ExprEngineCallAndReturn.cpp revision 8235f9c9c8b3d1737d1c6bd57f7ba3f616b92392
1//=-- ExprEngineCallAndReturn.cpp - Support for call/return -----*- C++ -*-===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file defines ExprEngine's support for calls and returns. 11// 12//===----------------------------------------------------------------------===// 13 14#include "clang/StaticAnalyzer/Core/CheckerManager.h" 15#include "clang/StaticAnalyzer/Core/PathSensitive/ExprEngine.h" 16#include "clang/StaticAnalyzer/Core/PathSensitive/ObjCMessage.h" 17#include "llvm/Support/SaveAndRestore.h" 18#include "clang/AST/DeclCXX.h" 19 20using namespace clang; 21using namespace ento; 22 23void ExprEngine::processCallEnter(CallEnter CE, ExplodedNode *Pred) { 24 // Get the entry block in the CFG of the callee. 25 const StackFrameContext *calleeCtx = CE.getCalleeContext(); 26 const CFG *CalleeCFG = calleeCtx->getCFG(); 27 const CFGBlock *Entry = &(CalleeCFG->getEntry()); 28 29 // Validate the CFG. 30 assert(Entry->empty()); 31 assert(Entry->succ_size() == 1); 32 33 // Get the solitary sucessor. 34 const CFGBlock *Succ = *(Entry->succ_begin()); 35 36 // Construct an edge representing the starting location in the callee. 37 BlockEdge Loc(Entry, Succ, calleeCtx); 38 39 // Construct a new state which contains the mapping from actual to 40 // formal arguments. 41 const LocationContext *callerCtx = Pred->getLocationContext(); 42 ProgramStateRef state = Pred->getState()->enterStackFrame(callerCtx, 43 calleeCtx); 44 45 // Construct a new node and add it to the worklist. 46 bool isNew; 47 ExplodedNode *Node = G.getNode(Loc, state, false, &isNew); 48 Node->addPredecessor(Pred, G); 49 if (isNew) 50 Engine.getWorkList()->enqueue(Node); 51} 52 53static const ReturnStmt *getReturnStmt(const ExplodedNode *Node) { 54 while (Node) { 55 const ProgramPoint &PP = Node->getLocation(); 56 // Skip any BlockEdges. 57 if (isa<BlockEdge>(PP) || isa<CallExit>(PP)) { 58 assert(Node->pred_size() == 1); 59 Node = *Node->pred_begin(); 60 continue; 61 } 62 if (const StmtPoint *SP = dyn_cast<StmtPoint>(&PP)) { 63 const Stmt *S = SP->getStmt(); 64 return dyn_cast<ReturnStmt>(S); 65 } 66 break; 67 } 68 return 0; 69} 70 71void ExprEngine::processCallExit(ExplodedNode *Pred) { 72 ProgramStateRef state = Pred->getState(); 73 const StackFrameContext *calleeCtx = 74 Pred->getLocationContext()->getCurrentStackFrame(); 75 const LocationContext *callerCtx = calleeCtx->getParent(); 76 const Stmt *CE = calleeCtx->getCallSite(); 77 78 // If the callee returns an expression, bind its value to CallExpr. 79 if (const ReturnStmt *RS = getReturnStmt(Pred)) { 80 const LocationContext *LCtx = Pred->getLocationContext(); 81 SVal V = state->getSVal(RS, LCtx); 82 state = state->BindExpr(CE, callerCtx, V); 83 } 84 85 // Bind the constructed object value to CXXConstructExpr. 86 if (const CXXConstructExpr *CCE = dyn_cast<CXXConstructExpr>(CE)) { 87 const CXXThisRegion *ThisR = 88 getCXXThisRegion(CCE->getConstructor()->getParent(), calleeCtx); 89 90 SVal ThisV = state->getSVal(ThisR); 91 // Always bind the region to the CXXConstructExpr. 92 state = state->BindExpr(CCE, Pred->getLocationContext(), ThisV); 93 } 94 95 static SimpleProgramPointTag returnTag("ExprEngine : Call Return"); 96 PostStmt Loc(CE, callerCtx, &returnTag); 97 bool isNew; 98 ExplodedNode *N = G.getNode(Loc, state, false, &isNew); 99 N->addPredecessor(Pred, G); 100 if (!isNew) 101 return; 102 103 // Perform the post-condition check of the CallExpr. 104 ExplodedNodeSet Dst; 105 NodeBuilderContext Ctx(Engine, calleeCtx->getCallSiteBlock(), N); 106 SaveAndRestore<const NodeBuilderContext*> NBCSave(currentBuilderContext, 107 &Ctx); 108 SaveAndRestore<unsigned> CBISave(currentStmtIdx, calleeCtx->getIndex()); 109 110 getCheckerManager().runCheckersForPostStmt(Dst, N, CE, *this); 111 112 // Enqueue the next element in the block. 113 for (ExplodedNodeSet::iterator I = Dst.begin(), E = Dst.end(); I != E; ++I) { 114 Engine.getWorkList()->enqueue(*I, 115 calleeCtx->getCallSiteBlock(), 116 calleeCtx->getIndex()+1); 117 } 118} 119 120static unsigned getNumberStackFrames(const LocationContext *LCtx) { 121 unsigned count = 0; 122 while (LCtx) { 123 if (isa<StackFrameContext>(LCtx)) 124 ++count; 125 LCtx = LCtx->getParent(); 126 } 127 return count; 128} 129 130bool ExprEngine::InlineCall(ExplodedNodeSet &Dst, 131 const CallExpr *CE, 132 ExplodedNode *Pred) { 133 ProgramStateRef state = Pred->getState(); 134 const Expr *Callee = CE->getCallee(); 135 const FunctionDecl *FD = 136 state->getSVal(Callee, Pred->getLocationContext()).getAsFunctionDecl(); 137 if (!FD || !FD->hasBody(FD)) 138 return false; 139 140 switch (CE->getStmtClass()) { 141 default: 142 // FIXME: Handle C++. 143 break; 144 case Stmt::CallExprClass: { 145 if (getNumberStackFrames(Pred->getLocationContext()) 146 == AMgr.InlineMaxStackDepth) 147 return false; 148 149 AnalysisDeclContext *CalleeADC = AMgr.getAnalysisDeclContext(FD); 150 const CFG *CalleeCFG = CalleeADC->getCFG(); 151 if (CalleeCFG->getNumBlockIDs() > AMgr.InlineMaxFunctionSize) 152 return false; 153 154 // Construct a new stack frame for the callee. 155 const StackFrameContext *CallerSFC = 156 Pred->getLocationContext()->getCurrentStackFrame(); 157 const StackFrameContext *CalleeSFC = 158 CalleeADC->getStackFrame(CallerSFC, CE, 159 currentBuilderContext->getBlock(), 160 currentStmtIdx); 161 162 CallEnter Loc(CE, CalleeSFC, Pred->getLocationContext()); 163 bool isNew; 164 ExplodedNode *N = G.getNode(Loc, state, false, &isNew); 165 N->addPredecessor(Pred, G); 166 if (isNew) 167 Engine.getWorkList()->enqueue(N); 168 return true; 169 } 170 } 171 return false; 172} 173 174static bool isPointerToConst(const ParmVarDecl *ParamDecl) { 175 QualType PointeeTy = ParamDecl->getOriginalType()->getPointeeType(); 176 if (PointeeTy != QualType() && PointeeTy.isConstQualified() && 177 !PointeeTy->isAnyPointerType() && !PointeeTy->isReferenceType()) { 178 return true; 179 } 180 return false; 181} 182 183// Try to retrieve the function declaration and find the function parameter 184// types which are pointers/references to a non-pointer const. 185// We do not invalidate the corresponding argument regions. 186static void findPtrToConstParams(llvm::SmallSet<unsigned, 1> &PreserveArgs, 187 const CallOrObjCMessage &Call) { 188 const Decl *CallDecl = Call.getDecl(); 189 if (!CallDecl) 190 return; 191 192 if (const FunctionDecl *FDecl = dyn_cast<FunctionDecl>(CallDecl)) { 193 const IdentifierInfo *II = FDecl->getIdentifier(); 194 195 // List the cases, where the region should be invalidated even if the 196 // argument is const. 197 if (II) { 198 StringRef FName = II->getName(); 199 // - 'int pthread_setspecific(ptheread_key k, const void *)' stores a 200 // value into thread local storage. The value can later be retrieved with 201 // 'void *ptheread_getspecific(pthread_key)'. So even thought the 202 // parameter is 'const void *', the region escapes through the call. 203 // - funopen - sets a buffer for future IO calls. 204 // - ObjC functions that end with "NoCopy" can free memory, of the passed 205 // in buffer. 206 // - Many CF containers allow objects to escape through custom 207 // allocators/deallocators upon container construction. 208 if (FName == "pthread_setspecific" || 209 FName == "funopen" || 210 FName.endswith("NoCopy") || 211 Call.isCFCGAllowingEscape(FName)) 212 return; 213 } 214 215 for (unsigned Idx = 0, E = Call.getNumArgs(); Idx != E; ++Idx) { 216 if (FDecl && Idx < FDecl->getNumParams()) { 217 if (isPointerToConst(FDecl->getParamDecl(Idx))) 218 PreserveArgs.insert(Idx); 219 } 220 } 221 return; 222 } 223 224 if (const ObjCMethodDecl *MDecl = dyn_cast<ObjCMethodDecl>(CallDecl)) { 225 assert(MDecl->param_size() <= Call.getNumArgs()); 226 unsigned Idx = 0; 227 for (clang::ObjCMethodDecl::param_const_iterator 228 I = MDecl->param_begin(), E = MDecl->param_end(); I != E; ++I, ++Idx) { 229 if (isPointerToConst(*I)) 230 PreserveArgs.insert(Idx); 231 } 232 return; 233 } 234} 235 236ProgramStateRef 237ExprEngine::invalidateArguments(ProgramStateRef State, 238 const CallOrObjCMessage &Call, 239 const LocationContext *LC) { 240 SmallVector<const MemRegion *, 8> RegionsToInvalidate; 241 242 if (Call.isObjCMessage()) { 243 // Invalidate all instance variables of the receiver of an ObjC message. 244 // FIXME: We should be able to do better with inter-procedural analysis. 245 if (const MemRegion *MR = Call.getInstanceMessageReceiver(LC).getAsRegion()) 246 RegionsToInvalidate.push_back(MR); 247 248 } else if (Call.isCXXCall()) { 249 // Invalidate all instance variables for the callee of a C++ method call. 250 // FIXME: We should be able to do better with inter-procedural analysis. 251 // FIXME: We can probably do better for const versus non-const methods. 252 if (const MemRegion *Callee = Call.getCXXCallee().getAsRegion()) 253 RegionsToInvalidate.push_back(Callee); 254 255 } else if (Call.isFunctionCall()) { 256 // Block calls invalidate all captured-by-reference values. 257 SVal CalleeVal = Call.getFunctionCallee(); 258 if (const MemRegion *Callee = CalleeVal.getAsRegion()) { 259 if (isa<BlockDataRegion>(Callee)) 260 RegionsToInvalidate.push_back(Callee); 261 } 262 } 263 264 // Indexes of arguments whose values will be preserved by the call. 265 llvm::SmallSet<unsigned, 1> PreserveArgs; 266 findPtrToConstParams(PreserveArgs, Call); 267 268 for (unsigned idx = 0, e = Call.getNumArgs(); idx != e; ++idx) { 269 if (PreserveArgs.count(idx)) 270 continue; 271 272 SVal V = Call.getArgSVal(idx); 273 274 // If we are passing a location wrapped as an integer, unwrap it and 275 // invalidate the values referred by the location. 276 if (nonloc::LocAsInteger *Wrapped = dyn_cast<nonloc::LocAsInteger>(&V)) 277 V = Wrapped->getLoc(); 278 else if (!isa<Loc>(V)) 279 continue; 280 281 if (const MemRegion *R = V.getAsRegion()) { 282 // Invalidate the value of the variable passed by reference. 283 284 // Are we dealing with an ElementRegion? If the element type is 285 // a basic integer type (e.g., char, int) and the underlying region 286 // is a variable region then strip off the ElementRegion. 287 // FIXME: We really need to think about this for the general case 288 // as sometimes we are reasoning about arrays and other times 289 // about (char*), etc., is just a form of passing raw bytes. 290 // e.g., void *p = alloca(); foo((char*)p); 291 if (const ElementRegion *ER = dyn_cast<ElementRegion>(R)) { 292 // Checking for 'integral type' is probably too promiscuous, but 293 // we'll leave it in for now until we have a systematic way of 294 // handling all of these cases. Eventually we need to come up 295 // with an interface to StoreManager so that this logic can be 296 // appropriately delegated to the respective StoreManagers while 297 // still allowing us to do checker-specific logic (e.g., 298 // invalidating reference counts), probably via callbacks. 299 if (ER->getElementType()->isIntegralOrEnumerationType()) { 300 const MemRegion *superReg = ER->getSuperRegion(); 301 if (isa<VarRegion>(superReg) || isa<FieldRegion>(superReg) || 302 isa<ObjCIvarRegion>(superReg)) 303 R = cast<TypedRegion>(superReg); 304 } 305 // FIXME: What about layers of ElementRegions? 306 } 307 308 // Mark this region for invalidation. We batch invalidate regions 309 // below for efficiency. 310 RegionsToInvalidate.push_back(R); 311 } else { 312 // Nuke all other arguments passed by reference. 313 // FIXME: is this necessary or correct? This handles the non-Region 314 // cases. Is it ever valid to store to these? 315 State = State->unbindLoc(cast<Loc>(V)); 316 } 317 } 318 319 // Invalidate designated regions using the batch invalidation API. 320 321 // FIXME: We can have collisions on the conjured symbol if the 322 // expression *I also creates conjured symbols. We probably want 323 // to identify conjured symbols by an expression pair: the enclosing 324 // expression (the context) and the expression itself. This should 325 // disambiguate conjured symbols. 326 unsigned Count = currentBuilderContext->getCurrentBlockCount(); 327 StoreManager::InvalidatedSymbols IS; 328 329 // NOTE: Even if RegionsToInvalidate is empty, we may still invalidate 330 // global variables. 331 return State->invalidateRegions(RegionsToInvalidate, 332 Call.getOriginExpr(), Count, LC, 333 &IS, &Call); 334 335} 336 337// For now, skip inlining variadic functions. 338// We also don't inline blocks. 339static bool shouldInlineCall(const CallExpr *CE, ExprEngine &Eng) { 340 if (!Eng.getAnalysisManager().shouldInlineCall()) 341 return false; 342 QualType callee = CE->getCallee()->getType(); 343 const FunctionProtoType *FT = 0; 344 if (const PointerType *PT = callee->getAs<PointerType>()) 345 FT = dyn_cast<FunctionProtoType>(PT->getPointeeType()); 346 else if (const BlockPointerType *BT = callee->getAs<BlockPointerType>()) { 347 // FIXME: inline blocks. 348 // FT = dyn_cast<FunctionProtoType>(BT->getPointeeType()); 349 (void) BT; 350 return false; 351 } 352 353 // If we have no prototype, assume the function is okay. 354 if (!FT) 355 return true; 356 357 // Skip inlining of variadic functions. 358 return !FT->isVariadic(); 359} 360 361void ExprEngine::VisitCallExpr(const CallExpr *CE, ExplodedNode *Pred, 362 ExplodedNodeSet &dst) { 363 // Perform the previsit of the CallExpr. 364 ExplodedNodeSet dstPreVisit; 365 getCheckerManager().runCheckersForPreStmt(dstPreVisit, Pred, CE, *this); 366 367 // Now evaluate the call itself. 368 class DefaultEval : public GraphExpander { 369 ExprEngine &Eng; 370 const CallExpr *CE; 371 public: 372 373 DefaultEval(ExprEngine &eng, const CallExpr *ce) 374 : Eng(eng), CE(ce) {} 375 virtual void expandGraph(ExplodedNodeSet &Dst, ExplodedNode *Pred) { 376 // Should we inline the call? 377 if (shouldInlineCall(CE, Eng) && 378 Eng.InlineCall(Dst, CE, Pred)) { 379 return; 380 } 381 382 // First handle the return value. 383 StmtNodeBuilder Bldr(Pred, Dst, *Eng.currentBuilderContext); 384 385 // Get the callee. 386 const Expr *Callee = CE->getCallee()->IgnoreParens(); 387 ProgramStateRef state = Pred->getState(); 388 SVal L = state->getSVal(Callee, Pred->getLocationContext()); 389 390 // Figure out the result type. We do this dance to handle references. 391 QualType ResultTy; 392 if (const FunctionDecl *FD = L.getAsFunctionDecl()) 393 ResultTy = FD->getResultType(); 394 else 395 ResultTy = CE->getType(); 396 397 if (CE->isLValue()) 398 ResultTy = Eng.getContext().getPointerType(ResultTy); 399 400 // Conjure a symbol value to use as the result. 401 SValBuilder &SVB = Eng.getSValBuilder(); 402 unsigned Count = Eng.currentBuilderContext->getCurrentBlockCount(); 403 const LocationContext *LCtx = Pred->getLocationContext(); 404 SVal RetVal = SVB.getConjuredSymbolVal(0, CE, LCtx, ResultTy, Count); 405 406 // Generate a new state with the return value set. 407 state = state->BindExpr(CE, LCtx, RetVal); 408 409 // Invalidate the arguments. 410 state = Eng.invalidateArguments(state, CallOrObjCMessage(CE, state, LCtx), 411 LCtx); 412 413 // And make the result node. 414 Bldr.generateNode(CE, Pred, state); 415 } 416 }; 417 418 // Finally, evaluate the function call. We try each of the checkers 419 // to see if the can evaluate the function call. 420 ExplodedNodeSet dstCallEvaluated; 421 DefaultEval defEval(*this, CE); 422 getCheckerManager().runCheckersForEvalCall(dstCallEvaluated, 423 dstPreVisit, 424 CE, *this, &defEval); 425 426 // Finally, perform the post-condition check of the CallExpr and store 427 // the created nodes in 'Dst'. 428 getCheckerManager().runCheckersForPostStmt(dst, dstCallEvaluated, CE, 429 *this); 430} 431 432void ExprEngine::VisitReturnStmt(const ReturnStmt *RS, ExplodedNode *Pred, 433 ExplodedNodeSet &Dst) { 434 435 ExplodedNodeSet dstPreVisit; 436 getCheckerManager().runCheckersForPreStmt(dstPreVisit, Pred, RS, *this); 437 438 StmtNodeBuilder B(dstPreVisit, Dst, *currentBuilderContext); 439 440 if (RS->getRetValue()) { 441 for (ExplodedNodeSet::iterator it = dstPreVisit.begin(), 442 ei = dstPreVisit.end(); it != ei; ++it) { 443 B.generateNode(RS, *it, (*it)->getState()); 444 } 445 } 446} 447