ThreadSafety.cpp revision dd0a1f58a505ced9674f6d31a1ff65cb87774d67
1c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru//===- ThreadSafety.cpp ----------------------------------------*- C++ --*-===// 2c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru// 3c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru// The LLVM Compiler Infrastructure 4c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru// 5c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru// This file is distributed under the University of Illinois Open Source 6c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru// License. See LICENSE.TXT for details. 7c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru// 8c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru//===----------------------------------------------------------------------===// 9c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru// 10c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru// A intra-procedural analysis for thread safety (e.g. deadlocks and race 11c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru// conditions), based off of an annotation system. 12c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru// 13c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru// See http://clang.llvm.org/docs/LanguageExtensions.html#threadsafety for more 14c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru// information. 15c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru// 16c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru//===----------------------------------------------------------------------===// 17c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru 18c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include "clang/Analysis/Analyses/ThreadSafety.h" 19c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include "clang/AST/Attr.h" 20c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include "clang/AST/DeclCXX.h" 21c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include "clang/AST/ExprCXX.h" 22c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include "clang/AST/StmtCXX.h" 23c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include "clang/AST/StmtVisitor.h" 24c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include "clang/Analysis/Analyses/PostOrderCFGView.h" 25c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include "clang/Analysis/AnalysisContext.h" 26c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include "clang/Analysis/CFG.h" 27c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include "clang/Analysis/CFGStmtMap.h" 28c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include "clang/Basic/OperatorKinds.h" 29c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include "clang/Basic/SourceLocation.h" 30c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include "clang/Basic/SourceManager.h" 31c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include "llvm/ADT/BitVector.h" 32c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include "llvm/ADT/FoldingSet.h" 33c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include "llvm/ADT/ImmutableMap.h" 34c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include "llvm/ADT/PostOrderIterator.h" 35c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include "llvm/ADT/SmallVector.h" 36c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include "llvm/ADT/StringRef.h" 37c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include "llvm/Support/raw_ostream.h" 38c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include <algorithm> 39c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include <utility> 40c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru#include <vector> 41c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru 42c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queruusing namespace clang; 43c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queruusing namespace thread_safety; 44c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru 45c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru// Key method definition 46c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste QueruThreadSafetyHandler::~ThreadSafetyHandler() {} 47c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru 48c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Querunamespace { 49c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru 50c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/// SExpr implements a simple expression language that is used to store, 51c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/// compare, and pretty-print C++ expressions. Unlike a clang Expr, a SExpr 52c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/// does not capture surface syntax, and it does not distinguish between 53c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/// C++ concepts, like pointers and references, that have no real semantic 54c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/// differences. This simplicity allows SExprs to be meaningfully compared, 55c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/// e.g. 56c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/// (x) = x 57c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/// (*this).foo = this->foo 58c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/// *&a = a 59c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/// 60c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/// Thread-safety analysis works by comparing lock expressions. Within the 61c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/// body of a function, an expression such as "x->foo->bar.mu" will resolve to 62c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/// a particular mutex object at run-time. Subsequent occurrences of the same 63c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/// expression (where "same" means syntactic equality) will refer to the same 64c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/// run-time object if three conditions hold: 65c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/// (1) Local variables in the expression, such as "x" have not changed. 66c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/// (2) Values on the heap that affect the expression have not changed. 67c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/// (3) The expression involves only pure function calls. 68c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/// 69c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/// The current implementation assumes, but does not verify, that multiple uses 70c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru/// of the same lock expression satisfies these criteria. 71c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queruclass SExpr { 72c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queruprivate: 73c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru enum ExprOp { 74c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru EOP_Nop, ///< No-op 75c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru EOP_Wildcard, ///< Matches anything. 76c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru EOP_Universal, ///< Universal lock. 77c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru EOP_This, ///< This keyword. 78c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru EOP_NVar, ///< Named variable. 79c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru EOP_LVar, ///< Local variable. 80c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru EOP_Dot, ///< Field access 81c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru EOP_Call, ///< Function call 82c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru EOP_MCall, ///< Method call 83c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru EOP_Index, ///< Array index 84c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru EOP_Unary, ///< Unary operation 85c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru EOP_Binary, ///< Binary operation 86c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru EOP_Unknown ///< Catchall for everything else 87c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru }; 88c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru 89c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru 90c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru class SExprNode { 91c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru private: 92c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru unsigned char Op; ///< Opcode of the root node 93c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru unsigned char Flags; ///< Additional opcode-specific data 94c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru unsigned short Sz; ///< Number of child nodes 95c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru const void* Data; ///< Additional opcode-specific data 96c559cd81139f97cecad1ad91a0b2e25a5936d53Jean-Baptiste Queru 97 public: 98 SExprNode(ExprOp O, unsigned F, const void* D) 99 : Op(static_cast<unsigned char>(O)), 100 Flags(static_cast<unsigned char>(F)), Sz(1), Data(D) 101 { } 102 103 unsigned size() const { return Sz; } 104 void setSize(unsigned S) { Sz = S; } 105 106 ExprOp kind() const { return static_cast<ExprOp>(Op); } 107 108 const NamedDecl* getNamedDecl() const { 109 assert(Op == EOP_NVar || Op == EOP_LVar || Op == EOP_Dot); 110 return reinterpret_cast<const NamedDecl*>(Data); 111 } 112 113 const NamedDecl* getFunctionDecl() const { 114 assert(Op == EOP_Call || Op == EOP_MCall); 115 return reinterpret_cast<const NamedDecl*>(Data); 116 } 117 118 bool isArrow() const { return Op == EOP_Dot && Flags == 1; } 119 void setArrow(bool A) { Flags = A ? 1 : 0; } 120 121 unsigned arity() const { 122 switch (Op) { 123 case EOP_Nop: return 0; 124 case EOP_Wildcard: return 0; 125 case EOP_Universal: return 0; 126 case EOP_NVar: return 0; 127 case EOP_LVar: return 0; 128 case EOP_This: return 0; 129 case EOP_Dot: return 1; 130 case EOP_Call: return Flags+1; // First arg is function. 131 case EOP_MCall: return Flags+1; // First arg is implicit obj. 132 case EOP_Index: return 2; 133 case EOP_Unary: return 1; 134 case EOP_Binary: return 2; 135 case EOP_Unknown: return Flags; 136 } 137 return 0; 138 } 139 140 bool operator==(const SExprNode& Other) const { 141 // Ignore flags and size -- they don't matter. 142 return (Op == Other.Op && 143 Data == Other.Data); 144 } 145 146 bool operator!=(const SExprNode& Other) const { 147 return !(*this == Other); 148 } 149 150 bool matches(const SExprNode& Other) const { 151 return (*this == Other) || 152 (Op == EOP_Wildcard) || 153 (Other.Op == EOP_Wildcard); 154 } 155 }; 156 157 158 /// \brief Encapsulates the lexical context of a function call. The lexical 159 /// context includes the arguments to the call, including the implicit object 160 /// argument. When an attribute containing a mutex expression is attached to 161 /// a method, the expression may refer to formal parameters of the method. 162 /// Actual arguments must be substituted for formal parameters to derive 163 /// the appropriate mutex expression in the lexical context where the function 164 /// is called. PrevCtx holds the context in which the arguments themselves 165 /// should be evaluated; multiple calling contexts can be chained together 166 /// by the lock_returned attribute. 167 struct CallingContext { 168 const NamedDecl* AttrDecl; // The decl to which the attribute is attached. 169 const Expr* SelfArg; // Implicit object argument -- e.g. 'this' 170 bool SelfArrow; // is Self referred to with -> or .? 171 unsigned NumArgs; // Number of funArgs 172 const Expr* const* FunArgs; // Function arguments 173 CallingContext* PrevCtx; // The previous context; or 0 if none. 174 175 CallingContext(const NamedDecl *D = 0, const Expr *S = 0, 176 unsigned N = 0, const Expr* const *A = 0, 177 CallingContext *P = 0) 178 : AttrDecl(D), SelfArg(S), SelfArrow(false), 179 NumArgs(N), FunArgs(A), PrevCtx(P) 180 { } 181 }; 182 183 typedef SmallVector<SExprNode, 4> NodeVector; 184 185private: 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 190private: 191 unsigned makeNop() { 192 NodeVec.push_back(SExprNode(EOP_Nop, 0, 0)); 193 return NodeVec.size()-1; 194 } 195 196 unsigned makeWildcard() { 197 NodeVec.push_back(SExprNode(EOP_Wildcard, 0, 0)); 198 return NodeVec.size()-1; 199 } 200 201 unsigned makeUniversal() { 202 NodeVec.push_back(SExprNode(EOP_Universal, 0, 0)); 203 return NodeVec.size()-1; 204 } 205 206 unsigned makeNamedVar(const NamedDecl *D) { 207 NodeVec.push_back(SExprNode(EOP_NVar, 0, D)); 208 return NodeVec.size()-1; 209 } 210 211 unsigned makeLocalVar(const NamedDecl *D) { 212 NodeVec.push_back(SExprNode(EOP_LVar, 0, D)); 213 return NodeVec.size()-1; 214 } 215 216 unsigned makeThis() { 217 NodeVec.push_back(SExprNode(EOP_This, 0, 0)); 218 return NodeVec.size()-1; 219 } 220 221 unsigned makeDot(const NamedDecl *D, bool Arrow) { 222 NodeVec.push_back(SExprNode(EOP_Dot, Arrow ? 1 : 0, D)); 223 return NodeVec.size()-1; 224 } 225 226 unsigned makeCall(unsigned NumArgs, const NamedDecl *D) { 227 NodeVec.push_back(SExprNode(EOP_Call, NumArgs, D)); 228 return NodeVec.size()-1; 229 } 230 231 // Grab the very first declaration of virtual method D 232 const CXXMethodDecl* getFirstVirtualDecl(const CXXMethodDecl *D) { 233 while (true) { 234 D = D->getCanonicalDecl(); 235 CXXMethodDecl::method_iterator I = D->begin_overridden_methods(), 236 E = D->end_overridden_methods(); 237 if (I == E) 238 return D; // Method does not override anything 239 D = *I; // FIXME: this does not work with multiple inheritance. 240 } 241 return 0; 242 } 243 244 unsigned makeMCall(unsigned NumArgs, const CXXMethodDecl *D) { 245 NodeVec.push_back(SExprNode(EOP_MCall, NumArgs, getFirstVirtualDecl(D))); 246 return NodeVec.size()-1; 247 } 248 249 unsigned makeIndex() { 250 NodeVec.push_back(SExprNode(EOP_Index, 0, 0)); 251 return NodeVec.size()-1; 252 } 253 254 unsigned makeUnary() { 255 NodeVec.push_back(SExprNode(EOP_Unary, 0, 0)); 256 return NodeVec.size()-1; 257 } 258 259 unsigned makeBinary() { 260 NodeVec.push_back(SExprNode(EOP_Binary, 0, 0)); 261 return NodeVec.size()-1; 262 } 263 264 unsigned makeUnknown(unsigned Arity) { 265 NodeVec.push_back(SExprNode(EOP_Unknown, Arity, 0)); 266 return NodeVec.size()-1; 267 } 268 269 /// Build an SExpr from the given C++ expression. 270 /// Recursive function that terminates on DeclRefExpr. 271 /// Note: this function merely creates a SExpr; it does not check to 272 /// ensure that the original expression is a valid mutex expression. 273 /// 274 /// NDeref returns the number of Derefence and AddressOf operations 275 /// preceeding the Expr; this is used to decide whether to pretty-print 276 /// SExprs with . or ->. 277 unsigned buildSExpr(const Expr *Exp, CallingContext* CallCtx, 278 int* NDeref = 0) { 279 if (!Exp) 280 return 0; 281 282 if (const DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(Exp)) { 283 const NamedDecl *ND = cast<NamedDecl>(DRE->getDecl()->getCanonicalDecl()); 284 const ParmVarDecl *PV = dyn_cast_or_null<ParmVarDecl>(ND); 285 if (PV) { 286 const FunctionDecl *FD = 287 cast<FunctionDecl>(PV->getDeclContext())->getCanonicalDecl(); 288 unsigned i = PV->getFunctionScopeIndex(); 289 290 if (CallCtx && CallCtx->FunArgs && 291 FD == CallCtx->AttrDecl->getCanonicalDecl()) { 292 // Substitute call arguments for references to function parameters 293 assert(i < CallCtx->NumArgs); 294 return buildSExpr(CallCtx->FunArgs[i], CallCtx->PrevCtx, NDeref); 295 } 296 // Map the param back to the param of the original function declaration. 297 makeNamedVar(FD->getParamDecl(i)); 298 return 1; 299 } 300 // Not a function parameter -- just store the reference. 301 makeNamedVar(ND); 302 return 1; 303 } else if (isa<CXXThisExpr>(Exp)) { 304 // Substitute parent for 'this' 305 if (CallCtx && CallCtx->SelfArg) { 306 if (!CallCtx->SelfArrow && NDeref) 307 // 'this' is a pointer, but self is not, so need to take address. 308 --(*NDeref); 309 return buildSExpr(CallCtx->SelfArg, CallCtx->PrevCtx, NDeref); 310 } 311 else { 312 makeThis(); 313 return 1; 314 } 315 } else if (const MemberExpr *ME = dyn_cast<MemberExpr>(Exp)) { 316 const NamedDecl *ND = ME->getMemberDecl(); 317 int ImplicitDeref = ME->isArrow() ? 1 : 0; 318 unsigned Root = makeDot(ND, false); 319 unsigned Sz = buildSExpr(ME->getBase(), CallCtx, &ImplicitDeref); 320 NodeVec[Root].setArrow(ImplicitDeref > 0); 321 NodeVec[Root].setSize(Sz + 1); 322 return Sz + 1; 323 } else if (const CXXMemberCallExpr *CMCE = dyn_cast<CXXMemberCallExpr>(Exp)) { 324 // When calling a function with a lock_returned attribute, replace 325 // the function call with the expression in lock_returned. 326 const CXXMethodDecl* MD = 327 cast<CXXMethodDecl>(CMCE->getMethodDecl()->getMostRecentDecl()); 328 if (LockReturnedAttr* At = MD->getAttr<LockReturnedAttr>()) { 329 CallingContext LRCallCtx(CMCE->getMethodDecl()); 330 LRCallCtx.SelfArg = CMCE->getImplicitObjectArgument(); 331 LRCallCtx.SelfArrow = 332 dyn_cast<MemberExpr>(CMCE->getCallee())->isArrow(); 333 LRCallCtx.NumArgs = CMCE->getNumArgs(); 334 LRCallCtx.FunArgs = CMCE->getArgs(); 335 LRCallCtx.PrevCtx = CallCtx; 336 return buildSExpr(At->getArg(), &LRCallCtx); 337 } 338 // Hack to treat smart pointers and iterators as pointers; 339 // ignore any method named get(). 340 if (CMCE->getMethodDecl()->getNameAsString() == "get" && 341 CMCE->getNumArgs() == 0) { 342 if (NDeref && dyn_cast<MemberExpr>(CMCE->getCallee())->isArrow()) 343 ++(*NDeref); 344 return buildSExpr(CMCE->getImplicitObjectArgument(), CallCtx, NDeref); 345 } 346 unsigned NumCallArgs = CMCE->getNumArgs(); 347 unsigned Root = makeMCall(NumCallArgs, CMCE->getMethodDecl()); 348 unsigned Sz = buildSExpr(CMCE->getImplicitObjectArgument(), CallCtx); 349 const Expr* const* CallArgs = CMCE->getArgs(); 350 for (unsigned i = 0; i < NumCallArgs; ++i) { 351 Sz += buildSExpr(CallArgs[i], CallCtx); 352 } 353 NodeVec[Root].setSize(Sz + 1); 354 return Sz + 1; 355 } else if (const CallExpr *CE = dyn_cast<CallExpr>(Exp)) { 356 const FunctionDecl* FD = 357 cast<FunctionDecl>(CE->getDirectCallee()->getMostRecentDecl()); 358 if (LockReturnedAttr* At = FD->getAttr<LockReturnedAttr>()) { 359 CallingContext LRCallCtx(CE->getDirectCallee()); 360 LRCallCtx.NumArgs = CE->getNumArgs(); 361 LRCallCtx.FunArgs = CE->getArgs(); 362 LRCallCtx.PrevCtx = CallCtx; 363 return buildSExpr(At->getArg(), &LRCallCtx); 364 } 365 // Treat smart pointers and iterators as pointers; 366 // ignore the * and -> operators. 367 if (const CXXOperatorCallExpr *OE = dyn_cast<CXXOperatorCallExpr>(CE)) { 368 OverloadedOperatorKind k = OE->getOperator(); 369 if (k == OO_Star) { 370 if (NDeref) ++(*NDeref); 371 return buildSExpr(OE->getArg(0), CallCtx, NDeref); 372 } 373 else if (k == OO_Arrow) { 374 return buildSExpr(OE->getArg(0), CallCtx, NDeref); 375 } 376 } 377 unsigned NumCallArgs = CE->getNumArgs(); 378 unsigned Root = makeCall(NumCallArgs, 0); 379 unsigned Sz = buildSExpr(CE->getCallee(), CallCtx); 380 const Expr* const* CallArgs = CE->getArgs(); 381 for (unsigned i = 0; i < NumCallArgs; ++i) { 382 Sz += buildSExpr(CallArgs[i], CallCtx); 383 } 384 NodeVec[Root].setSize(Sz+1); 385 return Sz+1; 386 } else if (const BinaryOperator *BOE = dyn_cast<BinaryOperator>(Exp)) { 387 unsigned Root = makeBinary(); 388 unsigned Sz = buildSExpr(BOE->getLHS(), CallCtx); 389 Sz += buildSExpr(BOE->getRHS(), CallCtx); 390 NodeVec[Root].setSize(Sz); 391 return Sz; 392 } else if (const UnaryOperator *UOE = dyn_cast<UnaryOperator>(Exp)) { 393 // Ignore & and * operators -- they're no-ops. 394 // However, we try to figure out whether the expression is a pointer, 395 // so we can use . and -> appropriately in error messages. 396 if (UOE->getOpcode() == UO_Deref) { 397 if (NDeref) ++(*NDeref); 398 return buildSExpr(UOE->getSubExpr(), CallCtx, NDeref); 399 } 400 if (UOE->getOpcode() == UO_AddrOf) { 401 if (DeclRefExpr* DRE = dyn_cast<DeclRefExpr>(UOE->getSubExpr())) { 402 if (DRE->getDecl()->isCXXInstanceMember()) { 403 // This is a pointer-to-member expression, e.g. &MyClass::mu_. 404 // We interpret this syntax specially, as a wildcard. 405 unsigned Root = makeDot(DRE->getDecl(), false); 406 makeWildcard(); 407 NodeVec[Root].setSize(2); 408 return 2; 409 } 410 } 411 if (NDeref) --(*NDeref); 412 return buildSExpr(UOE->getSubExpr(), CallCtx, NDeref); 413 } 414 unsigned Root = makeUnary(); 415 unsigned Sz = buildSExpr(UOE->getSubExpr(), CallCtx); 416 NodeVec[Root].setSize(Sz); 417 return Sz; 418 } else if (const ArraySubscriptExpr *ASE = 419 dyn_cast<ArraySubscriptExpr>(Exp)) { 420 unsigned Root = makeIndex(); 421 unsigned Sz = buildSExpr(ASE->getBase(), CallCtx); 422 Sz += buildSExpr(ASE->getIdx(), CallCtx); 423 NodeVec[Root].setSize(Sz); 424 return Sz; 425 } else if (const AbstractConditionalOperator *CE = 426 dyn_cast<AbstractConditionalOperator>(Exp)) { 427 unsigned Root = makeUnknown(3); 428 unsigned Sz = buildSExpr(CE->getCond(), CallCtx); 429 Sz += buildSExpr(CE->getTrueExpr(), CallCtx); 430 Sz += buildSExpr(CE->getFalseExpr(), CallCtx); 431 NodeVec[Root].setSize(Sz); 432 return Sz; 433 } else if (const ChooseExpr *CE = dyn_cast<ChooseExpr>(Exp)) { 434 unsigned Root = makeUnknown(3); 435 unsigned Sz = buildSExpr(CE->getCond(), CallCtx); 436 Sz += buildSExpr(CE->getLHS(), CallCtx); 437 Sz += buildSExpr(CE->getRHS(), CallCtx); 438 NodeVec[Root].setSize(Sz); 439 return Sz; 440 } else if (const CastExpr *CE = dyn_cast<CastExpr>(Exp)) { 441 return buildSExpr(CE->getSubExpr(), CallCtx, NDeref); 442 } else if (const ParenExpr *PE = dyn_cast<ParenExpr>(Exp)) { 443 return buildSExpr(PE->getSubExpr(), CallCtx, NDeref); 444 } else if (const ExprWithCleanups *EWC = dyn_cast<ExprWithCleanups>(Exp)) { 445 return buildSExpr(EWC->getSubExpr(), CallCtx, NDeref); 446 } else if (const CXXBindTemporaryExpr *E = dyn_cast<CXXBindTemporaryExpr>(Exp)) { 447 return buildSExpr(E->getSubExpr(), CallCtx, NDeref); 448 } else if (isa<CharacterLiteral>(Exp) || 449 isa<CXXNullPtrLiteralExpr>(Exp) || 450 isa<GNUNullExpr>(Exp) || 451 isa<CXXBoolLiteralExpr>(Exp) || 452 isa<FloatingLiteral>(Exp) || 453 isa<ImaginaryLiteral>(Exp) || 454 isa<IntegerLiteral>(Exp) || 455 isa<StringLiteral>(Exp) || 456 isa<ObjCStringLiteral>(Exp)) { 457 makeNop(); 458 return 1; // FIXME: Ignore literals for now 459 } else { 460 makeNop(); 461 return 1; // Ignore. FIXME: mark as invalid expression? 462 } 463 } 464 465 /// \brief Construct a SExpr from an expression. 466 /// \param MutexExp The original mutex expression within an attribute 467 /// \param DeclExp An expression involving the Decl on which the attribute 468 /// occurs. 469 /// \param D The declaration to which the lock/unlock attribute is attached. 470 void buildSExprFromExpr(const Expr *MutexExp, const Expr *DeclExp, 471 const NamedDecl *D, VarDecl *SelfDecl = 0) { 472 CallingContext CallCtx(D); 473 474 if (MutexExp) { 475 if (const StringLiteral* SLit = dyn_cast<StringLiteral>(MutexExp)) { 476 if (SLit->getString() == StringRef("*")) 477 // The "*" expr is a universal lock, which essentially turns off 478 // checks until it is removed from the lockset. 479 makeUniversal(); 480 else 481 // Ignore other string literals for now. 482 makeNop(); 483 return; 484 } 485 } 486 487 // If we are processing a raw attribute expression, with no substitutions. 488 if (DeclExp == 0) { 489 buildSExpr(MutexExp, 0); 490 return; 491 } 492 493 // Examine DeclExp to find SelfArg and FunArgs, which are used to substitute 494 // for formal parameters when we call buildMutexID later. 495 if (const MemberExpr *ME = dyn_cast<MemberExpr>(DeclExp)) { 496 CallCtx.SelfArg = ME->getBase(); 497 CallCtx.SelfArrow = ME->isArrow(); 498 } else if (const CXXMemberCallExpr *CE = 499 dyn_cast<CXXMemberCallExpr>(DeclExp)) { 500 CallCtx.SelfArg = CE->getImplicitObjectArgument(); 501 CallCtx.SelfArrow = dyn_cast<MemberExpr>(CE->getCallee())->isArrow(); 502 CallCtx.NumArgs = CE->getNumArgs(); 503 CallCtx.FunArgs = CE->getArgs(); 504 } else if (const CallExpr *CE = 505 dyn_cast<CallExpr>(DeclExp)) { 506 CallCtx.NumArgs = CE->getNumArgs(); 507 CallCtx.FunArgs = CE->getArgs(); 508 } else if (const CXXConstructExpr *CE = 509 dyn_cast<CXXConstructExpr>(DeclExp)) { 510 CallCtx.SelfArg = 0; // Will be set below 511 CallCtx.NumArgs = CE->getNumArgs(); 512 CallCtx.FunArgs = CE->getArgs(); 513 } else if (D && isa<CXXDestructorDecl>(D)) { 514 // There's no such thing as a "destructor call" in the AST. 515 CallCtx.SelfArg = DeclExp; 516 } 517 518 // Hack to handle constructors, where self cannot be recovered from 519 // the expression. 520 if (SelfDecl && !CallCtx.SelfArg) { 521 DeclRefExpr SelfDRE(SelfDecl, false, SelfDecl->getType(), VK_LValue, 522 SelfDecl->getLocation()); 523 CallCtx.SelfArg = &SelfDRE; 524 525 // If the attribute has no arguments, then assume the argument is "this". 526 if (MutexExp == 0) 527 buildSExpr(CallCtx.SelfArg, 0); 528 else // For most attributes. 529 buildSExpr(MutexExp, &CallCtx); 530 return; 531 } 532 533 // If the attribute has no arguments, then assume the argument is "this". 534 if (MutexExp == 0) 535 buildSExpr(CallCtx.SelfArg, 0); 536 else // For most attributes. 537 buildSExpr(MutexExp, &CallCtx); 538 } 539 540 /// \brief Get index of next sibling of node i. 541 unsigned getNextSibling(unsigned i) const { 542 return i + NodeVec[i].size(); 543 } 544 545public: 546 explicit SExpr(clang::Decl::EmptyShell e) { NodeVec.clear(); } 547 548 /// \param MutexExp The original mutex expression within an attribute 549 /// \param DeclExp An expression involving the Decl on which the attribute 550 /// occurs. 551 /// \param D The declaration to which the lock/unlock attribute is attached. 552 /// Caller must check isValid() after construction. 553 SExpr(const Expr* MutexExp, const Expr *DeclExp, const NamedDecl* D, 554 VarDecl *SelfDecl=0) { 555 buildSExprFromExpr(MutexExp, DeclExp, D, SelfDecl); 556 } 557 558 /// Return true if this is a valid decl sequence. 559 /// Caller must call this by hand after construction to handle errors. 560 bool isValid() const { 561 return !NodeVec.empty(); 562 } 563 564 bool shouldIgnore() const { 565 // Nop is a mutex that we have decided to deliberately ignore. 566 assert(NodeVec.size() > 0 && "Invalid Mutex"); 567 return NodeVec[0].kind() == EOP_Nop; 568 } 569 570 bool isUniversal() const { 571 assert(NodeVec.size() > 0 && "Invalid Mutex"); 572 return NodeVec[0].kind() == EOP_Universal; 573 } 574 575 /// Issue a warning about an invalid lock expression 576 static void warnInvalidLock(ThreadSafetyHandler &Handler, 577 const Expr *MutexExp, 578 const Expr *DeclExp, const NamedDecl* D) { 579 SourceLocation Loc; 580 if (DeclExp) 581 Loc = DeclExp->getExprLoc(); 582 583 // FIXME: add a note about the attribute location in MutexExp or D 584 if (Loc.isValid()) 585 Handler.handleInvalidLockExp(Loc); 586 } 587 588 bool operator==(const SExpr &other) const { 589 return NodeVec == other.NodeVec; 590 } 591 592 bool operator!=(const SExpr &other) const { 593 return !(*this == other); 594 } 595 596 bool matches(const SExpr &Other, unsigned i = 0, unsigned j = 0) const { 597 if (NodeVec[i].matches(Other.NodeVec[j])) { 598 unsigned ni = NodeVec[i].arity(); 599 unsigned nj = Other.NodeVec[j].arity(); 600 unsigned n = (ni < nj) ? ni : nj; 601 bool Result = true; 602 unsigned ci = i+1; // first child of i 603 unsigned cj = j+1; // first child of j 604 for (unsigned k = 0; k < n; 605 ++k, ci=getNextSibling(ci), cj = Other.getNextSibling(cj)) { 606 Result = Result && matches(Other, ci, cj); 607 } 608 return Result; 609 } 610 return false; 611 } 612 613 // A partial match between a.mu and b.mu returns true a and b have the same 614 // type (and thus mu refers to the same mutex declaration), regardless of 615 // whether a and b are different objects or not. 616 bool partiallyMatches(const SExpr &Other) const { 617 if (NodeVec[0].kind() == EOP_Dot) 618 return NodeVec[0].matches(Other.NodeVec[0]); 619 return false; 620 } 621 622 /// \brief Pretty print a lock expression for use in error messages. 623 std::string toString(unsigned i = 0) const { 624 assert(isValid()); 625 if (i >= NodeVec.size()) 626 return ""; 627 628 const SExprNode* N = &NodeVec[i]; 629 switch (N->kind()) { 630 case EOP_Nop: 631 return "_"; 632 case EOP_Wildcard: 633 return "(?)"; 634 case EOP_Universal: 635 return "*"; 636 case EOP_This: 637 return "this"; 638 case EOP_NVar: 639 case EOP_LVar: { 640 return N->getNamedDecl()->getNameAsString(); 641 } 642 case EOP_Dot: { 643 if (NodeVec[i+1].kind() == EOP_Wildcard) { 644 std::string S = "&"; 645 S += N->getNamedDecl()->getQualifiedNameAsString(); 646 return S; 647 } 648 std::string FieldName = N->getNamedDecl()->getNameAsString(); 649 if (NodeVec[i+1].kind() == EOP_This) 650 return FieldName; 651 652 std::string S = toString(i+1); 653 if (N->isArrow()) 654 return S + "->" + FieldName; 655 else 656 return S + "." + FieldName; 657 } 658 case EOP_Call: { 659 std::string S = toString(i+1) + "("; 660 unsigned NumArgs = N->arity()-1; 661 unsigned ci = getNextSibling(i+1); 662 for (unsigned k=0; k<NumArgs; ++k, ci = getNextSibling(ci)) { 663 S += toString(ci); 664 if (k+1 < NumArgs) S += ","; 665 } 666 S += ")"; 667 return S; 668 } 669 case EOP_MCall: { 670 std::string S = ""; 671 if (NodeVec[i+1].kind() != EOP_This) 672 S = toString(i+1) + "."; 673 if (const NamedDecl *D = N->getFunctionDecl()) 674 S += D->getNameAsString() + "("; 675 else 676 S += "#("; 677 unsigned NumArgs = N->arity()-1; 678 unsigned ci = getNextSibling(i+1); 679 for (unsigned k=0; k<NumArgs; ++k, ci = getNextSibling(ci)) { 680 S += toString(ci); 681 if (k+1 < NumArgs) S += ","; 682 } 683 S += ")"; 684 return S; 685 } 686 case EOP_Index: { 687 std::string S1 = toString(i+1); 688 std::string S2 = toString(i+1 + NodeVec[i+1].size()); 689 return S1 + "[" + S2 + "]"; 690 } 691 case EOP_Unary: { 692 std::string S = toString(i+1); 693 return "#" + S; 694 } 695 case EOP_Binary: { 696 std::string S1 = toString(i+1); 697 std::string S2 = toString(i+1 + NodeVec[i+1].size()); 698 return "(" + S1 + "#" + S2 + ")"; 699 } 700 case EOP_Unknown: { 701 unsigned NumChildren = N->arity(); 702 if (NumChildren == 0) 703 return "(...)"; 704 std::string S = "("; 705 unsigned ci = i+1; 706 for (unsigned j = 0; j < NumChildren; ++j, ci = getNextSibling(ci)) { 707 S += toString(ci); 708 if (j+1 < NumChildren) S += "#"; 709 } 710 S += ")"; 711 return S; 712 } 713 } 714 return ""; 715 } 716}; 717 718 719 720/// \brief A short list of SExprs 721class MutexIDList : public SmallVector<SExpr, 3> { 722public: 723 /// \brief Return true if the list contains the specified SExpr 724 /// Performs a linear search, because these lists are almost always very small. 725 bool contains(const SExpr& M) { 726 for (iterator I=begin(),E=end(); I != E; ++I) 727 if ((*I) == M) return true; 728 return false; 729 } 730 731 /// \brief Push M onto list, bud discard duplicates 732 void push_back_nodup(const SExpr& M) { 733 if (!contains(M)) push_back(M); 734 } 735}; 736 737 738 739/// \brief This is a helper class that stores info about the most recent 740/// accquire of a Lock. 741/// 742/// The main body of the analysis maps MutexIDs to LockDatas. 743struct LockData { 744 SourceLocation AcquireLoc; 745 746 /// \brief LKind stores whether a lock is held shared or exclusively. 747 /// Note that this analysis does not currently support either re-entrant 748 /// locking or lock "upgrading" and "downgrading" between exclusive and 749 /// shared. 750 /// 751 /// FIXME: add support for re-entrant locking and lock up/downgrading 752 LockKind LKind; 753 bool Managed; // for ScopedLockable objects 754 SExpr UnderlyingMutex; // for ScopedLockable objects 755 756 LockData(SourceLocation AcquireLoc, LockKind LKind, bool M = false) 757 : AcquireLoc(AcquireLoc), LKind(LKind), Managed(M), 758 UnderlyingMutex(Decl::EmptyShell()) 759 {} 760 761 LockData(SourceLocation AcquireLoc, LockKind LKind, const SExpr &Mu) 762 : AcquireLoc(AcquireLoc), LKind(LKind), Managed(false), 763 UnderlyingMutex(Mu) 764 {} 765 766 bool operator==(const LockData &other) const { 767 return AcquireLoc == other.AcquireLoc && LKind == other.LKind; 768 } 769 770 bool operator!=(const LockData &other) const { 771 return !(*this == other); 772 } 773 774 void Profile(llvm::FoldingSetNodeID &ID) const { 775 ID.AddInteger(AcquireLoc.getRawEncoding()); 776 ID.AddInteger(LKind); 777 } 778 779 bool isAtLeast(LockKind LK) { 780 return (LK == LK_Shared) || (LKind == LK_Exclusive); 781 } 782}; 783 784 785/// \brief A FactEntry stores a single fact that is known at a particular point 786/// in the program execution. Currently, this is information regarding a lock 787/// that is held at that point. 788struct FactEntry { 789 SExpr MutID; 790 LockData LDat; 791 792 FactEntry(const SExpr& M, const LockData& L) 793 : MutID(M), LDat(L) 794 { } 795}; 796 797 798typedef unsigned short FactID; 799 800/// \brief FactManager manages the memory for all facts that are created during 801/// the analysis of a single routine. 802class FactManager { 803private: 804 std::vector<FactEntry> Facts; 805 806public: 807 FactID newLock(const SExpr& M, const LockData& L) { 808 Facts.push_back(FactEntry(M,L)); 809 return static_cast<unsigned short>(Facts.size() - 1); 810 } 811 812 const FactEntry& operator[](FactID F) const { return Facts[F]; } 813 FactEntry& operator[](FactID F) { return Facts[F]; } 814}; 815 816 817/// \brief A FactSet is the set of facts that are known to be true at a 818/// particular program point. FactSets must be small, because they are 819/// frequently copied, and are thus implemented as a set of indices into a 820/// table maintained by a FactManager. A typical FactSet only holds 1 or 2 821/// locks, so we can get away with doing a linear search for lookup. Note 822/// that a hashtable or map is inappropriate in this case, because lookups 823/// may involve partial pattern matches, rather than exact matches. 824class FactSet { 825private: 826 typedef SmallVector<FactID, 4> FactVec; 827 828 FactVec FactIDs; 829 830public: 831 typedef FactVec::iterator iterator; 832 typedef FactVec::const_iterator const_iterator; 833 834 iterator begin() { return FactIDs.begin(); } 835 const_iterator begin() const { return FactIDs.begin(); } 836 837 iterator end() { return FactIDs.end(); } 838 const_iterator end() const { return FactIDs.end(); } 839 840 bool isEmpty() const { return FactIDs.size() == 0; } 841 842 FactID addLock(FactManager& FM, const SExpr& M, const LockData& L) { 843 FactID F = FM.newLock(M, L); 844 FactIDs.push_back(F); 845 return F; 846 } 847 848 bool removeLock(FactManager& FM, const SExpr& M) { 849 unsigned n = FactIDs.size(); 850 if (n == 0) 851 return false; 852 853 for (unsigned i = 0; i < n-1; ++i) { 854 if (FM[FactIDs[i]].MutID.matches(M)) { 855 FactIDs[i] = FactIDs[n-1]; 856 FactIDs.pop_back(); 857 return true; 858 } 859 } 860 if (FM[FactIDs[n-1]].MutID.matches(M)) { 861 FactIDs.pop_back(); 862 return true; 863 } 864 return false; 865 } 866 867 LockData* findLock(FactManager &FM, const SExpr &M) const { 868 for (const_iterator I = begin(), E = end(); I != E; ++I) { 869 const SExpr &Exp = FM[*I].MutID; 870 if (Exp.matches(M)) 871 return &FM[*I].LDat; 872 } 873 return 0; 874 } 875 876 LockData* findLockUniv(FactManager &FM, const SExpr &M) const { 877 for (const_iterator I = begin(), E = end(); I != E; ++I) { 878 const SExpr &Exp = FM[*I].MutID; 879 if (Exp.matches(M) || Exp.isUniversal()) 880 return &FM[*I].LDat; 881 } 882 return 0; 883 } 884 885 FactEntry* findPartialMatch(FactManager &FM, const SExpr &M) const { 886 for (const_iterator I=begin(), E=end(); I != E; ++I) { 887 const SExpr& Exp = FM[*I].MutID; 888 if (Exp.partiallyMatches(M)) return &FM[*I]; 889 } 890 return 0; 891 } 892}; 893 894 895 896/// A Lockset maps each SExpr (defined above) to information about how it has 897/// been locked. 898typedef llvm::ImmutableMap<SExpr, LockData> Lockset; 899typedef llvm::ImmutableMap<const NamedDecl*, unsigned> LocalVarContext; 900 901class LocalVariableMap; 902 903/// A side (entry or exit) of a CFG node. 904enum CFGBlockSide { CBS_Entry, CBS_Exit }; 905 906/// CFGBlockInfo is a struct which contains all the information that is 907/// maintained for each block in the CFG. See LocalVariableMap for more 908/// information about the contexts. 909struct CFGBlockInfo { 910 FactSet EntrySet; // Lockset held at entry to block 911 FactSet ExitSet; // Lockset held at exit from block 912 LocalVarContext EntryContext; // Context held at entry to block 913 LocalVarContext ExitContext; // Context held at exit from block 914 SourceLocation EntryLoc; // Location of first statement in block 915 SourceLocation ExitLoc; // Location of last statement in block. 916 unsigned EntryIndex; // Used to replay contexts later 917 bool Reachable; // Is this block reachable? 918 919 const FactSet &getSet(CFGBlockSide Side) const { 920 return Side == CBS_Entry ? EntrySet : ExitSet; 921 } 922 SourceLocation getLocation(CFGBlockSide Side) const { 923 return Side == CBS_Entry ? EntryLoc : ExitLoc; 924 } 925 926private: 927 CFGBlockInfo(LocalVarContext EmptyCtx) 928 : EntryContext(EmptyCtx), ExitContext(EmptyCtx), Reachable(false) 929 { } 930 931public: 932 static CFGBlockInfo getEmptyBlockInfo(LocalVariableMap &M); 933}; 934 935 936 937// A LocalVariableMap maintains a map from local variables to their currently 938// valid definitions. It provides SSA-like functionality when traversing the 939// CFG. Like SSA, each definition or assignment to a variable is assigned a 940// unique name (an integer), which acts as the SSA name for that definition. 941// The total set of names is shared among all CFG basic blocks. 942// Unlike SSA, we do not rewrite expressions to replace local variables declrefs 943// with their SSA-names. Instead, we compute a Context for each point in the 944// code, which maps local variables to the appropriate SSA-name. This map 945// changes with each assignment. 946// 947// The map is computed in a single pass over the CFG. Subsequent analyses can 948// then query the map to find the appropriate Context for a statement, and use 949// that Context to look up the definitions of variables. 950class LocalVariableMap { 951public: 952 typedef LocalVarContext Context; 953 954 /// A VarDefinition consists of an expression, representing the value of the 955 /// variable, along with the context in which that expression should be 956 /// interpreted. A reference VarDefinition does not itself contain this 957 /// information, but instead contains a pointer to a previous VarDefinition. 958 struct VarDefinition { 959 public: 960 friend class LocalVariableMap; 961 962 const NamedDecl *Dec; // The original declaration for this variable. 963 const Expr *Exp; // The expression for this variable, OR 964 unsigned Ref; // Reference to another VarDefinition 965 Context Ctx; // The map with which Exp should be interpreted. 966 967 bool isReference() { return !Exp; } 968 969 private: 970 // Create ordinary variable definition 971 VarDefinition(const NamedDecl *D, const Expr *E, Context C) 972 : Dec(D), Exp(E), Ref(0), Ctx(C) 973 { } 974 975 // Create reference to previous definition 976 VarDefinition(const NamedDecl *D, unsigned R, Context C) 977 : Dec(D), Exp(0), Ref(R), Ctx(C) 978 { } 979 }; 980 981private: 982 Context::Factory ContextFactory; 983 std::vector<VarDefinition> VarDefinitions; 984 std::vector<unsigned> CtxIndices; 985 std::vector<std::pair<Stmt*, Context> > SavedContexts; 986 987public: 988 LocalVariableMap() { 989 // index 0 is a placeholder for undefined variables (aka phi-nodes). 990 VarDefinitions.push_back(VarDefinition(0, 0u, getEmptyContext())); 991 } 992 993 /// Look up a definition, within the given context. 994 const VarDefinition* lookup(const NamedDecl *D, Context Ctx) { 995 const unsigned *i = Ctx.lookup(D); 996 if (!i) 997 return 0; 998 assert(*i < VarDefinitions.size()); 999 return &VarDefinitions[*i]; 1000 } 1001 1002 /// Look up the definition for D within the given context. Returns 1003 /// NULL if the expression is not statically known. If successful, also 1004 /// modifies Ctx to hold the context of the return Expr. 1005 const Expr* lookupExpr(const NamedDecl *D, Context &Ctx) { 1006 const unsigned *P = Ctx.lookup(D); 1007 if (!P) 1008 return 0; 1009 1010 unsigned i = *P; 1011 while (i > 0) { 1012 if (VarDefinitions[i].Exp) { 1013 Ctx = VarDefinitions[i].Ctx; 1014 return VarDefinitions[i].Exp; 1015 } 1016 i = VarDefinitions[i].Ref; 1017 } 1018 return 0; 1019 } 1020 1021 Context getEmptyContext() { return ContextFactory.getEmptyMap(); } 1022 1023 /// Return the next context after processing S. This function is used by 1024 /// clients of the class to get the appropriate context when traversing the 1025 /// CFG. It must be called for every assignment or DeclStmt. 1026 Context getNextContext(unsigned &CtxIndex, Stmt *S, Context C) { 1027 if (SavedContexts[CtxIndex+1].first == S) { 1028 CtxIndex++; 1029 Context Result = SavedContexts[CtxIndex].second; 1030 return Result; 1031 } 1032 return C; 1033 } 1034 1035 void dumpVarDefinitionName(unsigned i) { 1036 if (i == 0) { 1037 llvm::errs() << "Undefined"; 1038 return; 1039 } 1040 const NamedDecl *Dec = VarDefinitions[i].Dec; 1041 if (!Dec) { 1042 llvm::errs() << "<<NULL>>"; 1043 return; 1044 } 1045 Dec->printName(llvm::errs()); 1046 llvm::errs() << "." << i << " " << ((const void*) Dec); 1047 } 1048 1049 /// Dumps an ASCII representation of the variable map to llvm::errs() 1050 void dump() { 1051 for (unsigned i = 1, e = VarDefinitions.size(); i < e; ++i) { 1052 const Expr *Exp = VarDefinitions[i].Exp; 1053 unsigned Ref = VarDefinitions[i].Ref; 1054 1055 dumpVarDefinitionName(i); 1056 llvm::errs() << " = "; 1057 if (Exp) Exp->dump(); 1058 else { 1059 dumpVarDefinitionName(Ref); 1060 llvm::errs() << "\n"; 1061 } 1062 } 1063 } 1064 1065 /// Dumps an ASCII representation of a Context to llvm::errs() 1066 void dumpContext(Context C) { 1067 for (Context::iterator I = C.begin(), E = C.end(); I != E; ++I) { 1068 const NamedDecl *D = I.getKey(); 1069 D->printName(llvm::errs()); 1070 const unsigned *i = C.lookup(D); 1071 llvm::errs() << " -> "; 1072 dumpVarDefinitionName(*i); 1073 llvm::errs() << "\n"; 1074 } 1075 } 1076 1077 /// Builds the variable map. 1078 void traverseCFG(CFG *CFGraph, PostOrderCFGView *SortedGraph, 1079 std::vector<CFGBlockInfo> &BlockInfo); 1080 1081protected: 1082 // Get the current context index 1083 unsigned getContextIndex() { return SavedContexts.size()-1; } 1084 1085 // Save the current context for later replay 1086 void saveContext(Stmt *S, Context C) { 1087 SavedContexts.push_back(std::make_pair(S,C)); 1088 } 1089 1090 // Adds a new definition to the given context, and returns a new context. 1091 // This method should be called when declaring a new variable. 1092 Context addDefinition(const NamedDecl *D, Expr *Exp, Context Ctx) { 1093 assert(!Ctx.contains(D)); 1094 unsigned newID = VarDefinitions.size(); 1095 Context NewCtx = ContextFactory.add(Ctx, D, newID); 1096 VarDefinitions.push_back(VarDefinition(D, Exp, Ctx)); 1097 return NewCtx; 1098 } 1099 1100 // Add a new reference to an existing definition. 1101 Context addReference(const NamedDecl *D, unsigned i, Context Ctx) { 1102 unsigned newID = VarDefinitions.size(); 1103 Context NewCtx = ContextFactory.add(Ctx, D, newID); 1104 VarDefinitions.push_back(VarDefinition(D, i, Ctx)); 1105 return NewCtx; 1106 } 1107 1108 // Updates a definition only if that definition is already in the map. 1109 // This method should be called when assigning to an existing variable. 1110 Context updateDefinition(const NamedDecl *D, Expr *Exp, Context Ctx) { 1111 if (Ctx.contains(D)) { 1112 unsigned newID = VarDefinitions.size(); 1113 Context NewCtx = ContextFactory.remove(Ctx, D); 1114 NewCtx = ContextFactory.add(NewCtx, D, newID); 1115 VarDefinitions.push_back(VarDefinition(D, Exp, Ctx)); 1116 return NewCtx; 1117 } 1118 return Ctx; 1119 } 1120 1121 // Removes a definition from the context, but keeps the variable name 1122 // as a valid variable. The index 0 is a placeholder for cleared definitions. 1123 Context clearDefinition(const NamedDecl *D, Context Ctx) { 1124 Context NewCtx = Ctx; 1125 if (NewCtx.contains(D)) { 1126 NewCtx = ContextFactory.remove(NewCtx, D); 1127 NewCtx = ContextFactory.add(NewCtx, D, 0); 1128 } 1129 return NewCtx; 1130 } 1131 1132 // Remove a definition entirely frmo the context. 1133 Context removeDefinition(const NamedDecl *D, Context Ctx) { 1134 Context NewCtx = Ctx; 1135 if (NewCtx.contains(D)) { 1136 NewCtx = ContextFactory.remove(NewCtx, D); 1137 } 1138 return NewCtx; 1139 } 1140 1141 Context intersectContexts(Context C1, Context C2); 1142 Context createReferenceContext(Context C); 1143 void intersectBackEdge(Context C1, Context C2); 1144 1145 friend class VarMapBuilder; 1146}; 1147 1148 1149// This has to be defined after LocalVariableMap. 1150CFGBlockInfo CFGBlockInfo::getEmptyBlockInfo(LocalVariableMap &M) { 1151 return CFGBlockInfo(M.getEmptyContext()); 1152} 1153 1154 1155/// Visitor which builds a LocalVariableMap 1156class VarMapBuilder : public StmtVisitor<VarMapBuilder> { 1157public: 1158 LocalVariableMap* VMap; 1159 LocalVariableMap::Context Ctx; 1160 1161 VarMapBuilder(LocalVariableMap *VM, LocalVariableMap::Context C) 1162 : VMap(VM), Ctx(C) {} 1163 1164 void VisitDeclStmt(DeclStmt *S); 1165 void VisitBinaryOperator(BinaryOperator *BO); 1166}; 1167 1168 1169// Add new local variables to the variable map 1170void VarMapBuilder::VisitDeclStmt(DeclStmt *S) { 1171 bool modifiedCtx = false; 1172 DeclGroupRef DGrp = S->getDeclGroup(); 1173 for (DeclGroupRef::iterator I = DGrp.begin(), E = DGrp.end(); I != E; ++I) { 1174 if (VarDecl *VD = dyn_cast_or_null<VarDecl>(*I)) { 1175 Expr *E = VD->getInit(); 1176 1177 // Add local variables with trivial type to the variable map 1178 QualType T = VD->getType(); 1179 if (T.isTrivialType(VD->getASTContext())) { 1180 Ctx = VMap->addDefinition(VD, E, Ctx); 1181 modifiedCtx = true; 1182 } 1183 } 1184 } 1185 if (modifiedCtx) 1186 VMap->saveContext(S, Ctx); 1187} 1188 1189// Update local variable definitions in variable map 1190void VarMapBuilder::VisitBinaryOperator(BinaryOperator *BO) { 1191 if (!BO->isAssignmentOp()) 1192 return; 1193 1194 Expr *LHSExp = BO->getLHS()->IgnoreParenCasts(); 1195 1196 // Update the variable map and current context. 1197 if (DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(LHSExp)) { 1198 ValueDecl *VDec = DRE->getDecl(); 1199 if (Ctx.lookup(VDec)) { 1200 if (BO->getOpcode() == BO_Assign) 1201 Ctx = VMap->updateDefinition(VDec, BO->getRHS(), Ctx); 1202 else 1203 // FIXME -- handle compound assignment operators 1204 Ctx = VMap->clearDefinition(VDec, Ctx); 1205 VMap->saveContext(BO, Ctx); 1206 } 1207 } 1208} 1209 1210 1211// Computes the intersection of two contexts. The intersection is the 1212// set of variables which have the same definition in both contexts; 1213// variables with different definitions are discarded. 1214LocalVariableMap::Context 1215LocalVariableMap::intersectContexts(Context C1, Context C2) { 1216 Context Result = C1; 1217 for (Context::iterator I = C1.begin(), E = C1.end(); I != E; ++I) { 1218 const NamedDecl *Dec = I.getKey(); 1219 unsigned i1 = I.getData(); 1220 const unsigned *i2 = C2.lookup(Dec); 1221 if (!i2) // variable doesn't exist on second path 1222 Result = removeDefinition(Dec, Result); 1223 else if (*i2 != i1) // variable exists, but has different definition 1224 Result = clearDefinition(Dec, Result); 1225 } 1226 return Result; 1227} 1228 1229// For every variable in C, create a new variable that refers to the 1230// definition in C. Return a new context that contains these new variables. 1231// (We use this for a naive implementation of SSA on loop back-edges.) 1232LocalVariableMap::Context LocalVariableMap::createReferenceContext(Context C) { 1233 Context Result = getEmptyContext(); 1234 for (Context::iterator I = C.begin(), E = C.end(); I != E; ++I) { 1235 const NamedDecl *Dec = I.getKey(); 1236 unsigned i = I.getData(); 1237 Result = addReference(Dec, i, Result); 1238 } 1239 return Result; 1240} 1241 1242// This routine also takes the intersection of C1 and C2, but it does so by 1243// altering the VarDefinitions. C1 must be the result of an earlier call to 1244// createReferenceContext. 1245void LocalVariableMap::intersectBackEdge(Context C1, Context C2) { 1246 for (Context::iterator I = C1.begin(), E = C1.end(); I != E; ++I) { 1247 const NamedDecl *Dec = I.getKey(); 1248 unsigned i1 = I.getData(); 1249 VarDefinition *VDef = &VarDefinitions[i1]; 1250 assert(VDef->isReference()); 1251 1252 const unsigned *i2 = C2.lookup(Dec); 1253 if (!i2 || (*i2 != i1)) 1254 VDef->Ref = 0; // Mark this variable as undefined 1255 } 1256} 1257 1258 1259// Traverse the CFG in topological order, so all predecessors of a block 1260// (excluding back-edges) are visited before the block itself. At 1261// each point in the code, we calculate a Context, which holds the set of 1262// variable definitions which are visible at that point in execution. 1263// Visible variables are mapped to their definitions using an array that 1264// contains all definitions. 1265// 1266// At join points in the CFG, the set is computed as the intersection of 1267// the incoming sets along each edge, E.g. 1268// 1269// { Context | VarDefinitions } 1270// int x = 0; { x -> x1 | x1 = 0 } 1271// int y = 0; { x -> x1, y -> y1 | y1 = 0, x1 = 0 } 1272// if (b) x = 1; { x -> x2, y -> y1 | x2 = 1, y1 = 0, ... } 1273// else x = 2; { x -> x3, y -> y1 | x3 = 2, x2 = 1, ... } 1274// ... { y -> y1 (x is unknown) | x3 = 2, x2 = 1, ... } 1275// 1276// This is essentially a simpler and more naive version of the standard SSA 1277// algorithm. Those definitions that remain in the intersection are from blocks 1278// that strictly dominate the current block. We do not bother to insert proper 1279// phi nodes, because they are not used in our analysis; instead, wherever 1280// a phi node would be required, we simply remove that definition from the 1281// context (E.g. x above). 1282// 1283// The initial traversal does not capture back-edges, so those need to be 1284// handled on a separate pass. Whenever the first pass encounters an 1285// incoming back edge, it duplicates the context, creating new definitions 1286// that refer back to the originals. (These correspond to places where SSA 1287// might have to insert a phi node.) On the second pass, these definitions are 1288// set to NULL if the variable has changed on the back-edge (i.e. a phi 1289// node was actually required.) E.g. 1290// 1291// { Context | VarDefinitions } 1292// int x = 0, y = 0; { x -> x1, y -> y1 | y1 = 0, x1 = 0 } 1293// while (b) { x -> x2, y -> y1 | [1st:] x2=x1; [2nd:] x2=NULL; } 1294// x = x+1; { x -> x3, y -> y1 | x3 = x2 + 1, ... } 1295// ... { y -> y1 | x3 = 2, x2 = 1, ... } 1296// 1297void LocalVariableMap::traverseCFG(CFG *CFGraph, 1298 PostOrderCFGView *SortedGraph, 1299 std::vector<CFGBlockInfo> &BlockInfo) { 1300 PostOrderCFGView::CFGBlockSet VisitedBlocks(CFGraph); 1301 1302 CtxIndices.resize(CFGraph->getNumBlockIDs()); 1303 1304 for (PostOrderCFGView::iterator I = SortedGraph->begin(), 1305 E = SortedGraph->end(); I!= E; ++I) { 1306 const CFGBlock *CurrBlock = *I; 1307 int CurrBlockID = CurrBlock->getBlockID(); 1308 CFGBlockInfo *CurrBlockInfo = &BlockInfo[CurrBlockID]; 1309 1310 VisitedBlocks.insert(CurrBlock); 1311 1312 // Calculate the entry context for the current block 1313 bool HasBackEdges = false; 1314 bool CtxInit = true; 1315 for (CFGBlock::const_pred_iterator PI = CurrBlock->pred_begin(), 1316 PE = CurrBlock->pred_end(); PI != PE; ++PI) { 1317 // if *PI -> CurrBlock is a back edge, so skip it 1318 if (*PI == 0 || !VisitedBlocks.alreadySet(*PI)) { 1319 HasBackEdges = true; 1320 continue; 1321 } 1322 1323 int PrevBlockID = (*PI)->getBlockID(); 1324 CFGBlockInfo *PrevBlockInfo = &BlockInfo[PrevBlockID]; 1325 1326 if (CtxInit) { 1327 CurrBlockInfo->EntryContext = PrevBlockInfo->ExitContext; 1328 CtxInit = false; 1329 } 1330 else { 1331 CurrBlockInfo->EntryContext = 1332 intersectContexts(CurrBlockInfo->EntryContext, 1333 PrevBlockInfo->ExitContext); 1334 } 1335 } 1336 1337 // Duplicate the context if we have back-edges, so we can call 1338 // intersectBackEdges later. 1339 if (HasBackEdges) 1340 CurrBlockInfo->EntryContext = 1341 createReferenceContext(CurrBlockInfo->EntryContext); 1342 1343 // Create a starting context index for the current block 1344 saveContext(0, CurrBlockInfo->EntryContext); 1345 CurrBlockInfo->EntryIndex = getContextIndex(); 1346 1347 // Visit all the statements in the basic block. 1348 VarMapBuilder VMapBuilder(this, CurrBlockInfo->EntryContext); 1349 for (CFGBlock::const_iterator BI = CurrBlock->begin(), 1350 BE = CurrBlock->end(); BI != BE; ++BI) { 1351 switch (BI->getKind()) { 1352 case CFGElement::Statement: { 1353 CFGStmt CS = BI->castAs<CFGStmt>(); 1354 VMapBuilder.Visit(const_cast<Stmt*>(CS.getStmt())); 1355 break; 1356 } 1357 default: 1358 break; 1359 } 1360 } 1361 CurrBlockInfo->ExitContext = VMapBuilder.Ctx; 1362 1363 // Mark variables on back edges as "unknown" if they've been changed. 1364 for (CFGBlock::const_succ_iterator SI = CurrBlock->succ_begin(), 1365 SE = CurrBlock->succ_end(); SI != SE; ++SI) { 1366 // if CurrBlock -> *SI is *not* a back edge 1367 if (*SI == 0 || !VisitedBlocks.alreadySet(*SI)) 1368 continue; 1369 1370 CFGBlock *FirstLoopBlock = *SI; 1371 Context LoopBegin = BlockInfo[FirstLoopBlock->getBlockID()].EntryContext; 1372 Context LoopEnd = CurrBlockInfo->ExitContext; 1373 intersectBackEdge(LoopBegin, LoopEnd); 1374 } 1375 } 1376 1377 // Put an extra entry at the end of the indexed context array 1378 unsigned exitID = CFGraph->getExit().getBlockID(); 1379 saveContext(0, BlockInfo[exitID].ExitContext); 1380} 1381 1382/// Find the appropriate source locations to use when producing diagnostics for 1383/// each block in the CFG. 1384static void findBlockLocations(CFG *CFGraph, 1385 PostOrderCFGView *SortedGraph, 1386 std::vector<CFGBlockInfo> &BlockInfo) { 1387 for (PostOrderCFGView::iterator I = SortedGraph->begin(), 1388 E = SortedGraph->end(); I!= E; ++I) { 1389 const CFGBlock *CurrBlock = *I; 1390 CFGBlockInfo *CurrBlockInfo = &BlockInfo[CurrBlock->getBlockID()]; 1391 1392 // Find the source location of the last statement in the block, if the 1393 // block is not empty. 1394 if (const Stmt *S = CurrBlock->getTerminator()) { 1395 CurrBlockInfo->EntryLoc = CurrBlockInfo->ExitLoc = S->getLocStart(); 1396 } else { 1397 for (CFGBlock::const_reverse_iterator BI = CurrBlock->rbegin(), 1398 BE = CurrBlock->rend(); BI != BE; ++BI) { 1399 // FIXME: Handle other CFGElement kinds. 1400 if (Optional<CFGStmt> CS = BI->getAs<CFGStmt>()) { 1401 CurrBlockInfo->ExitLoc = CS->getStmt()->getLocStart(); 1402 break; 1403 } 1404 } 1405 } 1406 1407 if (!CurrBlockInfo->ExitLoc.isInvalid()) { 1408 // This block contains at least one statement. Find the source location 1409 // of the first statement in the block. 1410 for (CFGBlock::const_iterator BI = CurrBlock->begin(), 1411 BE = CurrBlock->end(); BI != BE; ++BI) { 1412 // FIXME: Handle other CFGElement kinds. 1413 if (Optional<CFGStmt> CS = BI->getAs<CFGStmt>()) { 1414 CurrBlockInfo->EntryLoc = CS->getStmt()->getLocStart(); 1415 break; 1416 } 1417 } 1418 } else if (CurrBlock->pred_size() == 1 && *CurrBlock->pred_begin() && 1419 CurrBlock != &CFGraph->getExit()) { 1420 // The block is empty, and has a single predecessor. Use its exit 1421 // location. 1422 CurrBlockInfo->EntryLoc = CurrBlockInfo->ExitLoc = 1423 BlockInfo[(*CurrBlock->pred_begin())->getBlockID()].ExitLoc; 1424 } 1425 } 1426} 1427 1428/// \brief Class which implements the core thread safety analysis routines. 1429class ThreadSafetyAnalyzer { 1430 friend class BuildLockset; 1431 1432 ThreadSafetyHandler &Handler; 1433 LocalVariableMap LocalVarMap; 1434 FactManager FactMan; 1435 std::vector<CFGBlockInfo> BlockInfo; 1436 1437public: 1438 ThreadSafetyAnalyzer(ThreadSafetyHandler &H) : Handler(H) {} 1439 1440 void addLock(FactSet &FSet, const SExpr &Mutex, const LockData &LDat); 1441 void removeLock(FactSet &FSet, const SExpr &Mutex, 1442 SourceLocation UnlockLoc, bool FullyRemove=false); 1443 1444 template <typename AttrType> 1445 void getMutexIDs(MutexIDList &Mtxs, AttrType *Attr, Expr *Exp, 1446 const NamedDecl *D, VarDecl *SelfDecl=0); 1447 1448 template <class AttrType> 1449 void getMutexIDs(MutexIDList &Mtxs, AttrType *Attr, Expr *Exp, 1450 const NamedDecl *D, 1451 const CFGBlock *PredBlock, const CFGBlock *CurrBlock, 1452 Expr *BrE, bool Neg); 1453 1454 const CallExpr* getTrylockCallExpr(const Stmt *Cond, LocalVarContext C, 1455 bool &Negate); 1456 1457 void getEdgeLockset(FactSet &Result, const FactSet &ExitSet, 1458 const CFGBlock* PredBlock, 1459 const CFGBlock *CurrBlock); 1460 1461 void intersectAndWarn(FactSet &FSet1, const FactSet &FSet2, 1462 SourceLocation JoinLoc, 1463 LockErrorKind LEK1, LockErrorKind LEK2, 1464 bool Modify=true); 1465 1466 void intersectAndWarn(FactSet &FSet1, const FactSet &FSet2, 1467 SourceLocation JoinLoc, LockErrorKind LEK1, 1468 bool Modify=true) { 1469 intersectAndWarn(FSet1, FSet2, JoinLoc, LEK1, LEK1, Modify); 1470 } 1471 1472 void runAnalysis(AnalysisDeclContext &AC); 1473}; 1474 1475 1476/// \brief Add a new lock to the lockset, warning if the lock is already there. 1477/// \param Mutex -- the Mutex expression for the lock 1478/// \param LDat -- the LockData for the lock 1479void ThreadSafetyAnalyzer::addLock(FactSet &FSet, const SExpr &Mutex, 1480 const LockData &LDat) { 1481 // FIXME: deal with acquired before/after annotations. 1482 // FIXME: Don't always warn when we have support for reentrant locks. 1483 if (Mutex.shouldIgnore()) 1484 return; 1485 1486 if (FSet.findLock(FactMan, Mutex)) { 1487 Handler.handleDoubleLock(Mutex.toString(), LDat.AcquireLoc); 1488 } else { 1489 FSet.addLock(FactMan, Mutex, LDat); 1490 } 1491} 1492 1493 1494/// \brief Remove a lock from the lockset, warning if the lock is not there. 1495/// \param Mutex The lock expression corresponding to the lock to be removed 1496/// \param UnlockLoc The source location of the unlock (only used in error msg) 1497void ThreadSafetyAnalyzer::removeLock(FactSet &FSet, 1498 const SExpr &Mutex, 1499 SourceLocation UnlockLoc, 1500 bool FullyRemove) { 1501 if (Mutex.shouldIgnore()) 1502 return; 1503 1504 const LockData *LDat = FSet.findLock(FactMan, Mutex); 1505 if (!LDat) { 1506 Handler.handleUnmatchedUnlock(Mutex.toString(), UnlockLoc); 1507 return; 1508 } 1509 1510 if (LDat->UnderlyingMutex.isValid()) { 1511 // This is scoped lockable object, which manages the real mutex. 1512 if (FullyRemove) { 1513 // We're destroying the managing object. 1514 // Remove the underlying mutex if it exists; but don't warn. 1515 if (FSet.findLock(FactMan, LDat->UnderlyingMutex)) 1516 FSet.removeLock(FactMan, LDat->UnderlyingMutex); 1517 } else { 1518 // We're releasing the underlying mutex, but not destroying the 1519 // managing object. Warn on dual release. 1520 if (!FSet.findLock(FactMan, LDat->UnderlyingMutex)) { 1521 Handler.handleUnmatchedUnlock(LDat->UnderlyingMutex.toString(), 1522 UnlockLoc); 1523 } 1524 FSet.removeLock(FactMan, LDat->UnderlyingMutex); 1525 return; 1526 } 1527 } 1528 FSet.removeLock(FactMan, Mutex); 1529} 1530 1531 1532/// \brief Extract the list of mutexIDs from the attribute on an expression, 1533/// and push them onto Mtxs, discarding any duplicates. 1534template <typename AttrType> 1535void ThreadSafetyAnalyzer::getMutexIDs(MutexIDList &Mtxs, AttrType *Attr, 1536 Expr *Exp, const NamedDecl *D, 1537 VarDecl *SelfDecl) { 1538 typedef typename AttrType::args_iterator iterator_type; 1539 1540 if (Attr->args_size() == 0) { 1541 // The mutex held is the "this" object. 1542 SExpr Mu(0, Exp, D, SelfDecl); 1543 if (!Mu.isValid()) 1544 SExpr::warnInvalidLock(Handler, 0, Exp, D); 1545 else 1546 Mtxs.push_back_nodup(Mu); 1547 return; 1548 } 1549 1550 for (iterator_type I=Attr->args_begin(), E=Attr->args_end(); I != E; ++I) { 1551 SExpr Mu(*I, Exp, D, SelfDecl); 1552 if (!Mu.isValid()) 1553 SExpr::warnInvalidLock(Handler, *I, Exp, D); 1554 else 1555 Mtxs.push_back_nodup(Mu); 1556 } 1557} 1558 1559 1560/// \brief Extract the list of mutexIDs from a trylock attribute. If the 1561/// trylock applies to the given edge, then push them onto Mtxs, discarding 1562/// any duplicates. 1563template <class AttrType> 1564void ThreadSafetyAnalyzer::getMutexIDs(MutexIDList &Mtxs, AttrType *Attr, 1565 Expr *Exp, const NamedDecl *D, 1566 const CFGBlock *PredBlock, 1567 const CFGBlock *CurrBlock, 1568 Expr *BrE, bool Neg) { 1569 // Find out which branch has the lock 1570 bool branch = 0; 1571 if (CXXBoolLiteralExpr *BLE = dyn_cast_or_null<CXXBoolLiteralExpr>(BrE)) { 1572 branch = BLE->getValue(); 1573 } 1574 else if (IntegerLiteral *ILE = dyn_cast_or_null<IntegerLiteral>(BrE)) { 1575 branch = ILE->getValue().getBoolValue(); 1576 } 1577 int branchnum = branch ? 0 : 1; 1578 if (Neg) branchnum = !branchnum; 1579 1580 // If we've taken the trylock branch, then add the lock 1581 int i = 0; 1582 for (CFGBlock::const_succ_iterator SI = PredBlock->succ_begin(), 1583 SE = PredBlock->succ_end(); SI != SE && i < 2; ++SI, ++i) { 1584 if (*SI == CurrBlock && i == branchnum) { 1585 getMutexIDs(Mtxs, Attr, Exp, D); 1586 } 1587 } 1588} 1589 1590 1591bool getStaticBooleanValue(Expr* E, bool& TCond) { 1592 if (isa<CXXNullPtrLiteralExpr>(E) || isa<GNUNullExpr>(E)) { 1593 TCond = false; 1594 return true; 1595 } else if (CXXBoolLiteralExpr *BLE = dyn_cast<CXXBoolLiteralExpr>(E)) { 1596 TCond = BLE->getValue(); 1597 return true; 1598 } else if (IntegerLiteral *ILE = dyn_cast<IntegerLiteral>(E)) { 1599 TCond = ILE->getValue().getBoolValue(); 1600 return true; 1601 } else if (ImplicitCastExpr *CE = dyn_cast<ImplicitCastExpr>(E)) { 1602 return getStaticBooleanValue(CE->getSubExpr(), TCond); 1603 } 1604 return false; 1605} 1606 1607 1608// If Cond can be traced back to a function call, return the call expression. 1609// The negate variable should be called with false, and will be set to true 1610// if the function call is negated, e.g. if (!mu.tryLock(...)) 1611const CallExpr* ThreadSafetyAnalyzer::getTrylockCallExpr(const Stmt *Cond, 1612 LocalVarContext C, 1613 bool &Negate) { 1614 if (!Cond) 1615 return 0; 1616 1617 if (const CallExpr *CallExp = dyn_cast<CallExpr>(Cond)) { 1618 return CallExp; 1619 } 1620 else if (const ParenExpr *PE = dyn_cast<ParenExpr>(Cond)) { 1621 return getTrylockCallExpr(PE->getSubExpr(), C, Negate); 1622 } 1623 else if (const ImplicitCastExpr *CE = dyn_cast<ImplicitCastExpr>(Cond)) { 1624 return getTrylockCallExpr(CE->getSubExpr(), C, Negate); 1625 } 1626 else if (const ExprWithCleanups* EWC = dyn_cast<ExprWithCleanups>(Cond)) { 1627 return getTrylockCallExpr(EWC->getSubExpr(), C, Negate); 1628 } 1629 else if (const DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(Cond)) { 1630 const Expr *E = LocalVarMap.lookupExpr(DRE->getDecl(), C); 1631 return getTrylockCallExpr(E, C, Negate); 1632 } 1633 else if (const UnaryOperator *UOP = dyn_cast<UnaryOperator>(Cond)) { 1634 if (UOP->getOpcode() == UO_LNot) { 1635 Negate = !Negate; 1636 return getTrylockCallExpr(UOP->getSubExpr(), C, Negate); 1637 } 1638 return 0; 1639 } 1640 else if (const BinaryOperator *BOP = dyn_cast<BinaryOperator>(Cond)) { 1641 if (BOP->getOpcode() == BO_EQ || BOP->getOpcode() == BO_NE) { 1642 if (BOP->getOpcode() == BO_NE) 1643 Negate = !Negate; 1644 1645 bool TCond = false; 1646 if (getStaticBooleanValue(BOP->getRHS(), TCond)) { 1647 if (!TCond) Negate = !Negate; 1648 return getTrylockCallExpr(BOP->getLHS(), C, Negate); 1649 } 1650 else if (getStaticBooleanValue(BOP->getLHS(), TCond)) { 1651 if (!TCond) Negate = !Negate; 1652 return getTrylockCallExpr(BOP->getRHS(), C, Negate); 1653 } 1654 return 0; 1655 } 1656 return 0; 1657 } 1658 // FIXME -- handle && and || as well. 1659 return 0; 1660} 1661 1662 1663/// \brief Find the lockset that holds on the edge between PredBlock 1664/// and CurrBlock. The edge set is the exit set of PredBlock (passed 1665/// as the ExitSet parameter) plus any trylocks, which are conditionally held. 1666void ThreadSafetyAnalyzer::getEdgeLockset(FactSet& Result, 1667 const FactSet &ExitSet, 1668 const CFGBlock *PredBlock, 1669 const CFGBlock *CurrBlock) { 1670 Result = ExitSet; 1671 1672 if (!PredBlock->getTerminatorCondition()) 1673 return; 1674 1675 bool Negate = false; 1676 const Stmt *Cond = PredBlock->getTerminatorCondition(); 1677 const CFGBlockInfo *PredBlockInfo = &BlockInfo[PredBlock->getBlockID()]; 1678 const LocalVarContext &LVarCtx = PredBlockInfo->ExitContext; 1679 1680 CallExpr *Exp = 1681 const_cast<CallExpr*>(getTrylockCallExpr(Cond, LVarCtx, Negate)); 1682 if (!Exp) 1683 return; 1684 1685 NamedDecl *FunDecl = dyn_cast_or_null<NamedDecl>(Exp->getCalleeDecl()); 1686 if(!FunDecl || !FunDecl->hasAttrs()) 1687 return; 1688 1689 1690 MutexIDList ExclusiveLocksToAdd; 1691 MutexIDList SharedLocksToAdd; 1692 1693 // If the condition is a call to a Trylock function, then grab the attributes 1694 AttrVec &ArgAttrs = FunDecl->getAttrs(); 1695 for (unsigned i = 0; i < ArgAttrs.size(); ++i) { 1696 Attr *Attr = ArgAttrs[i]; 1697 switch (Attr->getKind()) { 1698 case attr::ExclusiveTrylockFunction: { 1699 ExclusiveTrylockFunctionAttr *A = 1700 cast<ExclusiveTrylockFunctionAttr>(Attr); 1701 getMutexIDs(ExclusiveLocksToAdd, A, Exp, FunDecl, 1702 PredBlock, CurrBlock, A->getSuccessValue(), Negate); 1703 break; 1704 } 1705 case attr::SharedTrylockFunction: { 1706 SharedTrylockFunctionAttr *A = 1707 cast<SharedTrylockFunctionAttr>(Attr); 1708 getMutexIDs(SharedLocksToAdd, A, Exp, FunDecl, 1709 PredBlock, CurrBlock, A->getSuccessValue(), Negate); 1710 break; 1711 } 1712 default: 1713 break; 1714 } 1715 } 1716 1717 // Add and remove locks. 1718 SourceLocation Loc = Exp->getExprLoc(); 1719 for (unsigned i=0,n=ExclusiveLocksToAdd.size(); i<n; ++i) { 1720 addLock(Result, ExclusiveLocksToAdd[i], 1721 LockData(Loc, LK_Exclusive)); 1722 } 1723 for (unsigned i=0,n=SharedLocksToAdd.size(); i<n; ++i) { 1724 addLock(Result, SharedLocksToAdd[i], 1725 LockData(Loc, LK_Shared)); 1726 } 1727} 1728 1729 1730/// \brief We use this class to visit different types of expressions in 1731/// CFGBlocks, and build up the lockset. 1732/// An expression may cause us to add or remove locks from the lockset, or else 1733/// output error messages related to missing locks. 1734/// FIXME: In future, we may be able to not inherit from a visitor. 1735class BuildLockset : public StmtVisitor<BuildLockset> { 1736 friend class ThreadSafetyAnalyzer; 1737 1738 ThreadSafetyAnalyzer *Analyzer; 1739 FactSet FSet; 1740 LocalVariableMap::Context LVarCtx; 1741 unsigned CtxIndex; 1742 1743 // Helper functions 1744 const ValueDecl *getValueDecl(const Expr *Exp); 1745 1746 void warnIfMutexNotHeld(const NamedDecl *D, const Expr *Exp, AccessKind AK, 1747 Expr *MutexExp, ProtectedOperationKind POK); 1748 void warnIfMutexHeld(const NamedDecl *D, const Expr *Exp, Expr *MutexExp); 1749 1750 void checkAccess(const Expr *Exp, AccessKind AK); 1751 void checkPtAccess(const Expr *Exp, AccessKind AK); 1752 1753 void handleCall(Expr *Exp, const NamedDecl *D, VarDecl *VD = 0); 1754 1755public: 1756 BuildLockset(ThreadSafetyAnalyzer *Anlzr, CFGBlockInfo &Info) 1757 : StmtVisitor<BuildLockset>(), 1758 Analyzer(Anlzr), 1759 FSet(Info.EntrySet), 1760 LVarCtx(Info.EntryContext), 1761 CtxIndex(Info.EntryIndex) 1762 {} 1763 1764 void VisitUnaryOperator(UnaryOperator *UO); 1765 void VisitBinaryOperator(BinaryOperator *BO); 1766 void VisitCastExpr(CastExpr *CE); 1767 void VisitCallExpr(CallExpr *Exp); 1768 void VisitCXXConstructExpr(CXXConstructExpr *Exp); 1769 void VisitDeclStmt(DeclStmt *S); 1770}; 1771 1772 1773/// \brief Gets the value decl pointer from DeclRefExprs or MemberExprs 1774const ValueDecl *BuildLockset::getValueDecl(const Expr *Exp) { 1775 if (const ImplicitCastExpr *CE = dyn_cast<ImplicitCastExpr>(Exp)) 1776 return getValueDecl(CE->getSubExpr()); 1777 1778 if (const DeclRefExpr *DR = dyn_cast<DeclRefExpr>(Exp)) 1779 return DR->getDecl(); 1780 1781 if (const MemberExpr *ME = dyn_cast<MemberExpr>(Exp)) 1782 return ME->getMemberDecl(); 1783 1784 return 0; 1785} 1786 1787/// \brief Warn if the LSet does not contain a lock sufficient to protect access 1788/// of at least the passed in AccessKind. 1789void BuildLockset::warnIfMutexNotHeld(const NamedDecl *D, const Expr *Exp, 1790 AccessKind AK, Expr *MutexExp, 1791 ProtectedOperationKind POK) { 1792 LockKind LK = getLockKindFromAccessKind(AK); 1793 1794 SExpr Mutex(MutexExp, Exp, D); 1795 if (!Mutex.isValid()) { 1796 SExpr::warnInvalidLock(Analyzer->Handler, MutexExp, Exp, D); 1797 return; 1798 } else if (Mutex.shouldIgnore()) { 1799 return; 1800 } 1801 1802 LockData* LDat = FSet.findLockUniv(Analyzer->FactMan, Mutex); 1803 bool NoError = true; 1804 if (!LDat) { 1805 // No exact match found. Look for a partial match. 1806 FactEntry* FEntry = FSet.findPartialMatch(Analyzer->FactMan, Mutex); 1807 if (FEntry) { 1808 // Warn that there's no precise match. 1809 LDat = &FEntry->LDat; 1810 std::string PartMatchStr = FEntry->MutID.toString(); 1811 StringRef PartMatchName(PartMatchStr); 1812 Analyzer->Handler.handleMutexNotHeld(D, POK, Mutex.toString(), LK, 1813 Exp->getExprLoc(), &PartMatchName); 1814 } else { 1815 // Warn that there's no match at all. 1816 Analyzer->Handler.handleMutexNotHeld(D, POK, Mutex.toString(), LK, 1817 Exp->getExprLoc()); 1818 } 1819 NoError = false; 1820 } 1821 // Make sure the mutex we found is the right kind. 1822 if (NoError && LDat && !LDat->isAtLeast(LK)) 1823 Analyzer->Handler.handleMutexNotHeld(D, POK, Mutex.toString(), LK, 1824 Exp->getExprLoc()); 1825} 1826 1827/// \brief Warn if the LSet contains the given lock. 1828void BuildLockset::warnIfMutexHeld(const NamedDecl *D, const Expr* Exp, 1829 Expr *MutexExp) { 1830 SExpr Mutex(MutexExp, Exp, D); 1831 if (!Mutex.isValid()) { 1832 SExpr::warnInvalidLock(Analyzer->Handler, MutexExp, Exp, D); 1833 return; 1834 } 1835 1836 LockData* LDat = FSet.findLock(Analyzer->FactMan, Mutex); 1837 if (LDat) { 1838 std::string DeclName = D->getNameAsString(); 1839 StringRef DeclNameSR (DeclName); 1840 Analyzer->Handler.handleFunExcludesLock(DeclNameSR, Mutex.toString(), 1841 Exp->getExprLoc()); 1842 } 1843} 1844 1845 1846/// \brief Checks guarded_by and pt_guarded_by attributes. 1847/// Whenever we identify an access (read or write) to a DeclRefExpr that is 1848/// marked with guarded_by, we must ensure the appropriate mutexes are held. 1849/// Similarly, we check if the access is to an expression that dereferences 1850/// a pointer marked with pt_guarded_by. 1851void BuildLockset::checkAccess(const Expr *Exp, AccessKind AK) { 1852 Exp = Exp->IgnoreParenCasts(); 1853 1854 if (const UnaryOperator *UO = dyn_cast<UnaryOperator>(Exp)) { 1855 // For dereferences 1856 if (UO->getOpcode() == clang::UO_Deref) 1857 checkPtAccess(UO->getSubExpr(), AK); 1858 return; 1859 } 1860 1861 if (const MemberExpr *ME = dyn_cast<MemberExpr>(Exp)) { 1862 if (ME->isArrow()) 1863 checkPtAccess(ME->getBase(), AK); 1864 else 1865 checkAccess(ME->getBase(), AK); 1866 } 1867 1868 const ValueDecl *D = getValueDecl(Exp); 1869 if (!D || !D->hasAttrs()) 1870 return; 1871 1872 if (D->getAttr<GuardedVarAttr>() && FSet.isEmpty()) 1873 Analyzer->Handler.handleNoMutexHeld(D, POK_VarAccess, AK, 1874 Exp->getExprLoc()); 1875 1876 const AttrVec &ArgAttrs = D->getAttrs(); 1877 for (unsigned i = 0, Size = ArgAttrs.size(); i < Size; ++i) 1878 if (GuardedByAttr *GBAttr = dyn_cast<GuardedByAttr>(ArgAttrs[i])) 1879 warnIfMutexNotHeld(D, Exp, AK, GBAttr->getArg(), POK_VarAccess); 1880} 1881 1882/// \brief Checks pt_guarded_by and pt_guarded_var attributes. 1883void BuildLockset::checkPtAccess(const Expr *Exp, AccessKind AK) { 1884 Exp = Exp->IgnoreParenCasts(); 1885 1886 const ValueDecl *D = getValueDecl(Exp); 1887 if (!D || !D->hasAttrs()) 1888 return; 1889 1890 if (D->getAttr<PtGuardedVarAttr>() && FSet.isEmpty()) 1891 Analyzer->Handler.handleNoMutexHeld(D, POK_VarDereference, AK, 1892 Exp->getExprLoc()); 1893 1894 const AttrVec &ArgAttrs = D->getAttrs(); 1895 for (unsigned i = 0, Size = ArgAttrs.size(); i < Size; ++i) 1896 if (PtGuardedByAttr *GBAttr = dyn_cast<PtGuardedByAttr>(ArgAttrs[i])) 1897 warnIfMutexNotHeld(D, Exp, AK, GBAttr->getArg(), POK_VarDereference); 1898} 1899 1900 1901/// \brief Process a function call, method call, constructor call, 1902/// or destructor call. This involves looking at the attributes on the 1903/// corresponding function/method/constructor/destructor, issuing warnings, 1904/// and updating the locksets accordingly. 1905/// 1906/// FIXME: For classes annotated with one of the guarded annotations, we need 1907/// to treat const method calls as reads and non-const method calls as writes, 1908/// and check that the appropriate locks are held. Non-const method calls with 1909/// the same signature as const method calls can be also treated as reads. 1910/// 1911void BuildLockset::handleCall(Expr *Exp, const NamedDecl *D, VarDecl *VD) { 1912 const AttrVec &ArgAttrs = D->getAttrs(); 1913 MutexIDList ExclusiveLocksToAdd; 1914 MutexIDList SharedLocksToAdd; 1915 MutexIDList LocksToRemove; 1916 1917 for(unsigned i = 0; i < ArgAttrs.size(); ++i) { 1918 Attr *At = const_cast<Attr*>(ArgAttrs[i]); 1919 switch (At->getKind()) { 1920 // When we encounter an exclusive lock function, we need to add the lock 1921 // to our lockset with kind exclusive. 1922 case attr::ExclusiveLockFunction: { 1923 ExclusiveLockFunctionAttr *A = cast<ExclusiveLockFunctionAttr>(At); 1924 Analyzer->getMutexIDs(ExclusiveLocksToAdd, A, Exp, D, VD); 1925 break; 1926 } 1927 1928 // When we encounter a shared lock function, we need to add the lock 1929 // to our lockset with kind shared. 1930 case attr::SharedLockFunction: { 1931 SharedLockFunctionAttr *A = cast<SharedLockFunctionAttr>(At); 1932 Analyzer->getMutexIDs(SharedLocksToAdd, A, Exp, D, VD); 1933 break; 1934 } 1935 1936 // When we encounter an unlock function, we need to remove unlocked 1937 // mutexes from the lockset, and flag a warning if they are not there. 1938 case attr::UnlockFunction: { 1939 UnlockFunctionAttr *A = cast<UnlockFunctionAttr>(At); 1940 Analyzer->getMutexIDs(LocksToRemove, A, Exp, D, VD); 1941 break; 1942 } 1943 1944 case attr::ExclusiveLocksRequired: { 1945 ExclusiveLocksRequiredAttr *A = cast<ExclusiveLocksRequiredAttr>(At); 1946 1947 for (ExclusiveLocksRequiredAttr::args_iterator 1948 I = A->args_begin(), E = A->args_end(); I != E; ++I) 1949 warnIfMutexNotHeld(D, Exp, AK_Written, *I, POK_FunctionCall); 1950 break; 1951 } 1952 1953 case attr::SharedLocksRequired: { 1954 SharedLocksRequiredAttr *A = cast<SharedLocksRequiredAttr>(At); 1955 1956 for (SharedLocksRequiredAttr::args_iterator I = A->args_begin(), 1957 E = A->args_end(); I != E; ++I) 1958 warnIfMutexNotHeld(D, Exp, AK_Read, *I, POK_FunctionCall); 1959 break; 1960 } 1961 1962 case attr::LocksExcluded: { 1963 LocksExcludedAttr *A = cast<LocksExcludedAttr>(At); 1964 1965 for (LocksExcludedAttr::args_iterator I = A->args_begin(), 1966 E = A->args_end(); I != E; ++I) { 1967 warnIfMutexHeld(D, Exp, *I); 1968 } 1969 break; 1970 } 1971 1972 // Ignore other (non thread-safety) attributes 1973 default: 1974 break; 1975 } 1976 } 1977 1978 // Figure out if we're calling the constructor of scoped lockable class 1979 bool isScopedVar = false; 1980 if (VD) { 1981 if (const CXXConstructorDecl *CD = dyn_cast<const CXXConstructorDecl>(D)) { 1982 const CXXRecordDecl* PD = CD->getParent(); 1983 if (PD && PD->getAttr<ScopedLockableAttr>()) 1984 isScopedVar = true; 1985 } 1986 } 1987 1988 // Add locks. 1989 SourceLocation Loc = Exp->getExprLoc(); 1990 for (unsigned i=0,n=ExclusiveLocksToAdd.size(); i<n; ++i) { 1991 Analyzer->addLock(FSet, ExclusiveLocksToAdd[i], 1992 LockData(Loc, LK_Exclusive, isScopedVar)); 1993 } 1994 for (unsigned i=0,n=SharedLocksToAdd.size(); i<n; ++i) { 1995 Analyzer->addLock(FSet, SharedLocksToAdd[i], 1996 LockData(Loc, LK_Shared, isScopedVar)); 1997 } 1998 1999 // Add the managing object as a dummy mutex, mapped to the underlying mutex. 2000 // FIXME -- this doesn't work if we acquire multiple locks. 2001 if (isScopedVar) { 2002 SourceLocation MLoc = VD->getLocation(); 2003 DeclRefExpr DRE(VD, false, VD->getType(), VK_LValue, VD->getLocation()); 2004 SExpr SMutex(&DRE, 0, 0); 2005 2006 for (unsigned i=0,n=ExclusiveLocksToAdd.size(); i<n; ++i) { 2007 Analyzer->addLock(FSet, SMutex, LockData(MLoc, LK_Exclusive, 2008 ExclusiveLocksToAdd[i])); 2009 } 2010 for (unsigned i=0,n=SharedLocksToAdd.size(); i<n; ++i) { 2011 Analyzer->addLock(FSet, SMutex, LockData(MLoc, LK_Shared, 2012 SharedLocksToAdd[i])); 2013 } 2014 } 2015 2016 // Remove locks. 2017 // FIXME -- should only fully remove if the attribute refers to 'this'. 2018 bool Dtor = isa<CXXDestructorDecl>(D); 2019 for (unsigned i=0,n=LocksToRemove.size(); i<n; ++i) { 2020 Analyzer->removeLock(FSet, LocksToRemove[i], Loc, Dtor); 2021 } 2022} 2023 2024 2025/// \brief For unary operations which read and write a variable, we need to 2026/// check whether we hold any required mutexes. Reads are checked in 2027/// VisitCastExpr. 2028void BuildLockset::VisitUnaryOperator(UnaryOperator *UO) { 2029 switch (UO->getOpcode()) { 2030 case clang::UO_PostDec: 2031 case clang::UO_PostInc: 2032 case clang::UO_PreDec: 2033 case clang::UO_PreInc: { 2034 checkAccess(UO->getSubExpr(), AK_Written); 2035 break; 2036 } 2037 default: 2038 break; 2039 } 2040} 2041 2042/// For binary operations which assign to a variable (writes), we need to check 2043/// whether we hold any required mutexes. 2044/// FIXME: Deal with non-primitive types. 2045void BuildLockset::VisitBinaryOperator(BinaryOperator *BO) { 2046 if (!BO->isAssignmentOp()) 2047 return; 2048 2049 // adjust the context 2050 LVarCtx = Analyzer->LocalVarMap.getNextContext(CtxIndex, BO, LVarCtx); 2051 2052 checkAccess(BO->getLHS(), AK_Written); 2053} 2054 2055/// Whenever we do an LValue to Rvalue cast, we are reading a variable and 2056/// need to ensure we hold any required mutexes. 2057/// FIXME: Deal with non-primitive types. 2058void BuildLockset::VisitCastExpr(CastExpr *CE) { 2059 if (CE->getCastKind() != CK_LValueToRValue) 2060 return; 2061 checkAccess(CE->getSubExpr(), AK_Read); 2062} 2063 2064 2065void BuildLockset::VisitCallExpr(CallExpr *Exp) { 2066 if (CXXMemberCallExpr *CE = dyn_cast<CXXMemberCallExpr>(Exp)) { 2067 MemberExpr *ME = dyn_cast<MemberExpr>(CE->getCallee()); 2068 // ME can be null when calling a method pointer 2069 CXXMethodDecl *MD = CE->getMethodDecl(); 2070 2071 if (ME && MD) { 2072 if (ME->isArrow()) { 2073 if (MD->isConst()) { 2074 checkPtAccess(CE->getImplicitObjectArgument(), AK_Read); 2075 } else { // FIXME -- should be AK_Written 2076 checkPtAccess(CE->getImplicitObjectArgument(), AK_Read); 2077 } 2078 } else { 2079 if (MD->isConst()) 2080 checkAccess(CE->getImplicitObjectArgument(), AK_Read); 2081 else // FIXME -- should be AK_Written 2082 checkAccess(CE->getImplicitObjectArgument(), AK_Read); 2083 } 2084 } 2085 } else if (CXXOperatorCallExpr *OE = dyn_cast<CXXOperatorCallExpr>(Exp)) { 2086 switch (OE->getOperator()) { 2087 case OO_Equal: { 2088 const Expr *Target = OE->getArg(0); 2089 const Expr *Source = OE->getArg(1); 2090 checkAccess(Target, AK_Written); 2091 checkAccess(Source, AK_Read); 2092 break; 2093 } 2094 default: { 2095 const Expr *Source = OE->getArg(0); 2096 checkAccess(Source, AK_Read); 2097 break; 2098 } 2099 } 2100 } 2101 NamedDecl *D = dyn_cast_or_null<NamedDecl>(Exp->getCalleeDecl()); 2102 if(!D || !D->hasAttrs()) 2103 return; 2104 handleCall(Exp, D); 2105} 2106 2107void BuildLockset::VisitCXXConstructExpr(CXXConstructExpr *Exp) { 2108 const CXXConstructorDecl *D = Exp->getConstructor(); 2109 if (D && D->isCopyConstructor()) { 2110 const Expr* Source = Exp->getArg(0); 2111 checkAccess(Source, AK_Read); 2112 } 2113 // FIXME -- only handles constructors in DeclStmt below. 2114} 2115 2116void BuildLockset::VisitDeclStmt(DeclStmt *S) { 2117 // adjust the context 2118 LVarCtx = Analyzer->LocalVarMap.getNextContext(CtxIndex, S, LVarCtx); 2119 2120 DeclGroupRef DGrp = S->getDeclGroup(); 2121 for (DeclGroupRef::iterator I = DGrp.begin(), E = DGrp.end(); I != E; ++I) { 2122 Decl *D = *I; 2123 if (VarDecl *VD = dyn_cast_or_null<VarDecl>(D)) { 2124 Expr *E = VD->getInit(); 2125 // handle constructors that involve temporaries 2126 if (ExprWithCleanups *EWC = dyn_cast_or_null<ExprWithCleanups>(E)) 2127 E = EWC->getSubExpr(); 2128 2129 if (CXXConstructExpr *CE = dyn_cast_or_null<CXXConstructExpr>(E)) { 2130 NamedDecl *CtorD = dyn_cast_or_null<NamedDecl>(CE->getConstructor()); 2131 if (!CtorD || !CtorD->hasAttrs()) 2132 return; 2133 handleCall(CE, CtorD, VD); 2134 } 2135 } 2136 } 2137} 2138 2139 2140 2141/// \brief Compute the intersection of two locksets and issue warnings for any 2142/// locks in the symmetric difference. 2143/// 2144/// This function is used at a merge point in the CFG when comparing the lockset 2145/// of each branch being merged. For example, given the following sequence: 2146/// A; if () then B; else C; D; we need to check that the lockset after B and C 2147/// are the same. In the event of a difference, we use the intersection of these 2148/// two locksets at the start of D. 2149/// 2150/// \param FSet1 The first lockset. 2151/// \param FSet2 The second lockset. 2152/// \param JoinLoc The location of the join point for error reporting 2153/// \param LEK1 The error message to report if a mutex is missing from LSet1 2154/// \param LEK2 The error message to report if a mutex is missing from Lset2 2155void ThreadSafetyAnalyzer::intersectAndWarn(FactSet &FSet1, 2156 const FactSet &FSet2, 2157 SourceLocation JoinLoc, 2158 LockErrorKind LEK1, 2159 LockErrorKind LEK2, 2160 bool Modify) { 2161 FactSet FSet1Orig = FSet1; 2162 2163 for (FactSet::const_iterator I = FSet2.begin(), E = FSet2.end(); 2164 I != E; ++I) { 2165 const SExpr &FSet2Mutex = FactMan[*I].MutID; 2166 const LockData &LDat2 = FactMan[*I].LDat; 2167 2168 if (const LockData *LDat1 = FSet1.findLock(FactMan, FSet2Mutex)) { 2169 if (LDat1->LKind != LDat2.LKind) { 2170 Handler.handleExclusiveAndShared(FSet2Mutex.toString(), 2171 LDat2.AcquireLoc, 2172 LDat1->AcquireLoc); 2173 if (Modify && LDat1->LKind != LK_Exclusive) { 2174 FSet1.removeLock(FactMan, FSet2Mutex); 2175 FSet1.addLock(FactMan, FSet2Mutex, LDat2); 2176 } 2177 } 2178 } else { 2179 if (LDat2.UnderlyingMutex.isValid()) { 2180 if (FSet2.findLock(FactMan, LDat2.UnderlyingMutex)) { 2181 // If this is a scoped lock that manages another mutex, and if the 2182 // underlying mutex is still held, then warn about the underlying 2183 // mutex. 2184 Handler.handleMutexHeldEndOfScope(LDat2.UnderlyingMutex.toString(), 2185 LDat2.AcquireLoc, 2186 JoinLoc, LEK1); 2187 } 2188 } 2189 else if (!LDat2.Managed && !FSet2Mutex.isUniversal()) 2190 Handler.handleMutexHeldEndOfScope(FSet2Mutex.toString(), 2191 LDat2.AcquireLoc, 2192 JoinLoc, LEK1); 2193 } 2194 } 2195 2196 for (FactSet::const_iterator I = FSet1.begin(), E = FSet1.end(); 2197 I != E; ++I) { 2198 const SExpr &FSet1Mutex = FactMan[*I].MutID; 2199 const LockData &LDat1 = FactMan[*I].LDat; 2200 2201 if (!FSet2.findLock(FactMan, FSet1Mutex)) { 2202 if (LDat1.UnderlyingMutex.isValid()) { 2203 if (FSet1Orig.findLock(FactMan, LDat1.UnderlyingMutex)) { 2204 // If this is a scoped lock that manages another mutex, and if the 2205 // underlying mutex is still held, then warn about the underlying 2206 // mutex. 2207 Handler.handleMutexHeldEndOfScope(LDat1.UnderlyingMutex.toString(), 2208 LDat1.AcquireLoc, 2209 JoinLoc, LEK1); 2210 } 2211 } 2212 else if (!LDat1.Managed && !FSet1Mutex.isUniversal()) 2213 Handler.handleMutexHeldEndOfScope(FSet1Mutex.toString(), 2214 LDat1.AcquireLoc, 2215 JoinLoc, LEK2); 2216 if (Modify) 2217 FSet1.removeLock(FactMan, FSet1Mutex); 2218 } 2219 } 2220} 2221 2222 2223// Return true if block B never continues to its successors. 2224inline bool neverReturns(const CFGBlock* B) { 2225 if (B->hasNoReturnElement()) 2226 return true; 2227 if (B->empty()) 2228 return false; 2229 2230 CFGElement Last = B->back(); 2231 if (Optional<CFGStmt> S = Last.getAs<CFGStmt>()) { 2232 if (isa<CXXThrowExpr>(S->getStmt())) 2233 return true; 2234 } 2235 return false; 2236} 2237 2238 2239/// \brief Check a function's CFG for thread-safety violations. 2240/// 2241/// We traverse the blocks in the CFG, compute the set of mutexes that are held 2242/// at the end of each block, and issue warnings for thread safety violations. 2243/// Each block in the CFG is traversed exactly once. 2244void ThreadSafetyAnalyzer::runAnalysis(AnalysisDeclContext &AC) { 2245 CFG *CFGraph = AC.getCFG(); 2246 if (!CFGraph) return; 2247 const NamedDecl *D = dyn_cast_or_null<NamedDecl>(AC.getDecl()); 2248 2249 // AC.dumpCFG(true); 2250 2251 if (!D) 2252 return; // Ignore anonymous functions for now. 2253 if (D->getAttr<NoThreadSafetyAnalysisAttr>()) 2254 return; 2255 // FIXME: Do something a bit more intelligent inside constructor and 2256 // destructor code. Constructors and destructors must assume unique access 2257 // to 'this', so checks on member variable access is disabled, but we should 2258 // still enable checks on other objects. 2259 if (isa<CXXConstructorDecl>(D)) 2260 return; // Don't check inside constructors. 2261 if (isa<CXXDestructorDecl>(D)) 2262 return; // Don't check inside destructors. 2263 2264 BlockInfo.resize(CFGraph->getNumBlockIDs(), 2265 CFGBlockInfo::getEmptyBlockInfo(LocalVarMap)); 2266 2267 // We need to explore the CFG via a "topological" ordering. 2268 // That way, we will be guaranteed to have information about required 2269 // predecessor locksets when exploring a new block. 2270 PostOrderCFGView *SortedGraph = AC.getAnalysis<PostOrderCFGView>(); 2271 PostOrderCFGView::CFGBlockSet VisitedBlocks(CFGraph); 2272 2273 // Mark entry block as reachable 2274 BlockInfo[CFGraph->getEntry().getBlockID()].Reachable = true; 2275 2276 // Compute SSA names for local variables 2277 LocalVarMap.traverseCFG(CFGraph, SortedGraph, BlockInfo); 2278 2279 // Fill in source locations for all CFGBlocks. 2280 findBlockLocations(CFGraph, SortedGraph, BlockInfo); 2281 2282 // Add locks from exclusive_locks_required and shared_locks_required 2283 // to initial lockset. Also turn off checking for lock and unlock functions. 2284 // FIXME: is there a more intelligent way to check lock/unlock functions? 2285 if (!SortedGraph->empty() && D->hasAttrs()) { 2286 const CFGBlock *FirstBlock = *SortedGraph->begin(); 2287 FactSet &InitialLockset = BlockInfo[FirstBlock->getBlockID()].EntrySet; 2288 const AttrVec &ArgAttrs = D->getAttrs(); 2289 2290 MutexIDList ExclusiveLocksToAdd; 2291 MutexIDList SharedLocksToAdd; 2292 2293 SourceLocation Loc = D->getLocation(); 2294 for (unsigned i = 0; i < ArgAttrs.size(); ++i) { 2295 Attr *Attr = ArgAttrs[i]; 2296 Loc = Attr->getLocation(); 2297 if (ExclusiveLocksRequiredAttr *A 2298 = dyn_cast<ExclusiveLocksRequiredAttr>(Attr)) { 2299 getMutexIDs(ExclusiveLocksToAdd, A, (Expr*) 0, D); 2300 } else if (SharedLocksRequiredAttr *A 2301 = dyn_cast<SharedLocksRequiredAttr>(Attr)) { 2302 getMutexIDs(SharedLocksToAdd, A, (Expr*) 0, D); 2303 } else if (isa<UnlockFunctionAttr>(Attr)) { 2304 // Don't try to check unlock functions for now 2305 return; 2306 } else if (isa<ExclusiveLockFunctionAttr>(Attr)) { 2307 // Don't try to check lock functions for now 2308 return; 2309 } else if (isa<SharedLockFunctionAttr>(Attr)) { 2310 // Don't try to check lock functions for now 2311 return; 2312 } else if (isa<ExclusiveTrylockFunctionAttr>(Attr)) { 2313 // Don't try to check trylock functions for now 2314 return; 2315 } else if (isa<SharedTrylockFunctionAttr>(Attr)) { 2316 // Don't try to check trylock functions for now 2317 return; 2318 } 2319 } 2320 2321 // FIXME -- Loc can be wrong here. 2322 for (unsigned i=0,n=ExclusiveLocksToAdd.size(); i<n; ++i) { 2323 addLock(InitialLockset, ExclusiveLocksToAdd[i], 2324 LockData(Loc, LK_Exclusive)); 2325 } 2326 for (unsigned i=0,n=SharedLocksToAdd.size(); i<n; ++i) { 2327 addLock(InitialLockset, SharedLocksToAdd[i], 2328 LockData(Loc, LK_Shared)); 2329 } 2330 } 2331 2332 for (PostOrderCFGView::iterator I = SortedGraph->begin(), 2333 E = SortedGraph->end(); I!= E; ++I) { 2334 const CFGBlock *CurrBlock = *I; 2335 int CurrBlockID = CurrBlock->getBlockID(); 2336 CFGBlockInfo *CurrBlockInfo = &BlockInfo[CurrBlockID]; 2337 2338 // Use the default initial lockset in case there are no predecessors. 2339 VisitedBlocks.insert(CurrBlock); 2340 2341 // Iterate through the predecessor blocks and warn if the lockset for all 2342 // predecessors is not the same. We take the entry lockset of the current 2343 // block to be the intersection of all previous locksets. 2344 // FIXME: By keeping the intersection, we may output more errors in future 2345 // for a lock which is not in the intersection, but was in the union. We 2346 // may want to also keep the union in future. As an example, let's say 2347 // the intersection contains Mutex L, and the union contains L and M. 2348 // Later we unlock M. At this point, we would output an error because we 2349 // never locked M; although the real error is probably that we forgot to 2350 // lock M on all code paths. Conversely, let's say that later we lock M. 2351 // In this case, we should compare against the intersection instead of the 2352 // union because the real error is probably that we forgot to unlock M on 2353 // all code paths. 2354 bool LocksetInitialized = false; 2355 SmallVector<CFGBlock *, 8> SpecialBlocks; 2356 for (CFGBlock::const_pred_iterator PI = CurrBlock->pred_begin(), 2357 PE = CurrBlock->pred_end(); PI != PE; ++PI) { 2358 2359 // if *PI -> CurrBlock is a back edge 2360 if (*PI == 0 || !VisitedBlocks.alreadySet(*PI)) 2361 continue; 2362 2363 int PrevBlockID = (*PI)->getBlockID(); 2364 CFGBlockInfo *PrevBlockInfo = &BlockInfo[PrevBlockID]; 2365 2366 // Ignore edges from blocks that can't return. 2367 if (neverReturns(*PI) || !PrevBlockInfo->Reachable) 2368 continue; 2369 2370 // Okay, we can reach this block from the entry. 2371 CurrBlockInfo->Reachable = true; 2372 2373 // If the previous block ended in a 'continue' or 'break' statement, then 2374 // a difference in locksets is probably due to a bug in that block, rather 2375 // than in some other predecessor. In that case, keep the other 2376 // predecessor's lockset. 2377 if (const Stmt *Terminator = (*PI)->getTerminator()) { 2378 if (isa<ContinueStmt>(Terminator) || isa<BreakStmt>(Terminator)) { 2379 SpecialBlocks.push_back(*PI); 2380 continue; 2381 } 2382 } 2383 2384 FactSet PrevLockset; 2385 getEdgeLockset(PrevLockset, PrevBlockInfo->ExitSet, *PI, CurrBlock); 2386 2387 if (!LocksetInitialized) { 2388 CurrBlockInfo->EntrySet = PrevLockset; 2389 LocksetInitialized = true; 2390 } else { 2391 intersectAndWarn(CurrBlockInfo->EntrySet, PrevLockset, 2392 CurrBlockInfo->EntryLoc, 2393 LEK_LockedSomePredecessors); 2394 } 2395 } 2396 2397 // Skip rest of block if it's not reachable. 2398 if (!CurrBlockInfo->Reachable) 2399 continue; 2400 2401 // Process continue and break blocks. Assume that the lockset for the 2402 // resulting block is unaffected by any discrepancies in them. 2403 for (unsigned SpecialI = 0, SpecialN = SpecialBlocks.size(); 2404 SpecialI < SpecialN; ++SpecialI) { 2405 CFGBlock *PrevBlock = SpecialBlocks[SpecialI]; 2406 int PrevBlockID = PrevBlock->getBlockID(); 2407 CFGBlockInfo *PrevBlockInfo = &BlockInfo[PrevBlockID]; 2408 2409 if (!LocksetInitialized) { 2410 CurrBlockInfo->EntrySet = PrevBlockInfo->ExitSet; 2411 LocksetInitialized = true; 2412 } else { 2413 // Determine whether this edge is a loop terminator for diagnostic 2414 // purposes. FIXME: A 'break' statement might be a loop terminator, but 2415 // it might also be part of a switch. Also, a subsequent destructor 2416 // might add to the lockset, in which case the real issue might be a 2417 // double lock on the other path. 2418 const Stmt *Terminator = PrevBlock->getTerminator(); 2419 bool IsLoop = Terminator && isa<ContinueStmt>(Terminator); 2420 2421 FactSet PrevLockset; 2422 getEdgeLockset(PrevLockset, PrevBlockInfo->ExitSet, 2423 PrevBlock, CurrBlock); 2424 2425 // Do not update EntrySet. 2426 intersectAndWarn(CurrBlockInfo->EntrySet, PrevLockset, 2427 PrevBlockInfo->ExitLoc, 2428 IsLoop ? LEK_LockedSomeLoopIterations 2429 : LEK_LockedSomePredecessors, 2430 false); 2431 } 2432 } 2433 2434 BuildLockset LocksetBuilder(this, *CurrBlockInfo); 2435 2436 // Visit all the statements in the basic block. 2437 for (CFGBlock::const_iterator BI = CurrBlock->begin(), 2438 BE = CurrBlock->end(); BI != BE; ++BI) { 2439 switch (BI->getKind()) { 2440 case CFGElement::Statement: { 2441 CFGStmt CS = BI->castAs<CFGStmt>(); 2442 LocksetBuilder.Visit(const_cast<Stmt*>(CS.getStmt())); 2443 break; 2444 } 2445 // Ignore BaseDtor, MemberDtor, and TemporaryDtor for now. 2446 case CFGElement::AutomaticObjectDtor: { 2447 CFGAutomaticObjDtor AD = BI->castAs<CFGAutomaticObjDtor>(); 2448 CXXDestructorDecl *DD = const_cast<CXXDestructorDecl *>( 2449 AD.getDestructorDecl(AC.getASTContext())); 2450 if (!DD->hasAttrs()) 2451 break; 2452 2453 // Create a dummy expression, 2454 VarDecl *VD = const_cast<VarDecl*>(AD.getVarDecl()); 2455 DeclRefExpr DRE(VD, false, VD->getType(), VK_LValue, 2456 AD.getTriggerStmt()->getLocEnd()); 2457 LocksetBuilder.handleCall(&DRE, DD); 2458 break; 2459 } 2460 default: 2461 break; 2462 } 2463 } 2464 CurrBlockInfo->ExitSet = LocksetBuilder.FSet; 2465 2466 // For every back edge from CurrBlock (the end of the loop) to another block 2467 // (FirstLoopBlock) we need to check that the Lockset of Block is equal to 2468 // the one held at the beginning of FirstLoopBlock. We can look up the 2469 // Lockset held at the beginning of FirstLoopBlock in the EntryLockSets map. 2470 for (CFGBlock::const_succ_iterator SI = CurrBlock->succ_begin(), 2471 SE = CurrBlock->succ_end(); SI != SE; ++SI) { 2472 2473 // if CurrBlock -> *SI is *not* a back edge 2474 if (*SI == 0 || !VisitedBlocks.alreadySet(*SI)) 2475 continue; 2476 2477 CFGBlock *FirstLoopBlock = *SI; 2478 CFGBlockInfo *PreLoop = &BlockInfo[FirstLoopBlock->getBlockID()]; 2479 CFGBlockInfo *LoopEnd = &BlockInfo[CurrBlockID]; 2480 intersectAndWarn(LoopEnd->ExitSet, PreLoop->EntrySet, 2481 PreLoop->EntryLoc, 2482 LEK_LockedSomeLoopIterations, 2483 false); 2484 } 2485 } 2486 2487 CFGBlockInfo *Initial = &BlockInfo[CFGraph->getEntry().getBlockID()]; 2488 CFGBlockInfo *Final = &BlockInfo[CFGraph->getExit().getBlockID()]; 2489 2490 // Skip the final check if the exit block is unreachable. 2491 if (!Final->Reachable) 2492 return; 2493 2494 // FIXME: Should we call this function for all blocks which exit the function? 2495 intersectAndWarn(Initial->EntrySet, Final->ExitSet, 2496 Final->ExitLoc, 2497 LEK_LockedAtEndOfFunction, 2498 LEK_NotLockedAtEndOfFunction, 2499 false); 2500} 2501 2502} // end anonymous namespace 2503 2504 2505namespace clang { 2506namespace thread_safety { 2507 2508/// \brief Check a function's CFG for thread-safety violations. 2509/// 2510/// We traverse the blocks in the CFG, compute the set of mutexes that are held 2511/// at the end of each block, and issue warnings for thread safety violations. 2512/// Each block in the CFG is traversed exactly once. 2513void runThreadSafetyAnalysis(AnalysisDeclContext &AC, 2514 ThreadSafetyHandler &Handler) { 2515 ThreadSafetyAnalyzer Analyzer(Handler); 2516 Analyzer.runAnalysis(AC); 2517} 2518 2519/// \brief Helper function that returns a LockKind required for the given level 2520/// of access. 2521LockKind getLockKindFromAccessKind(AccessKind AK) { 2522 switch (AK) { 2523 case AK_Read : 2524 return LK_Shared; 2525 case AK_Written : 2526 return LK_Exclusive; 2527 } 2528 llvm_unreachable("Unknown AccessKind"); 2529} 2530 2531}} // end namespace clang::thread_safety 2532