AnalysisBasedWarnings.cpp revision b4d0a9678f8c592990593233e64c59247f40a74a
1//=- AnalysisBasedWarnings.cpp - Sema warnings based on libAnalysis -*- 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 analysis_warnings::[Policy,Executor]. 11// Together they are used by Sema to issue warnings based on inexpensive 12// static analysis algorithms in libAnalysis. 13// 14//===----------------------------------------------------------------------===// 15 16#include "clang/Sema/AnalysisBasedWarnings.h" 17#include "clang/Sema/SemaInternal.h" 18#include "clang/Sema/ScopeInfo.h" 19#include "clang/Basic/SourceManager.h" 20#include "clang/Lex/Preprocessor.h" 21#include "clang/AST/DeclObjC.h" 22#include "clang/AST/DeclCXX.h" 23#include "clang/AST/ExprObjC.h" 24#include "clang/AST/ExprCXX.h" 25#include "clang/AST/StmtObjC.h" 26#include "clang/AST/StmtCXX.h" 27#include "clang/AST/EvaluatedExprVisitor.h" 28#include "clang/AST/StmtVisitor.h" 29#include "clang/Analysis/AnalysisContext.h" 30#include "clang/Analysis/CFG.h" 31#include "clang/Analysis/Analyses/ReachableCode.h" 32#include "clang/Analysis/Analyses/CFGReachabilityAnalysis.h" 33#include "clang/Analysis/CFGStmtMap.h" 34#include "clang/Analysis/Analyses/UninitializedValues.h" 35#include "llvm/ADT/BitVector.h" 36#include "llvm/ADT/FoldingSet.h" 37#include "llvm/ADT/ImmutableMap.h" 38#include "llvm/ADT/PostOrderIterator.h" 39#include "llvm/ADT/SmallVector.h" 40#include "llvm/Support/Casting.h" 41#include <algorithm> 42#include <vector> 43 44using namespace clang; 45 46//===----------------------------------------------------------------------===// 47// Unreachable code analysis. 48//===----------------------------------------------------------------------===// 49 50namespace { 51 class UnreachableCodeHandler : public reachable_code::Callback { 52 Sema &S; 53 public: 54 UnreachableCodeHandler(Sema &s) : S(s) {} 55 56 void HandleUnreachable(SourceLocation L, SourceRange R1, SourceRange R2) { 57 S.Diag(L, diag::warn_unreachable) << R1 << R2; 58 } 59 }; 60} 61 62/// CheckUnreachable - Check for unreachable code. 63static void CheckUnreachable(Sema &S, AnalysisContext &AC) { 64 UnreachableCodeHandler UC(S); 65 reachable_code::FindUnreachableCode(AC, UC); 66} 67 68//===----------------------------------------------------------------------===// 69// Check for missing return value. 70//===----------------------------------------------------------------------===// 71 72enum ControlFlowKind { 73 UnknownFallThrough, 74 NeverFallThrough, 75 MaybeFallThrough, 76 AlwaysFallThrough, 77 NeverFallThroughOrReturn 78}; 79 80/// CheckFallThrough - Check that we don't fall off the end of a 81/// Statement that should return a value. 82/// 83/// \returns AlwaysFallThrough iff we always fall off the end of the statement, 84/// MaybeFallThrough iff we might or might not fall off the end, 85/// NeverFallThroughOrReturn iff we never fall off the end of the statement or 86/// return. We assume NeverFallThrough iff we never fall off the end of the 87/// statement but we may return. We assume that functions not marked noreturn 88/// will return. 89static ControlFlowKind CheckFallThrough(AnalysisContext &AC) { 90 CFG *cfg = AC.getCFG(); 91 if (cfg == 0) return UnknownFallThrough; 92 93 // The CFG leaves in dead things, and we don't want the dead code paths to 94 // confuse us, so we mark all live things first. 95 llvm::BitVector live(cfg->getNumBlockIDs()); 96 unsigned count = reachable_code::ScanReachableFromBlock(&cfg->getEntry(), 97 live); 98 99 bool AddEHEdges = AC.getAddEHEdges(); 100 if (!AddEHEdges && count != cfg->getNumBlockIDs()) 101 // When there are things remaining dead, and we didn't add EH edges 102 // from CallExprs to the catch clauses, we have to go back and 103 // mark them as live. 104 for (CFG::iterator I = cfg->begin(), E = cfg->end(); I != E; ++I) { 105 CFGBlock &b = **I; 106 if (!live[b.getBlockID()]) { 107 if (b.pred_begin() == b.pred_end()) { 108 if (b.getTerminator() && isa<CXXTryStmt>(b.getTerminator())) 109 // When not adding EH edges from calls, catch clauses 110 // can otherwise seem dead. Avoid noting them as dead. 111 count += reachable_code::ScanReachableFromBlock(&b, live); 112 continue; 113 } 114 } 115 } 116 117 // Now we know what is live, we check the live precessors of the exit block 118 // and look for fall through paths, being careful to ignore normal returns, 119 // and exceptional paths. 120 bool HasLiveReturn = false; 121 bool HasFakeEdge = false; 122 bool HasPlainEdge = false; 123 bool HasAbnormalEdge = false; 124 125 // Ignore default cases that aren't likely to be reachable because all 126 // enums in a switch(X) have explicit case statements. 127 CFGBlock::FilterOptions FO; 128 FO.IgnoreDefaultsWithCoveredEnums = 1; 129 130 for (CFGBlock::filtered_pred_iterator 131 I = cfg->getExit().filtered_pred_start_end(FO); I.hasMore(); ++I) { 132 const CFGBlock& B = **I; 133 if (!live[B.getBlockID()]) 134 continue; 135 136 // Destructors can appear after the 'return' in the CFG. This is 137 // normal. We need to look pass the destructors for the return 138 // statement (if it exists). 139 CFGBlock::const_reverse_iterator ri = B.rbegin(), re = B.rend(); 140 bool hasNoReturnDtor = false; 141 142 for ( ; ri != re ; ++ri) { 143 CFGElement CE = *ri; 144 145 // FIXME: The right solution is to just sever the edges in the 146 // CFG itself. 147 if (const CFGImplicitDtor *iDtor = ri->getAs<CFGImplicitDtor>()) 148 if (iDtor->isNoReturn(AC.getASTContext())) { 149 hasNoReturnDtor = true; 150 HasFakeEdge = true; 151 break; 152 } 153 154 if (isa<CFGStmt>(CE)) 155 break; 156 } 157 158 if (hasNoReturnDtor) 159 continue; 160 161 // No more CFGElements in the block? 162 if (ri == re) { 163 if (B.getTerminator() && isa<CXXTryStmt>(B.getTerminator())) { 164 HasAbnormalEdge = true; 165 continue; 166 } 167 // A labeled empty statement, or the entry block... 168 HasPlainEdge = true; 169 continue; 170 } 171 172 CFGStmt CS = cast<CFGStmt>(*ri); 173 const Stmt *S = CS.getStmt(); 174 if (isa<ReturnStmt>(S)) { 175 HasLiveReturn = true; 176 continue; 177 } 178 if (isa<ObjCAtThrowStmt>(S)) { 179 HasFakeEdge = true; 180 continue; 181 } 182 if (isa<CXXThrowExpr>(S)) { 183 HasFakeEdge = true; 184 continue; 185 } 186 if (const AsmStmt *AS = dyn_cast<AsmStmt>(S)) { 187 if (AS->isMSAsm()) { 188 HasFakeEdge = true; 189 HasLiveReturn = true; 190 continue; 191 } 192 } 193 if (isa<CXXTryStmt>(S)) { 194 HasAbnormalEdge = true; 195 continue; 196 } 197 198 bool NoReturnEdge = false; 199 if (const CallExpr *C = dyn_cast<CallExpr>(S)) { 200 if (std::find(B.succ_begin(), B.succ_end(), &cfg->getExit()) 201 == B.succ_end()) { 202 HasAbnormalEdge = true; 203 continue; 204 } 205 const Expr *CEE = C->getCallee()->IgnoreParenCasts(); 206 QualType calleeType = CEE->getType(); 207 if (calleeType == AC.getASTContext().BoundMemberTy) { 208 calleeType = Expr::findBoundMemberType(CEE); 209 assert(!calleeType.isNull() && "analyzing unresolved call?"); 210 } 211 if (getFunctionExtInfo(calleeType).getNoReturn()) { 212 NoReturnEdge = true; 213 HasFakeEdge = true; 214 } else if (const DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(CEE)) { 215 const ValueDecl *VD = DRE->getDecl(); 216 if (VD->hasAttr<NoReturnAttr>()) { 217 NoReturnEdge = true; 218 HasFakeEdge = true; 219 } 220 } 221 } 222 // FIXME: Add noreturn message sends. 223 if (NoReturnEdge == false) 224 HasPlainEdge = true; 225 } 226 if (!HasPlainEdge) { 227 if (HasLiveReturn) 228 return NeverFallThrough; 229 return NeverFallThroughOrReturn; 230 } 231 if (HasAbnormalEdge || HasFakeEdge || HasLiveReturn) 232 return MaybeFallThrough; 233 // This says AlwaysFallThrough for calls to functions that are not marked 234 // noreturn, that don't return. If people would like this warning to be more 235 // accurate, such functions should be marked as noreturn. 236 return AlwaysFallThrough; 237} 238 239namespace { 240 241struct CheckFallThroughDiagnostics { 242 unsigned diag_MaybeFallThrough_HasNoReturn; 243 unsigned diag_MaybeFallThrough_ReturnsNonVoid; 244 unsigned diag_AlwaysFallThrough_HasNoReturn; 245 unsigned diag_AlwaysFallThrough_ReturnsNonVoid; 246 unsigned diag_NeverFallThroughOrReturn; 247 bool funMode; 248 SourceLocation FuncLoc; 249 250 static CheckFallThroughDiagnostics MakeForFunction(const Decl *Func) { 251 CheckFallThroughDiagnostics D; 252 D.FuncLoc = Func->getLocation(); 253 D.diag_MaybeFallThrough_HasNoReturn = 254 diag::warn_falloff_noreturn_function; 255 D.diag_MaybeFallThrough_ReturnsNonVoid = 256 diag::warn_maybe_falloff_nonvoid_function; 257 D.diag_AlwaysFallThrough_HasNoReturn = 258 diag::warn_falloff_noreturn_function; 259 D.diag_AlwaysFallThrough_ReturnsNonVoid = 260 diag::warn_falloff_nonvoid_function; 261 262 // Don't suggest that virtual functions be marked "noreturn", since they 263 // might be overridden by non-noreturn functions. 264 bool isVirtualMethod = false; 265 if (const CXXMethodDecl *Method = dyn_cast<CXXMethodDecl>(Func)) 266 isVirtualMethod = Method->isVirtual(); 267 268 if (!isVirtualMethod) 269 D.diag_NeverFallThroughOrReturn = 270 diag::warn_suggest_noreturn_function; 271 else 272 D.diag_NeverFallThroughOrReturn = 0; 273 274 D.funMode = true; 275 return D; 276 } 277 278 static CheckFallThroughDiagnostics MakeForBlock() { 279 CheckFallThroughDiagnostics D; 280 D.diag_MaybeFallThrough_HasNoReturn = 281 diag::err_noreturn_block_has_return_expr; 282 D.diag_MaybeFallThrough_ReturnsNonVoid = 283 diag::err_maybe_falloff_nonvoid_block; 284 D.diag_AlwaysFallThrough_HasNoReturn = 285 diag::err_noreturn_block_has_return_expr; 286 D.diag_AlwaysFallThrough_ReturnsNonVoid = 287 diag::err_falloff_nonvoid_block; 288 D.diag_NeverFallThroughOrReturn = 289 diag::warn_suggest_noreturn_block; 290 D.funMode = false; 291 return D; 292 } 293 294 bool checkDiagnostics(Diagnostic &D, bool ReturnsVoid, 295 bool HasNoReturn) const { 296 if (funMode) { 297 return (ReturnsVoid || 298 D.getDiagnosticLevel(diag::warn_maybe_falloff_nonvoid_function, 299 FuncLoc) == Diagnostic::Ignored) 300 && (!HasNoReturn || 301 D.getDiagnosticLevel(diag::warn_noreturn_function_has_return_expr, 302 FuncLoc) == Diagnostic::Ignored) 303 && (!ReturnsVoid || 304 D.getDiagnosticLevel(diag::warn_suggest_noreturn_block, FuncLoc) 305 == Diagnostic::Ignored); 306 } 307 308 // For blocks. 309 return ReturnsVoid && !HasNoReturn 310 && (!ReturnsVoid || 311 D.getDiagnosticLevel(diag::warn_suggest_noreturn_block, FuncLoc) 312 == Diagnostic::Ignored); 313 } 314}; 315 316} 317 318/// CheckFallThroughForFunctionDef - Check that we don't fall off the end of a 319/// function that should return a value. Check that we don't fall off the end 320/// of a noreturn function. We assume that functions and blocks not marked 321/// noreturn will return. 322static void CheckFallThroughForBody(Sema &S, const Decl *D, const Stmt *Body, 323 const BlockExpr *blkExpr, 324 const CheckFallThroughDiagnostics& CD, 325 AnalysisContext &AC) { 326 327 bool ReturnsVoid = false; 328 bool HasNoReturn = false; 329 330 if (const FunctionDecl *FD = dyn_cast<FunctionDecl>(D)) { 331 ReturnsVoid = FD->getResultType()->isVoidType(); 332 HasNoReturn = FD->hasAttr<NoReturnAttr>() || 333 FD->getType()->getAs<FunctionType>()->getNoReturnAttr(); 334 } 335 else if (const ObjCMethodDecl *MD = dyn_cast<ObjCMethodDecl>(D)) { 336 ReturnsVoid = MD->getResultType()->isVoidType(); 337 HasNoReturn = MD->hasAttr<NoReturnAttr>(); 338 } 339 else if (isa<BlockDecl>(D)) { 340 QualType BlockTy = blkExpr->getType(); 341 if (const FunctionType *FT = 342 BlockTy->getPointeeType()->getAs<FunctionType>()) { 343 if (FT->getResultType()->isVoidType()) 344 ReturnsVoid = true; 345 if (FT->getNoReturnAttr()) 346 HasNoReturn = true; 347 } 348 } 349 350 Diagnostic &Diags = S.getDiagnostics(); 351 352 // Short circuit for compilation speed. 353 if (CD.checkDiagnostics(Diags, ReturnsVoid, HasNoReturn)) 354 return; 355 356 // FIXME: Function try block 357 if (const CompoundStmt *Compound = dyn_cast<CompoundStmt>(Body)) { 358 switch (CheckFallThrough(AC)) { 359 case UnknownFallThrough: 360 break; 361 362 case MaybeFallThrough: 363 if (HasNoReturn) 364 S.Diag(Compound->getRBracLoc(), 365 CD.diag_MaybeFallThrough_HasNoReturn); 366 else if (!ReturnsVoid) 367 S.Diag(Compound->getRBracLoc(), 368 CD.diag_MaybeFallThrough_ReturnsNonVoid); 369 break; 370 case AlwaysFallThrough: 371 if (HasNoReturn) 372 S.Diag(Compound->getRBracLoc(), 373 CD.diag_AlwaysFallThrough_HasNoReturn); 374 else if (!ReturnsVoid) 375 S.Diag(Compound->getRBracLoc(), 376 CD.diag_AlwaysFallThrough_ReturnsNonVoid); 377 break; 378 case NeverFallThroughOrReturn: 379 if (ReturnsVoid && !HasNoReturn && CD.diag_NeverFallThroughOrReturn) 380 S.Diag(Compound->getLBracLoc(), 381 CD.diag_NeverFallThroughOrReturn); 382 break; 383 case NeverFallThrough: 384 break; 385 } 386 } 387} 388 389//===----------------------------------------------------------------------===// 390// -Wuninitialized 391//===----------------------------------------------------------------------===// 392 393namespace { 394/// ContainsReference - A visitor class to search for references to 395/// a particular declaration (the needle) within any evaluated component of an 396/// expression (recursively). 397class ContainsReference : public EvaluatedExprVisitor<ContainsReference> { 398 bool FoundReference; 399 const DeclRefExpr *Needle; 400 401public: 402 ContainsReference(ASTContext &Context, const DeclRefExpr *Needle) 403 : EvaluatedExprVisitor<ContainsReference>(Context), 404 FoundReference(false), Needle(Needle) {} 405 406 void VisitExpr(Expr *E) { 407 // Stop evaluating if we already have a reference. 408 if (FoundReference) 409 return; 410 411 EvaluatedExprVisitor<ContainsReference>::VisitExpr(E); 412 } 413 414 void VisitDeclRefExpr(DeclRefExpr *E) { 415 if (E == Needle) 416 FoundReference = true; 417 else 418 EvaluatedExprVisitor<ContainsReference>::VisitDeclRefExpr(E); 419 } 420 421 bool doesContainReference() const { return FoundReference; } 422}; 423} 424 425/// DiagnoseUninitializedUse -- Helper function for diagnosing uses of an 426/// uninitialized variable. This manages the different forms of diagnostic 427/// emitted for particular types of uses. Returns true if the use was diagnosed 428/// as a warning. If a pariticular use is one we omit warnings for, returns 429/// false. 430static bool DiagnoseUninitializedUse(Sema &S, const VarDecl *VD, 431 const Expr *E, bool isAlwaysUninit) { 432 bool isSelfInit = false; 433 434 if (const DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(E)) { 435 if (isAlwaysUninit) { 436 // Inspect the initializer of the variable declaration which is 437 // being referenced prior to its initialization. We emit 438 // specialized diagnostics for self-initialization, and we 439 // specifically avoid warning about self references which take the 440 // form of: 441 // 442 // int x = x; 443 // 444 // This is used to indicate to GCC that 'x' is intentionally left 445 // uninitialized. Proven code paths which access 'x' in 446 // an uninitialized state after this will still warn. 447 // 448 // TODO: Should we suppress maybe-uninitialized warnings for 449 // variables initialized in this way? 450 if (const Expr *Initializer = VD->getInit()) { 451 if (DRE == Initializer->IgnoreParenImpCasts()) 452 return false; 453 454 ContainsReference CR(S.Context, DRE); 455 CR.Visit(const_cast<Expr*>(Initializer)); 456 isSelfInit = CR.doesContainReference(); 457 } 458 if (isSelfInit) { 459 S.Diag(DRE->getLocStart(), 460 diag::warn_uninit_self_reference_in_init) 461 << VD->getDeclName() << VD->getLocation() << DRE->getSourceRange(); 462 } else { 463 S.Diag(DRE->getLocStart(), diag::warn_uninit_var) 464 << VD->getDeclName() << DRE->getSourceRange(); 465 } 466 } else { 467 S.Diag(DRE->getLocStart(), diag::warn_maybe_uninit_var) 468 << VD->getDeclName() << DRE->getSourceRange(); 469 } 470 } else { 471 const BlockExpr *BE = cast<BlockExpr>(E); 472 S.Diag(BE->getLocStart(), 473 isAlwaysUninit ? diag::warn_uninit_var_captured_by_block 474 : diag::warn_maybe_uninit_var_captured_by_block) 475 << VD->getDeclName(); 476 } 477 478 // Report where the variable was declared when the use wasn't within 479 // the initializer of that declaration. 480 if (!isSelfInit) 481 S.Diag(VD->getLocStart(), diag::note_uninit_var_def) 482 << VD->getDeclName(); 483 484 return true; 485} 486 487static void SuggestInitializationFixit(Sema &S, const VarDecl *VD) { 488 // Don't issue a fixit if there is already an initializer. 489 if (VD->getInit()) 490 return; 491 492 // Suggest possible initialization (if any). 493 const char *initialization = 0; 494 QualType VariableTy = VD->getType().getCanonicalType(); 495 496 if (VariableTy->isObjCObjectPointerType() || 497 VariableTy->isBlockPointerType()) { 498 // Check if 'nil' is defined. 499 if (S.PP.getMacroInfo(&S.getASTContext().Idents.get("nil"))) 500 initialization = " = nil"; 501 else 502 initialization = " = 0"; 503 } 504 else if (VariableTy->isRealFloatingType()) 505 initialization = " = 0.0"; 506 else if (VariableTy->isBooleanType() && S.Context.getLangOptions().CPlusPlus) 507 initialization = " = false"; 508 else if (VariableTy->isEnumeralType()) 509 return; 510 else if (VariableTy->isPointerType() || VariableTy->isMemberPointerType()) { 511 if (S.Context.getLangOptions().CPlusPlus0x) 512 initialization = " = nullptr"; 513 // Check if 'NULL' is defined. 514 else if (S.PP.getMacroInfo(&S.getASTContext().Idents.get("NULL"))) 515 initialization = " = NULL"; 516 else 517 initialization = " = 0"; 518 } 519 else if (VariableTy->isScalarType()) 520 initialization = " = 0"; 521 522 if (initialization) { 523 SourceLocation loc = S.PP.getLocForEndOfToken(VD->getLocEnd()); 524 S.Diag(loc, diag::note_var_fixit_add_initialization) 525 << FixItHint::CreateInsertion(loc, initialization); 526 } 527} 528 529typedef std::pair<const Expr*, bool> UninitUse; 530 531namespace { 532struct SLocSort { 533 bool operator()(const UninitUse &a, const UninitUse &b) { 534 SourceLocation aLoc = a.first->getLocStart(); 535 SourceLocation bLoc = b.first->getLocStart(); 536 return aLoc.getRawEncoding() < bLoc.getRawEncoding(); 537 } 538}; 539 540class UninitValsDiagReporter : public UninitVariablesHandler { 541 Sema &S; 542 typedef SmallVector<UninitUse, 2> UsesVec; 543 typedef llvm::DenseMap<const VarDecl *, UsesVec*> UsesMap; 544 UsesMap *uses; 545 546public: 547 UninitValsDiagReporter(Sema &S) : S(S), uses(0) {} 548 ~UninitValsDiagReporter() { 549 flushDiagnostics(); 550 } 551 552 void handleUseOfUninitVariable(const Expr *ex, const VarDecl *vd, 553 bool isAlwaysUninit) { 554 if (!uses) 555 uses = new UsesMap(); 556 557 UsesVec *&vec = (*uses)[vd]; 558 if (!vec) 559 vec = new UsesVec(); 560 561 vec->push_back(std::make_pair(ex, isAlwaysUninit)); 562 } 563 564 void flushDiagnostics() { 565 if (!uses) 566 return; 567 568 for (UsesMap::iterator i = uses->begin(), e = uses->end(); i != e; ++i) { 569 const VarDecl *vd = i->first; 570 UsesVec *vec = i->second; 571 572 // Sort the uses by their SourceLocations. While not strictly 573 // guaranteed to produce them in line/column order, this will provide 574 // a stable ordering. 575 std::sort(vec->begin(), vec->end(), SLocSort()); 576 577 for (UsesVec::iterator vi = vec->begin(), ve = vec->end(); vi != ve; 578 ++vi) { 579 if (!DiagnoseUninitializedUse(S, vd, vi->first, 580 /*isAlwaysUninit=*/vi->second)) 581 continue; 582 583 SuggestInitializationFixit(S, vd); 584 585 // Skip further diagnostics for this variable. We try to warn only on 586 // the first point at which a variable is used uninitialized. 587 break; 588 } 589 590 delete vec; 591 } 592 delete uses; 593 } 594}; 595} 596 597 598//===----------------------------------------------------------------------===// 599// -Wthread-safety 600//===----------------------------------------------------------------------===// 601 602namespace { 603/// \brief Implements a set of CFGBlocks using a BitVector. 604/// 605/// This class contains a minimal interface, primarily dictated by the SetType 606/// template parameter of the llvm::po_iterator template, as used with external 607/// storage. We also use this set to keep track of which CFGBlocks we visit 608/// during the analysis. 609class CFGBlockSet { 610 llvm::BitVector VisitedBlockIDs; 611 612public: 613 // po_iterator requires this iterator, but the only interface needed is the 614 // value_type typedef. 615 struct iterator { 616 typedef const CFGBlock *value_type; 617 }; 618 619 CFGBlockSet() {} 620 CFGBlockSet(const CFG *G) : VisitedBlockIDs(G->getNumBlockIDs(), false) {} 621 622 /// \brief Set the bit associated with a particular CFGBlock. 623 /// This is the important method for the SetType template parameter. 624 bool insert(const CFGBlock *Block) { 625 // Note that insert() is called by po_iterator, which doesn't check to make 626 // sure that Block is non-null. Moreover, the CFGBlock iterator will 627 // occasionally hand out null pointers for pruned edges, so we catch those 628 // here. 629 if (Block == 0) 630 return false; // if an edge is trivially false. 631 if (VisitedBlockIDs.test(Block->getBlockID())) 632 return false; 633 VisitedBlockIDs.set(Block->getBlockID()); 634 return true; 635 } 636 637 /// \brief Check if the bit for a CFGBlock has been already set. 638 /// This method is for tracking visited blocks in the main threadsafety loop. 639 /// Block must not be null. 640 bool alreadySet(const CFGBlock *Block) { 641 return VisitedBlockIDs.test(Block->getBlockID()); 642 } 643}; 644 645/// \brief We create a helper class which we use to iterate through CFGBlocks in 646/// the topological order. 647class TopologicallySortedCFG { 648 typedef llvm::po_iterator<const CFG*, CFGBlockSet, true> po_iterator; 649 650 std::vector<const CFGBlock*> Blocks; 651 652public: 653 typedef std::vector<const CFGBlock*>::reverse_iterator iterator; 654 655 TopologicallySortedCFG(const CFG *CFGraph) { 656 Blocks.reserve(CFGraph->getNumBlockIDs()); 657 CFGBlockSet BSet(CFGraph); 658 659 for (po_iterator I = po_iterator::begin(CFGraph, BSet), 660 E = po_iterator::end(CFGraph, BSet); I != E; ++I) { 661 Blocks.push_back(*I); 662 } 663 } 664 665 iterator begin() { 666 return Blocks.rbegin(); 667 } 668 669 iterator end() { 670 return Blocks.rend(); 671 } 672}; 673 674/// \brief A LockID object uniquely identifies a particular lock acquired, and 675/// is built from an Expr* (i.e. calling a lock function). 676/// 677/// Thread-safety analysis works by comparing lock expressions. Within the 678/// body of a function, an expression such as "x->foo->bar.mu" will resolve to 679/// a particular lock object at run-time. Subsequent occurrences of the same 680/// expression (where "same" means syntactic equality) will refer to the same 681/// run-time object if three conditions hold: 682/// (1) Local variables in the expression, such as "x" have not changed. 683/// (2) Values on the heap that affect the expression have not changed. 684/// (3) The expression involves only pure function calls. 685/// The current implementation assumes, but does not verify, that multiple uses 686/// of the same lock expression satisfies these criteria. 687/// 688/// Clang introduces an additional wrinkle, which is that it is difficult to 689/// derive canonical expressions, or compare expressions directly for equality. 690/// Thus, we identify a lock not by an Expr, but by the set of named 691/// declarations that are referenced by the Expr. In other words, 692/// x->foo->bar.mu will be a four element vector with the Decls for 693/// mu, bar, and foo, and x. The vector will uniquely identify the expression 694/// for all practical purposes. 695/// 696/// Note we will need to perform substitution on "this" and function parameter 697/// names when constructing a lock expression. 698/// 699/// For example: 700/// class C { Mutex Mu; void lock() EXCLUSIVE_LOCK_FUNCTION(this->Mu); }; 701/// void myFunc(C *X) { ... X->lock() ... } 702/// The original expression for the lock acquired by myFunc is "this->Mu", but 703/// "X" is substituted for "this" so we get X->Mu(); 704/// 705/// For another example: 706/// foo(MyList *L) EXCLUSIVE_LOCKS_REQUIRED(L->Mu) { ... } 707/// MyList *MyL; 708/// foo(MyL); // requires lock MyL->Mu to be held 709/// 710/// FIXME: In C++0x Mutexes are the objects that control access to shared 711/// variables, while Locks are the objects that acquire and release Mutexes. We 712/// may want to switch to this new terminology soon, in which case we should 713/// rename this class "Mutex" and rename "LockId" to "MutexId", as well as 714/// making sure that the terms Lock and Mutex throughout this code are 715/// consistent with C++0x 716/// 717/// FIXME: We should also pick one and canonicalize all usage of lock vs acquire 718/// and unlock vs release as verbs. 719class LockID { 720 SmallVector<NamedDecl*, 2> DeclSeq; 721 722 /// Build a Decl sequence representing the lock from the given expression. 723 /// Recursive function that bottoms out when the final DeclRefExpr is reached. 724 void buildLock(Expr *Exp) { 725 if (DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(Exp)) { 726 NamedDecl *ND = cast<NamedDecl>(DRE->getDecl()->getCanonicalDecl()); 727 DeclSeq.push_back(ND); 728 } else if (MemberExpr *ME = dyn_cast<MemberExpr>(Exp)) { 729 NamedDecl *ND = ME->getMemberDecl(); 730 DeclSeq.push_back(ND); 731 buildLock(ME->getBase()); 732 } else { 733 // FIXME: add diagnostic 734 llvm::report_fatal_error("Expected lock expression!"); 735 } 736 } 737 738public: 739 LockID(Expr *LExpr) { 740 buildLock(LExpr); 741 assert(!DeclSeq.empty()); 742 } 743 744 bool operator==(const LockID &other) const { 745 return DeclSeq == other.DeclSeq; 746 } 747 748 bool operator!=(const LockID &other) const { 749 return !(*this == other); 750 } 751 752 // SmallVector overloads Operator< to do lexicographic ordering. Note that 753 // we use pointer equality (and <) to compare NamedDecls. This means the order 754 // of LockIDs in a lockset is nondeterministic. In order to output 755 // diagnostics in a deterministic ordering, we must order all diagnostics to 756 // output by SourceLocation when iterating through this lockset. 757 bool operator<(const LockID &other) const { 758 return DeclSeq < other.DeclSeq; 759 } 760 761 /// \brief Returns the name of the first Decl in the list for a given LockID; 762 /// e.g. the lock expression foo.bar() has name "bar". 763 /// The caret will point unambiguously to the lock expression, so using this 764 /// name in diagnostics is a way to get simple, and consistent, lock names. 765 /// We do not want to output the entire expression text for security reasons. 766 StringRef getName() const { 767 return DeclSeq.front()->getName(); 768 } 769 770 void Profile(llvm::FoldingSetNodeID &ID) const { 771 for (SmallVectorImpl<NamedDecl*>::const_iterator I = DeclSeq.begin(), 772 E = DeclSeq.end(); I != E; ++I) { 773 ID.AddPointer(*I); 774 } 775 } 776}; 777 778/// \brief This is a helper class that stores info about the most recent 779/// accquire of a Lock. 780/// 781/// The main body of the analysis maps LockIDs to LockDatas. 782struct LockData { 783 SourceLocation AcquireLoc; 784 785 LockData(SourceLocation Loc) : AcquireLoc(Loc) {} 786 787 bool operator==(const LockData &other) const { 788 return AcquireLoc == other.AcquireLoc; 789 } 790 791 bool operator!=(const LockData &other) const { 792 return !(*this == other); 793 } 794 795 void Profile(llvm::FoldingSetNodeID &ID) const { 796 ID.AddInteger(AcquireLoc.getRawEncoding()); 797 } 798}; 799 800/// A Lockset maps each LockID (defined above) to information about how it has 801/// been locked. 802typedef llvm::ImmutableMap<LockID, LockData> Lockset; 803 804/// \brief We use this class to visit different types of expressions in 805/// CFGBlocks, and build up the lockset. 806/// An expression may cause us to add or remove locks from the lockset, or else 807/// output error messages related to missing locks. 808/// FIXME: In future, we may be able to not inherit from a visitor. 809class BuildLockset : public StmtVisitor<BuildLockset> { 810 Sema &S; 811 Lockset LSet; 812 Lockset::Factory &LocksetFactory; 813 814 // Helper functions 815 void removeLock(SourceLocation UnlockLoc, Expr *LockExp); 816 void addLock(SourceLocation LockLoc, Expr *LockExp); 817 818public: 819 BuildLockset(Sema &S, Lockset LS, Lockset::Factory &F) 820 : StmtVisitor<BuildLockset>(), S(S), LSet(LS), 821 LocksetFactory(F) {} 822 823 Lockset getLockset() { 824 return LSet; 825 } 826 827 void VisitDeclRefExpr(DeclRefExpr *Exp); 828 void VisitCXXMemberCallExpr(CXXMemberCallExpr *Exp); 829}; 830 831/// \brief Add a new lock to the lockset, warning if the lock is already there. 832/// \param LockExp The lock expression corresponding to the lock to be added 833/// \param LockLoc The source location of the acquire 834void BuildLockset::addLock(SourceLocation LockLoc, Expr *LockExp) { 835 LockID Lock(LockExp); 836 LockData NewLockData(LockLoc); 837 838 if (LSet.contains(Lock)) 839 S.Diag(LockLoc, diag::warn_double_lock) << Lock.getName(); 840 841 LSet = LocksetFactory.add(LSet, Lock, NewLockData); 842} 843 844/// \brief Remove a lock from the lockset, warning if the lock is not there. 845/// \param LockExp The lock expression corresponding to the lock to be removed 846/// \param UnlockLoc The source location of the unlock (only used in error msg) 847void BuildLockset::removeLock(SourceLocation UnlockLoc, Expr *LockExp) { 848 LockID Lock(LockExp); 849 850 Lockset NewLSet = LocksetFactory.remove(LSet, Lock); 851 if(NewLSet == LSet) 852 S.Diag(UnlockLoc, diag::warn_unlock_but_no_acquire) << Lock.getName(); 853 854 LSet = NewLSet; 855} 856 857void BuildLockset::VisitDeclRefExpr(DeclRefExpr *Exp) { 858 // FIXME: checking for guarded_by/var and pt_guarded_by/var 859} 860 861/// \brief When visiting CXXMemberCallExprs we need to examine the attributes on 862/// the method that is being called and add, remove or check locks in the 863/// lockset accordingly. 864void BuildLockset::VisitCXXMemberCallExpr(CXXMemberCallExpr *Exp) { 865 NamedDecl *D = dyn_cast_or_null<NamedDecl>(Exp->getCalleeDecl()); 866 867 SourceLocation ExpLocation = Exp->getExprLoc(); 868 Expr *Parent = Exp->getImplicitObjectArgument(); 869 870 if(!D || !D->hasAttrs()) 871 return; 872 873 AttrVec &ArgAttrs = D->getAttrs(); 874 for(unsigned i = 0; i < ArgAttrs.size(); ++i) { 875 Attr *Attr = ArgAttrs[i]; 876 switch (Attr->getKind()) { 877 // When we encounter an exclusive lock function, we need to add the lock 878 // to our lockset. 879 case attr::ExclusiveLockFunction: { 880 ExclusiveLockFunctionAttr *ELFAttr = 881 cast<ExclusiveLockFunctionAttr>(Attr); 882 883 if (ELFAttr->args_size() == 0) {// The lock held is the "this" object. 884 addLock(ExpLocation, Parent); 885 break; 886 } 887 888 for (ExclusiveLockFunctionAttr::args_iterator I = ELFAttr->args_begin(), 889 E = ELFAttr->args_end(); I != E; ++I) 890 addLock(ExpLocation, *I); 891 // FIXME: acquired_after/acquired_before annotations 892 break; 893 } 894 895 // When we encounter an unlock function, we need to remove unlocked locks 896 // from the lockset, and flag a warning if they are not there. 897 case attr::UnlockFunction: { 898 UnlockFunctionAttr *UFAttr = cast<UnlockFunctionAttr>(Attr); 899 900 if (UFAttr->args_size() == 0) { // The lock held is the "this" object. 901 removeLock(ExpLocation, Parent); 902 break; 903 } 904 905 for (UnlockFunctionAttr::args_iterator I = UFAttr->args_begin(), 906 E = UFAttr->args_end(); I != E; ++I) 907 removeLock(ExpLocation, *I); 908 break; 909 } 910 911 // Ignore other (non thread-safety) attributes 912 default: 913 break; 914 } 915 } 916} 917 918typedef std::pair<SourceLocation, PartialDiagnostic> DelayedDiag; 919typedef llvm::SmallVector<DelayedDiag, 4> DiagList; 920 921struct SortDiagBySourceLocation { 922 Sema &S; 923 924 SortDiagBySourceLocation(Sema &S) : S(S) {} 925 926 bool operator()(const DelayedDiag &left, const DelayedDiag &right) { 927 // Although this call will be slow, this is only called when outputting 928 // multiple warnings. 929 return S.getSourceManager().isBeforeInTranslationUnit(left.first, 930 right.first); 931 } 932}; 933} // end anonymous namespace 934 935/// \brief Emit all buffered diagnostics in order of sourcelocation. 936/// We need to output diagnostics produced while iterating through 937/// the lockset in deterministic order, so this function orders diagnostics 938/// and outputs them. 939static void EmitDiagnostics(Sema &S, DiagList &D) { 940 SortDiagBySourceLocation SortDiagBySL(S); 941 sort(D.begin(), D.end(), SortDiagBySL); 942 for (DiagList::iterator I = D.begin(), E = D.end(); I != E; ++I) 943 S.Diag(I->first, I->second); 944} 945 946/// \brief Compute the intersection of two locksets and issue warnings for any 947/// locks in the symmetric difference. 948/// 949/// This function is used at a merge point in the CFG when comparing the lockset 950/// of each branch being merged. For example, given the following sequence: 951/// A; if () then B; else C; D; we need to check that the lockset after B and C 952/// are the same. In the event of a difference, we use the intersection of these 953/// two locksets at the start of D. 954static Lockset intersectAndWarn(Sema &S, Lockset LSet1, Lockset LSet2, 955 Lockset::Factory &Fact) { 956 Lockset Intersection = LSet1; 957 DiagList Warnings; 958 959 for (Lockset::iterator I = LSet2.begin(), E = LSet2.end(); I != E; ++I) { 960 if (!LSet1.contains(I.getKey())) { 961 const LockID &MissingLock = I.getKey(); 962 const LockData &MissingLockData = I.getData(); 963 PartialDiagnostic Warning = 964 S.PDiag(diag::warn_lock_not_released_in_scope) << MissingLock.getName(); 965 Warnings.push_back(DelayedDiag(MissingLockData.AcquireLoc, Warning)); 966 } 967 } 968 969 for (Lockset::iterator I = LSet1.begin(), E = LSet1.end(); I != E; ++I) { 970 if (!LSet2.contains(I.getKey())) { 971 const LockID &MissingLock = I.getKey(); 972 const LockData &MissingLockData = I.getData(); 973 PartialDiagnostic Warning = 974 S.PDiag(diag::warn_lock_not_released_in_scope) << MissingLock.getName(); 975 Warnings.push_back(DelayedDiag(MissingLockData.AcquireLoc, Warning)); 976 Intersection = Fact.remove(Intersection, MissingLock); 977 } 978 } 979 980 EmitDiagnostics(S, Warnings); 981 return Intersection; 982} 983 984/// \brief Returns the location of the first Stmt in a Block. 985static SourceLocation getFirstStmtLocation(CFGBlock *Block) { 986 SourceLocation Loc; 987 for (CFGBlock::const_iterator BI = Block->begin(), BE = Block->end(); 988 BI != BE; ++BI) { 989 if (const CFGStmt *CfgStmt = dyn_cast<CFGStmt>(&(*BI))) { 990 Loc = CfgStmt->getStmt()->getLocStart(); 991 if (Loc.isValid()) return Loc; 992 } 993 } 994 if (Stmt *S = Block->getTerminator().getStmt()) { 995 Loc = S->getLocStart(); 996 if (Loc.isValid()) return Loc; 997 } 998 return Loc; 999} 1000 1001/// \brief Warn about different locksets along backedges of loops. 1002/// This function is called when we encounter a back edge. At that point, 1003/// we need to verify that the lockset before taking the backedge is the 1004/// same as the lockset before entering the loop. 1005/// 1006/// \param LoopEntrySet Locks held before starting the loop 1007/// \param LoopReentrySet Locks held in the last CFG block of the loop 1008static void warnBackEdgeUnequalLocksets(Sema &S, const Lockset LoopReentrySet, 1009 const Lockset LoopEntrySet, 1010 SourceLocation FirstLocInLoop) { 1011 assert(FirstLocInLoop.isValid()); 1012 DiagList Warnings; 1013 1014 // Warn for locks held at the start of the loop, but not the end. 1015 for (Lockset::iterator I = LoopEntrySet.begin(), E = LoopEntrySet.end(); 1016 I != E; ++I) { 1017 if (!LoopReentrySet.contains(I.getKey())) { 1018 const LockID &MissingLock = I.getKey(); 1019 // We report this error at the location of the first statement in a loop 1020 PartialDiagnostic Warning = 1021 S.PDiag(diag::warn_expecting_lock_held_on_loop) 1022 << MissingLock.getName(); 1023 Warnings.push_back(DelayedDiag(FirstLocInLoop, Warning)); 1024 } 1025 } 1026 1027 // Warn for locks held at the end of the loop, but not at the start. 1028 for (Lockset::iterator I = LoopReentrySet.begin(), E = LoopReentrySet.end(); 1029 I != E; ++I) { 1030 if (!LoopEntrySet.contains(I.getKey())) { 1031 const LockID &MissingLock = I.getKey(); 1032 const LockData &MissingLockData = I.getData(); 1033 PartialDiagnostic Warning = 1034 S.PDiag(diag::warn_lock_not_released_in_scope) << MissingLock.getName(); 1035 Warnings.push_back(DelayedDiag(MissingLockData.AcquireLoc, Warning)); 1036 } 1037 } 1038 1039 EmitDiagnostics(S, Warnings); 1040} 1041 1042/// \brief Check a function's CFG for thread-safety violations. 1043/// 1044/// We traverse the blocks in the CFG, compute the set of locks that are held 1045/// at the end of each block, and issue warnings for thread safety violations. 1046/// Each block in the CFG is traversed exactly once. 1047static void checkThreadSafety(Sema &S, AnalysisContext &AC) { 1048 CFG *CFGraph = AC.getCFG(); 1049 if (!CFGraph) return; 1050 1051 Lockset::Factory LocksetFactory; 1052 1053 // FIXME: Swith to SmallVector? Otherwise improve performance impact? 1054 std::vector<Lockset> EntryLocksets(CFGraph->getNumBlockIDs(), 1055 LocksetFactory.getEmptyMap()); 1056 std::vector<Lockset> ExitLocksets(CFGraph->getNumBlockIDs(), 1057 LocksetFactory.getEmptyMap()); 1058 1059 // We need to explore the CFG via a "topological" ordering. 1060 // That way, we will be guaranteed to have information about required 1061 // predecessor locksets when exploring a new block. 1062 TopologicallySortedCFG SortedGraph(CFGraph); 1063 CFGBlockSet VisitedBlocks(CFGraph); 1064 1065 for (TopologicallySortedCFG::iterator I = SortedGraph.begin(), 1066 E = SortedGraph.end(); I!= E; ++I) { 1067 const CFGBlock *CurrBlock = *I; 1068 int CurrBlockID = CurrBlock->getBlockID(); 1069 1070 VisitedBlocks.insert(CurrBlock); 1071 1072 // Use the default initial lockset in case there are no predecessors. 1073 Lockset &Entryset = EntryLocksets[CurrBlockID]; 1074 Lockset &Exitset = ExitLocksets[CurrBlockID]; 1075 1076 // Iterate through the predecessor blocks and warn if the lockset for all 1077 // predecessors is not the same. We take the entry lockset of the current 1078 // block to be the intersection of all previous locksets. 1079 // FIXME: By keeping the intersection, we may output more errors in future 1080 // for a lock which is not in the intersection, but was in the union. We 1081 // may want to also keep the union in future. As an example, let's say 1082 // the intersection contains Lock L, and the union contains L and M. 1083 // Later we unlock M. At this point, we would output an error because we 1084 // never locked M; although the real error is probably that we forgot to 1085 // lock M on all code paths. Conversely, let's say that later we lock M. 1086 // In this case, we should compare against the intersection instead of the 1087 // union because the real error is probably that we forgot to unlock M on 1088 // all code paths. 1089 bool LocksetInitialized = false; 1090 for (CFGBlock::const_pred_iterator PI = CurrBlock->pred_begin(), 1091 PE = CurrBlock->pred_end(); PI != PE; ++PI) { 1092 1093 // if *PI -> CurrBlock is a back edge 1094 if (*PI == 0 || !VisitedBlocks.alreadySet(*PI)) 1095 continue; 1096 1097 int PrevBlockID = (*PI)->getBlockID(); 1098 if (!LocksetInitialized) { 1099 Entryset = ExitLocksets[PrevBlockID]; 1100 LocksetInitialized = true; 1101 } else { 1102 Entryset = intersectAndWarn(S, Entryset, ExitLocksets[PrevBlockID], 1103 LocksetFactory); 1104 } 1105 } 1106 1107 BuildLockset LocksetBuilder(S, Entryset, LocksetFactory); 1108 for (CFGBlock::const_iterator BI = CurrBlock->begin(), 1109 BE = CurrBlock->end(); BI != BE; ++BI) { 1110 if (const CFGStmt *CfgStmt = dyn_cast<CFGStmt>(&*BI)) 1111 LocksetBuilder.Visit(const_cast<Stmt*>(CfgStmt->getStmt())); 1112 } 1113 Exitset = LocksetBuilder.getLockset(); 1114 1115 // For every back edge from CurrBlock (the end of the loop) to another block 1116 // (FirstLoopBlock) we need to check that the Lockset of Block is equal to 1117 // the one held at the beginning of FirstLoopBlock. We can look up the 1118 // Lockset held at the beginning of FirstLoopBlock in the EntryLockSets map. 1119 for (CFGBlock::const_succ_iterator SI = CurrBlock->succ_begin(), 1120 SE = CurrBlock->succ_end(); SI != SE; ++SI) { 1121 1122 // if CurrBlock -> *SI is *not* a back edge 1123 if (*SI == 0 || !VisitedBlocks.alreadySet(*SI)) 1124 continue; 1125 1126 CFGBlock *FirstLoopBlock = *SI; 1127 SourceLocation FirstLoopLocation = getFirstStmtLocation(FirstLoopBlock); 1128 1129 assert(FirstLoopLocation.isValid()); 1130 // Fail gracefully in release code. 1131 if (!FirstLoopLocation.isValid()) 1132 continue; 1133 1134 Lockset PreLoop = EntryLocksets[FirstLoopBlock->getBlockID()]; 1135 Lockset LoopEnd = ExitLocksets[CurrBlockID]; 1136 warnBackEdgeUnequalLocksets(S, LoopEnd, PreLoop, FirstLoopLocation); 1137 } 1138 } 1139 1140 Lockset FinalLockset = ExitLocksets[CFGraph->getExit().getBlockID()]; 1141 if (!FinalLockset.isEmpty()) { 1142 DiagList Warnings; 1143 for (Lockset::iterator I=FinalLockset.begin(), E=FinalLockset.end(); 1144 I != E; ++I) { 1145 const LockID &MissingLock = I.getKey(); 1146 const LockData &MissingLockData = I.getData(); 1147 1148 std::string FunName = "<unknown>"; 1149 if (const NamedDecl *ContextDecl = dyn_cast<NamedDecl>(AC.getDecl())) { 1150 FunName = ContextDecl->getDeclName().getAsString(); 1151 } 1152 1153 PartialDiagnostic Warning = 1154 S.PDiag(diag::warn_locks_not_released) 1155 << MissingLock.getName() << FunName; 1156 Warnings.push_back(DelayedDiag(MissingLockData.AcquireLoc, Warning)); 1157 } 1158 EmitDiagnostics(S, Warnings); 1159 } 1160} 1161 1162 1163//===----------------------------------------------------------------------===// 1164// AnalysisBasedWarnings - Worker object used by Sema to execute analysis-based 1165// warnings on a function, method, or block. 1166//===----------------------------------------------------------------------===// 1167 1168clang::sema::AnalysisBasedWarnings::Policy::Policy() { 1169 enableCheckFallThrough = 1; 1170 enableCheckUnreachable = 0; 1171 enableThreadSafetyAnalysis = 0; 1172} 1173 1174clang::sema::AnalysisBasedWarnings::AnalysisBasedWarnings(Sema &s) 1175 : S(s), 1176 NumFunctionsAnalyzed(0), 1177 NumFunctionsWithBadCFGs(0), 1178 NumCFGBlocks(0), 1179 MaxCFGBlocksPerFunction(0), 1180 NumUninitAnalysisFunctions(0), 1181 NumUninitAnalysisVariables(0), 1182 MaxUninitAnalysisVariablesPerFunction(0), 1183 NumUninitAnalysisBlockVisits(0), 1184 MaxUninitAnalysisBlockVisitsPerFunction(0) { 1185 Diagnostic &D = S.getDiagnostics(); 1186 DefaultPolicy.enableCheckUnreachable = (unsigned) 1187 (D.getDiagnosticLevel(diag::warn_unreachable, SourceLocation()) != 1188 Diagnostic::Ignored); 1189 DefaultPolicy.enableThreadSafetyAnalysis = (unsigned) 1190 (D.getDiagnosticLevel(diag::warn_double_lock, SourceLocation()) != 1191 Diagnostic::Ignored); 1192 1193} 1194 1195static void flushDiagnostics(Sema &S, sema::FunctionScopeInfo *fscope) { 1196 for (SmallVectorImpl<sema::PossiblyUnreachableDiag>::iterator 1197 i = fscope->PossiblyUnreachableDiags.begin(), 1198 e = fscope->PossiblyUnreachableDiags.end(); 1199 i != e; ++i) { 1200 const sema::PossiblyUnreachableDiag &D = *i; 1201 S.Diag(D.Loc, D.PD); 1202 } 1203} 1204 1205void clang::sema:: 1206AnalysisBasedWarnings::IssueWarnings(sema::AnalysisBasedWarnings::Policy P, 1207 sema::FunctionScopeInfo *fscope, 1208 const Decl *D, const BlockExpr *blkExpr) { 1209 1210 // We avoid doing analysis-based warnings when there are errors for 1211 // two reasons: 1212 // (1) The CFGs often can't be constructed (if the body is invalid), so 1213 // don't bother trying. 1214 // (2) The code already has problems; running the analysis just takes more 1215 // time. 1216 Diagnostic &Diags = S.getDiagnostics(); 1217 1218 // Do not do any analysis for declarations in system headers if we are 1219 // going to just ignore them. 1220 if (Diags.getSuppressSystemWarnings() && 1221 S.SourceMgr.isInSystemHeader(D->getLocation())) 1222 return; 1223 1224 // For code in dependent contexts, we'll do this at instantiation time. 1225 if (cast<DeclContext>(D)->isDependentContext()) 1226 return; 1227 1228 if (Diags.hasErrorOccurred() || Diags.hasFatalErrorOccurred()) { 1229 // Flush out any possibly unreachable diagnostics. 1230 flushDiagnostics(S, fscope); 1231 return; 1232 } 1233 1234 const Stmt *Body = D->getBody(); 1235 assert(Body); 1236 1237 AnalysisContext AC(D, 0); 1238 1239 // Don't generate EH edges for CallExprs as we'd like to avoid the n^2 1240 // explosion for destrutors that can result and the compile time hit. 1241 AC.getCFGBuildOptions().PruneTriviallyFalseEdges = true; 1242 AC.getCFGBuildOptions().AddEHEdges = false; 1243 AC.getCFGBuildOptions().AddInitializers = true; 1244 AC.getCFGBuildOptions().AddImplicitDtors = true; 1245 1246 // Force that certain expressions appear as CFGElements in the CFG. This 1247 // is used to speed up various analyses. 1248 // FIXME: This isn't the right factoring. This is here for initial 1249 // prototyping, but we need a way for analyses to say what expressions they 1250 // expect to always be CFGElements and then fill in the BuildOptions 1251 // appropriately. This is essentially a layering violation. 1252 if (P.enableCheckUnreachable) { 1253 // Unreachable code analysis requires a linearized CFG. 1254 AC.getCFGBuildOptions().setAllAlwaysAdd(); 1255 } 1256 else { 1257 AC.getCFGBuildOptions() 1258 .setAlwaysAdd(Stmt::BinaryOperatorClass) 1259 .setAlwaysAdd(Stmt::BlockExprClass) 1260 .setAlwaysAdd(Stmt::CStyleCastExprClass) 1261 .setAlwaysAdd(Stmt::DeclRefExprClass) 1262 .setAlwaysAdd(Stmt::ImplicitCastExprClass) 1263 .setAlwaysAdd(Stmt::UnaryOperatorClass); 1264 } 1265 1266 // Construct the analysis context with the specified CFG build options. 1267 1268 // Emit delayed diagnostics. 1269 if (!fscope->PossiblyUnreachableDiags.empty()) { 1270 bool analyzed = false; 1271 1272 // Register the expressions with the CFGBuilder. 1273 for (SmallVectorImpl<sema::PossiblyUnreachableDiag>::iterator 1274 i = fscope->PossiblyUnreachableDiags.begin(), 1275 e = fscope->PossiblyUnreachableDiags.end(); 1276 i != e; ++i) { 1277 if (const Stmt *stmt = i->stmt) 1278 AC.registerForcedBlockExpression(stmt); 1279 } 1280 1281 if (AC.getCFG()) { 1282 analyzed = true; 1283 for (SmallVectorImpl<sema::PossiblyUnreachableDiag>::iterator 1284 i = fscope->PossiblyUnreachableDiags.begin(), 1285 e = fscope->PossiblyUnreachableDiags.end(); 1286 i != e; ++i) 1287 { 1288 const sema::PossiblyUnreachableDiag &D = *i; 1289 bool processed = false; 1290 if (const Stmt *stmt = i->stmt) { 1291 const CFGBlock *block = AC.getBlockForRegisteredExpression(stmt); 1292 assert(block); 1293 if (CFGReverseBlockReachabilityAnalysis *cra = AC.getCFGReachablityAnalysis()) { 1294 // Can this block be reached from the entrance? 1295 if (cra->isReachable(&AC.getCFG()->getEntry(), block)) 1296 S.Diag(D.Loc, D.PD); 1297 processed = true; 1298 } 1299 } 1300 if (!processed) { 1301 // Emit the warning anyway if we cannot map to a basic block. 1302 S.Diag(D.Loc, D.PD); 1303 } 1304 } 1305 } 1306 1307 if (!analyzed) 1308 flushDiagnostics(S, fscope); 1309 } 1310 1311 1312 // Warning: check missing 'return' 1313 if (P.enableCheckFallThrough) { 1314 const CheckFallThroughDiagnostics &CD = 1315 (isa<BlockDecl>(D) ? CheckFallThroughDiagnostics::MakeForBlock() 1316 : CheckFallThroughDiagnostics::MakeForFunction(D)); 1317 CheckFallThroughForBody(S, D, Body, blkExpr, CD, AC); 1318 } 1319 1320 // Warning: check for unreachable code 1321 if (P.enableCheckUnreachable) 1322 CheckUnreachable(S, AC); 1323 1324 // Check for thread safety violations 1325 if (P.enableThreadSafetyAnalysis) 1326 checkThreadSafety(S, AC); 1327 1328 if (Diags.getDiagnosticLevel(diag::warn_uninit_var, D->getLocStart()) 1329 != Diagnostic::Ignored || 1330 Diags.getDiagnosticLevel(diag::warn_maybe_uninit_var, D->getLocStart()) 1331 != Diagnostic::Ignored) { 1332 if (CFG *cfg = AC.getCFG()) { 1333 UninitValsDiagReporter reporter(S); 1334 UninitVariablesAnalysisStats stats; 1335 std::memset(&stats, 0, sizeof(UninitVariablesAnalysisStats)); 1336 runUninitializedVariablesAnalysis(*cast<DeclContext>(D), *cfg, AC, 1337 reporter, stats); 1338 1339 if (S.CollectStats && stats.NumVariablesAnalyzed > 0) { 1340 ++NumUninitAnalysisFunctions; 1341 NumUninitAnalysisVariables += stats.NumVariablesAnalyzed; 1342 NumUninitAnalysisBlockVisits += stats.NumBlockVisits; 1343 MaxUninitAnalysisVariablesPerFunction = 1344 std::max(MaxUninitAnalysisVariablesPerFunction, 1345 stats.NumVariablesAnalyzed); 1346 MaxUninitAnalysisBlockVisitsPerFunction = 1347 std::max(MaxUninitAnalysisBlockVisitsPerFunction, 1348 stats.NumBlockVisits); 1349 } 1350 } 1351 } 1352 1353 // Collect statistics about the CFG if it was built. 1354 if (S.CollectStats && AC.isCFGBuilt()) { 1355 ++NumFunctionsAnalyzed; 1356 if (CFG *cfg = AC.getCFG()) { 1357 // If we successfully built a CFG for this context, record some more 1358 // detail information about it. 1359 NumCFGBlocks += cfg->getNumBlockIDs(); 1360 MaxCFGBlocksPerFunction = std::max(MaxCFGBlocksPerFunction, 1361 cfg->getNumBlockIDs()); 1362 } else { 1363 ++NumFunctionsWithBadCFGs; 1364 } 1365 } 1366} 1367 1368void clang::sema::AnalysisBasedWarnings::PrintStats() const { 1369 llvm::errs() << "\n*** Analysis Based Warnings Stats:\n"; 1370 1371 unsigned NumCFGsBuilt = NumFunctionsAnalyzed - NumFunctionsWithBadCFGs; 1372 unsigned AvgCFGBlocksPerFunction = 1373 !NumCFGsBuilt ? 0 : NumCFGBlocks/NumCFGsBuilt; 1374 llvm::errs() << NumFunctionsAnalyzed << " functions analyzed (" 1375 << NumFunctionsWithBadCFGs << " w/o CFGs).\n" 1376 << " " << NumCFGBlocks << " CFG blocks built.\n" 1377 << " " << AvgCFGBlocksPerFunction 1378 << " average CFG blocks per function.\n" 1379 << " " << MaxCFGBlocksPerFunction 1380 << " max CFG blocks per function.\n"; 1381 1382 unsigned AvgUninitVariablesPerFunction = !NumUninitAnalysisFunctions ? 0 1383 : NumUninitAnalysisVariables/NumUninitAnalysisFunctions; 1384 unsigned AvgUninitBlockVisitsPerFunction = !NumUninitAnalysisFunctions ? 0 1385 : NumUninitAnalysisBlockVisits/NumUninitAnalysisFunctions; 1386 llvm::errs() << NumUninitAnalysisFunctions 1387 << " functions analyzed for uninitialiazed variables\n" 1388 << " " << NumUninitAnalysisVariables << " variables analyzed.\n" 1389 << " " << AvgUninitVariablesPerFunction 1390 << " average variables per function.\n" 1391 << " " << MaxUninitAnalysisVariablesPerFunction 1392 << " max variables per function.\n" 1393 << " " << NumUninitAnalysisBlockVisits << " block visits.\n" 1394 << " " << AvgUninitBlockVisitsPerFunction 1395 << " average block visits per function.\n" 1396 << " " << MaxUninitAnalysisBlockVisitsPerFunction 1397 << " max block visits per function.\n"; 1398} 1399