SemaChecking.cpp revision 0acc311bf73c85fd34ce6f89a4e786b7ecd214aa
1//===--- SemaChecking.cpp - Extra Semantic Checking -----------------------===// 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 implements extra semantic analysis beyond what is enforced 11// by the C type system. 12// 13//===----------------------------------------------------------------------===// 14 15#include "Sema.h" 16#include "clang/AST/ASTContext.h" 17#include "clang/AST/DeclObjC.h" 18#include "clang/AST/ExprCXX.h" 19#include "clang/AST/ExprObjC.h" 20#include "clang/Lex/LiteralSupport.h" 21#include "clang/Lex/Preprocessor.h" 22#include <limits> 23using namespace clang; 24 25/// getLocationOfStringLiteralByte - Return a source location that points to the 26/// specified byte of the specified string literal. 27/// 28/// Strings are amazingly complex. They can be formed from multiple tokens and 29/// can have escape sequences in them in addition to the usual trigraph and 30/// escaped newline business. This routine handles this complexity. 31/// 32SourceLocation Sema::getLocationOfStringLiteralByte(const StringLiteral *SL, 33 unsigned ByteNo) const { 34 assert(!SL->isWide() && "This doesn't work for wide strings yet"); 35 36 // Loop over all of the tokens in this string until we find the one that 37 // contains the byte we're looking for. 38 unsigned TokNo = 0; 39 while (1) { 40 assert(TokNo < SL->getNumConcatenated() && "Invalid byte number!"); 41 SourceLocation StrTokLoc = SL->getStrTokenLoc(TokNo); 42 43 // Get the spelling of the string so that we can get the data that makes up 44 // the string literal, not the identifier for the macro it is potentially 45 // expanded through. 46 SourceLocation StrTokSpellingLoc = SourceMgr.getSpellingLoc(StrTokLoc); 47 48 // Re-lex the token to get its length and original spelling. 49 std::pair<FileID, unsigned> LocInfo = 50 SourceMgr.getDecomposedLoc(StrTokSpellingLoc); 51 std::pair<const char *,const char *> Buffer = 52 SourceMgr.getBufferData(LocInfo.first); 53 const char *StrData = Buffer.first+LocInfo.second; 54 55 // Create a langops struct and enable trigraphs. This is sufficient for 56 // relexing tokens. 57 LangOptions LangOpts; 58 LangOpts.Trigraphs = true; 59 60 // Create a lexer starting at the beginning of this token. 61 Lexer TheLexer(StrTokSpellingLoc, LangOpts, Buffer.first, StrData, 62 Buffer.second); 63 Token TheTok; 64 TheLexer.LexFromRawLexer(TheTok); 65 66 // Use the StringLiteralParser to compute the length of the string in bytes. 67 StringLiteralParser SLP(&TheTok, 1, PP); 68 unsigned TokNumBytes = SLP.GetStringLength(); 69 70 // If the byte is in this token, return the location of the byte. 71 if (ByteNo < TokNumBytes || 72 (ByteNo == TokNumBytes && TokNo == SL->getNumConcatenated())) { 73 unsigned Offset = 74 StringLiteralParser::getOffsetOfStringByte(TheTok, ByteNo, PP); 75 76 // Now that we know the offset of the token in the spelling, use the 77 // preprocessor to get the offset in the original source. 78 return PP.AdvanceToTokenCharacter(StrTokLoc, Offset); 79 } 80 81 // Move to the next string token. 82 ++TokNo; 83 ByteNo -= TokNumBytes; 84 } 85} 86 87/// CheckablePrintfAttr - does a function call have a "printf" attribute 88/// and arguments that merit checking? 89bool Sema::CheckablePrintfAttr(const FormatAttr *Format, CallExpr *TheCall) { 90 if (Format->getType() == "printf") return true; 91 if (Format->getType() == "printf0") { 92 // printf0 allows null "format" string; if so don't check format/args 93 unsigned format_idx = Format->getFormatIdx() - 1; 94 // Does the index refer to the implicit object argument? 95 if (isa<CXXMemberCallExpr>(TheCall)) { 96 if (format_idx == 0) 97 return false; 98 --format_idx; 99 } 100 if (format_idx < TheCall->getNumArgs()) { 101 Expr *Format = TheCall->getArg(format_idx)->IgnoreParenCasts(); 102 if (!Format->isNullPointerConstant(Context, Expr::NPC_ValueDependentIsNull)) 103 return true; 104 } 105 } 106 return false; 107} 108 109Action::OwningExprResult 110Sema::CheckBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall) { 111 OwningExprResult TheCallResult(Owned(TheCall)); 112 113 switch (BuiltinID) { 114 case Builtin::BI__builtin___CFStringMakeConstantString: 115 assert(TheCall->getNumArgs() == 1 && 116 "Wrong # arguments to builtin CFStringMakeConstantString"); 117 if (CheckObjCString(TheCall->getArg(0))) 118 return ExprError(); 119 break; 120 case Builtin::BI__builtin_stdarg_start: 121 case Builtin::BI__builtin_va_start: 122 if (SemaBuiltinVAStart(TheCall)) 123 return ExprError(); 124 break; 125 case Builtin::BI__builtin_isgreater: 126 case Builtin::BI__builtin_isgreaterequal: 127 case Builtin::BI__builtin_isless: 128 case Builtin::BI__builtin_islessequal: 129 case Builtin::BI__builtin_islessgreater: 130 case Builtin::BI__builtin_isunordered: 131 if (SemaBuiltinUnorderedCompare(TheCall)) 132 return ExprError(); 133 break; 134 case Builtin::BI__builtin_isfinite: 135 case Builtin::BI__builtin_isinf: 136 case Builtin::BI__builtin_isinf_sign: 137 case Builtin::BI__builtin_isnan: 138 case Builtin::BI__builtin_isnormal: 139 if (SemaBuiltinUnaryFP(TheCall)) 140 return ExprError(); 141 break; 142 case Builtin::BI__builtin_return_address: 143 case Builtin::BI__builtin_frame_address: 144 if (SemaBuiltinStackAddress(TheCall)) 145 return ExprError(); 146 break; 147 case Builtin::BI__builtin_eh_return_data_regno: 148 if (SemaBuiltinEHReturnDataRegNo(TheCall)) 149 return ExprError(); 150 break; 151 case Builtin::BI__builtin_shufflevector: 152 return SemaBuiltinShuffleVector(TheCall); 153 // TheCall will be freed by the smart pointer here, but that's fine, since 154 // SemaBuiltinShuffleVector guts it, but then doesn't release it. 155 case Builtin::BI__builtin_prefetch: 156 if (SemaBuiltinPrefetch(TheCall)) 157 return ExprError(); 158 break; 159 case Builtin::BI__builtin_object_size: 160 if (SemaBuiltinObjectSize(TheCall)) 161 return ExprError(); 162 break; 163 case Builtin::BI__builtin_longjmp: 164 if (SemaBuiltinLongjmp(TheCall)) 165 return ExprError(); 166 break; 167 case Builtin::BI__sync_fetch_and_add: 168 case Builtin::BI__sync_fetch_and_sub: 169 case Builtin::BI__sync_fetch_and_or: 170 case Builtin::BI__sync_fetch_and_and: 171 case Builtin::BI__sync_fetch_and_xor: 172 case Builtin::BI__sync_fetch_and_nand: 173 case Builtin::BI__sync_add_and_fetch: 174 case Builtin::BI__sync_sub_and_fetch: 175 case Builtin::BI__sync_and_and_fetch: 176 case Builtin::BI__sync_or_and_fetch: 177 case Builtin::BI__sync_xor_and_fetch: 178 case Builtin::BI__sync_nand_and_fetch: 179 case Builtin::BI__sync_val_compare_and_swap: 180 case Builtin::BI__sync_bool_compare_and_swap: 181 case Builtin::BI__sync_lock_test_and_set: 182 case Builtin::BI__sync_lock_release: 183 if (SemaBuiltinAtomicOverloaded(TheCall)) 184 return ExprError(); 185 break; 186 } 187 188 return move(TheCallResult); 189} 190 191/// CheckFunctionCall - Check a direct function call for various correctness 192/// and safety properties not strictly enforced by the C type system. 193bool Sema::CheckFunctionCall(FunctionDecl *FDecl, CallExpr *TheCall) { 194 // Get the IdentifierInfo* for the called function. 195 IdentifierInfo *FnInfo = FDecl->getIdentifier(); 196 197 // None of the checks below are needed for functions that don't have 198 // simple names (e.g., C++ conversion functions). 199 if (!FnInfo) 200 return false; 201 202 // FIXME: This mechanism should be abstracted to be less fragile and 203 // more efficient. For example, just map function ids to custom 204 // handlers. 205 206 // Printf checking. 207 if (const FormatAttr *Format = FDecl->getAttr<FormatAttr>()) { 208 if (CheckablePrintfAttr(Format, TheCall)) { 209 bool HasVAListArg = Format->getFirstArg() == 0; 210 if (!HasVAListArg) { 211 if (const FunctionProtoType *Proto 212 = FDecl->getType()->getAs<FunctionProtoType>()) 213 HasVAListArg = !Proto->isVariadic(); 214 } 215 CheckPrintfArguments(TheCall, HasVAListArg, Format->getFormatIdx() - 1, 216 HasVAListArg ? 0 : Format->getFirstArg() - 1); 217 } 218 } 219 220 for (const NonNullAttr *NonNull = FDecl->getAttr<NonNullAttr>(); NonNull; 221 NonNull = NonNull->getNext<NonNullAttr>()) 222 CheckNonNullArguments(NonNull, TheCall); 223 224 return false; 225} 226 227bool Sema::CheckBlockCall(NamedDecl *NDecl, CallExpr *TheCall) { 228 // Printf checking. 229 const FormatAttr *Format = NDecl->getAttr<FormatAttr>(); 230 if (!Format) 231 return false; 232 233 const VarDecl *V = dyn_cast<VarDecl>(NDecl); 234 if (!V) 235 return false; 236 237 QualType Ty = V->getType(); 238 if (!Ty->isBlockPointerType()) 239 return false; 240 241 if (!CheckablePrintfAttr(Format, TheCall)) 242 return false; 243 244 bool HasVAListArg = Format->getFirstArg() == 0; 245 if (!HasVAListArg) { 246 const FunctionType *FT = 247 Ty->getAs<BlockPointerType>()->getPointeeType()->getAs<FunctionType>(); 248 if (const FunctionProtoType *Proto = dyn_cast<FunctionProtoType>(FT)) 249 HasVAListArg = !Proto->isVariadic(); 250 } 251 CheckPrintfArguments(TheCall, HasVAListArg, Format->getFormatIdx() - 1, 252 HasVAListArg ? 0 : Format->getFirstArg() - 1); 253 254 return false; 255} 256 257/// SemaBuiltinAtomicOverloaded - We have a call to a function like 258/// __sync_fetch_and_add, which is an overloaded function based on the pointer 259/// type of its first argument. The main ActOnCallExpr routines have already 260/// promoted the types of arguments because all of these calls are prototyped as 261/// void(...). 262/// 263/// This function goes through and does final semantic checking for these 264/// builtins, 265bool Sema::SemaBuiltinAtomicOverloaded(CallExpr *TheCall) { 266 DeclRefExpr *DRE =cast<DeclRefExpr>(TheCall->getCallee()->IgnoreParenCasts()); 267 FunctionDecl *FDecl = cast<FunctionDecl>(DRE->getDecl()); 268 269 // Ensure that we have at least one argument to do type inference from. 270 if (TheCall->getNumArgs() < 1) 271 return Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_few_args) 272 << 0 << TheCall->getCallee()->getSourceRange(); 273 274 // Inspect the first argument of the atomic builtin. This should always be 275 // a pointer type, whose element is an integral scalar or pointer type. 276 // Because it is a pointer type, we don't have to worry about any implicit 277 // casts here. 278 Expr *FirstArg = TheCall->getArg(0); 279 if (!FirstArg->getType()->isPointerType()) 280 return Diag(DRE->getLocStart(), diag::err_atomic_builtin_must_be_pointer) 281 << FirstArg->getType() << FirstArg->getSourceRange(); 282 283 QualType ValType = FirstArg->getType()->getAs<PointerType>()->getPointeeType(); 284 if (!ValType->isIntegerType() && !ValType->isPointerType() && 285 !ValType->isBlockPointerType()) 286 return Diag(DRE->getLocStart(), 287 diag::err_atomic_builtin_must_be_pointer_intptr) 288 << FirstArg->getType() << FirstArg->getSourceRange(); 289 290 // We need to figure out which concrete builtin this maps onto. For example, 291 // __sync_fetch_and_add with a 2 byte object turns into 292 // __sync_fetch_and_add_2. 293#define BUILTIN_ROW(x) \ 294 { Builtin::BI##x##_1, Builtin::BI##x##_2, Builtin::BI##x##_4, \ 295 Builtin::BI##x##_8, Builtin::BI##x##_16 } 296 297 static const unsigned BuiltinIndices[][5] = { 298 BUILTIN_ROW(__sync_fetch_and_add), 299 BUILTIN_ROW(__sync_fetch_and_sub), 300 BUILTIN_ROW(__sync_fetch_and_or), 301 BUILTIN_ROW(__sync_fetch_and_and), 302 BUILTIN_ROW(__sync_fetch_and_xor), 303 BUILTIN_ROW(__sync_fetch_and_nand), 304 305 BUILTIN_ROW(__sync_add_and_fetch), 306 BUILTIN_ROW(__sync_sub_and_fetch), 307 BUILTIN_ROW(__sync_and_and_fetch), 308 BUILTIN_ROW(__sync_or_and_fetch), 309 BUILTIN_ROW(__sync_xor_and_fetch), 310 BUILTIN_ROW(__sync_nand_and_fetch), 311 312 BUILTIN_ROW(__sync_val_compare_and_swap), 313 BUILTIN_ROW(__sync_bool_compare_and_swap), 314 BUILTIN_ROW(__sync_lock_test_and_set), 315 BUILTIN_ROW(__sync_lock_release) 316 }; 317#undef BUILTIN_ROW 318 319 // Determine the index of the size. 320 unsigned SizeIndex; 321 switch (Context.getTypeSize(ValType)/8) { 322 case 1: SizeIndex = 0; break; 323 case 2: SizeIndex = 1; break; 324 case 4: SizeIndex = 2; break; 325 case 8: SizeIndex = 3; break; 326 case 16: SizeIndex = 4; break; 327 default: 328 return Diag(DRE->getLocStart(), diag::err_atomic_builtin_pointer_size) 329 << FirstArg->getType() << FirstArg->getSourceRange(); 330 } 331 332 // Each of these builtins has one pointer argument, followed by some number of 333 // values (0, 1 or 2) followed by a potentially empty varags list of stuff 334 // that we ignore. Find out which row of BuiltinIndices to read from as well 335 // as the number of fixed args. 336 unsigned BuiltinID = FDecl->getBuiltinID(); 337 unsigned BuiltinIndex, NumFixed = 1; 338 switch (BuiltinID) { 339 default: assert(0 && "Unknown overloaded atomic builtin!"); 340 case Builtin::BI__sync_fetch_and_add: BuiltinIndex = 0; break; 341 case Builtin::BI__sync_fetch_and_sub: BuiltinIndex = 1; break; 342 case Builtin::BI__sync_fetch_and_or: BuiltinIndex = 2; break; 343 case Builtin::BI__sync_fetch_and_and: BuiltinIndex = 3; break; 344 case Builtin::BI__sync_fetch_and_xor: BuiltinIndex = 4; break; 345 case Builtin::BI__sync_fetch_and_nand:BuiltinIndex = 5; break; 346 347 case Builtin::BI__sync_add_and_fetch: BuiltinIndex = 6; break; 348 case Builtin::BI__sync_sub_and_fetch: BuiltinIndex = 7; break; 349 case Builtin::BI__sync_and_and_fetch: BuiltinIndex = 8; break; 350 case Builtin::BI__sync_or_and_fetch: BuiltinIndex = 9; break; 351 case Builtin::BI__sync_xor_and_fetch: BuiltinIndex =10; break; 352 case Builtin::BI__sync_nand_and_fetch:BuiltinIndex =11; break; 353 354 case Builtin::BI__sync_val_compare_and_swap: 355 BuiltinIndex = 12; 356 NumFixed = 2; 357 break; 358 case Builtin::BI__sync_bool_compare_and_swap: 359 BuiltinIndex = 13; 360 NumFixed = 2; 361 break; 362 case Builtin::BI__sync_lock_test_and_set: BuiltinIndex = 14; break; 363 case Builtin::BI__sync_lock_release: 364 BuiltinIndex = 15; 365 NumFixed = 0; 366 break; 367 } 368 369 // Now that we know how many fixed arguments we expect, first check that we 370 // have at least that many. 371 if (TheCall->getNumArgs() < 1+NumFixed) 372 return Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_few_args) 373 << 0 << TheCall->getCallee()->getSourceRange(); 374 375 376 // Get the decl for the concrete builtin from this, we can tell what the 377 // concrete integer type we should convert to is. 378 unsigned NewBuiltinID = BuiltinIndices[BuiltinIndex][SizeIndex]; 379 const char *NewBuiltinName = Context.BuiltinInfo.GetName(NewBuiltinID); 380 IdentifierInfo *NewBuiltinII = PP.getIdentifierInfo(NewBuiltinName); 381 FunctionDecl *NewBuiltinDecl = 382 cast<FunctionDecl>(LazilyCreateBuiltin(NewBuiltinII, NewBuiltinID, 383 TUScope, false, DRE->getLocStart())); 384 const FunctionProtoType *BuiltinFT = 385 NewBuiltinDecl->getType()->getAs<FunctionProtoType>(); 386 ValType = BuiltinFT->getArgType(0)->getAs<PointerType>()->getPointeeType(); 387 388 // If the first type needs to be converted (e.g. void** -> int*), do it now. 389 if (BuiltinFT->getArgType(0) != FirstArg->getType()) { 390 ImpCastExprToType(FirstArg, BuiltinFT->getArgType(0), CastExpr::CK_BitCast); 391 TheCall->setArg(0, FirstArg); 392 } 393 394 // Next, walk the valid ones promoting to the right type. 395 for (unsigned i = 0; i != NumFixed; ++i) { 396 Expr *Arg = TheCall->getArg(i+1); 397 398 // If the argument is an implicit cast, then there was a promotion due to 399 // "...", just remove it now. 400 if (ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Arg)) { 401 Arg = ICE->getSubExpr(); 402 ICE->setSubExpr(0); 403 ICE->Destroy(Context); 404 TheCall->setArg(i+1, Arg); 405 } 406 407 // GCC does an implicit conversion to the pointer or integer ValType. This 408 // can fail in some cases (1i -> int**), check for this error case now. 409 CastExpr::CastKind Kind = CastExpr::CK_Unknown; 410 CXXMethodDecl *ConversionDecl = 0; 411 if (CheckCastTypes(Arg->getSourceRange(), ValType, Arg, Kind, 412 ConversionDecl)) 413 return true; 414 415 // Okay, we have something that *can* be converted to the right type. Check 416 // to see if there is a potentially weird extension going on here. This can 417 // happen when you do an atomic operation on something like an char* and 418 // pass in 42. The 42 gets converted to char. This is even more strange 419 // for things like 45.123 -> char, etc. 420 // FIXME: Do this check. 421 ImpCastExprToType(Arg, ValType, Kind, /*isLvalue=*/false); 422 TheCall->setArg(i+1, Arg); 423 } 424 425 // Switch the DeclRefExpr to refer to the new decl. 426 DRE->setDecl(NewBuiltinDecl); 427 DRE->setType(NewBuiltinDecl->getType()); 428 429 // Set the callee in the CallExpr. 430 // FIXME: This leaks the original parens and implicit casts. 431 Expr *PromotedCall = DRE; 432 UsualUnaryConversions(PromotedCall); 433 TheCall->setCallee(PromotedCall); 434 435 436 // Change the result type of the call to match the result type of the decl. 437 TheCall->setType(NewBuiltinDecl->getResultType()); 438 return false; 439} 440 441 442/// CheckObjCString - Checks that the argument to the builtin 443/// CFString constructor is correct 444/// FIXME: GCC currently emits the following warning: 445/// "warning: input conversion stopped due to an input byte that does not 446/// belong to the input codeset UTF-8" 447/// Note: It might also make sense to do the UTF-16 conversion here (would 448/// simplify the backend). 449bool Sema::CheckObjCString(Expr *Arg) { 450 Arg = Arg->IgnoreParenCasts(); 451 StringLiteral *Literal = dyn_cast<StringLiteral>(Arg); 452 453 if (!Literal || Literal->isWide()) { 454 Diag(Arg->getLocStart(), diag::err_cfstring_literal_not_string_constant) 455 << Arg->getSourceRange(); 456 return true; 457 } 458 459 const char *Data = Literal->getStrData(); 460 unsigned Length = Literal->getByteLength(); 461 462 for (unsigned i = 0; i < Length; ++i) { 463 if (!Data[i]) { 464 Diag(getLocationOfStringLiteralByte(Literal, i), 465 diag::warn_cfstring_literal_contains_nul_character) 466 << Arg->getSourceRange(); 467 break; 468 } 469 } 470 471 return false; 472} 473 474/// SemaBuiltinVAStart - Check the arguments to __builtin_va_start for validity. 475/// Emit an error and return true on failure, return false on success. 476bool Sema::SemaBuiltinVAStart(CallExpr *TheCall) { 477 Expr *Fn = TheCall->getCallee(); 478 if (TheCall->getNumArgs() > 2) { 479 Diag(TheCall->getArg(2)->getLocStart(), 480 diag::err_typecheck_call_too_many_args) 481 << 0 /*function call*/ << Fn->getSourceRange() 482 << SourceRange(TheCall->getArg(2)->getLocStart(), 483 (*(TheCall->arg_end()-1))->getLocEnd()); 484 return true; 485 } 486 487 if (TheCall->getNumArgs() < 2) { 488 return Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_few_args) 489 << 0 /*function call*/; 490 } 491 492 // Determine whether the current function is variadic or not. 493 bool isVariadic; 494 if (CurBlock) 495 isVariadic = CurBlock->isVariadic; 496 else if (getCurFunctionDecl()) { 497 if (FunctionProtoType* FTP = 498 dyn_cast<FunctionProtoType>(getCurFunctionDecl()->getType())) 499 isVariadic = FTP->isVariadic(); 500 else 501 isVariadic = false; 502 } else { 503 isVariadic = getCurMethodDecl()->isVariadic(); 504 } 505 506 if (!isVariadic) { 507 Diag(Fn->getLocStart(), diag::err_va_start_used_in_non_variadic_function); 508 return true; 509 } 510 511 // Verify that the second argument to the builtin is the last argument of the 512 // current function or method. 513 bool SecondArgIsLastNamedArgument = false; 514 const Expr *Arg = TheCall->getArg(1)->IgnoreParenCasts(); 515 516 if (const DeclRefExpr *DR = dyn_cast<DeclRefExpr>(Arg)) { 517 if (const ParmVarDecl *PV = dyn_cast<ParmVarDecl>(DR->getDecl())) { 518 // FIXME: This isn't correct for methods (results in bogus warning). 519 // Get the last formal in the current function. 520 const ParmVarDecl *LastArg; 521 if (CurBlock) 522 LastArg = *(CurBlock->TheDecl->param_end()-1); 523 else if (FunctionDecl *FD = getCurFunctionDecl()) 524 LastArg = *(FD->param_end()-1); 525 else 526 LastArg = *(getCurMethodDecl()->param_end()-1); 527 SecondArgIsLastNamedArgument = PV == LastArg; 528 } 529 } 530 531 if (!SecondArgIsLastNamedArgument) 532 Diag(TheCall->getArg(1)->getLocStart(), 533 diag::warn_second_parameter_of_va_start_not_last_named_argument); 534 return false; 535} 536 537/// SemaBuiltinUnorderedCompare - Handle functions like __builtin_isgreater and 538/// friends. This is declared to take (...), so we have to check everything. 539bool Sema::SemaBuiltinUnorderedCompare(CallExpr *TheCall) { 540 if (TheCall->getNumArgs() < 2) 541 return Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_few_args) 542 << 0 /*function call*/; 543 if (TheCall->getNumArgs() > 2) 544 return Diag(TheCall->getArg(2)->getLocStart(), 545 diag::err_typecheck_call_too_many_args) 546 << 0 /*function call*/ 547 << SourceRange(TheCall->getArg(2)->getLocStart(), 548 (*(TheCall->arg_end()-1))->getLocEnd()); 549 550 Expr *OrigArg0 = TheCall->getArg(0); 551 Expr *OrigArg1 = TheCall->getArg(1); 552 553 // Do standard promotions between the two arguments, returning their common 554 // type. 555 QualType Res = UsualArithmeticConversions(OrigArg0, OrigArg1, false); 556 557 // Make sure any conversions are pushed back into the call; this is 558 // type safe since unordered compare builtins are declared as "_Bool 559 // foo(...)". 560 TheCall->setArg(0, OrigArg0); 561 TheCall->setArg(1, OrigArg1); 562 563 if (OrigArg0->isTypeDependent() || OrigArg1->isTypeDependent()) 564 return false; 565 566 // If the common type isn't a real floating type, then the arguments were 567 // invalid for this operation. 568 if (!Res->isRealFloatingType()) 569 return Diag(OrigArg0->getLocStart(), 570 diag::err_typecheck_call_invalid_ordered_compare) 571 << OrigArg0->getType() << OrigArg1->getType() 572 << SourceRange(OrigArg0->getLocStart(), OrigArg1->getLocEnd()); 573 574 return false; 575} 576 577/// SemaBuiltinUnorderedCompare - Handle functions like __builtin_isnan and 578/// friends. This is declared to take (...), so we have to check everything. 579bool Sema::SemaBuiltinUnaryFP(CallExpr *TheCall) { 580 if (TheCall->getNumArgs() < 1) 581 return Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_few_args) 582 << 0 /*function call*/; 583 if (TheCall->getNumArgs() > 1) 584 return Diag(TheCall->getArg(1)->getLocStart(), 585 diag::err_typecheck_call_too_many_args) 586 << 0 /*function call*/ 587 << SourceRange(TheCall->getArg(1)->getLocStart(), 588 (*(TheCall->arg_end()-1))->getLocEnd()); 589 590 Expr *OrigArg = TheCall->getArg(0); 591 592 if (OrigArg->isTypeDependent()) 593 return false; 594 595 // This operation requires a floating-point number 596 if (!OrigArg->getType()->isRealFloatingType()) 597 return Diag(OrigArg->getLocStart(), 598 diag::err_typecheck_call_invalid_unary_fp) 599 << OrigArg->getType() << OrigArg->getSourceRange(); 600 601 return false; 602} 603 604bool Sema::SemaBuiltinStackAddress(CallExpr *TheCall) { 605 // The signature for these builtins is exact; the only thing we need 606 // to check is that the argument is a constant. 607 SourceLocation Loc; 608 if (!TheCall->getArg(0)->isTypeDependent() && 609 !TheCall->getArg(0)->isValueDependent() && 610 !TheCall->getArg(0)->isIntegerConstantExpr(Context, &Loc)) 611 return Diag(Loc, diag::err_stack_const_level) << TheCall->getSourceRange(); 612 613 return false; 614} 615 616/// SemaBuiltinShuffleVector - Handle __builtin_shufflevector. 617// This is declared to take (...), so we have to check everything. 618Action::OwningExprResult Sema::SemaBuiltinShuffleVector(CallExpr *TheCall) { 619 if (TheCall->getNumArgs() < 3) 620 return ExprError(Diag(TheCall->getLocEnd(), 621 diag::err_typecheck_call_too_few_args) 622 << 0 /*function call*/ << TheCall->getSourceRange()); 623 624 unsigned numElements = std::numeric_limits<unsigned>::max(); 625 if (!TheCall->getArg(0)->isTypeDependent() && 626 !TheCall->getArg(1)->isTypeDependent()) { 627 QualType FAType = TheCall->getArg(0)->getType(); 628 QualType SAType = TheCall->getArg(1)->getType(); 629 630 if (!FAType->isVectorType() || !SAType->isVectorType()) { 631 Diag(TheCall->getLocStart(), diag::err_shufflevector_non_vector) 632 << SourceRange(TheCall->getArg(0)->getLocStart(), 633 TheCall->getArg(1)->getLocEnd()); 634 return ExprError(); 635 } 636 637 if (!Context.hasSameUnqualifiedType(FAType, SAType)) { 638 Diag(TheCall->getLocStart(), diag::err_shufflevector_incompatible_vector) 639 << SourceRange(TheCall->getArg(0)->getLocStart(), 640 TheCall->getArg(1)->getLocEnd()); 641 return ExprError(); 642 } 643 644 numElements = FAType->getAs<VectorType>()->getNumElements(); 645 if (TheCall->getNumArgs() != numElements+2) { 646 if (TheCall->getNumArgs() < numElements+2) 647 return ExprError(Diag(TheCall->getLocEnd(), 648 diag::err_typecheck_call_too_few_args) 649 << 0 /*function call*/ << TheCall->getSourceRange()); 650 return ExprError(Diag(TheCall->getLocEnd(), 651 diag::err_typecheck_call_too_many_args) 652 << 0 /*function call*/ << TheCall->getSourceRange()); 653 } 654 } 655 656 for (unsigned i = 2; i < TheCall->getNumArgs(); i++) { 657 if (TheCall->getArg(i)->isTypeDependent() || 658 TheCall->getArg(i)->isValueDependent()) 659 continue; 660 661 llvm::APSInt Result(32); 662 if (!TheCall->getArg(i)->isIntegerConstantExpr(Result, Context)) 663 return ExprError(Diag(TheCall->getLocStart(), 664 diag::err_shufflevector_nonconstant_argument) 665 << TheCall->getArg(i)->getSourceRange()); 666 667 if (Result.getActiveBits() > 64 || Result.getZExtValue() >= numElements*2) 668 return ExprError(Diag(TheCall->getLocStart(), 669 diag::err_shufflevector_argument_too_large) 670 << TheCall->getArg(i)->getSourceRange()); 671 } 672 673 llvm::SmallVector<Expr*, 32> exprs; 674 675 for (unsigned i = 0, e = TheCall->getNumArgs(); i != e; i++) { 676 exprs.push_back(TheCall->getArg(i)); 677 TheCall->setArg(i, 0); 678 } 679 680 return Owned(new (Context) ShuffleVectorExpr(Context, exprs.begin(), 681 exprs.size(), exprs[0]->getType(), 682 TheCall->getCallee()->getLocStart(), 683 TheCall->getRParenLoc())); 684} 685 686/// SemaBuiltinPrefetch - Handle __builtin_prefetch. 687// This is declared to take (const void*, ...) and can take two 688// optional constant int args. 689bool Sema::SemaBuiltinPrefetch(CallExpr *TheCall) { 690 unsigned NumArgs = TheCall->getNumArgs(); 691 692 if (NumArgs > 3) 693 return Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_many_args) 694 << 0 /*function call*/ << TheCall->getSourceRange(); 695 696 // Argument 0 is checked for us and the remaining arguments must be 697 // constant integers. 698 for (unsigned i = 1; i != NumArgs; ++i) { 699 Expr *Arg = TheCall->getArg(i); 700 if (Arg->isTypeDependent()) 701 continue; 702 703 if (!Arg->getType()->isIntegralType()) 704 return Diag(TheCall->getLocStart(), diag::err_prefetch_invalid_arg_type) 705 << Arg->getSourceRange(); 706 707 ImpCastExprToType(Arg, Context.IntTy, CastExpr::CK_IntegralCast); 708 TheCall->setArg(i, Arg); 709 710 if (Arg->isValueDependent()) 711 continue; 712 713 llvm::APSInt Result; 714 if (!Arg->isIntegerConstantExpr(Result, Context)) 715 return Diag(TheCall->getLocStart(), diag::err_prefetch_invalid_arg_ice) 716 << SourceRange(Arg->getLocStart(), Arg->getLocEnd()); 717 718 // FIXME: gcc issues a warning and rewrites these to 0. These 719 // seems especially odd for the third argument since the default 720 // is 3. 721 if (i == 1) { 722 if (Result.getLimitedValue() > 1) 723 return Diag(TheCall->getLocStart(), diag::err_argument_invalid_range) 724 << "0" << "1" << Arg->getSourceRange(); 725 } else { 726 if (Result.getLimitedValue() > 3) 727 return Diag(TheCall->getLocStart(), diag::err_argument_invalid_range) 728 << "0" << "3" << Arg->getSourceRange(); 729 } 730 } 731 732 return false; 733} 734 735/// SemaBuiltinEHReturnDataRegNo - Handle __builtin_eh_return_data_regno, the 736/// operand must be an integer constant. 737bool Sema::SemaBuiltinEHReturnDataRegNo(CallExpr *TheCall) { 738 llvm::APSInt Result; 739 if (!TheCall->getArg(0)->isIntegerConstantExpr(Result, Context)) 740 return Diag(TheCall->getLocStart(), diag::err_expr_not_ice) 741 << TheCall->getArg(0)->getSourceRange(); 742 743 return false; 744} 745 746 747/// SemaBuiltinObjectSize - Handle __builtin_object_size(void *ptr, 748/// int type). This simply type checks that type is one of the defined 749/// constants (0-3). 750// For compatability check 0-3, llvm only handles 0 and 2. 751bool Sema::SemaBuiltinObjectSize(CallExpr *TheCall) { 752 Expr *Arg = TheCall->getArg(1); 753 if (Arg->isTypeDependent()) 754 return false; 755 756 QualType ArgType = Arg->getType(); 757 const BuiltinType *BT = ArgType->getAs<BuiltinType>(); 758 llvm::APSInt Result(32); 759 if (!BT || BT->getKind() != BuiltinType::Int) 760 return Diag(TheCall->getLocStart(), diag::err_object_size_invalid_argument) 761 << SourceRange(Arg->getLocStart(), Arg->getLocEnd()); 762 763 if (Arg->isValueDependent()) 764 return false; 765 766 if (!Arg->isIntegerConstantExpr(Result, Context)) { 767 return Diag(TheCall->getLocStart(), diag::err_object_size_invalid_argument) 768 << SourceRange(Arg->getLocStart(), Arg->getLocEnd()); 769 } 770 771 if (Result.getSExtValue() < 0 || Result.getSExtValue() > 3) { 772 return Diag(TheCall->getLocStart(), diag::err_argument_invalid_range) 773 << "0" << "3" << SourceRange(Arg->getLocStart(), Arg->getLocEnd()); 774 } 775 776 return false; 777} 778 779/// SemaBuiltinLongjmp - Handle __builtin_longjmp(void *env[5], int val). 780/// This checks that val is a constant 1. 781bool Sema::SemaBuiltinLongjmp(CallExpr *TheCall) { 782 Expr *Arg = TheCall->getArg(1); 783 if (Arg->isTypeDependent() || Arg->isValueDependent()) 784 return false; 785 786 llvm::APSInt Result(32); 787 if (!Arg->isIntegerConstantExpr(Result, Context) || Result != 1) 788 return Diag(TheCall->getLocStart(), diag::err_builtin_longjmp_invalid_val) 789 << SourceRange(Arg->getLocStart(), Arg->getLocEnd()); 790 791 return false; 792} 793 794// Handle i > 1 ? "x" : "y", recursivelly 795bool Sema::SemaCheckStringLiteral(const Expr *E, const CallExpr *TheCall, 796 bool HasVAListArg, 797 unsigned format_idx, unsigned firstDataArg) { 798 if (E->isTypeDependent() || E->isValueDependent()) 799 return false; 800 801 switch (E->getStmtClass()) { 802 case Stmt::ConditionalOperatorClass: { 803 const ConditionalOperator *C = cast<ConditionalOperator>(E); 804 return SemaCheckStringLiteral(C->getTrueExpr(), TheCall, 805 HasVAListArg, format_idx, firstDataArg) 806 && SemaCheckStringLiteral(C->getRHS(), TheCall, 807 HasVAListArg, format_idx, firstDataArg); 808 } 809 810 case Stmt::ImplicitCastExprClass: { 811 const ImplicitCastExpr *Expr = cast<ImplicitCastExpr>(E); 812 return SemaCheckStringLiteral(Expr->getSubExpr(), TheCall, HasVAListArg, 813 format_idx, firstDataArg); 814 } 815 816 case Stmt::ParenExprClass: { 817 const ParenExpr *Expr = cast<ParenExpr>(E); 818 return SemaCheckStringLiteral(Expr->getSubExpr(), TheCall, HasVAListArg, 819 format_idx, firstDataArg); 820 } 821 822 case Stmt::DeclRefExprClass: { 823 const DeclRefExpr *DR = cast<DeclRefExpr>(E); 824 825 // As an exception, do not flag errors for variables binding to 826 // const string literals. 827 if (const VarDecl *VD = dyn_cast<VarDecl>(DR->getDecl())) { 828 bool isConstant = false; 829 QualType T = DR->getType(); 830 831 if (const ArrayType *AT = Context.getAsArrayType(T)) { 832 isConstant = AT->getElementType().isConstant(Context); 833 } else if (const PointerType *PT = T->getAs<PointerType>()) { 834 isConstant = T.isConstant(Context) && 835 PT->getPointeeType().isConstant(Context); 836 } 837 838 if (isConstant) { 839 const VarDecl *Def = 0; 840 if (const Expr *Init = VD->getDefinition(Def)) 841 return SemaCheckStringLiteral(Init, TheCall, 842 HasVAListArg, format_idx, firstDataArg); 843 } 844 845 // For vprintf* functions (i.e., HasVAListArg==true), we add a 846 // special check to see if the format string is a function parameter 847 // of the function calling the printf function. If the function 848 // has an attribute indicating it is a printf-like function, then we 849 // should suppress warnings concerning non-literals being used in a call 850 // to a vprintf function. For example: 851 // 852 // void 853 // logmessage(char const *fmt __attribute__ (format (printf, 1, 2)), ...){ 854 // va_list ap; 855 // va_start(ap, fmt); 856 // vprintf(fmt, ap); // Do NOT emit a warning about "fmt". 857 // ... 858 // 859 // 860 // FIXME: We don't have full attribute support yet, so just check to see 861 // if the argument is a DeclRefExpr that references a parameter. We'll 862 // add proper support for checking the attribute later. 863 if (HasVAListArg) 864 if (isa<ParmVarDecl>(VD)) 865 return true; 866 } 867 868 return false; 869 } 870 871 case Stmt::CallExprClass: { 872 const CallExpr *CE = cast<CallExpr>(E); 873 if (const ImplicitCastExpr *ICE 874 = dyn_cast<ImplicitCastExpr>(CE->getCallee())) { 875 if (const DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(ICE->getSubExpr())) { 876 if (const FunctionDecl *FD = dyn_cast<FunctionDecl>(DRE->getDecl())) { 877 if (const FormatArgAttr *FA = FD->getAttr<FormatArgAttr>()) { 878 unsigned ArgIndex = FA->getFormatIdx(); 879 const Expr *Arg = CE->getArg(ArgIndex - 1); 880 881 return SemaCheckStringLiteral(Arg, TheCall, HasVAListArg, 882 format_idx, firstDataArg); 883 } 884 } 885 } 886 } 887 888 return false; 889 } 890 case Stmt::ObjCStringLiteralClass: 891 case Stmt::StringLiteralClass: { 892 const StringLiteral *StrE = NULL; 893 894 if (const ObjCStringLiteral *ObjCFExpr = dyn_cast<ObjCStringLiteral>(E)) 895 StrE = ObjCFExpr->getString(); 896 else 897 StrE = cast<StringLiteral>(E); 898 899 if (StrE) { 900 CheckPrintfString(StrE, E, TheCall, HasVAListArg, format_idx, 901 firstDataArg); 902 return true; 903 } 904 905 return false; 906 } 907 908 default: 909 return false; 910 } 911} 912 913void 914Sema::CheckNonNullArguments(const NonNullAttr *NonNull, 915 const CallExpr *TheCall) { 916 for (NonNullAttr::iterator i = NonNull->begin(), e = NonNull->end(); 917 i != e; ++i) { 918 const Expr *ArgExpr = TheCall->getArg(*i); 919 if (ArgExpr->isNullPointerConstant(Context, 920 Expr::NPC_ValueDependentIsNotNull)) 921 Diag(TheCall->getCallee()->getLocStart(), diag::warn_null_arg) 922 << ArgExpr->getSourceRange(); 923 } 924} 925 926/// CheckPrintfArguments - Check calls to printf (and similar functions) for 927/// correct use of format strings. 928/// 929/// HasVAListArg - A predicate indicating whether the printf-like 930/// function is passed an explicit va_arg argument (e.g., vprintf) 931/// 932/// format_idx - The index into Args for the format string. 933/// 934/// Improper format strings to functions in the printf family can be 935/// the source of bizarre bugs and very serious security holes. A 936/// good source of information is available in the following paper 937/// (which includes additional references): 938/// 939/// FormatGuard: Automatic Protection From printf Format String 940/// Vulnerabilities, Proceedings of the 10th USENIX Security Symposium, 2001. 941/// 942/// Functionality implemented: 943/// 944/// We can statically check the following properties for string 945/// literal format strings for non v.*printf functions (where the 946/// arguments are passed directly): 947// 948/// (1) Are the number of format conversions equal to the number of 949/// data arguments? 950/// 951/// (2) Does each format conversion correctly match the type of the 952/// corresponding data argument? (TODO) 953/// 954/// Moreover, for all printf functions we can: 955/// 956/// (3) Check for a missing format string (when not caught by type checking). 957/// 958/// (4) Check for no-operation flags; e.g. using "#" with format 959/// conversion 'c' (TODO) 960/// 961/// (5) Check the use of '%n', a major source of security holes. 962/// 963/// (6) Check for malformed format conversions that don't specify anything. 964/// 965/// (7) Check for empty format strings. e.g: printf(""); 966/// 967/// (8) Check that the format string is a wide literal. 968/// 969/// All of these checks can be done by parsing the format string. 970/// 971/// For now, we ONLY do (1), (3), (5), (6), (7), and (8). 972void 973Sema::CheckPrintfArguments(const CallExpr *TheCall, bool HasVAListArg, 974 unsigned format_idx, unsigned firstDataArg) { 975 const Expr *Fn = TheCall->getCallee(); 976 977 // The way the format attribute works in GCC, the implicit this argument 978 // of member functions is counted. However, it doesn't appear in our own 979 // lists, so decrement format_idx in that case. 980 if (isa<CXXMemberCallExpr>(TheCall)) { 981 // Catch a format attribute mistakenly referring to the object argument. 982 if (format_idx == 0) 983 return; 984 --format_idx; 985 if(firstDataArg != 0) 986 --firstDataArg; 987 } 988 989 // CHECK: printf-like function is called with no format string. 990 if (format_idx >= TheCall->getNumArgs()) { 991 Diag(TheCall->getRParenLoc(), diag::warn_printf_missing_format_string) 992 << Fn->getSourceRange(); 993 return; 994 } 995 996 const Expr *OrigFormatExpr = TheCall->getArg(format_idx)->IgnoreParenCasts(); 997 998 // CHECK: format string is not a string literal. 999 // 1000 // Dynamically generated format strings are difficult to 1001 // automatically vet at compile time. Requiring that format strings 1002 // are string literals: (1) permits the checking of format strings by 1003 // the compiler and thereby (2) can practically remove the source of 1004 // many format string exploits. 1005 1006 // Format string can be either ObjC string (e.g. @"%d") or 1007 // C string (e.g. "%d") 1008 // ObjC string uses the same format specifiers as C string, so we can use 1009 // the same format string checking logic for both ObjC and C strings. 1010 if (SemaCheckStringLiteral(OrigFormatExpr, TheCall, HasVAListArg, format_idx, 1011 firstDataArg)) 1012 return; // Literal format string found, check done! 1013 1014 // If there are no arguments specified, warn with -Wformat-security, otherwise 1015 // warn only with -Wformat-nonliteral. 1016 if (TheCall->getNumArgs() == format_idx+1) 1017 Diag(TheCall->getArg(format_idx)->getLocStart(), 1018 diag::warn_printf_nonliteral_noargs) 1019 << OrigFormatExpr->getSourceRange(); 1020 else 1021 Diag(TheCall->getArg(format_idx)->getLocStart(), 1022 diag::warn_printf_nonliteral) 1023 << OrigFormatExpr->getSourceRange(); 1024} 1025 1026void Sema::CheckPrintfString(const StringLiteral *FExpr, 1027 const Expr *OrigFormatExpr, 1028 const CallExpr *TheCall, bool HasVAListArg, 1029 unsigned format_idx, unsigned firstDataArg) { 1030 1031 const ObjCStringLiteral *ObjCFExpr = 1032 dyn_cast<ObjCStringLiteral>(OrigFormatExpr); 1033 1034 // CHECK: is the format string a wide literal? 1035 if (FExpr->isWide()) { 1036 Diag(FExpr->getLocStart(), 1037 diag::warn_printf_format_string_is_wide_literal) 1038 << OrigFormatExpr->getSourceRange(); 1039 return; 1040 } 1041 1042 // Str - The format string. NOTE: this is NOT null-terminated! 1043 const char *Str = FExpr->getStrData(); 1044 1045 // CHECK: empty format string? 1046 unsigned StrLen = FExpr->getByteLength(); 1047 1048 if (StrLen == 0) { 1049 Diag(FExpr->getLocStart(), diag::warn_printf_empty_format_string) 1050 << OrigFormatExpr->getSourceRange(); 1051 return; 1052 } 1053 1054 // We process the format string using a binary state machine. The 1055 // current state is stored in CurrentState. 1056 enum { 1057 state_OrdChr, 1058 state_Conversion 1059 } CurrentState = state_OrdChr; 1060 1061 // numConversions - The number of conversions seen so far. This is 1062 // incremented as we traverse the format string. 1063 unsigned numConversions = 0; 1064 1065 // numDataArgs - The number of data arguments after the format 1066 // string. This can only be determined for non vprintf-like 1067 // functions. For those functions, this value is 1 (the sole 1068 // va_arg argument). 1069 unsigned numDataArgs = TheCall->getNumArgs()-firstDataArg; 1070 1071 // Inspect the format string. 1072 unsigned StrIdx = 0; 1073 1074 // LastConversionIdx - Index within the format string where we last saw 1075 // a '%' character that starts a new format conversion. 1076 unsigned LastConversionIdx = 0; 1077 1078 for (; StrIdx < StrLen; ++StrIdx) { 1079 1080 // Is the number of detected conversion conversions greater than 1081 // the number of matching data arguments? If so, stop. 1082 if (!HasVAListArg && numConversions > numDataArgs) break; 1083 1084 // Handle "\0" 1085 if (Str[StrIdx] == '\0') { 1086 // The string returned by getStrData() is not null-terminated, 1087 // so the presence of a null character is likely an error. 1088 Diag(getLocationOfStringLiteralByte(FExpr, StrIdx), 1089 diag::warn_printf_format_string_contains_null_char) 1090 << OrigFormatExpr->getSourceRange(); 1091 return; 1092 } 1093 1094 // Ordinary characters (not processing a format conversion). 1095 if (CurrentState == state_OrdChr) { 1096 if (Str[StrIdx] == '%') { 1097 CurrentState = state_Conversion; 1098 LastConversionIdx = StrIdx; 1099 } 1100 continue; 1101 } 1102 1103 // Seen '%'. Now processing a format conversion. 1104 switch (Str[StrIdx]) { 1105 // Handle dynamic precision or width specifier. 1106 case '*': { 1107 ++numConversions; 1108 1109 if (!HasVAListArg) { 1110 if (numConversions > numDataArgs) { 1111 SourceLocation Loc = getLocationOfStringLiteralByte(FExpr, StrIdx); 1112 1113 if (Str[StrIdx-1] == '.') 1114 Diag(Loc, diag::warn_printf_asterisk_precision_missing_arg) 1115 << OrigFormatExpr->getSourceRange(); 1116 else 1117 Diag(Loc, diag::warn_printf_asterisk_width_missing_arg) 1118 << OrigFormatExpr->getSourceRange(); 1119 1120 // Don't do any more checking. We'll just emit spurious errors. 1121 return; 1122 } 1123 1124 // Perform type checking on width/precision specifier. 1125 const Expr *E = TheCall->getArg(format_idx+numConversions); 1126 if (const BuiltinType *BT = E->getType()->getAs<BuiltinType>()) 1127 if (BT->getKind() == BuiltinType::Int) 1128 break; 1129 1130 SourceLocation Loc = getLocationOfStringLiteralByte(FExpr, StrIdx); 1131 1132 if (Str[StrIdx-1] == '.') 1133 Diag(Loc, diag::warn_printf_asterisk_precision_wrong_type) 1134 << E->getType() << E->getSourceRange(); 1135 else 1136 Diag(Loc, diag::warn_printf_asterisk_width_wrong_type) 1137 << E->getType() << E->getSourceRange(); 1138 1139 break; 1140 } 1141 } 1142 1143 // Characters which can terminate a format conversion 1144 // (e.g. "%d"). Characters that specify length modifiers or 1145 // other flags are handled by the default case below. 1146 // 1147 // FIXME: additional checks will go into the following cases. 1148 case 'i': 1149 case 'd': 1150 case 'o': 1151 case 'u': 1152 case 'x': 1153 case 'X': 1154 case 'D': 1155 case 'O': 1156 case 'U': 1157 case 'e': 1158 case 'E': 1159 case 'f': 1160 case 'F': 1161 case 'g': 1162 case 'G': 1163 case 'a': 1164 case 'A': 1165 case 'c': 1166 case 'C': 1167 case 'S': 1168 case 's': 1169 case 'p': 1170 ++numConversions; 1171 CurrentState = state_OrdChr; 1172 break; 1173 1174 case 'm': 1175 // FIXME: Warn in situations where this isn't supported! 1176 CurrentState = state_OrdChr; 1177 break; 1178 1179 // CHECK: Are we using "%n"? Issue a warning. 1180 case 'n': { 1181 ++numConversions; 1182 CurrentState = state_OrdChr; 1183 SourceLocation Loc = getLocationOfStringLiteralByte(FExpr, 1184 LastConversionIdx); 1185 1186 Diag(Loc, diag::warn_printf_write_back)<<OrigFormatExpr->getSourceRange(); 1187 break; 1188 } 1189 1190 // Handle "%@" 1191 case '@': 1192 // %@ is allowed in ObjC format strings only. 1193 if (ObjCFExpr != NULL) 1194 CurrentState = state_OrdChr; 1195 else { 1196 // Issue a warning: invalid format conversion. 1197 SourceLocation Loc = 1198 getLocationOfStringLiteralByte(FExpr, LastConversionIdx); 1199 1200 Diag(Loc, diag::warn_printf_invalid_conversion) 1201 << std::string(Str+LastConversionIdx, 1202 Str+std::min(LastConversionIdx+2, StrLen)) 1203 << OrigFormatExpr->getSourceRange(); 1204 } 1205 ++numConversions; 1206 break; 1207 1208 // Handle "%%" 1209 case '%': 1210 // Sanity check: Was the first "%" character the previous one? 1211 // If not, we will assume that we have a malformed format 1212 // conversion, and that the current "%" character is the start 1213 // of a new conversion. 1214 if (StrIdx - LastConversionIdx == 1) 1215 CurrentState = state_OrdChr; 1216 else { 1217 // Issue a warning: invalid format conversion. 1218 SourceLocation Loc = 1219 getLocationOfStringLiteralByte(FExpr, LastConversionIdx); 1220 1221 Diag(Loc, diag::warn_printf_invalid_conversion) 1222 << std::string(Str+LastConversionIdx, Str+StrIdx) 1223 << OrigFormatExpr->getSourceRange(); 1224 1225 // This conversion is broken. Advance to the next format 1226 // conversion. 1227 LastConversionIdx = StrIdx; 1228 ++numConversions; 1229 } 1230 break; 1231 1232 default: 1233 // This case catches all other characters: flags, widths, etc. 1234 // We should eventually process those as well. 1235 break; 1236 } 1237 } 1238 1239 if (CurrentState == state_Conversion) { 1240 // Issue a warning: invalid format conversion. 1241 SourceLocation Loc = 1242 getLocationOfStringLiteralByte(FExpr, LastConversionIdx); 1243 1244 Diag(Loc, diag::warn_printf_invalid_conversion) 1245 << std::string(Str+LastConversionIdx, 1246 Str+std::min(LastConversionIdx+2, StrLen)) 1247 << OrigFormatExpr->getSourceRange(); 1248 return; 1249 } 1250 1251 if (!HasVAListArg) { 1252 // CHECK: Does the number of format conversions exceed the number 1253 // of data arguments? 1254 if (numConversions > numDataArgs) { 1255 SourceLocation Loc = 1256 getLocationOfStringLiteralByte(FExpr, LastConversionIdx); 1257 1258 Diag(Loc, diag::warn_printf_insufficient_data_args) 1259 << OrigFormatExpr->getSourceRange(); 1260 } 1261 // CHECK: Does the number of data arguments exceed the number of 1262 // format conversions in the format string? 1263 else if (numConversions < numDataArgs) 1264 Diag(TheCall->getArg(format_idx+numConversions+1)->getLocStart(), 1265 diag::warn_printf_too_many_data_args) 1266 << OrigFormatExpr->getSourceRange(); 1267 } 1268} 1269 1270//===--- CHECK: Return Address of Stack Variable --------------------------===// 1271 1272static DeclRefExpr* EvalVal(Expr *E); 1273static DeclRefExpr* EvalAddr(Expr* E); 1274 1275/// CheckReturnStackAddr - Check if a return statement returns the address 1276/// of a stack variable. 1277void 1278Sema::CheckReturnStackAddr(Expr *RetValExp, QualType lhsType, 1279 SourceLocation ReturnLoc) { 1280 1281 // Perform checking for returned stack addresses. 1282 if (lhsType->isPointerType() || lhsType->isBlockPointerType()) { 1283 if (DeclRefExpr *DR = EvalAddr(RetValExp)) 1284 Diag(DR->getLocStart(), diag::warn_ret_stack_addr) 1285 << DR->getDecl()->getDeclName() << RetValExp->getSourceRange(); 1286 1287 // Skip over implicit cast expressions when checking for block expressions. 1288 RetValExp = RetValExp->IgnoreParenCasts(); 1289 1290 if (BlockExpr *C = dyn_cast<BlockExpr>(RetValExp)) 1291 if (C->hasBlockDeclRefExprs()) 1292 Diag(C->getLocStart(), diag::err_ret_local_block) 1293 << C->getSourceRange(); 1294 1295 if (AddrLabelExpr *ALE = dyn_cast<AddrLabelExpr>(RetValExp)) 1296 Diag(ALE->getLocStart(), diag::warn_ret_addr_label) 1297 << ALE->getSourceRange(); 1298 1299 } else if (lhsType->isReferenceType()) { 1300 // Perform checking for stack values returned by reference. 1301 // Check for a reference to the stack 1302 if (DeclRefExpr *DR = EvalVal(RetValExp)) 1303 Diag(DR->getLocStart(), diag::warn_ret_stack_ref) 1304 << DR->getDecl()->getDeclName() << RetValExp->getSourceRange(); 1305 } 1306} 1307 1308/// EvalAddr - EvalAddr and EvalVal are mutually recursive functions that 1309/// check if the expression in a return statement evaluates to an address 1310/// to a location on the stack. The recursion is used to traverse the 1311/// AST of the return expression, with recursion backtracking when we 1312/// encounter a subexpression that (1) clearly does not lead to the address 1313/// of a stack variable or (2) is something we cannot determine leads to 1314/// the address of a stack variable based on such local checking. 1315/// 1316/// EvalAddr processes expressions that are pointers that are used as 1317/// references (and not L-values). EvalVal handles all other values. 1318/// At the base case of the recursion is a check for a DeclRefExpr* in 1319/// the refers to a stack variable. 1320/// 1321/// This implementation handles: 1322/// 1323/// * pointer-to-pointer casts 1324/// * implicit conversions from array references to pointers 1325/// * taking the address of fields 1326/// * arbitrary interplay between "&" and "*" operators 1327/// * pointer arithmetic from an address of a stack variable 1328/// * taking the address of an array element where the array is on the stack 1329static DeclRefExpr* EvalAddr(Expr *E) { 1330 // We should only be called for evaluating pointer expressions. 1331 assert((E->getType()->isAnyPointerType() || 1332 E->getType()->isBlockPointerType() || 1333 E->getType()->isObjCQualifiedIdType()) && 1334 "EvalAddr only works on pointers"); 1335 1336 // Our "symbolic interpreter" is just a dispatch off the currently 1337 // viewed AST node. We then recursively traverse the AST by calling 1338 // EvalAddr and EvalVal appropriately. 1339 switch (E->getStmtClass()) { 1340 case Stmt::ParenExprClass: 1341 // Ignore parentheses. 1342 return EvalAddr(cast<ParenExpr>(E)->getSubExpr()); 1343 1344 case Stmt::UnaryOperatorClass: { 1345 // The only unary operator that make sense to handle here 1346 // is AddrOf. All others don't make sense as pointers. 1347 UnaryOperator *U = cast<UnaryOperator>(E); 1348 1349 if (U->getOpcode() == UnaryOperator::AddrOf) 1350 return EvalVal(U->getSubExpr()); 1351 else 1352 return NULL; 1353 } 1354 1355 case Stmt::BinaryOperatorClass: { 1356 // Handle pointer arithmetic. All other binary operators are not valid 1357 // in this context. 1358 BinaryOperator *B = cast<BinaryOperator>(E); 1359 BinaryOperator::Opcode op = B->getOpcode(); 1360 1361 if (op != BinaryOperator::Add && op != BinaryOperator::Sub) 1362 return NULL; 1363 1364 Expr *Base = B->getLHS(); 1365 1366 // Determine which argument is the real pointer base. It could be 1367 // the RHS argument instead of the LHS. 1368 if (!Base->getType()->isPointerType()) Base = B->getRHS(); 1369 1370 assert (Base->getType()->isPointerType()); 1371 return EvalAddr(Base); 1372 } 1373 1374 // For conditional operators we need to see if either the LHS or RHS are 1375 // valid DeclRefExpr*s. If one of them is valid, we return it. 1376 case Stmt::ConditionalOperatorClass: { 1377 ConditionalOperator *C = cast<ConditionalOperator>(E); 1378 1379 // Handle the GNU extension for missing LHS. 1380 if (Expr *lhsExpr = C->getLHS()) 1381 if (DeclRefExpr* LHS = EvalAddr(lhsExpr)) 1382 return LHS; 1383 1384 return EvalAddr(C->getRHS()); 1385 } 1386 1387 // For casts, we need to handle conversions from arrays to 1388 // pointer values, and pointer-to-pointer conversions. 1389 case Stmt::ImplicitCastExprClass: 1390 case Stmt::CStyleCastExprClass: 1391 case Stmt::CXXFunctionalCastExprClass: { 1392 Expr* SubExpr = cast<CastExpr>(E)->getSubExpr(); 1393 QualType T = SubExpr->getType(); 1394 1395 if (SubExpr->getType()->isPointerType() || 1396 SubExpr->getType()->isBlockPointerType() || 1397 SubExpr->getType()->isObjCQualifiedIdType()) 1398 return EvalAddr(SubExpr); 1399 else if (T->isArrayType()) 1400 return EvalVal(SubExpr); 1401 else 1402 return 0; 1403 } 1404 1405 // C++ casts. For dynamic casts, static casts, and const casts, we 1406 // are always converting from a pointer-to-pointer, so we just blow 1407 // through the cast. In the case the dynamic cast doesn't fail (and 1408 // return NULL), we take the conservative route and report cases 1409 // where we return the address of a stack variable. For Reinterpre 1410 // FIXME: The comment about is wrong; we're not always converting 1411 // from pointer to pointer. I'm guessing that this code should also 1412 // handle references to objects. 1413 case Stmt::CXXStaticCastExprClass: 1414 case Stmt::CXXDynamicCastExprClass: 1415 case Stmt::CXXConstCastExprClass: 1416 case Stmt::CXXReinterpretCastExprClass: { 1417 Expr *S = cast<CXXNamedCastExpr>(E)->getSubExpr(); 1418 if (S->getType()->isPointerType() || S->getType()->isBlockPointerType()) 1419 return EvalAddr(S); 1420 else 1421 return NULL; 1422 } 1423 1424 // Everything else: we simply don't reason about them. 1425 default: 1426 return NULL; 1427 } 1428} 1429 1430 1431/// EvalVal - This function is complements EvalAddr in the mutual recursion. 1432/// See the comments for EvalAddr for more details. 1433static DeclRefExpr* EvalVal(Expr *E) { 1434 1435 // We should only be called for evaluating non-pointer expressions, or 1436 // expressions with a pointer type that are not used as references but instead 1437 // are l-values (e.g., DeclRefExpr with a pointer type). 1438 1439 // Our "symbolic interpreter" is just a dispatch off the currently 1440 // viewed AST node. We then recursively traverse the AST by calling 1441 // EvalAddr and EvalVal appropriately. 1442 switch (E->getStmtClass()) { 1443 case Stmt::DeclRefExprClass: { 1444 // DeclRefExpr: the base case. When we hit a DeclRefExpr we are looking 1445 // at code that refers to a variable's name. We check if it has local 1446 // storage within the function, and if so, return the expression. 1447 DeclRefExpr *DR = cast<DeclRefExpr>(E); 1448 1449 if (VarDecl *V = dyn_cast<VarDecl>(DR->getDecl())) 1450 if (V->hasLocalStorage() && !V->getType()->isReferenceType()) return DR; 1451 1452 return NULL; 1453 } 1454 1455 case Stmt::ParenExprClass: 1456 // Ignore parentheses. 1457 return EvalVal(cast<ParenExpr>(E)->getSubExpr()); 1458 1459 case Stmt::UnaryOperatorClass: { 1460 // The only unary operator that make sense to handle here 1461 // is Deref. All others don't resolve to a "name." This includes 1462 // handling all sorts of rvalues passed to a unary operator. 1463 UnaryOperator *U = cast<UnaryOperator>(E); 1464 1465 if (U->getOpcode() == UnaryOperator::Deref) 1466 return EvalAddr(U->getSubExpr()); 1467 1468 return NULL; 1469 } 1470 1471 case Stmt::ArraySubscriptExprClass: { 1472 // Array subscripts are potential references to data on the stack. We 1473 // retrieve the DeclRefExpr* for the array variable if it indeed 1474 // has local storage. 1475 return EvalAddr(cast<ArraySubscriptExpr>(E)->getBase()); 1476 } 1477 1478 case Stmt::ConditionalOperatorClass: { 1479 // For conditional operators we need to see if either the LHS or RHS are 1480 // non-NULL DeclRefExpr's. If one is non-NULL, we return it. 1481 ConditionalOperator *C = cast<ConditionalOperator>(E); 1482 1483 // Handle the GNU extension for missing LHS. 1484 if (Expr *lhsExpr = C->getLHS()) 1485 if (DeclRefExpr *LHS = EvalVal(lhsExpr)) 1486 return LHS; 1487 1488 return EvalVal(C->getRHS()); 1489 } 1490 1491 // Accesses to members are potential references to data on the stack. 1492 case Stmt::MemberExprClass: { 1493 MemberExpr *M = cast<MemberExpr>(E); 1494 1495 // Check for indirect access. We only want direct field accesses. 1496 if (!M->isArrow()) 1497 return EvalVal(M->getBase()); 1498 else 1499 return NULL; 1500 } 1501 1502 // Everything else: we simply don't reason about them. 1503 default: 1504 return NULL; 1505 } 1506} 1507 1508//===--- CHECK: Floating-Point comparisons (-Wfloat-equal) ---------------===// 1509 1510/// Check for comparisons of floating point operands using != and ==. 1511/// Issue a warning if these are no self-comparisons, as they are not likely 1512/// to do what the programmer intended. 1513void Sema::CheckFloatComparison(SourceLocation loc, Expr* lex, Expr *rex) { 1514 bool EmitWarning = true; 1515 1516 Expr* LeftExprSansParen = lex->IgnoreParens(); 1517 Expr* RightExprSansParen = rex->IgnoreParens(); 1518 1519 // Special case: check for x == x (which is OK). 1520 // Do not emit warnings for such cases. 1521 if (DeclRefExpr* DRL = dyn_cast<DeclRefExpr>(LeftExprSansParen)) 1522 if (DeclRefExpr* DRR = dyn_cast<DeclRefExpr>(RightExprSansParen)) 1523 if (DRL->getDecl() == DRR->getDecl()) 1524 EmitWarning = false; 1525 1526 1527 // Special case: check for comparisons against literals that can be exactly 1528 // represented by APFloat. In such cases, do not emit a warning. This 1529 // is a heuristic: often comparison against such literals are used to 1530 // detect if a value in a variable has not changed. This clearly can 1531 // lead to false negatives. 1532 if (EmitWarning) { 1533 if (FloatingLiteral* FLL = dyn_cast<FloatingLiteral>(LeftExprSansParen)) { 1534 if (FLL->isExact()) 1535 EmitWarning = false; 1536 } else 1537 if (FloatingLiteral* FLR = dyn_cast<FloatingLiteral>(RightExprSansParen)){ 1538 if (FLR->isExact()) 1539 EmitWarning = false; 1540 } 1541 } 1542 1543 // Check for comparisons with builtin types. 1544 if (EmitWarning) 1545 if (CallExpr* CL = dyn_cast<CallExpr>(LeftExprSansParen)) 1546 if (CL->isBuiltinCall(Context)) 1547 EmitWarning = false; 1548 1549 if (EmitWarning) 1550 if (CallExpr* CR = dyn_cast<CallExpr>(RightExprSansParen)) 1551 if (CR->isBuiltinCall(Context)) 1552 EmitWarning = false; 1553 1554 // Emit the diagnostic. 1555 if (EmitWarning) 1556 Diag(loc, diag::warn_floatingpoint_eq) 1557 << lex->getSourceRange() << rex->getSourceRange(); 1558} 1559 1560//===--- CHECK: Integer mixed-sign comparisons (-Wsign-compare) --------===// 1561//===--- CHECK: Lossy implicit conversions (-Wconversion) --------------===// 1562 1563namespace { 1564 1565/// Structure recording the 'active' range of an integer-valued 1566/// expression. 1567struct IntRange { 1568 /// The number of bits active in the int. 1569 unsigned Width; 1570 1571 /// True if the int is known not to have negative values. 1572 bool NonNegative; 1573 1574 IntRange() {} 1575 IntRange(unsigned Width, bool NonNegative) 1576 : Width(Width), NonNegative(NonNegative) 1577 {} 1578 1579 // Returns the range of the bool type. 1580 static IntRange forBoolType() { 1581 return IntRange(1, true); 1582 } 1583 1584 // Returns the range of an integral type. 1585 static IntRange forType(ASTContext &C, QualType T) { 1586 return forCanonicalType(C, T->getCanonicalTypeInternal().getTypePtr()); 1587 } 1588 1589 // Returns the range of an integeral type based on its canonical 1590 // representation. 1591 static IntRange forCanonicalType(ASTContext &C, const Type *T) { 1592 assert(T->isCanonicalUnqualified()); 1593 1594 if (const VectorType *VT = dyn_cast<VectorType>(T)) 1595 T = VT->getElementType().getTypePtr(); 1596 if (const ComplexType *CT = dyn_cast<ComplexType>(T)) 1597 T = CT->getElementType().getTypePtr(); 1598 if (const EnumType *ET = dyn_cast<EnumType>(T)) 1599 T = ET->getDecl()->getIntegerType().getTypePtr(); 1600 1601 const BuiltinType *BT = cast<BuiltinType>(T); 1602 assert(BT->isInteger()); 1603 1604 return IntRange(C.getIntWidth(QualType(T, 0)), BT->isUnsignedInteger()); 1605 } 1606 1607 // Returns the supremum of two ranges: i.e. their conservative merge. 1608 static IntRange join(const IntRange &L, const IntRange &R) { 1609 return IntRange(std::max(L.Width, R.Width), 1610 L.NonNegative && R.NonNegative); 1611 } 1612 1613 // Returns the infinum of two ranges: i.e. their aggressive merge. 1614 static IntRange meet(const IntRange &L, const IntRange &R) { 1615 return IntRange(std::min(L.Width, R.Width), 1616 L.NonNegative || R.NonNegative); 1617 } 1618}; 1619 1620IntRange GetValueRange(ASTContext &C, llvm::APSInt &value, unsigned MaxWidth) { 1621 if (value.isSigned() && value.isNegative()) 1622 return IntRange(value.getMinSignedBits(), false); 1623 1624 if (value.getBitWidth() > MaxWidth) 1625 value.trunc(MaxWidth); 1626 1627 // isNonNegative() just checks the sign bit without considering 1628 // signedness. 1629 return IntRange(value.getActiveBits(), true); 1630} 1631 1632IntRange GetValueRange(ASTContext &C, APValue &result, QualType Ty, 1633 unsigned MaxWidth) { 1634 if (result.isInt()) 1635 return GetValueRange(C, result.getInt(), MaxWidth); 1636 1637 if (result.isVector()) { 1638 IntRange R = GetValueRange(C, result.getVectorElt(0), Ty, MaxWidth); 1639 for (unsigned i = 1, e = result.getVectorLength(); i != e; ++i) { 1640 IntRange El = GetValueRange(C, result.getVectorElt(i), Ty, MaxWidth); 1641 R = IntRange::join(R, El); 1642 } 1643 return R; 1644 } 1645 1646 if (result.isComplexInt()) { 1647 IntRange R = GetValueRange(C, result.getComplexIntReal(), MaxWidth); 1648 IntRange I = GetValueRange(C, result.getComplexIntImag(), MaxWidth); 1649 return IntRange::join(R, I); 1650 } 1651 1652 // This can happen with lossless casts to intptr_t of "based" lvalues. 1653 // Assume it might use arbitrary bits. 1654 // FIXME: The only reason we need to pass the type in here is to get 1655 // the sign right on this one case. It would be nice if APValue 1656 // preserved this. 1657 assert(result.isLValue()); 1658 return IntRange(MaxWidth, Ty->isUnsignedIntegerType()); 1659} 1660 1661/// Pseudo-evaluate the given integer expression, estimating the 1662/// range of values it might take. 1663/// 1664/// \param MaxWidth - the width to which the value will be truncated 1665IntRange GetExprRange(ASTContext &C, Expr *E, unsigned MaxWidth) { 1666 E = E->IgnoreParens(); 1667 1668 // Try a full evaluation first. 1669 Expr::EvalResult result; 1670 if (E->Evaluate(result, C)) 1671 return GetValueRange(C, result.Val, E->getType(), MaxWidth); 1672 1673 // I think we only want to look through implicit casts here; if the 1674 // user has an explicit widening cast, we should treat the value as 1675 // being of the new, wider type. 1676 if (ImplicitCastExpr *CE = dyn_cast<ImplicitCastExpr>(E)) { 1677 if (CE->getCastKind() == CastExpr::CK_NoOp) 1678 return GetExprRange(C, CE->getSubExpr(), MaxWidth); 1679 1680 IntRange OutputTypeRange = IntRange::forType(C, CE->getType()); 1681 1682 bool isIntegerCast = (CE->getCastKind() == CastExpr::CK_IntegralCast); 1683 if (!isIntegerCast && CE->getCastKind() == CastExpr::CK_Unknown) 1684 isIntegerCast = CE->getSubExpr()->getType()->isIntegerType(); 1685 1686 // Assume that non-integer casts can span the full range of the type. 1687 if (!isIntegerCast) 1688 return OutputTypeRange; 1689 1690 IntRange SubRange 1691 = GetExprRange(C, CE->getSubExpr(), 1692 std::min(MaxWidth, OutputTypeRange.Width)); 1693 1694 // Bail out if the subexpr's range is as wide as the cast type. 1695 if (SubRange.Width >= OutputTypeRange.Width) 1696 return OutputTypeRange; 1697 1698 // Otherwise, we take the smaller width, and we're non-negative if 1699 // either the output type or the subexpr is. 1700 return IntRange(SubRange.Width, 1701 SubRange.NonNegative || OutputTypeRange.NonNegative); 1702 } 1703 1704 if (ConditionalOperator *CO = dyn_cast<ConditionalOperator>(E)) { 1705 // If we can fold the condition, just take that operand. 1706 bool CondResult; 1707 if (CO->getCond()->EvaluateAsBooleanCondition(CondResult, C)) 1708 return GetExprRange(C, CondResult ? CO->getTrueExpr() 1709 : CO->getFalseExpr(), 1710 MaxWidth); 1711 1712 // Otherwise, conservatively merge. 1713 IntRange L = GetExprRange(C, CO->getTrueExpr(), MaxWidth); 1714 IntRange R = GetExprRange(C, CO->getFalseExpr(), MaxWidth); 1715 return IntRange::join(L, R); 1716 } 1717 1718 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(E)) { 1719 switch (BO->getOpcode()) { 1720 1721 // Boolean-valued operations are single-bit and positive. 1722 case BinaryOperator::LAnd: 1723 case BinaryOperator::LOr: 1724 case BinaryOperator::LT: 1725 case BinaryOperator::GT: 1726 case BinaryOperator::LE: 1727 case BinaryOperator::GE: 1728 case BinaryOperator::EQ: 1729 case BinaryOperator::NE: 1730 return IntRange::forBoolType(); 1731 1732 // Operations with opaque sources are black-listed. 1733 case BinaryOperator::PtrMemD: 1734 case BinaryOperator::PtrMemI: 1735 return IntRange::forType(C, E->getType()); 1736 1737 // Bitwise-and uses the *infinum* of the two source ranges. 1738 case BinaryOperator::And: 1739 return IntRange::meet(GetExprRange(C, BO->getLHS(), MaxWidth), 1740 GetExprRange(C, BO->getRHS(), MaxWidth)); 1741 1742 // Left shift gets black-listed based on a judgement call. 1743 case BinaryOperator::Shl: 1744 return IntRange::forType(C, E->getType()); 1745 1746 // Right shift by a constant can narrow its left argument. 1747 case BinaryOperator::Shr: { 1748 IntRange L = GetExprRange(C, BO->getLHS(), MaxWidth); 1749 1750 // If the shift amount is a positive constant, drop the width by 1751 // that much. 1752 llvm::APSInt shift; 1753 if (BO->getRHS()->isIntegerConstantExpr(shift, C) && 1754 shift.isNonNegative()) { 1755 unsigned zext = shift.getZExtValue(); 1756 if (zext >= L.Width) 1757 L.Width = (L.NonNegative ? 0 : 1); 1758 else 1759 L.Width -= zext; 1760 } 1761 1762 return L; 1763 } 1764 1765 // Comma acts as its right operand. 1766 case BinaryOperator::Comma: 1767 return GetExprRange(C, BO->getRHS(), MaxWidth); 1768 1769 // Black-list pointer subtractions. 1770 case BinaryOperator::Sub: 1771 if (BO->getLHS()->getType()->isPointerType()) 1772 return IntRange::forType(C, E->getType()); 1773 // fallthrough 1774 1775 default: 1776 break; 1777 } 1778 1779 // Treat every other operator as if it were closed on the 1780 // narrowest type that encompasses both operands. 1781 IntRange L = GetExprRange(C, BO->getLHS(), MaxWidth); 1782 IntRange R = GetExprRange(C, BO->getRHS(), MaxWidth); 1783 return IntRange::join(L, R); 1784 } 1785 1786 if (UnaryOperator *UO = dyn_cast<UnaryOperator>(E)) { 1787 switch (UO->getOpcode()) { 1788 // Boolean-valued operations are white-listed. 1789 case UnaryOperator::LNot: 1790 return IntRange::forBoolType(); 1791 1792 // Operations with opaque sources are black-listed. 1793 case UnaryOperator::Deref: 1794 case UnaryOperator::AddrOf: // should be impossible 1795 case UnaryOperator::OffsetOf: 1796 return IntRange::forType(C, E->getType()); 1797 1798 default: 1799 return GetExprRange(C, UO->getSubExpr(), MaxWidth); 1800 } 1801 } 1802 1803 FieldDecl *BitField = E->getBitField(); 1804 if (BitField) { 1805 llvm::APSInt BitWidthAP = BitField->getBitWidth()->EvaluateAsInt(C); 1806 unsigned BitWidth = BitWidthAP.getZExtValue(); 1807 1808 return IntRange(BitWidth, BitField->getType()->isUnsignedIntegerType()); 1809 } 1810 1811 return IntRange::forType(C, E->getType()); 1812} 1813 1814/// Checks whether the given value, which currently has the given 1815/// source semantics, has the same value when coerced through the 1816/// target semantics. 1817bool IsSameFloatAfterCast(const llvm::APFloat &value, 1818 const llvm::fltSemantics &Src, 1819 const llvm::fltSemantics &Tgt) { 1820 llvm::APFloat truncated = value; 1821 1822 bool ignored; 1823 truncated.convert(Src, llvm::APFloat::rmNearestTiesToEven, &ignored); 1824 truncated.convert(Tgt, llvm::APFloat::rmNearestTiesToEven, &ignored); 1825 1826 return truncated.bitwiseIsEqual(value); 1827} 1828 1829/// Checks whether the given value, which currently has the given 1830/// source semantics, has the same value when coerced through the 1831/// target semantics. 1832/// 1833/// The value might be a vector of floats (or a complex number). 1834bool IsSameFloatAfterCast(const APValue &value, 1835 const llvm::fltSemantics &Src, 1836 const llvm::fltSemantics &Tgt) { 1837 if (value.isFloat()) 1838 return IsSameFloatAfterCast(value.getFloat(), Src, Tgt); 1839 1840 if (value.isVector()) { 1841 for (unsigned i = 0, e = value.getVectorLength(); i != e; ++i) 1842 if (!IsSameFloatAfterCast(value.getVectorElt(i), Src, Tgt)) 1843 return false; 1844 return true; 1845 } 1846 1847 assert(value.isComplexFloat()); 1848 return (IsSameFloatAfterCast(value.getComplexFloatReal(), Src, Tgt) && 1849 IsSameFloatAfterCast(value.getComplexFloatImag(), Src, Tgt)); 1850} 1851 1852} // end anonymous namespace 1853 1854/// \brief Implements -Wsign-compare. 1855/// 1856/// \param lex the left-hand expression 1857/// \param rex the right-hand expression 1858/// \param OpLoc the location of the joining operator 1859/// \param Equality whether this is an "equality-like" join, which 1860/// suppresses the warning in some cases 1861void Sema::CheckSignCompare(Expr *lex, Expr *rex, SourceLocation OpLoc, 1862 const PartialDiagnostic &PD, bool Equality) { 1863 // Don't warn if we're in an unevaluated context. 1864 if (ExprEvalContexts.back().Context == Unevaluated) 1865 return; 1866 1867 // If either expression is value-dependent, don't warn. We'll get another 1868 // chance at instantiation time. 1869 if (lex->isValueDependent() || rex->isValueDependent()) 1870 return; 1871 1872 QualType lt = lex->getType(), rt = rex->getType(); 1873 1874 // Only warn if both operands are integral. 1875 if (!lt->isIntegerType() || !rt->isIntegerType()) 1876 return; 1877 1878 // In C, the width of a bitfield determines its type, and the 1879 // declared type only contributes the signedness. This duplicates 1880 // the work that will later be done by UsualUnaryConversions. 1881 // Eventually, this check will be reorganized in a way that avoids 1882 // this duplication. 1883 if (!getLangOptions().CPlusPlus) { 1884 QualType tmp; 1885 tmp = Context.isPromotableBitField(lex); 1886 if (!tmp.isNull()) lt = tmp; 1887 tmp = Context.isPromotableBitField(rex); 1888 if (!tmp.isNull()) rt = tmp; 1889 } 1890 1891 // The rule is that the signed operand becomes unsigned, so isolate the 1892 // signed operand. 1893 Expr *signedOperand = lex, *unsignedOperand = rex; 1894 QualType signedType = lt, unsignedType = rt; 1895 if (lt->isSignedIntegerType()) { 1896 if (rt->isSignedIntegerType()) return; 1897 } else { 1898 if (!rt->isSignedIntegerType()) return; 1899 std::swap(signedOperand, unsignedOperand); 1900 std::swap(signedType, unsignedType); 1901 } 1902 1903 unsigned unsignedWidth = Context.getIntWidth(unsignedType); 1904 unsigned signedWidth = Context.getIntWidth(signedType); 1905 1906 // If the unsigned type is strictly smaller than the signed type, 1907 // then (1) the result type will be signed and (2) the unsigned 1908 // value will fit fully within the signed type, and thus the result 1909 // of the comparison will be exact. 1910 if (signedWidth > unsignedWidth) 1911 return; 1912 1913 // Otherwise, calculate the effective ranges. 1914 IntRange signedRange = GetExprRange(Context, signedOperand, signedWidth); 1915 IntRange unsignedRange = GetExprRange(Context, unsignedOperand, unsignedWidth); 1916 1917 // We should never be unable to prove that the unsigned operand is 1918 // non-negative. 1919 assert(unsignedRange.NonNegative && "unsigned range includes negative?"); 1920 1921 // If the signed operand is non-negative, then the signed->unsigned 1922 // conversion won't change it. 1923 if (signedRange.NonNegative) 1924 return; 1925 1926 // For (in)equality comparisons, if the unsigned operand is a 1927 // constant which cannot collide with a overflowed signed operand, 1928 // then reinterpreting the signed operand as unsigned will not 1929 // change the result of the comparison. 1930 if (Equality && unsignedRange.Width < unsignedWidth) 1931 return; 1932 1933 Diag(OpLoc, PD) 1934 << lt << rt << lex->getSourceRange() << rex->getSourceRange(); 1935} 1936 1937/// Diagnose an implicit cast; purely a helper for CheckImplicitConversion. 1938static void DiagnoseImpCast(Sema &S, Expr *E, QualType T, unsigned diag) { 1939 S.Diag(E->getExprLoc(), diag) << E->getType() << T << E->getSourceRange(); 1940} 1941 1942/// Implements -Wconversion. 1943void Sema::CheckImplicitConversion(Expr *E, QualType T) { 1944 // Don't diagnose in unevaluated contexts. 1945 if (ExprEvalContexts.back().Context == Sema::Unevaluated) 1946 return; 1947 1948 // Don't diagnose for value-dependent expressions. 1949 if (E->isValueDependent()) 1950 return; 1951 1952 const Type *Source = Context.getCanonicalType(E->getType()).getTypePtr(); 1953 const Type *Target = Context.getCanonicalType(T).getTypePtr(); 1954 1955 // Never diagnose implicit casts to bool. 1956 if (Target->isSpecificBuiltinType(BuiltinType::Bool)) 1957 return; 1958 1959 // Strip vector types. 1960 if (isa<VectorType>(Source)) { 1961 if (!isa<VectorType>(Target)) 1962 return DiagnoseImpCast(*this, E, T, diag::warn_impcast_vector_scalar); 1963 1964 Source = cast<VectorType>(Source)->getElementType().getTypePtr(); 1965 Target = cast<VectorType>(Target)->getElementType().getTypePtr(); 1966 } 1967 1968 // Strip complex types. 1969 if (isa<ComplexType>(Source)) { 1970 if (!isa<ComplexType>(Target)) 1971 return DiagnoseImpCast(*this, E, T, diag::warn_impcast_complex_scalar); 1972 1973 Source = cast<ComplexType>(Source)->getElementType().getTypePtr(); 1974 Target = cast<ComplexType>(Target)->getElementType().getTypePtr(); 1975 } 1976 1977 const BuiltinType *SourceBT = dyn_cast<BuiltinType>(Source); 1978 const BuiltinType *TargetBT = dyn_cast<BuiltinType>(Target); 1979 1980 // If the source is floating point... 1981 if (SourceBT && SourceBT->isFloatingPoint()) { 1982 // ...and the target is floating point... 1983 if (TargetBT && TargetBT->isFloatingPoint()) { 1984 // ...then warn if we're dropping FP rank. 1985 1986 // Builtin FP kinds are ordered by increasing FP rank. 1987 if (SourceBT->getKind() > TargetBT->getKind()) { 1988 // Don't warn about float constants that are precisely 1989 // representable in the target type. 1990 Expr::EvalResult result; 1991 if (E->Evaluate(result, Context)) { 1992 // Value might be a float, a float vector, or a float complex. 1993 if (IsSameFloatAfterCast(result.Val, 1994 Context.getFloatTypeSemantics(QualType(TargetBT, 0)), 1995 Context.getFloatTypeSemantics(QualType(SourceBT, 0)))) 1996 return; 1997 } 1998 1999 DiagnoseImpCast(*this, E, T, diag::warn_impcast_float_precision); 2000 } 2001 return; 2002 } 2003 2004 // If the target is integral, always warn. 2005 if ((TargetBT && TargetBT->isInteger())) 2006 // TODO: don't warn for integer values? 2007 return DiagnoseImpCast(*this, E, T, diag::warn_impcast_float_integer); 2008 2009 return; 2010 } 2011 2012 if (!Source->isIntegerType() || !Target->isIntegerType()) 2013 return; 2014 2015 IntRange SourceRange = GetExprRange(Context, E, Context.getIntWidth(E->getType())); 2016 IntRange TargetRange = IntRange::forCanonicalType(Context, Target); 2017 2018 // FIXME: also signed<->unsigned? 2019 2020 if (SourceRange.Width > TargetRange.Width) { 2021 // People want to build with -Wshorten-64-to-32 and not -Wconversion 2022 // and by god we'll let them. 2023 if (SourceRange.Width == 64 && TargetRange.Width == 32) 2024 return DiagnoseImpCast(*this, E, T, diag::warn_impcast_integer_64_32); 2025 return DiagnoseImpCast(*this, E, T, diag::warn_impcast_integer_precision); 2026 } 2027 2028 return; 2029} 2030 2031