /// \param FSet2 The second lockset. 2096 /// \param JoinLoc The location of the join point for error reporting 2097 /// \param LEK1 The error message to report if a mutex is missing from LSet1 2098 /// \param LEK2 The error message to report if a mutex is missing from Lset2 2099 void ThreadSafetyAnalyzer::intersectAndWarn(FactSet &FSet1, 2100 const FactSet &FSet2, 2101 SourceLocation JoinLoc, 2102 LockErrorKind LEK1, 2103 LockErrorKind LEK2, 2104 bool Modify) { 2105 FactSet FSet1Orig = FSet1; 2106 2107 for (FactSet::const_iterator I = FSet2.begin(), E = FSet2.end(); 2108 I != E; ++I) { 2109 const SExpr &FSet2Mutex = FactMan[*I].MutID; 2110 const LockData &LDat2 = FactMan[*I].LDat; 2111 2112 if (const LockData *LDat1 = FSet1.findLock(FactMan, FSet2Mutex)) { 2113 if (LDat1->LKind != LDat2.LKind) { 2114 Handler.handleExclusiveAndShared(FSet2Mutex.toString(), 2115 LDat2.AcquireLoc, 2116 LDat1->AcquireLoc); 2117 if (Modify && LDat1->LKind != LK_Exclusive) { 2118 FSet1.removeLock(FactMan, FSet2Mutex); 2119 FSet1.addLock(FactMan, FSet2Mutex, LDat2); 2120 } 2121 } 2122 } else { 2123 if (LDat2.UnderlyingMutex.isValid()) { 2124 if (FSet2.findLock(FactMan, LDat2.UnderlyingMutex)) { 2125 // If this is a scoped lock that manages another mutex, and if the 2126 // underlying mutex is still held, then warn about the underlying 2127 // mutex. 2128 Handler.handleMutexHeldEndOfScope(LDat2.UnderlyingMutex.toString(), 2129 LDat2.AcquireLoc, 2130 JoinLoc, LEK1); 2131 } 2132 } 2133 else if (!LDat2.Managed && !FSet2Mutex.isUniversal()) 2134 Handler.handleMutexHeldEndOfScope(FSet2Mutex.toString(), 2135 LDat2.AcquireLoc, 2136 JoinLoc, LEK1); 2137 } 2138 } 2139 2140 for (FactSet::const_iterator I = FSet1.begin(), E = FSet1.end(); 2141 I != E; ++I) { 2142 const SExpr &FSet1Mutex = FactMan[*I].MutID; 2143 const LockData &LDat1 = FactMan[*I].LDat; 2144 2145 if (!FSet2.findLock(FactMan, FSet1Mutex)) { 2146 if (LDat1.UnderlyingMutex.isValid()) { 2147 if (FSet1Orig.findLock(FactMan, LDat1.UnderlyingMutex)) { 2148 // If this is a scoped lock that manages another mutex, and if the 2149 // underlying mutex is still held, then warn about the underlying 2150 // mutex. 2151 Handler.handleMutexHeldEndOfScope(LDat1.UnderlyingMutex.toString(), 2152 LDat1.AcquireLoc, 2153 JoinLoc, LEK1); 2154 } 2155 } 2156 else if (!LDat1.Managed && !FSet1Mutex.isUniversal()) 2157 Handler.handleMutexHeldEndOfScope(FSet1Mutex.toString(), 2158 LDat1.AcquireLoc, 2159 JoinLoc, LEK2); 2160 if (Modify) 2161 FSet1.removeLock(FactMan, FSet1Mutex); 2162 } 2163 } 2164 } 2165 2166 2167 2168 /// \brief Check a function's CFG for thread-safety violations. 2169 /// 2170 /// We traverse the blocks in the CFG, compute the set of mutexes that are held 2171 /// at the end of each block, and issue warnings for thread safety violations. 2172 /// Each block in the CFG is traversed exactly once. 2173 void ThreadSafetyAnalyzer::runAnalysis(AnalysisDeclContext &AC) { 2174 CFG *CFGraph = AC.getCFG(); 2175 if (!CFGraph) return; 2176 const NamedDecl *D = dyn_cast_or_null<NamedDecl>(AC.getDecl()); 2177 2178 // AC.dumpCFG(true); 2179 2180 if (!D) 2181 return; // Ignore anonymous functions for now. 2182 if (D->getAttr<NoThreadSafetyAnalysisAttr>()) 2183 return; 2184 // FIXME: Do something a bit more intelligent inside constructor and 2185 // destructor code. Constructors and destructors must assume unique access 2186 // to 'this', so checks on member variable access is disabled, but we should 2187 // still enable checks on other objects. 2188 if (isa<CXXConstructorDecl>(D)) 2189 return; // Don't check inside constructors. 2190 if (isa<CXXDestructorDecl>(D)) 2191 return; // Don't check inside destructors. 2192 2193 BlockInfo.resize(CFGraph->getNumBlockIDs(), 2194 CFGBlockInfo::getEmptyBlockInfo(LocalVarMap)); 2195 2196 // We need to explore the CFG via a "topological" ordering. 2197 // That way, we will be guaranteed to have information about required 2198 // predecessor locksets when exploring a new block. 2199 PostOrderCFGView *SortedGraph = AC.getAnalysis<PostOrderCFGView>(); 2200 PostOrderCFGView::CFGBlockSet VisitedBlocks(CFGraph); 2201 2202 // Mark entry block as reachable 2203 BlockInfo[CFGraph->getEntry().getBlockID()].Reachable = true; 2204 2205 // Compute SSA names for local variables 2206 LocalVarMap.traverseCFG(CFGraph, SortedGraph, BlockInfo); 2207 2208 // Fill in source locations for all CFGBlocks. 2209 findBlockLocations(CFGraph, SortedGraph, BlockInfo); 2210 2211 // Add locks from exclusive_locks_required and shared_locks_required 2212 // to initial lockset. Also turn off checking for lock and unlock functions. 2213 // FIXME: is there a more intelligent way to check lock/unlock functions? 2214 if (!SortedGraph->empty() && D->hasAttrs()) { 2215 const CFGBlock *FirstBlock = *SortedGraph->begin(); 2216 FactSet &InitialLockset = BlockInfo[FirstBlock->getBlockID()].EntrySet; 2217 const AttrVec &ArgAttrs = D->getAttrs(); 2218 2219 MutexIDList ExclusiveLocksToAdd; 2220 MutexIDList SharedLocksToAdd; 2221 2222 SourceLocation Loc = D->getLocation(); 2223 for (unsigned i = 0; i < ArgAttrs.size(); ++i) { 2224 Attr *Attr = ArgAttrs[i]; 2225 Loc = Attr->getLocation(); 2226 if (ExclusiveLocksRequiredAttr *A 2227 = dyn_cast<ExclusiveLocksRequiredAttr>(Attr)) { 2228 getMutexIDs(ExclusiveLocksToAdd, A, (Expr*) 0, D); 2229 } else if (SharedLocksRequiredAttr *A 2230 = dyn_cast<SharedLocksRequiredAttr>(Attr)) { 2231 getMutexIDs(SharedLocksToAdd, A, (Expr*) 0, D); 2232 } else if (isa<UnlockFunctionAttr>(Attr)) { 2233 // Don't try to check unlock functions for now 2234 return; 2235 } else if (isa<ExclusiveLockFunctionAttr>(Attr)) { 2236 // Don't try to check lock functions for now 2237 return; 2238 } else if (isa<SharedLockFunctionAttr>(Attr)) { 2239 // Don't try to check lock functions for now 2240 return; 2241 } else if (isa<ExclusiveTrylockFunctionAttr>(Attr)) { 2242 // Don't try to check trylock functions for now 2243 return; 2244 } else if (isa<SharedTrylockFunctionAttr>(Attr)) { 2245 // Don't try to check trylock functions for now 2246 return; 2247 } 2248 } 2249 2250 // FIXME -- Loc can be wrong here. 2251 for (unsigned i=0,n=ExclusiveLocksToAdd.size(); i<n; ++i) { 2252 addLock(InitialLockset, ExclusiveLocksToAdd[i], 2253 LockData(Loc, LK_Exclusive)); 2254 } 2255 for (unsigned i=0,n=SharedLocksToAdd.size(); i<n; ++i) { 2256 addLock(InitialLockset, SharedLocksToAdd[i], 2257 LockData(Loc, LK_Shared)); 2258 } 2259 } 2260 2261 for (PostOrderCFGView::iterator I = SortedGraph->begin(), 2262 E = SortedGraph->end(); I!= E; ++I) { 2263 const CFGBlock *CurrBlock = *I; 2264 int CurrBlockID = CurrBlock->getBlockID(); 2265 CFGBlockInfo *CurrBlockInfo = &BlockInfo[CurrBlockID]; 2266 2267 // Use the default initial lockset in case there are no predecessors. 2268 VisitedBlocks.insert(CurrBlock); 2269 2270 // Iterate through the predecessor blocks and warn if the lockset for all 2271 // predecessors is not the same. We take the entry lockset of the current 2272 // block to be the intersection of all previous locksets. 2273 // FIXME: By keeping the intersection, we may output more errors in future 2274 // for a lock which is not in the intersection, but was in the union. We 2275 // may want to also keep the union in future. As an example, let's say 2276 // the intersection contains Mutex L, and the union contains L and M. 2277 // Later we unlock M. At this point, we would output an error because we 2278 // never locked M; although the real error is probably that we forgot to 2279 // lock M on all code paths. Conversely, let's say that later we lock M. 2280 // In this case, we should compare against the intersection instead of the 2281 // union because the real error is probably that we forgot to unlock M on 2282 // all code paths. 2283 bool LocksetInitialized = false; 2284 llvm::SmallVector<CFGBlock*, 8> SpecialBlocks; 2285 for (CFGBlock::const_pred_iterator PI = CurrBlock->pred_begin(), 2286 PE = CurrBlock->pred_end(); PI != PE; ++PI) { 2287 2288 // if *PI -> CurrBlock is a back edge 2289 if (*PI == 0 || !VisitedBlocks.alreadySet(*PI)) 2290 continue; 2291 2292 int PrevBlockID = (*PI)->getBlockID(); 2293 CFGBlockInfo *PrevBlockInfo = &BlockInfo[PrevBlockID]; 2294 2295 // Ignore edges from blocks that can't return. 2296 if ((*PI)->hasNoReturnElement() || !PrevBlockInfo->Reachable) 2297 continue; 2298 2299 // Okay, we can reach this block from the entry. 2300 CurrBlockInfo->Reachable = true; 2301 2302 // If the previous block ended in a 'continue' or 'break' statement, then 2303 // a difference in locksets is probably due to a bug in that block, rather 2304 // than in some other predecessor. In that case, keep the other 2305 // predecessor's lockset. 2306 if (const Stmt *Terminator = (*PI)->getTerminator()) { 2307 if (isa<ContinueStmt>(Terminator) || isa<BreakStmt>(Terminator)) { 2308 SpecialBlocks.push_back(*PI); 2309 continue; 2310 } 2311 } 2312 2313 2314 FactSet PrevLockset; 2315 getEdgeLockset(PrevLockset, PrevBlockInfo->ExitSet, *PI, CurrBlock); 2316 2317 if (!LocksetInitialized) { 2318 CurrBlockInfo->EntrySet = PrevLockset; 2319 LocksetInitialized = true; 2320 } else { 2321 intersectAndWarn(CurrBlockInfo->EntrySet, PrevLockset, 2322 CurrBlockInfo->EntryLoc, 2323 LEK_LockedSomePredecessors); 2324 } 2325 } 2326 2327 // Skip rest of block if it's not reachable. 2328 if (!CurrBlockInfo->Reachable) 2329 continue; 2330 2331 // Process continue and break blocks. Assume that the lockset for the 2332 // resulting block is unaffected by any discrepancies in them. 2333 for (unsigned SpecialI = 0, SpecialN = SpecialBlocks.size(); 2334 SpecialI < SpecialN; ++SpecialI) { 2335 CFGBlock *PrevBlock = SpecialBlocks[SpecialI]; 2336 int PrevBlockID = PrevBlock->getBlockID(); 2337 CFGBlockInfo *PrevBlockInfo = &BlockInfo[PrevBlockID]; 2338 2339 if (!LocksetInitialized) { 2340 CurrBlockInfo->EntrySet = PrevBlockInfo->ExitSet; 2341 LocksetInitialized = true; 2342 } else { 2343 // Determine whether this edge is a loop terminator for diagnostic 2344 // purposes. FIXME: A 'break' statement might be a loop terminator, but 2345 // it might also be part of a switch. Also, a subsequent destructor 2346 // might add to the lockset, in which case the real issue might be a 2347 // double lock on the other path. 2348 const Stmt *Terminator = PrevBlock->getTerminator(); 2349 bool IsLoop = Terminator && isa<ContinueStmt>(Terminator); 2350 2351 FactSet PrevLockset; 2352 getEdgeLockset(PrevLockset, PrevBlockInfo->ExitSet, 2353 PrevBlock, CurrBlock); 2354 2355 // Do not update EntrySet. 2356 intersectAndWarn(CurrBlockInfo->EntrySet, PrevLockset, 2357 PrevBlockInfo->ExitLoc, 2358 IsLoop ? LEK_LockedSomeLoopIterations 2359 : LEK_LockedSomePredecessors, 2360 false); 2361 } 2362 } 2363 2364 BuildLockset LocksetBuilder(this, *CurrBlockInfo); 2365 2366 // Visit all the statements in the basic block. 2367 for (CFGBlock::const_iterator BI = CurrBlock->begin(), 2368 BE = CurrBlock->end(); BI != BE; ++BI) { 2369 switch (BI->getKind()) { 2370 case CFGElement::Statement: { 2371 const CFGStmt *CS = cast<CFGStmt>(&*BI); 2372 LocksetBuilder.Visit(const_cast<Stmt*>(CS->getStmt())); 2373 break; 2374 } 2375 // Ignore BaseDtor, MemberDtor, and TemporaryDtor for now. 2376 case CFGElement::AutomaticObjectDtor: { 2377 const CFGAutomaticObjDtor *AD = cast<CFGAutomaticObjDtor>(&*BI); 2378 CXXDestructorDecl *DD = const_cast<CXXDestructorDecl*>( 2379 AD->getDestructorDecl(AC.getASTContext())); 2380 if (!DD->hasAttrs()) 2381 break; 2382 2383 // Create a dummy expression, 2384 VarDecl *VD = const_cast<VarDecl*>(AD->getVarDecl()); 2385 DeclRefExpr DRE(VD, false, VD->getType(), VK_LValue, 2386 AD->getTriggerStmt()->getLocEnd()); 2387 LocksetBuilder.handleCall(&DRE, DD); 2388 break; 2389 } 2390 default: 2391 break; 2392 } 2393 } 2394 CurrBlockInfo->ExitSet = LocksetBuilder.FSet; 2395 2396 // For every back edge from CurrBlock (the end of the loop) to another block 2397 // (FirstLoopBlock) we need to check that the Lockset of Block is equal to 2398 // the one held at the beginning of FirstLoopBlock. We can look up the 2399 // Lockset held at the beginning of FirstLoopBlock in the EntryLockSets map. 2400 for (CFGBlock::const_succ_iterator SI = CurrBlock->succ_begin(), 2401 SE = CurrBlock->succ_end(); SI != SE; ++SI) { 2402 2403 // if CurrBlock -> *SI is *not* a back edge 2404 if (*SI == 0 || !VisitedBlocks.alreadySet(*SI)) 2405 continue; 2406 2407 CFGBlock *FirstLoopBlock = *SI; 2408 CFGBlockInfo *PreLoop = &BlockInfo[FirstLoopBlock->getBlockID()]; 2409 CFGBlockInfo *LoopEnd = &BlockInfo[CurrBlockID]; 2410 intersectAndWarn(LoopEnd->ExitSet, PreLoop->EntrySet, 2411 PreLoop->EntryLoc, 2412 LEK_LockedSomeLoopIterations, 2413 false); 2414 } 2415 } 2416 2417 CFGBlockInfo *Initial = &BlockInfo[CFGraph->getEntry().getBlockID()]; 2418 CFGBlockInfo *Final = &BlockInfo[CFGraph->getExit().getBlockID()]; 2419 2420 // Skip the final check if the exit block is unreachable. 2421 if (!Final->Reachable) 2422 return; 2423 2424 // FIXME: Should we call this function for all blocks which exit the function? 2425 intersectAndWarn(Initial->EntrySet, Final->ExitSet, 2426 Final->ExitLoc, 2427 LEK_LockedAtEndOfFunction, 2428 LEK_NotLockedAtEndOfFunction, 2429 false); 2430 } 2431 2432 } // end anonymous namespace 2433 2434 2435 namespace clang { 2436 namespace thread_safety { 2437 2438 /// \brief Check a function's CFG for thread-safety violations. 2439 /// 2440 /// We traverse the blocks in the CFG, compute the set of mutexes that are held 2441 /// at the end of each block, and issue warnings for thread safety violations. 2442 /// Each block in the CFG is traversed exactly once. 2443 void runThreadSafetyAnalysis(AnalysisDeclContext &AC, 2444 ThreadSafetyHandler &Handler) { 2445 ThreadSafetyAnalyzer Analyzer(Handler); 2446 Analyzer.runAnalysis(AC); 2447 } 2448 2449 /// \brief Helper function that returns a LockKind required for the given level 2450 /// of access. 2451 LockKind getLockKindFromAccessKind(AccessKind AK) { 2452 switch (AK) { 2453 case AK_Read : 2454 return LK_Shared; 2455 case AK_Written : 2456 return LK_Exclusive; 2457 } 2458 llvm_unreachable("Unknown AccessKind"); 2459 } 2460 2461 }} // end namespace clang::thread_safety 2462