1 //===- ThreadSafety.cpp ----------------------------------------*- C++ --*-===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // A intra-procedural analysis for thread safety (e.g. deadlocks and race 11 // conditions), based off of an annotation system. 12 // 13 // See http://clang.llvm.org/docs/ThreadSafetyAnalysis.html 14 // for more information. 15 // 16 //===----------------------------------------------------------------------===// 17 18 #include "clang/AST/Attr.h" 19 #include "clang/AST/DeclCXX.h" 20 #include "clang/AST/ExprCXX.h" 21 #include "clang/AST/StmtCXX.h" 22 #include "clang/AST/StmtVisitor.h" 23 #include "clang/Analysis/Analyses/PostOrderCFGView.h" 24 #include "clang/Analysis/Analyses/ThreadSafety.h" 25 #include "clang/Analysis/Analyses/ThreadSafetyTIL.h" 26 #include "clang/Analysis/Analyses/ThreadSafetyTraverse.h" 27 #include "clang/Analysis/Analyses/ThreadSafetyCommon.h" 28 #include "clang/Analysis/AnalysisContext.h" 29 #include "clang/Analysis/CFG.h" 30 #include "clang/Analysis/CFGStmtMap.h" 31 #include "clang/Basic/OperatorKinds.h" 32 #include "clang/Basic/SourceLocation.h" 33 #include "clang/Basic/SourceManager.h" 34 #include "llvm/ADT/BitVector.h" 35 #include "llvm/ADT/FoldingSet.h" 36 #include "llvm/ADT/ImmutableMap.h" 37 #include "llvm/ADT/PostOrderIterator.h" 38 #include "llvm/ADT/SmallVector.h" 39 #include "llvm/ADT/StringRef.h" 40 #include "llvm/Support/raw_ostream.h" 41 #include <algorithm> 42 #include <utility> 43 #include <vector> 44 45 using namespace clang; 46 using namespace thread_safety; 47 48 // Key method definition 49 ThreadSafetyHandler::~ThreadSafetyHandler() {} 50 51 namespace { 52 53 /// SExpr implements a simple expression language that is used to store, 54 /// compare, and pretty-print C++ expressions. Unlike a clang Expr, a SExpr 55 /// does not capture surface syntax, and it does not distinguish between 56 /// C++ concepts, like pointers and references, that have no real semantic 57 /// differences. This simplicity allows SExprs to be meaningfully compared, 58 /// e.g. 59 /// (x) = x 60 /// (*this).foo = this->foo 61 /// *&a = a 62 /// 63 /// Thread-safety analysis works by comparing lock expressions. Within the 64 /// body of a function, an expression such as "x->foo->bar.mu" will resolve to 65 /// a particular mutex object at run-time. Subsequent occurrences of the same 66 /// expression (where "same" means syntactic equality) will refer to the same 67 /// run-time object if three conditions hold: 68 /// (1) Local variables in the expression, such as "x" have not changed. 69 /// (2) Values on the heap that affect the expression have not changed. 70 /// (3) The expression involves only pure function calls. 71 /// 72 /// The current implementation assumes, but does not verify, that multiple uses 73 /// of the same lock expression satisfies these criteria. 74 class SExpr { 75 private: 76 enum ExprOp { 77 EOP_Nop, ///< No-op 78 EOP_Wildcard, ///< Matches anything. 79 EOP_Universal, ///< Universal lock. 80 EOP_This, ///< This keyword. 81 EOP_NVar, ///< Named variable. 82 EOP_LVar, ///< Local variable. 83 EOP_Dot, ///< Field access 84 EOP_Call, ///< Function call 85 EOP_MCall, ///< Method call 86 EOP_Index, ///< Array index 87 EOP_Unary, ///< Unary operation 88 EOP_Binary, ///< Binary operation 89 EOP_Unknown ///< Catchall for everything else 90 }; 91 92 93 class SExprNode { 94 private: 95 unsigned char Op; ///< Opcode of the root node 96 unsigned char Flags; ///< Additional opcode-specific data 97 unsigned short Sz; ///< Number of child nodes 98 const void* Data; ///< Additional opcode-specific data 99 100 public: 101 SExprNode(ExprOp O, unsigned F, const void* D) 102 : Op(static_cast<unsigned char>(O)), 103 Flags(static_cast<unsigned char>(F)), Sz(1), Data(D) 104 { } 105 106 unsigned size() const { return Sz; } 107 void setSize(unsigned S) { Sz = S; } 108 109 ExprOp kind() const { return static_cast<ExprOp>(Op); } 110 111 const NamedDecl* getNamedDecl() const { 112 assert(Op == EOP_NVar || Op == EOP_LVar || Op == EOP_Dot); 113 return reinterpret_cast<const NamedDecl*>(Data); 114 } 115 116 const NamedDecl* getFunctionDecl() const { 117 assert(Op == EOP_Call || Op == EOP_MCall); 118 return reinterpret_cast<const NamedDecl*>(Data); 119 } 120 121 bool isArrow() const { return Op == EOP_Dot && Flags == 1; } 122 void setArrow(bool A) { Flags = A ? 1 : 0; } 123 124 unsigned arity() const { 125 switch (Op) { 126 case EOP_Nop: return 0; 127 case EOP_Wildcard: return 0; 128 case EOP_Universal: return 0; 129 case EOP_NVar: return 0; 130 case EOP_LVar: return 0; 131 case EOP_This: return 0; 132 case EOP_Dot: return 1; 133 case EOP_Call: return Flags+1; // First arg is function. 134 case EOP_MCall: return Flags+1; // First arg is implicit obj. 135 case EOP_Index: return 2; 136 case EOP_Unary: return 1; 137 case EOP_Binary: return 2; 138 case EOP_Unknown: return Flags; 139 } 140 return 0; 141 } 142 143 bool operator==(const SExprNode& Other) const { 144 // Ignore flags and size -- they don't matter. 145 return (Op == Other.Op && 146 Data == Other.Data); 147 } 148 149 bool operator!=(const SExprNode& Other) const { 150 return !(*this == Other); 151 } 152 153 bool matches(const SExprNode& Other) const { 154 return (*this == Other) || 155 (Op == EOP_Wildcard) || 156 (Other.Op == EOP_Wildcard); 157 } 158 }; 159 160 161 /// \brief Encapsulates the lexical context of a function call. The lexical 162 /// context includes the arguments to the call, including the implicit object 163 /// argument. When an attribute containing a mutex expression is attached to 164 /// a method, the expression may refer to formal parameters of the method. 165 /// Actual arguments must be substituted for formal parameters to derive 166 /// the appropriate mutex expression in the lexical context where the function 167 /// is called. PrevCtx holds the context in which the arguments themselves 168 /// should be evaluated; multiple calling contexts can be chained together 169 /// by the lock_returned attribute. 170 struct CallingContext { 171 const NamedDecl* AttrDecl; // The decl to which the attribute is attached. 172 const Expr* SelfArg; // Implicit object argument -- e.g. 'this' 173 bool SelfArrow; // is Self referred to with -> or .? 174 unsigned NumArgs; // Number of funArgs 175 const Expr* const* FunArgs; // Function arguments 176 CallingContext* PrevCtx; // The previous context; or 0 if none. 177 178 CallingContext(const NamedDecl *D) 179 : AttrDecl(D), SelfArg(0), SelfArrow(false), NumArgs(0), FunArgs(0), 180 PrevCtx(0) {} 181 }; 182 183 typedef SmallVector<SExprNode, 4> NodeVector; 184 185 private: 186 // A SExpr is a list of SExprNodes in prefix order. The Size field allows 187 // the list to be traversed as a tree. 188 NodeVector NodeVec; 189 190 private: 191 unsigned make(ExprOp O, unsigned F = 0, const void *D = 0) { 192 NodeVec.push_back(SExprNode(O, F, D)); 193 return NodeVec.size() - 1; 194 } 195 196 unsigned makeNop() { 197 return make(EOP_Nop); 198 } 199 200 unsigned makeWildcard() { 201 return make(EOP_Wildcard); 202 } 203 204 unsigned makeUniversal() { 205 return make(EOP_Universal); 206 } 207 208 unsigned makeNamedVar(const NamedDecl *D) { 209 return make(EOP_NVar, 0, D); 210 } 211 212 unsigned makeLocalVar(const NamedDecl *D) { 213 return make(EOP_LVar, 0, D); 214 } 215 216 unsigned makeThis() { 217 return make(EOP_This); 218 } 219 220 unsigned makeDot(const NamedDecl *D, bool Arrow) { 221 return make(EOP_Dot, Arrow ? 1 : 0, D); 222 } 223 224 unsigned makeCall(unsigned NumArgs, const NamedDecl *D) { 225 return make(EOP_Call, NumArgs, D); 226 } 227 228 // Grab the very first declaration of virtual method D 229 const CXXMethodDecl* getFirstVirtualDecl(const CXXMethodDecl *D) { 230 while (true) { 231 D = D->getCanonicalDecl(); 232 CXXMethodDecl::method_iterator I = D->begin_overridden_methods(), 233 E = D->end_overridden_methods(); 234 if (I == E) 235 return D; // Method does not override anything 236 D = *I; // FIXME: this does not work with multiple inheritance. 237 } 238 return 0; 239 } 240 241 unsigned makeMCall(unsigned NumArgs, const CXXMethodDecl *D) { 242 return make(EOP_MCall, NumArgs, getFirstVirtualDecl(D)); 243 } 244 245 unsigned makeIndex() { 246 return make(EOP_Index); 247 } 248 249 unsigned makeUnary() { 250 return make(EOP_Unary); 251 } 252 253 unsigned makeBinary() { 254 return make(EOP_Binary); 255 } 256 257 unsigned makeUnknown(unsigned Arity) { 258 return make(EOP_Unknown, Arity); 259 } 260 261 inline bool isCalleeArrow(const Expr *E) { 262 const MemberExpr *ME = dyn_cast<MemberExpr>(E->IgnoreParenCasts()); 263 return ME ? ME->isArrow() : false; 264 } 265 266 /// Build an SExpr from the given C++ expression. 267 /// Recursive function that terminates on DeclRefExpr. 268 /// Note: this function merely creates a SExpr; it does not check to 269 /// ensure that the original expression is a valid mutex expression. 270 /// 271 /// NDeref returns the number of Derefence and AddressOf operations 272 /// preceding the Expr; this is used to decide whether to pretty-print 273 /// SExprs with . or ->. 274 unsigned buildSExpr(const Expr *Exp, CallingContext* CallCtx, 275 int* NDeref = 0) { 276 if (!Exp) 277 return 0; 278 279 if (const DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(Exp)) { 280 const NamedDecl *ND = cast<NamedDecl>(DRE->getDecl()->getCanonicalDecl()); 281 const ParmVarDecl *PV = dyn_cast_or_null<ParmVarDecl>(ND); 282 if (PV) { 283 const 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 (const MemberExpr *ME = dyn_cast<MemberExpr>(Exp)) { 313 const 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 (const 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 const CXXMethodDecl *MD = CMCE->getMethodDecl()->getMostRecentDecl(); 324 if (LockReturnedAttr* At = MD->getAttr<LockReturnedAttr>()) { 325 CallingContext LRCallCtx(CMCE->getMethodDecl()); 326 LRCallCtx.SelfArg = CMCE->getImplicitObjectArgument(); 327 LRCallCtx.SelfArrow = isCalleeArrow(CMCE->getCallee()); 328 LRCallCtx.NumArgs = CMCE->getNumArgs(); 329 LRCallCtx.FunArgs = CMCE->getArgs(); 330 LRCallCtx.PrevCtx = CallCtx; 331 return buildSExpr(At->getArg(), &LRCallCtx); 332 } 333 // Hack to treat smart pointers and iterators as pointers; 334 // ignore any method named get(). 335 if (CMCE->getMethodDecl()->getNameAsString() == "get" && 336 CMCE->getNumArgs() == 0) { 337 if (NDeref && isCalleeArrow(CMCE->getCallee())) 338 ++(*NDeref); 339 return buildSExpr(CMCE->getImplicitObjectArgument(), CallCtx, NDeref); 340 } 341 unsigned NumCallArgs = CMCE->getNumArgs(); 342 unsigned Root = makeMCall(NumCallArgs, CMCE->getMethodDecl()); 343 unsigned Sz = buildSExpr(CMCE->getImplicitObjectArgument(), CallCtx); 344 const Expr* const* CallArgs = CMCE->getArgs(); 345 for (unsigned i = 0; i < NumCallArgs; ++i) { 346 Sz += buildSExpr(CallArgs[i], CallCtx); 347 } 348 NodeVec[Root].setSize(Sz + 1); 349 return Sz + 1; 350 } else if (const CallExpr *CE = dyn_cast<CallExpr>(Exp)) { 351 const FunctionDecl *FD = CE->getDirectCallee()->getMostRecentDecl(); 352 if (LockReturnedAttr* At = FD->getAttr<LockReturnedAttr>()) { 353 CallingContext LRCallCtx(CE->getDirectCallee()); 354 LRCallCtx.NumArgs = CE->getNumArgs(); 355 LRCallCtx.FunArgs = CE->getArgs(); 356 LRCallCtx.PrevCtx = CallCtx; 357 return buildSExpr(At->getArg(), &LRCallCtx); 358 } 359 // Treat smart pointers and iterators as pointers; 360 // ignore the * and -> operators. 361 if (const CXXOperatorCallExpr *OE = dyn_cast<CXXOperatorCallExpr>(CE)) { 362 OverloadedOperatorKind k = OE->getOperator(); 363 if (k == OO_Star) { 364 if (NDeref) ++(*NDeref); 365 return buildSExpr(OE->getArg(0), CallCtx, NDeref); 366 } 367 else if (k == OO_Arrow) { 368 return buildSExpr(OE->getArg(0), CallCtx, NDeref); 369 } 370 } 371 unsigned NumCallArgs = CE->getNumArgs(); 372 unsigned Root = makeCall(NumCallArgs, 0); 373 unsigned Sz = buildSExpr(CE->getCallee(), CallCtx); 374 const Expr* const* CallArgs = CE->getArgs(); 375 for (unsigned i = 0; i < NumCallArgs; ++i) { 376 Sz += buildSExpr(CallArgs[i], CallCtx); 377 } 378 NodeVec[Root].setSize(Sz+1); 379 return Sz+1; 380 } else if (const BinaryOperator *BOE = dyn_cast<BinaryOperator>(Exp)) { 381 unsigned Root = makeBinary(); 382 unsigned Sz = buildSExpr(BOE->getLHS(), CallCtx); 383 Sz += buildSExpr(BOE->getRHS(), CallCtx); 384 NodeVec[Root].setSize(Sz); 385 return Sz; 386 } else if (const UnaryOperator *UOE = dyn_cast<UnaryOperator>(Exp)) { 387 // Ignore & and * operators -- they're no-ops. 388 // However, we try to figure out whether the expression is a pointer, 389 // so we can use . and -> appropriately in error messages. 390 if (UOE->getOpcode() == UO_Deref) { 391 if (NDeref) ++(*NDeref); 392 return buildSExpr(UOE->getSubExpr(), CallCtx, NDeref); 393 } 394 if (UOE->getOpcode() == UO_AddrOf) { 395 if (DeclRefExpr* DRE = dyn_cast<DeclRefExpr>(UOE->getSubExpr())) { 396 if (DRE->getDecl()->isCXXInstanceMember()) { 397 // This is a pointer-to-member expression, e.g. &MyClass::mu_. 398 // We interpret this syntax specially, as a wildcard. 399 unsigned Root = makeDot(DRE->getDecl(), false); 400 makeWildcard(); 401 NodeVec[Root].setSize(2); 402 return 2; 403 } 404 } 405 if (NDeref) --(*NDeref); 406 return buildSExpr(UOE->getSubExpr(), CallCtx, NDeref); 407 } 408 unsigned Root = makeUnary(); 409 unsigned Sz = buildSExpr(UOE->getSubExpr(), CallCtx); 410 NodeVec[Root].setSize(Sz); 411 return Sz; 412 } else if (const ArraySubscriptExpr *ASE = 413 dyn_cast<ArraySubscriptExpr>(Exp)) { 414 unsigned Root = makeIndex(); 415 unsigned Sz = buildSExpr(ASE->getBase(), CallCtx); 416 Sz += buildSExpr(ASE->getIdx(), CallCtx); 417 NodeVec[Root].setSize(Sz); 418 return Sz; 419 } else if (const AbstractConditionalOperator *CE = 420 dyn_cast<AbstractConditionalOperator>(Exp)) { 421 unsigned Root = makeUnknown(3); 422 unsigned Sz = buildSExpr(CE->getCond(), CallCtx); 423 Sz += buildSExpr(CE->getTrueExpr(), CallCtx); 424 Sz += buildSExpr(CE->getFalseExpr(), CallCtx); 425 NodeVec[Root].setSize(Sz); 426 return Sz; 427 } else if (const ChooseExpr *CE = dyn_cast<ChooseExpr>(Exp)) { 428 unsigned Root = makeUnknown(3); 429 unsigned Sz = buildSExpr(CE->getCond(), CallCtx); 430 Sz += buildSExpr(CE->getLHS(), CallCtx); 431 Sz += buildSExpr(CE->getRHS(), CallCtx); 432 NodeVec[Root].setSize(Sz); 433 return Sz; 434 } else if (const CastExpr *CE = dyn_cast<CastExpr>(Exp)) { 435 return buildSExpr(CE->getSubExpr(), CallCtx, NDeref); 436 } else if (const ParenExpr *PE = dyn_cast<ParenExpr>(Exp)) { 437 return buildSExpr(PE->getSubExpr(), CallCtx, NDeref); 438 } else if (const ExprWithCleanups *EWC = dyn_cast<ExprWithCleanups>(Exp)) { 439 return buildSExpr(EWC->getSubExpr(), CallCtx, NDeref); 440 } else if (const CXXBindTemporaryExpr *E = dyn_cast<CXXBindTemporaryExpr>(Exp)) { 441 return buildSExpr(E->getSubExpr(), CallCtx, NDeref); 442 } else if (isa<CharacterLiteral>(Exp) || 443 isa<CXXNullPtrLiteralExpr>(Exp) || 444 isa<GNUNullExpr>(Exp) || 445 isa<CXXBoolLiteralExpr>(Exp) || 446 isa<FloatingLiteral>(Exp) || 447 isa<ImaginaryLiteral>(Exp) || 448 isa<IntegerLiteral>(Exp) || 449 isa<StringLiteral>(Exp) || 450 isa<ObjCStringLiteral>(Exp)) { 451 makeNop(); 452 return 1; // FIXME: Ignore literals for now 453 } else { 454 makeNop(); 455 return 1; // Ignore. FIXME: mark as invalid expression? 456 } 457 } 458 459 /// \brief Construct a SExpr from an expression. 460 /// \param MutexExp The original mutex expression within an attribute 461 /// \param DeclExp An expression involving the Decl on which the attribute 462 /// occurs. 463 /// \param D The declaration to which the lock/unlock attribute is attached. 464 void buildSExprFromExpr(const Expr *MutexExp, const Expr *DeclExp, 465 const NamedDecl *D, VarDecl *SelfDecl = 0) { 466 CallingContext CallCtx(D); 467 468 if (MutexExp) { 469 if (const StringLiteral* SLit = dyn_cast<StringLiteral>(MutexExp)) { 470 if (SLit->getString() == StringRef("*")) 471 // The "*" expr is a universal lock, which essentially turns off 472 // checks until it is removed from the lockset. 473 makeUniversal(); 474 else 475 // Ignore other string literals for now. 476 makeNop(); 477 return; 478 } 479 } 480 481 // If we are processing a raw attribute expression, with no substitutions. 482 if (DeclExp == 0) { 483 buildSExpr(MutexExp, 0); 484 return; 485 } 486 487 // Examine DeclExp to find SelfArg and FunArgs, which are used to substitute 488 // for formal parameters when we call buildMutexID later. 489 if (const MemberExpr *ME = dyn_cast<MemberExpr>(DeclExp)) { 490 CallCtx.SelfArg = ME->getBase(); 491 CallCtx.SelfArrow = ME->isArrow(); 492 } else if (const CXXMemberCallExpr *CE = 493 dyn_cast<CXXMemberCallExpr>(DeclExp)) { 494 CallCtx.SelfArg = CE->getImplicitObjectArgument(); 495 CallCtx.SelfArrow = isCalleeArrow(CE->getCallee()); 496 CallCtx.NumArgs = CE->getNumArgs(); 497 CallCtx.FunArgs = CE->getArgs(); 498 } else if (const CallExpr *CE = dyn_cast<CallExpr>(DeclExp)) { 499 CallCtx.NumArgs = CE->getNumArgs(); 500 CallCtx.FunArgs = CE->getArgs(); 501 } else if (const CXXConstructExpr *CE = 502 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(const Expr* MutexExp, const 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, 570 const Expr *MutexExp, const Expr *DeclExp, 571 const NamedDecl *D, StringRef Kind) { 572 SourceLocation Loc; 573 if (DeclExp) 574 Loc = DeclExp->getExprLoc(); 575 576 // FIXME: add a note about the attribute location in MutexExp or D 577 if (Loc.isValid()) 578 Handler.handleInvalidLockExp(Kind, Loc); 579 } 580 581 bool operator==(const SExpr &other) const { 582 return NodeVec == other.NodeVec; 583 } 584 585 bool operator!=(const SExpr &other) const { 586 return !(*this == other); 587 } 588 589 bool matches(const SExpr &Other, unsigned i = 0, unsigned j = 0) const { 590 if (NodeVec[i].matches(Other.NodeVec[j])) { 591 unsigned ni = NodeVec[i].arity(); 592 unsigned nj = Other.NodeVec[j].arity(); 593 unsigned n = (ni < nj) ? ni : nj; 594 bool Result = true; 595 unsigned ci = i+1; // first child of i 596 unsigned cj = j+1; // first child of j 597 for (unsigned k = 0; k < n; 598 ++k, ci=getNextSibling(ci), cj = Other.getNextSibling(cj)) { 599 Result = Result && matches(Other, ci, cj); 600 } 601 return Result; 602 } 603 return false; 604 } 605 606 // A partial match between a.mu and b.mu returns true a and b have the same 607 // type (and thus mu refers to the same mutex declaration), regardless of 608 // whether a and b are different objects or not. 609 bool partiallyMatches(const SExpr &Other) const { 610 if (NodeVec[0].kind() == EOP_Dot) 611 return NodeVec[0].matches(Other.NodeVec[0]); 612 return false; 613 } 614 615 /// \brief Pretty print a lock expression for use in error messages. 616 std::string toString(unsigned i = 0) const { 617 assert(isValid()); 618 if (i >= NodeVec.size()) 619 return ""; 620 621 const SExprNode* N = &NodeVec[i]; 622 switch (N->kind()) { 623 case EOP_Nop: 624 return "_"; 625 case EOP_Wildcard: 626 return "(?)"; 627 case EOP_Universal: 628 return "*"; 629 case EOP_This: 630 return "this"; 631 case EOP_NVar: 632 case EOP_LVar: { 633 return N->getNamedDecl()->getNameAsString(); 634 } 635 case EOP_Dot: { 636 if (NodeVec[i+1].kind() == EOP_Wildcard) { 637 std::string S = "&"; 638 S += N->getNamedDecl()->getQualifiedNameAsString(); 639 return S; 640 } 641 std::string FieldName = N->getNamedDecl()->getNameAsString(); 642 if (NodeVec[i+1].kind() == EOP_This) 643 return FieldName; 644 645 std::string S = toString(i+1); 646 if (N->isArrow()) 647 return S + "->" + FieldName; 648 else 649 return S + "." + FieldName; 650 } 651 case EOP_Call: { 652 std::string S = toString(i+1) + "("; 653 unsigned NumArgs = N->arity()-1; 654 unsigned ci = getNextSibling(i+1); 655 for (unsigned k=0; k<NumArgs; ++k, ci = getNextSibling(ci)) { 656 S += toString(ci); 657 if (k+1 < NumArgs) S += ","; 658 } 659 S += ")"; 660 return S; 661 } 662 case EOP_MCall: { 663 std::string S = ""; 664 if (NodeVec[i+1].kind() != EOP_This) 665 S = toString(i+1) + "."; 666 if (const NamedDecl *D = N->getFunctionDecl()) 667 S += D->getNameAsString() + "("; 668 else 669 S += "#("; 670 unsigned NumArgs = N->arity()-1; 671 unsigned ci = getNextSibling(i+1); 672 for (unsigned k=0; k<NumArgs; ++k, ci = getNextSibling(ci)) { 673 S += toString(ci); 674 if (k+1 < NumArgs) S += ","; 675 } 676 S += ")"; 677 return S; 678 } 679 case EOP_Index: { 680 std::string S1 = toString(i+1); 681 std::string S2 = toString(i+1 + NodeVec[i+1].size()); 682 return S1 + "[" + S2 + "]"; 683 } 684 case EOP_Unary: { 685 std::string S = toString(i+1); 686 return "#" + S; 687 } 688 case EOP_Binary: { 689 std::string S1 = toString(i+1); 690 std::string S2 = toString(i+1 + NodeVec[i+1].size()); 691 return "(" + S1 + "#" + S2 + ")"; 692 } 693 case EOP_Unknown: { 694 unsigned NumChildren = N->arity(); 695 if (NumChildren == 0) 696 return "(...)"; 697 std::string S = "("; 698 unsigned ci = i+1; 699 for (unsigned j = 0; j < NumChildren; ++j, ci = getNextSibling(ci)) { 700 S += toString(ci); 701 if (j+1 < NumChildren) S += "#"; 702 } 703 S += ")"; 704 return S; 705 } 706 } 707 return ""; 708 } 709 }; 710 711 /// \brief A short list of SExprs 712 class MutexIDList : public SmallVector<SExpr, 3> { 713 public: 714 /// \brief Push M onto list, but discard duplicates. 715 void push_back_nodup(const SExpr& M) { 716 if (end() == std::find(begin(), end(), M)) 717 push_back(M); 718 } 719 }; 720 721 /// \brief This is a helper class that stores info about the most recent 722 /// accquire of a Lock. 723 /// 724 /// The main body of the analysis maps MutexIDs to LockDatas. 725 struct LockData { 726 SourceLocation AcquireLoc; 727 728 /// \brief LKind stores whether a lock is held shared or exclusively. 729 /// Note that this analysis does not currently support either re-entrant 730 /// locking or lock "upgrading" and "downgrading" between exclusive and 731 /// shared. 732 /// 733 /// FIXME: add support for re-entrant locking and lock up/downgrading 734 LockKind LKind; 735 bool Asserted; // for asserted locks 736 bool Managed; // for ScopedLockable objects 737 SExpr UnderlyingMutex; // for ScopedLockable objects 738 739 LockData(SourceLocation AcquireLoc, LockKind LKind, bool M=false, 740 bool Asrt=false) 741 : AcquireLoc(AcquireLoc), LKind(LKind), Asserted(Asrt), Managed(M), 742 UnderlyingMutex(Decl::EmptyShell()) 743 {} 744 745 LockData(SourceLocation AcquireLoc, LockKind LKind, const SExpr &Mu) 746 : AcquireLoc(AcquireLoc), LKind(LKind), Asserted(false), Managed(false), 747 UnderlyingMutex(Mu) 748 {} 749 750 bool operator==(const LockData &other) const { 751 return AcquireLoc == other.AcquireLoc && LKind == other.LKind; 752 } 753 754 bool operator!=(const LockData &other) const { 755 return !(*this == other); 756 } 757 758 void Profile(llvm::FoldingSetNodeID &ID) const { 759 ID.AddInteger(AcquireLoc.getRawEncoding()); 760 ID.AddInteger(LKind); 761 } 762 763 bool isAtLeast(LockKind LK) { 764 return (LK == LK_Shared) || (LKind == LK_Exclusive); 765 } 766 }; 767 768 769 /// \brief A FactEntry stores a single fact that is known at a particular point 770 /// in the program execution. Currently, this is information regarding a lock 771 /// that is held at that point. 772 struct FactEntry { 773 SExpr MutID; 774 LockData LDat; 775 776 FactEntry(const SExpr& M, const LockData& L) 777 : MutID(M), LDat(L) 778 { } 779 }; 780 781 782 typedef unsigned short FactID; 783 784 /// \brief FactManager manages the memory for all facts that are created during 785 /// the analysis of a single routine. 786 class FactManager { 787 private: 788 std::vector<FactEntry> Facts; 789 790 public: 791 FactID newLock(const SExpr& M, const LockData& L) { 792 Facts.push_back(FactEntry(M,L)); 793 return static_cast<unsigned short>(Facts.size() - 1); 794 } 795 796 const FactEntry& operator[](FactID F) const { return Facts[F]; } 797 FactEntry& operator[](FactID F) { return Facts[F]; } 798 }; 799 800 801 /// \brief A FactSet is the set of facts that are known to be true at a 802 /// particular program point. FactSets must be small, because they are 803 /// frequently copied, and are thus implemented as a set of indices into a 804 /// table maintained by a FactManager. A typical FactSet only holds 1 or 2 805 /// locks, so we can get away with doing a linear search for lookup. Note 806 /// that a hashtable or map is inappropriate in this case, because lookups 807 /// may involve partial pattern matches, rather than exact matches. 808 class FactSet { 809 private: 810 typedef SmallVector<FactID, 4> FactVec; 811 812 FactVec FactIDs; 813 814 public: 815 typedef FactVec::iterator iterator; 816 typedef FactVec::const_iterator const_iterator; 817 818 iterator begin() { return FactIDs.begin(); } 819 const_iterator begin() const { return FactIDs.begin(); } 820 821 iterator end() { return FactIDs.end(); } 822 const_iterator end() const { return FactIDs.end(); } 823 824 bool isEmpty() const { return FactIDs.size() == 0; } 825 826 FactID addLock(FactManager& FM, const SExpr& M, const LockData& L) { 827 FactID F = FM.newLock(M, L); 828 FactIDs.push_back(F); 829 return F; 830 } 831 832 bool removeLock(FactManager& FM, const SExpr& M) { 833 unsigned n = FactIDs.size(); 834 if (n == 0) 835 return false; 836 837 for (unsigned i = 0; i < n-1; ++i) { 838 if (FM[FactIDs[i]].MutID.matches(M)) { 839 FactIDs[i] = FactIDs[n-1]; 840 FactIDs.pop_back(); 841 return true; 842 } 843 } 844 if (FM[FactIDs[n-1]].MutID.matches(M)) { 845 FactIDs.pop_back(); 846 return true; 847 } 848 return false; 849 } 850 851 // Returns an iterator 852 iterator findLockIter(FactManager &FM, const SExpr &M) { 853 for (iterator I = begin(), E = end(); I != E; ++I) { 854 const SExpr &Exp = FM[*I].MutID; 855 if (Exp.matches(M)) 856 return I; 857 } 858 return end(); 859 } 860 861 LockData* findLock(FactManager &FM, const SExpr &M) const { 862 for (const_iterator I = begin(), E = end(); I != E; ++I) { 863 const SExpr &Exp = FM[*I].MutID; 864 if (Exp.matches(M)) 865 return &FM[*I].LDat; 866 } 867 return 0; 868 } 869 870 LockData* findLockUniv(FactManager &FM, const SExpr &M) const { 871 for (const_iterator I = begin(), E = end(); I != E; ++I) { 872 const SExpr &Exp = FM[*I].MutID; 873 if (Exp.matches(M) || Exp.isUniversal()) 874 return &FM[*I].LDat; 875 } 876 return 0; 877 } 878 879 FactEntry* findPartialMatch(FactManager &FM, const SExpr &M) const { 880 for (const_iterator I=begin(), E=end(); I != E; ++I) { 881 const SExpr& Exp = FM[*I].MutID; 882 if (Exp.partiallyMatches(M)) return &FM[*I]; 883 } 884 return 0; 885 } 886 }; 887 888 889 890 /// A Lockset maps each SExpr (defined above) to information about how it has 891 /// been locked. 892 typedef llvm::ImmutableMap<SExpr, LockData> Lockset; 893 typedef llvm::ImmutableMap<const NamedDecl*, unsigned> LocalVarContext; 894 895 class LocalVariableMap; 896 897 /// A side (entry or exit) of a CFG node. 898 enum CFGBlockSide { CBS_Entry, CBS_Exit }; 899 900 /// CFGBlockInfo is a struct which contains all the information that is 901 /// maintained for each block in the CFG. See LocalVariableMap for more 902 /// information about the contexts. 903 struct CFGBlockInfo { 904 FactSet EntrySet; // Lockset held at entry to block 905 FactSet ExitSet; // Lockset held at exit from block 906 LocalVarContext EntryContext; // Context held at entry to block 907 LocalVarContext ExitContext; // Context held at exit from block 908 SourceLocation EntryLoc; // Location of first statement in block 909 SourceLocation ExitLoc; // Location of last statement in block. 910 unsigned EntryIndex; // Used to replay contexts later 911 bool Reachable; // Is this block reachable? 912 913 const FactSet &getSet(CFGBlockSide Side) const { 914 return Side == CBS_Entry ? EntrySet : ExitSet; 915 } 916 SourceLocation getLocation(CFGBlockSide Side) const { 917 return Side == CBS_Entry ? EntryLoc : ExitLoc; 918 } 919 920 private: 921 CFGBlockInfo(LocalVarContext EmptyCtx) 922 : EntryContext(EmptyCtx), ExitContext(EmptyCtx), Reachable(false) 923 { } 924 925 public: 926 static CFGBlockInfo getEmptyBlockInfo(LocalVariableMap &M); 927 }; 928 929 930 931 // A LocalVariableMap maintains a map from local variables to their currently 932 // valid definitions. It provides SSA-like functionality when traversing the 933 // CFG. Like SSA, each definition or assignment to a variable is assigned a 934 // unique name (an integer), which acts as the SSA name for that definition. 935 // The total set of names is shared among all CFG basic blocks. 936 // Unlike SSA, we do not rewrite expressions to replace local variables declrefs 937 // with their SSA-names. Instead, we compute a Context for each point in the 938 // code, which maps local variables to the appropriate SSA-name. This map 939 // changes with each assignment. 940 // 941 // The map is computed in a single pass over the CFG. Subsequent analyses can 942 // then query the map to find the appropriate Context for a statement, and use 943 // that Context to look up the definitions of variables. 944 class LocalVariableMap { 945 public: 946 typedef LocalVarContext Context; 947 948 /// A VarDefinition consists of an expression, representing the value of the 949 /// variable, along with the context in which that expression should be 950 /// interpreted. A reference VarDefinition does not itself contain this 951 /// information, but instead contains a pointer to a previous VarDefinition. 952 struct VarDefinition { 953 public: 954 friend class LocalVariableMap; 955 956 const NamedDecl *Dec; // The original declaration for this variable. 957 const Expr *Exp; // The expression for this variable, OR 958 unsigned Ref; // Reference to another VarDefinition 959 Context Ctx; // The map with which Exp should be interpreted. 960 961 bool isReference() { return !Exp; } 962 963 private: 964 // Create ordinary variable definition 965 VarDefinition(const NamedDecl *D, const Expr *E, Context C) 966 : Dec(D), Exp(E), Ref(0), Ctx(C) 967 { } 968 969 // Create reference to previous definition 970 VarDefinition(const NamedDecl *D, unsigned R, Context C) 971 : Dec(D), Exp(0), Ref(R), Ctx(C) 972 { } 973 }; 974 975 private: 976 Context::Factory ContextFactory; 977 std::vector<VarDefinition> VarDefinitions; 978 std::vector<unsigned> CtxIndices; 979 std::vector<std::pair<Stmt*, Context> > SavedContexts; 980 981 public: 982 LocalVariableMap() { 983 // index 0 is a placeholder for undefined variables (aka phi-nodes). 984 VarDefinitions.push_back(VarDefinition(0, 0u, getEmptyContext())); 985 } 986 987 /// Look up a definition, within the given context. 988 const VarDefinition* lookup(const NamedDecl *D, Context Ctx) { 989 const unsigned *i = Ctx.lookup(D); 990 if (!i) 991 return 0; 992 assert(*i < VarDefinitions.size()); 993 return &VarDefinitions[*i]; 994 } 995 996 /// Look up the definition for D within the given context. Returns 997 /// NULL if the expression is not statically known. If successful, also 998 /// modifies Ctx to hold the context of the return Expr. 999 const Expr* lookupExpr(const NamedDecl *D, Context &Ctx) { 1000 const unsigned *P = Ctx.lookup(D); 1001 if (!P) 1002 return 0; 1003 1004 unsigned i = *P; 1005 while (i > 0) { 1006 if (VarDefinitions[i].Exp) { 1007 Ctx = VarDefinitions[i].Ctx; 1008 return VarDefinitions[i].Exp; 1009 } 1010 i = VarDefinitions[i].Ref; 1011 } 1012 return 0; 1013 } 1014 1015 Context getEmptyContext() { return ContextFactory.getEmptyMap(); } 1016 1017 /// Return the next context after processing S. This function is used by 1018 /// clients of the class to get the appropriate context when traversing the 1019 /// CFG. It must be called for every assignment or DeclStmt. 1020 Context getNextContext(unsigned &CtxIndex, Stmt *S, Context C) { 1021 if (SavedContexts[CtxIndex+1].first == S) { 1022 CtxIndex++; 1023 Context Result = SavedContexts[CtxIndex].second; 1024 return Result; 1025 } 1026 return C; 1027 } 1028 1029 void dumpVarDefinitionName(unsigned i) { 1030 if (i == 0) { 1031 llvm::errs() << "Undefined"; 1032 return; 1033 } 1034 const NamedDecl *Dec = VarDefinitions[i].Dec; 1035 if (!Dec) { 1036 llvm::errs() << "<<NULL>>"; 1037 return; 1038 } 1039 Dec->printName(llvm::errs()); 1040 llvm::errs() << "." << i << " " << ((const void*) Dec); 1041 } 1042 1043 /// Dumps an ASCII representation of the variable map to llvm::errs() 1044 void dump() { 1045 for (unsigned i = 1, e = VarDefinitions.size(); i < e; ++i) { 1046 const Expr *Exp = VarDefinitions[i].Exp; 1047 unsigned Ref = VarDefinitions[i].Ref; 1048 1049 dumpVarDefinitionName(i); 1050 llvm::errs() << " = "; 1051 if (Exp) Exp->dump(); 1052 else { 1053 dumpVarDefinitionName(Ref); 1054 llvm::errs() << "\n"; 1055 } 1056 } 1057 } 1058 1059 /// Dumps an ASCII representation of a Context to llvm::errs() 1060 void dumpContext(Context C) { 1061 for (Context::iterator I = C.begin(), E = C.end(); I != E; ++I) { 1062 const NamedDecl *D = I.getKey(); 1063 D->printName(llvm::errs()); 1064 const unsigned *i = C.lookup(D); 1065 llvm::errs() << " -> "; 1066 dumpVarDefinitionName(*i); 1067 llvm::errs() << "\n"; 1068 } 1069 } 1070 1071 /// Builds the variable map. 1072 void traverseCFG(CFG *CFGraph, const PostOrderCFGView *SortedGraph, 1073 std::vector<CFGBlockInfo> &BlockInfo); 1074 1075 protected: 1076 // Get the current context index 1077 unsigned getContextIndex() { return SavedContexts.size()-1; } 1078 1079 // Save the current context for later replay 1080 void saveContext(Stmt *S, Context C) { 1081 SavedContexts.push_back(std::make_pair(S,C)); 1082 } 1083 1084 // Adds a new definition to the given context, and returns a new context. 1085 // This method should be called when declaring a new variable. 1086 Context addDefinition(const NamedDecl *D, Expr *Exp, Context Ctx) { 1087 assert(!Ctx.contains(D)); 1088 unsigned newID = VarDefinitions.size(); 1089 Context NewCtx = ContextFactory.add(Ctx, D, newID); 1090 VarDefinitions.push_back(VarDefinition(D, Exp, Ctx)); 1091 return NewCtx; 1092 } 1093 1094 // Add a new reference to an existing definition. 1095 Context addReference(const NamedDecl *D, unsigned i, Context Ctx) { 1096 unsigned newID = VarDefinitions.size(); 1097 Context NewCtx = ContextFactory.add(Ctx, D, newID); 1098 VarDefinitions.push_back(VarDefinition(D, i, Ctx)); 1099 return NewCtx; 1100 } 1101 1102 // Updates a definition only if that definition is already in the map. 1103 // This method should be called when assigning to an existing variable. 1104 Context updateDefinition(const NamedDecl *D, Expr *Exp, Context Ctx) { 1105 if (Ctx.contains(D)) { 1106 unsigned newID = VarDefinitions.size(); 1107 Context NewCtx = ContextFactory.remove(Ctx, D); 1108 NewCtx = ContextFactory.add(NewCtx, D, newID); 1109 VarDefinitions.push_back(VarDefinition(D, Exp, Ctx)); 1110 return NewCtx; 1111 } 1112 return Ctx; 1113 } 1114 1115 // Removes a definition from the context, but keeps the variable name 1116 // as a valid variable. The index 0 is a placeholder for cleared definitions. 1117 Context clearDefinition(const NamedDecl *D, Context Ctx) { 1118 Context NewCtx = Ctx; 1119 if (NewCtx.contains(D)) { 1120 NewCtx = ContextFactory.remove(NewCtx, D); 1121 NewCtx = ContextFactory.add(NewCtx, D, 0); 1122 } 1123 return NewCtx; 1124 } 1125 1126 // Remove a definition entirely frmo the context. 1127 Context removeDefinition(const NamedDecl *D, Context Ctx) { 1128 Context NewCtx = Ctx; 1129 if (NewCtx.contains(D)) { 1130 NewCtx = ContextFactory.remove(NewCtx, D); 1131 } 1132 return NewCtx; 1133 } 1134 1135 Context intersectContexts(Context C1, Context C2); 1136 Context createReferenceContext(Context C); 1137 void intersectBackEdge(Context C1, Context C2); 1138 1139 friend class VarMapBuilder; 1140 }; 1141 1142 1143 // This has to be defined after LocalVariableMap. 1144 CFGBlockInfo CFGBlockInfo::getEmptyBlockInfo(LocalVariableMap &M) { 1145 return CFGBlockInfo(M.getEmptyContext()); 1146 } 1147 1148 1149 /// Visitor which builds a LocalVariableMap 1150 class VarMapBuilder : public StmtVisitor<VarMapBuilder> { 1151 public: 1152 LocalVariableMap* VMap; 1153 LocalVariableMap::Context Ctx; 1154 1155 VarMapBuilder(LocalVariableMap *VM, LocalVariableMap::Context C) 1156 : VMap(VM), Ctx(C) {} 1157 1158 void VisitDeclStmt(DeclStmt *S); 1159 void VisitBinaryOperator(BinaryOperator *BO); 1160 }; 1161 1162 1163 // Add new local variables to the variable map 1164 void VarMapBuilder::VisitDeclStmt(DeclStmt *S) { 1165 bool modifiedCtx = false; 1166 DeclGroupRef DGrp = S->getDeclGroup(); 1167 for (DeclGroupRef::iterator I = DGrp.begin(), E = DGrp.end(); I != E; ++I) { 1168 if (VarDecl *VD = dyn_cast_or_null<VarDecl>(*I)) { 1169 Expr *E = VD->getInit(); 1170 1171 // Add local variables with trivial type to the variable map 1172 QualType T = VD->getType(); 1173 if (T.isTrivialType(VD->getASTContext())) { 1174 Ctx = VMap->addDefinition(VD, E, Ctx); 1175 modifiedCtx = true; 1176 } 1177 } 1178 } 1179 if (modifiedCtx) 1180 VMap->saveContext(S, Ctx); 1181 } 1182 1183 // Update local variable definitions in variable map 1184 void VarMapBuilder::VisitBinaryOperator(BinaryOperator *BO) { 1185 if (!BO->isAssignmentOp()) 1186 return; 1187 1188 Expr *LHSExp = BO->getLHS()->IgnoreParenCasts(); 1189 1190 // Update the variable map and current context. 1191 if (DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(LHSExp)) { 1192 ValueDecl *VDec = DRE->getDecl(); 1193 if (Ctx.lookup(VDec)) { 1194 if (BO->getOpcode() == BO_Assign) 1195 Ctx = VMap->updateDefinition(VDec, BO->getRHS(), Ctx); 1196 else 1197 // FIXME -- handle compound assignment operators 1198 Ctx = VMap->clearDefinition(VDec, Ctx); 1199 VMap->saveContext(BO, Ctx); 1200 } 1201 } 1202 } 1203 1204 1205 // Computes the intersection of two contexts. The intersection is the 1206 // set of variables which have the same definition in both contexts; 1207 // variables with different definitions are discarded. 1208 LocalVariableMap::Context 1209 LocalVariableMap::intersectContexts(Context C1, Context C2) { 1210 Context Result = C1; 1211 for (Context::iterator I = C1.begin(), E = C1.end(); I != E; ++I) { 1212 const NamedDecl *Dec = I.getKey(); 1213 unsigned i1 = I.getData(); 1214 const unsigned *i2 = C2.lookup(Dec); 1215 if (!i2) // variable doesn't exist on second path 1216 Result = removeDefinition(Dec, Result); 1217 else if (*i2 != i1) // variable exists, but has different definition 1218 Result = clearDefinition(Dec, Result); 1219 } 1220 return Result; 1221 } 1222 1223 // For every variable in C, create a new variable that refers to the 1224 // definition in C. Return a new context that contains these new variables. 1225 // (We use this for a naive implementation of SSA on loop back-edges.) 1226 LocalVariableMap::Context LocalVariableMap::createReferenceContext(Context C) { 1227 Context Result = getEmptyContext(); 1228 for (Context::iterator I = C.begin(), E = C.end(); I != E; ++I) { 1229 const NamedDecl *Dec = I.getKey(); 1230 unsigned i = I.getData(); 1231 Result = addReference(Dec, i, Result); 1232 } 1233 return Result; 1234 } 1235 1236 // This routine also takes the intersection of C1 and C2, but it does so by 1237 // altering the VarDefinitions. C1 must be the result of an earlier call to 1238 // createReferenceContext. 1239 void LocalVariableMap::intersectBackEdge(Context C1, Context C2) { 1240 for (Context::iterator I = C1.begin(), E = C1.end(); I != E; ++I) { 1241 const NamedDecl *Dec = I.getKey(); 1242 unsigned i1 = I.getData(); 1243 VarDefinition *VDef = &VarDefinitions[i1]; 1244 assert(VDef->isReference()); 1245 1246 const unsigned *i2 = C2.lookup(Dec); 1247 if (!i2 || (*i2 != i1)) 1248 VDef->Ref = 0; // Mark this variable as undefined 1249 } 1250 } 1251 1252 1253 // Traverse the CFG in topological order, so all predecessors of a block 1254 // (excluding back-edges) are visited before the block itself. At 1255 // each point in the code, we calculate a Context, which holds the set of 1256 // variable definitions which are visible at that point in execution. 1257 // Visible variables are mapped to their definitions using an array that 1258 // contains all definitions. 1259 // 1260 // At join points in the CFG, the set is computed as the intersection of 1261 // the incoming sets along each edge, E.g. 1262 // 1263 // { Context | VarDefinitions } 1264 // int x = 0; { x -> x1 | x1 = 0 } 1265 // int y = 0; { x -> x1, y -> y1 | y1 = 0, x1 = 0 } 1266 // if (b) x = 1; { x -> x2, y -> y1 | x2 = 1, y1 = 0, ... } 1267 // else x = 2; { x -> x3, y -> y1 | x3 = 2, x2 = 1, ... } 1268 // ... { y -> y1 (x is unknown) | x3 = 2, x2 = 1, ... } 1269 // 1270 // This is essentially a simpler and more naive version of the standard SSA 1271 // algorithm. Those definitions that remain in the intersection are from blocks 1272 // that strictly dominate the current block. We do not bother to insert proper 1273 // phi nodes, because they are not used in our analysis; instead, wherever 1274 // a phi node would be required, we simply remove that definition from the 1275 // context (E.g. x above). 1276 // 1277 // The initial traversal does not capture back-edges, so those need to be 1278 // handled on a separate pass. Whenever the first pass encounters an 1279 // incoming back edge, it duplicates the context, creating new definitions 1280 // that refer back to the originals. (These correspond to places where SSA 1281 // might have to insert a phi node.) On the second pass, these definitions are 1282 // set to NULL if the variable has changed on the back-edge (i.e. a phi 1283 // node was actually required.) E.g. 1284 // 1285 // { Context | VarDefinitions } 1286 // int x = 0, y = 0; { x -> x1, y -> y1 | y1 = 0, x1 = 0 } 1287 // while (b) { x -> x2, y -> y1 | [1st:] x2=x1; [2nd:] x2=NULL; } 1288 // x = x+1; { x -> x3, y -> y1 | x3 = x2 + 1, ... } 1289 // ... { y -> y1 | x3 = 2, x2 = 1, ... } 1290 // 1291 void LocalVariableMap::traverseCFG(CFG *CFGraph, 1292 const PostOrderCFGView *SortedGraph, 1293 std::vector<CFGBlockInfo> &BlockInfo) { 1294 PostOrderCFGView::CFGBlockSet VisitedBlocks(CFGraph); 1295 1296 CtxIndices.resize(CFGraph->getNumBlockIDs()); 1297 1298 for (const auto *CurrBlock : *SortedGraph) { 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 CFGStmt CS = BI->castAs<CFGStmt>(); 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 const PostOrderCFGView *SortedGraph, 1378 std::vector<CFGBlockInfo> &BlockInfo) { 1379 for (const auto *CurrBlock : *SortedGraph) { 1380 CFGBlockInfo *CurrBlockInfo = &BlockInfo[CurrBlock->getBlockID()]; 1381 1382 // Find the source location of the last statement in the block, if the 1383 // block is not empty. 1384 if (const Stmt *S = CurrBlock->getTerminator()) { 1385 CurrBlockInfo->EntryLoc = CurrBlockInfo->ExitLoc = S->getLocStart(); 1386 } else { 1387 for (CFGBlock::const_reverse_iterator BI = CurrBlock->rbegin(), 1388 BE = CurrBlock->rend(); BI != BE; ++BI) { 1389 // FIXME: Handle other CFGElement kinds. 1390 if (Optional<CFGStmt> CS = BI->getAs<CFGStmt>()) { 1391 CurrBlockInfo->ExitLoc = CS->getStmt()->getLocStart(); 1392 break; 1393 } 1394 } 1395 } 1396 1397 if (!CurrBlockInfo->ExitLoc.isInvalid()) { 1398 // This block contains at least one statement. Find the source location 1399 // of the first statement in the block. 1400 for (CFGBlock::const_iterator BI = CurrBlock->begin(), 1401 BE = CurrBlock->end(); BI != BE; ++BI) { 1402 // FIXME: Handle other CFGElement kinds. 1403 if (Optional<CFGStmt> CS = BI->getAs<CFGStmt>()) { 1404 CurrBlockInfo->EntryLoc = CS->getStmt()->getLocStart(); 1405 break; 1406 } 1407 } 1408 } else if (CurrBlock->pred_size() == 1 && *CurrBlock->pred_begin() && 1409 CurrBlock != &CFGraph->getExit()) { 1410 // The block is empty, and has a single predecessor. Use its exit 1411 // location. 1412 CurrBlockInfo->EntryLoc = CurrBlockInfo->ExitLoc = 1413 BlockInfo[(*CurrBlock->pred_begin())->getBlockID()].ExitLoc; 1414 } 1415 } 1416 } 1417 1418 /// \brief Class which implements the core thread safety analysis routines. 1419 class ThreadSafetyAnalyzer { 1420 friend class BuildLockset; 1421 1422 ThreadSafetyHandler &Handler; 1423 LocalVariableMap LocalVarMap; 1424 FactManager FactMan; 1425 std::vector<CFGBlockInfo> BlockInfo; 1426 1427 public: 1428 ThreadSafetyAnalyzer(ThreadSafetyHandler &H) : Handler(H) {} 1429 1430 void addLock(FactSet &FSet, const SExpr &Mutex, const LockData &LDat, 1431 StringRef DiagKind); 1432 void removeLock(FactSet &FSet, const SExpr &Mutex, SourceLocation UnlockLoc, 1433 bool FullyRemove, LockKind Kind, StringRef DiagKind); 1434 1435 template <typename AttrType> 1436 void getMutexIDs(MutexIDList &Mtxs, AttrType *Attr, Expr *Exp, 1437 const NamedDecl *D, VarDecl *SelfDecl=0); 1438 1439 template <class AttrType> 1440 void getMutexIDs(MutexIDList &Mtxs, AttrType *Attr, Expr *Exp, 1441 const NamedDecl *D, 1442 const CFGBlock *PredBlock, const CFGBlock *CurrBlock, 1443 Expr *BrE, bool Neg); 1444 1445 const CallExpr* getTrylockCallExpr(const Stmt *Cond, LocalVarContext C, 1446 bool &Negate); 1447 1448 void getEdgeLockset(FactSet &Result, const FactSet &ExitSet, 1449 const CFGBlock* PredBlock, 1450 const CFGBlock *CurrBlock); 1451 1452 void intersectAndWarn(FactSet &FSet1, const FactSet &FSet2, 1453 SourceLocation JoinLoc, 1454 LockErrorKind LEK1, LockErrorKind LEK2, 1455 bool Modify=true); 1456 1457 void intersectAndWarn(FactSet &FSet1, const FactSet &FSet2, 1458 SourceLocation JoinLoc, LockErrorKind LEK1, 1459 bool Modify=true) { 1460 intersectAndWarn(FSet1, FSet2, JoinLoc, LEK1, LEK1, Modify); 1461 } 1462 1463 void runAnalysis(AnalysisDeclContext &AC); 1464 }; 1465 1466 /// \brief Gets the value decl pointer from DeclRefExprs or MemberExprs. 1467 static const ValueDecl *getValueDecl(const Expr *Exp) { 1468 if (const auto *CE = dyn_cast<ImplicitCastExpr>(Exp)) 1469 return getValueDecl(CE->getSubExpr()); 1470 1471 if (const auto *DR = dyn_cast<DeclRefExpr>(Exp)) 1472 return DR->getDecl(); 1473 1474 if (const auto *ME = dyn_cast<MemberExpr>(Exp)) 1475 return ME->getMemberDecl(); 1476 1477 return nullptr; 1478 } 1479 1480 template <typename Ty> 1481 class has_arg_iterator_range { 1482 typedef char yes[1]; 1483 typedef char no[2]; 1484 1485 template <typename Inner> 1486 static yes& test(Inner *I, decltype(I->args()) * = nullptr); 1487 1488 template <typename> 1489 static no& test(...); 1490 1491 public: 1492 static const bool value = sizeof(test<Ty>(nullptr)) == sizeof(yes); 1493 }; 1494 1495 static StringRef ClassifyDiagnostic(const CapabilityAttr *A) { 1496 return A->getName(); 1497 } 1498 1499 static StringRef ClassifyDiagnostic(QualType VDT) { 1500 // We need to look at the declaration of the type of the value to determine 1501 // which it is. The type should either be a record or a typedef, or a pointer 1502 // or reference thereof. 1503 if (const auto *RT = VDT->getAs<RecordType>()) { 1504 if (const auto *RD = RT->getDecl()) 1505 if (const auto *CA = RD->getAttr<CapabilityAttr>()) 1506 return ClassifyDiagnostic(CA); 1507 } else if (const auto *TT = VDT->getAs<TypedefType>()) { 1508 if (const auto *TD = TT->getDecl()) 1509 if (const auto *CA = TD->getAttr<CapabilityAttr>()) 1510 return ClassifyDiagnostic(CA); 1511 } else if (VDT->isPointerType() || VDT->isReferenceType()) 1512 return ClassifyDiagnostic(VDT->getPointeeType()); 1513 1514 return "mutex"; 1515 } 1516 1517 static StringRef ClassifyDiagnostic(const ValueDecl *VD) { 1518 assert(VD && "No ValueDecl passed"); 1519 1520 // The ValueDecl is the declaration of a mutex or role (hopefully). 1521 return ClassifyDiagnostic(VD->getType()); 1522 } 1523 1524 template <typename AttrTy> 1525 static typename std::enable_if<!has_arg_iterator_range<AttrTy>::value, 1526 StringRef>::type 1527 ClassifyDiagnostic(const AttrTy *A) { 1528 if (const ValueDecl *VD = getValueDecl(A->getArg())) 1529 return ClassifyDiagnostic(VD); 1530 return "mutex"; 1531 } 1532 1533 template <typename AttrTy> 1534 static typename std::enable_if<has_arg_iterator_range<AttrTy>::value, 1535 StringRef>::type 1536 ClassifyDiagnostic(const AttrTy *A) { 1537 for (const auto *Arg : A->args()) { 1538 if (const ValueDecl *VD = getValueDecl(Arg)) 1539 return ClassifyDiagnostic(VD); 1540 } 1541 return "mutex"; 1542 } 1543 1544 /// \brief Add a new lock to the lockset, warning if the lock is already there. 1545 /// \param Mutex -- the Mutex expression for the lock 1546 /// \param LDat -- the LockData for the lock 1547 void ThreadSafetyAnalyzer::addLock(FactSet &FSet, const SExpr &Mutex, 1548 const LockData &LDat, StringRef DiagKind) { 1549 // FIXME: deal with acquired before/after annotations. 1550 // FIXME: Don't always warn when we have support for reentrant locks. 1551 if (Mutex.shouldIgnore()) 1552 return; 1553 1554 if (FSet.findLock(FactMan, Mutex)) { 1555 if (!LDat.Asserted) 1556 Handler.handleDoubleLock(DiagKind, Mutex.toString(), LDat.AcquireLoc); 1557 } else { 1558 FSet.addLock(FactMan, Mutex, LDat); 1559 } 1560 } 1561 1562 1563 /// \brief Remove a lock from the lockset, warning if the lock is not there. 1564 /// \param Mutex The lock expression corresponding to the lock to be removed 1565 /// \param UnlockLoc The source location of the unlock (only used in error msg) 1566 void ThreadSafetyAnalyzer::removeLock(FactSet &FSet, const SExpr &Mutex, 1567 SourceLocation UnlockLoc, 1568 bool FullyRemove, LockKind ReceivedKind, 1569 StringRef DiagKind) { 1570 if (Mutex.shouldIgnore()) 1571 return; 1572 1573 const LockData *LDat = FSet.findLock(FactMan, Mutex); 1574 if (!LDat) { 1575 Handler.handleUnmatchedUnlock(DiagKind, Mutex.toString(), UnlockLoc); 1576 return; 1577 } 1578 1579 // Generic lock removal doesn't care about lock kind mismatches, but 1580 // otherwise diagnose when the lock kinds are mismatched. 1581 if (ReceivedKind != LK_Generic && LDat->LKind != ReceivedKind) { 1582 Handler.handleIncorrectUnlockKind(DiagKind, Mutex.toString(), LDat->LKind, 1583 ReceivedKind, UnlockLoc); 1584 return; 1585 } 1586 1587 if (LDat->UnderlyingMutex.isValid()) { 1588 // This is scoped lockable object, which manages the real mutex. 1589 if (FullyRemove) { 1590 // We're destroying the managing object. 1591 // Remove the underlying mutex if it exists; but don't warn. 1592 if (FSet.findLock(FactMan, LDat->UnderlyingMutex)) 1593 FSet.removeLock(FactMan, LDat->UnderlyingMutex); 1594 } else { 1595 // We're releasing the underlying mutex, but not destroying the 1596 // managing object. Warn on dual release. 1597 if (!FSet.findLock(FactMan, LDat->UnderlyingMutex)) { 1598 Handler.handleUnmatchedUnlock( 1599 DiagKind, LDat->UnderlyingMutex.toString(), UnlockLoc); 1600 } 1601 FSet.removeLock(FactMan, LDat->UnderlyingMutex); 1602 return; 1603 } 1604 } 1605 FSet.removeLock(FactMan, Mutex); 1606 } 1607 1608 1609 /// \brief Extract the list of mutexIDs from the attribute on an expression, 1610 /// and push them onto Mtxs, discarding any duplicates. 1611 template <typename AttrType> 1612 void ThreadSafetyAnalyzer::getMutexIDs(MutexIDList &Mtxs, AttrType *Attr, 1613 Expr *Exp, const NamedDecl *D, 1614 VarDecl *SelfDecl) { 1615 if (Attr->args_size() == 0) { 1616 // The mutex held is the "this" object. 1617 SExpr Mu(0, Exp, D, SelfDecl); 1618 if (!Mu.isValid()) 1619 SExpr::warnInvalidLock(Handler, 0, Exp, D, ClassifyDiagnostic(Attr)); 1620 else 1621 Mtxs.push_back_nodup(Mu); 1622 return; 1623 } 1624 1625 for (const auto *Arg : Attr->args()) { 1626 SExpr Mu(Arg, Exp, D, SelfDecl); 1627 if (!Mu.isValid()) 1628 SExpr::warnInvalidLock(Handler, Arg, Exp, D, ClassifyDiagnostic(Attr)); 1629 else 1630 Mtxs.push_back_nodup(Mu); 1631 } 1632 } 1633 1634 1635 /// \brief Extract the list of mutexIDs from a trylock attribute. If the 1636 /// trylock applies to the given edge, then push them onto Mtxs, discarding 1637 /// any duplicates. 1638 template <class AttrType> 1639 void ThreadSafetyAnalyzer::getMutexIDs(MutexIDList &Mtxs, AttrType *Attr, 1640 Expr *Exp, const NamedDecl *D, 1641 const CFGBlock *PredBlock, 1642 const CFGBlock *CurrBlock, 1643 Expr *BrE, bool Neg) { 1644 // Find out which branch has the lock 1645 bool branch = 0; 1646 if (CXXBoolLiteralExpr *BLE = dyn_cast_or_null<CXXBoolLiteralExpr>(BrE)) { 1647 branch = BLE->getValue(); 1648 } 1649 else if (IntegerLiteral *ILE = dyn_cast_or_null<IntegerLiteral>(BrE)) { 1650 branch = ILE->getValue().getBoolValue(); 1651 } 1652 int branchnum = branch ? 0 : 1; 1653 if (Neg) branchnum = !branchnum; 1654 1655 // If we've taken the trylock branch, then add the lock 1656 int i = 0; 1657 for (CFGBlock::const_succ_iterator SI = PredBlock->succ_begin(), 1658 SE = PredBlock->succ_end(); SI != SE && i < 2; ++SI, ++i) { 1659 if (*SI == CurrBlock && i == branchnum) { 1660 getMutexIDs(Mtxs, Attr, Exp, D); 1661 } 1662 } 1663 } 1664 1665 1666 bool getStaticBooleanValue(Expr* E, bool& TCond) { 1667 if (isa<CXXNullPtrLiteralExpr>(E) || isa<GNUNullExpr>(E)) { 1668 TCond = false; 1669 return true; 1670 } else if (CXXBoolLiteralExpr *BLE = dyn_cast<CXXBoolLiteralExpr>(E)) { 1671 TCond = BLE->getValue(); 1672 return true; 1673 } else if (IntegerLiteral *ILE = dyn_cast<IntegerLiteral>(E)) { 1674 TCond = ILE->getValue().getBoolValue(); 1675 return true; 1676 } else if (ImplicitCastExpr *CE = dyn_cast<ImplicitCastExpr>(E)) { 1677 return getStaticBooleanValue(CE->getSubExpr(), TCond); 1678 } 1679 return false; 1680 } 1681 1682 1683 // If Cond can be traced back to a function call, return the call expression. 1684 // The negate variable should be called with false, and will be set to true 1685 // if the function call is negated, e.g. if (!mu.tryLock(...)) 1686 const CallExpr* ThreadSafetyAnalyzer::getTrylockCallExpr(const Stmt *Cond, 1687 LocalVarContext C, 1688 bool &Negate) { 1689 if (!Cond) 1690 return 0; 1691 1692 if (const CallExpr *CallExp = dyn_cast<CallExpr>(Cond)) { 1693 return CallExp; 1694 } 1695 else if (const ParenExpr *PE = dyn_cast<ParenExpr>(Cond)) { 1696 return getTrylockCallExpr(PE->getSubExpr(), C, Negate); 1697 } 1698 else if (const ImplicitCastExpr *CE = dyn_cast<ImplicitCastExpr>(Cond)) { 1699 return getTrylockCallExpr(CE->getSubExpr(), C, Negate); 1700 } 1701 else if (const ExprWithCleanups* EWC = dyn_cast<ExprWithCleanups>(Cond)) { 1702 return getTrylockCallExpr(EWC->getSubExpr(), C, Negate); 1703 } 1704 else if (const DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(Cond)) { 1705 const Expr *E = LocalVarMap.lookupExpr(DRE->getDecl(), C); 1706 return getTrylockCallExpr(E, C, Negate); 1707 } 1708 else if (const UnaryOperator *UOP = dyn_cast<UnaryOperator>(Cond)) { 1709 if (UOP->getOpcode() == UO_LNot) { 1710 Negate = !Negate; 1711 return getTrylockCallExpr(UOP->getSubExpr(), C, Negate); 1712 } 1713 return 0; 1714 } 1715 else if (const BinaryOperator *BOP = dyn_cast<BinaryOperator>(Cond)) { 1716 if (BOP->getOpcode() == BO_EQ || BOP->getOpcode() == BO_NE) { 1717 if (BOP->getOpcode() == BO_NE) 1718 Negate = !Negate; 1719 1720 bool TCond = false; 1721 if (getStaticBooleanValue(BOP->getRHS(), TCond)) { 1722 if (!TCond) Negate = !Negate; 1723 return getTrylockCallExpr(BOP->getLHS(), C, Negate); 1724 } 1725 TCond = false; 1726 if (getStaticBooleanValue(BOP->getLHS(), TCond)) { 1727 if (!TCond) Negate = !Negate; 1728 return getTrylockCallExpr(BOP->getRHS(), C, Negate); 1729 } 1730 return 0; 1731 } 1732 if (BOP->getOpcode() == BO_LAnd) { 1733 // LHS must have been evaluated in a different block. 1734 return getTrylockCallExpr(BOP->getRHS(), C, Negate); 1735 } 1736 if (BOP->getOpcode() == BO_LOr) { 1737 return getTrylockCallExpr(BOP->getRHS(), C, Negate); 1738 } 1739 return 0; 1740 } 1741 return 0; 1742 } 1743 1744 1745 /// \brief Find the lockset that holds on the edge between PredBlock 1746 /// and CurrBlock. The edge set is the exit set of PredBlock (passed 1747 /// as the ExitSet parameter) plus any trylocks, which are conditionally held. 1748 void ThreadSafetyAnalyzer::getEdgeLockset(FactSet& Result, 1749 const FactSet &ExitSet, 1750 const CFGBlock *PredBlock, 1751 const CFGBlock *CurrBlock) { 1752 Result = ExitSet; 1753 1754 const Stmt *Cond = PredBlock->getTerminatorCondition(); 1755 if (!Cond) 1756 return; 1757 1758 bool Negate = false; 1759 const CFGBlockInfo *PredBlockInfo = &BlockInfo[PredBlock->getBlockID()]; 1760 const LocalVarContext &LVarCtx = PredBlockInfo->ExitContext; 1761 StringRef CapDiagKind = "mutex"; 1762 1763 CallExpr *Exp = 1764 const_cast<CallExpr*>(getTrylockCallExpr(Cond, LVarCtx, Negate)); 1765 if (!Exp) 1766 return; 1767 1768 NamedDecl *FunDecl = dyn_cast_or_null<NamedDecl>(Exp->getCalleeDecl()); 1769 if(!FunDecl || !FunDecl->hasAttrs()) 1770 return; 1771 1772 MutexIDList ExclusiveLocksToAdd; 1773 MutexIDList SharedLocksToAdd; 1774 1775 // If the condition is a call to a Trylock function, then grab the attributes 1776 AttrVec &ArgAttrs = FunDecl->getAttrs(); 1777 for (unsigned i = 0; i < ArgAttrs.size(); ++i) { 1778 Attr *Attr = ArgAttrs[i]; 1779 switch (Attr->getKind()) { 1780 case attr::ExclusiveTrylockFunction: { 1781 ExclusiveTrylockFunctionAttr *A = 1782 cast<ExclusiveTrylockFunctionAttr>(Attr); 1783 getMutexIDs(ExclusiveLocksToAdd, A, Exp, FunDecl, 1784 PredBlock, CurrBlock, A->getSuccessValue(), Negate); 1785 CapDiagKind = ClassifyDiagnostic(A); 1786 break; 1787 } 1788 case attr::SharedTrylockFunction: { 1789 SharedTrylockFunctionAttr *A = 1790 cast<SharedTrylockFunctionAttr>(Attr); 1791 getMutexIDs(SharedLocksToAdd, A, Exp, FunDecl, 1792 PredBlock, CurrBlock, A->getSuccessValue(), Negate); 1793 CapDiagKind = ClassifyDiagnostic(A); 1794 break; 1795 } 1796 default: 1797 break; 1798 } 1799 } 1800 1801 // Add and remove locks. 1802 SourceLocation Loc = Exp->getExprLoc(); 1803 for (const auto &ExclusiveLockToAdd : ExclusiveLocksToAdd) 1804 addLock(Result, ExclusiveLockToAdd, LockData(Loc, LK_Exclusive), 1805 CapDiagKind); 1806 for (const auto &SharedLockToAdd : SharedLocksToAdd) 1807 addLock(Result, SharedLockToAdd, LockData(Loc, LK_Shared), CapDiagKind); 1808 } 1809 1810 /// \brief We use this class to visit different types of expressions in 1811 /// CFGBlocks, and build up the lockset. 1812 /// An expression may cause us to add or remove locks from the lockset, or else 1813 /// output error messages related to missing locks. 1814 /// FIXME: In future, we may be able to not inherit from a visitor. 1815 class BuildLockset : public StmtVisitor<BuildLockset> { 1816 friend class ThreadSafetyAnalyzer; 1817 1818 ThreadSafetyAnalyzer *Analyzer; 1819 FactSet FSet; 1820 LocalVariableMap::Context LVarCtx; 1821 unsigned CtxIndex; 1822 1823 // Helper functions 1824 1825 void warnIfMutexNotHeld(const NamedDecl *D, const Expr *Exp, AccessKind AK, 1826 Expr *MutexExp, ProtectedOperationKind POK, 1827 StringRef DiagKind); 1828 void warnIfMutexHeld(const NamedDecl *D, const Expr *Exp, Expr *MutexExp, 1829 StringRef DiagKind); 1830 1831 void checkAccess(const Expr *Exp, AccessKind AK); 1832 void checkPtAccess(const Expr *Exp, AccessKind AK); 1833 1834 void handleCall(Expr *Exp, const NamedDecl *D, VarDecl *VD = 0); 1835 1836 public: 1837 BuildLockset(ThreadSafetyAnalyzer *Anlzr, CFGBlockInfo &Info) 1838 : StmtVisitor<BuildLockset>(), 1839 Analyzer(Anlzr), 1840 FSet(Info.EntrySet), 1841 LVarCtx(Info.EntryContext), 1842 CtxIndex(Info.EntryIndex) 1843 {} 1844 1845 void VisitUnaryOperator(UnaryOperator *UO); 1846 void VisitBinaryOperator(BinaryOperator *BO); 1847 void VisitCastExpr(CastExpr *CE); 1848 void VisitCallExpr(CallExpr *Exp); 1849 void VisitCXXConstructExpr(CXXConstructExpr *Exp); 1850 void VisitDeclStmt(DeclStmt *S); 1851 }; 1852 1853 /// \brief Warn if the LSet does not contain a lock sufficient to protect access 1854 /// of at least the passed in AccessKind. 1855 void BuildLockset::warnIfMutexNotHeld(const NamedDecl *D, const Expr *Exp, 1856 AccessKind AK, Expr *MutexExp, 1857 ProtectedOperationKind POK, 1858 StringRef DiagKind) { 1859 LockKind LK = getLockKindFromAccessKind(AK); 1860 1861 SExpr Mutex(MutexExp, Exp, D); 1862 if (!Mutex.isValid()) { 1863 SExpr::warnInvalidLock(Analyzer->Handler, MutexExp, Exp, D, DiagKind); 1864 return; 1865 } else if (Mutex.shouldIgnore()) { 1866 return; 1867 } 1868 1869 LockData* LDat = FSet.findLockUniv(Analyzer->FactMan, Mutex); 1870 bool NoError = true; 1871 if (!LDat) { 1872 // No exact match found. Look for a partial match. 1873 FactEntry* FEntry = FSet.findPartialMatch(Analyzer->FactMan, Mutex); 1874 if (FEntry) { 1875 // Warn that there's no precise match. 1876 LDat = &FEntry->LDat; 1877 std::string PartMatchStr = FEntry->MutID.toString(); 1878 StringRef PartMatchName(PartMatchStr); 1879 Analyzer->Handler.handleMutexNotHeld(DiagKind, D, POK, Mutex.toString(), 1880 LK, Exp->getExprLoc(), 1881 &PartMatchName); 1882 } else { 1883 // Warn that there's no match at all. 1884 Analyzer->Handler.handleMutexNotHeld(DiagKind, D, POK, Mutex.toString(), 1885 LK, Exp->getExprLoc()); 1886 } 1887 NoError = false; 1888 } 1889 // Make sure the mutex we found is the right kind. 1890 if (NoError && LDat && !LDat->isAtLeast(LK)) 1891 Analyzer->Handler.handleMutexNotHeld(DiagKind, D, POK, Mutex.toString(), LK, 1892 Exp->getExprLoc()); 1893 } 1894 1895 /// \brief Warn if the LSet contains the given lock. 1896 void BuildLockset::warnIfMutexHeld(const NamedDecl *D, const Expr *Exp, 1897 Expr *MutexExp, 1898 StringRef DiagKind) { 1899 SExpr Mutex(MutexExp, Exp, D); 1900 if (!Mutex.isValid()) { 1901 SExpr::warnInvalidLock(Analyzer->Handler, MutexExp, Exp, D, DiagKind); 1902 return; 1903 } 1904 1905 LockData* LDat = FSet.findLock(Analyzer->FactMan, Mutex); 1906 if (LDat) 1907 Analyzer->Handler.handleFunExcludesLock( 1908 DiagKind, D->getNameAsString(), Mutex.toString(), Exp->getExprLoc()); 1909 } 1910 1911 /// \brief Checks guarded_by and pt_guarded_by attributes. 1912 /// Whenever we identify an access (read or write) to a DeclRefExpr that is 1913 /// marked with guarded_by, we must ensure the appropriate mutexes are held. 1914 /// Similarly, we check if the access is to an expression that dereferences 1915 /// a pointer marked with pt_guarded_by. 1916 void BuildLockset::checkAccess(const Expr *Exp, AccessKind AK) { 1917 Exp = Exp->IgnoreParenCasts(); 1918 1919 if (const UnaryOperator *UO = dyn_cast<UnaryOperator>(Exp)) { 1920 // For dereferences 1921 if (UO->getOpcode() == clang::UO_Deref) 1922 checkPtAccess(UO->getSubExpr(), AK); 1923 return; 1924 } 1925 1926 if (const ArraySubscriptExpr *AE = dyn_cast<ArraySubscriptExpr>(Exp)) { 1927 checkPtAccess(AE->getLHS(), AK); 1928 return; 1929 } 1930 1931 if (const MemberExpr *ME = dyn_cast<MemberExpr>(Exp)) { 1932 if (ME->isArrow()) 1933 checkPtAccess(ME->getBase(), AK); 1934 else 1935 checkAccess(ME->getBase(), AK); 1936 } 1937 1938 const ValueDecl *D = getValueDecl(Exp); 1939 if (!D || !D->hasAttrs()) 1940 return; 1941 1942 if (D->hasAttr<GuardedVarAttr>() && FSet.isEmpty()) 1943 Analyzer->Handler.handleNoMutexHeld("mutex", D, POK_VarAccess, AK, 1944 Exp->getExprLoc()); 1945 1946 for (const auto *I : D->specific_attrs<GuardedByAttr>()) 1947 warnIfMutexNotHeld(D, Exp, AK, I->getArg(), POK_VarAccess, 1948 ClassifyDiagnostic(I)); 1949 } 1950 1951 /// \brief Checks pt_guarded_by and pt_guarded_var attributes. 1952 void BuildLockset::checkPtAccess(const Expr *Exp, AccessKind AK) { 1953 while (true) { 1954 if (const ParenExpr *PE = dyn_cast<ParenExpr>(Exp)) { 1955 Exp = PE->getSubExpr(); 1956 continue; 1957 } 1958 if (const CastExpr *CE = dyn_cast<CastExpr>(Exp)) { 1959 if (CE->getCastKind() == CK_ArrayToPointerDecay) { 1960 // If it's an actual array, and not a pointer, then it's elements 1961 // are protected by GUARDED_BY, not PT_GUARDED_BY; 1962 checkAccess(CE->getSubExpr(), AK); 1963 return; 1964 } 1965 Exp = CE->getSubExpr(); 1966 continue; 1967 } 1968 break; 1969 } 1970 1971 const ValueDecl *D = getValueDecl(Exp); 1972 if (!D || !D->hasAttrs()) 1973 return; 1974 1975 if (D->hasAttr<PtGuardedVarAttr>() && FSet.isEmpty()) 1976 Analyzer->Handler.handleNoMutexHeld("mutex", D, POK_VarDereference, AK, 1977 Exp->getExprLoc()); 1978 1979 for (auto const *I : D->specific_attrs<PtGuardedByAttr>()) 1980 warnIfMutexNotHeld(D, Exp, AK, I->getArg(), POK_VarDereference, 1981 ClassifyDiagnostic(I)); 1982 } 1983 1984 /// \brief Process a function call, method call, constructor call, 1985 /// or destructor call. This involves looking at the attributes on the 1986 /// corresponding function/method/constructor/destructor, issuing warnings, 1987 /// and updating the locksets accordingly. 1988 /// 1989 /// FIXME: For classes annotated with one of the guarded annotations, we need 1990 /// to treat const method calls as reads and non-const method calls as writes, 1991 /// and check that the appropriate locks are held. Non-const method calls with 1992 /// the same signature as const method calls can be also treated as reads. 1993 /// 1994 void BuildLockset::handleCall(Expr *Exp, const NamedDecl *D, VarDecl *VD) { 1995 SourceLocation Loc = Exp->getExprLoc(); 1996 const AttrVec &ArgAttrs = D->getAttrs(); 1997 MutexIDList ExclusiveLocksToAdd, SharedLocksToAdd; 1998 MutexIDList ExclusiveLocksToRemove, SharedLocksToRemove, GenericLocksToRemove; 1999 StringRef CapDiagKind = "mutex"; 2000 2001 for(unsigned i = 0; i < ArgAttrs.size(); ++i) { 2002 Attr *At = const_cast<Attr*>(ArgAttrs[i]); 2003 switch (At->getKind()) { 2004 // When we encounter a lock function, we need to add the lock to our 2005 // lockset. 2006 case attr::AcquireCapability: { 2007 auto *A = cast<AcquireCapabilityAttr>(At); 2008 Analyzer->getMutexIDs(A->isShared() ? SharedLocksToAdd 2009 : ExclusiveLocksToAdd, 2010 A, Exp, D, VD); 2011 2012 CapDiagKind = ClassifyDiagnostic(A); 2013 break; 2014 } 2015 2016 // An assert will add a lock to the lockset, but will not generate 2017 // a warning if it is already there, and will not generate a warning 2018 // if it is not removed. 2019 case attr::AssertExclusiveLock: { 2020 AssertExclusiveLockAttr *A = cast<AssertExclusiveLockAttr>(At); 2021 2022 MutexIDList AssertLocks; 2023 Analyzer->getMutexIDs(AssertLocks, A, Exp, D, VD); 2024 for (const auto &AssertLock : AssertLocks) 2025 Analyzer->addLock(FSet, AssertLock, 2026 LockData(Loc, LK_Exclusive, false, true), 2027 ClassifyDiagnostic(A)); 2028 break; 2029 } 2030 case attr::AssertSharedLock: { 2031 AssertSharedLockAttr *A = cast<AssertSharedLockAttr>(At); 2032 2033 MutexIDList AssertLocks; 2034 Analyzer->getMutexIDs(AssertLocks, A, Exp, D, VD); 2035 for (const auto &AssertLock : AssertLocks) 2036 Analyzer->addLock(FSet, AssertLock, 2037 LockData(Loc, LK_Shared, false, true), 2038 ClassifyDiagnostic(A)); 2039 break; 2040 } 2041 2042 // When we encounter an unlock function, we need to remove unlocked 2043 // mutexes from the lockset, and flag a warning if they are not there. 2044 case attr::ReleaseCapability: { 2045 auto *A = cast<ReleaseCapabilityAttr>(At); 2046 if (A->isGeneric()) 2047 Analyzer->getMutexIDs(GenericLocksToRemove, A, Exp, D, VD); 2048 else if (A->isShared()) 2049 Analyzer->getMutexIDs(SharedLocksToRemove, A, Exp, D, VD); 2050 else 2051 Analyzer->getMutexIDs(ExclusiveLocksToRemove, A, Exp, D, VD); 2052 2053 CapDiagKind = ClassifyDiagnostic(A); 2054 break; 2055 } 2056 2057 case attr::RequiresCapability: { 2058 RequiresCapabilityAttr *A = cast<RequiresCapabilityAttr>(At); 2059 for (auto *Arg : A->args()) 2060 warnIfMutexNotHeld(D, Exp, A->isShared() ? AK_Read : AK_Written, Arg, 2061 POK_FunctionCall, ClassifyDiagnostic(A)); 2062 break; 2063 } 2064 2065 case attr::LocksExcluded: { 2066 LocksExcludedAttr *A = cast<LocksExcludedAttr>(At); 2067 for (auto *Arg : A->args()) 2068 warnIfMutexHeld(D, Exp, Arg, ClassifyDiagnostic(A)); 2069 break; 2070 } 2071 2072 // Ignore attributes unrelated to thread-safety 2073 default: 2074 break; 2075 } 2076 } 2077 2078 // Figure out if we're calling the constructor of scoped lockable class 2079 bool isScopedVar = false; 2080 if (VD) { 2081 if (const CXXConstructorDecl *CD = dyn_cast<const CXXConstructorDecl>(D)) { 2082 const CXXRecordDecl* PD = CD->getParent(); 2083 if (PD && PD->hasAttr<ScopedLockableAttr>()) 2084 isScopedVar = true; 2085 } 2086 } 2087 2088 // Add locks. 2089 for (const auto &M : ExclusiveLocksToAdd) 2090 Analyzer->addLock(FSet, M, LockData(Loc, LK_Exclusive, isScopedVar), 2091 CapDiagKind); 2092 for (const auto &M : SharedLocksToAdd) 2093 Analyzer->addLock(FSet, M, LockData(Loc, LK_Shared, isScopedVar), 2094 CapDiagKind); 2095 2096 // Add the managing object as a dummy mutex, mapped to the underlying mutex. 2097 // FIXME -- this doesn't work if we acquire multiple locks. 2098 if (isScopedVar) { 2099 SourceLocation MLoc = VD->getLocation(); 2100 DeclRefExpr DRE(VD, false, VD->getType(), VK_LValue, VD->getLocation()); 2101 SExpr SMutex(&DRE, 0, 0); 2102 2103 for (const auto &M : ExclusiveLocksToAdd) 2104 Analyzer->addLock(FSet, SMutex, LockData(MLoc, LK_Exclusive, M), 2105 CapDiagKind); 2106 for (const auto &M : SharedLocksToAdd) 2107 Analyzer->addLock(FSet, SMutex, LockData(MLoc, LK_Shared, M), 2108 CapDiagKind); 2109 } 2110 2111 // Remove locks. 2112 // FIXME -- should only fully remove if the attribute refers to 'this'. 2113 bool Dtor = isa<CXXDestructorDecl>(D); 2114 for (const auto &M : ExclusiveLocksToRemove) 2115 Analyzer->removeLock(FSet, M, Loc, Dtor, LK_Exclusive, CapDiagKind); 2116 for (const auto &M : SharedLocksToRemove) 2117 Analyzer->removeLock(FSet, M, Loc, Dtor, LK_Shared, CapDiagKind); 2118 for (const auto &M : GenericLocksToRemove) 2119 Analyzer->removeLock(FSet, M, Loc, Dtor, LK_Generic, CapDiagKind); 2120 } 2121 2122 2123 /// \brief For unary operations which read and write a variable, we need to 2124 /// check whether we hold any required mutexes. Reads are checked in 2125 /// VisitCastExpr. 2126 void BuildLockset::VisitUnaryOperator(UnaryOperator *UO) { 2127 switch (UO->getOpcode()) { 2128 case clang::UO_PostDec: 2129 case clang::UO_PostInc: 2130 case clang::UO_PreDec: 2131 case clang::UO_PreInc: { 2132 checkAccess(UO->getSubExpr(), AK_Written); 2133 break; 2134 } 2135 default: 2136 break; 2137 } 2138 } 2139 2140 /// For binary operations which assign to a variable (writes), we need to check 2141 /// whether we hold any required mutexes. 2142 /// FIXME: Deal with non-primitive types. 2143 void BuildLockset::VisitBinaryOperator(BinaryOperator *BO) { 2144 if (!BO->isAssignmentOp()) 2145 return; 2146 2147 // adjust the context 2148 LVarCtx = Analyzer->LocalVarMap.getNextContext(CtxIndex, BO, LVarCtx); 2149 2150 checkAccess(BO->getLHS(), AK_Written); 2151 } 2152 2153 2154 /// Whenever we do an LValue to Rvalue cast, we are reading a variable and 2155 /// need to ensure we hold any required mutexes. 2156 /// FIXME: Deal with non-primitive types. 2157 void BuildLockset::VisitCastExpr(CastExpr *CE) { 2158 if (CE->getCastKind() != CK_LValueToRValue) 2159 return; 2160 checkAccess(CE->getSubExpr(), AK_Read); 2161 } 2162 2163 2164 void BuildLockset::VisitCallExpr(CallExpr *Exp) { 2165 if (CXXMemberCallExpr *CE = dyn_cast<CXXMemberCallExpr>(Exp)) { 2166 MemberExpr *ME = dyn_cast<MemberExpr>(CE->getCallee()); 2167 // ME can be null when calling a method pointer 2168 CXXMethodDecl *MD = CE->getMethodDecl(); 2169 2170 if (ME && MD) { 2171 if (ME->isArrow()) { 2172 if (MD->isConst()) { 2173 checkPtAccess(CE->getImplicitObjectArgument(), AK_Read); 2174 } else { // FIXME -- should be AK_Written 2175 checkPtAccess(CE->getImplicitObjectArgument(), AK_Read); 2176 } 2177 } else { 2178 if (MD->isConst()) 2179 checkAccess(CE->getImplicitObjectArgument(), AK_Read); 2180 else // FIXME -- should be AK_Written 2181 checkAccess(CE->getImplicitObjectArgument(), AK_Read); 2182 } 2183 } 2184 } else if (CXXOperatorCallExpr *OE = dyn_cast<CXXOperatorCallExpr>(Exp)) { 2185 switch (OE->getOperator()) { 2186 case OO_Equal: { 2187 const Expr *Target = OE->getArg(0); 2188 const Expr *Source = OE->getArg(1); 2189 checkAccess(Target, AK_Written); 2190 checkAccess(Source, AK_Read); 2191 break; 2192 } 2193 case OO_Star: 2194 case OO_Arrow: 2195 case OO_Subscript: { 2196 const Expr *Obj = OE->getArg(0); 2197 checkAccess(Obj, AK_Read); 2198 checkPtAccess(Obj, AK_Read); 2199 break; 2200 } 2201 default: { 2202 const Expr *Obj = OE->getArg(0); 2203 checkAccess(Obj, AK_Read); 2204 break; 2205 } 2206 } 2207 } 2208 NamedDecl *D = dyn_cast_or_null<NamedDecl>(Exp->getCalleeDecl()); 2209 if(!D || !D->hasAttrs()) 2210 return; 2211 handleCall(Exp, D); 2212 } 2213 2214 void BuildLockset::VisitCXXConstructExpr(CXXConstructExpr *Exp) { 2215 const CXXConstructorDecl *D = Exp->getConstructor(); 2216 if (D && D->isCopyConstructor()) { 2217 const Expr* Source = Exp->getArg(0); 2218 checkAccess(Source, AK_Read); 2219 } 2220 // FIXME -- only handles constructors in DeclStmt below. 2221 } 2222 2223 void BuildLockset::VisitDeclStmt(DeclStmt *S) { 2224 // adjust the context 2225 LVarCtx = Analyzer->LocalVarMap.getNextContext(CtxIndex, S, LVarCtx); 2226 2227 DeclGroupRef DGrp = S->getDeclGroup(); 2228 for (DeclGroupRef::iterator I = DGrp.begin(), E = DGrp.end(); I != E; ++I) { 2229 Decl *D = *I; 2230 if (VarDecl *VD = dyn_cast_or_null<VarDecl>(D)) { 2231 Expr *E = VD->getInit(); 2232 // handle constructors that involve temporaries 2233 if (ExprWithCleanups *EWC = dyn_cast_or_null<ExprWithCleanups>(E)) 2234 E = EWC->getSubExpr(); 2235 2236 if (CXXConstructExpr *CE = dyn_cast_or_null<CXXConstructExpr>(E)) { 2237 NamedDecl *CtorD = dyn_cast_or_null<NamedDecl>(CE->getConstructor()); 2238 if (!CtorD || !CtorD->hasAttrs()) 2239 return; 2240 handleCall(CE, CtorD, VD); 2241 } 2242 } 2243 } 2244 } 2245 2246 2247 2248 /// \brief Compute the intersection of two locksets and issue warnings for any 2249 /// locks in the symmetric difference. 2250 /// 2251 /// This function is used at a merge point in the CFG when comparing the lockset 2252 /// of each branch being merged. For example, given the following sequence: 2253 /// A; if () then B; else C; D; we need to check that the lockset after B and C 2254 /// are the same. In the event of a difference, we use the intersection of these 2255 /// two locksets at the start of D. 2256 /// 2257 /// \param FSet1 The first lockset. 2258 /// \param FSet2 The second lockset. 2259 /// \param JoinLoc The location of the join point for error reporting 2260 /// \param LEK1 The error message to report if a mutex is missing from LSet1 2261 /// \param LEK2 The error message to report if a mutex is missing from Lset2 2262 void ThreadSafetyAnalyzer::intersectAndWarn(FactSet &FSet1, 2263 const FactSet &FSet2, 2264 SourceLocation JoinLoc, 2265 LockErrorKind LEK1, 2266 LockErrorKind LEK2, 2267 bool Modify) { 2268 FactSet FSet1Orig = FSet1; 2269 2270 // Find locks in FSet2 that conflict or are not in FSet1, and warn. 2271 for (FactSet::const_iterator I = FSet2.begin(), E = FSet2.end(); 2272 I != E; ++I) { 2273 const SExpr &FSet2Mutex = FactMan[*I].MutID; 2274 const LockData &LDat2 = FactMan[*I].LDat; 2275 FactSet::iterator I1 = FSet1.findLockIter(FactMan, FSet2Mutex); 2276 2277 if (I1 != FSet1.end()) { 2278 const LockData* LDat1 = &FactMan[*I1].LDat; 2279 if (LDat1->LKind != LDat2.LKind) { 2280 Handler.handleExclusiveAndShared("mutex", FSet2Mutex.toString(), 2281 LDat2.AcquireLoc, LDat1->AcquireLoc); 2282 if (Modify && LDat1->LKind != LK_Exclusive) { 2283 // Take the exclusive lock, which is the one in FSet2. 2284 *I1 = *I; 2285 } 2286 } 2287 else if (LDat1->Asserted && !LDat2.Asserted) { 2288 // The non-asserted lock in FSet2 is the one we want to track. 2289 *I1 = *I; 2290 } 2291 } else { 2292 if (LDat2.UnderlyingMutex.isValid()) { 2293 if (FSet2.findLock(FactMan, LDat2.UnderlyingMutex)) { 2294 // If this is a scoped lock that manages another mutex, and if the 2295 // underlying mutex is still held, then warn about the underlying 2296 // mutex. 2297 Handler.handleMutexHeldEndOfScope("mutex", 2298 LDat2.UnderlyingMutex.toString(), 2299 LDat2.AcquireLoc, JoinLoc, LEK1); 2300 } 2301 } 2302 else if (!LDat2.Managed && !FSet2Mutex.isUniversal() && !LDat2.Asserted) 2303 Handler.handleMutexHeldEndOfScope("mutex", FSet2Mutex.toString(), 2304 LDat2.AcquireLoc, JoinLoc, LEK1); 2305 } 2306 } 2307 2308 // Find locks in FSet1 that are not in FSet2, and remove them. 2309 for (FactSet::const_iterator I = FSet1Orig.begin(), E = FSet1Orig.end(); 2310 I != E; ++I) { 2311 const SExpr &FSet1Mutex = FactMan[*I].MutID; 2312 const LockData &LDat1 = FactMan[*I].LDat; 2313 2314 if (!FSet2.findLock(FactMan, FSet1Mutex)) { 2315 if (LDat1.UnderlyingMutex.isValid()) { 2316 if (FSet1Orig.findLock(FactMan, LDat1.UnderlyingMutex)) { 2317 // If this is a scoped lock that manages another mutex, and if the 2318 // underlying mutex is still held, then warn about the underlying 2319 // mutex. 2320 Handler.handleMutexHeldEndOfScope("mutex", 2321 LDat1.UnderlyingMutex.toString(), 2322 LDat1.AcquireLoc, JoinLoc, LEK1); 2323 } 2324 } 2325 else if (!LDat1.Managed && !FSet1Mutex.isUniversal() && !LDat1.Asserted) 2326 Handler.handleMutexHeldEndOfScope("mutex", FSet1Mutex.toString(), 2327 LDat1.AcquireLoc, JoinLoc, LEK2); 2328 if (Modify) 2329 FSet1.removeLock(FactMan, FSet1Mutex); 2330 } 2331 } 2332 } 2333 2334 2335 // Return true if block B never continues to its successors. 2336 inline bool neverReturns(const CFGBlock* B) { 2337 if (B->hasNoReturnElement()) 2338 return true; 2339 if (B->empty()) 2340 return false; 2341 2342 CFGElement Last = B->back(); 2343 if (Optional<CFGStmt> S = Last.getAs<CFGStmt>()) { 2344 if (isa<CXXThrowExpr>(S->getStmt())) 2345 return true; 2346 } 2347 return false; 2348 } 2349 2350 2351 /// \brief Check a function's CFG for thread-safety violations. 2352 /// 2353 /// We traverse the blocks in the CFG, compute the set of mutexes that are held 2354 /// at the end of each block, and issue warnings for thread safety violations. 2355 /// Each block in the CFG is traversed exactly once. 2356 void ThreadSafetyAnalyzer::runAnalysis(AnalysisDeclContext &AC) { 2357 // TODO: this whole function needs be rewritten as a visitor for CFGWalker. 2358 // For now, we just use the walker to set things up. 2359 threadSafety::CFGWalker walker; 2360 if (!walker.init(AC)) 2361 return; 2362 2363 // AC.dumpCFG(true); 2364 // threadSafety::printSCFG(walker); 2365 2366 CFG *CFGraph = walker.getGraph(); 2367 const NamedDecl *D = walker.getDecl(); 2368 2369 if (D->hasAttr<NoThreadSafetyAnalysisAttr>()) 2370 return; 2371 2372 // FIXME: Do something a bit more intelligent inside constructor and 2373 // destructor code. Constructors and destructors must assume unique access 2374 // to 'this', so checks on member variable access is disabled, but we should 2375 // still enable checks on other objects. 2376 if (isa<CXXConstructorDecl>(D)) 2377 return; // Don't check inside constructors. 2378 if (isa<CXXDestructorDecl>(D)) 2379 return; // Don't check inside destructors. 2380 2381 BlockInfo.resize(CFGraph->getNumBlockIDs(), 2382 CFGBlockInfo::getEmptyBlockInfo(LocalVarMap)); 2383 2384 // We need to explore the CFG via a "topological" ordering. 2385 // That way, we will be guaranteed to have information about required 2386 // predecessor locksets when exploring a new block. 2387 const PostOrderCFGView *SortedGraph = walker.getSortedGraph(); 2388 PostOrderCFGView::CFGBlockSet VisitedBlocks(CFGraph); 2389 2390 // Mark entry block as reachable 2391 BlockInfo[CFGraph->getEntry().getBlockID()].Reachable = true; 2392 2393 // Compute SSA names for local variables 2394 LocalVarMap.traverseCFG(CFGraph, SortedGraph, BlockInfo); 2395 2396 // Fill in source locations for all CFGBlocks. 2397 findBlockLocations(CFGraph, SortedGraph, BlockInfo); 2398 2399 MutexIDList ExclusiveLocksAcquired; 2400 MutexIDList SharedLocksAcquired; 2401 MutexIDList LocksReleased; 2402 2403 // Add locks from exclusive_locks_required and shared_locks_required 2404 // to initial lockset. Also turn off checking for lock and unlock functions. 2405 // FIXME: is there a more intelligent way to check lock/unlock functions? 2406 if (!SortedGraph->empty() && D->hasAttrs()) { 2407 const CFGBlock *FirstBlock = *SortedGraph->begin(); 2408 FactSet &InitialLockset = BlockInfo[FirstBlock->getBlockID()].EntrySet; 2409 const AttrVec &ArgAttrs = D->getAttrs(); 2410 2411 MutexIDList ExclusiveLocksToAdd; 2412 MutexIDList SharedLocksToAdd; 2413 StringRef CapDiagKind = "mutex"; 2414 2415 SourceLocation Loc = D->getLocation(); 2416 for (const auto *Attr : ArgAttrs) { 2417 Loc = Attr->getLocation(); 2418 if (const auto *A = dyn_cast<RequiresCapabilityAttr>(Attr)) { 2419 getMutexIDs(A->isShared() ? SharedLocksToAdd : ExclusiveLocksToAdd, A, 2420 0, D); 2421 CapDiagKind = ClassifyDiagnostic(A); 2422 } else if (const auto *A = dyn_cast<ReleaseCapabilityAttr>(Attr)) { 2423 // UNLOCK_FUNCTION() is used to hide the underlying lock implementation. 2424 // We must ignore such methods. 2425 if (A->args_size() == 0) 2426 return; 2427 // FIXME -- deal with exclusive vs. shared unlock functions? 2428 getMutexIDs(ExclusiveLocksToAdd, A, nullptr, D); 2429 getMutexIDs(LocksReleased, A, nullptr, D); 2430 CapDiagKind = ClassifyDiagnostic(A); 2431 } else if (const auto *A = dyn_cast<AcquireCapabilityAttr>(Attr)) { 2432 if (A->args_size() == 0) 2433 return; 2434 getMutexIDs(A->isShared() ? SharedLocksAcquired 2435 : ExclusiveLocksAcquired, 2436 A, nullptr, D); 2437 CapDiagKind = ClassifyDiagnostic(A); 2438 } else if (isa<ExclusiveTrylockFunctionAttr>(Attr)) { 2439 // Don't try to check trylock functions for now 2440 return; 2441 } else if (isa<SharedTrylockFunctionAttr>(Attr)) { 2442 // Don't try to check trylock functions for now 2443 return; 2444 } 2445 } 2446 2447 // FIXME -- Loc can be wrong here. 2448 for (const auto &ExclusiveLockToAdd : ExclusiveLocksToAdd) 2449 addLock(InitialLockset, ExclusiveLockToAdd, LockData(Loc, LK_Exclusive), 2450 CapDiagKind); 2451 for (const auto &SharedLockToAdd : SharedLocksToAdd) 2452 addLock(InitialLockset, SharedLockToAdd, LockData(Loc, LK_Shared), 2453 CapDiagKind); 2454 } 2455 2456 for (const auto *CurrBlock : *SortedGraph) { 2457 int CurrBlockID = CurrBlock->getBlockID(); 2458 CFGBlockInfo *CurrBlockInfo = &BlockInfo[CurrBlockID]; 2459 2460 // Use the default initial lockset in case there are no predecessors. 2461 VisitedBlocks.insert(CurrBlock); 2462 2463 // Iterate through the predecessor blocks and warn if the lockset for all 2464 // predecessors is not the same. We take the entry lockset of the current 2465 // block to be the intersection of all previous locksets. 2466 // FIXME: By keeping the intersection, we may output more errors in future 2467 // for a lock which is not in the intersection, but was in the union. We 2468 // may want to also keep the union in future. As an example, let's say 2469 // the intersection contains Mutex L, and the union contains L and M. 2470 // Later we unlock M. At this point, we would output an error because we 2471 // never locked M; although the real error is probably that we forgot to 2472 // lock M on all code paths. Conversely, let's say that later we lock M. 2473 // In this case, we should compare against the intersection instead of the 2474 // union because the real error is probably that we forgot to unlock M on 2475 // all code paths. 2476 bool LocksetInitialized = false; 2477 SmallVector<CFGBlock *, 8> SpecialBlocks; 2478 for (CFGBlock::const_pred_iterator PI = CurrBlock->pred_begin(), 2479 PE = CurrBlock->pred_end(); PI != PE; ++PI) { 2480 2481 // if *PI -> CurrBlock is a back edge 2482 if (*PI == nullptr || !VisitedBlocks.alreadySet(*PI)) 2483 continue; 2484 2485 int PrevBlockID = (*PI)->getBlockID(); 2486 CFGBlockInfo *PrevBlockInfo = &BlockInfo[PrevBlockID]; 2487 2488 // Ignore edges from blocks that can't return. 2489 if (neverReturns(*PI) || !PrevBlockInfo->Reachable) 2490 continue; 2491 2492 // Okay, we can reach this block from the entry. 2493 CurrBlockInfo->Reachable = true; 2494 2495 // If the previous block ended in a 'continue' or 'break' statement, then 2496 // a difference in locksets is probably due to a bug in that block, rather 2497 // than in some other predecessor. In that case, keep the other 2498 // predecessor's lockset. 2499 if (const Stmt *Terminator = (*PI)->getTerminator()) { 2500 if (isa<ContinueStmt>(Terminator) || isa<BreakStmt>(Terminator)) { 2501 SpecialBlocks.push_back(*PI); 2502 continue; 2503 } 2504 } 2505 2506 FactSet PrevLockset; 2507 getEdgeLockset(PrevLockset, PrevBlockInfo->ExitSet, *PI, CurrBlock); 2508 2509 if (!LocksetInitialized) { 2510 CurrBlockInfo->EntrySet = PrevLockset; 2511 LocksetInitialized = true; 2512 } else { 2513 intersectAndWarn(CurrBlockInfo->EntrySet, PrevLockset, 2514 CurrBlockInfo->EntryLoc, 2515 LEK_LockedSomePredecessors); 2516 } 2517 } 2518 2519 // Skip rest of block if it's not reachable. 2520 if (!CurrBlockInfo->Reachable) 2521 continue; 2522 2523 // Process continue and break blocks. Assume that the lockset for the 2524 // resulting block is unaffected by any discrepancies in them. 2525 for (const auto *PrevBlock : SpecialBlocks) { 2526 int PrevBlockID = PrevBlock->getBlockID(); 2527 CFGBlockInfo *PrevBlockInfo = &BlockInfo[PrevBlockID]; 2528 2529 if (!LocksetInitialized) { 2530 CurrBlockInfo->EntrySet = PrevBlockInfo->ExitSet; 2531 LocksetInitialized = true; 2532 } else { 2533 // Determine whether this edge is a loop terminator for diagnostic 2534 // purposes. FIXME: A 'break' statement might be a loop terminator, but 2535 // it might also be part of a switch. Also, a subsequent destructor 2536 // might add to the lockset, in which case the real issue might be a 2537 // double lock on the other path. 2538 const Stmt *Terminator = PrevBlock->getTerminator(); 2539 bool IsLoop = Terminator && isa<ContinueStmt>(Terminator); 2540 2541 FactSet PrevLockset; 2542 getEdgeLockset(PrevLockset, PrevBlockInfo->ExitSet, 2543 PrevBlock, CurrBlock); 2544 2545 // Do not update EntrySet. 2546 intersectAndWarn(CurrBlockInfo->EntrySet, PrevLockset, 2547 PrevBlockInfo->ExitLoc, 2548 IsLoop ? LEK_LockedSomeLoopIterations 2549 : LEK_LockedSomePredecessors, 2550 false); 2551 } 2552 } 2553 2554 BuildLockset LocksetBuilder(this, *CurrBlockInfo); 2555 2556 // Visit all the statements in the basic block. 2557 for (CFGBlock::const_iterator BI = CurrBlock->begin(), 2558 BE = CurrBlock->end(); BI != BE; ++BI) { 2559 switch (BI->getKind()) { 2560 case CFGElement::Statement: { 2561 CFGStmt CS = BI->castAs<CFGStmt>(); 2562 LocksetBuilder.Visit(const_cast<Stmt*>(CS.getStmt())); 2563 break; 2564 } 2565 // Ignore BaseDtor, MemberDtor, and TemporaryDtor for now. 2566 case CFGElement::AutomaticObjectDtor: { 2567 CFGAutomaticObjDtor AD = BI->castAs<CFGAutomaticObjDtor>(); 2568 CXXDestructorDecl *DD = const_cast<CXXDestructorDecl *>( 2569 AD.getDestructorDecl(AC.getASTContext())); 2570 if (!DD->hasAttrs()) 2571 break; 2572 2573 // Create a dummy expression, 2574 VarDecl *VD = const_cast<VarDecl*>(AD.getVarDecl()); 2575 DeclRefExpr DRE(VD, false, VD->getType(), VK_LValue, 2576 AD.getTriggerStmt()->getLocEnd()); 2577 LocksetBuilder.handleCall(&DRE, DD); 2578 break; 2579 } 2580 default: 2581 break; 2582 } 2583 } 2584 CurrBlockInfo->ExitSet = LocksetBuilder.FSet; 2585 2586 // For every back edge from CurrBlock (the end of the loop) to another block 2587 // (FirstLoopBlock) we need to check that the Lockset of Block is equal to 2588 // the one held at the beginning of FirstLoopBlock. We can look up the 2589 // Lockset held at the beginning of FirstLoopBlock in the EntryLockSets map. 2590 for (CFGBlock::const_succ_iterator SI = CurrBlock->succ_begin(), 2591 SE = CurrBlock->succ_end(); SI != SE; ++SI) { 2592 2593 // if CurrBlock -> *SI is *not* a back edge 2594 if (*SI == 0 || !VisitedBlocks.alreadySet(*SI)) 2595 continue; 2596 2597 CFGBlock *FirstLoopBlock = *SI; 2598 CFGBlockInfo *PreLoop = &BlockInfo[FirstLoopBlock->getBlockID()]; 2599 CFGBlockInfo *LoopEnd = &BlockInfo[CurrBlockID]; 2600 intersectAndWarn(LoopEnd->ExitSet, PreLoop->EntrySet, 2601 PreLoop->EntryLoc, 2602 LEK_LockedSomeLoopIterations, 2603 false); 2604 } 2605 } 2606 2607 CFGBlockInfo *Initial = &BlockInfo[CFGraph->getEntry().getBlockID()]; 2608 CFGBlockInfo *Final = &BlockInfo[CFGraph->getExit().getBlockID()]; 2609 2610 // Skip the final check if the exit block is unreachable. 2611 if (!Final->Reachable) 2612 return; 2613 2614 // By default, we expect all locks held on entry to be held on exit. 2615 FactSet ExpectedExitSet = Initial->EntrySet; 2616 2617 // Adjust the expected exit set by adding or removing locks, as declared 2618 // by *-LOCK_FUNCTION and UNLOCK_FUNCTION. The intersect below will then 2619 // issue the appropriate warning. 2620 // FIXME: the location here is not quite right. 2621 for (const auto &Lock : ExclusiveLocksAcquired) 2622 ExpectedExitSet.addLock(FactMan, Lock, 2623 LockData(D->getLocation(), LK_Exclusive)); 2624 for (const auto &Lock : SharedLocksAcquired) 2625 ExpectedExitSet.addLock(FactMan, Lock, 2626 LockData(D->getLocation(), LK_Shared)); 2627 for (const auto &Lock : LocksReleased) 2628 ExpectedExitSet.removeLock(FactMan, Lock); 2629 2630 // FIXME: Should we call this function for all blocks which exit the function? 2631 intersectAndWarn(ExpectedExitSet, Final->ExitSet, 2632 Final->ExitLoc, 2633 LEK_LockedAtEndOfFunction, 2634 LEK_NotLockedAtEndOfFunction, 2635 false); 2636 } 2637 2638 } // end anonymous namespace 2639 2640 2641 namespace clang { 2642 namespace thread_safety { 2643 2644 /// \brief Check a function's CFG for thread-safety violations. 2645 /// 2646 /// We traverse the blocks in the CFG, compute the set of mutexes that are held 2647 /// at the end of each block, and issue warnings for thread safety violations. 2648 /// Each block in the CFG is traversed exactly once. 2649 void runThreadSafetyAnalysis(AnalysisDeclContext &AC, 2650 ThreadSafetyHandler &Handler) { 2651 ThreadSafetyAnalyzer Analyzer(Handler); 2652 Analyzer.runAnalysis(AC); 2653 } 2654 2655 /// \brief Helper function that returns a LockKind required for the given level 2656 /// of access. 2657 LockKind getLockKindFromAccessKind(AccessKind AK) { 2658 switch (AK) { 2659 case AK_Read : 2660 return LK_Shared; 2661 case AK_Written : 2662 return LK_Exclusive; 2663 } 2664 llvm_unreachable("Unknown AccessKind"); 2665 } 2666 2667 }} // end namespace clang::thread_safety 2668