ThreadSafety.cpp revision 589190b322a255c8e8b15dfde6b77ef498db7548
1//===- ThreadSafety.cpp ----------------------------------------*- C++ --*-===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// A intra-procedural analysis for thread safety (e.g. deadlocks and race 11// conditions), based off of an annotation system. 12// 13// See http://clang.llvm.org/docs/LanguageExtensions.html#threadsafety for more 14// information. 15// 16//===----------------------------------------------------------------------===// 17 18#include "clang/Analysis/Analyses/ThreadSafety.h" 19#include "clang/Analysis/Analyses/PostOrderCFGView.h" 20#include "clang/Analysis/AnalysisContext.h" 21#include "clang/Analysis/CFG.h" 22#include "clang/Analysis/CFGStmtMap.h" 23#include "clang/AST/DeclCXX.h" 24#include "clang/AST/ExprCXX.h" 25#include "clang/AST/StmtCXX.h" 26#include "clang/AST/StmtVisitor.h" 27#include "clang/Basic/SourceManager.h" 28#include "clang/Basic/SourceLocation.h" 29#include "clang/Basic/OperatorKinds.h" 30#include "llvm/ADT/BitVector.h" 31#include "llvm/ADT/FoldingSet.h" 32#include "llvm/ADT/ImmutableMap.h" 33#include "llvm/ADT/PostOrderIterator.h" 34#include "llvm/ADT/SmallVector.h" 35#include "llvm/ADT/StringRef.h" 36#include "llvm/Support/raw_ostream.h" 37#include <algorithm> 38#include <utility> 39#include <vector> 40 41using namespace clang; 42using namespace thread_safety; 43 44// Key method definition 45ThreadSafetyHandler::~ThreadSafetyHandler() {} 46 47namespace { 48 49/// SExpr implements a simple expression language that is used to store, 50/// compare, and pretty-print C++ expressions. Unlike a clang Expr, a SExpr 51/// does not capture surface syntax, and it does not distinguish between 52/// C++ concepts, like pointers and references, that have no real semantic 53/// differences. This simplicity allows SExprs to be meaningfully compared, 54/// e.g. 55/// (x) = x 56/// (*this).foo = this->foo 57/// *&a = a 58/// 59/// Thread-safety analysis works by comparing lock expressions. Within the 60/// body of a function, an expression such as "x->foo->bar.mu" will resolve to 61/// a particular mutex object at run-time. Subsequent occurrences of the same 62/// expression (where "same" means syntactic equality) will refer to the same 63/// run-time object if three conditions hold: 64/// (1) Local variables in the expression, such as "x" have not changed. 65/// (2) Values on the heap that affect the expression have not changed. 66/// (3) The expression involves only pure function calls. 67/// 68/// The current implementation assumes, but does not verify, that multiple uses 69/// of the same lock expression satisfies these criteria. 70class SExpr { 71private: 72 enum ExprOp { 73 EOP_Nop, ///< No-op 74 EOP_Wildcard, ///< Matches anything. 75 EOP_Universal, ///< Universal lock. 76 EOP_This, ///< This keyword. 77 EOP_NVar, ///< Named variable. 78 EOP_LVar, ///< Local variable. 79 EOP_Dot, ///< Field access 80 EOP_Call, ///< Function call 81 EOP_MCall, ///< Method call 82 EOP_Index, ///< Array index 83 EOP_Unary, ///< Unary operation 84 EOP_Binary, ///< Binary operation 85 EOP_Unknown ///< Catchall for everything else 86 }; 87 88 89 class SExprNode { 90 private: 91 unsigned char Op; ///< Opcode of the root node 92 unsigned char Flags; ///< Additional opcode-specific data 93 unsigned short Sz; ///< Number of child nodes 94 const void* Data; ///< Additional opcode-specific data 95 96 public: 97 SExprNode(ExprOp O, unsigned F, const void* D) 98 : Op(static_cast<unsigned char>(O)), 99 Flags(static_cast<unsigned char>(F)), Sz(1), Data(D) 100 { } 101 102 unsigned size() const { return Sz; } 103 void setSize(unsigned S) { Sz = S; } 104 105 ExprOp kind() const { return static_cast<ExprOp>(Op); } 106 107 const NamedDecl* getNamedDecl() const { 108 assert(Op == EOP_NVar || Op == EOP_LVar || Op == EOP_Dot); 109 return reinterpret_cast<const NamedDecl*>(Data); 110 } 111 112 const NamedDecl* getFunctionDecl() const { 113 assert(Op == EOP_Call || Op == EOP_MCall); 114 return reinterpret_cast<const NamedDecl*>(Data); 115 } 116 117 bool isArrow() const { return Op == EOP_Dot && Flags == 1; } 118 void setArrow(bool A) { Flags = A ? 1 : 0; } 119 120 unsigned arity() const { 121 switch (Op) { 122 case EOP_Nop: return 0; 123 case EOP_Wildcard: return 0; 124 case EOP_Universal: return 0; 125 case EOP_NVar: return 0; 126 case EOP_LVar: return 0; 127 case EOP_This: return 0; 128 case EOP_Dot: return 1; 129 case EOP_Call: return Flags+1; // First arg is function. 130 case EOP_MCall: return Flags+1; // First arg is implicit obj. 131 case EOP_Index: return 2; 132 case EOP_Unary: return 1; 133 case EOP_Binary: return 2; 134 case EOP_Unknown: return Flags; 135 } 136 return 0; 137 } 138 139 bool operator==(const SExprNode& Other) const { 140 // Ignore flags and size -- they don't matter. 141 return (Op == Other.Op && 142 Data == Other.Data); 143 } 144 145 bool operator!=(const SExprNode& Other) const { 146 return !(*this == Other); 147 } 148 149 bool matches(const SExprNode& Other) const { 150 return (*this == Other) || 151 (Op == EOP_Wildcard) || 152 (Other.Op == EOP_Wildcard); 153 } 154 }; 155 156 157 /// \brief Encapsulates the lexical context of a function call. The lexical 158 /// context includes the arguments to the call, including the implicit object 159 /// argument. When an attribute containing a mutex expression is attached to 160 /// a method, the expression may refer to formal parameters of the method. 161 /// Actual arguments must be substituted for formal parameters to derive 162 /// the appropriate mutex expression in the lexical context where the function 163 /// is called. PrevCtx holds the context in which the arguments themselves 164 /// should be evaluated; multiple calling contexts can be chained together 165 /// by the lock_returned attribute. 166 struct CallingContext { 167 const NamedDecl* AttrDecl; // The decl to which the attribute is attached. 168 Expr* SelfArg; // Implicit object argument -- e.g. 'this' 169 bool SelfArrow; // is Self referred to with -> or .? 170 unsigned NumArgs; // Number of funArgs 171 Expr** FunArgs; // Function arguments 172 CallingContext* PrevCtx; // The previous context; or 0 if none. 173 174 CallingContext(const NamedDecl *D = 0, Expr *S = 0, 175 unsigned N = 0, Expr **A = 0, CallingContext *P = 0) 176 : AttrDecl(D), SelfArg(S), SelfArrow(false), 177 NumArgs(N), FunArgs(A), PrevCtx(P) 178 { } 179 }; 180 181 typedef SmallVector<SExprNode, 4> NodeVector; 182 183private: 184 // A SExpr is a list of SExprNodes in prefix order. The Size field allows 185 // the list to be traversed as a tree. 186 NodeVector NodeVec; 187 188private: 189 unsigned makeNop() { 190 NodeVec.push_back(SExprNode(EOP_Nop, 0, 0)); 191 return NodeVec.size()-1; 192 } 193 194 unsigned makeWildcard() { 195 NodeVec.push_back(SExprNode(EOP_Wildcard, 0, 0)); 196 return NodeVec.size()-1; 197 } 198 199 unsigned makeUniversal() { 200 NodeVec.push_back(SExprNode(EOP_Universal, 0, 0)); 201 return NodeVec.size()-1; 202 } 203 204 unsigned makeNamedVar(const NamedDecl *D) { 205 NodeVec.push_back(SExprNode(EOP_NVar, 0, D)); 206 return NodeVec.size()-1; 207 } 208 209 unsigned makeLocalVar(const NamedDecl *D) { 210 NodeVec.push_back(SExprNode(EOP_LVar, 0, D)); 211 return NodeVec.size()-1; 212 } 213 214 unsigned makeThis() { 215 NodeVec.push_back(SExprNode(EOP_This, 0, 0)); 216 return NodeVec.size()-1; 217 } 218 219 unsigned makeDot(const NamedDecl *D, bool Arrow) { 220 NodeVec.push_back(SExprNode(EOP_Dot, Arrow ? 1 : 0, D)); 221 return NodeVec.size()-1; 222 } 223 224 unsigned makeCall(unsigned NumArgs, const NamedDecl *D) { 225 NodeVec.push_back(SExprNode(EOP_Call, NumArgs, D)); 226 return NodeVec.size()-1; 227 } 228 229 unsigned makeMCall(unsigned NumArgs, const NamedDecl *D) { 230 NodeVec.push_back(SExprNode(EOP_MCall, NumArgs, D)); 231 return NodeVec.size()-1; 232 } 233 234 unsigned makeIndex() { 235 NodeVec.push_back(SExprNode(EOP_Index, 0, 0)); 236 return NodeVec.size()-1; 237 } 238 239 unsigned makeUnary() { 240 NodeVec.push_back(SExprNode(EOP_Unary, 0, 0)); 241 return NodeVec.size()-1; 242 } 243 244 unsigned makeBinary() { 245 NodeVec.push_back(SExprNode(EOP_Binary, 0, 0)); 246 return NodeVec.size()-1; 247 } 248 249 unsigned makeUnknown(unsigned Arity) { 250 NodeVec.push_back(SExprNode(EOP_Unknown, Arity, 0)); 251 return NodeVec.size()-1; 252 } 253 254 /// Build an SExpr from the given C++ expression. 255 /// Recursive function that terminates on DeclRefExpr. 256 /// Note: this function merely creates a SExpr; it does not check to 257 /// ensure that the original expression is a valid mutex expression. 258 /// 259 /// NDeref returns the number of Derefence and AddressOf operations 260 /// preceeding the Expr; this is used to decide whether to pretty-print 261 /// SExprs with . or ->. 262 unsigned buildSExpr(Expr *Exp, CallingContext* CallCtx, int* NDeref = 0) { 263 if (!Exp) 264 return 0; 265 266 if (DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(Exp)) { 267 NamedDecl *ND = cast<NamedDecl>(DRE->getDecl()->getCanonicalDecl()); 268 ParmVarDecl *PV = dyn_cast_or_null<ParmVarDecl>(ND); 269 if (PV) { 270 FunctionDecl *FD = 271 cast<FunctionDecl>(PV->getDeclContext())->getCanonicalDecl(); 272 unsigned i = PV->getFunctionScopeIndex(); 273 274 if (CallCtx && CallCtx->FunArgs && 275 FD == CallCtx->AttrDecl->getCanonicalDecl()) { 276 // Substitute call arguments for references to function parameters 277 assert(i < CallCtx->NumArgs); 278 return buildSExpr(CallCtx->FunArgs[i], CallCtx->PrevCtx, NDeref); 279 } 280 // Map the param back to the param of the original function declaration. 281 makeNamedVar(FD->getParamDecl(i)); 282 return 1; 283 } 284 // Not a function parameter -- just store the reference. 285 makeNamedVar(ND); 286 return 1; 287 } else if (isa<CXXThisExpr>(Exp)) { 288 // Substitute parent for 'this' 289 if (CallCtx && CallCtx->SelfArg) { 290 if (!CallCtx->SelfArrow && NDeref) 291 // 'this' is a pointer, but self is not, so need to take address. 292 --(*NDeref); 293 return buildSExpr(CallCtx->SelfArg, CallCtx->PrevCtx, NDeref); 294 } 295 else { 296 makeThis(); 297 return 1; 298 } 299 } else if (MemberExpr *ME = dyn_cast<MemberExpr>(Exp)) { 300 NamedDecl *ND = ME->getMemberDecl(); 301 int ImplicitDeref = ME->isArrow() ? 1 : 0; 302 unsigned Root = makeDot(ND, false); 303 unsigned Sz = buildSExpr(ME->getBase(), CallCtx, &ImplicitDeref); 304 NodeVec[Root].setArrow(ImplicitDeref > 0); 305 NodeVec[Root].setSize(Sz + 1); 306 return Sz + 1; 307 } else if (CXXMemberCallExpr *CMCE = dyn_cast<CXXMemberCallExpr>(Exp)) { 308 // When calling a function with a lock_returned attribute, replace 309 // the function call with the expression in lock_returned. 310 CXXMethodDecl* MD = 311 cast<CXXMethodDecl>(CMCE->getMethodDecl()->getMostRecentDecl()); 312 if (LockReturnedAttr* At = MD->getAttr<LockReturnedAttr>()) { 313 CallingContext LRCallCtx(CMCE->getMethodDecl()); 314 LRCallCtx.SelfArg = CMCE->getImplicitObjectArgument(); 315 LRCallCtx.SelfArrow = 316 dyn_cast<MemberExpr>(CMCE->getCallee())->isArrow(); 317 LRCallCtx.NumArgs = CMCE->getNumArgs(); 318 LRCallCtx.FunArgs = CMCE->getArgs(); 319 LRCallCtx.PrevCtx = CallCtx; 320 return buildSExpr(At->getArg(), &LRCallCtx); 321 } 322 // Hack to treat smart pointers and iterators as pointers; 323 // ignore any method named get(). 324 if (CMCE->getMethodDecl()->getNameAsString() == "get" && 325 CMCE->getNumArgs() == 0) { 326 if (NDeref && dyn_cast<MemberExpr>(CMCE->getCallee())->isArrow()) 327 ++(*NDeref); 328 return buildSExpr(CMCE->getImplicitObjectArgument(), CallCtx, NDeref); 329 } 330 unsigned NumCallArgs = CMCE->getNumArgs(); 331 unsigned Root = 332 makeMCall(NumCallArgs, CMCE->getMethodDecl()->getCanonicalDecl()); 333 unsigned Sz = buildSExpr(CMCE->getImplicitObjectArgument(), CallCtx); 334 Expr** CallArgs = CMCE->getArgs(); 335 for (unsigned i = 0; i < NumCallArgs; ++i) { 336 Sz += buildSExpr(CallArgs[i], CallCtx); 337 } 338 NodeVec[Root].setSize(Sz + 1); 339 return Sz + 1; 340 } else if (CallExpr *CE = dyn_cast<CallExpr>(Exp)) { 341 FunctionDecl* FD = 342 cast<FunctionDecl>(CE->getDirectCallee()->getMostRecentDecl()); 343 if (LockReturnedAttr* At = FD->getAttr<LockReturnedAttr>()) { 344 CallingContext LRCallCtx(CE->getDirectCallee()); 345 LRCallCtx.NumArgs = CE->getNumArgs(); 346 LRCallCtx.FunArgs = CE->getArgs(); 347 LRCallCtx.PrevCtx = CallCtx; 348 return buildSExpr(At->getArg(), &LRCallCtx); 349 } 350 // Treat smart pointers and iterators as pointers; 351 // ignore the * and -> operators. 352 if (CXXOperatorCallExpr *OE = dyn_cast<CXXOperatorCallExpr>(CE)) { 353 OverloadedOperatorKind k = OE->getOperator(); 354 if (k == OO_Star) { 355 if (NDeref) ++(*NDeref); 356 return buildSExpr(OE->getArg(0), CallCtx, NDeref); 357 } 358 else if (k == OO_Arrow) { 359 return buildSExpr(OE->getArg(0), CallCtx, NDeref); 360 } 361 } 362 unsigned NumCallArgs = CE->getNumArgs(); 363 unsigned Root = makeCall(NumCallArgs, 0); 364 unsigned Sz = buildSExpr(CE->getCallee(), CallCtx); 365 Expr** CallArgs = CE->getArgs(); 366 for (unsigned i = 0; i < NumCallArgs; ++i) { 367 Sz += buildSExpr(CallArgs[i], CallCtx); 368 } 369 NodeVec[Root].setSize(Sz+1); 370 return Sz+1; 371 } else if (BinaryOperator *BOE = dyn_cast<BinaryOperator>(Exp)) { 372 unsigned Root = makeBinary(); 373 unsigned Sz = buildSExpr(BOE->getLHS(), CallCtx); 374 Sz += buildSExpr(BOE->getRHS(), CallCtx); 375 NodeVec[Root].setSize(Sz); 376 return Sz; 377 } else if (UnaryOperator *UOE = dyn_cast<UnaryOperator>(Exp)) { 378 // Ignore & and * operators -- they're no-ops. 379 // However, we try to figure out whether the expression is a pointer, 380 // so we can use . and -> appropriately in error messages. 381 if (UOE->getOpcode() == UO_Deref) { 382 if (NDeref) ++(*NDeref); 383 return buildSExpr(UOE->getSubExpr(), CallCtx, NDeref); 384 } 385 if (UOE->getOpcode() == UO_AddrOf) { 386 if (DeclRefExpr* DRE = dyn_cast<DeclRefExpr>(UOE->getSubExpr())) { 387 if (DRE->getDecl()->isCXXInstanceMember()) { 388 // This is a pointer-to-member expression, e.g. &MyClass::mu_. 389 // We interpret this syntax specially, as a wildcard. 390 unsigned Root = makeDot(DRE->getDecl(), false); 391 makeWildcard(); 392 NodeVec[Root].setSize(2); 393 return 2; 394 } 395 } 396 if (NDeref) --(*NDeref); 397 return buildSExpr(UOE->getSubExpr(), CallCtx, NDeref); 398 } 399 unsigned Root = makeUnary(); 400 unsigned Sz = buildSExpr(UOE->getSubExpr(), CallCtx); 401 NodeVec[Root].setSize(Sz); 402 return Sz; 403 } else if (ArraySubscriptExpr *ASE = dyn_cast<ArraySubscriptExpr>(Exp)) { 404 unsigned Root = makeIndex(); 405 unsigned Sz = buildSExpr(ASE->getBase(), CallCtx); 406 Sz += buildSExpr(ASE->getIdx(), CallCtx); 407 NodeVec[Root].setSize(Sz); 408 return Sz; 409 } else if (AbstractConditionalOperator *CE = 410 dyn_cast<AbstractConditionalOperator>(Exp)) { 411 unsigned Root = makeUnknown(3); 412 unsigned Sz = buildSExpr(CE->getCond(), CallCtx); 413 Sz += buildSExpr(CE->getTrueExpr(), CallCtx); 414 Sz += buildSExpr(CE->getFalseExpr(), CallCtx); 415 NodeVec[Root].setSize(Sz); 416 return Sz; 417 } else if (ChooseExpr *CE = dyn_cast<ChooseExpr>(Exp)) { 418 unsigned Root = makeUnknown(3); 419 unsigned Sz = buildSExpr(CE->getCond(), CallCtx); 420 Sz += buildSExpr(CE->getLHS(), CallCtx); 421 Sz += buildSExpr(CE->getRHS(), CallCtx); 422 NodeVec[Root].setSize(Sz); 423 return Sz; 424 } else if (CastExpr *CE = dyn_cast<CastExpr>(Exp)) { 425 return buildSExpr(CE->getSubExpr(), CallCtx, NDeref); 426 } else if (ParenExpr *PE = dyn_cast<ParenExpr>(Exp)) { 427 return buildSExpr(PE->getSubExpr(), CallCtx, NDeref); 428 } else if (ExprWithCleanups *EWC = dyn_cast<ExprWithCleanups>(Exp)) { 429 return buildSExpr(EWC->getSubExpr(), CallCtx, NDeref); 430 } else if (CXXBindTemporaryExpr *E = dyn_cast<CXXBindTemporaryExpr>(Exp)) { 431 return buildSExpr(E->getSubExpr(), CallCtx, NDeref); 432 } else if (isa<CharacterLiteral>(Exp) || 433 isa<CXXNullPtrLiteralExpr>(Exp) || 434 isa<GNUNullExpr>(Exp) || 435 isa<CXXBoolLiteralExpr>(Exp) || 436 isa<FloatingLiteral>(Exp) || 437 isa<ImaginaryLiteral>(Exp) || 438 isa<IntegerLiteral>(Exp) || 439 isa<StringLiteral>(Exp) || 440 isa<ObjCStringLiteral>(Exp)) { 441 makeNop(); 442 return 1; // FIXME: Ignore literals for now 443 } else { 444 makeNop(); 445 return 1; // Ignore. FIXME: mark as invalid expression? 446 } 447 } 448 449 /// \brief Construct a SExpr from an expression. 450 /// \param MutexExp The original mutex expression within an attribute 451 /// \param DeclExp An expression involving the Decl on which the attribute 452 /// occurs. 453 /// \param D The declaration to which the lock/unlock attribute is attached. 454 void buildSExprFromExpr(Expr *MutexExp, Expr *DeclExp, const NamedDecl *D) { 455 CallingContext CallCtx(D); 456 457 458 if (MutexExp) { 459 if (StringLiteral* SLit = dyn_cast<StringLiteral>(MutexExp)) { 460 if (SLit->getString() == StringRef("*")) 461 // The "*" expr is a universal lock, which essentially turns off 462 // checks until it is removed from the lockset. 463 makeUniversal(); 464 else 465 // Ignore other string literals for now. 466 makeNop(); 467 return; 468 } 469 } 470 471 // If we are processing a raw attribute expression, with no substitutions. 472 if (DeclExp == 0) { 473 buildSExpr(MutexExp, 0); 474 return; 475 } 476 477 // Examine DeclExp to find SelfArg and FunArgs, which are used to substitute 478 // for formal parameters when we call buildMutexID later. 479 if (MemberExpr *ME = dyn_cast<MemberExpr>(DeclExp)) { 480 CallCtx.SelfArg = ME->getBase(); 481 CallCtx.SelfArrow = ME->isArrow(); 482 } else if (CXXMemberCallExpr *CE = dyn_cast<CXXMemberCallExpr>(DeclExp)) { 483 CallCtx.SelfArg = CE->getImplicitObjectArgument(); 484 CallCtx.SelfArrow = dyn_cast<MemberExpr>(CE->getCallee())->isArrow(); 485 CallCtx.NumArgs = CE->getNumArgs(); 486 CallCtx.FunArgs = CE->getArgs(); 487 } else if (CallExpr *CE = dyn_cast<CallExpr>(DeclExp)) { 488 CallCtx.NumArgs = CE->getNumArgs(); 489 CallCtx.FunArgs = CE->getArgs(); 490 } else if (CXXConstructExpr *CE = dyn_cast<CXXConstructExpr>(DeclExp)) { 491 CallCtx.SelfArg = 0; // FIXME -- get the parent from DeclStmt 492 CallCtx.NumArgs = CE->getNumArgs(); 493 CallCtx.FunArgs = CE->getArgs(); 494 } else if (D && isa<CXXDestructorDecl>(D)) { 495 // There's no such thing as a "destructor call" in the AST. 496 CallCtx.SelfArg = DeclExp; 497 } 498 499 // If the attribute has no arguments, then assume the argument is "this". 500 if (MutexExp == 0) { 501 buildSExpr(CallCtx.SelfArg, 0); 502 return; 503 } 504 505 // For most attributes. 506 buildSExpr(MutexExp, &CallCtx); 507 } 508 509 /// \brief Get index of next sibling of node i. 510 unsigned getNextSibling(unsigned i) const { 511 return i + NodeVec[i].size(); 512 } 513 514public: 515 explicit SExpr(clang::Decl::EmptyShell e) { NodeVec.clear(); } 516 517 /// \param MutexExp The original mutex expression within an attribute 518 /// \param DeclExp An expression involving the Decl on which the attribute 519 /// occurs. 520 /// \param D The declaration to which the lock/unlock attribute is attached. 521 /// Caller must check isValid() after construction. 522 SExpr(Expr* MutexExp, Expr *DeclExp, const NamedDecl* D) { 523 buildSExprFromExpr(MutexExp, DeclExp, D); 524 } 525 526 /// Return true if this is a valid decl sequence. 527 /// Caller must call this by hand after construction to handle errors. 528 bool isValid() const { 529 return !NodeVec.empty(); 530 } 531 532 bool shouldIgnore() const { 533 // Nop is a mutex that we have decided to deliberately ignore. 534 assert(NodeVec.size() > 0 && "Invalid Mutex"); 535 return NodeVec[0].kind() == EOP_Nop; 536 } 537 538 bool isUniversal() const { 539 assert(NodeVec.size() > 0 && "Invalid Mutex"); 540 return NodeVec[0].kind() == EOP_Universal; 541 } 542 543 /// Issue a warning about an invalid lock expression 544 static void warnInvalidLock(ThreadSafetyHandler &Handler, Expr* MutexExp, 545 Expr *DeclExp, const NamedDecl* D) { 546 SourceLocation Loc; 547 if (DeclExp) 548 Loc = DeclExp->getExprLoc(); 549 550 // FIXME: add a note about the attribute location in MutexExp or D 551 if (Loc.isValid()) 552 Handler.handleInvalidLockExp(Loc); 553 } 554 555 bool operator==(const SExpr &other) const { 556 return NodeVec == other.NodeVec; 557 } 558 559 bool operator!=(const SExpr &other) const { 560 return !(*this == other); 561 } 562 563 bool matches(const SExpr &Other, unsigned i = 0, unsigned j = 0) const { 564 if (NodeVec[i].matches(Other.NodeVec[j])) { 565 unsigned n = NodeVec[i].arity(); 566 bool Result = true; 567 unsigned ci = i+1; // first child of i 568 unsigned cj = j+1; // first child of j 569 for (unsigned k = 0; k < n; 570 ++k, ci=getNextSibling(ci), cj = Other.getNextSibling(cj)) { 571 Result = Result && matches(Other, ci, cj); 572 } 573 return Result; 574 } 575 return false; 576 } 577 578 /// \brief Pretty print a lock expression for use in error messages. 579 std::string toString(unsigned i = 0) const { 580 assert(isValid()); 581 if (i >= NodeVec.size()) 582 return ""; 583 584 const SExprNode* N = &NodeVec[i]; 585 switch (N->kind()) { 586 case EOP_Nop: 587 return "_"; 588 case EOP_Wildcard: 589 return "(?)"; 590 case EOP_Universal: 591 return "*"; 592 case EOP_This: 593 return "this"; 594 case EOP_NVar: 595 case EOP_LVar: { 596 return N->getNamedDecl()->getNameAsString(); 597 } 598 case EOP_Dot: { 599 if (NodeVec[i+1].kind() == EOP_Wildcard) { 600 std::string S = "&"; 601 S += N->getNamedDecl()->getQualifiedNameAsString(); 602 return S; 603 } 604 std::string FieldName = N->getNamedDecl()->getNameAsString(); 605 if (NodeVec[i+1].kind() == EOP_This) 606 return FieldName; 607 608 std::string S = toString(i+1); 609 if (N->isArrow()) 610 return S + "->" + FieldName; 611 else 612 return S + "." + FieldName; 613 } 614 case EOP_Call: { 615 std::string S = toString(i+1) + "("; 616 unsigned NumArgs = N->arity()-1; 617 unsigned ci = getNextSibling(i+1); 618 for (unsigned k=0; k<NumArgs; ++k, ci = getNextSibling(ci)) { 619 S += toString(ci); 620 if (k+1 < NumArgs) S += ","; 621 } 622 S += ")"; 623 return S; 624 } 625 case EOP_MCall: { 626 std::string S = ""; 627 if (NodeVec[i+1].kind() != EOP_This) 628 S = toString(i+1) + "."; 629 if (const NamedDecl *D = N->getFunctionDecl()) 630 S += D->getNameAsString() + "("; 631 else 632 S += "#("; 633 unsigned NumArgs = N->arity()-1; 634 unsigned ci = getNextSibling(i+1); 635 for (unsigned k=0; k<NumArgs; ++k, ci = getNextSibling(ci)) { 636 S += toString(ci); 637 if (k+1 < NumArgs) S += ","; 638 } 639 S += ")"; 640 return S; 641 } 642 case EOP_Index: { 643 std::string S1 = toString(i+1); 644 std::string S2 = toString(i+1 + NodeVec[i+1].size()); 645 return S1 + "[" + S2 + "]"; 646 } 647 case EOP_Unary: { 648 std::string S = toString(i+1); 649 return "#" + S; 650 } 651 case EOP_Binary: { 652 std::string S1 = toString(i+1); 653 std::string S2 = toString(i+1 + NodeVec[i+1].size()); 654 return "(" + S1 + "#" + S2 + ")"; 655 } 656 case EOP_Unknown: { 657 unsigned NumChildren = N->arity(); 658 if (NumChildren == 0) 659 return "(...)"; 660 std::string S = "("; 661 unsigned ci = i+1; 662 for (unsigned j = 0; j < NumChildren; ++j, ci = getNextSibling(ci)) { 663 S += toString(ci); 664 if (j+1 < NumChildren) S += "#"; 665 } 666 S += ")"; 667 return S; 668 } 669 } 670 return ""; 671 } 672}; 673 674 675 676/// \brief A short list of SExprs 677class MutexIDList : public SmallVector<SExpr, 3> { 678public: 679 /// \brief Return true if the list contains the specified SExpr 680 /// Performs a linear search, because these lists are almost always very small. 681 bool contains(const SExpr& M) { 682 for (iterator I=begin(),E=end(); I != E; ++I) 683 if ((*I) == M) return true; 684 return false; 685 } 686 687 /// \brief Push M onto list, bud discard duplicates 688 void push_back_nodup(const SExpr& M) { 689 if (!contains(M)) push_back(M); 690 } 691}; 692 693 694 695/// \brief This is a helper class that stores info about the most recent 696/// accquire of a Lock. 697/// 698/// The main body of the analysis maps MutexIDs to LockDatas. 699struct LockData { 700 SourceLocation AcquireLoc; 701 702 /// \brief LKind stores whether a lock is held shared or exclusively. 703 /// Note that this analysis does not currently support either re-entrant 704 /// locking or lock "upgrading" and "downgrading" between exclusive and 705 /// shared. 706 /// 707 /// FIXME: add support for re-entrant locking and lock up/downgrading 708 LockKind LKind; 709 bool Managed; // for ScopedLockable objects 710 SExpr UnderlyingMutex; // for ScopedLockable objects 711 712 LockData(SourceLocation AcquireLoc, LockKind LKind, bool M = false) 713 : AcquireLoc(AcquireLoc), LKind(LKind), Managed(M), 714 UnderlyingMutex(Decl::EmptyShell()) 715 {} 716 717 LockData(SourceLocation AcquireLoc, LockKind LKind, const SExpr &Mu) 718 : AcquireLoc(AcquireLoc), LKind(LKind), Managed(false), 719 UnderlyingMutex(Mu) 720 {} 721 722 bool operator==(const LockData &other) const { 723 return AcquireLoc == other.AcquireLoc && LKind == other.LKind; 724 } 725 726 bool operator!=(const LockData &other) const { 727 return !(*this == other); 728 } 729 730 void Profile(llvm::FoldingSetNodeID &ID) const { 731 ID.AddInteger(AcquireLoc.getRawEncoding()); 732 ID.AddInteger(LKind); 733 } 734 735 bool isAtLeast(LockKind LK) { 736 return (LK == LK_Shared) || (LKind == LK_Exclusive); 737 } 738}; 739 740 741/// \brief A FactEntry stores a single fact that is known at a particular point 742/// in the program execution. Currently, this is information regarding a lock 743/// that is held at that point. 744struct FactEntry { 745 SExpr MutID; 746 LockData LDat; 747 748 FactEntry(const SExpr& M, const LockData& L) 749 : MutID(M), LDat(L) 750 { } 751}; 752 753 754typedef unsigned short FactID; 755 756/// \brief FactManager manages the memory for all facts that are created during 757/// the analysis of a single routine. 758class FactManager { 759private: 760 std::vector<FactEntry> Facts; 761 762public: 763 FactID newLock(const SExpr& M, const LockData& L) { 764 Facts.push_back(FactEntry(M,L)); 765 return static_cast<unsigned short>(Facts.size() - 1); 766 } 767 768 const FactEntry& operator[](FactID F) const { return Facts[F]; } 769 FactEntry& operator[](FactID F) { return Facts[F]; } 770}; 771 772 773/// \brief A FactSet is the set of facts that are known to be true at a 774/// particular program point. FactSets must be small, because they are 775/// frequently copied, and are thus implemented as a set of indices into a 776/// table maintained by a FactManager. A typical FactSet only holds 1 or 2 777/// locks, so we can get away with doing a linear search for lookup. Note 778/// that a hashtable or map is inappropriate in this case, because lookups 779/// may involve partial pattern matches, rather than exact matches. 780class FactSet { 781private: 782 typedef SmallVector<FactID, 4> FactVec; 783 784 FactVec FactIDs; 785 786public: 787 typedef FactVec::iterator iterator; 788 typedef FactVec::const_iterator const_iterator; 789 790 iterator begin() { return FactIDs.begin(); } 791 const_iterator begin() const { return FactIDs.begin(); } 792 793 iterator end() { return FactIDs.end(); } 794 const_iterator end() const { return FactIDs.end(); } 795 796 bool isEmpty() const { return FactIDs.size() == 0; } 797 798 FactID addLock(FactManager& FM, const SExpr& M, const LockData& L) { 799 FactID F = FM.newLock(M, L); 800 FactIDs.push_back(F); 801 return F; 802 } 803 804 bool removeLock(FactManager& FM, const SExpr& M) { 805 unsigned n = FactIDs.size(); 806 if (n == 0) 807 return false; 808 809 for (unsigned i = 0; i < n-1; ++i) { 810 if (FM[FactIDs[i]].MutID.matches(M)) { 811 FactIDs[i] = FactIDs[n-1]; 812 FactIDs.pop_back(); 813 return true; 814 } 815 } 816 if (FM[FactIDs[n-1]].MutID.matches(M)) { 817 FactIDs.pop_back(); 818 return true; 819 } 820 return false; 821 } 822 823 LockData* findLock(FactManager& FM, const SExpr& M) const { 824 for (const_iterator I = begin(), E = end(); I != E; ++I) { 825 const SExpr &Exp = FM[*I].MutID; 826 if (Exp.matches(M)) 827 return &FM[*I].LDat; 828 } 829 return 0; 830 } 831 832 LockData* findLockUniv(FactManager& FM, const SExpr& M) const { 833 for (const_iterator I = begin(), E = end(); I != E; ++I) { 834 const SExpr &Exp = FM[*I].MutID; 835 if (Exp.matches(M) || Exp.isUniversal()) 836 return &FM[*I].LDat; 837 } 838 return 0; 839 } 840}; 841 842 843 844/// A Lockset maps each SExpr (defined above) to information about how it has 845/// been locked. 846typedef llvm::ImmutableMap<SExpr, LockData> Lockset; 847typedef llvm::ImmutableMap<const NamedDecl*, unsigned> LocalVarContext; 848 849class LocalVariableMap; 850 851/// A side (entry or exit) of a CFG node. 852enum CFGBlockSide { CBS_Entry, CBS_Exit }; 853 854/// CFGBlockInfo is a struct which contains all the information that is 855/// maintained for each block in the CFG. See LocalVariableMap for more 856/// information about the contexts. 857struct CFGBlockInfo { 858 FactSet EntrySet; // Lockset held at entry to block 859 FactSet ExitSet; // Lockset held at exit from block 860 LocalVarContext EntryContext; // Context held at entry to block 861 LocalVarContext ExitContext; // Context held at exit from block 862 SourceLocation EntryLoc; // Location of first statement in block 863 SourceLocation ExitLoc; // Location of last statement in block. 864 unsigned EntryIndex; // Used to replay contexts later 865 866 const FactSet &getSet(CFGBlockSide Side) const { 867 return Side == CBS_Entry ? EntrySet : ExitSet; 868 } 869 SourceLocation getLocation(CFGBlockSide Side) const { 870 return Side == CBS_Entry ? EntryLoc : ExitLoc; 871 } 872 873private: 874 CFGBlockInfo(LocalVarContext EmptyCtx) 875 : EntryContext(EmptyCtx), ExitContext(EmptyCtx) 876 { } 877 878public: 879 static CFGBlockInfo getEmptyBlockInfo(LocalVariableMap &M); 880}; 881 882 883 884// A LocalVariableMap maintains a map from local variables to their currently 885// valid definitions. It provides SSA-like functionality when traversing the 886// CFG. Like SSA, each definition or assignment to a variable is assigned a 887// unique name (an integer), which acts as the SSA name for that definition. 888// The total set of names is shared among all CFG basic blocks. 889// Unlike SSA, we do not rewrite expressions to replace local variables declrefs 890// with their SSA-names. Instead, we compute a Context for each point in the 891// code, which maps local variables to the appropriate SSA-name. This map 892// changes with each assignment. 893// 894// The map is computed in a single pass over the CFG. Subsequent analyses can 895// then query the map to find the appropriate Context for a statement, and use 896// that Context to look up the definitions of variables. 897class LocalVariableMap { 898public: 899 typedef LocalVarContext Context; 900 901 /// A VarDefinition consists of an expression, representing the value of the 902 /// variable, along with the context in which that expression should be 903 /// interpreted. A reference VarDefinition does not itself contain this 904 /// information, but instead contains a pointer to a previous VarDefinition. 905 struct VarDefinition { 906 public: 907 friend class LocalVariableMap; 908 909 const NamedDecl *Dec; // The original declaration for this variable. 910 const Expr *Exp; // The expression for this variable, OR 911 unsigned Ref; // Reference to another VarDefinition 912 Context Ctx; // The map with which Exp should be interpreted. 913 914 bool isReference() { return !Exp; } 915 916 private: 917 // Create ordinary variable definition 918 VarDefinition(const NamedDecl *D, const Expr *E, Context C) 919 : Dec(D), Exp(E), Ref(0), Ctx(C) 920 { } 921 922 // Create reference to previous definition 923 VarDefinition(const NamedDecl *D, unsigned R, Context C) 924 : Dec(D), Exp(0), Ref(R), Ctx(C) 925 { } 926 }; 927 928private: 929 Context::Factory ContextFactory; 930 std::vector<VarDefinition> VarDefinitions; 931 std::vector<unsigned> CtxIndices; 932 std::vector<std::pair<Stmt*, Context> > SavedContexts; 933 934public: 935 LocalVariableMap() { 936 // index 0 is a placeholder for undefined variables (aka phi-nodes). 937 VarDefinitions.push_back(VarDefinition(0, 0u, getEmptyContext())); 938 } 939 940 /// Look up a definition, within the given context. 941 const VarDefinition* lookup(const NamedDecl *D, Context Ctx) { 942 const unsigned *i = Ctx.lookup(D); 943 if (!i) 944 return 0; 945 assert(*i < VarDefinitions.size()); 946 return &VarDefinitions[*i]; 947 } 948 949 /// Look up the definition for D within the given context. Returns 950 /// NULL if the expression is not statically known. If successful, also 951 /// modifies Ctx to hold the context of the return Expr. 952 const Expr* lookupExpr(const NamedDecl *D, Context &Ctx) { 953 const unsigned *P = Ctx.lookup(D); 954 if (!P) 955 return 0; 956 957 unsigned i = *P; 958 while (i > 0) { 959 if (VarDefinitions[i].Exp) { 960 Ctx = VarDefinitions[i].Ctx; 961 return VarDefinitions[i].Exp; 962 } 963 i = VarDefinitions[i].Ref; 964 } 965 return 0; 966 } 967 968 Context getEmptyContext() { return ContextFactory.getEmptyMap(); } 969 970 /// Return the next context after processing S. This function is used by 971 /// clients of the class to get the appropriate context when traversing the 972 /// CFG. It must be called for every assignment or DeclStmt. 973 Context getNextContext(unsigned &CtxIndex, Stmt *S, Context C) { 974 if (SavedContexts[CtxIndex+1].first == S) { 975 CtxIndex++; 976 Context Result = SavedContexts[CtxIndex].second; 977 return Result; 978 } 979 return C; 980 } 981 982 void dumpVarDefinitionName(unsigned i) { 983 if (i == 0) { 984 llvm::errs() << "Undefined"; 985 return; 986 } 987 const NamedDecl *Dec = VarDefinitions[i].Dec; 988 if (!Dec) { 989 llvm::errs() << "<<NULL>>"; 990 return; 991 } 992 Dec->printName(llvm::errs()); 993 llvm::errs() << "." << i << " " << ((const void*) Dec); 994 } 995 996 /// Dumps an ASCII representation of the variable map to llvm::errs() 997 void dump() { 998 for (unsigned i = 1, e = VarDefinitions.size(); i < e; ++i) { 999 const Expr *Exp = VarDefinitions[i].Exp; 1000 unsigned Ref = VarDefinitions[i].Ref; 1001 1002 dumpVarDefinitionName(i); 1003 llvm::errs() << " = "; 1004 if (Exp) Exp->dump(); 1005 else { 1006 dumpVarDefinitionName(Ref); 1007 llvm::errs() << "\n"; 1008 } 1009 } 1010 } 1011 1012 /// Dumps an ASCII representation of a Context to llvm::errs() 1013 void dumpContext(Context C) { 1014 for (Context::iterator I = C.begin(), E = C.end(); I != E; ++I) { 1015 const NamedDecl *D = I.getKey(); 1016 D->printName(llvm::errs()); 1017 const unsigned *i = C.lookup(D); 1018 llvm::errs() << " -> "; 1019 dumpVarDefinitionName(*i); 1020 llvm::errs() << "\n"; 1021 } 1022 } 1023 1024 /// Builds the variable map. 1025 void traverseCFG(CFG *CFGraph, PostOrderCFGView *SortedGraph, 1026 std::vector<CFGBlockInfo> &BlockInfo); 1027 1028protected: 1029 // Get the current context index 1030 unsigned getContextIndex() { return SavedContexts.size()-1; } 1031 1032 // Save the current context for later replay 1033 void saveContext(Stmt *S, Context C) { 1034 SavedContexts.push_back(std::make_pair(S,C)); 1035 } 1036 1037 // Adds a new definition to the given context, and returns a new context. 1038 // This method should be called when declaring a new variable. 1039 Context addDefinition(const NamedDecl *D, Expr *Exp, Context Ctx) { 1040 assert(!Ctx.contains(D)); 1041 unsigned newID = VarDefinitions.size(); 1042 Context NewCtx = ContextFactory.add(Ctx, D, newID); 1043 VarDefinitions.push_back(VarDefinition(D, Exp, Ctx)); 1044 return NewCtx; 1045 } 1046 1047 // Add a new reference to an existing definition. 1048 Context addReference(const NamedDecl *D, unsigned i, Context Ctx) { 1049 unsigned newID = VarDefinitions.size(); 1050 Context NewCtx = ContextFactory.add(Ctx, D, newID); 1051 VarDefinitions.push_back(VarDefinition(D, i, Ctx)); 1052 return NewCtx; 1053 } 1054 1055 // Updates a definition only if that definition is already in the map. 1056 // This method should be called when assigning to an existing variable. 1057 Context updateDefinition(const NamedDecl *D, Expr *Exp, Context Ctx) { 1058 if (Ctx.contains(D)) { 1059 unsigned newID = VarDefinitions.size(); 1060 Context NewCtx = ContextFactory.remove(Ctx, D); 1061 NewCtx = ContextFactory.add(NewCtx, D, newID); 1062 VarDefinitions.push_back(VarDefinition(D, Exp, Ctx)); 1063 return NewCtx; 1064 } 1065 return Ctx; 1066 } 1067 1068 // Removes a definition from the context, but keeps the variable name 1069 // as a valid variable. The index 0 is a placeholder for cleared definitions. 1070 Context clearDefinition(const NamedDecl *D, Context Ctx) { 1071 Context NewCtx = Ctx; 1072 if (NewCtx.contains(D)) { 1073 NewCtx = ContextFactory.remove(NewCtx, D); 1074 NewCtx = ContextFactory.add(NewCtx, D, 0); 1075 } 1076 return NewCtx; 1077 } 1078 1079 // Remove a definition entirely frmo the context. 1080 Context removeDefinition(const NamedDecl *D, Context Ctx) { 1081 Context NewCtx = Ctx; 1082 if (NewCtx.contains(D)) { 1083 NewCtx = ContextFactory.remove(NewCtx, D); 1084 } 1085 return NewCtx; 1086 } 1087 1088 Context intersectContexts(Context C1, Context C2); 1089 Context createReferenceContext(Context C); 1090 void intersectBackEdge(Context C1, Context C2); 1091 1092 friend class VarMapBuilder; 1093}; 1094 1095 1096// This has to be defined after LocalVariableMap. 1097CFGBlockInfo CFGBlockInfo::getEmptyBlockInfo(LocalVariableMap &M) { 1098 return CFGBlockInfo(M.getEmptyContext()); 1099} 1100 1101 1102/// Visitor which builds a LocalVariableMap 1103class VarMapBuilder : public StmtVisitor<VarMapBuilder> { 1104public: 1105 LocalVariableMap* VMap; 1106 LocalVariableMap::Context Ctx; 1107 1108 VarMapBuilder(LocalVariableMap *VM, LocalVariableMap::Context C) 1109 : VMap(VM), Ctx(C) {} 1110 1111 void VisitDeclStmt(DeclStmt *S); 1112 void VisitBinaryOperator(BinaryOperator *BO); 1113}; 1114 1115 1116// Add new local variables to the variable map 1117void VarMapBuilder::VisitDeclStmt(DeclStmt *S) { 1118 bool modifiedCtx = false; 1119 DeclGroupRef DGrp = S->getDeclGroup(); 1120 for (DeclGroupRef::iterator I = DGrp.begin(), E = DGrp.end(); I != E; ++I) { 1121 if (VarDecl *VD = dyn_cast_or_null<VarDecl>(*I)) { 1122 Expr *E = VD->getInit(); 1123 1124 // Add local variables with trivial type to the variable map 1125 QualType T = VD->getType(); 1126 if (T.isTrivialType(VD->getASTContext())) { 1127 Ctx = VMap->addDefinition(VD, E, Ctx); 1128 modifiedCtx = true; 1129 } 1130 } 1131 } 1132 if (modifiedCtx) 1133 VMap->saveContext(S, Ctx); 1134} 1135 1136// Update local variable definitions in variable map 1137void VarMapBuilder::VisitBinaryOperator(BinaryOperator *BO) { 1138 if (!BO->isAssignmentOp()) 1139 return; 1140 1141 Expr *LHSExp = BO->getLHS()->IgnoreParenCasts(); 1142 1143 // Update the variable map and current context. 1144 if (DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(LHSExp)) { 1145 ValueDecl *VDec = DRE->getDecl(); 1146 if (Ctx.lookup(VDec)) { 1147 if (BO->getOpcode() == BO_Assign) 1148 Ctx = VMap->updateDefinition(VDec, BO->getRHS(), Ctx); 1149 else 1150 // FIXME -- handle compound assignment operators 1151 Ctx = VMap->clearDefinition(VDec, Ctx); 1152 VMap->saveContext(BO, Ctx); 1153 } 1154 } 1155} 1156 1157 1158// Computes the intersection of two contexts. The intersection is the 1159// set of variables which have the same definition in both contexts; 1160// variables with different definitions are discarded. 1161LocalVariableMap::Context 1162LocalVariableMap::intersectContexts(Context C1, Context C2) { 1163 Context Result = C1; 1164 for (Context::iterator I = C1.begin(), E = C1.end(); I != E; ++I) { 1165 const NamedDecl *Dec = I.getKey(); 1166 unsigned i1 = I.getData(); 1167 const unsigned *i2 = C2.lookup(Dec); 1168 if (!i2) // variable doesn't exist on second path 1169 Result = removeDefinition(Dec, Result); 1170 else if (*i2 != i1) // variable exists, but has different definition 1171 Result = clearDefinition(Dec, Result); 1172 } 1173 return Result; 1174} 1175 1176// For every variable in C, create a new variable that refers to the 1177// definition in C. Return a new context that contains these new variables. 1178// (We use this for a naive implementation of SSA on loop back-edges.) 1179LocalVariableMap::Context LocalVariableMap::createReferenceContext(Context C) { 1180 Context Result = getEmptyContext(); 1181 for (Context::iterator I = C.begin(), E = C.end(); I != E; ++I) { 1182 const NamedDecl *Dec = I.getKey(); 1183 unsigned i = I.getData(); 1184 Result = addReference(Dec, i, Result); 1185 } 1186 return Result; 1187} 1188 1189// This routine also takes the intersection of C1 and C2, but it does so by 1190// altering the VarDefinitions. C1 must be the result of an earlier call to 1191// createReferenceContext. 1192void LocalVariableMap::intersectBackEdge(Context C1, Context C2) { 1193 for (Context::iterator I = C1.begin(), E = C1.end(); I != E; ++I) { 1194 const NamedDecl *Dec = I.getKey(); 1195 unsigned i1 = I.getData(); 1196 VarDefinition *VDef = &VarDefinitions[i1]; 1197 assert(VDef->isReference()); 1198 1199 const unsigned *i2 = C2.lookup(Dec); 1200 if (!i2 || (*i2 != i1)) 1201 VDef->Ref = 0; // Mark this variable as undefined 1202 } 1203} 1204 1205 1206// Traverse the CFG in topological order, so all predecessors of a block 1207// (excluding back-edges) are visited before the block itself. At 1208// each point in the code, we calculate a Context, which holds the set of 1209// variable definitions which are visible at that point in execution. 1210// Visible variables are mapped to their definitions using an array that 1211// contains all definitions. 1212// 1213// At join points in the CFG, the set is computed as the intersection of 1214// the incoming sets along each edge, E.g. 1215// 1216// { Context | VarDefinitions } 1217// int x = 0; { x -> x1 | x1 = 0 } 1218// int y = 0; { x -> x1, y -> y1 | y1 = 0, x1 = 0 } 1219// if (b) x = 1; { x -> x2, y -> y1 | x2 = 1, y1 = 0, ... } 1220// else x = 2; { x -> x3, y -> y1 | x3 = 2, x2 = 1, ... } 1221// ... { y -> y1 (x is unknown) | x3 = 2, x2 = 1, ... } 1222// 1223// This is essentially a simpler and more naive version of the standard SSA 1224// algorithm. Those definitions that remain in the intersection are from blocks 1225// that strictly dominate the current block. We do not bother to insert proper 1226// phi nodes, because they are not used in our analysis; instead, wherever 1227// a phi node would be required, we simply remove that definition from the 1228// context (E.g. x above). 1229// 1230// The initial traversal does not capture back-edges, so those need to be 1231// handled on a separate pass. Whenever the first pass encounters an 1232// incoming back edge, it duplicates the context, creating new definitions 1233// that refer back to the originals. (These correspond to places where SSA 1234// might have to insert a phi node.) On the second pass, these definitions are 1235// set to NULL if the variable has changed on the back-edge (i.e. a phi 1236// node was actually required.) E.g. 1237// 1238// { Context | VarDefinitions } 1239// int x = 0, y = 0; { x -> x1, y -> y1 | y1 = 0, x1 = 0 } 1240// while (b) { x -> x2, y -> y1 | [1st:] x2=x1; [2nd:] x2=NULL; } 1241// x = x+1; { x -> x3, y -> y1 | x3 = x2 + 1, ... } 1242// ... { y -> y1 | x3 = 2, x2 = 1, ... } 1243// 1244void LocalVariableMap::traverseCFG(CFG *CFGraph, 1245 PostOrderCFGView *SortedGraph, 1246 std::vector<CFGBlockInfo> &BlockInfo) { 1247 PostOrderCFGView::CFGBlockSet VisitedBlocks(CFGraph); 1248 1249 CtxIndices.resize(CFGraph->getNumBlockIDs()); 1250 1251 for (PostOrderCFGView::iterator I = SortedGraph->begin(), 1252 E = SortedGraph->end(); I!= E; ++I) { 1253 const CFGBlock *CurrBlock = *I; 1254 int CurrBlockID = CurrBlock->getBlockID(); 1255 CFGBlockInfo *CurrBlockInfo = &BlockInfo[CurrBlockID]; 1256 1257 VisitedBlocks.insert(CurrBlock); 1258 1259 // Calculate the entry context for the current block 1260 bool HasBackEdges = false; 1261 bool CtxInit = true; 1262 for (CFGBlock::const_pred_iterator PI = CurrBlock->pred_begin(), 1263 PE = CurrBlock->pred_end(); PI != PE; ++PI) { 1264 // if *PI -> CurrBlock is a back edge, so skip it 1265 if (*PI == 0 || !VisitedBlocks.alreadySet(*PI)) { 1266 HasBackEdges = true; 1267 continue; 1268 } 1269 1270 int PrevBlockID = (*PI)->getBlockID(); 1271 CFGBlockInfo *PrevBlockInfo = &BlockInfo[PrevBlockID]; 1272 1273 if (CtxInit) { 1274 CurrBlockInfo->EntryContext = PrevBlockInfo->ExitContext; 1275 CtxInit = false; 1276 } 1277 else { 1278 CurrBlockInfo->EntryContext = 1279 intersectContexts(CurrBlockInfo->EntryContext, 1280 PrevBlockInfo->ExitContext); 1281 } 1282 } 1283 1284 // Duplicate the context if we have back-edges, so we can call 1285 // intersectBackEdges later. 1286 if (HasBackEdges) 1287 CurrBlockInfo->EntryContext = 1288 createReferenceContext(CurrBlockInfo->EntryContext); 1289 1290 // Create a starting context index for the current block 1291 saveContext(0, CurrBlockInfo->EntryContext); 1292 CurrBlockInfo->EntryIndex = getContextIndex(); 1293 1294 // Visit all the statements in the basic block. 1295 VarMapBuilder VMapBuilder(this, CurrBlockInfo->EntryContext); 1296 for (CFGBlock::const_iterator BI = CurrBlock->begin(), 1297 BE = CurrBlock->end(); BI != BE; ++BI) { 1298 switch (BI->getKind()) { 1299 case CFGElement::Statement: { 1300 const CFGStmt *CS = cast<CFGStmt>(&*BI); 1301 VMapBuilder.Visit(const_cast<Stmt*>(CS->getStmt())); 1302 break; 1303 } 1304 default: 1305 break; 1306 } 1307 } 1308 CurrBlockInfo->ExitContext = VMapBuilder.Ctx; 1309 1310 // Mark variables on back edges as "unknown" if they've been changed. 1311 for (CFGBlock::const_succ_iterator SI = CurrBlock->succ_begin(), 1312 SE = CurrBlock->succ_end(); SI != SE; ++SI) { 1313 // if CurrBlock -> *SI is *not* a back edge 1314 if (*SI == 0 || !VisitedBlocks.alreadySet(*SI)) 1315 continue; 1316 1317 CFGBlock *FirstLoopBlock = *SI; 1318 Context LoopBegin = BlockInfo[FirstLoopBlock->getBlockID()].EntryContext; 1319 Context LoopEnd = CurrBlockInfo->ExitContext; 1320 intersectBackEdge(LoopBegin, LoopEnd); 1321 } 1322 } 1323 1324 // Put an extra entry at the end of the indexed context array 1325 unsigned exitID = CFGraph->getExit().getBlockID(); 1326 saveContext(0, BlockInfo[exitID].ExitContext); 1327} 1328 1329/// Find the appropriate source locations to use when producing diagnostics for 1330/// each block in the CFG. 1331static void findBlockLocations(CFG *CFGraph, 1332 PostOrderCFGView *SortedGraph, 1333 std::vector<CFGBlockInfo> &BlockInfo) { 1334 for (PostOrderCFGView::iterator I = SortedGraph->begin(), 1335 E = SortedGraph->end(); I!= E; ++I) { 1336 const CFGBlock *CurrBlock = *I; 1337 CFGBlockInfo *CurrBlockInfo = &BlockInfo[CurrBlock->getBlockID()]; 1338 1339 // Find the source location of the last statement in the block, if the 1340 // block is not empty. 1341 if (const Stmt *S = CurrBlock->getTerminator()) { 1342 CurrBlockInfo->EntryLoc = CurrBlockInfo->ExitLoc = S->getLocStart(); 1343 } else { 1344 for (CFGBlock::const_reverse_iterator BI = CurrBlock->rbegin(), 1345 BE = CurrBlock->rend(); BI != BE; ++BI) { 1346 // FIXME: Handle other CFGElement kinds. 1347 if (const CFGStmt *CS = dyn_cast<CFGStmt>(&*BI)) { 1348 CurrBlockInfo->ExitLoc = CS->getStmt()->getLocStart(); 1349 break; 1350 } 1351 } 1352 } 1353 1354 if (!CurrBlockInfo->ExitLoc.isInvalid()) { 1355 // This block contains at least one statement. Find the source location 1356 // of the first statement in the block. 1357 for (CFGBlock::const_iterator BI = CurrBlock->begin(), 1358 BE = CurrBlock->end(); BI != BE; ++BI) { 1359 // FIXME: Handle other CFGElement kinds. 1360 if (const CFGStmt *CS = dyn_cast<CFGStmt>(&*BI)) { 1361 CurrBlockInfo->EntryLoc = CS->getStmt()->getLocStart(); 1362 break; 1363 } 1364 } 1365 } else if (CurrBlock->pred_size() == 1 && *CurrBlock->pred_begin() && 1366 CurrBlock != &CFGraph->getExit()) { 1367 // The block is empty, and has a single predecessor. Use its exit 1368 // location. 1369 CurrBlockInfo->EntryLoc = CurrBlockInfo->ExitLoc = 1370 BlockInfo[(*CurrBlock->pred_begin())->getBlockID()].ExitLoc; 1371 } 1372 } 1373} 1374 1375/// \brief Class which implements the core thread safety analysis routines. 1376class ThreadSafetyAnalyzer { 1377 friend class BuildLockset; 1378 1379 ThreadSafetyHandler &Handler; 1380 LocalVariableMap LocalVarMap; 1381 FactManager FactMan; 1382 std::vector<CFGBlockInfo> BlockInfo; 1383 1384public: 1385 ThreadSafetyAnalyzer(ThreadSafetyHandler &H) : Handler(H) {} 1386 1387 void addLock(FactSet &FSet, const SExpr &Mutex, const LockData &LDat); 1388 void removeLock(FactSet &FSet, const SExpr &Mutex, 1389 SourceLocation UnlockLoc, bool FullyRemove=false); 1390 1391 template <typename AttrType> 1392 void getMutexIDs(MutexIDList &Mtxs, AttrType *Attr, Expr *Exp, 1393 const NamedDecl *D); 1394 1395 template <class AttrType> 1396 void getMutexIDs(MutexIDList &Mtxs, AttrType *Attr, Expr *Exp, 1397 const NamedDecl *D, 1398 const CFGBlock *PredBlock, const CFGBlock *CurrBlock, 1399 Expr *BrE, bool Neg); 1400 1401 const CallExpr* getTrylockCallExpr(const Stmt *Cond, LocalVarContext C, 1402 bool &Negate); 1403 1404 void getEdgeLockset(FactSet &Result, const FactSet &ExitSet, 1405 const CFGBlock* PredBlock, 1406 const CFGBlock *CurrBlock); 1407 1408 void intersectAndWarn(FactSet &FSet1, const FactSet &FSet2, 1409 SourceLocation JoinLoc, 1410 LockErrorKind LEK1, LockErrorKind LEK2, 1411 bool Modify=true); 1412 1413 void intersectAndWarn(FactSet &FSet1, const FactSet &FSet2, 1414 SourceLocation JoinLoc, LockErrorKind LEK1, 1415 bool Modify=true) { 1416 intersectAndWarn(FSet1, FSet2, JoinLoc, LEK1, LEK1, Modify); 1417 } 1418 1419 void runAnalysis(AnalysisDeclContext &AC); 1420}; 1421 1422 1423/// \brief Add a new lock to the lockset, warning if the lock is already there. 1424/// \param Mutex -- the Mutex expression for the lock 1425/// \param LDat -- the LockData for the lock 1426void ThreadSafetyAnalyzer::addLock(FactSet &FSet, const SExpr &Mutex, 1427 const LockData &LDat) { 1428 // FIXME: deal with acquired before/after annotations. 1429 // FIXME: Don't always warn when we have support for reentrant locks. 1430 if (Mutex.shouldIgnore()) 1431 return; 1432 1433 if (FSet.findLock(FactMan, Mutex)) { 1434 Handler.handleDoubleLock(Mutex.toString(), LDat.AcquireLoc); 1435 } else { 1436 FSet.addLock(FactMan, Mutex, LDat); 1437 } 1438} 1439 1440 1441/// \brief Remove a lock from the lockset, warning if the lock is not there. 1442/// \param Mutex The lock expression corresponding to the lock to be removed 1443/// \param UnlockLoc The source location of the unlock (only used in error msg) 1444void ThreadSafetyAnalyzer::removeLock(FactSet &FSet, 1445 const SExpr &Mutex, 1446 SourceLocation UnlockLoc, 1447 bool FullyRemove) { 1448 if (Mutex.shouldIgnore()) 1449 return; 1450 1451 const LockData *LDat = FSet.findLock(FactMan, Mutex); 1452 if (!LDat) { 1453 Handler.handleUnmatchedUnlock(Mutex.toString(), UnlockLoc); 1454 return; 1455 } 1456 1457 if (LDat->UnderlyingMutex.isValid()) { 1458 // This is scoped lockable object, which manages the real mutex. 1459 if (FullyRemove) { 1460 // We're destroying the managing object. 1461 // Remove the underlying mutex if it exists; but don't warn. 1462 if (FSet.findLock(FactMan, LDat->UnderlyingMutex)) 1463 FSet.removeLock(FactMan, LDat->UnderlyingMutex); 1464 } else { 1465 // We're releasing the underlying mutex, but not destroying the 1466 // managing object. Warn on dual release. 1467 if (!FSet.findLock(FactMan, LDat->UnderlyingMutex)) { 1468 Handler.handleUnmatchedUnlock(LDat->UnderlyingMutex.toString(), 1469 UnlockLoc); 1470 } 1471 FSet.removeLock(FactMan, LDat->UnderlyingMutex); 1472 return; 1473 } 1474 } 1475 FSet.removeLock(FactMan, Mutex); 1476} 1477 1478 1479/// \brief Extract the list of mutexIDs from the attribute on an expression, 1480/// and push them onto Mtxs, discarding any duplicates. 1481template <typename AttrType> 1482void ThreadSafetyAnalyzer::getMutexIDs(MutexIDList &Mtxs, AttrType *Attr, 1483 Expr *Exp, const NamedDecl *D) { 1484 typedef typename AttrType::args_iterator iterator_type; 1485 1486 if (Attr->args_size() == 0) { 1487 // The mutex held is the "this" object. 1488 SExpr Mu(0, Exp, D); 1489 if (!Mu.isValid()) 1490 SExpr::warnInvalidLock(Handler, 0, Exp, D); 1491 else 1492 Mtxs.push_back_nodup(Mu); 1493 return; 1494 } 1495 1496 for (iterator_type I=Attr->args_begin(), E=Attr->args_end(); I != E; ++I) { 1497 SExpr Mu(*I, Exp, D); 1498 if (!Mu.isValid()) 1499 SExpr::warnInvalidLock(Handler, *I, Exp, D); 1500 else 1501 Mtxs.push_back_nodup(Mu); 1502 } 1503} 1504 1505 1506/// \brief Extract the list of mutexIDs from a trylock attribute. If the 1507/// trylock applies to the given edge, then push them onto Mtxs, discarding 1508/// any duplicates. 1509template <class AttrType> 1510void ThreadSafetyAnalyzer::getMutexIDs(MutexIDList &Mtxs, AttrType *Attr, 1511 Expr *Exp, const NamedDecl *D, 1512 const CFGBlock *PredBlock, 1513 const CFGBlock *CurrBlock, 1514 Expr *BrE, bool Neg) { 1515 // Find out which branch has the lock 1516 bool branch = 0; 1517 if (CXXBoolLiteralExpr *BLE = dyn_cast_or_null<CXXBoolLiteralExpr>(BrE)) { 1518 branch = BLE->getValue(); 1519 } 1520 else if (IntegerLiteral *ILE = dyn_cast_or_null<IntegerLiteral>(BrE)) { 1521 branch = ILE->getValue().getBoolValue(); 1522 } 1523 int branchnum = branch ? 0 : 1; 1524 if (Neg) branchnum = !branchnum; 1525 1526 // If we've taken the trylock branch, then add the lock 1527 int i = 0; 1528 for (CFGBlock::const_succ_iterator SI = PredBlock->succ_begin(), 1529 SE = PredBlock->succ_end(); SI != SE && i < 2; ++SI, ++i) { 1530 if (*SI == CurrBlock && i == branchnum) { 1531 getMutexIDs(Mtxs, Attr, Exp, D); 1532 } 1533 } 1534} 1535 1536 1537bool getStaticBooleanValue(Expr* E, bool& TCond) { 1538 if (isa<CXXNullPtrLiteralExpr>(E) || isa<GNUNullExpr>(E)) { 1539 TCond = false; 1540 return true; 1541 } else if (CXXBoolLiteralExpr *BLE = dyn_cast<CXXBoolLiteralExpr>(E)) { 1542 TCond = BLE->getValue(); 1543 return true; 1544 } else if (IntegerLiteral *ILE = dyn_cast<IntegerLiteral>(E)) { 1545 TCond = ILE->getValue().getBoolValue(); 1546 return true; 1547 } else if (ImplicitCastExpr *CE = dyn_cast<ImplicitCastExpr>(E)) { 1548 return getStaticBooleanValue(CE->getSubExpr(), TCond); 1549 } 1550 return false; 1551} 1552 1553 1554// If Cond can be traced back to a function call, return the call expression. 1555// The negate variable should be called with false, and will be set to true 1556// if the function call is negated, e.g. if (!mu.tryLock(...)) 1557const CallExpr* ThreadSafetyAnalyzer::getTrylockCallExpr(const Stmt *Cond, 1558 LocalVarContext C, 1559 bool &Negate) { 1560 if (!Cond) 1561 return 0; 1562 1563 if (const CallExpr *CallExp = dyn_cast<CallExpr>(Cond)) { 1564 return CallExp; 1565 } 1566 else if (const ParenExpr *PE = dyn_cast<ParenExpr>(Cond)) { 1567 return getTrylockCallExpr(PE->getSubExpr(), C, Negate); 1568 } 1569 else if (const ImplicitCastExpr *CE = dyn_cast<ImplicitCastExpr>(Cond)) { 1570 return getTrylockCallExpr(CE->getSubExpr(), C, Negate); 1571 } 1572 else if (const ExprWithCleanups* EWC = dyn_cast<ExprWithCleanups>(Cond)) { 1573 return getTrylockCallExpr(EWC->getSubExpr(), C, Negate); 1574 } 1575 else if (const DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(Cond)) { 1576 const Expr *E = LocalVarMap.lookupExpr(DRE->getDecl(), C); 1577 return getTrylockCallExpr(E, C, Negate); 1578 } 1579 else if (const UnaryOperator *UOP = dyn_cast<UnaryOperator>(Cond)) { 1580 if (UOP->getOpcode() == UO_LNot) { 1581 Negate = !Negate; 1582 return getTrylockCallExpr(UOP->getSubExpr(), C, Negate); 1583 } 1584 return 0; 1585 } 1586 else if (const BinaryOperator *BOP = dyn_cast<BinaryOperator>(Cond)) { 1587 if (BOP->getOpcode() == BO_EQ || BOP->getOpcode() == BO_NE) { 1588 if (BOP->getOpcode() == BO_NE) 1589 Negate = !Negate; 1590 1591 bool TCond = false; 1592 if (getStaticBooleanValue(BOP->getRHS(), TCond)) { 1593 if (!TCond) Negate = !Negate; 1594 return getTrylockCallExpr(BOP->getLHS(), C, Negate); 1595 } 1596 else if (getStaticBooleanValue(BOP->getLHS(), TCond)) { 1597 if (!TCond) Negate = !Negate; 1598 return getTrylockCallExpr(BOP->getRHS(), C, Negate); 1599 } 1600 return 0; 1601 } 1602 return 0; 1603 } 1604 // FIXME -- handle && and || as well. 1605 return 0; 1606} 1607 1608 1609/// \brief Find the lockset that holds on the edge between PredBlock 1610/// and CurrBlock. The edge set is the exit set of PredBlock (passed 1611/// as the ExitSet parameter) plus any trylocks, which are conditionally held. 1612void ThreadSafetyAnalyzer::getEdgeLockset(FactSet& Result, 1613 const FactSet &ExitSet, 1614 const CFGBlock *PredBlock, 1615 const CFGBlock *CurrBlock) { 1616 Result = ExitSet; 1617 1618 if (!PredBlock->getTerminatorCondition()) 1619 return; 1620 1621 bool Negate = false; 1622 const Stmt *Cond = PredBlock->getTerminatorCondition(); 1623 const CFGBlockInfo *PredBlockInfo = &BlockInfo[PredBlock->getBlockID()]; 1624 const LocalVarContext &LVarCtx = PredBlockInfo->ExitContext; 1625 1626 CallExpr *Exp = 1627 const_cast<CallExpr*>(getTrylockCallExpr(Cond, LVarCtx, Negate)); 1628 if (!Exp) 1629 return; 1630 1631 NamedDecl *FunDecl = dyn_cast_or_null<NamedDecl>(Exp->getCalleeDecl()); 1632 if(!FunDecl || !FunDecl->hasAttrs()) 1633 return; 1634 1635 1636 MutexIDList ExclusiveLocksToAdd; 1637 MutexIDList SharedLocksToAdd; 1638 1639 // If the condition is a call to a Trylock function, then grab the attributes 1640 AttrVec &ArgAttrs = FunDecl->getAttrs(); 1641 for (unsigned i = 0; i < ArgAttrs.size(); ++i) { 1642 Attr *Attr = ArgAttrs[i]; 1643 switch (Attr->getKind()) { 1644 case attr::ExclusiveTrylockFunction: { 1645 ExclusiveTrylockFunctionAttr *A = 1646 cast<ExclusiveTrylockFunctionAttr>(Attr); 1647 getMutexIDs(ExclusiveLocksToAdd, A, Exp, FunDecl, 1648 PredBlock, CurrBlock, A->getSuccessValue(), Negate); 1649 break; 1650 } 1651 case attr::SharedTrylockFunction: { 1652 SharedTrylockFunctionAttr *A = 1653 cast<SharedTrylockFunctionAttr>(Attr); 1654 getMutexIDs(ExclusiveLocksToAdd, A, Exp, FunDecl, 1655 PredBlock, CurrBlock, A->getSuccessValue(), Negate); 1656 break; 1657 } 1658 default: 1659 break; 1660 } 1661 } 1662 1663 // Add and remove locks. 1664 SourceLocation Loc = Exp->getExprLoc(); 1665 for (unsigned i=0,n=ExclusiveLocksToAdd.size(); i<n; ++i) { 1666 addLock(Result, ExclusiveLocksToAdd[i], 1667 LockData(Loc, LK_Exclusive)); 1668 } 1669 for (unsigned i=0,n=SharedLocksToAdd.size(); i<n; ++i) { 1670 addLock(Result, SharedLocksToAdd[i], 1671 LockData(Loc, LK_Shared)); 1672 } 1673} 1674 1675 1676/// \brief We use this class to visit different types of expressions in 1677/// CFGBlocks, and build up the lockset. 1678/// An expression may cause us to add or remove locks from the lockset, or else 1679/// output error messages related to missing locks. 1680/// FIXME: In future, we may be able to not inherit from a visitor. 1681class BuildLockset : public StmtVisitor<BuildLockset> { 1682 friend class ThreadSafetyAnalyzer; 1683 1684 ThreadSafetyAnalyzer *Analyzer; 1685 FactSet FSet; 1686 LocalVariableMap::Context LVarCtx; 1687 unsigned CtxIndex; 1688 1689 // Helper functions 1690 const ValueDecl *getValueDecl(Expr *Exp); 1691 1692 void warnIfMutexNotHeld(const NamedDecl *D, Expr *Exp, AccessKind AK, 1693 Expr *MutexExp, ProtectedOperationKind POK); 1694 void warnIfMutexHeld(const NamedDecl *D, Expr *Exp, Expr *MutexExp); 1695 1696 void checkAccess(Expr *Exp, AccessKind AK); 1697 void checkDereference(Expr *Exp, AccessKind AK); 1698 void handleCall(Expr *Exp, const NamedDecl *D, VarDecl *VD = 0); 1699 1700public: 1701 BuildLockset(ThreadSafetyAnalyzer *Anlzr, CFGBlockInfo &Info) 1702 : StmtVisitor<BuildLockset>(), 1703 Analyzer(Anlzr), 1704 FSet(Info.EntrySet), 1705 LVarCtx(Info.EntryContext), 1706 CtxIndex(Info.EntryIndex) 1707 {} 1708 1709 void VisitUnaryOperator(UnaryOperator *UO); 1710 void VisitBinaryOperator(BinaryOperator *BO); 1711 void VisitCastExpr(CastExpr *CE); 1712 void VisitCallExpr(CallExpr *Exp); 1713 void VisitCXXConstructExpr(CXXConstructExpr *Exp); 1714 void VisitDeclStmt(DeclStmt *S); 1715}; 1716 1717 1718/// \brief Gets the value decl pointer from DeclRefExprs or MemberExprs 1719const ValueDecl *BuildLockset::getValueDecl(Expr *Exp) { 1720 if (const DeclRefExpr *DR = dyn_cast<DeclRefExpr>(Exp)) 1721 return DR->getDecl(); 1722 1723 if (const MemberExpr *ME = dyn_cast<MemberExpr>(Exp)) 1724 return ME->getMemberDecl(); 1725 1726 return 0; 1727} 1728 1729/// \brief Warn if the LSet does not contain a lock sufficient to protect access 1730/// of at least the passed in AccessKind. 1731void BuildLockset::warnIfMutexNotHeld(const NamedDecl *D, Expr *Exp, 1732 AccessKind AK, Expr *MutexExp, 1733 ProtectedOperationKind POK) { 1734 LockKind LK = getLockKindFromAccessKind(AK); 1735 1736 SExpr Mutex(MutexExp, Exp, D); 1737 if (!Mutex.isValid()) { 1738 SExpr::warnInvalidLock(Analyzer->Handler, MutexExp, Exp, D); 1739 return; 1740 } else if (Mutex.shouldIgnore()) { 1741 return; 1742 } 1743 1744 LockData* LDat = FSet.findLockUniv(Analyzer->FactMan, Mutex); 1745 if (!LDat || !LDat->isAtLeast(LK)) 1746 Analyzer->Handler.handleMutexNotHeld(D, POK, Mutex.toString(), LK, 1747 Exp->getExprLoc()); 1748} 1749 1750/// \brief Warn if the LSet contains the given lock. 1751void BuildLockset::warnIfMutexHeld(const NamedDecl *D, Expr* Exp, 1752 Expr *MutexExp) { 1753 SExpr Mutex(MutexExp, Exp, D); 1754 if (!Mutex.isValid()) { 1755 SExpr::warnInvalidLock(Analyzer->Handler, MutexExp, Exp, D); 1756 return; 1757 } 1758 1759 LockData* LDat = FSet.findLock(Analyzer->FactMan, Mutex); 1760 if (LDat) 1761 Analyzer->Handler.handleFunExcludesLock(D->getName(), Mutex.toString(), 1762 Exp->getExprLoc()); 1763} 1764 1765 1766/// \brief This method identifies variable dereferences and checks pt_guarded_by 1767/// and pt_guarded_var annotations. Note that we only check these annotations 1768/// at the time a pointer is dereferenced. 1769/// FIXME: We need to check for other types of pointer dereferences 1770/// (e.g. [], ->) and deal with them here. 1771/// \param Exp An expression that has been read or written. 1772void BuildLockset::checkDereference(Expr *Exp, AccessKind AK) { 1773 UnaryOperator *UO = dyn_cast<UnaryOperator>(Exp); 1774 if (!UO || UO->getOpcode() != clang::UO_Deref) 1775 return; 1776 Exp = UO->getSubExpr()->IgnoreParenCasts(); 1777 1778 const ValueDecl *D = getValueDecl(Exp); 1779 if(!D || !D->hasAttrs()) 1780 return; 1781 1782 if (D->getAttr<PtGuardedVarAttr>() && FSet.isEmpty()) 1783 Analyzer->Handler.handleNoMutexHeld(D, POK_VarDereference, AK, 1784 Exp->getExprLoc()); 1785 1786 const AttrVec &ArgAttrs = D->getAttrs(); 1787 for(unsigned i = 0, Size = ArgAttrs.size(); i < Size; ++i) 1788 if (PtGuardedByAttr *PGBAttr = dyn_cast<PtGuardedByAttr>(ArgAttrs[i])) 1789 warnIfMutexNotHeld(D, Exp, AK, PGBAttr->getArg(), POK_VarDereference); 1790} 1791 1792/// \brief Checks guarded_by and guarded_var attributes. 1793/// Whenever we identify an access (read or write) of a DeclRefExpr or 1794/// MemberExpr, we need to check whether there are any guarded_by or 1795/// guarded_var attributes, and make sure we hold the appropriate mutexes. 1796void BuildLockset::checkAccess(Expr *Exp, AccessKind AK) { 1797 const ValueDecl *D = getValueDecl(Exp); 1798 if(!D || !D->hasAttrs()) 1799 return; 1800 1801 if (D->getAttr<GuardedVarAttr>() && FSet.isEmpty()) 1802 Analyzer->Handler.handleNoMutexHeld(D, POK_VarAccess, AK, 1803 Exp->getExprLoc()); 1804 1805 const AttrVec &ArgAttrs = D->getAttrs(); 1806 for(unsigned i = 0, Size = ArgAttrs.size(); i < Size; ++i) 1807 if (GuardedByAttr *GBAttr = dyn_cast<GuardedByAttr>(ArgAttrs[i])) 1808 warnIfMutexNotHeld(D, Exp, AK, GBAttr->getArg(), POK_VarAccess); 1809} 1810 1811/// \brief Process a function call, method call, constructor call, 1812/// or destructor call. This involves looking at the attributes on the 1813/// corresponding function/method/constructor/destructor, issuing warnings, 1814/// and updating the locksets accordingly. 1815/// 1816/// FIXME: For classes annotated with one of the guarded annotations, we need 1817/// to treat const method calls as reads and non-const method calls as writes, 1818/// and check that the appropriate locks are held. Non-const method calls with 1819/// the same signature as const method calls can be also treated as reads. 1820/// 1821void BuildLockset::handleCall(Expr *Exp, const NamedDecl *D, VarDecl *VD) { 1822 const AttrVec &ArgAttrs = D->getAttrs(); 1823 MutexIDList ExclusiveLocksToAdd; 1824 MutexIDList SharedLocksToAdd; 1825 MutexIDList LocksToRemove; 1826 1827 for(unsigned i = 0; i < ArgAttrs.size(); ++i) { 1828 Attr *At = const_cast<Attr*>(ArgAttrs[i]); 1829 switch (At->getKind()) { 1830 // When we encounter an exclusive lock function, we need to add the lock 1831 // to our lockset with kind exclusive. 1832 case attr::ExclusiveLockFunction: { 1833 ExclusiveLockFunctionAttr *A = cast<ExclusiveLockFunctionAttr>(At); 1834 Analyzer->getMutexIDs(ExclusiveLocksToAdd, A, Exp, D); 1835 break; 1836 } 1837 1838 // When we encounter a shared lock function, we need to add the lock 1839 // to our lockset with kind shared. 1840 case attr::SharedLockFunction: { 1841 SharedLockFunctionAttr *A = cast<SharedLockFunctionAttr>(At); 1842 Analyzer->getMutexIDs(SharedLocksToAdd, A, Exp, D); 1843 break; 1844 } 1845 1846 // When we encounter an unlock function, we need to remove unlocked 1847 // mutexes from the lockset, and flag a warning if they are not there. 1848 case attr::UnlockFunction: { 1849 UnlockFunctionAttr *A = cast<UnlockFunctionAttr>(At); 1850 Analyzer->getMutexIDs(LocksToRemove, A, Exp, D); 1851 break; 1852 } 1853 1854 case attr::ExclusiveLocksRequired: { 1855 ExclusiveLocksRequiredAttr *A = cast<ExclusiveLocksRequiredAttr>(At); 1856 1857 for (ExclusiveLocksRequiredAttr::args_iterator 1858 I = A->args_begin(), E = A->args_end(); I != E; ++I) 1859 warnIfMutexNotHeld(D, Exp, AK_Written, *I, POK_FunctionCall); 1860 break; 1861 } 1862 1863 case attr::SharedLocksRequired: { 1864 SharedLocksRequiredAttr *A = cast<SharedLocksRequiredAttr>(At); 1865 1866 for (SharedLocksRequiredAttr::args_iterator I = A->args_begin(), 1867 E = A->args_end(); I != E; ++I) 1868 warnIfMutexNotHeld(D, Exp, AK_Read, *I, POK_FunctionCall); 1869 break; 1870 } 1871 1872 case attr::LocksExcluded: { 1873 LocksExcludedAttr *A = cast<LocksExcludedAttr>(At); 1874 1875 for (LocksExcludedAttr::args_iterator I = A->args_begin(), 1876 E = A->args_end(); I != E; ++I) { 1877 warnIfMutexHeld(D, Exp, *I); 1878 } 1879 break; 1880 } 1881 1882 // Ignore other (non thread-safety) attributes 1883 default: 1884 break; 1885 } 1886 } 1887 1888 // Figure out if we're calling the constructor of scoped lockable class 1889 bool isScopedVar = false; 1890 if (VD) { 1891 if (const CXXConstructorDecl *CD = dyn_cast<const CXXConstructorDecl>(D)) { 1892 const CXXRecordDecl* PD = CD->getParent(); 1893 if (PD && PD->getAttr<ScopedLockableAttr>()) 1894 isScopedVar = true; 1895 } 1896 } 1897 1898 // Add locks. 1899 SourceLocation Loc = Exp->getExprLoc(); 1900 for (unsigned i=0,n=ExclusiveLocksToAdd.size(); i<n; ++i) { 1901 Analyzer->addLock(FSet, ExclusiveLocksToAdd[i], 1902 LockData(Loc, LK_Exclusive, isScopedVar)); 1903 } 1904 for (unsigned i=0,n=SharedLocksToAdd.size(); i<n; ++i) { 1905 Analyzer->addLock(FSet, SharedLocksToAdd[i], 1906 LockData(Loc, LK_Shared, isScopedVar)); 1907 } 1908 1909 // Add the managing object as a dummy mutex, mapped to the underlying mutex. 1910 // FIXME -- this doesn't work if we acquire multiple locks. 1911 if (isScopedVar) { 1912 SourceLocation MLoc = VD->getLocation(); 1913 DeclRefExpr DRE(VD, false, VD->getType(), VK_LValue, VD->getLocation()); 1914 SExpr SMutex(&DRE, 0, 0); 1915 1916 for (unsigned i=0,n=ExclusiveLocksToAdd.size(); i<n; ++i) { 1917 Analyzer->addLock(FSet, SMutex, LockData(MLoc, LK_Exclusive, 1918 ExclusiveLocksToAdd[i])); 1919 } 1920 for (unsigned i=0,n=SharedLocksToAdd.size(); i<n; ++i) { 1921 Analyzer->addLock(FSet, SMutex, LockData(MLoc, LK_Shared, 1922 SharedLocksToAdd[i])); 1923 } 1924 } 1925 1926 // Remove locks. 1927 // FIXME -- should only fully remove if the attribute refers to 'this'. 1928 bool Dtor = isa<CXXDestructorDecl>(D); 1929 for (unsigned i=0,n=LocksToRemove.size(); i<n; ++i) { 1930 Analyzer->removeLock(FSet, LocksToRemove[i], Loc, Dtor); 1931 } 1932} 1933 1934 1935/// \brief For unary operations which read and write a variable, we need to 1936/// check whether we hold any required mutexes. Reads are checked in 1937/// VisitCastExpr. 1938void BuildLockset::VisitUnaryOperator(UnaryOperator *UO) { 1939 switch (UO->getOpcode()) { 1940 case clang::UO_PostDec: 1941 case clang::UO_PostInc: 1942 case clang::UO_PreDec: 1943 case clang::UO_PreInc: { 1944 Expr *SubExp = UO->getSubExpr()->IgnoreParenCasts(); 1945 checkAccess(SubExp, AK_Written); 1946 checkDereference(SubExp, AK_Written); 1947 break; 1948 } 1949 default: 1950 break; 1951 } 1952} 1953 1954/// For binary operations which assign to a variable (writes), we need to check 1955/// whether we hold any required mutexes. 1956/// FIXME: Deal with non-primitive types. 1957void BuildLockset::VisitBinaryOperator(BinaryOperator *BO) { 1958 if (!BO->isAssignmentOp()) 1959 return; 1960 1961 // adjust the context 1962 LVarCtx = Analyzer->LocalVarMap.getNextContext(CtxIndex, BO, LVarCtx); 1963 1964 Expr *LHSExp = BO->getLHS()->IgnoreParenCasts(); 1965 checkAccess(LHSExp, AK_Written); 1966 checkDereference(LHSExp, AK_Written); 1967} 1968 1969/// Whenever we do an LValue to Rvalue cast, we are reading a variable and 1970/// need to ensure we hold any required mutexes. 1971/// FIXME: Deal with non-primitive types. 1972void BuildLockset::VisitCastExpr(CastExpr *CE) { 1973 if (CE->getCastKind() != CK_LValueToRValue) 1974 return; 1975 Expr *SubExp = CE->getSubExpr()->IgnoreParenCasts(); 1976 checkAccess(SubExp, AK_Read); 1977 checkDereference(SubExp, AK_Read); 1978} 1979 1980 1981void BuildLockset::VisitCallExpr(CallExpr *Exp) { 1982 NamedDecl *D = dyn_cast_or_null<NamedDecl>(Exp->getCalleeDecl()); 1983 if(!D || !D->hasAttrs()) 1984 return; 1985 handleCall(Exp, D); 1986} 1987 1988void BuildLockset::VisitCXXConstructExpr(CXXConstructExpr *Exp) { 1989 // FIXME -- only handles constructors in DeclStmt below. 1990} 1991 1992void BuildLockset::VisitDeclStmt(DeclStmt *S) { 1993 // adjust the context 1994 LVarCtx = Analyzer->LocalVarMap.getNextContext(CtxIndex, S, LVarCtx); 1995 1996 DeclGroupRef DGrp = S->getDeclGroup(); 1997 for (DeclGroupRef::iterator I = DGrp.begin(), E = DGrp.end(); I != E; ++I) { 1998 Decl *D = *I; 1999 if (VarDecl *VD = dyn_cast_or_null<VarDecl>(D)) { 2000 Expr *E = VD->getInit(); 2001 // handle constructors that involve temporaries 2002 if (ExprWithCleanups *EWC = dyn_cast_or_null<ExprWithCleanups>(E)) 2003 E = EWC->getSubExpr(); 2004 2005 if (CXXConstructExpr *CE = dyn_cast_or_null<CXXConstructExpr>(E)) { 2006 NamedDecl *CtorD = dyn_cast_or_null<NamedDecl>(CE->getConstructor()); 2007 if (!CtorD || !CtorD->hasAttrs()) 2008 return; 2009 handleCall(CE, CtorD, VD); 2010 } 2011 } 2012 } 2013} 2014 2015 2016 2017/// \brief Compute the intersection of two locksets and issue warnings for any 2018/// locks in the symmetric difference. 2019/// 2020/// This function is used at a merge point in the CFG when comparing the lockset 2021/// of each branch being merged. For example, given the following sequence: 2022/// A; if () then B; else C; D; we need to check that the lockset after B and C 2023/// are the same. In the event of a difference, we use the intersection of these 2024/// two locksets at the start of D. 2025/// 2026/// \param FSet1 The first lockset. 2027/// \param FSet2 The second lockset. 2028/// \param JoinLoc The location of the join point for error reporting 2029/// \param LEK1 The error message to report if a mutex is missing from LSet1 2030/// \param LEK2 The error message to report if a mutex is missing from Lset2 2031void ThreadSafetyAnalyzer::intersectAndWarn(FactSet &FSet1, 2032 const FactSet &FSet2, 2033 SourceLocation JoinLoc, 2034 LockErrorKind LEK1, 2035 LockErrorKind LEK2, 2036 bool Modify) { 2037 FactSet FSet1Orig = FSet1; 2038 2039 for (FactSet::const_iterator I = FSet2.begin(), E = FSet2.end(); 2040 I != E; ++I) { 2041 const SExpr &FSet2Mutex = FactMan[*I].MutID; 2042 const LockData &LDat2 = FactMan[*I].LDat; 2043 2044 if (const LockData *LDat1 = FSet1.findLock(FactMan, FSet2Mutex)) { 2045 if (LDat1->LKind != LDat2.LKind) { 2046 Handler.handleExclusiveAndShared(FSet2Mutex.toString(), 2047 LDat2.AcquireLoc, 2048 LDat1->AcquireLoc); 2049 if (Modify && LDat1->LKind != LK_Exclusive) { 2050 FSet1.removeLock(FactMan, FSet2Mutex); 2051 FSet1.addLock(FactMan, FSet2Mutex, LDat2); 2052 } 2053 } 2054 } else { 2055 if (LDat2.UnderlyingMutex.isValid()) { 2056 if (FSet2.findLock(FactMan, LDat2.UnderlyingMutex)) { 2057 // If this is a scoped lock that manages another mutex, and if the 2058 // underlying mutex is still held, then warn about the underlying 2059 // mutex. 2060 Handler.handleMutexHeldEndOfScope(LDat2.UnderlyingMutex.toString(), 2061 LDat2.AcquireLoc, 2062 JoinLoc, LEK1); 2063 } 2064 } 2065 else if (!LDat2.Managed && !FSet2Mutex.isUniversal()) 2066 Handler.handleMutexHeldEndOfScope(FSet2Mutex.toString(), 2067 LDat2.AcquireLoc, 2068 JoinLoc, LEK1); 2069 } 2070 } 2071 2072 for (FactSet::const_iterator I = FSet1.begin(), E = FSet1.end(); 2073 I != E; ++I) { 2074 const SExpr &FSet1Mutex = FactMan[*I].MutID; 2075 const LockData &LDat1 = FactMan[*I].LDat; 2076 2077 if (!FSet2.findLock(FactMan, FSet1Mutex)) { 2078 if (LDat1.UnderlyingMutex.isValid()) { 2079 if (FSet1Orig.findLock(FactMan, LDat1.UnderlyingMutex)) { 2080 // If this is a scoped lock that manages another mutex, and if the 2081 // underlying mutex is still held, then warn about the underlying 2082 // mutex. 2083 Handler.handleMutexHeldEndOfScope(LDat1.UnderlyingMutex.toString(), 2084 LDat1.AcquireLoc, 2085 JoinLoc, LEK1); 2086 } 2087 } 2088 else if (!LDat1.Managed && !FSet1Mutex.isUniversal()) 2089 Handler.handleMutexHeldEndOfScope(FSet1Mutex.toString(), 2090 LDat1.AcquireLoc, 2091 JoinLoc, LEK2); 2092 if (Modify) 2093 FSet1.removeLock(FactMan, FSet1Mutex); 2094 } 2095 } 2096} 2097 2098 2099 2100/// \brief Check a function's CFG for thread-safety violations. 2101/// 2102/// We traverse the blocks in the CFG, compute the set of mutexes that are held 2103/// at the end of each block, and issue warnings for thread safety violations. 2104/// Each block in the CFG is traversed exactly once. 2105void ThreadSafetyAnalyzer::runAnalysis(AnalysisDeclContext &AC) { 2106 CFG *CFGraph = AC.getCFG(); 2107 if (!CFGraph) return; 2108 const NamedDecl *D = dyn_cast_or_null<NamedDecl>(AC.getDecl()); 2109 2110 // AC.dumpCFG(true); 2111 2112 if (!D) 2113 return; // Ignore anonymous functions for now. 2114 if (D->getAttr<NoThreadSafetyAnalysisAttr>()) 2115 return; 2116 // FIXME: Do something a bit more intelligent inside constructor and 2117 // destructor code. Constructors and destructors must assume unique access 2118 // to 'this', so checks on member variable access is disabled, but we should 2119 // still enable checks on other objects. 2120 if (isa<CXXConstructorDecl>(D)) 2121 return; // Don't check inside constructors. 2122 if (isa<CXXDestructorDecl>(D)) 2123 return; // Don't check inside destructors. 2124 2125 BlockInfo.resize(CFGraph->getNumBlockIDs(), 2126 CFGBlockInfo::getEmptyBlockInfo(LocalVarMap)); 2127 2128 // We need to explore the CFG via a "topological" ordering. 2129 // That way, we will be guaranteed to have information about required 2130 // predecessor locksets when exploring a new block. 2131 PostOrderCFGView *SortedGraph = AC.getAnalysis<PostOrderCFGView>(); 2132 PostOrderCFGView::CFGBlockSet VisitedBlocks(CFGraph); 2133 2134 // Compute SSA names for local variables 2135 LocalVarMap.traverseCFG(CFGraph, SortedGraph, BlockInfo); 2136 2137 // Fill in source locations for all CFGBlocks. 2138 findBlockLocations(CFGraph, SortedGraph, BlockInfo); 2139 2140 // Add locks from exclusive_locks_required and shared_locks_required 2141 // to initial lockset. Also turn off checking for lock and unlock functions. 2142 // FIXME: is there a more intelligent way to check lock/unlock functions? 2143 if (!SortedGraph->empty() && D->hasAttrs()) { 2144 const CFGBlock *FirstBlock = *SortedGraph->begin(); 2145 FactSet &InitialLockset = BlockInfo[FirstBlock->getBlockID()].EntrySet; 2146 const AttrVec &ArgAttrs = D->getAttrs(); 2147 2148 MutexIDList ExclusiveLocksToAdd; 2149 MutexIDList SharedLocksToAdd; 2150 2151 SourceLocation Loc = D->getLocation(); 2152 for (unsigned i = 0; i < ArgAttrs.size(); ++i) { 2153 Attr *Attr = ArgAttrs[i]; 2154 Loc = Attr->getLocation(); 2155 if (ExclusiveLocksRequiredAttr *A 2156 = dyn_cast<ExclusiveLocksRequiredAttr>(Attr)) { 2157 getMutexIDs(ExclusiveLocksToAdd, A, (Expr*) 0, D); 2158 } else if (SharedLocksRequiredAttr *A 2159 = dyn_cast<SharedLocksRequiredAttr>(Attr)) { 2160 getMutexIDs(SharedLocksToAdd, A, (Expr*) 0, D); 2161 } else if (isa<UnlockFunctionAttr>(Attr)) { 2162 // Don't try to check unlock functions for now 2163 return; 2164 } else if (isa<ExclusiveLockFunctionAttr>(Attr)) { 2165 // Don't try to check lock functions for now 2166 return; 2167 } else if (isa<SharedLockFunctionAttr>(Attr)) { 2168 // Don't try to check lock functions for now 2169 return; 2170 } else if (isa<ExclusiveTrylockFunctionAttr>(Attr)) { 2171 // Don't try to check trylock functions for now 2172 return; 2173 } else if (isa<SharedTrylockFunctionAttr>(Attr)) { 2174 // Don't try to check trylock functions for now 2175 return; 2176 } 2177 } 2178 2179 // FIXME -- Loc can be wrong here. 2180 for (unsigned i=0,n=ExclusiveLocksToAdd.size(); i<n; ++i) { 2181 addLock(InitialLockset, ExclusiveLocksToAdd[i], 2182 LockData(Loc, LK_Exclusive)); 2183 } 2184 for (unsigned i=0,n=SharedLocksToAdd.size(); i<n; ++i) { 2185 addLock(InitialLockset, SharedLocksToAdd[i], 2186 LockData(Loc, LK_Shared)); 2187 } 2188 } 2189 2190 for (PostOrderCFGView::iterator I = SortedGraph->begin(), 2191 E = SortedGraph->end(); I!= E; ++I) { 2192 const CFGBlock *CurrBlock = *I; 2193 int CurrBlockID = CurrBlock->getBlockID(); 2194 CFGBlockInfo *CurrBlockInfo = &BlockInfo[CurrBlockID]; 2195 2196 // Use the default initial lockset in case there are no predecessors. 2197 VisitedBlocks.insert(CurrBlock); 2198 2199 // Iterate through the predecessor blocks and warn if the lockset for all 2200 // predecessors is not the same. We take the entry lockset of the current 2201 // block to be the intersection of all previous locksets. 2202 // FIXME: By keeping the intersection, we may output more errors in future 2203 // for a lock which is not in the intersection, but was in the union. We 2204 // may want to also keep the union in future. As an example, let's say 2205 // the intersection contains Mutex L, and the union contains L and M. 2206 // Later we unlock M. At this point, we would output an error because we 2207 // never locked M; although the real error is probably that we forgot to 2208 // lock M on all code paths. Conversely, let's say that later we lock M. 2209 // In this case, we should compare against the intersection instead of the 2210 // union because the real error is probably that we forgot to unlock M on 2211 // all code paths. 2212 bool LocksetInitialized = false; 2213 llvm::SmallVector<CFGBlock*, 8> SpecialBlocks; 2214 for (CFGBlock::const_pred_iterator PI = CurrBlock->pred_begin(), 2215 PE = CurrBlock->pred_end(); PI != PE; ++PI) { 2216 2217 // if *PI -> CurrBlock is a back edge 2218 if (*PI == 0 || !VisitedBlocks.alreadySet(*PI)) 2219 continue; 2220 2221 // Ignore edges from blocks that can't return. 2222 if ((*PI)->hasNoReturnElement()) 2223 continue; 2224 2225 // If the previous block ended in a 'continue' or 'break' statement, then 2226 // a difference in locksets is probably due to a bug in that block, rather 2227 // than in some other predecessor. In that case, keep the other 2228 // predecessor's lockset. 2229 if (const Stmt *Terminator = (*PI)->getTerminator()) { 2230 if (isa<ContinueStmt>(Terminator) || isa<BreakStmt>(Terminator)) { 2231 SpecialBlocks.push_back(*PI); 2232 continue; 2233 } 2234 } 2235 2236 int PrevBlockID = (*PI)->getBlockID(); 2237 CFGBlockInfo *PrevBlockInfo = &BlockInfo[PrevBlockID]; 2238 FactSet PrevLockset; 2239 getEdgeLockset(PrevLockset, PrevBlockInfo->ExitSet, *PI, CurrBlock); 2240 2241 if (!LocksetInitialized) { 2242 CurrBlockInfo->EntrySet = PrevLockset; 2243 LocksetInitialized = true; 2244 } else { 2245 intersectAndWarn(CurrBlockInfo->EntrySet, PrevLockset, 2246 CurrBlockInfo->EntryLoc, 2247 LEK_LockedSomePredecessors); 2248 } 2249 } 2250 2251 // Process continue and break blocks. Assume that the lockset for the 2252 // resulting block is unaffected by any discrepancies in them. 2253 for (unsigned SpecialI = 0, SpecialN = SpecialBlocks.size(); 2254 SpecialI < SpecialN; ++SpecialI) { 2255 CFGBlock *PrevBlock = SpecialBlocks[SpecialI]; 2256 int PrevBlockID = PrevBlock->getBlockID(); 2257 CFGBlockInfo *PrevBlockInfo = &BlockInfo[PrevBlockID]; 2258 2259 if (!LocksetInitialized) { 2260 CurrBlockInfo->EntrySet = PrevBlockInfo->ExitSet; 2261 LocksetInitialized = true; 2262 } else { 2263 // Determine whether this edge is a loop terminator for diagnostic 2264 // purposes. FIXME: A 'break' statement might be a loop terminator, but 2265 // it might also be part of a switch. Also, a subsequent destructor 2266 // might add to the lockset, in which case the real issue might be a 2267 // double lock on the other path. 2268 const Stmt *Terminator = PrevBlock->getTerminator(); 2269 bool IsLoop = Terminator && isa<ContinueStmt>(Terminator); 2270 2271 FactSet PrevLockset; 2272 getEdgeLockset(PrevLockset, PrevBlockInfo->ExitSet, 2273 PrevBlock, CurrBlock); 2274 2275 // Do not update EntrySet. 2276 intersectAndWarn(CurrBlockInfo->EntrySet, PrevLockset, 2277 PrevBlockInfo->ExitLoc, 2278 IsLoop ? LEK_LockedSomeLoopIterations 2279 : LEK_LockedSomePredecessors, 2280 false); 2281 } 2282 } 2283 2284 BuildLockset LocksetBuilder(this, *CurrBlockInfo); 2285 2286 // Visit all the statements in the basic block. 2287 for (CFGBlock::const_iterator BI = CurrBlock->begin(), 2288 BE = CurrBlock->end(); BI != BE; ++BI) { 2289 switch (BI->getKind()) { 2290 case CFGElement::Statement: { 2291 const CFGStmt *CS = cast<CFGStmt>(&*BI); 2292 LocksetBuilder.Visit(const_cast<Stmt*>(CS->getStmt())); 2293 break; 2294 } 2295 // Ignore BaseDtor, MemberDtor, and TemporaryDtor for now. 2296 case CFGElement::AutomaticObjectDtor: { 2297 const CFGAutomaticObjDtor *AD = cast<CFGAutomaticObjDtor>(&*BI); 2298 CXXDestructorDecl *DD = const_cast<CXXDestructorDecl*>( 2299 AD->getDestructorDecl(AC.getASTContext())); 2300 if (!DD->hasAttrs()) 2301 break; 2302 2303 // Create a dummy expression, 2304 VarDecl *VD = const_cast<VarDecl*>(AD->getVarDecl()); 2305 DeclRefExpr DRE(VD, false, VD->getType(), VK_LValue, 2306 AD->getTriggerStmt()->getLocEnd()); 2307 LocksetBuilder.handleCall(&DRE, DD); 2308 break; 2309 } 2310 default: 2311 break; 2312 } 2313 } 2314 CurrBlockInfo->ExitSet = LocksetBuilder.FSet; 2315 2316 // For every back edge from CurrBlock (the end of the loop) to another block 2317 // (FirstLoopBlock) we need to check that the Lockset of Block is equal to 2318 // the one held at the beginning of FirstLoopBlock. We can look up the 2319 // Lockset held at the beginning of FirstLoopBlock in the EntryLockSets map. 2320 for (CFGBlock::const_succ_iterator SI = CurrBlock->succ_begin(), 2321 SE = CurrBlock->succ_end(); SI != SE; ++SI) { 2322 2323 // if CurrBlock -> *SI is *not* a back edge 2324 if (*SI == 0 || !VisitedBlocks.alreadySet(*SI)) 2325 continue; 2326 2327 CFGBlock *FirstLoopBlock = *SI; 2328 CFGBlockInfo *PreLoop = &BlockInfo[FirstLoopBlock->getBlockID()]; 2329 CFGBlockInfo *LoopEnd = &BlockInfo[CurrBlockID]; 2330 intersectAndWarn(LoopEnd->ExitSet, PreLoop->EntrySet, 2331 PreLoop->EntryLoc, 2332 LEK_LockedSomeLoopIterations, 2333 false); 2334 } 2335 } 2336 2337 CFGBlockInfo *Initial = &BlockInfo[CFGraph->getEntry().getBlockID()]; 2338 CFGBlockInfo *Final = &BlockInfo[CFGraph->getExit().getBlockID()]; 2339 2340 // FIXME: Should we call this function for all blocks which exit the function? 2341 intersectAndWarn(Initial->EntrySet, Final->ExitSet, 2342 Final->ExitLoc, 2343 LEK_LockedAtEndOfFunction, 2344 LEK_NotLockedAtEndOfFunction, 2345 false); 2346} 2347 2348} // end anonymous namespace 2349 2350 2351namespace clang { 2352namespace thread_safety { 2353 2354/// \brief Check a function's CFG for thread-safety violations. 2355/// 2356/// We traverse the blocks in the CFG, compute the set of mutexes that are held 2357/// at the end of each block, and issue warnings for thread safety violations. 2358/// Each block in the CFG is traversed exactly once. 2359void runThreadSafetyAnalysis(AnalysisDeclContext &AC, 2360 ThreadSafetyHandler &Handler) { 2361 ThreadSafetyAnalyzer Analyzer(Handler); 2362 Analyzer.runAnalysis(AC); 2363} 2364 2365/// \brief Helper function that returns a LockKind required for the given level 2366/// of access. 2367LockKind getLockKindFromAccessKind(AccessKind AK) { 2368 switch (AK) { 2369 case AK_Read : 2370 return LK_Shared; 2371 case AK_Written : 2372 return LK_Exclusive; 2373 } 2374 llvm_unreachable("Unknown AccessKind"); 2375} 2376 2377}} // end namespace clang::thread_safety 2378